Branch HEAD

[fs/lustre-release.git] / lustre / kernel_patches / patches / qsnet-rhel-2.4.patch

Index: linux-2.4.21/arch/i386/kernel/i386_ksyms.c
===================================================================
--- linux-2.4.21.orig/arch/i386/kernel/i386_ksyms.c     2005-06-01 22:51:51.000000000 -0400
+++ linux-2.4.21/arch/i386/kernel/i386_ksyms.c  2005-06-01 23:12:54.521450960 -0400
@@ -220,3 +220,12 @@
 EXPORT_SYMBOL_GPL(__PAGE_KERNEL);
 extern unsigned long long __supported_pte_mask;
 EXPORT_SYMBOL_GPL(__supported_pte_mask);
+
+extern asmlinkage long sys_open(const char *, int, int);
+EXPORT_SYMBOL(sys_open);
+extern asmlinkage off_t sys_lseek(unsigned int, off_t, unsigned int);
+EXPORT_SYMBOL(sys_lseek);
+extern asmlinkage long sys_poll(struct pollfd *, unsigned int, long);
+EXPORT_SYMBOL(sys_poll);
+extern asmlinkage long sys_kill(int, int);
+EXPORT_SYMBOL(sys_kill);
Index: linux-2.4.21/arch/ia64/kernel/ia64_ksyms.c
===================================================================
--- linux-2.4.21.orig/arch/ia64/kernel/ia64_ksyms.c     2005-06-01 22:51:59.000000000 -0400
+++ linux-2.4.21/arch/ia64/kernel/ia64_ksyms.c  2005-06-01 23:14:43.773842072 -0400
@@ -207,3 +207,13 @@
 EXPORT_SYMBOL_GPL(show_state);
 EXPORT_SYMBOL_GPL(show_regs);
 EXPORT_SYMBOL(pm_power_off);
+
+#define __KERNEL_SYSCALLS__ 1
+#include <asm/unistd.h>
+EXPORT_SYMBOL(sys_open);
+extern asmlinkage off_t sys_lseek(unsigned int, off_t, unsigned int);
+EXPORT_SYMBOL(sys_lseek);
+extern asmlinkage long sys_poll(struct pollfd *, unsigned int, long);
+EXPORT_SYMBOL(sys_poll);
+extern asmlinkage long sys_kill(int, int);
+EXPORT_SYMBOL(sys_kill);
Index: linux-2.4.21/arch/x86_64/kernel/x8664_ksyms.c
===================================================================
--- linux-2.4.21.orig/arch/x86_64/kernel/x8664_ksyms.c  2005-06-01 22:51:51.000000000 -0400
+++ linux-2.4.21/arch/x86_64/kernel/x8664_ksyms.c       2005-06-01 23:12:54.522450808 -0400
@@ -215,6 +215,10 @@
 EXPORT_SYMBOL(sys_exit);
 EXPORT_SYMBOL(sys_open);
 EXPORT_SYMBOL(sys_lseek);
+extern asmlinkage long sys_poll(struct pollfd *, unsigned int, long);
+EXPORT_SYMBOL(sys_poll);
+extern asmlinkage long sys_kill(int, int);
+EXPORT_SYMBOL(sys_kill);
 EXPORT_SYMBOL(sys_delete_module);
 EXPORT_SYMBOL(sys_sync);
 EXPORT_SYMBOL(sys_pause);
Index: linux-2.4.21/Documentation/Configure.help
===================================================================
--- linux-2.4.21.orig/Documentation/Configure.help      2005-06-01 23:12:39.856680344 -0400
+++ linux-2.4.21/Documentation/Configure.help   2005-06-01 23:12:54.547447008 -0400
@@ -28030,6 +28030,54 @@
   kernel tree does. Such modules that use library CRC32 functions
   require M here.
 
+
+Enable support for Quadrics QsNet (QSNET)
+CONFIG_QSNET
+    Quadrics QsNet is a high bandwidth, ultra low latency cluster 
+    interconnect which provides both user and kernel programmers with
+    secure, direct access to the Quadrics network.
+         
+Elan 3 device driver (ELAN3)
+CONFIG_ELAN3
+   This is the main device driver for the Quadrics QsNet (Elan3) PCI
+   device. This is a high bandwidth, ultra low latency interconnect 
+   which provides both user and kernel programmers with secure, direct
+   access to the Quadrics network.
+         
+Elan 3 Kernel Comms (EP3)
+CONFIG_EP3
+   This modules implements the QsNet kernel communications layer. This
+   is used to layer kernel level facilities on top of the basic Elan3 
+   device driver. These can be used to implement subsystems such as
+   TCP/IP and remote filing systems over the QsNet interconnect.
+         
+Elan IP device (EIP)
+CONFIG_EIP
+   This is a network IP device driver for the Quadrics QsNet device.
+   It allows the TCP/IP protocol to be run over the Quadrics interconnect.
+
+Elan 4 device driver (ELAN4)
+CONFIG_ELAN4
+   This is the main device driver for the Quadrics QsNetII (Elan4) PCI-X 
+   device. This is a high bandwidth, ultra low latency interconnect which 
+   provides both user and kernel programmers with secure, direct access to 
+   the Quadrics network.
+Resource Management System support (RMS)
+CONFIG_RMS
+   This is a support module for the Quadrics RMS resource manager. It 
+   provides kernel services for monitoring and controlling user job 
+   execution, termination and cleanup.
+     
+Switch monitoring (JTAG)
+CONFIG_JTAG
+   The jtag interface is used to allow processes to send and retrieve jtag
+   information to a Quadrics QsNet Elite switch via the parallel port. 
+   The module requires a /dev/jtag[0-3] entry (usually there is only a 
+   /dev/jtag0) device and a particular device only allows one process at a
+   time to access this resource. 
+   For more information about JTag interface, please refer to the IEEE 
+   document on http://www.ieee.org 
+  
 #
 # A couple of things I keep forgetting:
 #   capitalize: AppleTalk, Ethernet, DOS, DMA, FAT, FTP, Internet,
Index: linux-2.4.21/drivers/net/Config.in
===================================================================
--- linux-2.4.21.orig/drivers/net/Config.in     2005-06-01 22:52:03.000000000 -0400
+++ linux-2.4.21/drivers/net/Config.in  2005-06-01 23:12:54.549446704 -0400
@@ -272,6 +272,9 @@
 
 endmenu
 
+# Quadrics QsNet
+source drivers/net/qsnet/Config.in
+
 if [ "$CONFIG_PPC_ISERIES" = "y" ]; then
    dep_tristate 'iSeries Virtual Ethernet driver support' CONFIG_VETH $CONFIG_PPC_ISERIES
 fi
Index: linux-2.4.21/drivers/net/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/Makefile      2005-06-01 22:52:03.000000000 -0400
+++ linux-2.4.21/drivers/net/Makefile   2005-06-01 23:12:54.550446552 -0400
@@ -8,7 +8,7 @@
 obj-n           :=
 obj-            :=
 
-mod-subdirs     := appletalk arcnet fc irda tokenring pcmcia wireless wireless_old wan
+mod-subdirs     := appletalk arcnet fc irda tokenring pcmcia wireless wireless_old wan qsnet
 
 O_TARGET := net.o
 
@@ -48,6 +48,7 @@
 subdir-$(CONFIG_DEV_APPLETALK) += appletalk
 subdir-$(CONFIG_SK98LIN) += sk98lin
 subdir-$(CONFIG_SKFP) += skfp
+subdir-$(CONFIG_QSNET) += qsnet
 subdir-$(CONFIG_E100) += e100
 subdir-$(CONFIG_E1000) += e1000
 subdir-$(CONFIG_BONDING) += bonding
Index: linux-2.4.21/drivers/net/qsnet/Config.in
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/Config.in       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/Config.in    2005-06-01 23:12:54.550446552 -0400
@@ -0,0 +1,25 @@
+#
+# Config.in for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/Config.in
+#
+
+mainmenu_option next_comment
+comment "Quadrics QsNet device support"
+
+dep_tristate "Enable support for Quadrics QsNet"        CONFIG_QSNET    $CONFIG_PCI
+
+dep_tristate "Elan 3 device driver"                    CONFIG_ELAN3    $CONFIG_QSNET
+dep_tristate "Elan 4 device driver"                    CONFIG_ELAN4    $CONFIG_QSNET
+
+if [ "$CONFIG_ELAN3" = "$CONFIG_QSNET" ] || [ "$CONFIG_ELAN4" = "$CONFIG_QSNET" ]; then
+       dep_tristate "Elan Kernel Comms"                        CONFIG_EP       $CONFIG_QSNET
+fi
+dep_tristate "Elan IP device"                             CONFIG_EIP      $CONFIG_NET  $CONFIG_EP
+
+dep_tristate "Resource Management System support"        CONFIG_RMS      $CONFIG_QSNET
+dep_tristate "Switch monitoring"                         CONFIG_JTAG     $CONFIG_QSNET
+
+endmenu
Index: linux-2.4.21/drivers/net/qsnet/eip/eip_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/eip/eip_linux.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/eip/eip_linux.c      2005-06-01 23:12:54.553446096 -0400
@@ -0,0 +1,1565 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: eip_linux.c,v 1.89.2.3 2004/12/20 16:54:05 mike Exp $"
+
+#include <qsnet/kernel.h>
+#include <qsnet/debug.h>
+
+#include <linux/module.h>
+
+#include <linux/init.h>
+#include <linux/list.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/skbuff.h>
+#include <linux/kernel.h>
+#include <linux/proc_fs.h>
+#include <linux/time.h>
+#include <linux/version.h>
+
+#include <asm/uaccess.h>
+#include <asm/unaligned.h>
+
+#undef ASSERT
+#include <net/sock.h>
+#include <net/ip.h>
+
+
+
+#include <elan/epcomms.h>
+#include <elan/epsvc.h>
+
+#include "eip_linux.h"
+#include "eip_stats.h"
+
+#ifdef UNUSED
+static void eip_skb_display(struct sk_buff *);
+#endif
+static void eip_iph_display(struct iphdr *);
+#ifdef UNUSED
+static void eip_eiph_display(EIP_HEADER *);
+static void eip_packet_display(unsigned char *);
+#endif
+static void eip_tmd_display(EIP_TMD *);
+static void eip_tmd_head_display(EIP_TMD_HEAD *);
+static void eip_rmd_display(EIP_RMD *);
+static void eip_rmd_head_display(EIP_RMD_HEAD *);
+
+static void eip_rmd_reclaim(EIP_RMD *);
+
+static inline EP_NMH *eip_dma_reserve(int, int);
+static inline void __eip_tmd_load(EIP_TMD *, EP_RAILMASK *);
+static inline void __eip_tmd_unload(EIP_TMD *);
+static inline unsigned long eip_buff_alloc(int, int);
+static inline void eip_buff_free(unsigned long, int);
+static struct iphdr *eip_ipfrag_get(char *);
+static inline void eip_rmd_free(EIP_RMD *);
+static inline void eip_skb_load(EIP_RMD *);
+static inline void eip_skb_unload(EIP_RMD *);
+static inline void eip_rmd_requeue(EIP_RMD *);
+static EIP_RMD *eip_rmd_alloc(int, int);
+static int eip_rmd_alloc_replace(EIP_RMD *, int, int);
+static int eip_rmd_alloc_queue(int, int, int, int);
+static int eip_rmds_alloc(void);
+static void eip_rxhandler(EP_RXD *);
+static void eip_rx_tasklet(unsigned long);
+static inline void eip_tmd_init(EIP_TMD *, unsigned long, EIP_TMD_HEAD *, unsigned long, int);
+static inline EIP_TMD *eip_tmd_get(int);
+static inline void eip_tmd_put(EIP_TMD *);
+static inline void eip_tmd_load(EIP_TMD *);
+static inline void eip_tmd_unload(EIP_TMD *);
+static inline EIP_TMD *eip_tmd_alloc_queue(EIP_TMD *, EIP_TMD_HEAD *, int);
+static inline EIP_TMD *eip_tmd_alloc_queue_copybreak(EIP_TMD_HEAD *, int);
+static inline EIP_TMD *eip_tmd_alloc_queue_aggreg(EIP_TMD_HEAD *, int);
+static int eip_tmds_alloc(void);
+int eip_hard_start_xmit(struct sk_buff *, struct net_device *);
+static inline int eip_do_xmit(EIP_TMD *, EP_NMD *i, EP_PAYLOAD *);
+static void eip_txhandler(EP_TXD *, void *, EP_STATUS);
+static void eip_tx_tasklet(unsigned long);
+void eip_stop_queue(void);
+void eip_start_queue(void);
+static int eip_open(struct net_device *);
+static int eip_close(struct net_device *);
+static struct net_device_stats *eip_get_stats(struct net_device *);
+static int eip_change_mtu(struct net_device *, int);
+
+static int eip_rx_dropping = 0;
+static int eip_rx_tasklet_locked = 1;
+
+/* Global */
+struct timer_list eip_rx_tasklet_timer;
+       
+EIP_RX *eip_rx = NULL;
+EIP_TX *eip_tx = NULL;
+int  eip_checksum_state=CHECKSUM_NONE;
+
+int tmd_max = EIP_TMD_MAX_NR;
+int rmd_max = EIP_RMD_MAX_NR;
+int rx_envelope_nr = EIP_RX_ENVELOPE_NR;
+int rx_granularity = EIP_RX_GRANULARITY;
+int tx_copybreak_max = EIP_TX_COPYBREAK_MAX;
+EP_RAILMASK tx_railmask = EP_RAILMASK_ALL;
+int eipdebug = 0;
+
+#ifdef UNUSED
+static void eip_skb_display(struct sk_buff *skb)
+{
+       if (skb) {
+               __EIP_DBG_PRINTF("SKB [%p] : len %d truesize %d  proto %x pkt type %x cloned %d users %d summed %d\n", 
+                       skb, skb->len, skb->truesize, skb->protocol, skb->pkt_type, skb->cloned, atomic_read(&skb->users), skb->ip_summed);
+               __EIP_DBG_PRINTF("SKB [%p] : skb_shinfo dataref %d nr_frags %d frag_list[%p] (device %p)\n", skb,
+                        atomic_read(&skb_shinfo(skb)->dataref), skb_shinfo(skb)->nr_frags, skb_shinfo(skb)->frag_list, skb->dev);
+               __EIP_DBG_PRINTF("SKB [%p] : head[%p] data[%p] tail [%p] end [%p] data_len [%d]\n", skb, skb->head, skb->data, 
+                               skb->tail, skb->end, skb->data_len);
+               __EIP_DBG_PRINTF("SKB [%p] : Transport Layer h.(th, uh, icmph, raw)[%p]\n", skb, skb->h.th);
+               __EIP_DBG_PRINTF("SKB [%p] : Network Layer      nh.(iph, arph, raw)[%p]\n", skb, skb->nh.iph);
+               __EIP_DBG_PRINTF("SKB [%p] : Link Layer         mac.(ethernet, raw)[%p]\n", skb, skb->mac.ethernet);
+               return;
+       }
+       EIP_ERR_PRINTF("SKB IS NULL - NO SKB TO DISPLAY\n");
+}
+#endif
+static void eip_iph_display(struct iphdr *iph)
+{
+       if (iph) {
+               __EIP_DBG_PRINTF("IPH [%p] : version %d header len %d TOS 0x%x Total len %d\n", 
+                       iph, iph->version, iph->ihl, htons(iph->tos), htons(iph->tot_len));
+               __EIP_DBG_PRINTF("IPH [%p] : id %d frag flags 0x%x offset %d\n",
+                               iph, htons(iph->id), (iph->frag_off & htons(IP_CE | IP_DF | IP_MF)) >> 4, 
+                               (htons(iph->frag_off) << 3) & IP_OFFSET);
+               __EIP_DBG_PRINTF("IPH [%p] : TTL %d proto %d header checksum 0x%x\n", iph, iph->ttl, iph->protocol, iph->check);
+               __EIP_DBG_PRINTF("IPH [%p] : IP src %u.%u.%u.%u dest %u.%u.%u.%u\n", iph, 
+                                ((unsigned char *)&(iph->saddr))[0],((unsigned char *)&(iph->saddr))[1], ((unsigned char *)&(iph->saddr))[2],((unsigned char *)&(iph->saddr))[3],
+                                ((unsigned char *)&(iph->daddr))[0],((unsigned char *)&(iph->daddr))[1], ((unsigned char *)&(iph->daddr))[2],((unsigned char *)&(iph->daddr))[3]);
+               return;
+       }
+       EIP_ERR_PRINTF("IPH IS NULL - NO IPH TO DISPLAY\n");
+}
+#ifdef UNUSED
+static void eip_eiph_display(EIP_HEADER * eiph)
+{
+       if (eiph) {
+               __EIP_DBG_PRINTF("EIPH [%p] : dhost %04x.%04x.%04x sap %x\n", eiph, eiph->h_dhost.ip_bcast, eiph->h_dhost.ip_inst, 
+                               eiph->h_dhost.ip_addr, eiph->h_sap);
+               __EIP_DBG_PRINTF("EIPH [%p] : shost %04x.%04x.%04x \n", eiph, eiph->h_shost.ip_bcast, eiph->h_shost.ip_inst,
+                                eiph->h_shost.ip_addr);
+               return;
+       }
+       EIP_ERR_PRINTF("EIPH IS NULL - NO EIPH TO DISPLAY\n");
+}
+static void eip_packet_display(unsigned char *data)
+{
+       eip_eiph_display((EIP_HEADER *) data);
+       eip_iph_display((struct iphdr *) (data + EIP_HEADER_PAD + ETH_HLEN));
+}
+#endif
+static void eip_tmd_display(EIP_TMD * tmd)
+{
+       if (tmd) {
+               __EIP_DBG_PRINTF("\t\tTMD [%p] : next[%p] skb[%p] DVMA[%d]\n", tmd, tmd->chain.next, tmd->skb, tmd->dvma_idx);
+               if (tmd->dma_base)
+                       __EIP_DBG_PRINTF("TMD [%p] : head[%p] *data 0x%lx\n", tmd, tmd->head, *((unsigned long *) tmd->dma_base));
+               else
+                       __EIP_DBG_PRINTF("TMD [%p] : head[%p] NO DATA !!!\n", tmd, tmd->head);
+               __EIP_DBG_PRINTF("TMD [%p] : DMA(%lx,%d,%d) ebase[%x]\n",tmd,  tmd->dma_base, tmd->dma_len, tmd->nmd.nmd_len,
+                                tmd->nmd.nmd_addr);
+               return;
+       }
+       EIP_ERR_PRINTF("TMD IS NULL - NO TMD TO DISPLAY\n");
+       
+}
+static void eip_ipf_display(EIP_IPFRAG * ipf)
+{
+       if (ipf) {
+               __EIP_DBG_PRINTF("IPF[%p] : datagram len %d dma correction %d uts %lx frag_nr %d\n", ipf, ipf->datagram_len,
+                               ipf->dma_correction, ipf->timestamp.tv_usec, ipf->frag_nr);
+               eip_tmd_display((EIP_TMD *) ipf);
+               return;
+       }
+       EIP_ERR_PRINTF("IPF IS NULL - NO IPF TO DISPLAY\n");
+}
+
+static void eip_tmd_head_display(EIP_TMD_HEAD * head)
+{
+       if (head) {
+               __EIP_DBG_PRINTF("TMD HEAD [%p] : handle[%p] tmds[%p] %3.3d/%3.3d/%3.3d\n", head, head->handle, head->tmd, 
+                       EIP_STAT_QUEUED_GET(&head->stats), EIP_STAT_ALLOC_GET(&head->stats),
+                       eip_tx->tmd_max_nr);
+               return;
+       }
+       EIP_ERR_PRINTF("TMD HEAD IS NULL - NO TMD HEAD TO DISPLAY\n");
+}
+static void eip_rmd_display(EIP_RMD * rmd)
+{
+       if (rmd) {
+               __EIP_DBG_PRINTF("RMD [%p] : next[%p] rxd[%p] DVMA[%d]\n", rmd, rmd->chain.next, rmd->rxd, rmd->dvma_idx);
+               __EIP_DBG_PRINTF("RMD [%p] : head[%p]\n", rmd, rmd->head); 
+               __EIP_DBG_PRINTF("RMD [%p] : ebase[%x]\n", rmd,  rmd->nmd.nmd_addr); 
+               return;
+       }
+       EIP_ERR_PRINTF("RMD IS NULL - NO RMD TO DISPLAY\n");
+}
+static void eip_rmd_head_display(EIP_RMD_HEAD * head)
+{
+       if (head) {
+               __EIP_DBG_PRINTF("RMD HEAD [%p] : rcvr[%p] handle[%p] busy list[%p]\n", head, head->rcvr, head->handle, head->busy_list);
+               __EIP_DBG_PRINTF("RMD HEAD [%p] : %3.3d/%3.3d/%3.3d\n", head, 
+                               EIP_STAT_QUEUED_GET(&head->stats), EIP_STAT_ALLOC_GET(&head->stats), eip_rx->rmd_max_nr);
+               return;
+       }
+       EIP_ERR_PRINTF("RMD HEAD IS NULL - NO RMD HEAD TO DISPLAY\n");
+}
+
+/* END  - DISPLAY FUNCTIONS */
+static inline EP_NMH *eip_dma_reserve(int pages_nr, int perm)
+{
+       EP_NMH *handle = ep_dvma_reserve(eip_tx->ep_system, pages_nr, perm);
+       
+       if (handle)
+               EIP_DBG_PRINTF(EIP_DBG_EP_DVMA, "HANDLE [%p] %d pages of elan address space reserved\n", 
+                               handle, pages_nr);
+       else
+               EIP_ERR_PRINTF("cannot reserve %d page(s) of elan address space\n", pages_nr);
+
+       return handle;
+}
+
+static inline void __eip_tmd_load(EIP_TMD * tmd, EP_RAILMASK *rmask)
+{
+       EIP_ASSERT(tmd->nmd.nmd_len > 0);
+       
+       ep_dvma_load(eip_tx->ep_system, NULL, (caddr_t) tmd->dma_base, tmd->nmd.nmd_len, tmd->head->handle,
+                       tmd->dvma_idx, rmask, &tmd->nmd);
+}
+
+static inline void __eip_tmd_unload(EIP_TMD * tmd)
+{
+       EIP_ASSERT(tmd->nmd.nmd_addr && tmd->head->handle);
+       
+       ep_dvma_unload(eip_tx->ep_system, tmd->head->handle, &tmd->nmd);
+       tmd->nmd.nmd_addr = 0;
+}
+static inline unsigned long eip_buff_alloc(int buff_len, int gfp)
+{
+       unsigned long buff_base = (buff_len < PAGE_SIZE) ? 
+                               (unsigned long) kmalloc(buff_len, gfp) :
+                               __get_dma_pages(gfp, get_order(buff_len));
+       
+       if (likely(buff_base))
+               return buff_base;
+
+       EIP_ERR_PRINTF("cannot allocate %db of memory\n", buff_len);
+       return 0;
+}
+static inline void eip_buff_free(unsigned long buff_base, int buff_len)
+{
+       (buff_len < PAGE_SIZE) ?  kfree((void *) buff_base) :
+               free_pages(buff_base, get_order(buff_len));
+}
+static struct iphdr *eip_ipfrag_get(char *data)
+{
+       struct ethhdr *eh = (struct ethhdr *) (data);
+       struct iphdr *iph;
+
+       if (eh->h_proto == htons(ETH_P_IP)) {
+               iph = (struct iphdr *) ((char *) eh + ETH_HLEN);
+
+               /* EIP_DBG(eip_iph_display(iph)); */
+
+               if ((iph->frag_off & htons(IP_MF | IP_OFFSET)))
+                       return iph;
+       }
+       return NULL;
+}
+
+static inline void eip_rmd_free(EIP_RMD * rmd)
+{
+       EIP_ASSERT2(rmd->nmd.nmd_addr == 0, eip_rmd_display, rmd);
+       
+       if ( rmd->skb != NULL) 
+               kfree_skb (rmd->skb);
+       
+       kfree(rmd);
+
+       EIP_DBG_PRINTF(EIP_DBG_MEMFREE, "RMD [%p] : FREED\n", rmd);
+}
+static inline void eip_skb_load(EIP_RMD * rmd)
+{
+       EP_RAILMASK rmask = rmd->rxd ? ep_rxd_railmask (rmd->rxd) : 0;
+
+       EIP_ASSERT(skb_tailroom(rmd->skb) > 0);
+
+       ep_dvma_load(eip_tx->ep_system, NULL, (caddr_t) rmd->skb->data, skb_tailroom(rmd->skb), rmd->head->handle,
+                    rmd->dvma_idx, &rmask, &rmd->nmd);
+       
+       EIP_DBG_PRINTF(EIP_DBG_RMD_EP_DVMA, "RMD [%p] : LOADED\n", rmd);
+}
+static inline void eip_skb_unload(EIP_RMD * rmd)
+{
+       EIP_ASSERT(rmd->nmd.nmd_addr && rmd->head->handle);
+       
+       ep_dvma_unload(eip_tx->ep_system, rmd->head->handle, &rmd->nmd);
+       rmd->nmd.nmd_addr = 0;
+       
+       EIP_DBG_PRINTF(EIP_DBG_RMD_EP_DVMA, "RMD [%p] : UNLOADED\n", rmd);
+}
+static inline void eip_rmd_requeue(EIP_RMD * rmd)
+{
+       EIP_ASSERT(rmd->rxd);
+
+       rmd->chain.next    = NULL;
+
+       ep_requeue_receive(rmd->rxd, eip_rxhandler, rmd, &rmd->nmd, EP_NO_ALLOC|EP_NO_SLEEP );
+
+       atomic_inc(&rmd->head->stats);
+       
+       EIP_DBG_PRINTF(EIP_DBG_RMD_QUEUE, "RMD [%p] : REQUEUED\n", rmd);
+}
+static EIP_RMD * eip_rmd_alloc(int svc, int gfp)
+{
+       int buff_len = EIP_SVC_SMALLEST_LEN << svc;
+       EIP_RMD *rmd;
+       struct sk_buff *skb;
+
+       if (!(skb = alloc_skb((buff_len -  EIP_EXTRA), gfp)))
+               return NULL;
+       
+       skb_reserve(skb, 2);
+
+       if (!(rmd = (EIP_RMD *) kmalloc(buff_len, gfp))) {
+               kfree_skb(skb);
+               return NULL;
+       }
+
+       rmd->skb = skb;
+
+       rmd->chain.next = NULL;
+       rmd->rxd = NULL;
+       rmd->head = &eip_rx->head[svc];
+
+       return rmd;
+}
+
+static int eip_rmd_alloc_replace(EIP_RMD *rmd, int svc, int gfp) 
+{
+       struct sk_buff *skb,*old;
+       int buff_len = EIP_SVC_SMALLEST_LEN << svc;
+
+       if (!(skb = alloc_skb(buff_len, gfp)))
+               return 1;
+       
+       skb_reserve(skb, 2);
+
+       eip_skb_unload(rmd);
+
+       old      = rmd->skb;
+       rmd->skb = skb;
+
+       eip_skb_load(rmd);
+
+       eip_rmd_requeue(rmd);
+
+       kfree_skb(old);
+
+       return 0;
+}
+
+static int eip_rmd_alloc_queue(int svc, int dvma_idx, int gfp, int attr)
+{
+       EIP_RMD * rmd = eip_rmd_alloc(svc, gfp);
+
+       if (!rmd)
+               return 1;
+
+       EIP_STAT_ALLOC_ADD(&rmd->head->stats, 1);
+
+       rmd->dvma_idx = dvma_idx;
+       eip_skb_load(rmd);
+
+       EIP_DBG2(EIP_DBG_RMD, eip_rmd_display, rmd, "RMD [%p] : ALLOCATED for SVC 0x%x\n", rmd, svc);
+
+       if (ep_queue_receive(rmd->head->rcvr, eip_rxhandler, (void *) rmd, &rmd->nmd, attr) == ESUCCESS) {
+               atomic_inc(&rmd->head->stats);
+               EIP_DBG_PRINTF(EIP_DBG_RMD_QUEUE, "RMD [%p] : QUEUED on SVC 0x%x\n", rmd, svc);
+               return 0;
+       }
+       
+       EIP_ERR_PRINTF("RMD [%p] : couldn't be QUEUED on SVC 0x%x\n", rmd, svc);
+
+       EIP_STAT_ALLOC_SUB(&rmd->head->stats, 1);
+
+       eip_skb_unload(rmd);
+       eip_rmd_free(rmd);
+
+       return 1;
+}
+
+static int eip_rmds_alloc(void)
+{
+       int idx, svc;
+
+       eip_rx->irq_list = NULL;
+       eip_rx->irq_list_nr = 0;
+
+       for (svc = 0; svc < EIP_SVC_NR; svc++) {
+               eip_rx->head[svc].rcvr = ep_alloc_rcvr(eip_tx->ep_system, EIP_SVC_EP(svc), rx_envelope_nr);
+               if (!eip_rx->head[svc].rcvr) {
+                       EIP_ERR_PRINTF("Cannot install receiver for SVC 0x%x - maybe cable is disconnected\n", svc);
+                       return -EAGAIN;
+               }
+
+               eip_rx->head[svc].handle =
+                   eip_dma_reserve(EIP_DVMA_PAGES((EIP_SVC_SMALLEST_LEN << svc)) * eip_rx->rmd_max_nr,
+                                   EP_PERM_WRITE);
+               if (!eip_rx->head[svc].handle)
+                       return -ENOMEM;
+               
+               EIP_DBG(EIP_DBG_RMD_HEAD, eip_rmd_head_display, &eip_rx->head[svc]);
+
+               for (idx = 0; idx < EIP_RMD_NR; idx++) {
+                       if (eip_rmd_alloc_queue(svc, idx * EIP_DVMA_PAGES((EIP_SVC_SMALLEST_LEN << svc)), 
+                                               GFP_KERNEL, EP_NO_SLEEP))
+                               return -ENOMEM;
+               }
+       }
+       return 0;
+}
+static void eip_rmds_free(void)
+{
+       unsigned long flags;
+       EIP_RMD *rmd;
+       int svc; 
+       
+       spin_lock_irqsave(&eip_rx->lock, flags);
+       rmd = eip_rx->irq_list;
+       eip_rx->irq_list = NULL;
+       eip_rx->irq_list_nr = 0;
+       spin_unlock_irqrestore(&eip_rx->lock, flags);
+
+       eip_rmd_reclaim(rmd);
+       
+       for (svc = 0; svc < EIP_SVC_NR ; svc++) {
+               
+               while ((rmd = eip_rx->head[svc].busy_list)) {
+                       eip_rx->head[svc].busy_list = NULL;
+                       eip_rmd_reclaim(rmd);
+                       if (eip_rx->head[svc].busy_list) {
+                               EIP_DBG_PRINTF(EIP_DBG_RMD_QUEUE, "Still RMD [%p] on BUSY list SVC 0x%d - Scheduling\n", rmd, svc);     
+                               schedule();
+                       }
+               }
+
+               EIP_ASSERT(EIP_STAT_QUEUED_GET(&eip_rx->head[svc].stats) == EIP_STAT_ALLOC_GET(&eip_rx->head[svc].stats));
+               
+               EIP_DBG_PRINTF(EIP_DBG_GEN, "HEAD[%p] : FREEING RCVR [%p]\n", &eip_rx->head[svc],
+                               eip_rx->head[svc].rcvr);
+               
+               ep_free_rcvr(eip_rx->head[svc].rcvr);
+
+               EIP_DBG_PRINTF(EIP_DBG_EP_DVMA, "HEAD[%p] : RELEASING DVMA [%p]\n", &eip_rx->head[svc], 
+                               eip_rx->head[svc].handle);
+
+               ep_dvma_release(eip_tx->ep_system, eip_rx->head[svc].handle);
+       }
+
+}
+static int eip_rx_queues_low (void) {
+       int svc;
+       for (svc = 0; svc < EIP_SVC_NR; svc++) 
+               if (EIP_STAT_QUEUED_GET(&eip_rx->head[svc].stats)  < EIP_RMD_ALLOC_THRESH) 
+                       return (1);
+       return (0);
+}
+static void eip_rxhandler(EP_RXD * rxd)
+{
+       EIP_RMD *rmd            = (EIP_RMD *) ep_rxd_arg(rxd);
+       EP_STATUS ret           = ep_rxd_status(rxd);
+       EP_PAYLOAD * payload    = ep_rxd_payload(rxd);
+       unsigned long data      = (unsigned long) rmd->skb->data; 
+       int frag_nr             = 0;
+       int len;
+
+       struct sk_buff *skb;
+       static char count = 0;
+
+       atomic_dec(&rmd->head->stats);
+       rmd->rxd = rxd;
+
+       if (likely(ret == EP_SUCCESS)) {
+
+               rmd->head->dma++;
+
+               if ( eip_rx_dropping) {
+                   eip_rmd_requeue(rmd);
+                   return;
+               }
+
+               len = (payload) ? payload->Data[frag_nr++] : ep_rxd_len(rxd);
+
+               EIP_DBG(EIP_DBG_RMD, eip_rmd_display, rmd);
+
+again:
+               if ( (skb = skb_clone(rmd->skb, GFP_ATOMIC)) ) {
+                       unsigned int off = (data - (unsigned long) rmd->skb->data);
+
+                       /* have to set the length before calling
+                        * skb pull as it will not allow you to
+                        * pull past the end */
+
+                       skb_put (skb, off + len);
+                       skb_pull (skb, off);
+
+                       skb->protocol = eth_type_trans(skb, eip_rx->net_device);
+                       skb->ip_summed = eip_checksum_state;
+                       skb->dev = eip_rx->net_device;
+
+                       /* Fabien/David/Mike this is a hack/fix to allow aggrigation of packets to work.
+                        * The problem is ip_frag looks at the truesize to see if it is caching too much space.
+                        * As we are reusing a large skb (cloned) for a number of small fragments, they appear to take up alot of space.
+                        * so ip_frag dropped them after 4 frags (not good). So we lie and set the truesize to just bigger than the data. 
+                        */
+                       if (payload) 
+                               skb->truesize = SKB_DATA_ALIGN(skb->len + EIP_HEADER_PAD) +sizeof(struct sk_buff);
+
+               }
+               if ( (skb) && 
+                    (netif_rx(skb) != NET_RX_DROP)){
+
+                       eip_rx->bytes += len;
+                       
+                       if (payload && payload->Data[frag_nr] ) {
+                               data += EIP_IP_ALIGN(len);
+                               len   = payload->Data[frag_nr++];
+                               goto again;
+                       }
+                       eip_rx->packets += ++frag_nr;
+               } else if ( (eip_rx->dropped++ % 20) == 0)
+                               __EIP_DBG_PRINTK("Packet dropped by the TCP/IP stack - increase /proc/sys/net/core/netdev_max_backlog\n");
+       } else if (ret == EP_SHUTDOWN ) {
+               EIP_DBG2(EIP_DBG_RMD, eip_rmd_display, rmd, "ABORTING\n");
+                ep_complete_receive(rxd);
+                eip_skb_unload(rmd);
+               EIP_STAT_ALLOC_SUB(&rmd->head->stats, 1);
+                eip_rmd_free(rmd);
+               return;
+       } else {
+               EP_ENVELOPE *env = ep_rxd_envelope(rxd);
+               EP_NMD *nmd ;
+               
+               EIP_ERR_PRINTF("RMD[%p] : RECEIVE ret = %d\n", rmd, ret);
+
+               for (len = 0 ; len < env->nFrags ; len++) {
+                       nmd = &env->Frags[len];
+                       EIP_ERR_PRINTF("RMD[%p] : ep_frag #%d nmd_addr [%x] nmd_len %d\n", rmd, len, 
+                                       (unsigned int) nmd->nmd_addr, nmd->nmd_len);
+               }
+               eip_rx->errors++;
+               EIP_ASSERT2(atomic_read(&skb_shinfo(rmd->skb)->dataref) == 1, eip_rmd_display, rmd);
+       }
+
+       /* data is used to store the irq flags */
+       spin_lock_irqsave(&eip_rx->lock, data);
+       rmd->chain.next = eip_rx->irq_list;
+       eip_rx->irq_list = rmd;
+       eip_rx->irq_list_nr++;
+       spin_unlock_irqrestore(&eip_rx->lock, data);
+
+       if ( !timer_pending (&eip_rx_tasklet_timer)      /* the timer not already set                     */
+            && ( (count++ % eip_rx->sysctl_granularity) /* and either we have passed up a number of them */
+                 || eip_rx_queues_low() ))              /* or we are low                                 */
+               mod_timer (&eip_rx_tasklet_timer, lbolt + 1);
+}
+
+/* dest ; if the buffer still reference on it mocve the rmd to the dest list */
+static void eip_rmd_reclaim(EIP_RMD *rmd) 
+{
+       EIP_RMD *rmd_next = rmd;
+       int dataref;
+
+       while (rmd_next) {
+               rmd = rmd_next;
+               rmd_next = rmd_next->chain.next;
+
+               dataref = atomic_read(&skb_shinfo(rmd->skb)->dataref);
+               EIP_ASSERT(dataref > 0);
+               
+               if (dataref == 1) {
+                       eip_rmd_requeue(rmd);
+               } else {
+                       rmd->chain.next = rmd->head->busy_list;
+                       rmd->head->busy_list = rmd;
+               }
+       }
+}
+static void eip_rx_tasklet(unsigned long arg)
+{
+       EIP_RMD *rmd, *rmd_next;
+       unsigned long flags;
+       short svc, queued;
+       int   needs_reschedule;
+
+       if (eip_rx_tasklet_locked) /* we dont want the tasklet to do anything when we are finishing */
+           return;
+
+       for (svc = 0; svc < EIP_SVC_NR; svc++) {
+               rmd = eip_rx->head[svc].busy_list;
+               eip_rx->head[svc].busy_list = NULL;
+               eip_rmd_reclaim(rmd);
+       }
+
+       spin_lock_irqsave(&eip_rx->lock, flags);
+       rmd = eip_rx->irq_list;
+       eip_rx->irq_list = NULL;
+       eip_rx->irq_list_nr = 0;
+       spin_unlock_irqrestore(&eip_rx->lock, flags);
+       
+       eip_rmd_reclaim(rmd);
+
+       needs_reschedule = 0;
+
+       for (svc = 0; svc < EIP_SVC_NR; svc++) {
+               /* the plan is : allocate some more if possible or steall some dvma space from those on the EIP_BUSY_LIST */
+               queued = EIP_STAT_QUEUED_GET(&eip_rx->head[svc].stats);
+
+               EIP_ASSERT(queued >= 0 && queued <= EIP_RMD_MAX_NR);    
+               
+               if (queued < EIP_RMD_ALLOC_THRESH) {
+                       short allocated = EIP_STAT_ALLOC_GET(&eip_rx->head[svc].stats);
+                       short how_many; 
+
+                       EIP_ASSERT(allocated >= 0 && allocated <= EIP_RMD_MAX_NR);
+                       
+                       if (likely(allocated < eip_rx->rmd_max_nr)) {
+
+                               how_many = (((allocated / EIP_RMD_ALLOC_STEP) + 1) * EIP_RMD_ALLOC_STEP);
+                               if (how_many > eip_rx->rmd_max_nr)
+                                       how_many = eip_rx->rmd_max_nr;
+
+                               for (; allocated < how_many &&  
+                                                       (eip_rmd_alloc_queue(svc, allocated * EIP_DVMA_PAGES((EIP_SVC_SMALLEST_LEN << svc)), 
+                                                                             GFP_ATOMIC, EP_NO_ALLOC|EP_NO_SLEEP) == 0) ; allocated++);
+                               if ( allocated != how_many ) {
+                                       eip_rx->reschedule++;
+                                       needs_reschedule = 1;
+                               }
+                       } else {
+                               /* steal how_many rmds and put them on the aside list */
+                               how_many = EIP_RMD_ALLOC_THRESH - queued;
+
+                               EIP_ASSERT(how_many >= 0 && how_many <= EIP_RMD_ALLOC_THRESH);
+                               
+                               rmd_next = eip_rx->head[svc].busy_list;
+                               eip_rx->head[svc].busy_list = NULL;
+
+                               while (how_many-- && rmd_next) {
+                                       rmd = rmd_next;
+                                       rmd_next = rmd_next->chain.next;
+
+                                       if (eip_rmd_alloc_replace(rmd, svc, GFP_ATOMIC)) {
+                                               rmd_next = rmd;
+                                               break;
+                                       }
+                               }
+                               eip_rx->head[svc].busy_list = rmd_next;
+                               if ( how_many )
+                                       needs_reschedule = 1;
+                       }
+               }
+       }
+       
+       if ( needs_reschedule && ( !timer_pending (&eip_rx_tasklet_timer)))
+               mod_timer (&eip_rx_tasklet_timer, lbolt + 2);
+}
+
+static inline void eip_tmd_init(EIP_TMD * tmd, unsigned long buff_base, EIP_TMD_HEAD * head, unsigned long buff_len,
+                               int dvma_idx)
+{
+       tmd->dvma_idx = dvma_idx;
+       tmd->dma_base = buff_base;
+       tmd->dma_len = -1;
+       tmd->skb = NULL;
+       tmd->head = head;
+       tmd->chain.next = NULL;
+
+       if (tmd->head != &eip_tx->head[EIP_TMD_STD]) {
+               tmd->nmd.nmd_len = buff_len;
+               eip_tmd_load(tmd);
+       } else  {
+               tmd->nmd.nmd_len  = -1;
+               tmd->nmd.nmd_addr = 0;
+       }
+}
+
+static inline EIP_TMD *eip_tmd_get(int id)
+{
+       unsigned long flags;
+       EIP_TMD *tmd = NULL;
+       spin_lock_irqsave(&eip_tx->lock, flags);
+       while ((tmd = eip_tx->head[id].tmd) == NULL) {
+               spin_unlock_irqrestore(&eip_tx->lock, flags);
+               if (ep_enable_txcallbacks(eip_tx->xmtr) == 0) {
+
+                       spin_lock_irqsave (&eip_tx->lock, flags);
+                       if (eip_tx->head[id].tmd == NULL) {
+                               __EIP_DBG_PRINTF("Cannot get a TMD on head %d ... stopping queue\n", id);
+                               
+                               eip_stop_queue ();
+                               
+                               spin_unlock_irqrestore (&eip_tx->lock, flags);
+
+                               return NULL;
+                       }
+                       spin_unlock_irqrestore (&eip_tx->lock, flags);
+               }
+
+               ep_disable_txcallbacks(eip_tx->xmtr);
+               spin_lock_irqsave(&eip_tx->lock, flags);
+       }
+       eip_tx->head[id].tmd = tmd->chain.next;
+       spin_unlock_irqrestore(&eip_tx->lock, flags);
+       atomic_dec(&tmd->head->stats);
+       return tmd;
+}
+
+static inline void eip_tmd_put(EIP_TMD * tmd)
+{
+       unsigned long flags;
+
+       tmd->skb = NULL;
+
+       spin_lock_irqsave(&eip_tx->lock, flags);
+       tmd->chain.next = tmd->head->tmd;
+       tmd->head->tmd = tmd;
+       spin_unlock_irqrestore(&eip_tx->lock, flags);
+       atomic_inc(&tmd->head->stats);
+
+       eip_start_queue();
+
+       EIP_DBG_PRINTF(EIP_DBG_TMD_QUEUE, "TMD [%p] : REQUEUED\n", tmd);
+}
+static inline void eip_tmd_load(EIP_TMD * tmd)
+{
+       EP_RAILMASK rmask = tx_railmask;
+
+       __eip_tmd_load(tmd, &rmask);
+       
+       EIP_DBG_PRINTF(EIP_DBG_EP_DVMA, "TMD [%p] : LOADED\n", tmd);
+}
+static inline void eip_tmd_unload(EIP_TMD * tmd)
+{
+       __eip_tmd_unload(tmd);
+       
+       EIP_DBG_PRINTF(EIP_DBG_EP_DVMA, "TMD [%p] : UNLOADED\n", tmd);
+}
+static inline void eip_tmd_free(EIP_TMD * tmd)
+{
+       eip_buff_free(tmd->dma_base, tmd->nmd.nmd_len);
+       
+       EIP_DBG_PRINTF(EIP_DBG_MEMFREE, "TMD [%p] : FREED\n", tmd);
+       
+       EIP_STAT_ALLOC_SUB(&tmd->head->stats, 1);
+}
+
+/* tmd on a separate block */
+static inline EIP_TMD *eip_tmd_alloc_queue(EIP_TMD * tmd, EIP_TMD_HEAD * head, int dvma_idx)
+{
+       eip_tmd_init(tmd, 0, head, -1, dvma_idx);
+
+       eip_tmd_put(tmd);
+
+       EIP_STAT_ALLOC_ADD(&tmd->head->stats, 1);
+       EIP_DBG(EIP_DBG_TMD, eip_tmd_display, tmd);
+       return tmd;
+}
+/* tmd on the buffer */
+static inline EIP_TMD *eip_tmd_alloc_queue_copybreak(EIP_TMD_HEAD * head, int dvma_idx)
+{
+       EIP_TMD *tmd;
+       unsigned long buff_base;
+
+       if (!(buff_base = eip_buff_alloc(tx_copybreak_max + sizeof(EIP_TMD), GFP_KERNEL)))
+               return NULL;
+
+       tmd = (EIP_TMD *) (buff_base + tx_copybreak_max);
+       eip_tmd_init(tmd, buff_base, head, tx_copybreak_max, dvma_idx);
+
+       eip_tmd_put(tmd);
+       EIP_STAT_ALLOC_ADD(&tmd->head->stats, 1);
+       EIP_DBG(EIP_DBG_TMD, eip_tmd_display, tmd);
+       return tmd;
+}
+
+/* ipf are on the buffer */
+static inline EIP_TMD *eip_tmd_alloc_queue_aggreg(EIP_TMD_HEAD * head, int dvma_idx)
+{
+       EIP_TMD *tmd;
+       unsigned long buff_base;
+
+       if (!(buff_base = eip_buff_alloc(EIP_SVC_BIGGEST_LEN, GFP_KERNEL)))
+               return NULL;
+
+       tmd = (EIP_TMD *) (buff_base + EIP_SVC_BIGGEST_LEN - sizeof(EIP_IPFRAG));
+       eip_tmd_init(tmd, buff_base, head, EIP_SVC_BIGGEST_LEN - sizeof(EIP_IPFRAG), dvma_idx);
+
+       eip_tmd_put(tmd);
+       EIP_STAT_ALLOC_ADD(&tmd->head->stats, 1);
+       EIP_DBG(EIP_DBG_TMD, eip_tmd_display, tmd);
+       return tmd;
+}
+
+static int eip_tmds_alloc()
+{
+       int i;
+       int page_nr;
+       EIP_TMD *tmd;
+
+       page_nr = EIP_DVMA_PAGES(tx_copybreak_max);
+
+       eip_tx->head[EIP_TMD_COPYBREAK].handle = eip_dma_reserve(page_nr * eip_tx->tmd_max_nr, EP_PERM_READ);
+       
+       EIP_DBG(EIP_DBG_TMD_HEAD, eip_tmd_head_display, &eip_tx->head[EIP_TMD_COPYBREAK]);
+
+       for (i = 0; i < EIP_TMD_NR; i++) {
+               if (!eip_tmd_alloc_queue_copybreak(&eip_tx->head[EIP_TMD_COPYBREAK], i * page_nr))
+                       return -ENOMEM;
+       }
+
+       eip_tx->head[EIP_TMD_STD].handle =
+           eip_dma_reserve(EIP_DVMA_PAGES(EIP_SVC_BIGGEST_LEN) * eip_tx->tmd_max_nr, EP_PERM_READ);
+       
+       EIP_DBG(EIP_DBG_TMD_HEAD, eip_tmd_head_display, &eip_tx->head[EIP_TMD_STD]);
+
+       tmd = kmalloc(sizeof(EIP_TMD) * EIP_TMD_NR, GFP_KERNEL);
+       if (!tmd) {
+               EIP_ERR_PRINTF("Cannot ALLOCATE %d of tmds\n", (int) sizeof(EIP_TMD) * EIP_TMD_NR);
+               return -ENOMEM;
+       }
+       
+       page_nr = EIP_DVMA_PAGES(EIP_SVC_BIGGEST_LEN);
+       
+       for (i = 0; i < EIP_TMD_NR; i++, tmd++) {
+               if (!eip_tmd_alloc_queue(tmd, &eip_tx->head[EIP_TMD_STD], i * page_nr))
+                       return -ENOMEM;
+       }
+
+       page_nr = EIP_DVMA_PAGES(EIP_SVC_BIGGEST_LEN);
+
+       eip_tx->head[EIP_TMD_AGGREG].handle = eip_dma_reserve(page_nr * eip_tx->tmd_max_nr, EP_PERM_READ);
+       EIP_DBG(EIP_DBG_TMD_HEAD, eip_tmd_head_display, &eip_tx->head[EIP_TMD_AGGREG]);
+
+       for (i = 0; i < EIP_TMD_NR; i++) {
+               if (!eip_tmd_alloc_queue_aggreg(&eip_tx->head[EIP_TMD_AGGREG], i * page_nr))
+                       return -ENOMEM;
+       }
+       return 0;
+}
+
+static void eip_tmds_free(void) 
+{
+       EIP_TMD *tmd;
+       EIP_TMD *tmd_next;
+       int i;
+       
+       ep_poll_transmits(eip_tx->xmtr);
+
+       for (i = 0 ; i < 3 ; i++) {
+again:
+               if (EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats) < EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats)) {
+                       EIP_DBG_PRINTF(EIP_DBG_TMD, "Polling XMTR [%p]\n", eip_tx->xmtr);       
+                       ep_poll_transmits(eip_tx->xmtr);
+                       goto again;
+               }
+       }
+       /* everything should be queued */
+        if ((tmd = eip_tx->head[EIP_TMD_COPYBREAK].tmd)) {
+            do {
+                       tmd_next = tmd->chain.next;
+                        eip_tmd_unload(tmd);
+                       
+                       EIP_DBG(EIP_DBG_TMD, eip_tmd_display, tmd);
+                       
+                        eip_tmd_free(tmd);
+            } while (tmd_next && (tmd = tmd_next));
+        }
+       
+       EIP_DBG_PRINTF(EIP_DBG_TMD_EP_DVMA, "HEAD[EIP_TMD_COPYBREAK] release DVMA [%p]\n",
+                       eip_tx->head[EIP_TMD_COPYBREAK].handle);        
+       
+        ep_dvma_release(eip_tx->ep_system, eip_tx->head[EIP_TMD_COPYBREAK].handle);
+       
+       /* these ones have been allocated as a block */
+       if ((tmd = eip_tx->head[EIP_TMD_STD].tmd)) {
+               do {
+                       if (tmd->dvma_idx == 0 ) {
+                               kfree(tmd);
+                               /* eip_tmd_free(tmd); */
+                               EIP_STAT_ALLOC_SUB(&tmd->head->stats, EIP_TMD_NR);
+                               tmd_next = NULL;
+                               EIP_DBG_PRINTF(EIP_DBG_TMD_EP_DVMA, "TMD HEAD[%p] : [EIP_TMD_STD] BLOCK FREED\n", tmd); 
+                       } else 
+                               tmd_next = tmd->chain.next;
+               } while (tmd_next && (tmd = tmd_next));
+       }
+       EIP_DBG_PRINTF(EIP_DBG_TMD_EP_DVMA, "HEAD[EIP_TMD_STD] release DVMA [%p]\n", 
+                       eip_tx->head[EIP_TMD_STD].handle);      
+       
+        ep_dvma_release(eip_tx->ep_system, eip_tx->head[EIP_TMD_STD].handle);
+       
+       if ((tmd = eip_tx->head[EIP_TMD_AGGREG].tmd)) {
+               do {
+                       tmd_next = tmd->chain.next;
+
+                       EIP_DBG(EIP_DBG_TMD, eip_tmd_display, tmd);
+                       
+                       eip_tmd_unload(tmd);
+                       eip_tmd_free(tmd);
+               } while (tmd_next && (tmd = tmd_next));
+       }
+       EIP_DBG_PRINTF(EIP_DBG_TMD_EP_DVMA, "TMD HEAD[%p] : [EIP_TMD_AGGREG] release DVMA\n", 
+                       eip_tx->head[EIP_TMD_AGGREG].handle);   
+       
+        ep_dvma_release(eip_tx->ep_system, eip_tx->head[EIP_TMD_AGGREG].handle);
+
+       ep_free_xmtr(eip_tx->xmtr);
+       EIP_DBG_PRINTF(EIP_DBG_TMD, "XMTR[%p] : FREED\n", eip_tx->xmtr);
+}
+
+static inline void eip_ipf_skb_add(EIP_IPFRAG * ipf, struct sk_buff *skb)
+{
+       int align = EIP_IP_ALIGN(skb->len);
+       
+       
+       if (ipf->dma_len == -1) {       /* like a virgin; touched for the very first time */
+               do_gettimeofday(&ipf->timestamp);
+               /* FIXE ME put that in release tmd code */
+               ipf->frag_nr            = 0;
+               ipf->dma_len            = 0;
+               ipf->datagram_len       = -1;
+               ipf->dma_correction     = 0;
+       }
+       
+       memcpy((void *) (ipf->dma_base + ipf->dma_len), skb->data, skb->len);
+       
+       if (ipf->datagram_len == -1) {
+               struct iphdr * iph = skb->nh.iph;
+               int offset = ntohs(iph->frag_off);
+
+               /* last one ? ;  offset & ~IP_OFFSET = IP fragment flags */
+               if (((offset & ~IP_OFFSET) & IP_MF) == 0) {
+                       offset &= IP_OFFSET;
+                       offset <<= 3;    
+                       ipf->datagram_len = offset + htons(iph->tot_len) - sizeof(struct iphdr);
+               }
+       }
+
+       skb->next                       = ipf->skb;
+       ipf->skb                        = skb;
+       ipf->payload.Data[ipf->frag_nr] = skb->len;
+       ipf->dma_len                   += align;
+       ipf->dma_correction            += align - skb->len  + ETH_HLEN + sizeof(struct iphdr);
+       /* FIXME ; Count got wrong if ip header has options */
+
+       ipf->frag_nr++;
+
+       EIP_DBG2(EIP_DBG_TMD, eip_ipf_display, ipf, "ADDED skb[%p] len %db ALIGNED(%db)\n", skb, skb->len, EIP_IP_ALIGN(skb->len));
+}
+
+#define eip_ipf_hasroom(ipf, skb) ((ipf->dma_len + EIP_IP_ALIGN(skb->len) < eip_tx->sysctl_ipfrag_copybreak))
+int eip_hard_start_xmit(struct sk_buff *skb, struct net_device *devnet) 
+{
+
+       EIP_TMD *tmd;
+       EP_NMD nmd;
+       struct iphdr *iph;
+       int j;
+
+       if (skb->destructor){
+               atomic_inc(&eip_tx->destructor);
+               tasklet_schedule(&eip_tx->tasklet);
+       } 
+
+       if (!(iph = eip_ipfrag_get(skb->data)) || (eip_tx->sysctl_aggregation == 0)) { /* not ip fragment */
+no_aggreg:
+               j = (skb->len < eip_tx->sysctl_copybreak) ? EIP_TMD_COPYBREAK : EIP_TMD_STD; /* j = head id */
+               
+               if (!(tmd = eip_tmd_get(j))) {
+                       if (skb->destructor)
+                               atomic_dec(&eip_tx->destructor);
+                       return 1;
+               }
+               
+               tmd->dma_len    = skb->len;
+               tmd->skb        = skb;
+               tmd->skb->next  = NULL;
+               tmd->chain.next = NULL;
+               
+               if (j == EIP_TMD_COPYBREAK) {
+                       memcpy((void *) tmd->dma_base, skb->data, skb->len);
+                       
+                       ep_nmd_subset(&nmd, &tmd->nmd, 0, skb->len);
+#ifdef EIP_MORE_STATS
+                       eip_tx->sent_copybreak++;
+#endif
+                       return eip_do_xmit(tmd, &nmd, NULL);
+               }
+               tmd->dma_base           = (unsigned long) skb->data;
+               tmd->nmd.nmd_len        = skb->len;
+               eip_tmd_load(tmd);
+
+#ifdef EIP_MORE_STATS
+               eip_tx->sent_std++;
+#endif
+               return eip_do_xmit(tmd, &tmd->nmd, NULL);
+       } else if ( skb->len > EIP_SVC_BIGGEST_LEN/2 ) { 
+               /* don't aggregate when we have a full mtu of data */
+               /* or more than 32k ; in this case it is cheaper   */
+               /* to just map the buffer and send it              */
+               goto no_aggreg;
+       } else {
+               EIP_IPFRAG *ipf = NULL;
+               unsigned long flags;
+               struct list_head *l;
+               struct iphdr *iph2;
+               int i;
+               __u16 id = iph->id;
+               __u32 saddr = iph->saddr;
+               __u32 daddr = iph->daddr;
+               __u8 protocol = iph->protocol;
+
+                       EIP_DBG(EIP_DBG_IPH, eip_iph_display, iph);
+
+               j = 0;
+
+               /* here we can't have full mtu size aggregated packet */
+               EIP_ASSERT_RET(skb->len < eip_tx->sysctl_ipfrag_copybreak, 0);
+
+               spin_lock_irqsave(&eip_tx->ipfraglock, flags);
+               list_for_each(l, &eip_tx->ipfrag) {
+                       ipf = list_entry(l, EIP_IPFRAG, list);
+                       iph2 = eip_ipfrag_get((char *) ipf->dma_base);
+                       
+                        EIP_ASSERT(iph2);
+                       
+                       if ((iph2->id == id) && 
+                                       (get_unaligned(&iph2->saddr) == saddr) && 
+                                       (get_unaligned(&iph2->daddr) == daddr) && 
+                                       (iph2->protocol == protocol)) {
+                               /* || timeout */
+                               if (eip_ipf_hasroom(ipf, skb)) {
+                                       
+                                       eip_ipf_skb_add(ipf, skb);
+                                       
+                                       if ((ipf->datagram_len != -1) && 
+                                                       (ipf->dma_len == (ipf->datagram_len + ipf->dma_correction) || 
+                                                        ipf->frag_nr == (128 / sizeof(uint32_t)))) {
+send_aggreg:
+                                               ipf->payload.Data[ipf->frag_nr] = 0;
+                                               list_del(&ipf->list);
+                                               eip_tx->ipfrag_count--;
+                                               spin_unlock_irqrestore(&eip_tx->ipfraglock, flags);
+                                       
+                                               ep_nmd_subset(&nmd, &ipf->nmd, 0, ipf->dma_len);
+                                               
+#ifdef EIP_MORE_STATS
+                                               eip_tx->sent_aggreg++;
+#endif
+                                               if ((i = eip_do_xmit((EIP_TMD *) ipf, &nmd, &ipf->payload)) != EP_SUCCESS)
+                                                       return i;
+                                               if (j)
+                                                       goto new;
+                                               return 0;
+                                       }
+                                       
+                                       spin_unlock_irqrestore(&eip_tx->ipfraglock, flags);
+                                       tasklet_schedule(&eip_tx->tasklet);
+                                       return 0;
+                               } else {
+                                       EIP_DBG_PRINTF(EIP_DBG_TMD, "IPF[%p] : FULL %db full - sending it\n", ipf, ipf->dma_len);
+                                       j = 1;
+                                       goto send_aggreg;
+                               }
+                       }
+               }
+               spin_unlock_irqrestore(&eip_tx->ipfraglock, flags);
+new:
+               if (!(ipf = (EIP_IPFRAG *) eip_tmd_get(EIP_TMD_AGGREG)))
+                       goto no_aggreg;
+
+               eip_ipf_skb_add(ipf, skb);
+               
+               spin_lock_irqsave(&eip_tx->ipfraglock, flags);
+               list_add_tail(&ipf->list, &eip_tx->ipfrag);
+               eip_tx->ipfrag_count++;
+               spin_unlock_irqrestore(&eip_tx->ipfraglock, flags);
+               tasklet_schedule(&eip_tx->tasklet);
+       }
+       return 0;
+}
+static int eip_do_xmit(EIP_TMD * tmd, EP_NMD *nmd, EP_PAYLOAD *payload)
+{
+       EIP_HEADER *eiph = (EIP_HEADER *) tmd->dma_base;
+       int         attr = EP_SET_DATA((EP_NO_SLEEP | EP_NO_INTERRUPT | EP_NO_FAILOVER), EP_TYPE_SVC_INDICATOR, EP_SVC_EIP);
+       unsigned long flags;
+       int svc, rnum;
+
+       SIZE_TO_SVC(nmd->nmd_len, svc);
+
+       EIP_DBG(EIP_DBG_TMD, eip_tmd_display, tmd);
+       /* EIP_DBG(eip_eiph_display(eiph)); */
+       
+       if (unlikely (eiph->h_dhost.ip_bcast))
+               rnum = ep_pickRail (EP_NMD_RAILMASK (nmd) & tx_railmask & ep_xmtr_availrails(eip_tx->xmtr));
+       else
+               rnum = ep_pickRail (EP_NMD_RAILMASK (nmd) & tx_railmask & ep_xmtr_noderails(eip_tx->xmtr, ntohs(eiph->h_dhost.ip_addr)));
+
+       if (rnum >= 0)
+               attr = EP_SET_PREFRAIL(attr, rnum);
+
+       /* add to inuse list  */
+       spin_lock_irqsave (&eip_tx->lock, flags);
+       list_add_tail (&tmd->chain.link, &eip_tx->inuse);
+       spin_unlock_irqrestore (&eip_tx->lock, flags);
+
+       /* ENOMEM EINVAL ECONNREFUSED ESUCCESS */
+       svc = (unlikely(eiph->h_dhost.ip_bcast)) ? 
+               ep_multicast_message(eip_tx->xmtr, -1, -1, NULL, EIP_SVC_EP(svc), attr | EP_NOT_MYSELF, eip_txhandler, tmd, payload, nmd, 1) :
+
+               ep_transmit_message(eip_tx->xmtr, ntohs(eiph->h_dhost.ip_addr), EIP_SVC_EP(svc),  attr, eip_txhandler, tmd, payload, nmd, 1);
+               
+       if (likely(svc == EP_SUCCESS))
+               return 0;
+       else if (svc == ENOMEM) {
+               EIP_ERR_PRINTF("%s", "Memory allocation error ...\n");
+               eip_tx->errors++;
+       }
+       else
+       {
+               /* EP_EINVAL occurs when the svc has a bad value or the iovec has too many frag; */
+               /* we don't use the latter option here                                        */
+               __EIP_DBG_PRINTF("TMD [%p] : DROPPED skb[%p] status = %d from ep_?_message\n", tmd, tmd->skb, svc);
+
+               eip_tx->dropped++;
+       }
+
+       eip_txhandler(NULL, tmd, -99);
+
+       /* Quadrics GNAT sw-elan/4397 - since we will "never" be able to send this packet to the */
+       /* destination node, we drop it and feign success - this has the same behaviour as an    */
+       /* ethernet where it sticks the packet on the wire, but no-one receives it.              */
+       return 0;
+}
+
+static void eip_txhandler(EP_TXD * txd, void *arg, EP_STATUS status)
+{
+       EIP_TMD *tmd = (EIP_TMD *) arg;
+       struct sk_buff *skb_next;
+       unsigned long flags;
+       int svc = 0;
+       
+       if (likely(status == EP_SUCCESS)) {
+               SIZE_TO_SVC(tmd->dma_len, svc);
+               eip_tx->dma[svc]++;
+               eip_tx->bytes += tmd->dma_len;
+               
+               if (tmd->head == &eip_tx->head[EIP_TMD_AGGREG]) {
+                       EIP_IPFRAG *ipf = (EIP_IPFRAG *) tmd;
+                       eip_tx->packets += ipf->frag_nr;
+               } else
+                       eip_tx->packets++;
+       } else {
+               if (tmd->head == &eip_tx->head[EIP_TMD_AGGREG]) {
+                       EIP_IPFRAG *ipf = (EIP_IPFRAG *) tmd;
+                       eip_tx->dropped += ipf->frag_nr;
+                       EIP_DBG_PRINTF(EIP_DBG_TMD, "txhandler aggreg packet dropped status = %d\n", status);
+               } else  {
+                       eip_tx->dropped++;
+                       EIP_DBG_PRINTF(EIP_DBG_TMD, "txhandler packet dropped status = %d\n", status);
+               }
+       }
+
+       if (tmd->head == &eip_tx->head[EIP_TMD_STD]) {
+               eip_tmd_unload(tmd);
+               tmd->dma_base = 0;
+               tmd->nmd.nmd_len = -1;
+       }
+               
+       tmd->dma_len = -1;
+       
+       svc = 0;
+       while (tmd->skb) {
+               svc++;
+               
+               if (tmd->skb->destructor)
+                       atomic_dec(&eip_tx->destructor);
+
+               skb_next = tmd->skb->next;
+               dev_kfree_skb_any(tmd->skb);
+               tmd->skb = skb_next;
+       }
+       EIP_DBG_PRINTF(EIP_DBG_TMD, "IPF/TMD [%p] : %d skb RELEASE/FREED\n", tmd, svc);
+
+       /* remove from inuse list  */
+       spin_lock_irqsave (&eip_tx->lock, flags);
+       list_del (&tmd->chain.link);
+       spin_unlock_irqrestore (&eip_tx->lock, flags);
+
+       eip_tmd_put(tmd);
+}
+
+static void eip_tx_tasklet(unsigned long arg)
+{
+       struct timeval now;
+       unsigned long flags;
+       EIP_IPFRAG *ipf, *ipfq = NULL;
+       EP_NMD nmd;
+       struct list_head *list;
+       struct list_head *tmp;
+       char resched = 0;
+       char poll = 1;
+       
+       do_gettimeofday(&now);
+       
+       spin_lock_irqsave(&eip_tx->ipfraglock, flags);
+       if (eip_tx->ipfrag_count) {
+               list_for_each_safe(list, tmp, &eip_tx->ipfrag) {
+                       ipf = list_entry(list, EIP_IPFRAG, list);
+                       /* delta = (((now.tv_sec - ipf->timestamp.tv_sec) * 1000000UL) + now.tv_usec) - ipf->timestamp.tv_usec; */
+                       if (((((now.tv_sec - ipf->timestamp.tv_sec) * 1000000UL) + now.tv_usec) - 
+                                       ipf->timestamp.tv_usec) >= (1000UL * eip_tx->sysctl_ipfrag_to)) {
+                               list_del(&ipf->list);
+                               eip_tx->ipfrag_count--;
+                               ipf->chain.next = (EIP_TMD *) ipfq;
+                               ipfq = ipf;
+                       }
+               }
+       }
+       if (eip_tx->ipfrag_count)
+               resched = 1;
+       spin_unlock_irqrestore(&eip_tx->ipfraglock, flags);
+
+       while (ipfq) {
+               poll = 0;
+
+               ep_nmd_subset(&nmd, &ipfq->nmd, 0, ipfq->dma_len);
+               
+               ipfq->payload.Data[ipfq->frag_nr] = 0;
+               
+#ifdef EIP_MORE_STATS
+               eip_tx->sent_aggreg++;
+#endif
+               ipf = (EIP_IPFRAG *) ipfq->chain.next;
+               eip_do_xmit((EIP_TMD *) ipfq, &nmd, &ipfq->payload);
+               ipfq = ipf;
+       }
+       
+       if (poll)
+               ep_poll_transmits(eip_tx->xmtr);
+
+       if (atomic_read(&eip_tx->destructor) || resched )
+               tasklet_schedule(&eip_tx->tasklet);
+}
+void eip_start_queue()
+{
+       if (netif_queue_stopped(eip_tx->net_device)) {
+               EIP_DBG_PRINTK(EIP_DBG_GEN, "Waking up %s queue\n", eip_tx->net_device->name);
+               netif_wake_queue(eip_tx->net_device);
+       }
+}
+void eip_stop_queue()
+{
+       EIP_DBG_PRINTK(EIP_DBG_GEN, "Stopping %s queue\n", eip_tx->net_device->name);
+       netif_stop_queue(eip_tx->net_device);
+}
+
+static int eip_open(struct net_device *devnet)
+{
+       if (devnet->flags & IFF_PROMISC)
+               EIP_DBG_PRINTK(EIP_DBG_GEN, "%s entering in promiscuous mode\n", devnet->name);
+
+       netif_start_queue(devnet);
+       EIP_DBG_PRINTK(EIP_DBG_GEN, "iface %s MAC %02x:%02x:%02x:%02x:%02x:%02x up\n",
+                       devnet->name, (devnet->dev_addr[0]) & 0xff,
+                       (devnet->dev_addr[1]) & 0xff, (devnet->dev_addr[2]) & 0xff, (devnet->dev_addr[3]) & 0xff,
+                       (devnet->dev_addr[4]) & 0xff, (devnet->dev_addr[5]) & 0xff);
+       return 0;
+}
+
+static int eip_close(struct net_device *devnet)
+{
+       if (devnet->flags & IFF_PROMISC)
+               EIP_DBG_PRINTK(EIP_DBG_GEN, "%s leaving promiscuous mode\n", devnet->name);
+
+       netif_stop_queue(devnet);
+
+       eip_rx_tasklet(0);
+
+       EIP_DBG_PRINTK(EIP_DBG_GEN, "iface %s MAC %02x:%02x:%02x:%02x:%02x:%02x down\n", 
+               devnet->name, (devnet->dev_addr[0]) & 0xff,
+               (devnet->dev_addr[1]) & 0xff, (devnet->dev_addr[2]) & 0xff, (devnet->dev_addr[3]) & 0xff,
+               (devnet->dev_addr[4]) & 0xff, (devnet->dev_addr[5]) & 0xff);
+       return 0;
+}
+
+static struct net_device_stats *eip_get_stats(struct net_device *devnet)
+{
+       static struct net_device_stats stats;
+
+       stats.rx_packets = eip_rx->packets;
+       stats.rx_bytes = eip_rx->bytes;
+       stats.rx_errors = eip_rx->errors;
+       stats.rx_dropped = eip_rx->dropped;
+
+       stats.tx_packets = eip_tx->packets;
+       stats.tx_bytes = eip_tx->bytes;
+       stats.tx_errors = eip_tx->errors;
+       stats.tx_dropped = eip_tx->dropped;
+       return &stats;
+}
+
+static int eip_change_mtu(struct net_device *devnet, int mtu)
+{
+       if (mtu <= EIP_MTU_MAX) {
+               EIP_DBG_PRINTK(EIP_DBG_GEN, "MTU size changed from %d to %d\n", devnet->mtu, mtu);
+               devnet->mtu = mtu;
+       }
+       return 0;
+}
+
+#ifdef MODULE
+int eip_init(void)
+{
+       struct net_device *devnet;
+       int errno = 0;
+
+       eip_rx_dropping = 0; 
+       eip_rx_tasklet_locked = 1;
+
+       /* timer up but not started */
+       init_timer (&eip_rx_tasklet_timer);
+       eip_rx_tasklet_timer.function = eip_rx_tasklet;
+       eip_rx_tasklet_timer.data     = (unsigned long) 0;
+       eip_rx_tasklet_timer.expires  = lbolt + hz;
+
+       devnet = alloc_etherdev(sizeof(EIP_RX) + sizeof(EIP_TX));
+       if (!devnet) {
+               EIP_ERR_PRINTF("Unable to ALLOCATE etherdev structure\n");
+               return -ENOMEM;
+       }
+       strcpy (devnet->name, "eip0");
+
+       EIP_DBG_PRINTK(EIP_DBG_GEN, "Enabling aggregation code\n");
+       devnet->change_mtu = eip_change_mtu;
+       devnet->mtu = EIP_MTU_MAX;
+       devnet->open = eip_open;
+       devnet->stop = eip_close;
+       devnet->hard_start_xmit = eip_hard_start_xmit;
+       devnet->get_stats = eip_get_stats;
+
+        /* devnet->features |= (NETIF_F_DYNALLOC); */
+        /* devnet->features = (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HIGHDMA); */
+        /* devnet->features |= (NETIF_F_SG|NETIF_F_FRAGLIST|NETIF_F_HIGHDMA|NETIF_F_HW_CSUM); */
+
+       eip_rx = (EIP_RX *) devnet->priv;
+       eip_tx = (EIP_TX *) (eip_rx + 1);
+
+       /* instance 0 */
+       eip_tx->ep_system = ep_system();
+       if (eip_tx->ep_system == NULL) {
+               EIP_ERR_PRINTF("kernel comms for iface %s does not exist\n", devnet->name);
+               errno = -ENXIO;
+               goto out;
+       }
+       if (ep_waitfor_nodeid(eip_tx->ep_system) == ELAN_INVALID_NODE) {
+               EIP_ERR_PRINTF("network position not found\n");
+               errno = -EAGAIN;
+               goto out;
+       }
+       eip_tx->xmtr = ep_alloc_xmtr(eip_tx->ep_system);
+       if (!eip_tx->xmtr) {
+               EIP_ERR_PRINTF("Cannot create allocated transmitter - maybe cable is disconnected\n");
+               errno = -EAGAIN;
+               goto out;
+       }
+       /* assign MAC address */
+       *((int *) &devnet->dev_addr[4]) = htons(ep_nodeid(eip_tx->ep_system));
+       eip_rx->net_device = devnet;
+       eip_tx->net_device = devnet;
+
+       atomic_set(&eip_tx->destructor, 0);
+
+       if ((tmd_max >= EIP_TMD_MIN_NR) && (tmd_max <= EIP_TMD_MAX_NR)) {
+               EIP_DBG_PRINTF(EIP_DBG_GEN, "Setting tmd_max_nr to %d\n", tmd_max);
+               eip_tx->tmd_max_nr = tmd_max;
+       } else {
+               EIP_ERR_PRINTF("parameter error : %d <= tmd_max(%d) <= %d using default %d\n", 
+                               EIP_TMD_MIN_NR, tmd_max, EIP_TMD_MAX_NR, EIP_TMD_MAX_NR);
+               eip_tx->tmd_max_nr = EIP_TMD_MAX_NR;
+       }
+
+       if ((rmd_max >= EIP_RMD_MIN_NR) && (rmd_max <= EIP_RMD_MAX_NR)) {
+               EIP_DBG_PRINTF(EIP_DBG_GEN, "Setting rmd_max_nr to %d\n", rmd_max);
+               eip_rx->rmd_max_nr = rmd_max;
+       } else {
+               EIP_ERR_PRINTF("parameter error : %d <= rmd_max(%d) <= %d using default %d\n", EIP_RMD_MIN_NR,
+                          rmd_max, EIP_RMD_MAX_NR, EIP_RMD_MAX_NR);
+               eip_rx->rmd_max_nr = EIP_RMD_MAX_NR;
+       }
+
+       if ((rx_envelope_nr > 0) && (rx_envelope_nr <= 1024)) { /* > 1024 don't be silly */
+               EIP_DBG_PRINTK(EIP_DBG_GEN, "Setting rx_envelope_nr to %d\n", rx_envelope_nr);
+       } else {
+               EIP_ERR_PRINTF("parameter error : 0 < rx_envelope_nr(%d) <= 1024 using default %d\n",
+                          rx_envelope_nr, EIP_RX_ENVELOPE_NR);
+               rx_envelope_nr = EIP_RX_ENVELOPE_NR;
+       }
+
+       if (tx_copybreak_max <= EIP_TX_COPYBREAK_MAX) {
+               EIP_DBG_PRINTF(EIP_DBG_GEN, "Setting tx_copybreak_max to %d\n", tx_copybreak_max);
+       } else {
+               EIP_ERR_PRINTF("parameter error : tx_copybreak_max > %d using default %d\n",
+                          EIP_TX_COPYBREAK_MAX, EIP_TX_COPYBREAK_MAX);
+               tx_copybreak_max = EIP_TX_COPYBREAK_MAX;
+       }
+#ifdef EIP_MORE_STATS
+       eip_tx->sent_copybreak = 0;
+       eip_tx->sent_std = 0;
+       eip_tx->sent_aggreg = 0;
+#endif
+
+       eip_tx->ipfrag_count = 0;
+       eip_aggregation_set(1);
+       eip_rx_granularity_set(rx_granularity);
+       eip_tx_copybreak_set(EIP_TX_COPYBREAK);
+       eip_ipfrag_to_set(EIP_IPFRAG_TO);
+       eip_ipfrag_copybreak_set(EIP_IPFRAG_COPYBREAK);
+
+       spin_lock_init(&eip_tx->lock);
+       spin_lock_init(&eip_tx->ipfraglock);
+       spin_lock_init(&eip_rx->lock);
+       tasklet_init(&eip_rx->tasklet, eip_rx_tasklet, 0);
+       tasklet_init(&eip_tx->tasklet, eip_tx_tasklet, 0);
+       INIT_LIST_HEAD(&eip_tx->ipfrag);
+       INIT_LIST_HEAD(&eip_tx->inuse);
+
+       /* if we fail here cannot do much yet; waiting for rcvr remove code in ep. */
+       errno = eip_tmds_alloc();
+       if (errno)
+               goto out;
+
+       errno = eip_rmds_alloc();
+       if (errno)
+               goto out;
+
+       errno = eip_stats_init();
+       if (errno)
+               goto out;
+
+       if (ep_svc_indicator_set(eip_tx->ep_system, EP_SVC_EIP) != EP_SUCCESS) {
+               EIP_ERR_PRINTF("Cannot set the service indicator\n");
+               errno = -EINVAL;
+               goto out;
+       }
+
+       eip_rx_tasklet_locked = 0;
+       tasklet_schedule(&eip_rx->tasklet);
+
+       SET_MODULE_OWNER(eip_tx->net_device);
+
+       if (register_netdev(devnet)) {
+               printk("eip: failed to register netdev\n");
+               goto out;
+       }
+
+       EIP_DBG_PRINTK(EIP_DBG_GEN, "iface %s MAC %02x:%02x:%02x:%02x:%02x:%02x ready\n", 
+               devnet->name, (devnet->dev_addr[0]) & 0xff,
+               (devnet->dev_addr[1]) & 0xff, (devnet->dev_addr[2]) & 0xff, (devnet->dev_addr[3]) & 0xff,
+               (devnet->dev_addr[4]) & 0xff, (devnet->dev_addr[5]) & 0xff);
+
+       return 0;
+      out:
+       unregister_netdev(devnet);
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 25)
+       kfree(devnet);
+#else
+       free_netdev(devnet);
+#endif
+
+       return errno;
+}
+void eip_exit(void)
+{
+       int i;
+
+       eip_rx_dropping = 1;                /* means that new messages wont be sent to tcp stack */
+       eip_rx_tasklet_locked = 1;
+
+       netif_stop_queue(eip_tx->net_device);
+
+       if (ep_svc_indicator_clear(eip_tx->ep_system, EP_SVC_EIP) != EP_SUCCESS) {
+               EIP_ERR_PRINTF("Cannot unset the service indicator\n");
+       }
+
+       schedule_timeout(10);
+       
+       del_timer_sync (&eip_rx_tasklet_timer);
+
+       tasklet_disable(&eip_rx->tasklet);
+       tasklet_disable(&eip_tx->tasklet);
+
+       tasklet_kill(&eip_tx->tasklet);
+       tasklet_kill(&eip_rx->tasklet);
+
+        eip_rmds_free();
+        eip_tmds_free();
+
+       /* that things freed */
+       for (i = 0 ; i < EIP_SVC_NR ; i++) {
+               if ( EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats) != 0 )
+                       EIP_ERR_PRINTF("%d RMDs not FREED on SVC[%d]\n", EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats), i);
+       }
+       for (i = 0 ; i < 3 ; i++) {
+               if ( EIP_STAT_ALLOC_GET(&eip_tx->head[i].stats) != 0 )
+                       EIP_ERR_PRINTF("%d TMDs not freed on TX HEAD[%d]\n", EIP_STAT_ALLOC_GET(&eip_tx->head[i].stats), i);
+               
+       }
+       unregister_netdev(eip_tx->net_device);
+       kfree(eip_tx->net_device);
+       
+       eip_stats_cleanup();
+}
+
+module_init(eip_init);
+module_exit(eip_exit);
+
+MODULE_PARM(eipdebug, "i");
+MODULE_PARM_DESC(eipdebug, "Set debug flags");
+
+MODULE_PARM(rx_envelope_nr, "i");
+MODULE_PARM_DESC(rx_enveloppe_nr, "Number of allocated enveloppe on the rx side");
+
+MODULE_PARM(tx_copybreak_max, "i");
+MODULE_PARM_DESC(tx_copybreak_max, "Maximum size of the tx copybreak limit (default 512)");
+
+MODULE_PARM(tmd_max, "i");
+MODULE_PARM(rmd_max, "i");
+MODULE_PARM_DESC(tmd_max, "Maximun number of transmit buffers (default 64)");
+MODULE_PARM_DESC(rmd_max, "Maximun number of receive buffers (default 64)");
+
+MODULE_PARM(tx_railmask, "i");
+MODULE_PARM_DESC(tx_railmask, "Mask of which rails transmits can be queued on");
+
+MODULE_AUTHOR("Quadrics Ltd.");
+MODULE_DESCRIPTION("Elan IP driver");
+MODULE_LICENSE("GPL");
+#endif /* MODULE */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/eip/eip_linux.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/eip/eip_linux.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/eip/eip_linux.h      2005-06-01 23:12:54.554445944 -0400
@@ -0,0 +1,399 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: eip_linux.h,v 1.46.2.1 2004/10/01 10:49:38 mike Exp $"
+
+#ifndef __EIP_LINUX_H
+#define __EIP_LINUX_H
+
+#define EIP_WATERMARK                  (0xfab1e)
+
+#define EIP_PAGES(s)                   (((s - 1) >> PAGE_SHIFT) + 1)
+#define EIP_DVMA_PAGES(s)              ((s < PAGE_SIZE) ? EIP_PAGES(s) + 1 : EIP_PAGES(s))
+
+#define EIP_SVC_SMALLEST_LEN           (1 << 9)        /* 512 */
+#define EIP_SVC_BIGGEST_LEN            (1 << 16)       /* 64k */
+
+#define EIP_SVC_SMALLEST               (0)
+#define EIP_SVC_BIGGEST                        (7)
+
+#define EIP_SVC_NR                     (8)
+#define EIP_SVC_EP(s)                  (s + EP_MSG_SVC_EIP512)
+
+#define EIP_STAT_ALLOC_SHIFT           (8)
+#define EIP_STAT_ALLOC_GET(atomicp)    ((int) atomic_read(atomicp) >> EIP_STAT_ALLOC_SHIFT)
+#define EIP_STAT_ALLOC_ADD(atomicp, v) (atomic_add((v << EIP_STAT_ALLOC_SHIFT), atomicp))
+#define EIP_STAT_ALLOC_SUB(atomicp, v) (atomic_sub((v << EIP_STAT_ALLOC_SHIFT), atomicp))
+
+#define EIP_STAT_QUEUED_MASK           (0xff)
+#define EIP_STAT_QUEUED_GET(atomicp)   ((int) atomic_read(atomicp) & EIP_STAT_QUEUED_MASK)
+
+#define EIP_RMD_NR                     (8)
+#define EIP_RMD_MIN_NR                 (8)
+#define EIP_RMD_MAX_NR                 (64)    /* should be < than (1 << EIP_STAT_ALLOC_SHIFT) */
+
+#define EIP_RMD_ALLOC_STEP             (8)
+#define EIP_RMD_ALLOC_THRESH           (16)
+
+#define EIP_RMD_ALLOC                  (1)
+#define EIP_RMD_REPLACE                        (0)
+
+#define EIP_TMD_NR                     (64)
+#define EIP_TMD_MIN_NR                 (16)
+#define EIP_TMD_MAX_NR                 (64)    /* should be < than (1 << EIP_STAT_ALLOC_SHIFT) */
+
+#define EIP_TMD_TYPE_NR                        (3)
+#define EIP_TMD_COPYBREAK              (0x0)
+#define EIP_TMD_STD                    (0x1)
+#define EIP_TMD_AGGREG                 (0x2)
+
+#define EIP_TX_COPYBREAK               (512)
+#define EIP_TX_COPYBREAK_MAX           (1024)
+
+#define EIP_IPFRAG_TO                  (50)    /* time out before a frag is sent in msec */
+#define EIP_IPFRAG_COPYBREAK           (EIP_SVC_BIGGEST_LEN - sizeof(EIP_IPFRAG) - EIP_HEADER_PAD)
+
+#define EIP_RX_ENVELOPE_NR             ((EIP_RMD_MAX_NR*EIP_SVC_NR)/2)
+#define EIP_RX_GRANULARITY             (1)
+
+#define EIP_IP_ALIGN(X)                        (((X) + (15)) & ~(15))
+#define EIP_EXTRA                      roundup (sizeof(EIP_RMD), 256)
+#define EIP_RCV_DMA_LEN(s)                     (s - EIP_EXTRA - EIP_HEADER_PAD)
+#define EIP_MTU_MAX                    (EIP_RCV_DMA_LEN(EIP_SVC_BIGGEST_LEN) - (ETH_HLEN))
+
+#define SIZE_TO_SVC(s, svc)                                                                    \
+       do {                                                                                    \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 9)))  {svc = 0;break;}   \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 10))) {svc = 1;break;}   \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 11))) {svc = 2;break;}   \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 12))) {svc = 3;break;}   \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 13))) {svc = 4;break;}   \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 14))) {svc = 5;break;}   \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 15))) {svc = 6;break;}   \
+                                       if (s <= EIP_RCV_DMA_LEN((1 << 16))) {svc = 7;break;}   \
+                                       svc = -666;                                             \
+                                       EIP_ASSERT(1 == 0);                                     \
+       } while (0)
+
+extern int eipdebug;
+#define EIP_ASSERT_ON 
+/* #define NO_DEBUG */
+
+
+/* ######################## */
+#ifdef NO_DEBUG
+#define __EIP_DBG_PRINTF(fmt, args...)
+#define EIP_DBG_PRINTF(flag, fmt, args...)
+#else
+
+#define EIP_DBG_RMD            0x1
+#define EIP_DBG_TMD            0x2
+#define EIP_DBG_RMD_HEAD       0x4
+#define EIP_DBG_TMD_HEAD       0x8
+#define EIP_DBG_EIPH           0x10
+#define EIP_DBG_IPH            0x20
+#define EIP_DBG_RMD_EP_DVMA    0x40
+#define EIP_DBG_TMD_EP_DVMA    0x80
+#define EIP_DBG_EP_DVMA                (EIP_DBG_RMD_EP_DVMA|EIP_DBG_TMD_EP_DVMA)
+#define EIP_DBG_MEMALLOC       0x100
+#define EIP_DBG_MEMFREE                0x200
+#define EIP_DBG_RMD_QUEUE      0x400
+#define EIP_DBG_TMD_QUEUE      0x800
+#define EIP_DBG_GEN            0x1000
+#define EIP_DBG_DEBUG          0x2000
+       
+#define __EIP_DBG_PRINTF(fmt, args...) (qsnet_debugf (QSNET_DEBUG_BUFFER, " CPU #%d %s: " fmt, smp_processor_id(), __func__, ## args))
+#define EIP_DBG_PRINTF(flag, fmt, args...) (unlikely(eipdebug & flag) ? __EIP_DBG_PRINTF(fmt, ## args):(void)0)
+
+#define __EIP_DBG_PRINTK(fmt, args...) (qsnet_debugf (QSNET_DEBUG_BUF_CON, " CPU #%d %s: " fmt, smp_processor_id(), __func__, ## args))
+#define EIP_DBG_PRINTK(flag, fmt, args...) (unlikely(eipdebug & flag) ? __EIP_DBG_PRINTF(fmt, ## args):(void)0)
+           
+#define EIP_ERR_PRINTF(fmt, args...)   __EIP_DBG_PRINTK("!!! ERROR !!! - " fmt, ## args)
+
+       
+#define EIP_DBG2(flag, fn, fn_arg, fmt, args...)                                                               \
+    if (unlikely(eipdebug & flag)) {                                                                           \
+           qsnet_debugf (QSNET_DEBUG_BUFFER, "+CPU #%d %s: " fmt, smp_processor_id(), __func__, ##args);       \
+            (void)(fn)(fn_arg);                                                                                \
+           qsnet_debugf (QSNET_DEBUG_BUFFER, "-CPU #%d %s: " fmt, smp_processor_id(), __func__, ##args);       \
+    }
+
+
+#define EIP_DBG(flag, fn, args...)                                                             \
+    if (unlikely(eipdebug & flag)) {                                                           \
+           qsnet_debugf (QSNET_DEBUG_BUFFER, "+CPU #%d %s\n", smp_processor_id(), __func__);   \
+            (void)(fn)(args);                                                                  \
+           qsnet_debugf (QSNET_DEBUG_BUFFER, "-CPU #%d %s :\n", smp_processor_id(), __func__); \
+    }
+#endif /* NO_DEBUG */
+
+
+#ifdef EIP_ASSERT_ON
+
+#define __EIP_ASSERT_PRINT(exp)                                \
+               eipdebug = 0xffff;                              \
+               EIP_ERR_PRINTF("ASSERT : %s, %s::%d\n",         \
+                      #exp, __BASE_FILE__, __LINE__);          
+
+#define EIP_ASSERT(exp)                                                        \
+               if (!(exp)) {                                           \
+                       __EIP_ASSERT_PRINT(exp);                        \
+                       netif_stop_queue(eip_tx->net_device);           \
+               }
+
+#define EIP_ASSERT2(exp, f, arg)                                       \
+       do {                                                            \
+               if (!(exp)) {                                           \
+                       __EIP_ASSERT_PRINT(exp);                        \
+                       f(arg);                                         \
+               }                                                       \
+       } while (0)
+
+#define EIP_ASSERT_BUG(exp)                                            \
+       do {                                                            \
+               if (!(exp)) {                                           \
+                       __EIP_ASSERT_PRINT(exp);                        \
+                       BUG();                                          \
+               }                                                       \
+       } while (0)
+
+#define EIP_ASSERT_GOTO(exp, label, f, arg)                            \
+       do {                                                            \
+               if (!(exp)) {                                           \
+                       __EIP_ASSERT_PRINT(exp);                        \
+                       f(arg);                                         \
+                       goto label;                                     \
+               }                                                       \
+       } while (0)
+
+#define EIP_ASSERT_RET(exp, ret)                                       \
+       do {                                                            \
+               if (!(exp)) {                                           \
+                       __EIP_ASSERT_PRINT(exp);                        \
+                       return ret;                                     \
+               }                                                       \
+       } while (0)
+
+#define EIP_ASSERT_RETURN(exp, f, arg)                                 \
+       do {                                                            \
+               if (!(exp)) {                                           \
+                       __EIP_ASSERT_PRINT(exp);                        \
+                       f(arg);                                         \
+                       return;                                         \
+               }                                                       \
+       } while (0)
+
+#define EIP_ASSERT_RETNULL(exp, f, arg)                                        \
+       do {                                                            \
+               if (!(exp)) {                                           \
+                       __EIP_ASSERT_PRINT(exp);                        \
+                       f(arg);                                         \
+                       return NULL;                                    \
+               }                                                       \
+       } while (0)
+
+#else
+
+#define EIP_ASSERT(exp)                do {} while(0)
+#define EIP_ASSERT_OUT(exp)            do {} while(0)
+#define EIP_ASSERT_RETURN(exp)                 do {} while(0)
+#define EIP_ASSERT_RETNULL(exp)                do {} while(0)
+#define EIP_ASSERT_BUG(exp)            do {} while(0)
+
+#endif /* EIP_ASSERT */
+
+
+
+typedef struct {
+       u_short ip_bcast;
+       u_short ip_inst;
+       u_short ip_addr;
+} EIP_ADDRESS;
+
+typedef struct {
+       EIP_ADDRESS h_dhost;
+       EIP_ADDRESS h_shost;
+       u_short h_sap;
+} EIP_HEADER;
+#define EIP_HEADER_PAD                 (2)
+
+typedef struct eip_proc_fs {
+       const char *name;
+       struct proc_dir_entry **parent;
+       read_proc_t *read;
+       write_proc_t *write;
+       unsigned char allocated;
+       struct proc_dir_entry *entry;
+} EIP_PROC_FS;
+
+#define EIP_PROC_ROOT_DIR              "eip"
+
+#define EIP_PROC_DEBUG_DIR             "debug"
+#define EIP_PROC_DEBUG_RX_FLUSH                "rx_flush"
+#define EIP_PROC_DEBUG_TX_FLUSH                "tx_flush"
+
+#define EIP_PROC_AGGREG_DIR            "aggregation"
+#define EIP_PROC_AGGREG_ONOFF          "enable"
+#define EIP_PROC_AGGREG_TO             "timeout"
+#define EIP_PROC_AGGREG_COPYBREAK      "copybreak"
+
+#define EIP_PROC_TX_COPYBREAK          "tx_copybreak"
+#define EIP_PROC_STATS                 "stats"
+#define EIP_PROC_RX_GRAN               "rx_granularity"
+#define EIP_PROC_TX_RAILMASK           "tx_railmask"
+#define EIP_PROC_TMD_INUSE             "tmd_inuse"
+#define EIP_PROC_EIPDEBUG              "eipdebug"
+#define EIP_PROC_CHECKSUM               "checksum"
+
+/* RX */
+/* dma_len is used to keep the len of a received packet */
+/* nmd.nmd_len is the max dma that can be received      */
+/*                                                      */
+struct eip_rmd {
+       struct sk_buff *skb;
+
+       EP_NMD nmd;
+       u16 dvma_idx;
+
+       EP_RXD *rxd;
+       struct eip_rmd_head *head;
+       union {
+               struct list_head link;                          /* when on "busy" list */
+               struct eip_rmd  *next;                          /* all other lists */
+       } chain;
+};
+typedef struct eip_rmd EIP_RMD;
+struct eip_rmd_head {
+       EP_NMH *handle;
+
+       EP_RCVR *rcvr;
+       EIP_RMD *busy_list;
+
+       /* stats */
+       atomic_t stats;
+       unsigned long dma;
+};
+
+typedef struct eip_rmd_head EIP_RMD_HEAD;
+typedef struct eip_rx {
+       struct eip_rmd_head head[EIP_SVC_NR];
+
+       EIP_RMD *irq_list;
+       short    irq_list_nr;   
+
+       /* stats */
+       unsigned long packets;
+       unsigned long bytes;
+       unsigned long errors;
+       unsigned long dropped;
+       unsigned long reschedule;
+
+       spinlock_t lock;
+       struct tasklet_struct tasklet;
+       unsigned char rmd_max_nr;
+       unsigned char sysctl_granularity;
+       struct net_device *net_device;
+} EIP_RX;
+
+/* TX */
+/* dma_len_max is the maximum len for a given DMA                      */
+/* where mnd.nmd_len is the len of the packet to send ~> than skb->len */
+typedef struct eip_ipfrag_handle {
+       /* common with tmd */
+       unsigned long dma_base;
+       int dma_len;
+       EP_NMD nmd;
+       u16 dvma_idx;
+
+       struct sk_buff *skb;
+       struct eip_tmd_head *head;
+       union {
+               struct list_head link;                          /* when on "busy" list */
+               struct eip_tmd  *next;                          /* all other lists */
+       } chain;
+
+       /* private */
+       struct list_head list;
+       struct timeval timestamp;
+       unsigned int frag_nr;
+       int datagram_len; /* Ip data */
+       int dma_correction;
+       EP_PAYLOAD payload;
+} EIP_IPFRAG;
+
+struct eip_tmd {
+       unsigned long dma_base;
+       int dma_len;
+       EP_NMD nmd;
+       u16 dvma_idx;
+
+       struct sk_buff *skb;
+       struct eip_tmd_head *head;
+       union {
+               struct list_head link;                          /* when on "busy" list */
+               struct eip_tmd  *next;                          /* all other lists */
+       } chain;
+};
+
+struct eip_tmd_head {
+       EP_NMH *handle;
+
+       struct eip_tmd *tmd;
+       atomic_t stats;
+};
+
+typedef struct eip_tmd EIP_TMD;
+typedef struct eip_tmd_head EIP_TMD_HEAD;
+
+/* #define EIP_MORE_STATS */
+
+typedef struct eip_tx {
+       struct net_device *net_device;
+       EP_XMTR *xmtr;
+       EP_SYS *ep_system;
+
+       struct eip_tmd_head head[EIP_TMD_TYPE_NR];
+       struct list_head inuse;
+       atomic_t destructor;
+
+       /* stats */
+       unsigned long packets;
+       unsigned long bytes;
+       unsigned long errors;
+       unsigned long dropped;
+       unsigned long dma[EIP_SVC_NR];
+       
+#ifdef EIP_MORE_STATS
+       unsigned long sent_copybreak;
+       unsigned long sent_std;
+       unsigned long sent_aggreg;
+#endif
+
+       unsigned char tmd_max_nr;
+
+       unsigned short sysctl_copybreak;
+       unsigned short sysctl_ipfrag_to;
+       unsigned short sysctl_ipfrag_copybreak;
+       unsigned short sysctl_aggregation;
+
+       unsigned short ipfrag_count;
+       struct list_head ipfrag;
+       spinlock_t ipfraglock;
+
+       spinlock_t lock;
+       struct tasklet_struct tasklet;
+} EIP_TX;
+
+/* =============================================== */
+    /* unsigned long   multicast; */
+#endif                         /* __EIP_LINUX_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/eip/eip_stats.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/eip/eip_stats.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/eip/eip_stats.c      2005-06-01 23:12:54.555445792 -0400
@@ -0,0 +1,374 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+/*
+ * $Id: eip_stats.c,v 1.34.2.1 2005/01/26 14:31:56 mike Exp $
+ * $Source: /cvs/master/quadrics/eipmod/eip_stats.c,v $
+ */
+
+#include <qsnet/kernel.h>
+#include <linux/module.h>
+
+#include <elan/epcomms.h>
+
+#include <linux/netdevice.h>
+
+#include <linux/kernel.h>
+#include <linux/proc_fs.h>
+
+#include <asm/atomic.h>
+
+#include <qsnet/procfs_linux.h>
+
+#include "eip_linux.h"
+#include "eip_stats.h"
+
+extern EIP_RX *eip_rx;
+extern EIP_TX *eip_tx;
+extern int tx_copybreak_max;
+extern EP_RAILMASK tx_railmask;
+extern int  eip_checksum_state;
+extern void eip_stop_queue(void);
+extern void eip_start_queue(void);
+
+static int eip_stats_read(char *buf, char **start, off_t off, int count, int *eof, void *data)
+{
+       int i, outlen = 0;
+
+       *buf = '\0';
+       strcat(buf, "\n");
+       strcat(buf, "--------------------------------------------+------------+-----------------+\n");
+       strcat(buf, "    SKB/DMA    |               | Rx         | Tx         |  TMD TYPE       |\n");
+       strcat(buf, "--------------------------------------------+------------|-----------------+\n");
+
+       i = 0;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld | #1[%3.3d/%3.3d/%3.3d] |\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i],
+               EIP_STAT_QUEUED_GET(&eip_tx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_tx->head[i].stats),
+                eip_tx->tmd_max_nr);
+
+       i++;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld | #2[%3.3d/%3.3d/%3.3d] |\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i],
+               EIP_STAT_QUEUED_GET(&eip_tx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_tx->head[i].stats),
+               eip_tx->tmd_max_nr);
+
+       i++;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld | #3[%3.3d/%3.3d/%3.3d] |\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i],
+               EIP_STAT_QUEUED_GET(&eip_tx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_tx->head[i].stats),
+               eip_tx->tmd_max_nr);
+
+       i++;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld +-----------------+\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i]);
+
+       i++;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld |\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i]);
+
+       i++;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld |\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i]);
+
+       i++;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld |\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i]);
+
+       i++;
+       sprintf(buf + strlen(buf), " [%5d/%5d] | [%3.3d/%3.3d/%3.3d] | %10ld | %10ld |\n",
+               EIP_SVC_SMALLEST_LEN << i, (int) EIP_RCV_DMA_LEN((EIP_SVC_SMALLEST_LEN << i)),
+               EIP_STAT_QUEUED_GET(&eip_rx->head[i].stats), EIP_STAT_ALLOC_GET(&eip_rx->head[i].stats),
+               eip_rx->rmd_max_nr, eip_rx->head[i].dma, eip_tx->dma[i]);
+
+       strcat(buf, "--------------------------------------------+------------+\n");
+       sprintf(buf + strlen(buf), " RMD IRQ %4.4d                    %10lu | %10lu |\n",
+               eip_rx->irq_list_nr, 
+               eip_rx->packets, eip_tx->packets);
+       strcat(buf, "--------------------------------------------+------------+\n");
+
+#ifdef EIP_MORE_STATS
+       strcat(buf, "\n");
+       sprintf(buf + strlen(buf), " Copybreak %10ld Std %10ld Aggreg %10ld\n",
+                       eip_tx->sent_copybreak, eip_tx->sent_std, eip_tx->sent_aggreg);
+#endif
+
+
+       strcat(buf, "\n");
+       sprintf(buf + strlen(buf), "Rx bytes: %lu (%lu Mb) errors: %lu dropped: %lu reschedule: %lu\n",
+               eip_rx->bytes, eip_rx->bytes / (1024 * 1024), eip_rx->errors, eip_rx->dropped, eip_rx->reschedule);
+       sprintf(buf + strlen(buf), "Tx bytes: %lu (%lu Mb) errors: %lu dropped: %lu\n",
+               eip_tx->bytes, eip_tx->bytes / (1024 * 1024), eip_tx->errors, eip_tx->dropped);
+       strcat(buf, "\n");
+
+       outlen = strlen(buf);
+       ASSERT(outlen < PAGE_SIZE);
+       *eof = 1;
+       return outlen;
+}
+
+void eip_stats_dump(void)
+{
+    int eof;
+
+    char *buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
+
+    if (buf == NULL)
+    {
+       printk("no memory to produce eip_stats\n");
+       return;
+    }
+
+    eip_stats_read(buf, NULL, 0, 0, &eof, NULL);
+
+    printk(buf);
+
+    kfree(buf);
+}
+
+static int eip_stats_write(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       int i;
+       unsigned long flags;
+
+       spin_lock_irqsave(&eip_rx->lock, flags);
+       eip_rx->packets = 0;
+       eip_rx->bytes = 0;
+       eip_rx->errors = 0;
+       eip_rx->dropped = 0;
+       eip_rx->reschedule = 0;
+       for (i = 0; i < EIP_SVC_NR; eip_rx->head[i].dma = 0, i++);
+       spin_unlock_irqrestore(&eip_rx->lock, flags);
+
+       spin_lock_irqsave(&eip_tx->lock, flags);
+       eip_tx->packets = 0;
+       eip_tx->bytes = 0;
+       eip_tx->errors = 0;
+       eip_tx->dropped = 0;
+#ifdef EIP_MORE_STATS
+       eip_tx->sent_copybreak = 0;
+       eip_tx->sent_std = 0;
+       eip_tx->sent_aggreg = 0;
+#endif
+       for (i = 0; i < EIP_SVC_NR; eip_tx->dma[i] = 0, i++);
+       spin_unlock_irqrestore(&eip_tx->lock, flags);
+
+       return count;
+}
+
+#define                eip_stats_var_write(name)                                                                       \
+static int eip_stats_##name##_write(struct file *file, const char *buf, unsigned long count, void *data)       \
+{                                                                                                              \
+       char * b = (char *) buf;                                                                                \
+       *(b + count) = '\0';                                                                                    \
+       eip_##name##_set((int) simple_strtoul(b, NULL, 10));                                                    \
+       return count;                                                                                           \
+}
+
+#define        eip_stats_var_read(name, var)                                                                   \
+static int eip_stats_##name##_read(char *buf, char **start, off_t off, int count, int *eof, void *data)                \
+{                                                                                                              \
+       sprintf(buf, "%d\n", var);                                                                              \
+       *eof = 1;                                                                                               \
+       return strlen(buf);                                                                                     \
+}
+
+
+#define                eip_stats_var_set(name, min, max, default, var)                                                                 \
+void eip_##name##_set(int i)                                                                                                   \
+{                                                                                                                              \
+       if ( (i >= min) && (i <= max)) {                                                                                        \
+               EIP_DBG_PRINTK(EIP_DBG_GEN, "Setting " #name " to %d\n", i);                                                    \
+               var =(unsigned short) i;                                                                                        \
+       }                                                                                                                       \
+       else {                                                                                                                  \
+               EIP_ERR_PRINTF("parameter error : %d <= " #name "(%d) <= %d using default %d\n", min, i, (int) max, (int) default);     \
+       }                                                                                                                       \
+}
+
+eip_stats_var_set(tx_copybreak, 0, tx_copybreak_max, EIP_TX_COPYBREAK, eip_tx->sysctl_copybreak);
+eip_stats_var_set(rx_granularity, 1, EIP_RMD_MIN_NR, EIP_RX_GRANULARITY, eip_rx->sysctl_granularity);
+eip_stats_var_set(tx_railmask, 0, EP_RAILMASK_ALL, EP_RAILMASK_ALL, tx_railmask);
+eip_stats_var_set(ipfrag_to, 0, (1 << 16), EIP_IPFRAG_TO, eip_tx->sysctl_ipfrag_to);
+eip_stats_var_set(aggregation, 0, 1, 1, eip_tx->sysctl_aggregation);
+eip_stats_var_set(ipfrag_copybreak, 0, EIP_IPFRAG_COPYBREAK, EIP_IPFRAG_COPYBREAK, eip_tx->sysctl_ipfrag_copybreak);
+/* eip_stats_var_set(eipdebug, 0, , 0, eipdebug); */
+
+eip_stats_var_read(aggregation, eip_tx->sysctl_aggregation);
+eip_stats_var_read(ipfrag_count, eip_tx->ipfrag_count);
+eip_stats_var_read(ipfrag_to, eip_tx->sysctl_ipfrag_to);
+eip_stats_var_read(ipfrag_copybreak, eip_tx->sysctl_ipfrag_copybreak);
+eip_stats_var_read(tx_copybreak, eip_tx->sysctl_copybreak);
+eip_stats_var_read(rx_granularity, eip_rx->sysctl_granularity);
+eip_stats_var_read(tx_railmask, tx_railmask);
+
+eip_stats_var_write(aggregation);
+eip_stats_var_write(ipfrag_to);
+eip_stats_var_write(ipfrag_copybreak);
+eip_stats_var_write(tx_copybreak);
+eip_stats_var_write(rx_granularity);
+eip_stats_var_write(tx_railmask);
+
+
+static int eip_checksum_write(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       char * b = (char *) buf;
+       int    value;
+
+       *(b + count) = '\0';
+
+       value = (int) simple_strtoul(b, NULL, 10);
+       if  ((value >= CHECKSUM_NONE) && (value <= CHECKSUM_UNNECESSARY)) 
+               eip_checksum_state = value;
+       else 
+               EIP_ERR_PRINTF("%d <= checksum(%d) <= %d using old value %d\n", CHECKSUM_NONE, value, CHECKSUM_UNNECESSARY, eip_checksum_state);
+
+       return count;
+}
+
+static int eip_checksum_read(char *buf, char **start, off_t off, int count, int *eof, void *data)
+{
+       switch ( eip_checksum_state ) 
+       {
+       case 0  : sprintf(buf, "0 CHECKSUM_NONE\n");                      break;
+       case 1  : sprintf(buf, "1 CHECKSUM_HW\n");                        break;
+       case 2  : sprintf(buf, "2 CHECKSUM_UNNECESSARY\n");               break;
+       default : sprintf(buf, "%d INVALID VALUE\n", eip_checksum_state); break;
+       }
+       *eof = 1;
+       return strlen(buf);
+}
+
+static int eip_stats_eipdebug_read(char *buf, char **start, off_t off, int count, int *eof, void *data)
+{
+       *buf = '\0';
+       sprintf(buf + strlen(buf), "0x%x\n", eipdebug);
+       *eof = 1;
+       return strlen(buf);
+}
+static int eip_stats_eipdebug_write(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       char * p = (char *) buf;
+       *(p + count - 1) = '\0';
+       eipdebug = simple_strtoul(p, NULL, 0);
+       __EIP_DBG_PRINTK("Setting eipdebug to 0x%x\n", eipdebug);
+       return count;
+}
+
+static int eip_stats_tmd_inuse_read(char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       struct list_head *lp;
+       unsigned long flags;
+       unsigned int len = 0;
+
+       spin_lock_irqsave(&eip_tx->lock, flags);
+       list_for_each (lp, &eip_tx->inuse) {
+               EIP_TMD *tmd = list_entry (lp, EIP_TMD, chain.link);
+               EIP_HEADER *eiph = (EIP_HEADER *) tmd->dma_base;
+               
+                len += sprintf(page+len, "tmd=%p id=%d len=%d\n",
+                              tmd, eiph ? ntohs(eiph->h_dhost.ip_addr) : -1,
+                              tmd->dma_len);
+
+                if (len + 40 >= count)
+                        break;
+        }
+        spin_unlock_irqrestore(&eip_tx->lock, flags);
+
+       return qsnet_proc_calc_metrics (page, start, off, count, eof, len);
+}
+
+static int eip_stats_debug_rx_flush(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       EIP_DBG_PRINTF(EIP_DBG_GEN, "Flushing rx ...\n");
+       tasklet_schedule(&eip_rx->tasklet);
+       return count;
+}
+static int eip_stats_debug_tx_flush(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       EIP_DBG_PRINTF(EIP_DBG_GEN, "Flushing tx ... %d tmds reclaimed\n", ep_enable_txcallbacks(eip_tx->xmtr));
+       ep_disable_txcallbacks(eip_tx->xmtr);
+       tasklet_schedule(&eip_tx->tasklet);
+       return count;
+}
+
+#define EIP_PROC_PARENT_NR     (3)
+/* NOTE : the parents should be declared b4 the children */
+static EIP_PROC_FS eip_procs[] = {
+       /* {name, parent, read fn, write fn, allocated, entry}, */
+       {EIP_PROC_ROOT_DIR, &qsnet_procfs_root, NULL, NULL, 0, NULL},
+       {EIP_PROC_DEBUG_DIR, &eip_procs[0].entry, NULL, NULL, 0, NULL},
+       {EIP_PROC_AGGREG_DIR, &eip_procs[0].entry, NULL, NULL, 0, NULL},        /* end of parents */
+       {EIP_PROC_STATS, &eip_procs[0].entry, eip_stats_read, eip_stats_write, 0, NULL},
+       {EIP_PROC_TX_COPYBREAK, &eip_procs[0].entry, eip_stats_tx_copybreak_read, eip_stats_tx_copybreak_write, 0, NULL},
+       {EIP_PROC_RX_GRAN, &eip_procs[0].entry, eip_stats_rx_granularity_read, eip_stats_rx_granularity_write, 0, NULL},
+       {EIP_PROC_TX_RAILMASK, &eip_procs[0].entry, eip_stats_tx_railmask_read, eip_stats_tx_railmask_write, 0, NULL},
+       {EIP_PROC_TMD_INUSE, &eip_procs[0].entry, eip_stats_tmd_inuse_read, NULL, 0, NULL},
+       {EIP_PROC_EIPDEBUG, &eip_procs[0].entry, eip_stats_eipdebug_read, eip_stats_eipdebug_write, 0, NULL},
+       {EIP_PROC_CHECKSUM, &eip_procs[0].entry, eip_checksum_read, eip_checksum_write, 0, NULL},
+       {EIP_PROC_DEBUG_RX_FLUSH, &eip_procs[1].entry, NULL, eip_stats_debug_rx_flush, 0, NULL},
+       {EIP_PROC_DEBUG_TX_FLUSH, &eip_procs[1].entry, NULL, eip_stats_debug_tx_flush, 0, NULL},
+       {"ipfrag_count", &eip_procs[2].entry, eip_stats_ipfrag_count_read, NULL, 0, NULL},
+       {EIP_PROC_AGGREG_TO, &eip_procs[2].entry, eip_stats_ipfrag_to_read, eip_stats_ipfrag_to_write, 0, NULL},
+       {EIP_PROC_AGGREG_ONOFF, &eip_procs[2].entry, eip_stats_aggregation_read, eip_stats_aggregation_write, 0, NULL},
+       {EIP_PROC_AGGREG_COPYBREAK, &eip_procs[2].entry, eip_stats_ipfrag_copybreak_read, eip_stats_ipfrag_copybreak_write, 0, NULL},
+       {NULL, NULL, NULL, NULL, 1, NULL},
+};
+
+int eip_stats_init(void)
+{
+       int p;
+
+       for (p = 0; !eip_procs[p].allocated; p++) {
+               if (p < EIP_PROC_PARENT_NR)
+                       eip_procs[p].entry = proc_mkdir(eip_procs[p].name, *eip_procs[p].parent);
+               else
+                       eip_procs[p].entry = create_proc_entry(eip_procs[p].name, 0, *eip_procs[p].parent);
+
+               if (!eip_procs[p].entry) {
+                       EIP_ERR_PRINTF("%s\n", "Cannot allocate proc entry");
+                       eip_stats_cleanup();
+                       return -ENOMEM;
+               }
+
+               eip_procs[p].entry->owner = THIS_MODULE;
+               eip_procs[p].entry->write_proc = eip_procs[p].write;
+               eip_procs[p].entry->read_proc = eip_procs[p].read;
+               eip_procs[p].allocated = 1;
+       }
+       eip_procs[p].allocated = 0;
+       return 0;
+}
+
+void eip_stats_cleanup(void)
+{
+       int p;
+       for (p = 0; eip_procs[p].allocated; p++) {
+               EIP_DBG_PRINTF(EIP_DBG_GEN, "Removing %s from proc\n", eip_procs[p].name);
+               remove_proc_entry(eip_procs[p].name, *eip_procs[p].parent);
+       }
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/eip/eip_stats.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/eip/eip_stats.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/eip/eip_stats.h      2005-06-01 23:12:54.555445792 -0400
@@ -0,0 +1,22 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: eip_stats.h,v 1.14 2004/05/10 14:47:47 daniel Exp $"
+
+#ifndef __EIP_STATS_H
+#define        __EIP_STATS_H
+
+int eip_stats_init(void);
+void eip_stats_cleanup(void);
+void eip_rx_granularity_set(int);
+void eip_tx_copybreak_set(int);
+void eip_ipfrag_to_set(int);
+void eip_aggregation_set(int);
+void eip_ipfrag_copybreak_set(int);
+void eip_stats_dump(void);
+
+#endif                         /* __EIP_STATS_H */
Index: linux-2.4.21/drivers/net/qsnet/eip/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/eip/Makefile    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/eip/Makefile 2005-06-01 23:12:54.555445792 -0400
@@ -0,0 +1,31 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/eip/Makefile
+#
+
+
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/eip/Makefile
+#
+
+list-multi             := eip.o
+eip-objs       := eip_linux.o eip_stats.o
+export-objs            := 
+obj-$(CONFIG_EIP)      := eip.o
+
+eip.o : $(eip-objs)
+       $(LD) -r -o $@ $(eip-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/eip/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/eip/Makefile.conf       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/eip/Makefile.conf    2005-06-01 23:12:54.555445792 -0400
@@ -0,0 +1,10 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       eip.o
+MODULENAME     =       eip
+KOBJFILES      =       eip_linux.o eip_stats.o
+EXPORT_KOBJS   =       
+CONFIG_NAME    =       CONFIG_EIP
+SGALFC         =       
+# EXTRALINES START
+
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/eip/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/eip/quadrics_version.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/eip/quadrics_version.h       2005-06-01 23:12:54.556445640 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/elan/bitmap.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/bitmap.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/bitmap.c        2005-06-01 23:12:54.556445640 -0400
@@ -0,0 +1,287 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: bitmap.c,v 1.5 2004/01/20 17:32:17 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/shared/bitmap.c,v $*/
+
+#if defined(__KERNEL__)
+#include <qsnet/kernel.h>
+#endif
+#include <qsnet/config.h>
+#include <elan/bitmap.h>
+
+/*
+ * Return the index of the first available bit in the 
+ * bitmap , or -1 for failure
+ */
+int
+bt_freebit (bitmap_t *bitmap, int nbits)
+{
+    int last = (--nbits) >> BT_ULSHIFT;
+    int maxbit;
+    int        i, j;
+
+    /* look for a word with a bit off */
+    for (i = 0; i <= last; i++)
+       if (bitmap[i] != ~((bitmap_t) 0))
+           break;
+
+    if (i <= last)
+    {
+       /* found an word with a bit off,  now see which bit it is */
+       maxbit = (i == last) ? (nbits & BT_ULMASK) : (BT_NBIPUL-1);
+       for (j = 0; j <= maxbit; j++)
+           if ((bitmap[i] & (1 << j)) == 0)
+               return ((i << BT_ULSHIFT) | j);
+    }
+    return (-1);
+    
+}
+
+/*
+ * bt_lowbit:
+ *     Return the index of the lowest set bit in the
+ *     bitmap, or -1 for failure.
+ */
+int
+bt_lowbit (bitmap_t *bitmap, int nbits)
+{
+    int last = (--nbits) >> BT_ULSHIFT;
+    int maxbit;
+    int i, j;
+    
+    /* look for a word with a bit on */
+    for (i = 0; i <= last; i++)
+       if (bitmap[i] != 0)
+           break;
+    if (i <= last)
+    {
+       /* found a word bit a bit on, now see which bit it is */
+       maxbit = (i == last) ? (nbits & BT_ULMASK) : (BT_NBIPUL-1);
+       for (j = 0; j <= maxbit; j++)
+           if (bitmap[i] & (1 << j))
+               return ((i << BT_ULSHIFT) | j);
+    }
+
+    return (-1);
+}
+
+/*
+ * Return the index of the first available bit in the 
+ * bitmap , or -1 for failure
+ */
+int
+bt_nextbit (bitmap_t *bitmap, int nbits, int last, int isset)
+{
+    int first = ((last+1) + BT_NBIPUL-1) >> BT_ULSHIFT;
+    int end   = (--nbits) >> BT_ULSHIFT;
+    int maxbit;
+    int        i, j;
+
+    /* look for bits before the first whole word */
+    if (((last+1) & BT_ULMASK) != 0)
+    {
+       maxbit = ((first-1) == last) ? (nbits & BT_ULMASK) : (BT_NBIPUL-1);
+       for (j = ((last+1) & BT_ULMASK); j <= maxbit; j++)
+           if ((bitmap[first-1] & (1 << j)) == (isset << j))
+               return (((first-1) << BT_ULSHIFT) | j);
+    }
+
+    /* look for a word with a bit off */
+    for (i = first; i <= end; i++)
+       if (bitmap[i] != (isset ? 0 : ~((bitmap_t) 0)))
+           break;
+
+    if (i <= end)
+    {
+       /* found an word with a bit off,  now see which bit it is */
+       maxbit = (i == end) ? (nbits & BT_ULMASK) : (BT_NBIPUL-1);
+       for (j = 0; j <= maxbit; j++)
+           if ((bitmap[i] & (1 << j)) == (isset << j))
+               return ((i << BT_ULSHIFT) | j);
+    }
+    return (-1);
+}
+
+void
+bt_copy (bitmap_t *a, bitmap_t *b, int nbits)
+{
+    int i;
+
+    for (i = 0; i < (nbits>>BT_ULSHIFT); i++)
+       b[i] = a[i];
+
+    for (i <<= BT_ULSHIFT; i < nbits; i++)
+       if (BT_TEST(a, i))
+           BT_SET(b,i);
+       else
+           BT_CLEAR(b,i);
+}
+
+void
+bt_zero (bitmap_t *bitmap, int nbits)
+{
+    int i;
+
+    for (i = 0; i < (nbits>>BT_ULSHIFT); i++)
+       bitmap[i] = 0;
+
+    for (i <<= BT_ULSHIFT; i < nbits; i++)
+       BT_CLEAR(bitmap,i);
+}
+
+void
+bt_fill (bitmap_t *bitmap, int nbits)
+{
+    int i;
+
+    for (i = 0; i < (nbits>>BT_ULSHIFT); i++)
+       bitmap[i] = ~((bitmap_t) 0);
+
+    for (i <<= BT_ULSHIFT; i < nbits; i++)
+       BT_SET(bitmap,i);
+}
+
+int
+bt_cmp (bitmap_t *a, bitmap_t *b, int nbits)
+{
+    int i;
+
+    for (i = 0; i < (nbits>>BT_ULSHIFT); i++)
+       if (a[i] != b[i])
+           return (1);
+
+    for (i <<= BT_ULSHIFT; i < nbits; i++)
+       if (BT_TEST (a, i) != BT_TEST(b, i))
+           return (1);
+    return (0);
+}
+
+void
+bt_intersect (bitmap_t *a, bitmap_t *b, int nbits)
+{
+    int i;
+    
+    for (i = 0; i < (nbits>>BT_ULSHIFT); i++)
+       a[i] &= b[i];
+
+    for (i <<= BT_ULSHIFT; i < nbits; i++)
+       if (BT_TEST (a, i) && BT_TEST (b, i))
+           BT_SET (a, i);
+       else
+           BT_CLEAR (a, i);
+}
+
+void
+bt_remove (bitmap_t *a, bitmap_t *b, int nbits)
+{
+    int i;
+
+    for (i = 0; i < (nbits>>BT_ULSHIFT); i++)
+       a[i] &= ~b[i];
+
+    for (i <<= BT_ULSHIFT; i < nbits; i++)
+       if (BT_TEST (b, i))
+           BT_CLEAR (a, i);
+}
+
+void
+bt_add (bitmap_t *a, bitmap_t *b, int nbits)
+{
+    int i;
+
+    for (i = 0; i < (nbits>>BT_ULSHIFT); i++)
+       a[i] |= b[i];
+
+    for (i <<= BT_ULSHIFT; i < nbits; i++)
+       if (BT_TEST(b, i))
+           BT_SET (a, i);
+}
+
+/*
+ * bt_spans : partition a spans partition b
+ *    == all bits set in 'b' are set in 'a'
+ */
+int
+bt_spans (bitmap_t *a, bitmap_t *b, int nbits)
+{
+    int i;
+    
+    for (i = 0; i < nbits; i++)
+       if (BT_TEST (b, i) && !BT_TEST (a, i))
+           return (0);
+    return (1);
+}
+
+/*
+ * bt_subset: copy [base,base+nbits-1] from 'a' to 'b'
+ */
+void
+bt_subset (bitmap_t *a, bitmap_t *b, int base, int nbits)
+{
+    int i;
+
+    for (i = 0; i < nbits; i++)
+    {
+       if (BT_TEST (a, base+i))
+           BT_SET(b,i);
+       else
+           BT_CLEAR (b,i);
+    }
+}
+
+void 
+bt_up (bitmap_t *a, bitmap_t *b, bitmap_t *c, int nbits)
+{
+    int i;
+    
+    for (i = 0; i < nbits; i++)
+    {
+       if (!BT_TEST (a, i) && BT_TEST (b, i))
+       {
+           BT_SET (c, i);
+        }
+       else
+       {
+           BT_CLEAR (c, i);
+        }
+    }
+}
+
+void 
+bt_down (bitmap_t *a, bitmap_t *b, bitmap_t *c, int nbits)
+{
+    int i;
+    
+    for (i = 0; i < nbits; i++)
+    {
+       if (BT_TEST (a, i) && !BT_TEST (b, i))
+       {
+           BT_SET (c, i);
+        }
+       else
+       {
+           BT_CLEAR (c, i);
+        }
+    }
+}
+
+int
+bt_nbits (bitmap_t *a, int nbits)
+{
+    int i, c;
+    for (i = 0, c = 0; i < nbits; i++)
+       if (BT_TEST (a, i))
+           c++;
+    return (c);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan/capability.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/capability.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/capability.c    2005-06-01 23:12:54.557445488 -0400
@@ -0,0 +1,628 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: capability.c,v 1.13 2004/07/20 10:15:33 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/capability.c,v $ */
+
+
+#include <qsnet/kernel.h>
+#include <elan/elanmod.h>
+
+static LIST_HEAD(elan_cap_list); 
+
+typedef struct elan_vp_struct
+{
+       struct list_head list;
+       ELAN_CAPABILITY  vp;
+} ELAN_VP_NODE_STRUCT;
+
+
+typedef struct elan_attached_struct
+{
+       void               *cb_args;
+       ELAN_DESTROY_CB  cb_func;
+} ELAN_ATTACHED_STRUCT;
+
+typedef struct elan_cap_node_struct
+{
+       struct list_head list;
+       ELAN_CAP_STRUCT     node;
+       ELAN_ATTACHED_STRUCT *attached[ELAN_MAX_RAILS];
+       struct list_head vp_list;
+} ELAN_CAP_NODE_STRUCT;
+
+
+ELAN_CAP_NODE_STRUCT *
+find_cap_node(ELAN_CAPABILITY *cap)
+{
+       struct list_head        *tmp;
+       ELAN_CAP_NODE_STRUCT *ptr=NULL;
+
+       list_for_each(tmp, &elan_cap_list) {
+               ptr = list_entry(tmp, ELAN_CAP_NODE_STRUCT , list);
+               /* is it an exact match */
+               if ( ELAN_CAP_TYPE_MATCH(&ptr->node.cap,cap) 
+                    && ELAN_CAP_GEOM_MATCH(&ptr->node.cap,cap)) {
+                       return ptr;
+               }
+       }
+       return ptr;
+};
+
+ELAN_VP_NODE_STRUCT *
+find_vp_node( ELAN_CAP_NODE_STRUCT *cap_node,ELAN_CAPABILITY *map)
+{
+       struct list_head       * tmp;
+       ELAN_VP_NODE_STRUCT * ptr = NULL;
+
+       list_for_each(tmp, &cap_node->vp_list) {
+               ptr = list_entry(tmp, ELAN_VP_NODE_STRUCT , list);
+               /* is it an exact match */
+               if ( ELAN_CAP_TYPE_MATCH(&ptr->vp,map) 
+                    && ELAN_CAP_GEOM_MATCH(&ptr->vp,map)){
+                       return ptr;
+               }
+       }
+       return ptr;
+}
+
+int 
+elan_validate_cap(ELAN_CAPABILITY *cap)
+{
+       char                      space[127];
+
+       ELAN_DEBUG1 (ELAN_DBG_VP,"elan_validate_cap %s\n",elan_capability_string(cap,space));
+
+       /* check versions */
+       if (cap->cap_version != ELAN_CAP_VERSION_NUMBER)
+       {
+               ELAN_DEBUG2 (ELAN_DBG_VP,"elan_validate_cap: (cap->Version != ELAN_CAP_VERSION) %d %d\n", cap->cap_version, ELAN_CAP_VERSION_NUMBER);
+               return (EINVAL);
+       }
+
+       /* check its not HWTEST */
+       if ( cap->cap_type & ELAN_CAP_TYPE_HWTEST )
+       {
+               ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_cap: failed type = ELAN_CAP_TYPE_HWTEST \n");   
+               return (EINVAL);
+       }
+       
+       /* check its type */
+       switch (cap->cap_type & ELAN_CAP_TYPE_MASK)
+       {
+       case ELAN_CAP_TYPE_KERNEL :     
+               ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_cap: failed type = ELAN_CAP_TYPE_KERNEL \n");   
+               return (EINVAL);
+
+               /* check it has a valid type */
+       case ELAN_CAP_TYPE_BLOCK:
+       case ELAN_CAP_TYPE_CYCLIC:
+               break;
+
+               /* all others are failed as well */
+       default:
+               ELAN_DEBUG1 (ELAN_DBG_VP,"elan_validate_cap: failed unknown type = %x \n", (cap->cap_type & ELAN_CAP_TYPE_MASK));       
+               return (EINVAL);
+       }
+       
+       if ((cap->cap_lowcontext == ELAN_CAP_UNINITIALISED) || (cap->cap_highcontext == ELAN_CAP_UNINITIALISED)
+           || (cap->cap_lownode == ELAN_CAP_UNINITIALISED) || (cap->cap_highnode    == ELAN_CAP_UNINITIALISED))
+       {
+               
+               ELAN_DEBUG4 (ELAN_DBG_VP,"elan_validate_cap: ELAN_CAP_UNINITIALISED   LowNode %d   HighNode %d   LowContext %d   highContext %d\n",
+                            cap->cap_lownode , cap->cap_highnode,
+                            cap->cap_lowcontext , cap->cap_highcontext);
+               return (EINVAL);
+       }       
+
+       if (cap->cap_lowcontext > cap->cap_highcontext)
+       {
+               ELAN_DEBUG2 (ELAN_DBG_VP,"elan_validate_cap: (cap->cap_lowcontext > cap->cap_highcontext) %d %d\n",cap->cap_lowcontext , cap->cap_highcontext);
+               return (EINVAL);
+       }
+       
+       if (cap->cap_lownode > cap->cap_highnode)
+       {
+               ELAN_DEBUG2 (ELAN_DBG_VP,"elan_validate_cap: (cap->cap_lownode > cap->cap_highnode) %d %d\n",cap->cap_lownode, cap->cap_highnode);
+               return (EINVAL);
+       }
+
+       if (cap->cap_mycontext != ELAN_CAP_UNINITIALISED) 
+       {
+               ELAN_DEBUG1 (ELAN_DBG_VP,"elan_validate_cap: failed cap->cap_mycontext is set %d  \n", cap->cap_mycontext);
+               return (EINVAL);
+       }
+
+
+       if ((ELAN_CAP_NUM_NODES(cap) * ELAN_CAP_NUM_CONTEXTS(cap)) > ELAN_MAX_VPS)
+       {
+               ELAN_DEBUG6 (ELAN_DBG_VP,"elan_validate_cap: too many vps  LowNode %d   HighNode %d   LowContext %d   highContext %d,  %d >% d\n",
+                            cap->cap_lownode , cap->cap_highnode,
+                            cap->cap_lowcontext , cap->cap_highcontext,
+                            (ELAN_CAP_NUM_NODES(cap) * ELAN_CAP_NUM_CONTEXTS(cap)),
+                            ELAN_MAX_VPS);
+               
+               return (EINVAL);
+       }
+
+       return (ESUCCESS);
+}
+
+int
+elan_validate_map(ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map)
+{
+       ELAN_CAP_NODE_STRUCT * ptr  = NULL;
+       ELAN_VP_NODE_STRUCT  * vptr = NULL;
+       char space[256];
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_map \n");
+       ELAN_DEBUG1 (ELAN_DBG_VP,"elan_validate_map cap = %s \n",elan_capability_string(cap,space));
+       ELAN_DEBUG1 (ELAN_DBG_VP,"elan_validate_map map = %s \n",elan_capability_string(map,space));
+
+       /* does cap exist    */
+       ptr = find_cap_node(cap);
+       if ( ptr == NULL ) 
+       {
+               ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_map: cap not found \n");
+               kmutex_unlock(&elan_mutex);
+               return EINVAL;
+       }
+       /* is it active */
+       if ( ! ptr->node.active ) 
+       {
+               ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_map: cap not active \n");
+               kmutex_unlock(&elan_mutex);
+               return EINVAL;
+       }
+
+       /* are they the same */
+       if ( ELAN_CAP_TYPE_MATCH(cap,map) 
+            && ELAN_CAP_GEOM_MATCH(cap,map)) 
+       {
+               ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_map: cap == map  passed\n");
+               kmutex_unlock(&elan_mutex);
+               return ESUCCESS;
+       }
+
+       /* is map in map list */
+       vptr = find_vp_node(ptr, map);
+       if ( vptr == NULL ) 
+       {
+               ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_map:  map not found\n");
+               kmutex_unlock(&elan_mutex);
+               return EINVAL;
+       }
+       
+       ELAN_DEBUG0 (ELAN_DBG_VP,"elan_validate_map:  map passed\n");
+       kmutex_unlock(&elan_mutex);
+       return ESUCCESS;
+}
+
+int
+elan_create_cap(ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap)
+{
+       char                      space[127];
+       struct list_head        * tmp;
+       ELAN_CAP_NODE_STRUCT * ptr = NULL;
+       int                       i, rail;
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG1 (ELAN_DBG_VP,"elan_create_cap %s\n",elan_capability_string(cap,space));     
+
+       /* need to check that the cap does not over lap another one 
+          or is an exact match with only the userkey changing */
+       list_for_each(tmp, &elan_cap_list) {
+               ptr = list_entry(tmp, ELAN_CAP_NODE_STRUCT , list);
+
+               /* is it an exact match */
+               if ( ELAN_CAP_TYPE_MATCH(&ptr->node.cap,cap) 
+                    && ELAN_CAP_GEOM_MATCH(&ptr->node.cap,cap)
+                    && (&ptr->node.owner == owner)) {
+                       if ( ptr->node.active ) {
+                               /* dont inc attached count as its like a create */
+                               ptr->node.cap.cap_userkey = cap->cap_userkey;
+                               kmutex_unlock(&elan_mutex);
+                               return ESUCCESS;
+                       }
+                       else
+                       {
+                               kmutex_unlock(&elan_mutex);
+                               return EINVAL;
+                       }
+               }
+               
+               /* does it overlap, even with ones being destroyed */
+               if (elan_cap_overlap(&ptr->node.cap,cap))
+               {
+                       kmutex_unlock(&elan_mutex);
+                       return  EACCES;
+               }
+       }
+
+       /* create it */
+       KMEM_ALLOC(ptr, ELAN_CAP_NODE_STRUCT *, sizeof(ELAN_CAP_NODE_STRUCT), 1);
+       if (ptr == NULL)
+       {
+               kmutex_unlock(&elan_mutex);
+               return  ENOMEM;
+       }
+
+       /* create space for the attached array */
+       for(rail=0;rail<ELAN_MAX_RAILS;rail++)
+       {
+               ptr->attached[rail]=NULL;
+               if ( ELAN_CAP_IS_RAIL_SET(cap,rail) ) 
+               {
+                       KMEM_ALLOC(ptr->attached[rail], ELAN_ATTACHED_STRUCT *, sizeof(ELAN_ATTACHED_STRUCT) *  ELAN_CAP_NUM_CONTEXTS(cap), 1);
+                       if (ptr->attached[rail] == NULL) 
+                       {
+                               for(;rail>=0;rail--)
+                                       if ( ptr->attached[rail] )
+                                               KMEM_FREE(ptr->attached[rail], sizeof(ELAN_ATTACHED_STRUCT) *  ELAN_CAP_NUM_CONTEXTS(cap));
+
+                               KMEM_FREE(ptr, sizeof(ELAN_CAP_NODE_STRUCT));
+                               kmutex_unlock(&elan_mutex);
+                               return  ENOMEM;
+                       }
+                       /* blank the attached array */
+                       for(i=0;i<ELAN_CAP_NUM_CONTEXTS(cap);i++)
+                               ptr->attached[rail][i].cb_func = NULL;
+               }
+       }       
+       
+       ptr->node.owner     = owner;
+       ptr->node.cap       = *cap;
+       ptr->node.attached  = 1;    /* creator counts as attached */
+       ptr->node.active    = 1;
+       ptr->vp_list.next   = &(ptr->vp_list);
+       ptr->vp_list.prev   = &(ptr->vp_list);
+
+       list_add_tail(&ptr->list, &elan_cap_list);      
+
+       kmutex_unlock(&elan_mutex);
+       return  ESUCCESS;
+}
+
+void
+elan_destroy_cap_test(ELAN_CAP_NODE_STRUCT *cap_ptr)
+{
+       /* called by someone holding the mutex   */
+       struct list_head       * vp_tmp;
+       ELAN_VP_NODE_STRUCT * vp_ptr = NULL;
+       int                      rail;
+
+       /* check to see if it can be deleted now */
+       if ( cap_ptr->node.attached == 0 ) {
+               
+               ELAN_DEBUG0(ELAN_DBG_CAP,"elan_destroy_cap_test: attached == 0\n");     
+               
+               /* delete the vp list */
+               list_for_each(vp_tmp, &(cap_ptr->vp_list)) {
+                       vp_ptr = list_entry(vp_tmp, ELAN_VP_NODE_STRUCT , list);
+                       list_del(&vp_ptr->list);
+                       KMEM_FREE( vp_ptr, sizeof(ELAN_VP_NODE_STRUCT));
+               }
+               
+               list_del(&cap_ptr->list);
+
+               /* delete space for the attached array */
+               for(rail=0;rail<ELAN_MAX_RAILS;rail++)
+                       if (cap_ptr->attached[rail]) 
+                               KMEM_FREE(cap_ptr->attached[rail], sizeof(ELAN_ATTACHED_STRUCT) * ELAN_CAP_NUM_CONTEXTS(&(cap_ptr->node.cap)));
+                       
+               KMEM_FREE(cap_ptr, sizeof(ELAN_CAP_NODE_STRUCT));               
+       }
+}
+
+int
+elan_destroy_cap(ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap)
+{
+       char                      space[127];
+       struct list_head        * el;
+       struct list_head        * nel;
+       ELAN_CAP_NODE_STRUCT * ptr = NULL;
+       int                       i, rail;
+       int                       found = 0;
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG1 (ELAN_DBG_CAP,"elan_destroy_cap %s\n",elan_capability_string(cap,space));   
+
+       list_for_each_safe (el, nel, &elan_cap_list) {
+               ptr = list_entry(el, ELAN_CAP_NODE_STRUCT , list);
+               
+               /* is it an exact match */
+               if ( (ptr->node.owner == owner )
+                    && (  (cap == NULL) 
+                          || (ELAN_CAP_TYPE_MATCH(&ptr->node.cap,cap) && ELAN_CAP_GEOM_MATCH(&ptr->node.cap,cap)))) {
+
+                       if ( ptr->node.active ) {
+
+                               /* mark as in active and dec attached count */
+                               ptr->node.active = 0;
+                               ptr->node.attached--;
+                               ptr->node.owner  = 0; /* no one own's it now */
+                               
+                               /* need to tell any one who was attached that this has been destroy'd */
+                               for(rail=0;rail<ELAN_MAX_RAILS;rail++)
+                                       if (ELAN_CAP_IS_RAIL_SET( &(ptr->node.cap), rail)) {
+                                               for(i=0;i< ELAN_CAP_NUM_CONTEXTS(&(ptr->node.cap));i++)
+                                                       if ( ptr->attached[rail][i].cb_func != NULL) 
+                                                               ptr->attached[rail][i].cb_func(ptr->attached[rail][i].cb_args, cap, NULL);
+                                       }
+                               
+                               /* now try to destroy it */
+                               elan_destroy_cap_test(ptr);
+                               
+                               /* found it */
+                               found = 1;
+                       }
+               }
+       }
+       
+       if ( found )
+       {
+               kmutex_unlock(&elan_mutex);
+               return ESUCCESS;
+       }
+
+       /* failed */
+       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_destroy_cap: didnt find it \n"); 
+
+       kmutex_unlock(&elan_mutex);
+       return EINVAL;
+}
+
+int 
+elan_get_caps(uint *number_of_results, uint array_size, ELAN_CAP_STRUCT *caps)
+{
+       uint                      results = 0;
+       struct list_head        * tmp;
+       ELAN_CAP_NODE_STRUCT * ptr = NULL;
+       
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_get_caps\n");    
+
+       list_for_each(tmp, &elan_cap_list) {
+               ptr = list_entry(tmp, ELAN_CAP_NODE_STRUCT , list);
+               
+               copyout(&ptr->node, &caps[results], sizeof (ELAN_CAP_STRUCT));
+               
+               results++;
+               
+               if ( results >= array_size )
+               {
+                       copyout(&results, number_of_results, sizeof(uint));     
+                       kmutex_unlock(&elan_mutex);
+                       return ESUCCESS;
+               }
+       }
+
+       copyout(&results, number_of_results, sizeof(uint));     
+
+       kmutex_unlock(&elan_mutex);
+       return ESUCCESS;
+}
+
+int
+elan_create_vp(ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map)
+{
+       ELAN_CAP_NODE_STRUCT * cap_ptr = NULL;
+       ELAN_VP_NODE_STRUCT  * vp_ptr  = NULL;
+       
+       kmutex_lock(&elan_mutex);
+
+
+       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_create_vp\n");
+
+       /* the railmasks must match */
+       if ( cap->cap_railmask != map->cap_railmask)
+       {
+               kmutex_unlock(&elan_mutex);
+               return  EINVAL;
+       }
+
+       /* does the cap exist */
+       cap_ptr = find_cap_node(cap);
+       if ((cap_ptr == NULL) || ( cap_ptr->node.owner != owner ) || (! cap_ptr->node.active) )
+       {
+               kmutex_unlock(&elan_mutex);
+               return  EINVAL;
+       }
+       
+       /* is there already a mapping */
+       vp_ptr = find_vp_node(cap_ptr,map);
+       if ( vp_ptr != NULL) 
+       {
+               kmutex_unlock(&elan_mutex);
+               return  EINVAL;
+       }
+
+       /* create space for mapping */
+       KMEM_ALLOC(vp_ptr, ELAN_VP_NODE_STRUCT *, sizeof(ELAN_VP_NODE_STRUCT), 1);
+       if (vp_ptr == NULL)
+       {
+               kmutex_unlock(&elan_mutex);
+               return  ENOMEM;
+       }
+                       
+       /* copy map */
+       vp_ptr->vp = *map;
+       list_add_tail(&vp_ptr->list, &(cap_ptr->vp_list));      
+       kmutex_unlock(&elan_mutex);
+       return  ESUCCESS;
+}
+
+int
+elan_destroy_vp(ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map)
+{
+       ELAN_CAP_NODE_STRUCT * cap_ptr = NULL;
+       ELAN_VP_NODE_STRUCT  * vp_ptr  = NULL;
+       int                       i, rail;
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_destroy_vp\n");  
+
+       cap_ptr = find_cap_node(cap);
+       if ((cap_ptr!=NULL) && (cap_ptr->node.owner == owner) && ( cap_ptr->node.active))
+       {               
+               vp_ptr = find_vp_node( cap_ptr, map );
+               if ( vp_ptr != NULL ) 
+               {
+                       list_del(&vp_ptr->list);
+                       KMEM_FREE(vp_ptr, sizeof(ELAN_VP_NODE_STRUCT));
+             
+                       /* need to tell those who are attached that map is nolonger in use */
+                       for(rail=0;rail<ELAN_MAX_RAILS;rail++)
+                               if (ELAN_CAP_IS_RAIL_SET(cap, rail))
+                               {
+                                       for(i=0;i< ELAN_CAP_NUM_CONTEXTS(&(cap_ptr->node.cap));i++)
+                                               if ( cap_ptr->attached[rail][i].cb_func != NULL) 
+                                                       cap_ptr->attached[rail][i].cb_func( cap_ptr->attached[rail][i].cb_args, cap, map);
+                               }
+
+                       kmutex_unlock(&elan_mutex);
+                       return  ESUCCESS;
+               }
+       }       
+       
+       /* didnt find it */
+       kmutex_unlock(&elan_mutex);
+       return  EINVAL;
+}
+
+int 
+elan_attach_cap(ELAN_CAPABILITY *cap, unsigned int rail, void *args, ELAN_DESTROY_CB func)
+{
+       char                  space[127];
+       struct list_head     *el;
+
+       ELAN_DEBUG1 (ELAN_DBG_CAP,"elan_attach_cap %s\n",elan_capability_string(cap,space));
+
+       /* currently must provide a call back, as null mean something */
+       if ( func == NULL)
+               return (EINVAL);
+
+       /* mycontext must be set and correct */
+       if ( ! ELAN_CAP_VALID_MYCONTEXT(cap))
+               return (EINVAL);
+
+       /* rail must be one of the rails in railmask */
+       if (((1 << rail) & cap->cap_railmask) == 0)
+               return (EINVAL);
+       
+       kmutex_lock(&elan_mutex);
+
+       list_for_each(el, &elan_cap_list) {
+               ELAN_CAP_NODE_STRUCT *cap_ptr = list_entry(el, ELAN_CAP_NODE_STRUCT , list);
+               
+               /* is it an exact match */
+               if (ELAN_CAP_MATCH(&cap_ptr->node.cap,cap) && cap_ptr->node.active) {
+                       unsigned int attached_index = cap->cap_mycontext - cap->cap_lowcontext;
+                       
+                       if ( cap_ptr->attached[rail][attached_index].cb_func != NULL ) /* only one per ctx per rail */
+                       {
+                               kmutex_unlock(&elan_mutex);
+                               return   EINVAL;
+                       }
+
+                       /* keep track of who attached as we might need to tell them when */
+                       /* cap or maps get destroyed                                     */
+                       cap_ptr->attached[rail][ attached_index ].cb_func = func;
+                       cap_ptr->attached[rail][ attached_index ].cb_args = args;
+                       cap_ptr->node.attached++;
+
+                       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_attach_cap: passed\n");
+                       kmutex_unlock(&elan_mutex);
+                       return ESUCCESS;
+               }
+       }
+       
+       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_attach_cap: failed to find \n");
+
+       /* didnt find one */
+       kmutex_unlock(&elan_mutex);
+       return EINVAL;
+}
+
+int 
+elan_detach_cap(ELAN_CAPABILITY *cap, unsigned int rail)
+{
+       struct list_head *el, *nel;
+       char              space[256];
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG1(ELAN_DBG_CAP,"elan_detach_cap %s\n",elan_capability_string(cap,space));
+       list_for_each_safe (el, nel, &elan_cap_list) {
+               ELAN_CAP_NODE_STRUCT *ptr = list_entry (el, ELAN_CAP_NODE_STRUCT, list);
+
+               /* is it an exact match */
+               if (ELAN_CAP_TYPE_MATCH(&ptr->node.cap,cap) &&
+                   ELAN_CAP_GEOM_MATCH(&ptr->node.cap,cap) &&
+                   (ptr->node.cap.cap_railmask & cap->cap_railmask) == cap->cap_railmask) {
+               
+                       unsigned int attached_index = cap->cap_mycontext - cap->cap_lowcontext;
+
+                       if ( ptr->attached[rail][ attached_index ].cb_func == NULL ) 
+                               ELAN_DEBUG0(ELAN_DBG_CAP,"elanmod_detach_cap already removed \n");
+
+                       ptr->attached[rail][ attached_index ].cb_func = NULL;
+                       ptr->attached[rail][ attached_index ].cb_args = (void *)0;
+
+                       ptr->node.attached--;
+
+                       ELAN_DEBUG1(ELAN_DBG_CAP,"elanmod_detach_cap new attach count%d \n", ptr->node.attached);
+
+                       elan_destroy_cap_test(ptr);
+
+                       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_detach_cap: success\n"); 
+
+                       kmutex_unlock(&elan_mutex);
+                       return  ESUCCESS;
+               }
+       }
+
+       ELAN_DEBUG0(ELAN_DBG_CAP,"elan_detach_cap: failed to find\n");
+       kmutex_unlock(&elan_mutex);
+       return  EINVAL;
+}
+
+int
+elan_cap_dump()
+{
+       struct list_head        * tmp;
+       ELAN_CAP_NODE_STRUCT * ptr = NULL;
+       
+       kmutex_lock(&elan_mutex);       
+       
+       list_for_each(tmp, &elan_cap_list) {
+               ptr = list_entry(tmp, ELAN_CAP_NODE_STRUCT , list);
+
+               ELAN_DEBUG2 (ELAN_DBG_ALL, "cap dump: owner %p type %x\n", ptr->node.owner, ptr->node.cap.cap_type);
+                       
+               ELAN_DEBUG5 (ELAN_DBG_ALL, "cap dump: LowNode %d   HighNode %d   LowContext %d   mycontext %d   highContext %d\n",
+                            ptr->node.cap.cap_lownode , ptr->node.cap.cap_highnode,
+                            ptr->node.cap.cap_lowcontext , ptr->node.cap.cap_mycontext, ptr->node.cap.cap_highcontext);
+
+       }
+
+       kmutex_unlock(&elan_mutex);
+       return  ESUCCESS;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan/capability_general.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/capability_general.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/capability_general.c    2005-06-01 23:12:54.558445336 -0400
@@ -0,0 +1,446 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: capability_general.c,v 1.10 2004/02/25 13:47:59 daniel Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/shared/capability_general.c,v $ */
+
+#if defined(__KERNEL__)
+
+#include <qsnet/kernel.h>
+
+#else
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <sys/param.h>
+
+#endif
+
+#include <elan/elanmod.h>
+
+
+void
+elan_nullcap (ELAN_CAPABILITY *cap)
+{
+       register int i;
+
+       for (i = 0; i < sizeof (cap->cap_userkey)/sizeof(cap->cap_userkey.key_values[0]); i++)
+               cap->cap_userkey.key_values[i] = ELAN_CAP_UNINITIALISED;
+    
+       cap->cap_lowcontext  = ELAN_CAP_UNINITIALISED;
+       cap->cap_highcontext = ELAN_CAP_UNINITIALISED;
+       cap->cap_mycontext   = ELAN_CAP_UNINITIALISED;
+       cap->cap_lownode     = ELAN_CAP_UNINITIALISED;
+       cap->cap_highnode    = ELAN_CAP_UNINITIALISED;
+       cap->cap_railmask    = ELAN_CAP_UNINITIALISED;
+       cap->cap_type        = ELAN_CAP_UNINITIALISED;
+       cap->cap_spare       = 0;
+       cap->cap_version     = ELAN_CAP_VERSION_NUMBER;
+       
+       for (i = 0; i < sizeof (cap->cap_bitmap)/sizeof (cap->cap_bitmap[0]); i++)
+               cap->cap_bitmap[i] = 0;
+}
+
+char *
+elan_capability_string (ELAN_CAPABILITY *cap, char *str)
+{
+       if (cap == NULL) 
+               sprintf (str, "[-.-.-.-] cap = NULL\n");
+       else
+               sprintf (str, "[%x.%x.%x.%x] Version %x Type %x \n"
+                        "Context %x.%x.%x Node %x.%x\n",
+                        cap->cap_userkey.key_values[0], cap->cap_userkey.key_values[1],
+                        cap->cap_userkey.key_values[2], cap->cap_userkey.key_values[3],
+                        cap->cap_version, cap->cap_type, 
+                        cap->cap_lowcontext, cap->cap_mycontext, cap->cap_highcontext,
+                        cap->cap_lownode, cap->cap_highnode);
+       
+       return (str);
+}
+
+ELAN_LOCATION
+elan_vp2location (u_int process, ELAN_CAPABILITY *cap)
+{
+       ELAN_LOCATION location;
+       int i, vp, node, context, nnodes, nctxs;
+
+       vp = 0;
+
+       location.loc_node    = ELAN_INVALID_NODE;
+       location.loc_context = -1;
+       
+       nnodes = cap->cap_highnode - cap->cap_lownode + 1;
+       nctxs  = cap->cap_highcontext - cap->cap_lowcontext + 1;
+       
+       switch (cap->cap_type & ELAN_CAP_TYPE_MASK)
+       {
+       case ELAN_CAP_TYPE_BLOCK:
+               for (node = 0, i = 0; node < nnodes; node++)
+               {
+                       for (context = 0; context < nctxs; context++)
+                       {
+                               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, context + (node * nctxs)))
+                               {
+                                       if (vp == process)
+                                       {
+                                               /* Return relative indices within the capability box */
+                                               location.loc_node    = node;
+                                               location.loc_context = context;
+
+                                               return (location);
+                                       }
+                      
+                                       vp++;
+                               }
+                       }
+               }
+               break;
+       
+       case ELAN_CAP_TYPE_CYCLIC:
+               for (context = 0, i = 0; context < nctxs; context++)
+               {
+                       for (node = 0; node < nnodes; node++)
+                       {
+                               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, node + (context * nnodes)))
+                               {
+                                       if (vp == process)
+                                       {
+                                               location.loc_node    = node;
+                                               location.loc_context = context;
+
+                                               return (location);
+                                       }
+                   
+                                       vp++;
+                               }
+                       }
+               }
+               break;
+       }
+    
+       return( location );
+}
+
+int
+elan_location2vp (ELAN_LOCATION location, ELAN_CAPABILITY *cap)
+{
+    int  vp, node, context, nnodes, nctxs;
+
+    nnodes = cap->cap_highnode - cap->cap_lownode + 1;
+    nctxs  = cap->cap_highcontext - cap->cap_lowcontext + 1;
+
+    vp = 0;
+    
+    switch (cap->cap_type & ELAN_CAP_TYPE_MASK)
+    {
+    case ELAN_CAP_TYPE_BLOCK:
+       for (node = 0 ; node < nnodes ; node++)
+       {
+           for (context = 0; context < nctxs; context++)
+           {
+               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, context + (node * nctxs)))
+               {
+                   if ((location.loc_node == node) && (location.loc_context == context))
+                   {
+                       /* Found it ! */
+                       return( vp );
+                   }
+                   
+                   vp++;
+               }
+           }
+       }
+       break;
+       
+    case ELAN_CAP_TYPE_CYCLIC:
+       for (context = 0; context < nctxs; context++)
+       {
+           for (node = 0; node < nnodes; node++)
+           {
+               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, node + (context * nnodes)))
+               {
+                   if ((location.loc_node == node) && (location.loc_context == context))
+                   {
+                       /* Found it ! */
+                       return( vp );
+                   }
+                   
+                   vp++;
+               }
+           }
+       }
+       break;
+    }
+    
+    /* Failed to find it */
+    return( -1 );
+}
+
+/* Return the number of processes as described by a capability */
+int
+elan_nvps (ELAN_CAPABILITY *cap)
+{
+       int i, c, nbits = ELAN_CAP_BITMAPSIZE(cap);
+
+       if (cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP)
+               return (nbits);
+
+       for (i = 0, c = 0; i < nbits; i++)
+               if (BT_TEST (cap->cap_bitmap, i))
+                       c++;
+
+       return (c);
+}
+
+/* Return the number of local processes on a given node as described by a capability */
+int
+elan_nlocal (int node, ELAN_CAPABILITY *cap)
+{
+       int vp;
+       ELAN_LOCATION loc;
+       int nLocal = 0;
+
+       for (vp = 0; vp < elan_nvps(cap); vp++)
+       {
+               loc = elan_vp2location(vp, cap);
+               if (loc.loc_node == node)
+                       nLocal++;
+       }
+
+       return (nLocal);
+}
+
+/* Return the maximum number of local processes on any node as described by a capability */
+int
+elan_maxlocal (ELAN_CAPABILITY *cap)
+{
+       return(cap->cap_highcontext - cap->cap_lowcontext + 1);
+}
+
+/* Return the vps of the local processes on a given node as described by a capability */
+int
+elan_localvps (int node, ELAN_CAPABILITY *cap, int *vps, int size)
+{
+       int context;
+       ELAN_LOCATION loc;
+       int nLocal = 0;
+    
+       loc.loc_node = node;
+
+       for (context = 0; context < MIN(size, elan_maxlocal(cap)); context++)
+       {
+               loc.loc_context = context;
+       
+               /* Should return -1 if none found */
+               if ( (vps[context] = elan_location2vp( loc, cap )) != -1)
+                       nLocal++;
+       }
+
+       return (nLocal);
+}
+
+/* Return the number of rails that this capability utilises */
+int
+elan_nrails (ELAN_CAPABILITY *cap)
+{
+       int nrails = 0;
+       unsigned int railmask;
+
+       /* Test for a multi-rail capability */
+       if (cap->cap_type & ELAN_CAP_TYPE_MULTI_RAIL)
+       {
+               /* Grab rail bitmask from capability */
+               railmask = cap->cap_railmask;
+       
+               while (railmask)
+               {
+                       if (railmask & 1)
+                               nrails++;
+           
+                       railmask >>= 1;
+               }
+       }
+       else 
+               /* Default to just one rail */
+               nrails = 1;
+       
+       return (nrails);
+}
+
+/* Fill out an array giving the physical rail numbers utilised by a capability */
+int
+elan_rails (ELAN_CAPABILITY *cap, int *rails)
+{
+       int nrails, rail;
+       unsigned int railmask;
+
+       /* Test for a multi-rail capability */
+       if (cap->cap_type & ELAN_CAP_TYPE_MULTI_RAIL)
+       {
+               /* Grab rail bitmask from capability */
+               railmask = cap->cap_railmask;
+       
+               nrails = rail = 0;
+               while (railmask)
+               {
+                       if (railmask & 1)
+                               rails[nrails++] = rail;
+           
+                       rail++;
+                       railmask >>= 1;
+               }
+       }
+       else
+       {
+               /* Default to just one rail */
+               rails[0] = 0;
+               nrails = 1;
+       }
+
+       return( nrails );
+}
+
+int 
+elan_cap_overlap(ELAN_CAPABILITY *cap1, ELAN_CAPABILITY *cap2)
+{
+       /* by context */
+       if ( cap1->cap_highcontext < cap2->cap_lowcontext ) return (0);
+       if ( cap1->cap_lowcontext  > cap2->cap_highcontext) return (0);
+       
+       /* by node */
+       if ( cap1->cap_highnode < cap2->cap_lownode ) return (0);
+       if ( cap1->cap_lownode  > cap2->cap_highnode) return (0);
+
+       /* by rail */
+       /* they overlap if they have a rail in common */
+       return (cap1->cap_railmask & cap2->cap_railmask);
+}
+
+#if !defined(__KERNEL__)
+
+/* Fill out an array that hints at the best use of the rails on a
+ * per process basis. The library user can then decide whether or not
+ * to take this into account (e.g. TPORTs)
+ * All processes calling this fn will be returned the same information.
+ */
+int
+elan_prefrails(ELAN_CAPABILITY *cap, int *pref, int nvp)
+{
+       int i;
+       int nrails = elan_nrails(cap);
+       int maxlocal = elan_maxlocal(cap);
+
+       /* Test for a multi-rail capability */
+       if (! (cap->cap_type & ELAN_CAP_TYPE_MULTI_RAIL))
+       {
+               /* Default to just one rail */
+               for (i = 0; i < nvp; i++)
+                       pref[i] = 0;
+
+               return( 0 );
+       }
+
+       /*
+        * We allocate rails on a per node basis sharing our the rails
+        * equally amongst the local processes. However, if there is only
+        * one process per node and multiple rails, then we use a different
+        * algorithm where rails are allocated across all the processes in 
+        * a round-robin fashion
+        */
+    
+       if (maxlocal == 1)
+       {
+               /* Allocate rails in a round-robin manner */
+               for (i = 0; i < nvp; i++)
+                       *pref++ = i % nrails;
+       }
+       else
+       {
+               int node;
+               int *vps;
+               int nnodes = cap->cap_highnode - cap->cap_lownode + 1;
+
+               vps = (int *) malloc(sizeof(int)*maxlocal);
+
+               /* Grab the local process info for each node and allocate
+                * rails to those vps on an equal basis
+                */
+               for (node = 0; node < nnodes; node++)
+               {
+                       int nlocal;
+                       int pprail;
+
+                       /* Grab an array of local vps */
+                       nlocal = elan_localvps(node, cap, vps, maxlocal);
+           
+                       /* Calculate the number processes per rail */
+                       if ((pprail = nlocal/nrails) == 0)
+                               pprail = 1;
+
+                       /* Allocate processes to rails */
+                       for (i = 0; i < nlocal; i++)
+                       {
+                               pref[vps[i]] = (i / pprail) % nrails;
+                       }
+               }
+       
+               free(vps);
+       }
+
+       return( 0 );
+}
+
+void 
+elan_get_random_key(ELAN_USERKEY *key)
+{
+    int i;
+    for (i = 0; i < sizeof(key->key_values) / sizeof(key->key_values[0]); i++)
+       key->key_values[i] = lrand48();
+}
+
+int elan_lowcontext(ELAN_CAPABILITY *cap)
+{
+    return(cap->cap_lowcontext);
+}
+
+int elan_mycontext(ELAN_CAPABILITY *cap)
+{
+    return(cap->cap_mycontext);
+}
+
+int elan_highcontext(ELAN_CAPABILITY *cap)
+{
+    return(cap->cap_highcontext);
+}
+
+int elan_lownode(ELAN_CAPABILITY *cap)
+{
+    return(cap->cap_lownode);
+}
+
+int elan_highnode(ELAN_CAPABILITY *cap)
+{
+    return(cap->cap_highnode);
+}
+
+int elan_captype(ELAN_CAPABILITY *cap)
+{
+    return(cap->cap_type);
+}
+
+int elan_railmask(ELAN_CAPABILITY *cap)
+{
+    return(cap->cap_railmask);
+}
+
+#endif
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan/device.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/device.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/device.c        2005-06-01 23:12:54.559445184 -0400
@@ -0,0 +1,147 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: device.c,v 1.5 2003/09/24 13:55:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/device.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan/elanmod.h>
+
+static LIST_HEAD(elan_dev_list);
+
+ELAN_DEV_STRUCT *
+elan_dev_find (ELAN_DEV_IDX devidx)
+{
+       struct list_head   *tmp;
+       ELAN_DEV_STRUCT *ptr=NULL;
+
+       list_for_each(tmp, &elan_dev_list) {
+               ptr = list_entry(tmp, ELAN_DEV_STRUCT , node);
+               if (ptr->devidx == devidx) 
+                       return ptr;
+               if (ptr->devidx > devidx)
+                       return ERR_PTR(-ENXIO);
+       }
+       
+       return ERR_PTR(-EINVAL);
+}
+
+ELAN_DEV_STRUCT *
+elan_dev_find_byrail (unsigned short deviceid, unsigned rail)
+{
+       struct list_head   *tmp;
+       ELAN_DEV_STRUCT *ptr=NULL;
+
+       list_for_each(tmp, &elan_dev_list) {
+               ptr = list_entry(tmp, ELAN_DEV_STRUCT , node);
+
+               ELAN_DEBUG5 (ELAN_DBG_ALL,"elan_dev_find_byrail devidx %d - %04x %04x,  %d %d \n", ptr->devidx, 
+                            ptr->devinfo->dev_device_id, deviceid, ptr->devinfo->dev_rail, rail);
+
+               if (ptr->devinfo->dev_device_id == deviceid && ptr->devinfo->dev_rail == rail)
+                       return ptr;
+       }
+       
+       return NULL;
+}
+
+ELAN_DEV_IDX
+elan_dev_register (ELAN_DEVINFO *devinfo, ELAN_DEV_OPS *ops, void * user_data)
+{
+       ELAN_DEV_STRUCT *ptr;
+       ELAN_DEV_IDX        devidx = 0;
+       struct list_head   *tmp;
+
+        kmutex_lock(&elan_mutex);
+
+       /* is it already registered */
+       if ((ptr = elan_dev_find_byrail(devinfo->dev_device_id, devinfo->dev_rail)) != NULL) 
+       {
+               kmutex_unlock(&elan_mutex);
+               return EINVAL;
+       }
+
+       /* find a free device idx */
+       list_for_each (tmp, &elan_dev_list) {
+               if (list_entry (tmp, ELAN_DEV_STRUCT, node)->devidx != devidx)
+                       break;
+               devidx++;
+       }
+
+       /* create it and add */
+       KMEM_ALLOC(ptr, ELAN_DEV_STRUCT *, sizeof(ELAN_DEV_STRUCT), 1);
+       if (ptr == NULL)
+       {
+               kmutex_unlock(&elan_mutex);
+               return ENOMEM;
+       }
+
+       ptr->devidx    = devidx;
+       ptr->ops       = ops;
+       ptr->devinfo   = devinfo;
+       ptr->user_data = user_data;
+
+       /* insert this entry *before* the last entry we've found */
+       list_add_tail(&ptr->node, tmp);
+
+       kmutex_unlock(&elan_mutex);
+       return  ESUCCESS;
+}
+
+int
+elan_dev_deregister (ELAN_DEVINFO *devinfo)
+{
+       ELAN_DEV_STRUCT *target;
+
+       kmutex_lock(&elan_mutex);
+
+       if ((target = elan_dev_find_byrail (devinfo->dev_device_id, devinfo->dev_rail)) == NULL)
+       {
+               kmutex_unlock(&elan_mutex);
+               return  EINVAL;
+       }
+
+       list_del(&target->node);
+
+       /* delete target entry */
+       KMEM_FREE(target, sizeof(ELAN_DEV_STRUCT));
+
+       kmutex_unlock(&elan_mutex);
+       return  ESUCCESS;
+}
+
+int
+elan_dev_dump ()
+{
+       struct list_head   *tmp;
+       ELAN_DEV_STRUCT *ptr=NULL;
+
+       kmutex_lock(&elan_mutex);       
+
+       list_for_each(tmp, &elan_dev_list) {
+               ptr = list_entry(tmp, ELAN_DEV_STRUCT , node);
+
+               ELAN_DEBUG3 (ELAN_DBG_ALL,"dev dump: index %u rail %u elan%c\n", 
+                            ptr->devidx, ptr->devinfo->dev_rail, '3' + ptr->devinfo->dev_device_id);
+               ELAN_DEBUG5 (ELAN_DBG_ALL,"dev dump: Vid %x   Did %x  Rid %x  DR %d  DVal %x\n",
+                            ptr->devinfo->dev_vendor_id,
+                            ptr->devinfo->dev_device_id,
+                            ptr->devinfo->dev_revision_id,
+                            ptr->devinfo->dev_driver_version,
+                            ptr->devinfo->dev_num_down_links_value);
+
+       }
+
+       kmutex_unlock(&elan_mutex);
+       return ESUCCESS;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan/devinfo.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/devinfo.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/devinfo.c       2005-06-01 23:12:54.559445184 -0400
@@ -0,0 +1,78 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: devinfo.c,v 1.5 2003/09/24 13:55:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/devinfo.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan/elanmod.h>
+
+int 
+elan_get_devinfo(ELAN_DEV_IDX devidx, ELAN_DEVINFO *devinfo)
+{
+       ELAN_DEV_STRUCT *target;
+       int                 res;
+
+       kmutex_lock(&elan_mutex);
+
+       target = elan_dev_find (devidx);
+
+       if (IS_ERR (target))
+               res = PTR_ERR(target);
+       else
+       {
+               copyout(target->devinfo, devinfo, sizeof(ELAN_DEVINFO));
+               res = ESUCCESS;
+       }
+       
+       kmutex_unlock(&elan_mutex);
+       return res;
+}
+
+int 
+elan_get_position(ELAN_DEV_IDX devidx, ELAN_POSITION *position)
+{
+       ELAN_DEV_STRUCT *target;
+       int                 res;
+
+       kmutex_lock(&elan_mutex);
+
+       target = elan_dev_find(devidx);
+
+       if (IS_ERR (target))
+               res = PTR_ERR(target);
+       else
+               res = target->ops->get_position(target->user_data, position);
+       
+       kmutex_unlock(&elan_mutex);
+       return res;
+}
+
+int 
+elan_set_position(ELAN_DEV_IDX devidx, unsigned short nodeId, unsigned short numNodes)
+{
+       ELAN_DEV_STRUCT *target;
+       int                 res;
+
+       kmutex_lock(&elan_mutex);
+
+       target = elan_dev_find(devidx);
+
+       if (IS_ERR (target))
+               res = PTR_ERR (target);
+       else
+               res = target->ops->set_position(target->user_data, nodeId, numNodes);
+       
+       kmutex_unlock(&elan_mutex);
+       return res;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan/elanmod.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/elanmod.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/elanmod.c       2005-06-01 23:12:54.559445184 -0400
@@ -0,0 +1,149 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+#ident "@(#)$Id: elanmod.c,v 1.11 2004/06/18 09:28:16 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/elanmod.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan/elanmod.h>
+
+kmutex_t  elan_mutex;
+
+int 
+elan_init()
+{
+       kmutex_init(&elan_mutex);
+       return (ESUCCESS);
+}
+
+int 
+elan_fini()
+{
+       kmutex_destroy(&elan_mutex);
+       return (ESUCCESS);
+}
+
+int 
+elanmod_classify_cap (ELAN_POSITION *position, ELAN_CAPABILITY *cap, unsigned use)
+{
+       if (cap->cap_version != ELAN_CAP_VERSION_NUMBER)
+       {
+               ELAN_DEBUG2 (ELAN_DBG_VP, "elanmod_classify_cap: (cap->Version != ELAN_CAP_VERSION) %d %d\n", cap->cap_version, ELAN_CAP_VERSION_NUMBER);
+               return (-EINVAL);
+       }
+       
+       if (cap->cap_lowcontext == ELAN_CAP_UNINITIALISED || cap->cap_highcontext == ELAN_CAP_UNINITIALISED)
+       {
+               ELAN_DEBUG3 (ELAN_DBG_VP, "elanmod_classify_cap: LowContext %d    HighContext %d MyContext %d\n",
+                            cap->cap_lowcontext , cap->cap_highcontext, cap->cap_mycontext);
+               return (-EINVAL);
+       }
+       
+       if (cap->cap_lowcontext > cap->cap_highcontext)
+       {
+               ELAN_DEBUG2 (ELAN_DBG_VP, "elanmod_classify_cap: (cap->cap_lowcontext > cap->cap_highcontext) %d %d\n",cap->cap_lowcontext , cap->cap_highcontext);
+               return (-EINVAL);
+       }
+       
+       
+       switch (cap->cap_type & ELAN_CAP_TYPE_MASK)
+       {
+       case ELAN_CAP_TYPE_BLOCK:
+       case ELAN_CAP_TYPE_CYCLIC:
+               if (position->pos_mode == ELAN_POS_UNKNOWN)
+               {
+                       ELAN_DEBUG0 (ELAN_DBG_VP, "elanmod_classify_cap: Position Unknown \n");
+                       return (-EAGAIN);
+               }
+               
+               if ( ! ( ELAN_USER_CONTEXT(cap->cap_lowcontext) && ELAN_USER_CONTEXT(cap->cap_highcontext)))
+               {
+                       ELAN_DEBUG4 (ELAN_DBG_VP, "elanmod_classify_cap:  USER_BASE_CONTEXT %d %d %d %d \n" ,  ELAN_USER_BASE_CONTEXT_NUM,cap->cap_lowcontext, cap->cap_highcontext ,ELAN_USER_TOP_CONTEXT_NUM);
+                       return (-EINVAL);
+               }
+               if (cap->cap_lownode == ELAN_CAP_UNINITIALISED)
+                       cap->cap_lownode = position->pos_nodeid;
+               if (cap->cap_highnode == ELAN_CAP_UNINITIALISED)
+                       cap->cap_highnode = position->pos_nodeid;
+               
+               if (cap->cap_lownode < 0 || cap->cap_highnode >= position->pos_nodes || cap->cap_lownode > cap->cap_highnode)
+               {
+                       ELAN_DEBUG3 ( ELAN_DBG_VP,"elanmod_classify_cap: low %d high %d pos %d \n" , cap->cap_lownode  ,cap->cap_highnode, position->pos_nodes);
+                       
+                       return (-EINVAL);
+               }
+               
+               if ((cap->cap_highnode < position->pos_nodeid) || (cap->cap_lownode > position->pos_nodeid))
+               {
+                       ELAN_DEBUG3 (ELAN_DBG_VP, "elanmod_classify_cap: node not i range low %d high %d this %d\n",
+                                    cap->cap_lownode, cap->cap_highnode, position->pos_nodeid);
+                       return (-EINVAL);
+               }
+
+               break;
+       default:
+               ELAN_DEBUG1 (ELAN_DBG_VP, "elanmod_classify_cap: cant decode type %x \n", cap->cap_type & ELAN_CAP_TYPE_MASK);
+               return (-EINVAL);
+
+       }
+
+       switch (use)
+       {
+       case ELAN_USER_ATTACH:
+       case ELAN_USER_DETACH:
+               if (cap->cap_mycontext == ELAN_CAP_UNINITIALISED)
+               {
+                       ELAN_DEBUG0 (ELAN_DBG_VP, "elanmod_classify_cap: cap->cap_mycontext == ELAN_CAP_UNINITIALISED");
+                       return (-EINVAL);
+               }
+       
+               if ((cap->cap_mycontext != ELAN_CAP_UNINITIALISED) && 
+                   (cap->cap_mycontext < cap->cap_lowcontext || cap->cap_mycontext > cap->cap_highcontext))
+               {
+                       ELAN_DEBUG3 (ELAN_DBG_VP, "elanmod_classify_cap: cap->cap_mycontext out of range %d %d %d \n", cap->cap_lowcontext,cap->cap_mycontext,cap->cap_highcontext);
+                       return (-EINVAL);
+               }   
+               break;
+
+       case ELAN_USER_P2P:
+               break;
+
+       case ELAN_USER_BROADCAST:
+               if (! (cap->cap_type & ELAN_CAP_TYPE_BROADCASTABLE)) {
+                       ELAN_DEBUG0 (ELAN_DBG_VP, "elanmod_classify_cap: use ELAN_USER_BROADCAST but cap not ELAN_CAP_TYPE_BROADCASTABLE\n");
+                       return (-EINVAL);
+               }
+               break;
+
+       default:
+               ELAN_DEBUG1 (ELAN_DBG_VP, "elanmod_classify_cap: unknown use (%d)\n",use);
+               return (-EINVAL);
+       }
+
+
+
+       /* is any ctxt an rms one ?? */
+       if (ELAN_RMS_CONTEXT(cap->cap_lowcontext) || ELAN_RMS_CONTEXT(cap->cap_highcontext))
+       {
+               /* so both low and high must be */
+               if (!(ELAN_RMS_CONTEXT(cap->cap_lowcontext) && ELAN_RMS_CONTEXT(cap->cap_highcontext))) 
+               {
+                       ELAN_DEBUG2 (ELAN_DBG_VP, "elanmod_classify_cap: not rms ctxt %x %x\n",cap->cap_lowcontext,cap->cap_highcontext );
+                       return (-EINVAL);
+               }
+               ELAN_DEBUG0 (ELAN_DBG_VP, "elanmod_classify_cap: returning ELAN_CAP_RMS\n");
+               return (ELAN_CAP_RMS);
+       }
+
+       ELAN_DEBUG0 (ELAN_DBG_VP, "elanmod_classify_cap: returning ELAN_CAP_OK\n");
+       return (ELAN_CAP_OK);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan/elanmod_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/elanmod_linux.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/elanmod_linux.c 2005-06-01 23:12:54.560445032 -0400
@@ -0,0 +1,410 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elanmod_linux.c,v 1.16 2004/06/14 15:45:37 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/elanmod_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/elanmod.h>
+#include <elan/elanmod_linux.h>
+
+#include <linux/module.h>
+
+#include <linux/sysctl.h>
+#include <linux/init.h>
+
+#include <qsnet/procfs_linux.h>
+
+MODULE_AUTHOR("Quadrics Ltd.");
+MODULE_DESCRIPTION("Elan support module");
+
+MODULE_LICENSE("GPL");
+
+/* elanmod.c */
+EXPORT_SYMBOL(elanmod_classify_cap);
+
+/* bitmap.c */
+#include <elan/bitmap.h>
+
+EXPORT_SYMBOL(bt_freebit);
+EXPORT_SYMBOL(bt_lowbit); 
+EXPORT_SYMBOL(bt_nextbit);
+EXPORT_SYMBOL(bt_copy);
+EXPORT_SYMBOL(bt_zero); 
+EXPORT_SYMBOL(bt_fill); 
+EXPORT_SYMBOL(bt_cmp); 
+EXPORT_SYMBOL(bt_intersect);
+EXPORT_SYMBOL(bt_remove); 
+EXPORT_SYMBOL(bt_add); 
+EXPORT_SYMBOL(bt_spans);
+EXPORT_SYMBOL(bt_subset);  
+EXPORT_SYMBOL(bt_up);
+EXPORT_SYMBOL(bt_down);
+EXPORT_SYMBOL(bt_nbits);
+
+/* capability.c */
+EXPORT_SYMBOL(elan_nullcap);
+EXPORT_SYMBOL(elan_detach_cap);
+EXPORT_SYMBOL(elan_attach_cap);
+EXPORT_SYMBOL(elan_validate_map);
+
+/* stats.c */
+EXPORT_SYMBOL(elan_stats_register);
+EXPORT_SYMBOL(elan_stats_deregister);
+
+/* device.c */
+EXPORT_SYMBOL(elan_dev_deregister);
+EXPORT_SYMBOL(elan_dev_register);
+
+/* debug */
+int  elan_debug_mode = QSNET_DEBUG_BUFFER; 
+int  elan_debug_mask;
+
+static struct proc_dir_entry *elan_procfs_root;
+
+extern void elan_procfs_init(void);
+extern void elan_procfs_fini(void);
+
+static int elan_open    (struct inode *ino, struct file *fp);
+static int elan_release (struct inode *ino, struct file *fp);
+static int elan_ioctl   (struct inode *ino, struct file *fp, unsigned int cmd, unsigned long arg);
+
+static struct file_operations elan_fops = 
+{
+       ioctl:   elan_ioctl,
+       open:    elan_open,
+       release: elan_release,
+};
+
+static int __init elan_start(void)
+{
+       int res;
+
+       elan_procfs_init(); 
+
+       if ((res = elan_init()) != ESUCCESS)
+       {
+               elan_procfs_fini();
+               return (-res);
+       }
+
+       return (0);
+}
+
+static void __exit elan_exit(void)
+{
+       elan_fini();
+       elan_procfs_fini();
+}
+
+
+/* Declare the module init and exit functions */
+void
+elan_procfs_init()
+{
+       struct proc_dir_entry  *p;
+       
+       elan_procfs_root = proc_mkdir("elan",   qsnet_procfs_root);
+       
+       qsnet_proc_register_hex(elan_procfs_root, "debug_mask", &elan_debug_mask, 0);
+       qsnet_proc_register_hex(elan_procfs_root, "debug_mode", &elan_debug_mode, 0);
+
+       if ((p = create_proc_entry ("ioctl", 0, elan_procfs_root)) != NULL)
+       {
+               p->proc_fops = &elan_fops;
+               p->data      = 0;
+               p->owner     = THIS_MODULE;
+       }   
+}
+
+void
+elan_procfs_fini()
+{
+       remove_proc_entry ("debug_mask", elan_procfs_root);
+       remove_proc_entry ("debug_mode", elan_procfs_root);
+       
+       remove_proc_entry ("ioctl",   elan_procfs_root); 
+       remove_proc_entry ("version", elan_procfs_root);  
+       
+       remove_proc_entry ("elan",   qsnet_procfs_root);
+}
+
+module_init(elan_start);
+module_exit(elan_exit);
+
+static int
+elan_open (struct inode *inode, struct file *fp)
+{
+       MOD_INC_USE_COUNT;
+       fp->private_data = NULL;
+       return (0);
+}
+
+static int
+elan_release (struct inode *inode, struct file *fp)
+{
+       /* mark all caps owned by fp to be destroyed */
+       elan_destroy_cap(fp,NULL);
+
+       MOD_DEC_USE_COUNT;
+       return (0);
+}
+
+static int 
+elan_ioctl(struct inode *inode, struct file *fp, unsigned int cmd, unsigned long arg)
+{
+       int rep = 0;
+
+       switch (cmd) 
+       {
+       case ELANCTRL_STATS_GET_NEXT :
+       {
+               ELANCTRL_STATS_GET_NEXT_STRUCT args;
+
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_STATS_GET_NEXT_STRUCT)))
+                       return (-EFAULT); 
+
+               /* uses copyin/copyout */
+               if (elan_stats_get_next_index(args.statidx, args.next_statidx) != 0 ) 
+                       return (-EINVAL);       
+
+               break;
+       }
+       case ELANCTRL_STATS_FIND_INDEX :
+       {
+               ELANCTRL_STATS_FIND_INDEX_STRUCT args;
+
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_STATS_FIND_INDEX_STRUCT)))
+                       return (-EFAULT); 
+
+               /* uses copyin/copyout */
+               if (elan_stats_find_index(args.block_name, args.statidx, args.num_entries) != 0 ) 
+                       return (-EINVAL);       
+
+               break;
+       }
+       case ELANCTRL_STATS_GET_BLOCK_INFO :
+       {
+               ELANCTRL_STATS_GET_BLOCK_INFO_STRUCT args;
+               
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_STATS_GET_BLOCK_INFO_STRUCT)))
+                       return (-EFAULT);
+
+               /* uses copyin/copyout */
+               if (elan_stats_get_block_info(args.statidx, args.block_name, args.num_entries) != 0 ) 
+                       return (-EINVAL);
+               break;          
+       }
+       case ELANCTRL_STATS_GET_INDEX_NAME :
+       {
+               ELANCTRL_STATS_GET_INDEX_NAME_STRUCT args;
+               
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_STATS_GET_INDEX_NAME_STRUCT)))
+                       return (-EFAULT);
+
+               /* uses copyin/copyout */
+               if (elan_stats_get_index_name(args.statidx, args.index, args.name) != 0 )
+                       return (-EINVAL);
+               break;
+       }
+       case ELANCTRL_STATS_CLEAR_BLOCK :
+       {
+               ELANCTRL_STATS_CLEAR_BLOCK_STRUCT args;
+               
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_STATS_CLEAR_BLOCK_STRUCT)))
+                       return (-EFAULT);
+
+               /* statidx is not a pointer */
+               if (elan_stats_clear_block(args.statidx) != 0 )
+                       return (-EINVAL);
+               break;
+       }
+       case ELANCTRL_STATS_GET_BLOCK :
+       {
+               ELANCTRL_STATS_GET_BLOCK_STRUCT args;
+               
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_STATS_GET_BLOCK_STRUCT)))
+                       return (-EFAULT);
+
+               /* uses copyin/copyout */
+               if (elan_stats_get_block(args.statidx, args.entries, args.values) != 0 )
+                       return (-EINVAL);
+               break;
+       }
+       case ELANCTRL_GET_DEVINFO :
+       {
+               ELANCTRL_GET_DEVINFO_STRUCT args;
+               
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_GET_DEVINFO_STRUCT)))
+                       return (-EFAULT);
+
+               /* uses copyin/copyout */
+               if (elan_get_devinfo(args.devidx, args.devinfo) != 0 )
+                       return (-EINVAL);
+               break;          
+       }
+       case ELANCTRL_GET_POSITION :
+       {
+               ELANCTRL_GET_POSITION_STRUCT args;
+               
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_GET_POSITION_STRUCT)))
+                       return (-EFAULT); 
+
+               /* uses copyin/copyout */
+               if (elan_get_position(args.devidx, args.position) != 0 )
+                       return (-EINVAL);
+               break;          
+       }
+       case ELANCTRL_SET_POSITION :
+       {
+               ELANCTRL_SET_POSITION_STRUCT args;
+               
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_SET_POSITION_STRUCT)))
+                       return (-EFAULT);
+
+               /* uses copyin/copyout */
+               if (elan_set_position(args.devidx, args.nodeId, args.numNodes) != 0 )
+                       return (-EINVAL);       
+               break;          
+       }
+       case ELANCTRL_CREATE_CAP  :
+       {
+               ELANCTRL_CREATE_CAP_STRUCT *args;
+
+               /* get space for args */
+               KMEM_ALLOC(args, ELANCTRL_CREATE_CAP_STRUCT *, sizeof(ELANCTRL_CREATE_CAP_STRUCT), 1);
+               if (args == NULL)
+                       return(-ENOMEM);        
+
+               /* copy them */
+               if (copy_from_user (args, (void *) arg, sizeof (ELANCTRL_CREATE_CAP_STRUCT)))
+                       return (-EFAULT);
+               else 
+               {
+                       if ((elan_validate_cap(&args->cap) != 0) || (elan_create_cap(fp,&args->cap) != 0 )) 
+                               rep = (-EINVAL);
+               }
+
+               /* free the space */
+               KMEM_FREE(args, sizeof(ELANCTRL_CREATE_CAP_STRUCT));
+
+               break;          
+       }
+       case ELANCTRL_DESTROY_CAP  :
+       {
+               ELANCTRL_DESTROY_CAP_STRUCT *args;
+
+               /* get space for args */
+               KMEM_ALLOC(args, ELANCTRL_DESTROY_CAP_STRUCT *, sizeof(ELANCTRL_DESTROY_CAP_STRUCT), 1);
+               if (args == NULL)
+                       return(-ENOMEM);        
+
+               /* copy them */
+               if (copy_from_user (args, (void *) arg, sizeof (ELANCTRL_DESTROY_CAP_STRUCT)))
+                       rep = (-EFAULT);
+               else 
+               {
+                       if (elan_destroy_cap(fp, &args->cap) != 0 )
+                               rep = (-EINVAL);
+               }
+
+               /* free the space */
+               KMEM_FREE(args, sizeof(ELANCTRL_DESTROY_CAP_STRUCT));
+
+               break;          
+       }
+       case ELANCTRL_CREATE_VP  :
+       {
+               ELANCTRL_CREATE_VP_STRUCT *args;
+
+               /* get space for args */
+               KMEM_ALLOC(args, ELANCTRL_CREATE_VP_STRUCT *, sizeof(ELANCTRL_CREATE_VP_STRUCT), 1);
+               if (args == NULL)
+                       return(-ENOMEM);        
+
+               /* copy them */
+               if (copy_from_user (args, (void *) arg, sizeof (ELANCTRL_CREATE_VP_STRUCT)))
+                       return (-EFAULT);
+               else
+               {
+                       if ((elan_validate_cap( &args->map) != 0) || (elan_create_vp(fp, &args->cap, &args->map) != 0 ))
+                               rep = (-EINVAL);        
+               }
+
+               KMEM_FREE(args, sizeof(ELANCTRL_CREATE_VP_STRUCT ));
+
+               break;          
+       }
+       case ELANCTRL_DESTROY_VP  :
+       {
+               ELANCTRL_DESTROY_VP_STRUCT *args;
+
+               /* get space for args */
+               KMEM_ALLOC(args, ELANCTRL_DESTROY_VP_STRUCT *, sizeof(ELANCTRL_DESTROY_VP_STRUCT), 1);
+               if (args == NULL)
+                       return(-ENOMEM);        
+               
+               /* copy them */
+               if (copy_from_user (args, (void *) arg, sizeof (ELANCTRL_DESTROY_VP_STRUCT)))
+                       rep = (-EFAULT);
+               else 
+               {
+                       if (elan_destroy_vp(fp, &args->cap, &args->map) != 0 )
+                               rep = (-EINVAL);        
+               }
+
+               KMEM_FREE(args, sizeof(ELANCTRL_DESTROY_VP_STRUCT ));
+
+               break;          
+       }
+
+       case ELANCTRL_GET_CAPS  :
+       {
+               ELANCTRL_GET_CAPS_STRUCT args;
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_GET_CAPS_STRUCT)))
+                       return (-EFAULT);
+
+               /* uses copyin/copyout */
+               if (elan_get_caps(args.number_of_results, args.array_size, args.caps) != 0 )
+                       return (-EINVAL);
+               break;          
+       }
+       case ELANCTRL_DEBUG_DUMP :
+       {
+               elan_cap_dump();
+               elan_dev_dump();
+
+               break;
+       }
+       case ELANCTRL_DEBUG_BUFFER :
+       {
+               ELANCTRL_DEBUG_BUFFER_STRUCT args;
+
+               if (copy_from_user (&args, (void *) arg, sizeof (ELANCTRL_DEBUG_BUFFER_STRUCT)))
+                       return (-EFAULT);
+
+               /* uses copyin/copyout */
+               if ((args.size = qsnet_debug_buffer (args.buffer, args.size)) != -1 &&
+                   copy_to_user ((void *) arg, &args, sizeof (ELANCTRL_DEBUG_BUFFER_STRUCT)))
+                       return (-EFAULT);
+               break;
+       }
+       default:
+               return (-EINVAL);
+               break;
+       }
+
+       return (rep);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/Makefile   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/Makefile        2005-06-01 23:12:54.560445032 -0400
@@ -0,0 +1,31 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/elan/Makefile
+#
+
+
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/elan/Makefile
+#
+
+list-multi             := elan.o
+elan-objs      := elanmod.o device.o stats.o devinfo.o capability.o elanmod_linux.o capability_general.o bitmap.o
+export-objs            := elanmod_linux.o
+obj-$(CONFIG_QSNET)    := elan.o
+
+elan.o : $(elan-objs)
+       $(LD) -r -o $@ $(elan-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/elan/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/Makefile.conf      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/Makefile.conf   2005-06-01 23:12:54.561444880 -0400
@@ -0,0 +1,10 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       elan.o
+MODULENAME     =       elan
+KOBJFILES      =       elanmod.o device.o stats.o devinfo.o capability.o elanmod_linux.o capability_general.o bitmap.o
+EXPORT_KOBJS   =       elanmod_linux.o 
+CONFIG_NAME    =       CONFIG_QSNET
+SGALFC         =       
+# EXTRALINES START
+
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/elan/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/quadrics_version.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/quadrics_version.h      2005-06-01 23:12:54.561444880 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/elan/stats.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan/stats.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan/stats.c 2005-06-01 23:12:54.562444728 -0400
@@ -0,0 +1,277 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: stats.c,v 1.6 2003/09/24 13:55:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/stats.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan/elanmod.h>
+
+static LIST_HEAD(elan_stats_list);
+static ELAN_STATS_IDX elan_next_statidx=0;
+
+ELAN_STATS_STRUCT *
+elan_stats_find(ELAN_STATS_IDX statidx)
+{
+       struct list_head     *tmp;
+       ELAN_STATS_STRUCT *ptr=NULL;
+
+       list_for_each(tmp, &elan_stats_list) {
+               ptr = list_entry(tmp, ELAN_STATS_STRUCT , node);
+               if ( ptr->statidx == statidx ) 
+                       return ptr;
+       }
+
+       ELAN_DEBUG1 (ELAN_DBG_CTRL, "elan_stats_find failed %d\n", statidx);    
+       return NULL;
+}
+
+ELAN_STATS_STRUCT *
+elan_stats_find_by_name(caddr_t block_name)
+{
+       struct list_head     *tmp;
+       ELAN_STATS_STRUCT *ptr=NULL;
+
+       list_for_each(tmp, &elan_stats_list)    {
+               ptr = list_entry(tmp, ELAN_STATS_STRUCT , node);
+               if (!strcmp(ptr->block_name, block_name)) 
+               {
+                       ELAN_DEBUG3 (ELAN_DBG_CTRL, "elan_stats_find_by_name found %s (%d,%d)\n", block_name, ptr->statidx, ptr->num_entries);  
+                       return ptr;
+               }
+       }
+
+       ELAN_DEBUG1 (ELAN_DBG_CTRL, "elan_stats_find_by_name failed %s\n", block_name);
+       return NULL;
+}
+
+ELAN_STATS_STRUCT *
+elan_stats_find_next(ELAN_STATS_IDX statidx)
+{
+       struct list_head     *tmp;
+       ELAN_STATS_STRUCT *ptr=NULL;
+
+       list_for_each(tmp, &elan_stats_list) {
+               ptr = list_entry(tmp, ELAN_STATS_STRUCT , node);
+         
+               if ( ptr->statidx > statidx ) 
+                       return ptr;       
+       }       
+
+       return NULL;
+}
+
+int 
+elan_stats_get_next_index (ELAN_STATS_IDX statidx, ELAN_STATS_IDX *next_block)
+{
+       ELAN_STATS_STRUCT *target;
+       ELAN_STATS_IDX        next = 0;
+
+       kmutex_lock(&elan_mutex);
+
+       if ((target = elan_stats_find_next(statidx)) != NULL)
+               next = target->statidx;
+
+       copyout(&next, next_block, sizeof(ELAN_STATS_IDX) );
+
+       kmutex_unlock(&elan_mutex);
+       return 0;
+}
+
+int 
+elan_stats_find_index  (caddr_t  block_name, ELAN_STATS_IDX *statidx,  uint *num_entries)
+
+{
+       ELAN_STATS_STRUCT *target;
+       ELAN_STATS_IDX        index   = 0;
+       uint                  entries = 0;
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG1(ELAN_DBG_CTRL, "elan_stats_find_index %s \n", block_name);
+
+       if ((target = elan_stats_find_by_name(block_name)) != NULL)
+       {
+               index   = target->statidx;
+               entries = target->num_entries;
+       }
+
+       ELAN_DEBUG3(ELAN_DBG_CTRL, "elan_stats_find_index found %d %d (target=%p)\n", index, entries, target);
+
+       copyout(&index,   statidx,     sizeof(ELAN_STATS_IDX));
+       copyout(&entries, num_entries, sizeof(uint));
+
+       kmutex_unlock(&elan_mutex);
+       return  ESUCCESS;
+}
+
+int 
+elan_stats_get_block_info (ELAN_STATS_IDX statidx, caddr_t  block_name, uint *num_entries)
+{
+       ELAN_STATS_STRUCT *target;
+       int                   res=EINVAL;
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG1(ELAN_DBG_CTRL, "elan_stats_get_block_info statidx %d\n",statidx);
+
+       if ((target = elan_stats_find(statidx)) != NULL)
+       {
+               ELAN_DEBUG2(ELAN_DBG_CTRL, "elan_stats_get_block_info name %s entries %d\n",block_name, *num_entries);
+               
+               copyout( target->block_name,  block_name,  ELAN_STATS_NAME_MAX_LEN);
+               copyout(&target->num_entries, num_entries, sizeof(uint));
+
+               res = ESUCCESS;
+       }
+
+       kmutex_unlock(&elan_mutex);
+       return res;
+}
+
+int 
+elan_stats_get_index_name (ELAN_STATS_IDX statidx, uint index, caddr_t name)
+{
+       ELAN_STATS_STRUCT *target;
+       int                   res=EINVAL;
+
+       kmutex_lock(&elan_mutex);
+
+       ELAN_DEBUG2(ELAN_DBG_CTRL, "elan_stats_get_index_name statidx %d index %d\n",statidx, index);
+
+       if ((target = elan_stats_find(statidx)) != NULL)
+       {
+               if ( target->ops->elan_stats_get_name== NULL) 
+               {
+                       ELAN_DEBUG0(ELAN_DBG_CTRL, "elan_stats_get_index_name no callback\n");  
+                       kmutex_unlock(&elan_mutex);
+                       return  res;
+               }
+
+               if ((res = target->ops->elan_stats_get_name(target->arg, index, name)) == 0)
+                       ELAN_DEBUG1(ELAN_DBG_CTRL, "elan_stats_get_index_name name %s\n",name); 
+
+       }
+       kmutex_unlock(&elan_mutex);
+       return  res;
+}
+
+int 
+elan_stats_get_block (ELAN_STATS_IDX statidx, uint entries, ulong *values)
+{
+       ELAN_STATS_STRUCT *target;
+       int                   res=EINVAL;
+
+       kmutex_lock(&elan_mutex);
+
+       
+       if ((target = elan_stats_find(statidx)) != NULL)
+       {
+               if ( target->ops->elan_stats_get_block == NULL) 
+               {
+                       kmutex_unlock(&elan_mutex);
+                       return  res;
+               }
+
+               res = target->ops->elan_stats_get_block(target->arg, entries, values);
+       }
+
+       kmutex_unlock(&elan_mutex);
+       return  res;
+}
+
+int 
+elan_stats_clear_block (ELAN_STATS_IDX statidx)
+{
+       ELAN_STATS_STRUCT *target;
+       int                   res=EINVAL;
+
+       kmutex_lock(&elan_mutex);
+
+       if ((target = elan_stats_find(statidx)) != NULL)
+       {
+               if ( target->ops->elan_stats_clear_block == NULL) 
+               {
+                       kmutex_unlock(&elan_mutex);
+                       return  res;
+               }
+       
+               res = target->ops->elan_stats_clear_block(target->arg);
+       }
+       kmutex_unlock(&elan_mutex);
+       return  res;
+}
+
+void
+elan_stats_next_statidx(void)
+{
+       /* XXXXX need to put not in use check here incase we loop MRH */
+       /* tho its a bigish loop :)                                   */
+       elan_next_statidx++;
+       if (!elan_next_statidx)
+               elan_next_statidx++;
+}
+
+int 
+elan_stats_register (ELAN_STATS_IDX    *statidx, 
+                       char              *block_name, 
+                       uint               num_entries,
+                       ELAN_STATS_OPS *ops,
+                       void              *arg)
+{
+       ELAN_STATS_STRUCT *target;
+
+       kmutex_lock(&elan_mutex);
+
+       /* create it and add */
+       KMEM_ALLOC(target, ELAN_STATS_STRUCT *, sizeof(ELAN_STATS_STRUCT), 1);
+       if (target == NULL)
+       {
+               kmutex_unlock(&elan_mutex);
+               return  ENOMEM;
+       }
+
+       elan_stats_next_statidx();
+
+       *statidx = elan_next_statidx;
+
+       target->statidx     = elan_next_statidx;
+       target->num_entries = num_entries;
+       target->ops         = ops;
+       target->arg         = arg;
+       strcpy(target->block_name, block_name);
+       
+       list_add_tail(&target->node, &elan_stats_list);
+
+       kmutex_unlock(&elan_mutex);
+       return  0;
+}
+
+int
+elan_stats_deregister (ELAN_STATS_IDX statidx)
+{
+       ELAN_STATS_STRUCT *target;
+
+       kmutex_lock(&elan_mutex);
+       if ((target = elan_stats_find(statidx)) != NULL)
+       {
+
+               list_del(&target->node);
+               
+               /* delete target entry */
+               KMEM_FREE(target, sizeof(ELAN_STATS_STRUCT));
+       }
+       kmutex_unlock(&elan_mutex);
+
+       return  target == NULL ? EINVAL : 0;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/context.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/context.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/context.c      2005-06-01 23:12:54.565444272 -0400
@@ -0,0 +1,2101 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: context.c,v 1.116.2.1 2004/11/12 14:24:18 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/context.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <qsnet/autoconf.h>
+#include <elan/elanmod.h>
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+#include <elan3/vmseg.h>
+#include <elan3/elan3ops.h>
+#include <elan3/elansyscall.h>
+/*
+ * Global variables configurable from /etc/system file
+ *     (OR /etc/sysconfigtab on Digital UNIX)
+ */
+int ntrapped_threads   = 64;
+int ntrapped_dmas      = 64;
+int ntrapped_events    = E3_NonSysCntxQueueSize + 128;
+int ntrapped_commands  = 64;
+int noverflow_commands = 1024;
+int nswapped_threads   = 64;
+int nswapped_dmas      = 64;
+
+#define NUM_HALTOPS    8
+
+void *SwapListsLockInfo;
+void *CmdLockInfo;
+
+static void HaltSwapContext (ELAN3_DEV *dev, void *arg);
+
+static char *OthersStateStrings[]  = {"others_running", "others_halting", "others_swapping", 
+                                     "others_halting_more", "others_swapping_more", "others_swapped"};
+
+ELAN3_CTXT *
+elan3_alloc (ELAN3_DEV *dev, int  kernel)
+{
+    ELAN3_CTXT    *ctxt;
+    int           i;
+    unsigned long flags;
+
+    PRINTF1 (DBG_DEVICE, DBG_FN, "elan3_alloc: %s\n", kernel ? "kernel" : "user");
+
+    KMEM_ZALLOC (ctxt, ELAN3_CTXT *, sizeof (ELAN3_CTXT), TRUE);
+    
+    if (ctxt == NULL)
+       return (NULL);
+
+    elan_nullcap (&ctxt->Capability);
+
+    ctxt->Device      = dev;
+    ctxt->OthersState = CTXT_OTHERS_SWAPPED;
+    ctxt->RefCnt      = 1;
+    ctxt->Position    = dev->Position;
+
+    if (kernel)
+       ctxt->Status = CTXT_DETACHED | CTXT_SWAPPED_OUT | CTXT_KERNEL;
+    else
+       ctxt->Status = CTXT_DETACHED | CTXT_SWAPPED_OUT | CTXT_NO_LWPS;
+
+    ctxt->Elan3mmu = elan3mmu_alloc (ctxt);
+
+    kcondvar_init (&ctxt->Wait);
+    kcondvar_init (&ctxt->CommandPortWait);
+    kcondvar_init (&ctxt->LwpWait);
+    kcondvar_init (&ctxt->HaltWait);
+
+    spin_lock_init (&ctxt->InputFaultLock);
+
+    kmutex_init (&ctxt->SwapListsLock);
+    kmutex_init (&ctxt->CmdPortLock);
+    kmutex_init (&ctxt->NetworkErrorLock);
+    kmutex_init (&ctxt->CmdLock);
+
+    krwlock_init (&ctxt->VpLock);
+
+    KMEM_GETPAGES (ctxt->FlagPage, ELAN3_FLAGSTATS *, 1, TRUE);
+    if (!ctxt->FlagPage)
+       goto error;
+    bzero ((char *) ctxt->FlagPage, PAGESIZE);
+
+    KMEM_ZALLOC (ctxt->CommandTraps, COMMAND_TRAP *,    sizeof (COMMAND_TRAP)    * ntrapped_commands, TRUE);
+    if (!ctxt->CommandTraps)
+       goto error;
+
+    KMEM_ZALLOC (ctxt->ThreadTraps,  THREAD_TRAP *,     sizeof (THREAD_TRAP)     * ntrapped_threads,  TRUE);
+    if (!ctxt->ThreadTraps)
+       goto error;
+
+    KMEM_ZALLOC (ctxt->DmaTraps,     DMA_TRAP *,        sizeof (DMA_TRAP)        * ntrapped_dmas,     TRUE);
+    if (!ctxt->DmaTraps)
+       goto error;
+
+    KMEM_ZALLOC (ctxt->EventCookies, EVENT_COOKIE *,    sizeof (EVENT_COOKIE)    * ntrapped_events,   TRUE);
+    if (!ctxt->EventCookies)
+       goto error;
+
+    KMEM_ZALLOC (ctxt->Commands,     CProcTrapBuf_BE *, sizeof (CProcTrapBuf_BE) * noverflow_commands,TRUE);
+    if (!ctxt->Commands)
+       goto error;
+
+    KMEM_ZALLOC (ctxt->SwapThreads,  E3_Addr *,         sizeof (E3_Addr)         * nswapped_threads,  TRUE);
+    if (!ctxt->SwapThreads)
+       goto error;
+
+    KMEM_ZALLOC (ctxt->SwapDmas,     E3_DMA_BE *,       sizeof (E3_DMA_BE)       * nswapped_dmas,     TRUE);
+    if (!ctxt->SwapDmas)
+       goto error;
+
+    /*
+     * "slop" is defined as follows :
+     *     number of entries REQUIRED to be left spare to consume all other traps
+     *     up until the time that the context can be swapped out.
+     *  
+     * CommandTrapQ : 1 command issued by main + 1 issued by the thread processor per elan
+     * ThreadTrapQ  : 2 from command + 2 input
+     * DmaTrapQ     : 2 from command + 2 input
+     * EventTrapQ   : 2 from command + 1 thread + 1 dma + 2 input + E3_NonSysCntxQueueSize
+     */
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    ELAN3_QUEUE_INIT (ctxt->CommandTrapQ, ntrapped_commands,  2);
+    ELAN3_QUEUE_INIT (ctxt->ThreadTrapQ,  ntrapped_threads,   4);
+    ELAN3_QUEUE_INIT (ctxt->DmaTrapQ,     ntrapped_dmas,      4);
+    ELAN3_QUEUE_INIT (ctxt->EventCookieQ, ntrapped_events,    MIN(E3_NonSysCntxQueueSize + 6, ntrapped_events - 6));
+    ELAN3_QUEUE_INIT (ctxt->CommandQ,     noverflow_commands, 0);
+    ELAN3_QUEUE_INIT (ctxt->SwapThreadQ,  nswapped_threads,   0);
+    ELAN3_QUEUE_INIT (ctxt->SwapDmaQ,     nswapped_dmas,      0);
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+#if defined(DIGITAL_UNIX)
+    /* Allocate the segelan for the command port */
+    if (! kernel && elan3_segelan3_create (ctxt) == NULL)
+    {
+       elan3_detach(ctxt);
+       elan3_free (ctxt);
+       return ((ELAN3_CTXT *) NULL);
+    }
+#endif
+
+    /*
+     * Initialise the Input Fault list 
+     */
+    spin_lock (&ctxt->InputFaultLock);
+    for (i = 0; i < NUM_INPUT_FAULT_SAVE; i++)
+       ctxt->InputFaults[i].Next = (i == (NUM_INPUT_FAULT_SAVE-1)) ? NULL : &ctxt->InputFaults[i+1];
+    ctxt->InputFaultList = &ctxt->InputFaults[0];
+    spin_unlock (&ctxt->InputFaultLock);
+
+    ReserveHaltOperations (dev, NUM_HALTOPS, TRUE);
+    
+    if ((ctxt->RouteTable = AllocateRouteTable (ctxt->Device, ELAN3_MAX_VPS)) == NULL)
+    {
+       PRINTF0 (DBG_DEVICE, DBG_FN, "elan3_alloc: cannot map route table\n");
+       elan3_detach(ctxt);
+       elan3_free (ctxt);
+       return ((ELAN3_CTXT *) NULL);
+    }  
+
+    return (ctxt);
+
+
+ error:
+
+    elan3_detach(ctxt);
+    elan3_free (ctxt);
+    if (ctxt->FlagPage)
+       KMEM_FREEPAGES ((void *) ctxt->FlagPage, 1);
+    if (ctxt->CommandTraps)
+       KMEM_FREE ((void *) ctxt->CommandTraps, sizeof (COMMAND_TRAP)    * ntrapped_commands);
+    if (ctxt->ThreadTraps)
+       KMEM_FREE ((void *) ctxt->ThreadTraps,  sizeof (THREAD_TRAP)     * ntrapped_threads);
+    if (ctxt->DmaTraps)
+       KMEM_FREE ((void *) ctxt->DmaTraps,     sizeof (DMA_TRAP)        * ntrapped_dmas);
+    if (ctxt->EventCookies)
+       KMEM_FREE ((void *) ctxt->EventCookies, sizeof (EVENT_COOKIE)    * ntrapped_events);
+    if (ctxt->Commands)
+       KMEM_FREE ((void *) ctxt->Commands,     sizeof (CProcTrapBuf_BE) * noverflow_commands);
+    if (ctxt->SwapThreads)
+       KMEM_FREE ((void *) ctxt->SwapThreads,  sizeof (E3_Addr)         * nswapped_threads);
+    if (ctxt->SwapDmas)
+       KMEM_FREE ((void *) ctxt->SwapDmas,     sizeof (E3_DMA_BE)       * nswapped_dmas);
+
+    kcondvar_destroy (&ctxt->Wait);
+    kcondvar_destroy (&ctxt->CommandPortWait);
+    kcondvar_destroy (&ctxt->LwpWait);
+    kcondvar_destroy (&ctxt->HaltWait);
+
+    kmutex_destroy (&ctxt->SwapListsLock);
+    kmutex_destroy (&ctxt->CmdLock);
+    kmutex_destroy (&ctxt->NetworkErrorLock);
+    spin_lock_destroy  (&ctxt->InputFaultLock);
+
+    krwlock_destroy (&ctxt->VpLock);
+
+    KMEM_FREE (ctxt, sizeof (ELAN3_CTXT));
+
+    return (NULL);
+}
+
+void
+elan3_free (ELAN3_CTXT *ctxt)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    NETERR_FIXUP *nef;
+    
+    PRINTF1 (ctxt, DBG_FN, "elan3_free: %p \n", ctxt);
+   
+    elan3_removevp (ctxt, ELAN3_INVALID_PROCESS);                      /* Remove any virtual process mappings */
+
+#if defined(DIGITAL_UNIX)
+    WaitForContext (ctxt);                                     /* wait for all references to this context to go away */
+#endif
+
+    if (ctxt->RouteTable)
+       FreeRouteTable (dev, ctxt->RouteTable);
+    ctxt->RouteTable = NULL;
+
+    elan3mmu_free (ctxt->Elan3mmu);                            /* free of our Elan3mmu  */
+
+    if (ctxt->Private)                                         /* Call back to "user" to free off  */
+       ELAN3_OP_FREE_PRIVATE (ctxt);                           /* private data */
+
+#if defined(DIGITAL_UNIX)
+    if (! CTXT_IS_KERNEL(ctxt))
+       elan3_segelan3_destroy (ctxt);                          /* Unmap the command port from the users address space. */
+#endif
+   
+    ReleaseHaltOperations (dev, NUM_HALTOPS);
+
+    if (ctxt->Input0Resolver)
+       CancelNetworkErrorResolver (ctxt->Input0Resolver);
+
+    if (ctxt->Input1Resolver)
+       CancelNetworkErrorResolver (ctxt->Input1Resolver);
+
+    while ((nef = ctxt->NetworkErrorFixups) != NULL)
+    {
+       ctxt->NetworkErrorFixups = nef->Next;
+
+       CompleteNetworkErrorFixup (ctxt, nef, ESRCH);
+    }
+
+    KMEM_FREEPAGES ((void *) ctxt->FlagPage, 1);
+
+    KMEM_FREE ((void *) ctxt->CommandTraps, sizeof (COMMAND_TRAP)    * ntrapped_commands);
+    KMEM_FREE ((void *) ctxt->ThreadTraps,  sizeof (THREAD_TRAP)     * ntrapped_threads);
+    KMEM_FREE ((void *) ctxt->DmaTraps,     sizeof (DMA_TRAP)        * ntrapped_dmas);
+    KMEM_FREE ((void *) ctxt->EventCookies, sizeof (EVENT_COOKIE)    * ntrapped_events);
+    KMEM_FREE ((void *) ctxt->Commands,     sizeof (CProcTrapBuf_BE) * noverflow_commands);
+    KMEM_FREE ((void *) ctxt->SwapThreads,  sizeof (E3_Addr)         * nswapped_threads);
+    KMEM_FREE ((void *) ctxt->SwapDmas,     sizeof (E3_DMA_BE)       * nswapped_dmas);
+
+    kcondvar_destroy (&ctxt->Wait);
+    kcondvar_destroy (&ctxt->CommandPortWait);
+    kcondvar_destroy (&ctxt->LwpWait);
+    kcondvar_destroy (&ctxt->HaltWait);
+
+    kmutex_destroy (&ctxt->SwapListsLock);
+    kmutex_destroy (&ctxt->CmdLock);
+    kmutex_destroy (&ctxt->NetworkErrorLock);
+    spin_lock_destroy  (&ctxt->InputFaultLock);
+
+    krwlock_destroy (&ctxt->VpLock);
+
+    KMEM_FREE (ctxt, sizeof (ELAN3_CTXT));
+}
+
+int 
+elan3_doattach(ELAN3_CTXT *ctxt, ELAN_CAPABILITY *cap)
+{
+    unsigned long pgnum = ((cap->cap_mycontext & MAX_ROOT_CONTEXT_MASK) * sizeof (E3_CommandPort)) / PAGE_SIZE;
+    unsigned long pgoff = ((cap->cap_mycontext & MAX_ROOT_CONTEXT_MASK) * sizeof (E3_CommandPort)) & (PAGE_SIZE-1);
+    ELAN3_DEV     *dev   = ctxt->Device;
+    int           res   = ESUCCESS;
+    unsigned long flags;
+
+    /* Map in the command port for this context */
+    if (MapDeviceRegister (dev, ELAN3_BAR_COMMAND_PORT, &ctxt->CommandPage, pgnum * PAGE_SIZE, PAGE_SIZE, &ctxt->CommandPageHandle) != ESUCCESS)
+    {
+       PRINTF0 (ctxt, DBG_FN, "elan3_doattach: MapDeviceRegister failed");
+       return (EINVAL);
+    }
+
+    ctxt->CommandPort = ctxt->CommandPage + pgoff;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    res = 0;
+    if (ELAN3_DEV_CTX_TABLE(dev,cap->cap_mycontext) != NULL)
+       res = EBUSY;
+    else
+    {
+       if ((res = elan3mmu_attach (ctxt->Device, cap->cap_mycontext, ctxt->Elan3mmu, 
+                                   ctxt->RouteTable->Table, ctxt->RouteTable->Size-1)) == 0)
+       {
+           ELAN3_DEV_CTX_TABLE(dev,cap->cap_mycontext) = ctxt;
+           ctxt->Capability                            = *cap;
+       }
+    }
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    if (res == ESUCCESS)
+       elan3_swapin (ctxt, CTXT_DETACHED);
+    else 
+    {
+       UnmapDeviceRegister (dev, &ctxt->CommandPageHandle);
+       ctxt->CommandPage = (ioaddr_t) 0; 
+       ctxt->CommandPort = (ioaddr_t) 0;
+    }
+
+    return (res);
+}
+
+void
+elan3_destroy_callback( void * args, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map)
+{
+    if (map == NULL) 
+    {
+       /* the cap is being destroyed */
+       PRINTF0 (NULL, DBG_VP, "elan3_destroy_callback: the cap is being destroyed \n");
+    }
+    else
+    {
+       /* the map is being destroyed */
+       PRINTF0 (NULL, DBG_VP, "elan3_destroy_callback: the map is being destroyed \n");
+    }
+}
+
+int
+elan3_attach (ELAN3_CTXT *ctxt, ELAN_CAPABILITY *cap)
+{
+    ELAN3_DEV *dev = ctxt->Device;
+    int type;
+    int res;
+
+    switch (type = elan3_validate_cap (dev, cap, ELAN_USER_ATTACH))
+    {
+    case ELAN_CAP_OK:
+       /* nothing */
+       break;
+
+    case ELAN_CAP_RMS:
+       if ((res = elan_attach_cap(cap, dev->Devinfo.dev_rail, ctxt, elan3_destroy_callback)) != 0)
+           return res;
+       break;
+
+    default:
+       return (EINVAL);
+    }
+
+    if (((res = elan3_doattach(ctxt,cap)) != ESUCCESS) && (type == ELAN_CAP_RMS))
+       elan_detach_cap(cap, dev->Devinfo.dev_rail);
+
+    return res;
+}
+
+void
+elan3_detach ( ELAN3_CTXT *ctxt )
+{
+    ELAN3_DEV   *dev                 = ctxt->Device;
+    int need_to_call_elanmod_detach = 0;
+    unsigned long flags;
+
+    PRINTF1 (ctxt, DBG_FN, "elan3_detach: %p \n", ctxt );
+    
+    if (ctxt->Capability.cap_mycontext == ELAN_CAP_UNINITIALISED)
+    {
+       PRINTF0 (ctxt, DBG_FN, "elan3_detach: context not attached \n");
+       return ;
+    }
+
+    /* must you be in the ctx_table ?? */
+    
+    switch (ctxt->Capability.cap_type & ELAN_CAP_TYPE_MASK)
+    {
+    case ELAN_CAP_TYPE_BLOCK:
+    case ELAN_CAP_TYPE_CYCLIC:
+    {
+       if (ELAN3_SYSTEM_CONTEXT (ctxt->Capability.cap_mycontext))
+           return ;
+
+       if (! (ctxt->Capability.cap_type & ELAN_CAP_TYPE_HWTEST))
+           need_to_call_elanmod_detach = 1;
+
+       break;
+    }  
+    default:
+       return ;
+    }
+
+    elan3_swapout (ctxt, CTXT_DETACHED);
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    elan3mmu_detach (dev, ctxt->Capability.cap_mycontext);
+    ELAN3_DEV_CTX_TABLE(dev,ctxt->Capability.cap_mycontext) = NULL;
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    if (ctxt->CommandPage)
+    {
+       UnmapDeviceRegister (dev, &ctxt->CommandPageHandle);
+       ctxt->CommandPage = (ioaddr_t) 0;
+    }
+    
+    if (need_to_call_elanmod_detach) 
+       elan_detach_cap(&ctxt->Capability, dev->Devinfo.dev_rail);
+
+    elan_nullcap (&ctxt->Capability);
+
+}
+
+void
+elan3_dodetach ( ELAN3_CTXT *ctxt )
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    unsigned long flags;
+
+    PRINTF1 (ctxt, DBG_FN, "elan3_dodetach: %p \n", ctxt );
+    
+    if (ctxt->Capability.cap_mycontext == ELAN_CAP_UNINITIALISED)
+    {
+       PRINTF0 (ctxt, DBG_FN, "elan3_dodetach: context not attached \n");
+       return ;
+    }
+
+    elan3_swapout (ctxt, CTXT_DETACHED);
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    elan3mmu_detach (dev, ctxt->Capability.cap_mycontext);
+    ELAN3_DEV_CTX_TABLE(dev,ctxt->Capability.cap_mycontext) = NULL;
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    if (ctxt->CommandPage)
+    {
+       UnmapDeviceRegister (dev, &ctxt->CommandPageHandle);
+       ctxt->CommandPage = (ioaddr_t) 0;
+    }
+    
+    elan_nullcap (&ctxt->Capability);
+}
+
+void
+elan3_swapin (ELAN3_CTXT *ctxt, int reason)
+{
+    ELAN3_DEV *dev = ctxt->Device;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    ASSERT (ctxt->Status & CTXT_SWAPPED_REASONS);
+
+    PRINTF3 (ctxt, DBG_SWAP, "elan3_swapin: status %x State %s reason %x\n", 
+            ctxt->Status, OthersStateStrings[ctxt->OthersState], reason);
+
+    while (ctxt->Status & CTXT_SWAPPING_OUT)                   /* In transition */
+       kcondvar_wait (&ctxt->LwpWait, &dev->IntrLock, &flags);
+
+    if (reason == CTXT_NO_LWPS && ctxt->LwpCount++ != 0)       /* Added another LWP */
+    {
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       return;
+    }
+
+    if ((ctxt->Status & ~reason) & CTXT_SWAPPED_REASONS)
+       ctxt->Status &= ~reason;
+    else
+    {
+       ASSERT (ctxt->Status & CTXT_SWAPPED_OUT);
+       ASSERT (ctxt->OthersState == CTXT_OTHERS_SWAPPED);
+       
+       /*
+        * Will not be swapped out anymore, so ask the "user" to perform 
+        * any swapping in he needs before letting the context run again.
+        */
+       
+       ctxt->Status &= ~(CTXT_SWAPPED_OUT | CTXT_QUEUES_EMPTY | reason);
+       ctxt->OthersState = CTXT_OTHERS_RUNNING;
+
+       if (ctxt->Input0Trap.State == CTXT_STATE_OK && ctxt->Input1Trap.State == CTXT_STATE_OK)
+           SetInputterStateForContext (ctxt, 0, NULL);
+       
+       kcondvar_wakeupall (&ctxt->Wait, &dev->IntrLock);
+    }
+
+    PRINTF2 (ctxt, DBG_SWAP, "elan3_swapin: all done - status %x state %s\n",
+            ctxt->Status, OthersStateStrings[ctxt->OthersState]);
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+
+void
+elan3_swapout (ELAN3_CTXT *ctxt, int reason)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    int           cansleep;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    PRINTF3 (ctxt, DBG_SWAP, "elan3_swapout: status %x state %s reason %x\n", 
+            ctxt->Status, OthersStateStrings[ctxt->OthersState], reason);
+
+    if (reason == CTXT_NO_LWPS)
+    {
+       if (--ctxt->LwpCount != 0)                              /* Still other LWPs running */
+       {
+           spin_unlock_irqrestore (&dev->IntrLock, flags);
+           return;
+       }
+
+       kcondvar_wakeupall (&ctxt->LwpWait, &dev->IntrLock);            /* Wakeup anyone waiting on LwpCount */
+    }
+    
+    ctxt->Status |= reason;
+    
+    while (ctxt->Status & CTXT_SWAPPING_OUT)                   /* wait for someone else to finish swapping */
+       kcondvar_wait (&ctxt->LwpWait, &dev->IntrLock, &flags);         /* out */
+
+    if (ctxt->Status & CTXT_SWAPPED_OUT)
+    {
+       if (reason == CTXT_NO_LWPS)                             /* Wakeup other thread waiting on LWP exit */
+           kcondvar_wakeupall (&ctxt->LwpWait, &dev->IntrLock);
+       
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       return;
+    }
+    
+    /*
+     * mark the context as swapping out.
+     */
+    ctxt->Status |= CTXT_SWAPPING_OUT;
+    
+    if (reason != CTXT_FIXUP_NETERR)
+    {
+       /*
+        * Stop all of the lwps.
+        */
+       while (ctxt->LwpCount)
+       {
+           kcondvar_wakeupall (&ctxt->Wait, &dev->IntrLock);           /* Wake up any lwps */
+           kcondvar_wait (&ctxt->LwpWait, &dev->IntrLock, &flags);             /* then wait for them to enter elan3_swapout */
+       }
+    }
+    
+    StartSwapoutContext (ctxt, 0, NULL);
+    for (;;)
+    {
+       PRINTF0 (ctxt, DBG_SWAP, "elan3_swapout: HandleExceptions\n");
+
+       cansleep = (HandleExceptions(ctxt, &flags) == ESUCCESS);
+
+       PRINTF2 (ctxt, DBG_SWAP, "elan3_swapout: OthersState=%d cansleep=%d\n", ctxt->OthersState, cansleep);
+
+       if (ctxt->OthersState == CTXT_OTHERS_SWAPPED)
+           break;
+
+       if (cansleep)
+           kcondvar_wait (&ctxt->Wait, &dev->IntrLock, &flags);
+    }
+    PRINTF0 (ctxt, DBG_SWAP, "elan3_swapout: swapped out\n");
+    
+    ASSERT (ELAN3_QUEUE_EMPTY (ctxt->DmaTrapQ));
+    ASSERT (ELAN3_QUEUE_EMPTY (ctxt->ThreadTrapQ));
+
+    ctxt->Status |=  CTXT_SWAPPED_OUT;
+    ctxt->Status &= ~CTXT_SWAPPING_OUT;
+
+    kcondvar_wakeupall (&ctxt->LwpWait, &dev->IntrLock);
+
+    PRINTF2 (ctxt, DBG_SWAP, "elan3_swapout: all done - status %x state %s\n",
+            ctxt->Status, OthersStateStrings[ctxt->OthersState]);
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+int
+elan3_pagefault (ELAN3_CTXT *ctxt, E3_FaultSave_BE *FaultSave, int npages)
+{
+    E3_Addr     elanAddr = FaultSave->s.FaultAddress;
+    int                writeable;
+    int                res;
+
+    PRINTF3 (ctxt, DBG_FAULT, "elan3_pagefault: elanAddr %08x FSR %08x : %s\n", elanAddr, FaultSave->s.FSR.Status,
+            FaultSave->s.FSR.s.ProtFault ? "protection fault" : "pte invalid");
+    
+    /* Look at the FSR to determine the fault type etc */
+    
+    if (FaultSave->s.FSR.Status == 0)                          /* this is a target abort/parity error, so look */
+    {                                                          /* at the PCI config space registers to determine  */
+       ElanBusError (ctxt->Device);
+       return (EFAULT);                                        
+    }
+    
+    if (FaultSave->s.FSR.s.AlignmentErr)                       /* Alignment errors are always fatal. */
+    {
+       PRINTF0 (ctxt, DBG_FAULT, "elan3_pagefault: Alignment error\n");
+       return (EFAULT);
+    }
+
+    if (FaultSave->s.FSR.s.WalkBadData)                                /* Memory ECC error during a walk */
+    {
+       PRINTF0 (ctxt, DBG_FAULT, "elan3_pagefault: Memory ECC error during walk\n");
+       return (EFAULT);
+    }
+
+    if (!FaultSave->s.FSR.s.ProtFault &&                       /* DMA memory type changed */
+       !FaultSave->s.FSR.s.Walking)
+    {
+       PRINTF0 (ctxt, DBG_FAULT, "elan3_pagefault: DMA memory type changed\n");
+       return (EFAULT);
+    }
+
+    ASSERT (FaultSave->s.FSR.s.ProtFault ?                     /* protection errors, should always have a valid pte */
+           (!FaultSave->s.FSR.s.Walking || !(FaultSave->s.FSR.s.Level==3) ||  FaultSave->s.FSR.s.FaultPte == ELAN3_ET_PTE) : 
+           FaultSave->s.FSR.s.FaultPte == ELAN3_ET_INVALID);   /* otherwise it must be an invalid pte */
+
+    /*
+     * Determine whether to fault for a 'write' from the access permissions we need, and not
+     * from the access type (WrAcc).
+     */
+    writeable = (FaultSave->s.FSR.s.AccTypePerm & (1 << FSR_WritePermBit));
+
+    /* Check that we have the right permissions for this access type. */
+    if ((res = elan3mmu_checkperm (ctxt->Elan3mmu, (elanAddr&PAGEMASK), npages*PAGESIZE, FaultSave->s.FSR.s.AccTypePerm)) != 0)
+    {
+       PRINTF1 (ctxt, DBG_FAULT, "elan3_pagefault: %s\n", (res == ENOMEM) ? "no protection mapping" : "protection error");
+       
+       return (res);
+    }
+
+    res = LoadElanTranslation (ctxt, (elanAddr&PAGEMASK), npages*PAGESIZE, FaultSave->s.FSR.s.ProtFault, writeable);
+
+    if (res == ESUCCESS)
+    {
+       BumpStat (ctxt->Device, PageFaults);
+       BumpUserStat (ctxt, PageFaults);
+    }
+
+    PRINTF1 (ctxt, DBG_FAULT, "elan3_pagefault: -> %d\n", res);
+
+    return (res);
+}
+
+void
+elan3_block_inputter (ELAN3_CTXT *ctxt, int block)
+{
+    ELAN3_DEV *dev = ctxt->Device;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    
+    if (block)
+       ctxt->Status |= CTXT_USER_FILTERING;
+    else
+       ctxt->Status &= ~CTXT_USER_FILTERING;
+
+    if (ctxt->Capability.cap_mycontext != ELAN_CAP_UNINITIALISED)
+       SetInputterStateForContext (ctxt, 0, NULL);
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+int
+FixupNetworkErrors (ELAN3_CTXT *ctxt, unsigned long *flags)
+{
+    ELAN3_DEV   *dev = ctxt->Device;
+    NETERR_FIXUP *nef;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+    
+    if (ctxt->NetworkErrorFixups == NULL)
+       return (ESUCCESS);
+
+    spin_unlock_irqrestore (&dev->IntrLock, *flags);
+    
+    kmutex_lock (&ctxt->NetworkErrorLock);                     /* single thread while fixing up errors */
+    elan3_swapout (ctxt, CTXT_FIXUP_NETERR);
+
+    spin_lock_irqsave (&dev->IntrLock, *flags);
+    while ((nef = ctxt->NetworkErrorFixups) != NULL)
+    {
+       ctxt->NetworkErrorFixups = nef->Next;
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+
+       if (ELAN3_OP_FIXUP_NETWORK_ERROR (ctxt, nef) == OP_FAILED)
+           CompleteNetworkErrorFixup (ctxt, nef, EINVAL);
+
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, *flags);
+
+    elan3_swapin (ctxt, CTXT_FIXUP_NETERR);
+
+    kmutex_unlock (&ctxt->NetworkErrorLock);
+    spin_lock_irqsave (&dev->IntrLock, *flags);
+    return (EAGAIN);
+}
+
+int
+CompleteNetworkErrorResolver (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, NETERR_RESOLVER *rvp)
+{
+    int state;
+
+    switch (rvp->Status)
+    {
+    case ESUCCESS:
+       /*
+        * the item still existed at the source - if it's a wait for EOP transaction
+        * then the source will retry - otherwise the remote event will have been
+        * cleared and we should execute it
+        */
+       PRINTF1 (ctxt, DBG_NETERR, "CompleteNetworkErrorResolver: ESUCCESS zero WaitForEopTransaction %p\n", trap->WaitForEopTransaction);
+
+       state = trap->WaitForEopTransaction ? CTXT_STATE_OK : CTXT_STATE_NEEDS_RESTART;
+
+       break;
+
+    case ESRCH:        
+       /*
+        * the item was not found at the source - we should always execute the transaction
+        * since it will never be resent
+        */
+       PRINTF1 (ctxt, DBG_NETERR, "CompleteNetworkErrorResolver: ESRCH execute WaitForEopTransaction %p\n", trap->WaitForEopTransaction);
+       state = CTXT_STATE_NEEDS_RESTART;
+       break;
+
+    default:                                                   /* other errors */
+       PRINTF1 (ctxt, DBG_NETERR, "CompleteNetworkErrorResolver: %d\n", rvp->Status);
+       if (ElanException (ctxt, EXCEPTION_NETWORK_ERROR, INPUT_PROC, trap, &rvp) == OP_HANDLED)
+           state = CTXT_STATE_NEEDS_RESTART;
+       else
+           state = CTXT_STATE_OK;
+       break;
+    }
+
+    FreeNetworkErrorResolver (rvp);
+
+    return (state);
+}
+
+int
+HandleExceptions (ELAN3_CTXT *ctxt, unsigned long *flags)
+{
+    ELAN3_DEV        *dev    = ctxt->Device;
+    THREAD_TRAP      tproc;
+    DMA_TRAP         dproc;
+    NETERR_RESOLVER *rvp;
+    int                     state;
+
+    if (ctxt->Status & CTXT_COMMAND_OVERFLOW_ERROR)
+    {
+       ctxt->Status &= ~CTXT_COMMAND_OVERFLOW_ERROR;
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       ElanException (ctxt, EXCEPTION_COMMAND_OVERFLOW, COMMAND_PROC, NULL);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       return (EAGAIN);
+    }
+    
+    if (! ELAN3_QUEUE_BACK_EMPTY (ctxt->CommandTrapQ))
+    {
+       /* XXXX: unmap translations to the command port */
+
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       ResolveCProcTrap (ctxt);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       return (EAGAIN);
+    }
+    
+    if (ctxt->Input0Trap.State == CTXT_STATE_TRAPPED)
+    {
+       ctxt->Input0Trap.State = CTXT_STATE_RESOLVING;
+
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       ResolveIProcTrap (ctxt, &ctxt->Input0Trap, &ctxt->Input0Resolver);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       return (EAGAIN);
+    }
+
+    if (ctxt->Input1Trap.State == CTXT_STATE_TRAPPED)
+    {
+       ctxt->Input1Trap.State = CTXT_STATE_RESOLVING;
+
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       ResolveIProcTrap (ctxt, &ctxt->Input1Trap, &ctxt->Input1Resolver);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       return (EAGAIN);
+    }
+
+    if ((rvp = ctxt->Input0Resolver) != NULL && rvp->Completed)
+    {
+       ASSERT (ctxt->Input0Trap.State == CTXT_STATE_NETWORK_ERROR);
+
+       ctxt->Input0Resolver = NULL;
+       
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       state = CompleteNetworkErrorResolver (ctxt, &ctxt->Input0Trap, rvp);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       ctxt->Input0Trap.State = state;
+       return (EAGAIN);
+    }
+
+    if ((rvp = ctxt->Input1Resolver) != NULL && rvp->Completed)
+    {
+       ASSERT (ctxt->Input1Trap.State == CTXT_STATE_NETWORK_ERROR);
+
+       ctxt->Input1Resolver = NULL;
+       
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       state = CompleteNetworkErrorResolver (ctxt,&ctxt->Input1Trap, rvp);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       ctxt->Input1Trap.State = state;
+       return (EAGAIN);
+    }
+
+    if (NextTProcTrap (ctxt, &tproc))
+    {
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       ResolveTProcTrap (ctxt, &tproc);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       return (EAGAIN);
+    }
+    ctxt->Status &= ~CTXT_THREAD_QUEUE_FULL;
+
+    if (NextDProcTrap (ctxt, &dproc))
+    {
+       spin_unlock_irqrestore (&dev->IntrLock, *flags);
+       ResolveDProcTrap (ctxt, &dproc);
+       spin_lock_irqsave (&dev->IntrLock, *flags);
+       return (EAGAIN);
+    }
+    ctxt->Status &= ~CTXT_DMA_QUEUE_FULL;
+
+    /* Handle all event interrupts. */
+    if (! ELAN3_QUEUE_EMPTY (ctxt->EventCookieQ))
+    {
+       while (! ELAN3_QUEUE_EMPTY (ctxt->EventCookieQ))
+       {
+           E3_uint32 cookie = *ELAN3_QUEUE_FRONT (ctxt->EventCookieQ, ctxt->EventCookies);
+
+           ELAN3_QUEUE_REMOVE (ctxt->EventCookieQ);
+
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           if (ELAN3_OP_EVENT (ctxt, cookie, OP_LWP) != OP_DEFER)
+               spin_lock_irqsave (&dev->IntrLock, *flags);
+           else
+           {
+               spin_lock_irqsave (&dev->IntrLock, *flags);     /* place the cookie back on the queue. */
+                                                               /* note we place it on the front to ensure  */
+               ELAN3_QUEUE_ADD_FRONT (ctxt->EventCookieQ);     /* event ordering. */
+               *ELAN3_QUEUE_FRONT (ctxt->EventCookieQ, ctxt->EventCookies) = cookie;
+           }
+       }
+       return (EAGAIN);
+    }
+    ctxt->Status &= ~CTXT_EVENT_QUEUE_FULL;
+
+    if (! ELAN3_QUEUE_EMPTY (ctxt->SwapDmaQ))
+    {
+       while (! ELAN3_QUEUE_EMPTY (ctxt->SwapDmaQ))
+       {
+           E3_DMA_BE DmaDesc = *ELAN3_QUEUE_FRONT (ctxt->SwapDmaQ, ctxt->SwapDmas);
+
+           ELAN3_QUEUE_REMOVE (ctxt->SwapDmaQ);
+
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           RestartDmaDesc (ctxt, &DmaDesc);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+       }
+       return (EAGAIN);
+    }
+    
+    if (! ELAN3_QUEUE_EMPTY (ctxt->SwapThreadQ))
+    {
+       while (! ELAN3_QUEUE_EMPTY (ctxt->SwapThreadQ))
+       {
+           E3_Addr StackPointer = *ELAN3_QUEUE_FRONT (ctxt->SwapThreadQ, ctxt->SwapThreads);
+
+           ELAN3_QUEUE_REMOVE (ctxt->SwapThreadQ);
+
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           ReissueStackPointer (ctxt, StackPointer);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+       }
+       return (EAGAIN);
+    }
+    
+    switch (ctxt->OthersState)
+    {
+    case CTXT_OTHERS_SWAPPING:
+       if (! (ctxt->Status & CTXT_OTHERS_REASONS))
+           ctxt->OthersState = CTXT_OTHERS_RUNNING;
+       else
+           ctxt->OthersState = CTXT_OTHERS_SWAPPED;
+
+       PRINTF1 (ctxt, DBG_LWP, "HandleExceptions: OthersState : swapping -> %s\n", OthersStateStrings[ctxt->OthersState]);
+           
+       break;
+
+    case CTXT_OTHERS_SWAPPING_MORE:
+       ctxt->OthersState = CTXT_OTHERS_HALTING_MORE;
+       QueueHaltOperation (dev, 0, NULL, INT_DProcHalted | INT_TProcHalted, HaltSwapContext, ctxt);
+
+       PRINTF1 (ctxt, DBG_LWP, "HandleExceptions: OthersState : swapping_more -> %s\n", OthersStateStrings[ctxt->OthersState]);
+       break;
+    }
+    return (ESUCCESS);
+}
+
+int
+RestartContext (ELAN3_CTXT *ctxt, unsigned long *flags)
+{
+    ELAN3_DEV *dev = ctxt->Device;
+    int       res;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+
+    PRINTF1 (ctxt, DBG_LWP, "RestartContext: status %x\n", ctxt->Status);
+
+    if (! (ctxt->Status & CTXT_OTHERS_REASONS))
+    {
+       if (! ELAN3_QUEUE_FRONT_EMPTY (ctxt->CommandTrapQ) || ! ELAN3_QUEUE_EMPTY(ctxt->CommandQ))
+       {
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           RestartCProcTrap (ctxt);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+           return (EAGAIN);
+       }
+
+       if (ctxt->Input0Trap.State == CTXT_STATE_NEEDS_RESTART)
+       {
+           ctxt->Input0Trap.State = CTXT_STATE_EXECUTING;
+
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           res = RestartIProcTrap (ctxt, &ctxt->Input0Trap);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+           
+           if (res == ESUCCESS)
+               ctxt->Input0Trap.State = CTXT_STATE_OK;
+           else
+               ctxt->Input0Trap.State = CTXT_STATE_NEEDS_RESTART;
+           return (EAGAIN);
+       }
+
+       if (ctxt->Input1Trap.State == CTXT_STATE_NEEDS_RESTART)
+       {
+           ctxt->Input1Trap.State = CTXT_STATE_EXECUTING;
+
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           res = RestartIProcTrap (ctxt, &ctxt->Input1Trap);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+
+           if (res == ESUCCESS)
+               ctxt->Input1Trap.State = CTXT_STATE_OK;
+           else
+               ctxt->Input1Trap.State = CTXT_STATE_NEEDS_RESTART;
+           return (EAGAIN);
+       }
+
+       if (SetEventsNeedRestart (ctxt))
+       {
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           RestartSetEvents (ctxt);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+           return (EAGAIN);
+       }
+
+       SetInputterStateForContext (ctxt, 0, NULL);
+
+       if (TProcNeedsRestart (ctxt))
+       {
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+
+           LoadCommandPortTranslation (ctxt);
+           RestartTProcItems (ctxt);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+           return (EAGAIN);
+       }
+
+       if (DProcNeedsRestart (ctxt))
+       {
+           spin_unlock_irqrestore (&dev->IntrLock, *flags);
+           RestartDProcItems (ctxt);
+           spin_lock_irqsave (&dev->IntrLock, *flags);
+           return (EAGAIN);
+       }
+
+       if (ELAN3_QUEUE_EMPTY (ctxt->CommandTrapQ))
+       {
+           PRINTF1 (ctxt, DBG_LWP, "RestartContext: setting Command Flag at %p to 0\n", &ctxt->FlagPage->CommandFlag);
+
+           ctxt->FlagPage->CommandFlag = 0;
+
+           if (ctxt->Status & CTXT_WAITING_COMMAND)
+           {
+               PRINTF0 (ctxt, DBG_LWP, "RestartContext: waking up threads waiting for commandport\n");
+               
+               ctxt->Status &= ~CTXT_WAITING_COMMAND;
+               
+               kcondvar_wakeupall (&ctxt->CommandPortWait, &dev->IntrLock);
+           }
+       }
+    }
+
+    return (ESUCCESS);
+}
+
+static void
+HaltSwapContext (ELAN3_DEV *dev, void *arg)
+{
+    ELAN3_CTXT        *ctxt    = (ELAN3_CTXT *) arg;
+    int                      SysCntx = (ctxt->Capability.cap_mycontext & SYS_CONTEXT_BIT);
+    E3_ThreadQueue_BE thread;
+    E3_DMA_BE         dma;
+    sdramaddr_t       FPtr, BPtr;
+    sdramaddr_t              Base, Top;
+    u_int           *runCount;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    ASSERT (ctxt->OthersState == CTXT_OTHERS_HALTING || ctxt->OthersState == CTXT_OTHERS_HALTING_MORE);
+
+    PRINTF2 (ctxt, DBG_SWAP, "HaltSwapContext: status %x state %s\n", ctxt->Status, OthersStateStrings[ctxt->OthersState]);
+
+    if (! (ctxt->Status & CTXT_OTHERS_REASONS))
+    {
+       if (ctxt->OthersState == CTXT_OTHERS_HALTING_MORE)
+       {
+           runCount = SysCntx ? &dev->HaltAllCount : &dev->HaltNonContext0Count;
+
+           if (--(*runCount) == 0)
+               SetSchedStatusRegister (dev, 0, NULL);
+       }
+       ctxt->OthersState = CTXT_OTHERS_RUNNING;
+       
+       PRINTF0 (ctxt, DBG_SWAP, "HaltSwapContext: no more reason to swap -> others_running\n");
+
+       kcondvar_wakeupall (&ctxt->Wait, &dev->IntrLock);
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       return;
+    }
+
+    /*
+     * Capture all other processors since we're not being responsive to 
+     * the command processor interrupt.
+     */
+    CAPTURE_CPUS();
+
+    if (SysCntx)
+    {
+       FPtr = read_reg32 (dev, TProc_SysCntx_FPtr);
+       BPtr = read_reg32 (dev, TProc_SysCntx_BPtr);
+       Base = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxThreadQueue[0]);
+       Top  = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxThreadQueue[E3_SysCntxQueueSize-1]);
+    }
+    else
+    {
+       FPtr  = read_reg32 (dev, TProc_NonSysCntx_FPtr);
+       BPtr  = read_reg32 (dev, TProc_NonSysCntx_BPtr);
+       Base  = dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxThreadQueue[0]);
+       Top   = dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxThreadQueue[E3_NonSysCntxQueueSize-1]);
+    }
+
+    while (FPtr != BPtr)
+    {
+       elan3_sdram_copyq_from_sdram (dev, FPtr, (void *) &thread, sizeof (E3_ThreadQueue_BE));
+       
+       if (thread.s.Context == ctxt->Capability.cap_mycontext)
+       {
+           if (ELAN3_QUEUE_FULL (ctxt->SwapThreadQ))
+               break;
+           
+           *ELAN3_QUEUE_BACK(ctxt->SwapThreadQ, ctxt->SwapThreads) = thread.s.Thread;
+           ELAN3_QUEUE_ADD (ctxt->SwapThreadQ);
+           
+           /*
+            * Remove this entry from the queue by replacing it with 
+            * the "magic" thread value.
+            *
+            * NOTE: we must preserve the SYS_CONTEXT_BIT since the Elan uses this
+            * to mark the approriate run queue as empty.
+            */
+           thread.s.Context = SysCntx ? SYS_CONTEXT_BIT : 0;
+           thread.s.Thread  = VanishingStackPointer;
+
+           elan3_sdram_copyq_to_sdram (dev, (void *) &thread, FPtr, sizeof (E3_ThreadQueue_BE));
+       }
+       
+       FPtr = (FPtr == Top) ? Base : FPtr + sizeof (E3_ThreadQueue);
+    }
+
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProc.s.FSR)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData0.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData1.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData2.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData3.s.FSR.Status)) == 0);
+
+    if (SysCntx)
+    {
+       FPtr  = read_reg32 (dev, DProc_SysCntx_FPtr);
+       BPtr  = read_reg32 (dev, DProc_SysCntx_BPtr);
+       Base  = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[0]);
+       Top   = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[E3_SysCntxQueueSize-1]);
+    }
+    else
+    {
+       FPtr  = read_reg32 (dev, DProc_NonSysCntx_FPtr);
+       BPtr  = read_reg32 (dev, DProc_NonSysCntx_BPtr);
+       Base  = dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxDmaQueue[0]);
+       Top   = dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxDmaQueue[E3_NonSysCntxQueueSize-1]);
+    }
+
+    while (FPtr != BPtr)
+    {
+       elan3_sdram_copyq_from_sdram (dev, FPtr, &dma, sizeof (E3_DMA_BE));
+           
+       if (dma.s.dma_u.s.Context == ctxt->Capability.cap_mycontext)
+       {
+           if (ELAN3_QUEUE_FULL (ctxt->SwapDmaQ))
+               break;
+           
+           *ELAN3_QUEUE_BACK (ctxt->SwapDmaQ, ctxt->SwapDmas) = dma;
+           ELAN3_QUEUE_ADD (ctxt->SwapDmaQ);
+
+           /*
+            * Remove the DMA from the queue by replacing it with one with
+            * zero size and no events.
+            *
+            * NOTE: we must preserve the SYS_CONTEXT_BIT since the Elan uses this
+            * to mark the approriate run queue as empty.
+            */
+           dma.s.dma_type            = ((SysCntx ? SYS_CONTEXT_BIT : 0) << 16);
+           dma.s.dma_size            = 0;
+           dma.s.dma_source          = (E3_Addr) 0;
+           dma.s.dma_dest            = (E3_Addr) 0;
+           dma.s.dma_destCookieVProc = (E3_Addr) 0;
+           dma.s.dma_srcEvent        = (E3_Addr) 0;
+           dma.s.dma_srcCookieVProc  = (E3_Addr) 0;
+
+           elan3_sdram_copyq_to_sdram (dev, &dma, FPtr, sizeof (E3_DMA_BE));
+       }
+
+       FPtr = (FPtr == Top) ? Base : FPtr + sizeof (E3_DMA);
+    }
+
+    /*
+     * Release the other processors now before signalling the LWP.
+     */
+    RELEASE_CPUS();
+
+    if (! ELAN3_QUEUE_FULL (ctxt->SwapDmaQ) && !ELAN3_QUEUE_FULL (ctxt->SwapThreadQ))
+    {
+       /*
+        * We've compleletly emptied the elan queues of items in this
+        * context, so we now mark it as fully swapped out.
+        */
+       if (ctxt->OthersState == CTXT_OTHERS_HALTING_MORE)
+       {
+           runCount = SysCntx ? &dev->HaltAllCount : &dev->HaltNonContext0Count;
+           
+           if (--(*runCount) == 0)
+               SetSchedStatusRegister (dev, 0, NULL);
+           
+       }
+       PRINTF0 (ctxt, DBG_SWAP, "HaltSwapContext: queues emptied -> others_swapping\n");
+
+       ctxt->OthersState = CTXT_OTHERS_SWAPPING;
+       kcondvar_wakeupall (&ctxt->Wait, &dev->IntrLock);
+    }
+    else
+    {
+       if (ctxt->OthersState == CTXT_OTHERS_HALTING)
+       {
+           runCount = SysCntx ? &dev->HaltAllCount : &dev->HaltNonContext0Count;
+           
+           if ((*runCount)++ == 0)
+               SetSchedStatusRegister (dev, 0, NULL);
+       }
+       PRINTF0 (ctxt, DBG_SWAP, "HaltSwapContext: queues not emptied -> others_swapping_more\n");
+
+       ctxt->OthersState = CTXT_OTHERS_SWAPPING_MORE;
+       kcondvar_wakeupone (&ctxt->Wait, &dev->IntrLock);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+void
+UnloadCommandPageMapping (ELAN3_CTXT *ctxt)
+{
+    /*
+     * Unload the Elan translations,  and flag the main processor to stall after 
+     * issueing its next command.
+     */
+    if (ctxt->CommandPageMapping != NULL && (ctxt->Status & CTXT_COMMAND_MAPPED_ELAN))
+    {
+       ELAN3MMU_RGN *rgn = elan3mmu_rgnat_main (ctxt->Elan3mmu, ctxt->CommandPageMapping);
+       
+       if (rgn != NULL)
+       {
+           E3_Addr eaddr = rgn->rgn_ebase + (ctxt->CommandPageMapping - rgn->rgn_mbase);
+           
+           PRINTF1 (ctxt, DBG_INTR, "UnloadCommandPageMapping: unmapping command port at addr %08x\n", eaddr);
+           
+           elan3mmu_unload (ctxt->Elan3mmu, eaddr, PAGESIZE, PTE_UNLOAD);
+       }
+       
+       ctxt->Status &= ~CTXT_COMMAND_MAPPED_ELAN;
+    }
+}
+
+void
+StartSwapoutContext (ELAN3_CTXT *ctxt, E3_uint32 Pend, E3_uint32 *Maskp)
+{
+    ELAN3_DEV   *dev     = ctxt->Device;
+    int                SysCntx = (ctxt->Capability.cap_mycontext & SYS_CONTEXT_BIT);
+    u_int      *runCount;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+
+    PRINTF2 (ctxt, DBG_SWAP, "StartSwapoutContext: Status %x OthersState %s\n",
+            ctxt->Status, OthersStateStrings [ctxt->OthersState]);
+    /*
+     * Disable the inputters,  we should already have a reason for it.
+     */
+    SetInputterStateForContext (ctxt, Pend, Maskp);
+
+    UnloadCommandPageMapping (ctxt);
+
+    /* 
+     * Flag main processor to stall after issueing next command
+     */
+    PRINTF1 (ctxt, DBG_SWAP, "StartSwapoutContext: setting Command Flag at %p to 1\n", &ctxt->FlagPage->CommandFlag);
+
+    ctxt->FlagPage->CommandFlag = 1;
+
+    PRINTF1 (ctxt, DBG_SWAP, "StartSwapoutContext: OthersState=%d\n", ctxt->OthersState);
+
+    /*
+     * And queue a haltop to stop the queues and clear it out.
+     */
+    switch (ctxt->OthersState)
+    {
+    case CTXT_OTHERS_RUNNING:
+       PRINTF0 (ctxt, DBG_SWAP, "StartSwapoutContext: -> others_halting\n");
+
+       ctxt->OthersState = CTXT_OTHERS_HALTING;
+
+       QueueHaltOperation (dev, Pend, Maskp, INT_DProcHalted | INT_TProcHalted, HaltSwapContext, ctxt);
+       break;
+       
+    case CTXT_OTHERS_SWAPPING:
+       PRINTF0 (ctxt, DBG_SWAP, "StartSwapoutContext: -> others_swapping_more\n");
+       ctxt->OthersState = CTXT_OTHERS_SWAPPING_MORE;
+
+       runCount = SysCntx ? &dev->HaltAllCount : &dev->HaltNonContext0Count;
+           
+       if ((*runCount)++ == 0)
+           SetSchedStatusRegister (dev, Pend, Maskp);
+       break;
+    default:
+       PRINTF1 (ctxt, DBG_SWAP, "StartSwapoutContext: OthersState=%d\n", ctxt->OthersState);
+       break;
+    }
+}
+
+#if defined(DIGITAL_UNIX)
+/* temporary tweaks to priority bump */
+int lwp_do_prio = 1;
+int lwp_do_nxm = 1;
+int lwp_prio = BASEPRI_USER-1;
+#elif defined(LINUX)
+/* This is the default nice level for the helper LWP */
+int LwpNice = -1;
+#endif
+
+int
+elan3_lwp (ELAN3_CTXT *ctxt)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    int                  res;
+    unsigned long flags;
+
+    PRINTF1 (ctxt, DBG_LWP, "elan3_lwp: started, context 0x%x\n", ctxt->Capability.cap_mycontext);
+
+#if defined(DIGITAL_UNIX)
+    {
+        thread_t mythread = current_thread();
+        if (lwp_do_prio && (lwp_do_nxm || !IS_NXM_TASK(mythread->task)))
+        {
+            mythread->priority = mythread->sched_pri = lwp_prio;
+            mythread->max_priority = BASEPRI_HIGHEST;
+            (void) thread_priority(mythread, lwp_prio, 0, 1);
+        }
+    }
+#elif defined(LINUX)
+    {
+       /* Do the priority trick for the helper LWP so that it
+        * runs in preferance to the user threads which may be
+        * burning CPU waiting for a trap to be fixed up
+        */
+#ifdef NO_O1_SCHED
+       if (LwpNice >= -20 && LwpNice < 20)
+           current->nice = LwpNice;
+#else
+       set_user_nice(current, LwpNice);
+#endif
+    }
+#endif
+
+    elan3_swapin (ctxt, CTXT_NO_LWPS);
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    /* If we're swapped out, and not detached (or exiting) then wait until we're swapped back in */
+    /* since otherwise we could "spin" forever continually calling elan3_lwp() */
+    if ((ctxt->Status & CTXT_SWAPPED_REASONS) && ! (ctxt->Status & (CTXT_DETACHED|CTXT_EXITING)))
+       kcondvar_waitsig (&ctxt->Wait, &dev->IntrLock, &flags);
+
+    for (;;)
+    {
+#if defined(DIGITAL_UNIX)
+        if (thread_should_halt(current_thread()) || 
+            CURSIG_CHECK(task_to_proc(current_thread()->task), u.np_uthread))
+       {
+           PRINTF1 (ctxt, DBG_LWP, "elan3_lwp: exiting on %s\n", 
+                    thread_should_halt(current_thread()) ? "halt" : "signal");
+            break;
+       }
+#endif
+
+       if (ctxt->Status & CTXT_SWAPPED_REASONS)
+       {
+           PRINTF0 (ctxt, DBG_LWP, "elan3_lwp: exiting on swapped reasons\n");
+           break;
+       }
+
+       if (! (ctxt->inhibit))
+       {
+           if (FixupNetworkErrors (ctxt, &flags) == ESUCCESS &&
+               HandleExceptions (ctxt, &flags) == ESUCCESS &&
+               RestartContext (ctxt, &flags) == ESUCCESS)
+               {
+                   if (kcondvar_waitsig (&ctxt->Wait, &dev->IntrLock, &flags) == 0)
+                   {
+                       PRINTF0 (ctxt, DBG_LWP, "elan3_lwp: exiting by kcondvar_wait_sig()\n");
+                       break;
+                   }
+               }
+       }
+       else
+       {
+           printk("elan3_lwp :: skipping as inhibited\n");
+           if (kcondvar_waitsig (&ctxt->Wait, &dev->IntrLock, &flags) == 0)
+           {
+               PRINTF0 (ctxt, DBG_LWP, "elan3_lwp: exiting by kcondvar_wait_sig()\n");
+               break;
+           }
+       }
+
+    }
+
+    /* Return EINVAL to elan3_syscall_lwp() when we want it to exit */
+    res = (ctxt->Status & (CTXT_DETACHED|CTXT_EXITING)) ? EINVAL : 0;
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    
+    elan3_swapout (ctxt, CTXT_NO_LWPS);
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    FixupNetworkErrors (ctxt, &flags);
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    return (res);
+}
+
+void
+SetInputterStateForContext (ELAN3_CTXT *ctxt, E3_uint32 Pend, E3_uint32 *Maskp)
+{
+    ELAN3_DEV  *dev          = NULL;
+    int        new_disabled = 0;
+    int               ctxnum;
+
+    ASSERT (ctxt != NULL);
+    dev  = ctxt->Device;
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+
+    new_disabled = (ctxt->Input0Trap.State != CTXT_STATE_OK ||
+                   ctxt->Input1Trap.State != CTXT_STATE_OK ||
+                   (ctxt->Status & CTXT_INPUTTER_REASONS) != 0);
+    
+
+    ctxnum   = ctxt->Capability.cap_mycontext;
+
+#ifndef __lock_lint  
+    PRINTF2 (ctxt , DBG_IPROC, "SetInputterState: ctxnum %x %s attached\n", ctxnum, ctxt->Disabled ? "disabled " : "");
+#endif /* __lock_lint */
+        
+    if (ctxt->Disabled != new_disabled)
+    {
+       PRINTF2 (ctxt, DBG_IPROC, "SetInputterState: ctxnum %x change %s\n", ctxnum, new_disabled ? "enabled to disabled" : "disabled to enabled");
+       
+       ctxt->Disabled = new_disabled;
+
+       /* synchronize the context filter for this context */
+       elan3mmu_set_context_filter (dev, ctxnum, new_disabled, Pend, Maskp);
+    }
+}
+
+int
+CheckCommandQueueFlushed (ELAN3_CTXT *ctxt, E3_uint32 cflags, int how, unsigned long *flags)
+{
+    ELAN3_DEV *dev    = ctxt->Device;
+    int       delay  = 1;
+    int i, SeenComQueueEmpty;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+    ASSERT (cflags != DmaComQueueNotEmpty || dev->HaltDmaDequeueCount != 0);
+
+    /*
+     * Flush the command processor queues and poll the queue to see it it empties.
+     */
+    if (dev->FlushCommandCount++ == 0)
+       SetSchedStatusRegister (dev, 0, NULL);
+
+    /* 
+     * Ensure previous writes have been flushed through the write buffers
+     */
+    wmb(); mmiob();
+
+    /*
+     * If the command processor traps,  or it's taking too long to observe
+     * the queue as emtpy,  then we need to force the interrupt handler to 
+     * run for us.  So queue a halt operation for the dma processor.
+     */
+    SeenComQueueEmpty = !(read_reg32 (dev, ComQueueStatus) & cflags);
+    for (i = 20; i > 0 || (how & ISSUE_COMMAND_CANT_WAIT); i--)
+    {
+       if (SeenComQueueEmpty || (read_reg32 (dev, Exts.InterruptReg) & (INT_CProc | INT_ComQueue)))
+           break;
+       
+       mb();
+       DELAY (delay);
+
+       if ((delay <<= 1) == 0) delay = 1;
+
+       SeenComQueueEmpty = !(read_reg32 (dev, ComQueueStatus) & cflags);
+    }
+
+    if (--dev->FlushCommandCount == 0)
+       SetSchedStatusRegister (dev, 0, NULL);
+
+    /*
+     * If we've seen the command queue that we're interested in with nothing in it
+     * and the command processor has not trapped then the commands we've
+     * issued have been successfully processed.
+     */
+    if (SeenComQueueEmpty && ! (read_reg32 (dev, Exts.InterruptReg) & (INT_CProc | INT_ComQueue)))
+    {
+       PRINTF0 (ctxt, DBG_CMD, "CheckCommandQueueFlushed: observed dma queue empty and command proc not trapped\n");
+
+       if (cflags == DmaComQueueNotEmpty && --dev->HaltDmaDequeueCount == 0)
+           SetSchedStatusRegister (dev, 0, NULL);
+
+       return (ISSUE_COMMAND_OK);
+    }
+
+    if ((how & ISSUE_COMMAND_CANT_WAIT) != 0)
+       return (ISSUE_COMMAND_WAIT);
+    
+    /*
+     * Halt the dma processor and wait for it to halt,  if the command we've issued has
+     * trapped then the interrupt handler will have moved it to the context structure.
+     */
+    PRINTF0 (ctxt, DBG_CMD, "CheckCommandQueueFlushed: waiting for dproc to halt\n");
+    QueueHaltOperation (dev, 0, NULL, INT_DProcHalted, WakeupLwp, ctxt);
+    while (! ctxt->Halted)
+    {
+       PRINTF1 (ctxt, DBG_CMD, "CheckCommandQueueFlushed: waiting for Halted - %d\n", ctxt->Halted);
+
+       kcondvar_wait (&ctxt->HaltWait, &dev->IntrLock, flags);
+
+       PRINTF1 (ctxt, DBG_CMD, "CheckCommandQueueFlushed: woken for Halted - %d\n", ctxt->Halted);
+    }
+    ctxt->Halted = 0;
+    
+    PRINTF0 (ctxt, DBG_CMD, "CheckCommandQueueFlushed: dproc halted, checking for trap\n");
+    
+    if (cflags == DmaComQueueNotEmpty && --dev->HaltDmaDequeueCount == 0)
+       SetSchedStatusRegister (dev, 0, NULL);
+
+    return (ELAN3_QUEUE_BACK_EMPTY (ctxt->CommandTrapQ) ? ISSUE_COMMAND_OK : ISSUE_COMMAND_TRAPPED);
+}
+
+int
+WaitForCommandPort (ELAN3_CTXT *ctxt)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    int                  res;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    if (ctxt->Status & CTXT_DETACHED)
+       res = EINVAL;
+    else 
+    {
+       if (! ELAN3_QUEUE_EMPTY (ctxt->CommandTrapQ) || (ctxt->Status & CTXT_OTHERS_REASONS))
+       {
+           ctxt->Status |= CTXT_WAITING_COMMAND;
+           if (CTXT_IS_KERNEL(ctxt))
+               kcondvar_wait (&ctxt->CommandPortWait, &dev->IntrLock, &flags);
+           else 
+               kcondvar_waitsig (&ctxt->CommandPortWait, &dev->IntrLock, &flags);
+       }
+       
+       res = (!ELAN3_QUEUE_EMPTY(ctxt->CommandTrapQ) || (ctxt->Status & CTXT_OTHERS_REASONS)) ? EAGAIN : 0;
+    }
+       
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    return (res);
+}
+
+static char *
+CommandName (int offset)
+{
+    switch (offset)
+    {
+    case offsetof (E3_CommandPort, PutDma):    return ("PutDma");
+    case offsetof (E3_CommandPort, GetDma):    return ("GetDma");
+    case offsetof (E3_CommandPort, RunThread): return ("RunThread");
+    case offsetof (E3_CommandPort, WaitEvent0):        return ("WaitEvent0");
+    case offsetof (E3_CommandPort, WaitEvent1):        return ("WaitEvent1");
+    case offsetof (E3_CommandPort, SetEvent):  return ("SetEvent");
+    default:                                   return ("Bad Command");
+    }
+}
+
+int
+IssueCommand (ELAN3_CTXT *ctxt, unsigned cmdoff, E3_Addr value, int cflags)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    int                  res;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    if ((! (cflags & ISSUE_COMMAND_FOR_CPROC) && !ELAN3_QUEUE_EMPTY (ctxt->CommandTrapQ)) || (ctxt->Status & CTXT_OTHERS_REASONS))
+    {
+       /*
+        * Cannot issue commands for non-cproc traps if command port is trapped, 
+        * nor if the dma/thread trap queues are full, or we're swapping out
+        */
+       PRINTF2 (ctxt, DBG_CMD, "IssueCommand: %s %08x -> ISSUE_COMMAND_RETRY\n",
+                CommandName (cmdoff), value);
+
+       res = ISSUE_COMMAND_RETRY;
+    }
+    else
+    {
+       PRINTF2 (ctxt, DBG_CMD, "IssueCommand: %s %08x -> ISSUE_COMMAND_OK\n",
+                CommandName (cmdoff), value);
+
+       mb();                                                   /* ensure writes to main memory completed */
+       writel (value, ctxt->CommandPort + cmdoff);             /* issue command */
+       mmiob();                                                /* and flush through IO writes */
+
+       res = ISSUE_COMMAND_OK;
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    
+    return (res);
+}
+
+int
+IssueDmaCommand (ELAN3_CTXT *ctxt, E3_Addr value, void *item, int how)
+{
+    ELAN3_DEV     *dev    = ctxt->Device;
+    int                  res;
+    unsigned long flags;
+
+    /*
+     * Since we may be issuing a command that could trap, and we're interested in
+     * the outcome, the command port trap resolving code must be locked out.
+     */
+    kmutex_lock (&ctxt->CmdLock);
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    if ((! (how & ISSUE_COMMAND_FOR_CPROC) && !ELAN3_QUEUE_EMPTY (ctxt->CommandTrapQ)) || (ctxt->Status & CTXT_OTHERS_REASONS))
+    {
+       PRINTF2 (ctxt, DBG_CMD, "IssueDmaCommand: PutDma %08x [%p] -> ISSUE_COMMAND_RETRY\n", value, item);
+
+       /*
+        * Cannot issue commands for non-cproc traps if command port is trapped, 
+        * nor if the dma/thread trap queues are full, or we're swapping out
+        */
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       kmutex_unlock (&ctxt->CmdLock);
+       return (ISSUE_COMMAND_RETRY);
+    }
+    
+    ASSERT (item == NULL || ctxt->CommandPortItem == NULL);
+
+    /*
+     * Stop the DMA processor from removing entries from the 
+     * command port, and force the command processor to do this.
+     * This means that if a trap occurs then it will be the command
+     * processor that traps.
+     */
+    if (dev->HaltDmaDequeueCount++ == 0)
+       SetSchedStatusRegister (dev, 0, NULL);
+
+    PRINTF2 (ctxt, DBG_CMD, "IssueDmaCommand: PutDma %08x [%p]\n", value, item);
+
+    /*
+     * Always issue the DMA to the 'write' command,  since we've asserted HaltDmaDequeue
+     * the command processor will read the descriptor and transfer it to the run queue. 
+     * The command processor looks at the dma_direction field to determine whether it is
+     * a read or a write and whether to alter the dma_souce of the descriptr on the run 
+     * queue
+     */
+    mb();                                                      /* ensure writes to main memory ccompleted */
+    writel (value, ctxt->CommandPort + offsetof (E3_CommandPort, PutDma));
+    mmiob();                                                   /* and flush through IO writes */
+    
+    res = CheckCommandQueueFlushed (ctxt, DmaComQueueNotEmpty, how, &flags);
+
+    if (res == ISSUE_COMMAND_TRAPPED)
+    {
+       PRINTF2 (ctxt, DBG_CMD, "IssueDmaCommand: PutDma %08x [%p] -> ISSUE_COMMAND_TRAPPED\n", value, item);
+       /*
+        * Remember the item we're issueing so that if the command port traps the item will not
+        * get freed off until the descriptor has been read after the command trap has been fixed
+        * up.
+        */
+       if (item != NULL)
+           ctxt->CommandPortItem = item;
+    }
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    kmutex_unlock (&ctxt->CmdLock);
+
+    return (res);
+}
+
+int
+WaitForDmaCommand (ELAN3_CTXT *ctxt, void *item, int how)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    int           res;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    res = CheckCommandQueueFlushed (ctxt, DmaComQueueNotEmpty, how, &flags);
+
+    if (res == ISSUE_COMMAND_TRAPPED && item != NULL)
+       ctxt->CommandPortItem = item;
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    
+    return (res);
+}
+
+void
+FixupEventTrap (ELAN3_CTXT *ctxt, int proc, void *trap, E3_uint32 TrapType, E3_FaultSave_BE *FaultSaveArea, int flags)
+{
+    ASSERT (! CTXT_IS_KERNEL (ctxt));
+
+    /*
+     * This code re-issues the part of the set event that trapped.
+     */
+    switch (TrapType)
+    {
+    case MI_ChainedEventError:
+       ElanException (ctxt, EXCEPTION_CHAINED_EVENT, proc, trap, FaultSaveArea->s.EventAddress);
+       break;
+       
+
+    case MI_SetEventReadWait:
+       /*
+        * Fault occured on the read for the event location. Just re-issue
+        * setevent using EventAddress in E3_FaultSave
+        */
+       PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_SetEventReadWait: re-issuing setevent %08x\n", 
+                FaultSaveArea->s.EventAddress);
+       
+       ReissueEvent (ctxt, (E3_Addr) FaultSaveArea->s.EventAddress, flags);
+       break;
+
+    case MI_DoSetEvent:
+    {
+       /*
+        * Fault occured because the block write of a block copy event trapped.
+        * Must grab the event type, source and dest then simulate the block copy and then
+        * perform the set. Once the block copy is started the event location cannot be read
+        * again.
+        */
+       E3_Event *EventPtr  = (E3_Event *) elan3mmu_mainaddr (ctxt->Elan3mmu, FaultSaveArea->s.EventAddress);
+       E3_uint32 EventType = fuword (&EventPtr->ev_Type);
+       
+       /*
+        * Check that the event has the block copy bit
+        * set in it,  since we couldn't trap here if it
+        * didn't
+        */
+       if ((EventType & EV_TYPE_BCOPY) != EV_TYPE_BCOPY)
+       {
+           PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_DoSetEvent: Unexpected type=%x\n", EventType);
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+           break;
+       }
+       
+       PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_DoSetEvent: RunEventType %x\n", EventType);
+
+       if (RunEventType (ctxt, FaultSaveArea, EventType))
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+
+       break;
+    }
+    
+    case MI_ThreadUpdateNonSysCntxBack:
+    case MI_ThreadUpdateSysCntxBack:
+    {
+       /*
+        * Fault occured because the block write of a block copy event trapped.
+        * Must grab the event type, source and dest then simulate the block copy and then
+        * run the thread. Once the block copy is started the event location cannot be read
+        * again.
+        */
+       E3_Event *EventPtr = (E3_Event *) elan3mmu_mainaddr (ctxt->Elan3mmu, FaultSaveArea->s.EventAddress);
+       E3_uint32 EventType = fuword (&EventPtr->ev_Type);
+
+       /*
+        * Check for the correct EventPtr type
+        */
+       if ((EventType & (EV_TYPE_MASK_THREAD|EV_TYPE_MASK_BCOPY)) != (EV_TYPE_BCOPY | EV_TYPE_THREAD))
+       {
+           PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_ThreadUpdateCntx0Back: Unexpected type=%x for setevent trap. Should be thread\n", EventType);
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+           break;
+       }
+       
+       PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_ThreadUpdateCntx0Back: RunEventType %x\n", EventType);
+       if (RunEventType (ctxt, FaultSaveArea, EventType))
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+       break;
+    }
+    
+    case MI_EventIntUpdateBPtr:
+    {
+       /*
+        * Fault occured because the block write of a block copy event trapped.
+        * Must grab the event type, source and dest then simulate the block copy and then
+        * run the dma. Once the block copy is started the event location cannot be read
+        * again.
+        */
+       E3_Event *EventPtr = (E3_Event *) elan3mmu_mainaddr (ctxt->Elan3mmu, FaultSaveArea->s.EventAddress);
+       E3_uint32 EventType = fuword (&EventPtr->ev_Type);
+
+       /*
+        * Check for the correct EventPtr type
+        */
+       if ((EventType & (EV_TYPE_MASK_EVIRQ|EV_TYPE_MASK_BCOPY)) != (EV_TYPE_BCOPY | EV_TYPE_EVIRQ))
+       {
+           PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_EventIntUpdateBPtr: Unexpected type=%x\n", EventType);
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+           break;
+       }
+
+       PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_EventIntUpdateBPtr: RunEventType %x\n", EventType);
+       if (RunEventType(ctxt, FaultSaveArea, EventType))
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+       break;
+    }
+    
+    case MI_RunDmaDesc:
+    {
+       /*
+        * Fault occured because the block write of a block copy event trapped.
+        * Must grab the event type, source and dest then simulate the block copy and then
+        * run the dma. Once the block copy is started the event location cannot be read
+        * again.
+        */
+       E3_Event *EventPtr = (E3_Event *) elan3mmu_mainaddr (ctxt->Elan3mmu, FaultSaveArea->s.EventAddress);
+       E3_uint32 EventType = fuword (&EventPtr->ev_Type);
+
+       /*
+        * Check for the correct EventPtr type
+        */
+       if ((EventType & (EV_TYPE_MASK_DMA|EV_TYPE_MASK_BCOPY)) != (EV_TYPE_BCOPY | EV_TYPE_DMA))
+       {
+           PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_RunDmaDesc: Unexpected type=%x\n", EventType);
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+           break;
+       }
+
+       PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_RunDmaDesc: RunEventType %x\n", EventType);
+       if (RunEventType(ctxt, FaultSaveArea, EventType))
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+       break;
+    }
+    
+    case MI_WaitForCntxDmaDescRead:
+    case MI_WaitForNonCntxDmaDescRead:
+       /*
+        * Fault occured on the read of the dma descriptor. Run dma using the
+        * Fault Address in FaultSave.
+        */
+       PRINTF1 (ctxt, DBG_EVENT, "FixupEventTrap: MI_WaitForCntxDmaDescRead: re-issue dma at %08x\n", FaultSaveArea->s.FaultAddress);
+       
+       RestartDmaPtr (ctxt, FaultSaveArea->s.FaultAddress);
+       break;
+    
+    case MI_FinishedSetEvent:
+       /*
+        * Fault occured because the block write of a block copy event trapped.
+        * Simulate the block copy.
+        */
+       if (SimulateBlockCopy (ctxt, FaultSaveArea->s.EventAddress))
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+       break;
+       
+    case MI_BlockCopyEvent:
+    case MI_BlockCopyWaitForReadData:
+    {
+       /*
+        * Fault occured on the read or write of the data for a block copy
+        * event. Simulate the block copy using EventAddress in E3_FaultSave. Must also sample
+        * the event type and then perform a run.
+        */
+       E3_Event *EventPtr = (E3_Event *) elan3mmu_mainaddr (ctxt->Elan3mmu, FaultSaveArea->s.EventAddress);
+       E3_uint32 EventType = fuword (&EventPtr->ev_Type);
+
+       PRINTF0 (ctxt, DBG_EVENT, "FixupEventTrap: MI_BlockCopyWaitForReadData: BCopy read fault in BCopy event. Simulating BCopy.\n");
+       
+       if (RunEventType(ctxt, FaultSaveArea, EventType))
+           ElanException (ctxt, EXCEPTION_BAD_EVENT, proc, trap, FaultSaveArea, TrapType);
+       break;
+    }
+    
+    case MI_EventQueueOverflow:
+    case MI_ThreadQueueOverflow:
+    case MI_DmaQueueOverflow:
+       /* XXXX: should handle queue overflow */
+       PRINTF0 (ctxt, DBG_EVENT, "FixupEventTrap: Queue overflow\n");
+
+       ElanException (ctxt, EXCEPTION_QUEUE_OVERFLOW, proc, trap, FaultSaveArea, TrapType);
+       break;
+
+    default:
+       ElanException (ctxt, EXCEPTION_BUS_ERROR, proc, trap, FaultSaveArea, TrapType);
+       break;
+    }
+}
+
+int
+SimulateBlockCopy (ELAN3_CTXT *ctxt, E3_Addr EventAddress)
+{
+    E3_Addr  SourcePtrElan;
+    E3_Addr  DestPtrElan;
+    unsigned DataType;
+    int      i;
+
+    if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+    {
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+       ElanException (ctxt, EXCEPTION_FAULTED, EVENT_PROC, NULL, EventAddress);
+       return (TRUE);
+    }
+
+    SourcePtrElan = ELAN3_OP_LOAD32 (ctxt, EventAddress + offsetof (E3_BlockCopyEvent, ev_Source));
+    DestPtrElan   = ELAN3_OP_LOAD32 (ctxt, EventAddress + offsetof (E3_BlockCopyEvent, ev_Dest));
+    DataType      = DestPtrElan & EV_BCOPY_DTYPE_MASK;
+    DestPtrElan  &= ~EV_BCOPY_DTYPE_MASK;
+
+
+    PRINTF3 (ctxt, DBG_EVENT, "SimulateBlockCopy: Event %08x SourcePtr %08x DestPtr %08x\n",
+            EventAddress, SourcePtrElan, DestPtrElan);
+
+    if (SourcePtrElan & EV_WCOPY)
+       ELAN3_OP_STORE32 (ctxt, DestPtrElan, SourcePtrElan);
+    else
+    {
+       /*
+        * NOTE: since the block copy could be to sdram, we issue the writes backwards,
+        *       except we MUST ensure that the last item in the block is written last.
+        */
+#if defined(__LITTLE_ENDIAN__)
+       /*
+        * For little endian cpu's we don't need to worry about the data type.
+        */
+       for (i = E3_BLK_SIZE-(2*sizeof (E3_uint64)); i >= 0; i -= sizeof (E3_uint64))
+           ELAN3_OP_STORE64 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD64 (ctxt, SourcePtrElan + i));
+
+       i = E3_BLK_SIZE - sizeof (E3_uint64);
+       ELAN3_OP_STORE64 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD64 (ctxt, SourcePtrElan + i));
+#else
+       switch (DataType)
+       {
+       case EV_TYPE_BCOPY_BYTE:
+           for (i = E3_BLK_SIZE-(2*sizeof (E3_uint8)); i >= 0; i -= sizeof (E3_uint8))
+               ELAN3_OP_STORE8 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD8 (ctxt, SourcePtrElan + i));
+           
+           i = E3_BLK_SIZE - sizeof (E3_uint8);
+           ELAN3_OP_STORE8 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD8 (ctxt, SourcePtrElan + i));
+           break;
+
+       case EV_TYPE_BCOPY_HWORD: 
+           for (i = E3_BLK_SIZE-(2*sizeof (E3_uint16)); i >= 0; i -= sizeof (E3_uint16))
+               ELAN3_OP_STORE16 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD16 (ctxt, SourcePtrElan + i));
+           
+           i = E3_BLK_SIZE - sizeof (E3_uint16);
+           ELAN3_OP_STORE16 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD16 (ctxt, SourcePtrElan + i));
+           break;
+           
+       case EV_TYPE_BCOPY_WORD:  
+           for (i = E3_BLK_SIZE-(2*sizeof (E3_uint32)); i >= 0; i -= sizeof (E3_uint32))
+               ELAN3_OP_STORE32 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD32 (ctxt, SourcePtrElan + i));
+           
+           i = E3_BLK_SIZE - sizeof (E3_uint32);
+           ELAN3_OP_STORE32 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD32 (ctxt, SourcePtrElan + i));
+           break;
+           
+       case EV_TYPE_BCOPY_DWORD: 
+           for (i = E3_BLK_SIZE-(2*sizeof (E3_uint64)); i >= 0; i -= sizeof (E3_uint64))
+               ELAN3_OP_STORE64 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD64 (ctxt, SourcePtrElan + i));
+           
+           i = E3_BLK_SIZE - sizeof (E3_uint64);
+           ELAN3_OP_STORE64 (ctxt, DestPtrElan + i, ELAN3_OP_LOAD64 (ctxt, SourcePtrElan + i));
+           break;
+       }
+#endif
+    }
+    ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+    return (FALSE);
+}
+
+void
+ReissueEvent (ELAN3_CTXT *ctxt, E3_Addr addr, int flags)
+{
+    PRINTF1 (ctxt, DBG_CMD, "ReissueEvent : Event=%08x\n", addr);
+
+    if (IssueCommand (ctxt, offsetof (E3_CommandPort, SetEvent), addr, flags) == ISSUE_COMMAND_RETRY)
+    {
+       PRINTF1 (ctxt, DBG_CMD, "ReissueEvent: queue event %08x\n", addr);
+
+       kmutex_lock (&ctxt->SwapListsLock);
+       ctxt->ItemCount[LIST_SETEVENT]++;
+       ELAN3_OP_PUT_WORD_ITEM (ctxt, LIST_SETEVENT, addr);
+       kmutex_unlock (&ctxt->SwapListsLock);
+    }
+}
+
+int
+SetEventsNeedRestart (ELAN3_CTXT *ctxt)
+{
+    return (ctxt->ItemCount[LIST_SETEVENT] != 0);
+}
+
+void
+RestartSetEvents (ELAN3_CTXT *ctxt)
+{
+    void     *item;
+    E3_uint32 EventPointer;
+
+    kmutex_lock (&ctxt->SwapListsLock);
+    
+    while (ctxt->ItemCount[LIST_SETEVENT])
+    {
+       if (! ELAN3_OP_GET_WORD_ITEM (ctxt, LIST_SETEVENT, &item, &EventPointer))
+           ctxt->ItemCount[LIST_SETEVENT] = 0;
+       else
+       {
+           if (IssueCommand (ctxt, offsetof (E3_CommandPort, SetEvent), EventPointer, FALSE) == ISSUE_COMMAND_RETRY)
+           {
+               ELAN3_OP_PUTBACK_ITEM (ctxt, LIST_SETEVENT, item);
+               kmutex_unlock (&ctxt->SwapListsLock);
+               return;
+           }
+           
+           ctxt->ItemCount[LIST_SETEVENT]--;
+           ELAN3_OP_FREE_WORD_ITEM (ctxt, item);
+       }
+    }
+    kmutex_unlock (&ctxt->SwapListsLock);
+}
+
+int
+RunEventType(ELAN3_CTXT *ctxt, E3_FaultSave_BE *FaultSaveArea, E3_uint32 EventType)
+{
+    int failed = FALSE;
+
+    if ((EventType & EV_TYPE_BCOPY) != 0)
+       failed = SimulateBlockCopy(ctxt, FaultSaveArea->s.EventAddress);
+    
+    if ((EventType & EV_TYPE_MASK) == EV_TYPE_THREAD)
+       ReissueStackPointer (ctxt, EventType & ~(EV_TYPE_MASK_THREAD|EV_TYPE_MASK_BCOPY));
+    else if ((EventType & EV_TYPE_MASK) == EV_TYPE_DMA)
+       RestartDmaPtr (ctxt, EventType & ~(EV_TYPE_MASK_DMA|EV_TYPE_MASK_BCOPY));
+    else if ((EventType & EV_TYPE_EVIRQ) != 0)
+       QueueEventInterrupt (ctxt, EventType & ~(EV_TYPE_MASK_EVIRQ|EV_TYPE_MASK_BCOPY));
+    else /* Chained event */
+    {
+       if ((EventType & ~EV_TYPE_BCOPY) != 0) /* not null setevent */
+           ReissueEvent (ctxt, EventType & ~(EV_TYPE_MASK_CHAIN|EV_TYPE_MASK_BCOPY), FALSE);
+    }
+
+    return (failed);
+}
+
+void
+WakeupLwp (ELAN3_DEV *dev, void *arg)
+{
+    ELAN3_CTXT    *ctxt = (ELAN3_CTXT *) arg;
+    unsigned long flags;
+
+    PRINTF1 (ctxt, DBG_INTR, "WakeupLwp: %d\n", SPINLOCK_HELD (&dev->IntrLock));
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    ctxt->Halted = 1;
+    kcondvar_wakeupone (&ctxt->HaltWait, &dev->IntrLock);
+
+    PRINTF0 (ctxt, DBG_INTR, "WakeupLwp: woken up context\n");
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+void
+QueueEventInterrupt (ELAN3_CTXT *ctxt, E3_uint32 cookie)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    unsigned long flags;
+
+    PRINTF1 (ctxt, DBG_EVENT, "QueueEventInterrupt: cookie %08x\n", cookie);
+
+    if (ELAN3_OP_EVENT (ctxt, cookie, OP_INTR) == OP_DEFER)
+    {
+       spin_lock_irqsave (&ctxt->Device->IntrLock, flags);
+
+       if (ELAN3_QUEUE_REALLY_FULL (ctxt->EventCookieQ))
+       {
+           ctxt->Status |= CTXT_COMMAND_OVERFLOW_ERROR;
+           StartSwapoutContext (ctxt, 0, NULL);
+       }
+       else
+       {
+           *(ELAN3_QUEUE_BACK (ctxt->EventCookieQ, ctxt->EventCookies)) = cookie;
+           
+           ELAN3_QUEUE_ADD (ctxt->EventCookieQ);
+           kcondvar_wakeupone (&ctxt->Wait, &dev->IntrLock);
+           if (ELAN3_QUEUE_FULL (ctxt->EventCookieQ))
+           {
+               ctxt->Status |= CTXT_EVENT_QUEUE_FULL;
+               StartSwapoutContext (ctxt, 0, NULL);
+           }
+       }
+       spin_unlock_irqrestore (&ctxt->Device->IntrLock, flags);
+    }
+}
+
+int
+ElanException (ELAN3_CTXT *ctxt, int type, int proc, void *trap, ...)
+{
+    int     res;
+    va_list ap;
+
+    va_start (ap, trap);
+
+    PRINTF2 (ctxt, DBG_FN, "ElanException: proc %d type %d\n", proc, type);
+
+    res = ELAN3_OP_EXCEPTION (ctxt, type, proc, trap, ap);
+
+    va_end (ap);
+    
+    return (res);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/context_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/context_linux.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/context_linux.c        2005-06-01 23:12:54.566444120 -0400
@@ -0,0 +1,228 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: context_linux.c,v 1.28.2.2 2004/10/28 11:54:56 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/context_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kpte.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+
+int
+LoadElanTranslation (ELAN3_CTXT *ctxt, E3_Addr addr, int len, int protFault, int writeable)
+{
+    ELAN3MMU           *elan3mmu = ctxt->Elan3mmu;
+    ELAN3MMU_RGN       *rgn;
+    caddr_t            mainAddr;
+    int                        perm;
+    unsigned int        off;
+    unsigned long       flags;
+
+    ASSERT (PAGE_ALIGNED (addr) && PAGE_ALIGNED (len));
+
+    PRINTF (ctxt, DBG_FAULT, "LoadElanTranslation: addr %08x len %08x%s%s\n", 
+        addr, len, protFault ? " prot fault" : "", writeable ? " writeable" : "");
+
+    /* Ensure there's enough elan mmu tables for us to use */
+    elan3mmu_expand (elan3mmu, addr, len, PTBL_LEVEL_3, 0);
+
+    while (len > 0) 
+    {
+       /*
+        * Retrieve permission region and calculate main address
+        */
+       spin_lock (&elan3mmu->elan3mmu_lock);
+
+       rgn = elan3mmu_rgnat_elan (elan3mmu, addr);
+       if (rgn == NULL) {
+           PRINTF (ctxt, DBG_FAULT, "LoadElanTranslation: no permission region at %lx %p\n", 
+               (u_long) addr, rgn);
+           spin_unlock (&elan3mmu->elan3mmu_lock);
+           return (EFAULT);
+       }
+       mainAddr = rgn->rgn_mbase + (addr - rgn->rgn_ebase);
+
+       ASSERT (PAGE_ALIGNED ((unsigned long)mainAddr));
+
+       spin_unlock (&elan3mmu->elan3mmu_lock);
+
+       /*
+        * If we're tying to load a translation to the elan command port, 
+        * then don't do it now, but mark the context to have it reloaded
+        * just before we restart any threads. We do this because we don't
+        * want to call into the segment driver since we could then block
+        * waiting for the command port to become available.
+        */
+       if (mainAddr == ctxt->CommandPageMapping)
+       {
+           PRINTF (ctxt, DBG_FAULT, "LoadElanTranslation: addr=%08x maps command port\n", addr);
+
+           spin_lock_irqsave (&ctxt->Device->IntrLock, flags);
+           UnloadCommandPageMapping (ctxt);
+           spin_unlock_irqrestore (&ctxt->Device->IntrLock, flags);
+       }
+       else 
+       {
+           struct vm_area_struct *area;
+           struct mm_struct *mm = current->mm;
+           pte_t *ptep_ptr;
+           pte_t  ptep_value;
+
+           down_read (&current->mm->mmap_sem);
+
+           if ((area = find_vma_intersection(mm, (unsigned long)mainAddr, (unsigned long)mainAddr + PAGESIZE)) == NULL)
+           {
+               PRINTF (ctxt, DBG_FAULT, "LoadElanTranslation: %p no vma\n", mainAddr);
+               up_read (&current->mm->mmap_sem);
+               return EFAULT;
+           }
+
+           if (writeable && !(area->vm_flags & VM_WRITE)) 
+           {
+               PRINTF (ctxt, DBG_FAULT, "LoadElanTranslation: %p not writeable\n", mainAddr);
+               up_read (&current->mm->mmap_sem);
+               return EFAULT;
+           }
+           
+           spin_lock (&mm->page_table_lock);
+
+           /* dont deference the pointer after the unmap */
+           ptep_ptr = find_pte_map (mm, (unsigned long)mainAddr);  
+           if (ptep_ptr) {
+               ptep_value = *ptep_ptr;
+               pte_unmap(ptep_ptr);
+           }
+
+           PRINTF (ctxt, DBG_FAULT, "LoadElanTranslation: %p %s %s\n", 
+                   mainAddr, writeable ? "writeable" : "readonly", 
+                   !ptep_ptr ? "invalid" : pte_none(ptep_value) ? "none " : !pte_present(ptep_value) ? "swapped " : 
+                   writeable && !pte_write(ptep_value) ? "COW" : "OK");
+           
+           if (!ptep_ptr || pte_none(ptep_value) || !pte_present(ptep_value) || (writeable && !pte_write(ptep_value))) 
+           {  
+               spin_unlock (&mm->page_table_lock);
+
+               make_pages_present((unsigned long)mainAddr, (unsigned long)mainAddr + PAGE_SIZE);
+
+               spin_lock (&mm->page_table_lock);
+
+               /* dont deference the pointer after the unmap */
+               ptep_ptr = find_pte_map (mm, (unsigned long)mainAddr);  
+               if (ptep_ptr) {
+                   ptep_value = *ptep_ptr;
+                   pte_unmap(ptep_ptr);
+               }
+
+               if (!ptep_ptr || pte_none(ptep_value) || !pte_present(ptep_value) || (writeable && !pte_write(ptep_value))) 
+               {
+                   spin_unlock (&mm->page_table_lock);
+                   up_read (&current->mm->mmap_sem);
+                   return EFAULT;
+               }
+           } 
+
+           /* don't allow user write access to kernel pages if not kernel */
+           if (!pte_read(ptep_value))
+           {
+               spin_unlock (&mm->page_table_lock);
+               up_read (&current->mm->mmap_sem);
+               return EFAULT;
+           }
+
+           if (writeable)
+               pte_mkdirty(ptep_value);
+           pte_mkyoung (ptep_value);
+
+           /* now load the elan pte */
+           if (writeable)
+               perm  = rgn->rgn_perm;
+           else
+               perm = ELAN3_PERM_READONLY(rgn->rgn_perm & ELAN3_PTE_PERM_MASK) | (rgn->rgn_perm & ~ELAN3_PTE_PERM_MASK);
+
+           for (off = 0; off < PAGE_SIZE; off += ELAN3_PAGE_SIZE)
+               elan3mmu_pteload (elan3mmu, PTBL_LEVEL_3, addr + off, pte_phys(ptep_value) + off, perm, PTE_LOAD | PTE_NO_SLEEP);
+
+           spin_unlock (&mm->page_table_lock);
+           up_read (&current->mm->mmap_sem);
+       }
+
+       len -= PAGESIZE;
+       addr += PAGESIZE;
+    }
+    return (ESUCCESS);
+}
+
+
+/*
+ * LoadCommandPortTranslation:
+ *    explicitly load an elan translation to the command port.
+ *    but only do it if the command port is accessible.
+ *
+ *    we call this function just after we have restarted
+ *    and trapped commands,  since when a command traps
+ *    the elan translation to the command port is unloaded.
+ */
+void
+LoadCommandPortTranslation (ELAN3_CTXT *ctxt)
+{
+    ELAN3MMU     *elan3mmu = ctxt->Elan3mmu;
+    ELAN3MMU_RGN *rgn;
+    E3_Addr       addr;
+    int                  perm;
+    physaddr_t    phys;
+    unsigned int  off;
+    unsigned long flags;
+
+    PRINTF (ctxt, DBG_FAULT, "LoadCommandPortTranslation: SegAddr=%p Status=%x\n", ctxt->CommandPageMapping, ctxt->Status);
+
+    if (ctxt->CommandPageMapping != NULL  && !(ctxt->Status & CTXT_COMMAND_MAPPED_ELAN))
+    {
+       spin_lock (&elan3mmu->elan3mmu_lock);
+       
+       rgn = elan3mmu_rgnat_main (elan3mmu, ctxt->CommandPageMapping);
+       if (rgn == (ELAN3MMU_RGN *) NULL) 
+       {
+           PRINTF(ctxt, DBG_FAULT, "LoadCommandPortTranslation: no permission for command port\n");
+           spin_unlock (&elan3mmu->elan3mmu_lock);
+           return;
+       }
+       
+       addr = rgn->rgn_ebase + (ctxt->CommandPageMapping - rgn->rgn_mbase);
+       perm = rgn->rgn_perm;
+       phys = kmem_to_phys((caddr_t) ctxt->CommandPage);
+
+       spin_lock_irqsave (&ctxt->Device->IntrLock, flags);
+       if (ELAN3_QUEUE_EMPTY(ctxt->CommandTrapQ) && !(ctxt->Status & CTXT_OTHERS_REASONS))
+       {
+           PRINTF(ctxt, DBG_FAULT, "LoadCommandPortTranslation: load xlation addr=%08x phys=%llx perm=%d\n", 
+                  addr, (unsigned long long)phys, perm);
+
+           ctxt->Status |= CTXT_COMMAND_MAPPED_ELAN;
+
+           for (off = 0; off < PAGESIZE; off += ELAN3_PAGE_SIZE)
+               elan3mmu_pteload (elan3mmu, PTBL_LEVEL_3, addr + off, phys + off, perm, PTE_LOAD | PTE_NO_SLEEP);
+       }
+       spin_unlock_irqrestore (&ctxt->Device->IntrLock, flags);
+       
+       spin_unlock (&elan3mmu->elan3mmu_lock);
+    }
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/cproc.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/cproc.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/cproc.c        2005-06-01 23:12:54.567443968 -0400
@@ -0,0 +1,539 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: cproc.c,v 1.46 2004/02/10 15:05:10 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/cproc.c,v $ */
+
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/vmseg.h>
+
+void
+HandleCProcTrap (ELAN3_DEV *dev, E3_uint32 Pend, E3_uint32 *Maskp)
+{
+    E3_FaultSave_BE     FaultSave;
+    CProcTrapBuf_BE    TrapBuf;
+    COMMAND_TRAP       *trap;
+    ELAN3_CTXT        *ctxt;
+    sdramaddr_t         CurrTrap;
+    sdramaddr_t         LastTrapAddr;
+    int                NTrapEntries;
+    int                        NewPend;
+    unsigned long       flags;
+
+    /* 
+     * Temporarily mask out the command processor interrupt, since
+     * we may cause it be re-asserted when we re-issue the commands
+     * from the overflow queue area.
+     */
+    DISABLE_INT_MASK (dev, INT_CProc | INT_ComQueue);
+
+    NewPend = read_reg32 (dev, Exts.InterruptReg);
+
+    do {
+       if (NewPend & INT_ComQueue)
+       {
+           if ((read_reg32 (dev, ComQueueStatus) & ComQueueError) != 0)
+           {
+               printk ("elan%d: InterruptReg=%x ComQueueStatus=%x\n", dev->Instance,
+                       read_reg32 (dev, Exts.InterruptReg), read_reg32 (dev, ComQueueStatus));
+               panic ("elan: command queue has overflowed !!");
+               /* NOTREACHED */
+           }
+
+           BumpStat (dev, ComQueueHalfFull);
+
+           /*
+            * Capture the other cpus and stop the threads processor then
+            * allow the command processor to eagerly flush the command queue.
+            */
+           dev->FlushCommandCount++; dev->HaltThreadCount++;
+           SetSchedStatusRegister (dev, Pend, Maskp);
+
+           CAPTURE_CPUS();
+
+           while ((read_reg32 (dev, ComQueueStatus) & ComQueueNotEmpty) != 0)
+               mb();
+           
+           /*
+            * Let the threads processor run again, and release the cross call.
+            */
+           RELEASE_CPUS();
+
+           dev->FlushCommandCount--; dev->HaltThreadCount--;
+           SetSchedStatusRegister (dev, Pend, Maskp);
+
+           /*
+            * Re-sample the interrupt register to see if the command processor
+            * has trapped while flushing the queue.  Preserve the INT_ComQueue
+            * bit, so we can clear the ComQueueStatus register later.
+            */
+           NewPend = (read_reg32 (dev, Exts.InterruptReg) | INT_ComQueue);
+       }
+       
+       CurrTrap = dev->CommandPortTraps[dev->CurrentCommandPortTrap];
+       
+       if (NewPend & INT_CProc)
+       {
+           BumpStat (dev, CProcTraps);
+
+           /*
+            * Copy the MMU Fault Save area and zero it out for future traps.
+            */
+           elan3_sdram_copyq_from_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, CProc), &FaultSave, sizeof (E3_FaultSave));
+           elan3_sdram_zeroq_sdram      (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, CProc), sizeof (E3_FaultSave));
+
+           /*
+            * First entry in the cproc trap save area is the value of Areg and Breg for the
+            * uWord before the address fault.
+            */
+           TrapBuf.Align64 = elan3_sdram_readq (dev, CurrTrap); CurrTrap += sizeof (TrapBuf.Align64);
+
+           ctxt = ELAN3_DEV_CTX_TABLE(dev, (TrapBuf.r.Breg >> 16));
+           if (ctxt == NULL)
+           {
+               PRINTF2 (DBG_DEVICE, DBG_INTR, "HandleCProcTrap: context invalid [%08x.%08x]\n", TrapBuf.r.Areg, TrapBuf.r.Breg);
+               BumpStat (dev, InvalidContext);
+           }
+           else
+           {
+               if (ELAN3_QUEUE_REALLY_FULL (ctxt->CommandTrapQ))
+               {
+                   if ((ctxt->Status & CTXT_COMMAND_OVERFLOW_ERROR) == 0)
+                   {
+                       ctxt->Status |= CTXT_COMMAND_OVERFLOW_ERROR;
+                       StartSwapoutContext (ctxt, Pend, Maskp);
+                   }
+               }
+               else
+               {
+                   trap = ELAN3_QUEUE_BACK (ctxt->CommandTrapQ, ctxt->CommandTraps);
+                   
+                   trap->FaultSave     = FaultSave;
+                   trap->Status.Status = read_reg32 (dev, Exts.CProcStatus.Status);
+                   trap->TrapBuf       = TrapBuf;
+                   
+                   /*
+                    * The command processor does not stop after it has trapped. It will continue
+                    * to save commands for other contexts into the commands port save area.
+                    * The valid context for the trap is held in FaultSave. As some of this
+                    * trap code uses the context in the status register the local copy must be
+                    * updated with the trap context.
+                    */
+                   trap->Status.s.Context = (TrapBuf.r.Breg >> 16);
+                   
+                   PRINTF4 (ctxt, DBG_INTR, "HandleCProcTrap: WakeupFnt=%x Cntx=%x SuspAddr=%x TrapType=%s\n",
+                            trap->Status.s.WakeupFunction, trap->Status.s.Context,
+                            trap->Status.s.SuspendAddr, MiToName(trap->Status.s.TrapType));
+                   PRINTF2 (ctxt, DBG_INTR, "HandleCProcTrap: Areg=%08x Breg=%08x\n", 
+                            trap->TrapBuf.r.Areg, trap->TrapBuf.r.Breg);
+                   
+                   if (ELAN3_OP_CPROC_TRAP (ctxt, trap) == OP_DEFER)
+                   {
+                       ELAN3_QUEUE_ADD (ctxt->CommandTrapQ);
+                       
+                       PRINTF1 (ctxt, DBG_INTR, "HandleCProcTrap: setting Command Flag at %p to 1\n", &ctxt->FlagPage->CommandFlag);
+                       
+                       ctxt->FlagPage->CommandFlag = 1;
+                       
+                       kcondvar_wakeupone (&ctxt->Wait, &dev->IntrLock);
+                   }
+               }
+
+               UnloadCommandPageMapping (ctxt);
+           }
+       }
+       
+       /*
+        * Now change the CommandPortTrap queue.
+        * Must stop the command processor, wait for it to stop, find the final
+        * entry in the current cproc trap save area, reset the comm port
+        * trap save address to the other queue, clear the command port interrupt and
+        * set it running normally again, and then let it go again. This is not very
+        * time critical but it would be a good idea to prevent a higher priority
+        * interrupt from slowing down the process to prevent to fifos filling.
+        */
+       spin_lock_irqsave (&dev->CProcLock, flags);
+
+       SET_SCHED_STATUS (dev, CProcStop);
+
+       while ((read_reg32 (dev, Exts.SchCntReg) & CProcStopped) == 0)
+       {
+           PRINTF0 (DBG_DEVICE, DBG_INTR, "HandleCProcTrap: waiting for command processor to stop\n");
+           mb();
+       }
+       
+       /*
+        * Remember how many entries are in the saved command queue,  and 
+        * re-initialise it, before restarting the command processor.
+        */
+       NTrapEntries = (read_reg32 (dev, CProc_TrapSave_Addr) - dev->CommandPortTraps[dev->CurrentCommandPortTrap])/sizeof (E3_uint64);
+       LastTrapAddr = dev->CommandPortTraps[dev->CurrentCommandPortTrap] + NTrapEntries*sizeof (TrapBuf);
+
+       dev->CurrentCommandPortTrap ^= 1;
+       write_reg32 (dev, CProc_TrapSave_Addr, dev->CommandPortTraps[dev->CurrentCommandPortTrap]);
+
+       PRINTF1 (DBG_DEVICE, DBG_INTR, "HandleCProcTrap: command trap queue has %d entries\n", NTrapEntries);
+
+       if (NTrapEntries > ELAN3_COMMAND_TRAP_SIZE/sizeof (E3_uint64))
+           panic ("HandleCProcTrap: command trap queue has overflowed\n");
+       
+       if (NewPend & INT_CProc)
+       {
+           /*
+            * Clear the CProc interrupt and set it running normally again. Nothing should
+            * be running now that could issue commands apart from this trap handler.
+            */
+           PULSE_SCHED_STATUS (dev, RestartCProc);
+       }
+       
+       if (NewPend & INT_ComQueue)
+       {
+           /*
+            * Write any value here to clear out the half full and error bits of the command
+            * overflow queues. This will also remove the overflow interrupt.
+            */
+           write_reg32 (dev, ComQueueStatus, 0);
+       }
+       
+       /*
+        * And let the command processor start again
+        */
+       CLEAR_SCHED_STATUS (dev, CProcStop);
+       
+       /*
+        * Now re-issue all the commands that were issued after the command port trapped.
+        * Should halt the dma processor and force command sto be put onto the run queues
+        * to ensure that a remote re-issued command is handled correctly. NOTE it is
+        * not necessary to wait for the dma processor to stop and this will reduce the
+        * performance impact. As CProcHalt is asserted all commands will be flushed
+        * to the queues.
+        */
+       dev->HaltDmaDequeueCount++; dev->FlushCommandCount++;
+       SetSchedStatusRegister (dev, Pend, Maskp);
+       
+       /*
+        * XXXX: should we do a capture/release if the trap overflow
+        *       area has a "large" number of commands in it,  since
+        *       we will just stuff them all back in, together with 
+        *       all those issued by the other cpus/thread processors.
+        */
+       while (CurrTrap != LastTrapAddr)
+       {
+           /* Read the next saved (but not trapped) command */
+           TrapBuf.Align64 = elan3_sdram_readq (dev, CurrTrap); CurrTrap += sizeof (TrapBuf);
+           
+
+           ctxt = ELAN3_DEV_CTX_TABLE(dev, (TrapBuf.s.ContextType >> 16));
+           
+           if (ctxt == NULL)
+           {
+               PRINTF1 (DBG_DEVICE, DBG_INTR, "HandleCProcTrap: context %x invalid\n", TrapBuf.s.ContextType >> 16);
+               BumpStat (dev, InvalidContext);
+           }
+           else
+           {
+               if (!ELAN3_QUEUE_EMPTY (ctxt->CommandTrapQ) || (ctxt->Status & CTXT_OTHERS_REASONS))
+               {
+                   PRINTF3 (ctxt, DBG_INTR, "HandleCProcTrap: save command %x context %x - %08x\n",
+                            (TrapBuf.s.ContextType>>3) & 0x3ff, TrapBuf.s.ContextType >> 17, TrapBuf.s.Addr);
+                   
+                   if (ELAN3_QUEUE_REALLY_FULL (ctxt->CommandQ))
+                   {
+                       ctxt->Status |= CTXT_COMMAND_OVERFLOW_ERROR;
+                       StartSwapoutContext (ctxt, Pend, Maskp);
+                   }
+                   else
+                   {
+                       *ELAN3_QUEUE_BACK(ctxt->CommandQ, ctxt->Commands) = TrapBuf;
+
+                       ELAN3_QUEUE_ADD (ctxt->CommandQ);
+                   }
+                   continue;
+               }
+               
+               /* Reissue the command to the command port for this context */
+               PRINTF2 (ctxt, DBG_INTR, "HandleCProcTrap: re-issue command %x - %08x\n",
+                        (TrapBuf.s.ContextType>>5) & 0xff, TrapBuf.s.Addr);
+               
+               mb();
+               if (ELAN3_OP_CPROC_REISSUE(ctxt, &TrapBuf) != OP_HANDLED)
+                   ((E3_uint32 *) ctxt->CommandPort)[(TrapBuf.s.ContextType>>5) & 0xff] = TrapBuf.s.Addr;
+               mmiob();
+           }
+       }
+       
+       while ((read_reg32 (dev, ComQueueStatus) & ComQueueNotEmpty) != 0)
+       {
+           PRINTF0 (DBG_DEVICE, DBG_INTR, "HandleCProcTrap: waiting for queues to empty after reissueing commands\n");
+           mb();
+       }
+       
+       dev->HaltDmaDequeueCount--; dev->FlushCommandCount--;
+       SetSchedStatusRegister (dev, Pend, Maskp);
+       
+       spin_unlock_irqrestore (&dev->CProcLock, flags);
+
+       /*
+        * Re-read the interrupt register and see if we've got another command
+        * port interrupt
+        */
+       NewPend = read_reg32 (dev, Exts.InterruptReg);
+    } while ((NewPend & (INT_CProc | INT_ComQueue)) != 0);
+
+
+    /*
+     * Re-enable the command processor interrupt as we've finished 
+     * polling it.
+     */
+    ENABLE_INT_MASK (dev, INT_CProc | INT_ComQueue);
+}
+
+void
+ResolveCProcTrap (ELAN3_CTXT *ctxt)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    COMMAND_TRAP *trap;
+    int                  res;
+    unsigned long flags;
+
+    kmutex_lock (&ctxt->CmdLock);
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    while (! ELAN3_QUEUE_BACK_EMPTY (ctxt->CommandTrapQ))
+    {
+       trap = ELAN3_QUEUE_MIDDLE(ctxt->CommandTrapQ, ctxt->CommandTraps);
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+       switch (trap->Status.s.TrapType)
+       {
+       case MI_EventIntUpdateBPtr:
+       case MI_ChainedEventError:
+       case MI_EventQueueOverflow:
+       case MI_ThreadQueueOverflow:
+       case MI_DmaQueueOverflow:
+           PRINTF1 (ctxt, DBG_CPROC, "ResolveCProcTrap: %s\n", MiToName (trap->Status.s.TrapType));
+           break;
+           
+       default:
+           /* All other traps are MMU related, we should have a fault address and FSR */
+           if ((res = elan3_pagefault (ctxt, &trap->FaultSave, 1)) != ESUCCESS)
+           {
+               PRINTF1 (ctxt, DBG_CPROC, "ResolveCProcTrap: elan3_pagefault failed for address %08x\n", 
+                        trap->FaultSave.s.FaultAddress);
+               ElanException (ctxt, EXCEPTION_INVALID_ADDR, COMMAND_PROC, trap, &trap->FaultSave, res);
+               
+               /* Set the trap type to 0 so the command does not get re-issued */
+               trap->Status.s.TrapType = 0;
+           }
+           break;
+       }
+       
+       spin_lock_irqsave (&dev->IntrLock, flags);
+
+       ELAN3_QUEUE_CONSUME (ctxt->CommandTrapQ);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    kmutex_unlock (&ctxt->CmdLock);
+}
+
+int
+RestartCProcTrap (ELAN3_CTXT *ctxt)
+{
+    ELAN3_DEV     *dev      = ctxt->Device;
+    COMMAND_TRAP  trap;
+    void        *item;
+    int                  res;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    while (! ELAN3_QUEUE_FRONT_EMPTY (ctxt->CommandTrapQ))
+    {
+       trap = (*ELAN3_QUEUE_FRONT (ctxt->CommandTrapQ, ctxt->CommandTraps));
+       ELAN3_QUEUE_REMOVE (ctxt->CommandTrapQ);
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       
+       BumpUserStat (ctxt, CProcTraps);
+
+       switch (trap.Status.s.TrapType)
+       {
+       case 0:
+           res = ISSUE_COMMAND_OK;
+           break;
+           
+       case MI_WaitForWaitEventDesc:
+           /*
+            * Fault occured on the read of wait event descriptor for wait event type 0.
+            * Fault already fixed. Just re-issue the wait command. Wait event descriptor addr
+            * is in the Areg save value.
+            */
+           PRINTF1 (ctxt, DBG_CPROC, "RestartCProcTrap: WaitEvent type0 desc read fault %08x\n", 
+                    trap.TrapBuf.r.Areg);
+           
+           res = IssueCommand (ctxt, offsetof (E3_CommandPort, WaitEvent0), trap.TrapBuf.r.Areg, ISSUE_COMMAND_FOR_CPROC);
+           break;
+
+       case MI_WaitForEventReadTy0:
+           /*
+            * Fault occured on the read of event location for wait event type 0.
+            * Fault already fixed. Just re-issue the wait command. Wait event descriptor addr
+            * is in the Areg save value.
+            */
+           PRINTF1 (ctxt, DBG_CPROC, "RestartCProcTrap: WaitEvent type0 event loc fault %08x\n",
+                    trap.TrapBuf.r.Areg);
+           
+           res = IssueCommand (ctxt, offsetof (E3_CommandPort, WaitEvent0), trap.TrapBuf.r.Areg, ISSUE_COMMAND_FOR_CPROC);
+           break;
+           
+       case MI_WaitForEventReadTy1:
+           /*
+            * Fault occured on the read of the event location for wait event type 1.
+            * Areg has the original ptr and count.
+            * Fault already fixed. Just re-issue the wait command using Areg and context.
+            */
+           PRINTF1 (ctxt, DBG_CPROC, "RestartCProcTrap: WaitEvent type1 event location read fault %08x\n",
+                    trap.TrapBuf.r.Areg);
+           res = IssueCommand (ctxt, offsetof (E3_CommandPort, WaitEvent1), trap.TrapBuf.r.Areg, ISSUE_COMMAND_FOR_CPROC);
+           break;
+           
+       case MI_WaitForCntxDmaDescRead:
+       case MI_WaitForNonCntxDmaDescRead:
+           /*
+            * Fault occured on the read of the dma descriptor. Run dma using the
+            * Fault Address in FaultSave.
+            */
+           PRINTF1 (ctxt, DBG_CPROC, "RestartCProcTrap: MI_WaitForCntxDmaDescRead: re-issue dma at %08x\n", 
+                    trap.FaultSave.s.FaultAddress);
+           
+           res = IssueDmaCommand (ctxt, trap.FaultSave.s.FaultAddress, NULL, ISSUE_COMMAND_FOR_CPROC);
+           break;
+           
+       default:
+           /*
+            * Assume the fault will be fixed by FixupEventTrap.
+            */
+           FixupEventTrap (ctxt, COMMAND_PROC, &trap, trap.Status.s.TrapType, &trap.FaultSave, ISSUE_COMMAND_FOR_CPROC);
+
+           res = ISSUE_COMMAND_OK;
+           break;
+       }
+
+       switch (res)
+       {
+       case ISSUE_COMMAND_OK:                                  /* command re-issued ok*/
+           break;
+
+       case ISSUE_COMMAND_TRAPPED:                             /* command trapped,  it will have been copied */
+           return (EAGAIN);                                    /* to the back of the trap queue */
+
+       case ISSUE_COMMAND_RETRY:                               /* didn't issue command, so place back at front for */
+           spin_lock_irqsave (&dev->IntrLock, flags);          /* later (after resolving other traps */
+
+           if (ELAN3_QUEUE_REALLY_FULL (ctxt->CommandTrapQ))
+               ctxt->Status |= CTXT_COMMAND_OVERFLOW_ERROR;
+           else
+           {
+               ELAN3_QUEUE_ADD_FRONT(ctxt->CommandTrapQ);
+               (*ELAN3_QUEUE_FRONT (ctxt->CommandTrapQ, ctxt->CommandTraps)) = trap;
+           }
+           spin_unlock_irqrestore (&dev->IntrLock, flags);
+           return (EAGAIN);
+
+       default:
+           return (EINVAL);
+       }
+       spin_lock_irqsave (&dev->IntrLock, flags);
+    }  
+
+    /*
+     * GNAT 5409 - if CommandPortItem was not NULL, but other reasons were set,
+     *             then we'd not free the CommandPortItem even though we'd re-
+     *             issued all trapped and overflowed commands.  Hence only return
+     *             without clearing CommandPortItem if we will be called again as
+     *             either CommandTrapQ or CommandQ is not empty.
+     */
+
+    /* Now run the overflowed commands for this context */
+    if (! ELAN3_QUEUE_EMPTY (ctxt->CommandQ))
+    {
+       if (! ELAN3_QUEUE_EMPTY (ctxt->CommandTrapQ) || (ctxt->Status & CTXT_OTHERS_REASONS))
+       {
+           PRINTF0 (ctxt, DBG_CPROC, "RestartCProcTrap: cannot issue overflowed commands\n");
+           spin_unlock_irqrestore (&dev->IntrLock, flags);
+           return (EAGAIN);
+       }
+
+       /*
+        * Just re-issue the commands,  if one traps then the remainder will 
+        * just get placed in the overflow queue again and the interrupt handler
+        * will copy them back in here.
+        *
+        * Stop the dma processor from taking commands,  since one of the commands
+        * could be a re-issued remote dma, which must be processed by the command
+        * processor.
+        */
+       
+       if (dev->HaltDmaDequeueCount++ == 0)
+           SetSchedStatusRegister (dev, 0, NULL);
+       
+       while (! ELAN3_QUEUE_EMPTY (ctxt->CommandQ))
+       {
+           CProcTrapBuf_BE *TrapBuf = ELAN3_QUEUE_FRONT (ctxt->CommandQ, ctxt->Commands);
+           
+           PRINTF2 (ctxt, DBG_CPROC, "RestartCProcTrap: re-issue command %x - %08x\n",
+                    (TrapBuf->s.ContextType>>5) & 0xff, TrapBuf->s.Addr);
+           mb();                                                       /* ensure writes to main memory completed */
+           ((E3_uint32 *) ctxt->CommandPort)[(TrapBuf->s.ContextType>>5) & 0xff] = TrapBuf->s.Addr;
+           mmiob();                                            /* and flush through IO writes */
+           
+           ELAN3_QUEUE_REMOVE (ctxt->CommandQ);
+       }
+       
+       /* observe the command processor having halted */
+       res = CheckCommandQueueFlushed (ctxt, DmaComQueueNotEmpty, 0, &flags);
+       
+       if (res != ISSUE_COMMAND_OK)
+       {
+           PRINTF0 (ctxt, DBG_CPROC, "RestartCProcTrap: trapped after issueing overflowed commands\n");
+           spin_unlock_irqrestore (&dev->IntrLock, flags);
+           return (EAGAIN);
+       }
+    }
+
+    /* remove the command port item, while holding the lock */
+    item = ctxt->CommandPortItem;
+    ctxt->CommandPortItem = NULL;
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+       
+    if (item != NULL)                                          /* Free of any item that may have been stored */
+    {                                                          /* because of the commandport trap */
+       PRINTF1 (ctxt, DBG_CPROC, "RestartCProcTrap: commandPortItem %p\n", item);
+
+       kmutex_lock (&ctxt->SwapListsLock);
+       ELAN3_OP_FREE_BLOCK_ITEM (ctxt, item);
+       kmutex_unlock (&ctxt->SwapListsLock);
+    }
+
+    return (ESUCCESS);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/dproc.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/dproc.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/dproc.c        2005-06-01 23:12:54.568443816 -0400
@@ -0,0 +1,553 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: dproc.c,v 1.52 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/dproc.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/intrinsics.h>
+#include <elan3/dma.h>
+#include <elan3/vmseg.h>
+
+#define DMA_RETRY_FAIL_COUNT   8
+
+static void PrintUserDma (ELAN3_CTXT *ctxt, E3_Addr addr);
+
+int
+HandleDProcTrap (ELAN3_DEV *dev, E3_uint32 *RestartBits)
+{
+    DMA_TRAP    *trap   = dev->DmaTrap;
+
+    ASSERT(SPINLOCK_HELD (&dev->IntrLock));
+
+    /* Scoop out the trap information, before restarting the Elan */
+    trap->Status.Status = read_reg32 (dev, Exts.DProcStatus.Status);
+    
+    ASSERT(trap->Status.s.WakeupFunction == WakeupNever);
+
+    /* copy the normal dma access fault type */
+    elan3_sdram_copyq_from_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProc), &trap->FaultSave, sizeof (E3_FaultSave_BE));
+    
+    /* copy all 4 of the dma data fault type */
+    elan3_sdram_copyq_from_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData0), &trap->Data0, 4*sizeof (E3_FaultSave_BE));
+    
+    /* Copy the DMA descriptor */
+    copy_dma_regs (dev, &trap->Desc);
+    
+    /* Copy the packet info */
+    trap->PacketInfo.Value = read_reg32 (dev, Exts.Dmas.DmaRds.DMA_PacketInfo.Value);
+
+    /* update device statistics */
+    BumpStat (dev, DProcTraps);
+    switch (trap->Status.s.TrapType)
+    {
+    case MI_DmaPacketTimedOutOrPacketError:
+       if (trap->PacketInfo.s.PacketTimeout)
+           BumpStat (dev, DmaOutputTimeouts);
+       else if (trap->PacketInfo.s.PacketAckValue == C_ACK_ERROR)
+           BumpStat (dev, DmaPacketAckErrors);
+       break;
+       
+    case MI_DmaFailCountError:
+       BumpStat (dev, DmaRetries);
+       break;
+    }
+
+    /* Must now zero all the FSRs so that a subsequent fault can be seen */
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProc), sizeof (E3_FaultSave));
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData0), 4*sizeof (E3_FaultSave));
+           
+    *RestartBits |= RestartDProc;
+    return (TRUE);
+}
+
+void
+DeliverDProcTrap (ELAN3_DEV *dev, DMA_TRAP *dmaTrap, E3_uint32 Pend)
+{
+    ELAN3_CTXT     *ctxt;
+    E3_FaultSave_BE *FaultArea;
+    DMA_TRAP       *trap;
+    register int     i;
+
+    ASSERT(SPINLOCK_HELD (&dev->IntrLock));
+
+    ctxt = ELAN3_DEV_CTX_TABLE(dev, dmaTrap->Status.s.Context);
+
+    if (ctxt == NULL)
+    {
+       PRINTF1 (DBG_DEVICE, DBG_INTR, "DeliverDProcTrap: context %x invalid\n", dmaTrap->Status.s.Context);
+       BumpStat (dev, InvalidContext);
+    }
+    else
+    {
+       if (ELAN3_OP_DPROC_TRAP (ctxt, dmaTrap) == OP_DEFER)
+       {
+           if (ELAN3_QUEUE_REALLY_FULL (ctxt->DmaTrapQ))
+           {
+               ctxt->Status |= CTXT_COMMAND_OVERFLOW_ERROR;
+               StartSwapoutContext (ctxt, Pend, NULL);
+           }
+           else
+           {
+               trap = ELAN3_QUEUE_BACK (ctxt->DmaTrapQ, ctxt->DmaTraps);
+               
+               bcopy (dmaTrap, trap, sizeof (DMA_TRAP));
+               
+               PRINTF5 (ctxt, DBG_INTR, "DeliverDProcTrap: WakeupFnt=%x Cntx=%x SuspAddr=%x PacketInfo=%x TrapType=%s\n",
+                        trap->Status.s.WakeupFunction, trap->Status.s.Context, 
+                        trap->Status.s.SuspendAddr, trap->PacketInfo.Value, MiToName (trap->Status.s.TrapType));
+               PRINTF3 (ctxt, DBG_INTR, "                    FaultAddr=%x EventAddr=%x FSR=%x\n",
+                        trap->FaultSave.s.FaultAddress, trap->FaultSave.s.EventAddress,
+                        trap->FaultSave.s.FSR.Status);
+               for (i = 0, FaultArea = &trap->Data0; i < 4; i++, FaultArea++)
+                   PRINTF4 (ctxt, DBG_INTR, "                  %d FaultAddr=%x EventAddr=%x FSR=%x\n", i,
+                            FaultArea->s.FaultAddress, FaultArea->s.EventAddress, FaultArea->s.FSR.Status);
+               
+               PRINTF4 (ctxt, DBG_INTR, "                 type %08x size %08x source %08x dest %08x\n",
+                        trap->Desc.s.dma_type, trap->Desc.s.dma_size, trap->Desc.s.dma_source, trap->Desc.s.dma_dest);
+               PRINTF2 (ctxt, DBG_INTR, "                 Dest event %08x cookie/proc %08x\n",
+                        trap->Desc.s.dma_destEvent, trap->Desc.s.dma_destCookieVProc);
+               PRINTF2 (ctxt, DBG_INTR, "                 Source event %08x cookie/proc %08x\n",
+                        trap->Desc.s.dma_srcEvent, trap->Desc.s.dma_srcCookieVProc);
+               ELAN3_QUEUE_ADD (ctxt->DmaTrapQ);
+               kcondvar_wakeupone (&ctxt->Wait, &dev->IntrLock);
+               
+               if (ELAN3_QUEUE_FULL (ctxt->DmaTrapQ))
+               {
+                   PRINTF0 (ctxt, DBG_INTR, "DeliverDProcTrap: dma queue full, must swap out\n");
+                   ctxt->Status |= CTXT_DMA_QUEUE_FULL;
+                   
+                   StartSwapoutContext (ctxt, Pend, NULL);
+               }
+           }
+       }
+    }
+}
+
+int
+NextDProcTrap (ELAN3_CTXT *ctxt, DMA_TRAP *trap)
+{
+    ELAN3_DEV *dev = ctxt->Device;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+    
+    if (ELAN3_QUEUE_EMPTY (ctxt->DmaTrapQ))
+       return (0);
+
+    *trap = *ELAN3_QUEUE_FRONT (ctxt->DmaTrapQ, ctxt->DmaTraps);
+    ELAN3_QUEUE_REMOVE (ctxt->DmaTrapQ);
+    
+    return (1);
+}
+
+void
+ResolveDProcTrap (ELAN3_CTXT *ctxt, DMA_TRAP *trap)
+{
+    E3_FaultSave_BE *FaultArea;
+    int                     FaultHandled = 0;
+    int                     res;
+    register int     i;
+    
+    PRINTF4 (ctxt, DBG_DPROC, "ResolveDProcTrap: WakeupFnt=%x Cntx=%x SuspAddr=%x TrapType=%s\n",
+            trap->Status.s.WakeupFunction, trap->Status.s.Context, 
+            trap->Status.s.SuspendAddr, MiToName (trap->Status.s.TrapType));
+    PRINTF3 (ctxt, DBG_DPROC, "                    FaultAddr=%x EventAddr=%x FSR=%x\n",
+            trap->FaultSave.s.FaultAddress, trap->FaultSave.s.EventAddress,
+            trap->FaultSave.s.FSR.Status);
+    for (i = 0, FaultArea = &trap->Data0; i < 4; i++, FaultArea++)
+       PRINTF4 (ctxt, DBG_DPROC, "                  %d FaultAddr=%x EventAddr=%x FSR=%x\n", i,
+                FaultArea->s.FaultAddress, FaultArea->s.EventAddress, FaultArea->s.FSR.Status);
+
+    PRINTF4 (ctxt, DBG_DPROC, "                  type %08x size %08x source %08x dest %08x\n",
+            trap->Desc.s.dma_type, trap->Desc.s.dma_size, trap->Desc.s.dma_source, trap->Desc.s.dma_dest);
+    PRINTF2 (ctxt, DBG_DPROC, "                  Dest event %08x cookie/proc %08x\n",
+            trap->Desc.s.dma_destEvent, trap->Desc.s.dma_destCookieVProc);
+    PRINTF2 (ctxt, DBG_DPROC, "                  Source event %08x cookie/proc %08x\n",
+            trap->Desc.s.dma_srcEvent, trap->Desc.s.dma_srcCookieVProc);
+    
+    BumpUserStat (ctxt, DProcTraps);
+
+    switch (trap->Status.s.TrapType)
+    {
+    case MI_DmaPacketTimedOutOrPacketError:
+       /*
+        * Faulted due to packet timeout or a PAckError.
+        * Reset fail count and reissue the same desc.
+        */
+       PRINTF0 (ctxt, DBG_DPROC, "ResolveDProcTrap: got a PAckError or the output timed out. Rescheduling dma.\n");
+       if (ElanException (ctxt, EXCEPTION_PACKET_TIMEOUT, DMA_PROC, trap) == OP_IGNORE)
+       {
+           BumpUserStat (ctxt, DmaRetries);
+
+           trap->Desc.s.dma_failCount = DMA_RETRY_FAIL_COUNT;
+
+           RestartDmaTrap (ctxt, trap);
+       }
+       return;
+
+    case MI_DmaFailCountError:
+       /*
+        * Faulted due to dma fail count.
+        * Reset fail count and reissue the same desc.
+        */
+       PRINTF1 (ctxt, DBG_DPROC, "ResolveDProcTrap: Reset dma fail count to %d\n", DMA_RETRY_FAIL_COUNT);
+       
+       if (ElanException (ctxt, EXCEPTION_DMA_RETRY_FAIL, DMA_PROC, trap) == OP_IGNORE)
+       {
+           BumpUserStat (ctxt, DmaRetries);
+
+           trap->Desc.s.dma_failCount = DMA_RETRY_FAIL_COUNT;
+
+           RestartDmaTrap (ctxt, trap);
+       }
+       return;
+
+    case MI_TimesliceDmaQueueOverflow:
+       PRINTF0 (ctxt, DBG_DPROC, "ResolveDProcTrap: dma timeslice queue overflow\n");
+       RestartDmaTrap (ctxt, trap);
+       return;
+       
+    case MI_UnimplementedError:
+       PRINTF0 (ctxt, DBG_DPROC, "ResolveDProcTrap: unimplemented dma trap\n");
+       if (ElanException (ctxt, EXCEPTION_UNIMPLEMENTED, DMA_PROC, trap) == OP_IGNORE)
+           RestartDmaTrap (ctxt, trap);
+       return;
+
+    case MI_EventQueueOverflow:
+    case MI_ThreadQueueOverflow:
+    case MI_DmaQueueOverflow:
+       PRINTF0 (ctxt, DBG_DPROC, "ResolveDProcTrap: trapped on a write set event.\n");
+       FixupEventTrap (ctxt, DMA_PROC, trap, trap->Status.s.TrapType, &trap->FaultSave, 0);
+       return;
+
+    case MI_RemoteDmaCommand:
+    case MI_RunDmaCommand:
+    case MI_DequeueNonSysCntxDma:
+    case MI_DequeueSysCntxDma:
+       /*
+        * The DMA processor has trapped due to outstanding prefetches from the previous 
+        * dma.  The "current" dma has not been consumed, so we just ignore the trap
+        */
+       return;
+
+    case MI_WaitForRemoteDescRead2:
+    case MI_ExecuteDmaDescriptorForRun:
+       /*
+        * The DMA processor has trapped while fetching the dma descriptor, so
+        * zero it out to not confuse the user on an error
+        */
+       bzero (&trap->Desc, sizeof (trap->Desc));
+       break;
+    }
+
+    /*
+     * All other uWords will have updated one of the fault areas,  so fix
+     * any faults found in them.  If there were no faults found then it 
+     * must have been a bus error
+     */
+    for (i = 0, FaultArea = &trap->Data0; i < 4; i++, FaultArea++)
+    {
+       if (FaultArea->s.FSR.Status != 0)
+       {
+           FaultHandled++;
+
+           ASSERT ((FaultArea->s.FSR.Status & FSR_SizeMask) == FSR_Block64 ||
+                   (FaultArea->s.FSR.Status & FSR_SizeMask) == FSR_Block32);
+           
+           ASSERT (FaultArea->s.FaultContext == trap->Status.s.Context);
+           
+           if (((trap->Desc.s.dma_source & PAGEOFFSET) >= (PAGESIZE-E3_BLK_SIZE)) &&
+               ((trap->Desc.s.dma_source & PAGEMASK) != ((trap->Desc.s.dma_source + trap->Desc.s.dma_size-1) & PAGEMASK)))
+           {
+               /* XXXX: dma started within last 64 bytes of the page
+                *       terminate the process if it has pagefaulted */
+               if (FaultArea->s.FaultAddress == (trap->Desc.s.dma_source & ~(E3_BLK_SIZE-1)))
+               {
+                   printk ("elan%d: invalid dma - context=%x source=%x\n", ctxt->Device->Instance, 
+                           ctxt->Capability.cap_mycontext, trap->Desc.s.dma_source);
+                   
+                   if (ElanException (ctxt, EXCEPTION_BAD_DMA, DMA_PROC, trap, NULL, 0) != OP_IGNORE)
+                       return;
+               }
+           }
+
+           if (trap->Desc.s.dma_size != 0 && (res = elan3_pagefault (ctxt, FaultArea, 1)) != ESUCCESS)
+           {
+               /* XXXX: Rev B Elans can prefetch data passed the end of the dma descriptor */
+               /*       if the fault relates to this, then just ignore it */
+               if (FaultArea->s.FaultAddress < (trap->Desc.s.dma_source+trap->Desc.s.dma_size) ||
+                   FaultArea->s.FaultAddress > (trap->Desc.s.dma_source+trap->Desc.s.dma_size+E3_BLK_SIZE*2))
+               {
+                   PRINTF1 (ctxt, DBG_DPROC, "ResolveDProcTrap: elan3_pagefault failed for address %x\n",
+                            FaultArea->s.FaultAddress);
+                   
+                   if (ElanException (ctxt, EXCEPTION_INVALID_ADDR, DMA_PROC, trap, FaultArea, res) != OP_IGNORE)
+                       return;
+               }
+           }
+       }
+    }
+    
+    if (trap->FaultSave.s.FSR.Status != 0)
+    {
+       FaultHandled++;
+
+       ASSERT (trap->FaultSave.s.FaultContext == trap->Status.s.Context);
+
+       if ((trap->FaultSave.s.FSR.Status & FSR_SizeMask) == FSR_RouteFetch)
+       {
+           res = ResolveVirtualProcess (ctxt, trap->FaultSave.s.FaultAddress & 0xffff); /* mask out cookie */
+
+           switch (res)
+           {
+           default:
+               if (ElanException (ctxt, EXCEPTION_INVALID_PROCESS, DMA_PROC, trap, trap->FaultSave.s.FaultAddress, res) != OP_IGNORE)
+                   return;
+               
+           case EAGAIN:
+               /* XXXX; wait on trail blazing code */
+
+           case 0:
+               break;
+           }
+       }
+       else
+       {
+           if ((res = elan3_pagefault (ctxt, &trap->FaultSave, 1)) != ESUCCESS)
+           {
+               PRINTF1 (ctxt, DBG_DPROC, "ResolveDProcTrap: elan3_pagefault failed for address %x\n",
+                        trap->FaultSave.s.FaultAddress);
+
+               if (ElanException (ctxt, EXCEPTION_INVALID_ADDR, DMA_PROC, trap, &trap->FaultSave, res) != OP_IGNORE)
+                   return;
+           }
+       }
+    }
+
+    if (! FaultHandled)
+    {
+       ElanBusError (ctxt->Device);
+
+       if (ElanException (ctxt, EXCEPTION_INVALID_ADDR, DMA_PROC, trap, &trap->FaultSave, EFAULT) != OP_IGNORE)
+           return;
+    }
+
+    switch (trap->Status.s.TrapType)
+    {
+    case MI_WaitForRemoteDescRead2:
+       /*
+        * Faulted while trying to read the dma descriptor for a read dma.
+        * Fix fault and re-issue using FaultAddress.
+        */
+       PRINTF1 (ctxt, DBG_DPROC, "ResolveDProcTrap: trapped reading a remote dma descriptor at %x.\n",
+                trap->FaultSave.s.FaultAddress);
+       
+       RestartDmaPtr (ctxt, trap->FaultSave.s.FaultAddress);
+       break;
+       
+    case MI_ExecuteDmaDescriptorForRun:
+       /*
+        * Faulted while trying to read the dma descriptor for a write dma.
+        * Fix fault and re-issue using FaultAddress.
+        */
+       PRINTF1 (ctxt, DBG_DPROC, "ResolveDProcTrap: trapped reading a write dma descriptor at %x.\n", 
+                trap->FaultSave.s.FaultAddress);
+       
+       RestartDmaPtr (ctxt, trap->FaultSave.s.FaultAddress);
+       break;
+       
+    case MI_WaitForRemoteRoutes1:
+    case MI_WaitForRemoteRoutes2:
+    case MI_SendRemoteDmaDesc:
+    case MI_SendDmaIdentify:
+    case MI_SendRemoteDmaRoutes2:
+    case MI_WaitForDmaRoutes1:
+    case MI_DmaLoop:
+    case MI_ExitDmaLoop:
+    case MI_GetDestEventValue:
+    case MI_SendFinalUnlockTrans:
+    case MI_SendNullSetEvent:
+    case MI_SendFinalSetEvent:
+    case MI_SendDmaEOP:
+       /*
+        * Faulted either fetching routes or fetching dma data.
+        * Fix fault and re-issue using FaultAddress.
+        */
+
+    case MI_SendEOPforRemoteDma:
+    case MI_LookAtRemoteAck:
+    case MI_FailedAckIfCCis0:
+       /*
+        * Possible fault when reading the remote desc into the dma data buffers
+        */
+       PRINTF0 (ctxt, DBG_DPROC, "ResolveDProcTrap:  trapped reading a dma data or fetching a route\n");
+       RestartDmaTrap (ctxt, trap);
+       break;
+       
+    case MI_DequeueSysCntxDma:
+    case MI_DequeueNonSysCntxDma:
+    case MI_RemoteDmaCommand:
+    case MI_RunDmaCommand:
+       /*
+        * It is possible that a dma can get back onto the queue while outstanding dma
+        * have not finished trapping. In this case the trap can be ignored as the dma
+        * state has been saved. It might trap again the next time it comes to the front
+        * of the queue and be fixed then.
+        */
+       PRINTF0 (ctxt, DBG_DPROC, "ResolveDProcTrap: trap after dma has finished. ignored\n");
+       break;
+       
+    default:
+       PRINTF0 (ctxt, DBG_DPROC, "ResolveDProcTrap: trapped on a write set event.\n");
+       FixupEventTrap (ctxt, DMA_PROC, trap, trap->Status.s.TrapType, &trap->FaultSave, 0);
+       break;
+    }
+}
+
+int
+DProcNeedsRestart (ELAN3_CTXT *ctxt)
+{
+    return (ctxt->ItemCount[LIST_DMA_PTR] != 0 ||
+           ctxt->ItemCount[LIST_DMA_DESC] != 0);
+}
+
+void
+RestartDProcItems (ELAN3_CTXT *ctxt)
+{
+    void      *item;
+    E3_Addr    value;
+    int               res;
+    
+    kmutex_lock (&ctxt->SwapListsLock);
+    while (ctxt->ItemCount[LIST_DMA_PTR])
+    {
+       if (! ELAN3_OP_GET_WORD_ITEM (ctxt, LIST_DMA_PTR, &item, &value))
+           ctxt->ItemCount[LIST_DMA_PTR] = 0;
+       else
+       {
+           PRINTF1 (ctxt, DBG_DPROC, "RestartDProc: issue write dma at %x\n", value);
+           PrintUserDma (ctxt, value);
+
+           res = IssueDmaCommand (ctxt, value, NULL, 0);
+           
+           if (res == ISSUE_COMMAND_RETRY)
+           {
+               ELAN3_OP_PUTBACK_ITEM (ctxt, LIST_DMA_PTR, item);
+               kmutex_unlock (&ctxt->SwapListsLock);
+               return;
+           }
+           
+           ctxt->ItemCount[LIST_DMA_PTR]--;
+           ELAN3_OP_FREE_WORD_ITEM (ctxt, item);
+       }
+    }
+    
+    while (ctxt->ItemCount[LIST_DMA_DESC])
+    {
+       if (! ELAN3_OP_GET_BLOCK_ITEM (ctxt, LIST_DMA_DESC, &item, &value))
+           ctxt->ItemCount[LIST_DMA_DESC] = 0;
+       else
+       {
+           PRINTF1 (ctxt, DBG_DPROC, "RestartDProc: issue dma desc at %x\n", value);
+           PrintUserDma (ctxt, value);
+
+           res = IssueDmaCommand (ctxt, value, item, 0);
+
+           switch (res)
+           {
+           case ISSUE_COMMAND_OK:
+               ctxt->ItemCount[LIST_DMA_DESC]--;
+               ELAN3_OP_FREE_BLOCK_ITEM (ctxt, item);
+               break;
+               
+           case ISSUE_COMMAND_RETRY:
+               ELAN3_OP_PUTBACK_ITEM (ctxt, LIST_DMA_DESC, item);
+               kmutex_unlock (&ctxt->SwapListsLock);
+               return;
+               
+           case ISSUE_COMMAND_TRAPPED:
+               ctxt->ItemCount[LIST_DMA_DESC]--;
+               /* The item will be freed off when the command port trap */
+               /* fixed up and the command successfully re-issued */
+               break;
+           }
+       }
+    }
+
+    kmutex_unlock (&ctxt->SwapListsLock);
+}
+
+void
+RestartDmaDesc(ELAN3_CTXT *ctxt, E3_DMA_BE *desc)
+{
+    kmutex_lock (&ctxt->SwapListsLock);
+    if (desc->s.dma_direction != DMA_WRITE)
+       desc->s.dma_direction = (desc->s.dma_direction & ~DMA_READ) | DMA_READ_REQUEUE;
+
+    ELAN3_OP_PUT_BLOCK_ITEM (ctxt, LIST_DMA_DESC, (E3_uint32 *) desc);
+    ctxt->ItemCount[LIST_DMA_DESC]++;
+
+    kmutex_unlock (&ctxt->SwapListsLock);
+}
+
+void
+RestartDmaTrap(ELAN3_CTXT *ctxt, DMA_TRAP *trap)
+{
+    /* Negative length DMAs are illegal, since they hangup the dma processor,
+     * if they got generated then they will have been spotted by PollForDmahungup,
+     * and delivered to us with a Dequeue  suspend address,
+     *
+     * GNAT sw-elan3/3908: Moved this check into this new function to avoid
+     * it sampling old or invalid register state
+     */
+    if (trap->Desc.s.dma_size > E3_MAX_DMA_SIZE)
+       ElanException (ctxt, EXCEPTION_BAD_DMA, DMA_PROC, trap, NULL, 0);
+    else
+       RestartDmaDesc (ctxt, &trap->Desc);
+}
+
+void
+RestartDmaPtr (ELAN3_CTXT *ctxt, E3_Addr ptr)
+{
+    kmutex_lock (&ctxt->SwapListsLock);
+    ELAN3_OP_PUT_WORD_ITEM (ctxt, LIST_DMA_PTR, ptr);
+    ctxt->ItemCount[LIST_DMA_PTR]++;
+    kmutex_unlock (&ctxt->SwapListsLock);
+}
+
+static void
+PrintUserDma (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    E3_DMA *dma;
+
+    /* Dont call a function which takes locks unless we need to */
+    if (!(elan3_debug & DBG_DPROC))
+        return;
+
+    dma = (E3_DMA *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    PRINTF4 (ctxt, DBG_DPROC, "DMA: type %08x size %08x source %08x dest %08x\n",
+            fuword ((int *) &dma->dma_type), fuword ((int *) &dma->dma_size), 
+            fuword ((int *) &dma->dma_source), fuword ((int *) &dma->dma_dest));
+    PRINTF4 (ctxt, DBG_DPROC, "DMA: Dest %08x %08x  Local %08x %08x\n",
+            fuword ((int *) &dma->dma_destEvent), fuword ((int *) &dma->dma_destCookieProc), 
+            fuword ((int *) &dma->dma_srcEvent), fuword ((int *) &dma->dma_srcCookieProc));
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/elan3mmu_generic.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/elan3mmu_generic.c        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/elan3mmu_generic.c     2005-06-01 23:12:54.573443056 -0400
@@ -0,0 +1,3255 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elan3mmu_generic.c,v 1.75.2.1 2004/12/14 10:19:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/vm/elan3mmu_generic.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+
+#ifdef CONFIG_MPSAS
+#  define zero_all_ptbls
+#endif
+
+/*
+ * Debugging
+ */
+int    elan3mmu_debug = 0;
+
+#define        N_L3PTBL_MTX    (0x20)
+#define        N_L2PTBL_MTX    (0x40)
+#define        N_L1PTBL_MTX    (0x20)
+
+#define        L3PTBL_MTX_HASH(p) \
+       ((((uintptr_t)(p) >> 12) ^ ((uintptr_t)(p) >> 2)) & (N_L3PTBL_MTX - 1))
+static spinlock_t l3ptbl_lock[N_L3PTBL_MTX];
+
+#define        L2PTBL_MTX_HASH(p)   \
+       ((((uintptr_t)(p) >> 12) ^ ((uintptr_t)(p) >> 2)) & (N_L2PTBL_MTX - 1))
+static spinlock_t l2ptbl_lock[N_L2PTBL_MTX];
+
+#define        L1PTBL_MTX_HASH(p)   \
+       ((((uintptr_t)(p) >> 12) ^ ((uintptr_t)(p) >> 2)) & (N_L1PTBL_MTX - 1))
+static spinlock_t l1ptbl_lock[N_L1PTBL_MTX];
+
+
+#define        BASE2VA(p)      ((E3_Addr)((p)->ptbl_base << 16))
+#define        VA2BASE(v)      ((u_short)(((uintptr_t)(v)) >> 16))
+
+ELAN3MMU_GLOBAL_STATS  elan3mmu_global_stats;
+
+static void          elan3mmu_flush_context_filter (ELAN3_DEV *dev, void *);
+static void          elan3mmu_unload_loop (ELAN3MMU *elan3mmu, ELAN3_PTBL *ptbl, int first_valid, int nptes, int flags);
+
+static ELAN3_PTBL    *elan3mmu_create_ptbls (ELAN3_DEV *dev, int level, int attr, int keep);
+static ELAN3_PTBL    *elan3mmu_ta_to_ptbl (ELAN3MMU *elan3mmu, ELAN3_PTP *ptp);
+
+static ELAN3_PTBL    *elan3mmu_alloc_pte    (ELAN3_DEV *dev, ELAN3MMU *elan3mmu, int *idx);
+void                 elan3mmu_free_lXptbl  (ELAN3_DEV *dev, ELAN3_PTBL *ptbl);
+
+void                 elan3mmu_free_pte  (ELAN3_DEV *dev,  ELAN3MMU *elan3mmu,  ELAN3_PTBL *ptbl_ptr, int idx);
+
+static ELAN3_PTBL    *elan3mmu_alloc_l1ptbl (ELAN3_DEV *dev, int attr, ELAN3MMU *elan3mmu);
+static ELAN3_PTBL    *elan3mmu_alloc_l2ptbl (ELAN3_DEV *dev, int attr, ELAN3_PTBL *parent, ELAN3MMU *elan3mmu,
+                                           E3_Addr base, spinlock_t **plock, unsigned long *flags);
+static ELAN3_PTBL    *elan3mmu_alloc_l3ptbl (ELAN3_DEV *dev, int attr, ELAN3_PTBL *parent, ELAN3MMU *elan3mmu,
+                                           E3_Addr base, spinlock_t **plock, unsigned long *flags);
+
+static int          elan3mmu_steal_this_ptbl (ELAN3_DEV *dev, ELAN3_PTBL *l3ptbl);
+static ELAN3_PTBL    *elan3mmu_steal_l3ptbl (ELAN3_DEV *dev, int attr);
+
+static spinlock_t   *elan3mmu_ptbl_to_lock (int level, ELAN3_PTBL *ptbl);
+
+/*
+ * Encoding of MMU permissions against access type,
+ * to allow quick permission checking against access 
+ * type.
+ */
+u_char elan3mmu_permissionTable[] =
+{
+    0xcc,      /* 11001100 ELAN3_PERM_NULL        */
+    0x01,      /* 00000001 ELAN3_PERM_LOCALREAD   */
+    0x05,      /* 00000101 ELAN3_PERM_READ        */
+    0x33,      /* 00110011 ELAN3_PERM_NOREMOTE    */
+    0x37,      /* 00110111 ELAN3_PERM_REMOTEREAD  */
+    0x3f,      /* 00111111 ELAN3_PERM_REMOTEWRITE */
+    0xf7,      /* 11110111 ELAN3_PERM_REMOTEEVENT */
+    0xff,      /* 11111111 ELAN3_PERM_REMOTEALL          */
+} ;
+
+void
+elan3mmu_init()
+{
+    register int i;
+
+    HAT_PRINTF0 (1, "elan3mmu_init: initialising elan mmu\n");
+
+    for (i = 0; i < N_L1PTBL_MTX; i++)
+       spin_lock_init (&l1ptbl_lock[i]);
+
+    for (i = 0; i < N_L2PTBL_MTX; i++)
+       spin_lock_init (&l2ptbl_lock[i]);
+
+    for (i = 0; i < N_L3PTBL_MTX; i++)
+       spin_lock_init (&l3ptbl_lock[i]);
+
+    elan3mmu_global_stats.version = ELAN3MMU_STATS_VERSION;
+
+    elan3mmu_init_osdep();
+}
+
+void
+elan3mmu_fini()
+{
+    register int i;
+
+    HAT_PRINTF0 (1, "elan3mmu_fini: finalising elan mmu\n");
+
+    for (i = 0; i < N_L1PTBL_MTX; i++)
+       spin_lock_destroy (&l1ptbl_lock[i]);
+
+    for (i = 0; i < N_L2PTBL_MTX; i++)
+       spin_lock_destroy (&l2ptbl_lock[i]);
+
+    for (i = 0; i < N_L3PTBL_MTX; i++)
+       spin_lock_destroy (&l3ptbl_lock[i]);
+
+    elan3mmu_fini_osdep();
+}
+
+ELAN3MMU *
+elan3mmu_alloc (ELAN3_CTXT *ctxt)
+{
+    ELAN3MMU  *elan3mmu;
+    ELAN3_PTBL *l1ptbl;
+
+    ALLOC_ELAN3MMU (elan3mmu, TRUE);
+    
+    spin_lock_init (&elan3mmu->elan3mmu_lock);
+
+    spin_lock (&elan3mmu->elan3mmu_lock);                      /* lock_lint */
+
+    elan3mmu->elan3mmu_ergns    = NULL;
+    elan3mmu->elan3mmu_etail    = NULL;
+    elan3mmu->elan3mmu_ergnlast = NULL;
+    elan3mmu->elan3mmu_mrgns    = NULL;
+    elan3mmu->elan3mmu_mtail    = NULL;
+    elan3mmu->elan3mmu_mrgnlast = NULL;
+    elan3mmu->elan3mmu_ctxt     = ctxt;
+
+    spin_lock_init (&elan3mmu->elan3mmu_lXptbl_lock);
+    elan3mmu->elan3mmu_lXptbl   = NULL;
+
+    spin_unlock (&elan3mmu->elan3mmu_lock);                    /* lock_lint */
+
+    l1ptbl = elan3mmu_alloc_l1ptbl(ctxt->Device, 0, elan3mmu);
+
+    elan3mmu->elan3mmu_ctp      = (sdramaddr_t) 0;
+    elan3mmu->elan3mmu_dev      = ctxt->Device;
+    elan3mmu->elan3mmu_l1ptbl   = l1ptbl;
+
+    /* Ensure that there are at least some level 3 page tables,  since if a level 2 and */
+    /* a level 3 table are allocated together, then the level 3 is allocated with the NO_ALLOC */
+    /* flag,  thus there MUST be at least one that can be stolen or on the free list */
+    if (elan3mmu->elan3mmu_dev->Level[PTBL_LEVEL_3].PtblFreeList == NULL)
+       elan3mmu_create_ptbls (elan3mmu->elan3mmu_dev, PTBL_LEVEL_3, 0, 0);
+
+    HAT_PRINTF1 (1, "elan3mmu_alloc: elan3mmu %p\n", elan3mmu);
+
+    elan3mmu_alloc_osdep (elan3mmu);
+
+    return (elan3mmu);
+}
+
+void 
+elan3mmu_free (ELAN3MMU *elan3mmu)
+{
+    ELAN3MMU_RGN   *rgn;
+    ELAN3_PTBL    *l1ptbl;
+    spinlock_t    *l1lock;
+    unsigned long   l1flags;
+    unsigned long   flags;
+
+    HAT_PRINTF1 (1, "elan3mmu_free : elan3mmu %p\n", elan3mmu);
+    
+    /*
+     * Invalidate the level1 page table,  since it's already removed
+     * from the context table, there is no need to flush the tlb.
+     */
+    l1ptbl = elan3mmu->elan3mmu_l1ptbl;
+    elan3mmu->elan3mmu_l1ptbl = NULL;
+    
+    if (elan3mmu_lock_ptbl (l1ptbl, LK_PTBL_FAILOK, elan3mmu, (E3_Addr) 0, PTBL_LEVEL_1, &l1lock, &l1flags) == LK_PTBL_OK)
+    {
+       elan3mmu_l1inval (elan3mmu, l1ptbl, PTE_UNLOAD_NOFLUSH);
+       elan3mmu_free_l1ptbl (elan3mmu->elan3mmu_dev, l1ptbl, l1lock, l1flags);
+    }
+
+    /*
+     * Free of any permission regions.
+     */
+    spin_lock (&elan3mmu->elan3mmu_lock);                                      /* lock_lint */
+    while ((rgn = elan3mmu->elan3mmu_mrgns) != NULL)
+    {
+       spin_lock_irqsave (&elan3mmu->elan3mmu_dev->IntrLock, flags);           /* lock_lint */
+       elan3mmu_removergn_elan (elan3mmu, rgn->rgn_ebase);
+       elan3mmu_removergn_main (elan3mmu, rgn->rgn_mbase);
+       spin_unlock_irqrestore (&elan3mmu->elan3mmu_dev->IntrLock, flags);      /* lock_lint */
+       
+       FREE_ELAN3MMU_RGN (rgn);
+    }
+    elan3mmu->elan3mmu_mrgnlast = NULL;
+    elan3mmu->elan3mmu_ergnlast = NULL;
+
+    /* 
+     * Free the lXptbl list
+     */
+    ASSERT (elan3mmu->elan3mmu_lXptbl == NULL); /* XXXX MRH need to add list removal */
+ 
+    elan3mmu->elan3mmu_lXptbl = NULL;
+    spin_lock_destroy (&elan3mmu->elan3mmu_lXptbl_lock);
+
+
+    spin_unlock (&elan3mmu->elan3mmu_lock);                                    /* lock_lint */
+
+    spin_lock_destroy (&elan3mmu->elan3mmu_lock);
+
+    FREE_ELAN3MMU (elan3mmu);
+}
+
+/*================================================================================*/
+/* Interface routines to device driver */
+static void
+elan3mmu_flush_context_filter (ELAN3_DEV *dev, void *arg)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    ASSERT ((read_reg32 (dev, Exts.InterruptReg) & (INT_DiscardingSysCntx | INT_DiscardingNonSysCntx)) == 
+           (INT_DiscardingSysCntx | INT_DiscardingNonSysCntx));
+
+    dev->FilterHaltQueued = 0;
+
+    write_reg32 (dev, Input_Context_Fil_Flush, 0);
+
+    HAT_PRINTF0 (1, "elan3mmu_flush_context_filter completed\n");
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+void
+elan3mmu_set_context_filter (ELAN3_DEV *dev, int ctx, int disabled, E3_uint32 Pend, E3_uint32 *Maskp)
+{
+    int         mctx = ctx & MAX_ROOT_CONTEXT_MASK;
+    sdramaddr_t ctp  = dev->ContextTable + mctx * sizeof (E3_ContextControlBlock);
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+
+    ASSERT ((mctx < 32 || mctx >= ELAN3_KCOMM_BASE_CONTEXT_NUM) ? (ctx & SYS_CONTEXT_BIT) : ! (ctx & SYS_CONTEXT_BIT));
+
+    elan3_sdram_writel (dev, ctp + offsetof (E3_ContextControlBlock, filter), 
+                 ((ctx & SYS_CONTEXT_BIT) ? E3_CCB_CNTX0 : 0) | (disabled ? E3_CCB_DISCARD_ALL : 0));
+
+    HAT_PRINTF4 (1, "elan3mmu_set_context_filter: ctx %x [%lx] -> %s (%x)\n", ctx, ctp,
+                disabled ? "up" : "down", elan3_sdram_readl (dev, ctp + offsetof (E3_ContextControlBlock, filter)));
+
+    /* queue a halt operation to flush the context filter while the inputter is halted */
+    if (dev->FilterHaltQueued == 0)
+    {
+       dev->FilterHaltQueued = 1;
+       QueueHaltOperation (dev, Pend, Maskp, INT_DiscardingSysCntx | INT_DiscardingNonSysCntx, 
+                           elan3mmu_flush_context_filter, NULL);
+    }
+}
+
+int
+elan3mmu_attach (ELAN3_DEV *dev, int ctx, ELAN3MMU *elan3mmu, sdramaddr_t routeTable, E3_uint32 routeMask)
+{
+    sdramaddr_t ctp;
+    ELAN3_PTP    trootptp;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+
+    ctx &= MAX_ROOT_CONTEXT_MASK;                                              /* Mask out all high bits in context */
+    
+    if (ctx < 0 || ctx >= dev->ContextTableSize)
+       return (EINVAL);
+
+    ctp = dev->ContextTable + ctx * sizeof (E3_ContextControlBlock);
+    
+    trootptp = elan3_readptp (dev, ctp + offsetof (E3_ContextControlBlock, rootPTP));
+    
+    if (ELAN3_PTP_TYPE(trootptp) != ELAN3_ET_INVALID)
+       return (EBUSY);
+
+    elan3mmu->elan3mmu_ctp = ctp;
+    
+    trootptp = PTBL_TO_PTADDR (elan3mmu->elan3mmu_l1ptbl) | ELAN3_ET_PTP;
+    
+    HAT_PRINTF4 (1, "elan3mmu_attach: ctp at %08lx : trootptp=%08x VPT_ptr=%08lx VPT_mask=%08x\n",
+                ctp, trootptp, routeTable, routeMask);
+    
+    elan3_writeptp (dev, ctp + offsetof (E3_ContextControlBlock, rootPTP), trootptp);
+    elan3_writeptp (dev, ctp + offsetof (E3_ContextControlBlock, VPT_ptr), routeTable);
+    elan3_writeptp (dev, ctp + offsetof (E3_ContextControlBlock, VPT_mask), routeMask);
+    
+    return (ESUCCESS);
+}
+
+void
+elan3mmu_detach (ELAN3_DEV *dev, int ctx)
+{
+    ELAN3_PTP    invalidptp = ELAN3_INVALID_PTP;
+    sdramaddr_t ctp;
+    
+    ctx &= MAX_ROOT_CONTEXT_MASK;                                              /* Mask out all high bits in context */
+    
+    if (ctx < 0 || ctx >= dev->ContextTableSize)
+       return;
+    
+    ctp = dev->ContextTable + ctx * sizeof (E3_ContextControlBlock);
+    
+    HAT_PRINTF1 (1, "elan3mmu_detach: clearing ptp at %lx\n", ctp);
+    
+    elan3_writeptp (dev, ctp + offsetof (E3_ContextControlBlock, rootPTP), invalidptp);
+    elan3_writeptp (dev, ctp + offsetof (E3_ContextControlBlock, VPT_mask), 0);
+    elan3_writeptp (dev, ctp + offsetof (E3_ContextControlBlock, VPT_ptr), 0);
+    
+    ElanFlushTlb (dev);
+}
+
+int
+elan3mmu_reference (ELAN3MMU *elan3mmu, int ctx)
+{
+    ELAN3_DEV              *dev = elan3mmu->elan3mmu_dev;
+    sdramaddr_t            ctp;
+    E3_ContextControlBlock ccb;
+    ELAN3_PTP               trootptp;
+
+    ctx &= MAX_ROOT_CONTEXT_MASK;                                              /* Mask out all high bits in context */
+    
+    if (ctx < 0 || ctx >= dev->ContextTableSize)
+       return (EINVAL);
+
+    ctp = dev->ContextTable + ctx * sizeof (E3_ContextControlBlock);
+
+    trootptp = elan3_readptp (dev, ctp + offsetof (E3_ContextControlBlock, rootPTP));
+    
+    if (ELAN3_PTP_TYPE(trootptp) != ELAN3_ET_INVALID)
+       return (EBUSY);
+    
+    elan3_sdram_copyl_from_sdram (dev, elan3mmu->elan3mmu_ctp, &ccb, sizeof (E3_ContextControlBlock));
+    elan3_sdram_copyl_to_sdram (dev, &ccb, ctp, sizeof (E3_ContextControlBlock));
+    
+    return (ESUCCESS);
+    
+}
+/*================================================================================*/
+/* Elan permission regions. */
+
+/* elan address region management */
+ELAN3MMU_RGN *
+elan3mmu_findrgn_elan (ELAN3MMU *elan3mmu,
+                      E3_Addr addr, int tail)
+{
+    ELAN3MMU_RGN *next = NULL;
+    ELAN3MMU_RGN *rgn;
+    ELAN3MMU_RGN *hirgn;
+    ELAN3MMU_RGN *lorgn;
+    E3_Addr       base;
+    E3_Addr       lastaddr;
+    int                  forward;
+
+    ASSERT (SPINLOCK_HELD (&elan3mmu->elan3mmu_dev->IntrLock) || SPINLOCK_HELD (&elan3mmu->elan3mmu_lock));
+
+    if (elan3mmu->elan3mmu_ergns == NULL)
+       return (NULL);
+
+    rgn = elan3mmu->elan3mmu_ergnlast;
+    if (rgn == NULL)
+       rgn = elan3mmu->elan3mmu_ergns;
+
+    forward = 0;
+    if ((u_long) (base = rgn->rgn_ebase) < (u_long)addr)
+    {
+       if ((u_long)addr <= ((u_long) base + rgn->rgn_len - 1))
+           return (rgn);                                       /* ergnlast contained addr */
+
+       hirgn = elan3mmu->elan3mmu_etail;
+
+       if ((u_long) (lastaddr = (hirgn->rgn_ebase + hirgn->rgn_len - 1)) < (u_long) addr)
+           return (tail ? hirgn : NULL);                       /* addr is out of range */
+       
+       if ((u_long) (addr - base) > (u_long) (lastaddr - addr))
+           rgn = hirgn;
+       else
+       {
+           rgn = rgn->rgn_enext;
+           forward++;
+       }
+    }
+    else
+    {
+       lorgn = elan3mmu->elan3mmu_ergns;
+
+       if ((u_long)lorgn->rgn_ebase > (u_long) addr)
+           return (lorgn);                                     /* lowest regions is higher than addr */
+       if ((u_long)(addr - lorgn->rgn_ebase) < (u_long) (base - addr))
+       {
+           rgn = lorgn;                                        /* search forward from head */
+           forward++;
+       }
+    }
+    if (forward)
+    {
+       while ((u_long)(rgn->rgn_ebase + rgn->rgn_len - 1) < (u_long)addr)
+           rgn = rgn->rgn_enext;
+
+       if ((u_long)rgn->rgn_ebase <= (u_long)addr)
+           elan3mmu->elan3mmu_ergnlast = rgn;
+       return (rgn);
+    }
+    else
+    {
+       while ((u_long)rgn->rgn_ebase > (u_long)addr)
+       {
+           next = rgn;
+           rgn = rgn->rgn_eprev;
+       }
+
+       if ((u_long) (rgn->rgn_ebase + rgn->rgn_len - 1) < (u_long)addr)
+           return (next);
+       else
+       {
+           elan3mmu->elan3mmu_ergnlast = rgn;
+           return (rgn);
+       }
+    }
+}
+
+int
+elan3mmu_addrgn_elan (ELAN3MMU *elan3mmu, ELAN3MMU_RGN *nrgn)
+{
+    ELAN3MMU_RGN *rgn   = elan3mmu_findrgn_elan (elan3mmu, nrgn->rgn_ebase, 1);
+    E3_Addr       nbase = nrgn->rgn_ebase;
+    E3_Addr      ntop  = nbase + nrgn->rgn_len - 1; /* avoid wrap */
+    E3_Addr      base;
+
+    ASSERT (SPINLOCK_HELD (&elan3mmu->elan3mmu_dev->IntrLock) && SPINLOCK_HELD (&elan3mmu->elan3mmu_lock));
+
+    if (rgn == NULL)
+    {
+       elan3mmu->elan3mmu_ergns = elan3mmu->elan3mmu_etail = nrgn;
+       nrgn->rgn_enext = nrgn->rgn_eprev = NULL;
+    }
+    else
+    {
+       base = rgn->rgn_ebase;
+
+       if ((u_long)(base + rgn->rgn_len - 1) < (u_long)nbase)  /* top of region below requested address */
+       {                                                       /* so insert after region (and hence at end */
+           nrgn->rgn_eprev = rgn;                              /* of list */
+           nrgn->rgn_enext = NULL;
+           rgn->rgn_enext = elan3mmu->elan3mmu_etail = nrgn;
+       }
+       else
+       {
+           if ((u_long)nbase >= (u_long)base || (u_long)ntop >= (u_long)base)
+               return (-1);                                    /* overlapping region */
+
+           nrgn->rgn_enext = rgn;                              /* insert before region */
+           nrgn->rgn_eprev = rgn->rgn_eprev;
+           rgn->rgn_eprev  = nrgn;
+           if (elan3mmu->elan3mmu_ergns == rgn)
+               elan3mmu->elan3mmu_ergns = nrgn;
+           else
+               nrgn->rgn_eprev->rgn_enext = nrgn;
+       }
+    }
+    elan3mmu->elan3mmu_ergnlast = nrgn;
+    
+    return (0);
+}
+
+ELAN3MMU_RGN *
+elan3mmu_removergn_elan (ELAN3MMU *elan3mmu, E3_Addr addr)
+{
+    ELAN3MMU_RGN *rgn = elan3mmu_findrgn_elan (elan3mmu, addr, 0);
+    
+    ASSERT (SPINLOCK_HELD (&elan3mmu->elan3mmu_dev->IntrLock) && SPINLOCK_HELD (&elan3mmu->elan3mmu_lock));
+
+    if (rgn == NULL || rgn->rgn_ebase != addr)
+       return (NULL);
+    
+    elan3mmu->elan3mmu_ergnlast = rgn->rgn_enext;
+    if (rgn == elan3mmu->elan3mmu_etail)
+       elan3mmu->elan3mmu_etail = rgn->rgn_eprev;
+    else
+       rgn->rgn_enext->rgn_eprev = rgn->rgn_eprev;
+    
+    if (rgn == elan3mmu->elan3mmu_ergns)
+       elan3mmu->elan3mmu_ergns = rgn->rgn_enext;
+    else
+       rgn->rgn_eprev->rgn_enext = rgn->rgn_enext;
+
+    return (rgn);
+}
+
+ELAN3MMU_RGN *
+elan3mmu_rgnat_elan (ELAN3MMU *elan3mmu, E3_Addr addr)
+{
+    ELAN3MMU_RGN *rgn = elan3mmu_findrgn_elan (elan3mmu, addr, 0);
+    E3_Addr       base;
+
+    if (rgn != NULL && (u_long)(base = rgn->rgn_ebase) <= (u_long)addr && (u_long)addr <= (u_long)(base + rgn->rgn_len - 1))
+       return (rgn);
+    return (NULL);
+}
+
+/* main address region management */
+ELAN3MMU_RGN *
+elan3mmu_findrgn_main (ELAN3MMU *elan3mmu,
+                      caddr_t addr, int tail)
+{
+    ELAN3MMU_RGN *next = NULL;
+    ELAN3MMU_RGN *rgn;
+    ELAN3MMU_RGN *hirgn;
+    ELAN3MMU_RGN *lorgn;
+    caddr_t       lastaddr;
+    caddr_t       base;
+    int                  forward;
+
+    ASSERT (SPINLOCK_HELD (&elan3mmu->elan3mmu_dev->IntrLock) || SPINLOCK_HELD (&elan3mmu->elan3mmu_lock));
+
+    if (elan3mmu->elan3mmu_mrgns == NULL)
+       return (NULL);
+
+    rgn = elan3mmu->elan3mmu_mrgnlast;
+    if (rgn == NULL)
+       rgn = elan3mmu->elan3mmu_mrgns;
+
+    forward = 0;
+    if ((base = rgn->rgn_mbase) < addr)
+    {
+       if (addr <= (base + rgn->rgn_len - 1))
+           return (rgn);                                       /* ergnlast contained addr */
+
+       hirgn = elan3mmu->elan3mmu_mtail;
+       if ((lastaddr = hirgn->rgn_mbase + hirgn->rgn_len - 1) < addr)
+           return (tail ? hirgn : NULL);                       /* addr is out of range */
+       
+       if ((addr - base) > (lastaddr - addr))
+           rgn = hirgn;
+       else
+       {
+           rgn = rgn->rgn_mnext;
+           forward++;
+       }
+    }
+    else
+    {
+       lorgn = elan3mmu->elan3mmu_mrgns;
+       if (lorgn->rgn_mbase > addr)
+           return (lorgn);                                     /* lowest regions is higher than addr */
+       if ((addr - lorgn->rgn_mbase) < (base - addr))
+       {
+           rgn = lorgn;                                        /* search forward from head */
+           forward++;
+       }
+    }
+    if (forward)
+    {
+       while ((rgn->rgn_mbase + rgn->rgn_len - 1) < addr)
+           rgn = rgn->rgn_mnext;
+
+       if (rgn->rgn_mbase <= addr)
+           elan3mmu->elan3mmu_mrgnlast = rgn;
+       return (rgn);
+    }
+    else
+    {
+       while (rgn->rgn_mbase > addr)
+       {
+           next = rgn;
+           rgn = rgn->rgn_mprev;
+       }
+       if ((rgn->rgn_mbase + rgn->rgn_len - 1) < addr)
+           return (next);
+       else
+       {
+           elan3mmu->elan3mmu_mrgnlast = rgn;
+           return (rgn);
+       }
+    }
+}
+
+int
+elan3mmu_addrgn_main (ELAN3MMU *elan3mmu, ELAN3MMU_RGN *nrgn)
+{
+    ELAN3MMU_RGN *rgn   = elan3mmu_findrgn_main (elan3mmu, nrgn->rgn_mbase, 1);
+    caddr_t       nbase = nrgn->rgn_mbase;
+    caddr_t      ntop  = nbase + nrgn->rgn_len - 1;
+    caddr_t      base;
+
+    ASSERT (SPINLOCK_HELD (&elan3mmu->elan3mmu_dev->IntrLock) && SPINLOCK_HELD (&elan3mmu->elan3mmu_lock));
+
+    if (rgn == NULL)
+    {
+       elan3mmu->elan3mmu_mrgns = elan3mmu->elan3mmu_mtail = nrgn;
+       nrgn->rgn_mnext = nrgn->rgn_mprev = NULL;
+    }
+    else
+    {
+       base = rgn->rgn_mbase;
+
+       if ((base + rgn->rgn_len - 1) < nbase)                  /* top of region below requested address */
+       {                                                       /* so insert after region (and hence at end */
+           nrgn->rgn_mprev = rgn;                              /* of list */
+           nrgn->rgn_mnext = NULL;
+           rgn->rgn_mnext = elan3mmu->elan3mmu_mtail = nrgn;
+       }
+       else
+       {
+           if (nbase >= base || ntop >= base)
+               return (-1);                                    /* overlapping region */
+
+           nrgn->rgn_mnext = rgn;                              /* insert before region */
+           nrgn->rgn_mprev = rgn->rgn_mprev;
+           rgn->rgn_mprev  = nrgn;
+           if (elan3mmu->elan3mmu_mrgns == rgn)
+               elan3mmu->elan3mmu_mrgns = nrgn;
+           else
+               nrgn->rgn_mprev->rgn_mnext = nrgn;
+       }
+    }
+    elan3mmu->elan3mmu_mrgnlast = nrgn;
+    
+    return (0);
+}
+
+ELAN3MMU_RGN *
+elan3mmu_removergn_main (ELAN3MMU *elan3mmu, caddr_t addr)
+{
+    ELAN3MMU_RGN *rgn = elan3mmu_findrgn_main (elan3mmu, addr, 0);
+    
+    ASSERT (SPINLOCK_HELD (&elan3mmu->elan3mmu_dev->IntrLock) && SPINLOCK_HELD (&elan3mmu->elan3mmu_lock));
+
+    if (rgn == NULL || rgn->rgn_mbase != addr)
+       return (NULL);
+    
+    elan3mmu->elan3mmu_mrgnlast = rgn->rgn_mnext;
+    if (rgn == elan3mmu->elan3mmu_mtail)
+       elan3mmu->elan3mmu_mtail = rgn->rgn_mprev;
+    else
+       rgn->rgn_mnext->rgn_mprev = rgn->rgn_mprev;
+    
+    if (rgn == elan3mmu->elan3mmu_mrgns)
+       elan3mmu->elan3mmu_mrgns = rgn->rgn_mnext;
+    else
+       rgn->rgn_mprev->rgn_mnext = rgn->rgn_mnext;
+
+    return (rgn);
+}
+
+ELAN3MMU_RGN *
+elan3mmu_rgnat_main (ELAN3MMU *elan3mmu, caddr_t addr)
+{
+    ELAN3MMU_RGN *rgn = elan3mmu_findrgn_main (elan3mmu, addr, 0);
+    caddr_t       base;
+
+    if (rgn != NULL && (base = rgn->rgn_mbase) <= addr && addr <= (base + rgn->rgn_len - 1))
+       return (rgn);
+    return (NULL);
+}
+
+int
+elan3mmu_setperm (ELAN3MMU *elan3mmu,
+                 caddr_t   maddr,
+                 E3_Addr   eaddr,
+                 u_int     len,
+                 u_int     perm)
+{
+    ELAN3_DEV     *dev = elan3mmu->elan3mmu_dev;
+    ELAN3MMU_RGN *nrgn;
+    unsigned long  flags;
+
+    HAT_PRINTF4 (1, "elan3mmu_setperm: user %p elan %08x len %x perm %x\n", maddr, eaddr, len, perm);
+
+    if ((((uintptr_t) maddr) & PAGEOFFSET) || (eaddr & PAGEOFFSET) || (len & PAGEOFFSET)) 
+    {
+        HAT_PRINTF0 (1, "elan3mmu_setperm:  alignment failure\n");
+       return (EINVAL);
+    }
+
+    if (((uintptr_t) maddr + len - 1) < (uintptr_t) maddr || ((u_long)eaddr + len - 1) < (u_long)eaddr) 
+    {
+       HAT_PRINTF0 (1, "elan3mmu_setperm:  range failure\n");
+       return (EINVAL);
+    }
+
+    ALLOC_ELAN3MMU_RGN(nrgn, TRUE);
+    
+    spin_lock (&elan3mmu->elan3mmu_lock);
+    nrgn->rgn_mbase = maddr;
+    nrgn->rgn_ebase = eaddr;
+    nrgn->rgn_len   = len;
+    nrgn->rgn_perm  = perm;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    if (elan3mmu_addrgn_elan (elan3mmu, nrgn) < 0)
+    {
+       HAT_PRINTF0 (1, "elan3mmu_setperm:  elan address exists\n");
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       spin_unlock (&elan3mmu->elan3mmu_lock);
+
+       FREE_ELAN3MMU_RGN (nrgn);
+       return (EINVAL);
+    }
+    
+    if (elan3mmu_addrgn_main (elan3mmu, nrgn) < 0)
+    {
+       HAT_PRINTF0 (1, "elan3mmu_setperm:  main address exists\n");
+       elan3mmu_removergn_elan (elan3mmu, eaddr);
+
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       spin_unlock (&elan3mmu->elan3mmu_lock);
+
+       FREE_ELAN3MMU_RGN (nrgn);
+       return (EINVAL);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+
+    return (ESUCCESS);
+}
+
+void
+elan3mmu_clrperm (ELAN3MMU *elan3mmu,
+                 E3_Addr   addr,
+                 u_int     len)
+{
+    E3_Addr       raddr;
+    E3_Addr       rtop;
+    ELAN3MMU_RGN *nrgn;
+    ELAN3MMU_RGN *rgn;
+    ELAN3MMU_RGN *rgn_next;
+    u_int        ssize;
+    unsigned long flags;
+    int                  res;
+
+    HAT_PRINTF2 (1, "elan3mmu_clrperm: elan %08x len %x\n", addr, len);
+
+    raddr = (addr & PAGEMASK);
+    rtop = ((addr + len - 1) & PAGEMASK) + PAGEOFFSET;
+
+    ALLOC_ELAN3MMU_RGN (nrgn, TRUE);
+
+    spin_lock (&elan3mmu->elan3mmu_lock);
+    
+    for (rgn = elan3mmu_findrgn_elan (elan3mmu, addr, 0); rgn != NULL; rgn = rgn_next)
+    {
+       if (rtop < rgn->rgn_ebase)                              /* rtop was in a gap */
+           break;
+       
+       rgn_next = rgn->rgn_enext;                              /* Save next region pointer */
+       
+       if (raddr <= rgn->rgn_ebase && rtop >= (rgn->rgn_ebase + rgn->rgn_len - 1)) 
+       {
+           /* whole region is cleared */
+           elan3mmu_unload (elan3mmu, rgn->rgn_ebase, rgn->rgn_len, PTE_UNLOAD);
+           
+           spin_lock_irqsave (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+           elan3mmu_removergn_elan (elan3mmu, rgn->rgn_ebase);
+           elan3mmu_removergn_main (elan3mmu, rgn->rgn_mbase);
+           spin_unlock_irqrestore (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+
+           FREE_ELAN3MMU_RGN (rgn);
+       }
+       else if (raddr <= rgn->rgn_ebase)
+       {
+           /* clearing at beginning, so shrink size and increment base ptrs */
+           ssize = rtop - rgn->rgn_ebase + 1;
+
+           elan3mmu_unload (elan3mmu, rgn->rgn_ebase, ssize, PTE_UNLOAD);
+           
+           spin_lock_irqsave (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+           rgn->rgn_mbase += ssize;
+           rgn->rgn_ebase += ssize;
+           rgn->rgn_len   -= ssize;
+           spin_unlock_irqrestore (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+           
+       }
+       else if (rtop >= (rgn->rgn_ebase + rgn->rgn_len - 1))
+       {
+           /* clearing at end, so just shrink length of region */
+           ssize = ((rgn->rgn_ebase + rgn->rgn_len - 1) - raddr) + 1;
+
+           elan3mmu_unload (elan3mmu, raddr, ssize, PTE_UNLOAD);
+
+           spin_lock_irqsave (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+           rgn->rgn_len -= ssize;
+           spin_unlock_irqrestore (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+       }
+       else
+       {
+           /* the section to go is in the middle,  so need to  */
+           /* split it into two regions */
+           elan3mmu_unload (elan3mmu, raddr, rtop - raddr + 1, PTE_UNLOAD);
+
+           spin_lock_irqsave (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+
+           ASSERT (nrgn != NULL);
+
+           nrgn->rgn_mbase = rgn->rgn_mbase + (rtop - rgn->rgn_ebase + 1);;
+           nrgn->rgn_ebase = rtop + 1;
+           nrgn->rgn_len   = ((rgn->rgn_ebase + rgn->rgn_len - 1) - rtop);
+           nrgn->rgn_perm  = rgn->rgn_perm;
+
+           rgn->rgn_len = (raddr - rgn->rgn_ebase);            /* shrink original region */
+
+           res = elan3mmu_addrgn_elan (elan3mmu, nrgn);        /* insert new region */
+           ASSERT (res == 0);                                  /* which cannot fail */
+
+           res = elan3mmu_addrgn_main (elan3mmu, nrgn);        
+           ASSERT (res == 0);
+           spin_unlock_irqrestore (&elan3mmu->elan3mmu_dev->IntrLock, flags);
+
+           nrgn = NULL;
+       }
+    }
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+
+    if (nrgn != NULL)
+       FREE_ELAN3MMU_RGN (nrgn);
+}
+
+int
+elan3mmu_checkperm (ELAN3MMU *elan3mmu,
+                   E3_Addr   addr,
+                   u_int     len,
+                   u_int     access)
+{
+    E3_Addr     raddr = (((E3_Addr) addr) & PAGEMASK);
+    u_int        rtop = ((addr + len - 1) & PAGEMASK) + PAGEOFFSET;
+    u_int       rsize = rtop - raddr + 1;
+    ELAN3MMU_RGN *rgn;
+
+    HAT_PRINTF3 (1, "elan3mmu_checkperm: user %08x len %x access %x\n", addr, len, access);
+    
+    
+    if ((raddr + rsize - 1) < raddr)
+       return (ENOMEM);
+    
+    spin_lock (&elan3mmu->elan3mmu_lock);
+    if ((rgn = elan3mmu_rgnat_elan (elan3mmu, raddr)) == (ELAN3MMU_RGN *) NULL)
+    {
+       spin_unlock (&elan3mmu->elan3mmu_lock);
+       return (ENOMEM);
+    }
+    else
+    {
+       register int ssize;
+       
+       for (; rsize != 0; rsize -= ssize, raddr += ssize)
+       {
+           if (raddr > (rgn->rgn_ebase + rgn->rgn_len - 1))
+           {
+               rgn  = rgn->rgn_enext;
+               
+               if (rgn == NULL || raddr != rgn->rgn_ebase)
+               {
+                   spin_unlock (&elan3mmu->elan3mmu_lock);
+                   return (ENOMEM);
+               }
+           }
+           if ((raddr + rsize - 1) > (rgn->rgn_ebase + rgn->rgn_len - 1))
+               ssize = ((rgn->rgn_ebase + rgn->rgn_len - 1) - raddr) + 1;
+           else
+               ssize = rsize;
+           
+           HAT_PRINTF4 (1, "elan3mmu_checkperm : rgn %x -> %x perm %x access %x\n",
+                        rgn->rgn_ebase, rgn->rgn_ebase + rgn->rgn_len, rgn->rgn_perm, access);
+
+           if (ELAN3_INCOMPAT_ACCESS (rgn->rgn_perm, access))
+           {
+               spin_unlock (&elan3mmu->elan3mmu_lock);
+               return (EACCES);
+           }
+       }
+    }
+    
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+    
+    return (ESUCCESS);
+}
+
+caddr_t
+elan3mmu_mainaddr (ELAN3MMU *elan3mmu, E3_Addr addr)
+{
+    ELAN3MMU_RGN *rgn;
+    caddr_t      raddr;
+    
+    spin_lock (&elan3mmu->elan3mmu_lock);
+    if ((rgn = elan3mmu_rgnat_elan (elan3mmu, addr)) == (ELAN3MMU_RGN *) NULL)
+       raddr = NULL;
+    else
+       raddr = rgn->rgn_mbase + (addr - rgn->rgn_ebase);
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+
+    return (raddr);
+}
+
+E3_Addr
+elan3mmu_elanaddr (ELAN3MMU *elan3mmu, caddr_t addr)
+{
+    ELAN3MMU_RGN *rgn;
+    E3_Addr       raddr;
+
+    spin_lock (&elan3mmu->elan3mmu_lock);
+    if ((rgn = elan3mmu_rgnat_main (elan3mmu, addr)) == (ELAN3MMU_RGN *) NULL)
+       raddr = (E3_Addr) 0;
+    else
+       raddr = rgn->rgn_ebase + (addr - rgn->rgn_mbase);
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+
+    return (raddr);
+}
+
+void
+elan3mmu_displayrgns(ELAN3MMU *elan3mmu)
+{
+    ELAN3MMU_RGN *rgn;
+
+    spin_lock (&elan3mmu->elan3mmu_lock);
+    HAT_PRINTF0 (1, "elan3mmu_displayrgns: main regions\n");
+    for (rgn = elan3mmu->elan3mmu_mrgns; rgn; rgn = (rgn->rgn_mnext == elan3mmu->elan3mmu_mrgns) ? NULL : rgn->rgn_mnext)
+       HAT_PRINTF5 (1, "    RGN %p ebase %08x mbase %p len %08x perm %08x\n", rgn, rgn->rgn_ebase, rgn->rgn_mbase, rgn->rgn_len, rgn->rgn_perm);
+    HAT_PRINTF0 (1, "elan3mmu_displayrgns: elan regions\n");
+    for (rgn = elan3mmu->elan3mmu_ergns; rgn; rgn = (rgn->rgn_enext == elan3mmu->elan3mmu_ergns) ? NULL : rgn->rgn_enext)
+       HAT_PRINTF5 (1, "    RGN %p ebase %08x mbase %p len %08x perm %08x\n", rgn, rgn->rgn_ebase, rgn->rgn_mbase, rgn->rgn_len, rgn->rgn_perm);
+
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+}
+
+/*============================================================================*/
+/* Private functions */
+#define ELAN3_PTE_IS_VALID(ptbl, pte)  \
+          ((ptbl->ptbl_flags & PTBL_KERNEL) ? \
+          (pte&(~ELAN3_PTE_REF)) != elan3mmu_kernel_invalid_pte(ptbl->ptbl_elan3mmu) : \
+          ELAN3_PTE_VALID(pte))
+
+void
+elan3mmu_expand (ELAN3MMU *elan3mmu, E3_Addr addr, int len, int level, int attr)
+{
+    ELAN3_PTBL          *ptbl;
+    sdramaddr_t                pte;
+    spinlock_t        *lock;
+    u_int              span;
+    unsigned long       flags;
+
+    HAT_PRINTF3 (1, "elan3mmu_expand: elan3mmu %p %08x to %08x\n", elan3mmu, 
+                addr, addr + len);
+
+    for ( ; len != 0; addr += span, len -= span)
+    {
+       /* as we asked for level 3 we know its a pte */
+       pte = elan3mmu_ptealloc (elan3mmu, addr, level, &ptbl, &lock, attr, &flags);
+
+       switch (level)
+       {
+       case PTBL_LEVEL_3:
+           span = MIN(len, ELAN3_L3_PTSIZE - ((E3_Addr) addr & ELAN3_L3_PTOFFSET));
+           break;
+       case PTBL_LEVEL_2:
+           span = MIN(len, ELAN3_L2_PTSIZE - ((E3_Addr) addr & ELAN3_L2_PTOFFSET));
+           break;
+       default:
+           span = len;
+           break;
+       }
+       
+       if (pte != (sdramaddr_t) 0)
+           elan3mmu_unlock_ptbl (ptbl, lock, flags);
+    }
+}
+
+void
+elan3mmu_reserve (ELAN3MMU *elan3mmu, E3_Addr addr, u_int npages, sdramaddr_t *ptes)
+{
+    ELAN3_PTBL          *ptbl;
+    sdramaddr_t                pte;
+    spinlock_t        *lock;
+    u_int              span;
+    int                        len;
+    int                        i;
+    unsigned long       flags;
+
+    HAT_PRINTF3 (1, "elan3mmu_reserve: elan3mmu %p %08x to %08x\n", elan3mmu, 
+                addr, addr + (npages << ELAN3_PAGE_SHIFT));
+
+    for (len = (npages << ELAN3_PAGE_SHIFT); len != 0; addr += span, len -= span)
+    {
+       /* as we asked for level 3 we know its a pte */
+       pte = elan3mmu_ptealloc (elan3mmu, addr, 3, &ptbl, &lock, 0, &flags);
+
+       span = MIN(len, ELAN3_L3_PTSIZE - ((E3_Addr) addr & ELAN3_L3_PTOFFSET));
+       
+       if (ptes != NULL)
+       {
+           for (i = 0; i < span; i += ELAN3_PAGE_SIZE, pte += ELAN3_PTE_SIZE)
+               *ptes++ = pte;
+           ptbl->ptbl_valid += (span >> ELAN3_PAGE_SHIFT);
+
+           HAT_PRINTF4 (2, "elan3mmu_reserve: inc valid for level %d ptbl %p to %d   (%d)\n", 
+                    PTBL_LEVEL(ptbl->ptbl_flags), ptbl, ptbl->ptbl_valid, (span >> ELAN3_PAGE_SHIFT));
+
+       }
+
+       elan3mmu_unlock_ptbl (ptbl, lock, flags);
+    }
+}
+
+void
+elan3mmu_release (ELAN3MMU *elan3mmu, E3_Addr addr, u_int npages, sdramaddr_t *ptes)
+{
+    ELAN3_DEV           *dev = elan3mmu->elan3mmu_dev;
+    ELAN3_PTBL          *ptbl;
+    sdramaddr_t                pte;
+    ELAN3_PTE          tpte;
+    spinlock_t        *lock;
+    u_int              span;
+    int                        len;
+    int                        i;
+    int                        level;
+    unsigned long       flags;
+    
+    HAT_PRINTF3 (1, "elan3mmu_release: elan3mmu %p %08x to %08x\n", elan3mmu, 
+                addr, addr + (npages << ELAN3_PAGE_SHIFT));
+
+    if (ptes == NULL)
+       return;
+
+    tpte = elan3mmu_kernel_invalid_pte (elan3mmu);
+
+    for (len = (npages << ELAN3_PAGE_SHIFT); len != 0; addr += span, len -= span)
+    {
+       /* as we asked for level 3 we know its a pte */
+       pte = elan3mmu_ptefind(elan3mmu, addr, &level, &ptbl, &lock, &flags);
+       ASSERT (level == PTBL_LEVEL_3);
+
+       span = MIN(len, ELAN3_L3_PTSIZE - ((E3_Addr) addr & ELAN3_L3_PTOFFSET));
+
+
+       for (i = 0 ; i < span; i += ELAN3_PAGE_SIZE, pte += ELAN3_PTE_SIZE)
+           elan3_writepte (dev, pte, tpte);
+       ptbl->ptbl_valid -= (span >> ELAN3_PAGE_SHIFT);
+
+       HAT_PRINTF3 (2, "elan3mmu_release: inc valid for level %d ptbl %p to %d\n", 
+                    PTBL_LEVEL(ptbl->ptbl_flags), ptbl, ptbl->ptbl_valid);
+
+       elan3mmu_unlock_ptbl (ptbl, lock, flags);
+    }
+    ElanFlushTlb (elan3mmu->elan3mmu_dev);
+}
+
+void
+elan3mmu_pteload (ELAN3MMU *elan3mmu, int level, E3_Addr addr, physaddr_t paddr, int perm, int attr)
+    
+{
+    ELAN3_DEV     *dev;
+    ELAN3_PTBL    *ptbl;
+    spinlock_t   *lock;
+    unsigned long flags;
+    ELAN3_PTE      newpte;
+    ELAN3_PTE      oldpte;
+    sdramaddr_t   pte;
+
+    ASSERT((level == PTBL_LEVEL_2) || (level == PTBL_LEVEL_3));
+
+    /* Generate the new pte which we're going to load */
+    dev = elan3mmu->elan3mmu_dev;
+
+    newpte = elan3mmu_phys_to_pte (dev, paddr, perm);
+    
+    if (attr & PTE_LOAD_BIG_ENDIAN)
+       newpte |= ELAN3_PTE_BIG_ENDIAN;
+
+    HAT_PRINTF4 (1, "elan3mmu_pteload: elan3mmu %p level %d addr %x pte %llx\n", elan3mmu, level, addr, (long long) newpte);
+    HAT_PRINTF5 (1, "elan3mmu_pteload:%s%s%s perm=%d phys=%llx\n",
+                (newpte & ELAN3_PTE_LOCAL)  ? " local" : "",
+                (newpte & ELAN3_PTE_64_BIT)     ? " 64 bit" : "",
+                (newpte & ELAN3_PTE_BIG_ENDIAN) ? " big-endian" : " little-endian",
+                (u_int) (newpte & ELAN3_PTE_PERM_MASK) >> ELAN3_PTE_PERM_SHIFT,
+                (unsigned long long) (newpte & ELAN3_PTE_PFN_MASK));
+                 
+    if (level == PTBL_LEVEL_3)
+       pte = elan3mmu_ptealloc (elan3mmu, addr, level, &ptbl, &lock, attr, &flags);
+    else
+    {
+       sdramaddr_t ptp = elan3mmu_ptealloc (elan3mmu, addr, level, &ptbl, &lock, attr, &flags);
+
+       pte = elan3mmu_ptp2pte (elan3mmu, ptp, level);
+
+       HAT_PRINTF3 (2, "elan3mmu_pteload: level %d ptp at %lx => pte at %lx\n", level, ptp, pte);
+    }
+
+    if (pte == (sdramaddr_t) 0)
+    {
+       ASSERT (level == PTBL_LEVEL_3 && (attr & (PTE_NO_SLEEP | PTE_NO_STEAL)) == (PTE_NO_SLEEP | PTE_NO_STEAL));
+       return;
+    }
+
+    ASSERT (ptbl->ptbl_elan3mmu == elan3mmu);
+    ASSERT (PTBL_LEVEL(ptbl->ptbl_flags) == level);
+    ASSERT (PTBL_IS_LOCKED (ptbl->ptbl_flags));
+    
+    oldpte = elan3_readpte (dev, pte);
+
+    HAT_PRINTF3 (2, "elan3mmu_pteload: modify pte at %lx from %llx to %llx\n", pte, (long long) oldpte, (long long) newpte);
+
+    if (ELAN3_PTE_IS_VALID(ptbl, oldpte))
+    {
+       ELAN3MMU_STAT(ptereload);
+
+       ASSERT ((newpte & ~((E3_uint64)ELAN3_PTE_PERM_MASK | ELAN3_RM_MASK)) == (oldpte & ~((E3_uint64)ELAN3_PTE_PERM_MASK | ELAN3_RM_MASK)));
+       
+       if ((newpte & ~ELAN3_RM_MASK) != (oldpte & ~ELAN3_RM_MASK))
+       {
+           /* We're modifying a valid translation, it must be mapping the same page */
+           /* so we use elan3_modifypte to not affect the referenced and modified bits */
+           elan3_modifypte (dev, pte, newpte);
+
+
+           ElanFlushTlb (elan3mmu->elan3mmu_dev);
+       }
+    }
+    else
+    {
+       ELAN3MMU_STAT(pteload);
+
+       ptbl->ptbl_valid++;
+
+       HAT_PRINTF3 (2, "elan3mmu_pteload: inc valid for level %d ptbl %p to %d\n", 
+                    PTBL_LEVEL(ptbl->ptbl_flags), ptbl, ptbl->ptbl_valid);
+
+       HAT_PRINTF2 (2, "elan3mmu_pteload: write pte %lx to %llx\n", pte, (long long) newpte);
+
+       elan3_writepte (dev, pte, newpte);
+
+       if (ptbl->ptbl_flags & PTBL_KERNEL)
+           ElanFlushTlb (elan3mmu->elan3mmu_dev);
+
+    }
+
+    elan3mmu_unlock_ptbl (ptbl, lock, flags);
+}
+
+void
+elan3mmu_unload (ELAN3MMU *elan3mmu, E3_Addr addr, u_int len, int attr)
+{
+    ELAN3_PTBL          *ptbl;
+    sdramaddr_t         ptp;
+    spinlock_t        *lock;
+    int                        level;
+    u_int              span;
+    unsigned long      flags;
+
+    HAT_PRINTF3(1, "elan3mmu_unload (elan3mmu %p addr %x -> %x)\n", elan3mmu, addr, addr+len-1);
+
+    for (; len != 0; addr += span, len -= span)
+    {
+       ptp  = elan3mmu_ptefind(elan3mmu, addr, &level, &ptbl, &lock, &flags);
+
+       span = MIN(len, ELAN3_L3_PTSIZE - ((E3_Addr) addr & ELAN3_L3_PTOFFSET));
+
+       if (ptp != (sdramaddr_t) 0)
+       {
+           HAT_PRINTF2 (2, "elan3mmu_unload: unload [%x,%x]\n", addr, addr + span);
+           
+           if ( level ==  PTBL_LEVEL_3 ) 
+               elan3mmu_unload_loop (elan3mmu, ptbl, ptp - PTBL_TO_PTADDR(ptbl), span >> ELAN3_PAGE_SHIFT, attr);
+           else
+           {
+               ELAN3_PTP    invalidptp = ELAN3_INVALID_PTP;
+               ELAN3_DEV   *dev = elan3mmu->elan3mmu_dev;
+               ELAN3_PTBL  *lXptbl;
+               ELAN3_PTP    tptp;
+               int         idx;
+
+               tptp = elan3_readptp (elan3mmu->elan3mmu_dev, ptp);
+
+               ASSERT (ELAN3_PTP_TYPE(tptp) == ELAN3_ET_PTE);
+
+               lXptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tptp);
+               idx    = (PTP_TO_PT_PADDR(tptp) - PTBL_TO_PTADDR(lXptbl))/ELAN3_PTE_SIZE;
+
+               if ( level == PTBL_LEVEL_1) 
+                   span = MIN(len, ELAN3_L2_PTSIZE - ((E3_Addr) addr & ELAN3_L2_PTOFFSET));
+               else
+                   span = MIN(len, ELAN3_L3_PTSIZE - ((E3_Addr) addr & ELAN3_L3_PTOFFSET));
+
+               /* invalidate the ptp. */
+               elan3_writeptp (dev, ptp, invalidptp);
+               if (! (attr & PTE_UNLOAD_NOFLUSH))
+                   ElanFlushTlb (dev);     
+    
+               elan3mmu_free_pte ( dev, elan3mmu, lXptbl, idx); 
+
+               ptbl->ptbl_valid--;
+
+               HAT_PRINTF3 (2, "elan3mmu_unload: dec valid for level %d ptbl %p to %d\n", 
+                            PTBL_LEVEL(ptbl->ptbl_flags), ptbl, ptbl->ptbl_valid);     
+
+           }
+           elan3mmu_unlock_ptbl (ptbl, lock, flags);
+       }
+    }
+}
+
+static void
+elan3mmu_unload_loop (ELAN3MMU *elan3mmu, ELAN3_PTBL *ptbl, int first_valid, int nptes, int flags)
+{
+    ELAN3_DEV   *dev = elan3mmu->elan3mmu_dev;
+    sdramaddr_t pte;
+    ELAN3_PTE    tpte;
+    int         last_valid = first_valid + nptes;
+    int                i;
+    
+    HAT_PRINTF3 (1, "elan3mmu_unloadloop: ptbl %p entries [%d->%d]\n", ptbl, first_valid, last_valid);
+
+    ASSERT (PTBL_IS_LOCKED (ptbl->ptbl_flags));
+    ASSERT (PTBL_LEVEL(ptbl->ptbl_flags) == PTBL_LEVEL_3);
+    
+    pte = PTBL_TO_PTADDR(ptbl) + first_valid;
+    
+    for (i = first_valid; i < last_valid; i++, pte += ELAN3_PTE_SIZE)
+    {
+       if (ptbl->ptbl_valid == 0)
+           break;
+
+       tpte = elan3_readpte (dev, pte);
+       if (! ELAN3_PTE_IS_VALID(ptbl, tpte))
+           continue;
+       
+       elan3mmu_pteunload (ptbl, pte, flags, NO_MLIST_LOCK);
+    }
+}
+
+void
+elan3mmu_pteunload (ELAN3_PTBL *ptbl, sdramaddr_t pte, int flags, int got_mlist_lock)
+{
+    ELAN3_DEV   *dev = ptbl->ptbl_elan3mmu->elan3mmu_dev;
+    ELAN3_PTE    tpte;
+
+    ASSERT (PTBL_LEVEL (ptbl->ptbl_flags) == PTBL_LEVEL_3);
+    ASSERT (PTBL_IS_LOCKED (ptbl->ptbl_flags));
+
+    HAT_PRINTF2 (1, "elan3mmu_pteunload: ptbl %p pte %lx\n", ptbl, pte);
+
+    ELAN3MMU_STAT (pteunload);
+
+    elan3_invalidatepte (dev, pte);
+
+    if (! (flags & PTE_UNLOAD_NOFLUSH))
+       ElanFlushTlb (dev);
+    
+    tpte = ELAN3_INVALID_PTE;
+    elan3_writepte (dev, pte, tpte);
+    
+    if (ptbl->ptbl_flags & PTBL_KERNEL)
+    {
+       tpte = elan3mmu_kernel_invalid_pte(ptbl->ptbl_elan3mmu);
+
+       elan3_writepte (dev, pte, tpte);
+    }
+
+    ptbl->ptbl_valid--;
+
+    HAT_PRINTF3 (2, "elan3mmu_pteunload: dec valid for level %d ptbl %p to %d\n", 
+                PTBL_LEVEL(ptbl->ptbl_flags), ptbl, ptbl->ptbl_valid);
+
+}
+
+void
+elan3mmu_ptesync (ELAN3_PTBL *ptbl, sdramaddr_t pte, int flags, int got_mlist_lock)
+{
+
+}
+
+/*
+ * Create more page tables at a given level for this Elan.
+ */
+static ELAN3_PTBL *
+elan3mmu_create_ptbls (ELAN3_DEV *dev, int level, int attr, int keep)
+{
+    sdramaddr_t          pts;
+    ELAN3_PTBL    *ptbl;
+    ELAN3_PTBL    *first;
+    ELAN3_PTBL    *last;
+    ELAN3_PTBL_GR *ptg;
+    register int  i;
+    register int  inc;
+    
+    HAT_PRINTF1 (2, "elan3mmu_create_ptbls: create level %d ptbls\n", level);
+
+    pts = elan3_sdram_alloc (dev, PTBL_GROUP_SIZE);
+    if (pts == (sdramaddr_t) 0)
+    {
+       HAT_PRINTF0 (2, "elan3mmu_create_ptbls: cannot map elan pages\n");
+
+       ELAN3MMU_STAT (create_ptbl_failed);
+       return (NULL);
+    }
+    
+    HAT_PRINTF1 (2, "elan3mmu_create_ptbls: pts at %lx\n", pts);
+    
+    ALLOC_PTBL_GR (ptg, !(attr & PTE_NO_SLEEP));               /* Allocate the group of page tables */
+    if (ptg == NULL)                                           /* for this page */
+    {
+       HAT_PRINTF0 (2, "elan3mmu_create_ptbls: cannot allocate page table group\n");
+
+       elan3_sdram_free (dev, pts, PTBL_GROUP_SIZE);
+
+       ELAN3MMU_STAT (create_ptbl_failed);
+       return (NULL);
+    }
+
+    HAT_PRINTF1 (2, "elan3mmu_create_ptbls: ptg is %p\n", ptg);
+    
+    ElanSetPtblGr (dev, pts, ptg);
+    
+    HAT_PRINTF4 (2, "elan3mmu_create_ptbls: zeroing %d bytes at %lx, %d bytes at %p\n",
+                PTBL_GROUP_SIZE, pts, (int) sizeof (ELAN3_PTBL_GR), ptg);
+
+#ifndef zero_all_ptbls
+    elan3_sdram_zeroq_sdram (dev, pts, PTBL_GROUP_SIZE);               /* Ensure that all PTEs/PTPs are invalid */
+#endif
+    bzero ((caddr_t) ptg, sizeof (ELAN3_PTBL_GR));
+    
+    ptg->pg_addr  = pts;
+    ptg->pg_level = level;
+
+    ptbl = ptg->pg_ptbls;                                      /* Initialise the index in all page tables */
+    for (i = 0; i < PTBLS_PER_GROUP_MAX; i++)
+    {
+       ptbl->ptbl_index = (u_char) i;
+       ptbl->ptbl_next  = (ELAN3_PTBL *) 0xdeaddead;
+       ptbl++;
+    }
+    
+    switch (level)                                             /* Determine the number of ptbls we can  */
+    {                                                          /* allocate from this page, by jumping  */
+    case PTBL_LEVEL_X: inc = PTBLS_PER_PTBL_LX; break;         /* multiples of the smallest. */
+    case PTBL_LEVEL_1: inc = PTBLS_PER_PTBL_L1; break;
+    case PTBL_LEVEL_2: inc = PTBLS_PER_PTBL_L2; break;
+    case PTBL_LEVEL_3: inc = PTBLS_PER_PTBL_L3; break;
+    default:           inc = PTBLS_PER_PTBL_L3; break;
+    }
+
+    ptbl = ptg->pg_ptbls;                                      /* Chain them together */
+    for (i = 0; i < PTBLS_PER_GROUP_MAX; i += inc, ptbl += inc)
+       ptbl->ptbl_next = ptbl + inc;
+
+    first = ptg->pg_ptbls;                                     /* Determine list of */
+    last  = first + PTBLS_PER_GROUP_MAX - inc;                 /* ptbls to add to free list */
+    if (! keep)
+       ptbl = NULL;
+    else
+    {
+       ptbl  = first;
+       first = first->ptbl_next;
+    }
+    
+    spin_lock (&dev->Level[level].PtblLock);
+    dev->Level[level].PtblTotal     += PTBLS_PER_GROUP_MAX/inc;                /* Increment the counts */
+    dev->Level[level].PtblFreeCount += PTBLS_PER_GROUP_MAX/inc;
+
+    ELAN3MMU_SET_STAT (num_ptbl_level[level], dev->Level[level].PtblTotal);
+
+    if (keep)
+       dev->Level[level].PtblFreeCount--;
+    
+    last->ptbl_next = dev->Level[level].PtblFreeList;                  /* And add to free list */
+    dev->Level[level].PtblFreeList = first;
+    spin_unlock (&dev->Level[level].PtblLock);
+    
+    spin_lock (&dev->PtblGroupLock);
+    ptg->pg_next = dev->Level[level].PtblGroupList;
+    dev->Level[level].PtblGroupList = ptg;
+    spin_unlock (&dev->PtblGroupLock);
+
+    HAT_PRINTF1 (2, "elan3mmu_create_ptbls: returning ptbl %p\n", ptbl);
+    
+    return (ptbl);
+}
+
+static ELAN3_PTBL *
+elan3mmu_ta_to_ptbl (ELAN3MMU *elan3mmu, ELAN3_PTP *ptp)
+{
+    E3_Addr      ptpa  = PTP_TO_PT_PADDR(*ptp);
+    ELAN3_PTBL_GR *pg    = ElanGetPtblGr (elan3mmu->elan3mmu_dev, (sdramaddr_t)ptpa & ~(PTBL_GROUP_SIZE-1));
+    
+    return (pg->pg_ptbls + ((ptpa - pg->pg_addr) >> ELAN3_PT_SHIFT));
+}
+
+static ELAN3_PTBL *
+elan3mmu_alloc_lXptbl (ELAN3_DEV *dev, int attr,  ELAN3MMU *elan3mmu)
+{
+    ELAN3_PTBL *ptbl = NULL;
+
+    spin_lock (&dev->Level[PTBL_LEVEL_X].PtblLock);
+    if (dev->Level[PTBL_LEVEL_X].PtblFreeList)
+    {
+       ptbl = dev->Level[PTBL_LEVEL_X].PtblFreeList;
+
+       HAT_PRINTF1 (2, "elan3mmu_alloc_lXptbl: found ptbl %p on free list\n", ptbl);
+
+       dev->Level[PTBL_LEVEL_X].PtblFreeList = ptbl->ptbl_next;
+       dev->Level[PTBL_LEVEL_X].PtblFreeCount--;
+    }
+    spin_unlock (&dev->Level[PTBL_LEVEL_X].PtblLock);
+    
+    if (ptbl == NULL) 
+    {
+       ptbl = elan3mmu_create_ptbls (dev, PTBL_LEVEL_X, attr, 1);
+
+       HAT_PRINTF1 (2, "elan3mmu_alloc_lXptbl: created level X ptbl %p\n", ptbl);
+    }
+
+    if (ptbl == NULL)
+    {
+       if ((attr & PTE_NO_STEAL))
+       {
+           HAT_PRINTF0 (2, "elan3mmu_alloc_lXptbl: not allowed to steal ptbl for use at level 2\n");
+           return NULL;
+       }
+
+       ELAN3MMU_STAT(lX_alloc_l3);
+
+       ptbl = elan3mmu_steal_l3ptbl (dev, attr);
+       
+       HAT_PRINTF1 (2, "elan3mmu_alloc_lXptbl: stolen level3 ptbl %p used as level 2\n", ptbl);
+    }
+
+    ptbl->ptbl_elan3mmu = elan3mmu;
+    ptbl->ptbl_base     = 0;
+    ptbl->ptbl_parent   = 0;
+    ptbl->ptbl_flags    = PTBL_LEVEL_X | PTBL_ALLOCED;
+    
+    HAT_PRINTF2 (2, "elan3mmu_alloc_lXptbl: ptbl %p dev %p\n", ptbl, dev);
+
+#ifdef zero_all_ptbls
+    elan3_sdram_zero_sdarm (dev, PTBL_TO_PTADDR(ptbl), ELAN3_LX_ENTRIES*ELAN3_PTE_SIZE);
+#endif
+
+    return (ptbl);
+}
+
+static ELAN3_PTBL *
+elan3mmu_alloc_pte (ELAN3_DEV *dev, ELAN3MMU *elan3mmu, int *idx)
+{
+    ELAN3_PTBL   * ptbl_ptr;
+    int           index;
+
+    /* lock whilst looking for space */
+    spin_lock (&elan3mmu->elan3mmu_lXptbl_lock);
+    
+    /* walk the lXptbl list */
+    ptbl_ptr = elan3mmu->elan3mmu_lXptbl;
+    while ( ptbl_ptr != NULL ) 
+    {
+       /* does this ptlb have any free ones */
+       if (  (index = ptbl_ptr->ptbl_valid) < ELAN3_LX_ENTRIES) 
+       {
+           /*  better to search  from valid count as its likly to be free */
+           index = ptbl_ptr->ptbl_valid; 
+           do {
+               if ((ptbl_ptr->ptbl_base & (1 << index)) == 0)
+                   goto found;
+
+               /* move index on and wrap back to start if needed */
+               if ((++index) == ELAN3_LX_ENTRIES) 
+                   index = 0;
+           } while (index != ptbl_ptr->ptbl_valid);
+
+           panic ("elan3mmu_alloc_pte: has ptbl valid < 32 when but no free pte's");
+       }
+       ptbl_ptr = ptbl_ptr->ptbl_parent;
+    }
+       
+    /* unlock so we can create space */
+    spin_unlock (&elan3mmu->elan3mmu_lXptbl_lock); 
+
+    /* if create some more */
+    ptbl_ptr = elan3mmu_alloc_lXptbl(dev, 0, elan3mmu);
+
+    /* get the lock again */
+    spin_lock (&elan3mmu->elan3mmu_lXptbl_lock);
+       
+    /* add to front of list as its obviously got free ones on it */
+    ptbl_ptr->ptbl_parent     = elan3mmu->elan3mmu_lXptbl;
+    elan3mmu->elan3mmu_lXptbl = ptbl_ptr;
+
+    /* grap the first one */
+    index = 0;
+    
+ found:
+    ptbl_ptr->ptbl_base |= (1 << index);
+    ptbl_ptr->ptbl_valid++;
+
+    HAT_PRINTF3 (2, "elan3mmu_alloc_pte: inc valid for level %d ptbl %p to %d\n", 
+                PTBL_LEVEL(ptbl_ptr->ptbl_flags), ptbl_ptr, ptbl_ptr->ptbl_valid);
+
+    /* release the loc and return it */
+    spin_unlock (&elan3mmu->elan3mmu_lXptbl_lock); 
+
+    *idx = index;
+    return (ptbl_ptr);
+}
+
+static ELAN3_PTBL *
+elan3mmu_alloc_l1ptbl (ELAN3_DEV *dev, int attr, ELAN3MMU *elan3mmu)
+{
+    ELAN3_PTBL *ptbl = NULL;
+    ELAN3_PTBL *p;
+    int i,j;
+    
+    spin_lock (&dev->Level[PTBL_LEVEL_1].PtblLock);
+    if (dev->Level[PTBL_LEVEL_1].PtblFreeList)
+    {
+       ptbl = dev->Level[PTBL_LEVEL_1].PtblFreeList;
+       dev->Level[PTBL_LEVEL_1].PtblFreeList = ptbl->ptbl_next;
+       dev->Level[PTBL_LEVEL_1].PtblFreeCount--;
+    }
+    spin_unlock (&dev->Level[PTBL_LEVEL_1].PtblLock);
+    
+    if (ptbl == NULL)
+       ptbl = elan3mmu_create_ptbls (dev, PTBL_LEVEL_1, attr, 1);
+    
+    if (ptbl == NULL)
+       panic ("elan3mmu_alloc_l1ptbl: cannot alloc ptbl");
+    
+    for (p = ptbl, j = i = 0; i < PTBLS_PER_PTBL_L1; i++, p++)
+    {
+       p->ptbl_elan3mmu = elan3mmu;
+       p->ptbl_base     = VA2BASE (j);
+       p->ptbl_flags    = PTBL_LEVEL_1 | PTBL_GROUPED;
+       p->ptbl_parent   = NULL;
+       
+       j += L1_VA_PER_PTBL;
+    }
+    
+    /* Now mark the real page table as allocated */
+    /* level 1 ptbls are returned unlocked */
+    ptbl->ptbl_flags = PTBL_LEVEL_1 | PTBL_ALLOCED;
+    
+    HAT_PRINTF2 (2, "elan3mmu_alloc_l1ptbl: ptbl %p dev %p\n", ptbl, dev);
+
+#ifdef zero_all_ptbls
+    elan3_sdram_zeroq_sdram (dev, PTBL_TO_PTADDR(ptbl), ELAN3_L1_ENTRIES*ELAN3_PTP_SIZE);
+#endif
+
+    return (ptbl);
+}
+
+static ELAN3_PTBL *
+elan3mmu_alloc_l2ptbl (ELAN3_DEV *dev, int attr, ELAN3_PTBL *parent, ELAN3MMU *elan3mmu, E3_Addr base, spinlock_t **plock, unsigned long *flags)
+{
+    ELAN3_PTBL *ptbl = NULL;
+    ELAN3_PTBL *p;
+    int        i;
+    int        j;
+    unsigned long ptbl_flags;
+
+    spin_lock_irqsave (&dev->Level[PTBL_LEVEL_2].PtblLock, ptbl_flags);
+    if (dev->Level[PTBL_LEVEL_2].PtblFreeList)
+    {
+       ptbl = dev->Level[PTBL_LEVEL_2].PtblFreeList;
+
+       HAT_PRINTF1 (2, "elan3mmu_alloc_l2ptbl: found ptbl %p on free list\n", ptbl);
+
+       dev->Level[PTBL_LEVEL_2].PtblFreeList = ptbl->ptbl_next;
+       dev->Level[PTBL_LEVEL_2].PtblFreeCount--;
+    }
+    spin_unlock_irqrestore (&dev->Level[PTBL_LEVEL_2].PtblLock, ptbl_flags);
+    
+    if (ptbl == NULL) 
+    {
+       ptbl = elan3mmu_create_ptbls (dev, PTBL_LEVEL_2, attr, 1);
+
+       HAT_PRINTF1 (2, "elan3mmu_alloc_l2ptbl: created level 2 ptbl %p\n", ptbl);
+    }
+
+    if (ptbl == NULL)
+    {
+       if ((attr & PTE_NO_STEAL))
+       {
+           HAT_PRINTF0 (2, "elan3mmu_alloc_l2ptbl: not allowted to steal ptbl for use at level 2\n");
+           return (NULL);
+       }
+
+       ELAN3MMU_STAT(l2_alloc_l3);
+
+       ptbl = elan3mmu_steal_l3ptbl (dev, attr);
+       
+       HAT_PRINTF1 (2, "elan3mmu_alloc_l2ptbl: stolen level3 ptbl %p used as level 2\n", ptbl);
+    }
+    
+    *plock = elan3mmu_ptbl_to_lock (PTBL_LEVEL_2, ptbl);
+    spin_lock_irqsave (*plock, *flags);
+    
+    for (p = ptbl, j = i = 0; i < PTBLS_PER_PTBL_L2; i++, p++)
+    {
+       p->ptbl_elan3mmu = elan3mmu;
+       p->ptbl_base     = VA2BASE (base + j);
+       p->ptbl_flags    = PTBL_LEVEL_2 | PTBL_GROUPED;
+       p->ptbl_parent   = parent;
+       
+       j += L2_VA_PER_PTBL;
+    }
+    
+    ptbl->ptbl_flags  = PTBL_LEVEL_2 | PTBL_ALLOCED | PTBL_LOCKED;
+    
+    HAT_PRINTF3 (2, "elan3mmu_alloc_l2ptbl: ptbl %p dev %p base %x\n", ptbl, dev, base);
+
+#ifdef zero_all_ptbls
+    elan3_sdram_zero_sdarm (dev, PTBL_TO_PTADDR(ptbl), ELAN3_L2_ENTRIES*ELAN3_PTP_SIZE);
+#endif
+
+    return (ptbl);
+}
+
+static ELAN3_PTBL *
+elan3mmu_alloc_l3ptbl (ELAN3_DEV *dev, int attr, ELAN3_PTBL *parent, ELAN3MMU *elan3mmu, E3_Addr base, spinlock_t **plock, unsigned long *flags)
+{
+    ELAN3_PTBL *ptbl = NULL;
+    ELAN3_PTBL *p;
+    int               i;
+    int               j;
+    unsigned long ptbl_flags;
+
+    spin_lock_irqsave (&dev->Level[PTBL_LEVEL_3].PtblLock, ptbl_flags);
+    if (dev->Level[PTBL_LEVEL_3].PtblFreeList)
+    {
+       HAT_PRINTF1 (2, "elan3mmu_alloc_l3ptbl: found ptbl %p on free list\n", ptbl);
+
+       ptbl = dev->Level[PTBL_LEVEL_3].PtblFreeList;
+       dev->Level[PTBL_LEVEL_3].PtblFreeList = ptbl->ptbl_next;
+       dev->Level[PTBL_LEVEL_3].PtblFreeCount--;
+    }
+    spin_unlock_irqrestore (&dev->Level[PTBL_LEVEL_3].PtblLock, ptbl_flags);
+    
+    if (ptbl == NULL)
+    {
+       ptbl = elan3mmu_create_ptbls (dev, PTBL_LEVEL_3, attr, 1);
+
+       HAT_PRINTF1 (2, "elan3mmu_alloc_l3ptbl: created level 3 ptbl %p\n", ptbl);
+    }
+
+    if (ptbl == NULL)
+    {
+       if ((attr & PTE_NO_STEAL))
+       {
+           HAT_PRINTF0 (2, "elan3mmu_alloc_l3ptbl: not allowed to steal ptbl for use at level 3\n");
+           return (NULL);
+       }
+
+       ptbl = elan3mmu_steal_l3ptbl (dev, attr);
+
+       HAT_PRINTF1 (2, "elan3mmu_alloc_l3ptbl: stolen level3 ptbl %p\n", ptbl);
+    }
+    
+    *plock = elan3mmu_ptbl_to_lock (PTBL_LEVEL_3, ptbl);
+    spin_lock_irqsave (*plock,*flags);
+    
+    for (p = ptbl, j = i = 0; i < PTBLS_PER_PTBL_L3; i++, p++)
+    {
+       p->ptbl_elan3mmu = elan3mmu;
+       p->ptbl_base     = VA2BASE (base + j);
+       p->ptbl_flags    = PTBL_LEVEL_3 | PTBL_GROUPED;
+       p->ptbl_parent   = parent;
+       
+       j += L3_VA_PER_PTBL;
+    }
+    
+    ptbl->ptbl_flags = PTBL_LEVEL_3 | PTBL_ALLOCED | PTBL_LOCKED;
+    
+    HAT_PRINTF3 (2, "elan3mmu_alloc_l3ptbl: ptbl %p dev %p base %x\n", ptbl, dev, base);
+
+#ifdef zero_all_ptbls
+    elan3_sdram_zeroq_sdram (dev, PTBL_TO_PTADDR(ptbl), ELAN3_L3_ENTRIES*ELAN3_PTE_SIZE);
+#endif
+
+    return (ptbl);
+}
+
+void 
+elan3mmu_free_pte  (ELAN3_DEV *dev,  ELAN3MMU *elan3mmu,  ELAN3_PTBL *ptbl_ptr, int idx)
+{  
+    sdramaddr_t pte  = PTBL_TO_PTADDR (ptbl_ptr) | (idx * sizeof (ELAN3_PTE));
+    ELAN3_PTE    tpte = ELAN3_INVALID_PTE;
+    ELAN3_PTBL *prev;
+
+    /* ensure that the pte is invalid when free */
+    elan3_writepte (dev, pte, tpte);
+
+    /* lock whilst removing */
+    spin_lock (&elan3mmu->elan3mmu_lXptbl_lock);
+
+    HAT_PRINTF4 (2, "elan3mmu_free_pte idx %d   ptbl_ptr %p ptbl_base  %x  ptbl_ptr->ptbl_valid %d \n", 
+                idx, ptbl_ptr, ptbl_ptr->ptbl_base, ptbl_ptr->ptbl_valid);
+    /* make sure it was set */
+    ASSERT ( ptbl_ptr->ptbl_base & (1 << idx) ); 
+    ASSERT ( ptbl_ptr->ptbl_valid > 0  );
+
+    ptbl_ptr->ptbl_base &= ~(1 << idx);
+    ptbl_ptr->ptbl_valid--;
+
+    HAT_PRINTF3 (2, "elan3mmu_free_pte: dec valid for level %d ptbl %p to %d\n", 
+                PTBL_LEVEL(ptbl_ptr->ptbl_flags), ptbl_ptr, ptbl_ptr->ptbl_valid); 
+ 
+    /* was that the last one on this page */
+    if ( ! ptbl_ptr->ptbl_valid ) 
+    {
+       /* so no bits should be set then */
+       ASSERT ( ptbl_ptr->ptbl_base == 0 );
+
+       /* is this the first page ?? */
+       if ( elan3mmu->elan3mmu_lXptbl == ptbl_ptr ) 
+       {
+           /* make the list start at the second element */
+            elan3mmu->elan3mmu_lXptbl = ptbl_ptr->ptbl_parent;
+
+            /* put ptbl back on free list */
+            elan3mmu_free_lXptbl(dev, ptbl_ptr);
+
+            /* unlock and return */
+            spin_unlock (&elan3mmu->elan3mmu_lXptbl_lock);
+            return ;
+       }
+
+       /* scan thro list looking for this page */
+       prev = elan3mmu->elan3mmu_lXptbl;
+       while ( prev->ptbl_parent != NULL ) 
+       {
+           if ( prev->ptbl_parent == ptbl_ptr ) /* its the next one */
+           {
+               /* remove element from chain */
+               prev->ptbl_parent =  ptbl_ptr->ptbl_parent;
+
+               /* put ptbl back on free list */
+               elan3mmu_free_lXptbl(dev, ptbl_ptr);
+
+               /* unlock and return */
+               spin_unlock (&elan3mmu->elan3mmu_lXptbl_lock);
+               return ;
+           }           
+           prev = prev->ptbl_parent;
+       }
+       
+               panic ("elan3mmu_free_pte: failed to find ptbl in chain");
+       /* NOTREACHED */
+    }
+    
+    spin_unlock (&elan3mmu->elan3mmu_lXptbl_lock);
+}
+
+void
+elan3mmu_free_lXptbl (ELAN3_DEV *dev, ELAN3_PTBL *ptbl)
+{
+    ELAN3_PTBL_GR *ptg;
+
+    HAT_PRINTF2 (2, "elan3mmu_free_lXptbl: dev %p ptbl %p\n", dev, ptbl);
+
+    ASSERT (ptbl->ptbl_flags & PTBL_ALLOCED);
+    ASSERT ((ptbl->ptbl_flags & PTBL_KEEP) == 0);
+    ASSERT (PTBL_LEVEL(ptbl->ptbl_flags) == PTBL_LEVEL_X);
+    ASSERT (ptbl->ptbl_valid == 0);
+   
+    ptbl->ptbl_flags = 0;
+
+    ptg = PTBL_TO_GR(ptbl);
+
+    if (ptg->pg_level == PTBL_LEVEL_3)
+    {
+       ELAN3MMU_STAT(lX_freed_l3);
+
+       HAT_PRINTF1 (2, "elan3mmu_free_lXptbl: freeing stolen level 3 ptbl %p\n", ptbl);
+
+       /* this was really a level 3 ptbl which we had to steal */
+       spin_lock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+       ptbl->ptbl_next = dev->Level[PTBL_LEVEL_3].PtblFreeList;
+       dev->Level[PTBL_LEVEL_3].PtblFreeList = ptbl;
+       dev->Level[PTBL_LEVEL_3].PtblFreeCount++;
+       spin_unlock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+    }
+    else
+    {
+       spin_lock (&dev->Level[PTBL_LEVEL_X].PtblLock);
+       ptbl->ptbl_next = dev->Level[PTBL_LEVEL_X].PtblFreeList;
+       dev->Level[PTBL_LEVEL_X].PtblFreeList = ptbl;
+       dev->Level[PTBL_LEVEL_X].PtblFreeCount++;
+       spin_unlock (&dev->Level[PTBL_LEVEL_X].PtblLock);
+    }
+}
+
+void
+elan3mmu_free_l1ptbl (ELAN3_DEV *dev, ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags)
+{
+    HAT_PRINTF3 (2, "elan3mmu_free_l1ptbl: dev %p ptbl %p ptbl->ptbl_valid %x \n", dev, ptbl, ptbl->ptbl_valid);
+
+    ASSERT (ptbl->ptbl_flags & PTBL_ALLOCED);
+    ASSERT ((ptbl->ptbl_flags & PTBL_KEEP) == 0);
+    ASSERT (PTBL_LEVEL(ptbl->ptbl_flags) == PTBL_LEVEL_1);
+    ASSERT (ptbl->ptbl_valid == 0);
+    
+    HAT_PRINTF2 (2, "elan3mmu_free_l1ptbl: dev %p ptbl %p\n", dev, ptbl);
+
+    ptbl->ptbl_flags = 0;
+    spin_unlock (lock);
+    
+    spin_lock (&dev->Level[PTBL_LEVEL_1].PtblLock);
+    ptbl->ptbl_next = dev->Level[PTBL_LEVEL_1].PtblFreeList;
+    dev->Level[PTBL_LEVEL_1].PtblFreeList = ptbl;
+    dev->Level[PTBL_LEVEL_1].PtblFreeCount++;
+    spin_unlock (&dev->Level[PTBL_LEVEL_1].PtblLock);
+
+    local_irq_restore (flags);
+}
+
+void
+elan3mmu_free_l2ptbl (ELAN3_DEV *dev, ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags)
+{
+    ELAN3_PTBL_GR *ptg;
+
+    HAT_PRINTF2 (2, "elan3mmu_free_l2ptbl: dev %p ptbl %p\n", dev, ptbl);
+
+    ASSERT (PTBL_IS_LOCKED(ptbl->ptbl_flags));
+    ASSERT (ptbl->ptbl_flags & PTBL_ALLOCED);
+    ASSERT ((ptbl->ptbl_flags & PTBL_KEEP) == 0);
+    ASSERT (PTBL_LEVEL(ptbl->ptbl_flags) == PTBL_LEVEL_2);
+    ASSERT (ptbl->ptbl_valid == 0);
+   
+    ptbl->ptbl_flags = 0;
+    spin_unlock (lock);
+
+    ptg = PTBL_TO_GR(ptbl);
+
+    if (ptg->pg_level == PTBL_LEVEL_3)
+    {
+       ELAN3MMU_STAT(l2_freed_l3);
+
+       HAT_PRINTF1 (2, "elan3mmu_free_l2ptbl: freeing stolen level 3 ptbl %p\n", ptbl);
+
+       /* this was really a level 3 ptbl which we had to steal */
+       spin_lock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+       ptbl->ptbl_next = dev->Level[PTBL_LEVEL_3].PtblFreeList;
+       dev->Level[PTBL_LEVEL_3].PtblFreeList = ptbl;
+       dev->Level[PTBL_LEVEL_3].PtblFreeCount++;
+       spin_unlock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+    }
+    else
+    {
+       spin_lock (&dev->Level[PTBL_LEVEL_2].PtblLock);
+       ptbl->ptbl_next = dev->Level[PTBL_LEVEL_2].PtblFreeList;
+       dev->Level[PTBL_LEVEL_2].PtblFreeList = ptbl;
+       dev->Level[PTBL_LEVEL_2].PtblFreeCount++;
+       spin_unlock (&dev->Level[PTBL_LEVEL_2].PtblLock);
+    }  
+    local_irq_restore (flags);
+}
+
+void
+elan3mmu_free_l3ptbl (ELAN3_DEV *dev, ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags)
+{
+    ASSERT (PTBL_IS_LOCKED(ptbl->ptbl_flags));
+    ASSERT (ptbl->ptbl_flags & PTBL_ALLOCED);
+    ASSERT ((ptbl->ptbl_flags & PTBL_KEEP) == 0);
+    ASSERT (PTBL_LEVEL(ptbl->ptbl_flags) == PTBL_LEVEL_3);
+    ASSERT (ptbl->ptbl_valid == 0);
+    
+    HAT_PRINTF2 (2, "elan3mmu_free_l3ptbl: dev %p ptbl %p\n", dev, ptbl);
+
+    if (ptbl->ptbl_flags & PTBL_KERNEL)                                /* if the ptbl has been used by the kernel */
+    {                                                          /* then zero all the pte's, since they will */
+       elan3_sdram_zeroq_sdram (dev, PTBL_TO_PTADDR(ptbl), ELAN3_L3_ENTRIES*ELAN3_PTE_SIZE);
+    }
+
+    ptbl->ptbl_flags = 0;
+    spin_unlock (lock);
+    
+    spin_lock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+    ptbl->ptbl_next = dev->Level[PTBL_LEVEL_3].PtblFreeList;
+    dev->Level[PTBL_LEVEL_3].PtblFreeList = ptbl;
+    dev->Level[PTBL_LEVEL_3].PtblFreeCount++;
+    spin_unlock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+
+    local_irq_restore (flags);
+}
+
+void
+elan3mmu_kernel_l3ptbl (ELAN3_PTBL *ptbl)
+{
+    ELAN3_DEV   *dev  = ptbl->ptbl_elan3mmu->elan3mmu_dev;
+    sdramaddr_t pte  = PTBL_TO_PTADDR(ptbl);
+    ELAN3_PTE    tpte = elan3mmu_kernel_invalid_pte(ptbl->ptbl_elan3mmu);
+    int                i;
+
+    ptbl->ptbl_flags |= PTBL_KERNEL;
+    for (i = 0; i < ELAN3_L3_ENTRIES; i++, pte += ELAN3_PTE_SIZE)
+    {
+       elan3_writepte (dev, pte, tpte);
+    }
+}
+       
+#define PTBL_CAN_STEAL(flag)   (((flag) & (PTBL_KERNEL|PTBL_KEEP)) == 0 && (((flag) & PTBL_ALLOCED) && PTBL_LEVEL(flag) == PTBL_LEVEL_3))
+#define PTBL_MAY_STEAL(flag)   (((flag) & (PTBL_KERNEL|PTBL_KEEP|PTBL_LOCKED)) == 0 && (((flag) & PTBL_ALLOCED) && PTBL_LEVEL(flag) == PTBL_LEVEL_3))
+
+static int
+elan3mmu_steal_this_ptbl (ELAN3_DEV *dev, ELAN3_PTBL *l3ptbl)
+{
+    ELAN3_PTBL  *l2ptbl     = l3ptbl->ptbl_parent;
+    E3_Addr     l2addr     = BASE2VA(l2ptbl);
+    E3_Addr     l3addr     = BASE2VA(l3ptbl);
+    ELAN3_PTP    invalidptp = ELAN3_INVALID_PTP;
+    sdramaddr_t l2ptp;
+    spinlock_t *l2lock;
+    unsigned long l2flags;
+
+    HAT_PRINTF5 (1, "elan3mmu_steal_this_ptbl: l3ptbl %p (%x) l2ptbl %p (%x) l2addr %x\n",
+                l3ptbl, l3ptbl->ptbl_flags, l2ptbl, l2ptbl->ptbl_flags, l2addr);
+
+    if (PTBL_CAN_STEAL (l3ptbl->ptbl_flags) &&
+       elan3mmu_lock_ptbl (l2ptbl, LK_PTBL_NOWAIT, l3ptbl->ptbl_elan3mmu, l2addr, PTBL_LEVEL_2, &l2lock, &l2flags) == LK_PTBL_OK)
+    {
+       ELAN3MMU_STAT(stolen_ptbls);
+
+       /* Locked both L3 and L2 page tables. */
+       l2ptp = PTBL_TO_PTADDR (l2ptbl) + ELAN3_L2_INDEX(l3addr)*ELAN3_PTP_SIZE;
+       
+       /* detach the level 3 page table */
+       elan3_writeptp (dev, l2ptp, invalidptp);
+       ElanFlushTlb (dev);
+
+       l2ptbl->ptbl_valid--;
+
+       HAT_PRINTF3 (2, "elan3mmu_steal_this_ptbl: dec valid for level %d ptbl %p to %d\n", PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid); 
+
+       elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+
+       elan3mmu_unload_loop (l3ptbl->ptbl_elan3mmu, l3ptbl, 0, ELAN3_L3_ENTRIES, PTE_UNLOAD_NOFLUSH);
+
+       ASSERT (l3ptbl->ptbl_valid == 0);
+
+       l3ptbl->ptbl_flags = 0;
+       return (1);
+    }
+    return (0);
+}
+
+static ELAN3_PTBL *
+elan3mmu_steal_l3ptbl (ELAN3_DEV *dev, int attr)
+{
+    ELAN3_PTBL_GR      *ptg;
+    ELAN3_PTBL         *ptbl;
+    spinlock_t         *lock;
+    unsigned long        group_flags;
+    unsigned long        ptbl_flags;
+    register int        i;
+
+    HAT_PRINTF1 (2, "elan3mmu_steal_l3ptbl: attr %x\n", attr);
+
+    spin_lock_irqsave (&dev->PtblGroupLock, group_flags);
+
+    ptg = dev->Level3PtblGroupHand;
+
+    if (ptg == NULL)
+       ptg = dev->Level[PTBL_LEVEL_3].PtblGroupList;
+    
+    for (;;)
+    {
+       while (ptg)
+       {
+           for (i = 0, ptbl = ptg->pg_ptbls; i < PTBLS_PER_GROUP_MAX; i++, ptbl++)
+           {
+               if (PTBL_MAY_STEAL (ptbl->ptbl_flags) &&
+                   elan3mmu_lock_this_ptbl (ptbl, LK_PTBL_NOWAIT, &lock, &ptbl_flags) == LK_PTBL_OK)
+               {
+                   if (elan3mmu_steal_this_ptbl (dev, ptbl ))
+                   {
+                       HAT_PRINTF1 (2, "elan3mmu_steal_l3ptbl: stolen ptbl %p\n", ptbl);
+
+                       elan3mmu_unlock_ptbl (ptbl, lock,ptbl_flags);
+
+                       dev->Level3PtblGroupHand = ptg->pg_next;
+
+                       spin_unlock_irqrestore (&dev->PtblGroupLock, group_flags);
+
+                       return (ptbl);
+                   }
+                   elan3mmu_unlock_ptbl (ptbl, lock, ptbl_flags);
+               }
+           }
+           ptg = ptg->pg_next;
+       }
+       
+       if (dev->Level[PTBL_LEVEL_3].PtblFreeList)
+       {
+           spin_lock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+           ptbl = dev->Level[PTBL_LEVEL_3].PtblFreeList;
+           if (ptbl != NULL)
+           {
+               dev->Level[PTBL_LEVEL_3].PtblFreeList = ptbl->ptbl_next;
+               dev->Level[PTBL_LEVEL_3].PtblFreeCount--;
+           }
+           spin_unlock (&dev->Level[PTBL_LEVEL_3].PtblLock);
+
+           if (ptbl != NULL)
+           {
+               HAT_PRINTF1 (2, "elan3mmu_steal_l3ptbl: found ptbl %p on free list\n", ptbl);
+               break;
+           }
+       }
+
+       ptbl = elan3mmu_create_ptbls (dev, PTBL_LEVEL_3, attr, 1);
+
+       if (ptbl != NULL)
+       {
+           HAT_PRINTF1 (2, "elan3mmu_steal_l3ptbl: created new ptbl %p\n", ptbl);
+           break;
+       }
+       
+       HAT_PRINTF0 (1, "elan3mmu_steal_l3ptbl: cannot find a ptbl, retrying\n");
+       ptg = dev->Level[PTBL_LEVEL_3].PtblGroupList;
+    }
+
+    spin_unlock (&dev->PtblGroupLock);
+    return (ptbl);
+}
+
+sdramaddr_t
+elan3mmu_ptefind (ELAN3MMU *elan3mmu, E3_Addr addr, int *level, 
+                 ELAN3_PTBL **pptbl, spinlock_t **plock, unsigned long *flags)
+{
+    ELAN3_DEV   *dev = elan3mmu->elan3mmu_dev;
+    ELAN3_PTBL  *l1ptbl;
+    sdramaddr_t l1ptp;
+    ELAN3_PTP    tl1ptp;
+    E3_Addr     l1base;
+    ELAN3_PTBL  *l2ptbl;
+    sdramaddr_t l2ptp;
+    ELAN3_PTP    tl2ptp;
+    E3_Addr     l2base;
+    ELAN3_PTBL  *l3ptbl;
+    sdramaddr_t l3pte;
+    spinlock_t *l1lock;
+    spinlock_t *l2lock;
+    spinlock_t *l3lock;
+    unsigned long l1flags;
+    unsigned long l2flags;
+    unsigned long l3flags;
+
+    HAT_PRINTF2 (2, "elan3mmu_ptefind: elan3mmu %p addr %x\n", elan3mmu, addr);
+
+    l1ptbl = elan3mmu->elan3mmu_l1ptbl;
+    *level = 0;
+
+    if (l1ptbl == NULL)
+       return ((sdramaddr_t) NULL);
+
+    l1ptp  = PTBL_TO_PTADDR(l1ptbl) + ELAN3_L1_INDEX(addr)*ELAN3_PTP_SIZE;
+    l1base = ELAN3_L1_BASE(addr);
+    
+retryl1:
+    tl1ptp = elan3_readptp (dev, l1ptp);
+    
+    HAT_PRINTF4 (2, "elan3mmu_ptefind: l1ptbl %p l1ptp %lx l1base %x : tl1ptp %x\n", l1ptbl, l1ptp, l1base, tl1ptp);
+
+    switch (ELAN3_PTP_TYPE(tl1ptp))
+    {
+    case ELAN3_ET_PTE:
+       elan3mmu_lock_ptbl (l1ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_1, &l1lock, &l1flags);
+
+       tl1ptp = elan3_readptp (dev, l1ptp);
+       if (ELAN3_PTP_TYPE(tl1ptp) != ELAN3_ET_PTE)
+       {
+           elan3mmu_unlock_ptbl (l1ptbl, l1lock, l1flags);
+           goto retryl1;
+       }
+       
+       *level = 1;
+       *pptbl = l1ptbl;
+       *plock = l1lock;
+       *flags = l1flags;
+       
+       /* return with l1lock */
+       return (l1ptp);  
+
+    case ELAN3_ET_INVALID:
+       return ((sdramaddr_t) 0);
+       
+    case ELAN3_ET_PTP:
+       break;
+
+    default:
+       panic ("elan3mmu_ptefind: found bad entry in level 1 page table");
+       /* NOTREACHED */
+    }
+    
+    HAT_PRINTF1 (2, "elan3mmu_ptefind: chain to level 2 ptbl from ptp %x\n", tl1ptp);
+
+    l2ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl1ptp);
+    l2ptp  = PTBL_TO_PTADDR(l2ptbl) + ELAN3_L2_INDEX(addr)*ELAN3_PTP_SIZE;
+    l2base = ELAN3_L2_BASE(addr);
+    
+    tl2ptp = elan3_readptp (dev, l2ptp);
+    
+    HAT_PRINTF4 (2, "elan3mmu_ptefind: l2ptbl %p l2ptp %lx l2base %x : tl2ptp %x\n", l2ptbl, l2ptp, l2base, tl2ptp);
+
+    switch (ELAN3_PTP_TYPE(tl2ptp))
+    {
+    case ELAN3_ET_PTE:
+       switch (elan3mmu_lock_ptbl (l2ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_2, &l2lock, &l2flags))
+       {
+       case LK_PTBL_OK:
+           tl2ptp = elan3_readptp (dev, l2ptp);
+           if (ELAN3_PTP_TYPE(tl2ptp) != ELAN3_ET_PTE)
+           {
+               elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+               goto retryl1;
+           }
+           
+           *level = 2;
+           *pptbl = l2ptbl;
+           *plock = l2lock;
+           *flags = l2flags;
+           
+           /* return with l2lock */
+           return (l2ptp); 
+           
+       case LK_PTBL_MISMATCH:
+           HAT_PRINTF6 (2, "elan3mmu_ptefind: PTBL_MISMATCH : ptbl %p flags %x elan3mmu %p base %x (%p %x)\n",
+                        l2ptbl, l2ptbl->ptbl_flags, l2ptbl->ptbl_elan3mmu, l2ptbl->ptbl_base, elan3mmu, addr);
+           
+           /*
+            * We've trogged down to this ptbl,  but someone has just
+            * stolen it,  so try all over again.
+            */
+           goto retryl1;
+           
+       default:
+           panic ("elan3mmu_ptefind: elan3mmu_lock_ptbl returned bad value");
+           /* NOTREACHED */
+       }
+    case ELAN3_ET_INVALID:
+       return ((sdramaddr_t) 0);
+       
+    case ELAN3_ET_PTP:
+       break;
+    default:
+       panic ("elan3mmu_ptefind: found bad entry in level 2 page table");
+       /* NOTREACHED */
+    }
+    
+    HAT_PRINTF1 (2, "elan3mmu_ptefind: chain to level 3 page table from ptp %x\n", tl2ptp);
+
+    l3ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl2ptp);
+    l3pte  = PTBL_TO_PTADDR(l3ptbl) + ELAN3_L3_INDEX(addr)*ELAN3_PTE_SIZE;
+    
+    HAT_PRINTF2 (2, "elan3mmu_ptefind: l3ptbl %p l3pte %lx\n", l3ptbl, l3pte);
+                
+    switch (elan3mmu_lock_ptbl (l3ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_3, &l3lock, &l3flags))
+    {
+    case LK_PTBL_OK:
+       *level = 3;
+       *plock = l3lock;
+       *pptbl = l3ptbl;
+       *flags = l3flags;
+
+       return (l3pte);
+       
+    case LK_PTBL_FAILED:
+       panic ("elan3mmu_ptefind: l3 lock failed");
+       /* NOTREACHED */
+
+    case LK_PTBL_MISMATCH:
+       HAT_PRINTF6 (2, "elan3mmu_ptefind: PTBL_MISMATCH : ptbl %p flags %x elan3mmu %p base %x (%p %x)\n",
+                    l3ptbl, l3ptbl->ptbl_flags, l3ptbl->ptbl_elan3mmu, l3ptbl->ptbl_base, elan3mmu, addr);
+                    
+       /*
+        * We've trogged down to this ptbl,  but someone has just
+        * stolen it,  so try all over again.
+        */
+       goto retryl1;
+       
+    default:
+       panic ("elan3mmu_ptefind: elan3mmu_lock_ptbl returned bad value");
+       /* NOTREACHED */
+    }
+    /* NOTREACHED */
+    return ((sdramaddr_t) 0);
+}
+
+sdramaddr_t 
+elan3mmu_ptp2pte (ELAN3MMU *elan3mmu, sdramaddr_t ptp, int level)
+{
+    ELAN3_PTP tptp = elan3_readptp (elan3mmu->elan3mmu_dev, ptp);
+
+    ASSERT (level != 3 && ELAN3_PTP_TYPE(tptp) == ELAN3_ET_PTE);
+
+    return PTP_TO_PT_PADDR(tptp);
+}
+
+sdramaddr_t
+elan3mmu_ptealloc (ELAN3MMU *elan3mmu, E3_Addr addr, int level, 
+                  ELAN3_PTBL **pptbl, spinlock_t **plock, int attr, unsigned long *flags)
+{
+    ELAN3_DEV   *dev     = elan3mmu->elan3mmu_dev;
+    ELAN3_PTBL  *l1ptbl;
+    ELAN3_PTBL  *lXptbl;
+    int         idx;
+    sdramaddr_t l1ptp;
+    ELAN3_PTP    tl1ptp;
+    E3_Addr     l1base;
+    spinlock_t *l1lock;
+    ELAN3_PTBL  *l2ptbl;
+    sdramaddr_t l2ptp;
+    ELAN3_PTP    tl2ptp;
+    E3_Addr     l2base;
+    spinlock_t *l2lock;
+    ELAN3_PTBL  *l3ptbl;
+    sdramaddr_t l3pte;
+    E3_Addr     l3base;
+    spinlock_t *l3lock;
+
+    unsigned long l1flags;
+    unsigned long l2flags;
+    unsigned long l3flags;
+
+    HAT_PRINTF2 (2, "elan3mmu_ptealloc: elan3mmu %p addr %x\n", elan3mmu, addr);
+
+    l1ptbl = elan3mmu->elan3mmu_l1ptbl;
+    if (l1ptbl == NULL)
+       return ((sdramaddr_t) 0);
+
+    l1ptp  = PTBL_TO_PTADDR(l1ptbl) + ELAN3_L1_INDEX(addr)*ELAN3_PTP_SIZE;
+    l1base = ELAN3_L1_BASE(addr);
+               
+retryl1:
+    tl1ptp = elan3_readptp (dev, l1ptp);
+
+    HAT_PRINTF5 (2, "elan3mmu_ptealloc: l1ptbl %p 1ptp %lx l1base %x (%x) : tl1ptp %x\n",
+                l1ptbl, l1ptp, l1base, l1ptbl->ptbl_base, tl1ptp);
+
+    switch (ELAN3_PTP_TYPE(tl1ptp))
+    {
+    case ELAN3_ET_PTE:
+       if (level == PTBL_LEVEL_1)
+       {
+           elan3mmu_lock_ptbl (l1ptbl, 0, elan3mmu, addr, PTBL_LEVEL_1, &l1lock, &l1flags);
+
+           tl1ptp = elan3_readptp (dev, l1ptp);
+           if (ELAN3_PTP_TYPE(tl1ptp) != ELAN3_ET_PTE)
+           {
+               elan3mmu_unlock_ptbl (l1ptbl, l1lock, l1flags);
+               goto retryl1;
+           }
+           
+           *pptbl = l1ptbl;
+           *plock = l1lock;
+           *flags = l1flags;
+
+           /* return holding l1lock */
+           return (l1ptp);
+       }
+       panic ("elan3mmu_ptealloc: found pte in level 1 page table");
+       /* NOTREACHED */
+
+    case ELAN3_ET_PTP:
+       if (level == PTBL_LEVEL_1)
+           panic ("elan3mmu_ptealloc: found PTP when loading a level 1 PTE\n");
+       break;
+
+    case ELAN3_ET_INVALID:
+       if (level == PTBL_LEVEL_1)
+       {
+           if ((lXptbl = elan3mmu_alloc_pte (dev, elan3mmu,  &idx)) == NULL)
+               return ((sdramaddr_t) 0);
+
+           elan3mmu_lock_ptbl (l1ptbl, 0, elan3mmu, addr, PTBL_LEVEL_1, &l1lock, &l1flags);
+
+           tl1ptp = elan3_readptp (dev, l1ptp);
+           if (ELAN3_PTP_TYPE(tl1ptp) != ELAN3_ET_INVALID)
+           {
+               /* raced with someone else, whose got there first */
+               elan3mmu_free_pte (dev, elan3mmu, lXptbl, idx);
+
+               /* drop the l1lock and retry */
+               elan3mmu_unlock_ptbl (l1ptbl, l1lock, l1flags);
+               goto retryl1;
+           }
+           
+           tl1ptp = PTBL_TO_PTADDR(lXptbl) | (idx * ELAN3_PTE_SIZE) | ELAN3_ET_PTE;
+           
+           elan3_writeptp (dev, l1ptp, tl1ptp);
+
+           *pptbl = l1ptbl;
+           *plock = l1lock;
+           *flags = l1flags;
+
+           /* return holding l1lock */
+           return (l1ptp);
+       }
+
+       if (level == PTBL_LEVEL_2)
+       {
+           if ((lXptbl = elan3mmu_alloc_pte (dev, elan3mmu, &idx)) == NULL)
+               return ((sdramaddr_t) 0);
+
+           if ((l2ptbl = elan3mmu_alloc_l2ptbl (dev, attr, l1ptbl, elan3mmu, ELAN3_L2_BASE(addr), &l2lock, &l2flags)) == NULL)
+           {
+               elan3mmu_free_pte (dev, elan3mmu, lXptbl, idx); 
+               return ((sdramaddr_t) 0);
+           }
+
+           /* Connect l2ptbl to the new LX pte */
+           l2ptp  = PTBL_TO_PTADDR(l2ptbl) + ELAN3_L2_INDEX(addr) * ELAN3_PTP_SIZE;
+           tl2ptp = PTBL_TO_PTADDR(lXptbl) | (idx * ELAN3_PTE_SIZE) | ELAN3_ET_PTE;
+
+           elan3_writeptp (dev, l2ptp, tl2ptp);
+
+           /* Now need to lock the l1 ptbl */
+           elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+
+           elan3mmu_lock_ptbl (l1ptbl, 0, elan3mmu, addr, PTBL_LEVEL_1, &l1lock, &l1flags);
+           elan3mmu_lock_ptbl (l2ptbl, 0, elan3mmu, addr, PTBL_LEVEL_2, &l2lock, &l2flags);
+
+           tl1ptp = elan3_readptp (dev, l1ptp);
+           if (ELAN3_PTP_TYPE(tl1ptp) != ELAN3_ET_INVALID)
+           {
+               HAT_PRINTF0 (2, "elan3mmu_ptealloc: beaten to it,  free l2 ptbl/lx pte\n");
+               
+               tl2ptp = ELAN3_INVALID_PTP;
+               elan3_writeptp (dev, l2ptp, tl2ptp);
+               
+               HAT_PRINTF2 (2, "elan3mmu_ptealloc: write level 2 ptp %lx to %x\n", l2ptp, tl2ptp);
+               HAT_PRINTF2 (2, "elan3mmu_ptealloc: freeing l2 ptbl %p (%x)\n", l2ptbl, l2ptbl->ptbl_flags);
+               
+               elan3mmu_free_l2ptbl (dev, l2ptbl, l2lock, l2flags);
+               elan3mmu_free_pte (dev, elan3mmu, lXptbl, idx);
+
+               elan3mmu_unlock_ptbl (l1ptbl, l1lock, l1flags);
+
+               goto retryl1;
+           }
+           
+           /* Now have L1 locked,  so install the L2 ptbl */
+           l1ptp  = PTBL_TO_PTADDR(l1ptbl) + ELAN3_L1_INDEX(addr)*ELAN3_PTP_SIZE;
+           tl1ptp = PTBL_TO_PTADDR(l2ptbl) | ELAN3_ET_PTP;
+           l1ptbl->ptbl_valid++;
+
+           HAT_PRINTF3 (2, "elan3mmu_ptealloc: inc valid for level %d ptbl %p to %d\n", 
+                        PTBL_LEVEL(l1ptbl->ptbl_flags), l1ptbl, l1ptbl->ptbl_valid);
+           
+           elan3_writeptp (dev, l1ptp, tl1ptp);
+           
+           HAT_PRINTF2 (2, "elan3mmu_ptealloc: write l1ptp %lx to %x\n", l1ptp, tl1ptp);
+
+           /* unordered unlock - lock l1ptbl, lock l2ptbl, unlock l1ptbl */
+           elan3mmu_unlock_ptbl (l1ptbl, l1lock, l2flags); /* need to unlock with the l2flags to keep irq order correct */
+
+           *pptbl = l2ptbl;
+           *plock = l2lock;
+           *flags = l1flags; /* return the l1flags here as we have released the l2flags already to keep order */
+
+           /* return holding l2lock */
+           return (l2ptp);
+       }
+
+       HAT_PRINTF0 (2, "elan3mmu_ptealloc: allocating level 2 and level 3 page tables\n");
+
+       /* Allocate a level 2 and level 3 page table and link them together */
+       if ((l2ptbl = elan3mmu_alloc_l2ptbl (dev, attr, l1ptbl, elan3mmu, ELAN3_L2_BASE(addr), &l2lock, &l2flags)) == NULL)
+           return ((sdramaddr_t) 0);
+
+       if ((l3ptbl = elan3mmu_alloc_l3ptbl (dev, attr | PTE_NO_SLEEP, l2ptbl, elan3mmu, ELAN3_L3_BASE(addr), &l3lock, &l3flags)) == NULL)
+       {
+           elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+           return ((sdramaddr_t) 0);
+       }
+
+       ASSERT (PTBL_IS_LOCKED (l2ptbl->ptbl_flags));
+       ASSERT (PTBL_LEVEL (l2ptbl->ptbl_flags) == PTBL_LEVEL_2);
+       ASSERT (PTBL_IS_LOCKED (l3ptbl->ptbl_flags));
+       ASSERT (PTBL_LEVEL (l3ptbl->ptbl_flags) == PTBL_LEVEL_3);
+
+       HAT_PRINTF6 (2, "elan3mmu_ptealloc: l2ptbl %p (%x,%x) l3ptbl %p (%x,%x)\n",
+                    l2ptbl, l2ptbl->ptbl_flags, l2ptbl->ptbl_base,
+                    l3ptbl, l3ptbl->ptbl_flags, l3ptbl->ptbl_base);
+
+       if (CTXT_IS_KERNEL (elan3mmu->elan3mmu_ctxt))
+       {
+           l2ptbl->ptbl_flags |= PTBL_KERNEL;
+           elan3mmu_kernel_l3ptbl (l3ptbl);
+       }
+       
+       /*
+        * Connect L3 ptbl to the new L2 ptbl.
+        */
+       l2ptp  = PTBL_TO_PTADDR(l2ptbl) + ELAN3_L2_INDEX(addr) * ELAN3_PTP_SIZE;
+       tl2ptp = PTBL_TO_PTADDR(l3ptbl) | ELAN3_ET_PTP;
+
+       l2ptbl->ptbl_valid = 1;
+
+       HAT_PRINTF3 (2, "elan3mmu_ptealloc: set valid for level %d ptbl %p to %d\n", 
+                    PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid);
+
+       HAT_PRINTF2 (2, "elan3mmu_ptealloc: write level 2 ptp %lx to %x\n", l2ptp, tl2ptp);
+
+       elan3_writeptp (dev, l2ptp, tl2ptp);
+
+       /* 
+        * Now need to lock the l1 ptbl - to maintain lock ordering
+        * we set the PTBL_KEEP bit to stop the l3 ptbl from being 
+        * stolen and drop the locks in the order we aquired them
+        */
+       l3ptbl->ptbl_flags |= PTBL_KEEP;
+
+       elan3mmu_unlock_ptbl (l3ptbl, l3lock, l3flags);
+       elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+
+       elan3mmu_lock_ptbl (l1ptbl, 0, elan3mmu, addr, PTBL_LEVEL_1, &l1lock, &l1flags);
+       elan3mmu_lock_ptbl (l3ptbl, 0, elan3mmu, addr, PTBL_LEVEL_3, &l3lock, &l3flags);
+           
+       l3ptbl->ptbl_flags &= ~PTBL_KEEP;
+          
+       /* Now have l1 and l3 ptbls locked,  so install the new l2 ptbl into the l1. */
+       tl1ptp = elan3_readptp (dev, l1ptp);
+
+       HAT_PRINTF2 (2, "elan3mmu_ptealloc: l1ptp %lx is %x\n", l1ptp, tl1ptp);
+
+       if (ELAN3_PTP_TYPE(tl1ptp) != ELAN3_ET_INVALID)
+       {
+           HAT_PRINTF0 (2, "elan3mmu_ptealloc: beaten to it,  free l2/l3 ptbls\n");
+
+           /* free off the level 3 page table */
+           HAT_PRINTF2 (2, "elan3mmu_ptealloc: freeing l3 ptbl %p (%x)\n", l3ptbl, l3ptbl->ptbl_flags);
+
+           l3ptbl->ptbl_flags &= ~PTBL_KEEP;
+           elan3mmu_free_l3ptbl (dev, l3ptbl, l3lock, l3flags);
+
+           /* and unlock the level 1 ptbl */
+           elan3mmu_unlock_ptbl (l1ptbl, l1lock, l1flags);
+           
+           /* lock the level 2 page table, and clear out the PTP, then free it */
+           (void) elan3mmu_lock_ptbl (l2ptbl, 0, elan3mmu, addr, PTBL_LEVEL_2, &l2lock, &l2flags);
+
+           HAT_PRINTF2 (2, "elan3mmu_ptealloc: locked l2 ptbl %p (%x)\n", l2ptbl, l2ptbl->ptbl_flags);
+           
+           tl2ptp = ELAN3_INVALID_PTP;
+           elan3_writeptp (dev, l2ptp, tl2ptp);
+           l2ptbl->ptbl_valid = 0;
+
+           HAT_PRINTF3 (2, "elan3mmu_ptealloc: set to 0 valid for level %d ptbl %p to %d\n", PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid); 
+
+           HAT_PRINTF2 (2, "elan3mmu_ptealloc: write level 2 ptp %lx to %x\n", l2ptp, tl2ptp);
+           HAT_PRINTF2 (2, "elan3mmu_ptealloc: freeing l2 ptbl %p (%x)\n", l2ptbl, l2ptbl->ptbl_flags);
+
+           elan3mmu_free_l2ptbl (dev, l2ptbl, l2lock, l2flags);
+
+           goto retryl1;
+       }
+       
+       HAT_PRINTF4 (2, "elan3mmu_ptealloc: l1ptbl is %p (%x), l3ptbl is %p (%x)\n", 
+                    l1ptbl, l1ptbl->ptbl_flags, l3ptbl, l3ptbl->ptbl_flags);
+
+       /* Now have L1 and L3 locked,  so install the L2 ptbl */
+       l1ptp  = PTBL_TO_PTADDR(l1ptbl) + ELAN3_L1_INDEX(addr)*ELAN3_PTP_SIZE;
+       tl1ptp = PTBL_TO_PTADDR(l2ptbl) | ELAN3_ET_PTP;
+       l1ptbl->ptbl_valid++;
+
+       HAT_PRINTF3 (2, "elan3mmu_ptealloc: inc valid for level %d ptbl %p to %d\n", 
+                    PTBL_LEVEL(l1ptbl->ptbl_flags), l1ptbl, l1ptbl->ptbl_valid);
+
+       elan3_writeptp (dev, l1ptp, tl1ptp);
+
+       HAT_PRINTF2 (2, "elan3mmu_ptealloc: write l1ptp %lx to %x\n", l1ptp, tl1ptp);
+
+       /* unordered unlock - lock l1ptbl, lock l3ptbl, unlock l1ptbl */
+       elan3mmu_unlock_ptbl (l1ptbl, l1lock, l3flags); /* free using l3flags to keep irq ordering */
+
+       l3pte = PTBL_TO_PTADDR (l3ptbl) + ELAN3_L3_INDEX(addr)*ELAN3_PTE_SIZE;
+
+       /* Level 3 ptbl is already locked,  so just return the pte */
+       *pptbl = l3ptbl;
+       *plock = l3lock;
+       *flags = l1flags; /* return l1flags to keep irq ordering */
+
+       return (l3pte);
+
+    default:
+       panic ("elan3mmu_ptealloc: found bad entry in level 1 page table");
+       /* NOTREACHED */
+    }
+
+    HAT_PRINTF1 (2, "elan3mmu_ptealloc: chain to level 2 ptbl from ptp %x\n", tl1ptp);
+
+    l2ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl1ptp);
+    l2ptp  = PTBL_TO_PTADDR(l2ptbl) + ELAN3_L2_INDEX(addr)*ELAN3_PTP_SIZE;
+    l2base = ELAN3_L2_BASE(addr);
+
+    tl2ptp = elan3_readptp (dev, l2ptp);
+
+    HAT_PRINTF5 (2, "elan3mmu_ptealloc: l2ptbl %p l2ptp %lx l2base %x (%x) : tl2ptp %x\n",
+                l2ptbl, l2ptp, l2base, l2ptbl->ptbl_base, tl2ptp);
+
+    switch (ELAN3_PTP_TYPE(tl2ptp))
+    {
+    case ELAN3_ET_PTE:
+       if (level == PTBL_LEVEL_2) {
+           /* this is a pointer to a pte,  we should just return it */
+
+           switch (elan3mmu_lock_ptbl (l2ptbl, 0, elan3mmu, addr, PTBL_LEVEL_2, &l2lock, &l2flags))
+           {
+           case LK_PTBL_OK:
+               break;
+       
+           case LK_PTBL_FAILED:
+               panic ("elan3mmu_ptealloc: l2 lock failed");
+               /* NOTREACHED */
+               
+           case LK_PTBL_MISMATCH:
+               HAT_PRINTF6 (2, "elan3mmu_ptealloc: PTBL_MISMATCH : ptbl %p flags %x elan3mmu %p base %x (%p %x)\n",
+                            l2ptbl, l2ptbl->ptbl_flags, l2ptbl->ptbl_elan3mmu, l2ptbl->ptbl_base, elan3mmu, addr);
+               
+               /*
+                * We've trogged down to this ptbl,  but someone has just
+                * stolen it,  so try all over again.
+                */
+               goto retryl1;
+               
+           default:
+               panic ("elan3mmu_ptealloc: elan3mmu_lock_ptbl returned bad value");
+               /* NOTREACHED */
+           }
+
+
+           tl2ptp = elan3_readptp (dev, l2ptp);
+           if (ELAN3_PTP_TYPE(tl2ptp) != ELAN3_ET_PTE)
+           {
+               elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+               goto retryl1;
+           }
+
+           *pptbl = l2ptbl;
+           *plock = l2lock;
+           *flags = l2flags;
+
+           /* return holdind l2lock */
+           return (l2ptp);
+       }
+       panic ("elan3mmu: found pte in level 2 page table");
+       /* NOTREACHED */
+
+    case ELAN3_ET_PTP:
+       break;
+
+    case ELAN3_ET_INVALID:
+       if (level == PTBL_LEVEL_2) 
+       {
+           if ((lXptbl = elan3mmu_alloc_pte (dev, elan3mmu, &idx)) == NULL)
+               return ((sdramaddr_t) 0);
+
+           switch (elan3mmu_lock_ptbl (l2ptbl, 0, elan3mmu, addr, PTBL_LEVEL_2, &l2lock, &l2flags))
+           {
+           case LK_PTBL_OK:
+               break;
+       
+           case LK_PTBL_FAILED:
+               panic ("elan3mmu_ptealloc: l2 lock failed");
+               /* NOTREACHED */
+               
+           case LK_PTBL_MISMATCH:
+               HAT_PRINTF6 (2, "elan3mmu_ptealloc: PTBL_MISMATCH : ptbl %p flags %x elan3mmu %p base %x (%p %x)\n",
+                            l2ptbl, l2ptbl->ptbl_flags, l2ptbl->ptbl_elan3mmu, l2ptbl->ptbl_base, elan3mmu, addr);
+               
+               /*
+                * We've trogged down to this ptbl,  but someone has just
+                * stolen it,  so try all over again.
+                */
+               goto retryl1;
+               
+           default:
+               panic ("elan3mmu_ptealloc: elan3mmu_lock_ptbl returned bad value");
+               /* NOTREACHED */
+           }
+
+           tl2ptp = elan3_readptp (dev, l2ptp);
+           if (ELAN3_PTP_TYPE(tl2ptp) != ELAN3_ET_INVALID)
+           {
+               HAT_PRINTF0 (2, "elan3mmu_ptealloc: beaten to it,  free lx pte\n");
+
+               elan3mmu_free_pte (dev, elan3mmu, lXptbl, idx);
+
+               elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+               goto retryl1;
+           }
+
+           /* Connect l2ptbl to the new LX pte */
+           tl2ptp = PTBL_TO_PTADDR(lXptbl) | (idx * ELAN3_PTE_SIZE) | ELAN3_ET_PTE;
+                  
+           HAT_PRINTF3 (2, "elan3mmu_ptealloc: inc valid for level %d ptbl %p to %d\n", 
+                        PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid);
+           
+           elan3_writeptp (dev, l2ptp, tl2ptp);
+           
+           HAT_PRINTF2 (2, "elan3mmu_ptealloc: write l2ptp %lx to %x\n", l2ptp, tl2ptp);
+
+           *pptbl = l2ptbl;
+           *plock = l2lock;
+           *flags = l2flags;
+
+           /* return holding l2lock */
+           return (l2ptp);
+       }
+       HAT_PRINTF0 (2, "elan3mmu_ptealloc: allocate level 3 page table\n");
+
+       if ((l3ptbl = elan3mmu_alloc_l3ptbl (dev, attr, l2ptbl, elan3mmu, ELAN3_L3_BASE(addr), &l3lock, &l3flags)) == NULL)
+           return ((sdramaddr_t) 0);
+
+       if (CTXT_IS_KERNEL (elan3mmu->elan3mmu_ctxt))
+           elan3mmu_kernel_l3ptbl (l3ptbl);
+
+       /* 
+        * Now need to lock the l2 ptbl - to maintain lock ordering
+        * we set the PTBL_KEEP bit to stop the l3 ptbl from being 
+        * stolen and drop the locks in the order we aquired them
+        */
+       l3ptbl->ptbl_flags |= PTBL_KEEP;
+
+       elan3mmu_unlock_ptbl (l3ptbl, l3lock, l3flags);
+
+       if (elan3mmu_lock_ptbl (l2ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_2, &l2lock, &l2flags) == LK_PTBL_MISMATCH)
+       {
+           HAT_PRINTF0 (2, "elan3mmu_ptealloc: l2ptbl freed, free l3 ptbl and try again\n");
+             
+           elan3mmu_lock_ptbl (l3ptbl, 0, elan3mmu, addr, PTBL_LEVEL_3, &l3lock, &l3flags);
+
+           /* free off the level 3 page table, and try again */
+           l3ptbl->ptbl_flags &= ~PTBL_KEEP;
+           elan3mmu_free_l3ptbl (dev, l3ptbl, l3lock, l3flags);
+           
+           goto retryl1;
+       }
+
+       elan3mmu_lock_ptbl (l3ptbl, 0, elan3mmu, addr, PTBL_LEVEL_3, &l3lock, &l3flags);
+
+       l3ptbl->ptbl_flags &= ~PTBL_KEEP;
+
+       /* Now have L2 and L3 ptbls locked, see if someone has beaten us to it. */
+       tl2ptp = elan3_readptp (dev, l2ptp);
+
+       HAT_PRINTF2 (2, "elan3mmu_ptealloc: l2ptp at %lx is %x\n", l2ptp, tl2ptp);
+
+       if (ELAN3_PTP_TYPE(tl2ptp) != ELAN3_ET_INVALID)
+       {
+           HAT_PRINTF0 (2, "elan3mmu_ptealloc: beaten to it, free l3 ptbl and try again\n");
+
+           /* free off the level 3 page table, and try again */
+           l3ptbl->ptbl_flags &= ~PTBL_KEEP;
+           elan3mmu_free_l3ptbl (dev, l3ptbl, l3lock, l3flags);
+           
+           /* Someone has allocated the ptbl before us */
+           elan3mmu_unlock_ptbl (l2ptbl, l2lock, l2flags);
+           
+           goto retryl1;
+       }
+
+       ASSERT (PTBL_IS_LOCKED (l2ptbl->ptbl_flags));
+
+       /* Install the L3 ptbl into the L2 one */
+       l2ptp  = PTBL_TO_PTADDR(l2ptbl) + ELAN3_L2_INDEX(addr)*ELAN3_PTP_SIZE;
+       tl2ptp = PTBL_TO_PTADDR(l3ptbl) | ELAN3_ET_PTP;
+       l2ptbl->ptbl_valid++;
+
+       HAT_PRINTF3 (2, "elan3mmu_ptealloc: inc valid for level %d ptbl %p to %d\n",
+                    PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid);
+
+       elan3_writeptp (dev, l2ptp, tl2ptp);
+
+       HAT_PRINTF2 (2, "elan3mmu_ptealloc: write level 2 ptp %lx to %x\n", l2ptp, tl2ptp);
+
+       /* unordered unlock - lock l2ptbl, lock l3ptbl, unlock l2ptbl */
+       elan3mmu_unlock_ptbl (l2ptbl, l2lock, l3flags); /* free with the l3flags to keep irq ordering */
+
+       l3pte = PTBL_TO_PTADDR(l3ptbl) + ELAN3_L3_INDEX(addr)*ELAN3_PTE_SIZE;
+       
+       /* Level 3 ptbl is already locked, so just return the pte */
+       *pptbl = l3ptbl;
+       *plock = l3lock;
+       *flags = l2flags; /* return l2flags to keep irq ordering */
+
+       return (l3pte);
+
+    default:
+       panic ("elan3mmu_ptealloc: found bad entry in level 2 page table");
+       /* NOTREACHED */
+    }
+
+    HAT_PRINTF1 (2, "elan3mmu_ptealloc: chain to level 3 page table from ptp %x\n", tl2ptp);
+
+    l3ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl2ptp);
+    l3pte  = PTBL_TO_PTADDR(l3ptbl) + ELAN3_L3_INDEX(addr)*ELAN3_PTE_SIZE;
+    l3base = ELAN3_L3_BASE(addr);
+
+    HAT_PRINTF4 (2, "elan3mmu_ptealloc: l3ptbl %p 3pte %lx l3base %x (%x)\n",
+                l3ptbl, l3pte, l3base, l3ptbl->ptbl_base);
+                
+    if (elan3mmu_lock_ptbl (l3ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_3, &l3lock, &l3flags) == LK_PTBL_OK)
+    {
+       *pptbl = l3ptbl;
+       *plock = l3lock;
+       *flags = l3flags;
+
+       return (l3pte);
+    }
+
+    /* got all the way down here,  but its been nicked before we could lock it */
+    /* so try all over again */
+    goto retryl1;
+}
+
+void
+elan3mmu_l1inval (ELAN3MMU *elan3mmu, ELAN3_PTBL *l1ptbl, int attr)
+{
+    ELAN3_DEV     *dev = elan3mmu->elan3mmu_dev;
+    ELAN3_PTP      invalidptp = ELAN3_INVALID_PTP;
+    ELAN3_PTP      tl1ptp;
+    sdramaddr_t   l1ptp;
+    E3_Addr       addr;
+    spinlock_t   *l2lock;
+    ELAN3_PTBL    *l2ptbl;
+    ELAN3_PTBL    *lXptbl;
+    int           idx;
+    int                  i;
+    int                  ret;
+    unsigned long flags;
+
+    l1ptp = PTBL_TO_PTADDR(l1ptbl);
+
+    HAT_PRINTF2 (1, "elan3mmu_l1inval: l1ptbl %p l1ptp %lx\n", l1ptbl, l1ptp);
+
+    for (i = 0, addr = 0; i < ELAN3_L1_ENTRIES; i++, l1ptp += ELAN3_PTP_SIZE)
+    {
+       tl1ptp = elan3_readptp (dev, l1ptp);
+       switch (ELAN3_PTP_TYPE(tl1ptp))
+       {
+       case ELAN3_ET_PTE:
+           lXptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl1ptp);
+           idx    = (PTP_TO_PT_PADDR(tl1ptp) - PTBL_TO_PTADDR(lXptbl))/ELAN3_PTE_SIZE;  
+
+           HAT_PRINTF3 (2, "elan3mmu_l1inval: l1ptbl %p : lXptbl %p idx %d\n",
+                        l1ptbl, lXptbl, idx);
+
+           /* invalidate the L1 pte. */
+           elan3_writeptp (dev, l1ptp, invalidptp);
+           if (! (attr & PTE_UNLOAD_NOFLUSH))
+               ElanFlushTlb (dev);         
+
+           l1ptbl->ptbl_valid--;
+           elan3mmu_free_pte ( dev, elan3mmu,  lXptbl, idx); 
+
+           HAT_PRINTF3 (2, "elan3mmu_l1inval: dec valid for level %d ptbl %p to %d\n",
+                    PTBL_LEVEL(l1ptbl->ptbl_flags), l1ptbl, l1ptbl->ptbl_valid);
+           
+           break;
+
+       case ELAN3_ET_PTP:
+           HAT_PRINTF5 (2, "elan3mmu_l1inval: l1ptbl %p : ptp %lx (%x) addr %x (%d)\n",
+                        l1ptbl, l1ptp, tl1ptp, addr, i);
+
+           /* invalidate the L1 ptp. */
+           elan3_writeptp (dev, l1ptp, invalidptp);
+           if (! (attr & PTE_UNLOAD_NOFLUSH))
+               ElanFlushTlb (dev);
+
+           /* invalidate the level 2 page table */
+           l2ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl1ptp);
+           ret    = elan3mmu_l2inval (elan3mmu, l2ptbl, attr | PTE_UNLOAD_NOFLUSH, addr, &l2lock, &flags);
+
+           ASSERT ((l2ptbl->ptbl_flags & PTBL_KEEP) == 0);
+           
+           if (ret == LK_PTBL_OK)
+           {
+               if (((l2ptbl->ptbl_flags & PTBL_KEEP) == 0) && l2ptbl->ptbl_valid == 0)
+               {
+                   HAT_PRINTF1 (2, "elan3mmu_l1inval: free l2ptbl %p\n", l2ptbl);
+                   
+                   l1ptbl->ptbl_valid--;
+                   elan3mmu_free_l2ptbl (elan3mmu->elan3mmu_dev, l2ptbl, l2lock, flags);
+
+                   HAT_PRINTF3 (2, "elan3mmu_l1inval: dec valid for level %d ptbl %p to %d\n", 
+                                PTBL_LEVEL(l1ptbl->ptbl_flags), l1ptbl, l1ptbl->ptbl_valid);
+               }
+               else
+               {
+                   /* need to keep this page table,  so even though its now empty, */
+                   /* chain it back in */
+                   HAT_PRINTF1 (2, "elan3mmu_l1inval: keep l2ptbl %p\n", l2ptbl);
+
+                   elan3_writeptp (dev, l1ptp, tl1ptp);
+                   elan3mmu_unlock_ptbl (l2ptbl, l2lock, flags);
+               }    
+           }
+           else
+           {
+               l1ptbl->ptbl_valid--;
+
+               HAT_PRINTF3 (2, "elan3mmu_l1inval: dec valid for level %d ptbl %p to %d\n", 
+                            PTBL_LEVEL(l1ptbl->ptbl_flags), l1ptbl, l1ptbl->ptbl_valid);
+           }
+           break;
+           
+       case ELAN3_ET_INVALID:
+           break;
+
+       default:
+           panic ("elan3mmu_l1inval: found invalid entry in level 1 page table");
+           /* NOTREACHED */
+       }
+
+       if (l1ptbl->ptbl_valid == 0)
+           break;
+
+       addr += ELAN3_L1_SIZE;
+    }
+}
+
+int
+elan3mmu_l2inval (ELAN3MMU *elan3mmu, ELAN3_PTBL *l2ptbl, int attr, E3_Addr addr, spinlock_t **pl2lock, unsigned long *flags)
+{
+    ELAN3_DEV   *dev = elan3mmu->elan3mmu_dev;
+    ELAN3_PTP    invalidptp = ELAN3_INVALID_PTP;
+    ELAN3_PTP    tl2ptp;
+    sdramaddr_t l2ptp;
+    spinlock_t *l3lock;
+    unsigned long l3flags;
+    ELAN3_PTBL  *l3ptbl;
+    ELAN3_PTBL  *lXptbl;
+    int         idx;
+    int                i;
+    int                ret;
+
+    HAT_PRINTF2 (1, "elan3mmu_l2inval: l2ptbl %p addr %x\n", l2ptbl, addr);
+
+    ASSERT (PTBL_LEVEL (l2ptbl->ptbl_flags) == PTBL_LEVEL_2);
+    ASSERT (PTBL_LEVEL (l2ptbl->ptbl_parent->ptbl_flags) == PTBL_LEVEL_1);
+
+    ret = elan3mmu_lock_ptbl (l2ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_2, pl2lock, flags);
+
+    ASSERT (ret == LK_PTBL_OK);
+    ASSERT (l2ptbl->ptbl_elan3mmu == elan3mmu);
+    ASSERT (l2ptbl->ptbl_parent->ptbl_elan3mmu == elan3mmu);
+
+    l2ptp = PTBL_TO_PTADDR(l2ptbl);
+
+    for (i = 0; i < ELAN3_L2_ENTRIES; i++, l2ptp += ELAN3_PTP_SIZE)
+    {
+       tl2ptp = elan3_readptp (dev, l2ptp);
+       switch (ELAN3_PTP_TYPE(tl2ptp))
+       {
+       case ELAN3_ET_PTE:
+           lXptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl2ptp);
+           idx    = (PTP_TO_PT_PADDR(tl2ptp) - PTBL_TO_PTADDR(lXptbl))/ELAN3_PTE_SIZE;  
+
+           HAT_PRINTF3 (2, "elan3mmu_l2inval: l2ptbl %p : lXptbl %p idx %d\n",
+                        l2ptbl, lXptbl, idx);
+
+           /* invalidate the L2 pte. */
+           elan3_writeptp (dev, l2ptp, invalidptp);
+           if (! (attr & PTE_UNLOAD_NOFLUSH))
+               ElanFlushTlb (dev);
+
+           l2ptbl->ptbl_valid--;
+           elan3mmu_free_pte ( dev, elan3mmu, lXptbl, idx); 
+
+           HAT_PRINTF3 (2, "elan3mmu_l2inval: dec valid for level %d ptbl %p to %d\n", PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid); 
+
+           break;
+           
+       case ELAN3_ET_PTP:
+           HAT_PRINTF5 (2, "elan3mmu_l2inval: l2ptbl %p : ptp %lx (%x) addr %x (%d)\n",
+                        l2ptbl, l2ptp, tl2ptp, addr, i);
+
+           /* invalidate the L2 ptp. */
+           elan3_writeptp (dev, l2ptp, invalidptp);
+           if (! (attr & PTE_UNLOAD_NOFLUSH))
+               ElanFlushTlb (dev);
+           
+           /* unload the level 3 page table */
+           l3ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl2ptp);
+           ret = elan3mmu_l3inval (elan3mmu, l3ptbl, attr | PTE_UNLOAD_NOFLUSH, addr, &l3lock, &l3flags);
+           
+           if (ret == LK_PTBL_OK)
+           {
+               if ((l3ptbl->ptbl_flags & PTBL_KEEP) == 0 && l3ptbl->ptbl_valid == 0)
+               {
+                   /* decrement the valid count of the level 2 page table, and */
+                   /* free off the level 3 page table */
+                   HAT_PRINTF1 (2, "elan3mmu_l2inval: free l3ptbl %p\n", l3ptbl);
+
+                   l2ptbl->ptbl_valid--;
+                   elan3mmu_free_l3ptbl (elan3mmu->elan3mmu_dev, l3ptbl, l3lock, l3flags);
+
+                   HAT_PRINTF3 (2, "elan3mmu_l2inval: dec valid for level %d ptbl %p to %d\n", 
+                                PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid);
+               }
+               else
+               {
+                   /* need to keep this page table,  so even though its now empty, */
+                   /* chain it back in */
+                   HAT_PRINTF1 (2, "elan3mmu_l2inval: keep l3ptbl %p\n", l3ptbl);
+
+                   elan3_writeptp (dev, l2ptp, tl2ptp);
+                   elan3mmu_unlock_ptbl (l3ptbl, l3lock, l3flags);
+               }
+           }
+           else
+           {
+               l2ptbl->ptbl_valid--;
+
+               HAT_PRINTF3 (2, "elan3mmu_l2inval: dec valid for level %d ptbl %p to %d\n", 
+                            PTBL_LEVEL(l2ptbl->ptbl_flags), l2ptbl, l2ptbl->ptbl_valid);
+           }
+           break;
+           
+       case ELAN3_ET_INVALID:
+           break;
+
+       default:
+           panic ("elan3mmu_l2inval: found pte in level 2 page table");
+           /* NOTREACHED */
+       }
+
+       if (l2ptbl->ptbl_valid == 0)
+           break;
+
+       addr += ELAN3_L2_SIZE;
+    }
+
+    ASSERT (PTBL_IS_LOCKED(l2ptbl->ptbl_flags));
+
+    return (ret);
+}
+
+int 
+elan3mmu_l3inval (ELAN3MMU *elan3mmu, ELAN3_PTBL *l3ptbl, int attr, E3_Addr addr, spinlock_t **pl3lock, unsigned long *flags)
+{
+    int ret;
+
+    HAT_PRINTF3 (2, "elan3mmu_l3inval: l3ptbl %p parent %p addr %x\n", l3ptbl, l3ptbl->ptbl_parent, addr);
+
+    ASSERT (PTBL_IS_LOCKED (l3ptbl->ptbl_parent->ptbl_flags));
+    ASSERT (PTBL_LEVEL (l3ptbl->ptbl_parent->ptbl_flags) == PTBL_LEVEL_2);
+    ASSERT (l3ptbl->ptbl_parent->ptbl_elan3mmu == elan3mmu);
+    ASSERT (l3ptbl->ptbl_parent->ptbl_base == VA2BASE (ELAN3_L2_BASE(addr)));
+    
+    ret = elan3mmu_lock_ptbl (l3ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_3, pl3lock, flags);
+
+    ASSERT (ret == LK_PTBL_OK);
+    ASSERT (PTBL_LEVEL (l3ptbl->ptbl_flags) == PTBL_LEVEL_3);
+
+    elan3mmu_unload_loop (elan3mmu, l3ptbl, 0, ELAN3_L3_ENTRIES, attr);
+
+    ASSERT (PTBL_IS_LOCKED (l3ptbl->ptbl_flags));
+
+    return (ret);
+ }
+
+int
+elan3mmu_lock_this_ptbl (ELAN3_PTBL *ptbl, int flag, spinlock_t **plock, unsigned long *flags)
+{
+    int         level = PTBL_LEVEL (ptbl->ptbl_flags);
+    spinlock_t *lock  = elan3mmu_ptbl_to_lock (level, ptbl);
+
+    local_irq_save (*flags);
+
+    if ((flag & LK_PTBL_NOWAIT) == 0)
+       spin_lock (lock);
+    else if (! spin_trylock (lock)) {
+       local_irq_restore (*flags);
+       return (LK_PTBL_FAILED);
+    }
+    
+    if (level != PTBL_LEVEL (ptbl->ptbl_flags))
+    {
+       spin_unlock (lock);     
+       local_irq_restore (*flags);
+       return (LK_PTBL_MISMATCH);
+    }
+
+    ptbl->ptbl_flags |= PTBL_LOCKED;
+    *plock = lock;
+    return (LK_PTBL_OK);
+}
+
+int
+elan3mmu_lock_ptbl (ELAN3_PTBL *ptbl, u_int flag, ELAN3MMU *elan3mmu, E3_Addr va, int level, spinlock_t **plock, unsigned long *flags)
+{
+    spinlock_t *lock = elan3mmu_ptbl_to_lock (level, ptbl);
+    int         res  = LK_PTBL_MISMATCH;
+
+    local_irq_save (*flags);
+    
+    if ((flag & LK_PTBL_NOWAIT) == 0)
+       spin_lock (lock);
+    else if (spin_trylock (lock) == 0) {
+       local_irq_restore(*flags);
+       return (LK_PTBL_FAILED);
+    }
+    
+    if (PTBL_LEVEL (ptbl->ptbl_flags) != level)
+    {
+       res = LK_PTBL_MISMATCH;
+       goto mismatch;
+    }
+    
+    /* We have the right mutex,  so check that its the ptbl we want. */
+    switch (level)
+    {
+    case PTBL_LEVEL_1: va = ELAN3_L1_BASE(va); break;
+    case PTBL_LEVEL_2: va = ELAN3_L2_BASE(va); break;
+    case PTBL_LEVEL_3: va = ELAN3_L3_BASE(va); break;
+    }
+
+    if (ptbl->ptbl_elan3mmu != elan3mmu || ptbl->ptbl_base != VA2BASE(va))
+    {
+       res = LK_PTBL_MISMATCH;
+       goto mismatch;
+    }
+
+    ASSERT ((ptbl->ptbl_flags & PTBL_LOCKED) == 0);
+    ptbl->ptbl_flags |= PTBL_LOCKED;
+
+    *plock = lock;
+    return (LK_PTBL_OK);
+
+mismatch:
+    if (! (flag & LK_PTBL_FAILOK))
+       panic ("elan3mmu: failed to lock ptbl\n");
+       
+    spin_unlock (lock);
+    local_irq_restore(*flags);
+    return (res);
+}
+
+void
+elan3mmu_unlock_ptbl (ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags)
+{
+    ptbl->ptbl_flags &= ~PTBL_LOCKED;
+    spin_unlock_irqrestore (lock,flags);
+}
+
+static spinlock_t *
+elan3mmu_ptbl_to_lock (int level, ELAN3_PTBL *ptbl)
+{
+    switch (level)
+    {
+    case PTBL_LEVEL_3: return (&l3ptbl_lock[L3PTBL_MTX_HASH(ptbl)]);
+    case PTBL_LEVEL_2: return (&l2ptbl_lock[L2PTBL_MTX_HASH(ptbl)]);
+    case PTBL_LEVEL_1: return (&l1ptbl_lock[L1PTBL_MTX_HASH(ptbl)]);
+    case PTBL_LEVEL_X: 
+       panic ("elan3mmu: ptbl_to_lock, bad level X");
+    default:
+       panic ("elan3mmu: ptbl_to_lock, bad level");
+       /* NOTREACHED */
+    }
+    return (NULL);
+}
+
+void
+elan3mmu_display (ELAN3MMU *elan3mmu, E3_Addr addr)
+{
+    ELAN3_DEV   *dev = elan3mmu->elan3mmu_dev;
+    ELAN3_PTBL  *l1ptbl;
+    sdramaddr_t l1ptp;
+    spinlock_t *l1lock;
+    ELAN3_PTE    tl1pte;
+    ELAN3_PTP    tl1ptp;
+    E3_Addr     l1base;
+    ELAN3_PTBL  *l2ptbl;
+    sdramaddr_t l2ptp;
+    ELAN3_PTE    tl2pte;
+    spinlock_t *l2lock;
+    ELAN3_PTP    tl2ptp;
+    E3_Addr     l2base;
+    ELAN3_PTBL  *l3ptbl;
+    sdramaddr_t l3pte;
+    ELAN3_PTE    tl3pte;
+    spinlock_t *l3lock;
+    ELAN3_PTBL  *lXptbl;
+    int         idx;
+    unsigned long flags;
+
+    elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: elan3mmu %p addr %x\n", elan3mmu, addr);
+
+    l1ptbl = elan3mmu->elan3mmu_l1ptbl;
+
+    if (l1ptbl == NULL)
+       return;
+
+    l1ptp  = PTBL_TO_PTADDR(l1ptbl) + ELAN3_L1_INDEX(addr)*ELAN3_PTP_SIZE;
+    l1base = ELAN3_L1_BASE(addr);
+    
+    tl1ptp = elan3_readptp (dev, l1ptp);
+    elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: l1ptbl %p l1ptp %lx l1base %x : tl1ptp %x\n", l1ptbl, l1ptp, l1base, tl1ptp);
+    
+    switch (ELAN3_PTP_TYPE(tl1ptp))
+    {
+    case ELAN3_ET_PTE:
+       elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: level 1 page table for pte %x\n", tl1ptp);
+    
+       lXptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl1ptp);
+       idx    = (PTP_TO_PT_PADDR(tl1ptp) - PTBL_TO_PTADDR(lXptbl))/ELAN3_PTE_SIZE;  
+       
+       elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: lXptbl %p idx %d\n",lXptbl, idx);
+
+       tl1pte = elan3_readpte (dev,(PTBL_TO_PTADDR (lXptbl) + idx * ELAN3_PTE_SIZE));
+
+       switch (elan3mmu_lock_ptbl (l1ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_1, &l1lock, &flags))
+       {
+       case LK_PTBL_OK:
+           elan3mmu_unlock_ptbl (l1ptbl, l1lock, flags);
+           elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: lvl 1 l1pte matches value %llx\n", (long long) tl1pte);
+           break;
+           
+       case LK_PTBL_FAILED:
+           panic ("elan3mmu_display: l1 lock failed");
+           /* NOTREACHED */
+           
+       case LK_PTBL_MISMATCH:
+           elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: PTBL_MISMATCH : lvl 1 ptbl %p flags %x elan3mmu %p base %x (%p %x) %llx\n",
+                         l1ptbl, l1ptbl->ptbl_flags, l1ptbl->ptbl_elan3mmu, l1ptbl->ptbl_base, elan3mmu, addr, (long long)tl1pte);
+           
+           break;
+       default:
+           panic ("elan3mmu_display: lvl 1 elan3mmu_lock_ptbl returned bad value");
+           /* NOTREACHED */
+       }
+       return;
+       
+    case ELAN3_ET_INVALID:
+       return;
+       
+    case ELAN3_ET_PTP:
+       break;
+       
+    default:
+       panic ("elan3mmu_display: found bad entry in level 1 page table");
+       /* NOTREACHED */
+    }
+    
+    elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: chain to level 2 ptbl from ptp %x\n", tl1ptp);
+    
+    l2ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl1ptp);
+    l2ptp  = PTBL_TO_PTADDR(l2ptbl) + ELAN3_L2_INDEX(addr)*ELAN3_PTP_SIZE;
+    l2base = ELAN3_L2_BASE(addr);
+    
+    tl2ptp = elan3_readptp (dev, l2ptp);
+    elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: l2ptbl %p l2ptp %lx l2base %x : tl2ptp %x\n",
+                l2ptbl, l2ptp, l2base, tl2ptp);
+    
+    switch (ELAN3_PTP_TYPE(tl2ptp))
+    {
+    case ELAN3_ET_PTE:
+       elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: level 2 page table for pte %x\n", tl2ptp);
+    
+       lXptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl2ptp);
+       idx    = (PTP_TO_PT_PADDR(tl2ptp) - PTBL_TO_PTADDR(lXptbl))/ELAN3_PTE_SIZE;  
+       
+       elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: lXptbl %p idx %d\n",lXptbl, idx);
+
+       tl2pte = elan3_readpte (dev,(PTBL_TO_PTADDR (lXptbl) + idx * ELAN3_PTE_SIZE));
+
+       switch (elan3mmu_lock_ptbl (l2ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_2, &l2lock, &flags))
+       {
+       case LK_PTBL_OK:
+           elan3mmu_unlock_ptbl (l2ptbl, l2lock, flags);
+           elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: lvl 2 l1pte matches value %llx\n", (long long)tl2pte);
+           break;
+           
+       case LK_PTBL_FAILED:
+           panic ("elan3mmu_display: l2 lock failed");
+           /* NOTREACHED */
+           
+       case LK_PTBL_MISMATCH:
+           elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: PTBL_MISMATCH : lvl 2 ptbl %p flags %x elan3mmu %p base %x (%p %x) %llx\n",
+                         l2ptbl, l2ptbl->ptbl_flags, l2ptbl->ptbl_elan3mmu, l2ptbl->ptbl_base, elan3mmu, addr, (long long) tl2pte);
+           
+           break;
+       default:
+           panic ("elan3mmu_display: lvl 2 elan3mmu_lock_ptbl returned bad value");
+           /* NOTREACHED */
+       }
+       return;
+       
+    case ELAN3_ET_INVALID:
+       return;
+       
+    case ELAN3_ET_PTP:
+       break;
+
+    default:
+       panic ("elan3mmu_display: found bad entry in level 2 page table");
+       /* NOTREACHED */
+    }
+    
+    elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: chain to level 3 page table from ptp %x\n", tl2ptp);
+    
+    l3ptbl = elan3mmu_ta_to_ptbl (elan3mmu, &tl2ptp);
+    l3pte  = PTBL_TO_PTADDR(l3ptbl) + ELAN3_L3_INDEX(addr)*ELAN3_PTE_SIZE;
+    
+    elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: l3ptbl %p l3pte %lx\n",l3ptbl, l3pte);
+    
+    tl3pte = elan3_readpte (dev, l3pte);
+    switch (elan3mmu_lock_ptbl (l3ptbl, LK_PTBL_FAILOK, elan3mmu, addr, PTBL_LEVEL_3, &l3lock, &flags))
+    {
+    case LK_PTBL_OK:
+       elan3mmu_unlock_ptbl (l3ptbl, l3lock, flags);
+       elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: l3pte matches value %llx\n", (long long) tl3pte);
+       break;
+       
+    case LK_PTBL_FAILED:
+       panic ("elan3mmu_display: l3 lock failed");
+       /* NOTREACHED */
+       
+    case LK_PTBL_MISMATCH:
+       elan3_debugf (NULL, DBG_HAT, "elan3mmu_display: PTBL_MISMATCH : ptbl %p flags %x elan3mmu %p base %x (%p %x) %llx\n",
+                    l3ptbl, l3ptbl->ptbl_flags, l3ptbl->ptbl_elan3mmu, l3ptbl->ptbl_base, elan3mmu, addr, (long long) tl3pte);
+       
+       break;
+       
+    default:
+       panic ("elan3mmu_display: elan3mmu_lock_ptbl returned bad value");
+       /* NOTREACHED */
+    }
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/elan3mmu_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/elan3mmu_linux.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/elan3mmu_linux.c       2005-06-01 23:12:54.574442904 -0400
@@ -0,0 +1,284 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elan3mmu_linux.c,v 1.50.2.3 2004/12/14 10:19:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/vm/elan3mmu_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kpte.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+
+/*
+ * Strategy for syncing main <-> elan pte's:
+ * 
+ *   Install callbacks for linux flush_tlb_page(), flush_tlb_range(),
+ *   flush_tlb_all(), and flush_tlb_mm() so when a main PTE changes,
+ *   the elan translations, if any, are invalidated.  They can then be
+ *   faulted in again with the correct physical page, perms, etc., on demand. 
+ * 
+ *   Callbacks are stacked on the mm_struct, one per context.  We also stack
+ *   a ctxt pointer so we don't have to do lookups on every call.
+ *
+ *   Sanity check -- we clearly want to flush the elan PTEs in these 
+ *   situations, all of which are covered by tlb_flush_{page,range}()
+ *
+ *     1) kernel/vmscan.c::try_to_swap_out() swaps out a page
+ *
+ *     2) kernel/mremap.c::copy_one_pte() moves a page as a result of the 
+ *     mremap system call
+ * 
+ *     3) kernel/mprotect.c::change_pte_range() changes the permissions of a 
+ *     page as the result of the mprotect system call
+ *
+ * Other Notes: 
+ * 
+ *   Dirty a page in the mains page tables when it is faulted into the elan.
+ *   This way it will not be thrown away by the swapper.
+ * 
+ *   Pages write protected for COW are copied by elan3mmu_main_pagefault()
+ *   when a writeable translation is loaded into the elan.
+ */
+
+caddr_t             elan3mmu_kernel_invalid_space;
+ELAN3_PTE     elan3mmu_kernel_invalid_pte_val;
+
+void
+elan3mmu_init_osdep (void)
+{
+    pte_t *pte;
+
+    KMEM_GETPAGES (elan3mmu_kernel_invalid_space, caddr_t, 1, TRUE);
+
+    ASSERT(elan3mmu_kernel_invalid_space != NULL);
+
+    pte = find_pte_kernel ((unsigned long) elan3mmu_kernel_invalid_space);
+
+    elan3mmu_kernel_invalid_pte_val = ELAN3_PTE_64_BIT | (pte_phys(*pte) & ELAN3_PTE_PFN_MASK) | ELAN3_PERM_REMOTEREAD | ELAN3_ET_PTE;
+
+#ifdef __alpha
+    /*
+     * NOTE: Elan sign-extends bit 48 of the physical address, so if we need to
+     *       set any of bits 63:48, then we will set them all by setting bit 48/
+     */
+    if (alpha_mv.pci_dac_offset & 0xFFFF000000000000ull)
+        elan3mmu_kernel_invalid_pte_val |= (1ull << 48);
+    else
+       elan3mmu_kernel_invalid_pte_val |= alpha_mv.pci_dac_offset;
+#endif
+
+    HAT_PRINTF(0x10, "elan3mmu_invalid_space at %p phys=%llx pte=%llx\n", elan3mmu_kernel_invalid_space, 
+              (unsigned long long) pte_phys(*pte), (unsigned long long) elan3mmu_kernel_invalid_pte_val);
+}
+
+void
+elan3mmu_fini_osdep()
+{
+    KMEM_FREEPAGES (elan3mmu_kernel_invalid_space, 1);
+}
+
+void
+elan3mmu_alloc_osdep (ELAN3MMU *elan3mmu)
+{
+    elan3mmu->elan3mmu_coproc_mm = current->mm;
+}
+
+/*
+ * Convert physical page frame number to elan pte.
+ */
+ELAN3_PTE
+elan3mmu_phys_to_pte (ELAN3_DEV *dev, physaddr_t paddr, int perm)
+{
+    ELAN3_PTE newpte;
+    
+    ASSERT (paddr != 0);
+    
+    if ((paddr & dev->SdramPhysMask) == dev->SdramPhysBase)            /* SDRAM, turn on PTE_LOCAL bit */
+    {
+       PRINTF(NULL, DBG_HAT, "elan3mmu_phys_to_pte: phys %llx SDRAM\n", (unsigned long long) paddr);
+       
+       newpte = ELAN3_PTE_LOCAL | (paddr & ELAN3_PTE_PFN_MASK & ~dev->SdramPhysMask) | perm | ELAN3_ET_PTE;
+    }
+#if defined(LINUX_ALPHA)
+    else if ((paddr & dev->PciPhysMask) == dev->PciPhysBase)
+    {
+       PRINTF(NULL, DBG_HAT, "elan3mmu_phys_to_pte: phys %llx PCI\n", (unsigned long long) paddr);
+       newpte = ELAN3_PTE_64_BIT | (paddr & ELAN3_PTE_PFN_MASK & ~dev->PciPhysMask) | perm | ELAN3_ET_PTE;
+    }
+#endif
+    else                                               /* main memory, must convert to PCI view */
+    {
+       PRINTF(NULL, DBG_HAT, "elan3mmu_phys_to_pte: phys %llx is main memory\n", (unsigned long long) paddr);
+
+       /* main memory, just set the architecture specific PTE_BYPASS bit */
+       /* This requires the Tsunami chipset being programmed to support
+        * the monster window option. This is in linux-2.4.5 and later kernels 
+        * and is also patched into the RH 7.1/2.4.3-12 Alpha kernel
+        */
+       newpte = ELAN3_PTE_64_BIT | (paddr & ELAN3_PTE_PFN_MASK) | perm | ELAN3_ET_PTE;
+
+#ifdef __alpha
+       /*
+        * NOTE: Elan sign-extends bit 48 of the physical address, so if we need to
+        *       set any of bits 63:48, then we will set them all by setting bit 48/
+        */
+       if (alpha_mv.pci_dac_offset & 0xFFFF000000000000ull)
+            newpte |= (1ull << 48);
+        else
+           newpte |= alpha_mv.pci_dac_offset;
+#endif
+    }
+
+    if ( ELAN3_PERM_WRITEABLE( perm )) 
+       newpte |= ( ELAN3_PTE_MOD | ELAN3_PTE_REF );
+    else
+       newpte |= ( ELAN3_PTE_REF ) ;
+
+    return (newpte);
+}
+
+ELAN3_PTE
+elan3mmu_kernel_invalid_pte (ELAN3MMU *elan3mmu)
+{
+    if (elan3mmu->elan3mmu_dev->Devinfo.dev_revision_id == PCI_REVISION_ID_ELAN3_REVB)
+       return (elan3mmu_kernel_invalid_pte_val);
+    return (ELAN3_INVALID_PTE);
+}
+
+/* 
+ * Invalidate a range of addresses for specified context.
+ */
+void
+elan3mmu_pte_range_unload (ELAN3MMU *elan3mmu, struct mm_struct *mm, caddr_t addr, unsigned long len)
+{
+    E3_Addr       eaddr;
+    ELAN3MMU_RGN *rgn;
+    unsigned long span;
+
+    spin_lock (&elan3mmu->elan3mmu_lock);
+
+    for (; len; len -= span, addr += span)
+    {
+       rgn = elan3mmu_findrgn_main (elan3mmu, addr, 0);
+
+       if (rgn == NULL || (rgn->rgn_mbase + rgn->rgn_len) < addr)
+           span = len;
+       else if (rgn->rgn_mbase > addr)
+           span = MIN(len, rgn->rgn_mbase - addr);
+       else
+       {
+           span  = MIN(len, (rgn->rgn_mbase + rgn->rgn_len) - addr);
+           eaddr = rgn->rgn_ebase + (addr - rgn->rgn_mbase);
+
+            HAT_PRINTF(0x10, "  unloading eaddr %x main %p (%ld pages)\n", 
+             eaddr, addr, btopr(span));
+           elan3mmu_unload (elan3mmu, eaddr, span, PTE_UNLOAD);
+       }                       /* takes care of elan tlb flush also */
+    }
+
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+}
+
+/*
+ *
+ */
+void
+elan3mmu_update_range (ELAN3MMU *elan3mmu, struct mm_struct *mm, caddr_t vaddr, E3_Addr eaddr, u_int len, u_int perm)
+{
+    u_int roperm = ELAN3_PERM_READONLY(perm & ELAN3_PTE_PERM_MASK) | (perm & ~ELAN3_PTE_PERM_MASK);
+    u_int off;
+
+    HAT_PRINTF3(1, "elan3mmu_update_range (elan3mmu %p addr %p -> %p)\n", elan3mmu, vaddr, vaddr+len-1);
+
+    while (len > 0)
+    {
+       pte_t *pte_ptr;
+       pte_t  pte_value;
+
+       pte_ptr = find_pte_map(mm, (unsigned long)vaddr);
+       if (pte_ptr) {
+           pte_value = *pte_ptr;
+           pte_unmap(pte_ptr);
+       }
+
+       HAT_PRINTF(0x10, "  elan3mmu_update_range %x (%p) %s\n", eaddr, vaddr, 
+               !pte_ptr ? "invalid" : pte_none(pte_value) ? "none " : !pte_present(pte_value) ? "swapped " : 
+               !pte_write(pte_value) ? "RO/COW" : "OK");
+       
+       if (pte_ptr && !pte_none(pte_value) && pte_present(pte_value))
+           for (off = 0; off < PAGE_SIZE; off += ELAN3_PAGE_SIZE)
+               elan3mmu_pteload (elan3mmu, PTBL_LEVEL_3, eaddr + off, pte_phys(pte_value) + off, pte_write(pte_value) ? perm : roperm, PTE_LOAD|PTE_NO_SLEEP|PTE_NO_STEAL);
+       vaddr += PAGESIZE;
+       eaddr += PAGESIZE;
+       len   -= PAGESIZE;
+    }
+}
+
+/* 
+ * Update a range of addresses for specified context.
+ */
+void
+elan3mmu_pte_range_update (ELAN3MMU *elan3mmu, struct mm_struct *mm,caddr_t vaddr, unsigned long len)
+{
+    E3_Addr       eaddr;
+    ELAN3MMU_RGN *rgn;
+    unsigned long span;
+
+    spin_lock (&elan3mmu->elan3mmu_lock);
+
+    for (; len; len -= span, vaddr += span)
+    {
+       rgn = elan3mmu_findrgn_main (elan3mmu, vaddr, 0);
+
+       if (rgn == NULL || (rgn->rgn_mbase + rgn->rgn_len) < vaddr)
+           span = len;
+       else if (rgn->rgn_mbase > vaddr)
+           span = MIN(len, rgn->rgn_mbase - vaddr);
+       else
+       {
+           span  = MIN(len, (rgn->rgn_mbase + rgn->rgn_len) - vaddr);
+           eaddr = rgn->rgn_ebase + (vaddr - rgn->rgn_mbase);
+
+            HAT_PRINTF(0x10, "  updating eaddr %u main %p (%ld pages)\n", 
+             eaddr, vaddr, btopr(span));
+           
+           elan3mmu_update_range(elan3mmu, mm, vaddr, eaddr, span, rgn->rgn_perm);
+       }                       
+    }
+
+    spin_unlock (&elan3mmu->elan3mmu_lock);
+}
+
+/* 
+ * Invalidate all ptes for the given context.
+ */
+void
+elan3mmu_pte_ctxt_unload(ELAN3MMU *elan3mmu)
+{
+    ELAN3_PTBL  *l1ptbl   = (elan3mmu ? elan3mmu->elan3mmu_l1ptbl : NULL);
+    spinlock_t *l1mtx;
+    unsigned long flags;
+
+    if (l1ptbl && elan3mmu_lock_ptbl (l1ptbl, LK_PTBL_FAILOK, elan3mmu, (E3_Addr) 0, 1, &l1mtx, &flags) == LK_PTBL_OK)
+    {
+       elan3mmu_l1inval(elan3mmu, elan3mmu->elan3mmu_l1ptbl, 0);
+       elan3mmu_unlock_ptbl (l1ptbl, l1mtx, flags);
+    }
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/elan3ops.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/elan3ops.c        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/elan3ops.c     2005-06-01 23:12:54.575442752 -0400
@@ -0,0 +1,170 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elan3ops.c,v 1.4 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/elan3ops.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan/elanmod.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elan3ops.h>
+
+extern ELAN_STATS_OPS elan3_device_stats_ops;
+
+ELAN_DEV_OPS elan3_dev_ops = {
+
+       get_position,
+       set_position,   
+
+       ELAN_DEV_OPS_VERSION
+};
+
+ELAN_STATS_OPS elan3_device_stats_ops = {
+       ELAN_STATS_OPS_VERSION,
+
+       stats_get_index_name,
+       stats_get_block,
+       stats_clear_block
+};
+
+static char *elan3_device_stats_names[ELAN3_NUM_STATS] = 
+{
+               "version field",                 /* not cleared */
+               "elan interrupts",
+               "tlb flushes",
+               "traps with invalid context",
+               "interrupts com queue half full",
+               "cproc traps",
+               "dproc traps",
+               "tproc traps",
+               "iproc traps",
+               "event interrupts",
+               "elan page faults",
+               "EopBadAcks",
+               "EopResets",
+               "InputterBadLength",
+               "InputterCRCDiscards",
+               "InputterCRCErrors",
+               "InputterCRCBad",       
+               "errors in dma data",
+               "errors after dma identify",
+               "errors after thread identify",
+               "dma retries",
+               "dma output timeouts",
+               "dma packet ack errors",
+               "forced tproc traps",
+               "too many instruction traps",
+               "output timeouts",
+               "packet ack errors",
+               "LockError",
+               "DeskewError",
+               "PhaseError",
+               "DataError",
+               "FifoOvFlow0",
+               "FifoOvFlow1",
+               "link error value on data error",
+               "correctable ecc errors",
+               "uncorrectable ecc errors",
+               "multiple ecc errors",
+               "sdram bytes free",              /* not cleared */
+               "longest interrupt in ticks",
+               "punts of event int's to thread",
+               "reschedules of event int's thread"
+};
+
+int 
+stats_get_index_name (void *arg, uint  index, caddr_t name)
+{
+       copyout (elan3_device_stats_names[index], name, strlen (elan3_device_stats_names[index]) + 1  /* with \0 */);
+
+       return (0);
+}
+
+int
+stats_get_block (void *arg, uint entries, ulong *value)
+{
+       ELAN3_DEV *dev = (ELAN3_DEV *) arg;
+
+       if ( entries >  ELAN3_NUM_STATS ) /* if space too big only send valid portion */
+               entries = ELAN3_NUM_STATS;
+       
+       copyout(&dev->Stats, value, sizeof(ulong) * entries);
+
+       return (0);
+}
+
+int 
+stats_clear_block (void *arg)
+{
+       ELAN3_DEV *dev = (ELAN3_DEV *) arg;
+       u_long   *ptr = (u_long *) &dev->Stats;
+       int                n;
+       
+       for (n = 0; n < ELAN3_NUM_STATS; n++)
+       {
+               switch (n) 
+               {
+               case offsetof (ELAN3_STATS, Version)/sizeof(u_long):
+               case offsetof (ELAN3_STATS, SdramBytesFree)/sizeof(u_long):
+                       break;
+               default:
+                       ptr[n] = (ulong)0;
+               }
+       }
+       return (0);
+}
+
+int 
+get_position (void *user_data, ELAN_POSITION *position)
+{
+       ELAN3_DEV *dev = (ELAN3_DEV *)user_data;
+
+       copyout(&dev->Position, position, sizeof(ELAN_POSITION));
+
+       return (0);     
+}
+
+int 
+set_position (void *user_data, unsigned short nodeId, unsigned short numNodes)
+{
+       ELAN3_DEV *dev = (ELAN3_DEV *)user_data;
+
+       if (ComputePosition (&dev->Position, nodeId, numNodes, dev->Devinfo.dev_num_down_links_value) != 0)
+               return (EINVAL);
+       
+       return (0);     
+}
+
+int
+elan3_register_dev_stats(ELAN3_DEV * dev) 
+{
+       char name[ELAN_STATS_NAME_MAX_LEN+1];
+
+       sprintf (name, ELAN3_STATS_DEV_FMT, dev->Instance);
+
+       elan_stats_register(&dev->StatsIndex,
+                              name,
+                              sizeof (elan3_device_stats_names)/sizeof (elan3_device_stats_names[0]),
+                              &elan3_device_stats_ops,
+                              (void *)dev);
+
+       return (0);
+}
+
+void
+elan3_deregister_dev_stats(ELAN3_DEV * dev) 
+{
+       elan_stats_deregister(dev->StatsIndex);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/elandebug.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/elandebug.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/elandebug.c    2005-06-01 23:12:54.575442752 -0400
@@ -0,0 +1,151 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elandebug.c,v 1.25 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/elandebug.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+
+
+void
+elan3_debugf (void *p, unsigned int mode, char *fmt,...)
+{
+    char prefix[128];
+
+#if defined (DIGITAL_UNIX)
+#define PREFIX_FMT     "[%lx.%08x]"
+#define PREFIX_VAL     (int)CURTHREAD()
+#else
+#define PREFIX_FMT     "[%lx.%04d]"
+#define PREFIX_VAL     (current->pid)
+#endif
+
+    if ((unsigned long) p > DBG_NTYPES)
+    {
+       ELAN3_CTXT *ctxt = (ELAN3_CTXT *) p;
+
+        if (elan3_debug_display_ctxt && (ctxt->Capability.cap_mycontext & MAX_ROOT_CONTEXT_MASK) != elan3_debug_display_ctxt)
+            return;
+        if (elan3_debug_ignore_ctxt  && (ctxt->Capability.cap_mycontext & MAX_ROOT_CONTEXT_MASK) == elan3_debug_ignore_ctxt)
+            return;
+ 
+       if (ctxt->Capability.cap_mycontext == ELAN_CAP_UNINITIALISED)
+           sprintf (prefix, PREFIX_FMT " (XXX) ", lbolt, PREFIX_VAL);
+       else
+           sprintf (prefix, PREFIX_FMT " (%03x) ", lbolt, PREFIX_VAL,
+                    ctxt->Capability.cap_mycontext & MAX_ROOT_CONTEXT_MASK);
+    }
+    else
+    {
+       char *what;
+
+       if (elan3_debug_ignore_dev & (1 << ((unsigned long) p)))
+           return;
+
+       switch ((unsigned long) p)
+       {
+       case (int) DBG_DEVICE: what = "dev"; break;
+       case (int) DBG_KCOMM:  what = "kcm"; break;
+       case (int) DBG_ICS:    what = "ics"; break;
+       case (int) DBG_USER:   what = "usr"; break;
+       default:               what = NULL; break;
+       }
+           
+       if (what)
+           sprintf (prefix, PREFIX_FMT " [%s] ", lbolt,  PREFIX_VAL, what);
+       else
+           sprintf (prefix, PREFIX_FMT " [%3d] ", lbolt,  PREFIX_VAL, (int)(long)what);
+    }
+
+    {
+       va_list       ap;
+
+       va_start (ap, fmt);
+       qsnet_vdebugf ((((mode & elan3_debug_buffer)?QSNET_DEBUG_BUFFER:0)|((mode & elan3_debug_console)?QSNET_DEBUG_CONSOLE:0)) , prefix, fmt, ap);
+       va_end (ap);
+    }
+}
+
+
+void
+elan3_alloc_panicstate (ELAN3_DEV *dev, int allocsdram)
+{
+    register int bank;
+
+    if (dev->PanicState.RegPtr == NULL)
+       KMEM_ZALLOC (dev->PanicState.RegPtr, E3_Regs *, sizeof (E3_Regs), 1);
+
+    if (allocsdram)
+       for (bank = 0; bank < ELAN3_SDRAM_NUM_BANKS; bank++)
+           if (dev->PanicState.Sdram[bank] == NULL && dev->SdramBanks[bank].Size)
+               KMEM_ZALLOC (dev->PanicState.Sdram[bank], char *, dev->SdramBanks[bank].Size, 1);
+}
+
+void
+elan3_free_panicstate (ELAN3_DEV *dev)
+{
+    register int bank;
+
+    if (dev->PanicState.RegPtr != NULL)
+       KMEM_FREE (dev->PanicState.RegPtr, sizeof (E3_Regs));
+
+    for (bank = 0; bank < ELAN3_SDRAM_NUM_BANKS; bank++)
+       if (dev->PanicState.Sdram[bank] != NULL && dev->SdramBanks[bank].Size)
+           KMEM_FREE (dev->PanicState.Sdram[bank], dev->SdramBanks[bank].Size);
+
+    bzero (&dev->PanicState, sizeof (dev->PanicState));
+}
+
+void
+elan3_save_panicstate (ELAN3_DEV *dev)
+{
+    register int bank;
+    
+    if (dev->PanicState.RegPtr)
+    {
+       printk ("elan%d: saving state on panic .....\n", dev->Devinfo.dev_instance);
+
+       bcopy ((void *) dev->RegPtr, (void *) dev->PanicState.RegPtr, sizeof (E3_Regs));
+       
+       for (bank = 0; bank < ELAN3_SDRAM_NUM_BANKS; bank++)
+           if (dev->SdramBanks[bank].Size && dev->PanicState.Sdram[bank])
+               elan3_sdram_copyq_from_sdram (dev, (bank << ELAN3_SDRAM_BANK_SHIFT), dev->PanicState.Sdram[bank], dev->SdramBanks[bank].Size);
+       
+    }
+}
+
+int
+elan3_assfail (ELAN3_DEV *dev, char *string, char *file, int line)
+{
+    if (panicstr)
+       return (0);
+
+    printk ("elan: assertion failed '%s' File '%s' Line %d\n", string, file, line);
+
+#if defined(LINUX)
+    elan3_save_panicstate (dev);
+
+    panic ("elan: assertion failed '%s' File '%s' Line %d\n", string, file, line);
+#else
+    cmn_err (CE_PANIC, "elan: assertion failed '%s' File '%s' Line %d\n", string, file, line);
+#endif
+    /*NOTREACHED*/
+    return (0);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/elandev_generic.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/elandev_generic.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/elandev_generic.c      2005-06-01 23:12:54.578442296 -0400
@@ -0,0 +1,1862 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elandev_generic.c,v 1.111.2.3 2004/11/15 11:12:36 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/elandev_generic.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+
+#include <elan3/dma.h>
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/elansyscall.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/elan3ops.h>
+
+/*
+ * Module globals, configurable from system file.
+ */
+u_int  elan3_debug                  = 0;
+u_int  elan3_debug_console       = 0;
+u_int  elan3_debug_buffer           = -1;
+u_int  elan3_debug_ignore_dev       = 0;
+u_int  elan3_debug_ignore_kcomm     = 0;
+u_int  elan3_debug_ignore_ctxt      = 0;
+u_int  elan3_debug_display_ctxt     = 0;
+
+int    eventint_punt_loops;
+int    eventint_punt_ticks;
+int    eventint_resched_ticks;
+
+static void InitialiseDmaBuffers (ELAN3_DEV *dev, ioaddr_t CmdPort);
+static int  ProbeSdram (ELAN3_DEV *dev);
+static void InitialiseSdram (ELAN3_DEV *dev);
+static void ReEnableErrorInterrupts (void *arg);
+void        PollForDmaHungup (void *arg);
+static void elan3_event_interrupt (ELAN3_DEV *dev);
+
+/*
+ * BaseAddr is ptr to the start of a table aligned on a power of two byte address.
+ * SizePower must be in the range of 6 to 12. It defines the number of valid contexts as
+ * shown below.
+ *
+ * SizePower   Valid Contexts  Table size in bytes.
+ *     6            64               1k
+ *     7           128               2k
+ *     8           256               4K
+ *     9           512               8k
+ *    10          1024              16k
+ *    11          2048              32k
+ *    12          4096              64k
+ */
+#define GEN_CONTEXT_PTR(BaseAddr, SizePower) (((E3_uint32) BaseAddr) | \
+                                             (~((1 << ((SizePower) - 6)) - 1) & 0x3f))
+
+int
+InitialiseElan (ELAN3_DEV *dev, ioaddr_t CmdPort)
+{
+    E3_IprocTrapHeader_BE   TrapCleanup[4];
+    E3_ContextControlBlock  ContextControlBlock;
+    sdramaddr_t             ptr;
+    int                            res;
+    int                            i;
+
+    eventint_punt_loops    = 100;
+    eventint_punt_ticks    = (hz/100);
+    eventint_resched_ticks = (hz/4);
+    
+    dev->Stats.Version     = ELAN3_STATS_VERSION;
+    dev->Position.pos_mode = ELAN_POS_UNKNOWN;
+
+    /*
+     * The elan should have already been reset, so the interrupt mask
+     * should be 0 and the schedule status register should be set to
+     * its initial state
+     */
+    ASSERT (dev->InterruptMask == 0);
+    ASSERT ((read_reg32 (dev, Exts.SchCntReg) & HaltStopAndExtTestMask) == Sched_Initial_Value);
+
+    /*
+     * Write any value here to clear out the half full and error bits of the command
+     * overflow queues.
+     */
+    write_reg32 (dev, ComQueueStatus, 0);
+
+    /* Initialise the cache tags before touching the SDRAM */
+    /* we initialise them to "map" the bottom of SDRAM */
+    for (i = 0; i < E3_NumCacheLines; i++)
+    {
+       write_cache_tag (dev, Tags[i][0].Value, 0x0000000000000000ULL);
+       write_cache_tag (dev, Tags[i][1].Value, 0x0000080000000000ULL);
+       write_cache_tag (dev, Tags[i][2].Value, 0x0000100000000000ULL);
+       write_cache_tag (dev, Tags[i][3].Value, 0x0000180000000000ULL);
+    }
+
+#ifndef CONFIG_MPSAS
+    for (i = 0; i < E3_NumCacheLines*(E3_CACHELINE_SIZE/sizeof(E3_uint64)); i++)
+    {
+       write_cache_set (dev, Set0[i], 0xcac1ecac1ecac1e0ULL);
+       write_cache_set (dev, Set1[i], 0xcac1ecac1ecac1e1ULL);
+       write_cache_set (dev, Set2[i], 0xcac1ecac1ecac1e2ULL);
+       write_cache_set (dev, Set3[i], 0xcac1ecac1ecac1e3ULL);
+    }
+#endif
+
+    if ((res = ProbeSdram(dev)) != ESUCCESS)
+       return (res);
+
+    /* Enable all cache sets before initialising the sdram allocators */
+    write_reg32 (dev, Cache_Control_Reg.ContReg, (dev->Cache_Control_Reg |= CONT_EN_ALL_SETS));
+
+    InitialiseSdram (dev);
+
+    dev->TAndQBase              = elan3_sdram_alloc (dev, ELAN3_TANDQ_SIZE);
+    dev->ContextTable           = elan3_sdram_alloc (dev, ELAN3_CONTEXT_SIZE);
+    dev->ContextTableSize       = ELAN3_NUM_CONTEXTS;
+    dev->CommandPortTraps[0]    = elan3_sdram_alloc (dev, ELAN3_COMMAND_TRAP_SIZE);
+    dev->CommandPortTraps[1]    = elan3_sdram_alloc (dev, ELAN3_COMMAND_TRAP_SIZE);
+    dev->CurrentCommandPortTrap = 0;
+
+    PRINTF3 (DBG_DEVICE, DBG_CONFIG, "InitialiseElan: ContextTable %08lx TAndQ %08lx CommandPortTrap %08lx\n",
+            dev->ContextTable, dev->TAndQBase, dev->CommandPortTraps[0]);
+
+    /* Allocate the thread amd dma trap areas */
+    KMEM_ZALLOC (dev->ThreadTrap, THREAD_TRAP *, sizeof (THREAD_TRAP), TRUE);
+    KMEM_ZALLOC (dev->DmaTrap, DMA_TRAP *, sizeof (DMA_TRAP), TRUE);
+
+    /* Allocate the ctxt table */
+    KMEM_ZALLOC (dev->CtxtTable,  ELAN3_CTXT **, dev->ContextTableSize * sizeof ( ELAN3_CTXT *), TRUE);
+
+    /* Initialise halt queue list */
+    dev->HaltOperationsTailpp   = &dev->HaltOperations;
+
+    /* From elan3/code/harness/elanstuff.c */
+    /* Init the clock. */
+    write_ureg64 (dev, Clock.NanoSecClock, 0);
+    
+    /* Init the instruction count reg. */
+    write_ureg32 (dev, InstCount.s.StatsCount, 0);
+    
+    /* Init the stats control reg. Must be done before the count regs.*/
+    write_ureg32 (dev, StatCont.StatsControl, 0);
+    
+    /* Init the stats count regs. */
+    write_ureg32 (dev, StatCounts[0].s.StatsCount, 0);
+    write_ureg32 (dev, StatCounts[1].s.StatsCount, 0);
+    write_ureg32 (dev, StatCounts[2].s.StatsCount, 0);
+    write_ureg32 (dev, StatCounts[3].s.StatsCount, 0);
+    write_ureg32 (dev, StatCounts[4].s.StatsCount, 0);
+    write_ureg32 (dev, StatCounts[5].s.StatsCount, 0);
+    write_ureg32 (dev, StatCounts[6].s.StatsCount, 0);
+    write_ureg32 (dev, StatCounts[7].s.StatsCount, 0);
+    
+    /*
+     * Initialise the Context_Ptr and Fault_Base_Ptr
+     */
+    write_reg32 (dev, Fault_Base_Ptr, dev->TAndQBase + offsetof(E3_TrapAndQueue, IProcSysCntx));
+    write_reg32 (dev, Context_Ptr, GEN_CONTEXT_PTR (dev->ContextTable, ELAN3_LN2_NUM_CONTEXTS));
+
+    /* scrub the TProc Registers */
+    for (i = 0; i < 8; i++)
+       write_reg32 (dev, Globals[i], 0xdeadbabe);
+    for (i = 0; i < 8; i++)
+       write_reg32 (dev, Outs[i], 0xdeadbabe);
+    for (i = 0; i < 8; i++)
+       write_reg32 (dev, Locals[i], 0xdeadbabe);
+    for (i = 0; i < 8; i++)
+       write_reg32 (dev, Ins[i], 0xdeadbabe);
+
+    /*
+     * Initialise the Queue pointers.  Arrange them so that the starting positions are
+     * farthest apart in one set of the cache. Thus 512 bytes apart,  but with cntx0
+     * thread the same as the interrupt queue.
+     */
+    write_reg32 (dev, TProc_NonSysCntx_FPtr, dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxThreadQueue[0xc0]));
+    write_reg32 (dev, TProc_NonSysCntx_BPtr, dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxThreadQueue[0xc0]));
+    write_reg32 (dev, TProc_SysCntx_FPtr,    dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxThreadQueue[0x80]));
+    write_reg32 (dev, TProc_SysCntx_BPtr,    dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxThreadQueue[0x80]));
+    
+    write_reg32 (dev, DProc_NonSysCntx_FPtr, dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxDmaQueue[0]));
+    write_reg32 (dev, DProc_NonSysCntx_BPtr, dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxDmaQueue[0]));
+    write_reg32 (dev, DProc_SysCntx_FPtr,    dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[0x10]));
+    write_reg32 (dev, DProc_SysCntx_BPtr,    dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[0x10]));
+    
+    dev->Event_Int_Queue_FPtr = dev->TAndQBase + offsetof (E3_TrapAndQueue, EventIntQueue[0x80]);
+    write_reg32 (dev, Event_Int_Queue_FPtr, dev->Event_Int_Queue_FPtr);
+    write_reg32 (dev, Event_Int_Queue_BPtr, dev->TAndQBase + offsetof (E3_TrapAndQueue, EventIntQueue[0x80]));
+    
+    
+    /* Initialise Input_Trap_Base to last 8 Kbytes of trap area, uCode adds the right offset */
+    write_reg32 (dev, Input_Trap_Base, dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxThreadQueue[0]));
+    
+    /* Ptr to word used to save the SP to when a thread deschedules */
+    write_reg32 (dev, Thread_SP_Save_Ptr, dev->TAndQBase + offsetof (E3_TrapAndQueue, Thread_SP_Save));
+    
+    /* Initialise the command trap base */
+    write_reg32 (dev, CProc_TrapSave_Addr, dev->CommandPortTraps[0]);
+    
+    /* Initialise the set event tracing registers */
+    write_reg32 (dev, Event_Trace_Ptr, 0);
+    write_reg32 (dev, Event_Trace_Mask, 0);
+    
+    /* Initialise Tlb_Line_Value to zero. The TLB cannot be read while either the */
+    /* uCode or thread proc might be running. Must be set to 0. */
+    write_reg64 (dev, Tlb_Line_Value, 0);
+
+    /* Control register. Cache everything, Enable MMU, RefreshRate=3, CasLatency=1, StartSDR */
+    dev->Cache_Control_Reg |= CONT_MMU_ENABLE | CONT_EN_ALL_SETS | CONT_CACHE_ALL | CONT_ENABLE_ECC;
+
+#if ELAN3_PAGE_SHIFT == 13
+    dev->Cache_Control_Reg |= CONT_ENABLE_8K_PAGES;
+#endif
+
+    write_reg32 (dev, Cache_Control_Reg.ContReg,  dev->Cache_Control_Reg);
+
+    /*
+     * Initialise the context table to be discard for all contexts
+     */
+    ContextControlBlock.rootPTP  = 0;
+    ContextControlBlock.filter   = E3_CCB_DISCARD_ALL;
+    ContextControlBlock.VPT_mask = 0;
+    ContextControlBlock.VPT_ptr  = 0;
+
+    for (i = 0, ptr = dev->ContextTable; i < ELAN3_NUM_CONTEXTS; i++, ptr += sizeof (E3_ContextControlBlock))
+       elan3_sdram_copyl_to_sdram (dev, &ContextControlBlock, ptr, sizeof (E3_ContextControlBlock));
+
+    /* From elan3/code/trap_handler/init.c */
+    /*
+     * Initialise the Trap And Queue area in Elan SDRAM.
+     */
+    TrapCleanup[0].s.TrTypeCntx.TypeContext = 0;
+    TrapCleanup[0].s.TrAddr                = 0;
+    TrapCleanup[0].s.IProcTrapStatus.Status = CRC_STATUS_GOOD;
+    TrapCleanup[0].s.TrData0               = 0;
+    TrapCleanup[1].s.TrTypeCntx.TypeContext = 0;
+    TrapCleanup[1].s.TrAddr                = 0;
+    TrapCleanup[1].s.IProcTrapStatus.Status = CRC_STATUS_GOOD;
+    TrapCleanup[1].s.TrData0               = 0;
+    TrapCleanup[2].s.TrTypeCntx.TypeContext = 0;
+    TrapCleanup[2].s.TrAddr                = 0;
+    TrapCleanup[2].s.IProcTrapStatus.Status = CRC_STATUS_GOOD;
+    TrapCleanup[2].s.TrData0               = 0;
+    TrapCleanup[3].s.TrTypeCntx.TypeContext = 0;
+    TrapCleanup[3].s.TrAddr                = 0;
+    TrapCleanup[3].s.IProcTrapStatus.Status = CRC_STATUS_GOOD;
+    TrapCleanup[3].s.TrData0               = 0;
+
+    elan3_sdram_writel (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcSysCntx.s.FaultContext),  0);
+    elan3_sdram_writel (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcSysCntx.s.FSR.Status), 0);
+    elan3_sdram_writel (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcNonSysCntx.s.FaultContext), 0);
+    elan3_sdram_writel (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcNonSysCntx.s.FSR.Status), 0);
+    
+    /* Must now zero all the FSRs so that a subsequent Fault can be seen */ 
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, CProc), 16);
+
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProc), 16);
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData0), 64);
+    
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, TProc), 16);
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcData), 16);
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcInst), 16);
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcOpen), 16);
+
+    elan3_sdram_copyq_to_sdram (dev, TrapCleanup, dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh0_C0_TrHead[0]), 64);
+    elan3_sdram_copyq_to_sdram (dev, TrapCleanup, dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh1_C0_TrHead[0]), 64);
+
+    elan3_sdram_copyq_to_sdram (dev, TrapCleanup, dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh0_NonC0_TrHead[0]), 64);
+    elan3_sdram_copyq_to_sdram (dev, TrapCleanup, dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh1_NonC0_TrHead[0]), 64);
+
+    InitialiseDmaBuffers(dev, CmdPort);
+
+    /* reserve a halt operation for flushing the context filter */
+    ReserveHaltOperations (dev, 1, TRUE);
+
+    /* Allow the Thread/Dma to run */
+    CLEAR_SCHED_STATUS (dev, HaltThread | HaltDmas);
+
+    /* Enable All Interrrupts */
+    SET_INT_MASK (dev, (INT_PciMemErr | INT_SDRamInt | INT_EventInterrupt | INT_LinkError | INT_ComQueue |
+                       INT_TProc | INT_CProc | INT_DProc | INT_IProcCh1NonSysCntx | 
+                       INT_IProcCh1SysCntx | INT_IProcCh0NonSysCntx | INT_IProcCh0SysCntx));
+
+    /* Take the link out of boundary scan */
+    SET_SCHED_LINK_VALUE (dev, 0, 0);
+    
+    /* And clear any link errors */
+    PULSE_SCHED_STATUS (dev, ClearLinkErrorInt);
+
+    /* XXXX: clear discard context 0,  AFTER setting up the kernel comms */
+    CLEAR_SCHED_STATUS (dev, DiscardSysCntxIn | DiscardNonSysCntxIn);
+
+    /* Start a thread to handle excessive Event Interrrupts */
+    if (kernel_thread_create (elan3_event_interrupt, (caddr_t) dev) == NULL)
+    {
+       panic ("InitialiseElan: cannot start elan3_event_interrupt\n");
+       return (EFAIL);
+    }
+    dev->EventInterruptThreadStarted = 1;
+
+    ReserveHaltOperations (dev, 1, TRUE);
+
+    PollForDmaHungup (dev);
+
+    /* register the device and stats with elanmod for RMS */
+    dev->DeviceIdx = elan_dev_register(&dev->Devinfo, &elan3_dev_ops, (void *) dev);
+    
+    elan3_register_dev_stats(dev);
+
+    return (ESUCCESS);
+}
+
+static void
+InitialiseDmaBuffers(ELAN3_DEV *dev, ioaddr_t CmdPort)
+{
+   register int i;
+
+   /* GNAT sw-elan3/3908:
+    * Clear down the power on state of the Dma_Desc registers to make sure we don't
+    * try and interpret them when a trap happens.
+    */
+   write_reg32 (dev, Dma_Desc.dma_type,            0);
+   write_reg32 (dev, Dma_Desc.dma_size,            0);
+   write_reg32 (dev, Dma_Desc.dma_source,          0);
+   write_reg32 (dev, Dma_Desc.dma_dest,            0);
+   write_reg32 (dev, Dma_Desc.dma_destEvent,       0);
+   write_reg32 (dev, Dma_Desc.dma_destCookieVProc, 0);
+   write_reg32 (dev, Dma_Desc.dma_srcEvent,        0);
+   write_reg32 (dev, Dma_Desc.dma_srcCookieVProc,  0);
+   
+   /*
+    * The following is a sequence of writes to remove X's from the dma buffers and 
+    * registers. It is only safe to write these registers after reset and before any
+    * dma's have been issued. The chip will NOT function corectly if they are written at
+    * any other time or in a different order.
+    */
+   write_reg64 (dev, Exts.Dmas.DmaWrs.LdAlignment, 0);
+   write_reg64 (dev, Exts.Dmas.DmaWrs.LdDmaType, 0);
+   write_reg64 (dev, Exts.Dmas.DmaWrs.ResetAckNLdBytesToWr, ((u_longlong_t)0x1000) << 32);
+   write_reg64 (dev, Exts.Dmas.DmaWrs.LdBytesToRd, ((u_longlong_t)0x100) << 32);
+
+   for (i=0;i<(4*8);i++)
+       write_reg64 (dev, Dma_Alignment_Port[0], 0);
+
+   /*
+    * This is used to clear out X's from some of the trap registers. This is required to
+    * prevent the first traps from possibly writting X's into the SDram and upsetting the
+    * ECC value. It requires that the trap save area registers have been set up but does
+    * not require any translations to be ready.
+    */
+   writel (-1, CmdPort + offsetof (E3_CommandPort, SetEvent));
+   while ((read_reg32 (dev, Exts.InterruptReg) & INT_CProc) == 0)
+   {
+       mb();
+       DELAY (1);
+   }
+
+   write_reg32 (dev, CProc_TrapSave_Addr, dev->CommandPortTraps[dev->CurrentCommandPortTrap]);
+   
+   PULSE_SCHED_STATUS(dev, RestartCProc);
+}
+
+void
+FinaliseElan (ELAN3_DEV *dev)
+{
+    ELAN3_PTBL_GR *ptg;
+    ELAN3_HALTOP  *op;
+    ELAN3_HALTOP  *chain = NULL;
+    int           bank;
+    int                  indx;
+    int                  size;
+    unsigned long flags;
+    int           level;
+
+    elan_stats_deregister (dev->StatsIndex);
+    elan_dev_deregister(&dev->Devinfo);
+
+    /* Cancel the dma poller */
+    cancel_timer_fn (&dev->DmaPollTimeoutId);
+
+    /* release it's halt operation */
+    ReleaseHaltOperations (dev, 1);
+
+    /* stop all kernel threads */
+    dev->ThreadsShouldStop = 1;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    while (dev->EventInterruptThreadStarted && !dev->EventInterruptThreadStopped)
+    {
+       kcondvar_wakeupall (&dev->IntrWait, &dev->IntrLock);
+       kcondvar_wait (&dev->IntrWait, &dev->IntrLock, &flags);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    /* Set the interrupt mask to 0 and the schedule control register to run nothing */
+    SET_INT_MASK (dev, 0);
+    SET_SCHED_STATUS (dev, DiscardNonSysCntxIn | DiscardSysCntxIn | HaltThread | HaltDmas);
+
+    /* Cancel any link error timeout */
+    if (timer_fn_queued(&dev->ErrorTimeoutId))
+       cancel_timer_fn (&dev->ErrorTimeoutId);
+
+    /* Free of and page tables that have been allocated */
+    spin_lock (&dev->PtblGroupLock);
+    for(level=0; level<4; level++) 
+    {
+       while ((ptg = dev->Level[level].PtblGroupList) != NULL)
+       {
+           dev->Level[level].PtblGroupList = ptg->pg_next;
+
+           elan3_sdram_free (dev, ptg->pg_addr, PTBL_GROUP_SIZE);
+           FREE_PTBL_GR(ptg);
+       }
+    }
+ 
+    spin_unlock (&dev->PtblGroupLock);
+
+    /* Free of all halt operations */
+    spin_lock_irqsave (&dev->FreeHaltLock, flags);
+    while ((op = dev->FreeHaltOperations) != NULL)
+    {
+       dev->FreeHaltOperations = op->Next;
+
+       /* Keep a list of 'freed' ops for later KMEM_FREE call */
+       op->Next = chain;
+       chain = op;
+    }
+    spin_unlock_irqrestore (&dev->FreeHaltLock, flags);
+
+    /* Have now dropped the spinlock - can call KMEM_FREE */
+    while ((op = chain) != NULL)
+    {
+       chain = op->Next;
+
+       KMEM_FREE (op, sizeof (ELAN3_HALTOP));
+    }
+       
+    /* Free of the ctxt table */
+    KMEM_FREE (dev->CtxtTable,  dev->ContextTableSize * sizeof (ELAN3_CTXT *));
+
+    /* Free of the thread and dma atrap areas */
+    KMEM_FREE (dev->ThreadTrap, sizeof (THREAD_TRAP));
+    KMEM_FREE (dev->DmaTrap, sizeof (DMA_TRAP));
+
+    /* Free of the memsegs and pages */
+    for (bank = 0; bank < ELAN3_SDRAM_NUM_BANKS; bank++)
+    {
+       if (dev->SdramBanks[bank].Size)
+       {
+           UnmapDeviceRegister (dev, &dev->SdramBanks[bank].Handle);
+
+           KMEM_FREE (dev->SdramBanks[bank].PtblGroups, sizeof (ELAN3_PTBL_GR *) * (dev->SdramBanks[bank].Size / PTBL_GROUP_SIZE));
+
+           for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size <= dev->SdramBanks[bank].Size; indx++, size <<= 1)
+               KMEM_FREE (dev->SdramBanks[bank].Bitmaps[indx], sizeof (bitmap_t)*BT_BITOUL(dev->SdramBanks[bank].Size/size));
+       }
+    }
+    elan3_sdram_fini (dev);
+}
+
+#define INIT_PATTERN(offset)   (0xBEEC000000000011ull | ((u_longlong_t)(offset)) << 16)
+#define FREE_PATTERN(offset)   (0xBEEC000000000022ull | ((u_longlong_t)(offset)) << 16)
+
+static int
+ProbeSdram (ELAN3_DEV *dev)
+{
+    int                          Instance;
+    u_int                Bank;
+    int                          MemSpaceSize;
+    int                          BankMaxSize;
+    int                          BankOffset;
+    int                          BankSize;
+    ioaddr_t             BankBase;
+    ioaddr_t             PageBase;
+    ioaddr_t             PageBase1;
+    ioaddr_t             PageBase2;
+    DeviceMappingHandle   BankHandle;
+    DeviceMappingHandle   PageHandle;
+    DeviceMappingHandle   PageHandle1;
+    DeviceMappingHandle   PageHandle2;
+    register int          i;
+    u_longlong_t         value;
+    extern int            sdram_bank_limit;
+
+    /* NOTE: The Cache control register is set to only enable cache set 0 */
+    /*       and has ECC disabled */
+    Instance = dev->Instance;
+    
+    /* Determine the size of the SDRAM from the BAR register */
+    if (DeviceRegisterSize (dev, ELAN3_BAR_SDRAM, &MemSpaceSize) != ESUCCESS)
+    {
+       printk ("elan%d: cannot determine SDRAM size\n", Instance);
+       return (EFAIL);
+    }
+
+    elan3_sdram_init (dev);
+
+    BankMaxSize = MemSpaceSize / ELAN3_SDRAM_NUM_BANKS;
+
+    for (Bank = 0; Bank < ELAN3_SDRAM_NUM_BANKS; Bank++)
+    {
+       BankOffset = Bank * BankMaxSize;
+       
+       PRINTF3 (DBG_DEVICE, DBG_CONFIG, "elan%d: Probing RAM Bank %d (max size %08x)\n", Instance, Bank, BankMaxSize);
+       
+       /* Probe the memory bank by mapping two pages that are the size of the cache apart */
+       /* this guarantees that when we store the second pattern we displace the first pattern */
+       /* from the cache, also store the second pattern again the size of the cache up again */
+       /* to ensure that the SDRAM wires don't stay floating at pattern1 */
+
+       if (MapDeviceRegister (dev, ELAN3_BAR_SDRAM, &BankBase, BankOffset, PAGESIZE, &BankHandle) != ESUCCESS)
+       {
+           printk ("elan%d: Cannot probe memory bank %d\n", Instance, Bank);
+           continue;
+       }
+       
+       if (MapDeviceRegister (dev, ELAN3_BAR_SDRAM, &PageBase1, BankOffset + ELAN3_MAX_CACHE_SIZE, PAGESIZE, &PageHandle1) != ESUCCESS)
+       {
+           printk ("elan%d: Cannot probe memory bank %d\n", Instance, Bank);
+           UnmapDeviceRegister (dev, &BankHandle);
+           continue;
+       }
+
+       if (MapDeviceRegister (dev, ELAN3_BAR_SDRAM, &PageBase2, BankOffset + 2*ELAN3_MAX_CACHE_SIZE, PAGESIZE, &PageHandle2) != ESUCCESS)
+       {
+           printk ("elan%d: Cannot probe memory bank %d\n", Instance, Bank);
+           UnmapDeviceRegister (dev, &BankHandle);
+           UnmapDeviceRegister (dev, &PageHandle1);
+           continue;
+       }
+
+#define PATTERN0       (0x5555555555555555L)
+#define PATTERN1       (0xAAAAAAAAAAAAAAAAL)
+       writeq (PATTERN0, (u_longlong_t *) BankBase);
+       writeq (PATTERN1, (u_longlong_t *) PageBase1);
+       writeq (PATTERN1, (u_longlong_t *) PageBase2);
+
+       mmiob();
+
+       value = readq ((u_longlong_t *) BankBase);
+
+       if (value != PATTERN0)
+       {
+           UnmapDeviceRegister (dev, &BankHandle);
+           UnmapDeviceRegister (dev, &PageHandle1);
+           UnmapDeviceRegister (dev, &PageHandle2);
+           continue;
+       }
+
+       writeq (PATTERN1, (u_longlong_t *) BankBase);
+       writeq (PATTERN0, (u_longlong_t *) PageBase1);
+       writeq (PATTERN0, (u_longlong_t *) PageBase2);
+
+       mmiob();
+       
+       value = readq ((u_longlong_t *) BankBase);
+       if (value != PATTERN1)
+       {
+           UnmapDeviceRegister (dev, &BankHandle);
+           UnmapDeviceRegister (dev, &PageHandle1);
+           UnmapDeviceRegister (dev, &PageHandle2);
+           continue;
+       }
+       UnmapDeviceRegister (dev, &PageHandle1);
+       UnmapDeviceRegister (dev, &PageHandle2);
+
+       /* Bank is present, so work out its size,  we store tha maximum size at the base */
+       /* and then store the address at each address  on every power of two address until */
+       /* we reach the minimum mappable size (PAGESIZE), we then read back the value at the */
+       /* base to determine the bank size */
+       writeq ((u_longlong_t) BankMaxSize, (u_longlong_t *) BankBase);
+
+       for (BankSize = (BankMaxSize>>1); BankSize > PAGESIZE; BankSize >>= 1)
+       {
+           if (MapDeviceRegister (dev, ELAN3_BAR_SDRAM, &PageBase, BankOffset + BankSize, PAGESIZE, &PageHandle) == ESUCCESS)
+           {
+               writeq (BankSize, (u_longlong_t *) PageBase);
+               UnmapDeviceRegister (dev, &PageHandle);
+           }
+       }
+       mmiob();
+
+       BankSize = (u_long) readq ((u_longlong_t *) BankBase);
+       
+       if (sdram_bank_limit == 0 || BankSize <= (sdram_bank_limit * 1024 * 1024))
+           printk ("elan%d: memory bank %d is %dK\n", Instance, Bank, BankSize / 1024);
+       else
+       {
+           BankSize = (sdram_bank_limit * 1024 * 1024);
+           printk ("elan%d: limit memory bank %d to %dK\n", Instance, Bank, BankSize / 1024);
+       }
+
+       UnmapDeviceRegister (dev, &BankHandle);
+       
+       /* Now map all of this bank into the kernel */
+       if (MapDeviceRegister (dev, ELAN3_BAR_SDRAM, &BankBase, BankOffset, BankSize, &BankHandle) != ESUCCESS)
+       {
+           printk ("elan%d: Cannot initialise memory bank %d\n", Instance, Bank);
+           continue;
+       }
+       
+       dev->SdramBanks[Bank].Size    = BankSize;
+       dev->SdramBanks[Bank].Mapping = BankBase;
+       dev->SdramBanks[Bank].Handle  = BankHandle;
+
+#ifndef CONFIG_MPSAS
+       /* Initialise it for ECC */
+       preemptable_start {
+           for (i = 0; i < BankSize; i += 8)
+           {
+               elan3_sdram_writeq (dev, (Bank << ELAN3_SDRAM_BANK_SHIFT) | i, INIT_PATTERN(BankOffset+i));
+
+               preemptable_check();
+           }
+       } preemptable_end;
+#endif
+    }
+    
+    return (ESUCCESS);
+}
+
+static void
+InitialiseSdram (ELAN3_DEV *dev)
+{
+    int indx, size, b;
+
+    for (b = 0; b < ELAN3_SDRAM_NUM_BANKS; b++)
+    {
+       ELAN3_SDRAM_BANK *bank = &dev->SdramBanks[b];
+
+       if (bank->Size == 0)
+           continue;
+
+       /* allocate a ptbl group pointer for each possible ptbl group in this bank */
+       KMEM_ZALLOC (bank->PtblGroups, ELAN3_PTBL_GR **, sizeof (ELAN3_PTBL_GR *) * bank->Size/PTBL_GROUP_SIZE, TRUE);
+           
+       /* allocate the buddy allocator bitmaps */
+       for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size <= bank->Size; indx++, size <<= 1)
+           KMEM_ZALLOC (bank->Bitmaps[indx], bitmap_t *, sizeof (bitmap_t)*BT_BITOUL(bank->Size/size), TRUE);
+           
+       /* and add it to the sdram buddy allocator */
+       elan3_sdram_add (dev, (b << ELAN3_SDRAM_BANK_SHIFT), (b << ELAN3_SDRAM_BANK_SHIFT) + bank->Size);
+    }
+}
+
+#include <elan3/vpd.h>
+
+int
+ReadVitalProductData (ELAN3_DEV *dev, int *CasLatency)
+{
+    DeviceMappingHandle RomHandle;
+    unsigned char      *RomBase;
+    unsigned char      *PCIDataPtr;
+    unsigned char      *VPDPtr;
+    unsigned char      *lim;
+    int                        type;
+    int                        i, len, len2;
+    char               name[3] = "XX";
+    char               value[256];
+    int                        finished = 0;
+
+    
+    /* default valud for CAS latency is 3 */
+    (*CasLatency) = CAS_LATENCY_3;
+
+    if (MapDeviceRegister (dev, ELAN3_BAR_EBUS, (ioaddr_t *) &RomBase, ELAN3_EBUS_ROM_OFFSET, ELAN3_EBUS_ROM_SIZE, &RomHandle) != ESUCCESS)
+    {
+       printk ("elan%d: Cannot map ROM\n", dev->Instance);
+       return (EFAIL);
+    }
+    
+    /* Check the ROM signature */
+    if (RomBase[0] != 0x55 || RomBase[1] != 0xAA)
+    {
+       printk ("elan%d: Invalid ROM signature %02x %02x\n", dev->Instance, RomBase[0], RomBase[1]);
+       return (ESUCCESS);
+    }
+    
+    PCIDataPtr = RomBase + ((RomBase[0x19] << 8) | RomBase[0x18]);
+
+    /* check the pci data structure */
+    if (PCIDataPtr[0] != 'P' || PCIDataPtr[1] != 'C' || PCIDataPtr[2] != 'I' || PCIDataPtr[3] != 'R')
+    {
+       printk ("elan%d: Invalid PCI Data structure\n", dev->Instance);
+       return (ESUCCESS);
+    }
+    
+    /* Extract the VPD pointer */
+    VPDPtr = RomBase + ((PCIDataPtr[9] << 8) | PCIDataPtr[8]);
+
+    if (VPDPtr == RomBase)
+    {
+       printk ("elan%d: No Vital Product Data\n", dev->Instance);
+       return (ESUCCESS);
+    }
+
+    while (! finished)
+    {
+       type = *VPDPtr++;
+       
+       if (type & LARGE_RESOURCE_BIT)
+       {
+           len = *(VPDPtr++);
+           len += *(VPDPtr++) << 8;
+
+           switch (type & ~LARGE_RESOURCE_BIT)
+           {
+           case LARGE_RESOURCE_STRING:
+               printk ("elan%d: ", dev->Instance);
+               for (i = 0; i < len; i++)
+                   printk ("%c", *VPDPtr++);
+               printk ("\n");
+               break;
+               
+           case LARGE_RESOURCE_VENDOR_DEFINED:
+               VPDPtr += len;
+               break;
+               
+           case LARGE_RESOURCE_VITAL_PRODUCT_DATA:
+               for (lim = VPDPtr + len; VPDPtr < lim; )
+               {
+                   name[0] = *VPDPtr++;
+                   name[1] = *VPDPtr++;
+                   len2    = *VPDPtr++;
+
+                   for (i = 0; i < len2 && VPDPtr < lim; i++)
+                       value[i] = *VPDPtr++;
+                   value[i] = '\0';
+
+                   if (! strcmp (name, "SN"))
+                       printk ("elan%d: Serial Number - %s\n", dev->Instance, value);
+
+                   if (! strcmp (name, "Z0"))
+                       (*CasLatency) = (strcmp (value, "CAS_LATENCY_2") ? CAS_LATENCY_3 : CAS_LATENCY_2);
+               }
+               break;
+               
+           default:
+               printk ("elan%d: unknown large resource %x\n", dev->Instance, type);
+               finished = 1;
+               break;
+           }
+       }
+       else
+       {
+           len = type & 0x7;
+
+           switch (type >> 3)
+           {
+           case SMALL_RESOURCE_COMPATIBLE_DEVICE_ID:
+               VPDPtr += len;
+               break;
+
+           case SMALL_RESOURCE_VENDOR_DEFINED:
+               VPDPtr += len;
+               break;
+               
+           case SMALL_RESOURCE_END_TAG:
+               finished = 1;
+               break;
+               
+           default:
+               printk ("elan%d: unknown small resource %x\n", dev->Instance, type >> 3);
+               finished = 1;
+               break;
+           }
+       }
+    }
+    
+    UnmapDeviceRegister (dev, &RomHandle);
+    return (ESUCCESS);
+}
+
+void
+ElanSetPtblGr (ELAN3_DEV *dev, sdramaddr_t offset, ELAN3_PTBL_GR *ptg)
+{
+    int bank = offset >> ELAN3_SDRAM_BANK_SHIFT;
+    
+    dev->SdramBanks[bank].PtblGroups[(offset & (ELAN3_SDRAM_BANK_SIZE-1)) / PTBL_GROUP_SIZE] = ptg;
+}
+
+ELAN3_PTBL_GR *
+ElanGetPtblGr (ELAN3_DEV *dev, sdramaddr_t offset)
+{
+    int bank = offset >> ELAN3_SDRAM_BANK_SHIFT;
+    
+    return (dev->SdramBanks[bank].PtblGroups[(offset & (ELAN3_SDRAM_BANK_SIZE-1)) / PTBL_GROUP_SIZE]);
+}
+
+void
+ElanFlushTlb (ELAN3_DEV *dev)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->TlbLock, flags);
+    BumpStat (dev, TlbFlushes);
+
+    write_reg32 (dev, Cache_Control_Reg.ContReg, dev->Cache_Control_Reg | MMU_FLUSH);
+    mmiob();
+    spin_unlock_irqrestore (&dev->TlbLock, flags);
+
+    while (! (read_reg32 (dev, Cache_Control_Reg.ContReg) & MMU_FLUSHED))
+       mb();
+}
+
+void
+KillNegativeDma (ELAN3_DEV *dev, void *arg)
+{
+    DMA_TRAP     *trap    = dev->DmaTrap;
+    E3_Status_Reg status;
+    sdramaddr_t   FPtr, BPtr;
+    sdramaddr_t   Base, Top;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    ASSERT (read_reg32 (dev, Exts.InterruptReg) & INT_DProcHalted);
+
+    /* Initialise the trap to deliver to the offending user process */
+    trap->Status.Status   = read_reg32 (dev, Exts.DProcStatus.Status);
+    trap->PacketInfo.Value = 0;
+    
+    bzero (&trap->FaultSave, sizeof (trap->FaultSave));
+    bzero (&trap->Data0, sizeof (trap->Data0));
+    bzero (&trap->Data1, sizeof (trap->Data1));
+    bzero (&trap->Data2, sizeof (trap->Data2));
+    bzero (&trap->Data3, sizeof (trap->Data3));
+
+    /* run down the kernel dma run queue and panic on a -ve length dma */
+    FPtr  = read_reg32 (dev, DProc_SysCntx_FPtr);
+    BPtr  = read_reg32 (dev, DProc_SysCntx_BPtr);
+    Base  = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[0]);
+    Top   = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[E3_SysCntxQueueSize-1]);
+    
+    while (FPtr != BPtr)
+    {
+       elan3_sdram_copyq_from_sdram (dev, FPtr, &trap->Desc, sizeof (E3_DMA_BE));
+       
+       if (trap->Desc.s.dma_size > E3_MAX_DMA_SIZE)
+           panic ("KillNegativeDma: -ve sized kernel dma\n");
+
+       FPtr = (FPtr == Top) ? Base : FPtr + sizeof (E3_DMA);
+    }
+
+    /* run down the user dma run queue and "remove" and -ve length dma's */
+    FPtr  = read_reg32 (dev, DProc_NonSysCntx_FPtr);
+    BPtr  = read_reg32 (dev, DProc_NonSysCntx_BPtr);
+    Base  = dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxDmaQueue[0]);
+    Top   = dev->TAndQBase + offsetof (E3_TrapAndQueue, NonSysCntxDmaQueue[E3_NonSysCntxQueueSize-1]);
+    
+    while (FPtr != BPtr)
+    {
+       elan3_sdram_copyq_from_sdram (dev, FPtr, &trap->Desc, sizeof (E3_DMA_BE));
+       
+       if (trap->Desc.s.dma_size > E3_MAX_DMA_SIZE)
+       {
+           PRINTF3 (NULL, DBG_INTR, "KillNegativeDma: remove dma - context %d size %d SuspendAddr %x\n", 
+                    trap->Desc.s.dma_u.s.Context, trap->Desc.s.dma_size, trap->Status.s.SuspendAddr);
+
+           trap->Status.s.TrapType = trap->Status.s.SuspendAddr;
+           trap->Status.s.Context  = trap->Desc.s.dma_u.s.Context;
+
+           DeliverDProcTrap (dev, trap, 0);
+
+           /*
+            * Remove the DMA from the queue by replacing it with one with
+            * zero size and no events.
+            *
+            * NOTE: we must preserve the SYS_CONTEXT_BIT since the Elan uses this
+            * to mark the approriate run queue as empty.
+            */
+           trap->Desc.s.dma_type            = 0;
+           trap->Desc.s.dma_size            = 0;
+           trap->Desc.s.dma_source          = (E3_Addr) 0;
+           trap->Desc.s.dma_dest            = (E3_Addr) 0;
+           trap->Desc.s.dma_destCookieVProc = (E3_Addr) 0;
+           trap->Desc.s.dma_srcEvent        = (E3_Addr) 0;
+           trap->Desc.s.dma_srcCookieVProc  = (E3_Addr) 0;
+
+           elan3_sdram_copyq_to_sdram (dev, &trap->Desc, FPtr, sizeof (E3_DMA_BE));
+       }
+
+       FPtr = (FPtr == Top) ? Base : FPtr + sizeof (E3_DMA);
+    }
+
+    status.Status = read_reg32 (dev, Exts.DProcStatus.Status);
+
+    if (status.s.SuspendAddr == MI_DequeueNonSysCntxDma || 
+       status.s.SuspendAddr == MI_DequeueSysCntxDma ||
+       status.s.SuspendAddr == MI_DmaLoop)
+    {
+       PRINTF0 (NULL, DBG_INTR, "KillNegativeDma: unlock dma processor\n");
+       write_reg32 (dev, Exts.Dmas.DmaWrs.LdAlignment, 0);
+       write_reg32 (dev, Exts.Dmas.DmaWrs.LdDmaType,   0);
+       mmiob();
+       
+       DELAY (10);
+       
+       write_reg32 (dev, Exts.Dmas.DmaWrs.LdAlignment, 0);
+       write_reg32 (dev, Exts.Dmas.DmaWrs.LdDmaType,   0);
+       mmiob();
+    }
+
+    PRINTF0 (NULL, DBG_INTR, "KillNegativeDma: dma processor restarted\n");
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    schedule_timer_fn (&dev->DmaPollTimeoutId, PollForDmaHungup, (void *) dev, 1);
+}
+
+void
+ForceTProcTrap (ELAN3_DEV *dev, void *arg)
+{
+    printk ("elan%d: forced tproc trap .....\n", dev->Instance);
+
+    schedule_timer_fn (&dev->DmaPollTimeoutId, PollForDmaHungup, (void *) dev, 1);
+}
+
+void
+PollForDmaHungup (void *arg)
+{
+    ELAN3_DEV     *dev   = (ELAN3_DEV *) arg;
+    unsigned long flags;
+    E3_Status_Reg status;
+    E3_uint32     insn1, insn3;
+    register int  i;
+
+    if (read_reg32 (dev, Dma_Desc.dma_size) > E3_MAX_DMA_SIZE)
+    {
+       status.Status = read_reg32 (dev, Exts.DProcStatus);
+
+       PRINTF2 (NULL, DBG_INTR, "PollForDmaHungup: size %x SuspendAddr %x\n", read_reg32 (dev, Dma_Desc.dma_size), status.s.SuspendAddr);
+
+       if (status.s.SuspendAddr == MI_DequeueNonSysCntxDma || 
+           status.s.SuspendAddr == MI_DequeueSysCntxDma ||
+           status.s.SuspendAddr == MI_DmaLoop)
+       {
+           printk ("elan%d: PollForDmaHungup: size %x context %d SuspendAddr %x\n", 
+                   dev->Instance, read_reg32 (dev, Dma_Desc.dma_size),
+                   status.s.Context, status.s.SuspendAddr);
+       
+           PRINTF2 (NULL, DBG_INTR, "PollForDmaHungup: dma_size %x status %x\n",
+                    read_reg32 (dev, Dma_Desc.dma_size), status.Status);
+           
+           spin_lock_irqsave (&dev->IntrLock, flags);
+           QueueHaltOperation (dev, 0, NULL, INT_DProcHalted, KillNegativeDma, NULL);
+           spin_unlock_irqrestore (&dev->IntrLock, flags);
+           
+           return;
+       }
+    }
+
+    status.Status = read_reg32 (dev, Exts.TProcStatus);
+    if (status.s.WakeupFunction == WakeupStopped)
+    {
+       E3_uint32 PC = read_reg32 (dev, ExecutePC);
+
+       /* See if it's likely that the thread is really "stuck" on a waitevent/break 
+        * instruction ......... */
+       for (i = 0; i < 10; i++)
+       {
+           status.Status = read_reg32 (dev, Exts.TProcStatus);
+           insn1         = read_reg32 (dev, IBufferReg[1]);
+           insn3         = read_reg32 (dev, IBufferReg[3]);
+           
+           if (! (status.s.WakeupFunction == WakeupStopped && read_reg32 (dev, ExecutePC) == PC &&     /* stopping and it could be a break/waitevent */
+                  (insn1 == 0x81a00000 || insn3 == 0x81a00000 ||                                       /* break instruction */
+                   insn1 == 0x81b00000 || insn3 == 0x81b00000)))                                       /* waitevent instruction  */
+               break;
+       }
+
+       if (i == 10)
+       {
+           printk ("elan%d: forcing tproc trap from %s instruction at pc %x\n", dev->Instance, 
+                   (insn1 == 0x81a00000 || insn3 == 0x81a00000) ? "break" : "waitevent", PC);
+
+           spin_lock_irqsave (&dev->IntrLock, flags);
+           QueueHaltOperation (dev, 0, NULL, INT_TProcHalted, ForceTProcTrap, NULL);
+           spin_unlock_irqrestore (&dev->IntrLock, flags);
+           return;
+       }
+    }
+
+    schedule_timer_fn (&dev->DmaPollTimeoutId, PollForDmaHungup, (void *) dev, 10);
+}
+
+/*=======================================================================================*/
+/*
+ * Interrupt handler.
+ */
+static void
+ReEnableErrorInterrupts (void *arg)
+{
+    ELAN3_DEV     *dev = (ELAN3_DEV *) arg;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    if ((dev->SchCntReg & LinkBoundaryScan) == 0)
+       ENABLE_INT_MASK (dev, INT_ErrorInterrupts);
+
+    PRINTF1 (DBG_DEVICE, DBG_INTR, "ReEnableErrorInterrupts: IntMask=%x\n", read_reg32 (dev, Exts.InterruptMask));
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+void
+CheckForExcessiveErrorRate (ELAN3_DEV *dev)
+{
+    if (dev->ErrorTime == (lbolt/hz))
+    {
+       if (dev->ErrorsPerTick++ > 100)
+       {
+           PRINTF0 (DBG_DEVICE, DBG_INTR, "CheckForExcessiveErrorRate: too many links errors, disabling interrupt\n");
+
+           DISABLE_INT_MASK (dev, INT_ErrorInterrupts);
+
+           schedule_timer_fn (&dev->ErrorTimeoutId, ReEnableErrorInterrupts, (void *) dev, hz);
+       }
+    }
+    else
+    {
+       dev->ErrorTime     = (lbolt/hz);
+       dev->ErrorsPerTick = 0;
+    }
+}
+/*=======================================================================================*/
+/*
+ * Interrupt handler.
+ */
+static void
+HandlePciMemErr (ELAN3_DEV *dev)
+{
+    PRINTF0 (DBG_DEVICE, DBG_INTR, "HandlePciMemErr : masking out interrupt\n");
+    
+    ElanBusError (dev);
+    panic ("elan pci memory error\n");
+}
+
+static void
+HandleSDRamInterrupt (ELAN3_DEV *dev)
+{
+    E3_uint32     EccStatus0 = read_reg32 (dev, ECC_STATUS0);
+    E3_uint32     EccStatus1 = read_reg32 (dev, ECC_STATUS1);
+    unsigned long flags;
+
+    PRINTF5 (DBG_DEVICE, DBG_INTR, "elan: ECC error - Addr=%x UE=%x CE=%x ME=%x Syn=%x\n",
+            EccStatus0 & ECC_ADDR_MASK, EccStatus0 & ECC_UE_MASK, 
+            EccStatus0 & ECC_CE_MASK, EccStatus0 & ECC_ME_MASK, 
+            EccStatus1 & ECC_SYN_MASK);
+
+    if (EccStatus0 & (ECC_UE_MASK|ECC_CE_MASK))
+    {
+       printk ("elan%d: ECC memory error (Address=%08x Syndrome=%02x %s%s%s)\n",
+               dev->Instance, 
+               (EccStatus0 & ECC_ADDR_MASK), (EccStatus1 & ECC_SYN_MASK), 
+               (EccStatus0 & ECC_UE_MASK) ? "Uncorrectable "   : "",
+               (EccStatus0 & ECC_CE_MASK) ? "Correctable "     : "",
+               (EccStatus0 & ECC_ME_MASK) ? "Multiple Errors " : "");
+    }
+
+    if (EccStatus0 & ECC_UE_MASK)
+       panic ("elan: Uncorrectable ECC memory error");
+    if (EccStatus0 & ECC_CE_MASK)
+       BumpStat (dev, CorrectableErrors);
+    if (EccStatus0 & ECC_ME_MASK)
+       BumpStat (dev, MultipleErrors);
+
+    /*
+     * Clear the interrupt and reset the error flags.
+     * Note. Might loose an UE or CE if it occurs between reading the status and
+     *       clearing the interrupt. I don't think this matters very much as the
+     *   status reg will only be used to identify a bad simm.
+     */
+
+    spin_lock_irqsave (&dev->TlbLock, flags);
+    write_reg32 (dev, Cache_Control_Reg.ContReg, dev->Cache_Control_Reg | CLEAR_SDRAM_ERROR);
+    mmiob();
+    spin_unlock_irqrestore (&dev->TlbLock, flags);
+
+    CheckForExcessiveErrorRate (dev);
+}
+
+static int
+HandleEventInterrupt (ELAN3_DEV *dev, int nticks, unsigned long *flags)
+{
+    E3_uint32 Fptr  = dev->Event_Int_Queue_FPtr;
+    E3_uint32 Bptr  = read_reg32 (dev, Event_Int_Queue_BPtr);                                          /* PCI read */
+    long      tlim  = lbolt + nticks;
+    long      count = 0;
+    ELAN3_CTXT *ctxt;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+    ASSERT ((dev->InterruptMask & INT_EventInterrupt) == 0);
+           
+    while (Fptr != Bptr)
+    {
+       while (Fptr != Bptr)
+       {
+           E3_EventInt_BE  EvInt;
+           E3_uint32       Context;
+
+           /* If we're running in the interrupt handler and have seen a high
+            * rate of event interrupts then punt to the thread  - however on 
+            * Linux the elan interrupt handler can block the timer interrupt,
+            * and so lbolt (jiffies) is not incremented, hence we punt after
+            a number of loops instead */
+#if defined(LINUX)
+           if (in_interrupt() && ++count > eventint_punt_loops)
+               return (EAGAIN);
+#endif
+
+           if (nticks && ((int) (lbolt - tlim)) > 0)
+           {
+               PRINTF2 (DBG_DEVICE, DBG_INTR, "HandleEventInterrupt: Fptr %x Bptr %x punting to thread\n", Fptr, Bptr);
+               return (EAGAIN);
+           }
+
+           elan3_sdram_copyq_from_sdram (dev, Fptr, (void *) &EvInt, 8);                               /* PCI read */
+           
+           /* The context number is held in the top 16 bits of the EventContext */
+           Context = (EvInt.s.EventContext >> 16) & MAX_ROOT_CONTEXT_MASK;
+           
+           PRINTF2 (DBG_DEVICE, DBG_INTR, "HandleEventInterrupt: Context %d : Cookie %x\n", Context, EvInt.s.IntCookie);
+           
+           ctxt = ELAN3_DEV_CTX_TABLE(dev, Context);
+
+           /* Work out new fptr, and store it in the device, since we'll be dropping the IntrLock */
+           Fptr = E3_EVENT_INTQ_NEXT(Fptr);
+           dev->Event_Int_Queue_FPtr = Fptr;
+
+           if (ctxt == NULL)
+           {
+               PRINTF3 (DBG_DEVICE, DBG_INTR, "HandleEventInterrupt: Fptr %x Bptr %x context %d invalid\n",
+                        Fptr, Bptr, Context);
+               BumpStat (dev, InvalidContext);
+           }
+           else
+           {
+               BumpStat (dev, EventInterrupts);
+               
+               spin_unlock_irqrestore (&dev->IntrLock, *flags);
+               QueueEventInterrupt (ctxt, EvInt.s.IntCookie);
+               spin_lock_irqsave (&dev->IntrLock, *flags);
+           }
+           
+           /* Re-read the FPtr, since we've dropped the IntrLock */
+           Fptr = dev->Event_Int_Queue_FPtr;
+           
+           /* Store the new FPtr to the elan, this also clears the interrupt. */
+           write_reg32 (dev, Event_Int_Queue_FPtr, Fptr);                                      /* PCI write */
+
+           mmiob();
+       }
+
+       mb();
+       Bptr = read_reg32 (dev, Event_Int_Queue_BPtr);                                          /* PCI read */
+    }
+
+    return (ESUCCESS);
+}
+
+int
+SetLinkBoundaryScan (ELAN3_DEV *dev)
+{
+    int           res = ESUCCESS;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    if ((dev->SchCntReg & LinkBoundaryScan) != 0)
+       res = EAGAIN;
+    else
+    {
+       PRINTF0 (DBG_DEVICE, DBG_BSCAN, "SetLinkBoundaryScan: setting link into boundary scan mode\n");
+
+       /*
+        * We're going to set the link into boundary scan mode,  so firstly
+        * set the inputters to discard everything.
+        */
+       if (dev->DiscardAllCount++ == 0)
+           SetSchedStatusRegister (dev, read_reg32 (dev, Exts.InterruptReg), NULL);
+
+       /*
+        * Now disable the error interrupts
+        */
+       DISABLE_INT_MASK (dev, INT_ErrorInterrupts);
+       
+       /*
+        * And set the link into boundary scan mode, and drive
+        * a reset token onto the link.
+        */
+       SET_SCHED_LINK_VALUE (dev, 1, LinkResetToken);
+    }
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    return (res);
+}
+
+void
+ClearLinkBoundaryScan (ELAN3_DEV *dev)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    if ((dev->SchCntReg & LinkBoundaryScan) != 0)
+    {
+       PRINTF0 (DBG_DEVICE, DBG_BSCAN, "ClearLinkBoundaryScan: taking link out of boundary scan mode\n");
+
+       /*
+        * Take the link out of boundary scan 
+        */
+       SET_SCHED_LINK_VALUE (dev, 0, 0);
+
+       /*
+        * Clear any link errors.
+        */
+       PULSE_SCHED_STATUS (dev, ClearLinkErrorInt);
+
+       /*
+        * Re-enable the error interrupts.
+        */
+       if (! timer_fn_queued(&dev->ErrorTimeoutId))
+           ENABLE_INT_MASK (dev, INT_ErrorInterrupts);
+
+       /*
+        * And stop the inputter from discarding all packets.
+        */
+       if (--dev->DiscardAllCount == 0)
+           SetSchedStatusRegister (dev, 0, NULL);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+int
+WriteBoundaryScanValue (ELAN3_DEV *dev, int value)
+{
+    int           res = 0;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    if ((dev->SchCntReg & LinkBoundaryScan) != 0)
+    {
+       PRINTF1 (DBG_DEVICE, DBG_BSCAN, "WriteBoundaryScanValue: driving value 0x%x onto link\n", value);
+       SET_SCHED_LINK_VALUE (dev, 1, value);
+
+       res = read_reg32 (dev, Exts.LinkState);
+
+       PRINTF1 (DBG_DEVICE, DBG_BSCAN, "WriteBoundaryScanValue: return 0x%x\n", res);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    return (res);
+}
+
+int
+ReadBoundaryScanValue(ELAN3_DEV *dev, int link)
+{
+    int           res;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    if ((dev->SchCntReg & LinkBoundaryScan) == 0)
+    {
+       PRINTF1 (DBG_DEVICE, DBG_BSCAN, "ReadBoundaryScanValue: set linkval 0x%x\n",  link);
+       SET_SCHED_LINK_VALUE (dev, 0, link);
+    }
+    res = read_reg32 (dev, Exts.LinkState);
+    PRINTF1 (DBG_DEVICE, DBG_BSCAN, "ReadBoundaryScanValue: return 0x%x\n", res);
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    return (res);
+}
+
+static int
+ReadLinkVal (ELAN3_DEV *dev, int link)
+{
+    if ((dev->SchCntReg & LinkBoundaryScan) == 0)
+       SET_SCHED_LINK_VALUE (dev, 0, link);
+    
+    return (read_reg32 (dev, Exts.LinkState));
+}
+
+static void
+HandleLinkError (ELAN3_DEV *dev)
+{
+    E3_uint32 value = read_reg32 (dev, Exts.LinkErrorTypes);
+
+    PRINTF1 (DBG_DEVICE, DBG_LINKERR, "HandleLinkError: LinkErrorTypes %08x - clearing\n", value);
+    
+    if (value & LS_LockError)   BumpStat (dev, LockError);
+    if (value & LS_DeskewError) BumpStat (dev, DeskewError);
+    if (value & LS_PhaseError)  BumpStat (dev, PhaseError);
+    if (value & LS_DataError)   BumpStat (dev, DataError);
+    if (value & LS_FifoOvFlow0) BumpStat (dev, FifoOvFlow0);
+    if (value & LS_FifoOvFlow1) BumpStat (dev, FifoOvFlow1);
+
+    if (value & LS_DataError)
+       dev->Stats.LinkErrorValue = ReadLinkVal (dev, 12) | (ReadLinkVal (dev, 13) << 9);
+
+    PULSE_SCHED_STATUS (dev, ClearLinkErrorInt);
+
+    CheckForExcessiveErrorRate (dev);
+}
+
+static void
+HandleErrorInterrupt (ELAN3_DEV *dev, E3_uint32 Pend)
+{
+    if (Pend & INT_PciMemErr)
+       HandlePciMemErr (dev);
+    
+    if (Pend & INT_SDRamInt)
+       HandleSDRamInterrupt (dev);
+    
+    if (Pend & INT_LinkError)
+       HandleLinkError (dev);
+}
+       
+static void
+HandleAnyIProcTraps (ELAN3_DEV *dev, E3_uint32 Pend)
+{
+    E3_uint32       RestartBits = 0;
+    
+    if (Pend & INT_IProcCh0SysCntx)
+    {
+       HandleIProcTrap (dev, 0, Pend,
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcSysCntx),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh0_C0_TrHead[0]),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh0_C0_TrData[0]));
+                        
+       RestartBits |= RestartCh0SysCntx;
+    }
+    
+    if (Pend & INT_IProcCh1SysCntx)
+    {
+       HandleIProcTrap (dev, 1, Pend,
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcSysCntx),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh1_C0_TrHead[0]),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh1_C0_TrData[0]));
+                        
+       RestartBits |= RestartCh1SysCntx;
+    }
+
+    if (Pend & INT_IProcCh0NonSysCntx)
+    {
+       HandleIProcTrap (dev, 0, Pend,
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcNonSysCntx),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh0_NonC0_TrHead[0]),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh0_NonC0_TrData[0]));
+
+       RestartBits |= RestartCh0NonSysCntx;
+    }
+    
+
+    if (Pend & INT_IProcCh1NonSysCntx)
+    {
+       HandleIProcTrap (dev, 1, Pend,
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, IProcNonSysCntx),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh1_NonC0_TrHead[0]),
+                        dev->TAndQBase + offsetof (E3_TrapAndQueue, VCh1_NonC0_TrData[0]));
+       RestartBits |= RestartCh1NonSysCntx;
+    }
+
+    PULSE_SCHED_STATUS (dev, RestartBits);
+}
+
+static void
+elan3_event_interrupt (ELAN3_DEV *dev)
+{
+    unsigned long flags;
+
+    kernel_thread_init("elan3_event_int");
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    for (;;)
+    {
+       /* Make sure we never sleep with the EventInterrupt disabled */
+       if (! (dev->InterruptMask & INT_EventInterrupt))
+       {
+           if (HandleEventInterrupt (dev, eventint_resched_ticks, &flags) != ESUCCESS)
+               BumpStat (dev, EventRescheds);
+           
+           ENABLE_INT_MASK (dev, INT_EventInterrupt);
+       }
+
+       if (dev->ThreadsShouldStop)
+           break;
+
+       kcondvar_wait (&dev->IntrWait, &dev->IntrLock, &flags);
+    }
+    
+    dev->EventInterruptThreadStopped = 1;
+    kcondvar_wakeupall (&dev->IntrWait, &dev->IntrLock);
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    kernel_thread_exit ();
+}
+
+int
+InterruptHandler (ELAN3_DEV *dev)
+{
+    E3_uint32     Mask;
+    E3_uint32     Pend;
+    E3_uint32     RestartBits;
+    int           deliverDProcTrap;
+    int                  deliverTProcTrap;
+    static long   lboltsave;
+    int           loop_count = 0; 
+    unsigned long flags;
+    int  tproc_delivered;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    BumpStat (dev, Interrupts);
+
+    Mask = dev->InterruptMask;
+    Pend = read_reg32 (dev, Exts.InterruptReg);                                                /* PCI read */
+
+    /* Save the lbolt so we know how long in do loop or in event handling */
+    lboltsave = lbolt;
+
+    if ((Pend & Mask) == INT_EventInterrupt)
+    {
+       DISABLE_INT_MASK (dev, INT_EventInterrupt);
+
+       if (HandleEventInterrupt (dev, eventint_punt_ticks, &flags) == ESUCCESS)
+           ENABLE_INT_MASK (dev, INT_EventInterrupt);
+       else
+       {
+           BumpStat (dev, EventPunts);
+
+           kcondvar_wakeupone (&dev->IntrWait, &dev->IntrLock);
+       }
+
+        if ((lbolt - lboltsave) > dev->Stats.LongestInterrupt)
+            dev->Stats.LongestInterrupt = (lbolt - lboltsave);
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       return (ESUCCESS);
+    }
+
+    if ((Pend & Mask) == 0)
+    {
+       PRINTF3 (DBG_DEVICE, DBG_INTR, "InterruptHandler: Spurious Pend %x Mask %x SchedStatus %x\n", 
+                Pend, Mask, read_reg32 (dev, Exts.SchCntReg));
+
+        if ((lbolt - lboltsave) > dev->Stats.LongestInterrupt)
+            dev->Stats.LongestInterrupt = (lbolt - lboltsave);
+       spin_unlock_irqrestore (&dev->IntrLock, flags);
+       return (EFAIL);
+    }
+
+    PRINTF3 (DBG_DEVICE, DBG_INTR, "InterruptHandler: Pend %x Mask %08x SchedStatus %x\n", 
+            Pend, Mask, read_reg32 (dev, Exts.SchCntReg));
+
+    do {
+       loop_count++;
+       RestartBits = 0;
+
+       if (Pend & Mask & (INT_CProc | INT_ComQueue))
+           HandleCProcTrap (dev, Pend, &Mask);
+
+       tproc_delivered = 0;
+
+       if (Pend & Mask & INT_TProc) {
+           ELAN_REG_REC(Pend);
+           tproc_delivered = 1;
+           deliverTProcTrap = HandleTProcTrap (dev, &RestartBits);
+       }
+       else
+           deliverTProcTrap = 0;
+
+       if (Pend & Mask & INT_DProc)
+           deliverDProcTrap = HandleDProcTrap (dev, &RestartBits);
+       else
+           deliverDProcTrap = 0;
+
+       ASSERT ((RestartBits & RestartDProc) == 0 || (read_reg32 (dev, Exts.DProcStatus.Status) >> 29) == 4);
+       ASSERT ((RestartBits & RestartDProc) == 0 || elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProc.s.FSR.Status))      == 0);
+       ASSERT ((RestartBits & RestartDProc) == 0 || elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData0.s.FSR.Status)) == 0);
+       ASSERT ((RestartBits & RestartDProc) == 0 || elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData1.s.FSR.Status)) == 0);
+       ASSERT ((RestartBits & RestartDProc) == 0 || elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData2.s.FSR.Status)) == 0);
+       ASSERT ((RestartBits & RestartDProc) == 0 || elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData3.s.FSR.Status)) == 0);
+
+       PULSE_SCHED_STATUS (dev, RestartBits);          /* Restart any processors which had trapped. */
+       SET_INT_MASK (dev, Mask);                       /* And install the new interrupt mask */
+
+       if ((Pend & Mask & INT_TProc) && deliverTProcTrap)
+           DeliverTProcTrap (dev, dev->ThreadTrap, Pend);
+
+       if ((Pend & Mask & INT_DProc) && deliverDProcTrap)
+           DeliverDProcTrap (dev, dev->DmaTrap, Pend);
+
+       if (Pend & Mask & INT_Inputters)
+           HandleAnyIProcTraps (dev, Pend);
+       
+       if (Pend & Mask & INT_EventInterrupt)
+       {
+           DISABLE_INT_MASK (dev, INT_EventInterrupt);
+           
+           if (loop_count == 1 && HandleEventInterrupt (dev, eventint_punt_ticks, &flags) == ESUCCESS) /* always punt to the thread if we've */
+               ENABLE_INT_MASK (dev, INT_EventInterrupt);                                              /* been round the loop once */
+           else
+           {
+               BumpStat (dev, EventPunts);
+
+               kcondvar_wakeupone (&dev->IntrWait, &dev->IntrLock);
+           }
+       }
+
+       if (Pend & (INT_Halted | INT_Discarding))
+           ProcessHaltOperations (dev, Pend);
+
+       if (Pend & Mask & INT_ErrorInterrupts)
+           HandleErrorInterrupt (dev, Pend);
+
+       Mask = dev->InterruptMask;
+       Pend = read_reg32 (dev, Exts.InterruptReg);     /* PCI read */
+       
+       if (tproc_delivered)
+           ELAN_REG_REC(Pend);
+
+       PRINTF3 (DBG_DEVICE, DBG_INTR, "InterruptHandler: Pend %x Mask %08x SchedStatus %x\n", 
+                Pend, Mask, read_reg32 (dev, Exts.SchCntReg));
+    }  while ((Pend & Mask) != 0);
+
+    if ((lbolt - lboltsave) > dev->Stats.LongestInterrupt)
+        dev->Stats.LongestInterrupt = (lbolt - lboltsave);
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    PRINTF2 (DBG_DEVICE, DBG_INTR, "InterruptHandler: lbolt is %lx; start lbolt is %lx\n", 
+            lbolt, lboltsave);
+
+    return (ESUCCESS);
+}
+
+void
+SetSchedStatusRegister (ELAN3_DEV *dev, E3_uint32 Pend, volatile E3_uint32 *Maskp)
+{
+    E3_uint32 HaltMask  = dev->HaltOperationsMask;
+    E3_uint32 Mask      = Maskp ? *Maskp : dev->InterruptMask;
+    E3_uint32 ClearBits = 0;
+    E3_uint32 SetBits   = 0;
+
+    PRINTF5 (DBG_DEVICE, DBG_INTR, "SetSchedStatusRegister: HaltOperationsMask=%x HaltAll=%d HaltDmaDequeue=%d HaltThread=%d DiscardAll=%d\n",
+            HaltMask, dev->HaltAllCount, dev->HaltDmaDequeueCount, dev->HaltThreadCount, dev->DiscardAllCount);
+
+    if (dev->FlushCommandCount)
+       SetBits |= FlushCommandQueues;
+    
+    if ((HaltMask & INT_DProcHalted) || dev->HaltAllCount)
+    {
+       SetBits |= HaltDmas | HaltDmaDequeue;
+       if (Pend & INT_DProcHalted)
+           Mask &= ~INT_DProcHalted;
+       else
+           Mask |= INT_DProcHalted;
+    }
+
+    if (dev->HaltDmaDequeueCount)
+    {
+       SetBits |= HaltDmaDequeue;
+       if (Pend & INT_DProcHalted)
+           Mask &= ~INT_DProcHalted;
+       else
+           Mask |= INT_DProcHalted;
+    }
+
+    if ((HaltMask & INT_TProcHalted) || dev->HaltAllCount || dev->HaltThreadCount)
+    {
+       SetBits |= HaltThread;
+       if (Pend & INT_TProcHalted)
+           Mask &= ~INT_TProcHalted;
+       else
+           Mask |= INT_TProcHalted;
+    }
+
+    if ((HaltMask & INT_DiscardingSysCntx) || dev->DiscardAllCount)
+    {
+       SetBits |= DiscardSysCntxIn;
+       if (Pend & INT_DiscardingSysCntx)
+           Mask &= ~INT_DiscardingSysCntx;
+       else
+           Mask |= INT_DiscardingSysCntx;
+    }
+
+    if ((HaltMask & INT_DiscardingNonSysCntx) || dev->DiscardNonContext0Count || dev->DiscardAllCount)
+    {
+       SetBits |= DiscardNonSysCntxIn;
+       if (Pend & INT_DiscardingNonSysCntx)
+           Mask &= ~INT_DiscardingNonSysCntx;
+       else
+           Mask |= INT_DiscardingNonSysCntx;
+    }
+
+    if (dev->HaltNonContext0Count)
+       SetBits |= StopNonSysCntxs;
+
+    ClearBits = SetBits ^ (FlushCommandQueues | HaltDmas | HaltDmaDequeue | HaltThread |
+                          DiscardSysCntxIn | DiscardNonSysCntxIn | StopNonSysCntxs);
+
+    PRINTF4 (DBG_DEVICE, DBG_INTR, "SetSchedStatusRegister: SetBits=%x InterruptMask=%x InterruptReg=%x Mask=%x\n",
+            SetBits, dev->InterruptMask, read_reg32 (dev, Exts.InterruptReg), Mask);
+
+    MODIFY_SCHED_STATUS (dev, SetBits, ClearBits);
+
+    if (Maskp)
+       *Maskp = Mask;                                          /* copyback new interrupt mask */
+    else
+       SET_INT_MASK(dev, Mask);
+}
+
+void
+FreeHaltOperation (ELAN3_DEV *dev, ELAN3_HALTOP *op)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->FreeHaltLock, flags);
+    op->Next = dev->FreeHaltOperations;
+    dev->FreeHaltOperations = op;
+    spin_unlock_irqrestore (&dev->FreeHaltLock, flags);
+}
+
+int
+ReserveHaltOperations (ELAN3_DEV *dev, int count, int cansleep)
+{
+    ELAN3_HALTOP   *op;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->FreeHaltLock, flags);
+    while ((dev->NumHaltOperations - dev->ReservedHaltOperations) < count)
+    {
+       spin_unlock_irqrestore (&dev->FreeHaltLock, flags);
+
+       KMEM_ZALLOC (op, ELAN3_HALTOP *, sizeof (ELAN3_HALTOP), cansleep);
+
+       if (op == NULL)
+           return (FALSE);
+
+       spin_lock_irqsave (&dev->FreeHaltLock, flags);
+
+       dev->NumHaltOperations++;
+
+       op->Next = dev->FreeHaltOperations;
+       dev->FreeHaltOperations = op;
+    }
+                   
+    dev->ReservedHaltOperations += count;
+    
+    spin_unlock_irqrestore (&dev->FreeHaltLock, flags);
+
+    return (TRUE);
+}
+
+void
+ReleaseHaltOperations (ELAN3_DEV *dev, int count)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->FreeHaltLock, flags);
+    dev->ReservedHaltOperations -= count;
+    spin_unlock_irqrestore (&dev->FreeHaltLock, flags);
+}
+
+void
+QueueHaltOperation (ELAN3_DEV *dev, E3_uint32 Pend, volatile E3_uint32 *Maskp, 
+                   E3_uint32 ReqMask, void (*Function)(ELAN3_DEV *, void *), void *Arguement)
+{
+    ELAN3_HALTOP *op;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+    
+    spin_lock (&dev->FreeHaltLock);
+    op = dev->FreeHaltOperations;
+
+    ASSERT (op != NULL);
+
+    dev->FreeHaltOperations = op->Next;
+    spin_unlock (&dev->FreeHaltLock);
+
+    op->Mask      = ReqMask;
+    op->Function  = (void (*)(void *, void *))Function;
+    op->Arguement = Arguement;
+
+    dev->HaltOperationsMask |= ReqMask;                                /* Add our bits to the global bits needed. */
+    SetSchedStatusRegister (dev, Pend, Maskp);                 /* Set the control register and the interrupt mask */
+
+    /*
+     * If the condition is already satisfied, then SetSchedStatusRegister will
+     * have masked out the interrupt, so re-enable it now to take it straight
+     * away
+     */
+    if (Maskp == NULL)
+    {
+       if ((read_reg32 (dev, Exts.InterruptReg) & ReqMask) == ReqMask)
+           ENABLE_INT_MASK (dev, ReqMask);
+    }
+    else
+    {
+       if ((Pend & ReqMask) == ReqMask)
+           *Maskp |= ReqMask;
+    }
+
+    *dev->HaltOperationsTailpp = op;                           /* Queue at end of list, since ProcessHaltOperations */
+    dev->HaltOperationsTailpp = &op->Next;                     /* drops the IntrLock while running down the list */
+    op->Next = NULL;
+}
+                   
+void
+ProcessHaltOperations (ELAN3_DEV *dev, E3_uint32 Pend)
+{
+    E3_uint32     Mask;
+    ELAN3_HALTOP  *op;
+    ELAN3_HALTOP **prevp;
+    E3_uint32     haltMask;
+    ELAN3_HALTOP  *next;
+
+    PRINTF1 (DBG_DEVICE, DBG_INTR, "ProcessHaltOperations: Pend %x\n", Pend);
+
+    for (;;)
+    {
+       ELAN3_HALTOP  *head = NULL;
+       ELAN3_HALTOP **tailp = &head;
+
+       /*
+        * Generate a list of halt operations which can be called now.
+        */
+       for (haltMask = 0, prevp = &dev->HaltOperations; (op = *prevp) != NULL; )
+       {
+           if ((Pend & op->Mask) != op->Mask)
+           {
+               haltMask |= op->Mask;
+               prevp = &op->Next;
+           }
+           else
+           {
+               *prevp = op->Next;                              /* remove from list */
+               if (op->Next == NULL)
+                   dev->HaltOperationsTailpp = prevp;
+               
+               *tailp = op;                                    /* add to local list */
+               op->Next = NULL;
+               tailp = &op->Next;
+           }
+       }
+
+       if (head == NULL)                                       /* nothing to do, so update */
+       {                                                       /* the schedule status register */
+           dev->HaltOperationsMask = haltMask;                 /* and the interrupt mask */
+           SetSchedStatusRegister (dev, Pend, NULL);
+           return;
+       }
+
+       /*
+        * flush the command queues, before calling any operations
+        */
+       Mask = dev->InterruptMask;
+       
+       if (dev->FlushCommandCount++ == 0)
+           SetSchedStatusRegister (dev, Pend, &Mask);
+       
+       if ((read_reg32 (dev, ComQueueStatus) & ComQueueNotEmpty) != 0)
+       {
+           if (dev->HaltThreadCount++ == 0)
+               SetSchedStatusRegister (dev, Pend, &Mask);
+
+           CAPTURE_CPUS();
+
+           while ((read_reg32 (dev, ComQueueStatus) & ComQueueNotEmpty) != 0)
+               mb();
+
+           RELEASE_CPUS();
+                   
+           if (--dev->HaltThreadCount == 0)
+               SetSchedStatusRegister (dev, Pend, &Mask);
+       }
+               
+       if (read_reg32 (dev, Exts.InterruptReg) & INT_CProc)
+       {
+           PRINTF0 (DBG_DEVICE, DBG_INTR, "ProcessHaltOperations: command processor has trapped\n");
+           HandleCProcTrap (dev, Pend, &Mask);
+       }
+       
+       if (--dev->FlushCommandCount == 0)
+           SetSchedStatusRegister (dev, Pend, &Mask);
+       
+       PRINTF2 (DBG_DEVICE, DBG_INTR, "ProcessHaltOperations: interrupt mask %08x -> %08x\n", 
+                dev->InterruptMask, Mask);
+       
+       SET_INT_MASK (dev, Mask);
+       spin_unlock (&dev->IntrLock);
+
+       /*
+        * now process the list of operations
+        * we have
+        */
+       for (op = head; op != NULL; op = next)
+       {
+           next = op->Next;
+
+           op->Function (dev, op->Arguement);
+           
+           FreeHaltOperation (dev, op);
+       }
+
+       spin_lock (&dev->IntrLock);
+    }
+}
+
+int
+ComputePosition (ELAN_POSITION *pos, unsigned nodeId, unsigned numNodes, unsigned numDownLinksVal)
+{
+    int i, lvl, n;
+    char numDownLinks[ELAN_MAX_LEVELS];
+
+    if (nodeId >= numNodes)
+       return (EINVAL);
+
+    for (i = 0; i < ELAN_MAX_LEVELS; i++, numDownLinksVal >>= 4)
+       numDownLinks[i] = numDownLinksVal & 7;
+    
+    for (lvl = 0, n = numNodes; n > ((lvl % 3) == 2 ? 8 : 4) && lvl < ELAN_MAX_LEVELS; lvl++)
+    {
+       if (numDownLinks[lvl] == 0)
+           numDownLinks[lvl] = 4;
+       
+       if ((n % numDownLinks[lvl]) != 0)
+           return (EINVAL);
+       
+       n /= numDownLinks[lvl];
+    }
+
+    if (numDownLinks[lvl] == 0)
+       numDownLinks[lvl] = n;
+
+    if (numDownLinks[lvl] != n)
+       return (EINVAL);
+
+    for (i = 0; i <= lvl; i++)
+       pos->pos_arity[i] = numDownLinks[lvl - i];
+
+    pos->pos_nodes  = numNodes;
+    pos->pos_levels = lvl + 1;
+    pos->pos_nodeid = nodeId;
+    pos->pos_mode   = ELAN_POS_MODE_SWITCHED;
+
+    return (0);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/elandev_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/elandev_linux.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/elandev_linux.c        2005-06-01 23:12:54.582441688 -0400
@@ -0,0 +1,2302 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: elandev_linux.c,v 1.102.2.4 2004/12/20 16:55:17 mike Exp $"
+/*     $Source: /cvs/master/quadrics/elan3mod/elan3/os/elandev_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kpte.h>
+
+#include <linux/config.h>
+#include <linux/mm.h>
+#include <linux/pci.h>
+#include <linux/reboot.h>
+#include <linux/notifier.h>
+
+#include <linux/init.h>
+#include <linux/module.h>
+
+#include <linux/pci.h>
+#include <linux/ptrack.h>
+
+#include <asm/uaccess.h>
+#include <asm/io.h>
+#include <asm/pgalloc.h>
+#include <asm/pgtable.h>
+
+#include <elan/devinfo.h>
+#include <elan/elanmod.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elanio.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/elansyscall.h>
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,2,0)
+#error please use a 2.2 series kernel or newer
+#endif
+
+/* Minor numbers encoded as :
+ *   [5:0]     device number
+ *   [15:6]    function number
+ */
+#define ELAN3_DEVICE_MASK          0x3F
+
+#define ELAN3_MINOR_CONTROL      0
+#define ELAN3_MINOR_MEM          1
+#define ELAN3_MINOR_USER        2
+#define ELAN3_MINOR_SHIFT        6
+
+#define ELAN3_DEVICE(inode)    (MINOR(inode->i_rdev) & ELAN3_DEVICE_MASK)
+#define ELAN3_MINOR(inode)     (MINOR(inode->i_rdev) >> ELAN3_MINOR_SHIFT)
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
+#      define SetPageReserved(page)    set_bit(PG_reserved, &(page)->flags)
+#      define ClearPageReserved(page)  clear_bit(PG_reserved, &(page)->flags)
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,4,23)
+typedef void irqreturn_t;
+#endif
+#       define IRQ_NONE
+#       define IRQ_HANDLED
+#       define IRQ_RETVAL(x)
+#endif
+
+
+/*
+ * Function prototypes.
+ */
+static int     elanattach(int instance, struct pci_dev *pcidev);
+static int     elandetach(int instance);
+
+static int     elan3_open (struct inode *inode, struct file *file);
+static int     elan3_ioctl (struct inode *inode, struct file *file, 
+                            unsigned int cmd, unsigned long arg);
+static int     elan3_mmap (struct file *file, struct vm_area_struct *vm_area);
+static int     elan3_release (struct inode *inode, struct file *file);
+
+static int      elan3_reboot_event (struct notifier_block *self, unsigned long event, void *buffer);
+static int      elan3_panic_event (struct notifier_block *self, unsigned long event, void *buffer);
+
+static irqreturn_t InterruptHandlerWrapper(int irq, void *dev_id, struct pt_regs *regs);
+
+static int     ConfigurePci(ELAN3_DEV *dev);
+static int     ResetElan(ELAN3_DEV *dev, ioaddr_t intPalAddr);
+
+static void     elan3_shutdown_devices(int panicing);
+
+/*
+ * Globals. 
+ */
+static ELAN3_DEV       *elan3_devices[ELAN3_MAX_CONTROLLER];
+static int       NodeId = ELAN3_INVALID_NODE;
+static int       NumNodes;
+static int       DownLinks;
+static int       RandomRoutingDisabled;
+int              BackToBackMaster;
+int              BackToBackSlave;
+int              enable_sdram_writecombining;
+int             sdram_bank_limit;
+extern int       LwpNice;
+
+char *    elan_reg_rec_file [ELAN_REG_REC_MAX];
+int       elan_reg_rec_line [ELAN_REG_REC_MAX];
+long      elan_reg_rec_lbolt[ELAN_REG_REC_MAX];
+int       elan_reg_rec_cpu  [ELAN_REG_REC_MAX];
+E3_uint32 elan_reg_rec_reg  [ELAN_REG_REC_MAX];
+int       elan_reg_rec_index;
+
+MODULE_AUTHOR("Quadrics Ltd.");
+MODULE_DESCRIPTION("Elan3 Device Driver");
+
+MODULE_LICENSE("GPL");
+
+MODULE_PARM(NodeId,"i");
+MODULE_PARM(NumNodes,"i");
+MODULE_PARM(RandomRoutingDisabled,"i");
+MODULE_PARM(DownLinks,"i");
+MODULE_PARM(BackToBackMaster,"i");
+MODULE_PARM(BackToBackSlave,"i");
+MODULE_PARM(LwpNice, "i");
+MODULE_PARM(elan3_debug, "i");
+MODULE_PARM(elan3_debug_console, "i");
+MODULE_PARM(elan3_debug_buffer, "i");
+MODULE_PARM(elan3mmu_debug, "i");
+MODULE_PARM(sdram_bank_limit, "i");
+
+/* elan3/os/context.c */
+EXPORT_SYMBOL(elan3_alloc);
+EXPORT_SYMBOL(elan3_attach);
+EXPORT_SYMBOL(elan3_doattach);
+EXPORT_SYMBOL(elan3_free);
+EXPORT_SYMBOL(elan3_detach);
+EXPORT_SYMBOL(elan3_dodetach);
+EXPORT_SYMBOL(elan3_block_inputter);
+EXPORT_SYMBOL(CheckCommandQueueFlushed);
+
+/* elan3/os/sdram.c */
+EXPORT_SYMBOL(elan3_sdram_alloc);
+EXPORT_SYMBOL(elan3_sdram_free);
+EXPORT_SYMBOL(elan3_sdram_to_phys);
+EXPORT_SYMBOL(elan3_sdram_writeb);
+EXPORT_SYMBOL(elan3_sdram_writew);
+EXPORT_SYMBOL(elan3_sdram_writel);
+EXPORT_SYMBOL(elan3_sdram_writeq);
+EXPORT_SYMBOL(elan3_sdram_readb);
+EXPORT_SYMBOL(elan3_sdram_readw);
+EXPORT_SYMBOL(elan3_sdram_readl);
+EXPORT_SYMBOL(elan3_sdram_readq);
+EXPORT_SYMBOL(elan3_sdram_zerob_sdram);
+EXPORT_SYMBOL(elan3_sdram_zerow_sdram);
+EXPORT_SYMBOL(elan3_sdram_zerol_sdram);
+EXPORT_SYMBOL(elan3_sdram_zeroq_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyb_to_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyw_to_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyl_to_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyq_to_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyb_from_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyw_from_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyl_from_sdram);
+EXPORT_SYMBOL(elan3_sdram_copyq_from_sdram);
+
+/* elan3/os/tproc.c */
+EXPORT_SYMBOL(DeliverTProcTrap);
+EXPORT_SYMBOL(HandleTProcTrap);
+EXPORT_SYMBOL(SaveThreadToStack);
+
+/* elan3/os/tprocinsts.c */
+EXPORT_SYMBOL(RollThreadToClose);
+
+/* elan3/os/iproc.c */
+EXPORT_SYMBOL(InspectIProcTrap);
+EXPORT_SYMBOL(IProcTrapString);
+EXPORT_SYMBOL(SimulateUnlockQueue);
+
+/* elan3/os/cproc.c */
+EXPORT_SYMBOL(HandleCProcTrap);
+
+/* elan3/os/route_table.c */
+EXPORT_SYMBOL(GenerateRoute);
+EXPORT_SYMBOL(LoadRoute);
+EXPORT_SYMBOL(InvalidateRoute);
+EXPORT_SYMBOL(ValidateRoute);
+EXPORT_SYMBOL(ClearRoute);
+EXPORT_SYMBOL(GenerateProbeRoute);
+EXPORT_SYMBOL(GenerateCheckRoute);
+
+/* elan3/os/elandev_generic.c */
+EXPORT_SYMBOL(elan3_debug);
+EXPORT_SYMBOL(QueueHaltOperation);
+EXPORT_SYMBOL(ReleaseHaltOperations);
+EXPORT_SYMBOL(ReserveHaltOperations);
+
+/* elan3/vm/elan3mmu_generic.c */
+EXPORT_SYMBOL(elan3mmu_pteload);
+EXPORT_SYMBOL(elan3mmu_unload);
+EXPORT_SYMBOL(elan3mmu_set_context_filter);
+EXPORT_SYMBOL(elan3mmu_reserve);
+EXPORT_SYMBOL(elan3mmu_attach);
+EXPORT_SYMBOL(elan3mmu_detach);
+EXPORT_SYMBOL(elan3mmu_release);
+/* elan3/vm/elan3mmu_linux.c */
+EXPORT_SYMBOL(elan3mmu_phys_to_pte);
+EXPORT_SYMBOL(elan3mmu_kernel_invalid_pte);
+
+/* elan3/os/elan3_debug.c */
+EXPORT_SYMBOL(elan3_debugf);
+
+/* elan3/os/minames.c */
+EXPORT_SYMBOL(MiToName);
+
+/* elan3/os/elandev_generic.c */
+EXPORT_SYMBOL(MapDeviceRegister);
+EXPORT_SYMBOL(UnmapDeviceRegister);
+
+EXPORT_SYMBOL(elan_reg_rec_lbolt);
+EXPORT_SYMBOL(elan_reg_rec_file);
+EXPORT_SYMBOL(elan_reg_rec_index);
+EXPORT_SYMBOL(elan_reg_rec_cpu);
+EXPORT_SYMBOL(elan_reg_rec_reg);
+EXPORT_SYMBOL(elan_reg_rec_line);
+
+/*
+ * Standard device entry points.
+ */
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+
+#include <linux/dump.h>
+
+static int      elan3_dump_event (struct notifier_block *self, unsigned long event, void *buffer);
+
+static struct notifier_block elan3_dump_notifier = 
+{
+    notifier_call:     elan3_dump_event,
+    priority:          0,
+};
+
+static int
+elan3_dump_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    if ( event == DUMP_BEGIN )
+       elan3_shutdown_devices (FALSE);
+
+    return (NOTIFY_DONE);
+}
+
+#endif
+
+static struct file_operations elan3_fops = {
+        ioctl:   elan3_ioctl,          /* ioctl */
+        mmap:    elan3_mmap,           /* mmap */
+        open:    elan3_open,           /* open */
+        release: elan3_release,                /* release */
+};
+
+static struct notifier_block elan3_reboot_notifier = 
+{
+    notifier_call:     elan3_reboot_event,
+    priority:          0,
+};
+
+static struct notifier_block elan3_panic_notifier = 
+{
+    notifier_call:     elan3_panic_event,
+    priority:          0,
+};
+
+ELAN3_DEV *
+elan3_device (int instance)
+{
+       if (instance < 0 || instance >= ELAN3_MAX_CONTROLLER)
+           return ((ELAN3_DEV *) NULL);
+       return elan3_devices[instance];
+}
+EXPORT_SYMBOL(elan3_device);
+
+/*
+ * Called at rmmod time.  elandetach() for each card + general cleanup.
+ */
+#ifdef MODULE
+static void __exit elan3_exit(void)
+{
+       int i;
+
+       printk("elan: preparing to remove module\n");
+
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+       unregister_dump_notifier (&elan3_dump_notifier);
+#endif
+       unregister_reboot_notifier (&elan3_reboot_notifier);
+       notifier_chain_unregister (&panic_notifier_list, &elan3_panic_notifier);
+
+       /* call elandetach() for each device configured. */
+       for (i = 0; i < ELAN3_MAX_CONTROLLER; i++)
+               if (elan3_devices[i] != NULL)
+                       elandetach(i);
+
+       FinaliseNetworkErrorResolver();
+       elan3mmu_fini();
+
+       cookie_fini();
+       unregister_chrdev(ELAN3_MAJOR, ELAN3_NAME);
+
+       elan3_procfs_fini();
+
+       printk("elan: module removed\n");
+}
+
+/*
+ * Called at insmod time.  First we perform general driver initialization,
+ * then call elanattach() for each card.
+ */
+#ifdef MODULE
+static int __init elan3_init(void)
+#else
+__initfunc(int elan3_init(void))
+#endif
+{
+       int e;
+       int boards;
+       struct pci_dev *dev;
+       char revid;
+
+       elan_reg_rec_index=0;
+       {
+           int i;
+           for(i=0;i<ELAN_REG_REC_MAX;i++)
+               elan_reg_rec_file[i] = NULL;
+       }       
+
+       /* register major/minor num */
+       e = register_chrdev(ELAN3_MAJOR, ELAN3_NAME, &elan3_fops);
+       if (e < 0)
+               return e;
+
+       elan3_procfs_init ();
+
+       cookie_init();
+       elan3mmu_init();
+       InitialiseNetworkErrorResolver();
+
+       /* call elanattach() for each device found on PCI */
+       memset(elan3_devices, 0, sizeof(elan3_devices));
+       boards = 0;
+       for (dev = NULL; (dev = pci_find_device(PCI_VENDOR_ID_QUADRICS, PCI_DEVICE_ID_ELAN3, dev)) != NULL ;) 
+       {
+           pci_read_config_byte (dev, PCI_REVISION_ID, &revid);
+
+           if (revid == PCI_REVISION_ID_ELAN3_REVA)
+               printk ("elan at pci %s - RevA device not supported\n", dev->slot_name);
+           else
+           {
+               if (boards < ELAN3_MAX_CONTROLLER)
+                       /* Count successfully attached devices */ 
+                       boards += ((elanattach(boards, dev) == 0) ? 1 : 0);
+               else
+               {
+                   printk ("elan: max controllers = %d\n", ELAN3_MAX_CONTROLLER);
+                   break;
+               }
+           }
+       }
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+       register_dump_notifier (&elan3_dump_notifier);
+#endif
+       register_reboot_notifier (&elan3_reboot_notifier);
+       notifier_chain_register (&panic_notifier_list, &elan3_panic_notifier);
+
+       return 0;
+}
+
+/* Declare the module init and exit functions */
+module_init(elan3_init);
+module_exit(elan3_exit);
+
+#endif
+
+static void
+elan3_shutdown_devices(int panicing)
+{
+    ELAN3_DEV *dev;
+    unsigned long flags;
+    register int i;
+
+    local_irq_save (flags);
+    for (i = 0; i < ELAN3_MAX_CONTROLLER; i++)
+    {
+       if ((dev = elan3_devices[i]) != NULL)
+       {
+           if (! panicing) spin_lock (&dev->IntrLock);
+           
+           printk(KERN_INFO "elan%d: forcing link into reset\n", dev->Instance);
+
+           /*
+            * We're going to set the link into boundary scan mode,  so firstly
+            * set the inputters to discard everything.
+            */
+           if (dev->DiscardAllCount++ == 0)
+               SetSchedStatusRegister (dev, read_reg32 (dev, Exts.InterruptReg), NULL);
+
+           dev->LinkShutdown = 1;
+           
+           /*
+            * Now disable the error interrupts
+            */
+           DISABLE_INT_MASK (dev, INT_ErrorInterrupts);
+           
+           /*
+            * And set the link into boundary scan mode, and drive
+            * a reset token onto the link.
+            */
+           SET_SCHED_LINK_VALUE (dev, 1, LinkResetToken);
+
+           if (! panicing) spin_unlock (&dev->IntrLock);
+       }
+    }
+    local_irq_restore (flags);
+}
+
+static int
+elan3_reboot_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    if (! (event == SYS_RESTART || event == SYS_HALT || event == SYS_POWER_OFF))
+       return (NOTIFY_DONE);
+
+    elan3_shutdown_devices (FALSE);
+
+    return (NOTIFY_DONE);
+}
+
+static int
+elan3_panic_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    elan3_shutdown_devices (TRUE);
+
+    return (NOTIFY_DONE);
+}
+
+#include <elan3/elan3ops.h>
+/*
+ * Called by init_module() for each card discovered on PCI.
+ */
+static int
+elanattach(int instance, struct pci_dev *pcidev)
+{
+       ELAN3_DEV *dev;
+       int ramSize;
+       int level;
+       ioaddr_t sdramAddr, cmdPortAddr, intPalAddr;
+       DeviceMappingHandle handle;
+
+       printk("elan%d: attach, irq=%d\n", instance, pcidev->irq);
+
+       /*
+        * Allocate the ELAN3_DEV structure.
+        */
+       KMEM_ZALLOC(dev, ELAN3_DEV *, sizeof(ELAN3_DEV), TRUE);
+       if (dev == NULL) {
+               printk ("elan%d: KMEM_ALLOC failed\n", instance);
+               return (-ENOMEM);
+       }
+       elan3_devices[instance] = dev;
+       dev->Osdep.pci = pcidev;
+
+       dev->Instance = instance;
+
+       /* Initialise the device information */
+       pci_read_config_word (pcidev, PCI_VENDOR_ID,   &dev->Devinfo.dev_vendor_id);
+       pci_read_config_word (pcidev, PCI_DEVICE_ID,   &dev->Devinfo.dev_device_id);
+       pci_read_config_byte (pcidev, PCI_REVISION_ID, &dev->Devinfo.dev_revision_id);
+
+       dev->Devinfo.dev_instance             = instance;
+       dev->Devinfo.dev_rail                 = instance;
+       dev->Devinfo.dev_driver_version       = 0;
+       dev->Devinfo.dev_num_down_links_value = DownLinks;
+
+       dev->Position.pos_mode                = ELAN_POS_UNKNOWN;
+       dev->Position.pos_random_disabled     = RandomRoutingDisabled;
+       
+       /*
+        * Set up PCI config regs.
+        */
+       if (ConfigurePci(dev) != ESUCCESS)
+           goto fail0;
+
+       /*
+        * Determine the PFnums of the SDRAM and command port
+        */
+       if (MapDeviceRegister(dev, ELAN3_BAR_SDRAM, &sdramAddr, 0, PAGESIZE, &handle) != ESUCCESS)
+           goto fail1;
+
+       DeviceRegisterSize(dev, ELAN3_BAR_SDRAM, &ramSize);
+       
+       dev->SdramPhysMask = ~((physaddr_t) ramSize - 1);
+       dev->SdramPhysBase = kmem_to_phys((void *) sdramAddr);
+
+       UnmapDeviceRegister (dev, &handle);
+
+#if defined(LINUX_ALPHA)
+       /*
+        * consider a physical address to be on the same pci bus
+        * as us if it's physical address is "close" to our sdram
+        * physical address.
+        * this is almost certainly incorrect for large memory (> 2Gb)
+        * i386 machines - and is only correct for alpha for 32 bit
+        * base address registers.
+        *
+        * Modified this to match the Tru64 driver value;
+        * i.e. PciPhysMask = 0xfffffffffffc0000
+        */
+#  define PCI_ADDR_MASK (0x7FFFFFFFl)
+
+       dev->PciPhysMask = ~PCI_ADDR_MASK;
+       dev->PciPhysBase = dev->SdramPhysBase & dev->PciPhysMask;
+#endif
+       /*
+        * Now reset the elan chip.
+        */
+       if (MapDeviceRegister(dev, ELAN3_BAR_REGISTERS, &dev->RegPtr, 0, 0, &dev->RegHandle) != ESUCCESS)
+           goto fail1;
+
+       if (MapDeviceRegister(dev, ELAN3_BAR_EBUS, &intPalAddr, ELAN3_EBUS_INTPAL_OFFSET, PAGESIZE,
+                             &handle) != ESUCCESS)
+           goto fail2;
+
+       ResetElan(dev, intPalAddr);     
+
+       UnmapDeviceRegister (dev, &handle);
+
+       /* 
+        * Initialise the device mutex's which must be accessible from the 
+        * interrupt handler.  
+        */
+       kcondvar_init (&dev->IntrWait);
+       spin_lock_init (&dev->IntrLock);
+       spin_lock_init (&dev->TlbLock);
+       spin_lock_init (&dev->CProcLock);
+       spin_lock_init (&dev->FreeHaltLock);
+       for(level=0; level<4; level++)
+           spin_lock_init (&dev->Level[level].PtblLock);
+       spin_lock_init (&dev->PtblGroupLock);
+
+       /*
+        * Add the interrupt handler,  
+        */
+       if (request_irq(dev->Osdep.pci->irq, InterruptHandlerWrapper, 
+           SA_SHIRQ, "elan3", dev) != 0) {
+               printk ("elan%d: request_irq failed\n", instance);
+               goto fail3;
+       }
+
+       if (MapDeviceRegister(dev, ELAN3_BAR_COMMAND_PORT, &cmdPortAddr, 0, PAGESIZE, &handle) != ESUCCESS)
+           goto fail4;
+       
+       if (InitialiseElan(dev, cmdPortAddr) == EFAIL) {
+               printk ("elan%d: InitialiseElan failed\n", instance);
+               UnmapDeviceRegister (dev, &handle);
+               goto fail4;
+       }
+       UnmapDeviceRegister (dev, &handle);
+
+       /* If our nodeid is defined, then set it now */
+       if (NodeId != ELAN3_INVALID_NODE && ComputePosition (&dev->Position, NodeId, NumNodes, DownLinks) == 0)
+       {
+           if (RandomRoutingDisabled & ((1 << (dev->Position.pos_levels-1))-1))
+               printk ("elan%d: NodeId=%d NodeLevel=%d NumNodes=%d (random routing disabled 0x%x)\n", 
+                       dev->Instance, dev->Position.pos_nodeid, dev->Position.pos_levels, dev->Position.pos_nodes, RandomRoutingDisabled);
+           else
+               printk ("elan%d: NodeId=%d NodeLevel=%d NumNodes=%d (random routing ok)\n",
+                       dev->Instance, dev->Position.pos_nodeid, dev->Position.pos_levels, dev->Position.pos_nodes);
+       }
+
+       if (BackToBackMaster || BackToBackSlave)
+       {
+           dev->Position.pos_mode     = ELAN_POS_MODE_BACKTOBACK;
+           dev->Position.pos_nodeid   = (BackToBackMaster == 0);
+           dev->Position.pos_nodes    = 2;
+           dev->Position.pos_levels   = 1;
+           dev->Position.pos_arity[0] = 2;
+
+           printk ("elan%d: back-to-back %s - elan node %d\n", dev->Instance,
+                   BackToBackMaster ? "master" : "slave", dev->Position.pos_nodeid);
+       }
+
+       elan3_procfs_device_init (dev);
+       
+       /* Success */
+       return (0);
+
+fail4:
+       free_irq(dev->Osdep.pci->irq, dev);
+
+fail3:
+       kcondvar_destroy (&dev->IntrWait);
+       spin_lock_destroy (&dev->IntrLock);
+       spin_lock_destroy (&dev->InfoLock);
+       spin_lock_destroy (&dev->TlbLock);
+       spin_lock_destroy (&dev->CProcLock);
+       spin_lock_destroy (&dev->FreeHaltLock);
+       spin_lock_destroy (&dev->Level1PtblLock);
+       spin_lock_destroy (&dev->Level2PtblLock);
+       spin_lock_destroy (&dev->Level3PtblLock);
+       spin_lock_destroy (&dev->PtblGroupLock);
+
+fail2:
+       UnmapDeviceRegister (dev, &dev->RegHandle);
+
+fail1:
+       pci_disable_device (dev->Osdep.pci);
+fail0:
+       KMEM_FREE(dev, sizeof(ELAN3_DEV));
+
+       elan3_devices[instance] = NULL;
+       
+       /* Failure */
+       return (-ENODEV);
+}
+
+/*
+ * Called by elan3_exit() for each board found on PCI.
+ */
+static int
+elandetach(int instance)
+{
+       ELAN3_DEV *dev = elan3_devices[instance];
+
+       printk("elan%d: detach\n", instance);
+
+       elan3_procfs_device_fini (dev);
+
+       FinaliseElan (dev);
+
+       UnmapDeviceRegister (dev, &dev->RegHandle);
+
+       free_irq(dev->Osdep.pci->irq, dev);
+
+       pci_disable_device(dev->Osdep.pci);
+
+       kcondvar_destroy (&dev->IntrWait);
+       spin_lock_destroy (&dev->IntrLock);
+       spin_lock_destroy (&dev->InfoLock);
+       spin_lock_destroy (&dev->TlbLock);
+       spin_lock_destroy (&dev->CProcLock);
+       spin_lock_destroy (&dev->FreeHaltLock);
+       spin_lock_destroy (&dev->Level1PtblLock);
+       spin_lock_destroy (&dev->Level2PtblLock);
+       spin_lock_destroy (&dev->Level3PtblLock);
+       spin_lock_destroy (&dev->PtblGroupLock);
+
+       KMEM_FREE(dev, sizeof(ELAN3_DEV));
+       elan3_devices[instance] = NULL; 
+
+       return 0;
+}
+
+/*
+ * generic ioctls - available on control and user devices.
+ */
+
+static int
+device_stats_ioctl (ELAN3_DEV *dev, unsigned long arg)
+{
+    ELAN3IO_STATS_STRUCT *args;
+
+    KMEM_ALLOC(args, ELAN3IO_STATS_STRUCT *, sizeof(ELAN3IO_STATS_STRUCT), TRUE);
+       
+    if (args == NULL)
+       return (-ENOMEM);
+
+    if (copy_from_user (args, (void *) arg, sizeof (ELAN3IO_STATS_STRUCT)))
+    {
+       KMEM_FREE(args, sizeof(ELAN3IO_STATS_STRUCT));
+       return (-EFAULT);
+    }
+
+    switch (args->which)
+    {
+    case ELAN3_SYS_STATS_DEVICE:
+       if (copy_to_user (args->ptr, &dev->Stats, sizeof (ELAN3_STATS)))
+       {
+           KMEM_FREE(args, sizeof(ELAN3IO_STATS_STRUCT));
+           return (-EFAULT);
+       }
+       KMEM_FREE(args, sizeof(ELAN3IO_STATS_STRUCT));
+       return (0);
+
+    case ELAN3_SYS_STATS_MMU:
+       if (copy_to_user (args->ptr, &elan3mmu_global_stats, sizeof (ELAN3MMU_GLOBAL_STATS)))
+       {
+           KMEM_FREE(args, sizeof(ELAN3IO_STATS_STRUCT));
+           return (-EFAULT);
+       }
+       KMEM_FREE(args, sizeof(ELAN3IO_STATS_STRUCT));
+       return (0);
+           
+    default:
+       KMEM_FREE(args, sizeof(ELAN3IO_STATS_STRUCT));
+       return (-EINVAL);
+    }
+}
+
+/*
+ * /dev/elan3/controlX - control device
+ *
+ */
+
+typedef struct control_private
+{
+    u_int              pr_boundary_scan;
+} CONTROL_PRIVATE;
+
+static int
+control_open (struct inode *inode, struct file *file)
+{
+    CONTROL_PRIVATE *pr;
+
+    KMEM_ALLOC(pr, CONTROL_PRIVATE *, sizeof (CONTROL_PRIVATE), TRUE);
+
+    if (pr == NULL)
+       return (-ENOMEM);
+
+    pr->pr_boundary_scan = 0;
+    
+    file->private_data = (void *) pr;
+
+    MOD_INC_USE_COUNT;
+
+    return (0);
+}
+
+static int
+control_release (struct inode *inode, struct file *file)
+{
+    ELAN3_DEV        *dev = elan3_devices[ELAN3_DEVICE(inode)];
+    CONTROL_PRIVATE *pr  = (CONTROL_PRIVATE *) file->private_data;
+
+    if (pr->pr_boundary_scan)
+       ClearLinkBoundaryScan(dev);
+
+    KMEM_FREE (pr, sizeof(CONTROL_PRIVATE));
+
+    MOD_DEC_USE_COUNT;
+    return (0);
+}
+
+static int
+control_ioctl (struct inode *inode, struct file *file, 
+              unsigned int cmd, unsigned long arg)
+{
+    ELAN3_DEV        *dev = elan3_devices[ELAN3_DEVICE(inode)];
+    CONTROL_PRIVATE *pr  = (CONTROL_PRIVATE *) file->private_data;
+    int                     res;
+
+    switch (cmd) 
+    {
+    case ELAN3IO_SET_BOUNDARY_SCAN:
+       if (SetLinkBoundaryScan (dev) == 0)
+           pr->pr_boundary_scan = 1;
+       return (0);
+
+    case ELAN3IO_CLEAR_BOUNDARY_SCAN:
+       if (pr->pr_boundary_scan == 0)
+           return (-EINVAL);
+
+       pr->pr_boundary_scan = 0;
+
+       ClearLinkBoundaryScan (dev);
+       return (0);
+
+    case ELAN3IO_READ_LINKVAL:
+    {
+       E3_uint32 val;
+
+       if (pr->pr_boundary_scan == 0)
+           return (-EINVAL);
+
+       if (copy_from_user(&val, (E3_uint32 *)arg, sizeof(E3_uint32)))
+           return (-EFAULT);
+
+       val = ReadBoundaryScanValue (dev, val);
+
+       if (copy_to_user((E3_uint32 *)arg, &val, sizeof(E3_uint32)))
+           return (-EFAULT);
+       return (0);
+    }
+       
+    case ELAN3IO_WRITE_LINKVAL:
+    {
+       E3_uint32 val;
+
+       if (pr->pr_boundary_scan == 0)
+           return (-EINVAL);
+
+       if (copy_from_user(&val, (E3_uint32 *)arg, sizeof(E3_uint32)))
+           return (-EFAULT);
+
+       val = WriteBoundaryScanValue (dev, val);
+
+       if (copy_to_user((E3_uint32 *)arg, &val, sizeof(E3_uint32)))
+           return (-EFAULT);
+       
+       return (0);
+    }
+
+    case ELAN3IO_SET_POSITION:
+    {
+       ELAN3IO_SET_POSITION_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_SET_POSITION_STRUCT)))
+           return (-EFAULT);
+       
+       if (ComputePosition (&dev->Position, args.nodeId, args.numNodes, dev->Devinfo.dev_num_down_links_value) != 0)
+           return (-EINVAL);
+
+       return (0);
+    }
+
+    case ELAN3IO_SET_DEBUG:
+    {
+       ELAN3IO_SET_DEBUG_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_SET_DEBUG_STRUCT)))
+           return (-EFAULT);
+
+       if (! strcmp (args.what, "elan3_debug"))
+           elan3_debug = args.value;
+       else if (! strcmp (args.what, "elan3_debug_console"))
+           elan3_debug_console = args.value;
+       else if (! strcmp (args.what, "elan3_debug_buffer"))
+           elan3_debug_buffer = args.value;
+       else if (! strcmp (args.what, "elan3_debug_ignore_dev"))
+           elan3_debug_ignore_dev = args.value;
+       else if (! strcmp (args.what, "elan3_debug_ignore_ctxt"))
+           elan3_debug_ignore_ctxt = args.value;
+       else if (! strcmp (args.what, "elan3mmu_debug"))
+           elan3mmu_debug = args.value;
+       
+       return (0);
+    }
+
+    case ELAN3IO_NETERR_SERVER:
+    {
+       ELAN3IO_NETERR_SERVER_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_NETERR_SERVER_STRUCT)))
+           return (-EFAULT);
+       
+       res = AddNeterrServerSyscall (args.elanid, args.addr, args.name, NULL);
+       return (set_errno (res));
+    }
+    
+    case ELAN3IO_NETERR_FIXUP:
+    {
+       NETERR_MSG *msg;
+
+       KMEM_ALLOC(msg, NETERR_MSG *, sizeof (NETERR_MSG), TRUE);
+
+       if (msg == NULL)
+           return (set_errno (ENOMEM));
+       
+       if (copy_from_user (msg, (void *) arg, sizeof (NETERR_MSG)))
+           res = EFAULT;
+       else
+           res = ExecuteNetworkErrorFixup (msg);
+
+       KMEM_FREE (msg, sizeof (NETERR_MSG));
+       return (set_errno (res));
+    }
+
+    case ELAN3IO_STATS:
+       return (device_stats_ioctl (dev, arg));
+
+    case ELAN3IO_GET_DEVINFO:
+    {
+       if (copy_to_user ((void *) arg, &dev->Devinfo, sizeof (ELAN_DEVINFO)))
+           return (-EFAULT);
+       return (0);
+    }
+
+    case ELAN3IO_GET_POSITION:
+    {
+       if (copy_to_user ((void *) arg, &dev->Position, sizeof (ELAN_POSITION)))
+           return (-EFAULT);
+       return (0);
+    }
+    default:
+       return (-EINVAL);
+    }
+}
+
+static int
+control_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    ELAN3_DEV         *dev   = elan3_devices[ELAN3_DEVICE(file->f_dentry->d_inode)];
+    int                space = OFF_TO_SPACE(vma->vm_pgoff << PAGE_SHIFT);
+    int                off   = OFF_TO_OFFSET(vma->vm_pgoff << PAGE_SHIFT);
+    int                size;
+    ioaddr_t           addr;
+    DeviceMappingHandle handle;
+    physaddr_t         phys;
+    
+    if (space < ELAN3_BAR_SDRAM || space > ELAN3_BAR_EBUS)
+       return (-EINVAL);
+
+    if (off < 0 || DeviceRegisterSize (dev, space, &size) != ESUCCESS || off > size)
+       return (-EINVAL);
+
+    if (MapDeviceRegister(dev, space, &addr, off, PAGESIZE, &handle) != ESUCCESS)
+       return (-EINVAL);
+
+    phys = kmem_to_phys((caddr_t) addr);
+    UnmapDeviceRegister(dev, &handle);
+
+#ifdef NO_RMAP
+    if (remap_page_range(vma->vm_start, phys, vma->vm_end - vma->vm_start, vma->vm_page_prot))
+#else
+    if (remap_page_range(vma, vma->vm_start, phys, vma->vm_end - vma->vm_start, vma->vm_page_prot))
+#endif
+       return (-EAGAIN);
+
+    return (0);
+}
+
+/*
+ * /dev/elan3/sdramX - sdram access device
+ */
+typedef struct mem_page
+{
+    struct mem_page *pg_next;
+    sdramaddr_t      pg_addr;
+    u_long          pg_pgoff;
+    u_int           pg_ref;
+} MEM_PAGE;
+
+#define MEM_HASH_SIZE  32
+#define MEM_HASH(pgoff)        ((pgoff) & (MEM_HASH_SIZE-1))
+
+typedef struct mem_private
+{
+    ELAN3_DEV   *pr_dev;
+    MEM_PAGE   *pr_pages[MEM_HASH_SIZE];
+    spinlock_t  pr_lock;
+} MEM_PRIVATE;
+
+static void 
+mem_freepage (MEM_PRIVATE *pr, MEM_PAGE *pg)
+{
+    PRINTF (DBG_DEVICE, DBG_SEG, "mem_freepage: pr=%p pgoff=%lx pg=%p ref=%d\n", pr, pg->pg_pgoff, pg, pg->pg_ref);
+
+    elan3_sdram_free (pr->pr_dev, pg->pg_addr, PAGE_SIZE);
+    KMEM_FREE (pg, sizeof(MEM_PAGE));
+}
+
+static MEM_PAGE *
+mem_getpage (MEM_PRIVATE *pr, u_long pgoff, virtaddr_t addr)
+{
+    int       hashval = MEM_HASH (pgoff);
+    MEM_PAGE *npg = NULL;
+    MEM_PAGE *pg;
+
+    PRINTF (DBG_DEVICE, DBG_SEG, "mem_getpage: pr=%p pgoff=%lx addr=%lx\n", pr, pgoff, addr);
+    
+ again:
+    spin_lock (&pr->pr_lock);
+    for (pg = pr->pr_pages[hashval]; pg; pg = pg->pg_next)
+       if (pg->pg_pgoff == pgoff)
+           break;
+    
+    if (pg != NULL)
+    {
+       PRINTF (DBG_DEVICE, DBG_SEG, "mem_getpage: pr=%p pgoff=%lx addr=%lx -> found %p addr=%lx\n", pr, pgoff, addr, pg, pg->pg_addr);
+
+       pg->pg_ref++;
+       spin_unlock (&pr->pr_lock);
+
+       if (npg != NULL)                                        /* we'd raced and someone else had created */
+           mem_freepage (pr, npg);                             /* this page - so free of our new one*/
+       return (pg);
+    }
+    
+    if (npg != NULL)                                           /* didn't find the page, so inset the */
+    {                                                          /* new one we've just created */
+       npg->pg_next = pr->pr_pages[hashval];
+       pr->pr_pages[hashval] = npg;
+       
+       spin_unlock (&pr->pr_lock);
+       return (npg);
+    }
+    
+    spin_unlock (&pr->pr_lock);                                        /* drop spinlock before creating a new page */
+    
+    KMEM_ALLOC(npg, MEM_PAGE *, sizeof (MEM_PAGE), TRUE);
+
+    if (npg == NULL)
+       return (NULL);
+
+    if ((npg->pg_addr = elan3_sdram_alloc (pr->pr_dev, PAGE_SIZE)) == 0)
+    {
+       KMEM_FREE (npg, sizeof (MEM_PAGE));
+       return (NULL);
+    }
+
+    /* zero the page before returning it to the user */
+    elan3_sdram_zeroq_sdram (pr->pr_dev, npg->pg_addr, PAGE_SIZE);
+    
+    npg->pg_pgoff = pgoff;
+    npg->pg_ref   = 1;
+    
+    /* created a new page - so have to rescan before inserting it */
+    goto again;
+}
+
+static void
+mem_droppage (MEM_PRIVATE *pr, u_long pgoff, int dontfree)
+{
+    MEM_PAGE **ppg;
+    MEM_PAGE  *pg;
+
+    spin_lock (&pr->pr_lock);
+    for (ppg = &pr->pr_pages[MEM_HASH(pgoff)]; *ppg; ppg = &(*ppg)->pg_next)
+       if ((*ppg)->pg_pgoff == pgoff)
+           break;
+
+    pg = *ppg;
+
+    ASSERT (*ppg != NULL);
+    
+    PRINTF (DBG_DEVICE, DBG_SEG, "mem_droppage: pr=%p pgoff=%lx pg=%p ref=%d dontfree=%d\n", pr, pgoff, (*ppg), (*ppg)->pg_ref, dontfree);
+
+    if (--pg->pg_ref == 0 && !dontfree)
+    {
+       *ppg = pg->pg_next;
+
+       mem_freepage (pr, pg);
+    }
+
+    spin_unlock (&pr->pr_lock);
+}
+
+static int
+mem_open (struct inode *inode, struct file *file)
+{
+    ELAN3_DEV    *dev = elan3_devices[ELAN3_DEVICE(inode)];
+    MEM_PRIVATE *pr;
+    register int i;
+
+    KMEM_ALLOC(pr, MEM_PRIVATE *, sizeof (MEM_PRIVATE), TRUE);
+
+    if (pr == NULL)
+       return (-ENOMEM);
+
+    spin_lock_init (&pr->pr_lock);
+    pr->pr_dev = dev;
+    for (i = 0; i < MEM_HASH_SIZE; i++)
+       pr->pr_pages[i] = NULL;
+
+    file->private_data = (void *) pr;
+    
+    MOD_INC_USE_COUNT;
+    return (0);
+}
+
+static int
+mem_release (struct inode *node, struct file *file)
+{
+    MEM_PRIVATE *pr = (MEM_PRIVATE *) file->private_data;
+    MEM_PAGE    *pg, *next;
+    int          i;
+
+    /* free off any pages that we'd allocated */
+    spin_lock (&pr->pr_lock);
+    for (i = 0; i < MEM_HASH_SIZE; i++)
+    {
+       for (pg = pr->pr_pages[i]; pg; pg = next)
+       {
+           next = pg->pg_next;
+           mem_freepage (pr, pg);
+       }
+    }
+    spin_unlock (&pr->pr_lock);
+
+    KMEM_FREE (pr, sizeof (MEM_PRIVATE));
+
+    MOD_DEC_USE_COUNT;
+    return (0);
+}
+
+static int
+mem_ioctl (struct inode *inode, struct file *file, 
+                 unsigned int cmd, unsigned long arg)
+{
+    return (-EINVAL);
+}
+
+static void mem_vma_open(struct vm_area_struct *vma)
+{
+    MEM_PRIVATE   *pr = (MEM_PRIVATE *) vma->vm_private_data;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (DBG_DEVICE, DBG_SEG, "mem_vma_open: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+
+    preemptable_start {
+       for (addr = vma->vm_start, pgoff = vma->vm_pgoff; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++) {
+           mem_getpage (pr, pgoff, addr);
+           preemptable_check();
+       }
+    } preemptable_end;
+}
+
+static void mem_vma_close(struct vm_area_struct *vma)
+{
+    MEM_PRIVATE  *pr  = (MEM_PRIVATE *) vma->vm_private_data;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (DBG_DEVICE, DBG_SEG, "mem_vma_close: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+
+    /* NOTE: the call to close may not have the same vm_start/vm_end values as 
+     *       were passed into mmap()/open() - since if an partial unmap had occured
+     *       then the vma could have been shrunk or even split.
+     *
+     *       if a the vma is split then an vma_open() will be called for the top
+     *       portion - thus causing the reference counts to become incorrect.
+     *
+     * We drop the reference to any pages we're notified about - so they get freed
+     * earlier than when the device is finally released.
+     */
+    for (pgoff = vma->vm_pgoff, addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+       mem_droppage (pr, pgoff, 0);
+}
+
+static struct vm_operations_struct mem_vm_ops = {
+    open:              mem_vma_open,
+    close:             mem_vma_close,
+};
+
+static int
+mem_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    MEM_PRIVATE  *pr = (MEM_PRIVATE *) file->private_data;
+    MEM_PAGE     *pg;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (DBG_DEVICE, DBG_SEG, "mem_mmap: vm_mm=%p start=%lx end=%lx pgoff=%lx prot=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_page_prot.pgprot , file);
+
+    preemptable_start {
+       for (addr = vma->vm_start, pgoff = vma->vm_pgoff; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+       {
+           if ((pg = mem_getpage (pr, pgoff, addr)) == NULL)
+               goto failed;
+           
+#ifdef LINUX_SPARC
+           pgprot_val(vma->vm_page_prot) &= ~(_PAGE_CACHE);
+           pgprot_val(vma->vm_page_prot) |= _PAGE_IE;
+#elif defined(pgprot_noncached)
+           vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+#endif
+           
+#if defined(__ia64__)
+           if (enable_sdram_writecombining)
+               vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
+#endif
+           PRINTF (DBG_DEVICE, DBG_SEG, "mem_mmap: addr %lx -> pg=%p addr=%lx phys=%lx flags=%lx prot=%lx\n",
+                   addr, pg, pg->pg_addr, elan3_sdram_to_phys (pr->pr_dev, pg->pg_addr), vma->vm_flags, vma->vm_page_prot.pgprot);
+           
+#ifdef NO_RMAP
+           if (remap_page_range (addr, elan3_sdram_to_phys (pr->pr_dev, pg->pg_addr), PAGE_SIZE, vma->vm_page_prot))
+#else
+           if (remap_page_range (vma, addr, elan3_sdram_to_phys (pr->pr_dev, pg->pg_addr), PAGE_SIZE, vma->vm_page_prot))
+#endif
+           {
+               mem_droppage (pr, pgoff, 0);                    /* drop our reference to this page */
+               goto failed;
+           }
+
+           preemptable_check();
+       }
+    } preemptable_end;
+
+    /* Don't try to swap out Elan SDRAM pages.. */
+    vma->vm_flags |= VM_RESERVED;
+    
+    /*
+     * Don't dump SDRAM pages to a core file 
+     * (Pity I would really like to do this but it crashes in elf_core_dump() as
+     * it can only handle pages that are in the mem_map area (addy 11/01/2002))
+     */
+    vma->vm_flags |= VM_IO;
+
+    vma->vm_ops          = &mem_vm_ops;
+    vma->vm_file         = file;
+    vma->vm_private_data = (void *) pr;
+
+    return (0);
+
+ failed:
+    PRINTF (DBG_DEVICE, DBG_SEG, "mem_mmap: failed\n");
+
+    /* free of any pages we've already allocated/referenced */
+    while ((--pgoff) >= vma->vm_pgoff)
+       mem_droppage (pr, pgoff, 0);
+
+    return (-ENOMEM);
+}
+
+/*
+ * /dev/elan3/userX - control device
+ *
+ * "user_private" can be referenced from a number of places
+ *   1) the "file" structure.
+ *   2) the "mm" coproc ops
+ *   3) the "mmap" of the command port.
+ *
+ */
+typedef struct user_private
+{
+    spinlock_t        pr_lock;
+    atomic_t         pr_mappings;
+    atomic_t          pr_ref;
+    ELAN3_CTXT        *pr_ctxt;
+    struct mm_struct *pr_mm;
+    coproc_ops_t      pr_coproc;
+} USER_PRIVATE;
+
+static void
+user_free (USER_PRIVATE *pr)
+{
+    /* Have to unreserve the FlagPage or else we leak memory like a sieve! */
+    ClearPageReserved(pte_page(*find_pte_kernel((unsigned long) pr->pr_ctxt->FlagPage)));
+
+    elan3_detach(pr->pr_ctxt);
+    elan3_free (pr->pr_ctxt);
+
+    KMEM_FREE (pr, sizeof(USER_PRIVATE));
+
+    MOD_DEC_USE_COUNT;
+}
+
+static void
+user_coproc_release (void *arg, struct mm_struct *mm)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF3 (pr->pr_ctxt, DBG_SEG, "user_coproc_release: ctxt=%p pr=%p ref=%d\n",
+            pr->pr_ctxt, pr, atomic_read (&pr->pr_ref));
+
+    elan3mmu_pte_ctxt_unload (pr->pr_ctxt->Elan3mmu);
+
+    pr->pr_mm = NULL;
+
+    if (atomic_dec_and_test (&pr->pr_ref))
+       user_free (pr);
+}
+
+static void
+user_coproc_sync_range (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF2 (pr->pr_ctxt, DBG_SEG, "user_coproc_sync_range: start=%lx end=%lx\n", start, end);
+
+    ASSERT(start <= end);
+
+    elan3mmu_pte_range_unload(pr->pr_ctxt->Elan3mmu, mm, (caddr_t) start, end-start);
+}
+
+static void
+user_coproc_invalidate_range (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF2 (pr->pr_ctxt, DBG_SEG, "user_coproc_invalidate_range: start=%lx end=%lx\n", start, end);
+
+    ASSERT(start <= end);
+
+    elan3mmu_pte_range_unload(pr->pr_ctxt->Elan3mmu, mm, (caddr_t) start, end-start);
+}
+
+static void
+user_coproc_update_range (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    ASSERT(start <= end && ((start & PAGEOFFSET) == 0) && ((end & PAGEOFFSET) == 0));
+
+    PRINTF2 (pr->pr_ctxt, DBG_SEG, "user_coproc_update_range: start=%lx end=%lx\n", start, end);
+
+    elan3mmu_pte_range_update (pr->pr_ctxt->Elan3mmu, mm,(caddr_t) start, end-start);
+}
+
+static void
+user_coproc_change_protection (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end, pgprot_t newprot)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF2 (pr->pr_ctxt, DBG_SEG, "user_coproc_change_protection: start=%lx end=%lx\n", start, end);
+
+    ASSERT(start <= end);
+
+    elan3mmu_pte_range_unload(pr->pr_ctxt->Elan3mmu, mm, (caddr_t) start, end-start);
+}
+
+static void
+user_coproc_sync_page (void *arg, struct vm_area_struct *vma, unsigned long addr)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF1 (pr->pr_ctxt, DBG_SEG, "user_coproc_sync_page: addr=%lx\n", addr);
+
+    elan3mmu_pte_range_unload(pr->pr_ctxt->Elan3mmu, vma->vm_mm, (caddr_t) (addr & PAGE_MASK), PAGE_SIZE);
+}
+
+static void
+user_coproc_invalidate_page (void *arg, struct vm_area_struct *vma, unsigned long addr)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+    
+    PRINTF1 (pr->pr_ctxt, DBG_SEG, "user_coproc_invalidate_page: addr=%lx\n", addr);
+
+    elan3mmu_pte_range_unload(pr->pr_ctxt->Elan3mmu, vma->vm_mm, (caddr_t) (addr & PAGE_MASK), PAGE_SIZE);
+}
+
+static void
+user_coproc_update_page (void *arg, struct vm_area_struct *vma, unsigned long addr)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF1 (pr->pr_ctxt, DBG_SEG, "user_coproc_update_page: addr=%lx\n", addr);
+
+    elan3mmu_pte_range_update (pr->pr_ctxt->Elan3mmu,vma->vm_mm, (caddr_t) (addr & PAGE_MASK), PAGE_SIZE);
+}
+
+int
+user_ptrack_handler (void *arg, int phase, struct task_struct *child)
+{
+    USER_PRIVATE *pr   = (USER_PRIVATE *) arg;
+    ELAN3_CTXT    *ctxt = pr->pr_ctxt;
+
+    PRINTF5 (pr->pr_ctxt, DBG_FN, "user_ptrack_handler: ctxt=%p pr=%p ref=%d phase %d mm->ref %d\n", 
+            pr->pr_ctxt, pr, atomic_read (&pr->pr_ref), phase, atomic_read (&current->mm->mm_count));
+
+    if (phase == PTRACK_PHASE_EXIT)
+    {
+       /* this will force the helper thread to exit */
+       elan3_swapout (ctxt, CTXT_EXITING);
+       
+       if (atomic_dec_and_test (&pr->pr_ref))
+           user_free (pr);
+    }  
+    return PTRACK_FINISHED;
+}
+
+static int
+user_open (struct inode *inode, struct file *file)
+{
+    ELAN3_DEV     *dev = elan3_devices[ELAN3_DEVICE(inode)];
+    USER_PRIVATE *pr;
+    ELAN3_CTXT    *ctxt;
+
+    if (dev == NULL)
+       return (-ENXIO);
+
+    KMEM_ALLOC(pr, USER_PRIVATE *, sizeof (USER_PRIVATE), TRUE);
+
+    if (pr == NULL)
+       return (-ENOMEM);
+    
+    if ((ctxt = elan3_alloc (dev, 0)) == NULL)
+    {
+       KMEM_FREE (pr, sizeof (USER_PRIVATE));
+       return (-ENOMEM);
+    }
+
+    if (sys_init (ctxt) == NULL)
+    {
+       elan3_detach(ctxt);
+       elan3_free (ctxt);
+       KMEM_FREE (pr, sizeof (USER_PRIVATE));
+       return (-ENOMEM);
+    }
+
+    /* initialise refcnt to 3 - one for "file", one for XA handler, one for the coproc ops */
+    atomic_set (&pr->pr_ref, 3);
+
+    atomic_set (&pr->pr_mappings, 0);
+    spin_lock_init (&pr->pr_lock);
+
+    pr->pr_ctxt = ctxt;
+    pr->pr_mm   = current->mm;
+
+    /* register an ptrack handler to force the helper thread to exit when we do */
+    if (ptrack_register (user_ptrack_handler, pr) < 0)
+    {
+       elan3_detach(ctxt);
+       elan3_free (ctxt);
+       KMEM_FREE (pr, sizeof (USER_PRIVATE));
+       return (-ENOMEM);
+    }
+
+    /* register a coproc callback to notify us of translation changes */
+    
+    pr->pr_coproc.arg               = (void *) pr;
+    pr->pr_coproc.release           = user_coproc_release;
+    pr->pr_coproc.sync_range        = user_coproc_sync_range;
+    pr->pr_coproc.invalidate_range  = user_coproc_invalidate_range;
+    pr->pr_coproc.update_range      = user_coproc_update_range;
+    pr->pr_coproc.change_protection = user_coproc_change_protection;
+    pr->pr_coproc.sync_page         = user_coproc_sync_page;
+    pr->pr_coproc.invalidate_page   = user_coproc_invalidate_page;
+    pr->pr_coproc.update_page       = user_coproc_update_page;
+    
+    spin_lock (&current->mm->page_table_lock);
+    register_coproc_ops (current->mm, &pr->pr_coproc);
+    spin_unlock (&current->mm->page_table_lock);
+
+    file->private_data = (void *) pr;
+
+    PRINTF2 (pr->pr_ctxt, DBG_FN, "user_open: done ctxt=%p pr=%p\n", ctxt, pr);
+
+    MOD_INC_USE_COUNT;
+    return (0);
+}
+
+static int
+user_release (struct inode *inode, struct file *file)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) file->private_data;
+    
+    PRINTF3 (pr->pr_ctxt, DBG_FN, "user_release: ctxt=%p pr=%p ref=%d\n", pr->pr_ctxt, pr,
+            atomic_read (&pr->pr_ref));
+
+    if (atomic_dec_and_test (&pr->pr_ref))
+       user_free (pr);
+
+    return (0);
+}
+
+static int
+user_ioctl (struct inode *inode, struct file *file, 
+           unsigned int cmd, unsigned long arg)
+{
+    USER_PRIVATE *pr   = (USER_PRIVATE *) file->private_data;
+    ELAN3_CTXT    *ctxt = pr->pr_ctxt;
+    SYS_CTXT     *sctx = (SYS_CTXT *) ctxt->Private;
+    int           res  = 0;
+
+    if (current->mm != pr->pr_mm)
+       return (-EINVAL);
+    
+    PRINTF4 (ctxt, DBG_FN, "user_ioctl: ctxt=%p cmd=%x(%d) arg=%lx\n", ctxt, cmd, _IOC_NR(cmd), arg);
+
+    switch (cmd)
+    {
+    case ELAN3IO_FREE:
+       if (atomic_read (&pr->pr_mappings) > 0)
+           return (-EINVAL);
+       
+       spin_lock (&current->mm->page_table_lock);
+       if (pr->pr_mm != current->mm)
+           spin_unlock (&current->mm->page_table_lock);
+       else
+       {
+           unregister_coproc_ops (current->mm, &pr->pr_coproc);
+           spin_unlock (&current->mm->page_table_lock);
+
+           user_coproc_release (pr, current->mm);
+       }
+
+       if (ptrack_registered (user_ptrack_handler, pr))
+       {
+           ptrack_deregister (user_ptrack_handler, pr);
+           user_ptrack_handler (pr, PTRACK_PHASE_EXIT, NULL);
+       }
+       break;
+       
+    case ELAN3IO_ATTACH:
+    {
+       ELAN_CAPABILITY *cap;
+
+       KMEM_ALLOC(cap, ELAN_CAPABILITY *, sizeof (ELAN_CAPABILITY), TRUE);
+
+       if (cap == NULL)
+           return (set_errno (EFAULT));
+
+       if (copy_from_user (cap, (void *) arg, sizeof (ELAN_CAPABILITY)))
+           res = EFAULT;
+       else
+       {
+           if ((res = elan3_attach (ctxt, cap)) == 0)
+           {
+               if (copy_to_user ((void *) arg, cap, sizeof (ELAN_CAPABILITY)))
+               {
+                   elan3_detach (ctxt);
+                   res = EFAULT;
+               }
+           }
+       }
+       KMEM_FREE (cap, sizeof(ELAN_CAPABILITY));
+       break;
+    }
+    
+    case ELAN3IO_DETACH:
+       spin_lock (&pr->pr_lock);
+       if (atomic_read (&pr->pr_mappings) > 0)
+           res = EINVAL;
+       else
+           elan3_detach (ctxt);
+       spin_unlock (&pr->pr_lock);
+       break;
+
+    case ELAN3IO_ADDVP:
+    {
+       ELAN3IO_ADDVP_STRUCT *args;
+
+       KMEM_ALLOC(args, ELAN3IO_ADDVP_STRUCT *, sizeof (ELAN3IO_ADDVP_STRUCT), TRUE);
+
+       if (args == NULL)
+           return (set_errno (ENOMEM));
+       
+       if (copy_from_user (args, (void *) arg, sizeof (ELAN3IO_ADDVP_STRUCT)))
+           res = EFAULT;
+       else
+       {
+           if ( (res=elan3_addvp (ctxt, args->process, &args->capability)) != 0)
+               PRINTF0 (ctxt, DBG_FN, "ELAN3IO_ADDVP elan3_addvp failed \n");  
+       }
+
+       KMEM_FREE (args, sizeof (ELAN3IO_ADDVP_STRUCT));
+       break;
+    }
+
+    case ELAN3IO_REMOVEVP:
+       res = elan3_removevp (ctxt, arg);
+       break;
+       
+    case ELAN3IO_BCASTVP:
+    {
+       ELAN3IO_BCASTVP_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_BCASTVP_STRUCT)))
+           return (-EFAULT);
+       
+       res = elan3_addbcastvp (ctxt, args.process, args.lowvp, args.highvp);
+       break;
+    }
+
+    case ELAN3IO_LOAD_ROUTE:
+    {
+       ELAN3IO_LOAD_ROUTE_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_LOAD_ROUTE_STRUCT)))
+           return (-EFAULT);
+       
+       res = elan3_load_route (ctxt, args.process, args.flits);
+       break;
+    }
+
+    case ELAN3IO_CHECK_ROUTE:
+    {
+       ELAN3IO_CHECK_ROUTE_STRUCT args;
+
+       args.routeError = 0;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_LOAD_ROUTE_STRUCT)))
+           return (-EFAULT);
+       
+       if ((res = elan3_check_route (ctxt, args.process, args.flits, & args.routeError)) ==  ESUCCESS)
+       {
+           if (copy_to_user ( (void *) arg, &args,sizeof (ELAN3IO_LOAD_ROUTE_STRUCT)))
+               return (-EFAULT);
+       }
+       break;
+    }
+
+    case ELAN3IO_PROCESS_2_LOCATION:
+    {
+       ELAN3IO_PROCESS_2_LOCATION_STRUCT args;
+       ELAN_LOCATION                    loc;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_PROCESS_2_LOCATION_STRUCT)))
+           return (-EFAULT);
+
+       krwlock_write (&ctxt->VpLock);
+       loc = ProcessToLocation (ctxt, NULL, args.process , NULL);
+       krwlock_done (&ctxt->VpLock);
+
+       args.loc = loc;
+
+       if (copy_to_user ( (void *) arg, &args,sizeof (ELAN3IO_PROCESS_2_LOCATION_STRUCT)))
+           return (-EFAULT);
+
+       break;
+    }
+
+    case ELAN3IO_GET_ROUTE:
+    {
+       ELAN3IO_GET_ROUTE_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_GET_ROUTE_STRUCT)))
+           return (-EFAULT);
+       
+       if ((res = elan3_get_route (ctxt, args.process, args.flits)) ==  ESUCCESS)
+       {
+           if (copy_to_user ( (void *) arg, &args,sizeof (ELAN3IO_GET_ROUTE_STRUCT)))
+               return (-EFAULT);
+       }
+       break;
+    }
+
+    case ELAN3IO_RESET_ROUTE:
+    {
+       ELAN3IO_RESET_ROUTE_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_RESET_ROUTE_STRUCT)))
+           return (-EFAULT);
+       
+       res = elan3_reset_route (ctxt, args.process);
+       break;
+    }
+
+    case ELAN3IO_VP2NODEID:
+    {
+       ELAN3IO_VP2NODEID_STRUCT *vp2nodeId;
+       ELAN_LOCATION           location;
+
+       KMEM_ALLOC (vp2nodeId, ELAN3IO_VP2NODEID_STRUCT *, sizeof(ELAN3IO_VP2NODEID_STRUCT), TRUE);
+       if (vp2nodeId == NULL) 
+           return (set_errno (ENOMEM));
+
+       if (copy_from_user (vp2nodeId, (void *) arg, sizeof (ELAN3IO_VP2NODEID_STRUCT))) {
+           KMEM_FREE (vp2nodeId, sizeof(ELAN3IO_VP2NODEID_STRUCT));
+           return (-EFAULT);
+       }
+
+       krwlock_write (&ctxt->VpLock);
+       location = ProcessToLocation (ctxt, NULL, vp2nodeId->process , NULL);
+       krwlock_done (&ctxt->VpLock);
+
+       vp2nodeId->nodeId = location.loc_node;
+       if (copy_to_user ( (void *) arg, vp2nodeId, sizeof (ELAN3IO_VP2NODEID_STRUCT))) {
+           KMEM_FREE (vp2nodeId, sizeof(ELAN3IO_VP2NODEID_STRUCT));
+           return (-EFAULT);
+       }
+
+       KMEM_FREE (vp2nodeId, sizeof(ELAN3IO_VP2NODEID_STRUCT));
+
+       break;
+    }
+
+    case ELAN3IO_PROCESS:
+       return (elan3_process (ctxt));
+
+    case ELAN3IO_SETPERM:
+    {
+       ELAN3IO_SETPERM_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_SETPERM_STRUCT)))
+           return (-EFAULT);
+
+       res = elan3mmu_setperm (ctxt->Elan3mmu, args.maddr, args.eaddr, args.len, args.perm);
+       break;
+    }
+
+    case ELAN3IO_CLEARPERM:
+    {
+       ELAN3IO_CLEARPERM_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_CLEARPERM_STRUCT)))
+           return (-EFAULT);
+
+       elan3mmu_clrperm (ctxt->Elan3mmu, args.eaddr, args.len);
+       break;
+    }
+
+    case ELAN3IO_CHANGEPERM:
+    {
+       ELAN3IO_CHANGEPERM_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_CHANGEPERM_STRUCT)))
+           return (-EFAULT);
+
+       res = EINVAL;
+       break;
+    }
+
+    case ELAN3IO_HELPER_THREAD:
+       res = elan3_lwp (ctxt);
+       break;
+       
+    case ELAN3IO_WAITCOMMAND:
+       res = WaitForCommandPort (ctxt);
+       break;
+
+    case ELAN3IO_BLOCK_INPUTTER:
+       elan3_block_inputter (ctxt, arg);
+       break;
+
+    case ELAN3IO_SET_FLAGS:
+       sctx->Flags = arg;
+       break;
+
+    case ELAN3IO_SET_SIGNAL:
+       sctx->signal = arg;
+       break;
+
+    case ELAN3IO_WAITEVENT:
+       res = sys_waitevent (ctxt, (E3_Event *) arg);
+       break;
+
+    case ELAN3IO_ALLOC_EVENTCOOKIE:
+       res = cookie_alloc_cookie (sctx->Table, arg);
+       break;
+
+    case ELAN3IO_FREE_EVENTCOOKIE:
+       res = cookie_free_cookie (sctx->Table, arg);
+       break;
+
+    case ELAN3IO_ARM_EVENTCOOKIE:
+       res = cookie_arm_cookie (sctx->Table, arg);
+       break;
+
+    case ELAN3IO_WAIT_EVENTCOOKIE:
+       res = cookie_wait_cookie (sctx->Table, arg);
+       break;
+
+    case ELAN3IO_SWAPSPACE:
+       if (fuword (&((SYS_SWAP_SPACE *) arg)->Magic) != SYS_SWAP_MAGIC)
+           return (set_errno (EINVAL));
+       
+       ((SYS_CTXT *) ctxt->Private)->Swap = (SYS_SWAP_SPACE *) arg;
+       break;
+
+    case ELAN3IO_EXCEPTION_SPACE:
+       if (fuword (&((SYS_EXCEPTION_SPACE *) arg)->Magic) != SYS_EXCEPTION_MAGIC)
+           return (set_errno (EINVAL));
+
+       ((SYS_CTXT *) ctxt->Private)->Exceptions = (SYS_EXCEPTION_SPACE *) arg;
+       break;
+
+    case ELAN3IO_GET_EXCEPTION:
+    {
+       SYS_EXCEPTION *exception;
+
+       if (((SYS_CTXT *) ctxt->Private)->Exceptions == NULL)
+           return (set_errno (EINVAL));
+       
+       KMEM_ALLOC(exception, SYS_EXCEPTION *, sizeof (SYS_EXCEPTION), TRUE);
+
+       if (exception == NULL)
+           return (set_errno (ENOMEM));
+
+       if ((res = sys_getException (((SYS_CTXT *) ctxt->Private), exception)) == 0 &&
+           copy_to_user ((void *) arg, exception, sizeof (SYS_EXCEPTION)))
+           res = EFAULT;
+       
+       KMEM_FREE (exception, sizeof (SYS_EXCEPTION));
+       break;
+    }
+    
+    case ELAN3IO_UNLOAD:
+    {
+       ELAN3MMU             *elan3mmu = ctxt->Elan3mmu;
+       ELAN3IO_UNLOAD_STRUCT args;
+       int                   span;
+       unsigned long         flags;
+       E3_Addr               eaddr;
+       caddr_t               addr;
+       size_t                len;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_UNLOAD_STRUCT)))
+           return (-EFAULT);
+
+       addr = (caddr_t) args.addr;
+       len  = args.len;
+
+       if (((unsigned long) addr & PAGEMASK) || (len & PAGEMASK) || (len < 0))
+           return -EINVAL;
+
+       spin_lock_irqsave (&elan3mmu->elan3mmu_lock, flags);
+       for (; len; len -= span, addr += span)
+       {
+           ELAN3MMU_RGN *rgn = elan3mmu_findrgn_main (elan3mmu, addr, 0);
+           
+           if (rgn == NULL || (rgn->rgn_mbase + rgn->rgn_len) < addr)
+               span = len;
+           else if (rgn->rgn_mbase > addr)
+               span = MIN(len, rgn->rgn_mbase - addr);
+           else
+           {
+               span  = MIN(len, (rgn->rgn_mbase + rgn->rgn_len) - addr);
+               eaddr = rgn->rgn_ebase + (addr - rgn->rgn_mbase);
+               
+               elan3mmu_unload (elan3mmu, eaddr, span, PTE_UNLOAD);
+           }
+       }
+       spin_unlock_irqrestore (&elan3mmu->elan3mmu_lock, flags);
+       
+       return 0;
+    }
+
+    case ELAN3IO_GET_DEVINFO:
+    {
+       ELAN3IO_GET_DEVINFO_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_GET_DEVINFO_STRUCT)))
+           return (-EFAULT);
+       
+       if (copy_to_user ((void *) args.devinfo, &ctxt->Device->Devinfo, sizeof (ELAN_DEVINFO))) 
+           res = EFAULT;
+       break;
+    }
+
+    case ELAN3IO_GET_POSITION:
+    {
+       ELAN3IO_GET_POSITION_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN3IO_GET_POSITION_STRUCT)))
+           return (-EFAULT);   
+
+       if (copy_to_user ((void *) args.position, &ctxt->Device->Position, sizeof (ELAN_POSITION)))
+           res = EFAULT;
+       break;
+    }
+
+    default:
+       return (-EINVAL);
+    }
+
+    return (res ? set_errno (res) : 0);
+}
+
+static void user_vma_open(struct vm_area_struct *vma)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) vma->vm_private_data;
+
+    PRINTF (DBG_DEVICE, DBG_SEG, "user_vma_open: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+
+    if (vma->vm_pgoff == ELAN3IO_OFF_COMMAND_PAGE)
+       if (atomic_dec_and_test (&pr->pr_mappings))
+           pr->pr_ctxt->CommandPageMapping = NULL;
+}
+
+static void user_vma_close(struct vm_area_struct *vma)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) vma->vm_private_data;
+
+    PRINTF (DBG_DEVICE, DBG_SEG, "user_vma_close: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+
+    if (vma->vm_pgoff == ELAN3IO_OFF_COMMAND_PAGE)
+       atomic_inc (&pr->pr_mappings);
+}
+
+static struct vm_operations_struct user_vm_ops = {
+    open:              user_vma_open,
+    close:             user_vma_close,
+};
+
+static int
+user_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    USER_PRIVATE  *pr   = (USER_PRIVATE *) file->private_data;
+    ELAN3_CTXT     *ctxt = pr->pr_ctxt; 
+    ioaddr_t       ioaddr;
+
+    /* 
+     * NOTE - since we need to maintain the reference count on
+     *        the user_private we only permit single page 
+     *        mmaps - this means that we will certainly see
+     *        the correct number of closes to maintain the
+     *        the reference count correctly.
+     */
+    
+    if ((vma->vm_end - vma->vm_start) != PAGE_SIZE)
+       return (-EINVAL);
+
+    PRINTF (DBG_DEVICE, DBG_SEG, "user_mmap: vm_mm=%p start=%lx end=%lx pgoff=%lx flags=%lx prot=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_flags, vma->vm_page_prot.pgprot, vma->vm_file);
+
+    switch (vma->vm_pgoff)
+    {
+    default:
+       return (-EINVAL);
+       
+    case ELAN3IO_OFF_COMMAND_PAGE:
+       spin_lock (&pr->pr_lock);
+       if (ctxt->CommandPage == (ioaddr_t) 0 || atomic_read (&pr->pr_mappings) != 0)
+       {
+           PRINTF (DBG_DEVICE, DBG_SEG, "user_mmap: command port - %s\n", ctxt->CommandPort ? "already mapped" : "not attached");
+           spin_unlock (&pr->pr_lock);
+           return (-EINVAL);
+       }
+#ifdef LINUX_SPARC
+       pgprot_val(vma->vm_page_prot) &= ~(_PAGE_CACHE);
+       pgprot_val(vma->vm_page_prot) |= _PAGE_IE;
+#elif defined(pgprot_noncached)
+       vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+#endif
+
+       PRINTF (DBG_DEVICE, DBG_SEG, "user_mmap: commandport at %lx phys %llx prot %lx\n", 
+               vma->vm_start, (unsigned long long) kmem_to_phys ((void *) ctxt->CommandPort), vma->vm_page_prot.pgprot);
+
+       /* Don't try to swap out physical pages.. */
+       vma->vm_flags |= VM_RESERVED;
+    
+       /*
+        * Don't dump addresses that are not real memory to a core file.
+        */
+       vma->vm_flags |= VM_IO;
+
+#ifdef NO_RMAP
+       if (remap_page_range (vma->vm_start, kmem_to_phys ((void *) ctxt->CommandPage), vma->vm_end - vma->vm_start, vma->vm_page_prot))
+#else 
+       if (remap_page_range (vma, vma->vm_start, kmem_to_phys ((void *) ctxt->CommandPage), vma->vm_end - vma->vm_start, vma->vm_page_prot))
+#endif
+       {
+           spin_unlock (&pr->pr_lock);
+           return (-EAGAIN);
+       }
+       ctxt->CommandPageMapping = (void *) vma->vm_start;
+       
+       atomic_inc (&pr->pr_mappings);
+       
+       spin_unlock (&pr->pr_lock);
+       break;
+
+    case ELAN3IO_OFF_UREG_PAGE:
+#ifdef LINUX_SPARC
+       pgprot_val(vma->vm_page_prot) &= ~(_PAGE_CACHE);
+       pgprot_val(vma->vm_page_prot) |= _PAGE_IE;
+#elif defined(pgprot_noncached)
+       vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+#endif
+       ioaddr = ctxt->Device->RegPtr + (offsetof (E3_Regs, URegs) & PAGEMASK);
+
+       PRINTF (DBG_DEVICE, DBG_SEG, "user_mmap: user_regs at %lx phys %llx prot %lx\n", vma->vm_start, 
+               (unsigned long long) kmem_to_phys ((void *) ioaddr), vma->vm_page_prot.pgprot);
+
+       /* Don't try to swap out physical pages.. */
+       vma->vm_flags |= VM_RESERVED;
+    
+       /*
+        * Don't dump addresses that are not real memory to a core file.
+        */
+       vma->vm_flags |= VM_IO;
+
+#ifdef NO_RMAP
+       if (remap_page_range (vma->vm_start, kmem_to_phys ((void *) ioaddr),
+#else
+       if (remap_page_range (vma, vma->vm_start, kmem_to_phys ((void *) ioaddr),
+#endif
+                             vma->vm_end - vma->vm_start, vma->vm_page_prot))
+           return (-EAGAIN);
+       break;
+       
+    case ELAN3IO_OFF_FLAG_PAGE:
+       PRINTF (DBG_DEVICE, DBG_SEG, "user_mmap: flag page at %lx phys %llx\n", vma->vm_start, 
+               (unsigned long long) kmem_to_phys ((void *) ctxt->FlagPage));
+
+       /* we do not want to have this area swapped out, lock it */
+       vma->vm_flags |= VM_LOCKED;
+
+       /* Mark the page as reserved or else the remap_page_range() doesn't remap it */
+       SetPageReserved(pte_page(*find_pte_kernel((unsigned long) ctxt->FlagPage)));
+       
+#ifdef NO_RMAP
+       if (remap_page_range (vma->vm_start, kmem_to_phys ((void *) ctxt->FlagPage),
+#else
+       if (remap_page_range (vma, vma->vm_start, kmem_to_phys ((void *) ctxt->FlagPage),
+#endif
+                             vma->vm_end - vma->vm_start, vma->vm_page_prot))
+           return (-EAGAIN);
+       break;
+    }
+
+    ASSERT (vma->vm_ops == NULL);
+    
+    vma->vm_ops          = &user_vm_ops;
+    vma->vm_file         = file;
+    vma->vm_private_data = (void *) pr;
+    
+    return (0);
+}
+
+/* driver entry points */
+static int
+elan3_open (struct inode *inode, struct file *file)
+{
+    if (elan3_devices[ELAN3_DEVICE(inode)] == NULL)
+       return (-ENXIO);
+
+    PRINTF (DBG_DEVICE, DBG_FN, "elan3_open: device %d minor %d file=%p\n", ELAN3_DEVICE(inode), ELAN3_MINOR(inode), file);
+    
+    switch (ELAN3_MINOR (inode))
+    {
+    case ELAN3_MINOR_CONTROL:
+       return (control_open (inode, file));
+    case ELAN3_MINOR_MEM:
+       return (mem_open (inode, file));
+    case ELAN3_MINOR_USER:
+       return (user_open (inode, file));
+    default:
+       return (-ENXIO);
+    }
+}
+
+static int
+elan3_release (struct inode *inode, struct file *file)
+{
+    PRINTF (DBG_DEVICE, DBG_FN, "elan3_release: device %d minor %d file=%p\n", ELAN3_DEVICE(inode), ELAN3_MINOR(inode), file);
+    
+    switch (ELAN3_MINOR (inode))
+    {
+    case ELAN3_MINOR_CONTROL:
+       return (control_release (inode, file));
+    case ELAN3_MINOR_MEM:
+       return (mem_release (inode, file));
+    case ELAN3_MINOR_USER:
+       return (user_release (inode, file));
+    default:
+       return (-ENXIO);
+    }
+}
+
+static int
+elan3_ioctl (struct inode *inode, struct file *file, 
+            unsigned int cmd, unsigned long arg)
+{
+    switch (ELAN3_MINOR (inode))
+    {
+    case ELAN3_MINOR_CONTROL:
+       return (control_ioctl (inode, file, cmd, arg));
+    case ELAN3_MINOR_MEM:
+       return (mem_ioctl (inode, file, cmd, arg));
+    case ELAN3_MINOR_USER:
+       return (user_ioctl (inode, file, cmd, arg));
+    default:
+       return (-ENXIO);
+    }
+}
+
+
+static int
+elan3_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    PRINTF (DBG_DEVICE, DBG_SEG, "elan3_mmap: instance %d minor %d start=%lx end=%lx pgoff=%lx flags=%lx prot=%lx\n", 
+           ELAN3_DEVICE (file->f_dentry->d_inode), ELAN3_MINOR (file->f_dentry->d_inode),
+           vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_flags, vma->vm_page_prot.pgprot);
+
+    switch (ELAN3_MINOR (file->f_dentry->d_inode))
+    {
+    case ELAN3_MINOR_CONTROL:
+       return (control_mmap (file, vma));
+    case ELAN3_MINOR_MEM:
+       return (mem_mmap (file, vma));
+    case ELAN3_MINOR_USER:
+       return (user_mmap (file, vma));
+    default:
+       return (-ENXIO);
+    }
+}
+
+static irqreturn_t
+InterruptHandlerWrapper(int irq, void *dev_id, struct pt_regs *regs)
+{
+    if (InterruptHandler ((ELAN3_DEV *)dev_id) == 0)
+       return IRQ_HANDLED;
+    else
+       return IRQ_NONE;
+}
+
+
+/* 
+ * Elan specific PCI configuration registers.
+ */
+
+#define PCI_CONF_PARITY_PHYS_LO         0x40
+#define PCI_CONF_PARITY_PHYS_HI         0x44
+#define PCI_CONF_PARITY_PHASE_ADDR      0x46
+#define PCI_CONF_PARITY_MASTER_TYPE     0x47
+#define PCI_CONF_ELAN3_CTRL              0x48
+ 
+#define ECTRL_EXTEND_LATENCY            (1 << 0)
+#define ECTRL_ENABLE_PREFETCH           (1 << 1)
+#define ECTRL_SOFTWARE_INTERNAL_RESET   (1 << 2)
+#define ECTRL_REDUCED_RETRY_RATE        (1 << 3)
+#define ECTRL_CLOCK_DIVIDE_RATE_SHIFT   4
+#define ECTRL_COMMS_DIVIDE_RATE_SHIFT   10
+#define ECTRL_FORCE_COMMSCLK_LOCAL      (1 << 14)
+
+/*
+ * Configure PCI.
+ */
+static int
+ConfigurePci(ELAN3_DEV *dev)
+{
+       struct pci_dev *pci = dev->Osdep.pci;
+       u32 rom_address;
+
+       if (pci_enable_device(pci))
+           return (ENXIO);
+
+       /* disable ROM */
+       pci_read_config_dword(pci, PCI_ROM_ADDRESS, &rom_address);
+       rom_address &= ~PCI_ROM_ADDRESS_ENABLE;
+       pci_write_config_dword(pci, PCI_ROM_ADDRESS, rom_address);
+       mb();
+
+       /* this is in 32-bit WORDS */
+       pci_write_config_byte(pci, PCI_CACHE_LINE_SIZE, (64 >> 2));
+       mb();
+
+       /* allow 40 ticks to respond, 16 data phases */
+       pci_write_config_byte(pci, PCI_LATENCY_TIMER, 255);
+       mb();
+
+       /* don't enable PCI_COMMAND_SERR--see note in elandev_dunix.c */
+       pci_write_config_word(pci, PCI_COMMAND, PCI_COMMAND_MEMORY 
+           | PCI_COMMAND_MASTER | PCI_COMMAND_INVALIDATE | PCI_COMMAND_PARITY);
+       mb();
+
+       return ESUCCESS;
+}
+
+/* 
+ * Reset chip to a known state.
+ */
+static int
+ResetElan(ELAN3_DEV *dev, ioaddr_t intPalAddr)
+{
+       struct pci_dev *pci = dev->Osdep.pci;
+       int instance = dev->Instance;
+       u32 val;
+       u8 revid;
+       int CasLatency;
+       int res;
+
+       /* determine rev of board */
+       pci_read_config_byte(pci, PCI_REVISION_ID, &revid);
+
+       /* GNAT 2328 - don't set ECTRL_ENABLE_PREFETCH on Elan rev A */
+       val = ECTRL_EXTEND_LATENCY | (39 << ECTRL_CLOCK_DIVIDE_RATE_SHIFT)
+           | (6 << ECTRL_COMMS_DIVIDE_RATE_SHIFT);
+       switch (revid) 
+       {
+               case PCI_REVISION_ID_ELAN3_REVA:
+                       printk("elan%d: is an elan3 (revision a) - not supported\n", instance);
+                       return (EFAIL);
+
+               case PCI_REVISION_ID_ELAN3_REVB:        
+                       val |= ECTRL_ENABLE_PREFETCH;
+                       if (BackToBackMaster)
+                               val |= ECTRL_FORCE_COMMSCLK_LOCAL;
+                       printk("elan%d: is an elan3 (revision b)\n", instance);
+                       break;
+               default:
+                       printk("elan%d: unsupported elan3 revision %d\n", 
+                           instance, revid);
+                       return EFAIL;
+       }
+       pci_write_config_dword(pci, PCI_CONF_ELAN3_CTRL, val);
+       mb();
+
+       /*
+        * GNAT: 2474
+        * Hit reset on the Elan, then we MUST initialise the schedule status
+        * register to drive reset on the link before the link can come out
+        * of reset (15 uS). We need to keep it like this until we've 
+        * initialised SDRAM
+        */
+       pci_read_config_dword(pci, PCI_CONF_ELAN3_CTRL, &val);
+       pci_write_config_dword(pci, PCI_CONF_ELAN3_CTRL, 
+           val | ECTRL_SOFTWARE_INTERNAL_RESET);
+       mb();
+
+       /* Read the Vital Product Data to determine the cas latency */
+       if ((res = ReadVitalProductData (dev, &CasLatency)) != ESUCCESS)
+           return (res);
+
+       /*
+        * Now clear the Software internal reset bit, and start the sdram
+        */
+       pci_write_config_dword(pci, PCI_CONF_ELAN3_CTRL, val);
+       mb();
+
+       /* 
+        * Enable SDRAM before sizing and initalising it for ECC.
+        * NOTE: don't enable all sets of the cache (yet), nor ECC 
+        */
+       dev->Cache_Control_Reg = (CasLatency | REFRESH_RATE_16US);
+
+       write_reg32 (dev, Cache_Control_Reg.ContReg, (dev->Cache_Control_Reg | SETUP_SDRAM));
+       mb();
+
+       INIT_SCHED_STATUS(dev, Sched_Initial_Value);
+
+       /*
+        * Set the interrupt mask to 0 and enable the interrupt PAL
+        * by writing any value to it.
+        */
+       SET_INT_MASK (dev, 0);
+       writeb (0, intPalAddr);
+ 
+       return ESUCCESS;
+}
+
+/*
+ * Determine the size of elan PCI address spaces.  EFAIL is returned if 
+ * unused or invalid BAR is specified, or if board reports I/O mapped space.
+ */
+int
+DeviceRegisterSize(ELAN3_DEV *dev, int rnumber, int *sizep)
+{
+       struct pci_dev *pdev = dev->Osdep.pci;
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,4,0)
+       *sizep = pci_resource_size(pdev, rnumber);
+#else
+       *sizep = pci_resource_end(pdev, rnumber) - pci_resource_start(pdev, rnumber) + 1;
+#endif
+       return ESUCCESS;
+}
+
+/*
+ * Map PCI memory into kernel virtual address space.  On the alpha, 
+ * we just return appropriate kseg address, and Unmap is a no-op.
+ */
+int
+MapDeviceRegister(ELAN3_DEV *dev, int rnumber, ioaddr_t *addrp,
+                 int off, int len, DeviceMappingHandle *handlep)
+{      
+       struct pci_dev *pdev = dev->Osdep.pci;
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,4,0)
+       u64 base = pci_get_base_address(pdev, rnumber);
+       *addrp = (ioaddr_t) pci_base_to_kseg(base + off, pdev->bus->number);
+
+#else
+       if (len == 0)
+           len = pci_resource_end(pdev, rnumber) - pci_resource_start(pdev, rnumber) + 1;
+       
+       if (len == 0)
+           return (EINVAL);
+
+       *addrp = (ioaddr_t) ioremap_nocache (pci_resource_start(pdev, rnumber) + off, len);
+#endif
+
+       *handlep = (void *) *addrp;
+
+       return (*addrp ? ESUCCESS : ENOMEM);
+}
+void
+UnmapDeviceRegister(ELAN3_DEV *dev, DeviceMappingHandle *handlep)
+{
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,4,0)
+    iounmap (*handlep);
+#endif
+}
+
+void
+ElanBusError (ELAN3_DEV *dev)
+{
+       struct pci_dev  *pci = dev->Osdep.pci;  
+       u8  phaseaddr, type;
+       u16 status, cmd, physhi;
+       u32 physlo;
+ 
+       printk("elan%d: bus error occured\n", dev->Instance);
+
+       pci_read_config_word (pci, PCI_STATUS,                  &status);
+       pci_read_config_word (pci, PCI_COMMAND,                 &cmd);
+       pci_read_config_dword(pci, PCI_CONF_PARITY_PHYS_LO,     &physlo);
+       pci_read_config_word (pci, PCI_CONF_PARITY_PHYS_HI,     &physhi);
+       pci_read_config_byte (pci, PCI_CONF_PARITY_PHASE_ADDR,  &phaseaddr); 
+       pci_read_config_byte (pci, PCI_CONF_PARITY_MASTER_TYPE, &type);
+
+#define PCI_CONF_STAT_FORMAT   "\20" \
+       "\6SIXTY_SIX_MHZ\7UDF\10FAST_BACK\11PARITY" \
+       "\14SIG_TARGET_ABORT\15REC_TARGET_ABORT\16REC_MASTER_ABORT" \
+       "\17SIG_SYSTEM_ERROR\20DETECTED_PARITY"
+
+       printk ("elan%d: status %x cmd %4x physaddr %04x%08x phase %x type %x\n",
+               dev->Instance, status, cmd, physhi, physlo, phaseaddr, type);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/elansyscall.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/elansyscall.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/elansyscall.c  2005-06-01 23:12:54.584441384 -0400
@@ -0,0 +1,1230 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elansyscall.c,v 1.99.2.1 2004/10/28 17:08:56 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/elansyscall.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/autoconf.h>
+
+#include <elan/elanmod.h>
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/elansyscall.h>
+#include <elan/devinfo.h>
+
+static int       sys_exception (ELAN3_CTXT *ctxt, int type, int proc, void *trap, va_list ap);
+static int       sys_getWordItem (ELAN3_CTXT *ctxt, int list, void **itemp, E3_uint32 *valuep);
+static int       sys_getBlockItem (ELAN3_CTXT *ctxt, int list, void **itemp, E3_Addr *valuep);
+static void      sys_putWordItem (ELAN3_CTXT *ctxt, int list, E3_uint32 value);
+static void      sys_putBlockItem (ELAN3_CTXT *ctxt, int list, E3_uint32 *ptr);
+static void      sys_putbackItem (ELAN3_CTXT *ctxt, int list, void *item);
+static void      sys_freeWordItem (ELAN3_CTXT *ctxt, void *item);
+static void      sys_freeBlockItem (ELAN3_CTXT *ctxt, void *item);
+static int       sys_countItems (ELAN3_CTXT *ctxt, int list);
+static int       sys_event (ELAN3_CTXT *ctxt, E3_uint32 cookie, int flag);
+static void      sys_swapin (ELAN3_CTXT *ctxt);
+static void      sys_swapout (ELAN3_CTXT *ctxt);
+static void      sys_freePrivate (ELAN3_CTXT *ctxt);
+static int       sys_fixupNetworkError (ELAN3_CTXT *ctxt, NETERR_FIXUP *nef);
+static int       sys_startFaultCheck (ELAN3_CTXT *ctxt);
+static void      sys_endFaultCheck (ELAN3_CTXT *ctxt);
+static E3_uint8  sys_load8 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void      sys_store8 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint8 val);
+static E3_uint16 sys_load16 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void      sys_store16 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint16 val);
+static E3_uint32 sys_load32 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void      sys_store32 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint32 val);
+static E3_uint64 sys_load64 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void      sys_store64 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint64 val);
+
+static ELAN3_OPS elan3_sys_ops = {
+    ELAN3_OPS_VERSION,         /* Version */
+
+    sys_exception,             /* Exception */
+    sys_getWordItem,           /* GetWordItem */
+    sys_getBlockItem,          /* GetBlockItem */
+    sys_putWordItem,           /* PutWordItem */
+    sys_putBlockItem,          /* PutBlockItem */
+    sys_putbackItem,           /* PutbackItem */
+    sys_freeWordItem,          /* FreeWordItem */
+    sys_freeBlockItem,         /* FreeBlockItem */
+    sys_countItems,            /* CountItems */
+    sys_event,                 /* Event */
+    sys_swapin,                        /* Swapin */
+    sys_swapout,               /* Swapout */
+    sys_freePrivate,           /* FreePrivate */
+    sys_fixupNetworkError,     /* FixupNetworkError */
+    NULL,                      /* DProcTrap */
+    NULL,                      /* TProcTrap */
+    NULL,                      /* IProcTrap */
+    NULL,                      /* CProcTrap */
+    NULL,                      /* CProcReissue */
+    sys_startFaultCheck,       /* StartFaultCheck */
+    sys_endFaultCheck,          /* EndFaultCheck */
+    sys_load8,                 /* Load8 */
+    sys_store8,                        /* Store8 */
+    sys_load16,                        /* Load16 */
+    sys_store16,               /* Store16 */
+    sys_load32,                        /* Load32 */
+    sys_store32,               /* Store32 */
+    sys_load64,                        /* Load64 */
+    sys_store64                        /* Store64 */
+};
+
+va_list null_valist;
+
+SYS_CTXT *
+sys_init (ELAN3_CTXT *ctxt)
+{
+    SYS_CTXT *sctx;
+
+    /* Allocate and initialise the context private data */
+    KMEM_ZALLOC (sctx, SYS_CTXT *, sizeof  (SYS_CTXT), TRUE);
+
+    if (sctx == NULL)
+       return ((SYS_CTXT *) NULL);
+
+    sctx->Swap    = NULL;
+    sctx->Armed   = 0;
+    sctx->Backoff = 1;
+    sctx->Table   = cookie_alloc_table ((unsigned long) ELAN3_MY_TASK_HANDLE(), 0);
+    sctx->signal  = SIGSEGV;
+
+    if (sctx->Table == NULL)
+    {
+       KMEM_FREE (sctx, sizeof (SYS_CTXT));
+       return ((SYS_CTXT *) NULL);
+    }
+
+    kmutex_init  (&sctx->Lock);
+    spin_lock_init (&sctx->WaitLock);
+    kcondvar_init (&sctx->NetworkErrorWait);
+    
+    /* Install my context operations and private data */
+    ctxt->Operations = &elan3_sys_ops;
+    ctxt->Private    = (void *) sctx;
+    
+    return (sctx);
+}
+
+/* returns -ve on error or ELAN_CAP_OK or ELAN_CAP_RMS */
+/* use = ELAN_USER_ATTACH, ELAN_USER_P2P, ELAN_USER_BROADCAST */
+int 
+elan3_validate_cap(ELAN3_DEV *dev, ELAN_CAPABILITY *cap ,int use)
+{
+     /* Don't allow a user process to attach to system context */
+    if (ELAN3_SYSTEM_CONTEXT (cap->cap_lowcontext) || ELAN3_SYSTEM_CONTEXT (cap->cap_highcontext)
+       || cap->cap_highcontext <= ELAN_USER_BASE_CONTEXT_NUM  || cap->cap_highcontext <= ELAN_USER_BASE_CONTEXT_NUM)
+    {
+       PRINTF2 (DBG_DEVICE, DBG_VP,"elan3_validate_cap: lctx %x hctx %x \n",cap->cap_lowcontext,  cap->cap_highcontext);
+       PRINTF3 (DBG_DEVICE, DBG_VP,"elan3_validate_cap: bit %x  low %x high %x\n", ((cap->cap_lowcontext) & SYS_CONTEXT_BIT),
+                E3_NUM_CONTEXT_0, ELAN3_KCOMM_BASE_CONTEXT_NUM);
+
+
+       PRINTF0 (DBG_DEVICE, DBG_VP,"elan3_validate_cap: user process cant attach to system cap\n");
+       return (-EINVAL);
+    }
+ 
+    if (cap->cap_type & ELAN_CAP_TYPE_HWTEST)
+    {
+       if (!(cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP)) /* cant have a bit map */
+       {
+           PRINTF0 (DBG_DEVICE, DBG_VP, "elanmod_classify_cap: ELAN_CAP_TYPE_HWTEST must have ELAN_CAP_TYPE_NO_BITMAP\n");
+           return (-EINVAL);
+       }
+       
+       if (cap->cap_lowcontext != cap->cap_highcontext) 
+       {
+           PRINTF2 (DBG_DEVICE, DBG_VP, "elanmod_classify_cap: ELAN_CAP_TYPE_HWTEST (cap->cap_lowcontext != cap->cap_highcontext) %d %d\n",cap->cap_lowcontext , cap->cap_highcontext) ;
+           return (-EINVAL);
+       }
+       
+       if ( ! (ELAN3_HWTEST_CONTEXT(cap->cap_lowcontext) && ELAN3_HWTEST_CONTEXT(cap->cap_highcontext)))
+       {
+           PRINTF3 (DBG_DEVICE, DBG_VP, "elanmod_classify_cap: ELAN_CAP_TYPE_HWTEST HWTEST_BASE_CONTEXT %d %d %d \n" , ELAN3_HWTEST_BASE_CONTEXT_NUM,cap->cap_lowcontext ,ELAN3_HWTEST_TOP_CONTEXT_NUM);
+           return (-EINVAL);
+       }
+       
+       if (cap->cap_lownode != ELAN_CAP_UNINITIALISED || cap->cap_highnode != ELAN_CAP_UNINITIALISED)
+       {
+           PRINTF0 (DBG_DEVICE, DBG_VP, "elanmod_classify_cap: ELAN_CAP_TYPE_HWTEST nodes != ELAN_CAP_UNINITIALISED\n");
+           return (-EINVAL);
+       }
+
+       return ELAN_CAP_OK;
+    }
+
+    return elanmod_classify_cap(&dev->Position, cap, use);
+}
+
+int
+sys_waitevent (ELAN3_CTXT *ctxt, E3_Event *event)
+{
+    SYS_CTXT    *sctx = (SYS_CTXT *) ctxt->Private;
+    EVENT_COOKIE cookie;
+
+    if (ctxt->Device->Devinfo.dev_revision_id == PCI_REVISION_ID_ELAN3_REVA)
+       return (EINVAL);
+
+    cookie = fuword ((int *) &event->ev_Type) & ~(EV_TYPE_MASK_EVIRQ | EV_TYPE_MASK_BCOPY);
+
+    if (cookie_alloc_cookie (sctx->Table, cookie) != ESUCCESS)
+       return (EINVAL);
+
+    cookie_arm_cookie (sctx->Table, cookie);
+
+    if (fuword ((int *) &event->ev_Count) > 0)
+       cookie_wait_cookie (sctx->Table, cookie);
+    
+    cookie_free_cookie (sctx->Table, cookie);
+    
+    return (ESUCCESS);
+}
+
+static void *
+sys_getItem (SYS_SWAP_SPACE *sp, int list)
+{
+    void *itemp = (void *) fuptr_noerr ((void **) &sp->ItemListsHead[list]);
+    void *next;
+    
+    PRINTF4 (DBG_DEVICE, DBG_SYSCALL, "sys_getItem: sp=%p list=%d head=%p itemp=%p\n",
+            sp, list, &sp->ItemListsHead[list], itemp);
+    
+    if (itemp == NULL)
+       return (NULL);
+
+    next = (void *) fuptr_noerr ((void *) itemp);
+
+    suptr_noerr ((void *) &sp->ItemListsHead[list], (void *) next);
+    if (next == NULL)
+       suptr_noerr ((void *) &sp->ItemListsTailp[list], (void *)&sp->ItemListsHead[list]);
+    return (itemp);
+}
+
+static void
+sys_putItemBack (SYS_SWAP_SPACE *sp, int list, void *itemp)
+{
+    PRINTF4 (DBG_DEVICE, DBG_SYSCALL, "sys_putItemBack: sp=%p list=%d itemp=%p value=%08x\n",
+            sp, list, itemp, fuword_noerr ((int *) &((SYS_WORD_ITEM *) itemp)->Value));
+
+    suptr_noerr ((void **) itemp, NULL);                                                       /* item->Next = NULL */
+    suptr_noerr ((void **) fuptr_noerr ((void **) &sp->ItemListsTailp[list]), (void *)itemp);  /* *Tailp = item */
+    suptr_noerr ((void **) &sp->ItemListsTailp[list], (void *) itemp);                         /* Tailp = &item->Next */
+}
+
+static void
+sys_putItemFront (SYS_SWAP_SPACE *sp, int list, void *itemp)
+{
+    PRINTF4 (DBG_DEVICE, DBG_SYSCALL, "sys_putItemFront: sp=%p list=%d itemp=%p value=%08x\n",
+            sp, list, itemp, fuword_noerr ((int *) &((SYS_WORD_ITEM *) itemp)->Value));
+
+    suptr_noerr ((void **) itemp, fuptr_noerr ((void **) &sp->ItemListsHead[list]));           /* item->Next = Head */
+    suptr_noerr ((void **) &sp->ItemListsHead[list], (void *) itemp);                          /* Head = item */
+
+    if (fuptr_noerr ((void **) &sp->ItemListsTailp[list]) == (void *) &sp->ItemListsHead[list])        /* if (Tailp == &Head) */
+       suptr_noerr ((void **) &sp->ItemListsTailp[list], (void *) itemp);                      /*    Tailp = &Item->Next */
+}
+
+
+static int
+sys_getWordItem (ELAN3_CTXT *ctxt, int list, void **itemp, E3_uint32 *valuep)
+{
+    SYS_CTXT      *sctx = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp   = sctx->Swap;
+    SYS_WORD_ITEM  *item;
+    int                    res;
+    label_t        ljb;
+
+    kmutex_lock (&sctx->Lock);
+    
+    if (on_fault (&ljb))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, list, (void *) NULL, null_valist);
+       return (0);
+    }
+
+    item = (SYS_WORD_ITEM *) sys_getItem (sp, list);
+
+    if (item == NULL)
+       res = 0;
+    else
+    {
+       if (list == LIST_DMA_PTR)
+           sctx->Armed = TRUE;
+
+       *itemp  = (void *) item;
+       *valuep = (E3_Addr) fuword_noerr ((E3_int32 *) &item->Value);
+
+       PRINTF3 (ctxt, DBG_SYSCALL, "sys_getWordItem: list=%d -> item=%p value=%08x\n", list, *itemp, *valuep);
+
+       res = 1;
+    }
+    
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+
+    return (res);
+}
+
+static int
+sys_getBlockItem (ELAN3_CTXT *ctxt, int list, void **itemp, E3_Addr *valuep)
+{
+    SYS_CTXT      *sctx = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp   = sctx->Swap;
+    SYS_BLOCK_ITEM *item;
+    int                    res;
+    label_t        ljb;
+
+    kmutex_lock (&sctx->Lock);
+    
+    if (on_fault (&ljb))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, list, (void *) NULL, null_valist);
+       return (0);
+    }
+
+    item = sys_getItem (sp, list);
+
+    if (item == NULL)
+       res = 0;
+    else
+    {
+       E3_uint32 *dest = fuptr_noerr ((void **) &item->Pointer);
+
+       if (list == LIST_DMA_DESC)
+           sctx->Armed = TRUE;
+
+       *itemp  = (void *) item;
+       *valuep = elan3mmu_elanaddr (ctxt->Elan3mmu, (caddr_t) dest);
+
+       PRINTF3 (ctxt, DBG_SYSCALL, "sys_getBlockItem: list=%d -> item=%p addr=%08x\n", list, *itemp, *valuep);
+       PRINTF4 (ctxt, DBG_SYSCALL, "                  %08x %08x %08x %08x\n",
+                fuword_noerr ((int *) &dest[0]), fuword_noerr ((int *) &dest[1]), 
+                fuword_noerr ((int *) &dest[2]), fuword_noerr ((int *) &dest[3]));
+       PRINTF4 (ctxt, DBG_SYSCALL, "                  %08x %08x %08x %08x\n",
+                fuword_noerr ((int *) &dest[4]), fuword_noerr ((int *) &dest[5]),
+                fuword_noerr ((int *) &dest[6]), fuword_noerr ((int *) &dest[7]));
+
+       
+       res = 1;
+    }
+    
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+
+    return (res);
+}
+
+static void
+sys_putWordItem (ELAN3_CTXT *ctxt, int list, E3_Addr value)
+{
+    SYS_CTXT      *sctx = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp   = sctx->Swap;
+    SYS_WORD_ITEM  *item;
+    label_t        ljp;
+
+    kmutex_lock (&sctx->Lock);
+
+    PRINTF2 (ctxt,DBG_SYSCALL, "sys_putWordItem: list=%x value=%x\n", list, value);
+
+    if (on_fault (&ljp))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, list, (void *) NULL, null_valist);
+       return;
+    }
+
+    item = sys_getItem (sp, LIST_FREE_WORD);
+
+    PRINTF1 (ctxt, DBG_SYSCALL, "sys_putWordItem: item=%p\n", item);
+
+    if (item == NULL)
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       
+       sys_exception (ctxt, EXCEPTION_SWAP_FAILED, list, (void *) NULL, null_valist);
+       return;
+    }
+    
+    PRINTF2 (ctxt, DBG_SYSCALL, "sys_putWordItem: storing value=%08x at %p\n", value, &item->Value);
+
+    PRINTF2 (ctxt, DBG_SYSCALL, "sys_putWordItem: item=%p value=%08x\n", item, value);
+
+    suword_noerr ((E3_int32 *) &item->Value, value);                                           /* write "value" into item */
+
+    sys_putItemBack (sp, list, item);
+
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+}
+
+static void
+sys_putBlockItem (ELAN3_CTXT *ctxt, int list, E3_uint32 *ptr)
+{
+    SYS_CTXT      *sctx = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp   = sctx->Swap;
+    SYS_BLOCK_ITEM *item;
+    label_t        ljp;
+    E3_uint32      *source;
+    E3_uint32      *dest;
+
+    PRINTF2 (ctxt, DBG_SYSCALL, "sys_putBlockItem: list=%x ptr=%p\n", list, ptr);
+
+    kmutex_lock (&sctx->Lock);
+    
+    if (on_fault (&ljp))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, list, (void *) NULL, null_valist);
+       return;
+    }
+
+    item = sys_getItem (sp, LIST_FREE_BLOCK);                  /* get an item from the freelist. */
+
+    if (item == NULL)
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       
+       sys_exception (ctxt, EXCEPTION_SWAP_FAILED, list, (void *) NULL, null_valist);
+       return;
+    }
+
+    /*
+     * The block will have been read using 64 bit reads,  since we have
+     * to write it to user memory using 32 bit writes, we need to perform
+     * an endian swap on the Ultrasparc.
+     */
+    dest   = (E3_uint32 *) fuptr_noerr ((void **) &item->Pointer);
+    source = (E3_uint32 *) ptr;
+
+    PRINTF2 (ctxt, DBG_SYSCALL, "sys_putBlockItem: item=%p dest=%p\n",item, dest);
+    PRINTF4 (ctxt, DBG_SYSCALL, "                  %08x %08x %08x %08x\n",
+           source[0^WordEndianFlip], source[1^WordEndianFlip], source[2^WordEndianFlip], source[3^WordEndianFlip]);
+    PRINTF4 (ctxt, DBG_SYSCALL, "                  %08x %08x %08x %08x\n",
+            source[4^WordEndianFlip], source[5^WordEndianFlip], source[6^WordEndianFlip], source[7^WordEndianFlip]);
+
+    suword_noerr ((E3_int32 *) &dest[7], (E3_int32) source[7^WordEndianFlip]);
+    suword_noerr ((E3_int32 *) &dest[6], (E3_int32) source[6^WordEndianFlip]);
+    suword_noerr ((E3_int32 *) &dest[5], (E3_int32) source[5^WordEndianFlip]);
+    suword_noerr ((E3_int32 *) &dest[4], (E3_int32) source[4^WordEndianFlip]);
+    suword_noerr ((E3_int32 *) &dest[3], (E3_int32) source[3^WordEndianFlip]);
+    suword_noerr ((E3_int32 *) &dest[2], (E3_int32) source[2^WordEndianFlip]);
+    suword_noerr ((E3_int32 *) &dest[1], (E3_int32) source[1^WordEndianFlip]);
+    suword_noerr ((E3_int32 *) &dest[0], (E3_int32) source[0^WordEndianFlip]);
+
+    sys_putItemBack (sp, list, item);                          /* chain onto list of items. */
+
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+}
+
+static void
+sys_freeWordItem (ELAN3_CTXT *ctxt, void *itemp)
+{
+    SYS_CTXT      *sctx = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp   = sctx->Swap;
+    label_t        ljp;
+
+    kmutex_lock (&sctx->Lock);
+    
+    if (on_fault (&ljp))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, LIST_FREE_WORD, (void *) NULL, null_valist);
+       return;
+    }
+
+    sys_putItemBack (sp, LIST_FREE_WORD, itemp);
+
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+}
+
+static void
+sys_freeBlockItem (ELAN3_CTXT *ctxt, void *itemp)
+{
+    SYS_CTXT       *sctx = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp   = sctx->Swap;
+    SYS_BLOCK_ITEM *item = (SYS_BLOCK_ITEM *)itemp;
+    E3_uint32      *dest;
+    label_t        ljp;
+
+    kmutex_lock (&sctx->Lock);
+    
+    if (on_fault (&ljp))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, LIST_FREE_BLOCK, (void *) NULL, null_valist);
+       return;
+    }
+#ifdef DEBUG_PRINTF
+    dest = (E3_uint32 *) fuptr_noerr ((void **) &item->Pointer);
+
+    PRINTF2 (ctxt, DBG_SYSCALL, "sys_freeBlockItem: item=%p dest=%p\n", item, dest);
+    PRINTF4 (ctxt, DBG_SYSCALL, "                  %08x %08x %08x %08x\n",
+            fuword_noerr ((int *) &dest[0]), fuword_noerr ((int *) &dest[1]), 
+            fuword_noerr ((int *) &dest[2]), fuword_noerr ((int *) &dest[3]));
+    PRINTF4 (ctxt, DBG_SYSCALL, "                  %08x %08x %08x %08x\n",
+            fuword_noerr ((int *) &dest[4]), fuword_noerr ((int *) &dest[5]),
+            fuword_noerr ((int *) &dest[6]), fuword_noerr ((int *) &dest[7]));
+#endif
+
+    sys_putItemBack (sp, LIST_FREE_BLOCK, itemp);
+
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+}
+
+static void
+sys_putbackItem (ELAN3_CTXT *ctxt, int list, void *itemp)
+{
+    SYS_CTXT       *sctx = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp   = sctx->Swap;
+    label_t        ljp;
+
+    kmutex_lock (&sctx->Lock);
+    
+    if (on_fault (&ljp))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, list, (void *) NULL, null_valist);
+       return;
+    }
+
+    sys_putItemFront (sp, list, itemp);
+
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+}
+
+static int
+sys_countItems (ELAN3_CTXT *ctxt, int list)
+{
+    SYS_CTXT      *sctx  = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp    = sctx->Swap;
+    int                    count = 0;
+    void          *item;
+    label_t        ljb;
+
+    kmutex_lock (&sctx->Lock);
+    
+    if (on_fault (&ljb))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       sys_exception (ctxt, EXCEPTION_SWAP_FAULT, list, (void *) NULL, null_valist);
+       return (0);
+    }
+
+    for (item = (void *) fuptr_noerr ((void **) &sp->ItemListsHead[list]); 
+        item != NULL;
+        item = (void *) fuptr_noerr ((void **) item))
+    {
+       count++;
+    }
+
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+
+    return (count);
+}
+
+
+long sys_longTime;
+long sys_shortTime;
+int  sys_waitTicks;
+int  sys_maxBackoff;
+
+#define SYS_LONG_TIME          MAX((hz * 5) / 1000, 1)         /* 5 ms */
+#define SYS_SHORT_TIME         MAX((hz * 2) / 1000, 1)         /* 2 ms */
+#define SYS_WAIT_TICKS         MAX((hz * 1) / 1000, 1)         /* 1 ms  - backoff granularity */
+#define SYS_MAX_BACKOFF                MAX((hz * 5) / 1000, 1)         /* 5 ms  - max backoff for "nacked" packets*/
+#define SYS_TIMEOUT_BACKOFF    MAX((hz * 10) / 1000, 1)        /* 10 ms - backoff for output timeout (point to point) */
+#define SYS_BCAST_BACKOFF      MAX((hz * 50) / 1000, 1)        /* 50 ms - backoff for output timeout (broadcast) */
+#define SYS_NETERR_BACKOFF     MAX((hz * 10) / 1000, 1)        /* 10 ms - delay for network error in dma data */
+
+static void
+sys_backoffWait (ELAN3_CTXT *ctxt, int ticks)
+{
+    SYS_CTXT *sctx = (SYS_CTXT *) ctxt->Private;
+    long      t;
+
+    spin_lock (&sctx->WaitLock);
+
+    t = lbolt - sctx->Time;
+
+    if (sys_longTime   == 0) sys_longTime   = SYS_LONG_TIME;
+    if (sys_shortTime  == 0) sys_shortTime  = SYS_SHORT_TIME;
+    if (sys_waitTicks  == 0) sys_waitTicks  = SYS_WAIT_TICKS;
+    if (sys_maxBackoff == 0) sys_maxBackoff = SYS_MAX_BACKOFF;
+
+    if (t > sys_longTime)                                      /* It's a long time since the last trap */
+       sctx->Backoff = 0;                                      /* so set the backoff back down to 0 */
+
+    if (ticks)
+    {
+       PRINTF2 (ctxt, DBG_DPROC, "sys_backoffWait : Waiting - %d ticks [%lx]\n", ticks, t);
+       kcondvar_timedwait (&sctx->NetworkErrorWait, &sctx->WaitLock, NULL, lbolt + ticks);
+    }
+    else if (sctx->Armed)
+    {
+       if (t < sys_shortTime)                                  /* It's been a short time since the last */
+       {                                                       /* trap, so increase the backoff */
+           sctx->Backoff++;
+           
+           if (sctx->Backoff > sys_maxBackoff)
+               sctx->Backoff = sys_maxBackoff;
+       }
+
+       PRINTF2 (ctxt, DBG_DPROC, "sys_backoffWait : Waiting - %d [%lx]\n", sctx->Backoff, t);
+
+       if (sctx->Backoff)
+           kcondvar_timedwaitsig (&sctx->NetworkErrorWait, &sctx->WaitLock, NULL, lbolt + sctx->Backoff * sys_waitTicks);
+
+       sctx->Armed = 0;
+    }
+    else
+    {
+       PRINTF1 (ctxt, DBG_DPROC, "sys_backoffWait : Not Waiting - %d\n", sctx->Backoff);
+
+    }
+    sctx->Time = lbolt;
+
+    spin_unlock (&sctx->WaitLock);
+}
+
+static int
+trapSize (int proc)
+{
+    switch (proc)
+    {
+    case DMA_PROC:     return (sizeof (DMA_TRAP));
+    case THREAD_PROC:  return (sizeof (THREAD_TRAP));
+    case COMMAND_PROC: return (sizeof (COMMAND_TRAP));
+    case INPUT_PROC:   return (sizeof (INPUT_TRAP));
+    default:           return (0);
+    }
+}
+
+static int
+sys_exception (ELAN3_CTXT *ctxt, int type, int proc, void *trapp, va_list ap)
+{
+    SYS_CTXT *sctx  = (SYS_CTXT *) ctxt->Private;
+    int              res;
+
+    PRINTF2 (ctxt, DBG_SYSCALL, "sys_exception: type %d proc %d\n", type, proc);
+
+    switch (type)
+    {
+    case EXCEPTION_INVALID_ADDR:
+    {
+       E3_FaultSave_BE *faultSave = va_arg (ap, E3_FaultSave_BE *);
+       int              res       = va_arg (ap, int);
+       
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), faultSave, res, 0);
+       break;
+    }
+    
+    case EXCEPTION_UNIMP_INSTR:
+    {
+       E3_uint32 instr = va_arg (ap, E3_uint32);
+       
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, 0, instr);
+       break;
+    }
+    
+    case EXCEPTION_INVALID_PROCESS:
+    {
+       E3_uint32 vproc = va_arg (ap, E3_uint32);
+       int       res  = va_arg (ap, int);
+
+       switch (proc)
+       {
+       case DMA_PROC:
+           if (sctx->Flags & ELAN3_SYS_FLAG_DMA_BADVP)
+           {
+               DMA_TRAP *trap = (DMA_TRAP *) trapp;
+
+               if (trap->Desc.s.dma_direction != DMA_WRITE)
+                   trap->Desc.s.dma_srcEvent = trap->Desc.s.dma_destEvent;
+
+               trap->Desc.s.dma_direction       = DMA_WRITE;
+               trap->Desc.s.dma_size            = 0;
+               trap->Desc.s.dma_source          = (E3_Addr) 0;
+               trap->Desc.s.dma_dest            = (E3_Addr) 0;
+               trap->Desc.s.dma_destEvent       = (E3_Addr) 0;
+               trap->Desc.s.dma_destCookieVProc = 0;
+               trap->Desc.s.dma_srcCookieVProc  = 0;
+               
+               return (OP_IGNORE);
+           }
+           break;
+
+       case THREAD_PROC:
+           if (sctx->Flags & ELAN3_SYS_FLAG_THREAD_BADVP)
+           {
+               THREAD_TRAP *trap = (THREAD_TRAP *) trapp;
+
+               trap->TrapBits.s.PacketAckValue = E3_PAckError;
+               
+               return (OP_IGNORE);
+           }
+           break;
+       }
+           
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, res, vproc);
+       break;
+    }
+    
+    case EXCEPTION_FAULTED:
+    {
+       E3_Addr addr = va_arg (ap, E3_Addr);
+
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, 0, addr);
+       break;
+    }
+    
+    case EXCEPTION_QUEUE_OVERFLOW:
+    {
+       E3_FaultSave_BE *faultSave = va_arg (ap, E3_FaultSave_BE *);
+       int              trapType  = va_arg (ap, int);
+       
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), faultSave, 0, trapType);
+       break;
+    }
+    
+    case EXCEPTION_COMMAND_OVERFLOW:
+    {
+       int count = va_arg (ap, int);
+       
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, 0, count);
+       break;
+    }
+    
+    case EXCEPTION_CHAINED_EVENT:
+    {
+       E3_Addr addr = va_arg (ap, E3_Addr);
+       
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, 0, addr);
+       break;
+    }
+    
+    case EXCEPTION_DMA_RETRY_FAIL:
+    case EXCEPTION_PACKET_TIMEOUT:
+       if (proc != DMA_PROC)
+           sys_backoffWait (ctxt, SYS_TIMEOUT_BACKOFF);
+       else
+       {
+           DMA_TRAP *trap = (DMA_TRAP *) trapp;
+           
+           if (sctx->Flags & ELAN3_SYS_FLAG_DMAFAIL)
+           {
+               E3_BlockCopyEvent *event;
+
+               if (trap->Desc.s.dma_direction != DMA_WRITE)
+                   trap->Desc.s.dma_srcEvent = trap->Desc.s.dma_destEvent;
+
+               /* change the source word to be E3_EVENT_FAILED */
+               if ((event = (E3_BlockCopyEvent *) elan3mmu_mainaddr (ctxt->Elan3mmu, trap->Desc.s.dma_srcEvent)) == NULL)
+               {
+                   sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, 0, 0);
+                   break;
+               }
+
+               suword (&event->ev_Source, E3_EVENT_FAILED);
+               wmb(); mmiob();
+               
+               trap->Desc.s.dma_direction       = DMA_WRITE;
+               trap->Desc.s.dma_size            = 0;
+               trap->Desc.s.dma_source          = (E3_Addr) 0;
+               trap->Desc.s.dma_dest            = (E3_Addr) 0;
+               trap->Desc.s.dma_destEvent       = (E3_Addr) 0;
+               trap->Desc.s.dma_destCookieVProc = 0;
+               trap->Desc.s.dma_srcCookieVProc  = 0;
+               
+               return (OP_IGNORE);
+           }
+
+           if (type == EXCEPTION_DMA_RETRY_FAIL)
+               sys_backoffWait (ctxt, 0);
+           else
+           {
+               ELAN_LOCATION location;
+
+               krwlock_read (&ctxt->VpLock);
+               location = ProcessToLocation (ctxt, NULL, trap->Desc.s.dma_direction == DMA_WRITE ? 
+                                             trap->Desc.s.dma_destVProc : trap->Desc.s.dma_srcVProc, NULL);
+               krwlock_done (&ctxt->VpLock);
+               
+               sys_backoffWait (ctxt, location.loc_node == ELAN3_INVALID_NODE ? SYS_BCAST_BACKOFF : SYS_TIMEOUT_BACKOFF);
+           }
+       }
+       return (OP_IGNORE);
+       
+    case EXCEPTION_NETWORK_ERROR:
+    {
+       INPUT_TRAP       *trap  = (INPUT_TRAP *) trapp;
+       NETERR_RESOLVER **rvpp  = va_arg (ap, NETERR_RESOLVER **);
+
+       ASSERT (trap->State == CTXT_STATE_NETWORK_ERROR);
+
+       if (! (sctx->Flags & ELAN3_SYS_FLAG_NETERR) && (trap->DmaIdentifyTransaction || trap->ThreadIdentifyTransaction))
+       {
+           if ((*rvpp) != (NETERR_RESOLVER *) NULL)
+               res = (*rvpp)->Status;
+           else if ((res = QueueNetworkErrorResolver (ctxt, trap, rvpp)) == ESUCCESS)
+           {
+               /* Successfully queued the network error resolver */
+               return (OP_HANDLED);
+           }
+
+           /* network error resolution has failed - either a bad cookie or */
+           /* an rpc error has occured */
+           sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, res, 0);
+       }
+       else
+       {
+           /* Must be an overlaped dma packet. Must wait long enough to
+            * ensure that the sending dma'er has tried to send the next
+            * packet and had it discarded. In the real world this should
+            * be greater than an output timeout. (About 8mSec) */
+           
+           sys_backoffWait (ctxt, SYS_NETERR_BACKOFF);
+           
+           /* set this inputter state to be ok, since we've been called 
+            * by the lwp it will lower the context filter for us, so 
+            * re-enabling the inputter,  note we don't need to execute
+            * any of the packet since the dma process will re-transmit
+            * it after receiving a nack for the next packet */
+           trap->State = CTXT_STATE_OK;
+           
+           return (OP_HANDLED);
+       }
+       break;
+    }
+    
+    default:
+       sys_addException (sctx, type, proc, trapp, trapSize(proc), NULL, 0, 0);
+       break;
+    }
+    
+    if (type != EXCEPTION_DEBUG)
+#ifdef LINUX
+#ifdef NO_NPTL
+       psignal (CURPROC()->p_opptr, sctx->signal);
+#else
+       psignal (CURPROC()->parent, sctx->signal);
+#endif
+#else
+       psignal (CURPROC(), sctx->signal);
+#endif
+    return (OP_HANDLED);
+}
+
+static int
+sys_event (ELAN3_CTXT *ctxt, E3_uint32 cookie, int flag)
+{
+    SYS_CTXT *sctx = (SYS_CTXT *) ctxt->Private;
+
+    cookie_fire_cookie (sctx->Table, cookie);
+
+    return (OP_HANDLED); 
+}
+
+static void
+sys_swapin (ELAN3_CTXT *ctxt)
+{
+    PRINTF0 (ctxt, DBG_SYSCALL, "sys_swapin\n");
+}
+
+static void
+sys_swapout (ELAN3_CTXT *ctxt)
+{
+    PRINTF0 (ctxt, DBG_SYSCALL, "sys_swapout\n");
+}
+
+static void
+sys_freePrivate (ELAN3_CTXT *ctxt)
+{
+    SYS_CTXT *sctx = (SYS_CTXT *) ctxt->Private;
+
+    cookie_free_table (sctx->Table);
+
+    kmutex_destroy (&sctx->Lock);
+    spin_lock_destroy (&sctx->WaitLock);
+    kcondvar_destroy (&sctx->NetworkErrorWait);
+
+    KMEM_FREE (sctx, sizeof (SYS_CTXT));
+    ctxt->Private = NULL;
+}
+
+static int
+sys_checkThisDma (ELAN3_CTXT *ctxt, NETERR_FIXUP *nef, E3_DMA *dma)
+{
+    E3_DmaType type;
+    E3_uint32  cookie;
+    E3_uint32  cvproc;
+    int               ignore;
+    int               match;
+
+    type.type = fuword_noerr ((int *) &dma->dma_type);
+
+    if (type.s.direction == DMA_WRITE)
+    {
+       cookie = fuword_noerr ((int *) &dma->dma_srcCookieVProc);
+       cvproc = fuword_noerr ((int *) &dma->dma_destCookieVProc);
+    }
+    else
+    {
+       cookie = fuword_noerr ((int *) &dma->dma_destCookieVProc);
+       cvproc = fuword_noerr ((int *) &dma->dma_srcCookieVProc);
+    }
+
+    PRINTF5 (ctxt, DBG_NETERR, "sys_checkThisDma: dir = %d cookie = %08x cvproc = %08x CookieVProc %08x DstProcess %04x\n",
+            type.s.direction, cookie, cvproc, nef->Message.CookieVProc, nef->Message.DstProcess);
+
+    /* A DMA matches a network errror fixup if it's going to the right place (or is a broadcast)
+     * and the approriate cookie matches, except that we ignore DMA's which don't have a destEvent
+     * since they don't have any atomic behaviour (though they still send the identify) */
+
+    ignore = (type.s.direction == DMA_WRITE && cookie == 0 &&
+             fuword_noerr ((int *) &dma->dma_destEvent) == 0);
+    match  = (nef->Message.CookieVProc == cookie &&
+             (nef->Message.DstProcess == (cvproc & DMA_PROCESS_MASK) || nef->Message.WaitForEop));
+
+    PRINTF2 (ctxt, DBG_NETERR, "  -> %s %s\n", ignore ? "ignore" : match ? "matched" : "not-matched", nef->Message.WaitForEop ? "wait for eop" : "");
+
+    if (match && !ignore && !nef->Message.WaitForEop)
+    {
+       PRINTF0 (ctxt, DBG_NETERR, "sys_checkThisDma: nuking the dma\n");
+
+       /* NOTE - we access the dma descriptor backwards since it could exist in sdram */
+       if (type.s.direction != DMA_WRITE)
+           suword_noerr ((int *) &dma->dma_srcEvent, 0);
+
+       suword_noerr ((int *) &dma->dma_destEvent, 0);
+       suword_noerr ((int *) &dma->dma_dest,      0);
+       suword_noerr ((int *) &dma->dma_source,    0);
+       suword_noerr ((int *) &dma->dma_size,      0);
+
+       if (type.s.direction != DMA_WRITE)
+           suword_noerr ((int *) &dma->dma_type, fuword_noerr ((int *) &dma->dma_type) & E3_DMA_CONTEXT_MASK);
+
+       wmb(); mmiob();
+    }
+
+    return (match && !ignore);
+}
+
+static int
+sys_fixupNetworkError (ELAN3_CTXT *ctxt, NETERR_FIXUP *nef)
+{
+    SYS_CTXT       *sctx    = (SYS_CTXT *) ctxt->Private;
+    SYS_SWAP_SPACE *sp      = sctx->Swap;
+    int                    matched = 0;
+    SYS_WORD_ITEM  *wordp;
+    SYS_BLOCK_ITEM *blockp;
+    label_t        ljb;
+    int                    res;
+
+    PRINTF3 (ctxt, DBG_NETERR, "sys_fixupnetworkError %08x %08x %08x\n", 
+            nef->Message.CookieAddr, nef->Message.CookieVProc, nef->Message.NextCookie);
+
+    if (nef->Message.CookieAddr == (E3_Addr) 0)                        /* It's a DMA which requires fixing up */
+    {
+       kmutex_lock (&sctx->Lock);
+
+       if (on_fault (&ljb))
+           res = EFAULT;
+       else
+       {
+           /* scan the dma ptr list */
+           for (wordp = (SYS_WORD_ITEM *) fuptr_noerr ((void **) &sp->ItemListsHead[LIST_DMA_PTR]);
+                wordp != NULL; 
+                wordp = (SYS_WORD_ITEM *) fuptr_noerr ((void **) &wordp->Next))
+           {
+               E3_uint32 value = fuword_noerr ((int *) &wordp->Value);
+               E3_DMA    *dma  = (E3_DMA *) elan3mmu_mainaddr (ctxt->Elan3mmu, value);
+
+               PRINTF3 (ctxt, DBG_NETERR, "sys_fixupnetworkError: check block item %p Value %08x dma %p\n", wordp, value, dma);
+
+               matched += sys_checkThisDma (ctxt, nef, dma);
+           }
+       
+           /* scan the dma desc list */
+           for (blockp = (SYS_BLOCK_ITEM *) fuptr_noerr ((void **) &sp->ItemListsHead[LIST_DMA_DESC]);
+                blockp != NULL; 
+                blockp = (SYS_BLOCK_ITEM *) fuptr_noerr ((void **) &blockp->Next))
+           {
+               E3_DMA *dma = (E3_DMA *) fuptr_noerr ((void *) &blockp->Pointer);
+               
+               PRINTF2 (ctxt, DBG_NETERR, "sys_fixupnetworkError: check block item %p Pointer %p\n", blockp, dma);
+
+               matched += sys_checkThisDma (ctxt, nef, dma);
+           }
+           
+           /* If we've still not found it, then check the command port item */
+           /* it MUST be present as a command waiting to be executed, as */
+           /* otherwise it could have already happened and we will claim to */
+           /* have found it, but not realy */
+           if (ctxt->CommandPortItem != NULL)
+           {
+               E3_DMA *dma = (E3_DMA *) fuptr_noerr ((void *) &((SYS_BLOCK_ITEM *) ctxt->CommandPortItem)->Pointer);
+
+               if (sys_checkThisDma (ctxt, nef, dma))
+               {
+                   printk ("!!! it's the command port item - need to ensure that the command exists\n");
+                   matched++;
+               }
+           }
+
+           res = matched ? ESUCCESS : ESRCH;
+       }
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+
+       if (matched > 1)
+           ElanException (ctxt, EXCEPTION_COOKIE_ERROR, DMA_PROC, NULL, NULL, nef->Message.CookieVProc);
+    }
+    else                                                       /* It's a thread which requires fixing up */
+    {
+       E3_int32  *cookiePtr = (E3_int32 *) elan3mmu_mainaddr (ctxt->Elan3mmu, nef->Message.CookieAddr);
+       E3_uint32  curval    = fuword_noerr (cookiePtr);
+
+       if (curval == nef->Message.CookieVProc)         /* thread doesn't think it's been done */
+       {
+           if (! nef->Message.WaitForEop)
+           {
+               suword_noerr (cookiePtr, nef->Message.NextCookie);
+               mb(); mmiob();
+           }
+           
+           res = ESUCCESS;
+       }
+       else                                                    /* thread thinks that it's been executed */
+       {
+           res = ESRCH;
+       }
+    }
+    
+    CompleteNetworkErrorFixup (ctxt, nef, res);
+
+    return (OP_HANDLED);
+}
+
+
+static int
+sys_startFaultCheck (ELAN3_CTXT *ctxt)
+{
+    return (0);
+}
+
+static void
+sys_endFaultCheck (ELAN3_CTXT *ctxt)
+{
+    wmb();
+}
+
+static E3_uint8
+sys_load8 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    E3_uint8 *maddr = (E3_uint8 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    return (fubyte_noerr (maddr));
+}
+
+static void
+sys_store8 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint8 val)
+{
+    E3_uint8 *maddr = (E3_uint8 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    subyte_noerr (maddr, val);
+    wmb(); mmiob();
+}
+
+static E3_uint16
+sys_load16 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    E3_uint16 *maddr = (E3_uint16 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    return (fusword_noerr (maddr));
+}
+
+static void
+sys_store16 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint16 val)
+{
+    E3_uint16 *maddr = (E3_uint16 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    susword_noerr (maddr, val);
+    wmb(); mmiob();
+}
+
+static E3_uint32
+sys_load32 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    E3_uint32 *maddr = (E3_uint32 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    return (fuword_noerr (maddr));
+}
+
+static void
+sys_store32 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint32 val)
+{
+    E3_uint32 *maddr = (E3_uint32 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    suword_noerr (maddr, val);
+    wmb(); mmiob();
+}
+
+static E3_uint64
+sys_load64 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    E3_uint64 *maddr = (E3_uint64 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    return (fulonglong_noerr ((long long *) maddr));
+}
+
+static void
+sys_store64 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint64 val)
+{
+    E3_uint64 *maddr = (E3_uint64 *) elan3mmu_mainaddr (ctxt->Elan3mmu, addr);
+
+    sulonglong_noerr ((long long *) maddr, val);
+    wmb(); mmiob();
+}
+
+
+void
+sys_addException (SYS_CTXT *sctx, int type, int proc, caddr_t trapp, int size,
+                 E3_FaultSave_BE *faultSave, u_long res, u_long value)
+{
+    SYS_EXCEPTION      *ex_ptr;
+    int                        front;
+    int                        back;
+    int                        count;
+    label_t            ljp;
+
+    PRINTF4 (DBG_DEVICE, DBG_FN, "sys_addException: type %d proc %d res %ld value %ld\n",
+            type, proc, res, value);
+
+    KMEM_ZALLOC (ex_ptr, SYS_EXCEPTION *, sizeof  (SYS_EXCEPTION), TRUE);
+
+    if (ex_ptr != NULL)
+    {
+       bzero ((caddr_t) ex_ptr, sizeof (SYS_EXCEPTION));
+
+       ex_ptr->Type  = type;
+       ex_ptr->Proc  = proc;
+       ex_ptr->Res   = res;
+       ex_ptr->Value = value;
+       
+       if (trapp && size)
+           bcopy (trapp, (caddr_t) &ex_ptr->Union, size);
+       if (faultSave)
+           bcopy ((caddr_t) faultSave, (caddr_t) &ex_ptr->FaultArea, sizeof (E3_FaultSave_BE));
+    }
+
+    kmutex_lock (&sctx->Lock);
+    if (! on_fault (&ljp))
+    {
+       front = fuword_noerr (&sctx->Exceptions->Front);
+       back  = fuword_noerr (&sctx->Exceptions->Back);
+       count = fuword_noerr (&sctx->Exceptions->Count);
+
+       if (count <= 0 || front < 0 || back < 0 || front >= count || back >= count)
+           suword_noerr (&sctx->Exceptions->Overflow, fuword_noerr (&sctx->Exceptions->Overflow) + 1);
+       else if (((front+1) % count ) == back)
+           suword_noerr (&sctx->Exceptions->Overflow, fuword_noerr (&sctx->Exceptions->Overflow) + 1);
+       else
+       {
+           if (ex_ptr != NULL)
+               copyout_noerr ((caddr_t) ex_ptr, (caddr_t) &sctx->Exceptions->Exceptions[front], sizeof (SYS_EXCEPTION));
+           else
+           {
+               suword_noerr (&sctx->Exceptions->Exceptions[front].Type, EXCEPTION_ENOMEM);
+               suword_noerr (&sctx->Exceptions->Exceptions[front].Proc, 0);
+           }
+           suword_noerr (&sctx->Exceptions->Front, (front + 1) % count);
+       }
+
+       /* always reset the magic number in case it's been overwritten */
+       /* so that 'edb' can find the exception page in the core file */
+       suword_noerr (&sctx->Exceptions->Magic, SYS_EXCEPTION_MAGIC);
+    }
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+    
+    if (ex_ptr != NULL)
+       KMEM_FREE (ex_ptr, sizeof  (SYS_EXCEPTION));
+}
+
+int
+sys_getException (SYS_CTXT *sctx, SYS_EXCEPTION *ex)
+{
+    int     front;
+    int     back;
+    int     count;
+    int     res;
+    label_t ljp;
+
+    if (sctx->Exceptions == NULL)
+       return (EINVAL);
+
+    kmutex_lock (&sctx->Lock);
+    if (on_fault (&ljp))
+    {
+       no_fault();
+       kmutex_unlock (&sctx->Lock);
+       return (EFAULT);
+    }
+    
+    front = fuword_noerr (&sctx->Exceptions->Front);
+    back  = fuword_noerr (&sctx->Exceptions->Back);
+    count = fuword_noerr (&sctx->Exceptions->Count);
+
+    if (count <= 0 || front < 0 || back < 0 || front >= count || back >= count || back == front)
+       res = EINVAL;
+    else
+    {
+       copyin_noerr ((caddr_t) &sctx->Exceptions->Exceptions[back], (caddr_t) ex, sizeof (SYS_EXCEPTION));
+       suword_noerr (&sctx->Exceptions->Back, (back+1) % count);
+
+       res = ESUCCESS;
+    }
+    no_fault();
+    kmutex_unlock (&sctx->Lock);
+
+    return (res);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/eventcookie.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/eventcookie.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/eventcookie.c  2005-06-01 23:12:54.585441232 -0400
@@ -0,0 +1,324 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: eventcookie.c,v 1.7 2003/08/13 10:03:03 fabien Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/eventcookie.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+#include <elan3/vmseg.h>
+
+static EVENT_COOKIE_TABLE *cookie_tables;
+static spinlock_t         cookie_table_lock;
+
+/*
+ * cookie_drop_entry:
+ *   drop the reference to a cookie held 
+ *   by the cookie table
+ */
+static void
+cookie_drop_entry (EVENT_COOKIE_ENTRY *ent)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&ent->ent_lock, flags);
+    if (--ent->ent_ref != 0)
+    {
+       ent->ent_fired = ent->ent_cookie;
+       kcondvar_wakeupall (&ent->ent_wait, &ent->ent_lock);
+
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+    }
+    else
+    {
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+
+       spin_lock_destroy (&ent->ent_lock);
+       kcondvar_destroy (&ent->ent_wait);
+
+       KMEM_FREE (ent, sizeof (EVENT_COOKIE_ENTRY));
+    }
+}
+
+void
+cookie_init()
+{
+    spin_lock_init (&cookie_table_lock);
+}
+
+void
+cookie_fini()
+{
+    spin_lock_destroy (&cookie_table_lock);
+}
+
+EVENT_COOKIE_TABLE *
+cookie_alloc_table (unsigned long task, unsigned long handle)
+{
+    EVENT_COOKIE_TABLE *tbl, *ntbl;
+
+    KMEM_ZALLOC (ntbl, EVENT_COOKIE_TABLE *, sizeof (EVENT_COOKIE_TABLE), TRUE);
+
+    if (ntbl == NULL)
+       return (NULL);
+
+    spin_lock (&cookie_table_lock);
+    
+    for (tbl = cookie_tables; tbl; tbl = tbl->tbl_next)
+       if (tbl->tbl_task == task && tbl->tbl_handle == handle)
+           break;
+    
+    if (tbl != NULL)
+       tbl->tbl_ref++;
+    else
+    {
+       spin_lock_init (&ntbl->tbl_lock);
+
+       ntbl->tbl_task    = task;
+       ntbl->tbl_handle  = handle;
+       ntbl->tbl_ref     = 1;
+       ntbl->tbl_entries = NULL;
+
+       if ((ntbl->tbl_next = cookie_tables) != NULL)
+           cookie_tables->tbl_prev = ntbl;
+       cookie_tables = ntbl;
+       ntbl->tbl_prev = NULL;
+    }
+    spin_unlock (&cookie_table_lock);
+
+    if (tbl == NULL)
+       return (ntbl);
+    else
+    {
+       KMEM_FREE (ntbl, sizeof (EVENT_COOKIE_TABLE));
+       return (tbl);
+    }    
+}
+
+void
+cookie_free_table (EVENT_COOKIE_TABLE *tbl)
+{
+    EVENT_COOKIE_ENTRY *ent;
+
+    spin_lock (&cookie_table_lock);
+    if (tbl->tbl_ref > 1)
+    {
+       tbl->tbl_ref--;
+       spin_unlock (&cookie_table_lock);
+       return;
+    }
+    
+    if (tbl->tbl_prev)
+       tbl->tbl_prev->tbl_next = tbl->tbl_next;
+    else
+       cookie_tables = tbl->tbl_next;
+    if (tbl->tbl_next)
+       tbl->tbl_next->tbl_prev = tbl->tbl_prev;
+    
+    spin_unlock (&cookie_table_lock);
+    
+    /* NOTE - table no longer visible to other threads
+     *        no need to aquire tbl_lock */
+    while ((ent = tbl->tbl_entries) != NULL)
+    {
+       if ((tbl->tbl_entries = ent->ent_next) != NULL)
+           ent->ent_next->ent_prev = NULL;
+       
+       cookie_drop_entry (ent);
+    }
+    spin_lock_destroy (&tbl->tbl_lock);
+
+    KMEM_FREE (tbl, sizeof (EVENT_COOKIE_TABLE));
+}
+
+int
+cookie_alloc_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie)
+{
+    EVENT_COOKIE_ENTRY *ent, *nent;
+    unsigned long flags;
+
+    KMEM_ZALLOC (nent, EVENT_COOKIE_ENTRY *, sizeof (EVENT_COOKIE_ENTRY), TRUE);
+    
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+
+    if (ent == NULL)
+    {
+       kcondvar_init (&nent->ent_wait);
+       spin_lock_init (&nent->ent_lock);
+
+       nent->ent_ref    = 1;
+       nent->ent_cookie = cookie;
+
+       if ((nent->ent_next = tbl->tbl_entries) != NULL)
+           tbl->tbl_entries->ent_prev = nent;
+       tbl->tbl_entries = nent;
+       nent->ent_prev = NULL;
+    }
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    if (ent == NULL)
+       return (ESUCCESS);
+    else
+    {
+       KMEM_FREE (nent, sizeof (EVENT_COOKIE_ENTRY));
+       return (EINVAL);
+    }
+}
+
+int
+cookie_free_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie)
+{
+    EVENT_COOKIE_ENTRY *ent;
+    unsigned long flags;
+
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (EINVAL);
+    }
+
+    if (ent->ent_prev == NULL)
+       tbl->tbl_entries = ent->ent_next;
+    else
+       ent->ent_prev->ent_next = ent->ent_next;
+
+    if (ent->ent_next != NULL)
+       ent->ent_next->ent_prev = ent->ent_prev;
+    
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    cookie_drop_entry (ent);
+
+    return (ESUCCESS);
+}
+
+/*
+ * cookie_fire_cookie:
+ *    fire the cookie - this is called from the event interrupt.
+ */
+int
+cookie_fire_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie)
+{
+    EVENT_COOKIE_ENTRY *ent;
+    unsigned long flags;
+
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (EINVAL);
+    }
+           
+    spin_lock (&ent->ent_lock);
+    ent->ent_fired = cookie;
+    kcondvar_wakeupall (&ent->ent_wait, &ent->ent_lock);
+    spin_unlock (&ent->ent_lock);
+
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    return (ESUCCESS);
+}    
+
+/*
+ * cookie_wait_cookie:
+ *    deschedule on a cookie if it has not already fired.
+ *    note - if the cookie is removed from the table, then
+ *           we free it off when we're woken up.
+ */
+int
+cookie_wait_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie)
+{
+    EVENT_COOKIE_ENTRY *ent;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (EINVAL);
+    }
+
+    spin_lock (&ent->ent_lock);
+    spin_unlock (&tbl->tbl_lock);
+
+    if (ent->ent_fired != 0)
+    {
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+       return (ESUCCESS);
+    }
+
+    ent->ent_ref++;
+    kcondvar_waitsig (&ent->ent_wait, &ent->ent_lock, &flags);
+    
+    if (--ent->ent_ref > 0)
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+    else
+    {
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+       
+       spin_lock_destroy (&ent->ent_lock);
+       kcondvar_destroy (&ent->ent_wait);
+
+       KMEM_FREE (ent, sizeof (EVENT_COOKIE_ENTRY));
+    }
+    return (ESUCCESS);
+}
+
+int
+cookie_arm_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie)
+{
+    EVENT_COOKIE_ENTRY *ent;
+    unsigned long flags;
+
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (EINVAL);
+    }
+           
+    spin_lock (&ent->ent_lock);
+    ent->ent_fired = 0;
+    spin_unlock (&ent->ent_lock);
+
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    return (ESUCCESS);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/iproc.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/iproc.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/iproc.c        2005-06-01 23:12:54.586441080 -0400
@@ -0,0 +1,925 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: iproc.c,v 1.47 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/iproc.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/trtype.h>
+#include <elan3/vmseg.h>
+
+
+static int TrSizeTable[] = {0, 8, 16, 32, 64};
+
+static void  ConvertTransactionToSetEvent (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_Addr Addr);
+static void  SimulateBlockWrite  (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap);
+static void  SimulateWriteWord   (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap);
+static void  SimulateWriteDWord  (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap);
+static void  SimulateTraceRoute  (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap);
+static void  BumpInputterStats (ELAN3_DEV *dev, E3_IprocTrapHeader_BE *hdrp);
+
+void
+HandleIProcTrap (ELAN3_DEV           *dev, 
+                int                 Channel,
+                E3_uint32           Pend,
+                sdramaddr_t         FaultSaveOff,
+                sdramaddr_t         TransactionsOff,
+                sdramaddr_t         DataOff)
+{
+    E3_IprocTrapHeader_BE Transaction0;
+    ELAN3_CTXT          *ctxt;
+    INPUT_TRAP           *trap;
+    register int          i;
+
+    /*
+     * Read the 1st set of transactions, so we can determine the 
+     * context for the trap 
+     */
+    elan3_sdram_copyq_from_sdram (dev, TransactionsOff, (void *) &Transaction0, 16);
+    
+    BumpStat (dev, IProcTraps);
+    BumpInputterStats (dev, &Transaction0);
+
+    if (Transaction0.s.TrTypeCntx.s.TypeCntxInvalid)
+    {
+       /*
+        * The context is not valid. This will occur if the packet
+        * trapped for an EopError with no IdentTrans or an error corrupted the context
+        * giving a CRC error on the first transaction and the Ack had not been returned.
+        */
+       if (Transaction0.s.TrTypeCntx.s.LastTrappedTrans)
+       {
+           PRINTF0 (DBG_DEVICE, DBG_IPROC, "iproc: Error on EOP without a good context, ignoring trap\n");
+       }
+       else
+       {
+           /* Check that only crap has been received.  If not then die. */
+           if (! Transaction0.s.IProcTrapStatus.s.BadLength &&
+               (Transaction0.s.IProcTrapStatus.Status & CRC_MASK) == CRC_STATUS_GOOD)
+           {
+               printk ("iproc: Did not have a valid context for the trap area.\n");
+               printk ("iproc: TrTypeCntx=%x TrAddr=%x TrData0=%x IProcTrapStatus=%x\n",
+                        Transaction0.s.TrTypeCntx.TypeContext, Transaction0.s.TrAddr,
+                        Transaction0.s.TrData0, Transaction0.s.IProcTrapStatus.Status);
+               panic ("elan3: iproc did not have a valid context");
+               /* NOTREACHED */
+           }
+           PRINTF0 (DBG_DEVICE, DBG_IPROC, "iproc: First transaction is bad, ignoring trap\n");
+       }
+    }
+    else
+    {
+       ctxt = ELAN3_DEV_CTX_TABLE(dev, Transaction0.s.TrTypeCntx.s.Context);
+       
+       if (ctxt == NULL)
+       {
+           PRINTF1 (DBG_DEVICE, DBG_INTR, "HandleIProcTrap: context %x invalid\n", 
+                    Transaction0.s.TrTypeCntx.s.Context);
+
+           BumpStat (dev, InvalidContext);
+       }
+       else
+       {
+           trap = (Channel == 0) ? &ctxt->Input0Trap : &ctxt->Input1Trap;
+
+           ASSERT (trap->State == CTXT_STATE_OK);
+           
+           trap->Transactions[0] = Transaction0;
+
+           PRINTF1 (ctxt, DBG_INTR, "HandleIProcTrap: %s\n", IProcTrapString (&trap->Transactions[0], NULL));
+           /*
+            * Copy the rest of the transactions into the trap area.
+            */
+           for (i = 0; !(trap->Transactions[i].s.TrTypeCntx.s.LastTrappedTrans);)
+           {
+               if (++i >= MAX_TRAPPED_TRANS)
+               {
+                   trap->Overflow = 1;
+                   break;
+               }
+
+               elan3_sdram_copyq_from_sdram (dev, TransactionsOff + i*sizeof (E3_IprocTrapHeader), (void *) &trap->Transactions[i], 16);
+
+               PRINTF1 (ctxt, DBG_INTR, "                 %s\n", IProcTrapString (&trap->Transactions[i], NULL));
+
+               BumpInputterStats (dev, &trap->Transactions[i]);
+           }
+           
+           /*
+            * Remember the number of transactions we've copied.
+            */
+           trap->NumTransactions = i+1;
+
+           PRINTF1 (ctxt, DBG_INTR, "                 NumTransactions = %d\n", trap->NumTransactions);
+           
+           /*
+            * Copy all the data blocks in one go to let the Elan prefetcher work 
+            */
+           elan3_sdram_copyq_from_sdram (dev, DataOff, trap->DataBuffers, trap->NumTransactions*sizeof (E3_IprocTrapData));
+
+           /*
+            * Copy fault save area and clear out for next time round.
+            */
+           elan3_sdram_copyq_from_sdram (dev, FaultSaveOff, (void *) &trap->FaultSave, 16);
+           elan3_sdram_zeroq_sdram (dev, FaultSaveOff, 16);
+
+           if (ELAN3_OP_IPROC_TRAP (ctxt, trap, Channel) == OP_DEFER)
+           {
+               /*
+                * Mark the trap as valid and set the inputter state to 
+                * raise the context filter.
+                */
+               trap->State = CTXT_STATE_TRAPPED;
+               kcondvar_wakeupone (&ctxt->Wait, &dev->IntrLock);
+               
+               SetInputterStateForContext (ctxt, Pend, NULL);
+           }
+       }
+    }
+}
+
+void
+InspectIProcTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap)
+{
+    int              i;
+    int              StatusValid;
+
+    trap->AckSent                  = 0;
+    trap->BadTransaction            = 0;
+    
+    trap->TrappedTransaction        = NULL;
+    trap->TrappedDataBuffer        = NULL;
+    trap->WaitForEopTransaction     = NULL;
+    trap->WaitForEopDataBuffer      = NULL;
+    trap->DmaIdentifyTransaction    = NULL;
+    trap->ThreadIdentifyTransaction = NULL;
+    trap->LockQueuePointer          = (E3_Addr) 0;
+    trap->UnlockQueuePointer        = (E3_Addr) 0;
+
+    /*
+     * Now scan all the transactions received 
+     */
+    for (i = 0; i < trap->NumTransactions ; i++)
+    {
+       E3_IprocTrapHeader_BE *hdrp = &trap->Transactions[i];
+       E3_IprocTrapData_BE   *datap = &trap->DataBuffers[i];
+
+       StatusValid = hdrp->s.TrTypeCntx.s.StatusRegValid != 0;
+       
+       if (StatusValid && hdrp->s.IProcTrapStatus.s.AckSent)   /* Remember if we've sent the ack back */
+           trap->AckSent = 1;
+       
+       if (hdrp->s.TrTypeCntx.s.LastTrappedTrans)              /* Check for EOP */
+       {
+           ASSERT (i == trap->NumTransactions - 1);
+
+           switch (hdrp->s.IProcTrapStatus.Status & E3_IPS_EopType)
+           {
+           case EOP_GOOD:
+               /* if we get an EOP_GOOD then the outputer should have received a PAckOk. */  
+               /* unless it was a flood, in which case someone must have sent an ack */
+               /* but not necessarily us */
+               break;
+
+           case EOP_BADACK:
+               BumpUserStat (ctxt, EopBadAcks);
+
+               /* if we get an EOP_BADACK then the outputer did not receive a PAckOk even if
+                * we sent a PAckOk. We can clear tinfo.AckSent. */
+               if (trap->AckSent == 1)
+               {
+                   PRINTF0 (ctxt, DBG_IPROC, "InspectIProcTrap: Network error destroyed PAckOk\n");
+                   trap->AckSent = 0;
+               }
+               break;
+
+           case EOP_ERROR_RESET:
+               BumpUserStat (ctxt, EopResets);
+
+               /* if we get an EOP_ERROR_RESET then the outputer may or may not have got a PAckOk. */
+               trap->BadTransaction = 1;
+               break;
+
+           default:
+               panic ("InspectIProcTrap: invalid EOP type in status register\n");
+               /* NOTREACHED */
+           }
+           continue;
+       }
+
+       PRINTF2 (ctxt, DBG_IPROC, "InspectIProcTrap: %2d: %s\n", i, IProcTrapString (hdrp, datap));
+       
+       if (! StatusValid)                                      /* We're looking at transactions stored before the trap */
+       {                                                       /* these should only be identifies and lock transactions */
+
+           if (hdrp->s.TrTypeCntx.s.Type & TR_WRITEBLOCK_BIT)
+               panic ("InspectIProcTrap: writeblock transaction found in input trap header before trap occured\n");
+
+           switch (hdrp->s.TrTypeCntx.s.Type & TR_OPCODE_TYPE_MASK)
+           {
+           case TR_LOCKQUEUE & TR_OPCODE_TYPE_MASK:
+               if (trap->LockQueuePointer)                             /* Already seen a LOCKQUEUE transaction in this packet, */
+               {                                               /* the user program should not have done this !! */
+                   ElanException (ctxt, EXCEPTION_BAD_PACKET, INPUT_PROC, trap);
+                   return;
+               }
+
+               trap->LockQueuePointer = (E3_Addr) hdrp->s.TrAddr;      /* Remember the queue pointer in case we need to unlock it */
+               break;
+
+           case TR_DMAIDENTIFY & TR_OPCODE_TYPE_MASK:
+               if (trap->DmaIdentifyTransaction ||             /* Already seen an identify transaction in this packet */
+                   trap->ThreadIdentifyTransaction)            /* the user program should not have done this */
+               {                                                       
+                   ElanException (ctxt, EXCEPTION_BAD_PACKET, INPUT_PROC, trap);
+                   return;
+               }
+               trap->DmaIdentifyTransaction = hdrp;
+               break;
+
+           case TR_THREADIDENTIFY & TR_OPCODE_TYPE_MASK:
+               if (trap->DmaIdentifyTransaction ||             /* Already seen an identify transaction in this packet */
+                   trap->ThreadIdentifyTransaction)            /* the user program should not have done this */
+               {                                                       
+                   ElanException (ctxt, EXCEPTION_BAD_PACKET, INPUT_PROC, trap);
+                   return;
+               }
+               trap->ThreadIdentifyTransaction = hdrp;
+               break;
+               
+           default:
+               panic ("InspectIProcTrap: invalid transaction found in input trap header before trap occured\n");
+               /* NOTREACHED */
+           }
+           continue;
+       }
+
+       if (StatusValid && trap->TrappedTransaction == NULL)    /* Remember the transaction which caused the */
+       {                                                       /* trap */
+           trap->TrappedTransaction = hdrp;
+           trap->TrappedDataBuffer  = datap;
+       }
+
+       if(hdrp->s.IProcTrapStatus.s.BadLength ||
+          ((hdrp->s.IProcTrapStatus.Status & CRC_MASK) == CRC_STATUS_ERROR) ||
+          ((hdrp->s.IProcTrapStatus.Status & CRC_MASK) == CRC_STATUS_BAD))
+       {
+           int j;
+           PRINTF0 (ctxt, DBG_IPROC, "InspectIProcTrap: transaction has a bad crc\n");
+           for (j=0; j<TRANS_DATA_WORDS; j+=4)
+              PRINTF5 (ctxt, DBG_IPROC, "InspectIProcTrap: Data %0d %8x %8x %8x %8x\n",
+                       j, datap->TrData[j], datap->TrData[j+1], datap->TrData[j+2], datap->TrData[j+3]);
+           trap->BadTransaction = 1;
+           continue;
+       }
+       
+       /* No more to do if it's a writeblock transaction */
+       if (hdrp->s.TrTypeCntx.s.Type & TR_WRITEBLOCK_BIT)
+           continue;
+
+       
+       if (GET_STATUS_TRAPTYPE(hdrp->s.IProcTrapStatus) == MI_InputDoTrap &&
+           (hdrp->s.TrTypeCntx.s.Type & TR_WAIT_FOR_EOP) != 0)
+       {
+           /*
+            * This is a wait for eop transaction that has trapped because the inputer
+            * then received a EopError. The next transaction saved should always be an
+            * EopError.
+            */
+           PRINTF0 (ctxt, DBG_IPROC, "InspectIProcTrap: got a trapped WaitForEop transaction due to EopError\n");
+           
+           trap->WaitForEopTransaction = hdrp;
+           trap->WaitForEopDataBuffer  = datap;
+           continue;
+       }
+
+       switch (hdrp->s.TrTypeCntx.s.Type & TR_OPCODE_TYPE_MASK)
+       {
+       case TR_UNLOCKQUEUE & TR_OPCODE_TYPE_MASK:
+           if (trap->UnlockQueuePointer)
+           {
+               ElanException (ctxt, EXCEPTION_BAD_PACKET, INPUT_PROC, trap);
+               return;
+           }
+           trap->UnlockQueuePointer = (E3_Addr) hdrp->s.TrAddr;
+           break;
+       }
+    }
+}
+
+void
+ResolveIProcTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, NETERR_RESOLVER **rvpp)
+{
+    ELAN3_DEV     *dev = ctxt->Device;
+    int           res;
+    unsigned long flags;
+
+    ASSERT (! CTXT_IS_KERNEL (ctxt));
+
+    BumpUserStat (ctxt, IProcTraps);
+
+    InspectIProcTrap (ctxt, trap);
+
+    /*
+     * fixup page fault if we've trapped because of one.
+     */
+    if (trap->FaultSave.s.FaultContext != 0)
+    {
+       /*
+        * If it's a WRITEBLOCK transaction, then see if we remember faulting
+        * before it, and try and prefault in a sensible amount past it.
+        */
+       int                fixedFault = FALSE;
+       INPUT_FAULT_SAVE  *entry;
+       INPUT_FAULT_SAVE **predp;
+       int                npages;
+
+       if ((trap->TrappedTransaction->s.TrTypeCntx.s.Type & TR_WRITEBLOCK_BIT) != 0 && /* a DMA packet */
+           trap->LockQueuePointer == (E3_Addr) 0 &&                                    /* but not a queueing DMA */
+           trap->TrappedTransaction->s.TrAddr != 0)                                    /* and not a DMA to 0 */
+       {
+           spin_lock (&ctxt->InputFaultLock);
+           
+           for (predp = &ctxt->InputFaultList; (entry = *predp)->Next != NULL ; predp = &entry->Next)
+           {
+               if (entry->Addr == trap->TrappedTransaction->s.TrAddr)
+                   break;
+           }
+           
+           *predp = entry->Next;
+           entry->Next = ctxt->InputFaultList;
+           ctxt->InputFaultList = entry;
+           
+           if (entry->Addr == trap->TrappedTransaction->s.TrAddr)
+           {
+               if ((entry->Count <<= 1) > MAX_INPUT_FAULT_PAGES)
+                   entry->Count = MAX_INPUT_FAULT_PAGES;
+           }
+           else
+           {
+               entry->Count = MIN_INPUT_FAULT_PAGES;
+           }
+           
+           entry->Addr = trap->TrappedTransaction->s.TrAddr + (entry->Count * PAGESIZE);
+           npages = entry->Count;
+           
+           spin_unlock (&ctxt->InputFaultLock);
+           
+           if (elan3_pagefault (ctxt, &trap->FaultSave, npages) != ESUCCESS)
+           {
+               PRINTF2 (ctxt, DBG_IPROC, "ResolveIProcTrap: pagefaulting %d pages at %08x - failed\n", 
+                        npages, trap->TrappedTransaction->s.TrAddr);
+           }
+           else
+           {
+               PRINTF2 (ctxt, DBG_IPROC, "ResolveIProcTrap: pagefaulting %d pages at %08x - succeeded\n", 
+                        npages, trap->TrappedTransaction->s.TrAddr);
+               
+               fixedFault = TRUE;
+           }
+       }
+
+       /* Workaround WRITEBLOCK transaction executed when LOCKQUEUE transaction missed */
+       /* the packet will have been nacked */
+       if ((trap->TrappedTransaction->s.TrTypeCntx.s.Type & TR_WRITEBLOCK_BIT) &&      /* a DMA packet */
+           trap->LockQueuePointer == 0 && trap->UnlockQueuePointer &&                  /* a queueing DMA */
+           trap->TrappedTransaction->s.TrAddr == trap->FaultSave.s.FaultAddress)       /* and missed lockqueue */
+       {
+           fixedFault = TRUE;
+       }
+
+       if (! fixedFault)
+       {
+           if ((res = elan3_pagefault (ctxt, &trap->FaultSave, 1)) != ESUCCESS)
+           {
+               PRINTF1 (ctxt, DBG_IPROC, "ResolveIProcTrap: elan3_pagefault failed at %x\n", 
+                        trap->FaultSave.s.FaultAddress);
+               ElanException (ctxt, EXCEPTION_INVALID_ADDR, INPUT_PROC, trap, &trap->FaultSave, res);
+               return;
+           }
+       }
+    }
+
+    if (! trap->AckSent && trap->LockQueuePointer)                     /* Queued DMA */
+    {                                                                  /* The ack was not sent, so the queue will be locked. */
+       SimulateUnlockQueue (ctxt, trap->LockQueuePointer, FALSE);      /* We must unlock it. */
+    }
+
+    if (trap->AckSent && trap->BadTransaction)
+    {
+       if (trap->DmaIdentifyTransaction)
+       {
+           PRINTF0 (ctxt, DBG_IPROC, "ResolveIProcTrap: Dma identify needs network resultion\n");
+
+           BumpStat (dev, DmaIdentifyNetworkErrors);
+           BumpUserStat (ctxt, DmaIdentifyNetworkErrors);
+
+           if (trap->WaitForEopTransaction)
+               PRINTF0 (ctxt, DBG_IPROC, "ResolveIProcTrap: have delayed wait for eop transaction\n");
+       }
+       else if (trap->ThreadIdentifyTransaction)
+       {
+           PRINTF0 (ctxt, DBG_IPROC, "ResolveIProcTrap: Thread identify needs network resolution\n");
+
+           BumpStat (dev, ThreadIdentifyNetworkErrors);
+           BumpUserStat (ctxt, ThreadIdentifyNetworkErrors);
+
+           if (trap->WaitForEopTransaction)
+               PRINTF0 (ctxt, DBG_IPROC, "ResolveIProcTrap: have delayed wait for eop transaction\n");
+       }
+       else
+       {
+           BumpStat (dev, DmaNetworkErrors);
+           BumpUserStat (ctxt, DmaNetworkErrors);
+       }
+    }
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    
+    if (! trap->AckSent)
+    {
+       PRINTF0 (ctxt, DBG_IPROC, "ResolveIProcTrap: ack not sent, lowering context filter\n");
+
+       trap->State = CTXT_STATE_OK;
+    }
+    else
+    {
+       if (trap->BadTransaction)
+       {
+           PRINTF0 (ctxt, DBG_IPROC, "ResolveIProcTrap: ack sent, waiting on bad transaction\n");
+           trap->State = CTXT_STATE_NETWORK_ERROR;
+       }
+       else
+       {
+           PRINTF0 (ctxt, DBG_IPROC, "ResolveIProcTrap: ack sent, waiting on packet to be re-executed\n");
+           trap->State = CTXT_STATE_NEEDS_RESTART;
+       }
+    }
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    if (trap->AckSent && trap->BadTransaction)
+       ElanException (ctxt, EXCEPTION_NETWORK_ERROR, INPUT_PROC, trap, rvpp);
+}
+
+int
+RestartIProcTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap)
+{
+    PRINTF1 (ctxt, DBG_IPROC, "RestartIProc: %d transactions\n", trap->NumTransactions);
+
+    if (trap->TrappedTransaction == NULL)                      /* No transaction trapped - probably a network */
+       return (ESUCCESS);                                      /* error */
+
+    while (! trap->TrappedTransaction->s.TrTypeCntx.s.LastTrappedTrans)
+    {
+       E3_IprocTrapHeader_BE *hdrp = trap->TrappedTransaction;
+       E3_IprocTrapData_BE   *datap = trap->TrappedDataBuffer;
+       
+       ASSERT (hdrp->s.TrTypeCntx.s.StatusRegValid != 0);
+
+       PRINTF2 (ctxt, DBG_IPROC, "RestartIProc: TrType=0x%x Status=0x%x\n",
+                hdrp->s.TrTypeCntx.TypeContext, hdrp->s.IProcTrapStatus.Status);
+       
+       if ((hdrp->s.TrTypeCntx.s.Type & TR_WRITEBLOCK_BIT) != 0)
+       {
+           PRINTF1 (ctxt, DBG_IPROC, "RestartIProc: WRITEBLOCK : Addr %x\n", hdrp->s.TrAddr);
+           SimulateBlockWrite (ctxt, hdrp, datap);
+       }
+       else
+       {
+           switch (hdrp->s.TrTypeCntx.s.Type & TR_OPCODE_TYPE_MASK)
+           {
+           case TR_SETEVENT & TR_OPCODE_TYPE_MASK:
+               PRINTF1 (ctxt, DBG_IPROC, "RestartIProc: SETEVENT : %x\n", hdrp->s.TrAddr);
+
+               if (GET_STATUS_TRAPTYPE(hdrp->s.IProcTrapStatus) != MI_InputDoTrap)
+                   FixupEventTrap (ctxt, INPUT_PROC, trap, GET_STATUS_TRAPTYPE(hdrp->s.IProcTrapStatus), &trap->FaultSave, FALSE);
+               else if (hdrp->s.TrAddr)
+               {
+                   if (IssueCommand (ctxt, offsetof (E3_CommandPort, SetEvent), hdrp->s.TrAddr, FALSE) != ISSUE_COMMAND_OK)
+                       return (EAGAIN);
+               }
+               break;
+
+           case TR_WRITEWORD & TR_OPCODE_TYPE_MASK:
+               SimulateWriteWord (ctxt, hdrp, datap);
+               break;
+
+           case TR_WRITEDOUBLEWORD & TR_OPCODE_TYPE_MASK:
+               SimulateWriteDWord (ctxt, hdrp, datap);
+               break;
+               
+           case TR_UNLOCKQUEUE & TR_OPCODE_TYPE_MASK:
+               if (GET_STATUS_TRAPTYPE(hdrp->s.IProcTrapStatus) == MI_InputDoTrap)
+                   ElanException (ctxt, EXCEPTION_BAD_PACKET, INPUT_PROC, trap);
+               else
+               {
+                   switch (GET_STATUS_TRAPTYPE (hdrp->s.IProcTrapStatus))
+                   {
+                   case MI_WaitForUnLockDescRead:
+                       /*
+                        * Fault occured on the read of the queue descriptor - since the ack
+                        * has been sent we need to move the queue on one slot.
+                        */
+                       PRINTF0 (ctxt, DBG_IPROC, "RestartIProc: TR_UNLOCKQUEUE : desc read fault\n");
+
+                       SimulateUnlockQueue (ctxt, trap->LockQueuePointer, TRUE);
+                       
+                       if (IssueCommand (ctxt, offsetof (E3_CommandPort, SetEvent),
+                                         hdrp->s.TrAddr + E3_QUEUE_EVENT_OFFSET, FALSE) != ISSUE_COMMAND_OK)
+                       {
+                           /* Failed to issue setevent to complete queue unlock, since we've already unlocked */
+                           /* the queue, we should "convert" this transaction into a setevent transaction that */
+                           /* hasn't trapped */
+                           PRINTF0 (ctxt, DBG_IPROC, "RestartIProc: could not issue setevent for SimulateUnlockQueue\n");
+
+                           ConvertTransactionToSetEvent (ctxt, hdrp, hdrp->s.TrAddr + E3_QUEUE_EVENT_OFFSET);
+                           return (EAGAIN);
+                       }
+                       break;
+                       
+                   case MI_DoSetEvent:
+                       /*
+                        * Fault occured on either the write to unlock the queue or during 
+                        * processing of the event.  Test the fault address against the
+                        * queue address to find out which - in this case, since the ack
+                        * has been sent we need to move the queue on one slot.
+                        */
+                       if (trap->FaultSave.s.FaultAddress == trap->LockQueuePointer)
+                       {
+                           PRINTF0 (ctxt, DBG_IPROC, "RestartIProc: fixed unlock queue write to unlock fault\n");
+
+                           SimulateUnlockQueue (ctxt, trap->LockQueuePointer, TRUE);
+                           
+                           if (IssueCommand (ctxt, offsetof (E3_CommandPort, SetEvent),
+                                             hdrp->s.TrAddr + E3_QUEUE_EVENT_OFFSET, FALSE) != ISSUE_COMMAND_OK)
+                           {
+                               /* Failed to issue setevent to complete queue unlock, since we've already unlocked */
+                               /* the queue, we should "convert" this transaction into a setevent transaction that */
+                               /* hasn't trapped */
+                               PRINTF0 (ctxt, DBG_IPROC, "RestartIProc: could not issue setevent for SimulateUnlockQueue\n");
+                               
+                               ConvertTransactionToSetEvent (ctxt, hdrp, hdrp->s.TrAddr + E3_QUEUE_EVENT_OFFSET);
+                               return (EFAIL);
+                           }
+                           break;
+                       }
+                       /*DROPTHROUGH*/
+                       
+                   default:
+                       FixupEventTrap (ctxt, INPUT_PROC, trap, GET_STATUS_TRAPTYPE (hdrp->s.IProcTrapStatus),
+                                       &trap->FaultSave, FALSE);
+                       break;
+                   }
+                   trap->LockQueuePointer = trap->UnlockQueuePointer = 0;
+               }
+               break;
+
+           case TR_SENDDISCARD & TR_OPCODE_TYPE_MASK:
+               /* Just ignore send-discard transactions */
+               PRINTF0 (ctxt, DBG_IPROC, "RestartIProc: ignore SENDDISCARD\n");
+               break;
+
+           case TR_REMOTEDMA & TR_OPCODE_TYPE_MASK:
+               PRINTF0 (ctxt, DBG_IPROC, "RestartIProc: REMOTEDMA\n");         
+
+               /* modify the dma type since it will still be a "read" dma */
+               ((E3_DMA_BE *) datap)->s.dma_type &= ~(DMA_TYPE_READ | E3_DMA_CONTEXT_MASK);
+               ((E3_DMA_BE *) datap)->s.dma_type |= DMA_TYPE_ISREMOTE;
+
+               RestartDmaDesc (ctxt, (E3_DMA_BE *) datap);
+               break;
+
+           case TR_TRACEROUTE & TR_OPCODE_TYPE_MASK:
+               PRINTF0 (ctxt, DBG_IPROC, "RestartIProc: TRACEROUTE\n");
+               SimulateTraceRoute (ctxt, hdrp, datap);
+               break;
+
+           default:
+               ElanException (ctxt, EXCEPTION_BAD_PACKET, INPUT_PROC, trap);
+               break;
+           }
+       }
+
+       /*
+        * We've successfully processed this transaction, so move onto the 
+        * next one.
+        */
+       trap->TrappedTransaction++;
+       trap->TrappedDataBuffer++;
+    }
+    
+    return (ESUCCESS);
+}
+
+static void
+ConvertTransactionToSetEvent (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_Addr Addr)
+{
+    hdrp->s.TrTypeCntx.s.Type           = TR_SETEVENT;
+    hdrp->s.TrTypeCntx.s.StatusRegValid = 0;
+    hdrp->s.TrAddr                      = Addr;
+}
+
+void
+SimulateBlockWrite (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap)
+{
+    void     *saddr  = (void *) ((unsigned long) datap + (hdrp->s.TrAddr & 0x3f));
+    unsigned  nbytes = (hdrp->s.TrTypeCntx.s.Type) & TR_PARTSIZE_MASK;
+    int       i;
+
+    if (nbytes == 0)
+       nbytes = sizeof (E3_IprocTrapData_BE);
+
+    if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+    {
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+       PRINTF1 (ctxt, DBG_IPROC, "SimulateBlockWrite: faulted at %x\n", hdrp->s.TrAddr);
+       ElanException (ctxt, EXCEPTION_FAULTED, INPUT_PROC, NULL, hdrp->s.TrAddr);
+       return;
+    }
+
+    /*
+     * NOTE: since the block copy could be to sdram, we issue the writes backwards,
+     *       except we MUST ensure that the last item in the block is written last.
+     */
+    switch (((hdrp->s.TrTypeCntx.s.Type) >> TR_TYPE_SHIFT) & TR_TYPE_MASK)
+    {
+    case TR_TYPE_BYTE:                                         /* 8 bit */
+       for (i = nbytes - (2*sizeof (E3_uint8)); i >= 0; i -= sizeof (E3_uint8))
+           ELAN3_OP_STORE8 (ctxt, hdrp->s.TrAddr + i, ((E3_uint8 *) saddr)[i]);
+       i = nbytes - sizeof (E3_uint8);
+       ELAN3_OP_STORE8 (ctxt, hdrp->s.TrAddr + i, ((E3_uint8 *) saddr)[i]);
+       break;
+       
+    case TR_TYPE_SHORT:                                                /* 16 bit */
+       for (i = nbytes - (2*sizeof (E3_uint16)); i >= 0; i -= sizeof (E3_uint16))
+       ELAN3_OP_STORE16 (ctxt, hdrp->s.TrAddr + i, ((E3_uint16 *) saddr)[i]);
+       i = nbytes - sizeof (E3_uint16);
+       ELAN3_OP_STORE16 (ctxt, hdrp->s.TrAddr + i, ((E3_uint16 *) saddr)[i]);
+       break;
+       
+    case TR_TYPE_WORD:                                         /* 32 bit */
+       for (i = nbytes - (2*sizeof (E3_uint32)); i >= 0; i -= sizeof (E3_uint32))
+           ELAN3_OP_STORE32 (ctxt, hdrp->s.TrAddr + i, ((E3_uint32 *) saddr)[i]);
+       i = nbytes - sizeof (E3_uint32);
+       ELAN3_OP_STORE32 (ctxt, hdrp->s.TrAddr + i, ((E3_uint32 *) saddr)[i]);
+       break;
+       
+    case TR_TYPE_DWORD:                                                /* 64 bit  */
+       for (i = nbytes - (2*sizeof (E3_uint64)); i >= 0; i -= sizeof (E3_uint64))
+           ELAN3_OP_STORE64 (ctxt, hdrp->s.TrAddr + i, ((E3_uint64 *) saddr)[i]);
+       i = nbytes - sizeof (E3_uint64);
+       ELAN3_OP_STORE64 (ctxt, hdrp->s.TrAddr + i, ((E3_uint64 *) saddr)[i]);
+       break;
+    }
+    ELAN3_OP_END_FAULT_CHECK (ctxt);
+}
+
+void
+SimulateWriteWord (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap)
+{
+    if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+    {
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+       PRINTF1 (ctxt, DBG_IPROC, "SimulateWriteWord: faulted at %x\n", hdrp->s.TrAddr);
+       ElanException (ctxt, EXCEPTION_FAULTED, INPUT_PROC, NULL, hdrp->s.TrAddr);
+       return;
+    }
+
+    ELAN3_OP_STORE32 (ctxt, hdrp->s.TrAddr, ((E3_uint32 *) datap)[WordEndianFlip]);
+    ELAN3_OP_END_FAULT_CHECK (ctxt);
+}
+
+void
+SimulateWriteDWord (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap)
+{
+    if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+    {
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+       PRINTF1 (ctxt, DBG_IPROC, "SimulateWriteDWord: faulted at %x\n", hdrp->s.TrAddr);
+       ElanException (ctxt, EXCEPTION_FAULTED, INPUT_PROC, NULL, hdrp->s.TrAddr);
+       return;
+    }
+
+    ELAN3_OP_STORE64 (ctxt, hdrp->s.TrAddr, ((E3_uint64 *) datap)[0]);
+    ELAN3_OP_END_FAULT_CHECK (ctxt);
+}
+
+void
+SimulateTraceRoute (ELAN3_CTXT *ctxt, E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap)
+{
+    E3_uint32 *saddr  = (E3_uint32 *) ((unsigned long) datap + (hdrp->s.TrAddr & 0x3f));
+    unsigned   nwords = TrSizeTable[(hdrp->s.TrTypeCntx.s.Type >> TR_SIZE_SHIFT) & TR_SIZE_MASK] / sizeof (E3_uint32);
+    int        i;
+
+    if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+    {
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+       PRINTF1 (ctxt, DBG_IPROC, "SimulateTraceRoute: faulted at %x\n", hdrp->s.TrAddr);
+       ElanException (ctxt, EXCEPTION_FAULTED, INPUT_PROC, NULL, hdrp->s.TrAddr);
+       return;
+    }
+    
+    for (i = nwords-2; i >= 0; i--)
+       ELAN3_OP_STORE32 (ctxt, hdrp->s.TrAddr + (i * sizeof (E3_uint32)), saddr[i ^ WordEndianFlip]);
+
+    i = nwords-1;
+    ELAN3_OP_STORE32 (ctxt, hdrp->s.TrAddr + (i * sizeof (E3_uint32)), saddr[i ^ WordEndianFlip]);
+
+    ELAN3_OP_END_FAULT_CHECK (ctxt);
+}
+
+void
+SimulateUnlockQueue (ELAN3_CTXT *ctxt, E3_Addr QueuePointer, int SentAck)
+{
+    E3_uint32 QueueLock;
+    E3_Addr   QueueBPTR;
+    E3_Addr   QueueFPTR;
+    E3_uint64 QueueStateAndBPTR;
+
+    if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+    {
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+       PRINTF1 (ctxt, DBG_IPROC, "UnlockQueue: faulted with QueuePointer %x\n", QueuePointer);
+       ElanException (ctxt, EXCEPTION_FAULTED, INPUT_PROC, NULL, QueuePointer);
+       return;
+    }
+    
+    if (SentAck)
+    {
+       QueueBPTR = ELAN3_OP_LOAD32 (ctxt, QueuePointer + offsetof (E3_Queue, q_bptr));
+       QueueFPTR = ELAN3_OP_LOAD32 (ctxt, QueuePointer + offsetof (E3_Queue, q_fptr));
+
+       if (QueueBPTR == ELAN3_OP_LOAD32 (ctxt, QueuePointer + offsetof (E3_Queue, q_top)))     /* move on back pointer */
+           QueueBPTR = ELAN3_OP_LOAD32 (ctxt, QueuePointer + offsetof (E3_Queue, q_base));
+       else
+           QueueBPTR += ELAN3_OP_LOAD32 (ctxt, QueuePointer + offsetof (E3_Queue, q_size));
+       
+       QueueLock = ELAN3_OP_LOAD32 (ctxt, QueuePointer + offsetof (E3_Queue, q_state));
+
+       if (QueueBPTR == QueueFPTR)                             /* and set full bit if fptr == bptr */
+           QueueLock |= E3_QUEUE_FULL;
+       
+       QueueLock &= ~E3_QUEUE_LOCKED;
+       
+       QueueStateAndBPTR = (E3_uint64)QueueLock << 32 | QueueBPTR;
+
+       ELAN3_OP_STORE64 (ctxt, QueuePointer + offsetof (E3_Queue, q_state), QueueStateAndBPTR);
+    }
+    else
+    {
+       QueueLock = ELAN3_OP_LOAD32 (ctxt, QueuePointer + offsetof (E3_Queue, q_state));
+
+       QueueLock &= ~E3_QUEUE_LOCKED;
+       
+       ELAN3_OP_STORE32 (ctxt, QueuePointer + offsetof (E3_Queue, q_state), QueueLock);
+    }
+
+    no_fault();
+}
+
+static void
+BumpInputterStats (ELAN3_DEV *dev, E3_IprocTrapHeader_BE *hdrp)
+{
+    if (hdrp->s.TrTypeCntx.s.LastTrappedTrans)                 /* EOP */
+    {
+       switch (hdrp->s.IProcTrapStatus.Status & E3_IPS_EopType)
+       {
+       case EOP_BADACK:
+           BumpStat (dev, EopBadAcks);
+           break;
+       case EOP_ERROR_RESET:
+           BumpStat (dev, EopResets);
+           break;
+       }
+    }
+    else if (hdrp->s.TrTypeCntx.s.StatusRegValid)
+    {
+       /*
+        * Errors are tested in order of badness. i.e. badlength will prevent a BadCrc and so on...
+        */
+       if (hdrp->s.IProcTrapStatus.s.BadLength)
+           BumpStat (dev, InputterBadLength);
+       else if ((hdrp->s.IProcTrapStatus.Status & CRC_MASK) == CRC_STATUS_BAD)
+           BumpStat (dev, InputterCRCBad);
+       else if ((hdrp->s.IProcTrapStatus.Status & CRC_MASK) == CRC_STATUS_ERROR)
+           BumpStat (dev, InputterCRCErrors);
+       else if ((hdrp->s.IProcTrapStatus.Status & CRC_MASK) == CRC_STATUS_DISCARD)
+           BumpStat (dev, InputterCRCDiscards);
+    }
+}
+
+char *
+IProcTrapString (E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData_BE *datap)
+{
+    static char buffer[256];
+    static char typeString[256];
+    static char statusString[256];
+    char *ptr;
+    E3_Addr     Addr        = hdrp->s.TrAddr;
+    E3_uint32   Type        = hdrp->s.TrTypeCntx.s.Type;
+    E3_uint32   Context     = hdrp->s.TrTypeCntx.s.Context;
+    E3_uint32   StatusValid = hdrp->s.TrTypeCntx.s.StatusRegValid;
+    
+    if (hdrp->s.TrTypeCntx.s.LastTrappedTrans)
+    {
+       switch (hdrp->s.IProcTrapStatus.Status & E3_IPS_EopType)
+       {
+       case EOP_GOOD:          sprintf (typeString, "EOP GOOD"); break;
+       case EOP_BADACK:        sprintf (typeString, "EOP BADACK"); break;
+       case EOP_ERROR_RESET:   sprintf (typeString, "EOP ERROR RESET"); break;
+       default:                sprintf (typeString, "EOP - bad status"); break;
+       }
+       sprintf (buffer, "%15s Cntx=%08x", typeString, Context);
+    }
+    else
+    {
+       if (Type & TR_WRITEBLOCK_BIT)
+       {
+           switch ((Type >> TR_TYPE_SHIFT) & TR_TYPE_MASK)
+           {
+           case TR_TYPE_BYTE:  ptr = "Byte";    break;
+           case TR_TYPE_SHORT: ptr = "Short";   break;
+           case TR_TYPE_WORD:  ptr = "Word";    break;
+           case TR_TYPE_DWORD: ptr = "Double";  break;
+           default:            ptr = "Unknown"; break;
+           }
+           
+           sprintf (typeString, "WriteBlock Type=%s Size=%2d", ptr, Type & TR_PARTSIZE_MASK);
+       }
+       else
+       {
+           switch (Type & TR_OPCODE_TYPE_MASK)
+           {
+           case TR_SETEVENT & TR_OPCODE_TYPE_MASK:             sprintf (typeString, "Setevent"); break;
+           case TR_REMOTEDMA & TR_OPCODE_TYPE_MASK:            sprintf (typeString, "Remote DMA"); break;
+           case TR_LOCKQUEUE & TR_OPCODE_TYPE_MASK:            sprintf (typeString, "Lock Queue"); break;
+           case TR_UNLOCKQUEUE & TR_OPCODE_TYPE_MASK:          sprintf (typeString, "Unlock Queue"); break;
+           case TR_SENDDISCARD & TR_OPCODE_TYPE_MASK:          sprintf (typeString, "Send Discard"); break;
+           case TR_DMAIDENTIFY & TR_OPCODE_TYPE_MASK:          sprintf (typeString, "DMA Identify"); break;
+           case TR_THREADIDENTIFY & TR_OPCODE_TYPE_MASK:       sprintf (typeString, "Thread Identify"); break;
+           case TR_GTE & TR_OPCODE_TYPE_MASK:                  sprintf (typeString, "GTE"); break;
+           case TR_LT & TR_OPCODE_TYPE_MASK:                   sprintf (typeString, "LT"); break;
+           case TR_EQ & TR_OPCODE_TYPE_MASK:                   sprintf (typeString, "EQ"); break;
+           case TR_NEQ & TR_OPCODE_TYPE_MASK:                  sprintf (typeString, "NEQ"); break;
+           case TR_WRITEWORD & TR_OPCODE_TYPE_MASK:            sprintf (typeString, "Write Word"); break;
+           case TR_WRITEDOUBLEWORD & TR_OPCODE_TYPE_MASK:      sprintf (typeString, "Write Double"); break;
+           case TR_ATOMICADDWORD & TR_OPCODE_TYPE_MASK:        sprintf (typeString, "Atomic Add"); break;
+           case TR_TESTANDWRITE & TR_OPCODE_TYPE_MASK:         sprintf (typeString, "Test and Write"); break;
+           default:                                            sprintf (typeString, "Type=%d", Type & TR_OPCODE_TYPE_MASK); break;
+           }
+       }
+       sprintf (buffer, "%15s Addr=%08x Cntx=%08x", typeString, Addr, Context);
+       /*(Type & TR_SENDACK)      ? " Sendack" : "", */
+       /*(Type & TR_LAST_TRANS)   ? " LastTrans" : "", */
+       /*(Type & TR_WAIT_FOR_EOP) ? " WaitForEop" : ""); */
+    }
+    
+    if (StatusValid)
+    {
+       sprintf (statusString, " Type=%s %x", MiToName (hdrp->s.IProcTrapStatus.s.TrapType), hdrp->s.IProcTrapStatus.Status);
+       strcat (buffer, statusString);
+
+       if (hdrp->s.IProcTrapStatus.s.BadLength)
+           strcat (buffer, " BadLength");
+       switch (hdrp->s.IProcTrapStatus.Status & CRC_MASK)
+       {
+       case CRC_STATUS_DISCARD:
+           strcat (buffer, " CRC Discard");
+           break;
+       case CRC_STATUS_ERROR:
+           strcat (buffer, " CRC Error");
+           break;
+
+       case CRC_STATUS_BAD:
+           strcat (buffer, " CRC Bad");
+           break;
+       }
+    }
+
+    return (buffer);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/Makefile  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/Makefile       2005-06-01 23:12:54.587440928 -0400
@@ -0,0 +1,31 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/elan3/Makefile
+#
+
+
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/elan3/Makefile
+#
+
+list-multi             := elan3.o
+elan3-objs     := context.o cproc.o dproc.o elandebug.o elandev_generic.o elansyscall.o eventcookie.o iproc.o sdram.o minames.o network_error.o route_table.o tproc.o tprocinsts.o routecheck.o virtual_process.o elan3ops.o context_linux.o elandev_linux.o procfs_linux.o tproc_linux.o elan3mmu_generic.o elan3mmu_linux.o
+export-objs            := elandev_linux.o procfs_linux.o
+obj-$(CONFIG_ELAN3)    := elan3.o
+
+elan3.o : $(elan3-objs)
+       $(LD) -r -o $@ $(elan3-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/elan3/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/Makefile.conf     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/Makefile.conf  2005-06-01 23:12:54.587440928 -0400
@@ -0,0 +1,10 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       elan3.o
+MODULENAME     =       elan3
+KOBJFILES      =       context.o cproc.o dproc.o elandebug.o elandev_generic.o elansyscall.o eventcookie.o iproc.o sdram.o minames.o network_error.o route_table.o tproc.o tprocinsts.o routecheck.o virtual_process.o elan3ops.o context_linux.o elandev_linux.o procfs_linux.o tproc_linux.o elan3mmu_generic.o elan3mmu_linux.o
+EXPORT_KOBJS   =       elandev_linux.o procfs_linux.o
+CONFIG_NAME    =       CONFIG_ELAN3
+SGALFC         =       
+# EXTRALINES START
+
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/elan3/minames.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/minames.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/minames.c      2005-06-01 23:12:54.587440928 -0400
@@ -0,0 +1,38 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: minames.c,v 1.12 2003/06/07 15:57:49 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/minames.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan3/urom_addrs.h>
+
+caddr_t
+MiToName (int mi)
+{
+    static char space[32];
+    static struct {
+       int   mi;
+       char *name;
+    } info[] = {
+#include <elan3/minames.h>
+    };
+    register int i;
+
+
+    for (i = 0; i < sizeof(info)/sizeof(info[0]); i++)
+       if (info[i].mi == mi)
+           return (info[i].name);
+    sprintf (space, "MI %x", mi);
+    return (space);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/network_error.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/network_error.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/network_error.c        2005-06-01 23:12:54.589440624 -0400
@@ -0,0 +1,777 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: network_error.c,v 1.32.2.1 2004/10/28 11:54:57 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/network_error.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elandebug.h>
+
+#ifdef DIGITAL_UNIX
+#include <sys/cred.h>
+#include <sys/mbuf.h>
+#include <sys/utsname.h>
+#include <net/if.h>
+#include <netinet/in.h>
+#include <netinet/in_var.h>
+
+#include <rpc/types.h>
+#include <rpc/auth.h>
+#include <rpc/xdr.h>
+#include <rpc/clnt.h>
+
+typedef xdrproc_t kxdrproc_t;
+#endif
+
+#ifdef LINUX
+#include <linux/sunrpc/types.h>
+#include <linux/sunrpc/auth.h>
+#include <linux/sunrpc/xdr.h>
+#include <linux/sunrpc/clnt.h>
+
+#include <linux/utsname.h>
+#define SYS_NMLN       __NEW_UTS_LEN
+#endif
+
+#include <elan3/neterr_rpc.h>
+
+spinlock_t       ResolveRequestLock;
+kcondvar_t       ResolveRequestWait;
+
+NETERR_RESOLVER  *ResolveRequestHead;
+NETERR_RESOLVER **ResolveRequestTailp = &ResolveRequestHead;
+int              ResolveRequestCount;
+int              ResolveRequestThreads;
+int              ResolveRequestMaxThreads = 4;
+int              ResolveRequestTimeout = 60;
+
+typedef struct neterr_server
+{
+    struct neterr_server *Next;
+    struct neterr_server *Prev;
+    unsigned             ElanId;
+
+    char                *Name;
+    int                          RefCount;
+    struct sockaddr_in    Addr;
+} NETERR_SERVER;
+
+#define NETERR_HASH_ENTRIES    64
+#define NETERR_HASH(elanid)    (((unsigned) elanid) % NETERR_HASH_ENTRIES)
+NETERR_SERVER *NeterrServerHash[NETERR_HASH_ENTRIES];
+kmutex_t       NeterrServerLock;
+
+static NETERR_SERVER *FindNeterrServer (int elanId);
+static void           DereferenceNeterrServer (NETERR_SERVER *server);
+static int            CallNeterrServer (NETERR_SERVER *server, NETERR_MSG *msg);
+
+void
+InitialiseNetworkErrorResolver ()
+{
+    spin_lock_init (&ResolveRequestLock);
+    kcondvar_init (&ResolveRequestWait);
+    
+    ResolveRequestHead  = NULL;
+    ResolveRequestTailp = &ResolveRequestHead;
+
+    kmutex_init (&NeterrServerLock);
+}
+
+void
+FinaliseNetworkErrorResolver ()
+{
+    spin_lock_destroy (&ResolveRequestLock);
+    kcondvar_destroy (&ResolveRequestWait);
+    
+    kmutex_destroy (&NeterrServerLock);
+}
+
+static NETERR_RESOLVER *
+AllocateNetworkErrorResolver (void)
+{
+    NETERR_RESOLVER *rvp;
+
+    KMEM_ZALLOC (rvp, NETERR_RESOLVER *, sizeof (NETERR_RESOLVER), TRUE);
+    spin_lock_init (&rvp->Lock);
+
+    return (rvp);
+}
+
+void
+FreeNetworkErrorResolver (NETERR_RESOLVER *rvp)
+{
+    spin_lock_destroy (&rvp->Lock);
+    KMEM_FREE (rvp, sizeof (NETERR_RESOLVER));
+}
+
+static void
+elan3_neterr_resolver (void)
+{
+    NETERR_RESOLVER *rvp;
+    NETERR_SERVER   *server;
+    int                     status;
+    unsigned long    flags;
+
+    kernel_thread_init("elan3_neterr_resolver");
+    spin_lock (&ResolveRequestLock);
+
+    while ((rvp = ResolveRequestHead) != NULL)
+    {
+       if ((ResolveRequestHead = rvp->Next) == NULL)
+           ResolveRequestTailp = &ResolveRequestHead;
+       
+       spin_unlock (&ResolveRequestLock);
+
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "elan3_neterr_resolver: rvp = %p\n", rvp);
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      Rail          %d\n", rvp->Message.Rail);
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      SrcCapability %s\n", CapabilityString (&rvp->Message.SrcCapability));
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      DstCapability %s\n", CapabilityString (&rvp->Message.DstCapability));
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      CookieAddr    %08x\n", rvp->Message.CookieAddr);
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      CookieVProc   %08x\n", rvp->Message.CookieVProc);
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      NextCookie    %08x\n", rvp->Message.NextCookie);
+       PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      WaitForEop    %08x\n", rvp->Message.WaitForEop);
+       
+       if ((server = FindNeterrServer (rvp->Location.loc_node)) == NULL)
+           status = ECONNREFUSED;
+       else if (ResolveRequestTimeout && ((int)(lbolt - rvp->Timestamp)) > (ResolveRequestTimeout*HZ))
+       {
+           printk ("elan_neterr: rpc to '%s' timedout - context %d killed\n", server->Name, rvp->Message.SrcCapability.cap_mycontext);
+           status = ECONNABORTED;
+       }
+       else
+       {
+           status = CallNeterrServer (server, &rvp->Message);
+
+           DereferenceNeterrServer (server);
+       }
+       
+       if ((status == EINTR || status == ETIMEDOUT) && rvp->Ctxt != NULL)
+       {
+           PRINTF1 (DBG_DEVICE, DBG_NETERR, "elan3_neterr_resolver: retry rvp=%p\n", rvp);
+           spin_lock (&ResolveRequestLock);
+           rvp->Next = NULL;
+           *ResolveRequestTailp = rvp;
+           ResolveRequestTailp = &rvp->Next;
+       }
+       else
+       {
+           rvp->Status = status;
+           
+           spin_lock (&rvp->Lock);
+           
+           if (rvp->Ctxt != NULL)
+           {
+               PRINTF2 (rvp->Ctxt, DBG_NETERR, "elan3_neterr_resolver: completing rvp %p for ctxt %p\n", rvp, rvp->Ctxt);
+               spin_lock_irqsave (&rvp->Ctxt->Device->IntrLock, flags);
+               
+               rvp->Completed = TRUE;
+               
+               kcondvar_wakeupall (&rvp->Ctxt->Wait, &rvp->Ctxt->Device->IntrLock);
+               
+               /*
+                * drop the locks out of order since the rvp can get freeed
+                * as soon as we drop the IntrLock - so cannot reference the
+                * rvp after this.
+                */
+               
+               spin_unlock (&rvp->Lock);
+               spin_unlock_irqrestore (&rvp->Ctxt->Device->IntrLock, flags);
+           }
+           else
+           {
+               PRINTF2 (DBG_DEVICE, DBG_NETERR, "elan3_neterr_resolver: completing rvp %p for deceased ctxt %p\n", rvp, rvp->Ctxt);
+               spin_unlock (&rvp->Lock);
+               FreeNetworkErrorResolver (rvp);
+           }
+           
+           spin_lock (&ResolveRequestLock);
+           ResolveRequestCount--;
+       }
+    }
+
+    ResolveRequestThreads--;
+
+    spin_unlock (&ResolveRequestLock);
+    kernel_thread_exit();
+}
+
+int
+QueueNetworkErrorResolver (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, NETERR_RESOLVER **rvpp)
+{
+    int                           isdma   = trap->DmaIdentifyTransaction != NULL;
+    E3_IprocTrapHeader_BE *hdrp    = isdma ? trap->DmaIdentifyTransaction : trap->ThreadIdentifyTransaction;
+    E3_uint32              process = isdma ? (hdrp->s.TrAddr & 0xFFFF) : (hdrp->s.TrData0 & 0xFFFF);
+    NETERR_RESOLVER       *rvp;
+
+    PRINTF2 (ctxt, DBG_NETERR, "QueueNetworkErrorResolver: process = %d %s\n", process, isdma ? "(dma)" : "(thread)");
+
+    if ((rvp = AllocateNetworkErrorResolver()) == NULL)
+    {
+       PRINTF0 (ctxt, DBG_NETERR, "QueueNetworkErrorResolver: cannot allocate resolver\n");
+       return (ENOMEM);
+    }
+
+    rvp->Message.Rail = ctxt->Device->Devinfo.dev_rail;
+
+    krwlock_read (&ctxt->VpLock);
+    rvp->Location = ProcessToLocation (ctxt, NULL, process, &rvp->Message.SrcCapability);
+    krwlock_done (&ctxt->VpLock);
+
+    if (rvp->Location.loc_node == ELAN3_INVALID_NODE)
+    {
+       PRINTF0 (ctxt, DBG_NETERR, "QueueNetworkErrorResolver: invalid elan id\n");
+
+       FreeNetworkErrorResolver (rvp);
+       return (EINVAL);
+    }
+
+    rvp->Message.DstCapability = ctxt->Capability;
+    rvp->Message.DstProcess    = elan3_process (ctxt);
+    rvp->Message.WaitForEop    = (trap->WaitForEopTransaction != NULL);
+
+    if (isdma)
+    {
+       rvp->Message.CookieAddr  = 0;
+       rvp->Message.CookieVProc = hdrp->s.TrAddr;
+       rvp->Message.NextCookie  = 0;
+    }
+    else
+    {
+       rvp->Message.CookieAddr  = hdrp->s.TrAddr;
+       rvp->Message.CookieVProc = hdrp->s.TrData0;
+       rvp->Message.NextCookie  = hdrp->s.TrData1;
+    }
+
+    rvp->Completed = FALSE;
+    rvp->Ctxt      = ctxt;
+    rvp->Timestamp = lbolt;
+
+    spin_lock (&ResolveRequestLock);
+
+    rvp->Next = NULL;
+    *ResolveRequestTailp = rvp;
+    ResolveRequestTailp = &rvp->Next;
+    ResolveRequestCount++;
+
+    kcondvar_wakeupone (&ResolveRequestWait, &ResolveRequestLock);
+
+    if (ResolveRequestCount < ResolveRequestThreads || ResolveRequestThreads >= ResolveRequestMaxThreads)
+       spin_unlock (&ResolveRequestLock);
+    else
+    {
+       ResolveRequestThreads++;
+
+       spin_unlock (&ResolveRequestLock);
+       if (kernel_thread_create (elan3_neterr_resolver, NULL) == NULL)
+       {
+           spin_lock (&ResolveRequestLock);
+           ResolveRequestThreads--;
+           spin_unlock (&ResolveRequestLock);
+           
+           if (ResolveRequestThreads == 0)
+           {
+               PRINTF0 (ctxt, DBG_NETERR, "QueueNetworkErrorResolver: cannot thread pool\n");
+
+               FreeNetworkErrorResolver (rvp);
+               return (ENOMEM);
+           }
+       }
+    }
+
+    *rvpp = rvp;
+    return (ESUCCESS);
+}
+
+void
+CancelNetworkErrorResolver (NETERR_RESOLVER *rvp)
+{
+    spin_lock (&rvp->Lock);
+
+    PRINTF2 (rvp->Ctxt, DBG_NETERR, "CancelNetworkErrorResolver: rvp=%p %s\n", rvp, rvp->Completed ? "Completed" : "Pending");
+
+    if (rvp->Completed)
+    {
+       spin_unlock (&rvp->Lock);
+       FreeNetworkErrorResolver (rvp);
+    }
+    else
+    {
+       rvp->Ctxt = NULL;
+       spin_unlock (&rvp->Lock);
+    }
+}
+
+static NETERR_FIXUP *
+AllocateNetworkErrorFixup (void)
+{
+    NETERR_FIXUP *nef;
+
+    KMEM_ZALLOC (nef, NETERR_FIXUP *, sizeof (NETERR_FIXUP), TRUE);
+
+    if (nef == (NETERR_FIXUP *) NULL)
+       return (NULL);
+
+    kcondvar_init (&nef->Wait);
+
+    return (nef);
+}
+
+static void
+FreeNetworkErrorFixup (NETERR_FIXUP *nef)
+{
+    kcondvar_destroy (&nef->Wait);
+    KMEM_FREE (nef, sizeof (NETERR_FIXUP));
+}
+
+int
+ExecuteNetworkErrorFixup (NETERR_MSG *msg)
+{
+    ELAN3_DEV      *dev;
+    ELAN3_CTXT   *ctxt;
+    NETERR_FIXUP  *nef;
+    NETERR_FIXUP **predp;
+    int                   rc;
+    unsigned long  flags;
+
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "ExecuteNetworkErrorFixup: msg = %p\n", msg);
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      Rail          %d\n", msg->Rail);
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      SrcCapability %s\n", CapabilityString (&msg->SrcCapability));
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      DstCapability %s\n", CapabilityString (&msg->DstCapability));
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      CookieAddr    %08x\n", msg->CookieAddr);
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      CookieVProc   %08x\n", msg->CookieVProc);
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      NextCookie    %08x\n", msg->NextCookie);
+    PRINTF1 (DBG_DEVICE, DBG_NETERR, "                      WaitForEop    %08x\n", msg->WaitForEop);
+       
+    if ((dev = elan3_device (msg->Rail)) == NULL)
+       return (ESRCH);
+
+    if ((nef = AllocateNetworkErrorFixup()) == NULL)
+       return (ENOMEM);
+
+    if (nef == (NETERR_FIXUP *) NULL)
+       return (ENOMEM);
+    
+    bcopy (msg, &nef->Message, sizeof (NETERR_MSG));
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    
+    ctxt = ELAN3_DEV_CTX_TABLE(dev, msg->SrcCapability.cap_mycontext);
+
+    if (ctxt == NULL)
+       rc = ESRCH;
+    else if (!ELAN_CAP_MATCH (&msg->SrcCapability, &ctxt->Capability))
+       rc = EPERM;
+    else
+    {  
+       if (ctxt->Status & CTXT_NO_LWPS)
+           rc = EAGAIN;
+       else
+       {
+           for (predp = &ctxt->NetworkErrorFixups; *predp != NULL; predp = &(*predp)->Next)
+               ;
+           nef->Next = NULL;
+           *predp = nef;
+           
+           kcondvar_wakeupone (&ctxt->Wait, &dev->IntrLock);
+
+           while (! nef->Completed)
+               kcondvar_wait (&nef->Wait, &dev->IntrLock, &flags);
+
+           rc = nef->Status;
+       }
+    }
+    
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    FreeNetworkErrorFixup (nef);
+
+    return (rc);
+}
+
+void
+CompleteNetworkErrorFixup (ELAN3_CTXT *ctxt, NETERR_FIXUP *nef, int status)
+{
+    ELAN3_DEV *dev = ctxt->Device;
+    unsigned long flags;
+
+    PRINTF2 (ctxt, DBG_NETERR, "CompleteNetworkErrorFixup: %p %d\n", nef, status);
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    nef->Status = status;
+    nef->Completed = TRUE;
+    kcondvar_wakeupone (&nef->Wait, &dev->IntrLock);
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+
+static NETERR_SERVER *
+NewNeterrServer (int elanId, struct sockaddr_in *addr, char *name)
+{
+    NETERR_SERVER *server;
+
+    KMEM_ZALLOC (server, NETERR_SERVER *, sizeof (NETERR_SERVER), TRUE);
+    KMEM_ALLOC  (server->Name, char *, strlen (name)+1, TRUE);
+
+    bcopy (addr, &server->Addr, sizeof (struct sockaddr_in));
+    bcopy (name, server->Name, strlen (name)+1);
+
+    server->ElanId   = elanId;
+    server->RefCount = 1;
+    
+    return (server);
+}
+
+static void
+DeleteNeterrServer (NETERR_SERVER *server)
+{
+    KMEM_FREE (server->Name, strlen(server->Name)+1);
+    KMEM_FREE (server, sizeof (NETERR_SERVER));
+}
+
+static NETERR_SERVER *
+FindNeterrServer (int elanId)
+{
+    NETERR_SERVER *server;
+    
+    kmutex_lock (&NeterrServerLock);
+    
+    for (server = NeterrServerHash[NETERR_HASH(elanId)]; server != NULL; server = server->Next)
+       if (server->ElanId == elanId)
+           break;
+
+    if (server != NULL)
+       server->RefCount++;
+    kmutex_unlock (&NeterrServerLock);
+
+    return (server);
+}
+
+static void
+DereferenceNeterrServer (NETERR_SERVER *server)
+{
+    kmutex_lock (&NeterrServerLock);
+    if ((--server->RefCount) == 0)
+       DeleteNeterrServer (server);
+    kmutex_unlock  (&NeterrServerLock);
+}
+
+int
+AddNeterrServer (int elanId, struct sockaddr_in *addr, char *name)
+{
+    NETERR_SERVER *server;
+    NETERR_SERVER *old;
+    int            hashval = NETERR_HASH(elanId);
+
+    server = NewNeterrServer (elanId, addr, name);
+    
+    if (server == NULL)
+       return (ENOMEM);
+    
+    kmutex_lock (&NeterrServerLock);
+    for (old = NeterrServerHash[hashval]; old != NULL; old = old->Next)
+       if (old->ElanId == elanId)
+           break;
+    
+    /* remove "old" server from hash table */
+    if (old != NULL)
+    {
+       if (old->Prev)
+           old->Prev->Next = old->Next;
+       else
+           NeterrServerHash[hashval] = old->Next;
+       if (old->Next)
+           old->Next->Prev = old->Prev;
+    }
+
+    /* insert "new" server into hash table */
+    if ((server->Next = NeterrServerHash[hashval]) != NULL)
+       server->Next->Prev = server;
+    server->Prev = NULL;
+    NeterrServerHash[hashval] = server;
+
+    kmutex_unlock (&NeterrServerLock);
+
+    if (old != NULL)
+       DereferenceNeterrServer (old);
+    
+    return (ESUCCESS);
+}
+
+int
+AddNeterrServerSyscall (int elanId, void *addrp, void *namep, char *unused)
+{
+    struct sockaddr_in addr;
+    char              *name;
+    int                error;
+    int                nob;
+
+    /* Sanity check the supplied elanId argument */
+    if (elanId < 0)
+       return ( set_errno(EINVAL) );
+
+    KMEM_ALLOC (name, caddr_t, SYS_NMLN, TRUE);
+    
+    if (copyin ((caddr_t) addrp, (caddr_t) &addr, sizeof (addr)) ||
+       copyinstr ((caddr_t) namep, name, SYS_NMLN, &nob))
+    {
+       error = EFAULT;
+    }
+    else
+    {
+       PRINTF2 (DBG_DEVICE, DBG_NETERR, "AddNeterrServer: '%s' at elanid %d\n", name, elanId);
+
+       error = AddNeterrServer (elanId, &addr, name);
+    }
+    KMEM_FREE (name, SYS_NMLN);
+
+    return (error ? set_errno(error) : ESUCCESS);
+}
+
+
+#if defined(DIGITAL_UNIX)
+static int
+CallNeterrServer (NETERR_SERVER *server, NETERR_MSG *msg)
+{
+    cred_t        *cr = crget();
+    struct rpc_err  rpcerr;
+    extern cred_t  *kcred;
+    struct timeval  wait;
+    enum clnt_stat  rc;
+    int                    status;
+    CLIENT         *clnt;
+    int             error;
+
+    PRINTF4 (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s) - family=%d port=%d addr=%08x\n", server->Name,
+            server->Addr.sin_family, server->Addr.sin_port, server->Addr.sin_addr.s_addr);
+
+    if ((clnt = clntkudp_create (&server->Addr, (struct sockaddr_in *)0, NETERR_PROGRAM, NETERR_VERSION, 1, cr)) == NULL)
+    {
+       PRINTF1 (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s): clntkudp_create error\n", server->Name);
+
+       return (ENOMEM);
+    }
+    
+    wait.tv_sec  = NETERR_RPC_TIMEOUT;
+    wait.tv_usec = 0;
+    
+    PRINTF2 (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s): CLNT_CALL timeout = %d\n", server->Name, NETERR_RPC_TIMEOUT);
+   
+    rc = CLNT_CALL(clnt, NETERR_FIXUP_RPC, xdr_neterr_msg, (void *)msg, xdr_int, (void *) &status, wait);
+
+    PRINTF3 (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s): CLNT_CALL -> %d (%s)\n", server->Name, rc, clnt_sperrno(rc));;
+
+    switch (rc)
+    {
+    case RPC_SUCCESS:
+       break;
+
+    case RPC_INTR:
+       status = EINTR;
+       break;
+
+    case RPC_TIMEDOUT:
+       status = ETIMEDOUT;
+       break;
+
+    default:
+       printf ("CallNeterrServer(%s): %s\n", server->Name, clnt_sperrno(status));
+       status = ENOENT;
+       break;
+    }
+
+    CLNT_DESTROY(clnt);
+
+    crfree(cr);
+    
+    ASSERT(rc == RPC_SUCCESS || status != 0);
+
+    PRINTF2 (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s): status=%d\n", server->Name, status);
+
+    return (status);
+}
+#endif
+
+#if defined(LINUX)
+
+#define xdrsize(type) ((sizeof(type) + 3) >> 2)
+
+static int
+xdr_error(struct rpc_rqst *req, u32 *p, void *dummy)
+{
+    return -EIO;
+}
+
+static int
+xdr_decode_int(struct rpc_rqst *req, u32 *p, int *res)
+{ 
+    *res = ntohl(*p++);
+    return 0;
+}
+
+#define XDR_capability_sz ((12 + BT_BITOUL(ELAN3_MAX_VPS)) * sizeof (u32))
+
+static int
+xdr_encode_capability(u32 *p, ELAN_CAPABILITY *cap)
+{
+    u32 *pp = p;
+
+    /* basic xdr unit is u32 - for opaque types we must round up to that */
+    memcpy(p, &cap->cap_userkey, sizeof(cap->cap_userkey));
+    p += xdrsize(cap->cap_userkey);
+
+    *p++ = htonl(cap->cap_version);
+    ((u16 *) (p++))[1] = htons(cap->cap_type);
+    *p++ = htonl(cap->cap_lowcontext);
+    *p++ = htonl(cap->cap_highcontext);
+    *p++ = htonl(cap->cap_mycontext);
+    *p++ = htonl(cap->cap_lownode);
+    *p++ = htonl(cap->cap_highnode);
+    *p++ = htonl(cap->cap_railmask);
+
+    memcpy(p, &cap->cap_bitmap[0], sizeof(cap->cap_bitmap));
+    p += xdrsize(cap->cap_bitmap);
+
+    ASSERT (((unsigned long) p - (unsigned long) pp) == XDR_capability_sz);
+
+    return (p - pp);
+}
+
+
+#define XDR_neterr_sz  (((1 + 5) * sizeof (u32)) + (2*XDR_capability_sz))
+
+static int
+xdr_encode_neterr_msg(struct rpc_rqst *req, u32 *p, NETERR_MSG *msg)
+{
+    u32 *pp = p;
+
+    *p++ = htonl(msg->Rail);
+
+    p += xdr_encode_capability(p, &msg->SrcCapability);
+    p += xdr_encode_capability(p, &msg->DstCapability);
+
+    *p++ = htonl(msg->DstProcess);
+    *p++ = htonl(msg->CookieAddr);
+    *p++ = htonl(msg->CookieVProc);
+    *p++ = htonl(msg->NextCookie);
+    *p++ = htonl(msg->WaitForEop);
+
+    ASSERT (((unsigned long) p - (unsigned long) pp) == XDR_neterr_sz);
+
+    req->rq_slen = xdr_adjust_iovec(req->rq_svec, p);
+
+    return 0;
+}
+
+static struct rpc_procinfo neterr_procedures[2] = 
+{
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
+#      define RPC_ID_NULL      "neterr_null"
+#      define RPC_ID_FIXUP_RPC "neterr_fixup_rpc"
+#else
+#      define RPC_ID_NULL      NETERR_NULL_RPC
+#      define RPC_ID_FIXUP_RPC NETERR_FIXUP_RPC
+#endif
+    {  
+       RPC_ID_NULL,                    /* procedure name or number*/
+       (kxdrproc_t) xdr_error,         /* xdr encode fun */
+        (kxdrproc_t) xdr_error,        /* xdr decode fun */
+       0,                              /* req buffer size */
+       0,                              /* call count */
+    },
+    {  
+       RPC_ID_FIXUP_RPC,
+        (kxdrproc_t) xdr_encode_neterr_msg,
+        (kxdrproc_t) xdr_decode_int,
+       XDR_neterr_sz,
+       0,                      
+    },
+};
+
+static struct rpc_version neterr_version1 = 
+{
+    1,                         /* version */
+    2,                         /* number of procedures */
+    neterr_procedures  /* procedures */
+};
+
+static struct rpc_version *neterr_version[] = 
+{
+    NULL,
+    &neterr_version1,
+};
+
+static struct rpc_stat neterr_stats;
+
+static struct rpc_program neterr_program = 
+{
+    NETERR_SERVICE,
+    NETERR_PROGRAM,
+    sizeof(neterr_version)/sizeof(neterr_version[0]),
+    neterr_version,
+    &neterr_stats,
+};
+
+static int
+CallNeterrServer (NETERR_SERVER *server, NETERR_MSG *msg)
+{
+    struct rpc_xprt   *xprt;
+    struct rpc_clnt   *clnt;
+    struct rpc_timeout to;
+    int                rc, status;
+    
+    PRINTF (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s)\n", server->Name);
+
+    xprt_set_timeout(&to, 1, NETERR_RPC_TIMEOUT * HZ);
+
+    if ((xprt = xprt_create_proto(IPPROTO_UDP, &server->Addr, &to)) == NULL)
+    {
+       PRINTF (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s) xprt_create_proto failed\n", server->Name);
+       return EFAIL;
+    }
+
+    if ((clnt = rpc_create_client(xprt, server->Name, &neterr_program, NETERR_VERSION, RPC_AUTH_NULL)) == NULL)
+    {
+       PRINTF (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s) rpc_create_client failed\n", server->Name);
+       xprt_destroy (xprt);
+       
+       return EFAIL;
+    }
+
+    clnt->cl_softrtry = 1;
+    clnt->cl_chatty   = 0;
+    clnt->cl_oneshot  = 1;
+    clnt->cl_intr     = 0;
+
+    if ((rc = rpc_call(clnt, NETERR_FIXUP_RPC, msg, &status, 0)) < 0)
+    {
+       /* RPC error has occured - determine whether we should retry */
+
+       status = ETIMEDOUT;
+    }
+
+    PRINTF (DBG_DEVICE, DBG_NETRPC, "CallNeterrServer(%s): -> %d\n", server->Name, status);
+
+    return (status);
+}
+
+#endif /* defined(LINUX) */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/procfs_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/procfs_linux.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/procfs_linux.c 2005-06-01 23:12:54.589440624 -0400
@@ -0,0 +1,195 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: procfs_linux.c,v 1.21 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/procfs_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elandebug.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanvp.h>
+
+#include <linux/module.h>
+#include <linux/ctype.h>
+
+#include <qsnet/procfs_linux.h>
+
+struct proc_dir_entry *elan3_procfs_root;
+struct proc_dir_entry *elan3_config_root;
+
+static int
+proc_read_position (char *page, char **start, off_t off,
+                   int count, int *eof, void *data)
+{
+    ELAN3_DEV *dev = (ELAN3_DEV *) data;
+    int       len;
+
+    if (dev->Position.pos_mode == ELAN_POS_UNKNOWN)
+       len = sprintf (page, "<unknown>\n");
+    else
+       len = sprintf (page, 
+                      "NodeId                 %d\n"
+                      "NumLevels              %d\n"
+                      "NumNodes               %d\n",
+                      dev->Position.pos_nodeid, dev->Position.pos_levels, dev->Position.pos_nodes);
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+static int
+proc_write_position (struct file *file, const char *buf, unsigned long count, void *data)
+{
+    ELAN3_DEV *dev      = (ELAN3_DEV *) data;
+    unsigned  nodeid   = ELAN3_INVALID_NODE;
+    unsigned  numnodes = 0;
+    char     *page, *p;
+    int       res;
+
+    if (count == 0)
+       return (0);
+
+    if (count >= PAGE_SIZE)
+       return (-EINVAL);
+
+    if ((page = (char *) __get_free_page (GFP_KERNEL)) == NULL)
+       return (-ENOMEM);
+
+    MOD_INC_USE_COUNT;
+
+    if (copy_from_user (page, buf, count))
+       res = -EFAULT;
+    else
+    {
+       page[count] = '\0';
+       
+       if (page[count-1] == '\n')
+           page[count-1] = '\0';
+
+       if (! strcmp (page, "<unknown>"))
+       {
+           dev->Position.pos_mode      = ELAN_POS_UNKNOWN;
+           dev->Position.pos_nodeid    = ELAN3_INVALID_NODE;
+           dev->Position.pos_nodes     = 0;
+           dev->Position.pos_levels    = 0;
+       }
+       else
+       {
+           for (p = page; *p; )
+           {
+               while (isspace (*p))
+                   p++;
+               
+               if (! strncmp (p, "NodeId=", strlen("NodeId=")))
+                   nodeid   = simple_strtoul (p + strlen ("NodeId="), NULL, 0);
+               if (! strncmp (p, "NumNodes=", strlen ("NumNodes=")))
+                   numnodes = simple_strtoul (p + strlen ("NumNodes="), NULL, 0);
+               
+               while (*p && !isspace(*p))
+                   p++;
+           }
+
+           if (ComputePosition (&dev->Position, nodeid, numnodes, dev->Devinfo.dev_num_down_links_value) != 0)
+               printk ("elan%d: invalid values for NodeId=%d NumNodes=%d\n", dev->Instance, nodeid, numnodes);
+           else
+               printk ("elan%d: setting NodeId=%d NumNodes=%d NumLevels=%d\n", dev->Instance, dev->Position.pos_nodeid,
+                       dev->Position.pos_nodes, dev->Position.pos_levels);
+       }
+    }
+
+    MOD_DEC_USE_COUNT;
+    free_page ((unsigned long) page);
+
+    return (count);
+}
+
+
+void
+elan3_procfs_device_init (ELAN3_DEV *dev)
+{
+    struct proc_dir_entry *dir, *p;
+    char name[NAME_MAX];
+
+    sprintf (name, "device%d", dev->Instance);
+    dir = dev->Osdep.procdir = proc_mkdir (name, elan3_procfs_root);
+
+    if ((p = create_proc_entry ("position", 0, dir)) != NULL)
+    {
+       p->read_proc  = proc_read_position;
+       p->write_proc = proc_write_position;
+       p->data       = dev;
+       p->owner      = THIS_MODULE;
+    }
+
+}
+
+void
+elan3_procfs_device_fini (ELAN3_DEV *dev)
+{
+    struct proc_dir_entry *dir = dev->Osdep.procdir;
+    char name[NAME_MAX];
+
+    remove_proc_entry ("position", dir);
+
+    sprintf (name, "device%d", dev->Instance);
+    remove_proc_entry (name, elan3_procfs_root);
+}
+
+void
+elan3_procfs_init()
+{
+    extern int eventint_punt_loops;
+    extern int ResolveRequestTimeout;
+
+    elan3_procfs_root = proc_mkdir("elan3",  qsnet_procfs_root);
+
+    elan3_config_root = proc_mkdir("config", elan3_procfs_root);
+
+    qsnet_proc_register_hex (elan3_config_root, "elan3_debug",           &elan3_debug,           0);
+    qsnet_proc_register_hex (elan3_config_root, "elan3_debug_console",   &elan3_debug_console,   0);
+    qsnet_proc_register_hex (elan3_config_root, "elan3_debug_buffer",    &elan3_debug_buffer,    0);
+    qsnet_proc_register_hex (elan3_config_root, "elan3mmu_debug",      &elan3mmu_debug,      0);
+    qsnet_proc_register_int (elan3_config_root, "eventint_punt_loops", &eventint_punt_loops, 0);
+    qsnet_proc_register_int (elan3_config_root, "neterr_timeout",      &ResolveRequestTimeout, 0);
+
+#if defined(__ia64__)
+    {
+       extern int enable_sdram_writecombining;
+       qsnet_proc_register_int (elan3_config_root, "enable_sdram_writecombining", &enable_sdram_writecombining, 0);
+    }
+#endif
+}
+
+void
+elan3_procfs_fini()
+{
+#if defined(__ia64__)
+    remove_proc_entry ("enable_sdram_writecombining", elan3_config_root);
+#endif
+    remove_proc_entry ("neterr_timeout",      elan3_config_root);
+    remove_proc_entry ("eventint_punt_loops", elan3_config_root);
+    remove_proc_entry ("elan3mmu_debug",      elan3_config_root);
+    remove_proc_entry ("elan3_debug_buffer",    elan3_config_root);
+    remove_proc_entry ("elan3_debug_console",   elan3_config_root);
+    remove_proc_entry ("elan3_debug",           elan3_config_root);
+
+    remove_proc_entry ("config",  elan3_procfs_root);
+    remove_proc_entry ("version", elan3_procfs_root);
+ 
+    remove_proc_entry ("elan3",  qsnet_procfs_root);
+}
+
+EXPORT_SYMBOL(elan3_procfs_root);
+EXPORT_SYMBOL(elan3_config_root);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/quadrics_version.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/quadrics_version.h     2005-06-01 23:12:54.589440624 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/elan3/routecheck.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/routecheck.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/routecheck.c   2005-06-01 23:12:54.590440472 -0400
@@ -0,0 +1,313 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+/* ------------------------------------------------------------- */
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+#include <elan3/vmseg.h>
+
+/* ---------------------------------------------------------------------- */
+typedef struct elan3_net_location {
+    int netid;
+    int plane;
+    int level;
+} ELAN3_NET_LOCATION;
+/* ---------------------------------------------------------------------- */
+#define FLIT_LINK_ARRAY_MAX (ELAN3_MAX_LEVELS*2)
+/* ---------------------------------------------------------------------- */
+int 
+elan3_route_follow_link( ELAN3_CTXT *ctxt, ELAN3_NET_LOCATION *loc, int link)
+{
+    ELAN_POSITION *pos = &ctxt->Position;
+
+    if ((link<0) || (link>7)) 
+    {
+       PRINTF1 (ctxt, DBG_VP, "elan3_route_follow_link: link (%d) out of range \n",link);
+       return (ELAN3_ROUTE_INVALID);
+    }   
+
+    /* going up or down ? */
+    if ( link >= pos->pos_arity[loc->level] ) 
+    {
+       /* Up */
+       if (loc->level >= pos->pos_levels)
+           loc->plane = 0;
+       else
+       {
+           if ((loc->level == 1) && (pos->pos_arity[0]  == 8)) /* oddness in some machines ie 512 */
+               loc->plane = (16 * ( loc->plane / 8 )) + (4 * ( loc->plane % 4)) 
+                   +(link - pos->pos_arity[loc->level]);
+           else
+               loc->plane = (loc->plane * (8 - pos->pos_arity[loc->level]))
+                   +(link - pos->pos_arity[loc->level]);
+       }
+       loc->level--; 
+       if ( loc->level < 0 )
+       {
+           PRINTF0 (ctxt, DBG_VP, "elan3_route_follow_link: link goes off the top\n");
+           return (ELAN3_ROUTE_INVALID_LEVEL);
+       }
+       loc->netid = loc->netid / pos->pos_arity[loc->level];
+    }
+    else
+    {
+       /* going down */
+       if ((loc->level == 0) && (pos->pos_arity[0] == 8)) /* oddness in some machines ie 512 */
+           loc->netid = link % 2;
+       else
+           loc->netid =(loc->netid * pos->pos_arity[loc->level])+link;
+
+       loc->level++;
+       if (loc->level > pos->pos_levels)
+       {
+           PRINTF0 (ctxt, DBG_VP, "elan3_route_follow_link: link goes off the bottom\n");
+           return (ELAN3_ROUTE_INVALID_LEVEL); 
+       }
+
+       if ( loc->level >= (pos->pos_levels-1))
+           loc->plane = 0;
+       else
+           if ((loc->level == 1) && (pos->pos_arity[0] == 8)) /* oddness in some machines ie 512 */
+               loc->plane = (((loc->plane)>>2)*2) - ( ((loc->plane)>>2) & 3  ) + ((link<2)?0:4); /* ((p/4) % 4) */
+           else 
+               loc->plane = loc->plane/(8-pos->pos_arity[loc->level]);
+    }
+    return (ELAN3_ROUTE_SUCCESS);
+}
+/* ---------------------------------------------------------------------- */
+int /* assumes they are connected, really only used for finding the MyLink */
+elan3_route_get_mylink (ELAN_POSITION *pos, ELAN3_NET_LOCATION *locA, ELAN3_NET_LOCATION *locB)
+{
+    /* whats the My Link for locA to LocB */
+    if ( locA->level > locB->level ) 
+       return locB->plane - (locA->plane * (8 - pos->pos_arity[locA->level])) + pos->pos_arity[locA->level];
+    
+    return  locB->netid - (locA->netid * pos->pos_arity[locA->level]);
+}
+/* ---------------------------------------------------------------------- */
+#define FIRST_GET_HIGH_PRI(FLIT)    (FLIT & FIRST_HIGH_PRI)
+#define FIRST_GET_AGE(FLIT)         ((FLIT & FIRST_AGE(15))>>11)
+#define FIRST_GET_TIMEOUT(FLIT)     ((FLIT & FIRST_TIMEOUT(3))>>9)
+#define FIRST_GET_NEXT(FLIT)        ((FLIT & FIRST_PACKED(3))>>7)
+#define FIRST_GET_ROUTE(FLIT)       (FLIT & 0x7f)
+#define FIRST_GET_BCAST(FLIT)       (FLIT & 0x40)
+#define FIRST_GET_IS_INVALID(FLIT)  ((FLIT & 0x78) == 0x08)
+#define FIRST_GET_TYPE(FLIT)        ((FLIT & 0x30)>>4)
+#define PRF_GET_ROUTE(FLIT,N)       ((FLIT >> (N*4)) & 0x0F)
+#define PRF_GET_IS_MYLINK(ROUTE)    (ROUTE == PACKED_MYLINK)
+#define PRF_GET_IS_NORMAL(ROUTE)    (ROUTE & 0x8)
+#define PRF_GET_NORMAL_LINK(ROUTE)  (ROUTE & 0x7)
+#define PRF_MOVE_ON(INDEX,NEXT)     do { if (NEXT==3) {NEXT=0;INDEX++;} else {NEXT++; }} while (0);
+/* ---------------------------------------------------------------------- */
+int /* turn level needed or -1 if not possible */
+elan3_route_get_min_turn_level( ELAN_POSITION *pos, int nodeId)
+{
+    int l,range = 1;
+
+    for(l=pos->pos_levels-1;l>=0;l--)
+    {
+       range = range * pos->pos_arity[l];
+       
+       if ( ((pos->pos_nodeid - (pos->pos_nodeid % range)) <= nodeId ) 
+            && (nodeId <= (pos->pos_nodeid - (pos->pos_nodeid % range)+range -1))) 
+           return l;
+    }
+    return -1;
+}
+/* ---------------------------------------------------------------------- */
+int  
+elan3_route_check(ELAN3_CTXT *ctxt, E3_uint16 *flits, int destNodeId)
+{
+    ELAN3_NET_LOCATION lastLoc,currLoc;
+    int               err;
+    int               turnLevel;
+    int               goingDown;
+    int               lnk,index,next,val;
+    ELAN_POSITION    *pos = &ctxt->Position;
+   
+    /* is the dest possible */
+    if ( (destNodeId <0 ) || (destNodeId >= pos->pos_nodes))
+       return  (ELAN3_ROUTE_PROC_RANGE);
+
+    /* 
+     * walk the route, 
+     * - to see if we get there 
+     * - checking we dont turn around 
+     */
+    currLoc.netid = pos->pos_nodeid;         /* the elan */
+    currLoc.plane = 0;
+    currLoc.level = pos->pos_levels;
+
+    turnLevel = currLoc.level; /* track the how far the route goes in */
+    goingDown = 0;             /* once set we cant go up again ie only one change of direction */
+
+    /* move onto the network from the elan */
+    if ((err=elan3_route_follow_link(ctxt,&currLoc,4)) != ELAN3_ROUTE_SUCCESS) 
+    {
+       PRINTF0 (ctxt, DBG_VP, "elan3_route_check: initial elan3_route_follow_link failed\n");
+       return err;
+    }
+    /* do the first part of flit */
+    switch ( FIRST_GET_TYPE(flits[0]) ) 
+    {
+    case 0  /* sent */   : { lnk = (flits[0] & 0x7);                                 break; }    
+    case PACKED_MYLINK  : { lnk = pos->pos_nodeid % pos->pos_arity[pos->pos_levels-1];    break; }
+    case PACKED_ADAPTIVE : { lnk = 7; /* all routes are the same just check one */    break; }
+    default : 
+       PRINTF1 (ctxt, DBG_VP, "elan3_route_check: unexpected first flit (%d)\n",flits[0]);
+       return (ELAN3_ROUTE_INVALID); 
+    }
+    
+    /* move along this link and check new location */
+    memcpy(&lastLoc,&currLoc,sizeof(ELAN3_NET_LOCATION)); /* keep track of last loc */
+    if ((err=elan3_route_follow_link(ctxt,&currLoc,lnk)) != ELAN3_ROUTE_SUCCESS ) 
+    {
+       PRINTF0 (ctxt, DBG_VP, "elan3_route_check: elan3_route_follow_link failed\n");
+       return err;
+    }
+    if ((currLoc.level > pos->pos_levels) || (currLoc.level < 0 )) 
+    { 
+       PRINTF0 (ctxt, DBG_VP, "elan3_route_check: route leaves machine\n");
+       return (ELAN3_ROUTE_INVALID_LEVEL);
+    }
+    if ( lastLoc.level < currLoc.level ) 
+    {
+       turnLevel = lastLoc.level;
+       goingDown = 1;
+    }
+    else 
+    {
+       if (turnLevel > currLoc.level)
+           turnLevel =  currLoc.level;
+       if  (goingDown) 
+       {
+           PRINTF0 (ctxt, DBG_VP, "elan3_route_check: route ocilated\n");
+           return (ELAN3_ROUTE_OCILATES);
+       }
+    }   
+
+    /* loop on doing the remaining flits */
+    index = 1;
+    next  = FIRST_GET_NEXT(flits[0]);
+    val   = PRF_GET_ROUTE(flits[index],next);
+    while(val)
+    {
+       if (PRF_GET_IS_NORMAL(val) ) 
+           lnk = PRF_GET_NORMAL_LINK(val);
+       else
+       {
+         switch ( val ) 
+         {
+         case  PACKED_MYLINK : 
+         {
+             lnk = elan3_route_get_mylink(pos, &currLoc,&lastLoc);
+             break;
+         }
+         default : 
+             PRINTF1 (ctxt, DBG_VP, "elan3_route_check: unexpected packed flit (%d)\n",val);
+             return (ELAN3_ROUTE_INVALID);
+         }
+       }
+
+       /* move along this link and check new location */
+       memcpy(&lastLoc,&currLoc,sizeof(ELAN3_NET_LOCATION)); /* keep track of last loc */
+       if ((err=elan3_route_follow_link(ctxt,&currLoc,lnk)) != ELAN3_ROUTE_SUCCESS) 
+           return err;
+       
+       if ((currLoc.level > pos->pos_levels ) || ( currLoc.level < 0 ))
+       { 
+           PRINTF0 (ctxt, DBG_VP, "elan3_route_check: route leaves machine\n");
+           return (ELAN3_ROUTE_INVALID_LEVEL);
+       }
+
+       if ( lastLoc.level < currLoc.level ) 
+           goingDown = 1;
+       else 
+       {
+           if (turnLevel > currLoc.level)
+               turnLevel =  currLoc.level;
+           if  (goingDown) 
+           {
+               PRINTF0 (ctxt, DBG_VP, "elan3_route_check: route ocilated\n");
+               return (ELAN3_ROUTE_OCILATES);
+           }
+       }   
+
+       /* move to next part of flit */
+       PRF_MOVE_ON(index,next);
+       if ( index >= MAX_FLITS)
+       {
+           PRINTF0 (ctxt, DBG_VP, "elan3_route_check: route too long\n");
+           return (ELAN3_ROUTE_TOO_LONG);
+       }
+       /* extract the new value */
+       val = PRF_GET_ROUTE(flits[index],next);
+    }
+
+    /* have we got to where we want ? */
+    if ((currLoc.level != pos->pos_levels) || (currLoc.netid != destNodeId))
+    {
+       PRINTF2 (ctxt, DBG_VP, "elan3_route_check: goes to %d instead of %d\n",currLoc.netid , destNodeId );
+       return (ELAN3_ROUTE_WRONG_DEST);
+    }
+
+    /*
+     * there is the case of src == dest 
+     * getTurnLevel returns pos->pos_levels, and turnLevel is (pos->pos_levels -1) 
+     * then we assume they really want to  go onto the network.
+     * otherwise we check that the turn at the appriate level
+     */
+    if ( (pos->pos_nodeid != destNodeId) || ( turnLevel != (pos->pos_levels -1)) )
+    {
+       int lev;
+       if ((lev = elan3_route_get_min_turn_level(pos,destNodeId)) == -1) 
+       {
+           PRINTF0 (ctxt, DBG_VP, "elan3_route_check: cant calculate turn level\n");
+           return (ELAN3_ROUTE_INVALID); /* not sure this can happen here as checks above should protect me */
+       }
+       if (turnLevel != lev) 
+       {
+           PRINTF2 (ctxt, DBG_VP, "elan3_route_check: turn level should be %d but is %d \n", lev, turnLevel);
+           return (ELAN3_ROUTE_TURN_LEVEL);
+       }
+    }
+    return (ELAN3_ROUTE_SUCCESS);
+}
+/* ---------------------------------------------------------------------- */
+int
+elan3_route_broadcast_check(ELAN3_CTXT *ctxt , E3_uint16 *flits, int lowNode, int highNode ) 
+{
+    E3_uint16 flitsTmp[MAX_FLITS];
+    int       nflits,i;
+    
+    nflits = GenerateRoute (&ctxt->Position, flitsTmp, lowNode, highNode, DEFAULT_ROUTE_TIMEOUT, DEFAULT_ROUTE_PRIORITY);
+    
+    for(i=0;i<nflits;i++)
+       if ( flitsTmp[i] != flits[i] ) 
+       {
+           PRINTF3 (ctxt, DBG_VP, "elan3_route_broadcast_check:  flit[%d] %d (should be %d)\n",i,flits[i],flitsTmp[i]);
+           return (ELAN3_ROUTE_INVALID);   
+       }
+    
+    return (ELAN3_ROUTE_SUCCESS);
+}
+/* ---------------------------------------------------------------------- */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/route_table.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/route_table.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/route_table.c  2005-06-01 23:12:54.591440320 -0400
@@ -0,0 +1,560 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: route_table.c,v 1.23 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/route_table.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+
+static sdramaddr_t
+AllocateLargeRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int ctxnum, E3_uint64 *smallRoute)
+{
+    int                  bit = -1;
+    ELAN3_ROUTES  *rent;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&tbl->Lock, flags);
+    
+    for (rent = tbl->LargeRoutes; rent; rent = rent->Next)
+    {
+       if ((bit = bt_freebit (rent->Bitmap, NROUTES_PER_BLOCK)) != -1)
+           break;
+    }
+    
+    if (bit == -1)                                             /* No spare entries in large routes */
+    {                                                          /* so allocate a new page */
+       PRINTF0 (DBG_DEVICE, DBG_VP, "AllocateLargeRoute: allocate route entries\n");
+       
+       spin_unlock_irqrestore (&tbl->Lock, flags);
+
+       KMEM_ZALLOC(rent, ELAN3_ROUTES *, sizeof (ELAN3_ROUTES), TRUE);
+       
+       if (rent == (ELAN3_ROUTES *) NULL)
+           return ((sdramaddr_t) 0);
+       
+       rent->Routes = elan3_sdram_alloc (dev, PAGESIZE);
+       if (rent->Routes == (sdramaddr_t) 0)
+       {
+           KMEM_FREE (rent, sizeof (ELAN3_ROUTES));
+           return ((sdramaddr_t) 0);
+       }
+
+       spin_lock_irqsave (&tbl->Lock, flags);
+
+       /* Add to list of large routes */
+       rent->Next       = tbl->LargeRoutes;
+       tbl->LargeRoutes = rent;
+
+       /* and use entry 0 */
+       bit = 0;
+    }
+    
+    /* Set the bit in the bitmap to mark this route as allocated */
+    BT_SET (rent->Bitmap, bit);
+    
+    /* And generate the small route pointer and the pointer to the large routes */
+    (*smallRoute) = BIG_ROUTE_PTR(rent->Routes + (bit*NBYTES_PER_LARGE_ROUTE), ctxnum);
+
+    PRINTF4 (DBG_DEVICE, DBG_VP, "AllocateLargeRoute: rent %p using entry %d at %lx with route pointer %llx\n",
+            rent, bit, rent->Routes + (bit * NBYTES_PER_LARGE_ROUTE), (long long) (*smallRoute));
+
+    /* Invalidate the large route */
+    elan3_sdram_zeroq_sdram (dev, rent->Routes + (bit * NBYTES_PER_LARGE_ROUTE), NBYTES_PER_LARGE_ROUTE);
+
+    spin_unlock_irqrestore (&tbl->Lock, flags);
+
+    return (rent->Routes + (bit * NBYTES_PER_LARGE_ROUTE));
+}
+
+static void
+FreeLargeRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, E3_uint64 smallRoute)
+{
+    E3_Addr      addr = (E3_Addr) (smallRoute & ((1ULL << ROUTE_CTXT_SHIFT)-1));
+    ELAN3_ROUTES *rent;
+
+    PRINTF1 (DBG_DEVICE, DBG_VP, "FreeLargeRoute: free route %llx\n", (long long) smallRoute);
+
+    ASSERT (SPINLOCK_HELD (&tbl->Lock));
+
+    for (rent = tbl->LargeRoutes; rent; rent = rent->Next)
+    {
+       if (rent->Routes <= addr && (rent->Routes + ROUTE_BLOCK_SIZE) > addr)
+       {
+           int indx = (addr - rent->Routes)/NBYTES_PER_LARGE_ROUTE;
+           
+           PRINTF2 (DBG_DEVICE, DBG_VP, "FreeLargeRoute: rent=%p indx=%d\n", rent, indx);
+           
+           BT_CLEAR(rent->Bitmap, indx);
+           return;
+       }
+    }
+
+    panic ("elan: FreeLargeRoute - route not found in large route tables");
+}
+
+static void
+FreeLargeRoutes (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl)
+{
+    ELAN3_ROUTES *rent;
+
+    while ((rent = tbl->LargeRoutes) != NULL)
+    {
+       PRINTF1 (DBG_DEVICE, DBG_VP, "FreeLargeRoutes: free rent %p\n", rent);
+
+       tbl->LargeRoutes = rent->Next;
+
+       elan3_sdram_free (dev, rent->Routes, PAGESIZE);
+       
+       KMEM_FREE (rent, sizeof(ELAN3_ROUTES));
+    }
+}
+
+int
+GetRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int process, E3_uint16 *flits)
+{
+    E3_uint64  routeValue;
+    sdramaddr_t largeRouteOff;
+  
+    if (process < 0 || process >= tbl->Size)
+       return (EINVAL);
+
+    routeValue = elan3_sdram_readq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE);
+
+    if (routeValue & ROUTE_PTR)
+    {
+       largeRouteOff = (routeValue & ROUTE_PTR_MASK);
+       
+       routeValue = elan3_sdram_readq (dev, largeRouteOff + 0);
+       flits[0] = routeValue & 0xffff;
+       flits[1] = (routeValue >> 16)  & 0xffff;
+       flits[2] = (routeValue >> 32)  & 0xffff;
+       flits[3] = (routeValue >> 48)  & 0xffff;
+       
+       routeValue = elan3_sdram_readq (dev, largeRouteOff + 8);
+       flits[4] = routeValue & 0xffff;
+       flits[5] = (routeValue >> 16)  & 0xffff;
+       flits[6] = (routeValue >> 32)  & 0xffff;
+       flits[6] = (routeValue >> 48)  & 0xffff;
+    }
+    else
+    {
+       flits[0] = routeValue & 0xffff;
+       flits[1] = (routeValue >> 16)  & 0xffff;
+       flits[2] = (routeValue >> 32)  & 0xffff;
+    }
+
+    return (ESUCCESS);
+}
+
+ELAN3_ROUTE_TABLE *
+AllocateRouteTable (ELAN3_DEV *dev, int size)
+{
+    ELAN3_ROUTE_TABLE *tbl;
+
+    KMEM_ZALLOC (tbl, ELAN3_ROUTE_TABLE *, sizeof (ELAN3_ROUTE_TABLE), TRUE);
+
+    if (tbl == (ELAN3_ROUTE_TABLE *) NULL)
+       return (NULL);
+    
+    tbl->Size  = size;
+    tbl->Table = elan3_sdram_alloc (dev, size*NBYTES_PER_SMALL_ROUTE);
+
+    if (tbl->Table == 0)
+    {
+       KMEM_FREE (tbl, sizeof (ELAN3_ROUTE_TABLE));
+       return (NULL);
+    }
+    spin_lock_init (&tbl->Lock);
+
+    /* zero the route table */
+    elan3_sdram_zeroq_sdram (dev, tbl->Table, size*NBYTES_PER_SMALL_ROUTE);
+
+    return (tbl);
+}
+
+void
+FreeRouteTable (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl)
+{
+    elan3_sdram_free (dev, tbl->Table, tbl->Size*NBYTES_PER_SMALL_ROUTE);
+
+    FreeLargeRoutes (dev, tbl);
+
+    spin_lock_destroy (&tbl->Lock);
+
+    KMEM_FREE (tbl, sizeof (ELAN3_ROUTE_TABLE));
+}
+
+int
+LoadRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int process, int ctxnum, int nflits, E3_uint16 *flits)
+{
+    E3_uint64    routeValue;
+    E3_uint64    largeRouteValue;
+    sdramaddr_t   largeRouteOff;
+    unsigned long flags;
+
+    if (process < 0 || process >= tbl->Size)
+       return (EINVAL);
+
+    PRINTF3 (DBG_DEVICE, DBG_VP, "LoadRoute: table %lx process %d ctxnum %x\n", tbl->Table ,process, ctxnum);
+
+    if (nflits < 4)
+    {
+       spin_lock_irqsave (&tbl->Lock, flags);
+
+       /* See if we're replacing a "large" route */
+       routeValue = elan3_sdram_readq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE);
+       if (routeValue & ROUTE_PTR)
+           FreeLargeRoute (dev, tbl, routeValue);
+
+       routeValue = SMALL_ROUTE(flits, ctxnum);
+
+       if ( routeValue &  ROUTE_PTR)
+           PRINTF0 (DBG_DEVICE, DBG_VP, "SHOULD BE  A SMALL ROUTE !!!!!!!\n");
+
+       PRINTF2 (DBG_DEVICE, DBG_VP, "LoadRoute: loading small route %d  %llx\n", process, (long long) routeValue);
+       elan3_sdram_writeq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE, routeValue);
+    }
+    else
+    {
+       E3_uint64 value0 = BIG_ROUTE0(flits);
+       E3_uint64 value1 = BIG_ROUTE1(flits);
+
+       if ((largeRouteOff = AllocateLargeRoute (dev, tbl, ctxnum, &largeRouteValue)) == (sdramaddr_t) 0)
+           return (ENOMEM);
+
+       spin_lock_irqsave (&tbl->Lock, flags);
+           
+       routeValue = elan3_sdram_readq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE);
+
+       if ((routeValue & ROUTE_PTR) == 0)
+           elan3_sdram_writeq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE, largeRouteValue);
+       else
+       {
+           FreeLargeRoute (dev, tbl, largeRouteValue);
+
+           largeRouteOff   = (routeValue & ROUTE_PTR_MASK);
+       }
+
+       PRINTF3 (DBG_DEVICE, DBG_VP, "LoadRoute: loading large route %d - %llx %llx\n", process, 
+                (long long) value0, (long long) value1);
+
+       elan3_sdram_writeq (dev, largeRouteOff + 0, value0);
+       elan3_sdram_writeq (dev, largeRouteOff + 8, value1);
+    }
+
+    spin_unlock_irqrestore (&tbl->Lock, flags);
+    return (ESUCCESS);
+}
+void
+InvalidateRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int process)
+{
+    E3_uint64 routeValue;
+    unsigned long flags;
+
+    if (process < 0 || process >= tbl->Size)
+       return;
+
+    spin_lock_irqsave (&tbl->Lock, flags);
+
+    /* unset ROUTE_VALID
+     * does not matter if its short or long, will check when we re-use it
+     */
+    routeValue = elan3_sdram_readq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE);
+    elan3_sdram_writeq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE, (routeValue & (~ROUTE_VALID)));
+
+    spin_unlock_irqrestore (&tbl->Lock, flags);
+}
+void
+ValidateRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int process)
+{
+    E3_uint64 routeValue;
+    unsigned long flags;
+
+    if (process < 0 || process >= tbl->Size)
+       return;
+
+    PRINTF2 (DBG_DEVICE, DBG_VP, "ValidateRoute: table %ld process %d  \n", tbl->Table ,process);
+
+    spin_lock_irqsave (&tbl->Lock, flags);
+
+    /* set ROUTE_VALID
+     */
+    routeValue = elan3_sdram_readq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE);
+    elan3_sdram_writeq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE, (routeValue | ROUTE_VALID));
+
+    spin_unlock_irqrestore (&tbl->Lock, flags);
+}
+void
+ClearRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int process)
+{
+    E3_uint64 routeValue;
+    unsigned long flags;
+
+    if (process < 0 || process >= tbl->Size)
+       return;
+
+    spin_lock_irqsave (&tbl->Lock, flags);
+
+    PRINTF2 (DBG_DEVICE, DBG_VP, "ClearRoute: table %ld process %d  \n", tbl->Table ,process);
+
+    routeValue = elan3_sdram_readq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE);
+
+    elan3_sdram_writeq (dev, tbl->Table + process * NBYTES_PER_SMALL_ROUTE, 0);
+
+    if (routeValue & ROUTE_PTR)
+       FreeLargeRoute (dev, tbl, routeValue);
+
+    spin_unlock_irqrestore (&tbl->Lock, flags);
+}
+
+static int
+ElanIdEqual (ELAN_POSITION *pos, int level, int ida, int idb)
+{
+    int l;
+
+    for (l = pos->pos_levels-1; l >= level; l--)
+    {
+       ida /= pos->pos_arity[l];
+       idb /= pos->pos_arity[l];
+    }
+       
+    return (ida == idb);
+}
+
+static int
+RouteDown (ELAN_POSITION *pos, int level, int elanid)
+{
+    int l;
+
+    for (l = (pos->pos_levels - 1); level < pos->pos_levels - 1; level++, l--)
+    {  
+       if (  pos->pos_arity[l] )
+           elanid /= pos->pos_arity[l];    
+    }
+    elanid %= pos->pos_arity[l];
+
+    return elanid;
+}
+
+static int
+InitPackedAndFlits (u_char *packed, E3_uint16 *flits)
+{
+    int rb = 0;
+
+    bzero ((caddr_t) packed, MAX_PACKED+4);
+    bzero ((caddr_t) flits, MAX_FLITS * sizeof (E3_uint16));
+
+    /* Initialise 4 bytes of packed, so that the "padding" */
+    /* NEVER terminates with 00, as this is recognised as */
+    /* as CRC flit */
+    packed[rb++] = 0xF;
+    packed[rb++] = 0xF;
+    packed[rb++] = 0xF;
+    packed[rb++] = 0xF;
+
+    return (rb);
+}
+
+static int
+PackThemRoutesUp (E3_uint16 *flits, u_char *packed, int rb, int timeout, int highPri)
+{
+    int i, nflits;
+
+    flits[0] |= FIRST_TIMEOUT(timeout);
+    if (highPri)
+       flits[0] |= FIRST_HIGH_PRI;
+
+    /* round up the number of route bytes to flits */
+    /* and subtract the 4 extra we've padded out with */
+    nflits = (rb-1)/4;
+    
+    for (i = nflits; i > 0; i--)
+    {
+       flits[i] = (packed[rb-1] << 12 |
+                   packed[rb-2] << 8  |
+                   packed[rb-3] << 4  |
+                   packed[rb-4] << 0);
+       rb -= 4;
+    }
+    
+    /* Now set the position of the first packed route  */
+    /* byte in the 2nd 16 bit flit, taking account of the */
+    /* 4 byte padding */
+    flits[0] |= FIRST_PACKED (4-rb);
+    
+    return (nflits+1);
+}
+
+int
+GenerateRoute (ELAN_POSITION *pos, E3_uint16 *flits, int lowid, int highid, int timeout, int highPri)
+{
+    int     broadcast  = (lowid != highid);
+    int     rb         = 0;
+    int            first      = 1;
+    int     noRandom   = 0;
+    int     level;
+    u_char  packed[MAX_PACKED+4];
+    int     numDownLinks;
+
+    rb = InitPackedAndFlits (packed, flits);
+
+    for (level = pos->pos_levels-1;                            /* Move up out of the elan */
+        level > 0 && ! (ElanIdEqual (pos, level, pos->pos_nodeid, lowid) &&
+                        ElanIdEqual (pos, level, pos->pos_nodeid, highid)); level--)
+    {
+       noRandom |= pos->pos_random_disabled & (1 << (pos->pos_levels-1-level));
+    }
+    
+    for (level = pos->pos_levels-1;                            /* Move up out of the elan */
+        level > 0 && ! (ElanIdEqual (pos, level, pos->pos_nodeid, lowid) &&
+                        ElanIdEqual (pos, level, pos->pos_nodeid, highid)); level--)
+    {
+       numDownLinks = pos->pos_arity [level];
+       if (first)
+       {
+           if (broadcast || noRandom)
+               flits[0] = FIRST_BCAST_TREE;
+           else
+           {
+               if (numDownLinks == 4) 
+                   flits[0] = FIRST_ADAPTIVE;
+               else
+                   flits[0] = FIRST_ROUTE( numDownLinks + ( lowid % (8-numDownLinks) ));
+           }
+           first = 0;
+       }
+       else
+       {
+           if (broadcast || noRandom)
+               packed[rb++] = PACKED_BCAST_TREE;
+           else 
+           {
+               if (numDownLinks == 4) 
+                   packed[rb++] = PACKED_ADAPTIVE;
+               else
+                   packed[rb++] = PACKED_ROUTE( numDownLinks + ( lowid % (8-numDownLinks) ));
+           }               
+       }
+    }
+    
+    while (level < pos->pos_levels)
+    {
+       int lowRoute  = RouteDown (pos, level, lowid);
+       int highRoute = RouteDown (pos, level, highid);
+
+       if (first)
+       {
+           if (broadcast)
+               flits[0] = FIRST_BCAST(highRoute, lowRoute);
+           else
+               flits[0] = FIRST_ROUTE(lowRoute);
+
+           first = 0;
+       }
+       else
+       {
+           if (broadcast)
+           {
+               packed[rb++] = PACKED_BCAST0(highRoute, lowRoute);
+               packed[rb++] = PACKED_BCAST1(highRoute, lowRoute);
+           }
+           else
+               packed[rb++] = PACKED_ROUTE(lowRoute);
+       }
+       
+       level++;
+    }
+
+#ifdef ELITE_REVA_SUPPORTED
+    if (broadcast && (pos->pos_levels == 3))
+    {
+      packed[rb++] = PACKED_BCAST0(0, 0);
+      packed[rb++] = PACKED_BCAST1(0, 0);
+    }
+#endif
+
+    return (PackThemRoutesUp (flits, packed, rb, timeout, highPri));
+}
+
+int
+GenerateCheckRoute (ELAN_POSITION *pos, E3_uint16 *flits, int level, int adaptive)
+{
+    int     notfirst = 0;
+    int     l, rb;
+    u_char  packed[MAX_PACKED+4];
+
+    rb = InitPackedAndFlits (packed, flits);
+
+    for (l = pos->pos_levels-1; l > level; l--)
+       if (! notfirst++)
+           flits[0] = adaptive ? FIRST_ADAPTIVE : FIRST_BCAST_TREE;
+       else
+           packed[rb++] = adaptive ? PACKED_ADAPTIVE : PACKED_BCAST_TREE;
+
+    if (! notfirst++ ) 
+       flits[0] = FIRST_MYLINK;
+    else
+       packed[rb++] = PACKED_MYLINK;
+
+    for (l++ /* consume mylink */; l < pos->pos_levels; l++)
+       if (! notfirst++)
+           flits[0] = FIRST_ROUTE (RouteDown (pos, l, pos->pos_nodeid));
+       else
+           packed[rb++] = PACKED_ROUTE (RouteDown (pos, l, pos->pos_nodeid));
+
+
+    return (PackThemRoutesUp (flits, packed, rb, DEFAULT_ROUTE_TIMEOUT, HIGH_ROUTE_PRIORITY));
+}
+
+
+/*
+ * In this case "level" is the number of levels counted from the bottom.
+ */
+int
+GenerateProbeRoute (E3_uint16 *flits, int nodeid, int level, int *linkup, int *linkdown, int adaptive )
+{
+    int            first = 1;
+    int     i, rb;
+    u_char  packed[MAX_PACKED+4];
+
+    rb = InitPackedAndFlits (packed, flits);
+
+    /* Generate "up" routes */
+    for (i = 0; i < level; i++)
+    {
+       if (first)
+           flits[0] = linkup ? FIRST_ROUTE(linkup[i]) : adaptive ? FIRST_ADAPTIVE : FIRST_BCAST_TREE;
+       else
+           packed[rb++] = linkup ? PACKED_ROUTE(linkup[i]) : adaptive ? PACKED_ADAPTIVE : PACKED_BCAST_TREE;
+       first = 0;
+    }
+
+    /* Generate a "to-me" route down */
+    if (first)
+       flits[0] = FIRST_MYLINK;
+    else
+       packed[rb++] = PACKED_MYLINK;
+
+    for (i = level-1; i >= 0; i--)
+       packed[rb++] =  PACKED_ROUTE(linkdown[i]);
+
+    return (PackThemRoutesUp (flits, packed, rb, DEFAULT_ROUTE_TIMEOUT, HIGH_ROUTE_PRIORITY));
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/sdram.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/sdram.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/sdram.c        2005-06-01 23:12:54.593440016 -0400
@@ -0,0 +1,807 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: sdram.c,v 1.17 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/sdram.c,v $*/
+
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elandebug.h>
+
+/* sdram access functions */
+#define sdram_off_to_bank(dev,off)     (&dev->SdramBanks[(off) >> ELAN3_SDRAM_BANK_SHIFT])
+#define sdram_off_to_offset(dev,off)   ((off) & (ELAN3_SDRAM_BANK_SIZE-1))
+#define sdram_off_to_bit(dev,indx,off) (sdram_off_to_offset(dev,off) >> (SDRAM_MIN_BLOCK_SHIFT+(indx)))
+
+#define sdram_off_to_mapping(dev,off)  (sdram_off_to_bank(dev,off)->Mapping + sdram_off_to_offset(dev,off))
+    
+unsigned char
+elan3_sdram_readb (ELAN3_DEV *dev, sdramaddr_t off)
+{
+    return (readb ((unsigned char *) sdram_off_to_mapping(dev, off)));
+}
+
+unsigned short
+elan3_sdram_readw (ELAN3_DEV *dev, sdramaddr_t off)
+{
+    return (readw ((unsigned short *) sdram_off_to_mapping(dev, off)));
+}
+
+unsigned int
+elan3_sdram_readl (ELAN3_DEV *dev, sdramaddr_t off)
+{
+    return (readl ((unsigned int *) sdram_off_to_mapping(dev, off)));
+}
+
+unsigned long long
+elan3_sdram_readq (ELAN3_DEV *dev, sdramaddr_t off)
+{
+    return (readq ((unsigned long long *) sdram_off_to_mapping(dev, off)));
+}
+
+void
+elan3_sdram_writeb (ELAN3_DEV *dev, sdramaddr_t off, unsigned char val)
+{
+    writeb (val, (unsigned char *) sdram_off_to_mapping(dev, off));
+    wmb();
+}
+
+void
+elan3_sdram_writew (ELAN3_DEV *dev, sdramaddr_t off, unsigned short val)
+{
+    writew (val, (unsigned short *) sdram_off_to_mapping(dev, off));
+    wmb();
+}
+
+void
+elan3_sdram_writel (ELAN3_DEV *dev, sdramaddr_t off, unsigned int val)
+{
+    writel (val, (unsigned int *) sdram_off_to_mapping(dev, off));
+    wmb();
+}
+
+void
+elan3_sdram_writeq (ELAN3_DEV *dev, sdramaddr_t off, unsigned long long val)
+{
+    writeq (val, (unsigned long long *) sdram_off_to_mapping(dev, off));
+    wmb();
+}
+
+void
+elan3_sdram_copyb_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+    bcopy ((void *)sdram_off_to_mapping(dev, from), to, nbytes);
+}
+
+void
+elan3_sdram_copyw_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+#ifdef __LITTLE_ENDIAN__
+    bcopy ((void *)sdram_off_to_mapping(dev, from), to, nbytes);
+#else
+#error incorrect for big endian
+#endif
+}
+
+void
+elan3_sdram_copyl_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+#ifdef __LITTLE_ENDIAN__
+    bcopy ((void *)sdram_off_to_mapping(dev, from), to, nbytes);
+#else
+#error incorrect for big endian
+#endif
+}
+
+void
+elan3_sdram_copyq_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+#ifdef __LITTLE_ENDIAN__
+    bcopy ((void *)sdram_off_to_mapping(dev, from), to, nbytes);
+#else
+#error incorrect for big endian
+#endif
+}
+
+#define E3_WRITEBUFFER_SIZE            16
+#define E3_WRITEBUFFER_OFFSET(x)       (((unsigned long) x) & (E3_WRITEBUFFER_SIZE-1))
+#define E3_WRITEBUFFER_BASE(x)         (((unsigned long) x) & ~((unsigned long) (E3_WRITEBUFFER_SIZE-1)))
+
+void
+elan3_sdram_copyb_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    virtaddr_t slim  = (virtaddr_t) from + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint8_t)) + sizeof (uint8_t);
+    int        i;
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       for (i = 0; i < nbytes/sizeof(uint8_t); i++)
+           writeb (((uint8_t *) from)[i], &((uint8_t *) dbase)[i]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           slim -= ntop;
+           dlim -= ntop;
+           
+           for (i = 0; i < ntop/sizeof(uint8_t); i++)
+               writeb (((uint8_t *) slim)[i], &((uint8_t *) dlim)[i]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+           slim -= E3_WRITEBUFFER_SIZE;
+
+           for (i = 0; i < E3_WRITEBUFFER_SIZE/sizeof (uint8_t); i++)
+               writeb (((uint8_t *) slim)[i], &((uint8_t *) dlim)[i]);
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           for (i = 0; i < nbase/sizeof(uint8_t); i++)
+               writeb (((uint8_t *) from)[i], &((uint8_t *) dbase)[i]);
+           wmb();
+       }
+    }
+}
+
+void
+elan3_sdram_zerob_sdram (ELAN3_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint8_t)) + sizeof (uint8_t);
+    int        i;
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       for (i = 0; i < nbytes/sizeof(uint8_t); i++)
+           writeb (0, &((uint8_t *) dbase)[i]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           dlim -= ntop;
+           
+           for (i = 0; i < ntop/sizeof(uint8_t); i++)
+               writeb (0, &((uint8_t *) dlim)[i]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+
+           writeq (0, &((uint64_t *) dlim)[0]);
+           writeq (0, &((uint64_t *) dlim)[1]);
+
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           for (i = 0; i < nbase/sizeof(uint8_t); i++)
+               writeb (0, &((uint8_t *) dbase)[i]);
+           wmb();
+       }
+    }
+}
+
+void
+elan3_sdram_copyw_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    virtaddr_t slim  = (virtaddr_t) from + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint16_t)) + sizeof (uint16_t);
+    int        i;
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       for (i = 0; i < nbytes/sizeof(uint16_t); i++)
+           writew (((uint16_t *) from)[i], &((uint16_t *) dbase)[i]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           slim -= ntop;
+           dlim -= ntop;
+
+           for (i = 0; i < ntop/sizeof(uint16_t); i++)
+               writew (((uint16_t *) slim)[i], &((uint16_t *) dlim)[i]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+           slim -= E3_WRITEBUFFER_SIZE;
+
+           writew (((uint16_t *) slim)[0], &((uint16_t *) dlim)[0]);
+           writew (((uint16_t *) slim)[1], &((uint16_t *) dlim)[1]);
+           writew (((uint16_t *) slim)[2], &((uint16_t *) dlim)[2]);
+           writew (((uint16_t *) slim)[3], &((uint16_t *) dlim)[3]);
+           writew (((uint16_t *) slim)[4], &((uint16_t *) dlim)[4]);
+           writew (((uint16_t *) slim)[5], &((uint16_t *) dlim)[5]);
+           writew (((uint16_t *) slim)[6], &((uint16_t *) dlim)[6]);
+           writew (((uint16_t *) slim)[7], &((uint16_t *) dlim)[7]);
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           for (i = 0; i < nbase/sizeof(uint16_t); i++)
+               writew (((uint16_t *) from)[i], &((uint16_t *) dbase)[i]);
+           wmb();
+       }
+    }
+}
+
+void
+elan3_sdram_zerow_sdram (ELAN3_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint16_t)) + sizeof (uint16_t);
+    int        i;
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       for (i = 0; i < nbytes/sizeof(uint16_t); i++)
+           writew (0, &((uint16_t *) dbase)[i]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           dlim -= ntop;
+           
+           for (i = 0; i < ntop/sizeof(uint16_t); i++)
+               writew (0, &((uint16_t *) dlim)[i]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+
+           writeq (0, &((uint64_t *) dlim)[0]);
+           writeq (0, &((uint64_t *) dlim)[1]);
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           for (i = 0; i < nbase/sizeof(uint16_t); i++)
+               writew (0, &((uint16_t *) dbase)[i]);
+           wmb();
+       }
+    }
+}
+
+void
+elan3_sdram_copyl_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    virtaddr_t slim  = (virtaddr_t) from + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint32_t)) + sizeof (uint32_t);
+    int        i;
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       for (i = 0; i < nbytes/sizeof(uint32_t); i++)
+           writel (((uint32_t *) from)[i], &((uint32_t *) dbase)[i]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           slim -= ntop;
+           dlim -= ntop;
+
+           for (i = 0; i < ntop/sizeof(uint32_t); i++)
+               writel (((uint32_t *) slim)[i], &((uint32_t *) dlim)[i]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+           slim -= E3_WRITEBUFFER_SIZE;
+
+           writel (((uint32_t *) slim)[0], &((uint32_t *) dlim)[0]);
+           writel (((uint32_t *) slim)[1], &((uint32_t *) dlim)[1]);
+           writel (((uint32_t *) slim)[2], &((uint32_t *) dlim)[2]);
+           writel (((uint32_t *) slim)[3], &((uint32_t *) dlim)[3]);
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           for (i = 0; i < nbase/sizeof(uint32_t); i++)
+               writel (((uint32_t *) from)[i], &((uint32_t *) dbase)[i]);
+           wmb();
+       }
+    }
+}
+
+void
+elan3_sdram_zerol_sdram (ELAN3_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint32_t)) + sizeof (uint32_t);
+    int        i;
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       for (i = 0; i < nbytes/sizeof(uint32_t); i++)
+           writel (0, &((uint32_t *) dbase)[i]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           dlim -= ntop;
+
+           for (i = 0; i < ntop/sizeof(uint32_t); i++)
+               writel (0, &((uint32_t *) dlim)[i]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+
+           writeq (0, &((uint64_t *) dlim)[0]);
+           writeq (0, &((uint64_t *) dlim)[1]);
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           for (i = 0; i < nbase/sizeof(uint32_t); i++)
+               writel (0, &((uint32_t *) dbase)[i]);
+           wmb();
+       }
+    }
+}
+
+void
+elan3_sdram_copyq_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    virtaddr_t slim  = (virtaddr_t) from + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint64_t)) + sizeof (uint64_t);
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       writeq (((uint64_t *) from)[0], &((uint64_t *) dbase)[0]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           slim -= ntop;
+           dlim -= ntop;
+
+           writeq (((uint64_t *) slim)[0], &((uint64_t *) dlim)[0]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+           slim -= E3_WRITEBUFFER_SIZE;
+
+           writeq (((uint64_t *) slim)[0], &((uint64_t *) dlim)[0]);
+           writeq (((uint64_t *) slim)[1], &((uint64_t *) dlim)[1]);
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           writeq (((uint64_t *) from)[0], &((uint64_t *) dbase)[0]);
+           wmb();
+       }
+    }
+}
+
+void
+elan3_sdram_zeroq_sdram (ELAN3_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    virtaddr_t dbase = (virtaddr_t) sdram_off_to_mapping (dev, to);
+    virtaddr_t dlim  = (virtaddr_t) dbase + nbytes;
+    unsigned   nbase = E3_WRITEBUFFER_SIZE - E3_WRITEBUFFER_OFFSET (dbase);
+    unsigned   ntop  = E3_WRITEBUFFER_OFFSET (dlim - sizeof (uint64_t)) + sizeof (uint64_t);
+
+    if (E3_WRITEBUFFER_BASE(dbase) == E3_WRITEBUFFER_BASE(dlim))
+    {
+       writeq (0, &((uint64_t *) dbase)[0]);
+       wmb();
+    }
+    else
+    {
+       if (ntop < E3_WRITEBUFFER_SIZE)
+       {
+           dlim -= ntop;
+
+           writeq (0, &((uint64_t *) dlim)[0]);
+           wmb();
+       }
+       
+       while (dlim >= (dbase + E3_WRITEBUFFER_SIZE))
+       {
+           dlim -= E3_WRITEBUFFER_SIZE;
+           
+           writeq (0, &((uint64_t *) dlim)[0]);
+           writeq (0, &((uint64_t *) dlim)[1]);
+           wmb();
+       }
+       
+       if (nbase < E3_WRITEBUFFER_SIZE)
+       {
+           writeq (0, &((uint64_t *) dbase)[0]);
+           wmb();
+       }
+    }
+}
+
+physaddr_t
+elan3_sdram_to_phys (ELAN3_DEV *dev, sdramaddr_t off)
+{
+#if defined(DIGITAL_UNIX)
+    return (KSEG_TO_PHYS (sdram_off_to_mapping (dev, off)));
+#elif defined(LINUX)
+    return (kmem_to_phys ((void *) sdram_off_to_mapping (dev, off)));
+#endif    
+}
+
+/* sdram buddy allocator */
+#define read_next(dev, block)          elan3_sdram_readl(dev, block + 0)
+#define read_prev(dev, block)          elan3_sdram_readl(dev, block + 4)
+#define write_next(dev, block, val)    (elan3_sdram_writel(dev, block + 0, val), val)
+#define write_prev(dev, block, val)    (elan3_sdram_writel(dev, block + 4, val), val)
+
+#define freelist_insert(dev,idx,block)\
+do {\
+    sdramaddr_t next = dev->SdramFreeLists[(idx)];\
+\
+    /*\
+     * block->prev = NULL;\
+     * block->next = next;\
+     * if (next != NULL)\
+     *    next->prev = block;\
+     * freelist = block;\
+     */\
+    write_prev (dev, block, (sdramaddr_t) 0);\
+    write_next (dev, block, next);\
+    if (next != (sdramaddr_t) 0)\
+       write_prev (dev, next, block);\
+    dev->SdramFreeLists[idx] = block;\
+\
+    dev->SdramFreeCounts[idx]++;\
+    dev->Stats.SdramBytesFree += (SDRAM_MIN_BLOCK_SIZE << idx);\
+} while (0)
+
+#define freelist_remove(dev,idx,block)\
+do {\
+    /*\
+     * if (block->prev)\
+     *     block->prev->next = block->next;\
+     * else\
+     *     dev->SdramFreeLists[idx] = block->next;\
+     * if (block->next)\
+     *     block->next->prev = block->prev;\
+     */\
+    sdramaddr_t blocknext = read_next (dev, block);\
+    sdramaddr_t blockprev = read_prev (dev, block);\
+\
+    if (blockprev)\
+       write_next (dev, blockprev, blocknext);\
+    else\
+       dev->SdramFreeLists[idx] = blocknext;\
+    if (blocknext)\
+       write_prev (dev, blocknext, blockprev);\
+\
+    dev->SdramFreeCounts[idx]--;\
+    dev->Stats.SdramBytesFree -= (SDRAM_MIN_BLOCK_SIZE << idx);\
+} while (0)
+
+#define freelist_removehead(dev,idx,block)\
+do {\
+    sdramaddr_t blocknext = read_next (dev, block);\
+\
+    if ((dev->SdramFreeLists[idx] = blocknext) != 0)\
+       write_prev (dev, blocknext, 0);\
+\
+    dev->SdramFreeCounts[idx]--;\
+    dev->Stats.SdramBytesFree -= (SDRAM_MIN_BLOCK_SIZE << idx);\
+} while (0)
+
+#if defined(DEBUG)
+static int
+display_blocks (ELAN3_DEV *dev, int indx, char *string)
+{
+    sdramaddr_t block;
+    int nbytes = 0;
+
+    printk ("%s - indx %d\n", string, indx);
+    for (block = dev->SdramFreeLists[indx]; block != (sdramaddr_t) 0; block = read_next (dev, block))
+    {
+       printk ("  %lx", block);
+       nbytes += (SDRAM_MIN_BLOCK_SIZE << indx);
+    }
+    printk ("\n");
+
+    return (nbytes);
+}
+
+
+void
+elan3_sdram_display (ELAN3_DEV *dev, char *string)
+{
+    int indx;
+    int nbytes = 0;
+
+    printk ("elan3_sdram_display: dev=%p\n", dev);
+    for (indx = 0; indx < SDRAM_NUM_FREE_LISTS; indx++)
+       if (dev->SdramFreeLists[indx] != (sdramaddr_t) 0)
+           nbytes += display_blocks (dev, indx, string);
+    printk ("\n%d bytes free\n", nbytes);
+}
+
+void
+elan3_sdram_verify (ELAN3_DEV *dev)
+{
+    int indx, size, nbits, i, b;
+    sdramaddr_t block;
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; indx < SDRAM_NUM_FREE_LISTS; indx++, size <<= 1)
+    {
+       unsigned count = 0;
+
+       for (block = dev->SdramFreeLists[indx]; block; block = read_next (dev, block), count++)
+       {
+           ELAN3_SDRAM_BANK *bank = sdram_off_to_bank (dev, block);
+           unsigned         off  = sdram_off_to_offset (dev, block);
+           int              bit  = sdram_off_to_bit (dev, indx, block);
+
+           if ((block & (size-1)) != 0)
+               printk ("elan3_sdram_verify: block=%lx indx=%x - not aligned\n", block, indx);
+           
+           if (bank == NULL || off > bank->Size)
+               printk ("elan3_sdram_verify: block=%lx indx=%x - outside bank\n", block, indx);
+           else if (BT_TEST (bank->Bitmaps[indx], bit) == 0)
+               printk ("elan3_sdram_verify: block=%lx indx=%x - bit not set\n", block, indx);
+           else
+           {
+               for (i = indx-1, nbits = 2; i >= 0; i--, nbits <<= 1)
+               {
+                   bit = sdram_off_to_bit (dev, i, block);
+
+                   for (b = 0; b < nbits; b++)
+                       if (BT_TEST(bank->Bitmaps[i], bit + b))
+                           printk ("elan3_sdram_verify: block=%lx indx=%x - also free i=%d bit=%x\n", block, indx, i, bit+b);
+               }
+           }
+       }
+
+       if (dev->SdramFreeCounts[indx] != count)
+           printk ("elan3_sdram_verify: indx=%x expected %d got %d\n", indx, dev->SdramFreeCounts[indx], count);
+    }
+}
+
+#endif /* defined(DEBUG) */
+
+static void
+free_block (ELAN3_DEV *dev, sdramaddr_t block, int indx)
+{
+    ELAN3_SDRAM_BANK *bank = sdram_off_to_bank (dev, block);
+    unsigned        bit  = sdram_off_to_bit(dev, indx, block);
+    unsigned         size = SDRAM_MIN_BLOCK_SIZE << indx;
+
+    PRINTF3 (DBG_DEVICE, DBG_SDRAM, "free_block: block=%lx indx=%d bit=%x\n", block, indx, bit);
+
+    ASSERT ((block & (size-1)) == 0);
+    ASSERT (BT_TEST (bank->Bitmaps[indx], bit) == 0);
+    
+    while (BT_TEST (bank->Bitmaps[indx], bit ^ 1))
+    {
+       sdramaddr_t buddy = block ^ size;
+       
+       PRINTF3 (DBG_DEVICE, DBG_SDRAM, "free_block: merge block=%lx buddy=%lx indx=%d\n", block, buddy, indx);
+
+       BT_CLEAR (bank->Bitmaps[indx], bit ^ 1);
+
+       freelist_remove (dev, indx, buddy);
+       
+       block = (block < buddy) ? block : buddy;
+       indx++;
+       size <<= 1;
+       bit >>= 1;
+    }
+
+    PRINTF3 (DBG_DEVICE, DBG_SDRAM, "free_block: free block=%lx indx=%d bit=%x\n", block, indx, bit);
+
+    freelist_insert (dev, indx, block);
+
+    BT_SET (bank->Bitmaps[indx], bit);
+}
+
+void
+elan3_sdram_init (ELAN3_DEV *dev)
+{
+    int indx;
+
+    spin_lock_init (&dev->SdramLock);
+
+    for (indx = 0; indx < SDRAM_NUM_FREE_LISTS; indx++)
+    {
+       dev->SdramFreeLists[indx]  = (sdramaddr_t) 0;
+       dev->SdramFreeCounts[indx] = 0;
+    }
+}
+
+void
+elan3_sdram_fini (ELAN3_DEV *dev)
+{
+    spin_lock_destroy (&dev->SdramLock);
+}
+
+void
+elan3_sdram_add (ELAN3_DEV *dev, sdramaddr_t base, sdramaddr_t top)
+{
+    register int indx;
+    register unsigned long size;
+
+    /* align to the minimum block size */
+    base = (base + SDRAM_MIN_BLOCK_SIZE - 1) & ~((sdramaddr_t) SDRAM_MIN_BLOCK_SIZE-1);
+    top &= ~((sdramaddr_t) SDRAM_MIN_BLOCK_SIZE-1);
+
+    /* don't allow 0 as a valid "base" */
+    if (base == 0)
+       base = E3_CACHE_SIZE;
+
+    /* carve the bottom to the biggest boundary */
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; indx < SDRAM_NUM_FREE_LISTS; indx++, size <<= 1)
+    {
+       if ((base & size) == 0)
+           continue;
+
+       if ((base + size) > top)
+           break;
+
+       free_block (dev, base, indx);
+       
+       base += size;
+    }
+
+    /* carve the top down to the biggest boundary */
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; indx < SDRAM_NUM_FREE_LISTS; indx++, size <<= 1)
+    {
+       if ((top & size) == 0)
+           continue;
+
+       if ((top - size) < base)
+           break;
+
+       free_block (dev, (top - size), indx);
+       
+       top -= size;
+    }
+
+    /* now free of the space in between */
+    while (base < top)
+    {
+       free_block (dev, base, (SDRAM_NUM_FREE_LISTS-1));
+
+       base += SDRAM_MAX_BLOCK_SIZE;
+    }
+}
+
+sdramaddr_t
+elan3_sdram_alloc (ELAN3_DEV *dev, int nbytes)
+{
+    sdramaddr_t block;
+    register int i, indx;
+    unsigned long size;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->SdramLock, flags);
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size < nbytes; indx++, size <<= 1)
+       ;
+
+    PRINTF2 (DBG_DEVICE, DBG_SDRAM, "elan3_sdram_alloc: nbytes=%d indx=%d\n", nbytes, indx);
+
+    /* find the smallest block which is big enough for this allocation */
+    for (i = indx; i < SDRAM_NUM_FREE_LISTS; i++, size <<= 1)
+       if (dev->SdramFreeLists[i])
+           break;
+    
+    if (i == SDRAM_NUM_FREE_LISTS)
+    {
+       spin_unlock_irqrestore (&dev->SdramLock, flags);
+       return ((sdramaddr_t) 0);
+    }
+    
+    PRINTF2 (DBG_DEVICE, DBG_SDRAM, "elan3_sdram_alloc: use block=%lx indx=%d\n", dev->SdramFreeLists[i], i);
+
+    /* remove the block from the free list */
+    freelist_removehead (dev, i, (block = dev->SdramFreeLists[i]));
+
+    /* clear the approriate bit in the bitmap */
+    BT_CLEAR (sdram_off_to_bank (dev, block)->Bitmaps[i], sdram_off_to_bit (dev,i, block));
+
+    /* and split it up as required */
+    while (i-- > indx)
+       free_block (dev, block + (size >>= 1), i);
+
+    PRINTF1 (DBG_DEVICE, DBG_SDRAM, "elan3_sdram_alloc: return block=%lx\n", block);
+
+    spin_unlock_irqrestore (&dev->SdramLock, flags);
+
+    ASSERT ((block & ((SDRAM_MIN_BLOCK_SIZE << (indx))-1)) == 0);
+
+    return ((sdramaddr_t) block);
+}
+
+void
+elan3_sdram_free (ELAN3_DEV *dev, sdramaddr_t block, int nbytes)
+{
+    register int indx;
+    unsigned long size;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->SdramLock, flags);
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size < nbytes; indx++, size <<= 1)
+       ;
+
+    PRINTF2 (DBG_DEVICE, DBG_SDRAM, "elan3_sdram_free: indx=%d block=%lx\n", indx, block);
+    
+    free_block (dev, block, indx);
+
+    spin_unlock_irqrestore (&dev->SdramLock, flags);
+}
+
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/tproc.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/tproc.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/tproc.c        2005-06-01 23:12:54.594439864 -0400
@@ -0,0 +1,778 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: tproc.c,v 1.51.2.1 2004/11/15 11:12:36 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/tproc.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+#include <elan3/elansyscall.h>
+#include <elan3/threadsyscall.h>
+#include <elan3/intrinsics.h>
+#include <elan3/vmseg.h>
+
+int
+HandleTProcTrap (ELAN3_DEV *dev, E3_uint32 *RestartBits)
+{
+    THREAD_TRAP  *trap  = dev->ThreadTrap;
+    int           delay = 1;
+
+    ASSERT(SPINLOCK_HELD (&dev->IntrLock));
+
+    trap->Status.Status  = read_reg32 (dev, Exts.TProcStatus);
+    trap->sp             = read_reg32 (dev, Thread_Desc_SP);
+    trap->pc             = read_reg32 (dev, ExecutePC);
+    trap->npc            = read_reg32 (dev, ExecuteNPC);
+    trap->StartPC        = read_reg32 (dev, StartPC);
+    trap->mi             = GET_STATUS_TRAPTYPE(trap->Status);
+    trap->TrapBits.Bits  = read_reg32 (dev, TrapBits.Bits);
+    trap->DirtyBits.Bits = read_reg32 (dev, DirtyBits.Bits);
+
+    if ( ! (trap->Status.s.WakeupFunction == SleepOneTick) ) {
+       int p,i;
+       E3_uint32 reg = read_reg32 (dev, Exts.InterruptReg);    
+
+       ELAN_REG_REC(reg);
+       p = elan_reg_rec_index;
+       for(i=0;i<ELAN_REG_REC_MAX;i++) {
+           if (elan_reg_rec_file[i] != NULL ) 
+               printk("Elan Reg Record[%2d](%ld): cpu %d  reg %x [%d:%s]\n", p, elan_reg_rec_lbolt[p], elan_reg_rec_cpu[p], elan_reg_rec_reg[p],
+                      elan_reg_rec_line[p], elan_reg_rec_file[p]);
+           p = ( (p+1) % ELAN_REG_REC_MAX);
+       }
+    }
+    
+    ASSERT(trap->Status.s.WakeupFunction == SleepOneTick);
+
+    /* copy the four access fault areas */
+    elan3_sdram_copyq_from_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, TProc),          (void *) &trap->FaultSave, 16);
+    elan3_sdram_copyq_from_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcData), (void *) &trap->DataFaultSave, 16);
+    elan3_sdram_copyq_from_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcInst), (void *) &trap->InstFaultSave, 16);
+    elan3_sdram_copyq_from_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcOpen), (void *) &trap->OpenFaultSave, 16);
+    
+    /* copy the registers,  note the endian swap flips the odd registers into the even registers
+       and visa versa. */
+    copy_thread_regs (dev, trap->Registers);
+
+    /*
+     * If the output was open then the ack may not have returned yet. Must wait for the
+     * ack to become valid and update trap_dirty with the new value. Will simulate the
+     * instructions later.
+     */
+    if (trap->TrapBits.s.OutputWasOpen)
+    {
+       trap->TrapBits.Bits = read_reg32 (dev, TrapBits.Bits);
+       while (! trap->TrapBits.s.AckBufferValid)
+       {
+           PRINTF0 (DBG_DEVICE, DBG_INTR, "tproc: waiting for ack to become valid\n");
+           trap->TrapBits.Bits = read_reg32 (dev, TrapBits.Bits);
+           DELAY (delay);
+
+           if ((delay <<= 1) == 0) delay = 1;
+       }
+    }
+    
+    /* update device statistics */
+    BumpStat (dev, TProcTraps);
+    switch (trap->mi)
+    {
+    case MI_UnimplementedError:
+       if (trap->TrapBits.s.ForcedTProcTrap)
+           BumpStat (dev, ForcedTProcTraps);
+       if (trap->TrapBits.s.ThreadTimeout)
+       {
+           if (trap->TrapBits.s.PacketTimeout)
+               BumpStat (dev, ThreadOutputTimeouts);
+           else if (trap->TrapBits.s.PacketAckValue == E3_PAckError)
+               BumpStat (dev, ThreadPacketAckErrors);
+       }
+       if (trap->TrapBits.s.TrapForTooManyInsts)
+           BumpStat (dev, TrapForTooManyInsts);
+       break;
+    }
+
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, TProc), 16);
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcData), 16);
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcInst), 16);
+    elan3_sdram_zeroq_sdram (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, ThreadProcOpen), 16);
+
+    *RestartBits |= RestartTProc;
+
+    return (TRUE);
+}
+
+void
+DeliverTProcTrap (ELAN3_DEV *dev, THREAD_TRAP *threadTrap, E3_uint32 Pend)
+{
+    ELAN3_CTXT      *ctxt;
+    THREAD_TRAP      *trap;
+
+    ASSERT(SPINLOCK_HELD (&dev->IntrLock));
+
+    ctxt = ELAN3_DEV_CTX_TABLE(dev, threadTrap->Status.s.Context);
+
+    if (ctxt == NULL)
+    {
+       PRINTF1 (DBG_DEVICE, DBG_INTR, "DeliverTProcTrap: context %x invalid\n", threadTrap->Status.s.Context);
+       BumpStat (dev, InvalidContext);
+    }
+    else
+    {
+       if (ELAN3_OP_TPROC_TRAP (ctxt, threadTrap) == OP_DEFER)
+       {
+           if (ELAN3_QUEUE_REALLY_FULL (ctxt->ThreadTrapQ))
+           {
+               ctxt->Status |= CTXT_COMMAND_OVERFLOW_ERROR;
+               StartSwapoutContext (ctxt, Pend, NULL);
+           }
+           else
+           {
+               trap = ELAN3_QUEUE_BACK (ctxt->ThreadTrapQ, ctxt->ThreadTraps);
+               
+               bcopy (threadTrap, trap, sizeof (THREAD_TRAP));
+               
+               PRINTF4 (ctxt, DBG_INTR, "DeliverTProcTrap: SP=%08x PC=%08x NPC=%08x StartPC %08x\n",
+                        trap->sp, trap->pc, trap->npc, trap->StartPC);
+               PRINTF3 (ctxt, DBG_INTR, "       mi=%s trap=%08x dirty=%08x\n",
+                        MiToName (trap->mi), trap->TrapBits.Bits, trap->DirtyBits.Bits);
+               PRINTF3 (ctxt, DBG_INTR, "       FaultSave : FaultAddress %08x EventAddress %08x FSR %08x\n",
+                        trap->FaultSave.s.FaultAddress, trap->FaultSave.s.EventAddress, trap->FaultSave.s.FSR.Status);
+               PRINTF3 (ctxt, DBG_INTR, "       DataFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+                        trap->DataFaultSave.s.FaultAddress, trap->DataFaultSave.s.EventAddress, trap->DataFaultSave.s.FSR.Status);
+               PRINTF3 (ctxt, DBG_INTR, "       InstFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+                        trap->InstFaultSave.s.FaultAddress, trap->InstFaultSave.s.EventAddress, trap->InstFaultSave.s.FSR.Status);
+               PRINTF3 (ctxt, DBG_INTR, "       OpenFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+                        trap->OpenFaultSave.s.FaultAddress, trap->OpenFaultSave.s.EventAddress, trap->OpenFaultSave.s.FSR.Status);
+               
+               PRINTF4 (ctxt, DBG_INTR, "       g0=%08x g1=%08x g2=%08x g3=%08x\n", 
+                        trap->Registers[REG_GLOBALS+(0^WordEndianFlip)], trap->Registers[REG_GLOBALS+(1^WordEndianFlip)], 
+                        trap->Registers[REG_GLOBALS+(2^WordEndianFlip)], trap->Registers[REG_GLOBALS+(3^WordEndianFlip)]);
+               PRINTF4 (ctxt, DBG_INTR, "       g4=%08x g5=%08x g6=%08x g7=%08x\n", 
+                        trap->Registers[REG_GLOBALS+(4^WordEndianFlip)], trap->Registers[REG_GLOBALS+(5^WordEndianFlip)], 
+                        trap->Registers[REG_GLOBALS+(6^WordEndianFlip)], trap->Registers[REG_GLOBALS+(7^WordEndianFlip)]);
+               PRINTF4 (ctxt, DBG_INTR, "       o0=%08x o1=%08x o2=%08x o3=%08x\n", 
+                        trap->Registers[REG_OUTS+(0^WordEndianFlip)], trap->Registers[REG_OUTS+(1^WordEndianFlip)], 
+                        trap->Registers[REG_OUTS+(2^WordEndianFlip)], trap->Registers[REG_OUTS+(3^WordEndianFlip)]);
+               PRINTF4 (ctxt, DBG_INTR, "       o4=%08x o5=%08x o6=%08x o7=%08x\n", 
+                        trap->Registers[REG_OUTS+(4^WordEndianFlip)], trap->Registers[REG_OUTS+(5^WordEndianFlip)], 
+                        trap->Registers[REG_OUTS+(6^WordEndianFlip)], trap->Registers[REG_OUTS+(7^WordEndianFlip)]);
+               PRINTF4 (ctxt, DBG_INTR, "       l0=%08x l1=%08x l2=%08x l3=%08x\n", 
+                        trap->Registers[REG_LOCALS+(0^WordEndianFlip)], trap->Registers[REG_LOCALS+(1^WordEndianFlip)],
+                        trap->Registers[REG_LOCALS+(2^WordEndianFlip)], trap->Registers[REG_LOCALS+(3^WordEndianFlip)]);
+               PRINTF4 (ctxt, DBG_INTR, "       l4=%08x l5=%08x l6=%08x l7=%08x\n", 
+                        trap->Registers[REG_LOCALS+(4^WordEndianFlip)], trap->Registers[REG_LOCALS+(5^WordEndianFlip)],
+                        trap->Registers[REG_LOCALS+(6^WordEndianFlip)], trap->Registers[REG_LOCALS+(7^WordEndianFlip)]);
+               PRINTF4 (ctxt, DBG_INTR, "       i0=%08x i1=%08x i2=%08x i3=%08x\n", 
+                        trap->Registers[REG_INS+(0^WordEndianFlip)], trap->Registers[REG_INS+(1^WordEndianFlip)],
+                        trap->Registers[REG_INS+(2^WordEndianFlip)], trap->Registers[REG_INS+(3^WordEndianFlip)]);
+               PRINTF4 (ctxt, DBG_INTR, "       i4=%08x i5=%08x i6=%08x i7=%08x\n", 
+                        trap->Registers[REG_INS+(4^WordEndianFlip)], trap->Registers[REG_INS+(5^WordEndianFlip)],
+                        trap->Registers[REG_INS+(6^WordEndianFlip)], trap->Registers[REG_INS+(7^WordEndianFlip)]);
+               
+               ELAN3_QUEUE_ADD (ctxt->ThreadTrapQ);
+               kcondvar_wakeupone (&ctxt->Wait, &dev->IntrLock);
+               
+               if (ELAN3_QUEUE_FULL (ctxt->ThreadTrapQ))
+               {
+                   PRINTF0 (ctxt, DBG_INTR, "DeliverTProcTrap: thread queue full,  must swap out\n");
+                   ctxt->Status |= CTXT_THREAD_QUEUE_FULL;
+                   
+                   StartSwapoutContext (ctxt, Pend, NULL);
+               }
+           }
+       }
+    }
+}
+
+int
+NextTProcTrap (ELAN3_CTXT *ctxt, THREAD_TRAP *trap)
+{
+    ELAN3_DEV *dev = ctxt->Device;
+
+    ASSERT (SPINLOCK_HELD (&dev->IntrLock));
+    
+    if (ELAN3_QUEUE_EMPTY (ctxt->ThreadTrapQ))
+       return (0);
+
+    *trap = *ELAN3_QUEUE_FRONT (ctxt->ThreadTrapQ, ctxt->ThreadTraps);
+    ELAN3_QUEUE_REMOVE (ctxt->ThreadTrapQ);
+    
+    return (1);
+}
+
+void
+ResolveTProcTrap (ELAN3_CTXT *ctxt, THREAD_TRAP *trap)
+{
+    int       i;
+    int       res;
+    E3_Addr   StackPointer;
+
+    PRINTF4 (ctxt, DBG_TPROC, "ResolveTProcTrap: SP=%08x PC=%08x NPC=%08x StartPC %08x\n",
+            trap->sp, trap->pc, trap->npc, trap->StartPC);
+    PRINTF3 (ctxt, DBG_TPROC, "       mi=%s trap=%08x dirty=%08x\n",
+            MiToName (trap->mi), trap->TrapBits.Bits, trap->DirtyBits.Bits);
+    PRINTF3 (ctxt, DBG_TPROC, "       FaultSave : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->FaultSave.s.FaultAddress, trap->FaultSave.s.EventAddress, trap->FaultSave.s.FSR.Status);
+    PRINTF3 (ctxt, DBG_TPROC, "       DataFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->DataFaultSave.s.FaultAddress, trap->DataFaultSave.s.EventAddress, trap->DataFaultSave.s.FSR.Status);
+    PRINTF3 (ctxt, DBG_TPROC, "       InstFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->InstFaultSave.s.FaultAddress, trap->InstFaultSave.s.EventAddress, trap->InstFaultSave.s.FSR.Status);
+    PRINTF3 (ctxt, DBG_TPROC, "       OpenFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->OpenFaultSave.s.FaultAddress, trap->OpenFaultSave.s.EventAddress, trap->OpenFaultSave.s.FSR.Status);
+    
+    PRINTF4 (ctxt, DBG_TPROC, "       g0=%08x g1=%08x g2=%08x g3=%08x\n", 
+            trap->Registers[REG_GLOBALS+(0^WordEndianFlip)], trap->Registers[REG_GLOBALS+(1^WordEndianFlip)], 
+            trap->Registers[REG_GLOBALS+(2^WordEndianFlip)], trap->Registers[REG_GLOBALS+(3^WordEndianFlip)]);
+    PRINTF4 (ctxt, DBG_TPROC, "       g4=%08x g5=%08x g6=%08x g7=%08x\n", 
+            trap->Registers[REG_GLOBALS+(4^WordEndianFlip)], trap->Registers[REG_GLOBALS+(5^WordEndianFlip)], 
+            trap->Registers[REG_GLOBALS+(6^WordEndianFlip)], trap->Registers[REG_GLOBALS+(7^WordEndianFlip)]);
+    PRINTF4 (ctxt, DBG_TPROC, "       o0=%08x o1=%08x o2=%08x o3=%08x\n", 
+            trap->Registers[REG_OUTS+(0^WordEndianFlip)], trap->Registers[REG_OUTS+(1^WordEndianFlip)], 
+            trap->Registers[REG_OUTS+(2^WordEndianFlip)], trap->Registers[REG_OUTS+(3^WordEndianFlip)]);
+    PRINTF4 (ctxt, DBG_TPROC, "       o4=%08x o5=%08x o6=%08x o7=%08x\n", 
+            trap->Registers[REG_OUTS+(4^WordEndianFlip)], trap->Registers[REG_OUTS+(5^WordEndianFlip)], 
+            trap->Registers[REG_OUTS+(6^WordEndianFlip)], trap->Registers[REG_OUTS+(7^WordEndianFlip)]);
+    PRINTF4 (ctxt, DBG_TPROC, "       l0=%08x l1=%08x l2=%08x l3=%08x\n", 
+            trap->Registers[REG_LOCALS+(0^WordEndianFlip)], trap->Registers[REG_LOCALS+(1^WordEndianFlip)],
+            trap->Registers[REG_LOCALS+(2^WordEndianFlip)], trap->Registers[REG_LOCALS+(3^WordEndianFlip)]);
+    PRINTF4 (ctxt, DBG_TPROC, "       l4=%08x l5=%08x l6=%08x l7=%08x\n", 
+            trap->Registers[REG_LOCALS+(4^WordEndianFlip)], trap->Registers[REG_LOCALS+(5^WordEndianFlip)],
+            trap->Registers[REG_LOCALS+(6^WordEndianFlip)], trap->Registers[REG_LOCALS+(7^WordEndianFlip)]);
+    PRINTF4 (ctxt, DBG_TPROC, "       i0=%08x i1=%08x i2=%08x i3=%08x\n", 
+            trap->Registers[REG_INS+(0^WordEndianFlip)], trap->Registers[REG_INS+(1^WordEndianFlip)],
+            trap->Registers[REG_INS+(2^WordEndianFlip)], trap->Registers[REG_INS+(3^WordEndianFlip)]);
+    PRINTF4 (ctxt, DBG_TPROC, "       i4=%08x i5=%08x i6=%08x i7=%08x\n", 
+            trap->Registers[REG_INS+(4^WordEndianFlip)], trap->Registers[REG_INS+(5^WordEndianFlip)],
+            trap->Registers[REG_INS+(6^WordEndianFlip)], trap->Registers[REG_INS+(7^WordEndianFlip)]);
+           
+
+    BumpUserStat (ctxt, TProcTraps);
+
+    switch (trap->mi)
+    {
+    case MI_UnimplementedError:
+    {
+       /*
+        * This occurs if the threads processor trapped. All other cases will be for the ucode
+        * thread trapping.
+        */
+       int restart = 1;
+       int skip    = 0;
+       
+       PRINTF1 (ctxt, DBG_TPROC, "TProc: Mi=Unimp. Using trap->TrapBits=%x\n", trap->TrapBits.Bits);
+       
+       /*
+        * Data Access Exception.
+        */
+       if (trap->TrapBits.s.DataAccessException)
+       {
+           ASSERT (CTXT_IS_KERNEL(ctxt) || trap->DataFaultSave.s.FSR.Status == 0 ||
+                   ctxt->Capability.cap_mycontext == trap->DataFaultSave.s.FaultContext);
+
+           PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: DataAccessException %08x\n", trap->DataFaultSave.s.FaultAddress);
+
+           if ((res = elan3_pagefault (ctxt, &trap->DataFaultSave, 1)) != ESUCCESS)
+           {
+               PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: elan3_pagefault failed for data %08x\n",
+                        trap->DataFaultSave.s.FaultAddress);
+
+               if (ElanException (ctxt, EXCEPTION_INVALID_ADDR, THREAD_PROC, trap, &trap->DataFaultSave, res) != OP_IGNORE)
+                   restart = 0;
+           }
+       }
+       
+       /* 
+        * Instruction Access Exception.
+        */
+       if (trap->TrapBits.s.InstAccessException)
+       {
+           ASSERT (CTXT_IS_KERNEL (ctxt) || trap->InstFaultSave.s.FSR.Status == 0 ||
+                   ctxt->Capability.cap_mycontext == trap->InstFaultSave.s.FaultContext);
+           
+           PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: InstAccessException %08x\n", trap->InstFaultSave.s.FaultAddress);
+
+           if ((res = elan3_pagefault (ctxt, &trap->InstFaultSave, 1)) != ESUCCESS)
+           {
+               PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: elan3_pagefault failed for inst %08x\n",
+                        trap->InstFaultSave.s.FaultAddress);
+
+               ElanException (ctxt, EXCEPTION_INVALID_ADDR, THREAD_PROC, trap, &trap->InstFaultSave, res);
+               restart = 0;
+           }
+       }
+       
+       /*
+        * Forced TProc trap/Unimplemented instruction
+        *
+        *  If there is a force tproc trap then don't look at 
+        *  the unimplemented instruction bit - since it can
+        *  be set in obscure circumstances.
+        */
+       if (trap->TrapBits.s.ForcedTProcTrap)
+           PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: forced tproc trap, restarting\n");
+       else if (trap->TrapBits.s.Unimplemented)
+       {
+           E3_uint32 instr = ELAN3_OP_LOAD32 (ctxt, trap->pc & PC_MASK);
+
+           PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: unimplemented instruction %08x\n", instr);
+
+           if ((instr & OPCODE_MASK) == OPCODE_Ticc &&
+               (instr & OPCODE_IMM)  == OPCODE_IMM &&
+               (Ticc_COND(instr)     == Ticc_TA))
+           {
+               switch (INSTR_IMM(instr))
+               {
+               case ELAN3_ELANCALL_TRAPNUM:
+                   /*
+                    * Since the thread cannot easily access the global variable which holds
+                    * the elan system call number, we provide a different trap for the elan
+                    * system call, and copy the system call number into %g1 before calling
+                    * ThreadSyscall().
+                    */
+                   BumpUserStat (ctxt, ThreadElanCalls);
+
+                   if (ThreadElancall (ctxt, trap, &skip) != ESUCCESS)
+                   {
+                       ElanException (ctxt, EXCEPTION_BAD_SYSCALL, THREAD_PROC, trap);
+                       restart = 0;
+                   }
+                   break;
+
+               case ELAN3_SYSCALL_TRAPNUM:
+                   BumpUserStat (ctxt, ThreadSystemCalls);
+
+                   if (ThreadSyscall (ctxt, trap, &skip) != ESUCCESS)
+                   {
+                       ElanException (ctxt, EXCEPTION_BAD_SYSCALL, THREAD_PROC, trap);
+                       restart = 0;
+                   }
+                   break;
+
+               case ELAN3_DEBUG_TRAPNUM:
+                   ElanException (ctxt, EXCEPTION_DEBUG, THREAD_PROC, trap);
+                   skip = 1;
+                   break;
+                   
+               case ELAN3_ABORT_TRAPNUM:
+               default:
+                   ElanException (ctxt, EXCEPTION_UNIMP_INSTR, THREAD_PROC, trap, instr);
+                   restart = 0;
+                   break;
+               }
+                   
+           }
+           else
+           {
+               ElanException (ctxt, EXCEPTION_UNIMP_INSTR, THREAD_PROC, trap, instr);
+               restart = 0;
+           }
+       }
+       
+       /*
+        * Faulted fetching routes.
+        */
+       if (trap->TrapBits.s.OpenRouteFetch)
+       {
+           PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: OpenRouteFetch %08x\n", trap->OpenFaultSave.s.FaultAddress);
+           
+           if ((res = ResolveVirtualProcess (ctxt, trap->OpenFaultSave.s.FaultAddress)) != ESUCCESS &&
+               ElanException (ctxt, EXCEPTION_INVALID_PROCESS, THREAD_PROC, trap, trap->DataFaultSave.s.FaultAddress, res) != OP_IGNORE)
+           {
+               restart = 0;
+           }
+           else if (RollThreadToClose (ctxt, trap, E3_PAckDiscard) != ESUCCESS)        /* Force a discard */
+           {
+               restart = 0;
+           }
+       }
+       
+       /*
+        * Thread Timeout
+        */
+       if (trap->TrapBits.s.ThreadTimeout)
+       {
+           if (ElanException (ctxt, EXCEPTION_PACKET_TIMEOUT, THREAD_PROC, trap) != OP_IGNORE)
+               restart = 0;
+           else
+           {
+               PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: timeout or PAckError!\n");
+               
+               /* Might deschedule the thread for a while or mark the link error here. */
+               if (! trap->TrapBits.s.OutputWasOpen && RollThreadToClose (ctxt, trap, trap->TrapBits.s.PacketAckValue) != ESUCCESS)
+               {
+                   restart = 0;
+               }
+           }
+       }
+       
+       /*
+        * Open exception
+        */
+       if (trap->TrapBits.s.OpenException)
+       {
+           PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: open exception\n");
+           if (ElanException (ctxt, EXCEPTION_THREAD_KILLED, THREAD_PROC, trap) != OP_IGNORE)
+               restart = 0;
+       }
+       
+       /*
+        * Too many instructions.
+        */
+       if (trap->TrapBits.s.TrapForTooManyInsts)
+       {
+           PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: too many instructions\n");
+           if (ElanException (ctxt, EXCEPTION_THREAD_KILLED, THREAD_PROC, trap) != OP_IGNORE)
+               restart = 0;
+       }
+       
+       if (restart)
+       {
+           /*
+            * If the output was open when the trap was taken then the trap code must move
+            * the PC on past the close instruction and simulate the effect of all the instructions
+            * that do not output onto the link. The value of the ack received is then used to
+            * simulate the close instruction.
+            */
+           if (trap->TrapBits.s.OutputWasOpen && RollThreadToClose(ctxt, trap, trap->TrapBits.s.PacketAckValue) != ESUCCESS)
+           {
+               /*
+                * Don't restart if we couldn't roll it forweards 
+                * to a close instruction.
+                */
+               break;
+           }
+
+           /*
+            * We must check back 3 instructions from the PC,  and if we see the
+            * c_close_cookie() sequence then we must execute the instructions to
+            * the end of it.
+            */
+           /* XXXX: code to be written */
+           
+           StackPointer = SaveThreadToStack (ctxt, trap, skip);
+           
+           ReissueStackPointer (ctxt, StackPointer);
+       }
+       
+       break;
+    }
+    
+    /*
+     * This case is different from the others as %o6 has been overwritten with
+     * the SP. The real PC can be read from StartPC and written back
+     * into %o6 on the stack.
+     */
+    case MI_TProcNext:                 /* Reading the outs block */
+    {
+       E3_Addr stack = (trap->sp & SP_MASK) - sizeof (E3_Stack);
+
+       if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+       {
+           ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+           PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: faulted writing StartPc to o6\n");
+           ElanException (ctxt, EXCEPTION_CANNOT_SAVE_THREAD, THREAD_PROC, NULL);
+           break;
+       }
+       ELAN3_OP_STORE32 (ctxt, stack + offsetof (E3_Stack, Outs[6]), trap->StartPC & PC_MASK);
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+       /* DROPTHROUGH */
+    }
+    /*
+     * all of these will be generated when starting up a thread.
+     * Just re-issue the command after fixing the trap. The ucode keeps the startup
+     * from trap information in Thread_Desc_SP while it is still loading the regs.
+     */
+    case MI_WaitForGlobalsRead:                /* Reading the globals block (trap restart) */
+    case MI_WaitForNPCRead:            /* Reading the nPC, V and C (trap restart) */
+    case MI_WaitForPCload:             /* Reading the PC, N and Z (trap restart) */
+    case MI_WaitForInsRead:            /* Reading the ins block (trap restart) */
+    case MI_WaitForLocals:             /* Reading the ins block (trap restart) */
+    case MI_WaitForPCload2:            /* Reading the PC (normal thread start) */
+    case MI_WaitForSpStore:            /* Writing the SP to the outs block */
+       PRINTF2 (ctxt, DBG_TPROC, "ResolveTProcTrap: %s %08x\n", MiToName (trap->mi), trap->InstFaultSave.s.FaultAddress);
+
+       if ((res = elan3_pagefault (ctxt, &trap->FaultSave, 1)) != ESUCCESS)
+       {
+           PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: elan3_pagefault failed at %08x\n",
+                    trap->FaultSave.s.FaultAddress);
+           if (ElanException (ctxt, EXCEPTION_INVALID_ADDR, THREAD_PROC, &trap->FaultSave, trap, res) != OP_IGNORE)
+               break;
+       }
+
+       ReissueStackPointer (ctxt, trap->sp);
+       break;
+       
+       /*
+        * These traps could occur after the threads proc has stopped (either for a wait,
+        * break, or suspend, but not a trap). Must simulate the uCode's job.
+        */
+    case MI_WaitForOutsWrite:          /* Writing the outs block */
+    case MI_WaitForNPCWrite:           /* Writing the nPC block */
+    {
+       E3_uint32 DeschedBits = (trap->TrapBits.Bits & E3_TProcDescheduleMask);
+       E3_Addr   stack       = (trap->sp & SP_MASK) - sizeof (E3_Stack);
+       
+       PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: trapped on %s while stopping a thread\n", MiToName(trap->mi));
+       
+       /*
+        * Copy npc into o6.
+        */
+       trap->Registers[REG_OUTS+(6^WordEndianFlip)] = trap->npc;
+       
+       if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+       {
+           ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+           PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: faulted writing outs to stack\n");
+           ElanException (ctxt, EXCEPTION_CANNOT_SAVE_THREAD, THREAD_PROC, NULL);
+           break;
+       }
+       
+       /*
+        * Now write the outs back to the stack. NOTE then endian flip is undone.
+        */
+       for (i = 0; i < 8; i++)
+           ELAN3_OP_STORE32 (ctxt, stack + offsetof (E3_Stack, Outs[i]), trap->Registers[REG_OUTS+(i^WordEndianFlip)]);
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+       
+       /*
+        * thread has been saved. Now find out why the thread proc stopped.
+        */
+       if (DeschedBits == E3_TProcDescheduleSuspend)
+       {
+           PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: suspend instruction executed\n");
+           break;
+       }
+       
+       /*
+        * Break. Just reissue the command.
+        */
+       if (DeschedBits == E3_TProcDescheduleBreak)
+       {
+           PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: break instruction, reissue sp %08x\n", trap->sp);
+           ReissueStackPointer (ctxt, trap->sp);
+           break;
+       }
+       
+       ASSERT (DeschedBits == E3_TProcDescheduleWait);
+        
+       /* DROPTHROUGH to fix up a wait event */
+    }
+    
+    /*
+     * Trapped here trying to execute a wait instruction. All the thread state has already
+     * been saved and the trap has been fixed so simplest thing to do is to start the
+     * thread up at the wait instruction again.
+     */
+    case MI_WaitForEventWaitAddr:      /* Reading back the %o0,%o1 pair for a
+                                          wait event instr. */
+    case MI_WaitForWaitEventAccess:    /* Locked dword read of the event location.
+                                          Note that this read is done with write
+                                          permissions so we never get a trap on the write */
+    {
+       E3_Addr stack = (trap->sp & SP_MASK) - sizeof (E3_Stack);
+       
+       if ((res = elan3_pagefault (ctxt, &trap->FaultSave, 1)) != ESUCCESS)
+       {
+           PRINTF1 (ctxt, DBG_TPROC, "ResolveTProcTrap: elan3_pagefault failed at %08x\n", 
+                    trap->FaultSave.s.FaultAddress);
+           if (ElanException (ctxt, EXCEPTION_INVALID_ADDR, THREAD_PROC, trap, &trap->DataFaultSave, res) != OP_IGNORE)
+               break;
+       }
+
+       if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+       {
+           ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+           PRINTF0 (ctxt, DBG_TPROC, "ResolveTProcTrap: faulted writing pc to stack\n");
+           ElanException (ctxt, EXCEPTION_CANNOT_SAVE_THREAD, THREAD_PROC, NULL);
+           break;
+       }
+
+       ELAN3_OP_STORE32 (ctxt, stack + offsetof (E3_Stack, Outs[6]), trap->pc);
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+       
+       ReissueStackPointer (ctxt, trap->sp);
+       break;
+    }
+    
+    /*
+     * Assume the fault will be fixed by FixupEventTrap.
+     */
+    default:
+       FixupEventTrap (ctxt, THREAD_PROC, trap, trap->mi, &trap->FaultSave, 0);
+       break;
+    }
+}
+
+int
+TProcNeedsRestart (ELAN3_CTXT *ctxt)
+{
+    return (ctxt->ItemCount[LIST_THREAD] != 0);
+}
+
+void
+RestartTProcItems (ELAN3_CTXT *ctxt)
+{
+    void     *item;
+    E3_uint32 StackPointer;
+
+    kmutex_lock (&ctxt->SwapListsLock);
+    
+    while (ctxt->ItemCount[LIST_THREAD])
+    {
+       if (! ELAN3_OP_GET_WORD_ITEM (ctxt, LIST_THREAD, &item, &StackPointer))
+           ctxt->ItemCount[LIST_THREAD] = 0;
+       else
+       {
+           if (IssueCommand (ctxt, offsetof (E3_CommandPort, RunThread), StackPointer, 0) == ISSUE_COMMAND_RETRY)
+           {
+               ELAN3_OP_PUTBACK_ITEM (ctxt, LIST_THREAD, item);
+               kmutex_unlock (&ctxt->SwapListsLock);
+               return;
+           }
+           
+           ctxt->ItemCount[LIST_THREAD]--;
+           ELAN3_OP_FREE_WORD_ITEM (ctxt, item);
+       }
+    }
+    kmutex_unlock (&ctxt->SwapListsLock);
+}
+
+E3_Addr
+SaveThreadToStack (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, int SkipInstruction)
+{
+    E3_Addr      stack = (trap->sp & SP_MASK) - sizeof (E3_Stack);
+    E3_Addr      orflag;
+    register int i;
+
+    /*
+     * When the thread deschedules normally, the N & Z flags are written 
+     * to the stack in o6, and the V & C flags are lost.
+     * Since the Elan will store the NPC into o6 (to skip the instruction), 
+     * the CC flags are visible to the trap handler in the trapped PC and NPC.   
+     * If the instruction needs to be re-executed then the CC flags need to be
+     * kept in the right place to be read in when the thread re-starts.
+     *
+     * PC  has N & Z from trapped NPC.
+     * NPC has V & C from trapped PC.
+     */
+    if (SkipInstruction)
+    {
+       trap->Registers[REG_OUTS+(6^WordEndianFlip)]    = trap->npc;
+       trap->Registers[REG_GLOBALS+(0^WordEndianFlip)] = ((trap->npc & PC_MASK) + 4) | (trap->pc & CC_MASK);
+    }
+    else
+    {
+       trap->Registers[REG_OUTS+(6^WordEndianFlip)]    = (trap->pc & PC_MASK) | (trap->npc & CC_MASK);
+       trap->Registers[REG_GLOBALS+(0^WordEndianFlip)] = (trap->npc & PC_MASK) | (trap->pc & CC_MASK);
+    }
+    
+    if (ELAN3_OP_START_FAULT_CHECK(ctxt))
+    {
+       PRINTF0 (ctxt, DBG_TPROC, "RestartThread: faulted writing out thread\n");
+       ELAN3_OP_END_FAULT_CHECK(ctxt);
+
+       ElanException (ctxt, EXCEPTION_CANNOT_SAVE_THREAD, THREAD_PROC, NULL);
+       return ((E3_Addr) 0);
+    }
+
+
+#ifdef DEBUG_PRINTF
+    PRINTF4 (ctxt, DBG_TPROC, "SaveThreadToStack: SP=%08x PC=%08x NPC=%08x DIRTY=%08x\n",
+            trap->sp, trap->pc, trap->npc, trap->DirtyBits.Bits);
+    if (trap->DirtyBits.s.GlobalsDirty)
+    {
+       PRINTF4 (ctxt, DBG_TPROC, "       g0=%08x g1=%08x g2=%08x g3=%08x\n", 
+                trap->Registers[REG_GLOBALS+(0^WordEndianFlip)], trap->Registers[REG_GLOBALS+(1^WordEndianFlip)], 
+                trap->Registers[REG_GLOBALS+(2^WordEndianFlip)], trap->Registers[REG_GLOBALS+(3^WordEndianFlip)]);
+       PRINTF4 (ctxt, DBG_TPROC, "       g4=%08x g5=%08x g6=%08x g7=%08x\n", 
+                trap->Registers[REG_GLOBALS+(4^WordEndianFlip)], trap->Registers[REG_GLOBALS+(5^WordEndianFlip)], 
+                trap->Registers[REG_GLOBALS+(6^WordEndianFlip)], trap->Registers[REG_GLOBALS+(7^WordEndianFlip)]);
+    }
+    if (trap->DirtyBits.s.OutsDirty)
+    {
+       PRINTF4 (ctxt, DBG_TPROC, "       o0=%08x o1=%08x o2=%08x o3=%08x\n", 
+                trap->Registers[REG_OUTS+(0^WordEndianFlip)], trap->Registers[REG_OUTS+(1^WordEndianFlip)], 
+                trap->Registers[REG_OUTS+(2^WordEndianFlip)], trap->Registers[REG_OUTS+(3^WordEndianFlip)]);
+       PRINTF4 (ctxt, DBG_TPROC, "       o4=%08x o5=%08x o6=%08x o7=%08x\n", 
+                trap->Registers[REG_OUTS+(4^WordEndianFlip)], trap->Registers[REG_OUTS+(5^WordEndianFlip)], 
+                trap->Registers[REG_OUTS+(6^WordEndianFlip)], trap->Registers[REG_OUTS+(7^WordEndianFlip)]);
+    }
+    if (trap->DirtyBits.s.LocalsDirty)
+    {
+       PRINTF4 (ctxt, DBG_TPROC, "       l0=%08x l1=%08x l2=%08x l3=%08x\n", 
+                trap->Registers[REG_LOCALS+(0^WordEndianFlip)], trap->Registers[REG_LOCALS+(1^WordEndianFlip)], 
+                trap->Registers[REG_LOCALS+(2^WordEndianFlip)], trap->Registers[REG_LOCALS+(3^WordEndianFlip)]);
+       PRINTF4 (ctxt, DBG_TPROC, "       l4=%08x l5=%08x l6=%08x l7=%08x\n", 
+                trap->Registers[REG_LOCALS+(4^WordEndianFlip)], trap->Registers[REG_LOCALS+(5^WordEndianFlip)], 
+                trap->Registers[REG_LOCALS+(6^WordEndianFlip)], trap->Registers[REG_LOCALS+(7^WordEndianFlip)]);
+    }
+    if (trap->DirtyBits.s.InsDirty)
+    {
+       PRINTF4 (ctxt, DBG_TPROC, "       i0=%08x i1=%08x i2=%08x i3=%08x\n", 
+                trap->Registers[REG_INS+(0^WordEndianFlip)], trap->Registers[REG_INS+(1^WordEndianFlip)], 
+                trap->Registers[REG_INS+(2^WordEndianFlip)], trap->Registers[REG_INS+(3^WordEndianFlip)]);
+       PRINTF4 (ctxt, DBG_TPROC, "       i4=%08x i5=%08x i6=%08x i7=%08x\n", 
+                trap->Registers[REG_INS+(4^WordEndianFlip)], trap->Registers[REG_INS+(5^WordEndianFlip)], 
+                trap->Registers[REG_INS+(6^WordEndianFlip)], trap->Registers[REG_INS+(7^WordEndianFlip)]);
+    }
+#endif 
+    
+    PRINTF1 (ctxt, DBG_TPROC, "flushing registers to stack %08x\n", stack);
+
+    /* 
+     * NOTE - store the register to the stack in reverse order, since the stack 
+     * will be allocated in sdram, and we cannot use the sdram accessing functions 
+     * here, as it is "mapped" in user-space.
+     */
+    for (i = 0; i < 8; i++)
+    {
+       if (trap->DirtyBits.s.GlobalsDirty & (1 << i))
+           ELAN3_OP_STORE32 (ctxt, stack + offsetof (E3_Stack, Globals[i]), trap->Registers[REG_GLOBALS+(i^WordEndianFlip)]);
+       if (trap->DirtyBits.s.OutsDirty & (1 << i))
+           ELAN3_OP_STORE32 (ctxt, stack + offsetof (E3_Stack, Outs[i]),    trap->Registers[REG_OUTS+(i^WordEndianFlip)]);
+       if (trap->DirtyBits.s.LocalsDirty & (1 << i))
+           ELAN3_OP_STORE32 (ctxt, stack + offsetof (E3_Stack, Locals[i]),  trap->Registers[REG_LOCALS+(i^WordEndianFlip)]);
+       if (trap->DirtyBits.s.InsDirty & (1 << i))
+           ELAN3_OP_STORE32 (ctxt, stack + offsetof (E3_Stack, Ins[i]),     trap->Registers[REG_INS+(i^WordEndianFlip)]);
+    }
+
+    /* always restore all registers */
+    orflag = ThreadRestartFromTrapBit | ThreadReloadAllRegs;
+    
+    ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+    return (trap->sp | orflag);
+}
+
+void
+ReissueStackPointer (ELAN3_CTXT *ctxt, E3_Addr StackPointer)
+{
+    PRINTF1 (ctxt, DBG_TPROC, "ReissueStackPointer : Queue SP %08x\n", StackPointer);
+    
+    kmutex_lock (&ctxt->SwapListsLock);
+    ctxt->ItemCount[LIST_THREAD]++;
+    ELAN3_OP_PUT_WORD_ITEM (ctxt, LIST_THREAD, StackPointer);
+    kmutex_unlock (&ctxt->SwapListsLock);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/tprocinsts.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/tprocinsts.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/tprocinsts.c   2005-06-01 23:12:54.595439712 -0400
@@ -0,0 +1,401 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: tprocinsts.c,v 1.20 2003/09/24 13:57:25 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/tprocinsts.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+#include <elan3/vmseg.h>
+#include <elan3/elan3mmu.h>
+
+#define MAXINSTR       256             /* # Instructions to look at while looking for close */
+
+static E3_uint32 ALU (ELAN3_CTXT *ctxt,
+                     E3_uint32 fcode, E3_uint32 X, E3_uint32 Y,
+                     E3_uint32 *Z, E3_uint32 *N, E3_uint32 *C, E3_uint32 *V);
+
+char *OpcodeNames[] =
+{
+   "ADD   ",
+   "AND   ",
+   "OR    ",
+   "XOR   ",
+   "SUB   ",
+   "ANDN  ",
+   "ORN   ",
+   "XNOR  ",
+   "ADDX  ",
+   "UNIP  ",
+   "UMUL  ",
+   "SMUL  ",
+   "SUBX  ",
+   "UNIP  ",
+   "UDIV  ",
+   "SDIV  ",
+   "ADDcc ",
+   "ANDcc ",
+   "ORcc  ",
+   "XORcc ",
+   "SUBcc ",
+   "ANDNcc",
+   "ORNcc ",
+   "XNORcc",
+   "ADDXcc",
+   "UNIPcc",
+   "UMULcc",
+   "SMULcc",
+   "SUBXcc",
+   "UNIPcc",
+   "UDIVcc",
+   "SDIVcc"
+};
+
+#define REGISTER_VALUE(trap, rN)               (((rN) == 0) ? 0 : (trap)->Registers[(rN)^WordEndianFlip])
+#define ASSIGN_REGISTER(trap, rN, value)       ((rN) != 0 ? trap->Registers[(rN)^WordEndianFlip] = (value) : 0)
+
+int
+RollThreadToClose (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, E3_uint32 PAckVal)
+{
+    E3_Addr   pc      = (trap->pc & PC_MASK);
+    E3_Addr   npc     = (trap->npc & PC_MASK);
+    E3_uint32 Z       = (trap->npc & PSR_Z_BIT) ? 1 : 0;
+    E3_uint32 N       = (trap->npc & PSR_N_BIT) ? 1 : 0;
+    E3_uint32 C       = (trap->pc  & PSR_C_BIT) ? 1 : 0;
+    E3_uint32 V       = (trap->pc  & PSR_V_BIT) ? 1 : 0;
+    E3_uint32 instr;
+    E3_Addr   addr;
+
+    if (ELAN3_OP_START_FAULT_CHECK (ctxt))
+    {
+    failed:
+       ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+       ElanException (ctxt, EXCEPTION_SIMULATION_FAILED, THREAD_PROC, trap);
+       return (EFAULT);
+    }
+
+    /*
+     * Thread trapped with output open, or while closing,
+     * so roll the PC forwards to the instruction after the
+     * next c_close, and execute that with the register
+     * specified in c_close set to the trap which occured.
+     * (This is not 1 which means an ACK)
+     */
+    PRINTF1 (ctxt, DBG_TPROC, "RollThreadToClose: roll pc %x to c_close\n", pc);
+    
+    for (;;)
+    {
+       instr = ELAN3_OP_LOAD32 (ctxt, pc);
+
+       PRINTF2 (ctxt, DBG_TPROC, "RollThreadToClose: PC=%x INSTR=%x\n", pc, instr);
+
+       switch (OPCODE_CLASS(instr))
+       {
+       case OPCODE_CLASS_0:
+           switch ((instr) & OPCODE_CLASS0_MASK)
+           {
+           case OPCODE_SETHI:
+               PRINTF3 (ctxt, DBG_TPROC, "PC %x : sethi r%d = %x\n", pc, INSTR_RD(instr), instr << 10);
+
+               ASSIGN_REGISTER (trap, INSTR_RD(instr), instr << 10);
+               break;
+
+           case OPCODE_SENDREG:
+               PRINTF1 (ctxt, DBG_TPROC, "PC %x : sendreg\n", pc);
+               break;
+               
+           case OPCODE_SENDMEM:
+               PRINTF1 (ctxt, DBG_TPROC, "PC %x : sendmem\n", pc);
+               break;
+               
+           case OPCODE_BICC:
+           {
+               int     DoBranch   = (instr >> 28) & 1;
+               int     CondBranch = 1;
+               E3_Addr OldnPC     = npc;
+
+               PRINTF5 (ctxt, DBG_TPROC, "PC %x : Bicc Z=%x N=%x C=%x V=%x ", pc, Z, N, C, V);
+               switch (instr & OPCODE_BICC_MASK)
+               {
+               case OPCODE_BICC_BN:    CondBranch = 0;                 break;
+               case OPCODE_BICC_BE:    DoBranch ^= Z;                  break;
+               case OPCODE_BICC_BLE:   DoBranch ^= Z | (N ^ V);        break;
+               case OPCODE_BICC_BL:    DoBranch ^= N ^ V;              break;
+               case OPCODE_BICC_BLEU:  DoBranch ^= C | Z;              break;
+               case OPCODE_BICC_BCS:   DoBranch ^= C;                  break;
+               case OPCODE_BICC_BNEG:  DoBranch ^= N;                  break;
+               case OPCODE_BICC_BVS:   DoBranch ^= V;                  break;
+               }
+
+               /* Do the branch */
+               if (DoBranch != 0)
+               {
+                   npc = pc + (((instr & 0x3fffff) << 2) |
+                               (((instr & 0x200000) != 0) ? 0xff000000 : 0));
+                   
+                   PRINTF2 (ctxt, DBG_TPROC, "PC %x : branch taken to %x\n", pc, npc);
+               }
+               else
+               {
+                   npc = npc + 4;
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : branch not taken\n", pc);
+               }
+               pc = OldnPC;
+
+               /* Test if the next is annuled */
+               if (((instr & OPCODE_BICC_ANNUL) != 0) &
+                   ((DoBranch == 0) | (CondBranch == 0)))
+               {
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : branch annulled\n", pc);
+
+                   pc = npc;
+                   npc += 4;
+               }
+
+               /*
+                * we've already consumed the instruction - so continue rather
+                * than break;
+                */
+               continue;
+           }
+           
+           default:
+               PRINTF2 (ctxt, DBG_TPROC, "PC %x : unknown class 0 instr %x\n", pc, instr);
+               goto failed;
+           }
+           break;
+
+       case OPCODE_CLASS_1:
+               PRINTF2 (ctxt, DBG_TPROC, "PC %x : unknown class 1 instr %x\n", pc, instr);
+               goto failed;
+               
+       case OPCODE_CLASS_2:
+       {
+           E3_uint32 X = REGISTER_VALUE (trap, INSTR_RS1(instr));
+           E3_uint32 Y = (instr & OPCODE_IMM) ? INSTR_IMM(instr) : REGISTER_VALUE (trap, INSTR_RS2(instr));
+           
+           if ((instr & OPCODE_NOT_ALUOP) == 0)
+           {
+               E3_uint32 fcode  = (instr >> OPCODE_FCODE_SHIFT) & OPCODE_FCODE_MASK;
+               E3_uint32 result = ALU (ctxt, fcode, X, Y, &Z, &N, &C, &V);
+
+               PRINTF5 (ctxt, DBG_TPROC, "PC %x : %s %x %x -> %x", pc, OpcodeNames[fcode], X, Y, result);
+               PRINTF4 (ctxt, DBG_TPROC, "        Z=%x N=%x C=%x V=%x\n", Z, N, C, V);
+               
+               ASSIGN_REGISTER (trap, INSTR_RD(instr), result);
+           }
+           else
+           {
+               switch (instr & OPCODE_MASK)
+               {
+               case OPCODE_OPEN:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : c_open\n", pc);
+                   break;
+
+               case OPCODE_CLOSE:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : c_close\n", pc);
+                   goto found_close;
+
+               case OPCODE_SLL:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : SLL\n", pc);
+
+                   ASSIGN_REGISTER (trap, INSTR_RD(instr), X << Y);
+                   break;
+                   
+               case OPCODE_SRL:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : SRL\n", pc);
+                   
+                   ASSIGN_REGISTER (trap, INSTR_RD(instr), X >> Y);
+                   break;
+                   
+               case OPCODE_SRA:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : SRA\n", pc);
+                   
+                   ASSIGN_REGISTER (trap, INSTR_RD(instr), X >> Y);
+                   break;
+                   
+               case OPCODE_BREAKTEST:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : BREAKTEST  not allowed while open\n", pc);
+                   goto failed;
+                   
+               case OPCODE_BREAK:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : BREAK not allowed while open\n", pc);
+                   goto failed;
+
+               case OPCODE_SUSPEND:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : SUSPEND not allowed while open\n", pc);
+                   goto failed;
+                   
+               case OPCODE_WAIT:
+                   PRINTF1 (ctxt, DBG_TPROC, "PC %x : WAIT not allowed while open\n", pc);
+                   goto failed;
+
+               default:
+                   PRINTF2 (ctxt, DBG_TPROC, "PC %x : unknown class 2 instr %x\n", pc, instr);
+                   goto failed;
+               }
+           }
+           break;
+       }
+       
+       case OPCODE_CLASS_3:
+       {
+           if ((instr & OPCODE_IMM) != 0)
+               addr = REGISTER_VALUE (trap, INSTR_RS1(instr)) + INSTR_IMM(instr);
+           else
+               addr = (REGISTER_VALUE (trap, INSTR_RS1(instr)) + 
+                       REGISTER_VALUE (trap, INSTR_RS2(instr)));
+
+           switch (instr & OPCODE_MASK)
+           {
+           case OPCODE_LD:
+               PRINTF3 (ctxt, DBG_TPROC, "PC %x : LD [%x], r%d\n", pc, addr, INSTR_RD(instr));
+               
+               ASSIGN_REGISTER (trap, INSTR_RD(instr), ELAN3_OP_LOAD32 (ctxt, addr));
+               break;
+               
+           case OPCODE_LDD:
+           case OPCODE_LDBLOCK16:
+           case OPCODE_LDBLOCK32:
+           case OPCODE_LDBLOCK64:
+               PRINTF2 (ctxt, DBG_TPROC, "PC %x : LDBLOCKx @ %x is not possible while output open\n", pc, addr);
+               goto failed;
+           
+           case OPCODE_ST:
+               PRINTF2 (ctxt, DBG_TPROC, "PC %x : ST @ %x\n", pc, addr);
+               
+               ELAN3_OP_STORE32 (ctxt, addr, REGISTER_VALUE (trap, INSTR_RD(instr)));
+               break;
+                             
+           case OPCODE_STD:
+           case OPCODE_STBLOCK16:
+           case OPCODE_STBLOCK32:
+           case OPCODE_STBLOCK64:
+               PRINTF2 (ctxt, DBG_TPROC, "PC %x : STD @ %x is not posisble while output open\n", pc, addr);
+               goto failed;
+
+           case OPCODE_SWAP:
+               PRINTF2 (ctxt, DBG_TPROC, "PC %x : SWAP @ %x is not posible while output open\n", pc, addr);
+               goto failed;
+               
+           default:
+               PRINTF2 (ctxt, DBG_TPROC, "PC %x : unknown class 3 instr %x\n", pc, instr);
+               goto failed;
+           }
+           break;
+       }}
+
+       pc = npc;
+       npc += 4;
+    }
+    
+found_close:
+    ELAN3_OP_END_FAULT_CHECK (ctxt);
+
+    PRINTF1 (ctxt, DBG_TPROC, "PC %x : c_close\n", pc);
+    
+    /*
+     * Found the new pc, and have the close instruction in *instr
+     */
+    ASSIGN_REGISTER (trap, INSTR_RD(instr), PAckVal);
+    
+    /*
+     * Move to instruction after close.
+    */
+    trap->pc = npc;
+    
+    /* Insert the value of Z and N from the close inst */
+    trap->npc = (npc + 4) | ((PAckVal == E3_PAckOk) ? 1 :
+                            (PAckVal == E3_PAckTestFail) ? 2 : 0);
+
+    return (ESUCCESS);
+}
+
+E3_uint32
+ALU (ELAN3_CTXT *ctxt,
+     E3_uint32 fcode, E3_uint32 X, E3_uint32 Y,
+     E3_uint32 *Z, E3_uint32 *N, E3_uint32 *C, E3_uint32 *V)
+{
+    E3_uint32 XMSB, YMSB, ZMSB, Cprime;
+    E3_uint32 Yprime;
+    E3_uint32 Result=0;
+    
+    Yprime = ((fcode >> 2) & 1) ? ~Y : Y;
+    Cprime = ((fcode >> 2) & 1) ^ (*C & ((fcode >> 3) & 1));
+    XMSB             = (X >> 31) & 1;
+    YMSB             = (Yprime >> 31) & 1;
+    /* mul or div */
+    if ((fcode & 0xa) == 0xa)
+    {
+       PRINTF0 (ctxt, DBG_TPROC, "ALU: tried a multiply or a divide\n");
+       return (0);
+    }
+
+    switch (fcode & 3)
+    {
+       /*ADD */
+    case 0:
+       Result = X + Yprime + Cprime ;
+       if ((fcode & 0x10) == 0)
+           return (Result);
+       
+       ZMSB   = Result >> 31;
+       *V = ((XMSB & YMSB & ~ZMSB) | (~XMSB &~YMSB &  ZMSB));
+       *C = ((fcode >> 2) & 1) ^ ( (XMSB & YMSB) | (~ZMSB & (XMSB | YMSB)));
+       break;
+       
+       /*AND */
+    case 1:
+       Result = X & Yprime ;
+       if ((fcode & 0x10) == 0)
+           return (Result);
+       
+       *V = 0;
+       *C = 0;
+       break;
+       
+       /*OR  */
+    case 2:
+       Result = X | Yprime ;
+       if ((fcode & 0x10) == 0)
+           return (Result);
+       
+       *V = 0;
+       *C = 0;
+       break;
+       
+       /*XOR */
+    case 3:
+       Result = X ^ Yprime ;
+       if ((fcode & 0x10) == 0)
+           return (Result);
+       
+       *V = 0;
+       *C = 0;
+       break;
+    }
+    
+    *Z = (Result == 0) ? 1 : 0;
+    *N = (Result >> 31) & 1;
+
+    return (Result);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/tproc_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/tproc_linux.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/tproc_linux.c  2005-06-01 23:12:54.596439560 -0400
@@ -0,0 +1,215 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: tproc_linux.c,v 1.19.2.1 2004/10/28 17:08:56 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/tproc_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/autoconf.h>
+
+#include <asm/mman.h>
+#include <linux/file.h>
+
+#ifdef NO_ABI
+#include <asm/poll.h>
+extern asmlinkage long sys_open(const char *, int, int);
+extern asmlinkage ssize_t sys_write(unsigned int, const char *, size_t);
+extern asmlinkage ssize_t sys_read(unsigned int, char *, size_t);
+extern asmlinkage off_t sys_lseek(unsigned int, off_t, unsigned int);
+extern asmlinkage long sys_poll(struct pollfd *, unsigned int, long);
+extern asmlinkage long sys_kill(int, int); 
+#else
+#      include <linux/syscall.h>
+#endif
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+#include <elan3/elansyscall.h>
+#include <elan3/threadsyscall.h>
+
+/*
+ * NOTE: system calls from kernel on Linux are different on alpha and i386 
+ *       on alpha they return -errno on failure 
+ *       on i386  they return -1 on failure and set errno 
+ */
+
+static void
+ReturnSyscall (THREAD_TRAP *trap, unsigned long rc, int *skip)
+{
+    if (rc >= (unsigned long) (-130))
+    {
+       trap->pc |= PSR_C_BIT;  /* clear carry to indicate failure */
+
+       trap->Registers[REG_OUTS+(0^WordEndianFlip)] = -rc;
+    } 
+    else 
+    {
+       trap->pc &= ~PSR_C_BIT; /* set carry to indicate success */
+       trap->Registers[REG_OUTS+(0^WordEndianFlip)] = rc;
+    }
+    trap->Registers[REG_OUTS+(1^WordEndianFlip)] = 0;
+    *skip = 1;
+}
+
+static void 
+dump_regs(ELAN3_CTXT *ctxt, THREAD_TRAP *trap)
+{
+    PRINTF (ctxt, DBG_TPROC, "               OUTS %08x %08x %08x %08x\n",
+      trap->Registers[REG_OUTS+(0^WordEndianFlip)], 
+      trap->Registers[REG_OUTS+(1^WordEndianFlip)],
+      trap->Registers[REG_OUTS+(2^WordEndianFlip)], 
+      trap->Registers[REG_OUTS+(3^WordEndianFlip)]);
+    PRINTF (ctxt, DBG_TPROC, "                    %08x %08x %08x %08x\n",
+      trap->Registers[REG_OUTS+(4^WordEndianFlip)], 
+      trap->Registers[REG_OUTS+(5^WordEndianFlip)],
+      trap->Registers[REG_OUTS+(6^WordEndianFlip)], 
+      trap->Registers[REG_OUTS+(7^WordEndianFlip)]);
+}
+
+int
+ThreadSyscall (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, int *skip)
+{
+    int                  code;
+    caddr_t       maddr;
+    struct file  *file;
+    unsigned long rc;
+    int           i;
+    uintptr_t     av[6];
+    uintptr_t     ptr;
+   
+    PRINTF (ctxt, DBG_TPROC, "ThreadSyscall: PC %08x G1 %08x\n", 
+      trap->pc, trap->Registers[REG_GLOBALS+(1^WordEndianFlip)]);
+    dump_regs(ctxt, trap);
+
+    code = trap->Registers[REG_GLOBALS+(1^WordEndianFlip)];
+
+    /* Copy the system call arguments from %o0-%o5 */
+    for (i = 0; i < 6; i++)
+       av[i] = trap->Registers[REG_OUTS+(i^WordEndianFlip)];
+    
+    rc = (unsigned long) -EINVAL;
+
+    switch (code) {
+       case ELAN3_SYS_open:
+           maddr = elan3mmu_mainaddr (ctxt->Elan3mmu, (E3_Addr) av[0]);
+           if (maddr != NULL)
+               rc = sys_open((const char *)maddr, av[1], av[2]);
+           break;
+
+       case ELAN3_SYS_close:
+           rc = sys_close(av[0]);
+           break;
+
+       case ELAN3_SYS_write:
+           maddr = elan3mmu_mainaddr (ctxt->Elan3mmu, (E3_Addr) av[1]);
+           if (maddr != NULL)
+               rc = sys_write(av[0], (const char *)maddr, av[2]);
+           break;
+
+       case ELAN3_SYS_read:
+           maddr = elan3mmu_mainaddr (ctxt->Elan3mmu, (E3_Addr) av[1]);
+           if (maddr != NULL)
+               rc = sys_read(av[0], (char *)maddr, av[2]);
+           break;
+
+       case ELAN3_SYS_poll:
+           maddr = elan3mmu_mainaddr (ctxt->Elan3mmu, (E3_Addr) av[0]);
+           if (maddr != NULL)
+               rc = sys_poll((struct pollfd *)maddr, av[1], av[2]);
+           break;
+       
+       case ELAN3_SYS_lseek:
+           rc = sys_lseek(av[0], av[1], av[2]);
+           break;
+           
+       case ELAN3_SYS_mmap:
+           if ((E3_Addr) av[0] == (E3_Addr) 0)
+               maddr = NULL;
+           else if ((maddr = elan3mmu_mainaddr (ctxt->Elan3mmu, (E3_Addr) av[0])) == NULL)
+               break;
+       
+           file = NULL;
+           /* GNAT 5515: If *not* anonymous memory need to do fget */
+           if ((av[3] & MAP_ANONYMOUS) == 0 && (file = fget (av[4])) == NULL)
+           {
+               rc = -EBADF;
+               break;
+           }
+
+           down_write (&current->mm->mmap_sem);
+           ptr = do_mmap_pgoff (file, (unsigned long) maddr, av[1], av[2], av[3], av[5] >>PAGE_SHIFT);
+           up_write (&current->mm->mmap_sem);
+
+           if (file)
+               fput (file);
+           
+           if (IS_ERR((void *) ptr))
+               rc = PTR_ERR((void *) ptr);
+           else
+               rc = elan3mmu_elanaddr (ctxt->Elan3mmu, (caddr_t)ptr);
+
+           break;
+       
+       case ELAN3_SYS_munmap:
+           maddr = elan3mmu_mainaddr (ctxt->Elan3mmu, (E3_Addr) av[0]);
+
+#ifdef AC
+           if (maddr != NULL)
+               rc = do_munmap(current->mm, (unsigned long) maddr, av[1], 1);
+#else
+           if (maddr != NULL)
+               rc = do_munmap(current->mm, (unsigned long) maddr, av[1]);
+#endif
+           break;
+
+       case ELAN3_SYS_kill:
+           rc = sys_kill(av[0], av[1]);
+           break;
+
+       case ELAN3_SYS_getpid:
+           rc = current->pid;
+           break;
+
+       default:
+           return EINVAL;
+    }
+    ReturnSyscall(trap, rc, skip);
+    return ESUCCESS;
+}
+
+
+int
+ThreadElancall (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, int *skip)
+{
+       int ret = ESUCCESS;
+
+       PRINTF (ctxt, DBG_TPROC, "ThreadElancall: PC %08x\n", trap->pc);
+       dump_regs(ctxt, trap);
+
+       /*
+        * Elan system call 'type' is passed in o0
+        */
+       switch (trap->Registers[REG_OUTS+(0^WordEndianFlip)]) 
+       {
+       default:
+               ret = EINVAL;
+               break;
+       }
+       return ret;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan3/virtual_process.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan3/virtual_process.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan3/virtual_process.c      2005-06-01 23:12:54.597439408 -0400
@@ -0,0 +1,884 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: virtual_process.c,v 1.68 2004/06/07 13:50:10 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/os/virtual_process.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/elanmod.h>
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#include <elan3/urom_addrs.h>
+#include <elan3/thread.h>
+#include <elan3/vmseg.h>
+#include <elan3/elansyscall.h>
+
+static ELAN3_VPSEG *
+InstallSegment (ELAN3_CTXT *ctxt, int process, int entries)
+{
+    ELAN3_VPSEG **prevSeg, *seg;
+    int lastTop = -1;
+    int        top     = process + entries-1;
+
+    ASSERT (krwlock_is_write_locked (&ctxt->VpLock));
+
+    for (prevSeg = &ctxt->VpSegs; (seg = (*prevSeg)) != NULL; prevSeg = &seg->Next)
+    {
+       int thisTop = seg->Process + seg->Entries - 1;
+
+        if (process < seg->Process && (process <= lastTop || top >= seg->Process))
+       {
+           /*
+            * Overlaps with last segment, or this one 
+            */
+           return (NULL);
+       }
+       if (seg->Process > process)
+           break;
+       
+       lastTop = thisTop;
+    }
+    
+    KMEM_ZALLOC (seg, ELAN3_VPSEG *, sizeof (ELAN3_VPSEG), TRUE);
+    
+    if (seg == (ELAN3_VPSEG *) NULL)
+       return (NULL);
+    
+    seg->Process = process;
+    seg->Entries = entries;
+
+
+    PRINTF2 (ctxt, DBG_VP, "InstallSegment: add seg %p before %p\n", seg, *prevSeg);
+
+    seg->Next = *prevSeg;
+    *prevSeg = seg;
+
+    return (seg);
+}
+
+static int 
+RemoveSegment (ELAN3_CTXT *ctxt, ELAN3_VPSEG *seg)
+{
+    ELAN3_VPSEG **prevSeg, *thisSeg;
+
+    ASSERT (krwlock_is_write_locked (&ctxt->VpLock));
+
+    for (prevSeg = &ctxt->VpSegs; (thisSeg = (*prevSeg)) != NULL; prevSeg = &thisSeg->Next)
+    {
+       if (thisSeg == seg)
+           break;
+    }
+
+    if (thisSeg == (ELAN3_VPSEG *) NULL)
+       return (EINVAL);
+
+
+    PRINTF2 (ctxt, DBG_VP, "RemoveSegment: remove seg %p next %p\n", thisSeg, thisSeg->Next);
+
+    *prevSeg = thisSeg->Next;
+    
+    KMEM_FREE ((caddr_t) seg, sizeof (ELAN3_VPSEG));
+
+    return (ESUCCESS);
+}
+
+static ELAN3_VPSEG *
+FindSegment (ELAN3_CTXT *ctxt, int low, int high)
+{
+    ELAN3_VPSEG *seg;
+
+    ASSERT(krwlock_is_locked (&ctxt->VpLock));
+    
+    for (seg = ctxt->VpSegs; seg; seg = seg->Next)
+    {
+       if (seg->Process <= low && (seg->Process + seg->Entries) > high)
+           return (seg);
+    }
+
+    return ((ELAN3_VPSEG *) NULL);
+}
+
+ELAN_LOCATION
+ProcessToLocation (ELAN3_CTXT *ctxt, ELAN3_VPSEG *seg, int process, ELAN_CAPABILITY *cap)
+{
+    ELAN_LOCATION location;
+    int           nnodes,nctxs;
+    int           node,ctx,i;
+
+    ASSERT(krwlock_is_locked (&ctxt->VpLock));
+
+    location.loc_node    = ELAN3_INVALID_NODE;
+    location.loc_context = -1;
+
+    PRINTF3 (ctxt, DBG_VP, "ProcessToLocation: process %d seg %p cap %p\n", process, seg, cap);
+
+    if (seg == NULL)
+       seg = FindSegment (ctxt, process, process);
+
+    if (!seg || (seg->Type != ELAN3_VPSEG_P2P))
+       return (location);
+
+    cap    = &seg->SegCapability;
+    nnodes = ELAN_CAP_NUM_NODES (cap);
+    nctxs  = ELAN_CAP_NUM_CONTEXTS (cap);
+
+    switch (seg->SegCapability.cap_type & ELAN_CAP_TYPE_MASK)
+    {
+    case ELAN_CAP_TYPE_BLOCK:
+    {
+       int entries = ELAN_CAP_ENTRIES(cap);
+
+       for (node = 0, i = 0; node < nnodes && i < entries; node++)
+       {
+           for (ctx = 0; ctx < nctxs && i < entries; ctx++)
+           {
+               if (( seg->SegCapability.cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (seg->SegCapability.cap_bitmap, ctx + (node * nctxs)))
+               {
+                   if (i++ == (process - seg->Process))
+                   { 
+                       location.loc_node    = seg->SegCapability.cap_lownode    + node;
+                       location.loc_context = seg->SegCapability.cap_lowcontext + ctx;
+                       goto found;
+                   }
+               }
+           }
+       }
+       break;
+    }
+    case ELAN_CAP_TYPE_CYCLIC:
+    {
+       int entries = ELAN_CAP_ENTRIES(cap);
+
+       for (ctx = 0, i = 0; ctx < nctxs && i < entries; ctx++)
+       {
+           for (node = 0; node < nnodes && i < entries; node++)
+           {
+               if ((seg->SegCapability.cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (seg->SegCapability.cap_bitmap, node + (ctx * nnodes)))
+               {                                   
+                   if (i++ ==  (process - seg->Process))
+                   { 
+                       location.loc_node    = seg->SegCapability.cap_lownode    + node;
+                       location.loc_context = seg->SegCapability.cap_lowcontext + ctx;
+                       goto found;
+                   }
+               }
+           }
+       }
+       break;  
+    }
+    default:
+       break;
+    }
+       
+ found:
+    
+    PRINTF3 (ctxt, DBG_VP, "ProcessToLocation: process %d -> Node %d Context %d\n", process, location.loc_node,  location.loc_context);
+
+    if (cap != NULL)
+    {
+       bcopy ((caddr_t) &seg->SegCapability, (caddr_t) cap, sizeof (ELAN_CAPABILITY));
+       cap->cap_mycontext = location.loc_context;
+    }
+
+    return (location);
+}
+
+int
+LocationToProcess (ELAN3_CTXT *ctxt, ELAN3_VPSEG *seg, ELAN_LOCATION loc, ELAN_CAPABILITY *cap)
+{
+    int nnodes,nctxs;
+    int node,ctx,i;
+
+    if (seg == NULL)
+       return ELAN3_INVALID_PROCESS;
+
+    if (!seg || (seg->Type != ELAN3_VPSEG_P2P))
+       return ELAN3_INVALID_PROCESS;
+
+    nnodes = cap->cap_highnode - cap->cap_lownode + 1;
+    nctxs  = cap->cap_highcontext - cap->cap_lowcontext + 1;
+
+    switch (cap->cap_type & ELAN_CAP_TYPE_MASK)
+    {
+    case ELAN_CAP_TYPE_BLOCK:
+    {
+       int entries = ELAN_CAP_ENTRIES(cap);
+
+       for (node = 0, i = 0; node < nnodes && i < entries; node++)
+       {
+           for (ctx = 0; ctx < nctxs && i < entries; ctx++)
+           {
+               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, ctx + (node * nctxs)))
+               {
+                   if ((loc.loc_node    == (cap->cap_lownode + node) ) 
+                       && (loc.loc_context == (cap->cap_lowcontext + ctx) ))
+                   {
+                       return (i + seg->Process);
+                   }
+                   i++;
+               }
+           }
+       }
+       break;
+    }  
+    case ELAN_CAP_TYPE_CYCLIC:
+    {
+       int entries = ELAN_CAP_ENTRIES(cap);
+
+       for (ctx = 0, i = 0; ctx < nctxs && i < entries; ctx++)
+       {
+           for (node = 0; node < nnodes && i < entries; node++)
+           {
+               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, node + (ctx * nnodes)))
+               {
+                   if ((loc.loc_node   == (cap->cap_lownode + node) ) 
+                       && (loc.loc_context == (cap->cap_lowcontext + ctx) ))
+                   {
+                       return (i + seg->Process);
+                   }
+                   i++;
+                   
+               }
+           }
+       }
+       break;
+    }  
+    default:
+       break;
+    }
+       
+    return ELAN3_INVALID_PROCESS;
+}
+
+int
+elan3_addvp (ELAN3_CTXT *ctxt, int process, ELAN_CAPABILITY *cap)
+{
+    ELAN3_DEV       *dev = ctxt->Device;
+    ELAN_POSITION    *pos = &ctxt->Position;
+    ELAN3_VPSEG       *seg;
+    int                      i;
+    int                      nodeOff;
+    int                      ctxOff;
+    int                      nnodes;
+    int                      nctxs;
+    E3_uint16         flits[MAX_FLITS];
+    int               nflits;
+    int               entries;
+
+    PRINTF2 (ctxt, DBG_VP, "elan3_addvp: %d -> %s\n", process, CapabilityString (cap));
+
+    entries = ELAN_CAP_ENTRIES(cap);
+    if (entries <= 0 || (process + entries) > ELAN3_MAX_VPS)
+       return (EINVAL);
+
+    /*
+     * Scan the virtual process segment list, to add this entry, and ensure that
+     * the ranges don't overlap.
+     */
+    krwlock_write (&ctxt->VpLock);
+
+    /* check cap. */
+    switch (elan3_validate_cap (ctxt->Device, cap, ELAN_USER_P2P))
+    {
+    case ELAN_CAP_OK:
+       /* nothing */
+       break;
+
+    case ELAN_CAP_RMS:
+       if ( elan_validate_map(cap, cap) != ESUCCESS)
+       {
+           krwlock_done (&ctxt->VpLock);
+           return (EINVAL);
+       }
+       break;
+
+    default:
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    if ((seg = InstallSegment (ctxt, process, entries)) == NULL)
+    {
+       PRINTF0 (ctxt, DBG_VP, "elan3_addvp: failed to find a seg\n");
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+    
+    seg->Type                        = ELAN3_VPSEG_P2P;
+    seg->SegCapability               = *cap;
+    seg->SegCapability.cap_mycontext = ELAN_CAP_UNINITIALISED;
+
+    PRINTF3 (ctxt, DBG_VP, "elan3_addvp: segment type %x  %d %d\n",
+           seg->SegCapability.cap_type, seg->Process, entries);
+
+
+    nnodes = cap->cap_highnode - cap->cap_lownode + 1;
+    nctxs  = cap->cap_highcontext - cap->cap_lowcontext + 1;
+
+    /* position not determined, so cannot load any routes, the hwtest
+     * process must explicitly set it's own routes */
+    
+    if (!(cap->cap_type & ELAN_CAP_TYPE_HWTEST) && (pos->pos_mode != ELAN_POS_UNKNOWN))
+    {
+       switch (cap->cap_type & ELAN_CAP_TYPE_MASK)
+       {
+       case ELAN_CAP_TYPE_BLOCK:
+           for (nodeOff = 0, i = 0; nodeOff < nnodes && i < entries; nodeOff++)
+           {
+               for (ctxOff = 0; ctxOff < nctxs && i < entries; ctxOff++)
+               {
+                   if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, ctxOff + (nodeOff * nctxs)))
+                   {
+                       /* Don't load a route if there's no switch and trying to talk to myself */
+                       if (pos->pos_mode == ELAN_POS_MODE_SWITCHED ||
+                           (pos->pos_mode == ELAN_POS_MODE_LOOPBACK && cap->cap_lownode + nodeOff == pos->pos_nodeid) ||
+                           (pos->pos_mode == ELAN_POS_MODE_BACKTOBACK && cap->cap_lownode + nodeOff != pos->pos_nodeid))
+                       {
+                           PRINTF3 (ctxt, DBG_VP, "elan3_addvp: virtual process %d -> node %d context %d\n",
+                                    seg->Process + i, cap->cap_lownode +nodeOff, cap->cap_lowcontext +ctxOff);
+                           
+                           nflits = GenerateRoute (pos, flits, cap->cap_lownode + nodeOff, cap->cap_lownode + nodeOff,
+                                                   DEFAULT_ROUTE_TIMEOUT, DEFAULT_ROUTE_PRIORITY);
+                           
+
+
+                           LoadRoute (dev, ctxt->RouteTable, seg->Process+i, cap->cap_lowcontext + ctxOff, nflits, flits);  
+                       }
+                       
+                       i++;
+                   }
+               }
+           }
+           break;
+           
+       case ELAN_CAP_TYPE_CYCLIC:
+           for (ctxOff = 0, i = 0; ctxOff < nctxs && i < entries; ctxOff++)
+           {
+               for (nodeOff = 0; nodeOff < nnodes && i < entries; nodeOff++)
+               {
+                   if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, nodeOff + (ctxOff * nnodes)))
+                   {
+                       /* Don't load a route if there's no switch and trying to talk to myself */
+                       if (pos->pos_mode == ELAN_POS_MODE_SWITCHED ||
+                           (pos->pos_mode == ELAN_POS_MODE_LOOPBACK && cap->cap_lownode + nodeOff == pos->pos_nodeid) ||
+                           (pos->pos_mode == ELAN_POS_MODE_BACKTOBACK && cap->cap_lownode + nodeOff != pos->pos_nodeid))
+                       {
+                           PRINTF3 (ctxt, DBG_VP, "elan3_addvp: virtual process %d -> node %d context %d\n",
+                                    seg->Process + i, cap->cap_lownode + nodeOff, cap->cap_lowcontext +ctxOff);
+                       
+                           nflits = GenerateRoute (pos, flits, cap->cap_lownode + nodeOff, cap->cap_lownode + nodeOff,
+                                                   DEFAULT_ROUTE_TIMEOUT, DEFAULT_ROUTE_PRIORITY);
+                           
+
+                           LoadRoute (dev, ctxt->RouteTable, seg->Process+i, cap->cap_lowcontext +ctxOff, nflits, flits);  
+                       } 
+                       i++;                
+                   }
+               }
+           }
+           break;      
+       default:
+           break;
+       }
+    }
+  
+    krwlock_done (&ctxt->VpLock);
+
+    return (ESUCCESS);
+}
+
+int
+elan3_removevp (ELAN3_CTXT *ctxt, int process)
+{
+    ELAN3_VPSEG *seg;
+    ELAN3_VPSEG *next;
+    int                i;
+
+    krwlock_write (&ctxt->VpLock);
+
+    PRINTF1 (ctxt, DBG_VP, "elan3_removevp: remove process %d\n", process);
+
+    if (process == ELAN3_INVALID_PROCESS)
+       seg = ctxt->VpSegs;
+    else
+       seg = FindSegment (ctxt, process, process);
+
+    if (seg == (ELAN3_VPSEG *) NULL)
+    {
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+    
+    do {
+       PRINTF3 (ctxt, DBG_VP, "elan3_removevp: segment is %p [%x,%x]\n",
+                seg, seg->Process, seg->Process+seg->Entries);
+
+       for (i = 0; i < seg->Entries; i++)
+           ClearRoute (ctxt->Device, ctxt->RouteTable, seg->Process+i);
+
+        /* get Next pointer value before structure is free'd */
+        next = seg->Next;      
+       RemoveSegment (ctxt, seg);
+
+    } while (process == ELAN3_INVALID_PROCESS && (seg = next) != NULL);
+    
+    krwlock_done (&ctxt->VpLock);
+
+    return (ESUCCESS);
+}
+
+int
+elan3_addbcastvp (ELAN3_CTXT *ctxt, int process, int lowProc, int highProc)
+{
+    ELAN_POSITION *pos = &ctxt->Position;
+    ELAN3_VPSEG    *seg;
+    ELAN3_VPSEG    *aseg;
+    int            virtualProcess;
+    E3_uint64     routeValue;
+
+    PRINTF3 (ctxt, DBG_VP, "elan3_addbcastvp: process %d [%d,%d]\n", process, lowProc, highProc);
+
+    if (lowProc > highProc || pos->pos_mode != ELAN_POS_MODE_SWITCHED)
+       return (EINVAL);
+    
+    krwlock_write (&ctxt->VpLock);
+
+    if ((aseg = FindSegment (ctxt, lowProc, highProc)) == NULL || (aseg->Type != ELAN3_VPSEG_P2P))
+    {
+       PRINTF2 (ctxt, DBG_VP, "elan3_addbcastvp: process [%d,%d] does not map to p2p segment\n", lowProc, highProc);
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    /* check aseg->SegCapability */    
+    switch (elan3_validate_cap (ctxt->Device, &aseg->SegCapability, ELAN_USER_BROADCAST))
+    {
+    case ELAN_CAP_OK:
+       /* nothing */
+       break;
+       
+    case ELAN_CAP_RMS:
+       if ( elan_validate_map(&ctxt->Capability, &aseg->SegCapability) != ESUCCESS )
+       {
+           krwlock_done (&ctxt->VpLock);
+           return (EINVAL);
+       }
+       break;
+
+    default:
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    if ( ProcessToLocation (ctxt, aseg, lowProc,  NULL).loc_context != 
+        ProcessToLocation (ctxt, aseg, highProc, NULL).loc_context)
+    {
+       PRINTF2 (ctxt, DBG_VP, "elan3_addbcastvp: process [%d,%d] does not map to single context\n", lowProc, highProc);
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+    
+    if ((seg = InstallSegment (ctxt, process, 1)) == NULL)
+    {
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    seg->Type        = ELAN3_VPSEG_BROADCAST;
+    seg->SegLowProc  = lowProc;
+    seg->SegHighProc = highProc;
+
+    PRINTF4 (ctxt, DBG_VP, "elan3_addbcastvp: installed seg %p Type %d LowProc %d HighProc %d\n",
+           seg, seg->Type, seg->SegLowProc, seg->SegHighProc);
+
+    for (virtualProcess = lowProc; virtualProcess <= highProc; virtualProcess++)
+    {
+       if (virtualProcess < 0 || virtualProcess >= ctxt->RouteTable->Size)
+           routeValue = 0;
+       else
+           routeValue = elan3_sdram_readq ( ctxt->Device, ctxt->RouteTable->Table + virtualProcess * NBYTES_PER_SMALL_ROUTE);
+       
+       if (! (routeValue & ROUTE_VALID))
+       {
+           PRINTF2 (ctxt, DBG_VP, "loadvp[%x]: broadcast %x not valid\n", 
+                    ctxt->Capability.cap_mycontext, virtualProcess);
+           break;
+       }
+    }
+           
+    if (virtualProcess > highProc)                     /* All vps now present */
+    {                                          /* so load up broadcast route */
+       E3_uint16     flits[MAX_FLITS];
+       ELAN_LOCATION low    = ProcessToLocation (ctxt, aseg, lowProc,   NULL);
+       ELAN_LOCATION high   = ProcessToLocation (ctxt, aseg, highProc,  NULL);
+       int           nflits = GenerateRoute (pos, flits, low.loc_node, high.loc_node, DEFAULT_ROUTE_TIMEOUT, DEFAULT_ROUTE_PRIORITY);
+       
+       PRINTF6 (ctxt, DBG_VP, "loadvp[%x]: broadcast %d -> %x.%x [%x.%x]\n", ctxt->Capability.cap_mycontext,
+                seg->Process, low.loc_node, high.loc_node, 
+                low.loc_context, high.loc_context);
+       
+       LoadRoute ( ctxt->Device, ctxt->RouteTable, seg->Process, low.loc_context, nflits, flits);
+    }
+
+    krwlock_done (&ctxt->VpLock);
+
+    return (ESUCCESS);
+}
+
+int
+elan3_process (ELAN3_CTXT *ctxt)
+{
+    int           res = ELAN3_INVALID_PROCESS;
+    ELAN3_VPSEG   *seg;
+    ELAN_LOCATION loc;
+
+    krwlock_write (&ctxt->VpLock);
+
+    loc.loc_node    = ctxt->Position.pos_nodeid;
+    loc.loc_context = ctxt->Capability.cap_mycontext;
+
+    for (seg = ctxt->VpSegs ; seg; seg = seg->Next)
+    {
+       if (seg->Type == ELAN3_VPSEG_P2P &&
+           seg->SegCapability.cap_lowcontext  <= ctxt->Capability.cap_mycontext &&
+           seg->SegCapability.cap_highcontext >= ctxt->Capability.cap_mycontext &&
+           seg->SegCapability.cap_lownode     <= ctxt->Position.pos_nodeid &&
+           seg->SegCapability.cap_highnode    >= ctxt->Position.pos_nodeid)
+       {
+           if ((res=LocationToProcess (ctxt,seg,loc,&ctxt->Capability)) != ELAN3_INVALID_PROCESS)
+           {
+                krwlock_done (&ctxt->VpLock);
+                return res;
+           }
+       }
+    }
+
+    krwlock_done (&ctxt->VpLock);
+
+    return (res);
+}
+
+int
+elan3_check_route (ELAN3_CTXT *ctxt, int process, E3_uint16 *flits, E3_uint32 *routeError)
+{
+    PRINTF5 (ctxt, DBG_VP, "elan3_check_route: vp=%d flits=%04x %04x %04x %04x\n",
+            process, flits[0], flits[1], flits[2], flits[3]);
+    PRINTF4 (ctxt, DBG_VP, "                            %04x %04x %04x %04x\n",
+            flits[4], flits[5], flits[6], flits[7]);
+
+    krwlock_read (&ctxt->VpLock);
+    *routeError=elan3_route_check(ctxt,flits,ProcessToLocation (ctxt, NULL, process, NULL).loc_node);
+    krwlock_done (&ctxt->VpLock);
+
+    return (ESUCCESS); /* the call is a success tho the errorcode may be set */
+}
+
+int
+elan3_load_route (ELAN3_CTXT *ctxt, int process, E3_uint16 *flits)
+{
+    ELAN3_VPSEG *seg;
+    int                res = 0;
+    int                nflits;
+    int         err;
+
+    PRINTF5 (ctxt, DBG_VP, "elan3_load_route: vp=%d flits=%04x %04x %04x %04x\n",
+            process, flits[0], flits[1], flits[2], flits[3]);
+    PRINTF4 (ctxt, DBG_VP, "                            %04x %04x %04x %04x\n",
+            flits[4], flits[5], flits[6], flits[7]);
+
+    krwlock_write (&ctxt->VpLock);
+
+    /* check the route is valid */
+    if (!(ctxt->Capability.cap_type & ELAN_CAP_TYPE_HWTEST))
+    {
+       /* must have already attached to define my context number */
+       if (ctxt->Capability.cap_mycontext == ELAN_CAP_UNINITIALISED)
+       {
+           krwlock_done (&ctxt->VpLock);
+           return (EINVAL);
+       }
+
+       if ((err=elan3_route_check(ctxt,flits,ProcessToLocation (ctxt, NULL, process, NULL).loc_node)) != ELAN3_ROUTE_SUCCESS)
+       {
+           krwlock_done (&ctxt->VpLock);
+           return (EINVAL);
+       }
+    }
+
+    if ((seg = FindSegment (ctxt, process, process)) == NULL || seg->Type != ELAN3_VPSEG_P2P)
+    {
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    /* Calculate number of flits in this route */
+    for (nflits = 0; nflits < MAX_FLITS && flits[nflits]; nflits++)
+       ;
+    
+    res = LoadRoute (ctxt->Device, ctxt->RouteTable, process, ProcessToLocation (ctxt, seg, process, NULL).loc_context, nflits, flits);
+
+    krwlock_done (&ctxt->VpLock);
+
+    return (res);
+}
+
+int
+elan3_get_route (ELAN3_CTXT *ctxt, int process, E3_uint16 *flits)
+{
+    ELAN3_VPSEG *seg;
+    int                res = 0;
+
+    PRINTF1 (ctxt, DBG_VP, "elan3_get_route: vp=%d \n",  process);
+
+    krwlock_write (&ctxt->VpLock);
+
+    if (ctxt->RouteTable == NULL)  /* is there a route table */
+    {
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    if ((seg = FindSegment (ctxt, process, process)) != NULL && seg->Type != ELAN3_VPSEG_P2P)
+    {
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+    
+    if (seg == NULL)
+    {
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+    
+    res = GetRoute (ctxt->Device, ctxt->RouteTable, process, flits);
+    
+    krwlock_done (&ctxt->VpLock);
+
+    return (res);
+}
+
+int
+elan3_reset_route (ELAN3_CTXT *ctxt, int process)
+{
+    E3_uint16     flits[MAX_FLITS];
+
+    PRINTF1 (ctxt, DBG_VP, "elan3_reset_route: vp=%d \n",  process);
+ 
+    GenerateRoute (&ctxt->Position, flits, process, process, DEFAULT_ROUTE_TIMEOUT, DEFAULT_ROUTE_PRIORITY);
+    
+    return elan3_load_route(ctxt,process,flits);
+}
+
+int
+ResolveVirtualProcess (ELAN3_CTXT *ctxt, int process)
+{
+    E3_uint16    flits[MAX_FLITS];
+    ELAN3_DEV     *dev = ctxt->Device;
+    int                  res = ESUCCESS;
+    ELAN3_VPSEG   *seg;
+    ELAN3_VPSEG   *aseg;
+    E3_uint64    routeValue;
+
+    krwlock_read (&ctxt->VpLock);
+
+    PRINTF1 (ctxt, DBG_VP, "ResolveVirtualProcess: vp=%d \n",  process);
+
+    if (ctxt->RouteTable == NULL || process < 0 || process >= ctxt->RouteTable->Size)
+    {
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    if (! (seg = FindSegment (ctxt, process, process)))
+    {
+       PRINTF1 (ctxt, DBG_VP, "ResolveVirtualProcess: cannot find segment for virtual process %d\n", process);
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+    
+    /* check cap. */
+    switch (elan3_validate_cap (ctxt->Device, &seg->SegCapability, ((seg->Type == ELAN3_VPSEG_P2P) ? ELAN_USER_P2P : ELAN_USER_BROADCAST)))
+    {
+    case ELAN_CAP_OK:
+       /* nothing */
+       break;
+
+    case ELAN_CAP_RMS:
+       if ( elan_validate_map(&ctxt->Capability, &seg->SegCapability) != ESUCCESS)
+       {
+           krwlock_done (&ctxt->VpLock);
+           return (EINVAL);
+       }
+       break;
+
+    default:
+       krwlock_done (&ctxt->VpLock);
+       return (EINVAL);
+    }
+
+    BumpUserStat (ctxt, LoadVirtualProcess);
+
+    routeValue = elan3_sdram_readq (dev, ctxt->RouteTable->Table + process * NBYTES_PER_SMALL_ROUTE);
+    if (routeValue & ROUTE_VALID)                              /* Virtual process already */
+    {                                                          /* loaded */
+       krwlock_done (&ctxt->VpLock);
+       return (ESUCCESS);                      
+    }
+    
+    switch (seg->Type)
+    {
+    case ELAN3_VPSEG_P2P:
+       switch (seg->SegCapability.cap_type & ELAN_CAP_TYPE_MASK)
+       {
+       case ELAN_CAP_TYPE_BLOCK:
+       case ELAN_CAP_TYPE_CYCLIC:
+           if ((res = elan_validate_map (&ctxt->Capability,&seg->SegCapability)) == ESUCCESS &&
+               (res = GetRoute(dev, ctxt->RouteTable ,process,  flits)) == ESUCCESS)
+           {
+               if (elan3_route_check(ctxt, flits, ProcessToLocation (ctxt, seg, process, NULL).loc_node))
+                   res = EINVAL;
+               else
+                   ValidateRoute(dev, ctxt->RouteTable, process);
+           }
+           break;
+       default:
+           res = EINVAL;
+           break;
+       }
+       break;
+
+    case ELAN3_VPSEG_BROADCAST:
+       /* Find the segment that this broadcast range spans. */
+       aseg = FindSegment (ctxt, seg->SegLowProc, seg->SegHighProc);
+       
+       if (aseg == NULL || (aseg->Type != ELAN3_VPSEG_P2P) || !(aseg->SegCapability.cap_type & ELAN_CAP_TYPE_BROADCASTABLE))
+       {
+           PRINTF2 (ctxt, DBG_VP, "resolveVirtualProcess: %d -> EINVAL (%s)\n", process, 
+                    (aseg == NULL ? "no segment" : ((seg->Type != ELAN3_VPSEG_P2P) ? "not point to point" :
+                                                    "not broadcastable")));
+           res = EINVAL;
+           break;
+       }
+       
+       switch (aseg->SegCapability.cap_type & ELAN_CAP_TYPE_MASK)
+       {
+       case ELAN_CAP_TYPE_BLOCK:
+       case ELAN_CAP_TYPE_CYCLIC:
+       {
+           ELAN_LOCATION lowNode  = ProcessToLocation (ctxt,aseg,seg->SegLowProc  , NULL);
+           ELAN_LOCATION highNode = ProcessToLocation (ctxt,aseg,seg->SegHighProc , NULL);
+
+
+           if ((res = elan_validate_map (&ctxt->Capability,&aseg->SegCapability)) == ESUCCESS &&
+               (res=GetRoute(dev, ctxt->RouteTable ,process,  flits)) == ESUCCESS)
+           {
+               if (elan3_route_broadcast_check(ctxt,flits, lowNode.loc_node , highNode.loc_node ) != ELAN3_ROUTE_SUCCESS )
+                   res = EINVAL;
+               else
+                   ValidateRoute(dev, ctxt->RouteTable, process);
+           }
+           break;
+       }
+
+       default:
+           res = EINVAL;
+           break;
+       }
+    default:
+       res  = EINVAL;
+       break;
+    }
+
+    krwlock_done (&ctxt->VpLock);
+    return (res);
+}        
+
+void
+UnloadVirtualProcess (ELAN3_CTXT *ctxt, ELAN_CAPABILITY *cap)
+{
+    ELAN3_DEV        *dev  = ctxt->Device;
+    ELAN3_VPSEG      *seg;
+    ELAN_CAPABILITY *scap;
+    int              i;
+
+    for (seg = ctxt->VpSegs; seg; seg = seg->Next)
+    {
+       switch (seg->Type)
+       {
+       case ELAN3_VPSEG_P2P:
+           scap = &seg->SegCapability;
+           
+           if (cap == NULL || ELAN_CAP_MATCH (scap, cap))
+           {
+               PRINTF2 (ctxt, DBG_VP, "unloadvp: segment [%x.%x]\n", 
+                        seg->Process, seg->Process + seg->Entries-1);
+               
+               for (i = 0; i < seg->Entries; i++)
+                   InvalidateRoute (dev, ctxt->RouteTable, seg->Process+i);
+           }
+           break;
+
+       case ELAN3_VPSEG_BROADCAST:
+           for (i = 0; i < seg->Entries; i++)
+           {
+               ELAN3_VPSEG *aseg = FindSegment (ctxt, seg->SegLowProc, seg->SegHighProc);
+               
+               if (aseg != NULL && ELAN_CAP_MATCH(&aseg->SegCapability, cap))
+               {
+                   PRINTF1 (ctxt, DBG_VP, "unloadvp: broadcast vp %d\n", seg->Process);
+               
+                   InvalidateRoute (dev, ctxt->RouteTable, seg->Process+i);
+               }
+           }
+       }
+    }
+}
+
+caddr_t
+CapabilityString (ELAN_CAPABILITY *cap)
+{
+#define CAPSTR_LEN     200
+#define NCAPSTRS       4
+    static char       space[CAPSTR_LEN*NCAPSTRS];
+    static int        bufnum;
+    static spinlock_t lock;
+    static int       lockinitialised;
+    int                      num;
+    unsigned long     flags;
+
+    if (! lockinitialised)
+    {
+       spin_lock_init (&lock);
+       lockinitialised = 1;
+    }
+
+    spin_lock_irqsave (&lock, flags);
+    
+    if ((num = ++bufnum) == NCAPSTRS)
+       num = bufnum = 0;
+    spin_unlock_irqrestore (&lock, flags);
+
+    sprintf (space + (num * CAPSTR_LEN), "%4x %4x %4x %4x %4x %4x %4x [%x.%x.%x.%x]", cap->cap_type,
+            cap->cap_lownode, cap->cap_highnode, 
+            cap->cap_lowcontext, cap->cap_mycontext, cap->cap_highcontext,  ELAN_CAP_ENTRIES(cap),
+            cap->cap_userkey.key_values[0],  cap->cap_userkey.key_values[1],
+            cap->cap_userkey.key_values[2],  cap->cap_userkey.key_values[3]);
+
+    return (space + (num * CAPSTR_LEN));
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/debug.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/debug.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/debug.c        2005-06-01 23:12:54.597439408 -0400
@@ -0,0 +1,94 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: debug.c,v 1.16 2004/07/07 11:22:33 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/debug.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+
+unsigned       elan4_debug           = 0;
+unsigned       elan4_debug_toconsole = 0;
+unsigned       elan4_debug_tobuffer  = DBG_ALL;
+
+unsigned       elan4_debug_display_ctxt;
+unsigned       elan4_debug_ignore_ctxt;
+unsigned       elan4_debug_ignore_type;
+
+void
+elan4_debug_init()
+{
+    if ((elan4_debug & elan4_debug_tobuffer) != 0)
+       qsnet_debug_alloc();
+}
+
+void
+elan4_debug_fini()
+{
+}
+
+void
+elan4_debugf (void *type, int mode, char *fmt,...)
+{
+    char    prefix[128];
+    int     where = 0;
+    va_list ap;
+
+    if ((mode & elan4_debug_tobuffer) != 0 || type == DBG_BUFFER)
+       where |= QSNET_DEBUG_BUFFER;
+    if ((mode & elan4_debug_toconsole) != 0 || type == DBG_CONSOLE)
+       where |= QSNET_DEBUG_CONSOLE;
+
+    if (where == 0)
+       return;
+    
+    if ((unsigned long) type > DBG_NTYPES)
+    {
+       ELAN4_CTXT *ctxt = (ELAN4_CTXT *) type;
+
+        if (elan4_debug_display_ctxt && ctxt->ctxt_num != elan4_debug_display_ctxt)
+            return;
+        if (elan4_debug_ignore_ctxt  && ctxt->ctxt_num == elan4_debug_ignore_ctxt)
+            return;
+
+       sprintf (prefix, "[%08ld.%04d] elan4 (%03x) ", lbolt,  current->pid, ctxt->ctxt_num);
+    }
+    else if ((unsigned long) type == (int) DBG_CONSOLE)
+       prefix[0] = '\0';
+    else
+    {
+       char *what;
+
+       if (elan4_debug_ignore_type & (1 << ((unsigned long) type)))
+           return;
+
+       switch ((unsigned long) type)
+       {
+       case (int) DBG_DEVICE: what = "dev"; break;
+       case (int) DBG_USER:   what = "usr"; break;
+       default:               what = NULL; break;
+       }
+           
+       if (what)
+           sprintf (prefix, "[%08ld.%04d] elan4 [%s] ", lbolt, current->pid, what);
+       else
+           sprintf (prefix, "[%08ld.%04d] elan4 [%3d] ", lbolt, current->pid, (int)(long)type);
+    }
+
+    va_start(ap,fmt);
+    qsnet_vdebugf (where, prefix, fmt, ap);
+    va_end (ap);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/device.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/device.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/device.c       2005-06-01 23:12:54.602438648 -0400
@@ -0,0 +1,2805 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: device.c,v 1.87.6.3 2005/01/18 14:25:35 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/device.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+
+#include <elan4/sdram.h>
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/commands.h>
+#include <elan4/trtype.h>
+#include <elan4/neterr.h>
+
+#include <elan4/i2c.h>
+#include <elan3/vpd.h>
+
+/* allow this code to compile against an Eagle elanmod */
+#ifdef __ELANMOD_DEVICE_H
+#define ELAN_DEV_OPS           ELANMOD_DEV_OPS
+#define ELAN_DEV_OPS_VERSION   ELANMOD_DEV_OPS_VERSION
+#define elan_dev_register      elanmod_dev_register
+#define elan_dev_deregister    elanmod_dev_deregister
+#endif
+
+/* XXXX configurational defines */
+
+#if defined (CONFIG_MPSAS)
+#define HASH_0_SIZE_VAL                        (12 + 6)
+#define HASH_1_SIZE_VAL                        (2 + 6)
+#define CTXT_TABLE_SHIFT               8
+#define LN2_MAX_CQS                    8               /* 256 */
+#else
+#define HASH_0_SIZE_VAL                        (13 + 6)
+#define HASH_1_SIZE_VAL                        (2 + 6)
+#define CTXT_TABLE_SHIFT               12
+#define LN2_MAX_CQS                    10              /* 1024 */
+#endif
+
+unsigned int elan4_hash_0_size_val       = HASH_0_SIZE_VAL;
+unsigned int elan4_hash_1_size_val       = HASH_1_SIZE_VAL;
+unsigned int elan4_ctxt_table_shift      = CTXT_TABLE_SHIFT;
+unsigned int elan4_ln2_max_cqs           = LN2_MAX_CQS;
+unsigned int elan4_dmaq_highpri_size     = 2;                  /* 8192 entries */
+unsigned int elan4_threadq_highpri_size  = 1;                  /* 1024 entries */
+unsigned int elan4_dmaq_lowpri_size      = 2;                  /* 8192 entries */
+unsigned int elan4_threadq_lowpri_size   = 1;                  /* 1024 entries */
+unsigned int elan4_interruptq_size       = 0;                  /* 1024 entries */
+unsigned int elan4_mainint_punt_loops    = 1;
+unsigned int elan4_mainint_resched_ticks = 0;
+
+static int 
+elan4_op_get_position (void *arg, ELAN_POSITION *ptr)
+{
+    ELAN4_DEV     *dev = (ELAN4_DEV *)arg;
+    ELAN_POSITION  pos;
+
+    elan4_get_position (dev, &pos);
+
+    return copyout (&pos, ptr, sizeof (ELAN_POSITION));
+}
+
+static int 
+elan4_op_set_position (void *arg, unsigned short nodeid, unsigned short numnodes)
+{
+    /* XXXXX 
+
+       ELAN4_DEV *dev = (ELAN4_DEV *) arg;
+
+       compute_position (&pos, nodeid, numnode, num_down_links_value);
+
+       return elan4_set_position (dev, pos);
+    */
+    return EINVAL;
+}
+
+ELAN_DEV_OPS elan4_dev_ops = 
+{
+    elan4_op_get_position,
+    elan4_op_set_position,
+
+    ELAN_DEV_OPS_VERSION
+};
+
+static E4_uint32
+elan4_read_filter (ELAN4_DEV *dev, unsigned networkctx)
+{
+    return (elan4_sdram_readl (dev, dev->dev_ctxtable + (networkctx * sizeof (E4_ContextControlBlock)) + 
+                              offsetof (E4_ContextControlBlock, Filter)));
+}
+
+static void
+elan4_write_filter (ELAN4_DEV *dev, unsigned networkctx, E4_uint32 value)
+{
+    elan4_sdram_writel (dev, (dev->dev_ctxtable + (networkctx * sizeof (E4_ContextControlBlock)) +
+                       offsetof (E4_ContextControlBlock, Filter)), value);
+    pioflush_sdram(dev);
+}
+
+void
+elan4_set_schedstatus (ELAN4_DEV *dev, E4_uint32 intreg)
+{
+    E4_uint32 setbits  = 0;
+    E4_uint32 intmask  = 0;
+    E4_uint32 haltmask;
+    E4_uint32 next_sched;
+    E4_uint32 next_intmask;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_intmask_lock, flags);
+
+    haltmask = (dev->dev_haltop_mask | dev->dev_haltop_active);
+
+    if ((haltmask & INT_DProcHalted) || dev->dev_halt_all_count || dev->dev_halt_dproc_count)
+       setbits |= SCH_DProcHalt;
+    
+    if ((haltmask & INT_TProcHalted) || dev->dev_halt_all_count || dev->dev_halt_tproc_count)
+       setbits |= SCH_TProcHalt;
+
+    if ((haltmask & INT_CProcHalted) || dev->dev_halt_all_count || dev->dev_halt_cproc_count)
+       setbits |= SCH_CProcHalt;
+
+    if ((haltmask & INT_DiscardingLowPri) || dev->dev_discard_all_count || dev->dev_discard_lowpri_count)
+       setbits |= SCH_DiscardLowPriInput;
+    
+    if ((haltmask & INT_DiscardingHighPri) || dev->dev_discard_all_count || dev->dev_discard_highpri_count)
+       setbits |= SCH_DiscardHighPriInput;
+    
+    if (dev->dev_halt_lowpri_count)
+       setbits |= SCH_StopLowPriQueues;
+    
+    if (haltmask & INT_DProcHalted) intmask |= INT_DProcHalted;
+    if (haltmask & INT_TProcHalted) intmask |= INT_TProcHalted;
+    if (haltmask & INT_CProcHalted) intmask |= INT_CProcHalted;
+    if (haltmask & INT_DiscardingLowPri) intmask |= INT_DiscardingLowPri;
+    if (haltmask & INT_DiscardingHighPri) intmask |= INT_DiscardingHighPri;
+
+    next_intmask = (dev->dev_intmask     & ~(INT_Halted | INT_Discarding)) | (intmask & ~intreg);
+    next_sched   = (dev->dev_schedstatus & ~(SCH_Halt | SCH_Discard))      | setbits;
+
+    PRINTF5 (DBG_DEVICE, DBG_REGISTER, "elan4_set_schedstatus: haltmask=%x setbits=%x intmask=%x next_sched=%x next_intmask=%x\n",
+            haltmask, setbits, intmask, next_sched, next_intmask);
+
+    CHANGE_INT_MASK (dev, next_intmask);
+    CHANGE_SCHED_STATUS (dev, next_sched);
+
+    spin_unlock_irqrestore (&dev->dev_intmask_lock, flags);
+}
+
+void
+elan4_queue_haltop (ELAN4_DEV *dev, ELAN4_HALTOP *op)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_haltop_lock, flags);
+
+    /* add to the end of the halt operations list */
+    list_add_tail (&op->op_link, &dev->dev_haltop_list);
+
+    if ((dev->dev_haltop_mask & op->op_mask) != op->op_mask)
+    {
+       dev->dev_haltop_mask |= op->op_mask;
+       
+       elan4_set_schedstatus (dev, 0);
+    }
+    
+    spin_unlock_irqrestore (&dev->dev_haltop_lock, flags);
+}
+
+void
+elan4_queue_intop (ELAN4_DEV *dev, ELAN4_CQ *cq, ELAN4_INTOP *op)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_intop_lock, flags);
+
+    op->op_cookie = INTOP_ONESHOT | ((dev->dev_intop_cookie++) & INTOP_VALUE_MASK);
+
+    list_add_tail (&op->op_link, &dev->dev_intop_list);
+
+    writeq ((op->op_cookie << E4_MAIN_INT_SHIFT) | INTERRUPT_CMD, cq->cq_mapping);
+
+    spin_unlock_irqrestore (&dev->dev_intop_lock, flags);
+}
+
+void
+elan4_register_intop (ELAN4_DEV *dev, ELAN4_INTOP *op)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_intop_lock, flags);
+
+    op->op_cookie = INTOP_PERSISTENT | ((dev->dev_intop_cookie++) & INTOP_VALUE_MASK);
+
+    list_add_tail (&op->op_link, &dev->dev_intop_list);
+
+    spin_unlock_irqrestore (&dev->dev_intop_lock, flags);
+}
+
+void
+elan4_deregister_intop (ELAN4_DEV *dev, ELAN4_INTOP *op)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_intop_lock, flags);
+    list_del (&op->op_link);
+    spin_unlock_irqrestore (&dev->dev_intop_lock, flags);
+}
+
+static __inline__ void
+__issue_dma_flushop_cmd (ELAN4_CQ *cq)
+{
+    writeq (DMA_ShMemWrite | RUN_DMA_CMD, cq->cq_mapping);
+    writeq (0 /* cookie */,               cq->cq_mapping);
+    writeq (0 /* vproc */,                cq->cq_mapping);
+    writeq (0 /* srcAddr */,              cq->cq_mapping);
+    writeq (0 /* dstAddr */,              cq->cq_mapping);
+    writeq (0 /* srcEvent */,             cq->cq_mapping);
+    writeq (0 /* dstEvent */,             cq->cq_mapping);
+    writeq (SET_EVENT_CMD,                cq->cq_mapping);
+}
+
+static void
+handle_dma_flushops_intop (ELAN4_DEV *dev, void *arg)
+{
+    unsigned int  hipri        = ((unsigned long) arg & 1);
+    E4_uint64     status       = dev->dev_dma_flushop[hipri].status;
+    ELAN4_CQ     *cq           = dev->dev_dma_flushop[hipri].cq;
+    sdramaddr_t   cqdesc       = dev->dev_cqaddr + (elan4_cq2num(cq) * sizeof (E4_CommandQueueDesc));
+    E4_uint64     queuePtrs    = elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs));
+    E4_uint32     completedPtr = CQ_CompletedPtr(queuePtrs);
+    E4_uint32     size         = CQ_Size ((queuePtrs >> CQ_SizeShift) & CQ_SizeMask);
+    unsigned long flags;
+
+    /*
+     * Since we're called from a main interrupt which was issued through the approriate
+     * flushcq the command queue descriptor for dma flushing can no longer be in the 
+     * insert cache, nor can it be in the extractor (as it's trapped), hence it is
+     * safe to modify the completed pointer
+     */
+
+    spin_lock_irqsave (&dev->dev_haltop_lock, flags);
+
+    ASSERT (status != 0);
+
+    /* skip over either the DMA/SETEVENT or just the SETEVENT depending on the trap type */
+    if (CPROC_TrapType (status) == CommandProcDmaQueueOverflow)
+       completedPtr = (completedPtr & ~(size-1)) | ((completedPtr + 64) & (size - 1));
+    else
+       completedPtr = (completedPtr & ~(size-1)) | ((completedPtr + 8) & (size - 1));
+    
+    elan4_sdram_writel (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs) + 4,
+                       ((queuePtrs >> 32) & ~CQ_PtrOffsetMask) | (completedPtr & CQ_PtrOffsetMask));
+    
+    elan4_restartcq (dev, dev->dev_dma_flushop[hipri].cq);
+
+    if (! list_empty (&dev->dev_dma_flushop[hipri].list))
+       __issue_dma_flushop_cmd (dev->dev_dma_flushop[hipri].cq);
+
+    dev->dev_dma_flushop[hipri].status = 0;
+    
+    spin_unlock_irqrestore (&dev->dev_haltop_lock, flags);
+
+}
+
+static void
+handle_dma_flushops (ELAN4_DEV *dev, E4_uint64 status, int cqnum)
+{
+    unsigned int       hipri  = (cqnum == elan4_cq2num(dev->dev_dma_flushop[1].cq) ? 1 : 0);
+    ELAN4_CQ          *cq     = dev->dev_dma_flushop[hipri].cq;
+    ELAN4_CQ          *flushq = dev->dev_flush_cq[elan4_cq2num(cq) & (COMMAND_INSERTER_CACHE_ENTRIES-1)];
+    struct list_head  *ops;
+    unsigned long      flags;
+    int                       qfull,count;
+    E4_uint64         queuePtrs;
+    LIST_HEAD(list);
+    
+    spin_lock_irqsave (&dev->dev_haltop_lock, flags);
+    
+    ASSERT (cqnum == elan4_cq2num (dev->dev_dma_flushop[hipri].cq));
+    ASSERT (! list_empty (&dev->dev_dma_flushop[hipri].list));
+    ASSERT (dev->dev_dma_flushop[hipri].status == 0);
+    
+    /* remove the whole list */
+    ops = dev->dev_dma_flushop[hipri].list.next;
+
+    list_del_init (&dev->dev_dma_flushop[hipri].list);
+    
+    /* and add it to our local list */
+    list_add_tail (&list, ops);
+    
+    /* now determine whether the queue was full - since it cannot be empty 
+     * then if the front and back pointers are the same then it is full */
+    queuePtrs = hipri ? read_reg64 (dev, DProcHighPriPtrs) : read_reg64 (dev, DProcLowPriPtrs);
+    qfull     = (E4_QueueFrontPointer (queuePtrs) == E4_QueueBackPointer (queuePtrs));
+    
+    if (CPROC_TrapType(status) == CommandProcDmaQueueOverflow && !qfull)
+       printk (" ******* queue overflow trap - but queue not full\n");
+
+    if (qfull && CPROC_TrapType(status) != CommandProcDmaQueueOverflow)
+       printk (" ****** queue full - but not overflow trap : %llx %llx %x\n", 
+               read_reg64 (dev, DProcLowPriPtrs), read_reg64 (dev, DProcHighPriPtrs), CPROC_TrapType(status));
+
+    /* Store the status register, this also indicates that the intop is pending */
+    dev->dev_dma_flushop[hipri].status = status;
+
+    spin_unlock_irqrestore (&dev->dev_haltop_lock, flags);
+
+    /* Issue a main interrupt command to the approriate flush command queue,
+     * which will then safely update the completed pointer to skip over the
+     * command which has trapped, also prevent any new commands to be issued
+     * to the command queue.
+     */
+    dev->dev_dma_flushop[hipri].intop.op_function = handle_dma_flushops_intop;
+    dev->dev_dma_flushop[hipri].intop.op_arg      = (void *) (unsigned long) hipri;
+
+    elan4_queue_intop (dev, flushq, &dev->dev_dma_flushop[hipri].intop);
+    
+    /* now execute all operations */
+    for (count = 0; ! list_empty (&list); count++)
+    {
+       ELAN4_DMA_FLUSHOP *op = list_entry (list.next, ELAN4_DMA_FLUSHOP, op_link);
+       
+       list_del (&op->op_link);
+       
+       (*op->op_function) (dev, op->op_arg, qfull);
+    }
+
+    /* finally release the "reasons" for halting */
+    spin_lock_irqsave (&dev->dev_haltop_lock, flags);
+    if ((dev->dev_halt_dproc_count -= count) == 0)
+       elan4_set_schedstatus (dev, 0);
+    spin_unlock_irqrestore (&dev->dev_haltop_lock, flags);
+       
+    return;
+}
+
+void
+elan4_queue_dma_flushop (ELAN4_DEV *dev, ELAN4_DMA_FLUSHOP *op, int hipri)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_haltop_lock, flags);
+
+    if (dev->dev_halt_dproc_count++ == 0)                      /* ensure that the DMA processor cannot */
+       elan4_set_schedstatus (dev, 0);                         /* execute the DMA we issue. */
+
+    if (list_empty (&dev->dev_dma_flushop[hipri].list) && dev->dev_dma_flushop[hipri].status == 0)
+       __issue_dma_flushop_cmd (dev->dev_dma_flushop[hipri].cq);
+       
+    list_add_tail (&op->op_link, &dev->dev_dma_flushop[hipri].list);
+
+    spin_unlock_irqrestore (&dev->dev_haltop_lock, flags);
+}
+
+static void
+enable_elan_errors (void *arg)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) arg;
+
+    ENABLE_INT_MASK (dev, INT_ErrorInterrupts);
+}
+
+#define ERROR_DISABLE_PERIOD   (hz/2)
+#define ERROR_SAMPLE_PERIOD    (hz/10)
+#define ERROR_LIMIT            (100)
+
+static __inline__ void
+check_error_rate (ELAN4_DEV *dev)
+{
+    if (dev->dev_error_time == (lbolt/ERROR_SAMPLE_PERIOD))
+    {
+        if (++dev->dev_errors_per_period >= ERROR_LIMIT && (dev->dev_intmask & INT_ErrorInterrupts))
+       {
+           DISABLE_INT_MASK (dev, INT_ErrorInterrupts);
+           
+           schedule_timer_fn (&dev->dev_error_timeoutid, enable_elan_errors, (void *) dev, ERROR_DISABLE_PERIOD);
+       }
+    }
+    else
+    {
+       dev->dev_error_time        = (lbolt/ERROR_SAMPLE_PERIOD);
+       dev->dev_errors_per_period = 0;
+    }
+}
+
+static __inline__ int
+handle_mainints (ELAN4_DEV *dev, int nticks, int nintr)
+{
+    E4_uint32 nfptr = dev->dev_interruptq_nfptr;
+    E4_uint32 bptr  = read_reg32 (dev, MainIntQueuePtrs.s.Back);
+    E4_uint32 qsize = E4_QueueSize(elan4_interruptq_size);
+    E4_uint32 qmask = qsize - 1;
+    long      tlim  = lbolt + nticks;
+    int       done = 0;
+    unsigned long flags;
+
+    do {
+       int todo  = ((bptr - nfptr) & qmask) / E4_MainIntEntrySize;
+
+       ASSERT (todo > 0);
+
+       PRINTF4 (DBG_DEVICE, DBG_MAININT, "handle_mainints: fptr %x nfptr %x bptr %x : %d todo\n", 
+                read_reg32 (dev, MainIntQueuePtrs.s.Front), nfptr, bptr, todo);
+
+       if (nintr >= 0 && (done + todo) > nintr)                /* punt because too may to do in interrupt */
+       {
+           PRINTF4 (DBG_DEVICE, DBG_MAININT, "handle_mainints: punting (done %d todo %d) (bptr %x fptr %x)\n",
+                    done, todo, bptr, read_reg32 (dev, MainIntQueuePtrs.s.Front));
+
+           return 1;
+       }
+
+       BucketDevStat (dev, s_mainints, todo, MainIntBuckets);
+
+       /* consume all the entries in the queue which we think are there */
+       do {
+           E4_uint64   value = elan4_sdram_readq (dev, nfptr);
+           ELAN4_CTXT *ctxt  = elan4_localctxt (dev, E4_MAIN_INT_CTX (value));
+           E4_uint32   fptr  = nfptr;
+
+           PRINTF2 (DBG_DEVICE, DBG_MAININT, "handle_mainints: process cookie %llx - write fptr=%x\n", value, nfptr);
+
+           if (ctxt == NULL)
+               PRINTF1 (DBG_DEVICE, DBG_INTR, "handle_mainints: context %d invalid\n", E4_MAIN_INT_CTX (value));
+           else
+               ctxt->ctxt_ops->op_interrupt (ctxt, E4_MAIN_INT_COOKIE(value));
+
+           /* compute the next queue front pointer, before updating the front pointer
+            * since we need to ensure that elan4_queue_mainintop doesn't see the queue
+            * as being empty if an extra interrupt is queued in between */
+           dev->dev_interruptq_nfptr = nfptr = (nfptr & ~qmask) | ((nfptr + sizeof (E4_uint64)) & qmask);
+    
+           /* update the queue front pointer, doing this will clear the
+            * interrupt for *all* interrupt cookies which have previously 
+            * been added to the queue */
+           write_reg32 (dev, MainIntQueuePtrs.s.Front, E4_QueueFrontValue (fptr, elan4_interruptq_size));
+           pioflush_reg (dev);
+       } while (bptr != nfptr);
+       
+       /* re-sample the back pointer and if it's different from the previous
+        * queue front pointer, then the queue has something on it again */
+       done += todo;
+       
+       if ((nticks > 0 && ((int) (lbolt - tlim)) > 0))         /* been executing for too long in thread */
+           return 1;
+
+       bptr = read_reg32 (dev, MainIntQueuePtrs.s.Back);
+
+       PRINTF3 (DBG_DEVICE, DBG_MAININT, "handle_mainints: resample : fptr %x nfptr %x bptr %x\n", 
+                read_reg32 (dev, MainIntQueuePtrs.s.Front), nfptr, bptr);
+
+       /* at this point we've made some space in the interrupt queue,
+        * so check to see if we've got anything to restart */
+       spin_lock_irqsave (&dev->dev_mainint_lock, flags);
+       while (! list_empty (&dev->dev_interruptq_list))
+       {
+           ELAN4_INTOP *op = list_entry (dev->dev_interruptq_list.next, ELAN4_INTOP, op_link);
+           
+           list_del (&op->op_link);
+
+           op->op_function (dev, op->op_arg);
+       }
+       spin_unlock_irqrestore (&dev->dev_mainint_lock, flags);
+
+    } while (bptr != nfptr);
+
+    return 0;
+}
+
+static void
+elan4_mainint_thread (ELAN4_DEV *dev)
+{
+    unsigned long flags;
+
+    kernel_thread_init ("elan4_mainint");
+    
+    spin_lock_irqsave (&dev->dev_mainint_lock, flags);
+    for (;;)
+    {
+       if (dev->dev_stop_threads)
+           break;
+       
+       if (! (dev->dev_intmask & INT_MainInterrupt))
+       {
+           spin_unlock_irqrestore (&dev->dev_mainint_lock, flags);
+           
+           if (handle_mainints (dev, elan4_mainint_resched_ticks, -1))
+               BumpDevStat (dev, s_mainint_rescheds);
+
+           spin_lock_irqsave (&dev->dev_mainint_lock, flags);
+           ENABLE_INT_MASK (dev, INT_MainInterrupt);
+       }
+       
+       kcondvar_wait (&dev->dev_mainint_wait, &dev->dev_mainint_lock, &flags);
+    }
+
+    dev->dev_mainint_stopped = 1;
+    kcondvar_wakeupall (&dev->dev_mainint_wait, &dev->dev_mainint_lock);
+
+    spin_unlock_irqrestore (&dev->dev_mainint_lock, flags);
+
+    kernel_thread_exit();
+}
+
+void
+elan4_queue_mainintop (ELAN4_DEV *dev, ELAN4_INTOP *op)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_mainint_lock, flags);
+    if (dev->dev_interruptq_nfptr == read_reg32 (dev, MainIntQueuePtrs.s.Back))
+       op->op_function (dev, op->op_arg);
+    else
+       list_add_tail (&op->op_link, &dev->dev_interruptq_list);
+    spin_unlock_irqrestore (&dev->dev_mainint_lock, flags);
+}
+
+static __inline__ E4_uint32
+handle_cproc_trap (ELAN4_DEV *dev)
+{
+    E4_uint32   cqptr   = read_reg32 (dev, CommandControl.CommandQueueDescsBase) & E4_QueueDescPtrMask;
+    unsigned    cqnum   = ((cqptr - dev->dev_cqaddr) / sizeof (E4_CommandQueueDesc));
+    sdramaddr_t cqdesc  = dev->dev_cqaddr + (cqnum * sizeof (E4_CommandQueueDesc));
+    E4_uint64   control = elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_Control));
+    E4_uint64   status  = read_reg64 (dev, CProcStatus);
+    ELAN4_CTXT *ctxt    = elan4_localctxt (dev, CQ_Context (control));
+
+    PRINTF4 (DBG_DEVICE, DBG_INTR, "handle_cproc_trap: cqnum=%d status=%016llx control=%016llx TrapType\n", 
+            cqnum, status, control, CPROC_TrapType (status));
+    PRINTF4 (DBG_DEVICE, DBG_INTR, "                   %016llx %016llx %016llx %016llx\n",
+            elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs)),
+            elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_HoldingValue)),
+            elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_AckBuffers)),
+            elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_Control)));
+
+    BumpDevStat (dev, s_cproc_traps);
+
+    ctxt->ctxt_ops->op_cproc_trap (ctxt, status, cqnum);
+
+    return (CPROC_TrapType (status) == CommandProcWaitTrap ? SCH_RestartCProc | SCH_RestartEProc : SCH_RestartCProc);
+}
+
+static __inline__ E4_uint32
+handle_dproc_trap (ELAN4_DEV *dev, int unit)
+{
+    E4_uint64   status  = (unit == 0) ? read_reg64 (dev, DProc0Status) : read_reg64 (dev, DProc1Status);
+    E4_uint32   restart = (unit == 0) ? SCH_RestartDma0Proc : SCH_RestartDma1Proc;
+    ELAN4_CTXT *ctxt    = elan4_localctxt (dev, DPROC_Context (status));
+    
+    PRINTF3 (DBG_DEVICE, DBG_INTR, "handle_dproc_trap: unit %d context %d%s\n", unit, DPROC_Context(status),
+            DPROC_PrefetcherFault(status) ? " (prefetcher)" : "");
+
+    if (DPROC_PrefetcherFault (status))
+       restart |= SCH_RestartDmaPrefetchProc;
+                     
+    BumpDevStat (dev, s_dproc_traps);
+
+    ctxt->ctxt_ops->op_dproc_trap (ctxt, status, unit);
+
+    return (restart);
+}
+
+static __inline__ E4_uint32
+handle_eproc_trap (ELAN4_DEV *dev)
+{
+    E4_uint64   status = read_reg64 (dev, EProcStatus);
+    ELAN4_CTXT *ctxt   = elan4_localctxt (dev, EPROC_Context (status));
+
+    BumpDevStat (dev, s_eproc_traps);
+
+    ctxt->ctxt_ops->op_eproc_trap (ctxt, status);
+
+    return (SCH_RestartEProc);
+}
+
+static __inline__ E4_uint32
+handle_tproc_trap (ELAN4_DEV *dev)
+{
+    E4_uint64   status = read_reg64 (dev, TProcStatus);
+    ELAN4_CTXT *ctxt   = elan4_localctxt (dev, TPROC_Context (status));
+
+    BumpDevStat (dev, s_tproc_traps);
+
+    ctxt->ctxt_ops->op_tproc_trap (ctxt, status);
+    
+    return (SCH_RestartTProc);
+}
+
+static __inline__ void
+handle_haltints (ELAN4_DEV *dev, E4_uint32 intreg)
+{
+    struct list_head  list   = LIST_HEAD_INIT(list);
+    E4_uint32         mask   = 0;
+    E4_uint32         active = 0;
+    struct list_head *entry;
+    struct list_head *next;
+    unsigned long     flags;
+
+    BumpDevStat (dev, s_haltints);
+
+    spin_lock_irqsave (&dev->dev_haltop_lock, flags);
+
+    list_for_each_safe (entry, next, &dev->dev_haltop_list) {
+       ELAN4_HALTOP *op = list_entry (entry, ELAN4_HALTOP, op_link);
+
+       PRINTF (DBG_DEVICE, DBG_INTR, "handle_haltints: op=%p op_mask=%x intreg=%x\n", op, op->op_mask, intreg);
+
+       if ((op->op_mask & intreg) != op->op_mask)
+           mask |= op->op_mask;
+       else
+       {
+           list_del (&op->op_link);                            /* remove from list */
+           list_add_tail (&op->op_link, &list);                /* add to local list */
+
+           active |= op->op_mask;
+       }
+    }
+
+    ASSERT (dev->dev_haltop_mask == (mask | active));
+
+    dev->dev_haltop_mask = mask;
+
+    if (list_empty (&list))
+       elan4_set_schedstatus (dev, intreg);
+    else
+    {
+       dev->dev_haltop_active = active;
+       spin_unlock_irqrestore (&dev->dev_haltop_lock, flags);
+
+       while (! list_empty (&list)) 
+       {
+           ELAN4_HALTOP *op = list_entry (list.next, ELAN4_HALTOP, op_link);
+           
+           list_del (&op->op_link);
+
+           (*op->op_function) (dev, op->op_arg);
+       }
+
+       spin_lock_irqsave (&dev->dev_haltop_lock, flags);
+       dev->dev_haltop_active = 0;
+
+       elan4_set_schedstatus (dev, 0);
+    }
+
+    spin_unlock_irqrestore (&dev->dev_haltop_lock, flags);
+}
+
+static __inline__ E4_uint32
+handle_iproc_trap (ELAN4_DEV *dev, unsigned unit)
+{
+    sdramaddr_t hdroff = dev->dev_inputtraparea + offsetof (E4_IprocTrapState, TrHeader[0][unit]);
+    E4_uint64   status = elan4_sdram_readq (dev, hdroff + offsetof (E4_IprocTrapHeader, IProcStatusCntxAndTrType));
+    E4_uint32   filter = elan4_read_filter (dev, IPROC_NetworkContext (status));
+    ELAN4_CTXT *ctxt   = elan4_localctxt (dev, filter & E4_FILTER_CONTEXT_MASK);
+
+    /*
+     * The context is not valid in the following case :
+     *     ack not been sent AND bad CRC/bad length.
+     *
+     *  NOTE TransCRCStatus and BadLength only valid if NOT an EopTrap.
+     */
+    ASSERT ((IPROC_GoodAckSent (status) & (1 << IPROC_InputterChan (status))) || IPROC_EOPTrap (status) ||
+           (IPROC_TransCRCStatus (status) == CRC_STATUS_GOOD && !IPROC_BadLength (status)));
+    
+    BumpDevStat (dev, s_iproc_traps);
+
+    ctxt->ctxt_ops->op_iproc_trap (ctxt, status, unit);
+
+    return (SCH_RestartCh0LowPriInput << unit);
+}
+
+void
+handle_pcimemerr (ELAN4_DEV *dev)
+{
+    elan4_pcierror (dev);
+
+    check_error_rate (dev);
+}
+
+void
+handle_sdramint (ELAN4_DEV *dev)
+{
+    E4_uint64 status = read_reg64 (dev, SDRamECCStatus);
+    char      errstr[200];
+    int              i;
+    int              Found = 0;
+
+    PRINTF0 (DBG_DEVICE, DBG_INTR, "handle_sdramint\n");
+
+    /* search for this error already being logged */
+    for (i = sizeof (dev->dev_sdramerrs)/sizeof (dev->dev_sdramerrs[0]) - 1; i >= 0; i--)
+        if ((dev->dev_sdramerrs[i] & 0x000fffffffffffffULL) == status)
+       {
+            Found = 1;
+           dev->dev_sdramerrs[i] += 10000000000000ULL; // Keep a count.
+           break;
+       }
+
+    /* stash the status for /proc */
+    if (!Found)
+    {
+       for (i = sizeof (dev->dev_sdramerrs)/sizeof (dev->dev_sdramerrs[0]) - 1; i > 0; i--)
+           dev->dev_sdramerrs[i] = dev->dev_sdramerrs[i-1];
+       dev->dev_sdramerrs[0] = status;
+    }
+
+    printk ("elan%d: ECC Error %s\n", dev->dev_instance, elan4_sdramerr2str (dev, status, errstr));
+
+    if (!ECC_UncorrectableErr(status) && !ECC_MultUncorrectErrs(status))
+       printk ("elan%d: ECC error data=%016llx\n", dev->dev_instance, elan4_sdram_readq (dev, ECC_Addr(status)));
+
+    if (ECC_CorrectableErr (status))
+       BumpDevStat (dev, s_correctable_errors);
+    if (ECC_MultCorrectErrs (status))
+       BumpDevStat (dev, s_multiple_errors);
+
+    if (ECC_UncorrectableErr(status))
+       panic ("elan%d: uncorrectable ECC error\n", dev->dev_instance);
+    if (ECC_MultUncorrectErrs(status))
+       panic ("elan%d: muliple uncorrectable ECC error\n", dev->dev_instance);
+    
+    PULSE_SYSCONTROL (dev, CONT_CLEAR_SDRAM_ERROR);
+
+    check_error_rate (dev);
+}
+
+static void
+clear_linkerr_led (void *arg)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) arg;
+
+    write_i2c (dev, I2cStatus, read_i2c (dev, I2cStatus) | I2cCntl_ClearLinkError);
+}
+
+void
+handle_linkerror (ELAN4_DEV *dev)
+{
+    E4_uint32 LinkState;
+    E4_uint32 CurrState = read_reg32 (dev, LinkControlReg);
+
+    /* Set for reading errors. */
+    write_reg32 (dev, LinkControlReg,
+                 (CurrState = CurrState & ~((LCONT_TEST_CONTROL_MASK << LCONT_TEST_CONTROL_SHIFT) |
+                                           (LCONT_TEST_VALUE_MASK << LCONT_TEST_VALUE_SHIFT))));
+    LinkState = LCONT_LINK_STATE(CurrState = read_reg32 (dev, LinkControlReg));
+
+#ifdef DEBUG
+    {
+       E4_uint8 ErrorMsg[256], DataErrorVal[64];
+
+       strcpy (ErrorMsg, "handle_linkerror:");
+       if (LinkState & LS_LockError)    strcat (ErrorMsg, " LockError");
+       if (LinkState & LS_DeskewError)  strcat (ErrorMsg, " DeskewError");
+       if (LinkState & LS_PhaseError)   strcat (ErrorMsg, " PhaseError");
+       if (LinkState & LS_DataError)
+       {
+           E4_uint32 error[4];
+           E4_uint32 i;
+           strcat (ErrorMsg, " DataError");
+           /* Errors */
+           for(i = LRS_ErrorVal8to0; i <= LRS_ErrorVal35to27; i++)
+           {
+               write_reg32 (dev, LinkControlReg,
+                            CurrState | LCONT_TEST_VALUE(i) | (LCONT_READ_STATE << LCONT_TEST_CONTROL_SHIFT));
+               error[i - LRS_ErrorVal8to0] = LCONT_LINK_STATE(read_reg32 (dev, LinkControlReg));
+           }
+           sprintf (DataErrorVal, " Link State Error Val: %09llx %03x %03x %03x %03x", 
+                    (unsigned long long) ((error[0] & 0x1ffUL) | ((error[1] & 0x1ffUL) << 9)  |
+                                 ((error[2] & 0x1ffUL) << 18) | ((error[3] & 0x1ffUL) << 27)),
+                    error[3], error[2], error[1], error[0]);
+           strcat (ErrorMsg, DataErrorVal);
+       }
+       if (LinkState & LS_FifoOvFlow0)  strcat (ErrorMsg, " FifoOvFlow0");
+       if (LinkState & LS_FifoOvFlow1)  strcat (ErrorMsg, " FifoOvFlow1");
+       if (LinkState & LS_Mod45Changed)         strcat (ErrorMsg, " Mod45Changed");
+       if (LinkState & LS_PAckNotSeenError) strcat (ErrorMsg, " PAckNotSeenError");
+       strcat (ErrorMsg, "\n");
+       PRINTF0 (DBG_DEVICE, DBG_INTR, ErrorMsg);
+    }
+#endif
+
+    BumpDevStat (dev, s_link_errors);
+    
+    if (LinkState & LS_LockError)       BumpDevStat (dev, s_lock_errors);
+    if (LinkState & LS_DeskewError)     BumpDevStat (dev, s_deskew_errors);
+    if (LinkState & LS_PhaseError)      BumpDevStat (dev, s_phase_errors);
+    if (LinkState & LS_DataError)       BumpDevStat (dev, s_data_errors);
+    if (LinkState & LS_FifoOvFlow0)     BumpDevStat (dev, s_fifo_overflow0);
+    if (LinkState & LS_FifoOvFlow1)     BumpDevStat (dev, s_fifo_overflow1);
+    if (LinkState & LS_Mod45Changed)    BumpDevStat (dev, s_mod45changed);
+    if (LinkState & LS_PAckNotSeenError) BumpDevStat (dev, s_pack_not_seen);
+
+    PULSE_SCHED_RESTART (dev, SCH_ClearLinkErrorInt);
+    
+    /* schedule a timer to clear the link error LED, so that it stays on 
+     * for a second for every link error that occurs */
+    if (dev->dev_devinfo.dev_revision_id != PCI_REVISION_ID_ELAN4_REVA && !timer_fn_queued (&dev->dev_linkerr_timeoutid))
+       schedule_timer_fn (&dev->dev_linkerr_timeoutid, clear_linkerr_led, (void *) dev, HZ);
+      
+    check_error_rate (dev);
+}
+
+void
+handle_linkportkeyfail (ELAN4_DEV *dev)
+{
+    PRINTF0 (DBG_DEVICE, DBG_INTR, "handle_linkportkeyfail\n");
+
+    BumpDevStat (dev, s_linkport_keyfail);
+
+    PULSE_SYSCONTROL (dev, CONT_CLEAR_LINKPORT_INT);
+    
+    check_error_rate (dev);
+}
+
+
+static __inline__ void
+__elan4_4msi0 (ELAN4_DEV *dev, E4_uint32 intreg, E4_uint32 intmask)
+{
+    unsigned long flags;
+
+    if (intreg & intmask & INT_MainInterrupt)
+    {
+       DISABLE_INT_MASK (dev, INT_MainInterrupt);
+
+       if (handle_mainints (dev, -1, elan4_mainint_punt_loops) == 0)
+           ENABLE_INT_MASK (dev, INT_MainInterrupt);
+       else
+       {
+           BumpDevStat (dev, s_mainint_punts);
+           
+           spin_lock_irqsave (&dev->dev_mainint_lock, flags);
+           kcondvar_wakeupone (&dev->dev_mainint_wait, &dev->dev_mainint_lock);
+           spin_unlock_irqrestore (&dev->dev_mainint_lock, flags);
+       }
+    }
+}
+
+static __inline__ void
+__elan4_4msi1 (ELAN4_DEV *dev, E4_uint32 intreg, E4_uint32 intmask)
+{
+    E4_uint32 restart = 0;
+
+    PRINTF1 (DBG_DEVICE, DBG_INTR, "__elan4_4msi1: %x\n", intreg);
+    
+    spin_lock (&dev->dev_trap_lock);
+    
+    if (intreg & intmask & INT_CProc)
+       restart |= handle_cproc_trap (dev);
+    if (intreg & intmask & INT_EProc) 
+       restart |= handle_eproc_trap (dev);
+    if (intreg & intmask & INT_Dma0Proc) 
+       restart |= handle_dproc_trap (dev, 0);
+    if (intreg & intmask & INT_Dma1Proc) 
+       restart |= handle_dproc_trap (dev, 1);
+    if (intreg & intmask & INT_TProc)
+       restart |= handle_tproc_trap (dev);
+    
+    PULSE_SCHED_RESTART (dev, restart);
+    
+    spin_unlock (&dev->dev_trap_lock);
+    
+    if (intreg & (INT_Halted|INT_Discarding))
+       handle_haltints (dev, intreg);
+}
+
+static __inline__ void
+__elan4_4msi2 (ELAN4_DEV *dev, E4_uint32 intreg, E4_uint32 intmask)
+{
+    E4_uint32 restart = 0;
+
+    PRINTF1 (DBG_DEVICE, DBG_INTR, "__elan4_4msi2: %x\n", intreg);
+    
+    spin_lock (&dev->dev_trap_lock);
+    if (intreg & intmask & INT_IProcCh0LowPri)
+       restart |= handle_iproc_trap (dev, 0);
+    
+    if (intreg & intmask & INT_IProcCh1LowPri)
+       restart |= handle_iproc_trap (dev, 1);
+    
+    if (intreg & intmask & INT_IProcCh0HighPri)
+       restart |= handle_iproc_trap (dev, 2);
+    
+    if (intreg & intmask & INT_IProcCh1HighPri)
+       restart |= handle_iproc_trap (dev, 3);
+    
+    PULSE_SCHED_RESTART (dev, restart);
+    
+    spin_unlock (&dev->dev_trap_lock);
+}
+
+static __inline__ void
+__elan4_4msi3 (ELAN4_DEV *dev, E4_uint32 intreg, E4_uint32 intmask)
+{
+    PRINTF1 (DBG_DEVICE, DBG_INTR, "__elan4_4msi3: %x\n", intreg);
+    
+    if (intreg & intmask & INT_PciMemErr)
+       handle_pcimemerr (dev);
+    
+    if (intreg & intmask & INT_SDRamInt)
+       handle_sdramint (dev);
+    
+    if (intreg & intmask & INT_LinkError)
+       handle_linkerror (dev);
+    
+    if (intreg & intmask & INT_LinkPortKeyFail)
+       handle_linkportkeyfail (dev);
+}
+
+int
+elan4_1msi0 (ELAN4_DEV *dev)
+{
+    E4_uint32 intmask = dev->dev_intmask;
+    E4_uint32 intreg;
+
+    if (intmask == 0 || ((intreg = read_reg32 (dev, InterruptReg)) & intmask) == 0)
+       return (0);
+
+    BumpDevStat (dev, s_interrupts);
+    
+    do {
+       PRINTF1 (DBG_DEVICE, DBG_INTR, "elan4_1msi0: %x\n", intreg);
+
+       if (intreg & intmask & INT_MSI0)
+           __elan4_4msi0(dev, intreg, intmask);
+       if (intreg & intmask & INT_MSI1)
+           __elan4_4msi1(dev, intreg, intmask);
+       if (intreg & intmask & INT_MSI2)
+           __elan4_4msi2(dev, intreg, intmask); 
+       if (intreg & intmask & INT_MSI3)
+           __elan4_4msi3(dev, intreg, intmask);
+
+       /* must ensure that the read of the interrupt mask
+        * completes before the read of the interrupt register
+        * since the main interrupt thread clears it's interrupt
+        * and then re-enables it in the interrupt mask. */
+       intmask = dev->dev_intmask;
+       mb();
+       intreg = read_reg32 (dev, InterruptReg);
+
+    } while ((intreg & intmask) != 0);
+
+    return (1);
+}
+
+/* local context management */
+int
+elan4_insertctxt (ELAN4_DEV *dev, ELAN4_CTXT *ctxt, ELAN4_TRAP_OPS *ops)
+{
+    unsigned long flags;
+    int tbl;
+
+    ctxt->ctxt_dev = dev;
+    ctxt->ctxt_ops = ops;
+
+    INIT_LIST_HEAD (&ctxt->ctxt_cqalist);
+    spin_lock_init (&ctxt->ctxt_mmulock);
+
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+    {
+       KMEM_ZALLOC (ctxt->ctxt_mmuhash[tbl], ELAN4_HASH_ENTRY **,  dev->dev_hashsize[tbl] * sizeof (ELAN4_HASH_ENTRY *), 1);
+       
+       if (ctxt->ctxt_mmuhash[tbl] == NULL)
+       {
+           if (tbl != 0)
+               KMEM_FREE (ctxt->ctxt_mmuhash[0], dev->dev_hashsize[0] * sizeof (ELAN4_HASH_ENTRY *));
+           spin_lock_destroy (&ctxt->ctxt_mmulock);
+           return (-ENOMEM);
+       }
+    }
+
+    spin_lock_irqsave (&dev->dev_ctxt_lock, flags);
+
+    if ((ctxt->ctxt_num = bt_freebit (dev->dev_ctxmap, (1 << dev->dev_ctxtableshift))) >= 0)
+    {
+       /* chain onto the lists of all contexts */
+       list_add (&ctxt->ctxt_link, &dev->dev_ctxt_list);
+
+       BT_SET (dev->dev_ctxmap, ctxt->ctxt_num);
+    }
+    
+    spin_unlock_irqrestore (&dev->dev_ctxt_lock, flags);
+
+    return (ctxt->ctxt_num < 0 ? -ENOMEM : 0);
+}
+
+void
+elan4_removectxt (ELAN4_DEV *dev, ELAN4_CTXT *ctxt)
+{
+    unsigned long flags;
+    int tbl;
+
+    /* remove from list of contexts */
+    spin_lock_irqsave (&dev->dev_ctxt_lock, flags);
+
+    list_del (&ctxt->ctxt_link);
+
+    BT_CLEAR (dev->dev_ctxmap, ctxt->ctxt_num);
+
+    spin_unlock_irqrestore (&dev->dev_ctxt_lock, flags);
+
+    spin_lock_destroy (&ctxt->ctxt_info_lock);
+
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+       KMEM_FREE (ctxt->ctxt_mmuhash[tbl],  dev->dev_hashsize[tbl] * sizeof (ELAN4_HASH_ENTRY *));
+
+    spin_lock_destroy (&ctxt->ctxt_mmulock);
+}
+
+ELAN4_CTXT *
+elan4_localctxt (ELAN4_DEV *dev, unsigned num)
+{
+    struct list_head *entry;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_ctxt_lock, flags);
+
+    list_for_each (entry, &dev->dev_ctxt_list) {
+       ELAN4_CTXT *ctxt = list_entry (entry, ELAN4_CTXT, ctxt_link);
+       
+       if (ctxt->ctxt_num == num)
+       {
+           spin_unlock_irqrestore (&dev->dev_ctxt_lock, flags);
+           return (ctxt);
+       }
+    }
+    spin_unlock_irqrestore (&dev->dev_ctxt_lock, flags);
+
+    return ((ELAN4_CTXT *) NULL);
+}
+
+ELAN4_CTXT *
+elan4_networkctxt (ELAN4_DEV *dev, unsigned num)
+{
+    E4_uint32 filter = elan4_read_filter (dev, num);
+    
+    if ((filter & E4_FILTER_CONTEXT_MASK) == INVALID_CONTEXT)
+       return NULL;
+    else
+       return elan4_localctxt (dev, filter & E4_FILTER_CONTEXT_MASK);
+}
+
+/* network context management */
+int
+elan4_attach_filter (ELAN4_CTXT *ctxt, unsigned int ctxnum)
+{
+    ELAN4_DEV        *dev = ctxt->ctxt_dev;
+    int               res = 0;
+    E4_uint32         filter;
+    unsigned long     flags;
+    
+    spin_lock_irqsave (&dev->dev_ctxt_lock, flags);
+    
+    filter = elan4_read_filter (dev, ctxnum);
+    if ((filter & E4_FILTER_CONTEXT_MASK) != INVALID_CONTEXT)
+    {
+       PRINTF2 (ctxt, DBG_NETWORK_CTX, "elan4_attach_filter: ctx=%d filter=%x -> EBUSY\n", ctxnum, filter);
+       res = -EBUSY;
+    }
+    else
+    {
+       PRINTF1 (ctxt, DBG_NETWORK_CTX, "elan4_attach_filter: ctx=%d - SUCCESS\n", ctxnum);
+
+       elan4_write_filter (dev, ctxnum, ctxt->ctxt_num | E4_FILTER_DISCARD_ALL);
+       PULSE_SCHED_RESTART (dev, SCH_ContextFilterFlush);
+    }
+    spin_unlock_irqrestore (&dev->dev_ctxt_lock, flags);
+    
+    return (res);
+}
+
+void
+elan4_detach_filter (ELAN4_CTXT *ctxt, unsigned int ctxnum)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+
+    PRINTF1 (ctxt, DBG_NETWORK_CTX, "elan4_detach_filter: detach from network context %d\n", ctxnum);
+           
+    elan4_write_filter (dev, ctxnum, INVALID_CONTEXT | E4_FILTER_DISCARD_ALL);
+    PULSE_SCHED_RESTART (dev, SCH_ContextFilterFlush);
+}
+
+void
+elan4_set_filter (ELAN4_CTXT *ctxt, unsigned int ctxnum, E4_uint32 state)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+
+    PRINTF6 (ctxt, DBG_NETWORK_CTX, "elan4_set_filter: set filter state %x for network context %d <%s%s%s%s>\n", state, ctxnum,
+            (state & E4_FILTER_DISCARD_ALL) ? "discard,"  : "",
+            (state & E4_FILTER_ACKOK_ALL)   ? "ack-ok,"   : "",
+            (state & E4_FILTER_HIGH_PRI)    ? "high-pri," : "",
+            (state & E4_FILTER_STATS)       ? "stats,"    : "");
+           
+    elan4_write_filter (dev, ctxnum, ctxt->ctxt_num | state);
+    PULSE_SCHED_RESTART (dev, SCH_ContextFilterFlush);
+}
+
+void
+elan4_set_routetable (ELAN4_CTXT *ctxt, ELAN4_ROUTE_TABLE *tbl)
+{
+    ELAN4_DEV *dev   = ctxt->ctxt_dev;
+    E4_uint32  value = tbl ? (E4_VPT_VALID | E4_VPT_VALUE(tbl->tbl_entries, tbl->tbl_size)) : 0;
+    
+    /* and insert into the vp table */
+    elan4_sdram_writel (dev, (dev->dev_ctxtable + (ctxt->ctxt_num * sizeof (E4_ContextControlBlock)) +
+                       offsetof (E4_ContextControlBlock, VirtualProcessTable)), value);
+    pioflush_sdram(dev);
+
+    PULSE_SYSCONTROL (dev, CONT_ROUTE_FLUSH);
+}
+
+/* command queue management */
+ELAN4_CQA *
+elan4_getcqa (ELAN4_CTXT *ctxt, unsigned int idx)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+    struct list_head *el;
+
+    spin_lock (&dev->dev_cqlock);
+    list_for_each (el, &ctxt->ctxt_cqalist) {
+       ELAN4_CQA *cqa = list_entry (el, ELAN4_CQA, cqa_link);
+       
+       if (cqa->cqa_idx == idx)
+       {
+           cqa->cqa_ref++;
+
+           spin_unlock (&dev->dev_cqlock);
+           return cqa;
+       }
+    }
+    spin_unlock (&dev->dev_cqlock);
+    return NULL;
+}
+
+void
+elan4_putcqa (ELAN4_CTXT *ctxt, unsigned int idx)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+    struct list_head *el, *nel;
+
+    spin_lock (&dev->dev_cqlock);
+    list_for_each_safe (el, nel, &ctxt->ctxt_cqalist) {
+       ELAN4_CQA *cqa = list_entry (el, ELAN4_CQA, cqa_link);
+       
+       if (cqa->cqa_idx == idx)
+       {
+           if (--cqa->cqa_ref || bt_lowbit (cqa->cqa_bitmap, ELAN4_CQ_PER_CQA) != -1)
+               spin_unlock (&dev->dev_cqlock);
+           else
+           {
+               list_del (&cqa->cqa_link);
+               
+               BT_CLEAR (ctxt->ctxt_cqamap, cqa->cqa_idx);
+               BT_CLEAR (dev->dev_cqamap, cqa->cqa_cqnum/ELAN4_CQ_PER_CQA);
+               spin_unlock (&dev->dev_cqlock);
+               
+               KMEM_FREE (cqa, sizeof (ELAN4_CQA));
+           }
+           return;
+       }
+    }
+    spin_unlock (&dev->dev_cqlock);
+
+    printk ("elan4_putcqa: idx %d not found\n", idx);
+    BUG();
+}
+
+static ELAN4_CQ *
+elan4_getcq (ELAN4_CTXT *ctxt, unsigned int type)
+{
+    ELAN4_DEV        *dev = ctxt->ctxt_dev;
+    ELAN4_CQA        *cqa;
+    struct list_head *el;
+    int                      cidx, didx;
+
+    spin_lock (&dev->dev_cqlock);
+    list_for_each (el, &ctxt->ctxt_cqalist) {
+       cqa = list_entry (el, ELAN4_CQA, cqa_link);
+
+       if (cqa->cqa_type == type && (cidx = bt_freebit (cqa->cqa_bitmap, ELAN4_CQ_PER_CQA)) >=0)
+       {
+           BT_SET (cqa->cqa_bitmap, cidx);
+           
+           spin_unlock (&dev->dev_cqlock);
+           return &cqa->cqa_cq[cidx];
+       }
+    }
+    spin_unlock (&dev->dev_cqlock);
+
+    /* allocate a new cqa and it's chunk of command queue descriptors */
+    KMEM_ZALLOC (cqa, ELAN4_CQA *, sizeof (ELAN4_CQA), 1);
+    if (cqa == NULL)
+       return NULL;
+
+    spin_lock (&dev->dev_cqlock);
+    cidx = bt_freebit (ctxt->ctxt_cqamap, ELAN4_MAX_CQA);
+
+    /* On architectures which have MTRR registers for write-combinig
+     * the top command queues from dev->dev_cqreorder upwards are
+     * used for reordered queues.  Without MTRR registers any page
+     * sized group can use write combinig through the ptes. */
+#ifdef CONFIG_MTRR
+    if ((type & CQ_Reorder) != 0)
+       didx = bt_nextbit (dev->dev_cqamap, dev->dev_cqcount, dev->dev_cqreorder - 1, 0);
+    else
+       didx = bt_freebit (dev->dev_cqamap, dev->dev_cqreorder);
+#else
+    didx = bt_freebit (dev->dev_cqamap, dev->dev_cqcount);
+#endif
+
+    if (cidx < 0 || didx < 0)
+    {
+       spin_unlock (&dev->dev_cqlock);
+       KMEM_FREE (cqa, sizeof (ELAN4_CQA));
+       return NULL;
+    }
+
+    BT_SET (ctxt->ctxt_cqamap, cidx);
+    BT_SET (dev->dev_cqamap, didx);
+
+    cqa->cqa_idx   = cidx;
+    cqa->cqa_type  = type;
+    cqa->cqa_cqnum = (didx * ELAN4_CQ_PER_CQA);
+    
+    list_add_tail (&cqa->cqa_link, &ctxt->ctxt_cqalist);
+    
+    /* initialise the cqa struct */
+    for (cidx = 0; cidx < ELAN4_CQ_PER_CQA; cidx++)
+    {
+       cqa->cqa_cq[cidx].cq_idx = cidx;
+       cqa->cqa_cq[cidx].cq_cqa = cqa;
+    }
+
+    /* no mappings yet */
+    cqa->cqa_ref = 0;
+
+    /* we're going to return entry zero */
+    BT_SET (cqa->cqa_bitmap, 0);
+    spin_unlock (&dev->dev_cqlock);
+    
+    return &cqa->cqa_cq[0];
+}
+
+static void
+elan4_putcq (ELAN4_CTXT *ctxt, ELAN4_CQ *cq)
+{
+    ELAN4_DEV        *dev = ctxt->ctxt_dev;
+    ELAN4_CQA        *cqa = cq->cq_cqa;
+
+    spin_lock (&dev->dev_cqlock);
+
+    BT_CLEAR (cqa->cqa_bitmap, cq->cq_idx);
+
+    if (bt_lowbit (cqa->cqa_bitmap, ELAN4_CQ_PER_CQA) != -1 || cqa->cqa_ref)
+       spin_unlock (&dev->dev_cqlock);
+    else
+    {
+       list_del (&cqa->cqa_link);
+       
+       BT_CLEAR (ctxt->ctxt_cqamap, cqa->cqa_idx);
+       BT_CLEAR (dev->dev_cqamap, cqa->cqa_cqnum/ELAN4_CQ_PER_CQA);
+       spin_unlock (&dev->dev_cqlock);
+       
+       KMEM_FREE (cqa, sizeof (ELAN4_CQA));
+    }
+}
+
+ELAN4_CQ *
+elan4_alloccq (ELAN4_CTXT *ctxt, unsigned cqsize, unsigned perm, unsigned cqtype)
+{
+    ELAN4_DEV   *dev = ctxt->ctxt_dev;
+    ELAN4_CQ    *cq;
+    int         cqnum;
+    sdramaddr_t cqdesc;
+    unsigned    offset;
+    E4_uint64   value;
+
+    if ((cq = elan4_getcq (ctxt, cqtype)) == NULL)
+       return NULL;
+
+    cqnum = elan4_cq2num(cq);
+    
+    cq->cq_space = elan4_sdram_alloc (dev, CQ_Size(cqsize));
+    if (cq->cq_space == (virtaddr_t) 0)
+    {
+       elan4_putcq (ctxt, cq);
+       return (NULL);
+    }
+
+    cq->cq_size   = cqsize;
+    cq->cq_perm   = perm;
+    
+    /* and finally initialise the command queue descriptor */
+    cqdesc = dev->dev_cqaddr + (cqnum * sizeof (E4_CommandQueueDesc));
+
+    value  = CQ_QueuePtrsValue (cqsize, cq->cq_space, cq->cq_space);
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+       value |= ((cqtype & CQ_Priority) ? CQ_RevA_Priority : 0);
+    else
+       value |= (((cqtype & CQ_Priority) ? CQ_RevB_Priority : 0) | 
+                 ((cqtype & CQ_Reorder)  ? CQ_RevB_ReorderingQueue : CQ_RevB_32bitWriteQueue));
+
+    elan4_sdram_writeq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs), value);
+    elan4_sdram_writeq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_HoldingValue), 0);
+    elan4_sdram_writeq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_AckBuffers), 0);
+    elan4_sdram_writeq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_Control), CQ_ControlValue (ctxt->ctxt_num, 2, perm));
+    pioflush_sdram (dev);
+
+    offset = (cqnum + dev->dev_cqoffset) * CQ_CommandMappingSize;
+
+    cq->cq_mapping = elan4_map_device (dev, ELAN4_BAR_REGISTERS, (offset & ~(PAGE_SIZE-1)), 
+                                      PAGE_SIZE, &cq->cq_handle) + (offset & (PAGE_SIZE-1));
+#ifdef CONFIG_MPSAS
+    if (ctxt == &dev->dev_ctxt)
+       return (cq);
+#endif
+
+    elan4_sdram_flushcache (dev, cq->cq_space, CQ_Size(cqsize));
+
+    return (cq);
+}
+    
+void
+elan4_freecq (ELAN4_CTXT *ctxt, ELAN4_CQ *cq)
+{
+    ELAN4_DEV *dev    = ctxt->ctxt_dev;
+    unsigned   offset = (elan4_cq2num(cq) + dev->dev_cqoffset) * CQ_CommandMappingSize;
+
+    elan4_flushcq (dev, cq);
+
+    elan4_unmap_device (dev, cq->cq_mapping - (offset & (PAGE_SIZE-1)), PAGE_SIZE, &cq->cq_handle);
+    elan4_sdram_free (dev, cq->cq_space, CQ_Size (cq->cq_size));
+
+    elan4_putcq (ctxt, cq);
+}
+
+void
+elan4_restartcq (ELAN4_DEV *dev, ELAN4_CQ *cq)
+{
+    sdramaddr_t   cqdesc = dev->dev_cqaddr + (elan4_cq2num(cq) * sizeof (E4_CommandQueueDesc));
+    int           hipri;
+    unsigned long flags;
+    
+    PRINTF1 (DBG_DEVICE, DBG_CPROC, "restartcq: restarting cq %p\n", cq);
+    
+    spin_lock_irqsave (&dev->dev_requeue_lock, flags);
+
+    while (read_reg32 (dev, CommandControl.CommandRequeuePtr) & E4_CommandRequeueBusy)
+       ;
+    
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+       hipri = (elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs)) & CQ_RevA_Priority) != 0;
+    else
+       hipri = (elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs)) & CQ_RevB_Priority) != 0;
+    
+    if (hipri)
+    {
+       PRINTF1 (DBG_DEVICE, DBG_CPROC, "restartcq: restart cq %d as high pri\n", elan4_cq2num(cq));
+       write_reg32 (dev, CommandControl.CommandRequeuePtr, cqdesc | E4_CommandRequeueHighPri);
+    }
+    else
+    {
+       PRINTF1 (DBG_DEVICE, DBG_CPROC, "restartcq: restart cq %d as low pri\n", elan4_cq2num(cq));
+       write_reg32 (dev, CommandControl.CommandRequeuePtr, cqdesc);
+    }
+    pioflush_reg (dev);
+    
+    spin_unlock_irqrestore (&dev->dev_requeue_lock, flags);
+}
+
+static void
+flushcq_intop (ELAN4_DEV *dev, void *arg)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_flush_lock, flags);
+    dev->dev_flush_finished |= (1 << (unsigned long) arg);
+    kcondvar_wakeupall (&dev->dev_flush_wait, &dev->dev_flush_lock);
+    spin_unlock_irqrestore (&dev->dev_flush_lock, flags);
+}
+void
+elan4_flushcq (ELAN4_DEV *dev, ELAN4_CQ *cq)
+{
+    int                  flushqnum = elan4_cq2num(cq) & (COMMAND_INSERTER_CACHE_ENTRIES-1);
+    ELAN4_CQ     *flushq    = dev->dev_flush_cq[flushqnum];
+    unsigned long flags;
+
+    PRINTF (DBG_DEVICE, DBG_FLUSH, "elan4_flushcq: cqnum=%d\n", elan4_cq2num(cq));
+
+    spin_lock_irqsave (&dev->dev_flush_lock, flags);
+
+    while (! (dev->dev_flush_finished & (1 << flushqnum)))
+       kcondvar_wait (&dev->dev_flush_wait, &dev->dev_flush_lock, &flags);
+    
+    dev->dev_flush_finished &= ~(1 << flushqnum);
+
+    dev->dev_flush_op[flushqnum].op_function = flushcq_intop;
+    dev->dev_flush_op[flushqnum].op_arg      = (void *) (unsigned long) flushqnum;
+    
+    elan4_queue_intop (dev, flushq, &dev->dev_flush_op[flushqnum]);
+
+    while (! (dev->dev_flush_finished & (1 << flushqnum)))
+       kcondvar_wait (&dev->dev_flush_wait, &dev->dev_flush_lock, &flags);
+    
+    spin_unlock_irqrestore (&dev->dev_flush_lock, flags);
+}
+
+void
+elan4_updatecq (ELAN4_DEV *dev, ELAN4_CQ *cq, unsigned perm, unsigned restart)
+{
+    sdramaddr_t cqdesc  = dev->dev_cqaddr + (elan4_cq2num(cq) * sizeof (E4_CommandQueueDesc));
+    E4_uint32   control = elan4_sdram_readl (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_Control));
+
+    /* Write the command queues control word, but ensure that the ChannelNotCompleted fields
+     * are not modified.   We use this to just alter the RestartCount/Permissions fields */
+
+    elan4_sdram_writel (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_Control), 
+                       CQ_ControlValue (CQ_Context (control), restart ? restart : CQ_RestartCount (control), perm));
+}
+
+/* instruction cache flush */
+static __inline__ void
+elan4_flush_icache_locked (ELAN4_DEV *dev)
+{
+    int i, j;
+
+    PRINTF0 (DBG_DEVICE, DBG_FLUSH, "elan4_flush_icache_locked: flushing icache\n");
+
+    for (i = 0; i < (E4_ICacheLines/E4_ICachePortSize); i++)
+    {
+        write_reg64 (dev, ICachePort_Cntl_Addr, i << E4_ICacheTagAddrShift);
+        for (j = 0; j < E4_ICachePortSize; j++)
+           write_reg64 (dev, ICachePort[j], E4_InvalidTagValue);
+    }
+
+    /*
+     * Initialise the top of the ICache Set0 with a instruction which will
+     * cause a know trap fingerprint so that the application can identify it
+     * and ignore the trap.
+     */
+    write_reg64 (dev, ICachePort_Cntl_Addr, E4_ICacheFixupOffset | E4_AccessICacheRams);
+
+    /* Errata 24: must ensure that the DCache is flushed after loading 
+     *            code for the thread processor. */
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+       elan4_sdram_flushcache (dev, 0, E4_CacheSize);
+
+    pioflush_reg (dev);
+}
+
+static void
+device_iflush_haltop (ELAN4_DEV *dev, void *arg)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_flush_lock, flags);
+    
+    elan4_flush_icache_locked (dev);
+
+    dev->dev_iflush_queued = 0;
+
+    kcondvar_wakeupall (&dev->dev_flush_wait, &dev->dev_flush_lock);
+    spin_unlock_irqrestore (&dev->dev_flush_lock, flags);
+}
+
+void
+elan4_flush_icache_halted (ELAN4_CTXT *ctxt)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_flush_lock, flags);
+    
+    elan4_flush_icache_locked (dev);
+
+    spin_unlock_irqrestore (&dev->dev_flush_lock, flags);
+}
+
+void
+elan4_flush_icache (ELAN4_CTXT *ctxt)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&dev->dev_flush_lock, flags);
+
+    PRINTF1 (DBG_DEVICE, DBG_FLUSH, "elan4_flush_icache: queued=%d\n", dev->dev_iflush_queued);
+
+    if (! dev->dev_iflush_queued)
+    {
+       dev->dev_iflush_queued = 1;
+       
+       elan4_queue_haltop (dev, &dev->dev_iflush_haltop);
+    }
+
+    while (dev->dev_iflush_queued)
+       kcondvar_wait (&dev->dev_flush_wait, &dev->dev_flush_lock, &flags);
+
+    spin_unlock_irqrestore (&dev->dev_flush_lock, flags);
+}
+
+/* device context operations */
+static void
+device_cproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned cqnum)
+{
+    ELAN4_DEV        *dev  = ctxt->ctxt_dev;
+    ELAN4_CPROC_TRAP *trap = &dev->dev_cproc_trap;
+
+    elan4_extract_cproc_trap (dev, status, trap, cqnum);
+
+    DBGCMD (DBG_DEVICE, DBG_FLUSH, elan4_display_cproc_trap (DBG_DEVICE, DBG_FLUSH, "device_cproc_trap", trap));
+
+    switch (CPROC_TrapType (trap->tr_status))
+    {
+    case CommandProcInterruptQueueOverflow:
+       PRINTF (ctxt, DBG_FLUSH, "device_cproc_trap: cqnum=%d\n", cqnum);
+
+       /* XXXX: we could either just hit restart (and hope) - or we could extract
+        *       the event interrupt cookie out and "complete" the command before
+        *       restarting it */
+       elan4_restartcq (dev, dev->dev_flush_cq[cqnum]);
+       return;
+
+    case CommandProcDmaQueueOverflow:
+    case CommandProcPermissionTrap:
+       handle_dma_flushops (dev, status, cqnum);
+       return;
+       
+    default:
+       printk ("device_cproc_trap: status=%llx control=%llx TrapType=%x cqnum=%d\n", (long long) trap->tr_status,
+               elan4_sdram_readq (dev, dev->dev_cqaddr + cqnum * sizeof (E4_CommandQueueDesc) +
+                                  offsetof (E4_CommandQueueDesc, CQ_Control)),
+               (int) CPROC_TrapType(trap->tr_status), cqnum);
+       panic ("device_cproc_trap");
+    }
+}
+
+static void
+device_tproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status)
+{
+    ELAN4_TPROC_TRAP trap;
+
+    elan4_extract_tproc_trap (ctxt->ctxt_dev, status, &trap);
+
+    elan4_display_tproc_trap (DBG_CONSOLE, DBG_TRAP, "device_tproc_trap", &trap);
+    panic ("device_tproc_trap");
+}
+
+static void
+device_dproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit)
+{
+    ELAN4_DPROC_TRAP trap;
+
+    elan4_extract_dproc_trap (ctxt->ctxt_dev, status, &trap, unit);
+
+    elan4_display_dproc_trap (DBG_CONSOLE, DBG_TRAP, "device_dproc_trap", &trap);
+    panic ("device_dproc_trap");
+}
+
+static void
+device_interrupt (ELAN4_CTXT *ctxt, E4_uint64 cookie)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) ctxt;
+    struct list_head *el,*nel;
+    unsigned long flags;
+
+    PRINTF (ctxt, DBG_FLUSH, "device_interrupt: cookie=%llx\n", cookie);
+
+    spin_lock_irqsave (&dev->dev_intop_lock, flags);
+    list_for_each_safe (el, nel, &dev->dev_intop_list) {
+       ELAN4_INTOP *op = list_entry (el, ELAN4_INTOP, op_link);
+
+       if (op->op_cookie == cookie)
+       {
+           if ((op->op_cookie & INTOP_TYPE_MASK) == INTOP_ONESHOT)
+               list_del (&op->op_link);
+
+           spin_unlock_irqrestore (&dev->dev_intop_lock, flags);
+           
+           (*op->op_function)(dev, op->op_arg);
+           return;
+       }
+    }
+    spin_unlock_irqrestore (&dev->dev_intop_lock, flags);
+
+    panic ("device_interrupt: interrupt cookie %llx not found\n", cookie);
+}
+
+static void
+device_iproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+    ELAN4_IPROC_TRAP *trap = &dev->dev_iproc_trap;
+
+    elan4_extract_iproc_trap (dev, status, trap, unit);
+    elan4_inspect_iproc_trap (trap);
+
+    DBGCMD (ctxt, DBG_IPROC, elan4_display_iproc_trap (ctxt, DBG_IPROC, "device_iproc_trap", trap));
+
+    if (elan4_neterr_iproc_trap (dev, trap))
+       return;
+
+    elan4_display_iproc_trap (DBG_CONSOLE, DBG_TRAP, "device_iproc_trap", trap);
+    panic ("device_iproc_trap: unexpected trap\n");
+}
+
+ELAN4_TRAP_OPS device_trap_ops = 
+{
+    NULL,
+    device_cproc_trap,
+    device_dproc_trap,
+    device_tproc_trap,
+    device_iproc_trap,
+    device_interrupt,
+};
+
+/*
+ * elan4_initialise_device
+ *    initialise the ELAN4_DEV struct - spinlocks,cvs etc.
+ *    map the registers, sdram etc
+ */
+int
+elan4_initialise_device (ELAN4_DEV *dev)
+{
+    int i, bit;
+
+    if (elan4_mainint_resched_ticks == 0)
+       elan4_mainint_resched_ticks = (hz/4);
+
+    /* map the registers */
+    switch (dev->dev_devinfo.dev_revision_id)
+    {
+    case PCI_REVISION_ID_ELAN4_REVA:
+       dev->dev_regs = elan4_map_device (dev, ELAN4_BAR_REGISTERS, ELAN4_REVA_REG_OFFSET, ELAN4_REG_SIZE, &dev->dev_regs_handle);
+       
+       dev->dev_rom  = elan4_map_device (dev, ELAN4_BAR_REGISTERS, ELAN4_REVA_EBUS_OFFSET + ELAN4_REVA_EBUS_ROM_OFFSET, 
+                                         ELAN4_REVA_EBUS_ROM_SIZE, &dev->dev_rom_handle);
+       break;
+       
+    case PCI_REVISION_ID_ELAN4_REVB:
+       dev->dev_regs = elan4_map_device (dev, ELAN4_BAR_REGISTERS, ELAN4_REVB_REG_OFFSET, ELAN4_REG_SIZE, &dev->dev_regs_handle);
+       dev->dev_rom  = (ioaddr_t) 0;
+       dev->dev_i2c  = elan4_map_device (dev, ELAN4_BAR_REGISTERS, ELAN4_REVB_I2C_OFFSET, ELAN4_REVB_I2C_SIZE, &dev->dev_i2c_handle);
+       break;
+
+    default:
+       return -EINVAL;
+    }
+
+    /* XXXX: parse the ebus rom to determine the sdram configuration */
+    {
+       extern long long       sdram_cfg;
+
+       if (sdram_cfg == 0)
+           dev->dev_sdram_cfg = SDRAM_STARTUP_VALUE;
+       else
+           dev->dev_sdram_cfg = sdram_cfg;
+    }
+
+    for (bit = 0; ((1 << bit) & elan4_resource_len (dev, ELAN4_BAR_SDRAM)) == 0; bit++)
+       ;
+
+    switch ((dev->dev_sdram_cfg >> SDRAM_RamSize_SH) & 3)
+    {
+    case 0:                    /* 64Mbit, 128Mbit, 256Mbit, 512Mbit or 1Gbit (16-bit output) */
+       dev->dev_sdram_numbanks = 4; bit -= 2;
+       for (i = 0; i < dev->dev_sdram_numbanks; i++)
+       {
+           dev->dev_sdram_banks[i].b_base = (i << bit);
+           dev->dev_sdram_banks[i].b_size = (1 << bit);
+       }
+       break;
+
+    case 1:                    /*  64Mbit, 128Mbit, 256Mbit or 512Mbit (8-bit output) */
+       dev->dev_sdram_numbanks = 4; bit -= 2;
+       for (i = 0; i < dev->dev_sdram_numbanks; i++)
+       {
+           dev->dev_sdram_banks[i].b_base = ((i & 2) << (bit)) | ((i & 1) << (bit-1));
+           dev->dev_sdram_banks[i].b_size = (1 << bit);
+       }
+       break;
+       
+    case 2:                    /* 2Gbit (16-bit output) or 1Gbit (8-bit output) */
+       dev->dev_sdram_numbanks = 2; bit--;
+       for (i = 0; i < dev->dev_sdram_numbanks; i++)
+       {
+           dev->dev_sdram_banks[i].b_base = (i << bit);
+           dev->dev_sdram_banks[i].b_size = (1 << bit);
+       }
+       break;
+
+    case 3:                    /* 4Gbit (16-bit output) or 2Gbit (8-bit output) */
+       dev->dev_sdram_numbanks = 1;
+       dev->dev_sdram_banks[0].b_base = 0;
+       dev->dev_sdram_banks[0].b_size = (1 << bit);
+       break;
+    }
+
+    elan4_sdram_init (dev);
+
+    /* initialise locks for classes of interrupts */
+    spin_lock_init (&dev->dev_trap_lock);
+    spin_lock_init (&dev->dev_intop_lock);
+    spin_lock_init (&dev->dev_haltop_lock);
+    spin_lock_init (&dev->dev_mainint_lock);
+
+    /* initialise other locks */
+    spin_lock_init (&dev->dev_i2c_lock);
+
+    spin_lock_init (&dev->dev_mmulock);
+    spin_lock_init (&dev->dev_cqlock);
+    spin_lock_init (&dev->dev_ctxlock);
+
+    spin_lock_init (&dev->dev_intmask_lock);
+    spin_lock_init (&dev->dev_syscontrol_lock);
+
+    spin_lock_init (&dev->dev_ctxt_lock);
+    spin_lock_init (&dev->dev_flush_lock);
+    spin_lock_init (&dev->dev_requeue_lock);
+
+    kmutex_init (&dev->dev_lock);
+
+    kcondvar_init (&dev->dev_mainint_wait);
+    kcondvar_init (&dev->dev_flush_wait);
+
+    /* initialsie lists */
+    INIT_LIST_HEAD (&dev->dev_ctxt_list);
+    INIT_LIST_HEAD (&dev->dev_intop_list);
+    INIT_LIST_HEAD (&dev->dev_interruptq_list);
+    INIT_LIST_HEAD (&dev->dev_hc_list);
+    INIT_LIST_HEAD (&dev->dev_haltop_list);
+    INIT_LIST_HEAD (&dev->dev_dma_flushop[0].list);
+    INIT_LIST_HEAD (&dev->dev_dma_flushop[1].list);
+
+    dev->dev_state = ELAN4_STATE_STOPPED;
+
+    return (0);
+}
+
+void
+elan4_finalise_device (ELAN4_DEV *dev)
+{
+    kcondvar_destroy (&dev->dev_flush_wait);
+    kcondvar_destroy (&dev->dev_mainint_wait);
+
+    kmutex_destroy (&dev->dev_lock);
+
+    spin_lock_destroy (&dev->dev_requeue_lock);
+    spin_lock_destroy (&dev->dev_flush_lock);
+    spin_lock_destroy (&dev->dev_ctxt_lock);
+
+    spin_lock_destroy (&dev->dev_syscontrol_lock);
+    spin_lock_destroy (&dev->dev_intmask_lock);
+
+    spin_lock_destroy (&dev->dev_ctxlock);
+    spin_lock_destroy (&dev->dev_cqlock);
+    spin_lock_destroy (&dev->dev_mmulock);
+
+    spin_lock_destroy (&dev->dev_i2c_lock);
+
+    spin_lock_destroy (&dev->dev_mainint_lock);
+    spin_lock_destroy (&dev->dev_haltop_lock);
+    spin_lock_destroy (&dev->dev_intop_lock);
+    spin_lock_destroy (&dev->dev_trap_lock);
+
+    while (! list_empty (&dev->dev_hc_list))
+    {
+       ELAN4_HASH_CHUNK *hc = list_entry (dev->dev_hc_list.next, ELAN4_HASH_CHUNK, hc_link);
+       
+       list_del (&hc->hc_link);
+
+       KMEM_FREE(hc, sizeof (ELAN4_HASH_CHUNK));
+    }
+    
+    elan4_sdram_fini (dev);
+    
+    switch (dev->dev_devinfo.dev_revision_id)
+    {
+    case PCI_REVISION_ID_ELAN4_REVA:
+       elan4_unmap_device (dev, dev->dev_rom,  ELAN4_REVA_EBUS_ROM_SIZE, &dev->dev_rom_handle);
+       elan4_unmap_device (dev, dev->dev_regs, ELAN4_REG_SIZE, &dev->dev_regs_handle);
+       break;
+    case PCI_REVISION_ID_ELAN4_REVB:
+       elan4_unmap_device (dev, dev->dev_i2c,  ELAN4_REVB_I2C_SIZE, &dev->dev_i2c_handle);
+       elan4_unmap_device (dev, dev->dev_regs, ELAN4_REG_SIZE, &dev->dev_regs_handle);
+       break;
+    }
+}
+
+static void
+initialise_cache (ELAN4_DEV *dev)
+{
+    register int set, line;
+
+    /* Initialise the cache to "map" the bottom of sdram - we will use
+     * this space for cache flushing, so require the cache to be set
+     * up so that cachelines for this are in the correct set.
+     *
+     * XXXX: for MPSAS we set bit 28, to ensure that any access to 
+     *       sdram causes the line to be filled first to expunge any
+     *       Xs. */
+    for (set = 0; set < E4_NumCacheSets; set++)
+       for (line = 0; line < E4_NumCacheLines; line++)
+           write_tag (dev, Tags[set][line], (((E4_uint64) set) << 29) | (1 << 28) | (line << 16));
+}
+
+#ifndef CONFIG_MPSAS
+static void
+initialise_cache_tags (ELAN4_DEV *dev, unsigned addr)
+{
+    register int set, line;
+
+    /* Initialise the whole cache to hold sdram at "addr" as direct mapped */
+
+    for (set = 0; set < E4_NumCacheSets; set++)
+       for (line = 0; line < E4_NumCacheLines; line++)
+           write_tag (dev, Tags[set][line], addr | (set << 13) | (1 << 11));
+}
+
+static void
+initialise_ecc (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank)
+{
+    register int i, addr;
+
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+    {
+        initialise_cache_tags (dev, E4_CacheSize);
+        for (addr = 0; addr < bank->b_size; addr += E4_CacheSize)
+        {
+           for (i = 0; i < E4_CacheSize; i += sizeof (E4_uint64))
+               writeq (0xbeec000000000000ull | addr | i, bank->b_ioaddr + addr + i);
+           initialise_cache_tags (dev, addr);
+        }
+    }
+    else
+    {
+       /* Write the whole of this bank of sdram. */
+        for (addr = 0; addr < bank->b_size; addr += sizeof (E4_uint64))
+           writeq (0xbeec000000000000ull | addr, bank->b_ioaddr + addr);
+
+       /* Now flush out the top out of the cache */
+        for (addr = 0; addr < E4_CacheSize; addr += sizeof (E4_uint64))
+           writeq (0xbeec000000000000ull | addr, bank->b_ioaddr + addr);
+
+       /* Now read the top value of sdram to guarantee the write has occured before the ecc is enabled */
+       readq (bank->b_ioaddr + bank->b_size - sizeof (E4_uint64));
+    }
+}
+#endif
+
+#ifdef CONFIG_MPSAS
+static void
+do_initdma (ELAN4_DEV *dev)
+{
+#define VIRTUAL_ADDRESS        0x10000000ull
+    ELAN4_CQ  *cq  = dev->dev_flush_cq[0];
+    E4_uint64 value;
+    E4_uint32 intreg;
+    E4_uint64 status;
+
+    PRINTF (DBG_DEVICE, DBG_CONFIG, "elan: performing initialising dma\n");
+    
+    DISABLE_INT_MASK (dev, INT_Dma0Proc | INT_Dma1Proc);
+    
+    /* initialise the context filter */
+    elan4_attach_filter (&dev->dev_ctxt, 0);
+
+    /* now issue a DMA - we expect this to trap */
+    writeq (E4_DMA_TYPE_SIZE (128*4, DMA_DataTypeByte, 0, 0) | RUN_DMA_CMD, cq->cq_mapping + (0 << 3));
+    writeq (0,                                                              cq->cq_mapping + (1 << 3));
+    writeq (0,                                                              cq->cq_mapping + (2 << 3));
+    writeq (dev->dev_tproc_space,                                           cq->cq_mapping + (3 << 3));
+    writeq (dev->dev_tproc_space,                                           cq->cq_mapping + (4 << 3));
+    writeq (0,                                                              cq->cq_mapping + (5 << 3));
+    writeq (0,                                                              cq->cq_mapping + (6 << 3));
+    
+    /* spin waiting for it to trap - then restart the dma processor */
+    do {
+       value   = read_reg64 (dev, IntAndMaskReg);
+       intreg  = (value >> E4_INTERRUPT_REG_SHIFT);
+    } while ((intreg & (INT_Dma0Proc | INT_Dma1Proc)) == 0);
+    
+    /* check it trapped for the right reason */
+    status = (intreg & INT_Dma0Proc) ? read_reg64 (dev, DProc0Status) : read_reg64 (dev, DProc1Status);
+    
+    if (DPROC_PrefetcherFault (status) || (DPROC_TrapType(status) != DmaProcFailCountError && DPROC_TrapType(status) != DmaProcPacketAckError))
+    {
+       printk ("elan: bad dma trap, status = %lx\n", (long)status);
+       panic ("elan: bad dma trap\n");
+    }
+    
+    PULSE_SCHED_RESTART (dev, SCH_RestartDma0Proc | SCH_RestartDma1Proc | SCH_RestartDmaPrefetchProc);
+
+    elan4_detach _filter (&dev->dev_ctxt, 0);
+
+    ENABLE_INT_MASK (dev, INT_Dma0Proc | INT_Dma1Proc);
+#undef VIRTUAL_ADDRESS
+}
+#endif
+
+static int
+ebus_read_vpd (ELAN4_DEV *dev, unsigned char *data, unsigned int nob)
+{
+    unsigned int pci_data_ptr;
+    unsigned int vpd_ptr;
+    register int i;
+
+    if (read_ebus_rom (dev, 0) != 0x55 || read_ebus_rom (dev, 1) != 0xaa)
+    {
+       printk ("elan%d: invalid rom signature in ebus rom\n", dev->dev_instance);
+       return -EINVAL;
+    }
+
+    pci_data_ptr = (read_ebus_rom (dev, 0x19) << 8) | read_ebus_rom (dev, 0x18);
+
+    /* check the pci data structure */
+    if (read_ebus_rom (dev, pci_data_ptr + 0) != 'P' ||
+       read_ebus_rom (dev, pci_data_ptr + 1) != 'C' ||
+       read_ebus_rom (dev, pci_data_ptr + 2) != 'I' ||
+       read_ebus_rom (dev, pci_data_ptr + 3) != 'R')
+    {
+       printk ("elan%d: invalid pci data structure in ebus rom\n", dev->dev_instance);
+       return -EINVAL;
+    }
+    
+    /* extract the VPD pointer */
+    vpd_ptr = (read_ebus_rom (dev, pci_data_ptr + 9) << 8) | read_ebus_rom (dev, pci_data_ptr + 8);
+
+    if (vpd_ptr == 0)
+    {
+       printk ("elan%d: no vital product data in ebus rom\n", dev->dev_instance);
+       return -EINVAL;
+    }
+    
+    /* read the vpd data */
+    for (i = 0; i < nob; i++)
+       data[i] = read_ebus_rom (dev, vpd_ptr + i);
+
+    return 0;
+}
+
+int
+elan4_read_vpd (ELAN4_DEV *dev, unsigned char *tag, unsigned char *result) 
+{
+    unsigned char vpd[I2C_ELAN_EEPROM_VPD_SIZE];
+    unsigned char *ptr = vpd;
+    unsigned int   finished = 0;
+    unsigned char *lim;
+    unsigned char  name[3];
+    unsigned char  value[256];
+    unsigned char  type;
+    unsigned int   len, len2;
+    register int   i;
+
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+    {
+       if (ebus_read_vpd (dev, vpd, I2C_ELAN_EEPROM_VPD_SIZE) < 0)
+       {
+           PRINTF1 (DBG_DEVICE, DBG_CONFIG, "elan%d: elan4_read_vpd, unable to read serial number from EBUS rom\n", dev->dev_instance);
+           return -EINVAL ;
+       }       
+    }
+    else
+    {
+       if (i2c_read_rom (dev, I2C_ELAN_EEPROM_VPD_BASEADDR, I2C_ELAN_EEPROM_VPD_SIZE, vpd) < 0)
+       {
+           PRINTF1 (DBG_DEVICE, DBG_CONFIG, "elan%d: elan4_read_vpd, unable to read serial number from I2C rom\n", dev->dev_instance);
+           return  -EINVAL;
+       }
+    }
+
+    result[0] = 0;
+    while (! finished)
+    {
+       type = *ptr++;
+       
+       if (type & LARGE_RESOURCE_BIT)
+       {
+           len = *(ptr++);
+           len += *(ptr++) << 8;
+           
+           switch (type & ~LARGE_RESOURCE_BIT)
+           {
+           case LARGE_RESOURCE_STRING:
+           case LARGE_RESOURCE_VENDOR_DEFINED:
+               ptr += len;
+               break;
+               
+           case LARGE_RESOURCE_VITAL_PRODUCT_DATA:
+               for (lim = ptr + len; ptr < lim; )
+               {
+                   name[0] = *ptr++;
+                   name[1] = *ptr++;
+                   name[2] = '\0';
+                   len2    = *ptr++;
+                   
+                   for (i = 0; i < len2 && ptr < lim; i++)
+                       value[i] = *ptr++;
+                   value[i] = '\0';
+                                   
+                   PRINTF3 (DBG_DEVICE, DBG_CONFIG, "elan%d: elan4_read_vpd, %s: $s\n", dev->dev_instance, name, value);
+
+                   if (tag != NULL) 
+                   { /* looking for just one tag */
+                       if (!strcmp (name, tag))
+                           strcpy(result, value);
+                   } 
+                   else 
+                   { /* get all tags */
+                       strcat(result,name);
+                       strcat(result,": ");
+                       strcat(result,value);
+                       strcat(result,"\n");
+                   }
+               }
+               break;
+               
+           default:
+               PRINTF2 (DBG_DEVICE, DBG_CONFIG, "elan%d: elan4_read_vpd, unknown large resource %x\n", dev->dev_instance, type);
+               finished = 1;
+               break;
+           }
+       }
+       else
+       {
+           len = type & 0x7;
+           
+           switch (type >> 3)
+           {
+           case SMALL_RESOURCE_COMPATIBLE_DEVICE_ID:
+               ptr += len;
+               break;
+               
+           case SMALL_RESOURCE_VENDOR_DEFINED:
+               ptr += len;
+               break;
+               
+           case SMALL_RESOURCE_END_TAG:
+               finished = 1;
+               break;
+               
+           default:
+               PRINTF2 (DBG_DEVICE, DBG_CONFIG, "elan%d: elan4_read_vpd, unknown small resource %x\n", dev->dev_instance, type >> 3);
+               finished = 1;
+               break;
+           }
+       }
+    }
+
+    if ( result[0] == 0 ) {
+       if ( tag != 0 ) 
+           PRINTF2 (DBG_DEVICE, DBG_CONFIG, "elan%d: elan4_read_vpd, failed to find tag %s\n", dev->dev_instance, tag);
+       else
+           PRINTF1 (DBG_DEVICE, DBG_CONFIG, "elan%d: elan4_read_vpd, failed to find any tags\n", dev->dev_instance);
+       return -EINVAL;
+    }
+
+    return (0);
+}
+
+int
+elan4_start_device (ELAN4_DEV *dev)
+{
+    E4_VirtualProcessEntry entry;
+    unsigned              pagesizeval[2];
+    unsigned              hashsizeval[2];
+    register int           i, j, tbl, res;
+    unsigned               attempts = 0;
+    E4_PAGE_SIZE_TABLE;
+    unsigned char          serial[256];
+
+    PRINTF (DBG_DEVICE, DBG_ALL, "elan4_start_device: entered\n");
+
+    dev->dev_state = ELAN4_STATE_STARTING;
+
+ tryagain:
+    /* Initialise the pci config space */
+    if ((res = elan4_pciinit (dev)) < 0)
+       return (res);
+
+    /* Display the serial number */
+    if (elan4_read_vpd (dev, "SN", serial))
+       printk("elan%d: SN: failed to read\n", dev->dev_instance);
+    else
+       printk("elan%d: SN: %s\n", dev->dev_instance, serial);
+
+    /* initialise the interrupt mask to zero */
+    SET_INT_MASK (dev, 0);
+
+    /* Initialise the device registers */
+    write_reg64 (dev, TlbLineValue, 0);
+    write_reg64 (dev, SysControlReg, 0);
+
+    /* Initialise the SDRAM using the configuration value from the ROM */
+    write_reg64 (dev, SDRamConfigReg, dev->dev_sdram_cfg | SDRAM_SETUP);
+
+    /* Setup the linkport registers */
+    write_reg64 (dev, LinkPortLock, 0);
+    write_reg64 (dev, LinkPortKey,  LINK_PORT_LOCK_VALUE);
+
+    /* Setup the tick rates, start the clock, and init the stats registers */
+    write_ureg32 (dev, ClockTickRate.s.TickRates, ELAN4_CLOCK_TICK_RATE);
+    write_ureg64 (dev, Clock, 0);
+    write_ureg32 (dev, InstCount.s.StatsCount, 0);
+    for (i = 0; i < 8; i++)
+       write_ureg32 (dev, StatCounts[i].s.StatsCount, 0);
+
+    /* Initialise the Link Control register - disable the TLB prefetcher on RevB
+     * as it can cause very occasional data corruption. */
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVB)
+       write_reg32 (dev, LinkControlReg, LCONT_REVB_DISABLE_TLB_PREFETCH);
+    else
+       write_reg32 (dev, LinkControlReg, 0);
+
+    /* Initialise the Link Control Settings to set the PLL Reference Value */
+    write_reg32 (dev, LinkContSettings, 
+                (elan4_mod45disable ? LCONT_MOD45_DISABLE : 0) |
+                (3 << LCONT_CONFIG_PHASE_SHIFT) |
+                ((elan4_pll_div & LCONT_PLL_REF_VAL_BITS_MASK) << LCONT_PLL_REF_VAL_BITS_SHIFT) |
+                (LCONT_VOD_360 << LCONT_LVDS_VOLTAGE_BITS_SHIFT) |
+                (LCONT_TERM_AUTO_OHM << LCONT_LVDS_TERMINATION_SHIFT));
+
+    /* Clear the link error LED on RevB and above */
+    if (dev->dev_devinfo.dev_revision_id != PCI_REVISION_ID_ELAN4_REVA)
+       write_i2c (dev, I2cStatus, read_i2c (dev, I2cStatus) | I2cCntl_ClearLinkError);
+
+    initialise_cache (dev);
+
+    /* Initialise the MMU hash table parameters */
+    /* Select the largest elan pagesize which is spanned by the
+     * system pagesize for mmu table 0*/
+    for (i = 0; i < E4_PAGE_SIZE_TABLE_SIZE; i++)
+       if (PageSizeTable[i] > PAGE_SHIFT)
+           break;
+
+    pagesizeval[0] = i - 1;
+    hashsizeval[0] = elan4_hash_0_size_val;
+       
+    /* Select a suitable elan pagesize to match any "large" page
+     * support that the OS provides. */
+    pagesizeval[1] = PAGE_SIZE_4M;
+    hashsizeval[1] = elan4_hash_1_size_val;
+
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+    {
+       dev->dev_pagesizeval[tbl]   = pagesizeval[tbl];
+       dev->dev_pageshift[tbl]     = PageSizeTable[pagesizeval[tbl]];
+       dev->dev_hashsize[tbl]      = (1 << hashsizeval[tbl])/sizeof (E4_HashTableEntry);
+       dev->dev_rsvd_hashmask[tbl] = ((1 << (27 - dev->dev_pageshift[tbl]))-1) & ~((1 << hashsizeval[tbl])-1);
+       dev->dev_rsvd_hashval[tbl]  = 0xFFFFFFFF;
+    }
+
+    PRINTF2 (DBG_DEVICE, DBG_CONFIG, "elan4_start_device: pageshifts %d,%d\n", dev->dev_pageshift[0], 
+            NUM_HASH_TABLES == 2 ? dev->dev_pageshift[1] : 0);
+
+    /* Initialise the control register to the desired value */
+    dev->dev_syscontrol = (CONT_EN_ALL_SETS | CONT_MMU_ENABLE | CONT_CACHE_ALL | CONT_2K_NOT_1K_DMA_PACKETS |
+                          (pagesizeval[0] << CONT_TABLE0_PAGE_SIZE_SHIFT) | (hashsizeval[0] << CONT_TABLE0_MASK_SIZE_SHIFT));
+
+    if (NUM_HASH_TABLES == 2)
+       dev->dev_syscontrol |= CONT_TWO_HASH_TABLES | (pagesizeval[1] << CONT_TABLE1_PAGE_SIZE_SHIFT) | (hashsizeval[1] << CONT_TABLE1_MASK_SIZE_SHIFT);
+
+    write_reg64 (dev, SysControlReg, dev->dev_syscontrol);
+
+    /* use direct mapped pci writes during sdram initialisation, since for 
+     * cache flushing to work, we need to ensure that the cacheflush page
+     * never gets lines into the incorrect cache set. */
+    SET_SYSCONTROL (dev, dev_direct_map_pci_writes, CONT_DIRECT_MAP_PCI_WRITES);
+
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVB)
+       elan4_sdram_setup_delay_lines(dev);
+
+    for (i = res = 0; i < dev->dev_sdram_numbanks; i++)
+       if (dev->dev_sdram_banks[i].b_size)
+           res |= elan4_sdram_init_bank (dev, &dev->dev_sdram_banks[i]);
+
+    if (! res)
+    {
+       if (dev->dev_devinfo.dev_device_id == PCI_REVISION_ID_ELAN4_REVB && ++attempts < 5)
+       {
+           printk ("elan%d: sdram not working, resetting\n", dev->dev_instance);
+           goto tryagain;
+       }
+
+       printk ("elan%d: could not find any sdram banks\n", dev->dev_instance);
+       goto failed;
+    }
+
+#ifndef CONFIG_MPSAS
+    PRINTF0 (DBG_DEVICE, DBG_CONFIG, "elan4_start_device: initialising for ECC\n");
+
+    for (i = 0 ; i < dev->dev_sdram_numbanks; i++)
+       if (dev->dev_sdram_banks[i].b_ioaddr)
+           initialise_ecc (dev, &dev->dev_sdram_banks[i]);
+#endif
+
+    dev->dev_sdram_initial_ecc_val = read_reg64 (dev, SDRamECCStatus);
+
+    /* Now enable ECC after we've scrubbed the memory */
+    write_reg64 (dev, SDRamConfigReg, dev->dev_sdram_cfg | SDRAM_ENABLE_ECC);
+
+    /* clear any errors, and flush the tlb/route cache */
+    PULSE_SYSCONTROL (dev, CONT_TLB_FLUSH | CONT_ROUTE_FLUSH | CONT_CLEAR_LINKPORT_INT | CONT_CLEAR_SDRAM_ERROR);
+
+    write_ureg32 (dev, InstCount.s.StatsCount, 0);
+
+    /* Initialise the thread processor's register file */
+    for (i = 0; i < 64; i++)
+       write_reg64 (dev, TProcRegs[i], 0);
+
+    /* Initialise the thread processor's ICache tags */
+    for (i = 0; i < (E4_ICacheLines/E4_ICachePortSize); i++)
+    {
+        write_reg64 (dev, ICachePort_Cntl_Addr, i << E4_ICacheTagAddrShift);
+        for (j = 0; j < E4_ICachePortSize; j++)
+           write_reg64 (dev, ICachePort[j], E4_InvalidTagValue);
+    }
+
+    /*
+     * Initialise the ICache with a sethi %hi(addr << 7), %r0
+     * writing 8 64 bit values per loop of sethi %g0 values ending in 77 for something different??
+     */
+    for (i = 0; i < E4_ICacheSizeInBytes; i += (E4_ICachePortSize << 3))
+    {
+       write_reg64 (dev, ICachePort_Cntl_Addr, E4_AccessICacheRams | (i >> 3));
+
+       for (j = 0; j < E4_ICachePortSize; j++)
+           write_reg64 (dev, ICachePort[j], 
+                        (E4_uint64) (((E4_uint64)i << (4+7))    + ((E4_uint64)j << (1+7))    + (0x077)) |
+                        (E4_uint64) (((E4_uint64)i << (4+7+32)) + ((E4_uint64)j << (1+7+32)) + (0x0e7)) << 32);
+    }
+
+    /*
+     * Initialise the top of the ICache Set0 with a instruction which will
+     * cause a know trap fingerprint so that the application can identify it
+     * and ignore the trap.
+     */
+    write_reg64 (dev, ICachePort_Cntl_Addr, E4_ICacheFixupOffset | E4_AccessICacheRams);
+    for (i = 0; i < E4_ICachePortSize; i++)
+       write_reg64 (dev, ICachePort[i], E4_ICacheFixupInsn | (E4_ICacheFixupInsn << 32));
+
+    /* create the buddy allocator for SDRAM */
+    for (i = 0; i < dev->dev_sdram_numbanks; i++)
+       if (dev->dev_sdram_banks[i].b_ioaddr)
+           elan4_sdram_add_bank (dev, &dev->dev_sdram_banks[i]);
+
+    dev->dev_ctxtableshift        = elan4_ctxt_table_shift;
+    dev->dev_cqcount              = (1 << elan4_ln2_max_cqs);
+#ifdef CONFIG_MTRR
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVB)
+       dev->dev_cqreorder = dev->dev_cqcount >> 1;
+    else
+       dev->dev_cqreorder = dev->dev_cqcount;
+#endif
+
+    /* allocate the sdram for cache flushing whilst still in direct mapped mode */
+    dev->dev_cacheflush_space = elan4_sdram_alloc (dev, E4_CacheSize);
+
+    /* and longer need direct mapped pci writes */
+    CLEAR_SYSCONTROL (dev, dev_direct_map_pci_writes, CONT_DIRECT_MAP_PCI_WRITES);
+
+    /* allocate the hash tables, command queues, context tables etc */
+    PRINTF0 (DBG_DEVICE, DBG_CONFIG, "elan4_start_device: allocating hash tables, command queueus, context tables\n");
+
+    dev->dev_comqlowpri       = elan4_sdram_alloc (dev, (1 << COMMAND_RUN_QUEUE_BITS));
+    dev->dev_comqhighpri      = elan4_sdram_alloc (dev, (1 << COMMAND_RUN_QUEUE_BITS));
+    dev->dev_cqaddr           = elan4_sdram_alloc (dev, sizeof (E4_CommandQueueDesc) * dev->dev_cqcount);
+    dev->dev_dmaqhighpri      = elan4_sdram_alloc (dev, E4_QueueSize(elan4_dmaq_highpri_size));
+    dev->dev_dmaqlowpri       = elan4_sdram_alloc (dev, E4_QueueSize(elan4_dmaq_lowpri_size));
+    dev->dev_threadqhighpri   = elan4_sdram_alloc (dev, E4_QueueSize(elan4_threadq_highpri_size));
+    dev->dev_threadqlowpri    = elan4_sdram_alloc (dev, E4_QueueSize(elan4_threadq_lowpri_size));
+    dev->dev_interruptq       = elan4_sdram_alloc (dev, E4_QueueSize(elan4_interruptq_size));
+
+    dev->dev_ctxtable         = elan4_sdram_alloc (dev, (1 << dev->dev_ctxtableshift) * sizeof (E4_ContextControlBlock));
+    dev->dev_faultarea        = elan4_sdram_alloc (dev, CUN_Entries * sizeof (E4_FaultSave));
+    dev->dev_inputtraparea    = elan4_sdram_alloc (dev, sizeof (E4_IprocTrapState));
+
+    dev->dev_sdrampages[0]    = elan4_sdram_alloc (dev, SDRAM_PAGE_SIZE);
+    dev->dev_sdrampages[1]    = elan4_sdram_alloc (dev, SDRAM_PAGE_SIZE);
+
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+    {
+       dev->dev_hashtable[tbl] = elan4_sdram_alloc (dev, dev->dev_hashsize[tbl] * sizeof (E4_HashTableEntry));
+#ifndef CONFIG_MPSAS
+       /* Initialise hash tables to invalid (zero) */
+       elan4_sdram_zeroq_sdram (dev, dev->dev_hashtable[tbl], dev->dev_hashsize[tbl] * sizeof (E4_HashTableEntry));
+#endif
+    }
+
+    /* Initialise all context filters to discard */
+#ifdef CONFIG_MPSAS
+    if (sas_memset_dev (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM, dev->dev_ctxtable, 
+                       E4_FILTER_DISCARD_ALL, (1 << (dev->dev_ctxtableshift-1))) < 0)
+    {
+       for (i = 0; i < (1 << dev->dev_ctxtableshift); i++)
+           elan4_write_filter (dev, i, E4_FILTER_DISCARD_ALL);
+    }
+#else
+    for (i = 0; i < (1 << dev->dev_ctxtableshift); i++)
+       elan4_write_filter (dev, i, E4_FILTER_DISCARD_ALL);
+#endif
+
+    PRINTF4 (DBG_DEVICE, DBG_CONFIG, "elan4_start_device: hashtables %x,%x, %x,%x\n", dev->dev_hashtable[0], 
+           dev->dev_hashsize[0], dev->dev_hashtable[1], dev->dev_hashsize[1]);
+
+    /* install the hash table pointers */
+    PRINTF0 (DBG_DEVICE, DBG_CONFIG, "elan4_start_device: initialise registers with table addresses\n");
+    write_reg64 (dev, MmuTableBasePtrs, (((E4_uint64) dev->dev_hashtable[0]) | ((E4_uint64) dev->dev_hashtable[1]) << 32));
+    write_reg64 (dev, MmuFaultAndRootCntxPtr, (((E4_uint64) dev->dev_ctxtableshift) | 
+                                              ((E4_uint64) dev->dev_ctxtable) |
+                                              ((E4_uint64) dev->dev_faultarea) << 32));
+    write_reg64 (dev, InputTrapAndFilter, (((E4_uint64) dev->dev_ctxtableshift) | 
+                                          ((E4_uint64) dev->dev_ctxtable) |
+                                          ((E4_uint64) dev->dev_inputtraparea) << 32));
+    /*
+     * The run ptrs have this format: (Front << 32) | Back
+     * The base for both the front and back is uses the high bits of the back pointer.
+     * So writting just the base value is good enough.
+     */
+    write_reg64 (dev, CommandLowPriRunPtrs,  dev->dev_comqlowpri);
+    write_reg64 (dev, CommandHighPriRunPtrs, dev->dev_comqhighpri);
+
+    /* Initialise the run queues */
+    write_reg64 (dev, DProcHighPriPtrs,       E4_QueueValue (dev->dev_dmaqhighpri,    elan4_dmaq_highpri_size));
+    write_reg64 (dev, DProcLowPriPtrs,        E4_QueueValue (dev->dev_dmaqlowpri,     elan4_dmaq_lowpri_size));
+    write_reg64 (dev, TProcHighPriPtrs,       E4_QueueValue (dev->dev_threadqhighpri, elan4_threadq_highpri_size));
+    write_reg64 (dev, TProcLowPriPtrs,        E4_QueueValue (dev->dev_threadqlowpri,  elan4_threadq_lowpri_size));
+
+    /* Initialise the interrupt queue as "empty" - this is actually with one entry on it */
+    write_reg64 (dev, MainIntQueuePtrs.Value, (((E4_uint64) E4_QueueFrontValue (dev->dev_interruptq, elan4_interruptq_size) << 32) |
+                                              ((E4_uint64) E4_QueueBackPointer(dev->dev_interruptq + E4_MainIntEntrySize))));
+    
+    dev->dev_interruptq_nfptr = dev->dev_interruptq + E4_MainIntEntrySize;
+
+    /*
+     * Flush the context filter before dropping the Discard all bits in the schedule status register.
+     * Also hit the SCH_RestartTProc to clear out X's from the trap state and
+     * hit the SCH_RestartDmaPrefetchProc to clear out X's from the prev register.
+     */
+    PULSE_SCHED_RESTART (dev, SCH_ContextFilterFlush | SCH_RestartTProc | SCH_RestartDmaPrefetchProc);
+
+    /* setup the schedule status register. */
+    SET_SCHED_STATUS (dev, SCH_CProcTimeout6p2us | SCH_DProcTimeslice512us);
+
+    /*
+     * Now initialise the inserter cache.s
+     * Bit 31 of the first word of the descriptor is a valid bit. This must be cleared.
+     * Bit 31 becomes a used bit in the descriptors in memory.
+     */
+    for (i = 0; i < COMMAND_INSERTER_CACHE_ENTRIES; i++)
+    {
+       write_reg32 (dev, CommandControl.CommandQueueDescsBase, i);     /* select a cache line */
+       write_reg64 (dev, CommandCacheTestPort, 0);                     /* Mark it invalid */
+    }
+    
+    /* Setup the pointer to the command descriptors */
+    /*   the table must be aligned on a CQ_CommandDescsAlignement boundary */
+    /*   since we've allocated a small table - we work out the offset of the */
+    /*   first entry in our table for mapping in the command ports later */
+    dev->dev_cqoffset = (dev->dev_cqaddr & (CQ_CommandDescsAlignment-1)) / sizeof (E4_CommandQueueDesc);
+
+    write_reg32 (dev, CommandControl.CommandQueueDescsBase, (dev->dev_cqaddr & ~(CQ_CommandDescsAlignment-1)) | COM_ENABLE_DEQUEUE);
+
+    /* allocate the bitmaps for cq,ctxt allocation */
+    KMEM_ZALLOC (dev->dev_cqamap, bitmap_t *, BT_BITOUL(dev->dev_cqcount/ELAN4_CQ_PER_CQA) * sizeof (bitmap_t), 1);
+    KMEM_ZALLOC (dev->dev_ctxmap, bitmap_t *, BT_BITOUL(1 << dev->dev_ctxtableshift) * sizeof (bitmap_t), 1);
+
+    if (dev->dev_cqamap == NULL || dev->dev_ctxmap == NULL)
+       goto failed;
+
+    /* Make every fourth context be invalid for ICache fixup.
+     * context 0 is also invalid - since it is used to indicate 
+     * an invalid tag. */
+    for (i = 0; i < (1 << dev->dev_ctxtableshift); i += 4)
+       BT_SET (dev->dev_ctxmap, i);
+    
+    /* initialise the halt operations */
+    dev->dev_haltop_mask   = 0;
+    dev->dev_haltop_active = 0;
+
+    /* allocate the hash table shadow structures - and place all blocks on the free lists */
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+    {
+       KMEM_ZALLOC (dev->dev_mmuhash[tbl], ELAN4_HASH_ENTRY *,  dev->dev_hashsize[tbl] * sizeof (ELAN4_HASH_ENTRY), 1);
+       KMEM_ZALLOC (dev->dev_mmufree[tbl], ELAN4_HASH_ENTRY **, dev->dev_hashsize[tbl] * sizeof (ELAN4_HASH_ENTRY *), 1);
+
+       if (dev->dev_mmuhash[tbl] == NULL || dev->dev_mmufree[tbl] == NULL)
+           goto failed;
+
+       for (i = 0; i < dev->dev_hashsize[tbl]; i++)
+       {
+           dev->dev_mmuhash[tbl][i].he_entry = dev->dev_hashtable[tbl] + (i * sizeof (E4_HashTableEntry));
+           dev->dev_mmufree[tbl][i]          = &dev->dev_mmuhash[tbl][i];
+       }
+    }
+
+    /* setup the interrupt mask register */
+    SET_INT_MASK (dev, (INT_MSI0 | INT_MSI1 | INT_MSI2 | INT_MSI3) & ~(INT_Discarding | INT_Halted));
+
+    /* start a thread to handle excessive main interrupts */
+    if (kernel_thread_create (elan4_mainint_thread, (caddr_t) dev) == NULL)
+       goto failed;
+    dev->dev_mainint_started = 1;
+    
+    /* install the device context - and allocate the first 16 command queues */
+    if (elan4_insertctxt (dev, &dev->dev_ctxt, &device_trap_ops) != 0)
+       goto failed;
+
+    /* Allocate command queues, one for each entry in the inserter cache, 
+     * we'll use these queues to flush the insert cache */
+    for (i = 0; i < COMMAND_INSERTER_CACHE_ENTRIES; i++)
+    {
+       if ((dev->dev_flush_cq[i] = elan4_alloccq (&dev->dev_ctxt, CQ_Size1K, CQ_DmaStartEnableBit | CQ_InterruptEnableBit, 
+                                                  CQ_Priority)) == NULL)
+           goto failed;
+       
+       ASSERT (elan4_cq2num(dev->dev_flush_cq[i]) == i);
+
+       dev->dev_flush_finished |= (1 << i);
+    }
+
+    /* Allocate command queues for dma halt operations */
+    if ((dev->dev_dma_flushop[0].cq = elan4_alloccq (&dev->dev_ctxt, CQ_Size1K, CQ_DmaStartEnableBit, 0)) == NULL ||
+       (dev->dev_dma_flushop[1].cq = elan4_alloccq (&dev->dev_ctxt, CQ_Size1K, CQ_DmaStartEnableBit, CQ_Priority)) == NULL)
+       goto failed;
+
+#ifdef CONFIG_MPSAS
+    elan4_sdram_flushcache (dev, 0, E4_CacheSize);
+#endif
+
+    /* initialise halt operation for flushing the icache */
+    dev->dev_iflush_haltop.op_function = device_iflush_haltop;
+    dev->dev_iflush_haltop.op_arg      = dev;
+    dev->dev_iflush_haltop.op_mask     = INT_TProcHalted;
+
+    /* Allocate a route table, and create a valid route for vp==0, this is used
+     * when a DMA is removed from the dma run queue */
+    if ((dev->dev_routetable = elan4_alloc_routetable (dev, 0)) == NULL)
+       goto failed;
+
+    elan4_set_routetable (&dev->dev_ctxt, dev->dev_routetable);
+
+    entry.Values[0] = FIRST_MYLINK;
+    entry.Values[1] = 0;
+
+    elan4_write_route (dev, dev->dev_routetable, 0, &entry);
+
+    /* map the sdram pages into the elan */
+    dev->dev_tproc_suspend = DEVICE_TPROC_SUSPEND_ADDR;
+    dev->dev_tproc_space   = DEVICE_TPROC_SPACE_ADDR;
+
+    elan4mmu_pteload (&dev->dev_ctxt, 0, dev->dev_tproc_suspend, (dev->dev_sdrampages[0] >> PTE_PADDR_SHIFT) | PTE_SetPerm(PERM_LocExecute));
+    elan4mmu_pteload (&dev->dev_ctxt, 0, dev->dev_tproc_space,   (dev->dev_sdrampages[1] >> PTE_PADDR_SHIFT) | PTE_SetPerm(PERM_LocDataWrite));
+
+    /* and store the thread suspend sequence in it for use when a thread is removed from the run queue */
+    elan4_sdram_writel (dev, dev->dev_sdrampages[0], DEVICE_TPROC_SUSPEND_INSTR);
+
+#ifdef CONFIG_MPSAS
+    do_initdma (dev);
+#endif
+    
+    if (!elan4_neterr_init (dev))
+       goto failed;
+
+    elan4_configure_mtrr (dev);
+
+    /* finally register the device with elanmod for rms */
+    dev->dev_idx = elan_dev_register (&dev->dev_devinfo, &elan4_dev_ops, (void *) dev);
+
+    dev->dev_state = ELAN4_STATE_STARTED;
+
+    return (0);
+
+ failed:
+    printk ("elan%d: failed to start elan4 device - stopping\n", dev->dev_instance);
+
+    elan4_stop_device (dev);
+    return (-ENOMEM);
+}
+
+void
+elan4_stop_device (ELAN4_DEV *dev)
+{
+    unsigned long flags;
+    int i, tbl;
+
+    dev->dev_state = ELAN4_STATE_STOPPING;
+
+    elan_dev_deregister (&dev->dev_devinfo);
+
+    elan4_unconfigure_mtrr (dev);
+
+    elan4_neterr_destroy (dev);
+
+    if (dev->dev_tproc_suspend)
+       elan4mmu_unload_range (&dev->dev_ctxt, 0, dev->dev_tproc_suspend, 1 << dev->dev_pageshift[0]);
+
+    if (dev->dev_tproc_space)
+       elan4mmu_unload_range (&dev->dev_ctxt, 0, dev->dev_tproc_space,   1 << dev->dev_pageshift[0]);
+
+    if (dev->dev_routetable)
+    {
+       elan4_set_routetable (&dev->dev_ctxt, NULL);
+       elan4_free_routetable (dev, dev->dev_routetable);
+    }
+
+    for (i = 0; i < 2; i++)
+       if (dev->dev_dma_flushop[i].cq)
+           elan4_freecq (&dev->dev_ctxt, dev->dev_dma_flushop[i].cq);
+
+    /* free of the device context - and insert cache flushing command queues */
+    for (i = 0; i < COMMAND_INSERTER_CACHE_ENTRIES; i++)
+       if (dev->dev_flush_cq[i])
+           elan4_freecq (&dev->dev_ctxt, dev->dev_flush_cq[i]);
+
+    if (dev->dev_ctxt.ctxt_dev)
+       elan4_removectxt (dev, &dev->dev_ctxt);
+
+    /* stop the mainint thread */
+    spin_lock_irqsave (&dev->dev_mainint_lock, flags);
+    dev->dev_stop_threads = 1;
+
+    while (dev->dev_mainint_started && !dev->dev_mainint_stopped)
+    {
+       kcondvar_wakeupall (&dev->dev_mainint_wait, &dev->dev_mainint_lock);
+       kcondvar_wait (&dev->dev_mainint_wait, &dev->dev_mainint_lock, &flags);
+    }
+    dev->dev_mainint_started = dev->dev_mainint_stopped = 0;
+    spin_unlock_irqrestore (&dev->dev_mainint_lock, flags);
+
+    /* cancel any error interrupt timeouts */
+    if (timer_fn_queued (&dev->dev_error_timeoutid))
+       cancel_timer_fn (&dev->dev_error_timeoutid);
+
+    if (dev->dev_devinfo.dev_revision_id != PCI_REVISION_ID_ELAN4_REVA && timer_fn_queued (&dev->dev_linkerr_timeoutid))
+       cancel_timer_fn (&dev->dev_linkerr_timeoutid);
+    
+    /* reset the interrupt mask register to zero */
+    if (dev->dev_regs)
+       SET_INT_MASK (dev, 0);
+
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+    {
+       if (dev->dev_mmuhash[tbl])
+           KMEM_FREE (dev->dev_mmuhash[tbl], dev->dev_hashsize[tbl] * sizeof (ELAN4_HASH_ENTRY));
+       if (dev->dev_mmufree[tbl])
+           KMEM_FREE (dev->dev_mmufree[tbl], dev->dev_hashsize[tbl] * sizeof (ELAN4_HASH_ENTRY *));
+       if (dev->dev_hashtable[tbl])
+           elan4_sdram_free (dev, dev->dev_hashtable[tbl], dev->dev_hashsize[tbl] * sizeof (E4_HashTableEntry));
+    }
+
+    if (dev->dev_cqamap)
+       KMEM_FREE (dev->dev_cqamap, BT_BITOUL (dev->dev_cqcount/ELAN4_CQ_PER_CQA) * sizeof (bitmap_t));
+    if (dev->dev_ctxmap)
+       KMEM_FREE (dev->dev_ctxmap, BT_BITOUL(1 << dev->dev_ctxtableshift) * sizeof (bitmap_t));
+
+    if (dev->dev_comqlowpri)
+       elan4_sdram_free (dev, dev->dev_comqlowpri,     (1 << COMMAND_RUN_QUEUE_BITS));
+    if (dev->dev_comqhighpri)
+       elan4_sdram_free (dev, dev->dev_comqhighpri,    (1 << COMMAND_RUN_QUEUE_BITS));
+    if (dev->dev_cqaddr)
+       elan4_sdram_free (dev, dev->dev_cqaddr,         sizeof (E4_CommandQueueDesc) * dev->dev_cqcount);
+    if (dev->dev_dmaqhighpri)
+       elan4_sdram_free (dev, dev->dev_dmaqhighpri,    E4_QueueSize(elan4_dmaq_highpri_size));
+    if (dev->dev_dmaqlowpri)
+       elan4_sdram_free (dev, dev->dev_dmaqlowpri,     E4_QueueSize(elan4_dmaq_lowpri_size));
+    if (dev->dev_threadqhighpri)
+       elan4_sdram_free (dev, dev->dev_threadqhighpri, E4_QueueSize(elan4_threadq_highpri_size));
+    if (dev->dev_threadqlowpri)
+       elan4_sdram_free (dev, dev->dev_threadqlowpri,  E4_QueueSize(elan4_threadq_lowpri_size));
+    if (dev->dev_interruptq)
+       elan4_sdram_free (dev, dev->dev_interruptq,     E4_QueueSize(elan4_interruptq_size));
+    
+    if (dev->dev_ctxtable)
+       elan4_sdram_free (dev, dev->dev_ctxtable,       (1 << dev->dev_ctxtableshift) * sizeof (E4_ContextControlBlock));
+    if (dev->dev_faultarea)
+       elan4_sdram_free (dev, dev->dev_faultarea,      CUN_Entries * sizeof (E4_FaultSave));
+    if (dev->dev_inputtraparea)
+       elan4_sdram_free (dev, dev->dev_inputtraparea,  sizeof (E4_IprocTrapState));
+
+    if (dev->dev_sdrampages[0])
+       elan4_sdram_free (dev, dev->dev_sdrampages[0],  SDRAM_PAGE_SIZE);
+    if (dev->dev_sdrampages[1])
+       elan4_sdram_free (dev, dev->dev_sdrampages[1],  SDRAM_PAGE_SIZE);
+
+    for (i = 0; i < dev->dev_sdram_numbanks; i++)
+       if (dev->dev_sdram_banks[i].b_ioaddr)
+               elan4_sdram_fini_bank (dev, &dev->dev_sdram_banks[i]);
+
+    elan4_pcifini (dev);
+
+    dev->dev_state = ELAN4_STATE_STOPPED;
+
+    if (dev->dev_ack_errors)
+        kfree(dev->dev_ack_errors);
+    if (dev->dev_dproc_timeout)
+        kfree(dev->dev_dproc_timeout);
+    if (dev->dev_cproc_timeout)
+        kfree(dev->dev_cproc_timeout);
+}
+
+static __inline__ int
+compute_arity (int lvl, unsigned n, char *arity)
+{
+    if (arity[lvl] == 0)
+    {
+       if (n <= 8)
+           arity[lvl] = n;
+       else
+           arity[lvl] = 4;
+    }
+
+    return (arity[lvl]);
+}
+
+int
+elan4_compute_position (ELAN_POSITION *pos, unsigned nodeid, unsigned numnodes, unsigned arityval)
+{
+    int i, lvl, n;
+    char arity[ELAN_MAX_LEVELS];
+
+    if (nodeid >= numnodes)
+       return -EINVAL;
+
+    for (i = 0; i < ELAN_MAX_LEVELS; i++, arityval >>= 4)
+       arity[i] = arityval & 7;
+    
+    for (lvl = 0, n = numnodes; n > compute_arity(lvl, n, arity) && lvl < ELAN_MAX_LEVELS; lvl++)
+    {
+       if ((n % arity[lvl]) != 0)
+           return -EINVAL;
+       
+       n /= arity[lvl];
+    }
+
+    if (arity[lvl] != n)
+       return -EINVAL;
+
+    for (i = 0; i <= lvl; i++)
+       pos->pos_arity[i] = arity[lvl - i];
+
+    pos->pos_nodes  = numnodes;
+    pos->pos_levels = lvl + 1;
+    pos->pos_nodeid = nodeid;
+    pos->pos_mode   = ELAN_POS_MODE_SWITCHED;
+
+    return 0;
+}
+
+int
+elan4_get_position (ELAN4_DEV *dev, ELAN_POSITION *pos)
+{
+    kmutex_lock (&dev->dev_lock);
+    *pos = dev->dev_position;
+    kmutex_unlock (&dev->dev_lock);
+
+    return (pos->pos_mode);
+}
+
+int
+elan4_set_position (ELAN4_DEV *dev, ELAN_POSITION *pos)
+{
+    int forceLocal = 0;
+    int nnodes, i;
+    unsigned int *ack_errors;
+    unsigned int *dproc_timeout;
+    unsigned int *cproc_timeout;
+
+    switch (pos->pos_mode)
+    {
+    case ELAN_POS_UNKNOWN:
+       break;
+       
+    case ELAN_POS_MODE_SWITCHED:
+       if (pos->pos_levels > ELAN_MAX_LEVELS)
+           return (-EINVAL);
+       
+       for (i = 0, nnodes = 1; i < pos->pos_levels; i++)
+       {
+
+           if (pos->pos_arity[i] <= 0 || (i == 0 ? pos->pos_arity[i] > 8 : pos->pos_arity[i] >= 8))  /* allow an 8 way top-switch */
+               return (-EINVAL);
+           
+           nnodes *= pos->pos_arity[i];
+       }
+
+       if (pos->pos_nodes > nnodes || pos->pos_nodeid >= pos->pos_nodes)
+           return (-EINVAL);
+       break;
+       
+    case ELAN_POS_MODE_LOOPBACK:
+       if (pos->pos_levels != 1 || pos->pos_nodes != 1 || pos->pos_nodeid != 0 || pos->pos_arity[0] != 1)
+           return (-EINVAL);
+
+       forceLocal = 1;
+       break;
+
+    case ELAN_POS_MODE_BACKTOBACK:
+       if (pos->pos_levels != 1 || pos->pos_nodes != 2 || pos->pos_nodeid >= 2 || pos->pos_arity[0] != 2)
+           return (-EINVAL);
+
+       forceLocal = (pos->pos_nodeid == 0);
+       break;
+
+    default:
+       return (-EINVAL);
+    }
+
+    ack_errors = kmalloc(pos->pos_nodes * sizeof(unsigned int), GFP_KERNEL);
+    if (!ack_errors)
+       return (-EINVAL);
+    memset(ack_errors, 0, pos->pos_nodes * sizeof(unsigned int));
+    dproc_timeout = kmalloc(pos->pos_nodes * sizeof(unsigned int), GFP_KERNEL);
+    if (!dproc_timeout) 
+    {
+        kfree(ack_errors);
+        return (-EINVAL);
+    }
+    memset(dproc_timeout, 0, pos->pos_nodes * sizeof(unsigned int));
+    cproc_timeout = kmalloc(pos->pos_nodes * sizeof(unsigned int), GFP_KERNEL);
+    if (!cproc_timeout)
+    {
+        kfree(ack_errors);
+        kfree(dproc_timeout);
+        return (-EINVAL);
+    }
+    memset(cproc_timeout, 0, pos->pos_nodes * sizeof(unsigned int));
+       
+    kmutex_lock (&dev->dev_lock);
+    dev->dev_position = *pos;
+    dev->dev_ack_errors = ack_errors;
+    dev->dev_dproc_timeout = dproc_timeout;
+    dev->dev_cproc_timeout = cproc_timeout;
+    spin_lock_init(&dev->dev_error_routes_lock);
+
+    if (forceLocal)
+       write_reg32 (dev, LinkContSettings, read_reg32 (dev, LinkContSettings) | LCONT_FORCE_COMMSCLK_LOCAL);
+    else
+       write_reg32 (dev, LinkContSettings, read_reg32 (dev, LinkContSettings) & ~LCONT_FORCE_COMMSCLK_LOCAL);
+
+    pioflush_reg (dev);
+    kmutex_unlock (&dev->dev_lock);
+
+    return (0);
+}
+
+void
+elan4_get_params (ELAN4_DEV *dev, ELAN_PARAMS *params, unsigned short *mask)
+{
+    kmutex_lock (&dev->dev_lock);
+
+    *mask = dev->dev_devinfo.dev_params_mask;
+    memcpy (params, &dev->dev_devinfo.dev_params, sizeof (ELAN_PARAMS));
+    
+    kmutex_unlock (&dev->dev_lock);
+}
+
+void
+elan4_set_params (ELAN4_DEV *dev, ELAN_PARAMS *params, unsigned short mask)
+{      
+    int i;
+
+    kmutex_lock (&dev->dev_lock);
+    for (i = 0; i < ELAN4_PARAM_COUNT; i++)
+       if (mask & (1 << i))
+           dev->dev_devinfo.dev_params.values[i] = params->values[i];
+    
+    dev->dev_devinfo.dev_params_mask |= mask;
+    kmutex_unlock (&dev->dev_lock);
+}
+
+
+EXPORT_SYMBOL(elan4_get_position);
+EXPORT_SYMBOL(elan4_set_position);
+
+EXPORT_SYMBOL(elan4_queue_haltop);
+EXPORT_SYMBOL(elan4_queue_dma_flushop);
+EXPORT_SYMBOL(elan4_queue_mainintop);
+
+EXPORT_SYMBOL(elan4_insertctxt);
+EXPORT_SYMBOL(elan4_removectxt);
+
+EXPORT_SYMBOL(elan4_attach_filter);
+EXPORT_SYMBOL(elan4_detach_filter);
+EXPORT_SYMBOL(elan4_set_filter);
+EXPORT_SYMBOL(elan4_set_routetable);
+
+EXPORT_SYMBOL(elan4_alloccq);
+EXPORT_SYMBOL(elan4_freecq);
+EXPORT_SYMBOL(elan4_restartcq);
+
+EXPORT_SYMBOL(elan4_flush_icache);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/device_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/device_Linux.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/device_Linux.c 2005-06-01 23:12:54.606438040 -0400
@@ -0,0 +1,2625 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: device_Linux.c,v 1.74.6.9 2005/01/18 14:44:11 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/device_Linux.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+#include <qsnet/kpte.h>
+
+#include <asm/io.h>
+#include <asm/irq.h>
+#ifdef CONFIG_MTRR
+#include <asm/mtrr.h>
+#endif
+
+#include <linux/init.h>
+#include <linux/pci.h>
+#include <linux/module.h>
+#include <linux/reboot.h>
+#include <linux/notifier.h>
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
+#include <linux/wrapper.h>
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,4,23)
+typedef void irqreturn_t;
+#endif
+#       define IRQ_NONE
+#       define IRQ_HANDLED
+#endif
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/user.h>
+#include <elan4/ioctl.h>
+#include <elan4/intcookie.h>
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,4,0)
+#error please use a 2.4.0 series kernel or newer
+#endif
+
+
+#if defined(LINUX_SPARC) || defined(LINUX_PPC64)
+#define __io_remap_page_range(from,offset,size,prot)   remap_page_range(from,offset,size,prot)
+#define __remap_page_range(from,offset,size,prot)      remap_page_range(from,offset,size,prot)
+#elif defined(NO_RMAP)
+#define __io_remap_page_range(from,offset,size,prot)   io_remap_page_range(from,offset,size,prot)
+#define __remap_page_range(from,offset,size,prot)      remap_page_range(from,offset,size,prot)
+#else
+#define __io_remap_page_range(from,offset,size,prot)   io_remap_page_range(vma,from,offset,size,prot)
+#define __remap_page_range(from,offset,size,prot)      remap_page_range(vma,from,offset,size,prot)
+#endif
+
+#ifndef pgprot_noncached
+static inline pgprot_t pgprot_noncached(pgprot_t _prot)
+{
+       unsigned long prot = pgprot_val(_prot);
+#if defined(__powerpc__)
+       prot |= _PAGE_NO_CACHE | _PAGE_GUARDED;
+#elif defined(__sparc__)
+       prot &= ~(_PAGE_CACHE);
+       prot |= _PAGE_IE;
+#endif
+
+       return __pgprot(prot);
+}
+#endif
+
+#ifndef pgprot_writecombine
+static inline pgprot_t pgprot_writecombine (pgprot_t _prot)
+{
+    return _prot;
+}
+#endif
+
+#define ELAN4_DRIVER_VERSION           0x103           /* 16 bit value */
+
+/*
+ * Function prototypes.
+ */
+static int        elan4_attach_device (int instance, struct pci_dev *pdev);
+static void       elan4_detach_device (ELAN4_DEV *dev);
+
+static int        elan4_open (struct inode *inode, struct file *file);
+static int        elan4_release(struct inode *inode, struct file *file);
+static int        elan4_ioctl (struct inode *inode, struct file *file, 
+                               unsigned int cmd, unsigned long arg);
+static int        elan4_mmap (struct file *file, struct vm_area_struct *vm_area);
+
+static irqreturn_t elan4_irq (int irq, void *arg, struct pt_regs *regs);
+
+static void        elan4_shutdown_devices(int panicing);
+
+static int      disabled;                                      /* bitmask of which devices not to start */
+unsigned int   elan4_pll_cfg      = 0;
+int            elan4_pll_div      = 31;                        /* RevC PCB */
+int            elan4_mod45disable = 0;
+static int      optimise_pci_bus   = 1;                                /* 0 => don't, 1 => if ok, 2 => always */
+static int      default_features   = 0;                                /* default values for dev_features */
+
+long long       sdram_cfg = SDRAM_STARTUP_VALUE;
+static int      sdram_cfg_lo;
+static int     sdram_cfg_hi;
+int            sdram_bank_limit;
+
+MODULE_AUTHOR("Quadrics Ltd.");
+MODULE_DESCRIPTION("Elan 4 Device Driver");
+MODULE_LICENSE("GPL");
+
+MODULE_PARM(elan4_debug, "i");
+MODULE_PARM(elan4_debug_toconsole, "i");
+MODULE_PARM(elan4_debug_tobuffer, "i");
+MODULE_PARM(elan4_debug_mmu, "i");
+MODULE_PARM(elan4_pll_cfg, "i");
+MODULE_PARM(elan4_pll_div, "i");
+MODULE_PARM(elan4_mod45disable, "i");
+MODULE_PARM(optimise_pci_bus, "i");
+MODULE_PARM(default_features, "i");
+
+MODULE_PARM(disabled, "i");
+MODULE_PARM(sdram_cfg_lo, "i");
+MODULE_PARM(sdram_cfg_hi, "i");
+MODULE_PARM(sdram_bank_limit, "i");
+
+MODULE_PARM(elan4_hash_0_size_val, "i");
+MODULE_PARM(elan4_hash_1_size_val, "i");
+MODULE_PARM(elan4_ctxt_table_shift, "i");
+MODULE_PARM(elan4_ln2_max_cqs, "i");
+MODULE_PARM(elan4_dmaq_highpri_size, "i");
+MODULE_PARM(elan4_threadq_highpri_size, "i");
+MODULE_PARM(elan4_dmaq_lowpri_size, "i");
+MODULE_PARM(elan4_threadq_lowpri_size, "i");
+MODULE_PARM(elan4_interruptq_size, "i");
+
+MODULE_PARM(elan4_mainint_punt_loops, "i");
+MODULE_PARM(elan4_mainint_resched_ticks, "i");
+
+MODULE_PARM(user_p2p_route_options, "i");
+MODULE_PARM(user_bcast_route_options, "i");
+MODULE_PARM(user_dproc_retry_count, "i");
+MODULE_PARM(user_cproc_retry_count, "i");
+
+/*
+ * Standard device entry points.
+ */
+static struct file_operations elan4_fops = {
+    ioctl:   elan4_ioctl,
+    mmap:    elan4_mmap,
+    open:    elan4_open,
+    release: elan4_release,
+};
+
+ELAN4_DEV *elan4_devices[ELAN4_MAX_CONTROLLER];
+
+#if defined(CONFIG_DEVFS_FS)
+static devfs_handle_t devfs_handle;
+#endif
+
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+static int
+elan4_ioctl32_cmds[] =
+{      /* /dev/elan/control */
+    ELAN4IO_DEVINFO,
+    ELAN4IO_GET_POSITION,
+    ELAN4IO_SET_POSITION,
+    ELAN4IO_GET_PARAMS,
+    ELAN4IO_SET_PARAMS,
+
+    /* /dev/elan4/user */
+    ELAN4IO_POSITION,
+    ELAN4IO_FREE,
+    ELAN4IO_ATTACH,
+    ELAN4IO_DETACH,
+    ELAN4IO_BLOCK_INPUTTER,
+
+    ELAN4IO_ADD_P2PVP,
+    ELAN4IO_ADD_BCASTVP,
+    ELAN4IO_REMOVEVP,
+    ELAN4IO_SET_ROUTE,
+    ELAN4IO_RESET_ROUTE,
+    ELAN4IO_GET_ROUTE,
+    ELAN4IO_CHECK_ROUTE,
+
+    ELAN4IO_ALLOCCQ,
+    ELAN4IO_FREECQ,
+    ELAN4IO_SETPERM32,
+    ELAN4IO_CLRPERM32,
+    ELAN4IO_TRAPSIG,
+    ELAN4IO_TRAPHANDLER32,
+    ELAN4IO_REQUIRED_MAPPINGS,
+       
+    ELAN4IO_RESUME_EPROC_TRAP,
+    ELAN4IO_RESUME_CPROC_TRAP,
+    ELAN4IO_RESUME_DPROC_TRAP,
+    ELAN4IO_RESUME_TPROC_TRAP,
+    ELAN4IO_RESUME_IPROC_TRAP,
+
+    ELAN4IO_FLUSH_ICACHE,
+
+    ELAN4IO_STOP_CTXT,
+
+    ELAN4IO_ALLOC_INTCOOKIE,
+    ELAN4IO_FREE_INTCOOKIE,
+    ELAN4IO_ARM_INTCOOKIE,
+    ELAN4IO_WAIT_INTCOOKIE,
+
+    ELAN4IO_ALLOC_TRAP_QUEUES,
+    ELAN4IO_NETERR_MSG,
+    ELAN4IO_NETERR_TIMER,
+    ELAN4IO_NETERR_FIXUP,
+
+    ELAN4IO_DUMPCQ32,
+};
+
+static int      elan4_ioctl32 (unsigned int fd, unsigned int cmd, 
+                              unsigned long arg, struct file *file);
+#endif
+
+/*
+ * Standard device entry points.
+ */
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+
+#include <linux/dump.h>
+
+static int
+elan4_dump_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    if (event == DUMP_BEGIN)
+       elan4_shutdown_devices (FALSE);
+
+    return (NOTIFY_DONE);
+}
+static struct notifier_block elan4_dump_notifier = 
+{
+    notifier_call:     elan4_dump_event,
+    priority:          0,
+};
+
+#endif
+
+static int
+elan4_reboot_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    if ((event == SYS_RESTART || event == SYS_HALT || event == SYS_POWER_OFF))
+       elan4_shutdown_devices (0);
+
+    return (NOTIFY_DONE);
+}
+
+static struct notifier_block elan4_reboot_notifier = 
+{
+    notifier_call:     elan4_reboot_event,
+    priority:          0,
+};
+
+static int
+elan4_panic_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    elan4_shutdown_devices (1);
+
+    return (NOTIFY_DONE);
+}
+
+static struct notifier_block elan4_panic_notifier = 
+{
+    notifier_call:     elan4_panic_event,
+    priority:          0,
+};
+
+static int __init
+elan4_init (void)
+{
+    int             err;
+    struct pci_dev *pdev;
+    int                    count;
+#if defined(__ia64)
+    int             seenRevA = 0;
+#endif
+    
+    if ((err = register_chrdev (ELAN4_MAJOR, ELAN4_NAME, &elan4_fops)) < 0)
+       return (err);
+
+#if defined(CONFIG_DEVFS_FS)
+    devfs_handle = devfs_mk_dir (NULL, "elan4", NULL);
+#endif
+
+    intcookie_init();
+    elan4_debug_init();
+    elan4_procfs_init();
+    
+#ifdef CONFIG_MPSAS
+    sas_init();
+#endif
+
+    if (sdram_cfg_lo != 0 && sdram_cfg_hi != 0)
+       sdram_cfg = (((unsigned long long) sdram_cfg_hi) << 32) | ((unsigned long long) sdram_cfg_lo);
+
+    for (count = 0, pdev = NULL; (pdev = pci_find_device(PCI_VENDOR_ID_QUADRICS, PCI_DEVICE_ID_ELAN4, pdev)) != NULL ; count++)
+    {
+#if defined(__ia64)
+       unsigned char revid;
+       
+       pci_read_config_byte (pdev, PCI_REVISION_ID, &revid);
+
+       if (revid == PCI_REVISION_ID_ELAN4_REVA && seenRevA++ != 0 && pci_find_device (PCI_VENDOR_ID_HP, 0x122e, NULL))
+       {
+           printk ("elan: only a single elan4a supported on rx2600\n");
+           continue;
+       }
+#endif
+
+       if (count < ELAN4_MAX_CONTROLLER)
+           elan4_attach_device (count, pdev);
+    }
+
+    if (count >= ELAN4_MAX_CONTROLLER)
+       printk ("elan: found %d elan4 devices - only support %d\n", count, ELAN4_MAX_CONTROLLER);
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+    lock_kernel();
+    {
+       extern int register_ioctl32_conversion(unsigned int cmd, int (*handler)(unsigned int, unsigned int, unsigned long, struct file *));
+       register int i;
+       for (i = 0; i < sizeof (elan4_ioctl32_cmds)/sizeof(elan4_ioctl32_cmds[0]); i++)
+           register_ioctl32_conversion (elan4_ioctl32_cmds[i], elan4_ioctl32);
+    }
+    unlock_kernel();
+#endif
+
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+    register_dump_notifier (&elan4_dump_notifier);
+#endif
+    register_reboot_notifier (&elan4_reboot_notifier);
+
+#if !defined(NO_PANIC_NOTIFIER)
+    notifier_chain_register (&panic_notifier_list, &elan4_panic_notifier);
+#endif
+
+    return (0);
+}
+
+#ifdef MODULE
+static void __exit
+elan4_exit (void)
+{
+    int i;
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+    lock_kernel();
+    {
+       extern void unregister_ioctl32_conversion(unsigned int cmd);
+
+       for (i = 0; i < sizeof (elan4_ioctl32_cmds)/sizeof(elan4_ioctl32_cmds[0]); i++)
+           unregister_ioctl32_conversion (elan4_ioctl32_cmds[i]);
+    }
+    unlock_kernel();
+#endif
+
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+    unregister_dump_notifier (&elan4_dump_notifier);
+#endif
+    unregister_reboot_notifier (&elan4_reboot_notifier);
+
+#if !defined(NO_PANIC_NOTIFIER)
+    notifier_chain_unregister (&panic_notifier_list, &elan4_panic_notifier);
+#endif
+
+    for (i = 0; i < ELAN4_MAX_CONTROLLER; i++)
+       if (elan4_devices[i] != NULL)
+           elan4_detach_device (elan4_devices[i]);
+    
+    elan4_procfs_fini();
+    elan4_debug_fini();
+    intcookie_fini();
+
+#if defined(CONFIG_DEVFS_FS)
+    devfs_unregister (devfs_handle);
+#endif
+
+    unregister_chrdev(ELAN4_MAJOR, ELAN4_NAME);
+}
+
+module_init (elan4_init);
+module_exit (elan4_exit);
+
+#else
+__initcall (elan4_init);
+#endif
+
+/*
+ * Minor numbers encoded as :
+ *   [5:0]     device number
+ *   [15:6]    function number
+ */
+#define ELAN4_DEVICE_MASK      0x3F
+#define ELAN4_DEVICE(inode)    (MINOR((inode)->i_rdev) & ELAN4_DEVICE_MASK)
+
+#define ELAN4_MINOR_CONTROL    0
+#define ELAN4_MINOR_MEM                1
+#define ELAN4_MINOR_USER       2
+
+#define ELAN4_MINOR_SHIFT      6
+#define ELAN4_MINOR(inode)     (MINOR((inode)->i_rdev) >> ELAN4_MINOR_SHIFT)
+
+/*
+ * Called by init_module() for each card discovered on PCI.
+ */
+static int
+elan4_attach_device (int instance, struct pci_dev *pdev)
+{
+    ELAN4_DEV *dev;
+    int res;
+
+    if ((dev = (ELAN4_DEV *) kmalloc (sizeof (ELAN4_DEV), GFP_KERNEL)) == NULL)
+       return (-ENOMEM);
+    memset (dev, 0, sizeof (ELAN4_DEV));
+
+    /* setup os dependent section of ELAN4_DEV */
+    dev->dev_instance   = instance;
+    dev->dev_osdep.pdev = pdev;
+    dev->dev_features   = default_features;
+
+    /* initialise the devinfo */
+    pci_read_config_word (dev->dev_osdep.pdev, PCI_VENDOR_ID,   &dev->dev_devinfo.dev_vendor_id);
+    pci_read_config_word (dev->dev_osdep.pdev, PCI_DEVICE_ID,   &dev->dev_devinfo.dev_device_id);
+    pci_read_config_byte (dev->dev_osdep.pdev, PCI_REVISION_ID, &dev->dev_devinfo.dev_revision_id);
+
+    dev->dev_devinfo.dev_rail                 = instance;
+    dev->dev_devinfo.dev_driver_version       = ELAN4_DRIVER_VERSION;
+    dev->dev_devinfo.dev_num_down_links_value = 0;
+
+    dev->dev_position.pos_mode = ELAN_POS_UNKNOWN;
+
+    /* initialise the data structures and map the device */
+    if ((res = elan4_initialise_device (dev)) != 0)
+    {
+       kfree (dev);
+       return res;
+    }
+
+    /* add the interrupt handler */
+    if (request_irq (pdev->irq, elan4_irq, SA_SHIRQ, "elan4", dev) != 0)
+    {
+       elan4_finalise_device (dev);
+       kfree (dev);
+       return -ENXIO;
+    }
+
+    if (pci_request_regions(dev->dev_osdep.pdev, "elan4"))
+    {
+       free_irq (dev->dev_osdep.pdev->irq, dev);
+       kfree (dev);
+       return -ENODEV;
+    }
+
+#if defined(CONFIG_DEVFS_FS)
+    {
+       char name[16];
+       
+       sprintf (name, "control%d", dev->dev_instance);
+       dev->dev_osdep.devfs_control = devfs_register(devfs_handle, name, DEVFS_FL_NONE, ELAN4_MAJOR,
+                                                     dev->dev_instance | (ELAN4_MINOR_CONTROL << ELAN4_MINOR_SHIFT), S_IFCHR | S_IRUSR | S_IWUSR, 
+                                                     &elan4_fops, NULL);
+       sprintf (name, "sdram%d", dev->dev_instance);
+       dev->dev_osdep.devfs_sdram =  devfs_register(devfs_handle, name, DEVFS_FL_NONE, ELAN4_MAJOR,
+                                                    dev->dev_instance | (ELAN4_MINOR_MEM << ELAN4_MINOR_SHIFT), S_IFCHR | S_IRUSR|S_IWUSR | S_IRGRP|S_IWGRP | S_IROTH|S_IWOTH,
+                                                    &elan4_fops, NULL);
+       sprintf (name, "user%d", dev->dev_instance);
+       dev->dev_osdep.devfs_user =  devfs_register(devfs_handle, name, DEVFS_FL_NONE, ELAN4_MAJOR,
+                                                   dev->dev_instance | (ELAN4_MINOR_USER << ELAN4_MINOR_SHIFT), S_IFCHR | S_IRUSR|S_IWUSR | S_IRGRP|S_IWGRP | S_IROTH|S_IWOTH,
+                                                   &elan4_fops, NULL);
+    }
+#endif
+
+    /* add the procfs entry */
+    elan4_procfs_device_init (dev);
+
+    /* allow the device to be referenced now */
+    elan4_devices[instance] = dev;
+
+    if ((disabled & (1 << instance)) == 0)
+    {
+       if (elan4_start_device (dev) != 0)
+       {
+           printk ("elan%d: auto-start of device failed\n", dev->dev_instance);
+
+           elan4_detach_device (dev);
+           return (-ENXIO);
+       }
+       
+       dev->dev_state = ELAN4_STATE_STARTED;
+    }
+
+#if defined (__sparc)
+    printk ("elan%d: at pci %s (irq = %s)\n", instance, pdev->slot_name, __irq_itoa(pdev->irq));
+#else
+    printk ("elan%d: at pci %s (irq = %d)\n", instance, pdev->slot_name, pdev->irq);
+#endif
+
+    return (0);
+}
+
+/*
+ * Called by cleanup_module() for each board found on PCI.
+ */
+static void
+elan4_detach_device (ELAN4_DEV *dev)
+{
+    /* stop the chip and free of resources */
+    if (dev->dev_state == ELAN4_STATE_STARTED)
+       elan4_stop_device (dev);
+    
+    elan4_devices[dev->dev_instance] = NULL;
+
+#if defined(CONFIG_DEVFS_FS)
+    devfs_unregister (dev->dev_osdep.devfs_control);
+    devfs_unregister (dev->dev_osdep.devfs_sdram);
+    devfs_unregister (dev->dev_osdep.devfs_user);
+#endif
+
+    /* release the address space */
+    pci_release_regions (dev->dev_osdep.pdev);
+
+    /* release the interrupt */
+    free_irq (dev->dev_osdep.pdev->irq, dev);
+
+    /* remove the procfs entry */
+    elan4_procfs_device_fini (dev);
+
+    /* unmap the device and finalise the data structures */
+    elan4_finalise_device (dev);
+    
+    kfree (dev);
+}
+
+/*
+ * Maintain reference counts on the device
+ */
+ELAN4_DEV *
+elan4_reference_device (int instance, int state)
+{
+    ELAN4_DEV *dev = elan4_devices[instance];
+
+    if (dev == NULL)
+       return (NULL);
+
+    kmutex_lock (&dev->dev_lock);
+
+    if ((dev->dev_state & state) == 0)
+    {
+       kmutex_unlock (&dev->dev_lock);
+       return (NULL);
+    }
+
+    dev->dev_references++;
+    kmutex_unlock (&dev->dev_lock);
+
+#ifdef MODULE
+    MOD_INC_USE_COUNT;
+#endif
+
+#ifdef CONFIG_MPSAS
+    sas_set_position(dev);
+#endif
+
+    return (dev);
+}
+
+void
+elan4_dereference_device (ELAN4_DEV *dev)
+{
+    kmutex_lock (&dev->dev_lock);
+    dev->dev_references--;
+    kmutex_unlock (&dev->dev_lock);
+
+#ifdef MODULE
+    MOD_DEC_USE_COUNT;
+#endif
+}
+
+static void
+elan4_shutdown_devices(int panicing)
+{
+    ELAN4_DEV *dev;
+    unsigned long flags;
+    register int i;
+
+    local_irq_save (flags);
+    for (i = 0; i < ELAN4_MAX_CONTROLLER; i++)
+    {
+       if ((dev = elan4_devices[i]) != NULL)
+       {
+           printk(KERN_INFO "elan%d: forcing link into reset\n", dev->dev_instance);
+
+           /* set the inputters to discard everything */
+           if (! panicing) spin_lock (&dev->dev_haltop_lock);
+
+           if (dev->dev_discard_lowpri_count++ == 0)
+               elan4_set_schedstatus (dev, 0);
+           if (dev->dev_discard_highpri_count++ == 0)
+               elan4_set_schedstatus (dev, 0);
+
+           if (! panicing) spin_unlock (&dev->dev_haltop_lock);
+
+           /* ideally we'd like to halt all the outputters too,
+            * however this will prevent the kernel comms flushing
+            * to work correctly .....
+            */
+       }
+    }
+    local_irq_restore (flags);
+}
+
+/*
+ * /dev/elan4/controlX - control device
+ *
+ */
+static int
+control_open (struct inode *inode, struct file *file)
+{
+    ELAN4_DEV       *dev = elan4_reference_device (ELAN4_DEVICE(inode), ELAN4_STATE_STOPPED | ELAN4_STATE_STARTED);
+    CONTROL_PRIVATE *pr;
+    
+    if (dev == NULL)
+       return (-ENXIO);
+    
+    if ((pr = (CONTROL_PRIVATE *) kmalloc (sizeof (CONTROL_PRIVATE), GFP_KERNEL)) == NULL)
+    {
+       elan4_dereference_device (dev);
+       
+       return (-ENOMEM);
+    }
+
+    PRINTF (DBG_USER, DBG_FILE, "control_open: dev=%p pr=%p\n", dev, pr);
+
+    pr->pr_dev           = dev;
+    pr->pr_boundary_scan = 0;
+
+    file->private_data = (void *) pr;
+
+    return (0);
+}
+
+static int
+control_release (struct inode *inode, struct file *file)
+{
+    CONTROL_PRIVATE *pr  = (CONTROL_PRIVATE *) file->private_data;
+    ELAN4_DEV       *dev = pr->pr_dev;
+
+    PRINTF (DBG_DEVICE, DBG_FILE, "control_release: pr=%p\n", pr);
+
+    //if (pr->pr_boundary_scan)
+    //    elan4_clear_boundary_scan (dev, pr);
+
+    elan4_dereference_device (dev);
+
+    kfree (pr);
+
+    return (0);
+}
+
+static int
+control_ioctl (struct inode *inode, struct file *file, 
+                    unsigned int cmd, unsigned long arg)
+{
+    CONTROL_PRIVATE *pr  = (CONTROL_PRIVATE *) file->private_data;
+
+    PRINTF (DBG_DEVICE, DBG_FILE, "control_ioctl: cmd=%x arg=%lx\n", cmd, arg);
+
+    switch (cmd)
+    {
+    case ELAN4IO_DEVINFO:
+       if (copy_to_user ((void *) arg, &pr->pr_dev->dev_devinfo, sizeof (ELAN_DEVINFO)))
+           return (-EFAULT);
+       return (0);
+
+    case ELAN4IO_GET_POSITION:
+    {
+       ELAN_POSITION pos;
+
+       elan4_get_position (pr->pr_dev, &pos);
+
+       if (copy_to_user ((void *) arg, &pos, sizeof (ELAN_POSITION)))
+           return (-EFAULT);
+
+       return (0);
+    }
+
+    case ELAN4IO_SET_POSITION:
+    {
+       ELAN_POSITION pos;
+
+       if (copy_from_user (&pos, (void *) arg, sizeof (ELAN_POSITION)))
+           return (-EFAULT);
+       
+       return (elan4_set_position (pr->pr_dev, &pos));
+    }
+
+    case ELAN4IO_OLD_GET_PARAMS:
+    {
+       ELAN_PARAMS params;
+       unsigned short mask;
+
+       elan4_get_params (pr->pr_dev, &params, &mask);
+
+       if (copy_to_user ((void *) arg, &params, sizeof (ELAN_PARAMS)))
+           return (-EFAULT);
+
+       return (0);
+    }
+
+    case ELAN4IO_OLD_SET_PARAMS:
+    {
+       ELAN_PARAMS params;
+
+       if (copy_from_user (&params, (void *) arg, sizeof (ELAN_PARAMS)))
+           return (-EFAULT);
+       
+       elan4_set_params (pr->pr_dev, &params, 3);
+       
+       return (0);
+    }
+
+    case ELAN4IO_SET_PARAMS:
+    {
+       ELAN4IO_PARAMS_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_PARAMS_STRUCT)))
+           return (-EFAULT);
+       
+       elan4_set_params (pr->pr_dev, &args.p_params, args.p_mask);
+       
+       return (0);
+    }
+
+    case ELAN4IO_GET_PARAMS:
+    {
+       ELAN4IO_PARAMS_STRUCT args;
+
+       elan4_get_params (pr->pr_dev, &args.p_params, &args.p_mask);
+
+       if (copy_to_user ((void *) arg, &args, sizeof (ELAN_PARAMS)))
+           return (-EFAULT);
+
+       return (0);
+    }
+    }
+
+    return (-EINVAL);
+}
+
+static int
+control_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    CONTROL_PRIVATE *pr  = (CONTROL_PRIVATE *) file->private_data;
+    unsigned        bar = OFF_TO_BAR (vma->vm_pgoff << PAGE_SHIFT);
+    unsigned long    off = OFF_TO_OFFSET (vma->vm_pgoff << PAGE_SHIFT);
+    long            len = vma->vm_end - vma->vm_start;
+
+    PRINTF (DBG_USER, DBG_FILE, "control_mmap: pr=%p bar=%x off=%x\n", pr, bar, off);
+
+    /* check bar number and translate the standard psuedo bars */
+    switch (bar)
+    {
+    case ELAN4_BAR_SDRAM:
+    case ELAN4_BAR_REGISTERS:
+       break;
+
+    default:
+       return (-EINVAL);
+    }
+
+    if (off < 0 || (off + len) > pci_resource_len (pr->pr_dev->dev_osdep.pdev, bar))
+       return (-EINVAL);
+
+    vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+    if (__io_remap_page_range (vma->vm_start, pci_resource_start (pr->pr_dev->dev_osdep.pdev, bar) + off, len, vma->vm_page_prot))
+       return (-EAGAIN);
+
+    return (0);
+}
+
+/*
+ * /dev/elan4/sdramX - sdram access device
+ */
+static void 
+mem_freepage (MEM_PRIVATE *pr, MEM_PAGE *pg)
+{
+    PRINTF (DBG_USER, DBG_MEM, "mem_freepage: pr=%p pgoff=%lx pg=%p ref=%d\n", pr, pg->pg_pgoff, pg, pg->pg_ref);
+
+    elan4_sdram_free (pr->pr_dev, pg->pg_addr, SDRAM_PAGE_SIZE);
+    kfree (pg);
+}
+
+static MEM_PAGE *
+mem_getpage (MEM_PRIVATE *pr, unsigned long pgoff)
+{
+    int       hashval = MEM_HASH (pgoff);
+    MEM_PAGE *npg = NULL;
+    MEM_PAGE *pg;
+
+    ASSERT ((pgoff & SDRAM_PGOFF_OFFSET) == 0);
+
+    PRINTF (DBG_USER, DBG_MEM, "mem_getpage: pr=%p pgoff=%lx\n", pr, pgoff);
+    
+ again:
+    spin_lock (&pr->pr_lock);
+    for (pg = pr->pr_pages[hashval]; pg; pg = pg->pg_next)
+       if (pg->pg_pgoff == pgoff)
+           break;
+    
+    if (pg != NULL)
+    {
+       PRINTF (DBG_USER, DBG_MEM, "mem_getpage: pr=%p pgoff=%lx -> found %p addr=%x\n", pr, pgoff, pg, pg->pg_addr);
+
+       pg->pg_ref++;
+       spin_unlock (&pr->pr_lock);
+
+       if (npg != NULL)                                        /* we'd raced and someone else had created */
+           mem_freepage (pr, npg);                             /* this page - so free of our new one*/
+       return (pg);
+    }
+    
+    if (npg != NULL)                                           /* didn't find the page, so inset the */
+    {                                                          /* new one we've just created */
+       npg->pg_next = pr->pr_pages[hashval];
+       pr->pr_pages[hashval] = npg;
+       
+       spin_unlock (&pr->pr_lock);
+       return (npg);
+    }
+    
+    spin_unlock (&pr->pr_lock);                                        /* drop spinlock before creating a new page */
+    
+    if ((npg = (MEM_PAGE *) kmalloc (sizeof (MEM_PAGE), GFP_KERNEL)) == NULL)
+       return (NULL);
+    
+    if ((npg->pg_addr = elan4_sdram_alloc (pr->pr_dev, SDRAM_PAGE_SIZE)) == 0)
+    {
+       kfree (npg);
+       return (NULL);
+    }
+
+#ifndef CONFIG_MPSAS
+    /* zero the page before returning it to the user */
+    elan4_sdram_zeroq_sdram (pr->pr_dev, npg->pg_addr, SDRAM_PAGE_SIZE);
+#endif
+    
+    npg->pg_pgoff = pgoff;
+    npg->pg_ref   = 1;
+    
+    /* created a new page - so have to rescan before inserting it */
+    goto again;
+}
+
+static void
+mem_droppage (MEM_PRIVATE *pr, unsigned long pgoff, int dontfree)
+{
+    MEM_PAGE **ppg;
+    MEM_PAGE  *pg;
+
+    spin_lock (&pr->pr_lock);
+    for (ppg = &pr->pr_pages[MEM_HASH(pgoff)]; *ppg; ppg = &(*ppg)->pg_next)
+       if ((*ppg)->pg_pgoff == pgoff)
+           break;
+
+    pg = *ppg;
+
+    ASSERT (*ppg != NULL);
+
+    PRINTF (DBG_USER, DBG_MEM, "mem_droppage: pr=%p pgoff=%lx pg=%p ref=%d dontfree=%d\n", pr, pgoff, (*ppg), (*ppg)->pg_ref, dontfree);
+
+    if (--pg->pg_ref == 0 && !dontfree)
+    {
+       *ppg = pg->pg_next;
+
+       mem_freepage (pr, pg);
+    }
+
+    spin_unlock (&pr->pr_lock);
+}
+
+static int
+mem_open (struct inode *inode, struct file *file)
+{
+    ELAN4_DEV   *dev = elan4_reference_device (ELAN4_DEVICE(inode), ELAN4_STATE_STARTED);
+    MEM_PRIVATE *pr;
+    register int i;
+
+    if (dev == NULL)
+       return (-ENXIO);
+
+    if ((pr = (MEM_PRIVATE *) kmalloc (sizeof (MEM_PRIVATE), GFP_KERNEL)) == NULL)
+    {
+       elan4_dereference_device (dev);
+       return (-ENOMEM);
+    }
+
+    spin_lock_init (&pr->pr_lock);
+    pr->pr_dev = dev;
+    for (i = 0; i < MEM_HASH_SIZE; i++)
+       pr->pr_pages[i] = NULL;
+
+    file->private_data = (void *) pr;
+    
+    return (0);
+}
+
+static int
+mem_release (struct inode *node, struct file *file)
+{
+    MEM_PRIVATE *pr = (MEM_PRIVATE *) file->private_data;
+    MEM_PAGE    *pg, *next;
+    int          i;
+
+    /* free off any pages that we'd allocated */
+    spin_lock (&pr->pr_lock);
+    for (i = 0; i < MEM_HASH_SIZE; i++)
+    {
+       for (pg = pr->pr_pages[i]; pg; pg = next)
+       {
+           next = pg->pg_next;
+           mem_freepage (pr, pg);
+       }
+    }
+    spin_unlock (&pr->pr_lock);
+
+    elan4_dereference_device (pr->pr_dev);
+    kfree (pr);
+
+    return (0);
+}
+
+static int
+mem_ioctl (struct inode *inode, struct file *file, 
+                 unsigned int cmd, unsigned long arg)
+{
+    return (-EINVAL);
+}
+
+static void 
+mem_vma_open (struct vm_area_struct *vma)
+{
+    MEM_PRIVATE   *pr = (MEM_PRIVATE *) vma->vm_private_data;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (DBG_USER, DBG_MEM, "mem_vma_open: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+    
+    for (addr = vma->vm_start, pgoff = vma->vm_pgoff; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+       mem_getpage (pr, pgoff & SDRAM_PGOFF_MASK);
+}
+
+static void
+mem_vma_close (struct vm_area_struct *vma)
+{
+    MEM_PRIVATE  *pr  = (MEM_PRIVATE *) vma->vm_private_data;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (DBG_USER, DBG_MEM, "mem_vma_close: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+
+    /* NOTE: the call to close may not have the same vm_start/vm_end values as 
+     *       were passed into mmap()/open() - since if an partial unmap had occurred
+     *       then the vma could have been shrunk or even split.
+     *
+     *       if a the vma is split then an vma_open() will be called for the top
+     *       portion - thus causing the reference counts to become incorrect.
+     *
+     * We drop the reference to any pages we're notified about - so they get freed
+     * earlier than when the device is finally released.
+     */
+    for (pgoff = vma->vm_pgoff, addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+       mem_droppage (pr, pgoff & SDRAM_PGOFF_MASK, 0);
+}
+
+struct vm_operations_struct mem_vm_ops = {
+    open:              mem_vma_open,
+    close:             mem_vma_close,
+};
+
+static int
+mem_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    MEM_PRIVATE  *pr = (MEM_PRIVATE *) file->private_data;
+    MEM_PAGE     *pg;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (DBG_USER, DBG_MEM, "mem_mmap: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, file);
+
+    for (addr = vma->vm_start, pgoff = vma->vm_pgoff; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+    {
+       if ((pg = mem_getpage (pr, pgoff & SDRAM_PGOFF_MASK)) == NULL)
+           goto failed;
+
+       PRINTF (DBG_USER, DBG_MEM, "mem_mmap: addr %lx -> pg=%p sdram=%x+%x bar=%lx\n",
+               addr, pg, pg->pg_addr, (pgoff & SDRAM_PGOFF_OFFSET) * PAGE_SIZE,
+               pci_resource_start (pr->pr_dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+
+       vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+       if (! (pr->pr_dev->dev_features & ELAN4_FEATURE_NO_WRITE_COMBINE))
+           vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
+
+       if (__io_remap_page_range (addr, pci_resource_start (pr->pr_dev->dev_osdep.pdev, ELAN4_BAR_SDRAM) +
+                                pg->pg_addr + (pgoff & SDRAM_PGOFF_OFFSET) * PAGE_SIZE, PAGE_SIZE, vma->vm_page_prot))
+       {
+           mem_droppage (pr, pgoff & SDRAM_PGOFF_MASK, 0); /* drop our reference to this page */
+           goto failed;
+       }
+
+#if defined(conditional_schedule)
+       conditional_schedule();
+#endif
+    }
+
+    /* Don't allow these pages to be swapped out of dumped */
+    vma->vm_flags |= (VM_RESERVED | VM_IO);
+
+    vma->vm_ops          = &mem_vm_ops;
+    vma->vm_file         = file;
+    vma->vm_private_data = (void *) pr;
+
+    return (0);
+
+ failed:
+    /* free of any pages we've already allocated/referenced */
+    while (pgoff-- > vma->vm_pgoff)
+       mem_droppage (pr, pgoff & SDRAM_PGOFF_MASK, 0);
+
+    return (-ENOMEM);
+}
+
+/*
+ * /dev/elan4/userX - control device
+ *
+ */
+static inline void
+user_private_free (USER_PRIVATE *pr)
+{
+    ELAN4_DEV *dev = pr->pr_uctx->uctx_ctxt.ctxt_dev;
+
+    ASSERT (atomic_read (&pr->pr_ref) == 0);
+
+    user_free (pr->pr_uctx);
+    kfree (pr);
+
+    elan4_dereference_device (dev);
+}
+
+static void
+user_coproc_release (void *arg, struct mm_struct *mm)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_release: ref=%d\n", atomic_read (&pr->pr_ref));
+
+    elan4mmu_invalidate_ctxt (&pr->pr_uctx->uctx_ctxt);
+
+    pr->pr_mm = NULL;
+
+    if (atomic_dec_and_test (&pr->pr_ref))
+       user_private_free (pr);
+}
+
+static void
+user_coproc_sync_range (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_sync_range: start=%lx end=%lx\n", start, end);
+
+    /* XXXX: this is intended to sync the modified bit from our page tables,
+     *       into the main cpu's modified bits - however since we do not
+     *       syncronize our modified bit on a coproc_invalidate_page() call,
+     *       then it could get lost if we modify the page after the last
+     *       modification and writepage has occurred. Hence we invalidate
+     *       all translations and allow it to refault.
+     */
+
+    user_unload_main (pr->pr_uctx, start, end - start);
+}
+
+static void
+user_coproc_invalidate_range (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_invalidate_range: start=%lx end=%lx\n", start, end);
+
+    user_unload_main (pr->pr_uctx, start, end - start);
+}
+
+static void
+user_coproc_update_range (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_update_range: start=%lx end=%lx\n", start, end);
+
+#if defined(CONFIG_MPSAS)
+    if (((end - start) >> PAGE_SHIFT) > 16)
+       return;
+#endif
+
+    user_update_main (pr->pr_uctx, mm, start, end - start);
+}
+
+static void
+user_coproc_change_protection (void *arg, struct mm_struct *mm, unsigned long start, unsigned long end, pgprot_t newprot)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_change_protection: start=%lx end=%lx\n", start, end);
+
+    user_unload_main (pr->pr_uctx, start, end - start);
+}
+
+static void
+user_coproc_sync_page (void *arg, struct vm_area_struct *vma, unsigned long addr)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_sync_page: addr=%lx\n", addr);
+
+    user_unload_main (pr->pr_uctx, addr & PAGE_MASK, PAGE_SIZE);
+}
+
+static void
+user_coproc_invalidate_page (void *arg, struct vm_area_struct *vma, unsigned long addr)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_invalidate_page: addr=%lx\n", addr);
+
+    user_unload_main (pr->pr_uctx, addr & PAGE_MASK, PAGE_SIZE);
+}
+
+static void
+user_coproc_update_page (void *arg, struct vm_area_struct *vma, unsigned long addr)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) arg;
+
+    PRINTF (pr->pr_uctx, DBG_IOPROC, "user_coproc_update_page: addr=%lx\n", addr);
+
+    user_update_main (pr->pr_uctx, vma->vm_mm, addr & PAGE_MASK, PAGE_SIZE);
+}
+
+static int
+user_open (struct inode *inode, struct file *file)
+{
+    ELAN4_DEV    *dev;
+    USER_PRIVATE *pr;
+    USER_CTXT    *uctx;
+    
+    PRINTF (DBG_USER, DBG_FILE, "user_open: mm=%p users=%d count=%d\n", current->mm,
+           atomic_read (&current->mm->mm_users), atomic_read (&current->mm->mm_count));
+
+    if ((dev = elan4_reference_device (ELAN4_DEVICE(inode), ELAN4_STATE_STARTED)) == NULL)
+       return (-ENXIO);
+    
+    if ((pr = (USER_PRIVATE *) kmalloc (sizeof (USER_PRIVATE), GFP_KERNEL)) == NULL)
+    {
+       elan4_dereference_device (dev);
+       return (-ENOMEM);
+    }
+
+    uctx = user_alloc (dev);
+
+    if (IS_ERR(uctx))
+    {
+       elan4_dereference_device (dev);
+       kfree (pr);
+
+       return PTR_ERR(uctx);
+    }
+
+    /* initialise refcnt to 2 - one for "file", one for the coproc ops */
+    atomic_set (&pr->pr_ref, 2);
+
+    pr->pr_uctx = uctx;
+    pr->pr_mm   = current->mm;
+
+    /* register a coproc callback to notify us of translation changes */
+    pr->pr_coproc.arg               = (void *) pr;
+    pr->pr_coproc.release           = user_coproc_release;
+    pr->pr_coproc.sync_range        = user_coproc_sync_range;
+    pr->pr_coproc.invalidate_range  = user_coproc_invalidate_range;
+    pr->pr_coproc.update_range      = user_coproc_update_range;
+    pr->pr_coproc.change_protection = user_coproc_change_protection;
+    pr->pr_coproc.sync_page         = user_coproc_sync_page;
+    pr->pr_coproc.invalidate_page   = user_coproc_invalidate_page;
+    pr->pr_coproc.update_page       = user_coproc_update_page;
+    
+    spin_lock (&current->mm->page_table_lock);
+    register_coproc_ops (current->mm, &pr->pr_coproc);
+    spin_unlock (&current->mm->page_table_lock);
+
+    file->private_data = (void *) pr;
+
+    return (0);
+}
+
+static int
+user_release (struct inode *inode, struct file *file)
+{
+    USER_PRIVATE *pr = (USER_PRIVATE *) file->private_data;
+
+    PRINTF (pr->pr_uctx, DBG_FILE, "user_release: ref=%d\n", atomic_read (&pr->pr_ref));
+
+    if (atomic_dec_and_test (&pr->pr_ref))
+       user_private_free (pr);
+
+    return (0);
+}
+
+static int
+user_ioctl (struct inode *inode, struct file *file, 
+           unsigned int cmd, unsigned long arg)
+{
+    USER_PRIVATE *pr   = (USER_PRIVATE *) file->private_data;
+    USER_CTXT    *uctx = pr->pr_uctx;
+    int           res  = 0;
+
+    PRINTF (uctx, DBG_FILE, "user_ioctl: cmd=%x arg=%lx\n", cmd, arg);
+
+    if (current->mm != pr->pr_mm)
+       return (-EINVAL);
+    
+    switch (cmd)
+    {
+    case ELAN4IO_DEVINFO:
+       if (copy_to_user ((void *) arg, &uctx->uctx_ctxt.ctxt_dev->dev_devinfo, sizeof (ELAN_DEVINFO)))
+           return (-EFAULT);
+       return (0);
+
+    case ELAN4IO_POSITION:
+    {
+       ELAN4_DEV *dev = uctx->uctx_ctxt.ctxt_dev;
+
+       if (copy_to_user ((void *) arg, &dev->dev_position, sizeof (ELAN_POSITION)))
+           return (-EFAULT);
+       return (0);
+    }
+
+    case ELAN4IO_FREE:
+       spin_lock (&current->mm->page_table_lock);
+       if (pr->pr_mm != current->mm)
+           spin_unlock (&current->mm->page_table_lock);
+       else
+       {
+           unregister_coproc_ops (current->mm, &pr->pr_coproc);
+           spin_unlock (&current->mm->page_table_lock);
+
+           user_coproc_release (pr, current->mm);
+       }
+       return (0);
+
+    case ELAN4IO_ATTACH:
+    {
+       ELAN_CAPABILITY *cap;
+
+       if ((cap = kmalloc (sizeof (ELAN_CAPABILITY), GFP_KERNEL)) == NULL)
+           return (-ENOMEM);
+
+       if (copy_from_user (cap, (void *) arg, sizeof (ELAN_CAPABILITY)))
+           res = -EFAULT;
+       else if ((res = user_attach (uctx, cap)) == 0 && 
+                copy_to_user ((void *) arg, cap, sizeof (ELAN_CAPABILITY)))
+       {
+           user_detach (uctx, cap);
+           res = -EFAULT;
+       }
+
+       kfree (cap);
+       return (res);
+    }
+
+    case ELAN4IO_DETACH:
+    {
+       ELAN_CAPABILITY *cap;
+
+       if ((cap = kmalloc (sizeof (ELAN_CAPABILITY), GFP_KERNEL)) == NULL)
+           return (-ENOMEM);
+
+       if (copy_from_user (cap, (void *) arg, sizeof (ELAN_CAPABILITY)))
+           res = -EFAULT;
+       else
+           user_detach (uctx, cap);
+
+       kfree (cap);
+       return (res);
+    }
+
+    case ELAN4IO_BLOCK_INPUTTER:
+       user_block_inputter (uctx, arg);
+       return (0);
+
+    case ELAN4IO_ADD_P2PVP:
+    {
+       ELAN4IO_ADD_P2PVP_STRUCT *args;
+       
+       if ((args = kmalloc (sizeof (ELAN4IO_ADD_P2PVP_STRUCT), GFP_KERNEL)) == NULL)
+           return (-ENOMEM);
+
+       if (copy_from_user (args, (void *) arg, sizeof (ELAN4IO_ADD_P2PVP_STRUCT)))
+           res = -EFAULT;
+       else 
+           res = user_add_p2pvp (uctx, args->vp_process, &args->vp_capability);
+       
+       kfree (args);
+       return (res);
+    }
+
+    case ELAN4IO_ADD_BCASTVP:
+    {
+       ELAN4IO_ADD_BCASTVP_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_ADD_BCASTVP_STRUCT)))
+           return (-EFAULT);
+
+       return (user_add_bcastvp (uctx, args.vp_process, args.vp_lowvp, args.vp_highvp));
+    }
+
+    case ELAN4IO_REMOVEVP:
+       return (user_removevp (uctx, arg));
+
+    case ELAN4IO_SET_ROUTE:
+    {
+       ELAN4IO_ROUTE_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_ROUTE_STRUCT)))
+           return (-EFAULT);
+
+       return (user_set_route (uctx, args.rt_process, &args.rt_route));
+    }
+
+    case ELAN4IO_RESET_ROUTE:
+    {
+       ELAN4IO_ROUTE_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_ROUTE_STRUCT)))
+           return (-EFAULT);
+
+       return (user_reset_route (uctx, args.rt_process));
+    }
+
+    case ELAN4IO_GET_ROUTE:
+    {
+       ELAN4IO_ROUTE_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_ROUTE_STRUCT)))
+           return (-EFAULT);
+
+       if ((res = user_get_route (uctx, args.rt_process, &args.rt_route)) == 0 &&
+           copy_to_user ((void *) arg, &args, sizeof (ELAN4IO_ROUTE_STRUCT)))
+           res = -EFAULT;
+
+       return (res);
+    }
+
+    case ELAN4IO_CHECK_ROUTE:
+    {
+       ELAN4IO_ROUTE_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_ROUTE_STRUCT)))
+           return (-EFAULT);
+
+       if ((res = user_check_route (uctx, args.rt_process, &args.rt_route, &args.rt_error)) == 0 &&
+           copy_to_user ((void *) arg, &args, sizeof (ELAN4IO_ROUTE_STRUCT)))
+           res = -EFAULT;
+
+       return (res);
+    }
+       
+    case ELAN4IO_ALLOCCQ:
+    {
+       ELAN4IO_ALLOCCQ_STRUCT args;
+       USER_CQ              *ucq;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_ALLOCCQ_STRUCT)))
+           return (-EFAULT);
+       
+       ucq = user_alloccq (uctx, args.cq_size & CQ_SizeMask, args.cq_perm & CQ_PermissionMask,
+                           (args.cq_type & ELAN4IO_CQ_TYPE_REORDER) ? UCQ_REORDER : 0);
+       if (IS_ERR (ucq))
+           return PTR_ERR (ucq);
+       
+       args.cq_indx = elan4_cq2idx (ucq->ucq_cq);
+       
+       if (copy_to_user ((void *) arg, &args, sizeof (ELAN4IO_ALLOCCQ_STRUCT)))
+       {
+           user_dropcq (uctx, ucq);
+           return (-EFAULT);
+       }
+       
+       /* don't drop the reference on the cq until the context is freed,
+        * or the caller explicitly frees the cq */
+       return (0);
+    }
+       
+    case ELAN4IO_FREECQ:
+    {
+       USER_CQ *ucq;
+       unsigned indx;
+
+       if (copy_from_user (&indx, (void *) arg, sizeof (unsigned)))
+           return (-EFAULT);
+
+       if ((ucq = user_findcq (uctx, indx)) == NULL)           /* can't free unallocated cq */
+           return (-EINVAL);
+       
+       user_dropcq (uctx, ucq);                                /* drop the reference we've just taken */
+
+       if ((ucq->ucq_flags & UCQ_SYSTEM))                      /* can't free device driver cq */
+           return (-EINVAL);
+
+       user_dropcq (uctx, ucq);                                /* and the one held from the alloccq call */
+
+       return (0);
+    }
+
+    case ELAN4IO_DUMPCQ:
+    {
+       ELAN4IO_DUMPCQ_STRUCT args;
+       ELAN4_DEV *dev = uctx->uctx_ctxt.ctxt_dev;
+       USER_CQ *ucq;
+       void *buf;
+       int i;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof(ELAN4IO_DUMPCQ_STRUCT)))
+           return (-EFAULT);
+
+       if ((ucq = user_findcq (uctx, args.cq_indx)) == NULL)
+           return (-EINVAL);
+
+       if (args.bufsize)
+       {
+           E4_uint32 usedBufSize = min(args.cq_size, args.bufsize);
+
+           KMEM_ALLOC (buf, void *, usedBufSize, 0);
+
+           if (buf == NULL)
+               return (-ENOMEM);
+
+           for (i=0; i<usedBufSize; i+=sizeof(int))
+               ((int *)buf)[i/sizeof(int)] = elan4_sdram_readl(dev, ucq->ucq_cq->cq_space + i);
+
+           if (copy_to_user((void *)args.buffer, buf, usedBufSize))
+           {
+               KMEM_FREE(buf, args.bufsize);
+               return (-EFAULT);
+           }
+           KMEM_FREE(buf, usedBufSize);
+           args.bufsize = usedBufSize;
+       }
+
+       args.cq_size = CQ_Size(ucq->ucq_cq->cq_size);
+       args.cq_space = ucq->ucq_cq->cq_space;
+
+
+       if (copy_to_user((void *)arg, &args, sizeof(ELAN4IO_DUMPCQ_STRUCT)))
+       {
+           return (-EFAULT);
+       }
+       
+       user_dropcq (uctx, ucq); /* drop the reference we've just taken */
+
+       return (0);
+    }
+
+    case ELAN4IO_SETPERM:
+    {
+       ELAN4IO_PERM_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_PERM_STRUCT)))
+           return (-EFAULT);
+
+       return (user_setperm (uctx, args.ps_maddr, args.ps_eaddr, args.ps_len, args.ps_perm));
+    }
+
+    case ELAN4IO_CLRPERM:
+    {
+       ELAN4IO_PERM_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_PERM_STRUCT)))
+           return (-EFAULT);
+
+       user_clrperm (uctx, args.ps_eaddr, args.ps_len);
+       return (0);
+    }
+    
+    case ELAN4IO_TRAPSIG:
+    {
+       ELAN4IO_TRAPSIG_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_TRAPSIG_STRUCT)))
+           return (-EFAULT);
+
+       pr->pr_uctx->uctx_trap_pid   = current->pid;
+       pr->pr_uctx->uctx_trap_signo = args.ts_signo;
+       
+       return (0);
+    }
+    
+    case ELAN4IO_TRAPHANDLER:
+    {
+       ELAN4IO_TRAPHANDLER_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_TRAPHANDLER_STRUCT)))
+           return (-EFAULT);
+
+       return (user_trap_handler (pr->pr_uctx, (ELAN4_USER_TRAP *)args.th_trapp, args.th_nticks));
+    }
+
+    case ELAN4IO_REQUIRED_MAPPINGS:
+    {
+       ELAN4IO_REQUIRED_MAPPINGS_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_REQUIRED_MAPPINGS_STRUCT)))
+           return (-EFAULT);
+
+       pr->pr_uctx->uctx_upage_addr    = args.rm_upage_addr;
+       pr->pr_uctx->uctx_trestart_addr = args.rm_trestart_addr;
+
+       return (0);
+    }
+
+    case ELAN4IO_ALLOC_TRAP_QUEUES:
+    {
+       ELAN4IO_ALLOC_TRAP_QUEUES_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_ALLOC_TRAP_QUEUES_STRUCT)))
+           return (-EFAULT);
+
+       return (user_alloc_trap_queues (uctx, args.tq_ndproc_traps, args.tq_neproc_traps, 
+                                       args.tq_ntproc_traps, args.tq_nthreads, args.tq_ndmas));
+    }
+
+    case ELAN4IO_RESUME_EPROC_TRAP:
+    {
+       ELAN4IO_RESUME_EPROC_TRAP_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_RESUME_EPROC_TRAP_STRUCT)))
+           return (-EFAULT);
+
+       return (user_resume_eproc_trap (pr->pr_uctx, args.rs_addr));
+    }
+
+    case ELAN4IO_RESUME_CPROC_TRAP:
+    {
+       ELAN4IO_RESUME_CPROC_TRAP_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_RESUME_CPROC_TRAP_STRUCT)))
+           return (-EFAULT);
+
+       return (user_resume_cproc_trap (pr->pr_uctx, args.rs_indx));
+    }
+
+    case ELAN4IO_RESUME_DPROC_TRAP:
+    {
+       ELAN4IO_RESUME_DPROC_TRAP_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_RESUME_DPROC_TRAP_STRUCT)))
+           return (-EFAULT);
+
+       return (user_resume_dproc_trap (pr->pr_uctx, &args.rs_desc));
+    }
+
+    case ELAN4IO_RESUME_TPROC_TRAP:
+    {
+       ELAN4IO_RESUME_TPROC_TRAP_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_RESUME_TPROC_TRAP_STRUCT)))
+           return (-EFAULT);
+
+       return (user_resume_tproc_trap (pr->pr_uctx, &args.rs_regs));
+    }
+
+    case ELAN4IO_RESUME_IPROC_TRAP:
+    {
+       ELAN4IO_RESUME_IPROC_TRAP_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_RESUME_IPROC_TRAP_STRUCT)))
+           return (-EFAULT);
+
+       return (user_resume_iproc_trap (pr->pr_uctx, args.rs_channel, args.rs_trans, 
+                                       &args.rs_header, &args.rs_data));
+    }
+
+    case ELAN4IO_FLUSH_ICACHE:
+       elan4_flush_icache (&uctx->uctx_ctxt);
+       return (0);
+
+    case ELAN4IO_STOP_CTXT:
+       if (arg)
+           user_swapout (uctx, UCTX_USER_STOPPED);
+       else
+           user_swapin (uctx, UCTX_USER_STOPPED);
+       return (0);
+
+    case ELAN4IO_ALLOC_INTCOOKIE_TABLE:
+    {
+       ELAN_CAPABILITY *cap;
+       INTCOOKIE_TABLE *tbl;
+
+       if ((cap = kmalloc (sizeof (ELAN_CAPABILITY), GFP_KERNEL)) == NULL)
+           return (-ENOMEM);
+
+       if (copy_from_user (cap, (void *) arg, sizeof (ELAN_CAPABILITY)))
+           res = -EFAULT;
+       else
+       {
+           tbl = intcookie_alloc_table(cap);
+
+           if (tbl == NULL)
+               res = -ENOMEM;
+           else
+           {
+               /* Install the intcookie table we've just created */
+               spin_lock (&uctx->uctx_spinlock);
+               if (uctx->uctx_intcookie_table != NULL)
+                   res = -EBUSY;
+               else
+                   uctx->uctx_intcookie_table = tbl;
+               spin_unlock (&uctx->uctx_spinlock);
+               
+               /* drop the table we created if there already was one */
+               if (res != 0)
+                   intcookie_free_table (tbl);
+           }
+       }
+
+       kfree (cap);
+
+       return (res);
+    }
+
+    case ELAN4IO_FREE_INTCOOKIE_TABLE:
+    {
+       INTCOOKIE_TABLE *tbl;
+
+       spin_lock (&uctx->uctx_spinlock);
+       tbl = uctx->uctx_intcookie_table;
+       uctx->uctx_intcookie_table = NULL;
+       spin_unlock (&uctx->uctx_spinlock);
+
+       if (tbl != NULL)
+           intcookie_free_table (tbl);
+
+       return (tbl == NULL ? -EINVAL : 0);
+    }
+
+    case ELAN4IO_ALLOC_INTCOOKIE:
+    {
+       /* For backwards compatibility with the old libs (pre 1.8.0)
+        * we allocate an intcookie table on the first cookie
+        * alloc if one hasn't be created already
+        */
+       if (uctx->uctx_intcookie_table == NULL)
+       {
+           ELAN_CAPABILITY *cap;
+           INTCOOKIE_TABLE *tbl;
+           
+           if ((cap = kmalloc (sizeof (ELAN_CAPABILITY), GFP_KERNEL)) == NULL)
+               return (-ENOMEM);
+
+           /* Create a dummy capability */
+           elan_nullcap(cap);
+
+           /* Must be unique for each process on a node */
+           cap->cap_mycontext = (int) ELAN4_TASK_HANDLE();
+
+           /* Create a new intcookie table */
+           tbl = intcookie_alloc_table(cap);
+
+           /* Hang intcookie table off uctx */
+           spin_lock (&uctx->uctx_spinlock);
+           if (uctx->uctx_intcookie_table == NULL)
+           {
+               uctx->uctx_intcookie_table = tbl;
+               spin_unlock (&uctx->uctx_spinlock);
+           }
+           else
+           {
+               spin_unlock (&uctx->uctx_spinlock);
+               intcookie_free_table(tbl);
+           }
+
+           kfree(cap);
+       }
+       
+       return (intcookie_alloc (uctx->uctx_intcookie_table, arg));
+    }
+
+    case ELAN4IO_FREE_INTCOOKIE:
+       if (uctx->uctx_intcookie_table == NULL)
+           return -EINVAL;
+       else
+           return (intcookie_free (uctx->uctx_intcookie_table, arg));
+
+    case ELAN4IO_ARM_INTCOOKIE:
+       if (uctx->uctx_intcookie_table == NULL)
+           return -EINVAL;
+       else
+           return (intcookie_arm (uctx->uctx_intcookie_table, arg));
+
+    case ELAN4IO_WAIT_INTCOOKIE:
+       if (uctx->uctx_intcookie_table == NULL)
+           return -EINVAL;
+       else
+           return (intcookie_wait (uctx->uctx_intcookie_table, arg));
+
+    case ELAN4IO_FIRE_INTCOOKIE:
+    {
+       ELAN4IO_FIRECAP_STRUCT *args;
+
+       if ((args = kmalloc (sizeof (ELAN4IO_FIRECAP_STRUCT), GFP_KERNEL)) == NULL)
+           return (-ENOMEM);
+
+       if (copy_from_user (args, (void *) arg, sizeof (ELAN4IO_FIRECAP_STRUCT)))
+           res = -EFAULT;
+       else
+           res = intcookie_fire_cap (&args->fc_capability, args->fc_cookie);
+       
+       kfree (args);
+
+       return (res);
+    }
+
+    case ELAN4IO_NETERR_MSG:
+    {
+       ELAN4IO_NETERR_MSG_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_NETERR_MSG_STRUCT)))
+           return (-EFAULT);
+
+       return (user_send_neterr_msg (uctx, args.nm_vp, args.nm_nctx, args.nm_retries, &args.nm_msg));
+    }
+
+    case ELAN4IO_NETERR_TIMER:
+    {
+       unsigned long ticks = ((unsigned long) arg * HZ) / 1000;
+
+       PRINTF (uctx, DBG_NETERR, "elan4_neterr_timer: arg %ld inc %ld\n", arg, ticks);
+
+       mod_timer (&uctx->uctx_neterr_timer, (jiffies + (ticks > 0 ? ticks : 1)));
+       return 0;
+    }
+               
+    case ELAN4IO_NETERR_FIXUP:
+    {
+       ELAN4IO_NETERR_FIXUP_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_NETERR_FIXUP_STRUCT)))
+           return (-EFAULT);
+
+       if (args.nf_sten)
+           return (user_neterr_sten (uctx, args.nf_vp, args.nf_cookie, args.nf_waitforeop));
+       else
+           return (user_neterr_dma  (uctx, args.nf_vp, args.nf_cookie, args.nf_waitforeop));
+    }
+    default:
+       PRINTF (uctx, DBG_FILE, "user_ioctl: invalid ioctl %x\n", cmd);
+       return (-EINVAL);
+    }
+}
+
+static void
+user_vma_open (struct vm_area_struct *vma)
+{
+    USER_PRIVATE *pr   = (USER_PRIVATE *) vma->vm_private_data;
+    USER_CTXT    *uctx = pr->pr_uctx;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (uctx, DBG_FILE, "user_vma_open: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+
+    for (addr = vma->vm_start, pgoff = vma->vm_pgoff; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+       elan4_getcqa (&uctx->uctx_ctxt, pgoff);
+}
+
+static void 
+user_vma_close (struct vm_area_struct *vma)
+{
+    USER_PRIVATE *pr   = (USER_PRIVATE *) vma->vm_private_data;
+    USER_CTXT    *uctx = pr->pr_uctx;
+    unsigned long addr;
+    unsigned long pgoff;
+
+    PRINTF (uctx, DBG_FILE, "user_vma_close: vm_mm=%p start=%lx end=%lx pgoff=%lx file=%p\n",
+           vma->vm_mm, vma->vm_start, vma->vm_end, vma->vm_pgoff, vma->vm_file);
+
+    /* NOTE: the same comments apply as mem_vma_close */
+    for (addr = vma->vm_start, pgoff = vma->vm_pgoff; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+       if (elan4_getcqa (&uctx->uctx_ctxt, pgoff) != NULL)
+       {
+           elan4_putcqa (&uctx->uctx_ctxt, pgoff);                     /* drop the reference we've just taken */
+           elan4_putcqa (&uctx->uctx_ctxt, pgoff);                     /* and the one held by the mmap */
+       }
+}
+
+struct vm_operations_struct user_vm_ops = {
+    open:              user_vma_open,
+    close:             user_vma_close,
+};
+
+static int
+user_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    USER_PRIVATE *pr    = (USER_PRIVATE *) file->private_data;
+    USER_CTXT    *uctx  = pr->pr_uctx;
+    ELAN4_DEV     *dev   = uctx->uctx_ctxt.ctxt_dev;
+    ELAN4_CQA     *cqa;
+    unsigned long addr;
+    unsigned long pgoff;
+    int           res;
+    ioaddr_t      ioaddr;
+    
+    for (addr = vma->vm_start, pgoff = vma->vm_pgoff; addr < vma->vm_end; addr += PAGE_SIZE, pgoff++)
+    {
+       switch (pgoff)
+       {
+       default:
+           PRINTF (uctx, DBG_FILE, "user_mmap: command queue %ld mapping at %lx\n",  pgoff, addr);
+           
+           if ((cqa = elan4_getcqa (&uctx->uctx_ctxt, pgoff)) == NULL)
+           {
+               res = -EINVAL;
+               goto failed;
+           }
+
+           PRINTF (uctx, DBG_FILE, "user_mmap: cqa=%p idx=%d num=%d ref=%d\n", cqa, cqa->cqa_idx, cqa->cqa_cqnum, cqa->cqa_ref);
+    
+           vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+           if (! (dev->dev_features & ELAN4_FEATURE_NO_WRITE_COMBINE) && (cqa->cqa_type & CQ_Reorder) != 0)
+               vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
+
+           PRINTF (uctx, DBG_FILE, "user_mmap: remap_page_range (%lx, %lx, %lx, %lx)\n",
+                   addr, pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS) + 
+                   (cqa->cqa_cqnum + dev->dev_cqoffset) * CQ_CommandMappingSize, PAGE_SIZE,
+                   vma->vm_page_prot);
+
+           /* Don't allow these pages to be swapped out of dumped */
+           vma->vm_flags |= (VM_RESERVED | VM_IO);
+
+           if (__io_remap_page_range (addr, 
+                                      pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS) + 
+                                      (cqa->cqa_cqnum + dev->dev_cqoffset) * CQ_CommandMappingSize,
+                                      PAGE_SIZE, vma->vm_page_prot))
+           {
+               PRINTF (uctx, DBG_FILE, "user_mmap: remap_page_range failed\n");
+
+               elan4_putcqa (&uctx->uctx_ctxt, pgoff);
+               res = -ENOMEM;
+               goto failed;
+           }
+           break;
+           
+       case ELAN4_OFF_USER_REGS:
+           vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+           /* Don't allow these pages to be swapped out of dumped */
+           vma->vm_flags |= (VM_RESERVED | VM_IO);
+
+           switch (dev->dev_devinfo.dev_revision_id)
+           {
+           case PCI_REVISION_ID_ELAN4_REVA:
+               ioaddr = pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS) + ELAN4_REVA_REG_OFFSET + offsetof(E4_Registers, uRegs);
+               break;
+               
+           case PCI_REVISION_ID_ELAN4_REVB:
+               ioaddr = pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS) + ELAN4_REVB_REG_OFFSET + offsetof(E4_Registers, uRegs);
+               break;
+
+           default:
+               res = -EINVAL;
+               goto failed;
+           }
+
+           PRINTF (uctx, DBG_FILE, "user_mmap: user_regs at %lx ioaddr %lx prot %lx\n",
+                   addr, ioaddr, vma->vm_page_prot.pgprot);
+
+           if (__io_remap_page_range (addr,  (ioaddr & PAGEMASK), PAGE_SIZE, vma->vm_page_prot))
+           {                     
+               res = -EAGAIN;
+               goto failed;
+           }
+
+           break;
+           
+       case ELAN4_OFF_USER_PAGE:
+           PRINTF (uctx, DBG_FILE, "user_mmap: shared user page - kaddr=%lx uaddr=%lx phys=%lx\n", 
+                   uctx->uctx_upage, addr, kmem_to_phys (uctx->uctx_upage));
+
+           /* we do not want to have this area swapped out, lock it */
+           vma->vm_flags |= VM_LOCKED;
+           
+           /* Mark the page as reserved or else the remap_page_range() doesn't remap it */
+           SetPageReserved(pte_page(*find_pte_kernel((unsigned long) uctx->uctx_upage)));
+       
+           if (__remap_page_range (addr, kmem_to_phys (uctx->uctx_upage), PAGE_SIZE, vma->vm_page_prot))
+           {
+               PRINTF (uctx, DBG_FILE, "user_mmap: remap_page_range (user_page) failed\n");
+               res = -ENOMEM;
+               goto failed;
+           }
+           break;
+           
+       case ELAN4_OFF_TPROC_TRAMPOLINE:
+           vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+           PRINTF (uctx, DBG_FILE, "user_mmap: tproc trampoline - kaddr=%lx uaddr=%lx phys=%lx\n", uctx->uctx_trampoline, addr, 
+                   pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM) + uctx->uctx_trampoline + (addr & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT)));
+
+           /* Don't allow these pages to be swapped out of dumped */
+           vma->vm_flags |= (VM_RESERVED | VM_IO);
+
+           if (__io_remap_page_range (addr, pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM) + 
+                                      uctx->uctx_trampoline + (addr & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT)),
+                                      PAGE_SIZE, vma->vm_page_prot))
+           {
+               PRINTF (uctx, DBG_FILE, "user_mmap: remap_page_range (tproc_trampoline) failed\n");
+               res = -ENOMEM;
+               goto failed;
+           }
+           break;
+           
+       case ELAN4_OFF_DEVICE_STATS:
+           printk ("user_mmap: device_stats\n");
+           break;
+       }
+       
+    }
+
+    ASSERT (vma->vm_ops == NULL);
+    
+    /* Don't try to swap out physical pages.. */
+    vma->vm_flags |= VM_RESERVED;
+    
+    /*
+     * Don't dump addresses that are not real memory to a core file.
+     */
+    vma->vm_flags |= VM_IO;
+
+    vma->vm_ops          = &user_vm_ops;
+    vma->vm_file         = file;
+    vma->vm_private_data = (void *) pr;
+    
+    return (0);
+
+ failed:
+    for (addr -= PAGE_SIZE, pgoff--; addr >= vma->vm_start; addr -= PAGE_SIZE, pgoff--)
+       elan4_putcqa (&uctx->uctx_ctxt, pgoff);         /* drop the reference we've just taken */
+    return (res);
+}
+
+/* driver entry points */
+static int
+elan4_open (struct inode *inode, struct file *file)
+{
+    PRINTF (DBG_USER, DBG_FILE, "elan4_open: device %d minor %d file=%p\n", ELAN4_DEVICE(inode), ELAN4_MINOR(inode), file);
+    
+    switch (ELAN4_MINOR (inode))
+    {
+    case ELAN4_MINOR_CONTROL:
+       return (control_open (inode, file));
+    case ELAN4_MINOR_MEM:
+       return (mem_open (inode, file));
+    case ELAN4_MINOR_USER:
+       return (user_open (inode, file));
+    default:
+       return (-ENXIO);
+    }
+}
+
+static int
+elan4_release (struct inode *inode, struct file *file)
+{
+    PRINTF (DBG_USER, DBG_FILE, "elan4_release: device %d minor %d file=%p\n", ELAN4_DEVICE(inode), ELAN4_MINOR(inode), file);
+    
+    switch (ELAN4_MINOR (inode))
+    {
+    case ELAN4_MINOR_CONTROL:
+       return (control_release (inode, file));
+    case ELAN4_MINOR_MEM:
+       return (mem_release (inode, file));
+    case ELAN4_MINOR_USER:
+       return (user_release (inode, file));
+    default:
+       return (-ENXIO);
+    }
+}
+
+static int
+elan4_ioctl (struct inode *inode, struct file *file, 
+            unsigned int cmd, unsigned long arg)
+{
+    PRINTF (DBG_USER, DBG_FILE, "elan4_ioctl: device %d minor %d cmd %x\n", ELAN4_DEVICE(inode), ELAN4_MINOR(inode), cmd);
+    
+    switch (ELAN4_MINOR (inode))
+    {
+    case ELAN4_MINOR_CONTROL:
+       return (control_ioctl (inode, file, cmd, arg));
+    case ELAN4_MINOR_MEM:
+       return (mem_ioctl (inode, file, cmd, arg));
+    case ELAN4_MINOR_USER:
+       return (user_ioctl (inode, file, cmd, arg));
+    default:
+       return (-ENXIO);
+    }
+}
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+static int
+elan4_ioctl32 (unsigned int fd, unsigned int cmd, unsigned long arg, struct file *file)
+{
+    struct inode *inode = file->f_dentry->d_inode;
+    extern int sys_ioctl (unsigned int fd, unsigned int cmd, unsigned long arg);
+
+    PRINTF (DBG_USER, DBG_FILE, "elan4_ioctl32: device %d minor %d cmd %x\n", ELAN4_DEVICE(inode), ELAN4_MINOR(inode), cmd);
+    
+    if (ELAN4_MINOR (inode) == ELAN4_MINOR_USER)
+    {
+       USER_PRIVATE *pr    = (USER_PRIVATE *) file->private_data;
+       USER_CTXT    *uctx  = pr->pr_uctx;
+
+       if (current->mm != pr->pr_mm)
+           return -EINVAL;
+       
+       switch (cmd)
+       {
+       case ELAN4IO_SETPERM32:
+       {
+           ELAN4IO_PERM_STRUCT32 args;
+           
+           if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_PERM_STRUCT32)))
+               return (-EFAULT);
+           
+           PRINTF (DBG_USER, DBG_FILE, "user_ioctl32: setperm maddr=%x eaddr=%llx len=%llxx perm=%d\n",
+                   args.ps_maddr, args.ps_eaddr,args.ps_len, args.ps_perm);
+
+           return (user_setperm (uctx, args.ps_maddr, args.ps_eaddr, args.ps_len, args.ps_perm));
+       }
+       
+       case ELAN4IO_CLRPERM32:
+       {
+           ELAN4IO_PERM_STRUCT32 args;
+           
+           if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_PERM_STRUCT32)))
+               return (-EFAULT);
+           
+           PRINTF (DBG_USER, DBG_FILE, "user_ioctl32: clrperm eaddr=%llx len=%ll\n",
+                   args.ps_eaddr, args.ps_len);
+
+           user_clrperm (uctx, args.ps_eaddr, args.ps_len);
+           return (0);
+       }
+    
+       case ELAN4IO_TRAPHANDLER32:
+       {
+           ELAN4IO_TRAPHANDLER_STRUCT32 args;
+           
+           if (copy_from_user (&args, (void *) arg, sizeof (ELAN4IO_TRAPHANDLER_STRUCT32)))
+               return (-EFAULT);
+           
+           PRINTF (DBG_USER, DBG_FILE, "user_ioctl32: traphandler trapp=%x nticks=%d\n",
+                   args.th_trapp, args.th_nticks);
+
+           return (user_trap_handler (pr->pr_uctx, (ELAN4_USER_TRAP *)(unsigned long)args.th_trapp, args.th_nticks));
+       }
+       }
+    }
+
+    PRINTF (DBG_USER, DBG_FILE, "elan4_ioctl32: fd=%d cmd=%x arg=%lx file=%p\n", fd, cmd, arg, file);
+    return (sys_ioctl (fd, cmd, arg));
+}
+#endif
+
+
+
+static int
+elan4_mmap (struct file *file, struct vm_area_struct *vma)
+{
+    PRINTF (DBG_USER, DBG_FILE, "elan4_mmap: instance %d minor %d start=%lx end=%lx pgoff=%lx\n", 
+           ELAN4_DEVICE (file->f_dentry->d_inode), ELAN4_MINOR (file->f_dentry->d_inode),
+           vma->vm_start, vma->vm_end, vma->vm_pgoff);
+
+    switch (ELAN4_MINOR (file->f_dentry->d_inode))
+    {
+    case ELAN4_MINOR_CONTROL:
+       return (control_mmap (file, vma));
+    case ELAN4_MINOR_MEM:
+       return (mem_mmap (file, vma));
+    case ELAN4_MINOR_USER:
+       return (user_mmap (file, vma));
+    default:
+       return (-ENXIO);
+    }
+}
+
+void
+elan4_update_intel_p64h2 (ELAN4_DEV *dev, struct pci_dev *bridge)
+{
+    u16 cnf;
+    
+    pci_read_config_word (bridge, 0x40 /* CNF */, &cnf);
+    
+    /* We expect the CNF register to be configured as follows
+     *
+     * [8]   == 1      PMODE PCI Mode
+     * [7:6] == 2/3    PFREQ PCI Frequency (100/133)
+     * [5]   == 0      RSDIS Restreaming Disable
+     * [4:3] == 0x     PP    Prefetch Policy
+     * [2]   == 0       DTD   Delayed Transaction Depth
+     * [1:0] == 10      MDT   MaximumDelaedTransactions
+     */
+    
+    if ((cnf & (1 << 8)) == 0)
+       printk ("elan%d: strangeness - elan reports PCI-X but P64H2 reports PCI mode !\n", dev->dev_instance);
+    else if ((cnf & 0xb7) != 0x82 && (cnf & 0xb7) != 0x84 && optimise_pci_bus < 2)
+       printk ("elan%d: P64H2 CNF is not configured as expected : RSDIS=%d PP=%d DTD=%d MDT=%d\n",
+               dev->dev_instance, (cnf >> 5) & 1, (cnf >> 3) & 3, (cnf >> 2) & 1, cnf & 3);
+    else
+    {
+       switch ((cnf >> 6) & 3)
+       {
+       case 2:                                         /* PCI-X 100 */
+           pci_write_config_word (bridge, 0xfc /* PC100 */, 0x7777);
+           
+           printk ("elan%d: optimise P64H2 : setting MDT=0, DTD=1, PFC=777 for PCI-X 100\n", dev->dev_instance);
+           
+           break;
+           
+       case 3:                                         /* PCI-X 133 */
+           pci_write_config_word (bridge, 0xfe /* PC133 */, 0x7777);
+           
+           printk ("elan%d: optimise P64H2 : setting MDT=0, DTD=1, PFC=777 for PCI-X 133\n", dev->dev_instance);
+           break;
+       }
+       
+       pci_write_config_word (bridge, 0x40 /* CNF */, (cnf & 0xfff8) | 0x4);   /* DTD=1 MDT=0 */
+    }
+}
+
+int
+elan4_optimise_intel_p64h2 (ELAN4_DEV *dev, struct pci_dev *pdev)
+{
+    struct pci_bus   *bus      = pdev->bus;
+    struct pci_dev   *bridge   = bus->self;
+    unsigned int      devcount = 0;
+    u8                revision;
+    u32               ectrl;
+    struct list_head *el;
+    
+    pci_read_config_dword (pdev, PCI_ELAN_CONTROL, &ectrl);
+
+    /* We can only run in PCI-Xmode with a B1 stepping P64H2 because of P64H2 Errata 3 */
+    pci_read_config_byte (bridge, PCI_REVISION_ID, &revision);
+    if (revision < 0x04)
+    {
+       if ((ectrl & ECTRL_INITIALISATION_MODE) != Pci2_2)
+       {
+           static const char *p64h2_stepping[4] = {"UNKNOWN", "UNKNOWN", "UNKNOWN", "B0"};
+
+           printk ("elan%d: unable to use device because of P64H2 Errata 3 on\n"
+                   "       %s stepping part and running in a PCI-X slot\n", 
+                   dev->dev_instance, p64h2_stepping[revision]);
+           return -EINVAL;
+       }
+    }
+    
+    /* We can only alter the bus configuration registers if the Elan is the only device
+     * on the bus ... */
+    list_for_each (el, &bus->devices) {
+       struct pci_dev *pcip = list_entry (el, struct pci_dev, bus_list);
+
+       if (pcip == pdev || (pcip->vendor == PCI_VENDOR_ID_INTEL && pcip->device == 0x1462 /* P64H2 HOTPLUG */))
+           continue;
+           
+       devcount++;
+    }
+
+    if (devcount > 0 || !list_empty (&bus->children))
+    {
+       printk ("elan%d: unable to optimise P64H2 settings as %s%s\n", dev->dev_instance,
+               (devcount > 0) ? "more than one device on bus" :  "",
+               ! list_empty (&bus->children) ? "has child buses" : "");
+       return 0;
+    }
+
+#ifdef __ia64
+    if ((ectrl & ECTRL_INITIALISATION_MODE) == PciX100to133MHz)
+    {
+       struct pci_dev *pcip;
+       unsigned int sioh_good      = 0;
+       unsigned int sioh_downgrade = 0;
+       unsigned int snc_good       = 0;
+       unsigned int snc_downgrade  = 0;
+       
+       /* Search for the associated SIOH and SNC on ia64,
+        * if we have a C2 SIOH and a C0/C1 SNC, then we can
+        * reconfigure the P64H2 as follows:
+        *    CNF:MDT   = 0
+        *    CNF:DTD   = 1
+        *    CNF:PC133 = 7777
+        *
+        * if not, then issue a warning that down rev parts
+        * affect bandwidth.
+        */
+       for (pcip = NULL; (pcip = pci_find_device (PCI_VENDOR_ID_INTEL, 0x500, pcip)); )
+       {
+           pci_read_config_byte (pcip, PCI_REVISION_ID, &revision);
+           
+           if (revision >= 0x21)
+               snc_good++;
+           else
+           {
+               printk ("elan%d: SNC revision %x (%s)\n", dev->dev_instance, revision,
+                       revision == 0x00 ? "A0" : revision == 0x01 ? "A1" : 
+                       revision == 0x02 ? "A2" : revision == 0x03 ? "A3" :
+                       revision == 0x10 ? "B0" : revision == 0x20 ? "C0" : 
+                       revision == 0x21 ? "C1" : "UNKNOWN");
+           
+               snc_downgrade++;
+           }
+       }
+
+       for (pcip = NULL; (pcip = pci_find_device (PCI_VENDOR_ID_INTEL, 0x510, pcip)) != NULL; )
+       {
+           pci_read_config_byte (pcip, PCI_REVISION_ID, &revision);
+           
+           
+           if (revision >= 0x22)
+               sioh_good++;
+           else
+           {
+               printk ("elan%d: SIOH revsision %x (%s)\n", dev->dev_instance, revision,
+                       revision == 0x10 ? "C0" : revision == 0x20 ? "C0" : 
+                       revision == 0x21 ? "C1" : revision == 0x22 ? "C2" : "UNKNOWN");
+
+               sioh_downgrade++;
+           }
+       }
+
+       if (optimise_pci_bus < 2 && (sioh_downgrade || snc_downgrade))
+           printk ("elan%d: unable to optimise as SNC/SIOH below required C1/C2 steppings\n", dev->dev_instance);
+       else if (optimise_pci_bus < 2 && (sioh_good == 0 || snc_good == 0))
+           printk ("elan%d: unable to optimise as cannot determine SNC/SIOH revision\n", dev->dev_instance);
+       else
+           elan4_update_intel_p64h2 (dev, bridge);
+    }
+#endif
+    
+#ifdef __i386
+    if ((ectrl & ECTRL_INITIALISATION_MODE) == PciX100to133MHz)
+       elan4_update_intel_p64h2 (dev, bridge);
+#endif     
+    return 0;
+}
+
+int
+elan4_optimise_intel_pxh (ELAN4_DEV *dev, struct pci_dev *pdev)
+{
+#ifdef __i386
+    printk ("elan%d: unable to use device on this platform in 32 bit mode\n", dev->dev_instance);
+
+    return -EINVAL;
+#endif
+
+    dev->dev_features |= ELAN4_FEATURE_NO_DWORD_READ;
+
+    return 0;
+}
+
+void
+elan4_optimise_serverworks_ciobx2 (ELAN4_DEV *dev)
+{
+    struct pci_dev *pdev = dev->dev_osdep.pdev;
+    struct pci_dev *pcip;
+    unsigned char   bus;
+    unsigned int    dor;
+    
+    /* Find the CIOBX2 for our bus number */
+    for (pcip = NULL; (pcip = pci_find_device (PCI_VENDOR_ID_SERVERWORKS, 0x0101, pcip)) != NULL;)
+    {
+       pci_read_config_byte (pcip, 0x44 /* BUSNUM */, &bus);
+       
+       if (pdev->bus->number == bus)
+       {
+           printk ("elan%d: optimise CIOBX2 : setting DOR to disable read pipe lining\n", dev->dev_instance);
+
+           pci_read_config_dword (pcip, 0x78 /* DOR */, &dor);
+           pci_write_config_dword (pcip, 0x78 /* DOR */, dor | (1 << 16));
+       }
+    }
+}
+
+int
+elan4_optimise_bus (ELAN4_DEV *dev)
+{
+    struct pci_dev *pdev = dev->dev_osdep.pdev;
+
+    if (pdev->bus && pdev->bus->self) 
+    {
+       struct pci_dev *bridge = pdev->bus->self;
+       
+       if (bridge->vendor == PCI_VENDOR_ID_INTEL && bridge->device == 0x1460 /* Intel P64H2 */)
+           return elan4_optimise_intel_p64h2 (dev, pdev);
+
+       if ((bridge->vendor == PCI_VENDOR_ID_INTEL && bridge->device == 0x0329) /* Intel 6700PXH Fn 0 */ ||
+           (bridge->vendor == PCI_VENDOR_ID_INTEL && bridge->device == 0x032a) /* Intel 6700PXH Fn 2 */ ||
+           (bridge->vendor == PCI_VENDOR_ID_INTEL && bridge->device == 0x032c) /* Intel 6702PXH */ ||
+           (bridge->vendor == PCI_VENDOR_ID_INTEL && bridge->device == 0x0320) /* Intel PXH-D */)
+           return elan4_optimise_intel_pxh (dev, pdev);
+    }
+
+    if (pci_find_device (PCI_VENDOR_ID_HP, 0x122e, NULL) != NULL)              /* on HP ZX1 set the relaxed ordering  */
+       dev->dev_pteval = PTE_RelaxedOrder;                                     /* bit to get better DMA bandwidth. */
+
+    if (pci_find_device (PCI_VENDOR_ID_SERVERWORKS, 0x0101, NULL) != NULL)     /* ServerWorks CIOBX2 */
+       elan4_optimise_serverworks_ciobx2 (dev);
+
+    return 0;
+}
+
+int
+elan4_pciinit (ELAN4_DEV *dev)
+{
+    int res;
+    u32 value;
+    u16 command;
+    u8 cacheline;
+    unsigned long flags;
+
+    if (optimise_pci_bus && (res = elan4_optimise_bus (dev)) <0)
+       return (res);
+
+    if ((res = pci_enable_device (dev->dev_osdep.pdev)) < 0)
+       return (res);
+
+    pci_read_config_dword (dev->dev_osdep.pdev, PCI_ELAN_CONTROL, &value);
+    if ((value & ECTRL_INITIALISATION_MODE) == Pci2_2)
+       printk ("elan%d: is an elan4%c (PCI-2.2)\n", dev->dev_instance, 'a' + dev->dev_devinfo.dev_revision_id);
+    else
+    {
+       switch (value & ECTRL_INITIALISATION_MODE)
+       {
+       case PciX50To66MHz:
+           printk ("elan%d: is an elan4%c (PCI-X 50-66)\n", dev->dev_instance, 'a' + dev->dev_devinfo.dev_revision_id);
+           break;
+           
+       case PciX66to100MHz:
+           printk ("elan%d: is an elan4%c (PCI-X 66-100)\n", dev->dev_instance, 'a' + dev->dev_devinfo.dev_revision_id);
+           break;
+           
+       case PciX100to133MHz:
+           printk ("elan%d: is an elan4%c (PCI-X 100-133)\n", dev->dev_instance, 'a' + dev->dev_devinfo.dev_revision_id);
+           break;
+           
+       default:
+           printk ("elan%d: Invalid PCI-X mode\n", dev->dev_instance);
+           return (-EINVAL);
+       }
+    }
+
+    /* initialise the elan pll control register */
+    pci_read_config_dword (dev->dev_osdep.pdev, PCI_ELAN_PLL_CONTROL, &value);
+
+    if (elan4_pll_cfg)
+    {
+       printk ("elan%d: setting pll control to %08x\n", dev->dev_instance, elan4_pll_cfg);
+
+       pci_write_config_dword (dev->dev_osdep.pdev, PCI_ELAN_PLL_CONTROL, elan4_pll_cfg);
+    }
+    else
+    {
+       if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+           pci_write_config_dword (dev->dev_osdep.pdev, PCI_ELAN_PLL_CONTROL, 
+                                   (value & ~ECTRL_SYS_CLOCK_RATIO_MASK) | ECTRL_SYS_CLOCK_RATIO_4_3);
+       else
+           pci_write_config_dword (dev->dev_osdep.pdev, PCI_ELAN_PLL_CONTROL, 
+                                   (value & ~ECTRL_SYS_CLOCK_RATIO_MASK) | ECTRL_SYS_CLOCK_RATIO_6_5 | SysPll_FeedForwardISel0 | SysPll_FeedForwardISel1);
+    }  
+
+    /* initialise the elan control register */
+    pci_read_config_dword (dev->dev_osdep.pdev, PCI_ELAN_CONTROL, &value);
+
+    value = ((15 << ECTRL_IPROC_HIGH_PRI_TIME_SHIFT) |
+            (15 << ECTRL_OTHER_HIGH_PRI_TIME_SHIFT) |
+            (value & ECTRL_28_NOT_30_BIT_LOCAL_BAR) |
+            (dev->dev_topaddrmode ? ECTRL_ExtraMasterAddrBits : 0) |
+            ECTRL_ENABLE_LATENCY_RESET | 
+            ECTRL_ENABLE_WRITEBURSTS | 
+            ECTRL_ENABLE_2_2READBURSTS);
+
+#ifdef LINUX_SPARC
+    value &= ~(ECTRL_ENABLE_LATENCY_RESET | ECTRL_ENABLE_WRITEBURSTS);
+#endif
+
+    pci_write_config_dword (dev->dev_osdep.pdev, PCI_ELAN_CONTROL, value | ECTRL_SOFTWARE_INTERNAL_RESET);
+
+    switch (dev->dev_devinfo.dev_revision_id)
+    {
+    case PCI_REVISION_ID_ELAN4_REVA:
+       /* Delay 10ms here if we've changed the sysclock ratio */
+       /* to allow the PLL to stabalise before proceeding */
+       udelay (10000);
+       break;
+       
+    case PCI_REVISION_ID_ELAN4_REVB:
+    {
+       unsigned char val = read_i2c (dev, I2cLedsValue);
+
+       /* On RevB we have to explicitly reset the PLLs */
+       pci_read_config_word (dev->dev_osdep.pdev, PCI_COMMAND, &command);
+
+       write_i2c (dev, I2cLedsValue, val | 0x80);
+       udelay (1000);
+
+       /* Issue the PLL counter reset and immediately inhibit all pci interaction 
+        * while the PLL is recovering. The write to the PCI_COMMAND register has 
+        * to occur within 50uS of the write to the i2c registers */
+       local_irq_save (flags);
+       write_i2c (dev, I2cLedsValue, val & ~0x80);
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_COMMAND, (1 << 10) /* PCI_COMMAND_DISABLE_INT */);
+       local_irq_restore (flags);
+
+       /* Wait for the write to occur and for the PLL to regain lock */
+       udelay (20000); udelay (20000);
+
+       /* Re-enable pci interaction and clear any spurious errors deteced */
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_STATUS, PCI_STATUS_DETECTED_PARITY | PCI_STATUS_SIG_SYSTEM_ERROR);
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_COMMAND, command);
+       break;
+    }
+    }
+
+    pci_write_config_dword (dev->dev_osdep.pdev, PCI_ELAN_CONTROL, value);
+
+    /* Enable master accesses */
+    pci_set_master (dev->dev_osdep.pdev);
+
+    /* Verify that the memWrInvalidate bit is set */
+    pci_read_config_word (dev->dev_osdep.pdev, PCI_COMMAND, &command);
+    pci_read_config_byte (dev->dev_osdep.pdev, PCI_CACHE_LINE_SIZE, &cacheline);
+
+    if ((command & PCI_COMMAND_INVALIDATE) == 0)
+    {
+       printk ("elan%d: enable MemWrInvalidate (cacheline %d)\n",
+               dev->dev_instance, cacheline * 4);
+
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_COMMAND, command | PCI_COMMAND_INVALIDATE);
+    }
+
+    return (0);
+}
+
+void
+elan4_pcifini (ELAN4_DEV *dev)
+{
+    u32 value;
+
+    pci_read_config_dword (dev->dev_osdep.pdev, PCI_ELAN_CONTROL, &value);
+    pci_write_config_dword (dev->dev_osdep.pdev, PCI_ELAN_CONTROL, value | ECTRL_SOFTWARE_INTERNAL_RESET);
+    pci_write_config_dword (dev->dev_osdep.pdev, PCI_ELAN_CONTROL, value);
+
+    pci_disable_device (dev->dev_osdep.pdev);
+}
+
+void
+elan4_pcierror (ELAN4_DEV *dev)
+{
+    struct pci_dev *pci = dev->dev_osdep.pdev;
+    u8  type;
+    u16 status, cmd;
+    u32 physlo, physhi, control;
+    
+    printk("elan%d: pci error has occurred\n", dev->dev_instance);
+
+    pci_read_config_word  (pci, PCI_STATUS,             &status);
+    pci_read_config_word  (pci, PCI_COMMAND,             &cmd);
+    pci_read_config_dword (pci, PCI_ELAN_CONTROL,       &control);
+
+    if (control & ECTRL_REC_SPLIT_COMP_MESSAGE)
+    {
+       u32 message, attr;
+       
+       pci_write_config_dword (pci, PCI_ELAN_CONTROL, control & ~ECTRL_SELECT_SPLIT_MESS_ATTR);
+       pci_read_config_dword (pci, PCI_ELAN_SPLIT_MESSAGE_VALUE, &message);
+       pci_write_config_dword (pci, PCI_ELAN_CONTROL, control | ECTRL_SELECT_SPLIT_MESS_ATTR);
+       pci_read_config_dword (pci, PCI_ELAN_SPLIT_MESSAGE_VALUE, &attr);
+
+       printk ("elan%d: pcierror - received split completion message - attr=%08x, message=%08x\n", 
+               dev->dev_instance, attr, message);
+
+       pci_write_config_dword (pci, PCI_ELAN_CONTROL, control | ECTRL_REC_SPLIT_COMP_MESSAGE); /* clear the error */
+    }
+    else
+    {
+       pci_read_config_dword (pci, PCI_ELAN_PARITY_ADDR_LO, &physlo);
+       pci_read_config_dword (pci, PCI_ELAN_PARITY_ADDR_HI, &physhi);
+       pci_read_config_byte  (pci, PCI_ELAN_PARITY_TYPE,    &type);
+       
+       printk ("elan%d: pcierror - status %x cmd %4x physaddr %08x%08x type %x\n", 
+               dev->dev_instance, status, cmd, physhi, physlo, type);
+       
+       if (status & PCI_STATUS_PARITY)
+           printk ("elan%d: parity error signalled (PERR)\n", dev->dev_instance);
+       if (status & PCI_STATUS_DETECTED_PARITY)
+           printk ("elan%d: detected parity error\n", dev->dev_instance);
+       if (status & PCI_STATUS_REC_MASTER_ABORT)
+           printk ("elan%d: received master abort\n", dev->dev_instance);
+       if (status & PCI_STATUS_REC_TARGET_ABORT)
+           printk ("elan%d: received target abort\n", dev->dev_instance);
+       if (status & PCI_STATUS_SIG_SYSTEM_ERROR)
+           printk ("elan%d: signalled SERR\n", dev->dev_instance);
+       if (status & PCI_STATUS_SIG_TARGET_ABORT)
+           printk ("elan%d: signalled target abort\n", dev->dev_instance);
+
+       pci_write_config_word (pci, PCI_STATUS, status);        /* clear the errors */
+    }
+
+    DISABLE_INT_MASK (dev, INT_PciMemErr);
+
+#ifdef notdef
+    panic ("elan%d: pcierror\n", dev->dev_instance);           /* better panic ! */
+#endif
+}
+
+static irqreturn_t
+elan4_irq (int irq, void *arg, struct pt_regs *regs)
+{
+    if (elan4_1msi0 ((ELAN4_DEV *) arg))
+           return IRQ_HANDLED;
+    else
+           return IRQ_NONE;
+}
+
+ioaddr_t
+elan4_map_device (ELAN4_DEV *dev, unsigned bar, unsigned off, unsigned size, ELAN4_MAP_HANDLE *handle)
+{
+    return (ioaddr_t) ioremap_nocache (pci_resource_start (dev->dev_osdep.pdev, bar) + off, size);
+}
+
+void
+elan4_unmap_device (ELAN4_DEV *dev, ioaddr_t ptr, unsigned size, ELAN4_MAP_HANDLE *handle)
+{
+    iounmap ((void *) ptr);
+}
+
+unsigned long
+elan4_resource_len (ELAN4_DEV *dev, unsigned bar)
+{
+    return (pci_resource_len (dev->dev_osdep.pdev, bar));
+}
+
+void
+elan4_configure_mtrr (ELAN4_DEV *dev)
+{
+#ifdef CONFIG_MTRR
+    if (! (dev->dev_features & ELAN4_FEATURE_NO_WRITE_COMBINE))
+    {
+       /* try and initialise the MTRR registers to enable write-combining */
+       dev->dev_osdep.sdram_mtrr = mtrr_add (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM), 
+                                             pci_resource_len   (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM),
+                                             MTRR_TYPE_WRCOMB, 1);
+       if (dev->dev_osdep.sdram_mtrr < 0)
+           printk ("elan%d: cannot configure MTRR for sdram\n", dev->dev_instance);
+       
+       if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVB)
+       {
+           dev->dev_osdep.regs_mtrr = mtrr_add (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS) + 
+                                                (dev->dev_cqoffset + dev->dev_cqreorder) * CQ_CommandMappingSize,
+                                                CQ_CommandMappingSize * (dev->dev_cqcount >> 1), 
+                                                MTRR_TYPE_WRCOMB, 1);
+           
+           if (dev->dev_osdep.regs_mtrr < 0)
+               printk ("elan%d: cannot configure MTRR for command ports\n", dev->dev_instance);
+       }
+    }
+#endif
+}
+
+void
+elan4_unconfigure_mtrr (ELAN4_DEV *dev)
+{
+#ifdef CONFIG_MTRR
+    if (! (dev->dev_features & ELAN4_FEATURE_NO_WRITE_COMBINE))
+    {
+       if (dev->dev_osdep.sdram_mtrr >=0 )
+           mtrr_del (dev->dev_osdep.sdram_mtrr, pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM), 
+                     pci_resource_len   (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+       
+       if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVB && dev->dev_osdep.regs_mtrr >= 0)
+           mtrr_del (dev->dev_osdep.regs_mtrr, 
+                     pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS) + 
+                     (dev->dev_cqoffset + dev->dev_cqreorder) * CQ_CommandMappingSize,
+                     CQ_CommandMappingSize * (dev->dev_cqcount >> 1));
+    }
+#endif
+}
+
+EXPORT_SYMBOL(elan4_reference_device);
+EXPORT_SYMBOL(elan4_dereference_device);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/i2c.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/i2c.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/i2c.c  2005-06-01 23:12:54.607437888 -0400
@@ -0,0 +1,248 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: i2c.c,v 1.4 2004/01/07 13:37:45 jon Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/i2c.c,v $*/
+#include <qsnet/kernel.h>
+
+#include <elan4/sdram.h>
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/commands.h>
+
+#include <elan4/i2c.h>
+#include <elan4/pci.h>
+#include <elan4/ioctl.h>
+#include <elan4/registers.h>
+
+#define I2C_POLL_LIMIT         8
+
+static int
+i2c_poll_busy (ELAN4_DEV *dev)
+{
+    int t    = 100;
+    int loop = 0;
+    volatile unsigned char val;
+
+    /* wait for any led I2C operation to finish */
+    while (((val = read_i2c (dev, I2cPortControl)) & I2cCntl_I2cPortBusy) && loop++ < I2C_POLL_LIMIT)
+    {
+       DELAY (t);
+       
+       if (t < 500000)
+           t <<= 1;
+    }
+    if (loop >= I2C_POLL_LIMIT)
+    {
+       printk ("elan%d: I2c has timed out waiting for I2cPortBusy to clear!\n", dev->dev_instance);
+       printk ("elan%d: I2cPortControl=%x I2cLedBase=%x I2cStatus=%x\n",
+               dev->dev_instance, val, read_i2c (dev, I2cLedBase), read_i2c (dev, I2cStatus));
+    }
+
+    return val;
+}
+
+static int
+i2c_poll_stopped (ELAN4_DEV *dev)
+{
+    int t    = 100;
+    int loop = 0;
+    unsigned char val=0, newval;
+
+    /* wait for any led I2C operation to finish. Must see it stopped at least twice */
+    while (!(((newval = read_i2c (dev, I2cPortControl)) & I2cCntl_I2cStopped) &&
+             (val & I2cCntl_I2cStopped)) &&
+             (loop++ < I2C_POLL_LIMIT))
+    {
+       DELAY (t);
+       
+       if (t < 500000)
+           t <<= 1;
+       val = newval;
+    }
+
+    return val;
+}
+
+int
+i2c_disable_auto_led_update (ELAN4_DEV *dev)
+{
+    spin_lock (&dev->dev_i2c_lock);
+
+    if (dev->dev_i2c_led_disabled++ == 0)
+    {
+       write_i2c (dev, I2cLedBase, read_i2c (dev, I2cLedBase) & ~I2cCntl_I2cUpdatingLedReg);
+
+       if (! (i2c_poll_stopped (dev) & I2cCntl_I2cStopped))
+       {
+           write_i2c (dev, I2cLedBase, read_i2c (dev, I2cLedBase) | I2cCntl_I2cUpdatingLedReg);
+           
+           spin_unlock (&dev->dev_i2c_lock);
+           
+           return -EAGAIN;
+       }
+       
+       write_i2c (dev, I2cStatus, read_i2c (dev, I2cStatus) & ~I2cCntl_SampleNewLedValues);
+    }
+
+    spin_unlock (&dev->dev_i2c_lock);
+
+    return 0;
+}
+
+void
+i2c_enable_auto_led_update (ELAN4_DEV *dev)
+{
+    spin_lock (&dev->dev_i2c_lock);
+    if (--dev->dev_i2c_led_disabled == 0)
+    {
+       write_i2c (dev, I2cLedBase, read_i2c (dev, I2cLedBase) | I2cCntl_I2cUpdatingLedReg);
+       write_i2c (dev, I2cStatus, read_i2c (dev, I2cStatus) | I2cCntl_SampleNewLedValues);
+    }
+
+    spin_unlock (&dev->dev_i2c_lock);
+}
+
+int
+i2c_write (ELAN4_DEV *dev, unsigned int address, unsigned int count, unsigned char *data)
+{
+    int i;
+
+    if (! (i2c_poll_busy (dev) & I2cCntl_I2cStopped))
+       return -EAGAIN;
+    
+    write_i2c (dev, I2cWrData,      I2C_WRITE_ADDR(address));
+    write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite);
+    
+    if (i2c_poll_busy (dev) & I2cCntl_I2cPortAccFailed)
+       return -ENXIO;
+
+    for (i = 0; i < count; i++)
+    {
+       write_i2c (dev, I2cWrData, data[i]);
+       write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite | (i == (count-1) ? I2cCntl_I2cPortGenStopBit : 0));
+    }
+
+    return 0;
+}
+
+int
+i2c_read (ELAN4_DEV *dev, unsigned int address, unsigned int count, unsigned char *data)
+{
+    int i;
+
+    if (! (i2c_poll_busy (dev) & I2cCntl_I2cStopped))
+       return -EAGAIN; /* not idle */ 
+
+    write_i2c (dev, I2cWrData,      I2C_READ_ADDR(address));
+    write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite);
+
+    if (i2c_poll_busy (dev) & I2cCntl_I2cPortAccFailed)
+       return -ENXIO;
+    
+    for (i = 0; i < count; i++)
+    {
+       write_i2c (dev, I2cWrData, 0xff);
+       write_i2c (dev, I2cPortControl, I2cCntl_I2cPortRead | ((i == count-1) ? I2cCntl_I2cPortGenStopBit : 0));
+
+       i2c_poll_busy (dev);
+
+       data[i] = read_i2c (dev, I2cRdData);
+    }
+
+    return 0;
+}
+
+int
+i2c_writereg (ELAN4_DEV *dev, unsigned int address, unsigned int reg, unsigned int count, unsigned char *data)
+{
+    int i;
+
+    if (! (i2c_poll_busy (dev) & I2cCntl_I2cStopped))
+       return -EAGAIN; /* not idle */ 
+
+    write_i2c (dev, I2cWrData,      I2C_WRITE_ADDR(address));
+    write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite);
+
+    if (i2c_poll_busy (dev) & I2cCntl_I2cPortAccFailed)
+       return -ENXIO;
+    
+    write_i2c (dev, I2cWrData,      reg);
+    write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite);
+
+    if (i2c_poll_busy (dev) & I2cCntl_I2cPortAccFailed)
+       return -ENXIO;
+    
+    for (i = 0; i < count; i++)
+    {
+       write_i2c (dev, I2cWrData, data[i]);
+       write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite | ((i == count-1) ? I2cCntl_I2cPortGenStopBit : 0));
+
+       if (i2c_poll_busy (dev) & I2cCntl_I2cPortAccFailed)
+           printk (" i2c_writereg: off %d failed\n", i);
+    }
+
+    return 0;
+}
+
+int
+i2c_readreg (ELAN4_DEV *dev, unsigned int address, unsigned int reg, unsigned int count, unsigned char *data)
+{
+    if (! (i2c_poll_busy (dev) & I2cCntl_I2cStopped))
+       return -EAGAIN; /* not idle */ 
+
+    write_i2c (dev, I2cWrData,      I2C_WRITE_ADDR(address));
+    write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite);
+
+    if (i2c_poll_busy (dev) & I2cCntl_I2cPortAccFailed)
+       return -ENXIO;
+    
+    write_i2c (dev, I2cWrData,      reg);
+    write_i2c (dev, I2cPortControl, I2cCntl_I2cPortWrite | I2cCntl_I2cPortGenStopBit);
+
+    if (i2c_poll_busy (dev) & I2cCntl_I2cPortAccFailed)
+       return -ENXIO;
+
+    return i2c_read (dev, address, count, data);
+}
+
+int
+i2c_read_rom (ELAN4_DEV *dev, unsigned int addr, unsigned int len, unsigned char *data)
+{
+    unsigned int top = addr + len;
+    int res;
+
+    if ((res = i2c_disable_auto_led_update (dev)) == 0)
+    {
+       /* read the rom in chunks that don't span the block boundary */
+       while (addr < top)
+       {
+           unsigned int thisnob  = top - addr;
+           unsigned int blocknob = I2C_24LC16B_BLOCKSIZE - I2C_24LC16B_BLOCKOFFSET(addr);
+           
+           if (thisnob > blocknob)
+               thisnob = blocknob;
+
+           if ((res = i2c_readreg (dev, I2C_EEPROM_ADDR + I2C_24LC16B_BLOCKADDR(addr),
+                                   I2C_24LC16B_BLOCKOFFSET(addr), thisnob, data)) < 0)
+               break;
+           
+           addr += thisnob;
+           data += thisnob;
+       }
+
+       i2c_enable_auto_led_update (dev);
+    }
+    return res;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/intcookie.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/intcookie.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/intcookie.c    2005-06-01 23:12:54.608437736 -0400
@@ -0,0 +1,371 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: intcookie.c,v 1.14 2004/08/09 14:02:37 daniel Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/intcookie.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/debug.h>
+#include <elan4/types.h>
+#include <elan/capability.h>
+#include <elan4/intcookie.h>
+
+static INTCOOKIE_TABLE *intcookie_tables;
+static spinlock_t      intcookie_table_lock;
+
+/*
+ * intcookie_drop_entry:
+ *   drop the reference to a cookie held 
+ *   by the cookie table
+ */
+static void
+intcookie_drop_entry (INTCOOKIE_ENTRY *ent)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&ent->ent_lock, flags);
+    if (--ent->ent_ref != 0)
+    {
+       ent->ent_fired = ent->ent_cookie;
+       kcondvar_wakeupall (&ent->ent_wait, &ent->ent_lock);
+
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+    }
+    else
+    {
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+
+       spin_lock_destroy (&ent->ent_lock);
+       kcondvar_destroy (&ent->ent_wait);
+
+       KMEM_FREE (ent, sizeof (INTCOOKIE_ENTRY));
+    }
+}
+
+void
+intcookie_init()
+{
+    spin_lock_init (&intcookie_table_lock);
+}
+
+void
+intcookie_fini()
+{
+    spin_lock_destroy (&intcookie_table_lock);
+}
+
+INTCOOKIE_TABLE *
+intcookie_alloc_table (ELAN_CAPABILITY *cap)
+{
+    INTCOOKIE_TABLE *tbl, *ntbl;
+    ELAN_CAPABILITY *ncap;
+    
+    KMEM_ZALLOC (ntbl, INTCOOKIE_TABLE *, sizeof (INTCOOKIE_TABLE), 1);
+
+    if (ntbl == NULL)
+       return (NULL);
+
+    KMEM_ALLOC (ncap, ELAN_CAPABILITY *, ELAN_CAP_SIZE(cap), 1);
+
+    if (ncap == NULL)
+    {
+       KMEM_FREE (ntbl, sizeof (INTCOOKIE_TABLE));
+       return (NULL);
+    }
+
+    spin_lock (&intcookie_table_lock);
+    
+    for (tbl = intcookie_tables; tbl; tbl = tbl->tbl_next)
+       if (ELAN_CAP_MATCH (tbl->tbl_cap, cap) && tbl->tbl_cap->cap_mycontext == cap->cap_mycontext)
+           break;
+    
+    if (tbl != NULL)
+       tbl->tbl_ref++;
+    else
+    {
+       spin_lock_init (&ntbl->tbl_lock);
+
+       ntbl->tbl_cap     = ncap;
+       ntbl->tbl_ref     = 1;
+       ntbl->tbl_entries = NULL;
+       
+       /* Save supplied cap */
+       bcopy (cap, ncap, ELAN_CAP_SIZE(cap));
+
+       if ((ntbl->tbl_next = intcookie_tables) != NULL)
+           intcookie_tables->tbl_prev = ntbl;
+       intcookie_tables = ntbl;
+       ntbl->tbl_prev = NULL;
+    }
+    spin_unlock (&intcookie_table_lock);
+
+    if (tbl == NULL)
+       return (ntbl);
+    else
+    {
+       KMEM_FREE (ntbl, sizeof (INTCOOKIE_TABLE));
+       KMEM_FREE (ncap, ELAN_CAP_SIZE(cap));
+       return (tbl);
+    }    
+}
+
+void
+intcookie_free_table (INTCOOKIE_TABLE *tbl)
+{
+    INTCOOKIE_ENTRY *ent;
+
+    spin_lock (&intcookie_table_lock);
+    if (tbl->tbl_ref > 1)
+    {
+       tbl->tbl_ref--;
+       spin_unlock (&intcookie_table_lock);
+       return;
+    }
+    
+    if (tbl->tbl_prev)
+       tbl->tbl_prev->tbl_next = tbl->tbl_next;
+    else
+       intcookie_tables = tbl->tbl_next;
+    if (tbl->tbl_next)
+       tbl->tbl_next->tbl_prev = tbl->tbl_prev;
+    
+    spin_unlock (&intcookie_table_lock);
+    
+    /* NOTE - table no longer visible to other threads
+     *        no need to aquire tbl_lock */
+    while ((ent = tbl->tbl_entries) != NULL)
+    {
+       if ((tbl->tbl_entries = ent->ent_next) != NULL)
+           ent->ent_next->ent_prev = NULL;
+       
+       intcookie_drop_entry (ent);
+    }
+    spin_lock_destroy (&tbl->tbl_lock);
+
+    KMEM_FREE (tbl->tbl_cap, ELAN_CAP_SIZE(tbl->tbl_cap));
+    KMEM_FREE (tbl, sizeof (INTCOOKIE_TABLE));
+}
+
+int
+intcookie_alloc (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie)
+{
+    INTCOOKIE_ENTRY *ent, *nent;
+    unsigned long flags;
+
+    KMEM_ZALLOC (nent, INTCOOKIE_ENTRY *, sizeof (INTCOOKIE_ENTRY), 1);
+
+    if (nent == NULL)
+       return (-ENOMEM);
+    
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+
+    if (ent == NULL)
+    {
+       kcondvar_init (&nent->ent_wait);
+       spin_lock_init (&nent->ent_lock);
+
+       nent->ent_ref    = 1;
+       nent->ent_cookie = cookie;
+
+       if ((nent->ent_next = tbl->tbl_entries) != NULL)
+           tbl->tbl_entries->ent_prev = nent;
+       tbl->tbl_entries = nent;
+       nent->ent_prev = NULL;
+    }
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    if (ent == NULL)
+       return (0);
+    else
+    {
+       KMEM_FREE (nent, sizeof (INTCOOKIE_ENTRY));
+       return (-EINVAL);
+    }
+}
+
+int
+intcookie_free (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie)
+{
+    INTCOOKIE_ENTRY *ent;
+    unsigned long flags;
+
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (-EINVAL);
+    }
+
+    if (ent->ent_prev == NULL)
+       tbl->tbl_entries = ent->ent_next;
+    else
+       ent->ent_prev->ent_next = ent->ent_next;
+
+    if (ent->ent_next != NULL)
+       ent->ent_next->ent_prev = ent->ent_prev;
+    
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    intcookie_drop_entry (ent);
+
+    return (0);
+}
+
+/*
+ * intcookie_fire_cookie:
+ *    fire the cookie - this is called from the event interrupt.
+ */
+int
+intcookie_fire (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie)
+{
+    INTCOOKIE_ENTRY *ent;
+    unsigned long flags;
+
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (-EINVAL);
+    }
+           
+    spin_lock (&ent->ent_lock);
+    ent->ent_fired = cookie;
+    kcondvar_wakeupall (&ent->ent_wait, &ent->ent_lock);
+    spin_unlock (&ent->ent_lock);
+
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    return (0);
+}    
+
+int
+intcookie_fire_cap (ELAN_CAPABILITY *cap, ELAN4_INTCOOKIE cookie)
+{
+    int res;
+    INTCOOKIE_TABLE *tbl;
+
+    spin_lock (&intcookie_table_lock);
+    
+    for (tbl = intcookie_tables; tbl; tbl = tbl->tbl_next)
+       if (ELAN_CAP_MATCH (tbl->tbl_cap, cap) && tbl->tbl_cap->cap_mycontext == cap->cap_mycontext)
+           break;
+    
+    if (tbl != NULL)
+       tbl->tbl_ref++;
+
+    spin_unlock (&intcookie_table_lock);
+
+    /* No matching table found */
+    if (tbl == NULL)
+       return (-EINVAL);
+
+    /* Fire the correct cookie */
+    res = intcookie_fire (tbl, cookie);
+
+    /* Decrement reference count (and free if necessary) */
+    intcookie_free_table (tbl);
+
+    return (res);
+}
+
+/*
+ * intcookie_wait_cookie:
+ *    deschedule on a cookie if it has not already fired.
+ *    note - if the cookie is removed from the table, then
+ *           we free it off when we're woken up.
+ */
+int
+intcookie_wait (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie)
+{
+    INTCOOKIE_ENTRY *ent;
+    unsigned long flags;
+    int res;
+    
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (-EINVAL);
+    }
+
+    spin_lock (&ent->ent_lock);
+    spin_unlock (&tbl->tbl_lock);
+
+    if (ent->ent_fired != 0)
+    {
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+       return (0);
+    }
+
+    ent->ent_ref++;
+    kcondvar_waitsig (&ent->ent_wait, &ent->ent_lock, &flags);
+    
+    res = ent->ent_fired ? 0 : -EINTR;
+
+    if (--ent->ent_ref > 0)
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+    else
+    {
+       spin_unlock_irqrestore (&ent->ent_lock, flags);
+       
+       spin_lock_destroy (&ent->ent_lock);
+       kcondvar_destroy (&ent->ent_wait);
+
+       KMEM_FREE (ent, sizeof (INTCOOKIE_ENTRY));
+    }
+
+    return (res);
+}
+
+int
+intcookie_arm (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie)
+{
+    INTCOOKIE_ENTRY *ent;
+    unsigned long flags;
+
+    spin_lock_irqsave (&tbl->tbl_lock, flags);
+    for (ent = tbl->tbl_entries; ent; ent = ent->ent_next)
+       if (ent->ent_cookie == cookie)
+           break;
+    
+    if (ent == NULL)
+    {
+       spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+       return (-EINVAL);
+    }
+           
+    spin_lock (&ent->ent_lock);
+    ent->ent_fired = 0;
+    spin_unlock (&ent->ent_lock);
+
+    spin_unlock_irqrestore (&tbl->tbl_lock, flags);
+
+    return (0);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/Makefile  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/Makefile       2005-06-01 23:12:54.608437736 -0400
@@ -0,0 +1,31 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/elan4/Makefile
+#
+
+
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/elan4/Makefile
+#
+
+list-multi             := elan4.o
+elan4-objs     := device.o i2c.o mmu.o sdram.o debug.o routetable.o trap.o user.o user_ddcq.o regions.o intcookie.o neterr.o device_Linux.o user_Linux.o procfs_Linux.o mmu_Linux.o
+export-objs            := device.o device_Linux.o mmu.o mmu_Linux.o procfs_Linux.o routetable.o sdram.o trap.o
+obj-$(CONFIG_ELAN4)    := elan4.o
+
+elan4.o : $(elan4-objs)
+       $(LD) -r -o $@ $(elan4-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/elan4/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/Makefile.conf     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/Makefile.conf  2005-06-01 23:12:54.608437736 -0400
@@ -0,0 +1,10 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       elan4.o
+MODULENAME     =       elan4
+KOBJFILES      =       device.o i2c.o mmu.o sdram.o debug.o routetable.o trap.o user.o user_ddcq.o regions.o intcookie.o neterr.o device_Linux.o user_Linux.o procfs_Linux.o mmu_Linux.o
+EXPORT_KOBJS   =       device.o device_Linux.o mmu.o mmu_Linux.o procfs_Linux.o routetable.o sdram.o trap.o
+CONFIG_NAME    =       CONFIG_ELAN4
+SGALFC         =       
+# EXTRALINES START
+
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/elan4/mmu.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/mmu.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/mmu.c  2005-06-01 23:12:54.610437432 -0400
@@ -0,0 +1,854 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: mmu.c,v 1.29.6.2 2005/01/18 16:58:12 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/mmu.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kpte.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+
+int elan4_debug_mmu;
+
+/* Permission table - see ELAN4 MMU documentation */
+u_char elan4_permtable[] =
+{
+   0x00, /* 0x000000 - Disable */
+   0x00, /* 0x000000 - Unused  */
+   0x01, /* 0x000001 - Local Data Read */
+   0x03, /* 0x000011 - Local Data Write */
+   0x11, /* 0x010001 - Local Read */
+   0x10, /* 0x010000 - Local Execute */
+   0x05, /* 0x000101 - Read Only */
+   0x13, /* 0x010011 - Local Write */
+   0x20, /* 0x100000 - Local Event Access */
+   0x23, /* 0x100011 - Local Event Write Ac */
+   0xa3, /* 1x100011 - Remote Ev Loc Write */
+   0xaf, /* 1x101111 - Remote All */
+   0x07, /* 0x000111 - Remote Read Only */
+   0x0d, /* 0x001101 - Remote Write Only */
+   0x0f, /* 0x001111 - Remote Read/Write */
+   0xbf, /* 1x111111 - No Fault */
+};
+
+u_char elan4_permreadonly[] = 
+{
+    PERM_Disabled,             /* PERM_Disabled */
+    PERM_Disabled,             /* PERM_Unused */
+    PERM_LocDataRead,          /* PERM_LocDataRead */
+    PERM_LocDataRead,          /* PERM_LocDataWrite */
+    PERM_LocRead,              /* PERM_LocRead */
+    PERM_LocExecute,           /* PERM_LocExecute */
+    PERM_ReadOnly,             /* PERM_ReadOnly */
+    PERM_LocRead,              /* PERM_LocWrite */
+    PERM_LocEventOnly,         /* PERM_LocEventOnly */
+    PERM_LocDataRead,          /* PERM_LocEventWrite */
+    PERM_LocDataRead,          /* PERM_RemoteEvent */
+    PERM_ReadOnly,             /* PERM_RemoteAll */
+    PERM_RemoteReadOnly,       /* PERM_RemoteReadOnly */
+    PERM_ReadOnly,             /* PERM_RemoteWriteLocRead */
+    PERM_ReadOnly,             /* PERM_DataReadWrite */
+    PERM_ReadOnly,             /* PERM_NoFault */
+};
+
+static void
+elan4mmu_synctag (ELAN4_DEV *dev, ELAN4_HASH_ENTRY *he, int tagidx)
+{
+    E4_uint64 value = (he->he_tag[tagidx] & HE_TAG_VALID) ? he->he_tag[tagidx] & (TAG_ADDRESS_MASK | TAG_CONTEXT_MASK) : INVALID_CONTEXT;
+    
+    if (he->he_next)
+       value |= ((tagidx == 0) ? 
+                 ((he->he_next->he_entry >> TAG_CHAINPTR_HIGH_SHIFT) & TAG_CHAINPTR_30TO19_MASK) :
+                 ((he->he_next->he_entry << TAG_CHAINPTR_LOW_SHIFT) & TAG_CHAINPTR_18TO6_MASK));
+    else if (tagidx == 0)
+       value |= TAG_CHAINPTR_30TO19_MASK;
+    
+    MPRINTF (DBG_DEVICE, 4, "elan4mmu_synctag: he=%p tagidx=%d he->he_tag=%llx -> value=%llx\n", he, tagidx, he->he_tag[tagidx], value);
+
+    elan4_sdram_writeq (dev, he->he_entry + E4MMU_TAG_OFFSET(tagidx), value);
+}
+
+static void
+elan4mmu_chain_hents (ELAN4_DEV *dev, ELAN4_HASH_ENTRY *phe, ELAN4_HASH_ENTRY *he)
+{
+    ASSERT ((elan4_sdram_readq (dev, phe->he_entry + E4MMU_TAG_OFFSET(0)) & TAG_CHAINPTR_30TO19_MASK) == TAG_CHAINPTR_30TO19_MASK);
+
+    elan4_sdram_writeq (dev, phe->he_entry + E4MMU_TAG_OFFSET(1),
+                       ((phe->he_tag[1] & (TAG_ADDRESS_MASK | TAG_CONTEXT_MASK)) | ((he->he_entry << TAG_CHAINPTR_LOW_SHIFT) & TAG_CHAINPTR_18TO6_MASK)));
+    elan4_sdram_writeq (dev, phe->he_entry + E4MMU_TAG_OFFSET(0),
+                       ((phe->he_tag[0] & (TAG_ADDRESS_MASK | TAG_CONTEXT_MASK)) | ((he->he_entry >> TAG_CHAINPTR_HIGH_SHIFT) & TAG_CHAINPTR_30TO19_MASK)));
+}
+
+static void
+elan4mmu_writepte (ELAN4_DEV *dev, ELAN4_HASH_ENTRY *he, int tagidx, int pteidx, E4_uint64 value)
+{
+    /*
+     * NOTE - we can only change a valid PTE if we're upgrading it's permissions,
+     * any other changes should have invalidated it first. */
+
+    MPRINTF (DBG_DEVICE, 4, "elan4mmu_writepte: he=%p tagidx=%d pteidx=%x value=%llx\n", he, tagidx, pteidx, (unsigned long long) value);
+
+    if (pteidx == 3)
+    {
+       elan4_sdram_writew (dev, he->he_entry + E4MMU_PTE3_WORD1_OFFSET(tagidx), (value >> 16) & 0xFFFF);
+       elan4_sdram_writew (dev, he->he_entry + E4MMU_PTE3_WORD2_OFFSET(tagidx), (value >> 32) & 0xFFFF);
+       elan4_sdram_writew (dev, he->he_entry + E4MMU_PTE3_WORD0_OFFSET(tagidx), (value >> 0)  & 0xFFFF);
+    }
+    else
+    {
+       elan4_sdram_writew (dev, he->he_entry + E4MMU_PTE_HIGH_OFFSET(tagidx, pteidx), (value >> 32) & 0xFFFF);
+       elan4_sdram_writel (dev, he->he_entry + E4MMU_PTE_LOW_OFFSET(tagidx, pteidx), value & 0xFFFFFFFF);
+    }
+}
+
+static void
+elan4mmu_invalidatepte (ELAN4_DEV *dev, ELAN4_HASH_ENTRY *he, int tagidx, int pteidx)
+{
+    if (pteidx == 3)
+       elan4_sdram_writeb (dev, he->he_entry + E4MMU_PTE3_WORD0_OFFSET(tagidx), PTE_SetPerm (PERM_Disabled));
+    else
+       elan4_sdram_writeb (dev, he->he_entry + E4MMU_PTE_LOW_OFFSET(tagidx, pteidx), PTE_SetPerm (PERM_Disabled));
+}
+
+static E4_uint64
+elan4mmu_readpte (ELAN4_DEV *dev, ELAN4_HASH_ENTRY *he, int tagidx, int pteidx)
+{
+    if (pteidx == 3)
+       return (((E4_uint64) elan4_sdram_readw (dev, he->he_entry + E4MMU_PTE3_WORD0_OFFSET(tagidx)) << 0)  |
+               ((E4_uint64) elan4_sdram_readw (dev, he->he_entry + E4MMU_PTE3_WORD1_OFFSET(tagidx)) << 16) |
+               ((E4_uint64) elan4_sdram_readw (dev, he->he_entry + E4MMU_PTE3_WORD2_OFFSET(tagidx)) << 32));
+    else
+       return ((E4_uint64) elan4_sdram_readl (dev, he->he_entry + E4MMU_PTE_LOW_OFFSET(tagidx, pteidx)) |
+               ((E4_uint64) elan4_sdram_readw (dev, he->he_entry + E4MMU_PTE_HIGH_OFFSET(tagidx, pteidx)) << 32));
+}
+
+
+void
+elan4mmu_flush_tlb (ELAN4_DEV *dev)
+{
+    PULSE_SYSCONTROL (dev, CONT_TLB_FLUSH);
+
+    while (read_reg64 (dev, SysControlReg) & CONT_TLB_FLUSH)
+       DELAY (1);
+}
+
+/*
+ * elanmmu_flush_tlb_hash - this flushes the hash copy entries and the elan
+ * tlb.  However after the write to the hash copy entry if the elan was
+ * in the process of walking, then it could write the hash copy with a valid
+ * entry which we had just invalidated. However once we've seen the tlb flushed
+ * then if the walk engine had done a write - then we need to invaldate the
+ * hash copy entries again and reflush the tlb.
+ *
+ * If we're invalidating a lot of hash blocks, then the chances are that the
+ * walk engine will perform a write - so we flush the tlb first, then invalidate
+ * the hash copy entries, then flush the tlb again.
+ */
+static void
+elan4mmu_flush_tlb_hash (ELAN4_DEV *dev, int tbl, unsigned baseidx, unsigned topidx)
+{
+    int       notmany = (abs(topidx - baseidx) < 5) ? 1 : 0;
+    int       hashidx;
+    E4_uint32 reg;
+
+    if (notmany)
+       PULSE_SYSCONTROL (dev, CONT_CLEAR_WALK_WROTE_TABLES);
+    else
+       elan4mmu_flush_tlb(dev);
+
+    do {
+       for (hashidx = baseidx; hashidx <= topidx; hashidx++)
+           if (dev->dev_mmuhash[tbl][hashidx].he_tag[0] & HE_TAG_COPY)
+           {
+               ASSERT ((dev->dev_mmuhash[tbl][hashidx].he_tag[0] & HE_TAG_VALID) == 0);
+               ASSERT ((dev->dev_mmuhash[tbl][hashidx].he_tag[1] & HE_TAG_VALID) == 0);
+
+               elan4mmu_synctag (dev, &dev->dev_mmuhash[tbl][hashidx], 0);
+               elan4mmu_synctag (dev, &dev->dev_mmuhash[tbl][hashidx], 1);
+           }
+       
+       PULSE_SYSCONTROL (dev, CONT_TLB_FLUSH);
+       
+       while ((reg = read_reg64 (dev, SysControlReg)) & CONT_TLB_FLUSH)
+           DELAY (1);
+       
+    } while (notmany-- && (reg & CONT_CLEAR_WALK_WROTE_TABLES) != 0);
+}
+
+void
+elan4mmu_display_hent (ELAN4_DEV *dev, ELAN4_HASH_ENTRY *he, int hashidx)
+{
+    int tagidx;
+
+    elan4_debugf (DBG_DEVICE, DBG_MMU, "elan4mmu_display_hent: hashidx=%d he=%p entry at %lx\n", hashidx, he, he->he_entry);
+    elan4_debugf (DBG_DEVICE, DBG_MMU, "                       next=%p prev=%p chain=%p,%p\n", he->he_next, he->he_prev, he->he_chain[0], he->he_chain[1]);
+    for (tagidx = 0; tagidx < 2; tagidx++)
+    {
+       E4_uint64 tag  = elan4_sdram_readq (dev, he->he_entry + E4MMU_TAG_OFFSET(tagidx));
+       E4_uint64 pte0 = elan4_sdram_readq (dev, he->he_entry + E4MMU_PTE_LOW_OFFSET(tagidx, 0));
+       E4_uint64 pte1 = elan4_sdram_readq (dev, he->he_entry + E4MMU_PTE_LOW_OFFSET(tagidx, 1));
+       E4_uint64 pte2 = elan4_sdram_readq (dev, he->he_entry + E4MMU_PTE_LOW_OFFSET(tagidx, 2));
+       E4_uint64 pte3 = ((pte0 >> 48) | (pte1 >> 32) | (pte2 >> 16));
+       
+       elan4_debugf (DBG_DEVICE, DBG_MMU, "                       Tag %d (%llx,%08x) context=%04x vaddr=%llx\n", tagidx, he->he_tag[tagidx], he->he_pte[tagidx], (int) (tag & TAG_CONTEXT_MASK), (tag & TAG_ADDRESS_MASK));
+       elan4_debugf (DBG_DEVICE, DBG_MMU, "                       Pte 0 - PPN=%llx PERM=%x TYPE=%x%s%s\n", (pte0 & PTE_PPN_MASK) >> PTE_PPN_SHIFT, 
+                     (int) (pte0 & PTE_PERM_MASK) >> PTE_PERM_SHIFT, (int)(pte0 & PTE_TYPE_MASK), (pte0 & PTE_MOD_MASK) ? " mod" : "", (pte0 & PTE_REF_MASK) ? " ref" : "");
+       elan4_debugf (DBG_DEVICE, DBG_MMU, "                       Pte 1 - PPN=%llx PERM=%x TYPE=%x%s%s\n", (pte1 & PTE_PPN_MASK) >> PTE_PPN_SHIFT, 
+                     (int) (pte1 & PTE_PERM_MASK) >> PTE_PERM_SHIFT, (int)(pte1 & PTE_TYPE_MASK), (pte1 & PTE_MOD_MASK) ? " mod" : "", (pte1 & PTE_REF_MASK) ? " ref" : "");
+       elan4_debugf (DBG_DEVICE, DBG_MMU, "                       Pte 2 - PPN=%llx PERM=%x TYPE=%x%s%s\n", (pte2 & PTE_PPN_MASK) >> PTE_PPN_SHIFT, 
+                     (int) (pte2 & PTE_PERM_MASK) >> PTE_PERM_SHIFT, (int)(pte2 & PTE_TYPE_MASK), (pte2 & PTE_MOD_MASK) ? " mod" : "", (pte2 & PTE_REF_MASK) ? " ref" : "");
+       elan4_debugf (DBG_DEVICE, DBG_MMU, "                       Pte 3 - PPN=%llx PERM=%x TYPE=%x%s%s\n", (pte3 & PTE_PPN_MASK) >> PTE_PPN_SHIFT, 
+                     (int) (pte3 & PTE_PERM_MASK) >> PTE_PERM_SHIFT, (int)(pte3 & PTE_TYPE_MASK), (pte3 & PTE_MOD_MASK) ? " mod" : "", (pte3 & PTE_REF_MASK) ? " ref" : "");
+    }
+}
+
+static __inline__ ELAN4_HASH_ENTRY *
+he_ctxt_next (ELAN4_HASH_ENTRY *he, int ctxnum)
+{
+    return ((he->he_tag[0] & TAG_CONTEXT_MASK) == ctxnum) ? he->he_chain[0] : he->he_chain[1];
+}
+
+static __inline__ ELAN4_HASH_ENTRY *
+he_ctxt_unlink (ELAN4_CTXT *ctxt, int tbl, int hashidx, ELAN4_HASH_ENTRY *prevhe, ELAN4_HASH_ENTRY *he, ELAN4_HASH_ENTRY *next)
+{
+    /* Check whether either tag is in use by this context */
+    if ((he->he_tag[0] & TAG_CONTEXT_MASK) == ctxt->ctxt_num || (he->he_tag[1] & TAG_CONTEXT_MASK) == ctxt->ctxt_num)
+       return he;
+
+    if (prevhe == NULL)
+       ctxt->ctxt_mmuhash[tbl][hashidx] = next;
+    else
+    {
+       /* previous he, ensure that both chain pointers are changed is this ctxt is using both tags */
+       ASSERT ((prevhe->he_tag[0] & TAG_CONTEXT_MASK) == ctxt->ctxt_num || (prevhe->he_tag[1] & TAG_CONTEXT_MASK) == ctxt->ctxt_num);
+
+       if ((prevhe->he_tag[0] & TAG_CONTEXT_MASK) == ctxt->ctxt_num)
+           prevhe->he_chain[0] = next;
+       if ((prevhe->he_tag[1] & TAG_CONTEXT_MASK) == ctxt->ctxt_num)
+           prevhe->he_chain[1] = next;
+    }
+
+    return prevhe;
+}
+
+void
+elan4mmu_display (ELAN4_CTXT *ctxt, int tbl, const char *tag)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+    ELAN4_HASH_ENTRY *he;
+    int hashidx;
+
+    for (hashidx = 0; hashidx < dev->dev_hashsize[tbl]; hashidx++)
+       for (he = ctxt->ctxt_mmuhash[tbl][hashidx]; he != NULL; he = he_ctxt_next (he, ctxt->ctxt_num))
+       {
+           elan4_debugf (DBG_DEVICE, DBG_MMU, "%s: hashidx=%d he=%p tags <%llx,%llx>\n", tag, hashidx, he,
+                         (he->he_tag[0] & TAG_CONTEXT_MASK) == ctxt->ctxt_num ? E4MMU_TAG2VADDR (he->he_tag[0], hashidx, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1) : 0,
+                         (he->he_tag[1] & TAG_CONTEXT_MASK) == ctxt->ctxt_num ? E4MMU_TAG2VADDR (he->he_tag[1], hashidx, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1) : 0);
+           elan4mmu_display_hent (dev, he, hashidx);
+       }
+}
+
+static ELAN4_HASH_ENTRY *
+elan4mmu_alloc_hent (ELAN4_DEV *dev, int tbl, int hashidx, E4_uint64 newtag, int *tagidx)
+{
+    ELAN4_HASH_ENTRY *he, *phe;
+    unsigned long flags;
+    int i;
+
+    spin_lock_irqsave (&dev->dev_mmulock, flags);
+
+    /* 2nd see if there are any partial free blocks */
+    if ((he = dev->dev_mmufree[tbl][hashidx]) != NULL)
+    {
+       *tagidx = ((he->he_tag[0] & TAG_CONTEXT_MASK) == INVALID_CONTEXT) ? 0 : 1;
+       
+       MPRINTF (DBG_DEVICE, 3, "elan4mmu_alloc_hent: allocate he=%p idx=%d%s\n", he, *tagidx, (he == &dev->dev_mmuhash[tbl][hashidx]) ? " hash-block" : "");
+       
+       he->he_tag[*tagidx] = newtag | HE_TAG_VALID;
+
+       elan4mmu_synctag (dev, he, *tagidx);
+       
+       if ((he->he_tag[(*tagidx) ^ 1] & TAG_CONTEXT_MASK) != INVALID_CONTEXT)
+       {
+           MPRINTF (DBG_DEVICE, 3, "elan4mmu_alloc_hent: block full - remove from freelist\n");
+           dev->dev_mmufree[tbl][hashidx] = he->he_chain[*tagidx];
+       }
+       
+       spin_unlock_irqrestore (&dev->dev_mmulock, flags);
+       return (he);
+    }
+    
+    if ((he = dev->dev_mmufreelist) != NULL)
+       dev->dev_mmufreelist = he->he_next;
+    else
+    {
+       ELAN4_HASH_CHUNK *hc;
+       sdramaddr_t       entry;
+
+       KMEM_ALLOC (hc, ELAN4_HASH_CHUNK *, sizeof (ELAN4_HASH_CHUNK), 0);
+       
+       if (hc == NULL)
+       {
+           spin_unlock_irqrestore (&dev->dev_mmulock, flags);
+           return ((ELAN4_HASH_ENTRY *) NULL);
+       }
+       
+       if ((entry = elan4_sdram_alloc (dev, sizeof (E4_HashTableEntry) * ELAN4_HENT_CHUNKS)) == (sdramaddr_t) 0)
+       {
+           spin_unlock_irqrestore (&dev->dev_mmulock, flags);
+
+           KMEM_FREE (hc, sizeof (ELAN4_HASH_CHUNK));
+           return ((ELAN4_HASH_ENTRY *) NULL);
+       }
+
+       list_add_tail (&hc->hc_link, &dev->dev_hc_list);
+
+       elan4_sdram_zeroq_sdram (dev, entry, sizeof (E4_HashTableEntry) * ELAN4_HENT_CHUNKS);
+
+       /* no initialise all chunks and chain all but the first onto the freelist */
+       for (i = 0; i < ELAN4_HENT_CHUNKS; i++, entry += sizeof (E4_HashTableEntry))
+       {
+           hc->hc_hents[i].he_entry = entry;
+
+           if (i == 0)
+               he = &hc->hc_hents[0];
+           else
+           {
+               hc->hc_hents[i].he_next = dev->dev_mmufreelist;
+               dev->dev_mmufreelist = &hc->hc_hents[i];
+           }
+       }
+    }
+
+    /* Initialise hash entry, using slot 0 */
+    *tagidx = 0;
+
+    he->he_next     = NULL;
+    he->he_prev     = NULL;
+    he->he_chain[0] = NULL;
+    he->he_chain[1] = NULL;
+    he->he_tag[0]   = newtag | HE_TAG_VALID;
+    he->he_tag[1]   = E4MMU_TAG(0, INVALID_CONTEXT);
+    he->he_pte[0]   = 0;
+    he->he_pte[1]   = 0;
+    
+    elan4mmu_synctag (dev, he, 0);
+    
+    /* add slot 1 to freelist */
+    he->he_chain[1] = dev->dev_mmufree[tbl][hashidx];
+    dev->dev_mmufree[tbl][hashidx] = he;
+    
+    /* add to mmuhash lists */
+    for (phe = &dev->dev_mmuhash[tbl][hashidx]; phe->he_next; phe = phe->he_next)
+       ;
+    phe->he_next = he;
+    he->he_prev  = phe;
+    he->he_next  = NULL;
+    
+    /* finally chain the hash block into the hash tables */
+    elan4mmu_chain_hents (dev, phe, he);
+    
+    spin_unlock_irqrestore (&dev->dev_mmulock, flags);
+    return (he);
+}
+
+static void
+elan4mmu_free_hent (ELAN4_DEV *dev, int tbl, int hashidx, ELAN4_HASH_ENTRY *he, int tagidx)
+{
+    unsigned long flags;
+    int pteidx;
+
+    /* Invalidate the tag, and zero all ptes */
+    for (pteidx = 0; pteidx < 4; pteidx++)
+       if (HE_GET_PTE(he, tagidx, pteidx))
+           elan4mmu_writepte (dev, he, tagidx, pteidx, 0);
+
+    spin_lock_irqsave (&dev->dev_mmulock, flags);
+
+    he->he_tag[tagidx] = E4MMU_TAG(0, INVALID_CONTEXT);
+    he->he_pte[tagidx] = 0;
+
+    elan4mmu_synctag (dev, he, tagidx);
+
+    if ((he->he_tag[tagidx^1] & TAG_CONTEXT_MASK) == INVALID_CONTEXT) /* Both tags are now free */
+    {
+       if (he == &dev->dev_mmuhash[tbl][hashidx])              /* it's the hash block entry */
+       {                                                       /* so as it's already on the freelist */
+           he->he_chain[tagidx] = he->he_chain[tagidx^1];      /* just copy it's chain pointers */
+
+           MPRINTF (DBG_DEVICE, 3, "elan4mmu_free_hent: tbl=%d hashidx=%x tagidx=%d he=%p => all free but hashblk\n", tbl, hashidx, tagidx, he);
+       }
+       else
+       {
+           MPRINTF (DBG_DEVICE, 3, "elan4mmu_free_hent: tbl=%d hashidx=%x tagidx=%d he=%p => all free\n", tbl, hashidx, tagidx, he);
+           
+           /* XXXX - should remove it from the hash table, and 
+           *         place back on the anonymous freelist */
+           he->he_chain[tagidx] = he->he_chain[tagidx^1];
+       }
+    }
+    else
+    {
+       /* Other tag still in use */
+       he->he_chain[tagidx] = dev->dev_mmufree[tbl][hashidx];
+       dev->dev_mmufree[tbl][hashidx] = he;
+
+       MPRINTF (DBG_DEVICE, 3, "elan4mmu_free_hent: tbl=%d hashidx=%x tagidx=%d he=%p => other tag in use\n", tbl, hashidx, tagidx, he);
+    }
+    spin_unlock_irqrestore (&dev->dev_mmulock, flags);
+}
+
+ELAN4_HASH_ENTRY *
+elan4mmu_ptealloc (ELAN4_CTXT *ctxt, int tbl, E4_Addr vaddr, unsigned int *tagidxp)
+{
+    ELAN4_DEV        *dev     = ctxt->ctxt_dev;
+    unsigned         ctxnum  = ctxt->ctxt_num;
+    unsigned          hashidx = E4MMU_HASH_INDEX (ctxnum, vaddr, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1);
+    E4_uint64         newtag  = E4MMU_TAG(vaddr, ctxnum);
+    ELAN4_HASH_ENTRY *he      = &dev->dev_mmuhash[tbl][hashidx];
+    unsigned         tagidx;
+
+    MPRINTF (ctxt, 2, "elan4mmu_ptealloc: tbl=%d ctxnum=%d vaddr=%llx -> hashidx %d\n", tbl, ctxnum, vaddr, hashidx);
+
+    /* 1st) check whether we're reloading an existing entry */
+    for (he = ctxt->ctxt_mmuhash[tbl][hashidx]; he != NULL; he = he_ctxt_next (he, ctxnum))
+    {
+       ASSERT ((he->he_tag[0] & TAG_CONTEXT_MASK) == ctxnum || (he->he_tag[1] & TAG_CONTEXT_MASK) == ctxnum);
+
+       for (tagidx = 0; tagidx < 2; tagidx++)
+       {
+           if ((he->he_tag[tagidx] & (TAG_ADDRESS_MASK | TAG_CONTEXT_MASK | HE_TAG_VALID)) == (newtag | HE_TAG_VALID))
+           {
+               MPRINTF (ctxt, 2, "elan4mmu_ptealloc: return old he %p tagidx %d\n", he, tagidx);
+
+               *tagidxp = tagidx;
+               return he;
+           }
+       }
+    }
+
+    if ((he = elan4mmu_alloc_hent (dev, tbl, hashidx, newtag, &tagidx)) == NULL)
+       return NULL;
+
+    /* chain onto context hash */
+    if ((he->he_tag[tagidx ^ 1] & TAG_CONTEXT_MASK) == ctxnum) /* already chained using other link */
+    {                                                          /* so ensure both slots are chained the same */
+       he->he_chain[tagidx] = he->he_chain[tagidx^1];
+    }
+    else
+    {
+       he->he_chain[tagidx] = ctxt->ctxt_mmuhash[tbl][hashidx];
+       ctxt->ctxt_mmuhash[tbl][hashidx] = he;
+    }
+
+    MPRINTF (ctxt, 2, "elan4mmu_ptealloc: return new he %p tagidx %d\n", he, tagidx);
+
+    *tagidxp = tagidx;
+
+    return he;
+}
+
+int
+elan4mmu_pteload (ELAN4_CTXT *ctxt, int tbl, E4_Addr vaddr, E4_uint64 newpte)
+{
+    ELAN4_DEV        *dev     = ctxt->ctxt_dev;
+    unsigned          pteidx  = E4MMU_SHIFT_ADDR(vaddr, dev->dev_pageshift[tbl]) & 3;
+    unsigned         tagidx;
+    ELAN4_HASH_ENTRY *he;
+
+    MPRINTF (ctxt, 0, "elan4mmu_pteload: ctx=%d tbl=%d pteidx=%d vaddr=%llx pte=%llx\n", 
+           ctxt->ctxt_num, tbl, pteidx, (unsigned long long)vaddr, newpte);
+
+    spin_lock (&ctxt->ctxt_mmulock);
+
+    if ((he = elan4mmu_ptealloc (ctxt, tbl, vaddr, &tagidx)) == NULL)
+    {
+       spin_unlock (&ctxt->ctxt_mmulock);
+       return -ENOMEM;
+    }
+
+    MPRINTF (ctxt, 1, "elan4mmu_pteload: %s he=%p tagidx=%d pteidx=%d\n", HE_GET_PTE(he,0,pteidx) ? "reloading" : "loading", he, tagidx, pteidx);
+    
+    ASSERT (HE_GET_PTE(he,tagidx,pteidx) == 0 ||                                                       /* invalid -> valid */
+           (elan4mmu_readpte (dev, he, tagidx, pteidx) & PTE_PPN_MASK) == (newpte & PTE_PPN_MASK));    /* or same phys address */
+    
+    elan4mmu_writepte (dev, he, tagidx, pteidx, newpte);
+    
+    HE_SET_PTE(he, tagidx, pteidx, (newpte & PTE_PERM_TYPE_MASK));
+
+    spin_unlock (&ctxt->ctxt_mmulock);
+    return 0;
+}
+
+void
+elan4mmu_unload_range (ELAN4_CTXT *ctxt, int tbl, E4_Addr start, unsigned long len)
+{
+    ELAN4_DEV        *dev       = ctxt->ctxt_dev;
+    unsigned          ctxnum    = ctxt->ctxt_num;
+    unsigned long     tagspan   = (1 << (dev->dev_pageshift[tbl] + 2));
+    E4_Addr           end       = start + len - 1;
+    int                      needflush = 0;
+    unsigned          baseidx, topidx;
+    unsigned          hashidx, tagidx, pteidx;
+    ELAN4_HASH_ENTRY *he, *prevhe, *next;
+    
+    MPRINTF (ctxt, 0, "elan4mmu_unload_range: tbl=%d start=%llx end=%llx len=%lx\n", tbl, start, end, len);
+
+    /* determine how much of the hash table we've got to scan */
+    
+    /* GNAT 6760: When we have a Main page size which maps onto multiple Elan pages
+     * we need to do something a bit more clever here or else it takes ms per page invalidate
+     * This change helps in the meantime
+     */
+    /* if (len <= (1 << dev->dev_pageshift[tbl])) */
+    if (len <= PAGE_SIZE)
+    {
+       baseidx = E4MMU_HASH_INDEX (ctxnum, start, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1);
+       topidx  = E4MMU_HASH_INDEX (ctxnum, end,   dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1);
+
+       if (baseidx != topidx)
+       {
+           /* GNAT 6760: Need to search whole of the hash table (slow!) */
+           baseidx = 0;
+           topidx  = dev->dev_hashsize[tbl] - 1;
+       }
+    }
+    else
+    {
+       baseidx = 0;
+       topidx  = dev->dev_hashsize[tbl] - 1;
+    }
+
+    MPRINTF (ctxt, 1, "elan4mmu_unload_range: baseidx=%d topidx=%d\n", baseidx, topidx);
+
+    spin_lock (&ctxt->ctxt_mmulock);
+
+    /* 1st - invalidate the tag for all hash blocks which are completely invalidated,
+     *       and remember the first/last hash blocks */
+    for (hashidx = baseidx; hashidx <= topidx; hashidx++)
+       for (he = ctxt->ctxt_mmuhash[tbl][hashidx]; he != NULL; he = he_ctxt_next (he, ctxnum))
+           for (tagidx = 0; tagidx < 2; tagidx++)
+               if ((he->he_tag[tagidx] & TAG_CONTEXT_MASK) == ctxnum)
+               {
+                   E4_Addr base = E4MMU_TAG2VADDR (he->he_tag[tagidx], hashidx, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1);
+                   E4_Addr top  = base + (tagspan -1);
+                   
+                   if (start < top && end > base)
+                   {
+                       unsigned bidx = (start <= base) ? 0 : (start & (tagspan-1)) >> dev->dev_pageshift[tbl];
+                       unsigned tidx = (end   >= top)  ? 3 : (end   & (tagspan-1)) >> dev->dev_pageshift[tbl];
+                       
+                       MPRINTF (ctxt, 1, "elan4mmu_unload_range: he=%p base=%llx top=%llx hashidx=%d bidx=%d tidx=%d\n", he, base, top, hashidx, bidx, tidx);
+                       
+                       for (pteidx = bidx; pteidx <= tidx; pteidx++)
+                           if (HE_GET_PTE(he, tagidx, pteidx))
+                           {
+                               elan4mmu_invalidatepte (dev, he, tagidx, pteidx);
+                               needflush = 1;
+                           }
+                   }
+                   else if (base >= start && top <= end)               /* hash entry completely spanned */
+                   {                                                   /* so invalidate the tag */
+                       MPRINTF (ctxt, 1, "elan4mmu_unload_range: he=%p base=%llx top=%llx spanned\n", he, base, top);
+
+                       he->he_tag[tagidx] &= ~HE_TAG_VALID;
+                       
+                       elan4mmu_synctag (dev, he, tagidx);
+                       needflush = 1;
+                   }
+               }
+
+    if (needflush)
+    {
+       /* 2nd invalidate the first/last hash blocks if they are partially invalidated
+        * and flush the tlb/hash copy blocks */
+       elan4mmu_flush_tlb_hash (dev, tbl, baseidx, topidx);
+       
+       /* 3rd free off the hash entries which are completely invalidated */
+       for (hashidx = baseidx; hashidx <= topidx; hashidx++)
+           for (prevhe = NULL, he = ctxt->ctxt_mmuhash[tbl][hashidx]; he != NULL; he = next)
+           {
+               next = he_ctxt_next (he, ctxnum);
+               
+               for (tagidx = 0; tagidx < 2; tagidx++)
+                   if ((he->he_tag[tagidx] & TAG_CONTEXT_MASK) == ctxnum)
+                   {
+                       E4_Addr base = E4MMU_TAG2VADDR (he->he_tag[tagidx], hashidx, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1);
+                       E4_Addr top  = base + (tagspan -1);
+                       
+                       if (start < top && end > base)
+                       {
+                           unsigned bidx = (start <= base) ? 0 : (start & (tagspan-1)) >> dev->dev_pageshift[tbl];
+                           unsigned tidx = (end   >= top)  ? 3 : (end   & (tagspan-1)) >> dev->dev_pageshift[tbl];
+                           
+                           MPRINTF (ctxt, 1, "elan4mmu_unload_range: he=%p base=%llx top=%llx bidx=%d tidx=%d\n", he, base, top, bidx, tidx);
+                           
+                           for (pteidx = bidx; pteidx <= tidx; pteidx++)
+                               if (HE_GET_PTE(he, tagidx, pteidx))
+                               {
+                                   HE_SET_PTE(he, tagidx, pteidx, 0);
+                                   
+                                   elan4mmu_writepte (dev, he, tagidx, pteidx, 0);
+                               }
+                       }
+                       
+                       if ((base >= start && top <= end) || he->he_pte[tagidx] == 0)   /* hash entry completely spanned or all pte's cleared */
+                       {                                                                       /* so invalidate the pte's and free it */
+                           
+                           MPRINTF (ctxt, 1, "elan4mmu_unload_range: he=%p base=%llx top=%llx spanned or empty\n", he, base, top);
+                           
+                           elan4mmu_free_hent (dev, tbl, hashidx, he, tagidx);
+                       }
+                   }
+               
+               prevhe = he_ctxt_unlink (ctxt, tbl, hashidx, prevhe, he, next);
+           }
+    }
+    spin_unlock (&ctxt->ctxt_mmulock);
+}
+
+void
+elan4mmu_invalidate_ctxt (ELAN4_CTXT *ctxt)
+{
+    ELAN4_DEV *dev    = ctxt->ctxt_dev;
+    int        ctxnum = ctxt->ctxt_num;
+    ELAN4_HASH_ENTRY *he;
+    int tbl, hashidx, tagidx;
+
+    MPRINTF (ctxt, 0, "elan4mmu_invalidate_ctxt: invalidating ctxnum=%d\n", ctxnum);
+
+    spin_lock (&ctxt->ctxt_mmulock);
+
+    /* 1st invalidate all tags belonging to me */
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+       for (hashidx = 0; hashidx < dev->dev_hashsize[tbl]; hashidx++)
+           for (he = ctxt->ctxt_mmuhash[tbl][hashidx]; he != NULL; he = he_ctxt_next (he, ctxnum))
+               for (tagidx = 0; tagidx < 2; tagidx++)
+                   if ((he->he_tag[tagidx] & TAG_CONTEXT_MASK) == ctxnum) /* own tag block */
+                   {
+                       MPRINTF (ctxt, 1, "elan4mmu_invalidate_ctxt: he=%p addr=%llx hashidx=%d tagidx=%d\n", 
+                                he, he->he_tag[tagidx] & TAG_ADDRESS_MASK, hashidx, tagidx);
+
+                       he->he_tag[tagidx] &= ~HE_TAG_VALID;
+                       
+                       elan4mmu_synctag (dev, he, tagidx);
+                   }
+
+    /* 2nd flush the tlb & cached hash block */
+    elan4mmu_flush_tlb (dev);
+    
+    /* 3rd invalidate all pte's and free off the hash entries */
+    for (tbl = 0; tbl < NUM_HASH_TABLES; tbl++)
+       for (hashidx = 0; hashidx < dev->dev_hashsize[tbl]; hashidx++)
+           while ((he = ctxt->ctxt_mmuhash[tbl][hashidx]) != NULL)
+           {
+               ctxt->ctxt_mmuhash[tbl][hashidx] = he_ctxt_next (he, ctxnum);
+
+               for (tagidx = 0; tagidx < 2; tagidx++)
+                   if ((he->he_tag[tagidx] & TAG_CONTEXT_MASK) == ctxnum)
+                       elan4mmu_free_hent (dev, tbl, hashidx, he, tagidx);
+           }
+    spin_unlock (&ctxt->ctxt_mmulock);
+}
+
+ELAN4_HASH_CACHE *
+elan4mmu_reserve (ELAN4_CTXT *ctxt, int tbl, E4_Addr start, unsigned int npages, int cansleep)
+{
+    ELAN4_DEV        *dev      = ctxt->ctxt_dev;
+    E4_Addr           end      = start + (npages << dev->dev_pageshift[tbl]) - 1;
+    unsigned long     tagshift = dev->dev_pageshift[tbl] + 2;
+    E4_Addr           tagspan  = 1 << tagshift;
+    E4_Addr           base     = (start & ~(tagspan-1));
+    E4_Addr           top      = (end   & ~(tagspan-1)) + (tagspan-1);
+    unsigned int      nhes     = (top - base + 1) >> tagshift;
+    ELAN4_HASH_CACHE *hc;
+    unsigned int      tagidx,  pteidx;
+    E4_Addr           addr;
+    int                      i;
+    
+    MPRINTF (ctxt, 0, "elan4mmu_reserve: start=%llx npages=%d\n", start, npages);
+    MPRINTF (ctxt, 0, "         pageshift=%d tagspan=%lx base=%llx top=%llx end=%llx nhes=%d\n",
+            dev->dev_pageshift[tbl], tagspan, base, top, end, nhes);
+
+    KMEM_ALLOC (hc, ELAN4_HASH_CACHE *, offsetof (ELAN4_HASH_CACHE, hc_hes[nhes]), cansleep);
+
+    if (hc == NULL)
+       return NULL;
+
+    hc->hc_start = start;
+    hc->hc_end   = end;
+    hc->hc_tbl   = tbl;
+
+    spin_lock (&ctxt->ctxt_mmulock);
+    for (addr = base, i = 0; i < nhes; addr += tagspan, i++)
+    {
+       unsigned bidx = (i == 0)        ? (start & (tagspan-1)) >> dev->dev_pageshift[tbl] : 0;
+       unsigned tidx = (i == (nhes-1)) ? (end   & (tagspan-1)) >> dev->dev_pageshift[tbl] : 3;
+
+       
+       if ((hc->hc_hes[i] = elan4mmu_ptealloc (ctxt, tbl, addr & ~(tagspan-1), &tagidx)) == NULL)
+           goto failed;
+
+
+       MPRINTF (ctxt, 2, "elan4mmu_reserve: tbl=%d addr=%llx -> hashidx=%d tagidx=%d\n", tbl, addr & ~(tagspan-1), 
+                E4MMU_HASH_INDEX (ctxt->ctxt_num, (addr & ~(tagspan-1)), dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1), tagidx);
+                        
+       for (pteidx = bidx; pteidx <= tidx; pteidx++)
+       {
+           ASSERT (HE_GET_PTE (hc->hc_hes[i], tagidx, pteidx) == 0);
+
+           MPRINTF (ctxt, 2, "elan4mmu_reserve: i=%d addr=%llx he=%p (tagidx=%d pteidx=%d)\n",
+                    i, addr, hc->hc_hes[i], tagidx, pteidx);
+
+           HE_SET_PTE (hc->hc_hes[i], tagidx, pteidx, PTE_PERM_TYPE_MASK);
+       }
+    }
+    spin_unlock (&ctxt->ctxt_mmulock);
+
+    return hc;
+
+ failed:
+    for (i--, addr -= tagspan; i >= 0; i--, addr -= tagspan)
+    {
+       unsigned bidx    = (i == 0) ? (start & (tagspan-1)) >> dev->dev_pageshift[tbl] : 0;
+       unsigned tidx    = (i == (nhes-1)) ? (end   & (tagspan-1)) >> dev->dev_pageshift[tbl] : 3;
+       unsigned hashidx = E4MMU_HASH_INDEX (ctxt->ctxt_num, addr, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1);
+       unsigned tagidx  = (addr == E4MMU_TAG2VADDR (hc->hc_hes[i]->he_tag[0], hashidx, dev->dev_pageshift[tbl], dev->dev_hashsize[tbl]-1)) ? 0 : 1;
+           
+       for (pteidx = bidx; pteidx <= tidx; pteidx++)
+           HE_SET_PTE(hc->hc_hes[i], tagidx, pteidx, 0);
+
+       if (hc->hc_hes[i]->he_pte[tagidx] == 0)
+           elan4mmu_free_hent (dev, tbl, hashidx, hc->hc_hes[i], tagidx);
+    }
+    spin_unlock (&ctxt->ctxt_mmulock);
+
+    KMEM_FREE (hc, offsetof (ELAN4_HASH_CACHE, hc_hes[nhes]));
+    
+    return NULL;
+}
+
+void
+elan4mmu_release (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc)
+{
+    ELAN4_DEV        *dev      = ctxt->ctxt_dev;
+    E4_Addr          start    = hc->hc_start;
+    E4_Addr           end      = hc->hc_end;
+    unsigned long     tagshift = dev->dev_pageshift[hc->hc_tbl] + 2;
+    E4_Addr           tagspan  = 1 << tagshift;
+    E4_Addr           base     = (start & ~(tagspan-1));
+    E4_Addr           top      = (end   & ~(tagspan-1)) + (tagspan-1);
+    unsigned int      nhes     = (top - base + 1) >> tagshift;
+    ELAN4_HASH_ENTRY *prevhe, *he, *next;
+    E4_Addr           addr;
+    unsigned int      pteidx;
+    int                      i;
+
+    spin_lock (&ctxt->ctxt_mmulock);
+
+    MPRINTF (ctxt, 0, "elan4mmu_release: base=%llx top=%llx\n", base, top);
+
+    for (addr = base, i = 0; i < nhes; addr += tagspan, i++)
+    {
+       unsigned bidx    = (i == 0)        ? (start & (tagspan-1)) >> dev->dev_pageshift[hc->hc_tbl] : 0;
+       unsigned tidx    = (i == (nhes-1)) ? (end   & (tagspan-1)) >> dev->dev_pageshift[hc->hc_tbl] : 3;
+       unsigned hashidx = E4MMU_HASH_INDEX (ctxt->ctxt_num, addr, dev->dev_pageshift[hc->hc_tbl], dev->dev_hashsize[hc->hc_tbl]-1);
+       unsigned tagidx  = (addr == E4MMU_TAG2VADDR (hc->hc_hes[i]->he_tag[0], hashidx, dev->dev_pageshift[hc->hc_tbl], dev->dev_hashsize[hc->hc_tbl]-1)) ? 0 : 1;
+           
+       for (pteidx = bidx; pteidx <= tidx; pteidx++)
+       {
+           elan4mmu_invalidatepte (dev, hc->hc_hes[i], tagidx, pteidx);
+
+           HE_SET_PTE(hc->hc_hes[i], tagidx, pteidx, 0);
+       }
+
+       MPRINTF (ctxt, 2, "elan4mmu_release: i=%d addr=%llx he=%p (hashidx=%d tagidx=%d pteidx=%d) pte=%x\n",
+                i, addr, hc->hc_hes[i], hashidx, tagidx, pteidx, hc->hc_hes[i]->he_pte[tagidx]);
+
+       /* remove from context hash */
+       for (prevhe = NULL, he = ctxt->ctxt_mmuhash[hc->hc_tbl][hashidx], next = he_ctxt_next (he, ctxt->ctxt_num); he != hc->hc_hes[i]; he = next)
+           next = he_ctxt_next (he, ctxt->ctxt_num);
+       
+       if (hc->hc_hes[i]->he_pte[tagidx] == 0)
+           elan4mmu_free_hent (dev, hc->hc_tbl, hashidx, hc->hc_hes[i], tagidx);
+
+       prevhe = he_ctxt_unlink (ctxt, hc->hc_tbl, hashidx, prevhe, he, next);
+    }
+    spin_unlock (&ctxt->ctxt_mmulock);
+}
+
+void
+elan4mmu_set_pte (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc, unsigned int idx, E4_uint64 newpte)
+{
+    ELAN4_DEV        *dev      = ctxt->ctxt_dev;
+    unsigned int      tbl      = hc->hc_tbl;
+    unsigned int      tagshift = dev->dev_pageshift[tbl] + 2;
+    E4_Addr           tagspan  = 1 << tagshift;
+    E4_Addr           addr     = hc->hc_start + (idx << dev->dev_pageshift[tbl]);
+    ELAN4_HASH_ENTRY *he       = hc->hc_hes[(addr - (hc->hc_start & ~(tagspan-1))) >> tagshift];
+    unsigned          pteidx   = E4MMU_SHIFT_ADDR(addr, dev->dev_pageshift[tbl]) & 3;
+    unsigned          tagidx   = he->he_tag[0] == (E4MMU_TAG (addr, ctxt->ctxt_num) | HE_TAG_VALID) ? 0 : 1;
+
+    MPRINTF (ctxt, 2, "elan4mmu_set_pte: idx=%d addr=%llx he=%p (tagidx=%d pteidx=%d) newpte=%llx\n", idx, addr, he, tagidx, pteidx, newpte);
+
+    ASSERT (he->he_tag[tagidx] == (E4MMU_TAG (addr, ctxt->ctxt_num) | HE_TAG_VALID));
+
+    elan4mmu_writepte (dev, he, tagidx, pteidx, newpte);
+}
+
+E4_uint64
+elan4mmu_get_pte (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc, unsigned int idx)
+{
+    ELAN4_DEV        *dev      = ctxt->ctxt_dev;
+    unsigned int      tbl      = hc->hc_tbl;
+    unsigned int      tagshift = dev->dev_pageshift[tbl] + 2;
+    E4_Addr           tagspan  = 1 << tagshift;
+    E4_Addr           addr     = hc->hc_start + (idx << dev->dev_pageshift[tbl]);
+    ELAN4_HASH_ENTRY *he       = hc->hc_hes[(addr - (hc->hc_start & ~(tagspan-1))) >> tagshift];
+    unsigned          pteidx   = E4MMU_SHIFT_ADDR(addr, dev->dev_pageshift[tbl]) & 3;
+    unsigned          tagidx   = he->he_tag[0] == (E4MMU_TAG (addr, ctxt->ctxt_num) | HE_TAG_VALID) ? 0 : 1;
+
+    ASSERT (he->he_tag[tagidx] == (E4MMU_TAG (addr, ctxt->ctxt_num) | HE_TAG_VALID));
+
+    return elan4mmu_readpte (dev, he, tagidx, pteidx);
+}
+
+void
+elan4mmu_clear_pte (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc, unsigned int idx)
+{
+    ELAN4_DEV        *dev      = ctxt->ctxt_dev;
+    unsigned int      tbl      = hc->hc_tbl;
+    unsigned int      tagshift = dev->dev_pageshift[tbl] + 2;
+    E4_Addr           tagspan  = 1 << tagshift;
+    E4_Addr           addr     = hc->hc_start + (idx << dev->dev_pageshift[tbl]);
+    ELAN4_HASH_ENTRY *he       = hc->hc_hes[(addr - (hc->hc_start & ~(tagspan-1))) >> tagshift];
+    unsigned          pteidx   = E4MMU_SHIFT_ADDR(addr, dev->dev_pageshift[tbl]) & 3;
+    unsigned          tagidx   = he->he_tag[0] == (E4MMU_TAG (addr, ctxt->ctxt_num) | HE_TAG_VALID) ? 0 : 1;
+
+    MPRINTF (ctxt, 2, "elan4mmu_clear_pte: idx=%d addr=%llx he=%p (tagidx=%d pteidx=%d)\n", idx, addr, he, tagidx, pteidx);
+
+    ASSERT (he->he_tag[tagidx] == (E4MMU_TAG (addr, ctxt->ctxt_num) | HE_TAG_VALID));
+
+    elan4mmu_invalidatepte (dev, he, tagidx, pteidx);
+}
+
+EXPORT_SYMBOL(elan4mmu_flush_tlb);
+EXPORT_SYMBOL(elan4mmu_pteload);
+EXPORT_SYMBOL(elan4mmu_unload_range);
+EXPORT_SYMBOL(elan4mmu_reserve);
+EXPORT_SYMBOL(elan4mmu_release);
+EXPORT_SYMBOL(elan4mmu_set_pte);
+EXPORT_SYMBOL(elan4mmu_get_pte);
+EXPORT_SYMBOL(elan4mmu_clear_pte);
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/mmu_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/mmu_Linux.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/mmu_Linux.c    2005-06-01 23:12:54.611437280 -0400
@@ -0,0 +1,265 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: mmu_Linux.c,v 1.8 2004/05/10 14:10:46 daniel Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/mmu_Linux.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+
+#include <linux/pci.h>
+#include <linux/version.h>
+
+/*
+ * Convert a physical address into an pte.  This should generate a "local" pte for 
+ * physical addresses which are elan4 sdram or elan4 command queues.  For elan4
+ * registers and other addresses on the same bus, this should be the local pci 
+ * bus address.  All other addresses should access the physical address via the
+ * PCI bridge.
+ */
+
+#ifdef __alpha
+#define ioaddr2paddr(ioaddr)   virt_to_phys((void *) __ioremap(ioaddr, PAGE_SIZE))
+#elif defined(__ia64)
+#define ioaddr2paddr(ioaddr)   ((ioaddr) & ~__IA64_UNCACHED_OFFSET)
+#else
+#define ioaddr2paddr(ioaddr)   (ioaddr)
+#endif
+
+int
+elan4mmu_categorise_paddr (ELAN4_DEV *dev, physaddr_t *physp)
+{
+    physaddr_t sdram_base = ioaddr2paddr (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+    physaddr_t sdram_top  = ioaddr2paddr (pci_resource_end (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+    physaddr_t regs_base  = ioaddr2paddr (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS));
+    physaddr_t regs_top   = ioaddr2paddr (pci_resource_end (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS));
+    physaddr_t phys       = *physp;
+    int        iscommand;
+
+    if (phys >= sdram_base && phys <= sdram_top)
+    {
+       (*physp) = (phys ^ sdram_base);
+       return ELAN4MMU_PADDR_SDRAM;
+    }
+    
+    if (phys >= regs_base && phys < regs_top)
+    {
+       if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+           iscommand = (phys < (regs_base + ELAN4_REVA_REG_OFFSET));
+       else
+           iscommand = (phys < (regs_base + ELAN4_REVB_I2C_OFFSET));
+       
+       if (iscommand)
+       {
+           (*physp) = phys ^ regs_base;
+
+           return ELAN4MMU_PADDR_COMMAND;
+       }
+       else
+       {
+           // XXXX (*physp) = phys2bus (phys);
+
+           return ELAN4MMU_PADDR_LOCALPCI;
+       }
+    }
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 0)
+    if (VALID_PAGE (virt_to_page (phys_to_virt (phys))))
+#else
+    if (virt_addr_valid (phys_to_virt (phys)))
+#endif
+       return ELAN4MMU_PADDR_PAGE;
+    
+    return ELAN4MMU_PADDR_OTHER;
+}
+
+int
+elan4mmu_sdram_aliascheck (ELAN4_CTXT *ctxt, E4_Addr addr, physaddr_t phys)
+{
+    ELAN4_DEV *dev = ctxt->ctxt_dev;
+
+    /*
+     * On MPSAS we don't allocate a large enough context table, so 
+     * if we see an address/context pair which would "alias" because
+     * they differ in unchecked hash bits to a previous pteload, 
+     * then we kill the application.
+     */
+    unsigned hashval = (E4MMU_SHIFT_ADDR(addr, (dev->dev_pageshift[0]) + 2) ^ E4MMU_CONTEXT_SCRAMBLE(ctxt->ctxt_num));
+    
+    if (dev->dev_rsvd_hashval[0] == 0xFFFFFFFF)
+       dev->dev_rsvd_hashval[0] = hashval & dev->dev_rsvd_hashmask[0];
+    
+    if ((hashval & dev->dev_rsvd_hashmask[0]) != dev->dev_rsvd_hashval[0])
+    {
+       printk ("elan4mmu_sdram_aliascheck: vaddr=%016llx ctxnum=%x -> [%x] overlaps %x - %x [hashidx=%x]\n", (unsigned long long) addr, 
+               ctxt->ctxt_num, hashval, hashval & dev->dev_rsvd_hashmask[0], dev->dev_rsvd_hashval[0],
+               E4MMU_HASH_INDEX (ctxt->ctxt_num, addr, dev->dev_pageshift[0], dev->dev_hashsize[0]-1));
+       
+       return 0;
+    }
+
+    if (((addr & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT)) != (phys & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT))))
+    {
+       printk ("elan4mmu_sdram_aliascheck: vaddr=%016llx incorrectly alias sdram at %lx\n", (unsigned long long) addr, 
+               phys ^ pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+       return 0;
+    }
+
+    return 1;
+}
+
+int
+elan4mmu_alloc_topaddr (ELAN4_DEV *dev, physaddr_t paddr, unsigned type)
+{
+#if defined(__i386) && !defined(CONFIG_X86_PAE)
+    if (dev->dev_topaddrvalid == 0)
+    {
+       dev->dev_topaddrvalid = 1;
+
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_ELAN_TOPPHYSADDR(0), 0);
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_ELAN_TOPPHYSADDR(1), 0);
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_ELAN_TOPPHYSADDR(2), 0);
+       pci_write_config_word (dev->dev_osdep.pdev, PCI_ELAN_TOPPHYSADDR(3), 0);
+    }
+    return (0);
+#else
+    register int i;
+    E4_uint16 match;
+
+    if (dev->dev_topaddrmode)                                  /* ExtraMasterAddrBits=1 => match {paddr[63:50],type[3:2]} */
+       match = ((paddr >> 48) & ~3) | ((type >> 2) & 3);
+    else                                                       /* ExtraMasterAddrBits=0 => match {paddr[63:48]} */
+       match = (paddr >> 48);
+    
+    MPRINTF (DBG_DEVICE, 2, "elan4mmu_alloc_topaddr: mode=%d paddr=%lx type=%x match=%x [%x %x.%x.%x.%x]\n",
+            dev->dev_topaddrmode, paddr, type, match, dev->dev_topaddrvalid,
+            dev->dev_topaddr[0], dev->dev_topaddr[1], dev->dev_topaddr[2], dev->dev_topaddr[3]);
+    
+    for (i = 0; i < 4; i++)
+       if ((dev->dev_topaddrvalid & (1 << i)) && dev->dev_topaddr[i] == match)
+           return (i);
+    
+    for (i = 0; i < 4; i++)
+    {
+       if ((dev->dev_topaddrvalid & (1 << i)) == 0)
+       {
+           MPRINTF (DBG_DEVICE, 2, "elan4mmu_alloc_topaddr: allocate slot %d for %x\n", i, match);
+
+           dev->dev_topaddrvalid |= (1 << i);
+           dev->dev_topaddr[i] = match;
+
+           pci_write_config_word (dev->dev_osdep.pdev, PCI_ELAN_TOPPHYSADDR(i), match);
+           return (i);
+       }
+    }
+
+    panic ("elan4mmu_alloc_topaddr: all topaddrs in use\n");
+    return (0);
+#endif
+}
+
+E4_uint64
+elan4mmu_phys2pte (ELAN4_DEV *dev, physaddr_t phys, unsigned perm)
+{
+    physaddr_t sdram_base = ioaddr2paddr (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+    physaddr_t sdram_top  = ioaddr2paddr (pci_resource_end (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+    physaddr_t regs_base  = ioaddr2paddr (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS));
+    physaddr_t regs_top   = ioaddr2paddr (pci_resource_end (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS));
+    int        iscommand;
+    E4_uint64  pte;
+    unsigned   type;
+    
+    if (phys >= sdram_base && phys <= sdram_top)
+    {
+       phys ^= sdram_base;
+       type  = PTE_SetPerm (perm);
+    }
+    else if (phys >= regs_base && phys < regs_top)
+    {
+       if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+           iscommand = (phys < (regs_base + ELAN4_REVA_REG_OFFSET));
+       else
+           iscommand = (phys < (regs_base + ELAN4_REVB_I2C_OFFSET));
+       
+       if (iscommand)
+       {
+           phys ^= regs_base;
+           type  = PTE_SetPerm (perm) | PTE_CommandQueue;
+       }
+       else
+       {
+           type = PTE_SetPerm (perm) | PTE_PciNotLocal;
+           // phys = phys2bus (phys);
+       }
+    }
+    else
+    {
+       type = PTE_SetPerm (perm) | PTE_PciNotLocal | dev->dev_pteval;
+
+#ifdef LINUX_SPARC
+       /* XXXX if not local pci bus, then or in the bypass bit */
+       phys |= 0xfffe000000000000;
+       type |= PTE_BigEndian;
+#endif
+
+
+#if defined(__alpha)
+       phys |= alpha_mv.pci_dac_offset;
+#endif
+    }
+
+    if ((type & PTE_PciNotLocal) == 0)
+       pte = (phys >> PTE_PADDR_SHIFT) | type;
+    else
+    {
+       unsigned topaddr = elan4mmu_alloc_topaddr (dev, phys, type);
+       
+       if (dev->dev_topaddrmode)
+           pte = (phys >> PTE_PADDR_SHIFT) | (type & ~0xc) | (topaddr << 2);
+       else
+           pte = ((phys >> PTE_PADDR_SHIFT) & ~PTE_TOPADDR_MASK) | (((E4_uint64) topaddr) << 45) | type;
+    }
+
+    return pte;
+}
+
+physaddr_t
+elan4mmu_pte2phys (ELAN4_DEV *dev, E4_uint64 pte)
+{
+    physaddr_t sdram_base = ioaddr2paddr (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+    physaddr_t regs_base  = ioaddr2paddr (pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS));
+    physaddr_t phys;
+    
+    if (pte & PTE_PciNotLocal)
+    {
+       if (dev->dev_topaddrmode)
+           phys = ((physaddr_t)(dev->dev_topaddr[(pte >> 2) & 3] & 0xfffc) << 48) | ((pte & PTE_PPN_MASK) << PTE_PADDR_SHIFT);
+       else
+           phys = ((physaddr_t)(dev->dev_topaddr[(pte >> 45) & 3] & 0xffff) << 48)| ((pte & PTE_PPN_MASK & ~PTE_TOPADDR_MASK) << PTE_PADDR_SHIFT);
+
+#ifdef LINUX_SPARC     /* XXXX if not local pci bus, then or in the bypass bit */
+       phys ^= 0xfffe000000000000;
+#endif
+
+#if defined(__alpha)
+       phys ^= alpha_mv.pci_dac_offset;
+#endif
+       return phys;
+    }
+    
+    if (pte & PTE_CommandQueue)
+       return (regs_base | ((pte & PTE_PPN_MASK) << PTE_PADDR_SHIFT));
+    
+    /* sdram */
+    return (sdram_base | ((pte & PTE_PPN_MASK) << PTE_PADDR_SHIFT));
+}
+
+EXPORT_SYMBOL(elan4mmu_phys2pte);
+EXPORT_SYMBOL(elan4mmu_pte2phys);
Index: linux-2.4.21/drivers/net/qsnet/elan4/neterr.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/neterr.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/neterr.c       2005-06-01 23:12:54.612437128 -0400
@@ -0,0 +1,270 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: neterr.c,v 1.4.6.3 2004/11/05 13:11:17 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/neterr.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/sdram.h>
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/commands.h>
+#include <elan4/trtype.h>
+#include <elan4/neterr.h>
+
+typedef struct neterr_inputq
+{
+    E4_InputQueue      inputq;                                 /* input queue */
+    E4_Event32         qevent;                                 /* input queue event */
+    E4_uint64          sent;                                   /* # messages sent (cq flow control)*/
+} NETERR_INPUTQ;
+
+#define NETERR_NSLOTS  64                                      /* single page of queue space (4Kb) */
+
+#define NETERR_RETRIES 16
+#define NETERR_CQ_SIZE CQ_Size8K
+#define NETERR_CQ_MSGS (CQ_Size(NETERR_CQ_SIZE) / (21*8))
+#define NETERR_VP_COUNT        64                                      /* this *must* be > NETERR_CQ_MSGS */
+#define NETERR_VP_BASE 1                                       /* use vp 1 upwards */
+
+void
+elan4_neterr_interrupt (ELAN4_DEV *dev, void *arg)
+{
+    E4_Addr          qfptr  = elan4_sdram_readq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, inputq.q_fptr));
+    E4_Addr          qbptr  = elan4_sdram_readq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, inputq.q_bptr));
+    E4_Addr          qfirst = DEVICE_NETERR_SLOTS_ADDR;
+    E4_Addr          qlast  = qfirst + (NETERR_NSLOTS-1) * ELAN4_NETERR_MSG_SIZE;
+    ELAN4_CQ        *cq     = dev->dev_neterr_intcq;
+    int              count  = 0;
+    ELAN4_CTXT      *ctxt;
+    ELAN4_NETERR_MSG msg;
+
+    while (qfptr != qbptr)
+    {
+       elan4_sdram_copyq_from_sdram (dev, dev->dev_neterr_slots + (qfptr - qfirst), &msg, ELAN4_NETERR_MSG_SIZE);
+
+       ctxt = elan4_networkctxt (dev, msg.msg_context);
+
+       if (ctxt != NULL && ctxt->ctxt_ops->op_neterrmsg)
+           ctxt->ctxt_ops->op_neterrmsg (ctxt, &msg);
+       else
+           PRINTF (DBG_DEVICE, DBG_NETERR, "elan4_neterr_interrupt: no process - sender %d.%d\n", msg.msg_sender.loc_node, msg.msg_sender.loc_context);
+
+       count++;
+
+       /* move on the from pointer */
+       qfptr = (qfptr == qlast) ? qfirst : qfptr + ELAN4_NETERR_MSG_SIZE;
+
+       elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, inputq.q_fptr), qfptr);
+    }
+
+    if (count == 0)
+    {
+       printk ("elan4_neterr_interrupt: spurious\n");
+       return;
+    }
+
+    /* Issue the waitevent to the interrupt queue */
+    writeq (WAIT_EVENT_CMD | (DEVICE_NETERR_INPUTQ_ADDR + offsetof (NETERR_INPUTQ, qevent)),   cq->cq_mapping);
+    writeq (  E4_EVENT_INIT_VALUE (-32 * count, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0),       cq->cq_mapping);
+    writeq (  DEVICE_NETERR_INTCQ_ADDR,                                                                cq->cq_mapping);
+    writeq (INTERRUPT_CMD | (dev->dev_neterr_intop.op_cookie << E4_MAIN_INT_SHIFT),            cq->cq_mapping);
+
+    pioflush_reg (dev);
+}
+
+int
+elan4_neterr_init (ELAN4_DEV *dev)
+{
+    unsigned int intqaddr;
+    E4_Addr     qfirst, qlast;
+    
+    if ((dev->dev_neterr_inputq = elan4_sdram_alloc (dev, SDRAM_PAGE_SIZE)) == 0)
+       return 0;
+
+    if ((dev->dev_neterr_slots = elan4_sdram_alloc (dev, roundup (NETERR_NSLOTS * ELAN4_NETERR_MSG_SIZE, SDRAM_PAGE_SIZE))) == 0)
+       return 0;
+
+    if ((dev->dev_neterr_msgcq = elan4_alloccq (&dev->dev_ctxt, NETERR_CQ_SIZE, CQ_STENEnableBit | CQ_WriteEnableBit, CQ_Priority)) == NULL)
+       return 0;
+
+    if ((dev->dev_neterr_intcq = elan4_alloccq (&dev->dev_ctxt, CQ_Size1K, CQ_WaitEventEnableBit | CQ_InterruptEnableBit, CQ_Priority)) == NULL)
+       return 0;
+
+    intqaddr = (dev->dev_cqoffset + elan4_cq2num (dev->dev_neterr_intcq)) * CQ_CommandMappingSize;
+    qfirst   = DEVICE_NETERR_SLOTS_ADDR;
+    qlast    = qfirst + (NETERR_NSLOTS-1) * ELAN4_NETERR_MSG_SIZE;
+
+    spin_lock_init (&dev->dev_neterr_lock);
+
+    /* Register an interrupt operation */
+    dev->dev_neterr_intop.op_function = elan4_neterr_interrupt;
+    dev->dev_neterr_intop.op_arg      = NULL;
+
+    elan4_register_intop (dev, &dev->dev_neterr_intop);
+
+    /* Initialise the inputq descriptor and event */
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, inputq.q_fptr), qfirst);
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, inputq.q_bptr), qfirst);
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, inputq.q_control), E4_InputQueueControl (qfirst, qlast, ELAN4_NETERR_MSG_SIZE));
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, inputq.q_event), DEVICE_NETERR_INPUTQ_ADDR + offsetof (NETERR_INPUTQ, qevent));
+    
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, qevent.ev_CountAndType), E4_EVENT_INIT_VALUE (-32, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, qevent.ev_WritePtr), DEVICE_NETERR_INTCQ_ADDR);
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, qevent.ev_WriteValue), (dev->dev_neterr_intop.op_cookie << E4_MAIN_INT_SHIFT) | INTERRUPT_CMD);
+
+    elan4_sdram_writeq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, sent), 0);
+
+    /* Map them all into the device context */
+    elan4mmu_pteload (&dev->dev_ctxt, 0, DEVICE_NETERR_INPUTQ_ADDR, (dev->dev_neterr_inputq >> PTE_PADDR_SHIFT) | PTE_SetPerm(PERM_RemoteAll));
+    elan4mmu_pteload (&dev->dev_ctxt, 0, DEVICE_NETERR_INTCQ_ADDR,  (intqaddr >> PTE_PADDR_SHIFT) | PTE_SetPerm(PERM_LocDataWrite) | PTE_CommandQueue);
+    elan4mmu_pteload (&dev->dev_ctxt, 0, DEVICE_NETERR_SLOTS_ADDR, (dev->dev_neterr_slots >> PTE_PADDR_SHIFT) | PTE_SetPerm(PERM_DataReadWrite));
+
+    /* finally attach to the neterr context */
+    if (elan4_attach_filter (&dev->dev_ctxt, ELAN4_NETERR_CONTEXT_NUM) != 0)
+       panic ("elan4_neterr_init: failed to attach to neterr context\n");
+
+    /* and drop the context filter */
+    elan4_set_filter (&dev->dev_ctxt, ELAN4_NETERR_CONTEXT_NUM, E4_FILTER_HIGH_PRI);
+
+    return 1;
+}
+
+void
+elan4_neterr_destroy (ELAN4_DEV *dev)
+{
+    if (dev->dev_neterr_intcq)
+    {
+       elan4_detach_filter (&dev->dev_ctxt, ELAN4_NETERR_CONTEXT_NUM);
+       
+       elan4mmu_unload_range (&dev->dev_ctxt, 0, DEVICE_NETERR_SLOTS_ADDR,  1 << dev->dev_pageshift[0]);
+       elan4mmu_unload_range (&dev->dev_ctxt, 0, DEVICE_NETERR_INTCQ_ADDR,  1 << dev->dev_pageshift[0]);
+       elan4mmu_unload_range (&dev->dev_ctxt, 0, DEVICE_NETERR_INPUTQ_ADDR, 1 << dev->dev_pageshift[0]);
+
+       spin_lock_destroy (&dev->dev_neterr_lock);
+    }
+
+    if (dev->dev_neterr_intcq)
+       elan4_freecq (&dev->dev_ctxt, dev->dev_neterr_intcq);
+    dev->dev_neterr_intcq = NULL;
+
+    if (dev->dev_neterr_msgcq)
+       elan4_freecq (&dev->dev_ctxt, dev->dev_neterr_msgcq);
+    dev->dev_neterr_msgcq = NULL;
+
+    if (dev->dev_neterr_slots)
+       elan4_sdram_free (dev, dev->dev_neterr_slots, roundup (NETERR_NSLOTS * ELAN4_NETERR_MSG_SIZE, SDRAM_PAGE_SIZE));
+    dev->dev_neterr_slots = 0;
+    
+    if (dev->dev_neterr_inputq)
+       elan4_sdram_free (dev, dev->dev_neterr_inputq, SDRAM_PAGE_SIZE);
+    dev->dev_neterr_inputq = 0;
+}
+
+int
+elan4_neterr_sendmsg (ELAN4_DEV *dev, unsigned int nodeid, unsigned int retries, ELAN4_NETERR_MSG *msg)
+{
+    ELAN4_CQ  *cq = dev->dev_neterr_msgcq;
+    E4_uint64  sent;
+    E4_VirtualProcessEntry route;
+    unsigned int vp;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_neterr_lock, flags);
+
+    sent = elan4_sdram_readq (dev, dev->dev_neterr_inputq + offsetof (NETERR_INPUTQ, sent));
+
+    PRINTF (DBG_DEVICE, DBG_NETERR, "elan4_neterr_sendmsg: nodeid=%d retries=%d cookie=%llx sender=%d,%d%s\n", 
+           nodeid, retries, msg->msg_cookies[0], msg->msg_sender.loc_node, msg->msg_sender.loc_context,
+           (dev->dev_neterr_queued - sent) >= NETERR_CQ_MSGS ? " - no cq space" : "");
+
+    if ((dev->dev_neterr_queued - sent) >= NETERR_CQ_MSGS)
+    {
+       spin_unlock_irqrestore (&dev->dev_neterr_lock, flags);
+       return 0;
+    }
+
+    vp = NETERR_VP_BASE + (dev->dev_neterr_queued % NETERR_VP_COUNT);
+
+    if (elan4_generate_route (&dev->dev_position, &route, ELAN4_NETERR_CONTEXT_NUM, nodeid, nodeid, FIRST_SYSTEM_PACKET | FIRST_HIGH_PRI) < 0)
+    {
+       spin_unlock_irqrestore (&dev->dev_neterr_lock, flags);
+       return 0;
+    }
+
+    elan4_write_route (dev, dev->dev_routetable, vp, &route);
+
+    writeq ((GUARD_CMD | GUARD_CHANNEL(0) | GUARD_RESET(retries)),                             cq->cq_mapping);
+    writeq (NOP_CMD,                                                                           cq->cq_mapping);
+    
+    writeq (OPEN_STEN_PKT_CMD | OPEN_PACKET (0, PACK_OK | RESTART_COUNT_ZERO, vp),             cq->cq_mapping);
+    writeq (SEND_TRANS_CMD | (TR_INPUT_Q_GETINDEX << 16),                                      cq->cq_mapping);
+    writeq (  DEVICE_NETERR_INPUTQ_ADDR + offsetof (NETERR_INPUTQ, inputq),                    cq->cq_mapping);
+
+    writeq (SEND_TRANS_CMD | (TR_WRITE (64 >> 3, 0, TR_DATATYPE_DWORD) << 16),                 cq->cq_mapping);
+    writeq (  0 /* address */,                                                                 cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[0],                                                          cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[1],                                                          cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[2],                                                          cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[3],                                                          cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[4],                                                          cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[5],                                                          cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[6],                                                          cq->cq_mapping);
+    writeq (  ((E4_uint64 *) msg)[7],                                                          cq->cq_mapping);
+
+    writeq (SEND_TRANS_CMD | (TR_INPUT_Q_COMMIT << 16),                                                cq->cq_mapping);
+    writeq (  DEVICE_NETERR_INPUTQ_ADDR + offsetof (NETERR_INPUTQ, inputq),                    cq->cq_mapping);
+    writeq (  0 /* cookie */,                                                                  cq->cq_mapping);
+    
+    writeq (GUARD_CMD | GUARD_CHANNEL(0) | GUARD_RESET(NETERR_RETRIES),                                cq->cq_mapping);
+    writeq (WRITE_DWORD_CMD | (DEVICE_NETERR_INPUTQ_ADDR + offsetof (NETERR_INPUTQ, sent)),    cq->cq_mapping);
+    writeq (  ++dev->dev_neterr_queued,                                                                cq->cq_mapping);
+
+    pioflush_reg (dev);
+
+    spin_unlock_irqrestore (&dev->dev_neterr_lock, flags);
+
+    return 1;
+}
+
+int
+elan4_neterr_iproc_trap (ELAN4_DEV *dev, ELAN4_IPROC_TRAP *trap)
+{
+    E4_IprocTrapHeader *hdrp = &trap->tr_transactions[trap->tr_trappedTrans];
+    unsigned long flags;
+
+    switch (IPROC_TrapValue (hdrp->IProcStatusCntxAndTrType))
+    {
+    case InputEopErrorOnWaitForEop:
+    case InputEopErrorTrap:
+    case InputCrcErrorAfterPAckOk:
+       return 1;
+
+    case InputEventEngineTrapped:
+       printk ("elan%d: device_iproc_trap: InputEventEngineTrapped - Trans=%x TrAddr=%llx\n", 
+               dev->dev_instance, (int)IPROC_TransactionType (hdrp->IProcStatusCntxAndTrType), (long long) hdrp->TrAddr);
+
+       if ((IPROC_TransactionType (hdrp->IProcStatusCntxAndTrType) & TR_OPCODE_MASK) == (TR_INPUT_Q_COMMIT & TR_OPCODE_MASK) &&
+           hdrp->TrAddr == DEVICE_NETERR_INPUTQ_ADDR + offsetof (NETERR_INPUTQ, inputq))
+       {
+           spin_lock_irqsave (&dev->dev_neterr_lock, flags);
+           writeq ((DEVICE_NETERR_INPUTQ_ADDR + offsetof (NETERR_INPUTQ, qevent)) | SET_EVENT_CMD, dev->dev_neterr_msgcq->cq_mapping);
+           spin_unlock_irqrestore (&dev->dev_neterr_lock, flags);
+           return 1;
+       }
+       
+    default:
+       return 0;
+    }
+}
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/procfs_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/procfs_Linux.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/procfs_Linux.c 2005-06-01 23:12:54.613436976 -0400
@@ -0,0 +1,1041 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: procfs_Linux.c,v 1.27.2.5 2005/01/18 14:36:17 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/elan4mod/procfs_Linux.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <linux/module.h>
+#include <linux/proc_fs.h>
+#include <linux/ctype.h>
+
+#include <qsnet/procfs_linux.h>
+
+#include <elan4/i2c.h>
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/user.h>
+
+/*
+ *
+ * procfs format for elan4:
+ *
+ * /proc/qsnet/elan4/config
+ *    elan4_debug
+ *    elan4_debug_toconsole
+ *    elan4_debug_tobuffer
+ *    elan4_debug_display_ctxt
+ *    elan4_debug_ignore_ctxt
+ *    elan4_debug_ignore_type
+ *    elan4_debug_mmu
+ *    elan4_mainint_punt_loops
+ *    user_p2p_route_options
+ *    user_bcast_route_options
+ *
+ * /proc/qsnet/elan4/deviceN
+ *    stats
+ *    position
+ *    vpd
+ */
+
+struct proc_dir_entry *elan4_procfs_root;
+struct proc_dir_entry *elan4_config_root;
+
+/* borrowed from fs/proc/proc_misc - helper for proc_read_int */
+static int 
+proc_calc_metrics(char *page, char **start, off_t off, int count, int *eof, int len)
+{
+    if (len <= off+count) *eof = 1;
+    *start = page + off;
+    len -= off;
+    if (len>count) len = count;
+    if (len<0) len = 0;
+    return len;
+}
+
+static int
+proc_read_devinfo (char *page, char **start, off_t off,
+                   int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    int        len = 0;
+
+    if (! dev)
+       len = sprintf (page, "<unknown>\n");
+    else
+    {
+       len += sprintf (page + len, "dev_vendor_id            0x%x\n", dev->dev_devinfo.dev_vendor_id);
+       len += sprintf (page + len, "dev_device_id            0x%x\n", dev->dev_devinfo.dev_vendor_id);
+       len += sprintf (page + len, "dev_revision_id          0x%x\n", dev->dev_devinfo.dev_revision_id);
+       len += sprintf (page + len, "dev_instance             0x%x\n", dev->dev_devinfo.dev_instance);
+       len += sprintf (page + len, "dev_rail                 0x%x\n", dev->dev_devinfo.dev_rail);
+       len += sprintf (page + len, "dev_driver_version       0x%x\n", dev->dev_devinfo.dev_driver_version);
+       len += sprintf (page + len, "dev_params_mask          0x%x\n", dev->dev_devinfo.dev_params_mask);
+       len += sprintf (page + len, "dev_params:                  \n");
+       len += sprintf (page + len, " 0  - PciCmdQPadFlag     0x%x\n", dev->dev_devinfo.dev_params.values[0]);
+       len += sprintf (page + len, " 1  - EventCopyWinPt     0x%x\n", dev->dev_devinfo.dev_params.values[1]);
+       len += sprintf (page + len, " 2  - PciWriteCombining  0x%x\n", dev->dev_devinfo.dev_params.values[2]);
+       len += sprintf (page + len, " 3  -                    0x%x\n", dev->dev_devinfo.dev_params.values[3]);
+       len += sprintf (page + len, " 4  -                    0x%x\n", dev->dev_devinfo.dev_params.values[4]);
+       len += sprintf (page + len, " 5  -                    0x%x\n", dev->dev_devinfo.dev_params.values[5]);
+       len += sprintf (page + len, " 6  -                    0x%x\n", dev->dev_devinfo.dev_params.values[6]);
+       len += sprintf (page + len, " 7  -                    0x%x\n", dev->dev_devinfo.dev_params.values[7]);
+       len += sprintf (page + len, " 8  -                    0x%x\n", dev->dev_devinfo.dev_params.values[8]);
+       len += sprintf (page + len, " 9  -                    0x%x\n", dev->dev_devinfo.dev_params.values[9]);
+       len += sprintf (page + len, " 10 -                    0x%x\n", dev->dev_devinfo.dev_params.values[10]);
+       len += sprintf (page + len, " 11 -                    0x%x\n", dev->dev_devinfo.dev_params.values[11]);
+       len += sprintf (page + len, "dev_num_down_links_value 0x%x\n", dev->dev_devinfo.dev_num_down_links_value);
+
+       len += sprintf (page + len, "features                 0x%x\n", dev->dev_features);
+    }
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+static int
+proc_read_position (char *page, char **start, off_t off,
+                   int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    int        len;
+
+    if (dev->dev_position.pos_mode == ELAN_POS_UNKNOWN)
+       len = sprintf (page, "<unknown>\n");
+    else
+       len = sprintf (page, 
+                      "NodeId                 %d\n"
+                      "NumLevels              %d\n"
+                      "NumNodes               %d\n",
+                      dev->dev_position.pos_nodeid, 
+                      dev->dev_position.pos_levels, 
+                      dev->dev_position.pos_nodes);
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+static int
+proc_write_position (struct file *file, const char *buf, unsigned long count, void *data)
+{
+    ELAN4_DEV *dev     = (ELAN4_DEV *) data;
+    unsigned  nodeid   = ELAN_INVALID_NODE;
+    unsigned  numnodes = 0;
+    char     *page, *p;
+    int       res;
+    ELAN_POSITION pos;
+
+    if (count == 0)
+       return (0);
+
+    if (count >= PAGE_SIZE)
+       return (-EINVAL);
+
+    if ((page = (char *) __get_free_page (GFP_KERNEL)) == NULL)
+       return (-ENOMEM);
+
+    MOD_INC_USE_COUNT;
+
+    if (copy_from_user (page, buf, count))
+       res = -EFAULT;
+    else
+    {
+       page[count] = '\0';
+       
+       if (page[count-1] == '\n')
+           page[count-1] = '\0';
+
+       if (! strcmp (page, "<unknown>"))
+       {
+           pos.pos_mode      = ELAN_POS_UNKNOWN;
+           pos.pos_nodeid    = ELAN_INVALID_NODE;
+           pos.pos_nodes     = 0;
+           pos.pos_levels    = 0;
+       }
+       else
+       {
+           for (p = page; *p; )
+           {
+               while (isspace (*p))
+                   p++;
+               
+               if (! strncmp (p, "NodeId=", strlen("NodeId=")))
+                   nodeid   = simple_strtoul (p + strlen ("NodeId="), NULL, 0);
+               if (! strncmp (p, "NumNodes=", strlen ("NumNodes=")))
+                   numnodes = simple_strtoul (p + strlen ("NumNodes="), NULL, 0);
+               
+               while (*p && !isspace(*p))
+                   p++;
+           }
+
+           if (elan4_compute_position (&pos, nodeid, numnodes, dev->dev_devinfo.dev_num_down_links_value) != 0)
+               printk ("elan%d: invalid values for NodeId=%d NumNodes=%d\n", dev->dev_instance, nodeid, numnodes);
+           else
+           {
+               printk ("elan%d: setting NodeId=%d NumNodes=%d NumLevels=%d\n", dev->dev_instance, pos.pos_nodeid,
+                       pos.pos_nodes, pos.pos_levels);
+
+               if (elan4_set_position (dev, &pos) < 0)
+                   printk ("elan%d: failed to set device position\n", dev->dev_instance);
+           }
+       }
+    }
+
+    MOD_DEC_USE_COUNT;
+    free_page ((unsigned long) page);
+
+    return (count);
+}
+
+static int
+proc_read_temp (char *page, char **start, off_t off,
+               int count, int *eof, void *data)
+{
+    ELAN4_DEV    *dev = (ELAN4_DEV *) data;
+    unsigned char values[2];
+    int           len;
+
+    if (i2c_disable_auto_led_update (dev) < 0)
+       len = sprintf (page, "<unknown>");
+    else
+    {
+       if (i2c_read (dev, I2C_TEMP_ADDR, 2, values) < 0)
+           len = sprintf (page, "<not-present>");
+       else
+           len = sprintf (page, "%s%d%s\n", (values[0] & 0x80) ? "-" : "",
+                          (values[0] & 0x80) ? -((signed char)values[0]) - 1 : values[0],
+                          (values[1] & 0x80) ? ".5" : ".0");
+
+       i2c_enable_auto_led_update (dev);
+    }
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+static int
+proc_read_eccerr (char *page, char **start, off_t off,
+                 int count, int *eof, void *data)
+{
+    ELAN4_DEV    *dev = (ELAN4_DEV *) data;
+    char          errstr[200];
+    register int  i, len = 0;
+
+    *page = '\0';
+
+    for (i = 0; i < sizeof (dev->dev_sdramerrs)/sizeof(dev->dev_sdramerrs[0]); i++)
+       if (dev->dev_sdramerrs[i])
+           len += sprintf (page + len, "%s occured %0d times\n",
+                           elan4_sdramerr2str (dev, dev->dev_sdramerrs[i] & 0x000fffffffffffffULL, errstr),
+                           (int) (dev->dev_sdramerrs[i] >> 52) + 1);
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+static int
+proc_read_vpd (char *page, char **start, off_t off,
+              int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    int        len;
+
+    if ( elan4_read_vpd (dev, NULL, page) )
+       len = sprintf (page, "no vpd tags found\n");
+    else
+       len = strlen(page)+1;
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+static struct device_info 
+{
+    char *name;
+    int (*read_func) (char *page, char **start, off_t off, int count, int *eof, void *data);
+    int (*write_func) (struct file *file, const char *buf, unsigned long count, void *data);
+    unsigned minrev;
+} device_info[] = {
+    {"devinfo",   proc_read_devinfo,  NULL,                0},
+    {"position",  proc_read_position, proc_write_position, 0},
+    {"temp",      proc_read_temp,     NULL,                1},
+    {"eccerr",    proc_read_eccerr,   NULL,                0},
+    {"vpd",       proc_read_vpd,      NULL,                0},
+};
+
+static int
+proc_read_link_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+
+    p += sprintf (p, "%20s %ld\n", "link_errors", dev->dev_stats.s_link_errors);
+    p += sprintf (p, "%20s %ld\n", "lock_errors", dev->dev_stats.s_lock_errors);
+    p += sprintf (p, "%20s %ld\n", "deskew_errors", dev->dev_stats.s_deskew_errors);
+    p += sprintf (p, "%20s %ld\n", "phase_errors", dev->dev_stats.s_phase_errors);
+
+    p += sprintf (p, "%20s %ld\n", "data_errors", dev->dev_stats.s_data_errors);
+    p += sprintf (p, "%20s %ld\n", "fifo_overflow0", dev->dev_stats.s_fifo_overflow0);
+    p += sprintf (p, "%20s %ld\n", "fifo_overflow1", dev->dev_stats.s_fifo_overflow1);
+    p += sprintf (p, "%20s %ld\n", "mod45changed", dev->dev_stats.s_mod45changed);
+    p += sprintf (p, "%20s %ld\n", "pack_not_seen", dev->dev_stats.s_pack_not_seen);
+
+    p += sprintf (p, "%20s %ld\n", "linkport_keyfail", dev->dev_stats.s_linkport_keyfail);
+    p += sprintf (p, "%20s %ld\n", "eop_reset", dev->dev_stats.s_eop_reset);
+    p += sprintf (p, "%20s %ld\n", "bad_length", dev->dev_stats.s_bad_length);
+    p += sprintf (p, "%20s %ld\n", "crc_error", dev->dev_stats.s_crc_error);
+    p += sprintf (p, "%20s %ld\n", "crc_bad", dev->dev_stats.s_crc_bad);
+
+    p += sprintf (p, "%20s %ld\n", "cproc_timeout", dev->dev_stats.s_cproc_timeout);
+    p += sprintf (p, "%20s %ld\n", "dproc_timeout", dev->dev_stats.s_dproc_timeout);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static char *
+proc_sprintf_bucket_stat (char *p, char *name, unsigned long *stats, int *buckets)
+{
+    int i;
+
+    p += sprintf (p, "%20s ", name);
+
+    for (i = 0; i < ELAN4_DEV_STATS_BUCKETS-1; i++)
+       p += sprintf (p, "%ld(<=%d) ", stats[i], buckets[i]);
+    p += sprintf (p, "%ld(>%d)\n", stats[i], buckets[i-1]);
+
+    return p;
+}
+
+static int
+proc_read_intr_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+
+    p += sprintf (p, "%20s %ld\n", "interrupts",       dev->dev_stats.s_interrupts);
+    p += sprintf (p, "%20s %ld\n", "haltints",         dev->dev_stats.s_haltints);
+
+    p += sprintf (p, "%20s %ld\n", "mainint_punts",    dev->dev_stats.s_mainint_punts);
+    p += sprintf (p, "%20s %ld\n", "mainint_rescheds", dev->dev_stats.s_mainint_rescheds);
+
+    p  = proc_sprintf_bucket_stat (p, "mainints", dev->dev_stats.s_mainints, MainIntBuckets);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_trap_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+
+    p += sprintf (p, "%20s %ld\n", "cproc_traps", dev->dev_stats.s_cproc_traps);
+    p += sprintf (p, "%20s %ld\n", "dproc_traps", dev->dev_stats.s_dproc_traps);
+    p += sprintf (p, "%20s %ld\n", "eproc_traps", dev->dev_stats.s_eproc_traps);
+    p += sprintf (p, "%20s %ld\n", "iproc_traps", dev->dev_stats.s_iproc_traps);
+    p += sprintf (p, "%20s %ld\n", "tproc_traps", dev->dev_stats.s_tproc_traps);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_cproc_trap_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+    int        i;
+    extern char *const CProcTrapNames[];
+
+    for (i = 0; i < sizeof (dev->dev_stats.s_cproc_trap_types)/sizeof(dev->dev_stats.s_cproc_trap_types[0]); i++)
+       p += sprintf (p, "%-40s %ld\n", CProcTrapNames[i], dev->dev_stats.s_cproc_trap_types[i]);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_dproc_trap_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+    int        i;
+    extern char *const DProcTrapNames[];
+
+    for (i = 0; i < sizeof (dev->dev_stats.s_dproc_trap_types)/sizeof(dev->dev_stats.s_dproc_trap_types[0]); i++)
+       p += sprintf (p, "%-40s %ld\n", DProcTrapNames[i], dev->dev_stats.s_dproc_trap_types[i]);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_eproc_trap_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+    int        i;
+    extern char *const EProcTrapNames[];
+
+    for (i = 0; i < sizeof (dev->dev_stats.s_eproc_trap_types)/sizeof(dev->dev_stats.s_eproc_trap_types[0]); i++)
+       p += sprintf (p, "%-40s %ld\n", EProcTrapNames[i], dev->dev_stats.s_eproc_trap_types[i]);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_iproc_trap_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+    int        i;
+    extern char *const IProcTrapNames[];
+
+    for (i = 0; i < sizeof (dev->dev_stats.s_iproc_trap_types)/sizeof(dev->dev_stats.s_iproc_trap_types[0]); i++)
+       p += sprintf (p, "%-40s %ld\n", IProcTrapNames[i], dev->dev_stats.s_iproc_trap_types[i]);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_tproc_trap_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+    int        i;
+    extern char *const TProcTrapNames[];
+
+    for (i = 0; i < sizeof (dev->dev_stats.s_tproc_trap_types)/sizeof(dev->dev_stats.s_tproc_trap_types[0]); i++)
+       p += sprintf (p, "%-40s %ld\n", TProcTrapNames[i], dev->dev_stats.s_tproc_trap_types[i]);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_sdram_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev = (ELAN4_DEV *) data;
+    char      *p   = page;
+
+    p += sprintf (p, "%20s %ld\n",  "correctable_errors", dev->dev_stats.s_correctable_errors);
+    p += sprintf (p, "%20s %ld\n",  "multiple_errors",    dev->dev_stats.s_multiple_errors);
+    p += sprintf (p, "%20s %ldK\n", "sdram_bytes_free",   dev->dev_stats.s_sdram_bytes_free/1024);
+
+    return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+void
+elan4_ringbuf_store(ELAN4_ROUTE_RINGBUF *ringbuf, E4_VirtualProcessEntry *route, ELAN4_DEV *dev)
+{
+    int newend;
+    unsigned long flags;
+
+    spin_lock_irqsave(&dev->dev_error_routes_lock, flags);
+    bcopy(route, &ringbuf->routes[ringbuf->end], sizeof(E4_VirtualProcessEntry));
+    newend = ringbuf->end + 1;
+    if (newend >= DEV_STASH_ROUTE_COUNT)
+        newend -= DEV_STASH_ROUTE_COUNT;
+    if (newend == ringbuf->start)
+        ringbuf->start += 1;
+    if (ringbuf->start >= DEV_STASH_ROUTE_COUNT)
+        ringbuf->start -= DEV_STASH_ROUTE_COUNT;
+    ringbuf->end = newend;
+    spin_unlock_irqrestore(&dev->dev_error_routes_lock, flags);
+}
+       
+static int
+proc_read_dproc_timeout_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       ELAN4_DEV *dev = (ELAN4_DEV *) data;
+       char      *p   = page;
+       unsigned int *dproc_timeout;
+
+       dproc_timeout = dev->dev_dproc_timeout;
+
+       if (!dproc_timeout) 
+               p += sprintf (p, "No stats available\n");
+       else
+       {
+               int i;
+
+               for (i=0; i<dev->dev_position.pos_nodes; i++) 
+                       if (dproc_timeout[i] != 0) 
+                               p += sprintf (p, "Node %d: %u errors\n", i, dproc_timeout[i]);
+       }
+
+       return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+elan4_route2str (E4_VirtualProcessEntry *route, char *routeStr)
+{
+    int        part = 0;
+    int        shift;
+    int        broadcast;
+    E4_uint64  value;
+    char      *ptr = routeStr;
+    int        b;
+
+    /* unpack first */
+    value = route->Values[part] & 0x7f;
+    if ( (value & 0x78) == 0) {
+        /* empty route */
+        strcpy(routeStr,"Invalid lead route");
+        return (-EINVAL);
+    }
+
+    if ( value & 0x40 ) {
+        /* broad cast */
+       strcpy(routeStr,"Broadcast");
+       return (-EINVAL);
+    } else {
+        switch ((value  & 0x30) >> 4) {
+        case 0: { *ptr++ = '0' + (value & 0x7); break; }
+        case 1: { *ptr++ = 'M';                 break; }
+        case 2: { *ptr++ = 'U';                 break; }
+        case 3: { *ptr++ = 'A';                 break; }
+        }
+    }
+
+    shift = 16;
+    broadcast = 0;
+    while ( 1 ) {
+        b =  (route->Values[part] >> shift) & 0xf;
+
+        if ( broadcast ) {
+            /* about to pick up the second byte of a broadcast pair */
+            broadcast = 0;
+        } else {
+            if ( b & 0x8) {
+                /*  output link */
+                 *ptr++ = '0' + (b & 0x7);
+            } else {
+                if ( b & 0x4) {
+                    /* broad cast */
+                    broadcast = 1;
+                } else {
+                    switch ( b & 0x3 ) {
+                    case 0: { *ptr++ =  0 ; return (0);     break; }
+                    case 1: { *ptr++ = 'M';                 break; }
+                    case 2: { *ptr++ = 'U';                 break; }
+                    case 3: { *ptr++ = 'A';                 break; }
+                    }
+                }
+            }
+        }
+
+        shift += 4; 
+        if ( part != 0 ) {
+            if ( shift > 36) {
+                /* too far, now in the crc value */
+                strcpy(routeStr,"Invalid route length");
+                return (-EINVAL);
+            }
+        } else { 
+            if ( shift >= 64) { 
+                /* move to the next 64 bits */
+                part = 1;
+                shift = 2;
+            }
+        }
+    }
+
+    /* never reached */
+    return (-EINVAL);
+}
+
+
+static int
+proc_read_dproc_timeout_routes (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       ELAN4_DEV *dev = (ELAN4_DEV *) data;
+       char      *p   = page;
+       ELAN4_ROUTE_RINGBUF *ringbuf;
+       char      routestr[33];
+
+       ringbuf = &dev->dev_dproc_timeout_routes;
+
+       if (!ringbuf) 
+               p += sprintf (p, "No stats available\n");
+       else
+       {
+               int start;
+               int end;
+               int i;
+               unsigned long flags;
+
+               memset(&routestr, 0, 33);
+
+               spin_lock_irqsave(&dev->dev_error_routes_lock, flags);
+               
+               start = ringbuf->start;
+               end = ringbuf->end;
+
+               if (end < start)
+                       end = DEV_STASH_ROUTE_COUNT;
+
+               for (i=start; i<end; i++) 
+               {
+                       elan4_route2str (&ringbuf->routes[i], routestr);
+                       p += sprintf (p, "Route %llx %llx->%s\n", ringbuf->routes[i].Values[0], ringbuf->routes[i].Values[1], routestr);
+               }
+
+               if (ringbuf->end < start)
+               {
+                       start = 0;
+                       end = ringbuf->end;
+                       for (i=start; i<end; i++)
+                       {
+                               elan4_route2str (&ringbuf->routes[i], routestr);
+                               p += sprintf (p, "Route %llx %llx->%s\n", ringbuf->routes[i].Values[0], ringbuf->routes[i].Values[1], routestr);
+                       }
+               }
+
+               spin_unlock_irqrestore(&dev->dev_error_routes_lock, flags);
+       }
+
+       return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+
+static int
+proc_read_cproc_timeout_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       ELAN4_DEV *dev = (ELAN4_DEV *) data;
+       char      *p   = page;
+       unsigned int *cproc_timeout;
+
+       cproc_timeout = dev->dev_cproc_timeout;
+
+       if (!cproc_timeout) 
+               p += sprintf (p, "No stats available\n");
+       else
+       {
+               int i;
+
+               for (i=0; i<dev->dev_position.pos_nodes; i++) 
+                       if (cproc_timeout[i] != 0) 
+                               p += sprintf (p, "Node %d: %u errors\n", i, cproc_timeout[i]);
+       }
+
+       return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_cproc_timeout_routes (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       ELAN4_DEV *dev = (ELAN4_DEV *) data;
+       char      *p   = page;
+       ELAN4_ROUTE_RINGBUF *ringbuf;
+       char      routestr[33];
+
+       ringbuf = &dev->dev_cproc_timeout_routes;
+
+       if (!ringbuf) 
+               p += sprintf (p, "No stats available\n");
+       else
+       {
+               int start;
+               int end;
+               int i;
+               unsigned long flags;
+
+               memset(&routestr, 0, 33);
+
+               spin_lock_irqsave(&dev->dev_error_routes_lock, flags);
+               
+               start = ringbuf->start;
+               end = ringbuf->end;
+
+               if (end < start)
+                       end = DEV_STASH_ROUTE_COUNT;
+
+               for (i=start; i<end; i++) 
+               {
+                       elan4_route2str (&ringbuf->routes[i], routestr);
+                       p += sprintf (p, "Route %llx %llx->%s\n", ringbuf->routes[i].Values[0], ringbuf->routes[i].Values[1], routestr);
+               }
+
+               if (ringbuf->end < start)
+               {
+                       start = 0;
+                       end = ringbuf->end;
+                       for (i=start; i<end; i++)
+                       {
+                               elan4_route2str (&ringbuf->routes[i], routestr);
+                               p += sprintf (p, "Route %llx %llx->%s\n", ringbuf->routes[i].Values[0], ringbuf->routes[i].Values[1], routestr);
+                       }
+               }
+
+               spin_unlock_irqrestore(&dev->dev_error_routes_lock, flags);
+       }
+
+       return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_traperr_stats (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       ELAN4_DEV *dev = (ELAN4_DEV *) data;
+       char      *p   = page;
+       unsigned int *ack_errors;
+
+       ack_errors = dev->dev_ack_errors;
+
+       if (!ack_errors) 
+               p += sprintf (p, "No stats available\n");
+       else
+       {
+               int i;
+
+               for (i=0; i<dev->dev_position.pos_nodes; i++) 
+                       if (ack_errors[i] != 0) 
+                               p += sprintf (p, "Node %d: %u errors\n", i, ack_errors[i]);
+       }
+
+       return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static int
+proc_read_ackerror_routes (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       ELAN4_DEV *dev = (ELAN4_DEV *) data;
+       char      *p   = page;
+       ELAN4_ROUTE_RINGBUF *ringbuf;
+       char      routestr[33];
+
+       ringbuf = &dev->dev_ack_error_routes;
+
+       if (!ringbuf) 
+               p += sprintf (p, "No stats available\n");
+       else
+       {
+               int start;
+               int end;
+               int i;
+               unsigned long flags;
+
+               memset(&routestr, 0, 33);
+
+               spin_lock_irqsave(&dev->dev_error_routes_lock, flags);
+               
+               start = ringbuf->start;
+               end = ringbuf->end;
+
+               if (end < start)
+                       end = DEV_STASH_ROUTE_COUNT;
+
+               for (i=start; i<end; i++) 
+               {
+                       elan4_route2str (&ringbuf->routes[i], routestr);
+                       p += sprintf (p, "Route %llx %llx->%s\n", ringbuf->routes[i].Values[0], ringbuf->routes[i].Values[1], routestr);
+               }
+
+               if (ringbuf->end < start)
+               {
+                       start = 0;
+                       end = ringbuf->end;
+                       for (i=start; i<end; i++)
+                       {
+                               elan4_route2str (&ringbuf->routes[i], routestr);
+                               p += sprintf (p, "Route %llx %llx->%s\n", ringbuf->routes[i].Values[0], ringbuf->routes[i].Values[1], routestr);
+                       }
+               }
+
+               spin_unlock_irqrestore(&dev->dev_error_routes_lock, flags);
+       }
+
+       return (proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static struct stats_info 
+{
+    char *name;
+    int (*read_func) (char *page, char **start, off_t off, int count, int *eof, void *data);
+    int (*write_func) (struct file *file, const char *buf, unsigned long count, void *data); 
+} stats_info[] = {
+    {"link",     proc_read_link_stats, NULL},
+    {"intr",     proc_read_intr_stats, NULL},
+    {"trap",     proc_read_trap_stats, NULL},
+    {"cproc",    proc_read_cproc_trap_stats, NULL},
+    {"dproc",    proc_read_dproc_trap_stats, NULL},
+    {"eproc",    proc_read_eproc_trap_stats, NULL},
+    {"iproc",    proc_read_iproc_trap_stats, NULL},
+    {"tproc",    proc_read_tproc_trap_stats, NULL},
+    {"sdram",    proc_read_sdram_stats, NULL},
+    {"trapdmaerr", proc_read_traperr_stats, NULL},
+    {"dproctimeout", proc_read_dproc_timeout_stats, NULL},
+    {"cproctimeout", proc_read_cproc_timeout_stats, NULL},
+    {"dproctimeoutroutes", proc_read_dproc_timeout_routes, NULL},
+    {"cproctimeoutroutes", proc_read_cproc_timeout_routes, NULL},
+    {"ackerrroutes", proc_read_ackerror_routes, NULL},
+};
+
+static int
+proc_read_sysconfig (char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+    ELAN4_DEV *dev        = (ELAN4_DEV *) data;
+    E4_uint32  syscontrol = dev->dev_syscontrol;
+    int               len       = 0;
+
+   *eof = 1;
+   if (off != 0)
+      return (0);
+
+    if (syscontrol & CONT_EN_ALL_SETS)
+       len += sprintf (page + len, "%sEN_ALL_SETS", len == 0 ? "" : " ");
+    if (syscontrol & CONT_MMU_ENABLE)
+       len += sprintf (page + len, "%sMMU_ENABLE", len == 0 ? "" : " ");
+    if (syscontrol & CONT_CACHE_HASH_TABLE)
+       len += sprintf (page + len, "%sCACHE_HASH_TABLE", len == 0 ? "" : " ");
+    if (syscontrol & CONT_CACHE_CHAINS)
+       len += sprintf (page + len, "%sCACHE_CHAINS", len == 0 ? "" : " ");
+    if (syscontrol & CONT_CACHE_ROOT_CNTX)
+       len += sprintf (page + len, "%sCACHE_ROOT_CNTX", len == 0 ? "" : " ");
+    if (syscontrol & CONT_CACHE_STEN_ROUTES)
+       len += sprintf (page + len, "%sCACHE_STEN_ROUTES", len == 0 ? "" : " ");
+    if (syscontrol & CONT_CACHE_DMA_ROUTES)
+       len += sprintf (page + len, "%sCACHE_DMA_ROUTES", len == 0 ? "" : " ");
+    if (syscontrol & CONT_INHIBIT_MAX_CHAIN_ITEMS)
+       len += sprintf (page + len, "%sINHIBIT_MAX_CHAIN_ITEMS", len == 0 ? "" : " ");
+
+    len += sprintf (page + len, "%sTABLE0_MASK_SIZE=%d", len == 0 ? "" : " ", (syscontrol >> CONT_TABLE0_MASK_SIZE_SHIFT) & PAGE_MASK_MASK);
+    len += sprintf (page + len, "%sTABLE0_PAGE_SIZE=%d", len == 0 ? "" : " ", (syscontrol >> CONT_TABLE0_PAGE_SIZE_SHIFT) & PAGE_SIZE_MASK);
+    len += sprintf (page + len, "%sTABLE1_MASK_SIZE=%d", len == 0 ? "" : " ", (syscontrol >> CONT_TABLE1_MASK_SIZE_SHIFT) & PAGE_MASK_MASK);
+    len += sprintf (page + len, "%sTABLE1_PAGE_SIZE=%d", len == 0 ? "" : " ", (syscontrol >> CONT_TABLE1_PAGE_SIZE_SHIFT) & PAGE_SIZE_MASK);
+
+    if (syscontrol & CONT_2K_NOT_1K_DMA_PACKETS)
+       len += sprintf (page + len, "%s2K_NOT_1K_DMA_PACKETS", len == 0 ? "" : " ");
+    if (syscontrol & CONT_ALIGN_ALL_DMA_PACKETS)
+       len += sprintf (page + len, "%sALIGN_ALL_DMA_PACKETS", len == 0 ? "" : " ");
+    if (syscontrol & CONT_DIRECT_MAP_PCI_WRITES)
+       len += sprintf (page + len, "%sDIRECT_MAP_PCI_WRITES", len == 0 ? "" : " ");
+
+    len += sprintf (page + len, "\n");
+
+   *start = page;
+   return (len);
+}
+
+static int
+proc_write_sysconfig (struct file *file, const char *ubuffer, unsigned long count, void *data)
+{
+    ELAN4_DEV *dev       = (ELAN4_DEV *) data;
+    unsigned long page   = __get_free_page (GFP_KERNEL);
+    char         *buffer = (char *)page;
+    int            add   = 0;
+    int            sub   = 0;
+    
+    count = MIN (count, PAGE_SIZE - 1);
+    if (copy_from_user (buffer, ubuffer, count))
+    {
+       free_page (page);
+       return (-EFAULT);
+    }
+   
+    buffer[count] = 0;                         /* terminate string */
+
+    while (*buffer != 0)
+    {
+       char *ptr;
+       char *end;
+       int   ch;
+       int   val;
+       int   op;
+
+       ch = *buffer;
+       if (ch == ' ' || ch == '\t' || ch == '\r' || ch == '\n')
+       {
+           buffer++;
+           continue;
+       }
+      
+       op = *buffer;
+       if (op == '+' || op == '-')
+           buffer++;
+
+       for (end = buffer; *end != 0; end++)
+           if (*end == ' ' || *end == '\t' ||
+               *end == '\r' || *end == '\n')
+               break;
+      
+       if (end == buffer)
+           break;
+      
+       ch = *end;
+       *end = 0;
+
+       for (ptr = buffer; *ptr != 0; ptr++)
+           if ('a' <= *ptr && *ptr <= 'z')
+               *ptr = *ptr + 'A' - 'a';
+      
+       if (!strcmp (buffer, "EN_ALL_SETS"))
+           val = CONT_EN_ALL_SETS;
+       if (!strcmp (buffer, "CACHE_HASH_TABLE"))
+           val = CONT_CACHE_HASH_TABLE;
+       else if (!strcmp (buffer, "CACHE_CHAINS"))
+           val = CONT_CACHE_CHAINS;
+       else if (!strcmp (buffer, "CACHE_ROOT_CNTX"))
+           val = CONT_CACHE_ROOT_CNTX;
+       else if (!strcmp (buffer, "CACHE_STEN_ROUTES"))
+           val = CONT_CACHE_STEN_ROUTES;
+       else if (!strcmp (buffer, "CACHE_DMA_ROUTES"))
+           val = CONT_CACHE_DMA_ROUTES;
+       else if (!strcmp (buffer, "2K_NOT_1K_DMA_PACKETS"))
+           val = CONT_2K_NOT_1K_DMA_PACKETS;
+       else if (!strcmp (buffer, "ALIGN_ALL_DMA_PACKETS"))
+           val = CONT_ALIGN_ALL_DMA_PACKETS;
+       else
+           val = 0;
+
+       if (op == '+')
+           add |= val;
+       else if (op == '-')
+           sub |= val;
+
+       *end = ch;
+       buffer = end;
+    }
+
+    if ((add | sub) & CONT_EN_ALL_SETS)
+       elan4_sdram_flushcache (dev, 0, E4_CacheSize);
+
+    CHANGE_SYSCONTROL (dev, add, sub);
+   
+    if ((add | sub) & CONT_EN_ALL_SETS)
+       elan4_sdram_flushcache (dev, 0, E4_CacheSize);
+
+    free_page (page);
+    return (count);
+}
+
+static struct config_info 
+{
+    char *name;
+    int (*read_func) (char *page, char **start, off_t off, int count, int *eof, void *data);
+    int (*write_func) (struct file *file, const char *buf, unsigned long count, void *data); 
+} config_info[] = {
+    {"sysconfig",  proc_read_sysconfig, proc_write_sysconfig},
+};
+
+void
+elan4_procfs_device_init (ELAN4_DEV *dev)
+{
+    struct proc_dir_entry *p;
+    char name[NAME_MAX];
+    int i;
+
+    sprintf (name, "device%d", dev->dev_instance);
+    dev->dev_osdep.procdir  = proc_mkdir (name, elan4_procfs_root);
+
+    for (i = 0; i < sizeof (device_info)/sizeof (device_info[0]); i++)
+    {
+       if (dev->dev_devinfo.dev_revision_id < device_info[i].minrev)
+           continue;
+
+       if ((p = create_proc_entry (device_info[i].name, 0, dev->dev_osdep.procdir)) != NULL)
+       {
+           p->read_proc  = device_info[i].read_func;
+           p->write_proc = device_info[i].write_func;
+           p->data       = dev;
+           p->owner      = THIS_MODULE;
+       }
+    }
+
+    dev->dev_osdep.configdir = proc_mkdir ("config", dev->dev_osdep.procdir);
+    for (i = 0; i < sizeof (config_info)/sizeof (config_info[0]); i++)
+    {
+       if ((p = create_proc_entry (config_info[i].name, 0, dev->dev_osdep.configdir)) != NULL)
+       {
+           p->read_proc  = config_info[i].read_func;
+           p->write_proc = config_info[i].write_func;
+           p->data       = dev;
+           p->owner      = THIS_MODULE;
+       }
+    }
+
+    dev->dev_osdep.statsdir = proc_mkdir ("stats", dev->dev_osdep.procdir);
+    for (i = 0; i < sizeof (stats_info)/sizeof (stats_info[0]); i++)
+    {
+       if ((p = create_proc_entry (stats_info[i].name, 0, dev->dev_osdep.statsdir)) != NULL)
+       {
+           p->read_proc  = stats_info[i].read_func;
+           p->write_proc = stats_info[i].write_func;
+           p->data       = dev;
+           p->owner      = THIS_MODULE;
+       }
+    }
+}
+
+void
+elan4_procfs_device_fini (ELAN4_DEV *dev)
+{
+    char name[NAME_MAX];
+    int i;
+
+    for (i = 0; i < sizeof (stats_info)/sizeof (stats_info[0]); i++)
+       remove_proc_entry (stats_info[i].name, dev->dev_osdep.statsdir);
+    remove_proc_entry ("stats", dev->dev_osdep.procdir);
+
+    for (i = 0; i < sizeof (config_info)/sizeof (config_info[0]); i++)
+       remove_proc_entry (config_info[i].name, dev->dev_osdep.configdir);
+    remove_proc_entry ("config", dev->dev_osdep.procdir);
+
+    for (i = 0; i < sizeof (device_info)/sizeof (device_info[0]); i++)
+    {
+       if (dev->dev_devinfo.dev_revision_id < device_info[i].minrev)
+           continue;
+       
+       remove_proc_entry (device_info[i].name, dev->dev_osdep.procdir);
+    }
+
+    sprintf (name, "device%d", dev->dev_instance);
+    remove_proc_entry (name, elan4_procfs_root);
+}
+
+void
+elan4_procfs_init(void)
+{
+    elan4_procfs_root = proc_mkdir("elan4", qsnet_procfs_root);
+    elan4_config_root = proc_mkdir("config", elan4_procfs_root);
+
+    qsnet_proc_register_hex (elan4_config_root, "elan4_debug",              &elan4_debug,              0);
+    qsnet_proc_register_hex (elan4_config_root, "elan4_debug_toconsole",    &elan4_debug_toconsole,    0);
+    qsnet_proc_register_hex (elan4_config_root, "elan4_debug_tobuffer",     &elan4_debug_tobuffer,     0);
+    qsnet_proc_register_int (elan4_config_root, "elan4_debug_mmu",          &elan4_debug_mmu,          0);
+    qsnet_proc_register_int (elan4_config_root, "elan4_mainint_punt_loops", &elan4_mainint_punt_loops, 0);
+    qsnet_proc_register_hex (elan4_config_root, "user_p2p_route_options",   &user_p2p_route_options,   0);
+    qsnet_proc_register_hex (elan4_config_root, "user_bcast_route_options", &user_bcast_route_options, 0);
+    qsnet_proc_register_int (elan4_config_root, "user_dproc_retry_count",   &user_dproc_retry_count,    0);
+    qsnet_proc_register_int (elan4_config_root, "user_cproc_retry_count",   &user_cproc_retry_count,    0);
+    qsnet_proc_register_int (elan4_config_root, "num_fault_save",   &num_fault_save,    0);
+    qsnet_proc_register_int (elan4_config_root, "min_fault_pages",   &min_fault_pages,    0);
+    qsnet_proc_register_int (elan4_config_root, "max_fault_pages",   &max_fault_pages,    0);
+}
+
+void
+elan4_procfs_fini(void)
+{
+    remove_proc_entry ("max_fault_pages",          elan4_config_root);
+    remove_proc_entry ("min_fault_pages",          elan4_config_root);
+    remove_proc_entry ("num_fault_save",           elan4_config_root);
+    remove_proc_entry ("user_cproc_retry_count",   elan4_config_root);
+    remove_proc_entry ("user_dproc_retry_count",   elan4_config_root);
+    remove_proc_entry ("user_bcast_route_options", elan4_config_root);
+    remove_proc_entry ("user_p2p_route_options",   elan4_config_root);
+    remove_proc_entry ("elan4_mainint_punt_loops", elan4_config_root);
+    remove_proc_entry ("elan4_debug_mmu",          elan4_config_root);
+    remove_proc_entry ("elan4_debug_tobuffer",     elan4_config_root);
+    remove_proc_entry ("elan4_debug_toconsole",    elan4_config_root);
+    remove_proc_entry ("elan4_debug",              elan4_config_root);
+
+    remove_proc_entry ("config", elan4_procfs_root);
+    remove_proc_entry ("elan4", qsnet_procfs_root);
+}
+
+EXPORT_SYMBOL(elan4_procfs_root);
+EXPORT_SYMBOL(elan4_config_root);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/quadrics_version.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/quadrics_version.h     2005-06-01 23:12:54.614436824 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/elan4/regions.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/regions.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/regions.c      2005-06-01 23:12:54.615436672 -0400
@@ -0,0 +1,609 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: regions.c,v 1.18.2.1 2004/11/18 11:31:08 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/regions.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/user.h>
+
+/*================================================================================*/
+/* elan address region management */
+USER_RGN *
+user_findrgn_elan (USER_CTXT *uctx, E4_Addr addr, int tail)
+{
+    USER_RGN *rgn;
+    USER_RGN *hirgn;
+    USER_RGN *lorgn;
+    E4_Addr   base;
+    E4_Addr   lastaddr;
+    int              forward;
+    
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_rgnlock) || kmutex_is_locked (&uctx->uctx_rgnmutex));
+
+    if (uctx->uctx_ergns == NULL)
+       return (NULL);
+
+    rgn = uctx->uctx_ergnlast;
+    if (rgn == NULL)
+       rgn = uctx->uctx_ergns;
+
+    forward = 0;
+    if ((base = rgn->rgn_ebase) < addr)
+    {
+       if (addr <= (base + rgn->rgn_len - 1))
+           return (rgn);                                       /* ergnlast contained addr */
+       
+       hirgn = uctx->uctx_etail;
+       
+       if ((lastaddr = (hirgn->rgn_ebase + hirgn->rgn_len - 1)) < addr)
+           return (tail ? hirgn : NULL);                       /* addr is out of range */
+       
+       if ((addr - base) > (lastaddr - addr))
+           rgn = hirgn;
+       else
+       {
+           rgn = rgn->rgn_enext;
+           forward++;
+       }
+    }
+    else
+    {
+       lorgn = uctx->uctx_ergns;
+
+       if (lorgn->rgn_ebase > addr)
+           return (lorgn);                                     /* lowest regions is higher than addr */
+       if ((addr - lorgn->rgn_ebase) < (base - addr))
+       {
+           rgn = lorgn;                                        /* search forward from head */
+           forward++;
+       }
+    }
+    if (forward)
+    {
+       while ((rgn->rgn_ebase + rgn->rgn_len - 1) < addr)
+           rgn = rgn->rgn_enext;
+
+       if (rgn->rgn_ebase <= addr)
+           uctx->uctx_ergnlast = rgn;
+       return (rgn);
+    }
+    else
+    {
+       while (rgn->rgn_ebase > addr)
+           rgn = rgn->rgn_eprev;
+
+       if ((rgn->rgn_ebase + rgn->rgn_len - 1) < addr)
+           return (rgn->rgn_enext);
+       else
+       {
+           uctx->uctx_ergnlast = rgn;
+           return (rgn);
+       }
+    }
+}
+
+static int
+user_addrgn_elan (USER_CTXT *uctx, USER_RGN  *nrgn)
+{
+    USER_RGN *rgn   = user_findrgn_elan (uctx, nrgn->rgn_ebase, 1);
+    E4_Addr   nbase = nrgn->rgn_ebase;
+    E4_Addr   ntop  = nbase + nrgn->rgn_len - 1;
+    E4_Addr   base;
+
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_rgnlock) && kmutex_is_locked (&uctx->uctx_rgnmutex));
+
+    if (rgn == NULL)
+    {
+       uctx->uctx_ergns = uctx->uctx_etail = nrgn;
+       nrgn->rgn_enext = nrgn->rgn_eprev = NULL;
+    }
+    else
+    {
+       base = rgn->rgn_ebase;
+       
+       if ((base + rgn->rgn_len - 1) < nbase)                  /* top of region below requested address */
+       {                                                       /* so insert after region (and hence at end */
+           nrgn->rgn_eprev = rgn;                              /* of list */
+           nrgn->rgn_enext = NULL;
+           rgn->rgn_enext = uctx->uctx_etail = nrgn;
+       }
+       else
+       {
+           if (nbase >= base || ntop >= base)                  /* overlapping region */
+               return (-1);
+           
+           nrgn->rgn_enext = rgn;                              /* insert before region */
+           nrgn->rgn_eprev = rgn->rgn_eprev;
+           rgn->rgn_eprev  = nrgn;
+           if (uctx->uctx_ergns == rgn)
+               uctx->uctx_ergns = nrgn;
+           else
+               nrgn->rgn_eprev->rgn_enext = nrgn;
+       }
+    }
+    uctx->uctx_ergnlast = nrgn;
+    
+    return (0);
+}
+
+static USER_RGN *
+user_removergn_elan (USER_CTXT *uctx, USER_RGN  *rgn)
+{
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_rgnlock) && kmutex_is_locked (&uctx->uctx_rgnmutex));
+
+    uctx->uctx_ergnlast = rgn->rgn_enext;
+    if (rgn == uctx->uctx_etail)
+       uctx->uctx_etail = rgn->rgn_eprev;
+    else
+       rgn->rgn_enext->rgn_eprev = rgn->rgn_eprev;
+    
+    if (rgn == uctx->uctx_ergns)
+       uctx->uctx_ergns = rgn->rgn_enext;
+    else
+       rgn->rgn_eprev->rgn_enext = rgn->rgn_enext;
+
+    return (rgn);
+}
+
+USER_RGN *
+user_rgnat_elan (USER_CTXT *uctx, E4_Addr addr)
+{
+    USER_RGN *rgn = user_findrgn_elan (uctx, addr, 0);
+
+    if (rgn != NULL && rgn->rgn_ebase <= addr && addr <= (rgn->rgn_ebase + rgn->rgn_len - 1))
+       return (rgn);
+
+    return (NULL);
+}
+
+/* main address region management */
+USER_RGN *
+user_findrgn_main (USER_CTXT *uctx, virtaddr_t addr, int tail)
+{
+    USER_RGN  *rgn;
+    USER_RGN  *hirgn;
+    USER_RGN  *lorgn;
+    virtaddr_t lastaddr;
+    virtaddr_t base;
+    int               forward;
+    
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_rgnlock) || kmutex_is_locked (&uctx->uctx_rgnmutex));
+
+    if (uctx->uctx_mrgns == NULL)
+       return (NULL);
+    
+    rgn = uctx->uctx_mrgnlast;
+    if (rgn == NULL)
+       rgn = uctx->uctx_mrgns;
+    
+    forward = 0;
+    if ((base = rgn->rgn_mbase) < addr)
+    {
+       if (addr <= (base + rgn->rgn_len - 1))
+           return (rgn);                                       /* ergnlast contained addr */
+       
+       hirgn = uctx->uctx_mtail;
+       if ((lastaddr = hirgn->rgn_mbase + hirgn->rgn_len - 1) < addr)
+           return (tail ? hirgn : NULL);                       /* addr is out of range */
+       
+       if ((addr - base) > (lastaddr - addr))
+           rgn = hirgn;
+       else
+       {
+           rgn = rgn->rgn_mnext;
+           forward++;
+       }
+    }
+    else
+    {
+       lorgn = uctx->uctx_mrgns;
+       if (lorgn->rgn_mbase > addr)
+           return (lorgn);                                     /* lowest regions is higher than addr */
+       if ((addr - lorgn->rgn_mbase) < (base - addr))
+       {
+           rgn = lorgn;                                        /* search forward from head */
+           forward++;
+       }
+    }
+    if (forward)
+    {
+       while ((rgn->rgn_mbase + rgn->rgn_len - 1) < addr)
+           rgn = rgn->rgn_mnext;
+
+       if (rgn->rgn_mbase <= addr)
+           uctx->uctx_mrgnlast = rgn;
+       return (rgn);
+    }
+    else
+    {
+       while (rgn->rgn_mbase > addr)
+           rgn = rgn->rgn_mprev;
+
+       if ((rgn->rgn_mbase + rgn->rgn_len - 1) < addr)
+           return (rgn->rgn_mnext);
+       else
+       {
+           uctx->uctx_mrgnlast = rgn;
+           return (rgn);
+       }
+    }
+}
+
+static int
+user_addrgn_main (USER_CTXT *uctx, USER_RGN *nrgn)
+{
+    USER_RGN  *rgn   = user_findrgn_main (uctx, nrgn->rgn_mbase, 1);
+    virtaddr_t nbase = nrgn->rgn_mbase;
+    virtaddr_t ntop  = nbase + nrgn->rgn_len - 1;
+    virtaddr_t base;
+
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_rgnlock) && kmutex_is_locked (&uctx->uctx_rgnmutex));
+
+    if (rgn == NULL)
+    {
+       uctx->uctx_mrgns = uctx->uctx_mtail = nrgn;
+       nrgn->rgn_mnext = nrgn->rgn_mprev = NULL;
+    }
+    else
+    {
+       base = rgn->rgn_mbase;
+
+       if ((base + rgn->rgn_len - 1) < nbase)                  /* top of region below requested address */
+       {                                                       /* so insert after region (and hence at end */
+           nrgn->rgn_mprev = rgn;                              /* of list */
+           nrgn->rgn_mnext = NULL;
+           rgn->rgn_mnext = uctx->uctx_mtail = nrgn;
+       }
+       else
+       {
+           if (nbase >= base || ntop >= base)                  /* overlapping region */
+               return (-1);
+
+           nrgn->rgn_mnext = rgn;                              /* insert before region */
+           nrgn->rgn_mprev = rgn->rgn_mprev;
+           rgn->rgn_mprev  = nrgn;
+           if (uctx->uctx_mrgns == rgn)
+               uctx->uctx_mrgns = nrgn;
+           else
+               nrgn->rgn_mprev->rgn_mnext = nrgn;
+       }
+    }
+    uctx->uctx_mrgnlast = nrgn;
+    
+    return (0);
+}
+
+static USER_RGN *
+user_removergn_main (USER_CTXT *uctx, USER_RGN *rgn)
+{
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_rgnlock) && kmutex_is_locked (&uctx->uctx_rgnmutex));
+
+    uctx->uctx_mrgnlast = rgn->rgn_mnext;
+    if (rgn == uctx->uctx_mtail)
+       uctx->uctx_mtail = rgn->rgn_mprev;
+    else
+       rgn->rgn_mnext->rgn_mprev = rgn->rgn_mprev;
+    
+    if (rgn == uctx->uctx_mrgns)
+       uctx->uctx_mrgns = rgn->rgn_mnext;
+    else
+       rgn->rgn_mprev->rgn_mnext = rgn->rgn_mnext;
+
+    return (rgn);
+}
+
+/* Remove whole region from both lists */
+static void
+user_removergn (USER_CTXT *uctx, USER_RGN *rgn)
+{
+    spin_lock (&uctx->uctx_rgnlock);
+
+    elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* XXXX tbl */, rgn->rgn_ebase, rgn->rgn_len);
+           
+    user_removergn_elan (uctx, rgn);
+    user_removergn_main (uctx, rgn);
+    
+    spin_unlock (&uctx->uctx_rgnlock);
+    
+    KMEM_FREE (rgn, sizeof (USER_RGN));
+}
+
+/* Remove all allocated regions */
+void
+user_freergns (USER_CTXT *uctx)
+{
+    kmutex_lock (&uctx->uctx_rgnmutex);
+
+    while (uctx->uctx_mrgns)
+       user_removergn(uctx, uctx->uctx_mrgns);
+
+    kmutex_unlock (&uctx->uctx_rgnmutex);
+
+    ASSERT (uctx->uctx_ergns == NULL);
+}
+
+USER_RGN *
+user_rgnat_main (USER_CTXT *uctx, virtaddr_t addr)
+{
+    USER_RGN *rgn = user_findrgn_main (uctx, addr, 0);
+    
+    if (rgn != NULL && rgn->rgn_mbase <= addr && addr <= (rgn->rgn_mbase + rgn->rgn_len - 1))
+       return (rgn);
+    return (NULL);
+}
+
+int
+user_setperm (USER_CTXT *uctx, virtaddr_t maddr, E4_Addr eaddr, unsigned long len, unsigned perm)
+{
+    USER_RGN *nrgn;
+
+    PRINTF4 (uctx, DBG_PERM, "user_setperm: user %lx elan %llx len %lx perm %x\n", maddr, (long long) eaddr, len, perm);
+
+    if ((maddr & PAGEOFFSET) || (eaddr & PAGEOFFSET) || (len & PAGEOFFSET))
+    {
+        PRINTF0 (uctx, DBG_PERM, "user_setperm:  alignment failure\n");
+       return (-EINVAL);
+    }
+    
+    if ((maddr + len - 1) <= maddr || (eaddr + len - 1) <= eaddr) 
+    {
+       PRINTF0 (uctx, DBG_PERM, "user_setperm:  range failure\n");
+       return (-EINVAL);
+    }
+
+    KMEM_ALLOC (nrgn, USER_RGN *, sizeof (USER_RGN), 1);
+    
+    if (nrgn == NULL)
+       return (-ENOMEM);
+
+    nrgn->rgn_mbase = maddr;
+    nrgn->rgn_ebase = eaddr;
+    nrgn->rgn_len   = len;
+    nrgn->rgn_perm  = perm;
+
+    kmutex_lock (&uctx->uctx_rgnmutex);
+    spin_lock (&uctx->uctx_rgnlock);
+
+    if (user_addrgn_elan (uctx, nrgn) < 0)
+    {
+       PRINTF0 (uctx, DBG_PERM, "user_setperm:  elan address exists\n");
+       spin_unlock (&uctx->uctx_rgnlock);
+       kmutex_unlock (&uctx->uctx_rgnmutex);
+
+       KMEM_FREE (nrgn, sizeof (USER_RGN));
+       return (-EINVAL);
+    }
+    
+    if (user_addrgn_main (uctx, nrgn) < 0)
+    {
+       PRINTF0 (uctx, DBG_PERM, "user_setperm:  main address exists\n");
+       user_removergn_elan (uctx, nrgn);
+       
+       spin_unlock (&uctx->uctx_rgnlock);
+       kmutex_unlock (&uctx->uctx_rgnmutex);
+
+       KMEM_FREE (nrgn, sizeof (USER_RGN));
+       return (-EINVAL);
+    }
+    spin_unlock (&uctx->uctx_rgnlock);
+
+    if ((perm & PERM_Preload))
+       user_preload_main (uctx, maddr, len);
+
+    kmutex_unlock (&uctx->uctx_rgnmutex);
+
+    return (0);
+}
+
+void
+user_clrperm (USER_CTXT *uctx, E4_Addr addr, unsigned long len)
+{
+    E4_Addr       raddr;
+    E4_Addr       rtop;
+    USER_RGN     *nrgn;
+    USER_RGN     *rgn;
+    USER_RGN     *rgn_next;
+    unsigned long ssize;
+    int                  res;
+
+    PRINTF2 (uctx, DBG_PERM, "user_clrperm: elan %llx len %lx\n", addr, len);
+
+    raddr = (addr & PAGEMASK);
+    rtop = ((addr + len - 1) & PAGEMASK) + (PAGESIZE-1);
+
+    kmutex_lock (&uctx->uctx_rgnmutex);
+    
+    for (rgn = user_findrgn_elan (uctx, addr, 0); rgn != NULL; rgn = rgn_next)
+    {
+       if (rtop < rgn->rgn_ebase)                              /* rtop was in a gap */
+           break;
+       
+       rgn_next = rgn->rgn_enext;                              /* Save next region pointer */
+       
+       PRINTF (uctx, DBG_PERM, "              elan %llx->%llx main %p->%p\n", 
+               rgn->rgn_ebase, rgn->rgn_ebase + rgn->rgn_len-1,
+               rgn->rgn_mbase, rgn->rgn_mbase + rgn->rgn_len-1);
+
+       if (raddr <= rgn->rgn_ebase && rtop >= (rgn->rgn_ebase + rgn->rgn_len - 1))
+       {
+           /* whole region is cleared */
+
+           PRINTF (uctx, DBG_PERM, "              whole region\n");
+           PRINTF (uctx, DBG_PERM, "              unload elan %llx->%llx\n", rgn->rgn_ebase, rgn->rgn_ebase + rgn->rgn_len-1);
+           user_removergn (uctx, rgn);
+       }
+       else if (raddr <= rgn->rgn_ebase)
+       {
+           /* clearing at beginning, so shrink size and increment base ptrs */
+           ssize = rtop - rgn->rgn_ebase + 1;
+           
+           PRINTF (uctx, DBG_PERM, "              clear at beginning %x\n", ssize);
+
+           spin_lock (&uctx->uctx_rgnlock);
+
+           PRINTF (uctx, DBG_PERM, "              unload elan %llx->%llx\n", rgn->rgn_ebase, rgn->rgn_ebase + ssize-1);
+           elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* XXXX tbl */, rgn->rgn_ebase, ssize);
+
+           rgn->rgn_mbase += ssize;
+           rgn->rgn_ebase += ssize;
+           rgn->rgn_len   -= ssize;
+           
+           spin_unlock(&uctx->uctx_rgnlock);
+       }
+       else if (rtop >= (rgn->rgn_ebase + rgn->rgn_len - 1))
+       {
+           /* clearing at end, so just shrink length of region */
+           ssize = (rgn->rgn_ebase + rgn->rgn_len - 1) - raddr + 1;
+
+           PRINTF (uctx, DBG_PERM, "              clear at end %x\n", ssize);
+
+           spin_lock (&uctx->uctx_rgnlock);
+
+           PRINTF (uctx, DBG_PERM, "              unload elan %llx->%llx\n", raddr, raddr+ssize-1);
+           elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* XXXX tbl */, raddr, ssize);
+
+           rgn->rgn_len -= ssize;
+
+           spin_unlock(&uctx->uctx_rgnlock);
+       }
+       else
+       {
+           /* the section to go is in the middle,  so need to  */
+           /* split it into two regions */
+           KMEM_ALLOC (nrgn, USER_RGN *, sizeof (USER_RGN), 1);
+
+           spin_lock (&uctx->uctx_rgnlock);
+
+           PRINTF (uctx, DBG_PERM, "              unload elan %llx->%llx\n", raddr, rtop);
+           elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* XXXX tbl */, raddr, rtop - raddr + 1);
+
+           nrgn->rgn_mbase = rgn->rgn_mbase + (rtop - rgn->rgn_ebase + 1);
+           nrgn->rgn_ebase = rtop + 1;
+           nrgn->rgn_len   = (rgn->rgn_ebase + rgn->rgn_len - 1) - rtop;
+           nrgn->rgn_perm  = rgn->rgn_perm;
+
+           PRINTF (uctx, DBG_PERM, "              new elan %llx->%llx main %p->%p\n", 
+                   nrgn->rgn_ebase, nrgn->rgn_ebase + nrgn->rgn_len-1,
+                   nrgn->rgn_mbase, nrgn->rgn_mbase + nrgn->rgn_len-1);
+
+           rgn->rgn_len = (raddr - rgn->rgn_ebase);            /* shrink original region */
+
+           PRINTF (uctx, DBG_PERM, "              old elan %llx->%llx main %p->%p\n", 
+                   rgn->rgn_ebase, rgn->rgn_ebase + rgn->rgn_len-1,
+                   rgn->rgn_mbase, rgn->rgn_mbase + rgn->rgn_len-1);
+
+           res = user_addrgn_elan (uctx, nrgn);                /* insert new region */
+           ASSERT (res == 0);                                  /* which cannot fail */
+
+           res = user_addrgn_main (uctx, nrgn);        
+           ASSERT (res == 0);
+
+           spin_unlock(&uctx->uctx_rgnlock);
+       }
+    }
+    kmutex_unlock (&uctx->uctx_rgnmutex);
+}
+
+int
+user_checkperm (USER_CTXT *uctx, E4_Addr raddr, unsigned long rsize, unsigned access)
+{
+    USER_RGN *rgn;
+
+    PRINTF3 (uctx, DBG_PERM, "user_checkperm: elan %lx len %lx access %x\n", raddr, rsize, access);
+    
+    if ((raddr + rsize - 1) < raddr)
+       return (-ENOMEM);
+    
+    kmutex_lock (&uctx->uctx_rgnmutex);
+    if ((rgn = user_rgnat_elan (uctx, raddr)) == (USER_RGN *) NULL)
+    {
+       kmutex_unlock (&uctx->uctx_rgnmutex);
+       return (-ENOMEM);
+    }
+    else
+    {
+       register int ssize;
+       
+       for (; rsize != 0; rsize -= ssize, raddr += ssize)
+       {
+           if (raddr > (rgn->rgn_ebase + rgn->rgn_len - 1))
+           {
+               rgn  = rgn->rgn_enext;
+               
+               if (rgn == NULL || raddr != rgn->rgn_ebase)
+               {
+                   kmutex_unlock (&uctx->uctx_rgnmutex);
+                   return (-ENOMEM);
+               }
+           }
+           if ((raddr + rsize - 1) > (rgn->rgn_ebase + rgn->rgn_len - 1))
+               ssize = ((rgn->rgn_ebase + rgn->rgn_len - 1) - raddr) + 1;
+           else
+               ssize = rsize;
+           
+           PRINTF4 (uctx, DBG_PERM, "user_checkperm : rgn %lx -> %lx perm %x access %x\n",
+                    rgn->rgn_ebase, rgn->rgn_ebase + (E4_Addr)rgn->rgn_len, rgn->rgn_perm, access);
+
+           if (ELAN4_INCOMPAT_ACCESS (rgn->rgn_perm, access))
+           {
+               kmutex_unlock (&uctx->uctx_rgnmutex);
+               return (-EACCES);
+           }
+       }
+    }
+    
+    kmutex_unlock (&uctx->uctx_rgnmutex);
+    
+    return (0);
+}
+
+virtaddr_t
+user_elan2main (USER_CTXT *uctx, E4_Addr addr)
+{
+    USER_RGN  *rgn;
+    virtaddr_t raddr;
+    
+    spin_lock (&uctx->uctx_rgnlock);
+    
+    if ((rgn = user_rgnat_elan (uctx, addr)) == (USER_RGN *) NULL)
+       raddr = (virtaddr_t) 0;
+    else
+       raddr = rgn->rgn_mbase + (addr - rgn->rgn_ebase);
+
+    spin_unlock (&uctx->uctx_rgnlock);
+
+    return (raddr);
+}
+
+E4_Addr
+user_main2elan (USER_CTXT *uctx, virtaddr_t addr)
+{
+    USER_RGN *rgn;
+    E4_Addr   raddr;
+
+    spin_lock (&uctx->uctx_rgnlock);
+    
+    if ((rgn = user_rgnat_main (uctx, addr)) == (USER_RGN *) NULL)
+       raddr = (virtaddr_t) 0;
+    else
+       raddr = rgn->rgn_ebase + (addr - rgn->rgn_mbase);
+    
+    spin_unlock (&uctx->uctx_rgnlock);
+
+    return (raddr);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/routetable.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/routetable.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/routetable.c   2005-06-01 23:12:54.615436672 -0400
@@ -0,0 +1,249 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: routetable.c,v 1.15 2004/07/20 09:29:40 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/routetable.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/sdram.h>
+#include <elan4/debug.h>
+#include <elan4/device.h>
+
+ELAN4_ROUTE_TABLE *
+elan4_alloc_routetable (ELAN4_DEV *dev, unsigned size)
+{
+    ELAN4_ROUTE_TABLE *tbl;
+
+    KMEM_ZALLOC (tbl, ELAN4_ROUTE_TABLE *, sizeof (ELAN4_ROUTE_TABLE), 1);
+
+    if (tbl == (ELAN4_ROUTE_TABLE *) NULL)
+       return (NULL);
+    
+    tbl->tbl_size    = (size & E4_VPT_SIZE_MASK);
+    tbl->tbl_entries = elan4_sdram_alloc (dev, (E4_VPT_MIN_ENTRIES << tbl->tbl_size) * sizeof (E4_VirtualProcessEntry));
+
+    if (tbl->tbl_entries == 0)
+    {
+       KMEM_FREE (tbl, sizeof (ELAN4_ROUTE_TABLE));
+       return ((ELAN4_ROUTE_TABLE *) NULL);
+    }
+
+    spin_lock_init (&tbl->tbl_lock);
+
+    /* zero the route table */
+    elan4_sdram_zeroq_sdram (dev, tbl->tbl_entries, (E4_VPT_MIN_ENTRIES << tbl->tbl_size) * sizeof (E4_VirtualProcessEntry));
+
+    return (tbl);
+}
+
+void
+elan4_free_routetable (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl)
+{
+    elan4_sdram_free (dev, tbl->tbl_entries, (E4_VPT_MIN_ENTRIES << tbl->tbl_size) * sizeof (E4_VirtualProcessEntry));
+    
+    spin_lock_destroy (&tbl->tbl_lock);
+
+    KMEM_FREE (tbl, sizeof (ELAN4_ROUTE_TABLE));
+}
+
+void
+elan4_write_route (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl, unsigned vp, E4_VirtualProcessEntry *entry)
+{
+    ASSERT (vp < (E4_VPT_MIN_ENTRIES << tbl->tbl_size));
+    
+    elan4_sdram_writeq (dev, tbl->tbl_entries + (vp * sizeof (E4_VirtualProcessEntry)) + offsetof (E4_VirtualProcessEntry, Values[1]), entry->Values[1]);
+    elan4_sdram_writeq (dev, tbl->tbl_entries + (vp * sizeof (E4_VirtualProcessEntry)) + offsetof (E4_VirtualProcessEntry, Values[0]), entry->Values[0]);
+    pioflush_sdram (dev);
+}
+
+void
+elan4_read_route (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl, unsigned vp, E4_VirtualProcessEntry *entry)
+{
+    ASSERT (vp < (E4_VPT_MIN_ENTRIES << tbl->tbl_size));
+    
+    entry->Values[0] = elan4_sdram_readq (dev, tbl->tbl_entries + (vp * sizeof (E4_VirtualProcessEntry)) + offsetof (E4_VirtualProcessEntry, Values[0]));
+    entry->Values[1] = elan4_sdram_readq (dev, tbl->tbl_entries + (vp * sizeof (E4_VirtualProcessEntry)) + offsetof (E4_VirtualProcessEntry, Values[1]));
+}
+
+void
+elan4_invalidate_route (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl, unsigned vp)
+{
+    ASSERT (vp < (E4_VPT_MIN_ENTRIES << tbl->tbl_size));
+
+    elan4_sdram_writeq (dev, tbl->tbl_entries + (vp * sizeof (E4_VirtualProcessEntry)) + offsetof (E4_VirtualProcessEntry, Values[0]), 0);
+    elan4_sdram_writeq (dev, tbl->tbl_entries + (vp * sizeof (E4_VirtualProcessEntry)) + offsetof (E4_VirtualProcessEntry, Values[1]), 0);
+    pioflush_sdram (dev);
+}
+
+static void
+pack_them_routes (E4_VirtualProcessEntry *entry, E4_uint16 first, E4_uint8 *packed, unsigned ctx)
+{
+    E4_uint64 value0 = first;
+    E4_uint64 value1 = ROUTE_CTXT_VALUE(ctx);
+    E4_uint32 ThirdRouteBCastVal;
+    register int i;
+
+    for (i = 0; i < (ROUTE_NUM_PACKED >> 1); i++)
+    {
+       value0 |= ((E4_uint64) packed[i]) << ((i << 2) + ROUTE_PACKED_OFFSET);
+       value1 |= ((E4_uint64) packed[i+(ROUTE_NUM_PACKED >> 1)]) << ((i << 2));
+    }
+
+    /* DMA fix for large broadcast route values that fall into the double issue of route value 3 bug. */
+    /* NOTE - this is only required when the link is running in Mod45 mode, it could be automatically
+     *        disabled when Mod44 is detected */
+    
+    /* First seach for the alignment type. The bug is only sensitive to an odd bcast aligment on the 3rd word. */
+    for (i=4;i<16;i++)
+       if (((value0 >> (i*4)) & 0xc) == 4)
+           i++;
+    
+    if (i == 17)
+    {
+       ThirdRouteBCastVal = value1 & 0xcccccccc;
+       if      (((value1 & 0xfffff0000000ULL) == 0ULL) && (ThirdRouteBCastVal == 0x04444444))
+           value1 |= 0x140000000ULL;
+       else if (((value1 & 0xfffffff00000ULL) == 0ULL) && (ThirdRouteBCastVal == 0x00044444))
+           value1 |= 0x1400000ULL;
+       else if (((value1 & 0xfffffffff000ULL) == 0ULL) && (ThirdRouteBCastVal == 0x00000444))
+           value1 |= 0x14000ULL;
+       else if (((value1 & 0xfffffffffff0ULL) == 0ULL) && (ThirdRouteBCastVal == 0x00000004))
+           value1 |= 0x140ULL;
+    }
+    
+    entry->Values[0] = value0;
+    entry->Values[1] = value1;
+}
+
+int
+elan4_generate_route (ELAN_POSITION *pos, E4_VirtualProcessEntry *route, unsigned ctx, unsigned lowid, unsigned highid, unsigned options)
+{
+    unsigned int broadcast    = (lowid != highid);
+    unsigned int noadaptive   = 0;
+    int          padbcast     = 0;
+    E4_uint16    first;
+    int                 rb;
+    E4_uint8     packed[ROUTE_NUM_PACKED];
+    int                 level, llink, hlink;
+
+ regenerate_routes:
+    first = 0;
+    rb    = 0;
+
+    switch (pos->pos_mode)
+    {
+    case ELAN_POS_MODE_LOOPBACK:
+       if (lowid != highid || lowid != pos->pos_nodeid)
+           return (-EINVAL);
+       
+       route->Values[0] = FIRST_MYLINK;
+       route->Values[1] = ROUTE_CTXT_VALUE (ctx);
+       return (0);
+
+    case ELAN_POS_MODE_BACKTOBACK:
+       if (lowid != highid || lowid == pos->pos_nodeid)
+           return (-EINVAL);
+       
+       route->Values[0] = FIRST_MYLINK;
+       route->Values[1] = ROUTE_CTXT_VALUE (ctx);
+       return (0);
+
+    case ELAN_POS_MODE_SWITCHED:
+    {
+       unsigned char *arityp  = &pos->pos_arity[pos->pos_levels - 1];
+       unsigned int   spanned = *arityp;
+       unsigned int   broadcasting = 0;
+       
+       bzero (packed, sizeof (packed));
+
+       /* XXXX compute noadaptive ? */
+
+       for (level = 0; 
+            level < pos->pos_levels && ! ((pos->pos_nodeid / spanned) == (lowid / spanned) &&
+                                          (pos->pos_nodeid / spanned) ==  (highid / spanned)); 
+            level++, spanned *= *(--arityp))
+       {
+           if (first == 0)
+               first = (broadcast || noadaptive) ? FIRST_BCAST_TREE : FIRST_ADAPTIVE;
+           else if (broadcast && padbcast)
+           {
+               padbcast = 0;
+               packed[rb++] = PACKED_BCAST0(4, 4);
+               packed[rb++] = PACKED_BCAST1(4, 4);
+           }
+           else
+               packed[rb++] = (broadcast || noadaptive) ? PACKED_BCAST_TREE : PACKED_ADAPTIVE;    
+       }
+
+       while (level >= 0)
+       {
+           spanned /= *arityp;
+           
+           llink = (lowid  / spanned) % *arityp;
+           hlink = (highid / spanned) % *arityp;
+           
+           if (llink != hlink || broadcasting)
+           {
+               broadcasting = 1;
+               
+               if (first == 0)
+                   first = FIRST_BCAST (hlink, llink);
+               else
+               {
+                   packed[rb++] = PACKED_BCAST0(hlink, llink);
+                   
+                   if ((rb % 4) == 0 && PACKED_BCAST1(hlink, llink) == 0)
+                   {
+                       padbcast = 1;
+                       goto regenerate_routes;
+                   }
+                   
+                   packed[rb++] = PACKED_BCAST1(hlink, llink);
+               }
+           }
+           else
+           {
+               if (first == 0)
+                   first = FIRST_ROUTE(llink);
+               else
+                   packed[rb++] = PACKED_ROUTE(llink);
+           }
+           
+           level--;
+           arityp++;
+       }
+
+       pack_them_routes (route, first | (options & FIRST_OPTIONS_MASK), packed, ctx);
+       return (0);
+    }
+    }
+
+    return (-EINVAL);
+}
+
+int
+elan4_check_route (ELAN_POSITION *postiion, ELAN_LOCATION location, E4_VirtualProcessEntry *route, unsigned flags)
+{
+    /* XXXX - TBD */
+    return (0);
+}
+
+EXPORT_SYMBOL(elan4_alloc_routetable);
+EXPORT_SYMBOL(elan4_free_routetable);
+EXPORT_SYMBOL(elan4_write_route);
+EXPORT_SYMBOL(elan4_read_route);
+EXPORT_SYMBOL(elan4_invalidate_route);
+EXPORT_SYMBOL(elan4_generate_route);
+EXPORT_SYMBOL(elan4_check_route);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/sdram.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/sdram.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/sdram.c        2005-06-01 23:12:54.617436368 -0400
@@ -0,0 +1,1034 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: sdram.c,v 1.29.6.1 2004/11/29 11:39:13 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/sdram.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+
+EXPORT_SYMBOL_GPL(elan4_sdram_readb);
+EXPORT_SYMBOL_GPL(elan4_sdram_readw);
+EXPORT_SYMBOL_GPL(elan4_sdram_readl);
+EXPORT_SYMBOL_GPL(elan4_sdram_readq);
+EXPORT_SYMBOL_GPL(elan4_sdram_writeb);
+EXPORT_SYMBOL_GPL(elan4_sdram_writew);
+EXPORT_SYMBOL_GPL(elan4_sdram_writel);
+EXPORT_SYMBOL_GPL(elan4_sdram_writeq);
+EXPORT_SYMBOL_GPL(elan4_sdram_zerob_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_zerow_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_zerol_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_zeroq_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyb_from_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyw_from_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyl_from_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyq_from_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyb_to_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyw_to_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyl_to_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_copyq_to_sdram);
+EXPORT_SYMBOL_GPL(elan4_sdram_alloc);
+EXPORT_SYMBOL_GPL(elan4_sdram_free);
+EXPORT_SYMBOL_GPL(elan4_sdram_flushcache);
+
+#define SDRAM_MIN_BANK_SIZE            ((1 << 15) * 8)         /* 256 Kbytes */
+
+static inline ELAN4_SDRAM_BANK *
+sdramaddr_to_bank (ELAN4_DEV *dev, sdramaddr_t saddr)
+{
+    register int i;
+    
+    for (i = 0; i < dev->dev_sdram_numbanks; i++)
+    {
+       ELAN4_SDRAM_BANK *bank = &dev->dev_sdram_banks[i];
+
+       if (saddr >= bank->b_base && saddr < (bank->b_base + bank->b_size))
+           return (bank);
+    }
+    printk ("sdramaddr_to_bank: sdram address %lx not in a sdram bank\n", saddr);
+    BUG();
+
+    return (NULL);     /* NOTREACHED */
+}
+
+static inline int
+sdramaddr_to_bankoffset (ELAN4_DEV *dev, sdramaddr_t saddr)
+{
+    return (saddr & (sdramaddr_to_bank (dev, saddr)->b_size-1));
+}
+
+static inline int
+sdramaddr_to_bit(ELAN4_DEV *dev, int indx, sdramaddr_t saddr)
+{
+    return (sdramaddr_to_bankoffset(dev, saddr) >> (SDRAM_MIN_BLOCK_SHIFT+(indx)));
+}
+
+static inline ioaddr_t
+sdramaddr_to_ioaddr (ELAN4_DEV *dev, sdramaddr_t saddr)
+{
+    ELAN4_SDRAM_BANK *bank = sdramaddr_to_bank (dev, saddr);
+
+    return (bank->b_ioaddr + (saddr - bank->b_base));
+}
+
+unsigned char
+elan4_sdram_readb (ELAN4_DEV *dev, sdramaddr_t off)
+{
+    return (__elan4_readb (dev, sdramaddr_to_ioaddr(dev, off)));
+}
+
+unsigned short
+elan4_sdram_readw (ELAN4_DEV *dev, sdramaddr_t off)
+{
+    return (__elan4_readw (dev, sdramaddr_to_ioaddr(dev, off)));
+}
+
+unsigned int
+elan4_sdram_readl (ELAN4_DEV *dev, sdramaddr_t off)
+{
+    return (__elan4_readl (dev, sdramaddr_to_ioaddr(dev, off)));
+}
+
+unsigned long long
+elan4_sdram_readq (ELAN4_DEV *dev, sdramaddr_t off)
+{
+    return (readq (sdramaddr_to_ioaddr(dev, off)));
+}
+
+void
+elan4_sdram_writeb (ELAN4_DEV *dev, sdramaddr_t off, unsigned char val)
+{
+    writeb (val, sdramaddr_to_ioaddr(dev, off));
+
+    mb();
+}
+
+void
+elan4_sdram_writew (ELAN4_DEV *dev, sdramaddr_t off, unsigned short val)
+{
+    writew (val, sdramaddr_to_ioaddr(dev, off));
+
+    mb();
+}
+
+void
+elan4_sdram_writel (ELAN4_DEV *dev, sdramaddr_t off, unsigned int val)
+{
+    writel (val, sdramaddr_to_ioaddr(dev, off));
+
+    mb();
+}
+
+void
+elan4_sdram_writeq (ELAN4_DEV *dev, sdramaddr_t off, unsigned long long val)
+{
+    writeq (val, sdramaddr_to_ioaddr(dev, off));
+
+    mb();
+}
+
+void
+elan4_sdram_zerob_sdram (ELAN4_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    ioaddr_t lim  = dest + nbytes;
+
+    for (; dest < lim; dest += sizeof (u8))
+       writeb (0, dest);
+}
+
+void
+elan4_sdram_zerow_sdram (ELAN4_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    ioaddr_t lim  = dest + nbytes;
+
+    for (; dest < lim; dest += sizeof (u8))
+       writeb (0, dest);
+}
+
+void
+elan4_sdram_zerol_sdram (ELAN4_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    ioaddr_t lim  = dest + nbytes;
+
+    for (; dest < lim; dest += sizeof (u32))
+       writel (0, dest);
+}
+
+void
+elan4_sdram_zeroq_sdram (ELAN4_DEV *dev, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    ioaddr_t lim  = dest + nbytes;
+
+#ifdef CONFIG_MPSAS
+    if (sas_memset_dev (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM, to, 0, nbytes) == 0)
+       return;
+#endif
+
+    for (; dest < lim; dest += sizeof (u64))
+       writeq (0, dest);
+}
+
+void
+elan4_sdram_copyb_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+    ioaddr_t src  = sdramaddr_to_ioaddr (dev, from);
+    u8      *dest = (u8 *) to;
+    ioaddr_t lim  = src + nbytes;
+
+    for (; src < lim; src += sizeof (u8))
+       *dest++ = __elan4_readb (dev, src);
+}
+
+void
+elan4_sdram_copyw_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+    ioaddr_t src  = sdramaddr_to_ioaddr (dev, from);
+    u16     *dest = (u16 *) to;
+    ioaddr_t lim  = src + nbytes;
+
+    for (; src < lim; src += sizeof (u16))
+       *dest++ = __elan4_readw (dev, src);
+}
+
+void
+elan4_sdram_copyl_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+    ioaddr_t src  = sdramaddr_to_ioaddr (dev, from);
+    u32     *dest = (u32 *) to;
+    ioaddr_t lim  = src + nbytes;
+
+    for (; src < lim; src += sizeof (u32))
+       *dest++ = __elan4_readl (dev, src);
+}
+
+void
+elan4_sdram_copyq_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes)
+{
+    ioaddr_t src  = sdramaddr_to_ioaddr (dev, from);
+    u64     *dest = (u64 *) to;
+    ioaddr_t lim  = src + nbytes;
+
+#ifdef CONFIG_MPSAS
+    if (sas_copyfrom_dev (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM, from, (unsigned long) to, nbytes) == 0)
+       return;
+#endif
+
+    for (; src < lim; src += sizeof (u64))
+       *dest++ = readq (src);
+}
+
+void
+elan4_sdram_copyb_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    u8      *src  = (u8 *) from;
+    ioaddr_t lim  = dest + nbytes;
+
+    for (; dest < lim; dest += sizeof (u8))
+       writeb (*src++, dest);
+
+    mb();
+}
+
+void
+elan4_sdram_copyw_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    u16     *src  = (u16 *) from;
+    ioaddr_t lim  = dest + nbytes;
+
+    for (; dest < lim; dest += sizeof (u16))
+       writew (*src++, dest);
+
+    mb();
+}
+
+void
+elan4_sdram_copyl_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    u32     *src  = (u32 *) from;
+    ioaddr_t lim  = dest + nbytes;
+
+    for (; dest < lim; dest += sizeof (u16))
+       writew (*src++, dest);
+
+    mb();
+}
+
+void
+elan4_sdram_copyq_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes)
+{
+    ioaddr_t dest = sdramaddr_to_ioaddr (dev, to);
+    u64     *src  = (u64 *) from;
+    ioaddr_t lim  = dest + nbytes;
+
+#ifdef CONFIG_MPSAS
+    if (sas_copyto_dev (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM, to, (unsigned long) from, nbytes) == 0)
+       return;
+#endif
+
+    for (; dest < lim; dest += sizeof (u64))
+       writeq (*src++, dest);
+
+    mb();
+}
+
+/* sdram buddy allocator */
+typedef struct sdramblock
+{
+    sdramaddr_t        next;
+    sdramaddr_t prev;
+} sdramblock_t;
+
+static inline sdramaddr_t
+read_next (ELAN4_DEV *dev, sdramaddr_t block)
+{
+    return __elan4_readl (dev, sdramaddr_to_ioaddr (dev, block + offsetof (sdramblock_t, next)));
+}
+
+static inline sdramaddr_t
+read_prev (ELAN4_DEV *dev, sdramaddr_t block)
+{
+    return __elan4_readl (dev, sdramaddr_to_ioaddr (dev, block + offsetof (sdramblock_t, prev)));
+}
+
+static inline void
+write_next (ELAN4_DEV *dev, sdramaddr_t block, sdramaddr_t val)
+{
+    writel (val, sdramaddr_to_ioaddr (dev, block + offsetof (sdramblock_t, next)));
+}
+
+static inline void
+write_prev (ELAN4_DEV *dev, sdramaddr_t block, sdramaddr_t val)
+{
+    writel (val, sdramaddr_to_ioaddr (dev, block + offsetof (sdramblock_t, prev)));
+}
+
+static inline void
+freelist_insert (ELAN4_DEV *dev, int idx, sdramaddr_t block)
+{
+    sdramaddr_t next = dev->dev_sdram_freelists[(idx)];
+
+    /*
+     * block->prev = NULL;
+     * block->next = next;
+     * if (next != NULL)
+     *    next->prev = block;
+     * freelist = block;
+     */
+    write_prev (dev, block, (sdramaddr_t) 0);
+    write_next (dev, block, next);
+    if (next != (sdramaddr_t) 0)
+       write_prev (dev, next, block);
+    dev->dev_sdram_freelists[idx] = block;
+
+    dev->dev_sdram_freecounts[idx]++;
+    dev->dev_stats.s_sdram_bytes_free += (SDRAM_MIN_BLOCK_SIZE << idx);
+
+    mb();
+}
+
+static inline void
+freelist_remove (ELAN4_DEV *dev,int idx, sdramaddr_t block)
+{
+    /*
+     * if (block->prev)
+     *     block->prev->next = block->next;
+     * else
+     *     dev->dev_sdram_freelists[idx] = block->next;
+     * if (block->next)
+     *     block->next->prev = block->prev;
+     */
+    sdramaddr_t blocknext = read_next (dev, block);
+    sdramaddr_t blockprev = read_prev (dev, block);
+
+    if (blockprev)
+       write_next (dev, blockprev, blocknext);
+    else
+       dev->dev_sdram_freelists[idx] = blocknext;
+    if (blocknext)
+       write_prev (dev, blocknext, blockprev);
+
+    dev->dev_sdram_freecounts[idx]--;
+    dev->dev_stats.s_sdram_bytes_free -= (SDRAM_MIN_BLOCK_SIZE << idx);
+
+    mb();
+}
+
+static inline void
+freelist_removehead(ELAN4_DEV *dev, int idx, sdramaddr_t block)
+{
+    sdramaddr_t blocknext = read_next (dev, block);
+
+    if ((dev->dev_sdram_freelists[idx] = blocknext) != 0)
+       write_prev (dev, blocknext, 0);
+
+    dev->dev_sdram_freecounts[idx]--;
+    dev->dev_stats.s_sdram_bytes_free -= (SDRAM_MIN_BLOCK_SIZE << idx);
+
+    mb();
+}
+
+#ifdef DEBUG
+static int
+display_blocks (ELAN4_DEV *dev, int indx, char *string)
+{
+    sdramaddr_t block;
+    int nbytes = 0;
+
+    PRINTF (DBG_DEVICE, DBG_SDRAM, "%s - indx %d\n", string, indx);
+    for (block = dev->dev_sdram_freelists[indx]; block != (sdramaddr_t) 0; block = read_next (dev, block))
+    {
+       PRINTF (DBG_DEVICE, DBG_SDRAM, "  %x\n", block);
+       nbytes += (SDRAM_MIN_BLOCK_SIZE << indx);
+    }
+
+    return (nbytes);
+}
+
+void
+elan4_sdram_display (ELAN4_DEV *dev, char *string)
+{
+    int indx;
+    int nbytes = 0;
+    
+    PRINTF (DBG_DEVICE, DBG_SDRAM, "elan4_sdram_display: dev=%p\n", dev);
+    for (indx = 0; indx < SDRAM_NUM_FREE_LISTS; indx++)
+       if (dev->dev_sdram_freelists[indx] != (sdramaddr_t) 0)
+           nbytes += display_blocks (dev, indx, string);
+    PRINTF (DBG_DEVICE, DBG_SDRAM, "\n%d bytes free - %d pages free\n", nbytes, nbytes/SDRAM_PAGE_SIZE);
+}
+
+void
+elan4_sdram_verify (ELAN4_DEV *dev)
+{
+    int indx, size, nbits, i, b;
+    sdramaddr_t block;
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; indx < SDRAM_NUM_FREE_LISTS; indx++, size <<= 1)
+    {
+       unsigned count = 0;
+
+       for (block = dev->dev_sdram_freelists[indx]; block; block = read_next (dev, block), count++)
+       {
+           ELAN4_SDRAM_BANK *bank = sdramaddr_to_bank (dev, block);
+           unsigned         off  = sdramaddr_to_bankoffset (dev, block);
+           int              bit  = sdramaddr_to_bit (dev, indx, block);
+
+           if ((block & (size-1)) != 0)
+               printk ("elan4_sdram_verify: block=%lx indx=%x - not aligned\n", block, indx);
+           
+           if (bank == NULL || off > bank->b_size)
+               printk ("elan4_sdram_verify: block=%lx indx=%x - outside bank\n", block, indx);
+           else if (BT_TEST (bank->b_bitmaps[indx], bit) == 0)
+               printk ("elan4_sdram_verify: block=%lx indx=%x - bit not set\n", block, indx);
+           else
+           {
+               for (i = indx-1, nbits = 2; i >= 0; i--, nbits <<= 1)
+               {
+                   bit = sdramaddr_to_bit (dev, i, block);
+
+                   for (b = 0; b < nbits; b++)
+                       if (BT_TEST(bank->b_bitmaps[i], bit + b))
+                           printk ("elan4_sdram_verify: block=%lx indx=%x - also free i=%d bit=%x\n", block, indx, i, bit+b);
+               }
+           }
+       }
+
+       if (dev->dev_sdram_freecounts[indx] != count)
+           printk ("elan4_sdram_verify: indx=%x expected %d got %d\n", indx, dev->dev_sdram_freecounts[indx], count);
+    }
+}
+
+#endif
+
+static void
+free_block (ELAN4_DEV *dev, sdramaddr_t block, int indx)
+{
+    ELAN4_SDRAM_BANK *bank = sdramaddr_to_bank (dev, block);
+    unsigned         bit  = sdramaddr_to_bit (dev, indx, block);
+    unsigned         size = SDRAM_MIN_BLOCK_SIZE << indx;
+
+    PRINTF3 (DBG_DEVICE, DBG_SDRAM, "free_block: block=%x indx=%d bit=%x\n", block, indx, bit);
+
+    ASSERT ((block & (size-1)) == 0);
+    ASSERT (BT_TEST (bank->b_bitmaps[indx], bit) == 0);
+
+    while (BT_TEST (bank->b_bitmaps[indx], bit ^ 1))
+    {
+       sdramaddr_t buddy = block ^ size;
+       
+       PRINTF3 (DBG_DEVICE, DBG_SDRAM, "free_block: merge block=%x buddy=%x indx=%d\n", block, buddy, indx);
+       
+       BT_CLEAR (bank->b_bitmaps[indx], bit ^ 1);
+       
+       freelist_remove (dev, indx, buddy);
+       
+       block = (block < buddy) ? block : buddy;
+       indx++;
+       size <<= 1;
+       bit >>= 1;
+    }
+    
+    PRINTF3 (DBG_DEVICE, DBG_SDRAM, "free_block: free block=%x indx=%d bit=%x\n", block, indx, bit);
+    
+    freelist_insert (dev, indx, block);
+    
+    BT_SET (bank->b_bitmaps[indx], bit);
+}
+
+void
+elan4_sdram_init (ELAN4_DEV *dev)
+{
+    int indx;
+
+    spin_lock_init (&dev->dev_sdram_lock);
+
+    for (indx = 0; indx < SDRAM_NUM_FREE_LISTS; indx++)
+    {
+       dev->dev_sdram_freelists[indx]  = (sdramaddr_t) 0;
+       dev->dev_sdram_freecounts[indx] = 0;
+    }
+}
+
+void
+elan4_sdram_fini (ELAN4_DEV *dev)
+{
+    spin_lock_destroy (&dev->dev_sdram_lock);
+}
+
+#ifdef CONFIG_MPSAS
+/* size of Elan SDRAM in simulation */
+#define SDRAM_used_addr_bits           (16)
+#define SDRAM_SIMULATION_BANK_SIZE     ((1 << SDRAM_used_addr_bits) * 8)       /* 128 kbytes */
+
+static int
+elan4_sdram_probe_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank)
+{
+    printk ("elan%d: memory bank %d is %d Kb\n", dev->dev_instance, (int) (bank - dev->dev_sdram_banks), (int) (SDRAM_SIMULATION_BANK_SIZE / 1024));
+
+    bank->b_size = SDRAM_SIMULATION_BANK_SIZE;
+
+    return 1;
+}
+
+#else
+
+static void
+initialise_cache_tags (ELAN4_DEV *dev, unsigned addr)
+{
+    register int set, line;
+
+    mb();
+
+    /* Initialise the whole cache to hold sdram at "addr" as direct mapped */
+
+    for (set = 0; set < E4_NumCacheSets; set++)
+       for (line = 0; line < E4_NumCacheLines; line++)
+           write_tag (dev, Tags[set][line], addr | (set << 13) | (1 << 11));
+
+    read_tag (dev, Tags[set][line]);   /* read it back to guarantee the memory system is quite again */
+    mb();
+}
+
+static __inline__ int
+sdram_GreyToBinary(int GreyVal, int NoOfBits)
+{
+    int Bit;
+    int BinaryVal=0;
+    for (Bit=(1 << (NoOfBits-1)); Bit != 0; Bit >>= 1)
+       BinaryVal ^= (GreyVal & Bit) ^ ((BinaryVal >> 1) & Bit);
+    return (BinaryVal);
+}
+
+static __inline__ int
+sdram_BinaryToGrey(int BinaryVal)
+{
+    return (BinaryVal ^ (BinaryVal >> 1));
+}
+
+void
+elan4_sdram_setup_delay_lines (ELAN4_DEV *dev)
+{
+    /* This is used to fix the SDRAM delay line values */
+    int i, AutoGenDelayValue=0;
+    int NewDelayValue;
+
+    if (dev->dev_sdram_cfg & SDRAM_FIXED_DELAY_ENABLE)   /* already setup. */
+       return;
+
+    /* now get an average of 10 dll values */
+    for (i=0;i<10;i++)
+       AutoGenDelayValue += sdram_GreyToBinary(SDRAM_GET_DLL_DELAY(read_reg64 (dev, SDRamConfigReg)),
+                                              SDRAM_FIXED_DLL_DELAY_BITS);
+
+    NewDelayValue = SDRAM_DLL_CORRECTION_FACTOR + (AutoGenDelayValue / 10); /* Mean of 10 values */
+
+    dev->dev_sdram_cfg = (dev->dev_sdram_cfg & ~(SDRAM_FIXED_DLL_DELAY_MASK << SDRAM_FIXED_DLL_DELAY_SHIFT)) |
+                         SDRAM_FIXED_DELAY_ENABLE | SDRAM_FIXED_DLL_DELAY(sdram_BinaryToGrey(NewDelayValue));
+
+    write_reg64 (dev, SDRamConfigReg, dev->dev_sdram_cfg);     /* Put back the new value */
+
+    pioflush_reg (dev);
+}
+
+static int
+elan4_sdram_probe_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank)
+{
+    unsigned long      mappedsize = bank->b_size;
+    ioaddr_t           ioaddr;
+    unsigned long long value, size;
+    register int       i;
+    extern int         sdram_bank_limit;
+
+    if (mappedsize > SDRAM_MAX_BLOCK_SIZE)
+       mappedsize = SDRAM_MAX_BLOCK_SIZE;
+
+    while ((ioaddr = elan4_map_device (dev, ELAN4_BAR_SDRAM, bank->b_base, mappedsize, &bank->b_handle)) == 0)
+    {
+       if (mappedsize <= (64*1024*1024))                       /* boards normally populated with 64mb, so winge if we can't see this much */
+           printk ("elan%d: could not map bank %d size %dMb\n", dev->dev_instance, (int)(bank - dev->dev_sdram_banks), (int)mappedsize/(1024*1024));
+
+       if ((mappedsize >>= 1) < (1024*1024))
+           return 0;
+    }
+
+    /* first probe to see if the memory bank is present */
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+       initialise_cache_tags (dev, E4_CacheSize);
+
+    for (i = 0; i < 64; i++)
+    {
+       unsigned long long pattern = (1ull << i);
+
+       writeq (pattern, ioaddr);                                       /* write pattern at base  */
+
+        if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+           initialise_cache_tags (dev, 0);
+
+       writeq (~pattern, ioaddr + E4_CacheSize);                       /* write ~pattern at cachesize */
+
+        if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+          initialise_cache_tags (dev, E4_CacheSize);
+       
+       writeq (~pattern, ioaddr + 2*E4_CacheSize);                     /* write ~pattern at 2*cachesize */
+        if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+           initialise_cache_tags (dev, 2*E4_CacheSize);
+       
+       value = readq (ioaddr);                                 /* read pattern back at 0 */
+       
+       if (value != pattern)
+       {
+           printk ("elan%d: sdram bank %d not present\n", dev->dev_instance, (int) (bank - dev->dev_sdram_banks));
+           elan4_unmap_device (dev, ioaddr, mappedsize, &bank->b_handle);
+           return 0;
+       }
+    }
+    
+    /* sdram bank is present, so work out it's size.  We store the maximum size at the base
+     * and then store the address at each address on every power of two address until
+     * we reach the minimum mappable size (PAGESIZE), we then read back the value at the
+     * base to determine the bank size */
+    writeq (mappedsize, ioaddr);
+    if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+        initialise_cache_tags (dev, 0);
+
+    for (size = mappedsize >> 1; size > PAGE_SIZE; size >>= 1)
+    {
+       writeq (size, ioaddr + size);
+        if (dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA)
+           initialise_cache_tags (dev, size);
+    }
+
+    if ((size = readq (ioaddr)) < SDRAM_MIN_BANK_SIZE)
+    {
+       printk ("elan%d: memory bank %d dubious\n", dev->dev_instance, (int) (bank - dev->dev_sdram_banks));
+       elan4_unmap_device (dev, ioaddr, mappedsize, &bank->b_handle);
+       return 0;
+    }
+
+    if (sdram_bank_limit == 0 || size <= (sdram_bank_limit * 1024 * 1024))
+       printk ("elan%d: memory bank %d is %d Mb\n", dev->dev_instance, (int) (bank - dev->dev_sdram_banks), (int) (size / (1024*1024)));
+    else
+    {
+       size = (sdram_bank_limit * 1024 * 1024);
+       printk ("elan%d: limit bank %d to %d Mb\n", dev->dev_instance, (int) (bank - dev->dev_sdram_banks), (int) (size / (1024*1024)));
+    }
+
+    bank->b_size = size;
+
+    elan4_unmap_device (dev, ioaddr, mappedsize, &bank->b_handle);
+    return 1;
+}
+#endif
+
+int
+elan4_sdram_init_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank)
+{
+    int indx, size;
+
+    bank->b_ioaddr = 0;
+
+    if (! elan4_sdram_probe_bank (dev, bank))
+       return 0;
+
+    if ((bank->b_ioaddr = elan4_map_device (dev, ELAN4_BAR_SDRAM, bank->b_base, bank->b_size, &bank->b_handle)) == (ioaddr_t) 0)
+    {
+       printk ("elan%d: could not map sdrambank %d\n", dev->dev_instance, (int) (bank - dev->dev_sdram_banks));
+       return 0;
+    }
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size <= bank->b_size; indx++, size <<= 1) /* allocate the buddy allocator bitmaps */
+       KMEM_ZALLOC (bank->b_bitmaps[indx], bitmap_t *, sizeof (bitmap_t) * BT_BITOUL(bank->b_size/size), 1);
+    
+    return 1;
+}
+
+void
+elan4_sdram_fini_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank)
+{
+    int indx, size;
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size <= bank->b_size; indx++, size <<= 1)
+       KMEM_FREE (bank->b_bitmaps[indx], sizeof (bitmap_t) * BT_BITOUL(bank->b_size/size));
+    
+    elan4_unmap_device (dev, bank->b_ioaddr, bank->b_size, &bank->b_handle);
+}
+
+void
+elan4_sdram_add_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank)
+{
+    sdramaddr_t base = bank->b_base;
+    sdramaddr_t top  = bank->b_base + bank->b_size;
+    register int indx;
+    register unsigned long size;
+
+    /* align to the minimum block size */
+    base = (base + SDRAM_MIN_BLOCK_SIZE - 1) & ~((sdramaddr_t) SDRAM_MIN_BLOCK_SIZE-1);
+    top &= ~((sdramaddr_t) SDRAM_MIN_BLOCK_SIZE-1);
+
+    /* don't allow 0 as a valid "base" */
+    if (base == 0)
+       base = SDRAM_MIN_BLOCK_SIZE;
+
+    /* carve the bottom to the biggest boundary */
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; indx < SDRAM_NUM_FREE_LISTS; indx++, size <<= 1)
+    {
+       if ((base & size) == 0)
+           continue;
+
+       if ((base + size) > top)
+           break;
+
+       free_block (dev, base, indx);
+       
+       base += size;
+    }
+
+    /* carve the top down to the biggest boundary */
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; indx < SDRAM_NUM_FREE_LISTS; indx++, size <<= 1)
+    {
+       if ((top & size) == 0)
+           continue;
+
+       if ((top - size) < base)
+           break;
+
+       free_block (dev, (top - size), indx);
+       
+       top -= size;
+    }
+
+    /* now free of the space in between */
+    while (base < top)
+    {
+       free_block (dev, base, (SDRAM_NUM_FREE_LISTS-1));
+
+       base += SDRAM_MAX_BLOCK_SIZE;
+    }
+}
+
+sdramaddr_t
+elan4_sdram_alloc (ELAN4_DEV *dev, int nbytes)
+{
+    sdramaddr_t block;
+    register int i, indx;
+    unsigned long size;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_sdram_lock, flags);
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size < nbytes; indx++, size <<= 1)
+       ;
+
+    PRINTF2 (DBG_DEVICE, DBG_SDRAM, "elan4_sdram_alloc: nbytes=%d indx=%d\n", nbytes, indx);
+
+    /* need to split a bigger block up */
+    for (i = indx; i < SDRAM_NUM_FREE_LISTS; i++, size <<= 1)
+       if (dev->dev_sdram_freelists[i])
+           break;
+    
+    if (i == SDRAM_NUM_FREE_LISTS)
+    {
+       spin_unlock_irqrestore (&dev->dev_sdram_lock, flags);
+       printk ("elan4_sdram_alloc: %d bytes failed\n", nbytes);
+       return ((sdramaddr_t) 0);
+    }
+    
+    PRINTF2 (DBG_DEVICE, DBG_SDRAM, "elan4_sdram_alloc: use block=%x indx=%d\n", dev->dev_sdram_freelists[i], i);
+
+    /* remove the block from the free list */
+    freelist_removehead (dev, i, (block = dev->dev_sdram_freelists[i]));
+
+    /* clear the approriate bit in the bitmap */
+    BT_CLEAR (sdramaddr_to_bank (dev, block)->b_bitmaps[i], sdramaddr_to_bit (dev,i, block));
+
+    /* and split it up as required */
+    while (i-- > indx)
+       free_block (dev, block + (size >>= 1), i);
+
+    spin_unlock_irqrestore (&dev->dev_sdram_lock, flags);
+
+    ASSERT ((block & ((SDRAM_MIN_BLOCK_SIZE << (indx))-1)) == 0);
+
+#ifdef CONFIG_MPSAS
+    elan4_sdram_zeroq_sdram (dev, block, sizeof (sdramblock_t));
+#endif
+
+    return ((sdramaddr_t) block);
+}
+
+void
+elan4_sdram_free (ELAN4_DEV *dev, sdramaddr_t block, int nbytes)
+{
+    register int indx;
+    unsigned long size;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->dev_sdram_lock, flags);
+
+    for (indx = 0, size = SDRAM_MIN_BLOCK_SIZE; size < nbytes; indx++, size <<= 1)
+       ;
+
+    PRINTF2 (DBG_DEVICE, DBG_SDRAM, "elan4_sdram_free: indx=%d block=%x\n", indx, block);
+
+    free_block (dev, block, indx);
+
+    spin_unlock_irqrestore (&dev->dev_sdram_lock, flags);
+}
+
+void
+elan4_sdram_flushcache (ELAN4_DEV *dev, sdramaddr_t addr, int len)
+{
+    int set, off;
+
+    SET_SYSCONTROL (dev, dev_direct_map_pci_writes, CONT_DIRECT_MAP_PCI_WRITES);
+
+    /*
+     * if flushing more than a single set (8K), then you have to flush the whole cache.
+     *   NOTE - in the real world we will probably want to generate a burst across
+     *          the pci bus.
+     */
+    if (len >= E4_CacheSetSize)
+    {
+       PRINTF3 (DBG_DEVICE, DBG_SDRAM, "elan4_sdram_flushcache: addr=%x len=%x (%x) => whole cache\n", addr, len, addr + len);
+
+#ifdef CONFIG_MPSAS
+       elan4_sdram_zeroq_sdram (dev, dev->dev_cacheflush_space, E4_CacheSize);
+#else
+       for (set = 0; set < E4_NumCacheSets; set++)
+           for (off = 0; off < E4_CacheSetSize; off += E4_CacheLineSize)
+               elan4_sdram_writeq (dev, dev->dev_cacheflush_space + (set * E4_CacheSetSize) + off, 0);
+#endif
+    }
+    else
+    {
+       unsigned base    = addr & ~(E4_CACHELINE_SIZE-1);
+       unsigned top     = (addr + len + (E4_CACHELINE_SIZE-1)) & ~(E4_CACHELINE_SIZE-1);
+       unsigned baseoff = base & (E4_CacheSetSize-1);
+       unsigned topoff  = top  & (E4_CacheSetSize-1);
+
+       if ((base ^ top) & E4_CacheSetSize)                     /* wraps */
+       {
+           PRINTF7 (DBG_DEVICE, DBG_SDRAM, "elan4_sdram_flushcache: addr=%x len=%x (%x) => split cache (%x,%x %x,%x)\n", 
+                    addr, len, addr + len, 0, topoff, baseoff, E4_CacheSetSize);
+
+#ifdef CONFIG_MPSAS
+           for (set = 0; set < E4_NumCacheSets; set++)
+           {
+               elan4_sdram_zeroq_sdram (dev, dev->dev_cacheflush_space + (set * E4_CacheSetSize), topoff);
+               elan4_sdram_zeroq_sdram (dev, dev->dev_cacheflush_space + (set * E4_CacheSetSize) + baseoff, E4_CacheSetSize - baseoff);
+           }
+#else
+           for (set = 0; set < E4_NumCacheSets; set++)
+           {
+               for (off = 0; off < (top & (E4_CacheSetSize-1)); off += E4_CACHELINE_SIZE)
+                   elan4_sdram_writeq (dev, dev->dev_cacheflush_space + (set * E4_CacheSetSize) + off, 0);
+               
+               for (off = (base & (E4_CacheSetSize-1)); off < E4_CacheSetSize; off += E4_CACHELINE_SIZE)
+                   elan4_sdram_writeq (dev, dev->dev_cacheflush_space + (set * E4_CacheSetSize) + off, 0);
+           }
+#endif
+       }
+       else
+       {
+           PRINTF5 (DBG_DEVICE, DBG_SDRAM, "elan4_sdram_flushcache: addr=%x len=%x (%x) => part cache (%x,%x)\n", 
+                    addr, len, addr + len, baseoff, topoff);
+
+#ifdef CONFIG_MPSAS
+           for (set = 0; set < E4_NumCacheSets; set++)
+               elan4_sdram_zeroq_sdram (dev, dev->dev_cacheflush_space + (set * E4_CacheSetSize) + baseoff, topoff - baseoff);
+#else
+           for (set = 0; set < E4_NumCacheSets; set++)
+               for (off = (base & (E4_CacheSetSize-1)); off < (top & (E4_CacheSetSize-1)); off += E4_CACHELINE_SIZE)
+                   elan4_sdram_writeq (dev, dev->dev_cacheflush_space + (set * E4_CacheSetSize) + off, 0);
+#endif
+       }
+    }
+    pioflush_sdram (dev);
+    
+    CLEAR_SYSCONTROL (dev, dev_direct_map_pci_writes, CONT_DIRECT_MAP_PCI_WRITES);
+}
+
+static char *
+get_correctableErr_bitpos(uint SyndromeBits)
+{
+    switch (SyndromeBits)
+    {
+    case 0x00: return ("NoErr");
+    case 0x31: return ("00"); 
+    case 0x32: return ("01"); 
+    case 0xc4: return ("02"); 
+    case 0xc8: return ("03"); 
+    case 0x26: return ("04"); 
+    case 0x91: return ("05"); 
+    case 0x89: return ("06"); 
+    case 0x64: return ("07"); 
+    case 0xc1: return ("08"); 
+    case 0xf2: return ("09"); 
+    case 0x34: return ("10"); 
+    case 0xf8: return ("11"); 
+    case 0xf1: return ("12"); 
+    case 0xc2: return ("13"); 
+    case 0xf4: return ("14"); 
+    case 0x38: return ("15"); 
+    case 0xd6: return ("16"); 
+    case 0xa1: return ("17"); 
+    case 0x79: return ("18"); 
+    case 0xa4: return ("19"); 
+    case 0xd9: return ("20"); 
+    case 0xa2: return ("21"); 
+    case 0x76: return ("22"); 
+    case 0xa8: return ("23"); 
+    case 0xe6: return ("24"); 
+    case 0x51: return ("25"); 
+    case 0xb9: return ("26"); 
+    case 0x54: return ("27"); 
+    case 0xe9: return ("28"); 
+    case 0x52: return ("29"); 
+    case 0xb6: return ("30"); 
+    case 0x58: return ("31"); 
+    case 0x13: return ("32"); 
+    case 0x23: return ("33"); 
+    case 0x4c: return ("34"); 
+    case 0x8c: return ("35"); 
+    case 0x62: return ("36"); 
+    case 0x19: return ("37"); 
+    case 0x98: return ("38"); 
+    case 0x46: return ("39"); 
+    case 0x1c: return ("40"); 
+    case 0x2f: return ("41"); 
+    case 0x43: return ("42"); 
+    case 0x8f: return ("43"); 
+    case 0x1f: return ("44"); 
+    case 0x2c: return ("45"); 
+    case 0x4f: return ("46"); 
+    case 0x83: return ("47"); 
+    case 0x6d: return ("48"); 
+    case 0x1a: return ("49"); 
+    case 0x97: return ("50"); 
+    case 0x4a: return ("51"); 
+    case 0x9d: return ("52"); 
+    case 0x2a: return ("53"); 
+    case 0x67: return ("54"); 
+    case 0x8a: return ("55"); 
+    case 0x6e: return ("56"); 
+    case 0x15: return ("57"); 
+    case 0x9b: return ("58"); 
+    case 0x45: return ("59"); 
+    case 0x9e: return ("60"); 
+    case 0x25: return ("61"); 
+    case 0x6b: return ("62"); 
+    case 0x85: return ("63"); 
+    case 0x01: return ("C0"); 
+    case 0x02: return ("C1"); 
+    case 0x04: return ("C2"); 
+    case 0x08: return ("C3"); 
+    case 0x10: return ("C4"); 
+    case 0x20: return ("C5"); 
+    case 0x40: return ("C6"); 
+    case 0x80: return ("C7"); 
+
+    case 0x07: case 0x0b: case 0x0d: case 0x0e: case 0x3d: case 0x3e: case 0x70: case 0x7c: // T  
+    case 0xb0: case 0xbc: case 0xc7: case 0xcb: case 0xd0: case 0xd3: case 0xe0: case 0xe3: // T  
+       return ("triple");
+
+    case 0x0f: case 0x55: case 0x5a: case 0xa5: case 0xaa: case 0xf0: case 0xff: // Q  
+       return ("quadruple");
+
+    case 0x16: case 0x29: case 0x37: case 0x3b: case 0x49: case 0x57: case 0x5b: case 0x5d: case 0x5e: case 0x61: // M  
+    case 0x68: case 0x73: case 0x75: case 0x7a: case 0x7f: case 0x86: case 0x92: case 0x94: case 0xa7: case 0xab: // M  
+    case 0xad: case 0xae: case 0xb3: case 0xb5: case 0xba: case 0xbf: case 0xcd: case 0xce: case 0xd5: case 0xda: // M  
+    case 0xdc: case 0xdf: case 0xe5: case 0xea: case 0xec: case 0xef: case 0xf7: case 0xfb: case 0xfd: case 0xfe: // M  
+       return ("multiple");
+
+    default:   // all other cases
+       return ("double");
+    }
+}
+
+char *
+elan4_sdramerr2str (ELAN4_DEV *dev, E4_uint64 status, char *str)
+{
+    E4_uint64 StartupSyndrome    = dev->dev_sdram_initial_ecc_val;
+    int       RisingDQSsyndrome  = ((ECC_RisingDQSSyndrome(status) == ECC_RisingDQSSyndrome(StartupSyndrome)) ?
+                                   0 : ECC_RisingDQSSyndrome(status));
+    int              FallingDQSsyndrome = ((ECC_FallingDQSSyndrome(status) == ECC_FallingDQSSyndrome(StartupSyndrome)) ?
+                                   0 : ECC_FallingDQSSyndrome(status));
+    E4_uint64 Addr = ECC_Addr(status);
+    int       Bank = (Addr >> 6) & 3;
+    int       Cas  = ((Addr >> 3) & 7) | ((Addr >> (8 - 3)) & 0xf8) | ((Addr >> (25 - 8)) & 0x100) |
+                    ((Addr >> (27 - 9)) & 0x200) | ((Addr >> (29 - 10)) & 0xc00);
+    int       Ras  = ((Addr >> 13) & 0xfff) | ((Addr >> (26 - 12)) & 0x1000) | ((Addr >> (28 - 13)) & 0x2000) |
+                    ((Addr >> (30 - 14)) & 0x4000);
+
+    sprintf (str, "Addr=%07llx Bank=%x Ras=%x Cas=%x Falling DQS=%s Rising DQS=%s Syndrome=%x%s%s%s%s",                /* 41 + 16 + 8 + 15 + 24 + 13 + 22 + 10 + 10 == 151 */
+            (long long)Addr, Bank, Ras, Cas,
+            get_correctableErr_bitpos(FallingDQSsyndrome),
+            get_correctableErr_bitpos(RisingDQSsyndrome),
+            (int)ECC_Syndrome(status),
+            ECC_UncorrectableErr(status)   ? " Uncorrectable" : "",
+            ECC_MultUncorrectErrs(status)  ? " Multiple-Uncorrectable" : "",
+            ECC_CorrectableErr(status)     ? " Correctable" : "",
+            ECC_MultCorrectErrs(status)    ? " Multiple-Correctable" : "");
+
+    return str;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/trap.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/trap.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/trap.c 2005-06-01 23:12:54.619436064 -0400
@@ -0,0 +1,778 @@
+/*
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: trap.c,v 1.19.10.2 2004/11/03 14:24:32 duncant Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/trap.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+
+#include <elan4/trtype.h>
+#include <elan4/commands.h>
+
+char * const PermTypes[16] = 
+{
+    "Disabled",       "Unused",          "LocalDataRead", "LocalDataWrite",
+    "LocalRead",      "LocalExecute",    "ReadOnly",      "LocalWrite",
+    "LocalEventOnly", "LocalEventWrite", "RemoteEvent",   "RemoteAll",
+    "RemoteReadOnly", "RemoteWriteOnly", "DataReadWrite", "NoFault",
+};
+
+char * const AccTypes[] =
+{
+    "LocalDataRead ", "LocalDataWrite", "RemoteRead    ", "RemoteWrite   ",
+    "Execute       ", "LocalEvent    ", "Unused        ", "RemoteEvent   "
+};
+char * const DataTypes[] = {"Byte ", "HWord", "Word ", "DWord"};
+char * const PhysTypes[] = {"Special Read", "Special Write", "Physical Read", "Physical Write"};
+    
+char * const EProcTrapNames[] = {
+    "EventProcNoFault",
+    "EventProcAddressAlignment",
+    "EventProcMemoryFault",
+    "EventProcCountWrapError",
+};
+
+char * const CProcTrapNames[] = {
+    "CommandProcNoFault",
+    "CommandProcInserterError",
+    "CommandProcPermissionTrap",
+    "CommandProcSendTransInvalid",
+    "CommandProcSendTransExpected",
+    "CommandProcDmaQueueOverflow",
+    "CommandProcInterruptQueueOverflow",
+    "CommandProcMemoryFault",
+    "CommandProcRouteFetchFault",
+    "CommandProcFailCountZero",
+    "CommandProcAddressAlignment",
+    "CommandProcWaitTrap",
+    "CommandProcMultipleGuards",
+    "CommandProcOpenOnGuardedChan",
+    "CommandProcThreadQueueOverflow",
+    "CommandProcBadData",
+};
+
+char *const CProcInsertError[] = {
+    "No Error",
+    "Overflowed",
+    "Invalid Write Size",
+    "Invalid Write Order",
+};
+
+char * const DProcTrapNames[] = {
+    "DmaProcNoFault",
+    "DmaProcRouteFetchFault",
+    "DmaProcFailCountError",
+    "DmaProcPacketAckError",
+    "DmaProcRunQueueReadFault",
+    "DmaProcQueueOverFlow",
+};
+
+char *const IProcTrapNames[] = {
+    "InputNoFault",
+    "InputAddressAlignment",
+    "InputMemoryFault",
+    "InputInvalidTransType",
+    "InputDmaQueueOverflow",
+    "InputEventEngineTrapped",
+    "InputCrcErrorAfterPAckOk",
+    "InputEopErrorOnWaitForEop",
+    "InputEopErrorTrap",
+    "InputDiscardAfterAckOk",
+};
+
+char *const TProcTrapNames[] = {
+    "HaltThread",
+    "TrapForTooManyInstructions",
+    "InstAccessException",
+    "Unimplemented",
+    "DataAccessException",
+    "DataAlignmentError",
+    "TrapForUsingBadData",
+};
+
+#define declare_spaces(space, str)             char space[64]; do { int i; for (i = 0; i < strlen(str); i++) spaces[i] = ' '; space[i] = '\0'; } while (0)
+#define declare_prefix(space, spaces, str)     char space[64]; do { strcpy (space, spaces); strcat (space, str); } while (0)
+
+void
+elan4_display_farea (void *type, int mode, char *str, E4_FaultSave *farea)
+{
+    E4_uint32 FSR = FaultSaveFSR(farea->FSRAndFaultContext);
+
+    declare_spaces(spaces, str);
+    
+    elan4_debugf (type, mode, "%s Fault occurred at %016llx for context %4x\n", str,
+                 farea->FaultAddress, FaultSaveContext(farea->FSRAndFaultContext));
+    
+    if (FSR & AT_VirtualWriteAccBit)                           /* Virtual write access */
+       elan4_debugf (type, mode, "%s FSR=%x: Virtual Write. DWSize=0x%x EndP=0x%x Access=%s DT=%s\n",
+                     spaces, FSR, FSR & AT_VirtualWriteSizeMask,
+                     (FSR >> AT_VirtualWriteEndPtrShift) & AT_VirtualWriteEndPtrMask,
+                     AccTypes[(FSR >> AT_PermBitsShift) & AT_PermBitsMask],
+                     DataTypes[(FSR >> AT_BlkDataTyShift) & AT_BlkDataTyMask]);
+    else if (FSR & AT_VirtualReadAccBit)                       /* Virtual read access */
+       elan4_debugf (type, mode, "%s FSR=%x: Virtual Read. DWSize=0x%x Access=%s DT=%s\n",
+                     spaces, FSR, FSR & AT_VirtualReadSizeMask,
+                     AccTypes[(FSR >> AT_PermBitsShift) & AT_PermBitsMask],
+                     DataTypes[(FSR >> AT_BlkDataTyShift) & AT_BlkDataTyMask]);
+    else
+       elan4_debugf (type, mode, "%s FSR=%x: %s. Size=0x%x\n", spaces,
+                     FSR, PhysTypes[(FSR >> AT_SelBitsShift) & AT_SelBitsMask],
+                     FSR & AT_OtherSizeMask);
+    elan4_debugf (type, mode, "%s FSR: %s %s%s %sWalking\n", spaces,
+                 (FSR & AT_NonAlloc) ? "NonAlloc" : "Alloc",
+                 (FSR & AT_DmaData) ? "Dma " : "",
+                 (FSR & FSR_WalkForThread) ? "ThreadAcc" : "UnitsAcc",
+                 (FSR & FSR_Walking) ? "" : "Not");
+    PRINTF (type, mode, "%s FSR: %s%sHashTable=%s\n", spaces,
+           (FSR & FSR_NoTranslationsFound) ? "NoTranslationsFound " : "",
+           (FSR & FSR_WalkingProtectionFault) ? "WalkingProtectionFault " : "",
+           (FSR & FSR_HashTable1) ? "1" : "0");
+    if (FSR & (FSR_RouteVProcErr | FSR_FaultForBadData))
+       elan4_debugf (type, mode, "%s FSR: %s%s\n", spaces,
+                     (FSR & FSR_RouteVProcErr) ? "RouteVProcErr " : "",
+                     (FSR & FSR_FaultForBadData) ? "FaultForBadData " : "");
+}
+
+void
+elan4_display_eproc_trap (void *type, int mode, char *str, ELAN4_EPROC_TRAP *trap)
+{
+    declare_spaces (spaces, str);
+
+    elan4_debugf (type, mode, "%s Status=%016llx %s EventAddr=%016llx CountAndType=%016llx\n", str,
+                 trap->tr_status, EProcTrapNames[EPROC_TrapType(trap->tr_status)],
+                 trap->tr_eventaddr, trap->tr_event.ev_CountAndType);
+    elan4_debugf (type, mode, "%s Param=%016llx.%016llx\n", spaces,
+                 trap->tr_event.ev_Params[0], trap->tr_event.ev_Params[1]);
+
+    elan4_display_farea (type, mode, strcat (spaces, EPROC_Port0Fault(trap->tr_status) ? " EPROC0" : " EPROC1"), &trap->tr_faultarea);
+}
+
+void
+elan4_display_cproc_trap (void *type, int mode, char *str, ELAN4_CPROC_TRAP *trap)
+{
+    declare_spaces(spaces, str);
+
+    elan4_debugf (type, mode, "%s Status=%llx %s Command=%llx\n", str, trap->tr_status, 
+                 CProcTrapNames[CPROC_TrapType(trap->tr_status)], trap->tr_command);
+    elan4_debugf (type, mode, "%s Desc=%016llx %016llx %016llx %016llx\n", str,
+                 trap->tr_qdesc.CQ_QueuePtrs, trap->tr_qdesc.CQ_HoldingValue,
+                 trap->tr_qdesc.CQ_AckBuffers, trap->tr_qdesc.CQ_Control);
+
+    switch (CPROC_TrapType (trap->tr_status))
+    {
+    case CommandProcInserterError:
+       elan4_debugf (type, mode, "%s   %s\n", str, CProcInsertError[CQ_RevB_ErrorType(trap->tr_qdesc.CQ_QueuePtrs)]);
+       break;
+
+    case CommandProcWaitTrap:
+       elan4_display_eproc_trap (type, mode, spaces, &trap->tr_eventtrap);
+       break;
+
+    default:
+       elan4_display_farea (type, mode, spaces, &trap->tr_faultarea);
+       break;
+    }
+}
+
+void
+elan4_display_dproc_trap (void *type, int mode, char *str, ELAN4_DPROC_TRAP *trap)
+{
+    declare_spaces (spaces, str);
+
+    elan4_debugf (type, mode, "%s status %llx - %s\n", str,
+                 trap->tr_status, DProcTrapNames[DPROC_TrapType(trap->tr_status)]);
+
+    elan4_debugf (type, mode, "%s DESC %016llx %016llx %016llx %016llx\n", spaces, trap->tr_desc.dma_typeSize, 
+                 trap->tr_desc.dma_cookie, trap->tr_desc.dma_vproc, trap->tr_desc.dma_srcAddr);
+    elan4_debugf (type, mode, "%s      %016llx %016llx %016llx\n", spaces, trap->tr_desc.dma_dstAddr, 
+                 trap->tr_desc.dma_srcEvent, trap->tr_desc.dma_dstEvent);
+
+    if (DPROC_PrefetcherFault (trap->tr_status))
+       elan4_display_farea (type, mode, spaces, &trap->tr_prefetchFault);
+}
+
+void
+elan4_display_tproc_trap (void *type, int mode, char *str, ELAN4_TPROC_TRAP *trap)
+{
+    register int i;
+    declare_spaces (spaces, str);
+
+    elan4_debugf (type, mode, "%s PC=%016llx nPC=%016llx State=%016llx Status=%016llx -%s%s%s%s\n", str,
+                 trap->tr_pc, trap->tr_npc, trap->tr_state, trap->tr_status, 
+                 (trap->tr_state & TS_TrapForTooManyInstructions) ? " TrapForTooManyInstructions" : "",
+                 (trap->tr_state & TS_Unimplemented)              ? " Unimplemented"              : "",
+                 (trap->tr_state & TS_DataAlignmentError)         ? " DataAlignmentError"         : "",
+                 (trap->tr_state & TS_InstAccessException)        ? " InstAccessException"        : "",
+                 (trap->tr_state & TS_DataAccessException)        ? " DataAlignmentError"         : "");
+    
+    for (i = 0; i < 64; i += 4)
+       elan4_debugf (type, mode, "%s r%d - %016llx %016llx %016llx %016llx\n", spaces, i,
+                     trap->tr_regs[i], trap->tr_regs[i+1], trap->tr_regs[i+2], trap->tr_regs[i+3]);
+    
+    if (trap->tr_state & TS_InstAccessException)
+    {
+       declare_prefix (prefix, spaces, "Inst");
+
+       elan4_display_farea (type, mode, prefix, &trap->tr_instFault);
+    }
+
+    if (trap->tr_state & TS_DataAccessException)
+    {
+       declare_prefix (prefix, spaces, "Data");
+       elan4_display_farea (type, mode, prefix, &trap->tr_dataFault);
+    }
+}
+
+void
+elan4_display_iproc_trap (void *type, int mode, char *str, ELAN4_IPROC_TRAP *trap)
+{
+    register int i;
+    declare_spaces (spaces, str);
+
+    for (i = 0; i < trap->tr_numTransactions; i++)
+    {
+       E4_IprocTrapHeader *hdrp    = &trap->tr_transactions[i];
+       E4_uint64           status  = hdrp->IProcStatusCntxAndTrType;
+       E4_Addr             addr    = hdrp->TrAddr;
+       char               *typeString;
+       char                buffer[256];
+       char               *ptr = buffer;
+       
+       if (IPROC_EOPTrap(status))
+       {
+           switch (IPROC_EOPType(status))
+           {
+           case EOP_GOOD:        typeString = "EopGood";   break;
+           case EOP_BADACK:      typeString = "EopBadAck"; break;
+           case EOP_ERROR_RESET: typeString = "EopReset";  break;
+           default:              typeString = "EopBad";    break;
+           }
+           
+           ptr += sprintf (ptr, "%15s Cntx=%-6d", typeString, IPROC_NetworkContext(status));
+       }
+       else
+       {
+           if (IPROC_BadLength(status))
+               typeString = "BadLength";
+           else if (IPROC_TransCRCStatus(status) == CRC_STATUS_DISCARD)
+               typeString = "DiscardCrc";
+           else if (IPROC_TransCRCStatus(status) == CRC_STATUS_ERROR)
+               typeString = "ErrorCrc Remote Network error";
+           else if (IPROC_TransCRCStatus(status) == CRC_STATUS_BAD)
+               typeString = "BadCrc Cable error into this node.";
+           else
+           {
+               if ((IPROC_TransactionType(status) & TR_BLOCK_OPCODE_MASK) == TR_WRITEBLOCK)
+                   typeString = "WriteBlock";
+               else
+               {
+                   switch (IPROC_TransactionType(status) & TR_OPCODE_MASK)
+                   {
+                   case TR_SETEVENT_IDENTIFY & TR_OPCODE_MASK: typeString = "SetEvent";        break;
+                   case TR_REMOTEDMA & TR_OPCODE_MASK:         typeString = "RemoteDma";       break;
+                   case TR_SENDDISCARD & TR_OPCODE_MASK:       typeString = "SendDiscard";     break;
+                   case TR_GTE & TR_OPCODE_MASK:               typeString = "GTE";             break;
+                   case TR_LT & TR_OPCODE_MASK:                typeString = "LT";              break;
+                   case TR_EQ & TR_OPCODE_MASK:                typeString = "EQ";              break;
+                   case TR_NEQ & TR_OPCODE_MASK:               typeString = "NEQ";             break;
+                   case TR_IDENTIFY & TR_OPCODE_MASK:          typeString = "Idenfity";        break;
+                   case TR_ADDWORD & TR_OPCODE_MASK:           typeString = "AddWord";         break;
+                   case TR_INPUT_Q_COMMIT & TR_OPCODE_MASK:    typeString = "InputQCommit";    break;
+                   case TR_TESTANDWRITE & TR_OPCODE_MASK:      typeString = "TestAndWrite";    break;
+                   case TR_INPUT_Q_GETINDEX & TR_OPCODE_MASK:  typeString = "InputQGetIndex";  break;
+                   case TR_TRACEROUTE_TRANS & TR_OPCODE_MASK:  typeString = "TraceRoute";      break;
+                   default:                                    typeString = "Unknown";         break;
+                   }
+               }
+           }
+
+           ptr += sprintf (ptr, "%15s Cntx=%-6d Addr=%016llx", typeString, IPROC_NetworkContext(status), (unsigned long long) addr);
+       }
+       
+       
+       if (IPROC_TrapValue(status) != InputNoFault)
+       {
+           ptr += sprintf (ptr, " TrType=%2d ChanTrapped=%x GoodAck=%x BadAck=%x InputterChan=%d", IPROC_TrapValue(status),
+                           IPROC_ChannelTrapped(status), IPROC_GoodAckSent(status), IPROC_BadAckSent(status),
+                           IPROC_InputterChan(status));
+           if (IPROC_EOPTrap(status))
+               ptr += sprintf (ptr, " EOPType=%d", IPROC_EOPType(status));
+           else
+               ptr += sprintf (ptr, " %s%s%s%s", 
+                               IPROC_FirstTrans(status) ? " FirstTrans" : "",
+                               IPROC_LastTrans(status) ? " LastTrans" : "",
+                               (IPROC_TransactionType(status) & TR_WAIT_FOR_EOP) ? " WaitForEop" : "",
+                               (IPROC_GoodAckSent(status) &  (1 << IPROC_Channel(status))) ? " AckSent" : "");
+       }
+       
+       elan4_debugf (type, mode, "%s %s\n", str, buffer);
+
+       str = spaces;
+    }
+
+    elan4_display_farea (type, mode, spaces, &trap->tr_faultarea);
+}
+
+#define elan4_sdram_copy_faultarea(dev, unit, farea) \
+    elan4_sdram_copyq_from_sdram ((dev), (dev)->dev_faultarea + (unit) * sizeof (E4_FaultSave), (E4_uint64 *) farea, sizeof (E4_FaultSave));
+
+void
+elan4_extract_eproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_EPROC_TRAP *trap, int iswaitevent)
+{
+    /* only one of the memory ports can fault at a time */
+    ASSERT (EPROC_TrapType(status) != EventProcMemoryFault || (EPROC_Port0Fault(status) ^ EPROC_Port1Fault(status)) == 1);
+
+    trap->tr_status = status;
+    
+    if (EPROC_Port0Fault(status))
+       elan4_sdram_copy_faultarea (dev, CUN_EventProc0, &trap->tr_faultarea);
+    if (EPROC_Port1Fault(status))
+       elan4_sdram_copy_faultarea (dev, CUN_EventProc1, &trap->tr_faultarea);
+
+    if (iswaitevent)
+    {
+       /*
+        * for waitevents the Event address is always taken from the command processor
+        * 
+        * if we trapped during the copy then we take the "Event" from the event processor
+        * since we need to complete the copy.  Otherwise we'll be reissuing the original
+        * command again
+        */
+       E4_uint32 fsr = FaultSaveFSR(trap->tr_faultarea.FSRAndFaultContext);
+
+       trap->tr_eventaddr = read_reg64 (dev, CommandHold) ^ WAIT_EVENT_CMD;
+
+       if (EPROC_TrapType(trap->tr_status) == EventProcMemoryFault && 
+           (AT_Perm(fsr) == AT_PermLocalDataRead || AT_Perm(fsr) == AT_PermLocalDataWrite))
+       {
+           trap->tr_event.ev_CountAndType = read_reg64 (dev, EventCountAndType);
+           trap->tr_event.ev_Params[0]    = read_reg64 (dev, EventParameters[0]);
+           trap->tr_event.ev_Params[1]    = read_reg64 (dev, EventParameters[1]);
+       }
+       else
+       {
+           trap->tr_event.ev_Params[0]    = read_reg64 (dev, CommandCopy[5]);
+           trap->tr_event.ev_CountAndType = read_reg64 (dev, CommandCopy[4]);
+           trap->tr_event.ev_Params[1]    = read_reg64 (dev, CommandCopy[6]);
+
+       }
+    }
+    else
+    {
+       trap->tr_eventaddr             = read_reg64 (dev, EventAddress);
+       trap->tr_event.ev_CountAndType = read_reg64 (dev, EventCountAndType);
+       trap->tr_event.ev_Params[0]    = read_reg64 (dev, EventParameters[0]);
+       trap->tr_event.ev_Params[1]    = read_reg64 (dev, EventParameters[1]);
+    }
+
+    BumpDevStat (dev, s_eproc_trap_types[EPROC_TrapType(status)]);
+}
+
+int 
+cproc_open_extract_vp (ELAN4_DEV *dev, ELAN4_CQ *cq)
+{
+       /* cq = ucq->ucq_cq */
+       if ((cq->cq_perm & CQ_STENEnableBit) != 0)
+       {
+            sdramaddr_t   cqdesc       = dev->dev_cqaddr + (elan4_cq2num(cq) * sizeof (E4_CommandQueueDesc));
+           E4_uint64     queuePtrs    = elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs));
+           sdramaddr_t   insertPtr    = (queuePtrs & CQ_PtrMask);
+           sdramaddr_t   commandPtr   = CQ_CompletedPtr (queuePtrs);
+           unsigned int  cqSize       = CQ_Size ((queuePtrs >> CQ_SizeShift) & CQ_SizeMask);
+           E4_uint64     openCommand  = 0;
+
+           if (dev->dev_devinfo.dev_revision_id != PCI_REVISION_ID_ELAN4_REVA && (queuePtrs & CQ_RevB_ReorderingQueue))
+           {
+               E4_uint32 oooMask = elan4_sdram_readl (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_HoldingValue));
+
+               for (; (oooMask & 1) != 0; oooMask >>= 1)
+                   insertPtr = (insertPtr & ~(cqSize-1)) | ((insertPtr + sizeof (E4_uint64)) & (cqSize-1));
+           }
+
+           while (commandPtr != insertPtr)
+           {
+               E4_uint64    command = elan4_sdram_readq (dev, commandPtr);
+               unsigned int cmdSize;
+
+                switch (__categorise_command (command, &cmdSize))
+               {
+               case 0:
+                   (void) __whole_command (&commandPtr, insertPtr, cqSize, cmdSize);
+                   break;
+
+               case 1: /* open */
+                   return (command >> 32);
+                           
+                   break; /* Not reached */
+
+               case 2:
+                   if (openCommand == 0)
+                       (void) __whole_command (&commandPtr, insertPtr, cqSize, cmdSize);
+                   /* Else we should have stopped by now */
+                   else ASSERT(1==2);
+               case 3:
+                   printk ("cproc_open_extract_vp: invalid command %llx\n", command);
+                   return -1;
+               }
+           } /* while */
+       }
+
+       return -1;
+}
+
+void
+elan4_extract_cproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_CPROC_TRAP *trap, unsigned cqnum)
+{
+    /* extract the state from the device */
+    elan4_sdram_copy_faultarea (dev, CUN_CommandProc, &trap->tr_faultarea);
+
+    trap->tr_status  = status;
+    trap->tr_command = read_reg64 (dev, CommandHold);
+    
+    elan4_sdram_copyq_from_sdram (dev, dev->dev_cqaddr + (cqnum * sizeof (E4_CommandQueueDesc)), &trap->tr_qdesc, sizeof (E4_CommandQueueDesc));
+
+    if (CPROC_TrapType (status) == CommandProcWaitTrap)
+       elan4_extract_eproc_trap (dev, read_reg64 (dev, EProcStatus), &trap->tr_eventtrap, 1);
+
+    BumpDevStat (dev, s_cproc_trap_types[CPROC_TrapType(status)]);
+
+    if (PackValue(trap->tr_qdesc.CQ_AckBuffers, 0) == PackTimeout || PackValue(trap->tr_qdesc.CQ_AckBuffers, 1) == PackTimeout)
+       BumpDevStat (dev, s_cproc_timeout);
+}
+
+void
+elan4_extract_dproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_DPROC_TRAP *trap, unsigned unit)
+{
+    trap->tr_status = status;
+    
+    if (unit == 0)
+    {
+       trap->tr_desc.dma_typeSize   = read_reg64 (dev, Dma0Desc.dma_typeSize);
+       trap->tr_desc.dma_cookie     = read_reg64 (dev, Dma0Desc.dma_cookie);
+       trap->tr_desc.dma_vproc      = read_reg64 (dev, Dma0Desc.dma_vproc);
+       trap->tr_desc.dma_srcAddr    = read_reg64 (dev, Dma0Desc.dma_srcAddr);
+       trap->tr_desc.dma_dstAddr    = read_reg64 (dev, Dma0Desc.dma_dstAddr);
+       trap->tr_desc.dma_srcEvent   = read_reg64 (dev, Dma0Desc.dma_srcEvent);
+       trap->tr_desc.dma_dstEvent   = read_reg64 (dev, Dma0Desc.dma_dstEvent);
+       
+       elan4_sdram_copy_faultarea (dev, CUN_DProcPA0, &trap->tr_packAssemFault);
+    }
+    else
+    {
+       trap->tr_desc.dma_typeSize   = read_reg64 (dev, Dma1Desc.dma_typeSize);
+       trap->tr_desc.dma_cookie     = read_reg64 (dev, Dma1Desc.dma_cookie);
+       trap->tr_desc.dma_vproc      = read_reg64 (dev, Dma1Desc.dma_vproc);
+       trap->tr_desc.dma_srcAddr    = read_reg64 (dev, Dma1Desc.dma_srcAddr);
+       trap->tr_desc.dma_dstAddr    = read_reg64 (dev, Dma1Desc.dma_dstAddr);
+       trap->tr_desc.dma_srcEvent   = read_reg64 (dev, Dma1Desc.dma_srcEvent);
+       trap->tr_desc.dma_dstEvent   = read_reg64 (dev, Dma1Desc.dma_dstEvent);
+       
+       elan4_sdram_copy_faultarea (dev, CUN_DProcPA1, &trap->tr_packAssemFault);
+    }
+    
+    if (DPROC_PrefetcherFault (trap->tr_status))
+       elan4_sdram_copy_faultarea (dev, (CUN_DProcData0 | DPROC_FaultUnitNo(trap->tr_status)), &trap->tr_prefetchFault);
+
+    if (DPROC_PacketTimeout (trap->tr_status))
+       BumpDevStat (dev, s_dproc_timeout);
+
+    BumpDevStat (dev, s_dproc_trap_types[DPROC_TrapType(status)]);
+}    
+
+void
+elan4_extract_tproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_TPROC_TRAP *trap)
+{
+    int i;
+
+    trap->tr_status = status;
+    trap->tr_state  = read_reg64 (dev, Thread_Trap_State);
+    trap->tr_pc     = read_reg64 (dev, PC_W);
+    trap->tr_npc    = read_reg64 (dev, nPC_W);
+    trap->tr_dirty  = read_reg64 (dev, DirtyBits);
+    trap->tr_bad    = read_reg64 (dev, BadBits);
+
+#ifdef CONFIG_MPSAS
+    if (sas_copyfrom_dev (dev->dev_osdep.pdev, ELAN4_BAR_REGISTERS, 
+                         ((dev->dev_devinfo.dev_revision_id == PCI_REVISION_ID_ELAN4_REVA) ? ELAN4_REVA_REG_OFFSET : ELAN4_REVB_REG_OFFSET) +
+                         offsetof (E4_Registers, Regs.TProcRegs), (unsigned long) &trap->tr_regs, 64*sizeof (E4_uint64)) < 0)
+    {
+       for (i = 0; i < 64; i++)
+           if (trap->tr_dirty & ((E4_uint64) 1 << i))
+               trap->tr_regs[i] = read_reg64 (dev, TProcRegs[i]);
+    }
+
+    for (i = 0; i < 64; i++)
+       if (! (trap->tr_dirty & ((E4_uint64) 1 << i)))
+           trap->tr_regs[i] = 0xdeadbabedeadbabeULL;
+#else
+    for (i = 0; i < 64; i++)
+    {
+       if (trap->tr_dirty & ((E4_uint64) 1 << i))
+           trap->tr_regs[i] = read_reg64 (dev, TProcRegs[i]);
+       else
+           trap->tr_regs[i] = 0xdeadbabedeadbabeULL;
+    }
+#endif
+    
+    if (trap->tr_state & TS_DataAccessException)
+       elan4_sdram_copy_faultarea (dev, CUN_TProcData0 | TS_DataPortNo (trap->tr_state), &trap->tr_dataFault);
+
+    if (trap->tr_state & TS_InstAccessException)
+       elan4_sdram_copy_faultarea (dev, CUN_TProcInst, &trap->tr_instFault);
+
+    for (i = 0; i < 7; i++)
+       if (trap->tr_state & (1 << i))
+           BumpDevStat (dev, s_tproc_trap_types[i]);
+}
+
+void
+elan4_extract_iproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_IPROC_TRAP *trap, unsigned unit)
+{
+    sdramaddr_t hdroff  = dev->dev_inputtraparea + offsetof (E4_IprocTrapState, TrHeader[0][unit]);
+    sdramaddr_t dataoff = dev->dev_inputtraparea + offsetof (E4_IprocTrapState, TrData[0][unit]);
+    register int i, j;
+    int                  CurrUnitNo    = (unit >= 2) ? CUN_IProcHighPri : CUN_IProcLowPri;
+    sdramaddr_t CurrFaultArea = dev->dev_faultarea + (CurrUnitNo * sizeof (E4_FaultSave));
+
+    /* Finally copy the fault area */
+    elan4_sdram_copy_faultarea (dev, CurrUnitNo, &trap->tr_faultarea);
+
+    /*
+     * Clear out the fault save area after reading to allow a fault on the write of the back pointer of
+     * an InputQCommit to be obsurved if a simultaneous event proc trap occurs.
+     */
+    elan4_sdram_writeq (dev, CurrFaultArea + offsetof(E4_FaultSave, FSRAndFaultContext), 0x0ULL);
+    elan4_sdram_writeq (dev, CurrFaultArea + offsetof(E4_FaultSave, FaultAddress), 0x0ULL);
+
+    /* copy the transaction headers */
+    trap->tr_transactions[0].IProcStatusCntxAndTrType = status;
+    trap->tr_transactions[0].TrAddr                   = elan4_sdram_readq (dev, hdroff + offsetof (E4_IprocTrapHeader, TrAddr));
+    
+    for (i = 0; !IPROC_EOPTrap(trap->tr_transactions[i].IProcStatusCntxAndTrType);)
+    {
+       if (IPROC_BadLength (trap->tr_transactions[i].IProcStatusCntxAndTrType))
+           BumpDevStat (dev, s_bad_length);
+       else if (IPROC_TransCRCStatus (trap->tr_transactions[i].IProcStatusCntxAndTrType) == CRC_STATUS_BAD)
+           BumpDevStat (dev, s_crc_bad);
+       else if (IPROC_TransCRCStatus (trap->tr_transactions[i].IProcStatusCntxAndTrType) == CRC_STATUS_ERROR)
+           BumpDevStat (dev, s_crc_error);
+
+       BumpDevStat (dev, s_iproc_trap_types[IPROC_TrapValue (trap->tr_transactions[i].IProcStatusCntxAndTrType)]);
+
+       hdroff += NO_OF_INPUT_CHANNELS*sizeof (E4_IprocTrapHeader);
+
+       if (++i == MAX_TRAPPED_TRANS)
+           break;
+
+       elan4_sdram_copyq_from_sdram (dev, hdroff, &trap->tr_transactions[i], sizeof (E4_IprocTrapHeader));
+    }
+    
+    if (IPROC_EOPType (trap->tr_transactions[i].IProcStatusCntxAndTrType) == EOP_ERROR_RESET)
+       BumpDevStat (dev, s_eop_reset);
+
+    /* Remember the number of transactions we've copied */
+    trap->tr_numTransactions = i + 1;
+    
+    /* Copy all the data blocks in one go */
+    for (i = 0; i < MIN (trap->tr_numTransactions, MAX_TRAPPED_TRANS); i++, dataoff += NO_OF_INPUT_CHANNELS*sizeof (E4_IprocTrapData))
+    {
+       if (IPROC_BadLength(status) || IPROC_TransCRCStatus (status) != CRC_STATUS_GOOD)
+           elan4_sdram_copyq_from_sdram (dev, dataoff, trap->tr_dataBuffers[i].Data, TRANS_DATA_DWORDS*sizeof(E4_uint64));
+       else
+       {
+           int trtype  = IPROC_TransactionType(trap->tr_transactions[i].IProcStatusCntxAndTrType);
+           int ndwords = (trtype & TR_SIZE_MASK) >> TR_SIZE_SHIFT;
+
+           elan4_sdram_copyq_from_sdram (dev, dataoff, trap->tr_dataBuffers[i].Data, ndwords*sizeof(E4_uint64));
+
+           for (j = ndwords; j < TRANS_DATA_DWORDS; j++)
+               trap->tr_dataBuffers[i].Data[j] = 0xbeec0f212345678ull;
+       }
+    }
+    
+}
+
+void
+elan4_inspect_iproc_trap (ELAN4_IPROC_TRAP *trap)
+{
+    int i;
+
+    trap->tr_flags          = 0;
+    trap->tr_trappedTrans    = TR_TRANS_INVALID;
+    trap->tr_waitForEopTrans = TR_TRANS_INVALID;
+    trap->tr_identifyTrans   = TR_TRANS_INVALID;
+
+    if (trap->tr_numTransactions > MAX_TRAPPED_TRANS)
+       trap->tr_flags = TR_FLAG_TOOMANY_TRANS;
+
+    /*
+     * Now scan all the transactions received 
+     */
+    for (i = 0; i < MIN(trap->tr_numTransactions, MAX_TRAPPED_TRANS) ; i++)
+    {
+       E4_IprocTrapHeader *hdrp   = &trap->tr_transactions[i];
+       E4_uint64           status = hdrp->IProcStatusCntxAndTrType;
+
+       if (trap->tr_identifyTrans == TR_TRANS_INVALID)
+       {
+           switch (IPROC_TransactionType (status) & (TR_OPCODE_MASK | TR_SIZE_MASK))
+           {
+           case TR_IDENTIFY          & (TR_OPCODE_MASK | TR_SIZE_MASK):
+           case TR_REMOTEDMA         & (TR_OPCODE_MASK | TR_SIZE_MASK):
+           case TR_SETEVENT_IDENTIFY & (TR_OPCODE_MASK | TR_SIZE_MASK):
+           case TR_INPUT_Q_COMMIT    & (TR_OPCODE_MASK | TR_SIZE_MASK):
+           case TR_ADDWORD           & (TR_OPCODE_MASK | TR_SIZE_MASK):
+           case TR_TESTANDWRITE      & (TR_OPCODE_MASK | TR_SIZE_MASK):
+               trap->tr_identifyTrans = i;
+               break;
+           }
+       }
+
+       if (IPROC_TrapValue(status) == InputNoFault)            /* We're looking at transactions stored before the trap */
+           continue;                                           /* these should only be identifies */
+       
+       if (trap->tr_trappedTrans == TR_TRANS_INVALID)          /* Remember the transaction which caused the */
+           trap->tr_trappedTrans = i;                          /* trap */
+
+       if (IPROC_GoodAckSent (status) & (1 << IPROC_InputterChan (status)))
+           trap->tr_flags |= TR_FLAG_ACK_SENT;
+           
+       if (IPROC_EOPTrap(status))                              /* Check for EOP */
+       {
+           ASSERT (i == trap->tr_numTransactions - 1);
+
+           switch (IPROC_EOPType(status))
+           {
+           case EOP_GOOD:
+               /* if we get an EOP_GOOD then the outputer should have received a PAckOk. */  
+               /* unless it was a flood, in which case someone must have sent an ack */
+               /* but not necessarily us */
+               break;
+
+           case EOP_BADACK:
+               /* if we get an EOP_BADACK then the outputer did not receive a PAckOk even if
+                * we sent a PAckOk. WFlag this to ignore the AckSent. */
+               trap->tr_flags |= TR_FLAG_EOP_BAD;
+               break;
+
+           case EOP_ERROR_RESET:
+               /* if we get an EOP_ERROR_RESET then the outputer may or may not have got a PAckOk. */
+               trap->tr_flags |= TR_FLAG_EOP_ERROR;
+               break;
+
+           default:
+               printk ("elan4_inspect_iproc_trap: unknown eop type %d", IPROC_EOPType(status));
+               BUG();
+               /* NOTREACHED */
+           }
+           continue;
+       }
+       else
+       {
+           if (IPROC_BadLength(status) || (IPROC_TransCRCStatus (status) == CRC_STATUS_ERROR ||
+                                           IPROC_TransCRCStatus (status) == CRC_STATUS_BAD))
+           {
+               {
+                   register int j;
+                   if (IPROC_BadLength(status))
+                       PRINTF2 (DBG_DEVICE, DBG_INTR, "LinkError: Trapped on bad length data. status=%016llx Address=%016llx\n",
+                                status, hdrp->TrAddr);
+                   else
+                       PRINTF2 (DBG_DEVICE, DBG_INTR, "LinkError: Trapped with bad CRC. status=%016llx Address=%016llx\n",
+                                status, hdrp->TrAddr);
+                   for (j = 0; j < TRANS_DATA_DWORDS; j++)
+                       PRINTF2 (DBG_DEVICE, DBG_INTR, "LinkError: DataBuffers[%d] : %016llx\n", j, trap->tr_dataBuffers[i].Data[j]);
+               }
+
+               trap->tr_flags |= TR_FLAG_BAD_TRANS;
+               continue;
+           }
+           
+           if (IPROC_TransCRCStatus (status) == CRC_STATUS_DISCARD)
+               continue;
+
+           if ((((IPROC_TransactionType(status) & TR_BLOCK_OPCODE_MASK) == TR_WRITEBLOCK) ||
+                (IPROC_TransactionType(status) == TR_TRACEROUTE_TRANS)) &&
+               (trap->tr_flags & TR_FLAG_ACK_SENT) && trap->tr_identifyTrans == TR_TRANS_INVALID)
+           {
+               /* 
+                * Writeblock after the ack is sent without an identify transaction - this is 
+                * considered to be a DMA packet and requires the next packet to be nacked - since 
+                * the DMA processor will send this in a deterministic time and there's an upper 
+                * limit on the network latency (the output timeout) we just need to hold the context 
+                * filter up for a while.
+                */
+               trap->tr_flags |= TR_FLAG_DMA_PACKET;
+           }
+           
+           if (IPROC_LastTrans(status) && (IPROC_TransactionType(status) & TR_WAIT_FOR_EOP))
+           {
+               /*
+                * WaitForEop transactions - if we have to do network error fixup
+                * then we may need to execute/ignore this transaction dependant
+                * on whether the source will be resending it.
+                */
+               trap->tr_waitForEopTrans = i;
+           }
+
+           /*
+            * This is a special case caused by a minor input processor bug.
+            * If simultaneous InputMemoryFault and InputEventEngineTrapped occur then the chip will probably return
+            * InputEventEngineTrapped even though the write of the back pointer has not occured and must be done by
+            * the trap handler.
+            * In this case the fault address will equal q->q_bptr. If there has been only EventEngineTrap then the
+            * the fault address should be zero as the trap handler now always zeros this after every input trap.
+            */
+           if ((IPROC_TransactionType (status) & TR_OPCODE_MASK) == (TR_INPUT_Q_COMMIT & TR_OPCODE_MASK) &&
+               trap->tr_faultarea.FaultAddress == hdrp->TrAddr + offsetof(E4_InputQueue, q_bptr) &&
+               IPROC_TrapValue(status) == InputEventEngineTrapped)
+           {
+               hdrp->IProcStatusCntxAndTrType = (status & 0xFFFFFFF0FFFFFFFFull) | ((E4_uint64) InputMemoryFault << 32);
+           }
+       }
+
+       PRINTF (DBG_DEVICE, DBG_INTR, "inspect[%d] status=%llx TrapValue=%d -> flags %x\n", i, status, IPROC_TrapValue(status), trap->tr_flags);
+    }
+}
+
+E4_uint64
+elan4_trapped_open_command (ELAN4_DEV *dev, ELAN4_CQ *cq)
+{
+    sdramaddr_t cqdesc     = dev->dev_cqaddr + elan4_cq2num(cq) * sizeof (E4_CommandQueueDesc);
+    E4_uint64   cqcontrol  = elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_Control));
+    E4_uint32   extractOff = CQ_ExtractPtr (cqcontrol) & (CQ_Size(cq->cq_size)-1);
+    
+    if (extractOff == 0)
+       extractOff = CQ_Size(cq->cq_size) - sizeof (E4_uint64);
+    else
+       extractOff -= sizeof (E4_uint64);
+
+    return (elan4_sdram_readq (dev, cq->cq_space + extractOff));
+}
+
+EXPORT_SYMBOL(elan4_extract_eproc_trap);
+EXPORT_SYMBOL(elan4_display_eproc_trap);
+EXPORT_SYMBOL(elan4_extract_cproc_trap);
+EXPORT_SYMBOL(elan4_display_cproc_trap);
+EXPORT_SYMBOL(elan4_extract_dproc_trap);
+EXPORT_SYMBOL(elan4_display_dproc_trap);
+EXPORT_SYMBOL(elan4_extract_tproc_trap);
+EXPORT_SYMBOL(elan4_display_tproc_trap);
+EXPORT_SYMBOL(elan4_extract_iproc_trap);
+EXPORT_SYMBOL(elan4_inspect_iproc_trap);
+EXPORT_SYMBOL(elan4_display_iproc_trap);
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/user.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/user.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/user.c 2005-06-01 23:12:54.624435304 -0400
@@ -0,0 +1,3352 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: user.c,v 1.68.2.9 2004/12/20 16:56:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/user.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kpte.h>
+
+#include <elan/elanmod.h>
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/user.h>
+
+#include <elan4/trtype.h>
+#include <elan4/commands.h>
+
+#include <stdarg.h>
+
+/* allow this code to compile against an Eagle elanmod */
+#ifdef __ELANMOD_DEVICE_H
+#define elan_attach_cap(cap,rnum,args,func)    elanmod_attach_cap(cap,args,func)
+#define elan_detach_cap(cap,rnum)              elanmod_detach_cap(cap)
+#endif
+
+#define NETERR_MSGS    16
+
+int user_p2p_route_options   = FIRST_TIMEOUT(3);
+int user_bcast_route_options = FIRST_TIMEOUT(3);
+int user_dproc_retry_count   = 15;
+int user_cproc_retry_count   = 2;
+
+int num_fault_save           = 30;
+int min_fault_pages          = 1;
+int max_fault_pages          = 128;
+
+static int
+user_validate_cap (USER_CTXT *uctx, ELAN_CAPABILITY *cap, unsigned use)
+{
+    /* Don't allow a user process to attach to system context */
+    if (ELAN4_SYSTEM_CONTEXT (cap->cap_lowcontext) || ELAN4_SYSTEM_CONTEXT (cap->cap_highcontext))
+    {
+       PRINTF3 (DBG_DEVICE, DBG_VP,"user_validate_cap: lctx %x hctx %x high %x\n", cap->cap_lowcontext, cap->cap_highcontext, ELAN4_KCOMM_BASE_CONTEXT_NUM);
+       PRINTF0 (DBG_DEVICE, DBG_VP,"user_validate_cap: user process cant attach to system cap\n");
+       return (EINVAL);
+    }
+    
+    return elanmod_classify_cap(&uctx->uctx_position, cap, use);
+}
+
+static __inline__ void
+__user_signal_trap (USER_CTXT *uctx)
+{
+    switch (uctx->uctx_trap_state)
+    {
+    case UCTX_TRAP_IDLE:
+       PRINTF (uctx, DBG_TRAP, "user_signal_trap: deliver signal %d to pid %d\n", uctx->uctx_trap_signo, uctx->uctx_trap_pid);
+
+       if (uctx->uctx_trap_signo)
+           kill_proc (uctx->uctx_trap_pid, uctx->uctx_trap_signo, 1);
+       break;
+
+    case UCTX_TRAP_SLEEPING:
+       PRINTF (uctx, DBG_TRAP, "user_signal_trap: wakeup sleeping trap handler\n");
+
+       kcondvar_wakeupone (&uctx->uctx_wait, &uctx->uctx_spinlock);
+       break;
+    }
+    uctx->uctx_trap_state = UCTX_TRAP_SIGNALLED;
+}
+
+static void
+user_signal_timer (unsigned long arg)
+{
+    USER_CTXT    *uctx = (USER_CTXT *) arg;
+    unsigned long flags;
+
+    PRINTF (uctx, DBG_TRAP, "user_signal_timer: state=%d pid=%d signal=%d (now %d start %d)\n",
+           uctx->uctx_trap_state, uctx->uctx_trap_pid, uctx->uctx_trap_signo, jiffies,
+           uctx->uctx_int_start);
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    __user_signal_trap (uctx);
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+#define MAX_INTS_PER_TICK      50
+#define MIN_INTS_PER_TICK      20
+
+static void
+user_signal_trap (USER_CTXT *uctx)
+{
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_spinlock));
+
+    PRINTF (uctx, DBG_TRAP, "user_signal_trap: state=%d pid=%d signal=%d%s\n", uctx->uctx_trap_state,
+           uctx->uctx_trap_pid, uctx->uctx_trap_signo, timer_pending(&uctx->uctx_int_timer) ? " (timer-pending)" : "");
+
+    uctx->uctx_int_count++;
+
+    if (timer_pending (&uctx->uctx_int_timer))
+       return;
+
+    if (uctx->uctx_int_count > ((int)(jiffies - uctx->uctx_int_start) * MAX_INTS_PER_TICK))
+    {
+       PRINTF (uctx, DBG_TRAP, "user_signal_trap: deferring signal for %d ticks (count %d ticks %d -> %d)\n", 
+               uctx->uctx_int_delay + 1, uctx->uctx_int_count, (int) (jiffies - uctx->uctx_int_start),
+               ((int)(jiffies - uctx->uctx_int_start) * MAX_INTS_PER_TICK));
+
+       /* We're interrupting too fast, so defer this signal */
+       uctx->uctx_int_timer.expires = jiffies + (++uctx->uctx_int_delay);
+
+       add_timer (&uctx->uctx_int_timer);
+    }
+    else
+    {
+       __user_signal_trap (uctx);
+
+       PRINTF (uctx, DBG_TRAP, "user_signal_trap: check signal for %d ticks (count %d ticks %d -> %d)\n", 
+               uctx->uctx_int_delay + 1, uctx->uctx_int_count, (int) (jiffies - uctx->uctx_int_start),
+               (int)(jiffies - uctx->uctx_int_start) * MIN_INTS_PER_TICK);
+           
+       if (uctx->uctx_int_count < ((int) (jiffies - uctx->uctx_int_start)) * MIN_INTS_PER_TICK)
+       {
+           PRINTF (uctx, DBG_TRAP, "user_signal_trap: reset interrupt throttle (count %d ticks %d)\n", 
+                   uctx->uctx_int_count, (int) (jiffies - uctx->uctx_int_start));
+
+           uctx->uctx_int_start = jiffies;
+           uctx->uctx_int_count = 0;
+           uctx->uctx_int_delay = 0;
+       }
+    }
+}
+
+static void
+user_neterr_timer (unsigned long arg)
+{
+    USER_CTXT *uctx = (USER_CTXT *) arg;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    uctx->uctx_status |= UCTX_NETERR_TIMER;
+    
+    user_signal_trap (uctx);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_flush_dma_runqueue (ELAN4_DEV *dev, USER_CTXT *uctx, int qfull)
+{
+    E4_uint64          qptrs = read_reg64 (dev, DProcLowPriPtrs);
+    E4_uint32          qsize = E4_QueueSize (E4_QueueSizeValue (qptrs));
+    E4_uint32          qfptr = E4_QueueFrontPointer (qptrs);
+    E4_uint32          qbptr = E4_QueueBackPointer (qptrs);
+    E4_DProcQueueEntry qentry;
+
+    while ((qfptr != qbptr) || qfull)
+    {
+       E4_uint64 typeSize = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_typeSize));
+
+       if (DMA_Context (typeSize) == uctx->uctx_ctxt.ctxt_num)
+       {
+           elan4_sdram_copyq_from_sdram (dev, qfptr, &qentry, sizeof (E4_DProcQueueEntry));
+
+           PRINTF4 (uctx, DBG_SWAP, "user_flush_dma_runqueue: %016llx %016llx %016llx %016llx\n", qentry.Desc.dma_typeSize, 
+                    qentry.Desc.dma_cookie, qentry.Desc.dma_vproc, qentry.Desc.dma_srcAddr);
+           PRINTF3 (uctx, DBG_SWAP, "                         %016llx %016llx %016llx\n", qentry.Desc.dma_dstAddr, 
+                    qentry.Desc.dma_srcEvent, qentry.Desc.dma_dstEvent);
+
+           if (RING_QUEUE_REALLY_FULL (uctx->uctx_dmaQ))
+               uctx->uctx_status |= UCTX_DPROC_QUEUE_OVERFLOW;
+           else
+           {
+               *RING_QUEUE_BACK (uctx->uctx_dmaQ, uctx->uctx_dmas) = qentry.Desc;
+               (void) RING_QUEUE_ADD (uctx->uctx_dmaQ);
+           }
+           
+           qentry.Desc.dma_typeSize = DMA_ShMemWrite | dev->dev_ctxt.ctxt_num;
+           qentry.Desc.dma_cookie   = 0;
+           qentry.Desc.dma_vproc    = 0;
+           qentry.Desc.dma_srcAddr  = 0;
+           qentry.Desc.dma_dstAddr  = 0;
+           qentry.Desc.dma_srcEvent = 0;
+           qentry.Desc.dma_dstEvent = 0;
+
+           elan4_sdram_copyq_to_sdram (dev, &qentry, qfptr, sizeof (E4_DProcQueueEntry));
+       }
+
+       qfptr = (qfptr & ~(qsize-1)) | ((qfptr + sizeof (E4_DProcQueueEntry)) & (qsize-1));
+       qfull = 0;
+    }
+}
+
+static void
+user_flush_thread_runqueue (ELAN4_DEV *dev, USER_CTXT *uctx, int qfull)
+{
+    E4_uint64          qptrs = read_reg64 (dev, TProcLowPriPtrs);
+    E4_uint32          qsize = E4_QueueSize (E4_QueueSizeValue (qptrs));
+    E4_uint32          qfptr = E4_QueueFrontPointer (qptrs);
+    E4_uint32          qbptr = E4_QueueBackPointer (qptrs);
+    E4_TProcQueueEntry qentry;
+
+    while ((qfptr != qbptr) || qfull)
+    {
+       E4_uint64 context = elan4_sdram_readq (dev, qfptr + offsetof (E4_TProcQueueEntry, Context));
+
+       if (TPROC_Context (context) == uctx->uctx_ctxt.ctxt_num)
+       {
+           elan4_sdram_copyq_from_sdram (dev, qfptr, &qentry, sizeof (E4_TProcQueueEntry));
+
+           PRINTF (uctx, DBG_SWAP, "user_flush_thread_runqueue: %016llx %016llx %016llx %016llx\n", qentry.Regs.Registers[0],
+                   qentry.Regs.Registers[1], qentry.Regs.Registers[2], qentry.Regs.Registers[3]);
+           PRINTF (uctx, DBG_SWAP, "                            %016llx %016llx %016llx\n", 
+                   qentry.Regs.Registers[4], qentry.Regs.Registers[5], qentry.Regs.Registers[6]);
+
+           if (RING_QUEUE_REALLY_FULL (uctx->uctx_threadQ))
+               uctx->uctx_status |= UCTX_TPROC_QUEUE_OVERFLOW;
+           else
+           {
+               *RING_QUEUE_BACK (uctx->uctx_threadQ, uctx->uctx_threads) = qentry.Regs;
+               (void) RING_QUEUE_ADD (uctx->uctx_threadQ);
+           }
+           
+           /* change the thread to execute the suspend sequence */
+           qentry.Regs.Registers[0] = dev->dev_tproc_suspend;
+           qentry.Regs.Registers[1] = dev->dev_tproc_space;
+           qentry.Context           = dev->dev_ctxt.ctxt_num;
+
+           elan4_sdram_copyq_to_sdram (dev, &qentry, qfptr, sizeof (E4_TProcQueueEntry));
+       }
+       
+       qfptr = (qfptr & ~(qsize-1)) | ((qfptr + sizeof (E4_TProcQueueEntry)) & (qsize-1));
+       qfull = 0;
+    }
+}
+
+static void
+user_flush_dmas (ELAN4_DEV *dev, void *arg, int qfull)
+{
+    USER_CTXT        *uctx = (USER_CTXT *) arg;
+    unsigned long     flags;
+    
+    ASSERT ((read_reg32 (dev, InterruptReg) & INT_DProcHalted) != 0);
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    if ((uctx->uctx_status & (UCTX_SWAPPED_REASONS|UCTX_STOPPED_REASONS)) == 0)
+    {
+       PRINTF1 (uctx, DBG_SWAP, "user_flush_dmas: status %x - no more reasons\n", uctx->uctx_status);
+
+       uctx->uctx_status &= ~UCTX_STOPPING;
+
+       user_signal_trap (uctx);
+    }
+    else
+    {
+       user_flush_dma_runqueue (dev, uctx, qfull);
+
+       uctx->uctx_status = (uctx->uctx_status | UCTX_STOPPED) & ~UCTX_STOPPING;
+    
+       PRINTF1 (uctx, DBG_SWAP, "user_flush_dmas: statux %x - stopped\n", uctx->uctx_status);
+
+       kcondvar_wakeupall (&uctx->uctx_wait, &uctx->uctx_spinlock);
+    }
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_flush (ELAN4_DEV *dev, void *arg)
+{
+    USER_CTXT        *uctx = (USER_CTXT *) arg;
+    struct list_head *entry;
+    unsigned long     flags;
+
+    ASSERT ((read_reg32 (dev, InterruptReg) & (INT_Halted|INT_Discarding)) == (INT_Halted|INT_Discarding));
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    if ((uctx->uctx_status & (UCTX_SWAPPED_REASONS|UCTX_STOPPED_REASONS)) == 0)
+    {
+       PRINTF1 (uctx, DBG_SWAP, "user_flush: status %x - no more reasons\n", uctx->uctx_status);
+
+       uctx->uctx_status &= ~UCTX_STOPPING;
+
+       user_signal_trap (uctx);
+    }
+    else
+    {
+       PRINTF1 (uctx, DBG_SWAP, "user_flush: status %x - flushing context\n", uctx->uctx_status);
+
+       list_for_each (entry, &uctx->uctx_cqlist) {
+           USER_CQ *ucq = list_entry (entry, USER_CQ, ucq_link);
+
+           if (ucq->ucq_state == UCQ_RUNNING)
+           {
+               /* NOTE: since the inserter can still be running we modify the permissions
+                *       to zero then when the extractor starts up again it will trap */
+               PRINTF1 (uctx, DBG_SWAP, "user_flush: stopping cq indx=%d\n", elan4_cq2idx(ucq->ucq_cq));
+
+               elan4_updatecq (dev, ucq->ucq_cq, 0, 0);
+           }
+       }
+       
+       user_flush_thread_runqueue (dev, uctx, TPROC_LowRunQueueFull(read_reg64 (dev, TProcStatus)));
+
+       /* since we can't determine whether the dma run queue is full or empty, we use a dma
+        * halt operation to do the flushing - as the reason for halting the dma processor 
+        * will be released when we return, we keep it halted until the flush has completed */
+       elan4_queue_dma_flushop (dev, &uctx->uctx_dma_flushop, 0);
+
+       if (uctx->uctx_status & UCTX_EXITING)
+           elan4_flush_icache_halted (&uctx->uctx_ctxt);
+    }
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_set_filter (USER_CTXT *uctx, E4_uint32 state)
+{
+    struct list_head *entry;
+
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_spinlock));
+
+    list_for_each (entry, &uctx->uctx_cent_list) {
+       USER_CTXT_ENTRY *cent = list_entry (entry, USER_CTXT_ENTRY, cent_link);
+
+       elan4_set_filter (&uctx->uctx_ctxt, cent->cent_cap->cap_mycontext, state);
+    }
+}
+
+static void
+user_start_nacking (USER_CTXT *uctx, unsigned reason)
+{
+    PRINTF2 (uctx, DBG_SWAP, "user_start_nacking: status %x reason %x\n", uctx->uctx_status, reason);
+
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_spinlock));
+
+    if (UCTX_NACKING(uctx))
+       uctx->uctx_status |= reason;
+    else
+    {
+       uctx->uctx_status |= reason;
+
+       user_set_filter (uctx, E4_FILTER_STATS | E4_FILTER_DISCARD_ALL);
+    }
+}
+
+static void
+user_stop_nacking (USER_CTXT *uctx, unsigned reason)
+{
+    PRINTF2 (uctx, DBG_SWAP, "user_stop_nacking: status %x reason %x\n", uctx->uctx_status, reason);
+    
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_spinlock));
+    
+    uctx->uctx_status &= ~reason;
+    
+    if (! UCTX_NACKING (uctx))
+       user_set_filter (uctx, E4_FILTER_STATS);
+}
+
+static void
+user_start_stopping (USER_CTXT *uctx, unsigned reason)
+{
+    ELAN4_DEV *dev =uctx->uctx_ctxt.ctxt_dev;
+
+    PRINTF2 (uctx, DBG_SWAP, "user_start_stopping: status %x reason %x\n", uctx->uctx_status, reason);
+
+    ASSERT (! (uctx->uctx_status & UCTX_STOPPED));
+
+    user_start_nacking (uctx, reason);
+    
+    if ((uctx->uctx_status & UCTX_STOPPING) != 0)
+       return;
+    
+    uctx->uctx_status |= UCTX_STOPPING;
+
+    /* queue the halt operation to  remove all threads/dmas/cqs from the run queues */
+    /*    and also flush through the context filter change */
+    elan4_queue_haltop (dev, &uctx->uctx_haltop);
+}
+
+static void
+user_stop_stopping (USER_CTXT *uctx, unsigned reason)
+{
+    PRINTF2 (uctx, DBG_SWAP, "user_stop_stopping: status %x reason %x\n", uctx->uctx_status, reason);
+    
+    user_stop_nacking (uctx, reason);
+
+    if (UCTX_RUNNABLE (uctx))
+    {
+       uctx->uctx_status &= ~UCTX_STOPPED;
+
+       PRINTF1 (uctx, DBG_SWAP, "user_stop_stopping: no more reasons => %x\n", uctx->uctx_status);
+
+       user_signal_trap (uctx);
+    }
+}
+
+void
+user_swapout (USER_CTXT *uctx, unsigned reason)
+{
+    ELAN4_DEV *dev = uctx->uctx_ctxt.ctxt_dev;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    
+    PRINTF2 (uctx, DBG_SWAP, "user_swapout: status %x reason %x\n", uctx->uctx_status, reason);
+    
+    user_start_nacking (uctx, reason);
+    
+    while (uctx->uctx_status & (UCTX_SWAPPING|UCTX_STOPPING) &&                /* wait for someone else to finish */
+          uctx->uctx_trap_count > 0)                                           /* and for trap handlers to notice */
+    {                                                                          /* and exit */
+       PRINTF1 (uctx, DBG_SWAP, "user_swapout: waiting for %d trap handlers to exit/previous swapout\n", uctx->uctx_trap_count);
+
+       kcondvar_wakeupall (&uctx->uctx_wait, &uctx->uctx_spinlock);
+       kcondvar_wait (&uctx->uctx_wait, &uctx->uctx_spinlock, &flags);
+    }
+
+    if (uctx->uctx_status & UCTX_SWAPPED)                                      /* already swapped out */
+    {
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+       return;
+    }
+    
+    uctx->uctx_status |= (UCTX_SWAPPING|UCTX_STOPPING);                                /* mark the context as swapping & stopping */
+    
+    /* queue the halt operation to  remove all threads/dmas/cqs from the run queues */
+    /*    and also flush through the context filter change */
+    elan4_queue_haltop (dev, &uctx->uctx_haltop);
+    
+    while (! (uctx->uctx_status & UCTX_STOPPED))
+       kcondvar_wait (&uctx->uctx_wait, &uctx->uctx_spinlock, &flags);
+
+    /* all state has been removed from the elan - we can now "tidy" it up */
+
+    PRINTF0 (uctx, DBG_SWAP, "user_swapout: swapped out\n");
+    
+    uctx->uctx_status = (uctx->uctx_status & ~UCTX_SWAPPING) | UCTX_SWAPPED;
+    
+    kcondvar_wakeupall (&uctx->uctx_wait, &uctx->uctx_spinlock);
+
+    PRINTF1 (uctx, DBG_SWAP, "user_swapout: all done - status %x\n", uctx->uctx_status);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+void
+user_swapin (USER_CTXT *uctx, unsigned reason)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    ASSERT (uctx->uctx_status & UCTX_SWAPPED_REASONS);
+
+    PRINTF2 (uctx, DBG_SWAP, "user_swapin: status %x reason %x\n", uctx->uctx_status, reason);
+
+    while (uctx->uctx_status & (UCTX_SWAPPING|UCTX_STOPPING))                  /* wait until other threads have */
+       kcondvar_wait (&uctx->uctx_wait, &uctx->uctx_spinlock, &flags);         /* completed their swap operation */
+
+    ASSERT (uctx->uctx_status & (UCTX_SWAPPED | UCTX_STOPPED));
+
+    user_stop_nacking (uctx, reason);
+
+    if (! (uctx->uctx_status & UCTX_SWAPPED_REASONS))
+    {
+       uctx->uctx_status &= ~UCTX_SWAPPED;
+
+       /* no longer swapped out - wakeup anyone sleeping waiting for swapin */
+       kcondvar_wakeupall (&uctx->uctx_wait, &uctx->uctx_spinlock);
+
+       if (! (uctx->uctx_status & UCTX_STOPPED_REASONS))
+       {
+           uctx->uctx_status &= ~UCTX_STOPPED;
+           user_signal_trap (uctx);
+       }
+    }
+
+    PRINTF1 (uctx, DBG_SWAP, "user_swapin: all done - status %x\n", uctx->uctx_status);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+void
+user_destroy_callback (void *arg, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map)
+{
+    USER_CTXT *uctx = (USER_CTXT *) arg;
+
+    PRINTF (uctx, DBG_VP, "user_destroy_callback: %s\n", map == NULL ? "cap destoyed" : "map destroyed");
+}
+
+int
+user_attach (USER_CTXT *uctx, ELAN_CAPABILITY *cap)
+{
+    ELAN4_DEV       *dev = uctx->uctx_ctxt.ctxt_dev;
+    USER_CTXT_ENTRY *cent;
+    unsigned long flags;
+    int ctype, res;
+    
+    if ((ctype = user_validate_cap (uctx, cap, ELAN_USER_ATTACH)) < 0)
+       return ctype;
+
+    if ((ctype == ELAN_CAP_RMS) && (res = elan_attach_cap (cap, dev->dev_devinfo.dev_rail, uctx, user_destroy_callback)) != 0)
+    {
+       /* NOTE: elan_attach_cap returns +ve errnos */
+       return -res;
+    }
+
+    KMEM_ALLOC (cent, USER_CTXT_ENTRY *, sizeof (USER_CTXT_ENTRY), 1);
+    if (cent == NULL)
+    {
+       if (ctype == ELAN_CAP_RMS)
+           elan_detach_cap (cap, dev->dev_devinfo.dev_rail);
+
+       return -ENOMEM;
+    }
+
+    KMEM_ALLOC (cent->cent_cap, ELAN_CAPABILITY *, ELAN_CAP_SIZE(cap), 1);
+    if (cent->cent_cap == NULL)
+    {
+       if (ctype == ELAN_CAP_RMS)
+           elan_detach_cap (cap, dev->dev_devinfo.dev_rail);
+
+       KMEM_FREE (cent, sizeof (USER_CTXT_ENTRY));
+       return -ENOMEM;
+    }
+
+    bcopy (cap, cent->cent_cap, ELAN_CAP_SIZE(cap));
+
+    if ((res = elan4_attach_filter (&uctx->uctx_ctxt, cap->cap_mycontext)) != 0)
+    {
+       if (ctype == ELAN_CAP_RMS)
+           elan_detach_cap (cap, dev->dev_devinfo.dev_rail);
+       
+       KMEM_FREE (cent->cent_cap, ELAN_CAP_SIZE (cap));
+       KMEM_FREE (cent, sizeof (USER_CTXT_ENTRY));
+
+       return res;
+    }
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    list_add_tail (&cent->cent_link, &uctx->uctx_cent_list);
+
+    if (! UCTX_NACKING (uctx))
+       user_set_filter (uctx, E4_FILTER_STATS);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    return (0);
+    
+}
+
+void
+user_detach (USER_CTXT *uctx, ELAN_CAPABILITY *cap)
+{
+    ELAN4_DEV         *dev = uctx->uctx_ctxt.ctxt_dev;
+    struct list_head  *entry;
+    struct list_head  *next;
+    struct list_head   list;
+    unsigned long      flags;
+
+    INIT_LIST_HEAD (&list);
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    
+    PRINTF (uctx, DBG_NETWORK_CTX, cap ? "user_detach: network context %d\n" : "user_detach: all network contexts\n", cap ? cap->cap_mycontext : 0);
+
+    list_for_each_safe (entry, next, &uctx->uctx_cent_list) {
+       USER_CTXT_ENTRY *cent = list_entry (entry, USER_CTXT_ENTRY, cent_link);
+
+       if (cap == NULL || ELAN_CAP_MATCH (cap, cent->cent_cap))
+       {
+           PRINTF1 (uctx, DBG_NETWORK_CTX, "user_detach: detach from network context %d\n", cent->cent_cap->cap_mycontext);
+           
+           elan4_detach_filter (&uctx->uctx_ctxt, cent->cent_cap->cap_mycontext);
+
+           list_del (&cent->cent_link);
+           list_add_tail (&cent->cent_link, &list);
+       }
+    }
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    while (! list_empty (&list))
+    {
+       USER_CTXT_ENTRY *cent = list_entry (list.next, USER_CTXT_ENTRY, cent_link);
+
+       list_del (&cent->cent_link);
+
+       if (user_validate_cap (uctx, cent->cent_cap, ELAN_USER_DETACH) == ELAN_CAP_RMS)
+           elan_detach_cap (cent->cent_cap, dev->dev_devinfo.dev_rail); 
+       
+       KMEM_FREE (cent->cent_cap, ELAN_CAP_SIZE (cent->cent_cap));
+       KMEM_FREE (cent, sizeof (USER_CTXT_ENTRY));
+    }
+}
+
+void
+user_block_inputter (USER_CTXT *uctx, unsigned blocked)
+{
+    unsigned long flags;
+    int isblocked;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    
+    isblocked = (uctx->uctx_status & UCTX_USER_FILTERING);
+
+    if (blocked && !isblocked)
+       user_start_nacking (uctx, UCTX_USER_FILTERING);
+
+    if (!blocked && isblocked)
+       user_stop_nacking (uctx, UCTX_USER_FILTERING);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static USER_VPSEG *
+user_install_vpseg (USER_CTXT *uctx, unsigned process, unsigned entries)
+{
+    struct list_head *entry;
+    USER_VPSEG       *seg;
+
+    ASSERT (kmutex_is_locked (&uctx->uctx_vpseg_lock));
+
+    list_for_each (entry, &uctx->uctx_vpseg_list) {
+       seg = list_entry (entry, USER_VPSEG, vps_link);
+
+       if (process <= (seg->vps_process + seg->vps_entries-1) && 
+           (process + entries - 1) >= seg->vps_process)
+           return ((USER_VPSEG *) NULL);
+    }
+
+    KMEM_ZALLOC (seg, USER_VPSEG *, sizeof (USER_VPSEG), 1);
+    
+    if (seg == (USER_VPSEG *) NULL)
+       return ((USER_VPSEG *) NULL);
+
+    seg->vps_process = process;
+    seg->vps_entries = entries;
+
+    list_add_tail (&seg->vps_link, &uctx->uctx_vpseg_list);
+
+    return (seg);
+}
+
+static void
+user_remove_vpseg (USER_CTXT *uctx, USER_VPSEG *seg)
+{
+    ASSERT (kmutex_is_locked (&uctx->uctx_vpseg_lock));
+
+    list_del (&seg->vps_link);
+    
+    switch (seg->vps_type)
+    {
+    case USER_VPSEG_P2P:
+       /* These pointers (union) are only valid for P2P segs */
+       if (seg->vps_p2p_routes)
+           KMEM_FREE (seg->vps_p2p_routes, sizeof (E4_VirtualProcessEntry) * seg->vps_entries);
+       
+       if (seg->vps_p2p_cap)
+           KMEM_FREE (seg->vps_p2p_cap, ELAN_CAP_SIZE(seg->vps_p2p_cap));
+
+       break;
+       
+    case USER_VPSEG_BCAST:
+       ;
+    }
+
+    KMEM_FREE (seg, sizeof (USER_VPSEG));
+}
+
+static USER_VPSEG *
+user_find_vpseg (USER_CTXT *uctx, unsigned low, unsigned high)
+{
+    struct list_head *entry;
+
+    ASSERT (kmutex_is_locked (&uctx->uctx_vpseg_lock));
+
+    list_for_each (entry, &uctx->uctx_vpseg_list) {
+       USER_VPSEG *seg = list_entry (entry, USER_VPSEG, vps_link);
+
+       if (seg->vps_process <= low && (seg->vps_process + seg->vps_entries) > high)
+           return (seg);
+    }
+
+    return ((USER_VPSEG *) NULL);
+}
+
+static ELAN_LOCATION 
+user_process2location (USER_CTXT *uctx, USER_VPSEG *seg, unsigned process)
+{
+    ELAN_LOCATION location;
+    int           nnodes, nctxs;
+    int           nodeOff, ctxOff, vpOff;
+
+    location.loc_node    = ELAN_INVALID_NODE;
+    location.loc_context = -1;
+
+    if (seg == NULL)
+       seg = user_find_vpseg (uctx, process, process);
+
+    if (seg == NULL || (seg->vps_type != USER_VPSEG_P2P))
+       return (location);
+
+    nnodes = ELAN_CAP_NUM_NODES (seg->vps_p2p_cap);
+    nctxs  = ELAN_CAP_NUM_CONTEXTS (seg->vps_p2p_cap);
+
+    switch (seg->vps_p2p_cap->cap_type & ELAN_CAP_TYPE_MASK)
+    {
+    case ELAN_CAP_TYPE_BLOCK:
+       for (nodeOff = 0, vpOff = 0; nodeOff < nnodes; nodeOff++)
+       {
+           for (ctxOff = 0; ctxOff < nctxs; ctxOff++)
+           {
+               if ((seg->vps_p2p_cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (seg->vps_p2p_cap->cap_bitmap, ctxOff + (nodeOff * nctxs)))
+               {
+                   if (vpOff++ == (process - seg->vps_process))
+                   { 
+                       location.loc_node    = seg->vps_p2p_cap->cap_lownode + nodeOff;
+                       location.loc_context = seg->vps_p2p_cap->cap_lowcontext + ctxOff;
+                       goto found;
+                   }
+               }
+           }
+       }
+       break;
+       
+    case ELAN_CAP_TYPE_CYCLIC:
+       for (ctxOff = 0, vpOff = 0; ctxOff < nctxs; ctxOff++)
+       {
+           for (nodeOff = 0; nodeOff < nnodes; nodeOff++)
+           {
+               if ((seg->vps_p2p_cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (seg->vps_p2p_cap->cap_bitmap, nodeOff + (ctxOff * nnodes)))
+               {                                   
+                   if (vpOff++ ==  (process - seg->vps_process))
+                   { 
+                       location.loc_node    = seg->vps_p2p_cap->cap_lownode + nodeOff;
+                       location.loc_context = seg->vps_p2p_cap->cap_lowcontext + ctxOff;
+                       goto found;
+                   }
+               }
+           }
+       }
+       break;  
+    }
+       
+ found:
+    return (location);
+}
+
+static unsigned 
+user_location2process (USER_CTXT *uctx, ELAN_LOCATION location)
+{
+    unsigned int      process = ELAN_INVALID_PROCESS;
+    struct list_head *entry;
+    int               nnodes, nctxs;
+    int               nodeOff, ctxOff, vpOff;
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+    list_for_each (entry, &uctx->uctx_vpseg_list) {
+       USER_VPSEG *seg = list_entry (entry, USER_VPSEG, vps_link);
+
+       if (seg->vps_type != USER_VPSEG_P2P)
+           continue;
+
+       if (location.loc_node >= seg->vps_p2p_cap->cap_lownode && location.loc_node <= seg->vps_p2p_cap->cap_highnode &&
+           location.loc_context >= seg->vps_p2p_cap->cap_lowcontext && location.loc_context <= seg->vps_p2p_cap->cap_highcontext)
+       {
+           nnodes = ELAN_CAP_NUM_NODES (seg->vps_p2p_cap);
+           nctxs  = ELAN_CAP_NUM_CONTEXTS (seg->vps_p2p_cap);
+
+           switch (seg->vps_p2p_cap->cap_type & ELAN_CAP_TYPE_MASK)
+           {
+           case ELAN_CAP_TYPE_BLOCK:
+               for (nodeOff = 0, vpOff = 0; nodeOff < nnodes; nodeOff++)
+               {
+                   for (ctxOff = 0; ctxOff < nctxs; ctxOff++)
+                   {
+                       if ((seg->vps_p2p_cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (seg->vps_p2p_cap->cap_bitmap, ctxOff + (nodeOff * nctxs)))
+                       {
+                           if (location.loc_node == seg->vps_p2p_cap->cap_lownode + nodeOff &&
+                               location.loc_context == seg->vps_p2p_cap->cap_lowcontext + ctxOff)
+                           {
+                               process = seg->vps_process + vpOff;
+                               goto found;
+                           }
+                           vpOff++;
+                       }
+                   }
+               }
+               break;
+       
+           case ELAN_CAP_TYPE_CYCLIC:
+               for (ctxOff = 0, vpOff = 0; ctxOff < nctxs; ctxOff++)
+               {
+                   for (nodeOff = 0; nodeOff < nnodes; nodeOff++)
+                   {
+                       if ((seg->vps_p2p_cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (seg->vps_p2p_cap->cap_bitmap, nodeOff + (ctxOff * nnodes)))
+                       {
+                           if (location.loc_node == seg->vps_p2p_cap->cap_lownode + nodeOff &&
+                               location.loc_context == seg->vps_p2p_cap->cap_lowcontext + ctxOff)
+                           {
+                               process = seg->vps_process + vpOff;
+                               goto found;
+                           }
+                           vpOff++;
+                       }
+                   }
+               }
+               break;
+           }
+       }
+    }
+ found:
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+
+    return (process);
+}
+
+static void
+user_loadroute_vpseg (USER_CTXT *uctx, USER_VPSEG *seg, ELAN_POSITION *pos)
+{
+    ELAN4_DEV             *dev    = uctx->uctx_ctxt.ctxt_dev;
+    ELAN_CAPABILITY       *cap    = seg->vps_p2p_cap;
+    unsigned               nnodes = ELAN_CAP_NUM_NODES (cap);
+    unsigned               nctxs  = ELAN_CAP_NUM_CONTEXTS (cap);
+    E4_VirtualProcessEntry route;
+    unsigned              nodeOff;
+    unsigned              ctxOff;
+    unsigned              vpOff;
+
+    switch (cap->cap_type & ELAN_CAP_TYPE_MASK)
+    {
+    case ELAN_CAP_TYPE_BLOCK:
+       for (nodeOff = 0, vpOff = 0; nodeOff < nnodes; nodeOff++)
+       {
+           for (ctxOff = 0; ctxOff < nctxs; ctxOff++)
+           {
+               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, ctxOff + (nodeOff * nctxs)))
+               {
+                   if (seg->vps_p2p_routes != NULL)
+                       route = seg->vps_p2p_routes[vpOff];
+                   else if (elan4_generate_route (&uctx->uctx_position, &route, cap->cap_lowcontext + ctxOff,
+                                                  cap->cap_lownode + nodeOff, cap->cap_lownode + nodeOff, user_p2p_route_options) < 0)
+                   {
+                       vpOff++;
+                       continue;
+                   }
+
+                   PRINTF5 (uctx, DBG_VP, "user_loadroute_vpseg: virtual process %d -> node %d context %d [%016llx.%016llx]\n",
+                            seg->vps_process + vpOff, cap->cap_lownode + nodeOff, cap->cap_lowcontext + ctxOff,
+                            route.Values[0], route.Values[1]);
+                   
+                   elan4_write_route (dev, uctx->uctx_routetable, seg->vps_process + vpOff, &route);
+                                             
+                   vpOff++;
+               }
+           }
+       }
+       break;
+
+    case ELAN_CAP_TYPE_CYCLIC:
+       for (ctxOff = 0, vpOff = 0; ctxOff < nctxs; ctxOff++)
+       {
+           for (nodeOff = 0; nodeOff < nnodes; nodeOff++)
+           {
+               if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) || BT_TEST (cap->cap_bitmap, nodeOff + (ctxOff * nnodes)))
+               {
+                   if (seg->vps_p2p_routes != NULL)
+                       route = seg->vps_p2p_routes[vpOff];
+                   else if (elan4_generate_route (&uctx->uctx_position, &route, cap->cap_lowcontext + ctxOff,
+                                                  cap->cap_lownode + nodeOff, cap->cap_lownode + nodeOff, user_p2p_route_options) < 0)
+                   {
+                       vpOff++;
+                       continue;
+                   }
+
+                   PRINTF5 (uctx, DBG_VP, "user_loadroute_vpseg: virtual process %d -> node %d context %d [%016llx.%016llx]\n",
+                            seg->vps_process + vpOff, cap->cap_lownode + nodeOff, cap->cap_lowcontext + ctxOff,
+                            route.Values[0], route.Values[1]);
+                   
+                   elan4_write_route (dev, uctx->uctx_routetable, seg->vps_process + vpOff, &route);
+                                             
+                   vpOff++;
+               }
+           }
+       }
+       break;
+    }
+}
+
+static int
+user_loadroute_bcast (USER_CTXT *uctx, USER_VPSEG *seg)
+{
+    ELAN4_DEV             *dev = uctx->uctx_ctxt.ctxt_dev;
+    ELAN_POSITION         *pos = &uctx->uctx_position;
+    E4_VirtualProcessEntry route;
+    USER_VPSEG            *aseg;
+    int                    res;
+    ELAN_LOCATION          low;
+    ELAN_LOCATION          high;
+
+    if ((aseg = user_find_vpseg (uctx, seg->vps_bcast_lowvp, seg->vps_bcast_highvp)) == NULL || aseg->vps_type != USER_VPSEG_P2P)
+       return (-EINVAL);
+    
+#ifdef use_elanmod
+    if ((res = user_validate_cap (dev, aseg->vps_p2p_cap, ELAN_USER_BROADCAST)) < 0)
+       return (res);
+#endif
+    
+    low  = user_process2location (uctx, aseg, seg->vps_bcast_lowvp);
+    high = user_process2location (uctx, aseg, seg->vps_bcast_highvp);
+
+    if (low.loc_context != high.loc_context)
+       return (-EINVAL);
+
+    /* NOTE: if loopback can only broadcast to ourself - 
+     *       if back-to-back can only broadcast to other node */
+    if ((pos->pos_mode == ELAN_POS_MODE_LOOPBACK   && low.loc_node != high.loc_node && low.loc_node != pos->pos_nodeid) ||
+       (pos->pos_mode == ELAN_POS_MODE_BACKTOBACK && low.loc_node != high.loc_node && low.loc_node == pos->pos_nodeid))
+    {
+       return (-EINVAL);
+    }
+    
+    if ((res = elan4_generate_route (pos, &route, low.loc_context, low.loc_node, high.loc_node, user_bcast_route_options)) < 0)
+       return (res);
+
+    PRINTF (uctx, DBG_VP, "user_loadroute_bcast: virtual process %d -> nodes %d.%d context %d [%016llx.%016llx]\n",
+           seg->vps_process, low.loc_node, high.loc_node, low.loc_context, route.Values[0], route.Values[1]);
+    
+    elan4_write_route (dev, uctx->uctx_routetable, seg->vps_process, &route);
+    return (0);
+}
+
+int
+user_add_p2pvp (USER_CTXT *uctx, unsigned process, ELAN_CAPABILITY *cap)
+{
+    USER_VPSEG      *seg;
+    ELAN_CAPABILITY *ncap;
+    unsigned         entries;
+
+    if ((cap->cap_type & ELAN_CAP_TYPE_NO_BITMAP) == 0)
+       entries = bt_nbits (cap->cap_bitmap , ELAN_CAP_BITMAPSIZE(cap));
+    else
+       entries = ELAN_CAP_BITMAPSIZE(cap);
+    
+    if ((process + entries) > (E4_VPT_MIN_ENTRIES << uctx->uctx_routetable->tbl_size))
+       return (-EINVAL);
+
+    KMEM_ALLOC (ncap, ELAN_CAPABILITY *, ELAN_CAP_SIZE (cap), 1);
+
+    if (ncap == NULL)
+       return (-ENOMEM);
+    
+    bcopy (cap, ncap, ELAN_CAP_SIZE (cap));
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+
+    if ((seg = user_install_vpseg (uctx, process, entries)) == NULL)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+    
+    seg->vps_type       = USER_VPSEG_P2P;
+    seg->vps_p2p_cap    = ncap;
+    seg->vps_p2p_routes = NULL;
+
+    user_loadroute_vpseg (uctx, seg, &uctx->uctx_position);
+    
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+
+    return (0);
+}
+
+int
+user_add_bcastvp (USER_CTXT *uctx, unsigned process, unsigned lowvp, unsigned highvp)
+{
+    USER_VPSEG *seg;
+    int         res;
+
+    if (lowvp > highvp || process >= (E4_VPT_MIN_ENTRIES << uctx->uctx_routetable->tbl_size))
+       return (-EINVAL);
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+
+    if ((seg = user_install_vpseg (uctx, process, 1)) == NULL)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    seg->vps_type         = USER_VPSEG_BCAST;
+    seg->vps_bcast_lowvp  = lowvp;
+    seg->vps_bcast_highvp = highvp;
+
+    if ((res = user_loadroute_bcast (uctx, seg)) < 0)
+       user_remove_vpseg (uctx, seg);
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+    return (res);
+}
+
+int
+user_removevp (USER_CTXT *uctx, unsigned process)
+{
+    USER_VPSEG *seg;
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+    
+    if (process == ELAN_INVALID_PROCESS)
+       seg = list_entry (uctx->uctx_vpseg_list.next, USER_VPSEG, vps_link);
+    else
+       seg = user_find_vpseg (uctx, process, process);
+
+    if (seg == NULL)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    do {
+       ELAN4_DEV    *dev = uctx->uctx_ctxt.ctxt_dev;
+       int i;
+
+       for (i = 0; i < seg->vps_entries; i++)
+           elan4_invalidate_route (dev, uctx->uctx_routetable, seg->vps_process + i);
+
+       user_remove_vpseg (uctx, seg);
+
+    } while (process == ELAN_INVALID_PROCESS && (seg = list_entry (uctx->uctx_vpseg_list.next, USER_VPSEG, vps_link)) != NULL);
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+
+    return (0);
+}
+
+int
+user_set_route (USER_CTXT *uctx, unsigned process, E4_VirtualProcessEntry *route)
+{
+    ELAN4_DEV    *dev = uctx->uctx_ctxt.ctxt_dev;
+    USER_VPSEG   *seg;
+    ELAN_LOCATION location;
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+
+    if ((seg = user_find_vpseg (uctx, process, process)) == NULL || seg->vps_type != USER_VPSEG_P2P)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    /* check that the route supplied is valid and goes to the correct place */
+    location = user_process2location (uctx, seg, process);
+
+    if (elan4_check_route (&uctx->uctx_position, location, route, 0) != 0)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    if (seg->vps_p2p_routes == NULL)
+       KMEM_ZALLOC (seg->vps_p2p_routes, E4_VirtualProcessEntry *, sizeof (E4_VirtualProcessEntry) * seg->vps_entries, 1);
+    
+    if (seg->vps_p2p_routes == NULL)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-ENOMEM);
+    }
+    
+    seg->vps_p2p_routes[process - seg->vps_process].Values[0] = route->Values[0];
+    seg->vps_p2p_routes[process - seg->vps_process].Values[1] = ROUTE_CTXT_VALUE(location.loc_context) | (route->Values[1] & ~ROUTE_CTXT_MASK);
+    
+    PRINTF (uctx, DBG_ROUTE, "user_set_route: vp=%d -> %016llx%016llx\n", process, 
+           seg->vps_p2p_routes[process - seg->vps_process].Values[1], seg->vps_p2p_routes[process - seg->vps_process].Values[0]);
+
+    elan4_write_route (dev, uctx->uctx_routetable, process, &seg->vps_p2p_routes[process - seg->vps_process]);
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+
+    return (0);
+}
+
+int
+user_reset_route (USER_CTXT *uctx, unsigned process)
+{
+    ELAN4_DEV             *dev = uctx->uctx_ctxt.ctxt_dev;
+    E4_VirtualProcessEntry route;
+    ELAN_LOCATION          location;
+    USER_VPSEG            *seg;
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+
+    if ((seg = user_find_vpseg (uctx, process, process)) == NULL || seg->vps_type != USER_VPSEG_P2P)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    if (seg->vps_p2p_routes != NULL)
+    {
+       seg->vps_p2p_routes[process - seg->vps_process].Values[0] = 0;
+       seg->vps_p2p_routes[process - seg->vps_process].Values[1] = 0;
+    }
+    
+    /* generate the default route to this location */
+    location = user_process2location (uctx, seg, process);
+
+    PRINTF (uctx, DBG_ROUTE, "user_reset_route: vp=%d\n", process);
+
+    if (elan4_generate_route (&uctx->uctx_position, &route, location.loc_context, location.loc_node, location.loc_node, 0) < 0)
+       elan4_invalidate_route (dev, uctx->uctx_routetable, process);
+    else
+       elan4_write_route (dev, uctx->uctx_routetable, process, &route);
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+
+    return (0);
+}
+
+int
+user_get_route (USER_CTXT *uctx, unsigned process, E4_VirtualProcessEntry *route)
+{
+    ELAN4_DEV  *dev = uctx->uctx_ctxt.ctxt_dev;
+    USER_VPSEG   *seg;
+    
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+
+    if ((seg = user_find_vpseg (uctx, process, process)) == NULL || seg->vps_type != USER_VPSEG_P2P)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    elan4_read_route (dev, uctx->uctx_routetable, process, route);
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+    return (0);
+}
+
+int
+user_check_route (USER_CTXT *uctx, unsigned process, E4_VirtualProcessEntry *route, unsigned *error)
+{
+    ELAN4_DEV  *dev = uctx->uctx_ctxt.ctxt_dev;
+    USER_VPSEG *seg;
+    
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+
+    if ((seg = user_find_vpseg (uctx, process, process)) == NULL || seg->vps_type != USER_VPSEG_P2P)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    elan4_read_route (dev, uctx->uctx_routetable, process, route);
+
+    *error = elan4_check_route (&uctx->uctx_position, user_process2location (uctx, seg, process), route, 0);
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+    return (0);
+}
+
+int
+user_send_neterr_msg (USER_CTXT *uctx, unsigned int vp, unsigned int nctx, unsigned int retries, ELAN4_NETERR_MSG *msg)
+{
+    USER_VPSEG   *seg;
+    ELAN_LOCATION location;
+    unsigned long flags;
+    int                  res, found = 0;
+    struct list_head *el;
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+    /* determine the location of the virtual process */
+    if ((seg = user_find_vpseg (uctx, vp, vp)) == NULL)
+    {
+       PRINTF (uctx, DBG_NETERR, "user_send_neterr_msg: vp=%d has no vpseg\n", vp);
+
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return -EINVAL;
+    }
+
+    switch (seg->vps_type)
+    {
+    case USER_VPSEG_P2P:
+       location = user_process2location (uctx, seg, vp);
+       break;
+
+    case USER_VPSEG_BCAST:
+       PRINTF (uctx, DBG_NETERR, "user_send_neterr_msg: vp=%d is a bcast vp\n", vp);
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return -EINVAL;
+    }
+
+    /*  check that we're attached to the network context */
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    list_for_each (el , &uctx->uctx_cent_list) {
+       USER_CTXT_ENTRY *cent = list_entry (el, USER_CTXT_ENTRY, cent_link);
+       
+       if (cent->cent_cap->cap_mycontext == nctx)
+           found++;
+    }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    
+    if (! found)
+    {
+       PRINTF (uctx, DBG_NETERR, "user_send_neterr_msg: nctx=%d not attached\n", nctx);
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+
+       return -EINVAL;
+    }
+
+    /* Update the fields which the user might have "faked" */
+    msg->msg_context            = location.loc_context;
+    msg->msg_sender.loc_node    = uctx->uctx_position.pos_nodeid;
+    msg->msg_sender.loc_context = nctx;
+
+    res = elan4_neterr_sendmsg (uctx->uctx_ctxt.ctxt_dev, location.loc_node, retries, msg);
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+
+    return (res);
+}
+
+
+static int
+user_resolvevp (USER_CTXT *uctx, unsigned process)
+{
+    int                    res = 0;
+    USER_VPSEG            *seg;
+    ELAN_LOCATION          location;
+    E4_VirtualProcessEntry route;
+
+    PRINTF1 (uctx, DBG_VP, "user_resolvevp: process=%d\n", process);
+
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+
+    if ((seg = user_find_vpseg (uctx, process, process)) == NULL)
+    {
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       return (-EINVAL);
+    }
+
+    switch (seg->vps_type)
+    {
+    case USER_VPSEG_P2P:
+#ifdef use_elanmod
+       if ((res = user_validate_cap (uctx, seg->vps_p2p_cap, ELAN_USER_P2P)) != 0)
+           break;
+#endif
+
+       location = user_process2location (uctx, seg, process);
+
+       PRINTF (uctx, DBG_VP, "user_resolvevp: vp=%d -> node=%d ctx=%d\n", process, location.loc_node, location.loc_context);
+       
+       if (seg->vps_p2p_routes != NULL && seg->vps_p2p_routes[process - seg->vps_process].Values[0] != 0)
+           route = seg->vps_p2p_routes[process - seg->vps_process];
+       else if ((res = elan4_generate_route (&uctx->uctx_position, &route, location.loc_context, location.loc_node, location.loc_node, user_p2p_route_options)) < 0)
+           break;;
+       
+       elan4_write_route (uctx->uctx_ctxt.ctxt_dev, uctx->uctx_routetable, process, &route);
+       break;
+
+    case USER_VPSEG_BCAST:
+       res = user_loadroute_bcast (uctx, seg);
+       break;
+       
+    default:
+       res = -EINVAL;
+       break;
+    }
+
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+    return (res);
+}
+
+static void
+user_eproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status)
+{
+    USER_CTXT    *uctx = (USER_CTXT *) ctxt;
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    if (RING_QUEUE_REALLY_FULL (uctx->uctx_eprocTrapQ) || (uctx->uctx_status & UCTX_STOPPED))
+    {
+       PRINTF (uctx, DBG_EPROC, "user_eproc_trap: %s\n", (uctx->uctx_status & UCTX_STOPPED) ? "context stopped" : "trap queue overflow");
+
+       uctx->uctx_status |= UCTX_EPROC_QUEUE_ERROR;
+    }
+    else
+    {
+       elan4_extract_eproc_trap (ctxt->ctxt_dev, status, RING_QUEUE_BACK (uctx->uctx_eprocTrapQ, uctx->uctx_eprocTraps), 0);
+       
+       DBGCMD (ctxt, DBG_EPROC, elan4_display_eproc_trap (ctxt, DBG_EPROC, "user_eproc_trap", RING_QUEUE_BACK(uctx->uctx_eprocTrapQ, uctx->uctx_eprocTraps)));
+       
+       if (RING_QUEUE_ADD (uctx->uctx_eprocTrapQ))
+           user_start_stopping (uctx, UCTX_EPROC_QUEUE_FULL);
+    }
+
+    user_signal_trap (uctx);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_cproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned cqnum)
+{
+    USER_CTXT        *uctx = (USER_CTXT *) ctxt;
+    USER_CQ          *ucq  = NULL;
+    struct list_head *entry;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    
+    list_for_each (entry, &uctx->uctx_cqlist) {
+       ucq = list_entry (entry, USER_CQ, ucq_link);
+
+       if (elan4_cq2num(ucq->ucq_cq) == cqnum)
+           break;
+    }
+
+    ASSERT (ucq != NULL);
+
+    if (ucq->ucq_state != UCQ_RUNNING && CPROC_TrapType (status) == CommandProcInserterError)
+    {
+       PRINTF (ctxt, DBG_TRAP, "user_cproc_trap CommandProcInserterError\n");
+       ucq->ucq_errored++;
+    }
+    else
+    {
+       int vp;
+       
+       ASSERT (ucq->ucq_state == UCQ_RUNNING);
+
+       elan4_extract_cproc_trap (ctxt->ctxt_dev, status, &ucq->ucq_trap, cqnum);
+
+       DBGCMD (ctxt, DBG_CPROC, elan4_display_cproc_trap (ctxt, DBG_CPROC, "user_cproc_trap", &ucq->ucq_trap));
+
+       ucq->ucq_state = UCQ_TRAPPED;
+
+       if ((vp = cproc_open_extract_vp(uctx->uctx_ctxt.ctxt_dev, ucq->ucq_cq)) != -1)
+       {
+           E4_VirtualProcessEntry route;
+
+           elan4_read_route(uctx->uctx_ctxt.ctxt_dev, uctx->uctx_routetable, vp,  &route);
+           elan4_ringbuf_store(&uctx->uctx_ctxt.ctxt_dev->dev_cproc_timeout_routes, &route, uctx->uctx_ctxt.ctxt_dev);
+       }
+    }
+
+    user_signal_trap (uctx);
+       
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_dproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit)
+{
+    USER_CTXT *uctx = (USER_CTXT *) ctxt;
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    if (RING_QUEUE_REALLY_FULL (uctx->uctx_dprocTrapQ) || (uctx->uctx_status & UCTX_STOPPED))
+    {
+       PRINTF (uctx, DBG_DPROC, "user_dproc_trap: %s\n", (uctx->uctx_status & UCTX_STOPPED) ? "context stopped" : "trap queue overflow");
+
+       uctx->uctx_status |= UCTX_DPROC_QUEUE_ERROR;
+    }
+    else
+    {
+       ELAN4_DPROC_TRAP *trap = RING_QUEUE_BACK (uctx->uctx_dprocTrapQ, uctx->uctx_dprocTraps);
+       
+       elan4_extract_dproc_trap (ctxt->ctxt_dev, status, trap, unit);
+
+       if ((DPROC_PacketTimeout (trap->tr_status)) || (DPROC_TrapType (trap->tr_status) == DmaProcFailCountError))
+       {
+           E4_VirtualProcessEntry route;
+
+            elan4_read_route(ctxt->ctxt_dev, uctx->uctx_routetable, trap->tr_desc.dma_vproc, &route);
+
+           if ((route.Values[0] != 0) || (route.Values[1] != 0))
+            {
+               elan4_ringbuf_store(&uctx->uctx_ctxt.ctxt_dev->dev_dproc_timeout_routes, &route, uctx->uctx_ctxt.ctxt_dev);
+           }
+       }
+       else if (DPROC_TrapType (trap->tr_status) == DmaProcPacketAckError)
+       {
+           E4_VirtualProcessEntry route;
+
+            elan4_read_route(ctxt->ctxt_dev, uctx->uctx_routetable, trap->tr_desc.dma_vproc, &route);
+
+           if ((route.Values[0] != 0) || (route.Values[1] != 0))
+            {
+               elan4_ringbuf_store(&uctx->uctx_ctxt.ctxt_dev->dev_ack_error_routes, &route, uctx->uctx_ctxt.ctxt_dev);
+            }
+       } 
+       else 
+       
+       DBGCMD (ctxt, DBG_DPROC, elan4_display_dproc_trap (ctxt, DBG_DPROC, "user_dproc_trap", trap));
+
+       if (!DPROC_PrefetcherFault (status) && DPROC_TrapType(status) == DmaProcFailCountError && !RING_QUEUE_FULL (uctx->uctx_dmaQ))
+       {
+           trap->tr_desc.dma_typeSize |= DMA_FailCount (user_dproc_retry_count);
+
+           *RING_QUEUE_BACK (uctx->uctx_dmaQ, uctx->uctx_dmas) = trap->tr_desc;
+    
+           (void) RING_QUEUE_ADD (uctx->uctx_dmaQ);
+       }
+       else
+       {
+           if (RING_QUEUE_ADD (uctx->uctx_dprocTrapQ))
+               user_start_stopping (uctx, UCTX_DPROC_QUEUE_FULL);
+       }
+    }
+
+    user_signal_trap (uctx);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_tproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status)
+{
+    USER_CTXT *uctx = (USER_CTXT *) ctxt;
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    if (RING_QUEUE_REALLY_FULL (uctx->uctx_tprocTrapQ) || (uctx->uctx_status & UCTX_STOPPED))
+    {
+       PRINTF (uctx, DBG_TPROC, "user_tproc_trap: %s\n", (uctx->uctx_status & UCTX_STOPPED) ? "context stopped" : "trap queue overflow");
+
+       uctx->uctx_status |= UCTX_TPROC_QUEUE_ERROR;
+    }
+    else
+    {
+       elan4_extract_tproc_trap (ctxt->ctxt_dev, status, RING_QUEUE_BACK (uctx->uctx_tprocTrapQ, uctx->uctx_tprocTraps));
+       
+       DBGCMD (ctxt, DBG_TPROC, elan4_display_tproc_trap (ctxt, DBG_TPROC, "user_tproc_trap", RING_QUEUE_BACK (uctx->uctx_tprocTrapQ, uctx->uctx_tprocTraps)));
+       
+       if (RING_QUEUE_ADD (uctx->uctx_tprocTrapQ))
+           user_start_stopping (uctx, UCTX_TPROC_QUEUE_FULL);
+    }
+    user_signal_trap (uctx);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_iproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit)
+{
+    USER_CTXT       *uctx  = (USER_CTXT *) ctxt;
+    USER_IPROC_TRAP *utrap = &uctx->uctx_iprocTrap[unit & 1];
+    unsigned long    flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    ASSERT (utrap->ut_state == UTS_IPROC_RUNNING);
+
+    elan4_extract_iproc_trap (ctxt->ctxt_dev, status, &utrap->ut_trap, unit);
+    DBGCMD (ctxt, DBG_IPROC, elan4_display_iproc_trap (ctxt, DBG_IPROC, "user_iproc_trap", &utrap->ut_trap));
+
+    utrap->ut_state = UTS_IPROC_TRAPPED;
+
+    user_start_nacking (uctx, unit ? UCTX_IPROC_CH0_TRAPPED : UCTX_IPROC_CH1_TRAPPED);
+
+    user_signal_trap (uctx);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_interrupt (ELAN4_CTXT *ctxt, E4_uint64 cookie)
+{
+    USER_CTXT *uctx = (USER_CTXT *) ctxt;
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    PRINTF1 (uctx, DBG_TRAP, "user_interrupt: cookie=%llx\n", cookie);
+
+    switch (cookie)
+    {
+    case ELAN4_INT_COOKIE_DDCQ:
+       uctx->uctx_ddcq_intr--;
+
+       user_signal_trap (uctx);
+       break;
+
+    default:
+       if (uctx->uctx_intcookie_table == NULL || intcookie_fire (uctx->uctx_intcookie_table, cookie) != 0)
+       {
+           PRINTF2 (uctx, DBG_TRAP, "user_interrupt: cookie=%llx %s\n", cookie, uctx->uctx_intcookie_table ? "not found" : "no table");
+           uctx->uctx_status |= UCTX_EPROC_QUEUE_ERROR;
+           user_signal_trap (uctx);
+       }
+       break;
+    }
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+static void
+user_neterrmsg (ELAN4_CTXT *ctxt, ELAN4_NETERR_MSG *msg)
+{
+    USER_CTXT *uctx = (USER_CTXT *) ctxt;
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    
+    if (! RING_QUEUE_FULL (uctx->uctx_msgQ))
+    {
+       memcpy (RING_QUEUE_BACK (uctx->uctx_msgQ, uctx->uctx_msgs), msg, sizeof (ELAN4_NETERR_MSG));
+
+       (void) RING_QUEUE_ADD (uctx->uctx_msgQ);
+    
+       user_signal_trap (uctx);
+    }
+    
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+}
+
+ELAN4_TRAP_OPS user_trap_ops = 
+{
+    user_eproc_trap,
+    user_cproc_trap,
+    user_dproc_trap,
+    user_tproc_trap,
+    user_iproc_trap,
+    user_interrupt,
+    user_neterrmsg,
+};
+
+static int
+deliver_trap (ELAN4_USER_TRAP *utrapp, int type, unsigned proc, void *trap, ...)
+{
+    register int i, len;
+    va_list ap;
+
+    PRINTF (NULL, DBG_TRAP, "deliver_trap: type=%d proc=%d\n", type, proc);
+
+    switch (proc)
+    {
+    case UTS_CPROC:      len = sizeof (ELAN4_CPROC_TRAP); break;
+    case UTS_DPROC:      len = sizeof (ELAN4_DPROC_TRAP); break;
+    case UTS_EPROC:      len = sizeof (ELAN4_EPROC_TRAP); break;
+    case UTS_IPROC:      len = sizeof (ELAN4_IPROC_TRAP); break;
+    case UTS_TPROC:      len = sizeof (ELAN4_TPROC_TRAP); break;
+    case UTS_NETERR_MSG: len = sizeof (ELAN4_NETERR_MSG); break;
+    default:             len = 0; break;
+    }
+
+    if (put_user (type, &utrapp->ut_type) || put_user (proc, &utrapp->ut_proc) || copy_to_user (&utrapp->ut_trap, trap, len))
+       return (UTS_EFAULT);
+
+    va_start (ap, trap);
+    for (i = 0; i < sizeof (utrapp->ut_args)/sizeof (utrapp->ut_args[0]); i++)
+       if (put_user (va_arg (ap, unsigned long), &utrapp->ut_args[i]))
+           return (UTS_EFAULT);
+    va_end (ap);
+
+    return (type);
+}
+
+static int
+user_pagefault (USER_CTXT *uctx, E4_FaultSave *farea)
+{
+    E4_Addr      addr = farea->FaultAddress;
+    E4_uint32    fsr  = FaultSaveFSR(farea->FSRAndFaultContext);
+    FAULT_SAVE  *entry;
+    FAULT_SAVE **predp;
+    int count;
+
+    PRINTF2 (uctx, DBG_FAULT, "user_pagefault: addr=%llx fsr %x\n", (unsigned long long) addr, fsr);
+    
+    if ((fsr & FSR_FaultForBadData) != 0)                      /* Memory ECC error during walk */
+    {
+       PRINTF0 (uctx, DBG_FAULT, "user_pagefault: ECC error during walk\n");
+       return (-EFAULT);
+    }
+    
+    if ((fsr & FSR_FaultForMaxChainCount) != 0)                        /* Have walked a chain of 1024 items */
+    {
+       PRINTF0 (uctx, DBG_FAULT, "user_pagefault: pte chain too long\n");
+       return (-EFAULT);
+    }
+    
+    if (uctx->uctx_num_fault_save)
+    {
+        spin_lock (&uctx->uctx_fault_lock);
+        for( predp = &uctx->uctx_fault_list; (entry = *predp)->next != NULL; predp = &entry->next)
+        {
+           if (entry->addr == (addr & ~((E4_Addr) PAGE_SIZE-1)))
+               break;
+        }
+
+        *predp = entry->next;
+        entry->next = uctx->uctx_fault_list;
+        uctx->uctx_fault_list = entry;
+
+        if (entry->addr == (addr & ~((E4_Addr) PAGE_SIZE-1)))
+        {
+           if ((entry->count <<= 1) > max_fault_pages)
+               entry->count = max_fault_pages;
+        }
+        else
+           entry->count = min_fault_pages;
+
+        entry->addr = (addr & ~((E4_Addr) PAGE_SIZE-1))+(entry->count * PAGE_SIZE);
+        count = entry->count;
+        spin_unlock (&uctx->uctx_fault_lock);
+
+        if (user_load_range (uctx, addr & ~((E4_Addr) PAGE_SIZE-1), count * PAGESIZE, fsr) == 0)
+       return 0;
+
+       /* else pre-faulting has failed, try just this page */
+    }
+
+    return (user_load_range (uctx, addr & ~((E4_Addr) PAGE_SIZE-1), PAGE_SIZE, fsr));
+
+}
+
+static int
+queue_dma_for_retry (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, E4_DMA *dma)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    if (RING_QUEUE_FULL (uctx->uctx_dmaQ))
+    {
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+       
+       return (deliver_trap (utrapp, UTS_QUEUE_OVERFLOW, UTS_NOPROC, NULL, UCTX_DPROC_QUEUE_OVERFLOW));
+    }
+
+    *RING_QUEUE_BACK (uctx->uctx_dmaQ, uctx->uctx_dmas) = *dma;
+    
+    (void) RING_QUEUE_ADD (uctx->uctx_dmaQ);
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    return (UTS_FINISHED);
+}
+
+static int
+queue_thread_for_retry (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, E4_ThreadRegs *regs)
+{
+    unsigned long flags;
+    
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    if (RING_QUEUE_FULL (uctx->uctx_threadQ))
+    {
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+       return (deliver_trap (utrapp, UTS_QUEUE_OVERFLOW, UTS_NOPROC, NULL, UCTX_TPROC_QUEUE_OVERFLOW));
+    }
+
+    *RING_QUEUE_BACK (uctx->uctx_threadQ, uctx->uctx_threads) = *regs;
+    (void) RING_QUEUE_ADD (uctx->uctx_threadQ);
+    
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    return (UTS_FINISHED);
+}
+
+static int
+fixup_eproc_trap (USER_CTXT *uctx, ELAN4_EPROC_TRAP *trap, int waitevent)
+{
+    E4_FaultSave *farea = &trap->tr_faultarea;
+    E4_uint32     fsr   = FaultSaveFSR(farea->FSRAndFaultContext);
+    E4_uint64     CountAndType;
+    E4_uint64     CopySource;
+    E4_uint64     CopyDest;
+
+    /*
+     * Event processor can trap as follows :
+     *   1) Event location read                (faddr == event location & Event Permission)
+     *   2) Event location write       (faddr == event location & Event Permission)
+     *   3) Copy Source read           Read Access
+     *   4) Copy/Write dest write      other
+     *
+     *  NOTE - it is possible to see both 3) and 4) together - but only with physical errors.
+     */
+    if (AT_Perm(fsr) == AT_PermLocalDataRead || AT_Perm(fsr) == AT_PermLocalDataWrite)
+    {
+       /* 
+        * We complete the copy/write by issuing a waitevent 0 of the approriate type.
+        *   - NB mask off bottom bits of EventAddr in case of partial setevent
+        */
+       E4_uint64 EventAddr = trap->tr_eventaddr & ~((E4_uint64) E4_EVENT_ALIGN-1);
+
+       if (! user_ddcq_check (uctx, 4))
+           return (0);
+       
+       if ((trap->tr_event.ev_CountAndType & E4_EVENT_COPY_TYPE_MASK) == E4_EVENT_WRITE)
+       {
+           /* case 4) faulted on write word to destination */
+
+           CountAndType = trap->tr_event.ev_CountAndType & E4_EVENT_TYPE_MASK;
+           
+           PRINTF (uctx, DBG_TRAP, "fixup_eproc_trap: write Event=%llx CountAndType=%llx\n", EventAddr, CountAndType);
+           PRINTF (uctx, DBG_TRAP, "                  WritePtr=%llx WriteValue=%llx\n", 
+                   trap->tr_event.ev_WritePtr, trap->tr_event.ev_WriteValue);
+
+           user_ddcq_waitevent (uctx, EventAddr, CountAndType, trap->tr_event.ev_WritePtr, trap->tr_event.ev_WriteValue);
+       }
+       else
+       {
+           /* case 3) or case 4) faulted on read/write of copy */
+           if (AT_Perm (fsr) == AT_PermLocalDataRead)
+           {
+               CountAndType = (trap->tr_event.ev_CountAndType & E4_EVENT_DATA_TYPE_MASK) | EPROC_CopySize(trap->tr_status);
+               CopySource   = trap->tr_event.ev_CopySource - EVENT_COPY_BLOCK_SIZE;
+               CopyDest     = trap->tr_event.ev_CopyDest;
+           }
+           else
+           {
+               CountAndType = ((trap->tr_event.ev_CountAndType & E4_EVENT_DATA_TYPE_MASK) | 
+                               ((EPROC_CopySize(trap->tr_status) + EVENT_COPY_NDWORDS) & E4_EVENT_COPY_SIZE_MASK));
+               CopySource   = trap->tr_event.ev_CopySource - EVENT_COPY_BLOCK_SIZE;
+               CopyDest     = trap->tr_event.ev_CopyDest - EVENT_COPY_BLOCK_SIZE;
+           }
+           
+           PRINTF (uctx, DBG_TRAP, "fixup_eproc_trap: copy Event=%llx CountAndType=%llx\n", EventAddr, CountAndType);
+           PRINTF (uctx, DBG_TRAP, "                  CopySource=%llx CopyDest=%llx\n", CopySource, CopyDest);
+
+           user_ddcq_waitevent (uctx, EventAddr, CountAndType, CopySource, CopyDest);
+       }
+    }
+    else
+    {
+       E4_uint64 EventAddr  = trap->tr_eventaddr & ~((E4_uint64) E4_EVENT_ALIGN-1);
+       E4_uint32 EventCount = trap->tr_eventaddr & (E4_EVENT_ALIGN-1);
+
+       /* case 1) or 2) - just reissue the event */
+       if (! waitevent)
+           PRINTF (uctx, DBG_TRAP, "fixup_eproc_trap: setevent EventAddr=%llx EventCount=%x\n", EventAddr, EventCount);
+       else
+       {
+           PRINTF (uctx, DBG_TRAP, "fixup_eproc_trap: waitevent Event=%llx CountAndType=%llx\n", EventAddr, trap->tr_event.ev_CountAndType);
+           PRINTF (uctx, DBG_TRAP, "                  Param[0]=%llx Param[1]=%llx\n",
+                    trap->tr_event.ev_Params[0], trap->tr_event.ev_Params[1]);
+       }
+
+       if (! user_ddcq_check (uctx, waitevent ? 4 : 2))
+           return (0);
+       
+       if (waitevent)
+           user_ddcq_waitevent (uctx, EventAddr, trap->tr_event.ev_CountAndType, 
+                                 trap->tr_event.ev_Params[0], trap->tr_event.ev_Params[1]);
+       else
+           user_ddcq_seteventn (uctx, EventAddr, EventCount);
+    }
+
+    return (1);
+}
+
+
+static int
+resolve_eproc_trap (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, ELAN4_EPROC_TRAP *trap)
+{
+    switch (EPROC_TrapType (trap->tr_status))
+    {
+    case EventProcNoFault:
+       PRINTF (uctx, DBG_TRAP, "resolve_eproc_trap: EventProcNoFault\n");
+
+       return (UTS_FINISHED);
+       
+    case EventProcAddressAlignment:
+       return (deliver_trap (utrapp, UTS_ALIGNMENT_ERROR, UTS_EPROC, trap));
+
+    case EventProcMemoryFault:
+       PRINTF (uctx, DBG_TRAP, "resolve_eproc_trap: EventProcMemoryFault @ %llx\n", trap->tr_faultarea.FaultAddress);
+
+       if (user_pagefault (uctx, &trap->tr_faultarea) != 0)
+           return (deliver_trap (utrapp, UTS_INVALID_ADDR, UTS_EPROC, trap));
+
+       return (UTS_FINISHED);
+       
+    case EventProcCountWrapError:
+       return (deliver_trap (utrapp, UTS_BAD_TRAP, UTS_EPROC, trap));
+
+    default:
+       printk ("resolve_eproc_trap: bad trap type %d\n", EPROC_TrapType (trap->tr_status));
+       BUG();
+    }
+
+    return (UTS_FINISHED);
+}
+
+static int
+resolve_cproc_trap (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, USER_CQ *ucq)
+{
+    ELAN4_DEV        *dev    = uctx->uctx_ctxt.ctxt_dev;
+    ELAN4_CPROC_TRAP *trap   = &ucq->ucq_trap;
+    E4_uint64         command;
+    int               res;
+
+    PRINTF2 (uctx, DBG_CPROC, "resolve_cproc_trap: cq %p is trapped - Status %lx\n", ucq, trap->tr_status);
+    
+    switch (CPROC_TrapType (trap->tr_status))
+    {
+    case CommandProcDmaQueueOverflow:
+       PRINTF (uctx, DBG_CPROC, "resolve_cproc_trap: CommandProcDmaQueueOverflow\n");
+       /*
+        * XXXX: should wait for the queue to become emptier if we're 
+        *       responsible for it being very full
+        */
+       ucq->ucq_state = UCQ_NEEDS_RESTART;
+       break;
+
+    case CommandProcInterruptQueueOverflow:
+       PRINTF (uctx, DBG_CPROC, "resolve_cproc_trap: CommandProcInterruptQueueOverflow\n");
+       /*
+        * XXXX: should wait for the queue to become emptier if we're
+        *       responsible for it being very full
+        */
+       ucq->ucq_state = UCQ_NEEDS_RESTART;
+       break;
+       
+    case CommandProcWaitTrap:
+       PRINTF0 (uctx, DBG_CPROC, "resolve_cproc_trap: CommandProcWaitTrap\n");
+       
+       if ((res = resolve_eproc_trap (uctx, utrapp, &trap->tr_eventtrap)) != UTS_FINISHED)
+       {
+           ucq->ucq_state = UCQ_STOPPED;
+
+           return (res);
+       }
+       
+       if (fixup_eproc_trap (uctx, &trap->tr_eventtrap, 1) == 0)
+           return UTS_RESCHEDULE;
+
+       ucq->ucq_state = UCQ_NEEDS_RESTART;
+       break;
+       
+    case CommandProcMemoryFault:
+       PRINTF1 (uctx, DBG_CPROC, "resolve_cproc_trap: CommandProcMemoryFault at %llx\n", trap->tr_faultarea.FaultAddress);
+       if (user_pagefault (uctx, &trap->tr_faultarea) != 0)
+       {
+           ucq->ucq_state = UCQ_STOPPED;
+
+           return (deliver_trap (utrapp, UTS_INVALID_ADDR, UTS_CPROC, trap, elan4_cq2idx(ucq->ucq_cq)));
+       }
+       
+       ucq->ucq_state = UCQ_NEEDS_RESTART;
+       break;
+       
+    case CommandProcRouteFetchFault:
+       command = elan4_trapped_open_command (dev, ucq->ucq_cq);
+       
+       PRINTF1 (uctx, DBG_CPROC, "resolve_cproc_trap: CommandProcRouteFetchFault to vp %d\n", (int) (command >> 32));
+       
+       if (user_resolvevp (uctx, (unsigned) (command >> 32)) != 0)
+       {
+           ucq->ucq_state = UCQ_STOPPED;
+
+           return (deliver_trap (utrapp, UTS_INVALID_VPROC, UTS_CPROC, trap, elan4_cq2idx(ucq->ucq_cq), (long) (command >> 32)));
+       }
+
+       ucq->ucq_state = UCQ_NEEDS_RESTART;
+       break;
+       
+    case CommandProcFailCountZero:
+       PRINTF0 (uctx, DBG_CPROC, "resolve_cproc_trap: CommandProcFailCountZero - reset failcount\n");
+       
+       /* NOTE - we must not modify the ChannelNotCompleted bits - so modify */
+       /*        the restart count with a part-word store */
+       elan4_updatecq (dev, ucq->ucq_cq, ucq->ucq_cq->cq_perm, user_cproc_retry_count);
+
+       ucq->ucq_state = UCQ_NEEDS_RESTART;
+       break;
+
+    case CommandProcAddressAlignment:
+       ucq->ucq_state = UCQ_STOPPED;
+
+       return (deliver_trap (utrapp, UTS_ALIGNMENT_ERROR, UTS_CPROC, trap, elan4_cq2idx(ucq->ucq_cq)));
+
+    case CommandProcPermissionTrap:
+    {
+       sdramaddr_t cqdesc = dev->dev_cqaddr + (elan4_cq2num(ucq->ucq_cq) * sizeof (E4_CommandQueueDesc));
+       E4_uint64   control = elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_Control));
+
+       PRINTF (uctx, DBG_CPROC, "resolve_cproc_trap: CommandProcPermissionTrap - %s\n",
+               (control & CQ_PermissionMask) != ucq->ucq_cq->cq_perm ? "resume from stop" : "permission denied");
+       
+       if ((control & CQ_PermissionMask) == ucq->ucq_cq->cq_perm)
+           return (deliver_trap (utrapp, UTS_PERMISSION_DENIED, UTS_CPROC, trap, elan4_cq2idx(ucq->ucq_cq)));
+
+       elan4_updatecq (dev, ucq->ucq_cq, ucq->ucq_cq->cq_perm, 0);
+
+       ucq->ucq_state = UCQ_NEEDS_RESTART;
+       break;
+    }
+    
+    case CommandProcBadData:
+       ucq->ucq_state = UCQ_STOPPED;
+
+       return (deliver_trap (utrapp, UTS_INVALID_COMMAND, UTS_CPROC, trap, elan4_cq2idx(ucq->ucq_cq)));
+
+    default:
+       ucq->ucq_state = UCQ_STOPPED;
+
+       return (deliver_trap (utrapp, UTS_BAD_TRAP, UTS_CPROC, trap, elan4_cq2idx(ucq->ucq_cq)));
+    }
+
+    return (UTS_FINISHED);
+}
+
+static int
+resolve_dproc_trap (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, ELAN4_DPROC_TRAP *trap)
+{
+    ELAN_LOCATION location;
+    int node;
+
+    if (DPROC_PrefetcherFault (trap->tr_status))
+    {
+       PRINTF (uctx, DBG_DPROC, "resolve_dproc_trap: PrefetcherFault at %llx\n", trap->tr_prefetchFault.FaultAddress);
+
+       if (user_pagefault (uctx, &trap->tr_prefetchFault) != 0)
+           return (deliver_trap (utrapp, UTS_INVALID_ADDR, UTS_DPROC, trap));
+       
+       return (queue_dma_for_retry (uctx, utrapp, &trap->tr_desc));
+    }
+    
+    switch (DPROC_TrapType (trap->tr_status))
+    {
+    case DmaProcRouteFetchFault:
+       PRINTF (uctx, DBG_DPROC, "resolve_dproc_trap: DmaProcRouteFetchFault vp %d\n", trap->tr_desc.dma_vproc);
+
+       if (user_resolvevp (uctx, trap->tr_desc.dma_vproc) != 0)
+           return (deliver_trap (utrapp, UTS_INVALID_VPROC, UTS_DPROC, trap, trap->tr_desc.dma_vproc));
+       
+       return (queue_dma_for_retry (uctx, utrapp, &trap->tr_desc /* immediate */));
+       
+    case DmaProcFailCountError:
+       PRINTF (uctx, DBG_DPROC, "resolve_dproc_trap: DmaProcFailCountError - vp %d cookie %llx\n",
+               trap->tr_desc.dma_vproc, trap->tr_desc.dma_cookie);
+
+       trap->tr_desc.dma_typeSize |= DMA_FailCount (user_dproc_retry_count);
+
+       kmutex_lock (&uctx->uctx_vpseg_lock);
+       location = user_process2location(uctx, NULL, trap->tr_desc.dma_vproc);
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       node = location.loc_node;
+
+        if ((node >= 0) && (node <= uctx->uctx_ctxt.ctxt_dev->dev_position.pos_nodes))
+        { 
+           kmutex_lock(&uctx->uctx_ctxt.ctxt_dev->dev_lock);
+           uctx->uctx_ctxt.ctxt_dev->dev_dproc_timeout[node]++;
+           kmutex_unlock(&uctx->uctx_ctxt.ctxt_dev->dev_lock);
+        }
+       
+       return (queue_dma_for_retry (uctx, utrapp, &trap->tr_desc /* XXXX - backoff for some time later */));
+
+    case DmaProcPacketAckError:
+       PRINTF (uctx, DBG_DPROC, "resolve_dproc_trap: DmaProcPacketAckError - %d%s\n", DPROC_PacketAckValue (trap->tr_status), 
+               DPROC_PacketTimeout (trap->tr_status) ? " timeout" : "");
+
+       kmutex_lock (&uctx->uctx_vpseg_lock);
+       location = user_process2location(uctx, NULL, trap->tr_desc.dma_vproc);
+       kmutex_unlock (&uctx->uctx_vpseg_lock);
+       node = location.loc_node;
+
+       if ((node >= 0) && (node <= uctx->uctx_ctxt.ctxt_dev->dev_position.pos_nodes))
+       {
+           kmutex_lock(&uctx->uctx_ctxt.ctxt_dev->dev_lock);
+           uctx->uctx_ctxt.ctxt_dev->dev_ack_errors[node]++;
+           kmutex_unlock(&uctx->uctx_ctxt.ctxt_dev->dev_lock);
+       }
+
+       return (queue_dma_for_retry (uctx, utrapp, &trap->tr_desc /* XXXX - backoff for some time later */));
+
+    case DmaProcQueueOverflow:
+       PRINTF (uctx, DBG_DPROC, "resolve_dproc_trap: DmaProcQueueOverflow\n");
+       return (queue_dma_for_retry (uctx, utrapp, &trap->tr_desc /* XXXX - backoff for some time later */));
+       
+    case DmaProcRunQueueReadFault:
+       return (deliver_trap (utrapp, UTS_BAD_TRAP, UTS_DPROC, trap));
+       
+    default:
+       printk ("resolve_dproc_trap: unknown trap type : %d\n", DPROC_TrapType(trap->tr_status));
+       BUG();
+    }
+    return UTS_FINISHED;
+}
+
+int
+resolve_tproc_trap (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, ELAN4_TPROC_TRAP *trap)
+{
+    PRINTF (uctx, DBG_TPROC, "resolve_tproc_trap: trap state = %lx\n", trap->tr_state);
+
+    if (trap->tr_state & TS_TrapForTooManyInstructions)
+       return (deliver_trap (utrapp, UTS_BAD_TRAP, UTS_TPROC, trap));
+    
+    if (trap->tr_state & TS_Unimplemented)
+       return (deliver_trap (utrapp, UTS_UNIMP_INSTR, UTS_TPROC, trap));
+    
+    if (trap->tr_state & TS_DataAlignmentError)
+       return (deliver_trap (utrapp, UTS_ALIGNMENT_ERROR, UTS_TPROC, trap));
+    
+    if ((trap->tr_state & TS_InstAccessException) && user_pagefault (uctx, &trap->tr_instFault) != 0)
+       return (deliver_trap (utrapp, UTS_INVALID_ADDR, UTS_TPROC, trap));
+    
+    if ((trap->tr_state & TS_DataAccessException) && user_pagefault (uctx, &trap->tr_dataFault) != 0)
+       return (deliver_trap (utrapp, UTS_INVALID_ADDR, UTS_TPROC, trap));
+    
+    /* If we're restarting from trap - then just need to re-issue it */
+    if (trap->tr_pc == uctx->uctx_trestart_addr || (trap->tr_state & TS_TrappedFlag))
+    {
+       PRINTF (uctx, DBG_TPROC, "resolve_tproc_trap: trapped in trap code PC=%llx SP=%llx\n", trap->tr_pc, trap->tr_regs[1]);
+
+       trap->tr_regs[0] = uctx->uctx_trestart_addr;
+    }
+    else
+    {
+       E4_uint64 *sp = (E4_uint64 *) user_elan2main (uctx, trap->tr_regs[1]);
+       int        i, reload;
+
+       /* need to store the register on the stack see */
+       /*  lib_tproc_trampoline_elan4_thread.S for stack layout */
+#define TS_STACK_OFF(REG)      ((((REG)&7)) - (((REG)>>3)*8) - 8)
+       for (reload = 0, i = 0; i < 64; i++)
+       {
+           if (trap->tr_dirty & ((E4_uint64) 1 << i))
+           {
+               PRINTF (uctx, DBG_TPROC, "resolve_tproc_trap: %%r%d  [%016llx] -> %p\n", i, trap->tr_regs[i], &sp[TS_STACK_OFF(i)]);
+
+               sulonglong ((u64 *) &sp[TS_STACK_OFF(i)], trap->tr_regs[i]);
+               
+               reload |= (1 << (i >> 3));
+           }
+       }
+#undef TS_STACK_OFF
+
+       PRINTF (uctx, DBG_TPROC, "resolve_tproc_trap: pc %llx npc %llx\n", trap->tr_pc, trap->tr_npc);
+       PRINTF (uctx, DBG_TPROC, "resolve_tproc_trap: CC %x reload %x\n", (int) (trap->tr_state >> TS_XCCshift), reload);
+
+       trap->tr_regs[0] = uctx->uctx_trestart_addr;
+       trap->tr_regs[2] = trap->tr_pc;
+       trap->tr_regs[3] = trap->tr_npc;
+       trap->tr_regs[4] = (trap->tr_state >> TS_XCCshift) & TS_XCCmask;
+       trap->tr_regs[5] = reload;
+    }
+
+    return (queue_thread_for_retry (uctx, utrapp, (E4_ThreadRegs *) trap->tr_regs));
+}
+
+static int
+resolve_iproc_trap (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, int channel)
+{
+    USER_IPROC_TRAP  *utrap = &uctx->uctx_iprocTrap[channel];
+    ELAN4_IPROC_TRAP *trap  = &utrap->ut_trap;
+    unsigned long     flags;
+
+    elan4_inspect_iproc_trap (trap);
+
+    if (trap->tr_flags & TR_FLAG_TOOMANY_TRANS)
+       return (deliver_trap (utrapp, UTS_INVALID_TRANS, UTS_IPROC, trap, channel));
+
+    ASSERT (trap->tr_trappedTrans >= 0 && trap->tr_trappedTrans < trap->tr_numTransactions);
+
+    switch (IPROC_TrapValue (trap->tr_transactions[trap->tr_trappedTrans].IProcStatusCntxAndTrType))
+    {
+    case InputMemoryFault:
+       if (user_pagefault (uctx, &trap->tr_faultarea) != 0)
+       {
+           utrap->ut_state = UTS_IPROC_STOPPED;
+           
+           return (deliver_trap (utrapp, UTS_INVALID_ADDR, UTS_IPROC, trap, channel));
+       }
+       break;
+
+    case InputDmaQueueOverflow:
+    case InputEventEngineTrapped:
+       /* nothing to do for these 2 - restarting will simulate the transactions */
+       break;
+
+    case InputEopErrorOnWaitForEop:
+    case InputEopErrorTrap:
+       break;
+
+    case InputCrcErrorAfterPAckOk:
+       PRINTF (DBG_DEVICE, DBG_IPROC, "InputCrcErrorAfterPAckOk: flags %x\n", trap->tr_flags);
+
+       ASSERT ((trap->tr_flags & TR_FLAG_ACK_SENT) && ((trap->tr_flags & (TR_FLAG_DMA_PACKET|TR_FLAG_BAD_TRANS)) ||
+                                                       ((trap->tr_flags & TR_FLAG_EOP_ERROR) && trap->tr_identifyTrans == TR_TRANS_INVALID)));
+       break;
+
+    case InputDiscardAfterAckOk:
+       return (deliver_trap (utrapp, UTS_INVALID_TRANS, UTS_IPROC, trap, channel));
+
+    case InputAddressAlignment:
+       return (deliver_trap (utrapp, UTS_ALIGNMENT_ERROR, UTS_IPROC, trap, channel));
+
+    case InputInvalidTransType:
+       return (deliver_trap (utrapp, UTS_INVALID_TRANS, UTS_IPROC, trap, channel));
+
+    default:
+       printk ("resolve_iproc_trap: unknown trap type %d\n", IPROC_TrapValue (trap->tr_transactions[trap->tr_trappedTrans].IProcStatusCntxAndTrType));
+       BUG();
+       /* NOTREACHED */
+    }
+
+    if (! (trap->tr_flags & TR_FLAG_ACK_SENT) || (trap->tr_flags & TR_FLAG_EOP_BAD))
+    {
+       spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+       utrap->ut_state = UTS_IPROC_RUNNING;
+
+       user_stop_nacking (uctx, channel ? UCTX_IPROC_CH0_TRAPPED : UCTX_IPROC_CH1_TRAPPED);
+       
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    }
+    else if ((trap->tr_flags & (TR_FLAG_DMA_PACKET | TR_FLAG_BAD_TRANS)) || ((trap->tr_flags & TR_FLAG_EOP_ERROR) && (trap->tr_identifyTrans == TR_TRANS_INVALID)))
+    {
+       /* 
+        * TR_FLAG_DMA_PACKET   means a DMA packet has faulted.
+        *
+        * TR_FLAG_BAD_TRANS    means we have a transaction with a bad CRC after the transaction
+        *                      which sent the ack - this implies it's an overlapped ack DMA packet
+        *
+        * TR_FLAG_EOP_ERROR    means we've received an EOP reset - if we hadn't seen an identify
+        *                      transaction then it's a DMA packet.
+        *
+        * To ensure that the DMA processor works correctly the next packet must be NACKed to 
+        * cause it to resend this one.
+        */
+       PRINTF (uctx, DBG_IPROC, "resolve_iproc_trap: %s during DMA packet\n",
+               (trap->tr_flags & TR_FLAG_BAD_TRANS) ? "BadTransaction" : (trap->tr_flags & TR_FLAG_EOP_ERROR) ? "EopError" : "trap");
+
+       spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+       if (trap->tr_flags & TR_FLAG_DMA_PACKET)
+       {
+           if (! (trap->tr_flags & TR_FLAG_BAD_TRANS))
+               utrap->ut_state = UTS_IPROC_EXECUTE_PACKET;
+           else
+           {
+               kcondvar_t waithere;
+
+               /* We must ensure that the next packet is always nacked, so
+                * we wait here for an output timeout before dropping the 
+                * context filter - we just pause here for 4 mS */
+               kcondvar_init (&waithere);
+               kcondvar_timedwait (&waithere, &uctx->uctx_spinlock, &flags, lbolt + (HZ/250) + 1);;
+               kcondvar_destroy (&waithere);
+
+               utrap->ut_state = UTS_IPROC_RUNNING;
+               
+               user_stop_nacking (uctx, channel ? UCTX_IPROC_CH0_TRAPPED : UCTX_IPROC_CH1_TRAPPED);
+           }
+       }
+       else
+       {
+           utrap->ut_state = UTS_IPROC_RUNNING;
+
+           user_stop_nacking (uctx, channel ? UCTX_IPROC_CH0_TRAPPED : UCTX_IPROC_CH1_TRAPPED);
+       }
+
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    }
+    else if (trap->tr_flags & TR_FLAG_EOP_ERROR)
+    {
+       PRINTF (uctx, DBG_IPROC, "resolve_iproc_trap: EopError with identify\n");
+
+       utrap->ut_state = UTS_IPROC_NETWORK_ERROR;
+    }
+    else
+    {
+       PRINTF (uctx, DBG_IPROC, "resolve_iproc_trap: execute packet\n");
+
+       utrap->ut_state = UTS_IPROC_EXECUTE_PACKET;
+    }
+
+    return UTS_FINISHED;
+}
+
+
+static int
+resolve_cproc_traps (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    struct list_head *entry;
+    int res = UTS_FINISHED;
+
+    kmutex_lock (&uctx->uctx_cqlock);
+    list_for_each (entry, &uctx->uctx_cqlist) {
+       USER_CQ *ucq = list_entry (entry, USER_CQ, ucq_link);
+
+       if (ucq->ucq_state == UCQ_TRAPPED)
+       {
+           res = resolve_cproc_trap (uctx, utrapp, ucq);
+
+           if (res != UTS_FINISHED)
+               break;
+       }
+
+       if (ucq->ucq_errored)
+       {
+           ucq->ucq_errored = 0;
+           res = deliver_trap (utrapp, UTS_CPROC_ERROR, UTS_CPROC, &ucq->ucq_trap, elan4_cq2idx(ucq->ucq_cq));
+           break;
+       }
+    }
+    kmutex_unlock (&uctx->uctx_cqlock);
+
+    return (res);
+}
+
+static int
+resolve_eproc_traps (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    unsigned long flags;
+    int res;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    while (! RING_QUEUE_EMPTY (uctx->uctx_eprocTrapQ))
+    {
+       ELAN4_EPROC_TRAP trap = *RING_QUEUE_FRONT (uctx->uctx_eprocTrapQ, uctx->uctx_eprocTraps);
+
+       (void) RING_QUEUE_REMOVE (uctx->uctx_eprocTrapQ);
+
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+       if ((res = resolve_eproc_trap (uctx, utrapp, &trap)) != UTS_FINISHED)
+           return (res);
+
+       if (fixup_eproc_trap (uctx, &trap, 0) == 0)
+       {
+           PRINTF (uctx, DBG_EPROC, "resolve_eproc_trap: could not fixup eproc trap - requeue it\n");
+
+           spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+           if (RING_QUEUE_REALLY_FULL(uctx->uctx_eprocTrapQ))
+               uctx->uctx_status |= UCTX_EPROC_QUEUE_OVERFLOW;
+           else
+           {
+               *RING_QUEUE_FRONT(uctx->uctx_eprocTrapQ, uctx->uctx_eprocTraps) = trap;
+           
+               (void) RING_QUEUE_ADD_FRONT(uctx->uctx_eprocTrapQ);
+           }
+           spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+           return UTS_RESCHEDULE;
+       }
+       
+       spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    }
+
+    if (uctx->uctx_status & UCTX_EPROC_QUEUE_FULL)
+       user_stop_stopping (uctx, UCTX_EPROC_QUEUE_FULL);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    return (UTS_FINISHED);
+}
+           
+static int
+resolve_dproc_traps (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    unsigned long flags;
+    int res;
+    
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    while (! RING_QUEUE_EMPTY (uctx->uctx_dprocTrapQ))
+    {
+       ELAN4_DPROC_TRAP trap = *RING_QUEUE_FRONT(uctx->uctx_dprocTrapQ, uctx->uctx_dprocTraps);
+       
+       (void) RING_QUEUE_REMOVE (uctx->uctx_dprocTrapQ);
+
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+       if ((res = resolve_dproc_trap (uctx, utrapp, &trap)) != UTS_FINISHED)
+           return (res);
+       
+       spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    }
+
+    if (uctx->uctx_status & UCTX_DPROC_QUEUE_FULL)
+       user_stop_stopping (uctx, UCTX_DPROC_QUEUE_FULL);
+    
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    return (UTS_FINISHED);
+}
+
+static int
+resolve_tproc_traps (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    unsigned long flags;
+    int res;
+    
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    while (! RING_QUEUE_EMPTY (uctx->uctx_tprocTrapQ))
+    {
+       ELAN4_TPROC_TRAP trap = *RING_QUEUE_FRONT(uctx->uctx_tprocTrapQ, uctx->uctx_tprocTraps);
+       
+       (void) RING_QUEUE_REMOVE (uctx->uctx_tprocTrapQ);
+
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+       if ((res = resolve_tproc_trap (uctx, utrapp, &trap)) != UTS_FINISHED)
+           return (res);
+       
+       spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    }
+
+    if (uctx->uctx_status & UCTX_TPROC_QUEUE_FULL)
+       user_stop_stopping (uctx, UCTX_TPROC_QUEUE_FULL);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    return (UTS_FINISHED);
+}
+
+static int
+resolve_iproc_traps (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    unsigned long flags;
+    int i, res;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    for (i = 0; i < 2; i++)
+       if (uctx->uctx_iprocTrap[i].ut_state == UTS_IPROC_TRAPPED)
+       {
+           uctx->uctx_iprocTrap[i].ut_state = UTS_IPROC_RESOLVING;
+           spin_unlock_irqrestore(&uctx->uctx_spinlock, flags);
+           
+           if ((res = resolve_iproc_trap (uctx, utrapp, i)) != UTS_FINISHED)
+               return (res);
+           
+           spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+       }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    
+    return (UTS_FINISHED);
+}
+
+static int
+resolve_all_traps (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    int res;
+
+    if ((res = resolve_iproc_traps (uctx, utrapp)) != UTS_FINISHED ||
+       (res = resolve_cproc_traps (uctx, utrapp)) != UTS_FINISHED ||
+       (res = resolve_eproc_traps (uctx, utrapp)) != UTS_FINISHED ||
+       (res = resolve_dproc_traps (uctx, utrapp)) != UTS_FINISHED ||
+       (res = resolve_tproc_traps (uctx, utrapp)) != UTS_FINISHED)
+       return (res);
+
+    if (uctx->uctx_status & UCTX_OVERFLOW_REASONS)
+       return (deliver_trap (utrapp, UTS_QUEUE_OVERFLOW, UTS_NOPROC, NULL, uctx->uctx_status));
+
+    if (uctx->uctx_status & UCTX_ERROR_REASONS)
+       return (deliver_trap (utrapp, UTS_QUEUE_ERROR, UTS_NOPROC, NULL, uctx->uctx_status));
+
+    return (UTS_FINISHED);
+}
+
+static int
+execute_iproc_traps (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    unsigned long flags;
+    int i;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    for (i = 0; i < 2; i++)
+       switch (uctx->uctx_iprocTrap[i].ut_state)
+       {
+       case UTS_IPROC_EXECUTE_PACKET:
+           uctx->uctx_iprocTrap[i].ut_state = UTS_IPROC_EXECUTING;
+           spin_unlock_irqrestore(&uctx->uctx_spinlock, flags);
+           
+           return (deliver_trap (utrapp, UTS_EXECUTE_PACKET, UTS_IPROC, &uctx->uctx_iprocTrap[i].ut_trap, i));
+
+       case UTS_IPROC_NETWORK_ERROR:
+           spin_unlock_irqrestore(&uctx->uctx_spinlock, flags);
+           
+           return (deliver_trap (utrapp, UTS_NETWORK_ERROR_TRAP, UTS_IPROC, &uctx->uctx_iprocTrap[i].ut_trap, i));
+       }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    
+    return (UTS_FINISHED);
+}
+
+static int
+progress_neterr (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    if (! RING_QUEUE_EMPTY (uctx->uctx_msgQ))
+    {
+       ELAN4_NETERR_MSG msg = *RING_QUEUE_FRONT (uctx->uctx_msgQ, uctx->uctx_msgs);
+       
+       (void) RING_QUEUE_REMOVE (uctx->uctx_msgQ);
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+       
+       return deliver_trap (utrapp, UTS_NETWORK_ERROR_MSG, UTS_NETERR_MSG, &msg, user_location2process (uctx, msg.msg_sender));
+    }
+    
+    if (uctx->uctx_status & UCTX_NETERR_TIMER)
+    {
+       uctx->uctx_status &= ~UCTX_NETERR_TIMER;
+
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+       
+       return deliver_trap (utrapp, UTS_NETWORK_ERROR_TIMER, UTS_NOPROC, NULL);
+    }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    
+    return (UTS_FINISHED);
+}
+
+static void
+restart_command_queues (USER_CTXT *uctx)
+{
+    struct list_head *entry;
+
+    ASSERT (SPINLOCK_HELD (&uctx->uctx_spinlock));
+    
+    list_for_each (entry, &uctx->uctx_cqlist) {
+       USER_CQ *ucq = list_entry (entry, USER_CQ, ucq_link);
+       
+       if (ucq->ucq_state == UCQ_NEEDS_RESTART)
+       {
+           ucq->ucq_state = UCQ_RUNNING;
+           
+           elan4_restartcq (uctx->uctx_ctxt.ctxt_dev, ucq->ucq_cq);
+       }
+    }
+}
+
+static int
+restart_dmas (USER_CTXT *uctx)
+{
+    PRINTF (uctx, DBG_TRAP, "restart_dmas: back=%d front=%d\n", uctx->uctx_dmaQ.q_back, uctx->uctx_dmaQ.q_front);
+
+    while (! RING_QUEUE_EMPTY (uctx->uctx_dmaQ))
+    {
+       if (! user_ddcq_check (uctx, 7))
+           return (0);
+
+       user_ddcq_run_dma (uctx, RING_QUEUE_FRONT(uctx->uctx_dmaQ, uctx->uctx_dmas));
+       
+       (void) RING_QUEUE_REMOVE (uctx->uctx_dmaQ);
+    }
+
+    return (1);
+}
+
+static int
+restart_threads (USER_CTXT *uctx)
+{
+    PRINTF (uctx, DBG_TRAP, "restart_threads: back=%d front=%d\n", uctx->uctx_threadQ.q_back, uctx->uctx_threadQ.q_front);
+
+    while (! RING_QUEUE_EMPTY (uctx->uctx_threadQ))
+    {
+       if (! user_ddcq_check (uctx, 7))
+           return (0);
+
+       user_ddcq_run_thread (uctx, RING_QUEUE_FRONT(uctx->uctx_threadQ, uctx->uctx_threads));
+       
+       (void) RING_QUEUE_REMOVE (uctx->uctx_threadQ);
+    }
+
+    return (1);
+}
+
+int
+user_resume_eproc_trap (USER_CTXT *uctx, E4_Addr addr)
+{
+    PRINTF2 (uctx, DBG_RESUME, "user_resume_eproc_trap: addr=%llx -> %s\n", addr, user_ddcq_check(uctx, 2) ? "success" : "EAGAIN");
+
+    if (! user_ddcq_check (uctx, 2))
+       return (-EAGAIN);
+
+    user_ddcq_setevent (uctx, addr);
+
+    return (0);
+}
+
+int
+user_resume_cproc_trap (USER_CTXT *uctx, unsigned indx)
+{
+    struct list_head *entry;
+    unsigned long flags;
+
+    PRINTF1 (uctx, DBG_RESUME, "user_resume_cproc_trap: indx=%d\n", indx);
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    list_for_each (entry, &uctx->uctx_cqlist) {
+       USER_CQ *ucq = list_entry (entry, USER_CQ, ucq_link);
+       
+       if (elan4_cq2idx(ucq->ucq_cq) == indx && ucq->ucq_state == UCQ_STOPPED && !(ucq->ucq_flags & UCQ_SYSTEM))
+       {
+           ucq->ucq_state = UCQ_NEEDS_RESTART;
+           
+           user_signal_trap (uctx);
+
+           spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+           return (0);
+       }
+    }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    return (-EINVAL);
+}
+
+int
+user_resume_dproc_trap (USER_CTXT *uctx, E4_DMA *dma)
+{
+    unsigned long flags;
+    int res = 0;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    if (RING_QUEUE_FULL (uctx->uctx_dmaQ))
+       res = -ENOMEM;
+    else
+    {
+       *RING_QUEUE_BACK (uctx->uctx_dmaQ, uctx->uctx_dmas) = *dma;
+       (void) RING_QUEUE_ADD (uctx->uctx_dmaQ);
+
+       user_signal_trap (uctx);
+    }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    return (res);
+}
+
+int
+user_resume_tproc_trap (USER_CTXT *uctx, E4_ThreadRegs *regs)
+{
+    unsigned long flags;
+    int res = 0;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    if (RING_QUEUE_FULL (uctx->uctx_threadQ))
+       res = -ENOMEM;
+    else
+    {
+       *RING_QUEUE_BACK (uctx->uctx_threadQ, uctx->uctx_threads) = *regs;
+       (void) RING_QUEUE_ADD (uctx->uctx_threadQ);
+
+       user_signal_trap (uctx);
+    }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    return (res);
+}
+
+int
+user_resume_iproc_trap (USER_CTXT *uctx, unsigned channel, unsigned trans,
+                       E4_IprocTrapHeader *hdrp, E4_IprocTrapData *datap)
+{
+    unsigned long flags;
+    int res = 0;
+
+    if (channel >= 2)
+       return (-EINVAL);
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    if (uctx->uctx_iprocTrap[channel].ut_state != UTS_IPROC_STOPPED &&
+       uctx->uctx_iprocTrap[channel].ut_state != UTS_IPROC_EXECUTING &&
+       uctx->uctx_iprocTrap[channel].ut_state != UTS_IPROC_NETWORK_ERROR)
+       res = -EINVAL;
+    else
+    {
+       ELAN4_IPROC_TRAP *trap = &uctx->uctx_iprocTrap[channel].ut_trap;
+
+       if (trans < trap->tr_numTransactions)
+       {
+           PRINTF1 (uctx, DBG_RESUME, "user_resume_iproc_trap: trans=%d -> execute\n", trans);
+
+           uctx->uctx_iprocTrap[channel].ut_state = UTS_IPROC_EXECUTE_PACKET;
+           trap->tr_trappedTrans                  = trans;
+           trap->tr_transactions[trans]           = *hdrp;
+           trap->tr_dataBuffers[trans]            = *datap;
+       }
+       else
+       {
+           PRINTF1 (uctx, DBG_RESUME, "user_resume_iproc_trap: trans=%d -> running\n", trans);
+
+           uctx->uctx_iprocTrap[channel].ut_state = UTS_IPROC_RUNNING;
+       
+           user_stop_nacking (uctx, channel ? UCTX_IPROC_CH0_TRAPPED : UCTX_IPROC_CH1_TRAPPED);
+       }
+    }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    
+    return (res);
+}
+
+int
+__categorise_command (E4_uint64 command, int *cmdSize)
+{
+    switch (command & 0x3)
+    {
+    case RUN_THREAD_CMD: *cmdSize = 7; break;
+       
+    default:
+       switch (command & 0x7)
+       {
+       case WRITE_DWORD_CMD: *cmdSize = 2; break;
+       case ADD_DWORD_CMD:   *cmdSize = 2; break;
+           
+       default:
+           switch (command & 0xF)
+           {
+           case OPEN_STEN_PKT_CMD:
+               *cmdSize = 1;
+               return 1;
+               
+           case COPY64_CMD:    *cmdSize = 2; break;
+           case GUARD_CMD:     *cmdSize = 1; break;
+           case INTERRUPT_CMD: *cmdSize = 1; break;
+           case RUN_DMA_CMD:   *cmdSize = 7; break;
+               
+           default:
+               switch (command & 0x1f)
+               {
+               case SEND_TRANS_CMD:
+                   *cmdSize = 2 + (((command >> 16) & TR_SIZE_MASK) >> TR_SIZE_SHIFT);
+                   return 2;
+                   
+               case SET_EVENT_CMD:    *cmdSize = 1; break;
+               case SET_EVENTN_CMD:   *cmdSize = 2; break;
+               case WAIT_EVENT_CMD:   *cmdSize = 4; break;
+
+               default:
+                   switch (command & 0x3f)
+                   {
+                   case NOP_CMD:            *cmdSize = 1; break;
+                   case MAKE_EXT_CLEAN_CMD: *cmdSize = 1; break;
+                   default:
+                       return 3;
+                   }
+                   break;
+               }
+           }
+       }
+    }
+
+    return 0;
+}
+
+int
+__whole_command (sdramaddr_t *commandPtr, sdramaddr_t insertPtr, unsigned int cqSize, unsigned int cmdSize)
+{
+    /* Move onto next command */
+    while (cmdSize-- && (*commandPtr) != insertPtr)
+       *commandPtr = ((*commandPtr) & ~(cqSize-1)) | (((*commandPtr) + sizeof (E4_uint64)) & (cqSize-1));
+
+    return cmdSize == -1;
+}
+
+int
+user_neterr_sten (USER_CTXT *uctx, unsigned int vp, E4_uint64 cookie, int waitforeop)
+{
+    ELAN4_DEV        *dev   = uctx->uctx_ctxt.ctxt_dev;
+    int                      found = 0;
+    struct list_head *el;
+
+    user_swapout (uctx, UCTX_NETERR_FIXUP);
+
+    kmutex_lock (&uctx->uctx_cqlock);
+    list_for_each (el, &uctx->uctx_cqlist) {
+       USER_CQ *ucq = list_entry (el, USER_CQ, ucq_link);
+       
+       if ((ucq->ucq_cq->cq_perm & CQ_STENEnableBit) != 0)
+       {
+           sdramaddr_t   cqdesc       = dev->dev_cqaddr + (elan4_cq2num(ucq->ucq_cq) * sizeof (E4_CommandQueueDesc));
+           E4_uint64     queuePtrs    = elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_QueuePtrs));
+           sdramaddr_t   insertPtr    = (queuePtrs & CQ_PtrMask);
+           sdramaddr_t   commandPtr   = CQ_CompletedPtr (queuePtrs);
+           unsigned int  cqSize       = CQ_Size ((queuePtrs >> CQ_SizeShift) & CQ_SizeMask);
+           E4_uint64     openCommand  = 0;
+
+           if (dev->dev_devinfo.dev_revision_id != PCI_REVISION_ID_ELAN4_REVA && (queuePtrs & CQ_RevB_ReorderingQueue))
+           {
+               E4_uint32 oooMask = elan4_sdram_readl (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_HoldingValue));
+               
+               for (; (oooMask & 1) != 0; oooMask >>= 1)
+                   insertPtr = (insertPtr & ~(cqSize-1)) | ((insertPtr + sizeof (E4_uint64)) & (cqSize-1));
+           }
+
+           while (commandPtr != insertPtr)
+           {
+               E4_uint64    command = elan4_sdram_readq (dev, commandPtr);
+               sdramaddr_t  identifyPtr;
+               unsigned int cmdSize;
+               
+               switch (__categorise_command (command, &cmdSize))
+               {
+               case 0:
+                   (void) __whole_command (&commandPtr, insertPtr, cqSize, cmdSize);
+                   break;
+                   
+               case 1:
+                   PRINTF (uctx, DBG_NETERR, "user_neterr_sten: cq=%d OPEN %llx\n", elan4_cq2num (ucq->ucq_cq), command);
+                   
+                   if ((command >> 32) == vp)
+                       openCommand = command;
+
+                   (void) __whole_command (&commandPtr, insertPtr, cqSize, cmdSize);
+                   break;
+                   
+               case 2:
+                   PRINTF (uctx, DBG_NETERR, "user_neterr_sten: cq=%d SENDTRANS %llx\n", elan4_cq2num (ucq->ucq_cq), command);
+                   
+                   if (openCommand == 0)
+                       (void) __whole_command (&commandPtr, insertPtr, cqSize, cmdSize);
+                   else
+                   {
+                       switch ((command >> 16) & (TR_OPCODE_MASK | TR_SIZE_MASK))
+                       {
+                       case TR_IDENTIFY  & (TR_OPCODE_MASK | TR_SIZE_MASK):
+                       case TR_REMOTEDMA & (TR_OPCODE_MASK | TR_SIZE_MASK):
+                           PRINTF (uctx, DBG_NETERR, "user_neterr_sten: TR_IDENTIFY/TR_REMOTEDMA\n");
+                           identifyPtr = (commandPtr & ~(cqSize-1)) | ((commandPtr + sizeof (E4_uint64)) & (cqSize-1));
+                           break;
+                           
+                       case TR_SETEVENT_IDENTIFY & (TR_OPCODE_MASK | TR_SIZE_MASK):
+                       case TR_INPUT_Q_COMMIT    & (TR_OPCODE_MASK | TR_SIZE_MASK):
+                           PRINTF (uctx, DBG_NETERR, "user_neterr_sten: TR_SETEVENT_IDENTIFY/TR_INPUT_Q_COMMIT\n");
+                           identifyPtr = (commandPtr & ~(cqSize-1)) | ((commandPtr + 2*sizeof (E4_uint64)) & (cqSize-1));
+                           break;
+                           
+                       case TR_ADDWORD & (TR_OPCODE_MASK | TR_SIZE_MASK):
+                           PRINTF (uctx, DBG_NETERR, "user_neterr_sten: TR_ADDWORD\n");
+                           identifyPtr = (commandPtr & ~(cqSize-1)) | ((commandPtr + 3*sizeof (E4_uint64)) & (cqSize-1));
+                           break;
+                           
+                       case TR_TESTANDWRITE & (TR_OPCODE_MASK | TR_SIZE_MASK):
+                           PRINTF (uctx, DBG_NETERR, "user_neterr_sten: TR_TESTANDWRITE\n");
+                           identifyPtr = (commandPtr & ~(cqSize-1)) | ((commandPtr + 4*sizeof (E4_uint64)) & (cqSize-1));
+                           break;
+                           
+                       default:
+                           identifyPtr = 0;
+                       }
+                       
+                       if (! __whole_command (&commandPtr, insertPtr, cqSize, cmdSize))
+                       {
+                           PRINTF (uctx, DBG_NETERR, "user_neterr_sten: not whole command\n");
+                           openCommand = 0;
+                       }
+
+                       else if (identifyPtr)
+                       {
+                           E4_uint64 tcookie = elan4_sdram_readq (dev, identifyPtr);
+                           
+                           PRINTF (uctx, DBG_NETERR, "user_neterr_sten: cookie=%llx [%llx]\n", tcookie, cookie);
+                           
+                           if (tcookie == cookie)
+                           {
+                               unsigned int vchan = (openCommand >> 4) & 0x1f;
+                               
+                               PRINTF (uctx, DBG_NETERR, "user_neterr_sten: cookie matches - vchan=%d\n", vchan);
+                               
+                               if (! waitforeop)
+                               {
+                                   /* Alter the CQ_AckBuffer for this channel to indicate an
+                                    * ack was received */
+                                   E4_uint64 value  = elan4_sdram_readq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_AckBuffers));
+                                   E4_uint64 nvalue = ((value & ~((E4_uint64)0xf << ((vchan & 0xf) << 2))) |
+                                                       ((E4_uint64) PackOk << ((vchan & 0xf) << 2)));
+                                   
+                                   PRINTF (uctx, DBG_NETERR, "user_neterr_sten: CQ_AckBuffers %llx -> %llx\n", value, nvalue);
+                                   
+                                   elan4_sdram_writeq (dev, cqdesc + offsetof (E4_CommandQueueDesc, CQ_AckBuffers), nvalue);
+                                   pioflush_sdram (dev);
+                               }
+                               
+                               found++;
+                           }
+                           openCommand = 0;
+                       }
+                       
+                       if ((command >> 16) & TR_LAST_AND_SEND_ACK)
+                           openCommand = 0;
+                   }
+                   break;
+                   
+               case 3:
+                   PRINTF (uctx, DBG_NETERR, "user_neterr_sten: invalid command %llx\n", command);
+                   kmutex_unlock (&uctx->uctx_cqlock);
+                   return -EINVAL;
+               }
+               
+           }
+       }
+    }
+    kmutex_unlock (&uctx->uctx_cqlock);
+
+    user_swapin (uctx, UCTX_NETERR_FIXUP);
+
+    return found;
+}
+
+int
+user_neterr_dma (USER_CTXT *uctx, unsigned int vp, E4_uint64 cookie, int waitforeop)
+{
+    unsigned long flags;
+    int found = 0;
+    int idx;
+
+    user_swapout (uctx, UCTX_NETERR_FIXUP);
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    RING_QUEUE_ITERATE (uctx->uctx_dmaQ, idx) {
+       E4_DMA *dma = &uctx->uctx_dmas[idx];
+
+       if (dma->dma_vproc == vp && dma->dma_cookie == cookie)
+       {
+           PRINTF (uctx, DBG_NETERR, "user_neterr_dma: dmaQ matches %s\n", waitforeop ? "waitforeop" : "remove remoteness");
+
+           if (! waitforeop) 
+           {
+               dma->dma_dstEvent = 0;
+               dma->dma_typeSize = DMA_ShMemWrite | DMA_Context (dma->dma_typeSize);
+           }
+           found++;
+       }
+    }
+
+    RING_QUEUE_ITERATE (uctx->uctx_dprocTrapQ, idx) {
+       ELAN4_DPROC_TRAP *trap = &uctx->uctx_dprocTraps[idx];
+
+       if (trap->tr_desc.dma_vproc == vp && trap->tr_desc.dma_cookie == cookie)
+       {
+           PRINTF (uctx, DBG_NETERR, "user_neterr_dma: dmaTrapQ matches %s\n", waitforeop ? "waitforeop" : "remove remoteness");
+
+           if (! waitforeop) 
+           {
+               trap->tr_desc.dma_dstEvent = 0;
+               trap->tr_desc.dma_typeSize = DMA_ShMemWrite | DMA_Context (trap->tr_desc.dma_typeSize);
+           }
+           found++;
+       }
+    }
+
+    /* The device driver command queue should be empty at this point ! */
+    if (user_ddcq_flush (uctx) == 0)
+       found = -EAGAIN;
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    /* The device driver command queue should be empty at this point ! */
+    if (user_ddcq_flush (uctx) == 0)
+       found = -EAGAIN;
+    
+    user_swapin (uctx, UCTX_NETERR_FIXUP);
+
+    return found;
+}
+
+int
+user_trap_handler (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, int nticks)
+{
+    unsigned long entered = jiffies;
+    unsigned int  need_reenter = 0;
+    unsigned long flags;
+    int           res;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    PRINTF1 (uctx, DBG_TRAP, "user_trap_handler: entered state=%d\n", uctx->uctx_trap_state);
+
+    uctx->uctx_trap_count++;
+    
+    for (;;)
+    {
+       if (uctx->uctx_status & UCTX_SWAPPED_REASONS)
+       {
+           PRINTF0 (uctx, DBG_TRAP, "user_trap_handler: exiting on swapped reasons\n");
+           
+           res = UTS_FINISHED;
+           goto no_more_to_do;
+       }
+
+       if ((long) (jiffies - entered) > HZ)
+       {
+           PRINTF0 (uctx, DBG_TRAP, "user_trap_handler: exiting for reschedule\n");
+           res = UTS_RESCHEDULE;
+           goto no_more_to_do;
+       }
+       
+       switch (uctx->uctx_trap_state)
+       {
+       case UCTX_TRAP_ACTIVE:
+           uctx->uctx_trap_state = UCTX_TRAP_SLEEPING;
+           
+           if (nticks == 0 || need_reenter || kcondvar_timedwaitsig (&uctx->uctx_wait, &uctx->uctx_spinlock, &flags, lbolt + nticks) != CV_RET_NORMAL)
+           {
+               PRINTF0 (uctx, DBG_TRAP, "user_trap_handler: exiting by kcondvar_timedwaitsig\n");
+
+               res = UTS_FINISHED;
+               goto no_more_to_do;
+           }
+
+           /* Have slept above, so resample entered */
+           entered = jiffies;
+           
+           uctx->uctx_trap_state = UCTX_TRAP_SIGNALLED;
+           continue;
+
+       case UCTX_TRAP_IDLE:
+       case UCTX_TRAP_SIGNALLED:
+           uctx->uctx_trap_state = UCTX_TRAP_ACTIVE;
+           break;
+       }
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+       PRINTF2 (uctx, DBG_TRAP, "user_trap_handler: resolve traps - state=%d status=%x\n", uctx->uctx_trap_state, uctx->uctx_status);
+
+       switch ((res = resolve_all_traps (uctx, utrapp)))
+       {
+       case UTS_FINISHED:
+           break;
+           
+       case UTS_RESCHEDULE:
+           need_reenter++;
+           break;
+
+       default:
+           spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+           goto no_more_to_do;
+       }
+
+       spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+       if (! user_ddcq_flush (uctx))
+       {
+           PRINTF0 (uctx, DBG_TRAP, "user_trap_handler: ddcq not flushed - re-enter\n");
+           need_reenter++;
+           continue;
+       }
+       spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+       if ((res = progress_neterr (uctx, utrapp)) != UTS_FINISHED)
+       {
+           spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+           goto no_more_to_do;
+       }
+
+       if ((res = execute_iproc_traps (uctx, utrapp)) != UTS_FINISHED)
+       {
+           spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+           goto no_more_to_do;
+       }
+
+       PRINTF2 (uctx, DBG_TRAP, "user_trap_handler: restart items - state=%d status=%x\n", uctx->uctx_trap_state, uctx->uctx_status);
+
+       spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+       if (UCTX_RUNNABLE (uctx))
+       {
+           restart_command_queues (uctx);
+
+           if (! restart_threads (uctx) || ! restart_dmas (uctx))
+           {
+               PRINTF0 (uctx, DBG_TRAP, "user_trap_handler: ddcq full - re-enter\n");
+               need_reenter++;
+           }
+       }
+    }
+ no_more_to_do:
+    uctx->uctx_trap_state = UCTX_TRAP_IDLE;
+
+    /*
+     * Always ensure that the command queue is flushed with a flow control
+     * write, so that on the next trap we (hopefully) find it empty and so
+     * can immediately restart the context.   Also if we need to be re-enter
+     * the trap handler and don't have an interrupt outstanding, then issue
+     * one now.
+     */
+    user_ddcq_flush (uctx);
+    if (need_reenter && uctx->uctx_ddcq_intr == 0)
+    {
+       uctx->uctx_ddcq_intr++;
+       user_ddcq_intr (uctx);
+    }
+
+    if (--uctx->uctx_trap_count == 0 && (uctx->uctx_status & UCTX_SWAPPING))
+       kcondvar_wakeupall (&uctx->uctx_wait, &uctx->uctx_spinlock);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    PRINTF2 (uctx, DBG_TRAP, "user_trap_handler: finished state=%d res=%d\n", uctx->uctx_trap_state, res);
+
+    return (res == UTS_EFAULT ? -EFAULT : 0);
+}
+
+USER_CQ *
+user_alloccq (USER_CTXT *uctx, unsigned cqsize, unsigned perm, unsigned uflags)
+{
+    USER_CQ      *ucq;
+    unsigned long flags;
+
+    KMEM_ZALLOC (ucq, USER_CQ *, sizeof (USER_CQ), 1);
+
+    if (ucq == (USER_CQ *) NULL)
+       return ERR_PTR(-ENOMEM);
+    
+    /* NOTE - do not allow the user to create high-priority queues as we only flush through the low-priority run queues */
+    if ((ucq->ucq_cq = elan4_alloccq (&uctx->uctx_ctxt, cqsize, perm, (uflags & UCQ_REORDER) ? CQ_Reorder : 0)) == NULL)
+    {
+       KMEM_FREE (ucq, sizeof (USER_CQ));
+       
+       PRINTF2 (uctx, DBG_CQ, "user_alloccq: failed elan4_allocq cqsize %d uflags %x\n", cqsize, uflags);
+
+       return ERR_PTR(-ENOMEM);
+    }
+    
+    atomic_set (&ucq->ucq_ref, 1);
+
+    ucq->ucq_state = UCQ_RUNNING;
+    ucq->ucq_flags = uflags;
+    
+    PRINTF3 (uctx, DBG_CQ, "user_alloccq: ucq=%p idx=%d cqnum=%d\n", ucq, elan4_cq2idx (ucq->ucq_cq), elan4_cq2num(ucq->ucq_cq));
+
+    /* chain it onto the context */
+    kmutex_lock (&uctx->uctx_cqlock);
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    list_add (&ucq->ucq_link, &uctx->uctx_cqlist);
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    kmutex_unlock (&uctx->uctx_cqlock);
+
+    return (ucq);
+}
+
+USER_CQ *
+user_findcq (USER_CTXT *uctx, unsigned idx)
+{
+    struct list_head *entry;
+    unsigned long flags;
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    list_for_each (entry, &uctx->uctx_cqlist) {
+       USER_CQ *ucq = list_entry (entry, USER_CQ, ucq_link);
+
+       if (elan4_cq2idx(ucq->ucq_cq) == idx)
+       {
+           atomic_inc (&ucq->ucq_ref);
+           spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+           return (ucq);
+       }
+    }
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    return (NULL);
+}
+
+void
+user_dropcq (USER_CTXT *uctx, USER_CQ *ucq)
+{
+    unsigned long flags;
+
+    PRINTF2 (uctx, DBG_CQ, "user_dropcq: ucq=%p ref=%d\n", ucq, atomic_read (&ucq->ucq_ref));
+
+    kmutex_lock (&uctx->uctx_cqlock);
+    if (! atomic_dec_and_test (&ucq->ucq_ref))
+    {
+       kmutex_unlock (&uctx->uctx_cqlock);
+       return;
+    }
+
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+    list_del (&ucq->ucq_link);
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+
+    kmutex_unlock (&uctx->uctx_cqlock);
+
+    elan4_freecq (&uctx->uctx_ctxt, ucq->ucq_cq);
+
+    KMEM_FREE (ucq, sizeof (USER_CQ));
+}
+
+int
+user_alloc_trap_queues (USER_CTXT *uctx, unsigned ndproc_traps, unsigned neproc_traps, 
+                       unsigned ntproc_traps, unsigned nthreads, unsigned ndmas)
+{
+    ELAN4_DPROC_TRAP *dprocs;
+    ELAN4_EPROC_TRAP *eprocs;
+    ELAN4_TPROC_TRAP *tprocs;
+    E4_DMA           *dmas;
+    E4_ThreadRegs    *threads;
+    ELAN4_NETERR_MSG *msgs;
+    unsigned long     flags;
+
+    int nmsgs = NETERR_MSGS;
+
+    /* bounds check the values that have been passed in */
+    if (ndproc_traps < 2 || ndproc_traps > 10000 ||
+       ntproc_traps < 1 || ntproc_traps > 100   ||
+       neproc_traps < 6 || neproc_traps > 10000 ||
+       nthreads     < 2 || nthreads     > 10000 ||
+       ndmas        < 2 || ndmas        > 10000)
+       return -EINVAL;
+
+    if (uctx->uctx_dmas != NULL)
+       return -EBUSY;
+
+    KMEM_ZALLOC (dprocs, ELAN4_DPROC_TRAP *, ndproc_traps * sizeof (ELAN4_DPROC_TRAP), 1);
+    KMEM_ZALLOC (eprocs, ELAN4_EPROC_TRAP *, neproc_traps * sizeof (ELAN4_EPROC_TRAP), 1);
+    KMEM_ZALLOC (tprocs, ELAN4_TPROC_TRAP *, ntproc_traps * sizeof (ELAN4_TPROC_TRAP), 1);
+    KMEM_ZALLOC (threads, E4_ThreadRegs *, nthreads * sizeof (E4_ThreadRegs), 1);
+    KMEM_ZALLOC (dmas, E4_DMA *, ndmas * sizeof (E4_DMA), 1);
+    KMEM_ZALLOC (msgs, ELAN4_NETERR_MSG *, nmsgs * sizeof (ELAN4_NETERR_MSG), 1);
+
+    if (dprocs == NULL || eprocs == NULL || tprocs == NULL || dmas == NULL || threads == NULL || msgs == NULL)
+    {
+       if (dprocs != NULL) KMEM_FREE (dprocs, ndproc_traps * sizeof (ELAN4_DPROC_TRAP));
+       if (eprocs != NULL) KMEM_FREE (eprocs, neproc_traps * sizeof (ELAN4_EPROC_TRAP));
+       if (tprocs != NULL) KMEM_FREE (tprocs, ntproc_traps * sizeof (ELAN4_TPROC_TRAP));
+       if (threads != NULL) KMEM_FREE (threads, nthreads * sizeof (E4_ThreadRegs));
+       if (dmas != NULL) KMEM_FREE (dmas, ndmas * sizeof (E4_DMA));
+       if (msgs != NULL) KMEM_FREE (msgs, nmsgs * sizeof (ELAN4_NETERR_MSG));
+       
+       return -ENOMEM;
+    }
+    
+    spin_lock_irqsave (&uctx->uctx_spinlock, flags);
+
+    uctx->uctx_dprocTraps = dprocs;
+    uctx->uctx_eprocTraps = eprocs;
+    uctx->uctx_tprocTraps = tprocs;
+    uctx->uctx_threads    = threads;
+    uctx->uctx_dmas       = dmas;
+    uctx->uctx_msgs       = msgs;
+
+    RING_QUEUE_INIT (uctx->uctx_dprocTrapQ, ndproc_traps, 1 /* 1 for 2nd dma */);
+    RING_QUEUE_INIT (uctx->uctx_tprocTrapQ, ntproc_traps, 0);
+    RING_QUEUE_INIT (uctx->uctx_eprocTrapQ, neproc_traps, 5 /* 1 for command, 2 for dma, 2 for inputter */);
+    RING_QUEUE_INIT (uctx->uctx_threadQ,    nthreads,     1);
+    RING_QUEUE_INIT (uctx->uctx_dmaQ,       ndmas,        1);
+    RING_QUEUE_INIT (uctx->uctx_msgQ,       nmsgs,        0);
+
+    spin_unlock_irqrestore (&uctx->uctx_spinlock, flags);
+    
+    return 0;
+}
+
+USER_CTXT *
+user_alloc (ELAN4_DEV *dev)
+{
+    USER_CTXT *uctx;
+    int res;
+    int i;
+
+    /* Allocate and initialise the context private data */
+    KMEM_ZALLOC (uctx, USER_CTXT *, sizeof  (USER_CTXT), 1);
+
+    if (uctx == NULL)
+       return ERR_PTR(-ENOMEM);
+
+    if (elan4_get_position (dev, &uctx->uctx_position) == ELAN_POS_UNKNOWN)
+    {
+       KMEM_FREE (uctx, sizeof (USER_CTXT));
+       return ERR_PTR(-EAGAIN);
+    }
+    
+    if ((res = elan4_insertctxt (dev, &uctx->uctx_ctxt, &user_trap_ops)) != 0)
+    {
+       KMEM_FREE (uctx, sizeof (USER_CTXT));
+       return ERR_PTR(res);
+    }
+
+    KMEM_GETPAGES (uctx->uctx_upage, ELAN4_USER_PAGE *, btopr (sizeof (ELAN4_USER_PAGE)), 1);
+    if (uctx->uctx_upage == NULL)
+    {
+       elan4_removectxt (dev, &uctx->uctx_ctxt);
+       KMEM_FREE (uctx, sizeof (USER_CTXT));
+       return ERR_PTR(-ENOMEM);
+    }
+    
+    if ((uctx->uctx_trampoline = elan4_sdram_alloc (dev, SDRAM_PAGE_SIZE)) == 0)
+    {
+       KMEM_FREEPAGES (uctx->uctx_upage, btopr (sizeof (ELAN4_USER_PAGE)));
+       elan4_removectxt (dev, &uctx->uctx_ctxt);
+
+       KMEM_FREE (uctx, sizeof (USER_CTXT));
+       return ERR_PTR(-ENOMEM);
+    }
+    
+    if ((uctx->uctx_routetable = elan4_alloc_routetable (dev, 4 /* 512 << 4 == 8192 entries */)) == NULL)
+    {
+       elan4_sdram_free (dev, uctx->uctx_trampoline, SDRAM_PAGE_SIZE);
+       KMEM_FREEPAGES (uctx->uctx_upage, btopr (sizeof (ELAN4_USER_PAGE)));
+       elan4_removectxt (dev, &uctx->uctx_ctxt);
+
+       KMEM_FREE (uctx, sizeof (USER_CTXT));
+       return ERR_PTR(-ENOMEM);
+    }
+
+    elan4_set_routetable (&uctx->uctx_ctxt, uctx->uctx_routetable);
+
+    /* initialise the trap and swap queues to be really full */
+    RING_QUEUE_INIT (uctx->uctx_dprocTrapQ, 0, 1);
+    RING_QUEUE_INIT (uctx->uctx_tprocTrapQ, 0, 1);
+    RING_QUEUE_INIT (uctx->uctx_eprocTrapQ, 0, 1);
+    RING_QUEUE_INIT (uctx->uctx_threadQ, 0, 1);
+    RING_QUEUE_INIT (uctx->uctx_dmaQ, 0, 1);
+
+    INIT_LIST_HEAD (&uctx->uctx_cent_list);
+    INIT_LIST_HEAD (&uctx->uctx_vpseg_list);
+    INIT_LIST_HEAD (&uctx->uctx_cqlist);
+
+    uctx->uctx_haltop.op_function = user_flush;
+    uctx->uctx_haltop.op_arg      = uctx;
+    uctx->uctx_haltop.op_mask     = INT_Halted|INT_Discarding;
+
+    uctx->uctx_dma_flushop.op_function = user_flush_dmas;
+    uctx->uctx_dma_flushop.op_arg      = uctx;
+
+    kmutex_init (&uctx->uctx_vpseg_lock);
+    kmutex_init (&uctx->uctx_cqlock);
+    kmutex_init (&uctx->uctx_rgnmutex);
+
+    spin_lock_init (&uctx->uctx_spinlock);
+    spin_lock_init (&uctx->uctx_rgnlock);
+    spin_lock_init (&uctx->uctx_fault_lock);
+
+    kcondvar_init (&uctx->uctx_wait);
+
+    if ((uctx->uctx_ddcq = user_alloccq (uctx, CQ_Size1K, CQ_EnableAllBits, UCQ_SYSTEM)) == NULL)
+    {
+       user_free (uctx);
+       return ERR_PTR(-ENOMEM);
+    }
+
+    uctx->uctx_trap_count = 0;
+    uctx->uctx_trap_state = UCTX_TRAP_IDLE;
+    uctx->uctx_status     = 0 /* UCTX_DETACHED | UCTX_SWAPPED | UCTX_STOPPED */;
+
+    init_timer (&uctx->uctx_int_timer);
+
+    uctx->uctx_int_timer.function = user_signal_timer;
+    uctx->uctx_int_timer.data     = (unsigned long) uctx;
+    uctx->uctx_int_start          = jiffies;
+    uctx->uctx_int_count          = 0;
+    uctx->uctx_int_delay          = 0;
+
+    init_timer (&uctx->uctx_neterr_timer);
+    uctx->uctx_neterr_timer.function = user_neterr_timer;
+    uctx->uctx_neterr_timer.data     = (unsigned long) uctx;
+
+    uctx->uctx_upage->upage_ddcq_completed = 0;
+    uctx->uctx_ddcq_completed              = 0;
+    uctx->uctx_ddcq_insertcnt              = 0;
+
+    uctx->uctx_num_fault_save = num_fault_save;
+    if (uctx->uctx_num_fault_save) 
+    {  
+       KMEM_ZALLOC (uctx->uctx_faults, FAULT_SAVE *, (sizeof(FAULT_SAVE) * uctx->uctx_num_fault_save), 1);
+        if ( uctx->uctx_faults == NULL) 
+       {
+           user_free (uctx);
+           return ERR_PTR(-ENOMEM);
+        }
+    
+        for (i = 0; i < uctx->uctx_num_fault_save; i++)
+           uctx->uctx_faults[i].next = (i == (uctx->uctx_num_fault_save-1) ? NULL : &uctx->uctx_faults[i+1]);
+
+    }
+    uctx->uctx_fault_list = uctx->uctx_faults;
+
+    return (uctx);
+}
+
+void
+user_free (USER_CTXT *uctx)
+{
+    ELAN4_DEV *dev = uctx->uctx_ctxt.ctxt_dev;
+
+    user_swapout (uctx, UCTX_EXITING);
+
+    /* Detach from all input contexts */
+    user_detach (uctx, NULL);
+
+    /* since we're single threaded here - (called from close()) */
+    /* we don't need to hold the lock to drop the command queues */
+    /* since they cannot be mapped into user space */
+    while (! list_empty (&uctx->uctx_cqlist))
+       user_dropcq (uctx, list_entry (uctx->uctx_cqlist.next, USER_CQ, ucq_link));
+
+    /* Free off all of vpseg_list */
+    kmutex_lock (&uctx->uctx_vpseg_lock);
+    while (! list_empty (&uctx->uctx_vpseg_list))
+       user_remove_vpseg(uctx, list_entry (uctx->uctx_vpseg_list.next, USER_VPSEG, vps_link));
+    kmutex_unlock (&uctx->uctx_vpseg_lock);
+    
+    if (timer_pending (&uctx->uctx_int_timer))
+       del_timer_sync (&uctx->uctx_int_timer);
+    
+    if (timer_pending (&uctx->uctx_neterr_timer))
+       del_timer_sync (&uctx->uctx_neterr_timer);
+
+    if (uctx->uctx_dprocTraps)
+       KMEM_FREE (uctx->uctx_dprocTraps, uctx->uctx_dprocTrapQ.q_size * sizeof (ELAN4_DPROC_TRAP));
+    if (uctx->uctx_tprocTraps)
+       KMEM_FREE (uctx->uctx_tprocTraps, uctx->uctx_tprocTrapQ.q_size * sizeof (ELAN4_TPROC_TRAP));
+    if (uctx->uctx_eprocTraps)
+       KMEM_FREE (uctx->uctx_eprocTraps, uctx->uctx_eprocTrapQ.q_size * sizeof (ELAN4_EPROC_TRAP));
+    if (uctx->uctx_dmas)
+       KMEM_FREE (uctx->uctx_dmas, uctx->uctx_dmaQ.q_size * sizeof (E4_DMA));
+    if (uctx->uctx_msgs)
+       KMEM_FREE (uctx->uctx_msgs, NETERR_MSGS * sizeof (ELAN4_NETERR_MSG));
+    if (uctx->uctx_threads)
+       KMEM_FREE (uctx->uctx_threads, uctx->uctx_threadQ.q_size * sizeof (E4_ThreadRegs));
+    if (uctx->uctx_faults)
+       KMEM_FREE (uctx->uctx_faults, (sizeof(FAULT_SAVE) * uctx->uctx_num_fault_save));
+
+    if (uctx->uctx_intcookie_table)
+       intcookie_free_table (uctx->uctx_intcookie_table);
+
+    elan4_set_routetable (&uctx->uctx_ctxt, NULL);
+    elan4_free_routetable (dev, uctx->uctx_routetable);
+
+    /* Free off all USER_RGNs */
+    user_freergns(uctx);
+
+    elan4_sdram_free (dev, uctx->uctx_trampoline, SDRAM_PAGE_SIZE);
+
+    /* Clear the PG_Reserved bit before free to avoid a memory leak */
+    ClearPageReserved(pte_page(*find_pte_kernel((unsigned long) uctx->uctx_upage)));
+    KMEM_FREEPAGES (uctx->uctx_upage, btopr (sizeof (ELAN4_USER_PAGE)));
+
+    elan4_removectxt (dev, &uctx->uctx_ctxt);
+
+    kcondvar_destroy (&uctx->uctx_wait);
+
+    spin_lock_destroy (&uctx->uctx_rgnlock);
+    spin_lock_destroy (&uctx->uctx_spinlock);
+
+    kmutex_destroy (&uctx->uctx_rgnmutex);
+    kmutex_destroy (&uctx->uctx_cqlock);
+    kmutex_destroy (&uctx->uctx_vpseg_lock);
+
+    KMEM_FREE (uctx, sizeof (USER_CTXT));
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/user_ddcq.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/user_ddcq.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/user_ddcq.c    2005-06-01 23:12:54.625435152 -0400
@@ -0,0 +1,226 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: user_ddcq.c,v 1.15 2004/06/23 11:06:05 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/user_ddcq.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/user.h>
+#include <elan4/commands.h>
+
+#if PAGE_SIZE < CQ_CommandMappingSize
+#  define ELAN4_COMMAND_QUEUE_MAPPING  PAGE_SIZE
+#else
+#  define ELAN4_COMMAND_QUEUE_MAPPING  CQ_CommandMappingSize
+#endif
+
+/* The user device driver command queue is used for re-issuing 
+ * trapped items.  It is allocated as a 1K command queue, and
+ * we insert command flow writes event 256 words.
+ */
+#define USER_CTRLFLOW_COUNT    256
+
+/* Flow control of the device driver command queue is handled by periodically 
+ * inserting dword writes into the command stream.   When you need to know 
+ * that the queue has been flushed, then you insert an extra contorl flow
+ * write into the command queue.   Should the queue not be flushed, but the
+ * trap handler be returning to user space, then it will also insert and 
+ * extra interrupt command to ensure that it is re-entered after the queue
+ * has been flushed.
+ *
+ * Note - we account the space for the interrupt command on each control 
+ * flow write so that we do not overflow the queue even if we end up 
+ * inserting an interrupt for every command flow write.  In general only
+ * a single interrupt should get inserted....
+ */
+
+#define user_ddcq_command_write(value,off) do { \
+    PRINTF(uctx, DBG_DDCQ, "user_ddcq_command_write: cmdptr=%x off=%d value=%llx\n", cmdptr, off, value);\
+    writeq(value, cmdptr + (off << 3)); \
+} while (0)
+
+#define user_ddcq_command_space(uctx)  \
+    ((CQ_Size (uctx->uctx_ddcq->ucq_cq->cq_size)>>3) - ((uctx)->uctx_ddcq_insertcnt - (uctx)->uctx_upage->upage_ddcq_completed))
+
+#define user_ddcq_command_flow_write(uctx) do { \
+   E4_uint64 iptr   = (uctx)->uctx_ddcq_insertcnt; \
+   ioaddr_t  cmdptr = (uctx)->uctx_ddcq->ucq_cq->cq_mapping + ((iptr<<3) & ((ELAN4_COMMAND_QUEUE_MAPPING >> 1)-1));\
+\
+    (uctx)->uctx_ddcq_completed = ((uctx)->uctx_ddcq_insertcnt += 3);\
+\
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_command_flow_write: completed=%llx [%llx] addr=%llx\n", (uctx)->uctx_ddcq_completed, \
+           (uctx)->uctx_upage->upage_ddcq_completed, (uctx)->uctx_upage_addr); \
+    user_ddcq_command_write (GUARD_CMD       | GUARD_ALL_CHANNELS,      0);\
+    user_ddcq_command_write (WRITE_DWORD_CMD | (uctx)->uctx_upage_addr, 1);\
+    user_ddcq_command_write ((uctx)->uctx_ddcq_completed,               2);\
+} while (0)
+
+#define user_ddcq_command_flow_intr(uctx) do { \
+   E4_uint64 iptr   = (uctx)->uctx_ddcq_insertcnt; \
+   ioaddr_t  cmdptr = (uctx)->uctx_ddcq->ucq_cq->cq_mapping + ((iptr<<3) & ((ELAN4_COMMAND_QUEUE_MAPPING >> 1)-1));\
+\
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_command_flow_intr: completed=%llx [%llx] addr=%llx\n", (uctx)->uctx_ddcq_completed, \
+           (uctx)->uctx_upage->upage_ddcq_completed, (uctx)->uctx_upage_addr); \
+    user_ddcq_command_write (INTERRUPT_CMD   | ELAN4_INT_COOKIE_DDCQ,   3);\
+} while (0)
+
+#define user_ddcq_command_prologue(uctx, count) do { \
+   E4_uint64 iptr   = (uctx)->uctx_ddcq_insertcnt; \
+   ioaddr_t  cmdptr = (uctx)->uctx_ddcq->ucq_cq->cq_mapping + ((iptr<<3) & ((ELAN4_COMMAND_QUEUE_MAPPING >> 1)-1));\
+   PRINTF(uctx, DBG_DDCQ, "user_ddcq_command_prologue: iptr=%llx cmdptr=%x\n", iptr, cmdptr);
+
+#define user_ddcq_command_epilogue(uctx, count, extra) \
+   (uctx)->uctx_ddcq_insertcnt = iptr + (count);\
+\
+   PRINTF(uctx, DBG_DDCQ, "user_ddcq_command_epilogue: iptr=%llx + %x + %x - completed %llx\n", iptr, count, extra, (uctx)->uctx_ddcq_completed);\
+   if (((iptr) + (count) + (extra)) > ((uctx)->uctx_ddcq_completed + USER_CTRLFLOW_COUNT))\
+       user_ddcq_command_flow_write(uctx); \
+} while (0)
+
+int
+user_ddcq_check (USER_CTXT *uctx, unsigned num)
+{
+    PRINTF (uctx, DBG_DDCQ, "user_check_ddcq: insert=%llx completed=%llx num=%d\n", 
+           uctx->uctx_ddcq_insertcnt, uctx->uctx_upage->upage_ddcq_completed, num);
+
+    /* Ensure that there is enough space for the command we want to issue,
+     * PLUS the guard/writeword for the control flow flush.
+     * PLUS the interrupt command for rescheduling */
+    if (user_ddcq_command_space (uctx) > (num + 4))
+    {
+       PRINTF (uctx, DBG_DDCQ, "user_ddcq_check: loads of space\n");
+
+       return (1);
+    }
+    
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_check: not enough space - reschedule\n");
+
+    uctx->uctx_trap_state = UCTX_TRAP_SIGNALLED;
+    return (0);
+}
+
+int
+user_ddcq_flush (USER_CTXT *uctx)
+{
+    ELAN4_DEV *dev = uctx->uctx_ctxt.ctxt_dev;
+    USER_CQ   *ucq = uctx->uctx_ddcq;
+
+    switch (ucq->ucq_state)
+    {
+    case UCQ_TRAPPED:
+       PRINTF (uctx, DBG_DDCQ, "user_ddcq_flush: command queue is trapped\n");
+       return (0);
+       
+    case UCQ_NEEDS_RESTART:
+       PRINTF (uctx, DBG_DDCQ, "user_ddcq_flush: restarting command queue\n");
+
+       if (UCTX_RUNNABLE (uctx))
+       {
+           ucq->ucq_state = UCQ_RUNNING;
+           elan4_restartcq (dev, ucq->ucq_cq);
+       }
+       break;
+    }
+
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_flush: insertcnt=%llx completed=%llx [%llx]\n", 
+           uctx->uctx_ddcq_insertcnt, uctx->uctx_ddcq_completed, uctx->uctx_upage->upage_ddcq_completed);
+
+    if (uctx->uctx_ddcq_completed != uctx->uctx_ddcq_insertcnt)
+       user_ddcq_command_flow_write (uctx);
+
+    return (uctx->uctx_ddcq_completed == uctx->uctx_upage->upage_ddcq_completed);
+}
+
+void
+user_ddcq_intr (USER_CTXT *uctx)
+{
+    user_ddcq_command_flow_intr (uctx);
+}
+
+void
+user_ddcq_run_dma (USER_CTXT *uctx, E4_DMA *dma)
+{
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_run_dma: cookie=%llx vproc=%llx\n",  dma->dma_cookie, dma->dma_vproc);
+
+    user_ddcq_command_prologue(uctx, 7) {
+
+       user_ddcq_command_write ((dma->dma_typeSize & ~DMA_ContextMask) | RUN_DMA_CMD, 0);
+       user_ddcq_command_write (dma->dma_cookie,                                      1);
+       user_ddcq_command_write (dma->dma_vproc,                                       2);
+       user_ddcq_command_write (dma->dma_srcAddr,                                     3);
+       user_ddcq_command_write (dma->dma_dstAddr,                                     4);
+       user_ddcq_command_write (dma->dma_srcEvent,                                    5);
+       user_ddcq_command_write (dma->dma_dstEvent,                                    6);
+
+    } user_ddcq_command_epilogue (uctx, 7, 0);
+}
+
+void
+user_ddcq_run_thread (USER_CTXT *uctx, E4_ThreadRegs *regs)
+{
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_run_thread: PC=%llx SP=%llx\n", regs->Registers[0], regs->Registers[1]);
+
+    user_ddcq_command_prologue(uctx, 7) {
+
+       user_ddcq_command_write (regs->Registers[0] | RUN_THREAD_CMD, 0);
+       user_ddcq_command_write (regs->Registers[1],                  1);
+       user_ddcq_command_write (regs->Registers[2],                  2);
+       user_ddcq_command_write (regs->Registers[3],                  3);
+       user_ddcq_command_write (regs->Registers[4],                  4);
+       user_ddcq_command_write (regs->Registers[5],                  5);
+       user_ddcq_command_write (regs->Registers[6],                  6);
+       
+    } user_ddcq_command_epilogue (uctx, 7, 0);
+}
+
+void
+user_ddcq_setevent (USER_CTXT *uctx, E4_Addr addr)
+{
+    user_ddcq_command_prologue (uctx, 1) {
+
+       user_ddcq_command_write (SET_EVENT_CMD | addr, 0);
+    
+    } user_ddcq_command_epilogue (uctx, 1, 0);
+}
+
+void
+user_ddcq_seteventn (USER_CTXT *uctx, E4_Addr addr, E4_uint32 count)
+{
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_seteventn: addr=%llx count=%lx\n", addr, count);
+
+    user_ddcq_command_prologue (uctx, 2) {
+
+       user_ddcq_command_write (SET_EVENTN_CMD, 0);
+       user_ddcq_command_write (addr | count,   1);
+
+    } user_ddcq_command_epilogue (uctx, 2, 0);
+}
+
+void
+user_ddcq_waitevent (USER_CTXT *uctx, E4_Addr addr, E4_uint64 CountAndType, E4_uint64 Param0, E4_uint64 Param1)
+{
+    PRINTF (uctx, DBG_DDCQ, "user_ddcq_waitevent: addr=%llx CountAndType=%llx Param=%llx,%llx\n", addr, CountAndType, Param0, Param1);
+
+    user_ddcq_command_prologue (uctx, 4) {
+
+       user_ddcq_command_write (WAIT_EVENT_CMD | addr, 0);
+       user_ddcq_command_write (CountAndType,          1);
+       user_ddcq_command_write (Param0,                2);
+       user_ddcq_command_write (Param1,                3);
+
+    } user_ddcq_command_epilogue (uctx, 4, 0);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/elan4/user_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/elan4/user_Linux.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/elan4/user_Linux.c   2005-06-01 23:12:54.626435000 -0400
@@ -0,0 +1,377 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: user_Linux.c,v 1.25.2.4 2005/01/18 14:36:10 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/user_Linux.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kpte.h>
+
+#include <linux/pci.h>
+
+#include <elan4/debug.h>
+#include <elan4/device.h>
+#include <elan4/user.h>
+
+static int
+user_pteload (USER_CTXT *uctx, E4_Addr addr, physaddr_t phys, int perm)
+{
+    ELAN4_DEV *dev = uctx->uctx_ctxt.ctxt_dev;
+    E4_uint64  newpte = elan4mmu_phys2pte (dev, phys, perm);
+    
+    /*
+     * On MPSAS we don't allocate a large enough context table, so 
+     * if we see an address/context pair which would "alias" because
+     * they differ in unchecked hash bits to a previous pteload, 
+     * then we kill the application.
+     */
+    {
+       unsigned hashval = (E4MMU_SHIFT_ADDR(addr, (dev->dev_pageshift[0]) + 2) ^ E4MMU_CONTEXT_SCRAMBLE(uctx->uctx_ctxt.ctxt_num));
+       
+       if (dev->dev_rsvd_hashval[0] == 0xFFFFFFFF)
+           dev->dev_rsvd_hashval[0] = hashval & dev->dev_rsvd_hashmask[0];
+       
+       if ((hashval & dev->dev_rsvd_hashmask[0]) != dev->dev_rsvd_hashval[0])
+       {
+           printk ("user_pteload: vaddr=%016llx ctxnum=%x -> [%x] overlaps %x - %x [hashidx=%x]\n", (unsigned long long) addr, 
+                   uctx->uctx_ctxt.ctxt_num, hashval, hashval & dev->dev_rsvd_hashmask[0], dev->dev_rsvd_hashval[0],
+                   E4MMU_HASH_INDEX (uctx->uctx_ctxt.ctxt_num, addr, dev->dev_pageshift[0], dev->dev_hashsize[0]-1));
+           
+           return -EFAULT;
+       }
+    }
+
+    if ((newpte & (PTE_PciNotLocal | PTE_CommandQueue)) == 0 && 
+       ((addr & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT)) != (phys & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT))))
+    {
+       printk ("user_pteload: vaddr=%016llx incorrectly alias sdram at %lx\n", (unsigned long long) addr, 
+               phys ^ pci_resource_start (dev->dev_osdep.pdev, ELAN4_BAR_SDRAM));
+       return -EFAULT;
+    }
+
+    if (newpte & PTE_PciNotLocal)
+       PRINTF (uctx, DBG_FAULT, "user_pteload: addr=%llx -> pte=%llx (pci)\n", addr, newpte);
+    else if (newpte & PTE_CommandQueue)
+       PRINTF (uctx, DBG_FAULT, "user_pteload: addr=%llx -> pte=%llx (command)\n", addr, newpte);
+    else
+       PRINTF (uctx, DBG_FAULT, "user_pteload: addr=%llx -> pte=%llx (sdram)\n", addr, newpte);
+
+    elan4mmu_pteload (&uctx->uctx_ctxt, 0, addr, newpte);
+
+    return (0);
+}
+
+int
+user_load_range (USER_CTXT *uctx, E4_Addr eaddr, unsigned long nbytes, E4_uint32 fsr)
+{
+    ELAN4_DEV             *dev       = uctx->uctx_ctxt.ctxt_dev;
+    struct mm_struct      *mm        = current->mm;
+    int                    writeable = (AT_Perm(fsr) == AT_PermLocalDataWrite ||
+                                       AT_Perm(fsr) == AT_PermRemoteWrite    ||
+                                       AT_Perm(fsr) == AT_PermLocalEvent     ||
+                                       AT_Perm(fsr) == AT_PermRemoteEvent);
+    struct vm_area_struct *vma;
+    int                    i, perm;
+    unsigned long          len;
+    unsigned long          maddr;
+    physaddr_t            phys;
+
+    kmutex_lock (&uctx->uctx_rgnmutex);
+
+    while (nbytes > 0)
+    {
+       USER_RGN *rgn = user_rgnat_elan (uctx, eaddr);
+
+       if (rgn == NULL || ELAN4_INCOMPAT_ACCESS (rgn->rgn_perm, AT_Perm (fsr)))
+       {
+           PRINTF (uctx, DBG_FAULT, "user_load_range: eaddr=%llx -> %s\n", eaddr, rgn == NULL ? "no mapping" : "no permission");
+
+           kmutex_unlock (&uctx->uctx_rgnmutex);
+           return (rgn == NULL ? -EFAULT : -EPERM);
+       }
+
+       if (writeable)
+           perm = rgn->rgn_perm;
+/* This is the correct code but it breaks the Eagle libraries (1.6.X) - backed out (addy 24.08.04)
+       else if (AT_Perm(fsr) == AT_PermExecute && (rgn->rgn_perm & PERM_Mask) != PERM_LocExecute)
+*/
+       else if (AT_Perm(fsr) == AT_PermExecute)
+           perm = PERM_LocRead | (rgn->rgn_perm & ~PERM_Mask);
+       else
+           perm = ELAN4_PERM_READONLY (rgn->rgn_perm & PERM_Mask) | (rgn->rgn_perm & ~PERM_Mask);
+
+       PRINTF (uctx, DBG_FAULT, "user_load_range: rgn=%p [%llx.%lx.%x]\n", rgn, rgn->rgn_ebase, rgn->rgn_mbase, rgn->rgn_len);
+
+       len = ((rgn->rgn_ebase + rgn->rgn_len) - eaddr);
+       if (len > nbytes)
+           len = nbytes;
+       nbytes -= len;
+       
+       maddr = rgn->rgn_mbase + (eaddr - rgn->rgn_ebase);
+
+       PRINTF (uctx, DBG_FAULT, "user_load_range: eaddr=%llx->%llx -> %lx->%lx len=%x perm=%x\n", eaddr, 
+               eaddr + len, maddr, maddr + len, len, perm);
+
+       down_read (&mm->mmap_sem);
+       while (len > 0)
+       {
+           if ((vma = find_vma_intersection (mm, maddr, maddr + PAGE_SIZE)) == NULL ||
+               (writeable && !(vma->vm_flags & VM_WRITE)))
+           {
+               PRINTF (DBG_USER, DBG_FAULT, "ctxt_pagefault: %s %lx\n", vma ? "no writeble at" : "no vma for", maddr);
+               up_read (&mm->mmap_sem);
+               kmutex_unlock (&uctx->uctx_rgnmutex);
+               return (-EFAULT);
+           }
+
+           spin_lock (&mm->page_table_lock);
+           {
+               pte_t *ptep_ptr;
+               pte_t  ptep_value;
+
+               ptep_ptr = find_pte_map (mm, maddr);
+               if (ptep_ptr) {
+                   ptep_value = *ptep_ptr;
+                   pte_unmap(ptep_ptr);
+               }
+
+               PRINTF (uctx, DBG_FAULT, "user_load_range: %lx %s %s\n", maddr, writeable ? "writeable" : "readonly", 
+                       !ptep_ptr ? "invalid" : pte_none(ptep_value) ? "none " : !pte_present(ptep_value) ? "swapped " : 
+                       writeable && !pte_write(ptep_value) ? "COW" : "OK");
+               
+               if (ptep_ptr == NULL || pte_none(ptep_value) || !pte_present(ptep_value) || (writeable && !pte_write(ptep_value)) || !pte_read (ptep_value))
+               {
+                   spin_unlock (&mm->page_table_lock);
+                   
+                   make_pages_present(maddr, maddr + PAGE_SIZE);
+                   
+                   spin_lock (&mm->page_table_lock);
+
+                   ptep_ptr = find_pte_map (mm, maddr);
+                   if (ptep_ptr) {
+                       ptep_value = *ptep_ptr;
+                       pte_unmap(ptep_ptr);
+                   }
+                   
+                   if (ptep_ptr == NULL || pte_none(ptep_value) || !pte_present(ptep_value) || (writeable && !pte_write(ptep_value)) || !pte_read (ptep_value))
+                   {   
+                       spin_unlock (&mm->page_table_lock);
+                       up_read (&mm->mmap_sem);
+                       kmutex_unlock (&uctx->uctx_rgnmutex);
+                       return (-EFAULT);
+                   }
+               } 
+               
+               if (writeable)
+                   pte_mkdirty(ptep_value);
+               pte_mkyoung (ptep_value);
+
+               phys = pte_phys (ptep_value);
+
+               for (i = 0; i < PAGE_SIZE; i += (1 << dev->dev_pageshift[0]))
+               {
+                   if (user_pteload (uctx, eaddr, phys, perm) < 0)
+                   {
+                       spin_unlock (&mm->page_table_lock);
+                       up_read (&mm->mmap_sem);
+                       kmutex_unlock (&uctx->uctx_rgnmutex);
+                       return (-EFAULT);
+                   }
+                   
+                   eaddr += (1 << dev->dev_pageshift[0]);
+                   phys  += (1 << dev->dev_pageshift[0]);
+               }
+           }
+           spin_unlock (&mm->page_table_lock);
+               
+           maddr += PAGE_SIZE;
+           len   -= PAGE_SIZE;
+       }
+       up_read (&mm->mmap_sem);
+    }
+    kmutex_unlock (&uctx->uctx_rgnmutex);
+
+    PRINTF (uctx, DBG_FAULT, "user_load_range: alldone\n");
+
+    return (0);
+}
+
+void
+user_preload_main (USER_CTXT *uctx, virtaddr_t addr, unsigned long len)
+{
+    virtaddr_t             lim = addr + len - 1;
+    struct vm_area_struct *vma;
+
+    down_read (&current->mm->mmap_sem);
+
+    if ((vma = find_vma (current->mm, addr)) != NULL)
+    {
+       do {
+           unsigned long start = vma->vm_start;
+           unsigned long end   = vma->vm_end;
+
+           if ((start-1) >= lim)
+               break;
+
+           if (start < addr) start = addr;
+           if ((end-1) > lim) end = lim+1;
+               
+           if (vma->vm_flags & VM_IO)
+               continue;
+
+           user_unload_main (uctx, start, end - start);
+
+           make_pages_present (start, end);
+
+           user_update_main (uctx, current->mm, start, end - start);
+
+       } while ((vma = find_vma (current->mm, vma->vm_end)) != NULL);
+    }
+    up_read (&current->mm->mmap_sem);
+}
+
+static void
+user_update_range (USER_CTXT *uctx, int tbl, struct mm_struct *mm, virtaddr_t maddr, E4_Addr eaddr, unsigned long len, int perm)
+{
+    ELAN4_DEV *dev    = uctx->uctx_ctxt.ctxt_dev;
+    int        roperm = ELAN4_PERM_READONLY(perm & PERM_Mask) | (perm & ~PERM_Mask);
+    int        nbytes;
+
+    while (len > 0)
+    {
+       pte_t *ptep_ptr;
+       pte_t  ptep_value;
+       
+       ptep_ptr = find_pte_map (mm, maddr);
+       if (ptep_ptr) {
+           ptep_value = *ptep_ptr;
+           pte_unmap(ptep_ptr);
+       }
+
+       PRINTF (uctx, DBG_IOPROC, "user_update_range: %llx (%lx) %s\n", eaddr, maddr, 
+               !ptep_ptr ? "invalid" : pte_none(ptep_value) ? "none " : !pte_present(ptep_value) ? "swapped " : 
+               !pte_write(ptep_value) ? "RO/COW" : "OK");
+       
+       if (ptep_ptr && !pte_none(ptep_value) && pte_present(ptep_value) && pte_read (ptep_value)) {
+           physaddr_t phys_value = pte_phys(ptep_value);
+           for (nbytes = 0; nbytes < PAGE_SIZE; nbytes += (1 << dev->dev_pageshift[0]))
+           {
+               user_pteload (uctx, eaddr, phys_value, pte_write (ptep_value) ? perm : roperm);
+
+               eaddr       += (1 << dev->dev_pageshift[0]);
+               phys_value  += (1 << dev->dev_pageshift[0]);
+           }
+       }
+
+       maddr += PAGE_SIZE;
+       len   -= PAGE_SIZE;
+    }
+}
+
+void
+user_update_main (USER_CTXT *uctx, struct mm_struct *mm, virtaddr_t start, unsigned long len)
+{
+    USER_RGN     *rgn;
+    unsigned long ssize;
+    virtaddr_t    end = start + len - 1;
+
+    spin_lock (&uctx->uctx_rgnlock);
+
+    PRINTF (uctx, DBG_IOPROC, "user_update_main: start=%lx end=%lx\n", start, end);
+
+    for (rgn = user_findrgn_main (uctx, start, 0); rgn != NULL; rgn = rgn->rgn_mnext)
+    {
+       if (end < rgn->rgn_mbase)
+           break;
+       
+       if (start <= rgn->rgn_mbase && end >= (rgn->rgn_mbase + rgn->rgn_len - 1)) 
+       {
+           PRINTF (uctx, DBG_IOPROC, "user_update_main: whole %lx -> %lx\n", rgn->rgn_mbase, rgn->rgn_mbase + rgn->rgn_len - 1);
+
+           user_update_range (uctx, 0 /* tbl */, mm, rgn->rgn_mbase, rgn->rgn_ebase, rgn->rgn_len, rgn->rgn_perm);
+       }
+       else if (start <= rgn->rgn_mbase)
+       {
+           ssize = end - rgn->rgn_mbase + 1;
+
+           PRINTF (uctx, DBG_IOPROC, "user_update_main: start %lx -> %lx\n", rgn->rgn_mbase, rgn->rgn_mbase + ssize);
+
+           user_update_range (uctx, 0 /* tbl */, mm, rgn->rgn_mbase, rgn->rgn_ebase, ssize, rgn->rgn_perm);
+       }
+       else if (end >= (rgn->rgn_mbase + rgn->rgn_len - 1))
+       {
+           ssize = (rgn->rgn_mbase + rgn->rgn_len) - start;
+
+           PRINTF (uctx, DBG_IOPROC, "user_update_main: end   %lx -> %lx\n", start, start + ssize);
+
+           user_update_range (uctx, 0 /* tbl */, mm, start, rgn->rgn_ebase + (start - rgn->rgn_mbase), ssize, rgn->rgn_perm);
+       }
+       else
+       {
+           PRINTF (uctx, DBG_IOPROC, "user_update_main: middle %lx -> %lx\n", start, end);
+
+           user_update_range (uctx, 0 /* tbl */, mm, start, rgn->rgn_ebase + (start - rgn->rgn_mbase), len, rgn->rgn_perm);
+       }
+    }
+    spin_unlock (&uctx->uctx_rgnlock);
+}
+
+void
+user_unload_main (USER_CTXT *uctx, virtaddr_t start, unsigned long len)
+{
+    USER_RGN     *rgn;
+    unsigned long ssize;
+    virtaddr_t    end = start + len - 1;
+
+    spin_lock (&uctx->uctx_rgnlock);
+
+    PRINTF (uctx, DBG_IOPROC, "user_unload_main: start=%lx end=%lx\n", start, end);
+
+    for (rgn = user_findrgn_main (uctx, start, 0); rgn != NULL; rgn = rgn->rgn_mnext)
+    {
+       if (end < rgn->rgn_mbase)
+           break;
+       
+       if (start <= rgn->rgn_mbase && end >= (rgn->rgn_mbase + rgn->rgn_len - 1))
+       {
+           PRINTF (uctx, DBG_IOPROC, "user_unload_main: whole %lx -> %lx\n", rgn->rgn_mbase, rgn->rgn_mbase + rgn->rgn_len - 1);
+
+           elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* tbl */, rgn->rgn_ebase, rgn->rgn_len);
+       }
+       else if (start <= rgn->rgn_mbase)
+       {
+           ssize = end - rgn->rgn_mbase + 1;
+
+           PRINTF (uctx, DBG_IOPROC, "user_unload_main: start %lx -> %lx\n", rgn->rgn_mbase, rgn->rgn_mbase + ssize);
+
+           elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* tbl */, rgn->rgn_ebase, ssize);
+       }
+       else if (end >= (rgn->rgn_mbase + rgn->rgn_len - 1))
+       {
+           ssize = (rgn->rgn_mbase + rgn->rgn_len) - start;
+           
+           PRINTF (uctx, DBG_IOPROC, "user_unload_main: end   %lx -> %lx\n", start, start + ssize);
+           
+           elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* tbl */, rgn->rgn_ebase + (start - rgn->rgn_mbase), ssize);
+       }
+       else
+       {
+
+           PRINTF (uctx, DBG_IOPROC, "user_unload_main: middle %lx -> %lx\n", start, end);
+
+           elan4mmu_unload_range (&uctx->uctx_ctxt, 0 /* tbl */, rgn->rgn_ebase + (start - rgn->rgn_mbase), len);
+       }
+    }
+    spin_unlock (&uctx->uctx_rgnlock);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/asm_elan4_thread.S
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/asm_elan4_thread.S   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/asm_elan4_thread.S        2005-06-01 23:12:54.626435000 -0400
@@ -0,0 +1,78 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: asm_elan4_thread.S,v 1.1 2003/09/23 13:55:11 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/asm_elan4_thread.S,v $*/
+
+#include <elan4/events.h>
+#include <elan4/commands.h>
+
+/*
+ * c_reschedule (E4_uint64 *commandport)
+ */            
+       .global c_reschedule
+c_reschedule:
+       add             %sp, -128, %sp
+       st64            %r16, [%sp]                     // preserve call preserved registers
+       st64            %r24, [%sp + 64]                // - see CALL_USED_REGISTERS.
+       mov             %r16,%r16                       // BUG FIX: E4 RevA
+       mov             %r24,%r24                       // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       
+       mov             %r7, %r18                       // (%r2) return pc
+1:     call            2f
+        mov            %sp, %r17                       // (%r1) SP
+2:     add             %r7, (3f-1b), %r16              // (%r0) PC
+       mov             NOP_CMD, %r23                   // "nop" command
+       st64suspend     %r16, [%r8]
+3:     ld64            [%sp], %r16
+       ld64            [%sp + 64], %r24                // restore call preserved register
+       jmpl            %r2+8, %r0                      // and return
+        add            %sp, 128, %sp
+       
+
+/*
+ * c_waitevent (E4_uint64 *commandport, E4_Event *event, E4_uint64 count)
+ */
+       .global c_waitevent
+c_waitevent:
+       add             %sp, -192, %sp
+       st64            %r16, [%sp + 64]                // preserve call preserved registers
+       st64            %r24, [%sp + 128]               // - see CALL_USED_REGISTERS.
+       mov             %r16,%r16                       // BUG FIX: E4 RevA
+       mov             %r24,%r24                       // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+
+       mov             %r7, %r18                       // (%r2) return pc
+1:     call            2f
+        mov            %sp, %r17                       // (%r1) SP
+2:     add             %r7, (3f-1b), %r16              // (%r0) PC
+       st32            %r16, [%sp]                     // event source block
+       mov             MAKE_EXT_CLEAN_CMD, %r23        // "flush command queue desc" command
+       st8             %r23, [%sp+56]                  // event source block
+       mov             %r16,%r16                       // BUG FIX: E4 RevA
+       mov             %r23,%r23                       // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       
+
+       or              %r9, WAIT_EVENT_CMD, %r16
+       sll8            %r10, 32, %r17
+       or              %r17, E4_EVENT_TYPE_VALUE(E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, 8), %r17
+       mov             %sp, %r18
+       mov             %r8, %r19
+       
+       st32suspend     %r16, [%r8]
+       
+3:     ld64            [%sp + 64], %r16                // restore call preserved register
+       ld64            [%sp + 128], %r24
+       jmpl            %r2+8, %r0                      // and return
+        add            %sp, 192, %sp
+
Index: linux-2.4.21/drivers/net/qsnet/ep/assym_elan4.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/assym_elan4.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/assym_elan4.h     2005-06-01 23:12:54.627434848 -0400
@@ -0,0 +1,20 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: genassym_elan4.c,v 1.3 2004/04/25 11:26:07 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/genassym_elan4.c,v $*/
+
+/* Generated by genassym_elan4 - do not modify */
+
+#define EP4_RCVR_THREAD_STALL  0
+#define EP4_RCVR_PENDING_TAILP 128
+#define EP4_RCVR_PENDING_HEAD  136
+#define EP4_RCVR_DEBUG         176
+#define EP4_RXD_NEXT           664
+#define EP4_RXD_QUEUED         728
+#define EP4_RXD_DEBUG          944
Index: linux-2.4.21/drivers/net/qsnet/ep/cm.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/cm.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/cm.c      2005-06-01 23:12:54.632434088 -0400
@@ -0,0 +1,3000 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: cm.c,v 1.83.2.6 2005/01/13 12:37:57 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/cm.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "debug.h"
+#include "cm.h"
+#include <elan/epsvc.h>
+
+#include <qsnet/procfs_linux.h>
+
+#if defined(LINUX)
+#include "conf_linux.h"
+#endif
+
+int BranchingRatios[CM_MAX_LEVELS];
+
+int MachineId      = -1;
+int BrokenLevel    = -1;                       /* Simulates Broken Network */
+int RejoinCheck    = 1;
+int RejoinPanic    = 0;
+
+static int
+SegmentNo (CM_RAIL *cmRail, u_int nodeid, u_int lvl)
+{
+    int i;
+
+    ASSERT (lvl < cmRail->NumLevels);
+    
+    for (i = 0; i < lvl; i++)
+       nodeid /= cmRail->Levels[i].NumSegs;
+    
+    return (nodeid % cmRail->Levels[lvl].NumSegs);
+}
+
+static int
+ClusterIds (CM_RAIL *cmRail, int clvl, int *clmin, int *clmax)
+{
+    int clid  = cmRail->Rail->Position.pos_nodeid - cmRail->Levels[clvl].MinNodeId;
+
+    if (clvl == 0)
+       *clmin = *clmax = clid;
+    else
+    {
+       *clmin = cmRail->Levels[clvl - 1].MinNodeId - cmRail->Levels[clvl].MinNodeId;
+       *clmax = *clmin + cmRail->Levels[clvl - 1].NumNodes - 1;
+    }
+    return (clid);
+}
+
+#if defined(PER_CPU_TIMEOUT)
+static void
+__Schedule_Discovery (CM_RAIL *cmRail)         /* we urgently need to schedule discovery */
+{
+    cmRail->NextDiscoverTime = lbolt;
+
+    if (cmRail->NextRunTime == 0 || AFTER (cmRail->NextRunTime, cmRail->NextDiscoverTime))
+       cmRail->NextRunTime = cmRail->NextDiscoverTime;
+}
+
+static void
+__Schedule_Heartbeat (CM_RAIL *cmRail)
+{
+    cmRail->NextHeartbeatTime = lbolt;
+
+    if (cmRail->NextRunTime == 0 || AFTER (cmRail->NextRunTime, cmRail->NextHeartbeatTime))
+       cmRail->NextRunTime = cmRail->NextHeartbeatTime;
+}
+#else
+
+static void
+__Schedule_Timer (CM_RAIL *cmRail, long tick)
+{
+    if (! timer_pending (&cmRail->HeartbeatTimer) || AFTER (cmRail->NextRunTime, tick))
+    {
+       cmRail->NextRunTime = tick;
+
+       mod_timer (&cmRail->HeartbeatTimer, tick);
+    }
+}
+
+static void
+__Schedule_Discovery (CM_RAIL *cmRail)         /* we urgently need to schedule discovery */
+{
+    __Schedule_Timer (cmRail, cmRail->NextDiscoverTime = lbolt);
+}
+
+static void
+__Schedule_Heartbeat (CM_RAIL *cmRail)
+{
+    __Schedule_Timer (cmRail, cmRail->NextHeartbeatTime = lbolt);
+}
+#endif
+
+static int
+MsgBusy (CM_RAIL *cmRail, int msgNumber)
+{
+    switch (ep_outputq_state (cmRail->Rail, cmRail->MsgQueue, msgNumber))
+    {
+    case EP_OUTPUTQ_BUSY:                      /* still busy */
+       return 1;
+       
+    case EP_OUTPUTQ_FAILED:                    /* NACKed */
+    {
+#if defined(DEBUG_PRINTF)
+       CM_MSG  *msg  = ep_outputq_msg (cmRail->Rail, cmRail->MsgQueue, msgNumber);
+       uint8_t  type  = msg->Hdr.Type;
+       uint16_t nmaps = msg->Hdr.NumMaps;
+       int16_t  off   = msg->Payload.Statemaps[CM_MSG_MAP(0)].offset;
+       
+       CPRINTF4 (((type == CM_MSG_TYPE_DISCOVER_LEADER) || (type == CM_MSG_TYPE_DISCOVER_SUBORDINATE))  ? 6 : 3, /* we expect broadcasts to be NACKed */
+                 "%s: msg %d type %d  failed%s\n", cmRail->Rail->Name, msgNumber, type, 
+                 (type != CM_MSG_TYPE_HEARTBEAT) ? "" : nmaps == 0 ? ": null heartbeat" :
+                 off == STATEMAP_RESET ? ": heartbeat with R statemaps" : ": heartbeat with statemaps");
+#endif
+       return 0;
+    }
+    
+    case EP_OUTPUTQ_FINISHED:
+       return 0;
+
+    default:
+       panic ("MsgBusy - bad return code from ep_outputq_state\n");
+       /* NOTREACHED */
+    }
+    return 0;
+}
+
+static void
+LaunchMessage (CM_RAIL *cmRail, int msgNumber, int vp, int qnum, int retries, int type, int lvl, int nmaps)
+{
+    CM_MSG *msg = ep_outputq_msg (cmRail->Rail, cmRail->MsgQueue, msgNumber);
+    CM_HDR *hdr = &msg->Hdr;
+
+    ASSERT (nmaps >= 0 && nmaps <= CM_MSG_MAXMAPS);
+    ASSERT (SPINLOCK_HELD (&cmRail->Lock));
+
+    hdr->Version   = CM_MSG_VERSION;
+    hdr->ParamHash = cmRail->ParamHash;
+    hdr->Timestamp = cmRail->Timestamp;
+    hdr->Checksum  = 0;
+    hdr->NodeId    = cmRail->Rail->Position.pos_nodeid;
+    hdr->MachineId = MachineId;
+    hdr->NumMaps   = nmaps;
+    hdr->Level     = lvl;
+    hdr->Type      = type;
+    hdr->Checksum  = CheckSum ((char *)msg + CM_MSG_BASE(nmaps), CM_MSG_SIZE(nmaps));
+
+    if (BrokenLevel != -1 && (lvl >= ((BrokenLevel >> (cmRail->Rail->Number*4)) & 0xf)))                       /* Simulate broken network? */
+       return;
+
+    if (ep_outputq_send (cmRail->Rail, cmRail->MsgQueue, msgNumber, 
+                        CM_MSG_SIZE(nmaps), vp, qnum, retries));
+       IncrStat (cmRail, LaunchMessageFail);
+}
+
+static int
+SendMessage (CM_RAIL *cmRail, int nodeId, int lvl, int type)
+{
+    int        msgNumber = CM_NUM_NODE_MSG_BUFFERS + cmRail->NextSpareMsg;
+    int n         = CM_NUM_SPARE_MSG_BUFFERS;
+    int retries;
+
+    ASSERT (type == CM_MSG_TYPE_IMCOMING ||    /* other types must use SendToSgmt */
+           type == CM_MSG_TYPE_REJOIN);
+   
+    while (n-- > 0 && MsgBusy (cmRail, msgNumber))     /* search for idle "spare" buffer */
+    {
+       if (++(cmRail->NextSpareMsg) == CM_NUM_SPARE_MSG_BUFFERS)
+           cmRail->NextSpareMsg = 0;
+      
+       msgNumber = CM_NUM_NODE_MSG_BUFFERS + cmRail->NextSpareMsg;
+    }
+
+    if (n == 0)                                        /* all "spare" message buffers busy */
+    {
+       CPRINTF3 (3, "%s: all spare message buffers busy: trying to send type %d to %d\n",
+                 cmRail->Rail->Name, type, nodeId);
+       return (0);
+    }
+
+    /* NB IMCOMING may be echoed by MANY nodes, so we don't (and musn't) have any retries */
+    retries = (type == CM_MSG_TYPE_IMCOMING) ? 0 : CM_P2P_DMA_RETRIES;
+   
+    LaunchMessage (cmRail, msgNumber, EP_VP_NODE (nodeId), EP_SYSTEMQ_INTR, /* eager receive */
+                  retries, type, lvl, 0);
+   
+    if (++(cmRail->NextSpareMsg) == CM_NUM_SPARE_MSG_BUFFERS) /* check this one last next time */
+       cmRail->NextSpareMsg = 0;
+
+    return (1);
+}
+
+static int
+SendToSgmt (CM_RAIL *cmRail, CM_SGMT *sgmt, int type)
+{    
+    bitmap_t         seg;
+    int              offset;
+    int              nmaps;
+    int              sidx;
+    int              clvl;
+    
+    ASSERT (sgmt->Level <= cmRail->TopLevel);
+
+    if (MsgBusy (cmRail, sgmt->MsgNumber))             /* previous message still busy */
+    {
+       CPRINTF3 (3, "%s: node message buffer busy: trying to send type %d to %d\n",
+                 cmRail->Rail->Name, type, sgmt->NodeId);
+      
+       return (0);
+    }
+
+    switch (type)
+    {
+    case CM_MSG_TYPE_RESOLVE_LEADER:
+    case CM_MSG_TYPE_DISCOVER_LEADER:
+       ASSERT (sgmt->State == CM_SGMT_ABSENT);
+       ASSERT (sgmt->Level == ((cmRail->Role == CM_ROLE_LEADER_CANDIDATE) ? cmRail->TopLevel : cmRail->TopLevel - 1));
+       ASSERT (sgmt->Level < cmRail->NumLevels);
+       ASSERT (sgmt->Sgmt == cmRail->Levels[sgmt->Level].MySgmt);
+      
+       /* broadcast to me and all my peers at this level (== my segment in the level above) */
+       sidx = (sgmt->Level == cmRail->NumLevels - 1) ? 0 : cmRail->Levels[sgmt->Level + 1].MySgmt;
+
+       LaunchMessage (cmRail, sgmt->MsgNumber, EP_VP_BCAST (sgmt->Level + 1, sidx), 
+                      EP_SYSTEMQ_INTR, 0,              /* eager rx; no retries */
+                      type, sgmt->Level, 0);
+       return (1);
+      
+    case CM_MSG_TYPE_DISCOVER_SUBORDINATE:
+       ASSERT (sgmt->Sgmt != cmRail->Levels[sgmt->Level].MySgmt);
+       ASSERT (sgmt->State == CM_SGMT_WAITING);
+       ASSERT (sgmt->Level > 0);                       /* broadcasting just to subtree */
+      
+       LaunchMessage (cmRail, sgmt->MsgNumber, EP_VP_BCAST (sgmt->Level, sgmt->Sgmt), 
+                      EP_SYSTEMQ_INTR, 0,              /* eager rx; no retries */
+                      CM_MSG_TYPE_DISCOVER_SUBORDINATE, sgmt->Level, 0);
+       return (1);
+      
+    case CM_MSG_TYPE_NOTIFY:
+       ASSERT (sgmt->State == CM_SGMT_PRESENT);
+      
+       LaunchMessage (cmRail, sgmt->MsgNumber, EP_VP_NODE (sgmt->NodeId),
+                      EP_SYSTEMQ_INTR, CM_P2P_DMA_RETRIES, /* eager rx; lots of retries */
+                      CM_MSG_TYPE_NOTIFY, sgmt->Level, 0);
+       return (1);
+      
+    case CM_MSG_TYPE_HEARTBEAT:
+    {
+       CM_MSG *msg = ep_outputq_msg (cmRail->Rail, cmRail->MsgQueue, sgmt->MsgNumber);
+       CM_HDR *hdr = &msg->Hdr;
+
+       ASSERT (sgmt->State == CM_SGMT_PRESENT);
+       
+       hdr->AckSeq = sgmt->AckSeq;
+   
+       if (!sgmt->MsgAcked)                    /* Current message not acknowledged */
+       {
+           /* must have been something significant to require an ack */
+           ASSERT (sgmt->SendMaps);
+           ASSERT (sgmt->NumMaps > 0);
+           
+           CPRINTF3 (3, "%s: retrying heartbeat to %d (%d entries)\n", cmRail->Rail->Name, sgmt->NodeId, sgmt->NumMaps);
+
+           IncrStat (cmRail, RetryHeartbeat);
+
+           nmaps = sgmt->NumMaps;
+       }
+       else
+       {
+           nmaps = 0;
+      
+           if (sgmt->SendMaps)                 /* can send maps */
+           {
+               for (clvl = sgmt->Level; clvl < cmRail->NumLevels; clvl++)
+               {
+                   if (!sgmt->Maps[clvl].OutputMapValid)
+                       continue;
+                   
+                   while ((offset = statemap_findchange (sgmt->Maps[clvl].OutputMap, &seg, 1)) >= 0)
+                   {
+                       CM_STATEMAP_ENTRY *map = &msg->Payload.Statemaps[CM_MSG_MAP(nmaps)];
+
+                       sgmt->Maps[clvl].SentChanges = 1;
+                       
+                       map->level  = clvl;
+                       map->offset = offset;
+                       map->seg[0] = seg & 0xffff;
+                       map->seg[1] = (seg >> 16) & 0xffff;
+#if (BT_ULSHIFT == 6)
+                       map->seg[2] = (seg >> 32) & 0xffff;
+                       map->seg[3] = (seg >> 48) & 0xffff;
+#elif (BT_ULSHIFT != 5)
+#error "Bad value for BT_ULSHIFT"
+#endif
+                       if (++nmaps == CM_MSG_MAXMAPS)
+                           goto msg_full;
+                   }
+
+                   if (sgmt->Maps[clvl].SentChanges)
+                   {
+                       CM_STATEMAP_ENTRY *map = &msg->Payload.Statemaps[CM_MSG_MAP(nmaps)];
+
+                       sgmt->Maps[clvl].SentChanges = 0;
+
+                       map->level  = clvl;
+                       map->offset = STATEMAP_NOMORECHANGES;
+                       
+                       if (++nmaps == CM_MSG_MAXMAPS)
+                           goto msg_full;
+                   }
+               }
+           }
+           
+           ASSERT (nmaps < CM_MSG_MAXMAPS);
+
+       msg_full:
+           sgmt->NumMaps = nmaps;              /* remember how many incase we retry */
+
+           if (nmaps == 0)                     /* no changes to send */
+               hdr->Seq = sgmt->MsgSeq;        /* this one can be dropped */
+           else
+           {
+               hdr->Seq = ++(sgmt->MsgSeq);    /* on to next message number */
+               sgmt->MsgAcked = 0;             /* need this one to be acked before I can send another */
+
+               IncrStat (cmRail, MapChangesSent);
+           }
+       }
+
+       LaunchMessage (cmRail, sgmt->MsgNumber, EP_VP_NODE (sgmt->NodeId), 
+                      EP_SYSTEMQ_POLLED,  CM_P2P_DMA_RETRIES, /* polled receive, lots of retries */
+                      CM_MSG_TYPE_HEARTBEAT, sgmt->Level, nmaps);
+
+       IncrStat (cmRail, HeartbeatsSent);
+
+       return (1);
+    }
+
+    default:                                   /* other types must use SendMessage */
+       printk ("SendToSgmt: invalid type %d\n", type);
+       ASSERT (0);
+
+       return (1);
+    }
+}
+
+static char *
+GlobalStatusString (statemap_t *map, int idx)
+{
+    char *strings[] = {"....", "S...", "C...", "R...", 
+                      ".s..", "Ss..", "Cs..", "Rs..", 
+                      "..r.", "S.r.", "C.r.", "R.r.", 
+                      ".sr.", "Ssr.", "Csr.", "Rsr.", 
+                      "...R", "S..R", "C..R", "R..R", 
+                      ".s.R", "Ss.R", "Cs.R", "Rs.R", 
+                      "..rR", "S.rR", "C.rR", "R.rR", 
+                      ".srR", "SsrR", "CsrR", "RsrR"};
+    
+    return (strings[statemap_getbits (map, idx * CM_GSTATUS_BITS, CM_GSTATUS_BITS)]);
+}
+
+static char *
+MapString (char *name, statemap_t *map, int nnodes, char *trailer)
+{
+    static char *space;
+    int          i;
+
+    if (space == NULL)
+       KMEM_ALLOC (space, char *, EP_MAX_NODES*(CM_GSTATUS_BITS+1), 0);
+
+    if (space == NULL)
+       return ("<cannot allocate memory>");
+    else
+    {
+       char *ptr = space;
+
+       sprintf (space, "%s ", name); ptr += strlen (ptr);
+       for (i = 0; i < nnodes; i++, ptr += strlen (ptr))
+           sprintf (ptr, "%s%s", i == 0 ? "" : ",", GlobalStatusString (map, i));
+       sprintf (ptr, " %s", trailer);
+       return (space);
+    }
+}
+
+void
+DisplayMap (DisplayInfo *di, CM_RAIL *cmRail, char *name, statemap_t *map, int nnodes, char *trailer)
+{
+    char  linebuf[256];
+    char *ptr = linebuf;
+    int   i;
+
+#define NODES_PER_LINE 32
+    for (i = 0; i < nnodes; i++)
+    {
+       if (ptr == linebuf)
+       {
+           sprintf (ptr, "%4d", i);
+           ptr += strlen (ptr);
+       }
+       
+       sprintf (ptr, ",%s", GlobalStatusString (map, i));
+       ptr += strlen (ptr);
+
+       if ((i % NODES_PER_LINE) == (NODES_PER_LINE-1) || (i == (nnodes-1)))
+       {
+           (di->func)(di->arg, "%s: %s %s %s\n", cmRail->Rail->Name, name, linebuf, trailer);
+           ptr = linebuf;
+       }
+    }
+#undef NODES_PER_LINE
+}
+
+void
+DisplayNodeMaps (DisplayInfo *di, CM_RAIL *cmRail)
+{
+    int   lvl;
+    int   clvl;
+    char  mapname[128];
+    
+    (di->func)(di->arg, "%s: Node %d maps...\n", cmRail->Rail->Name, cmRail->Rail->Position.pos_nodeid);
+
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+    {
+       int nnodes = cmRail->Levels[clvl].NumNodes;
+
+       (di->func)(di->arg, "%s: Cluster level %d: Connected %ld - %s%s\n", 
+                  cmRail->Rail->Name, clvl, cmRail->Levels[clvl].Connected,
+                  cmRail->Levels[clvl].Online     ? "Online" : "Offline",
+                  cmRail->Levels[clvl].Restarting ? ", Restarting" : "");
+
+       for (lvl = 0; lvl < cmRail->TopLevel && lvl <= clvl; lvl++)
+       {
+           CM_LEVEL *level = &cmRail->Levels[lvl];
+
+           sprintf (mapname, "%10s%2d", "Level", lvl);
+           DisplayMap (di, cmRail, mapname, level->SubordinateMap[clvl], nnodes,
+                       level->SubordinateMapValid[clvl] ? "" : "(invalid)");
+       }
+
+       sprintf (mapname, "%12s", "Local");
+       DisplayMap (di, cmRail, mapname, cmRail->Levels[clvl].LocalMap, nnodes, "");
+
+       sprintf (mapname, "%12s", "Subtree");
+       DisplayMap (di, cmRail, mapname, cmRail->Levels[clvl].SubTreeMap, nnodes, 
+                   cmRail->Levels[clvl].SubTreeMapValid ? "" : "(invalid)");
+
+       sprintf (mapname, "%12s", "Global");
+       DisplayMap (di, cmRail, mapname, cmRail->Levels[clvl].GlobalMap, nnodes, 
+                   cmRail->Levels[clvl].GlobalMapValid ? "" : "(invalid)");
+
+       sprintf (mapname, "%12s", "LastGlobal");
+       DisplayMap (di, cmRail, mapname, cmRail->Levels[clvl].LastGlobalMap, nnodes, "");
+    }
+}
+
+void
+DisplayNodeSgmts (DisplayInfo *di, CM_RAIL *cmRail)
+{
+    int   lvl;
+    int   sidx;
+    
+    (di->func)(di->arg, "%s: Node %d segments...\n", cmRail->Rail->Name, cmRail->NodeId);
+    
+    for (lvl = 0; lvl <= cmRail->TopLevel && lvl < cmRail->NumLevels; lvl++)
+    {
+       (di->func)(di->arg, "   level %d: ", lvl);
+       
+       for (sidx = 0; sidx < ((lvl == cmRail->TopLevel) ? 1 : cmRail->Levels[lvl].NumSegs); sidx++)
+       {
+           CM_SGMT *sgmt = &cmRail->Levels[lvl].Sgmts[sidx];
+               
+           if (sgmt->State == CM_SGMT_PRESENT)
+               (di->func)(di->arg, "[%d, in: %d out: %d %s%s]", 
+                       sgmt->NodeId,
+                       sgmt->AckSeq,
+                       sgmt->MsgSeq,
+                       sgmt->MsgAcked ? "A" : "-",
+                       sgmt->SendMaps ? "!" : "-");
+           else
+               (di->func)(di->arg, "[%s]", (sgmt->State == CM_SGMT_ABSENT ? "absent" :
+                                sgmt->State == CM_SGMT_WAITING ? "waiting" :
+                                sgmt->State == CM_SGMT_COMING ? "coming" : "UNKNOWN"));
+       }
+       (di->func)(di->arg, "\n");
+    }
+}
+
+
+static void
+StartConnecting (CM_RAIL *cmRail, CM_SGMT *sgmt, int NodeId, int Timestamp)
+{
+    int clvl;
+
+    CPRINTF4 (2, "%s: lvl %d subtree %d node %d -> connecting\n", cmRail->Rail->Name, sgmt->Level, sgmt->Sgmt, NodeId);
+
+    /* Only reconnect the same guy if he was reborn */
+    ASSERT (sgmt->State != CM_SGMT_PRESENT ||
+           (sgmt->NodeId == NodeId && sgmt->Timestamp != Timestamp));
+   
+    /* After we've connected to a new peer, we wait to receive
+     * STATEMAP_RESET before we accumulate changes and we wait for a
+     * complete map to be received before we propagate changes to other
+     * nodes.
+     *
+     * If I'm the subordinate, I can start sending maps right away, since
+     * the leader is ready for them already.  If I'm the leader, I hold off
+     * sending maps until I've seen the subordinate's first heartbeat,
+     * because the subordinate might miss my NOTIFY message, still think
+     * she's a leader candidate and ignore my heartbeats.
+     */
+    sgmt->SendMaps = (sgmt->Level == cmRail->TopLevel); /* I can send maps to my leader (she NOTIFIED me) */
+
+    for (clvl = sgmt->Level; clvl < cmRail->NumLevels; clvl++)
+    {
+       statemap_reset (sgmt->Maps[clvl].CurrentInputMap);
+       statemap_reset (sgmt->Maps[clvl].InputMap);
+       statemap_reset (sgmt->Maps[clvl].OutputMap);
+       
+       sgmt->Maps[clvl].InputMapValid = 0;
+       sgmt->Maps[clvl].OutputMapValid = 0;
+       sgmt->Maps[clvl].SentChanges = 0;
+
+       if (sgmt->Level == cmRail->TopLevel)    /* connection to leader */
+       {
+           ASSERT (sgmt->Sgmt == 0);
+           ASSERT (cmRail->Role == CM_ROLE_SUBORDINATE);
+
+           if (cmRail->Levels[clvl].SubTreeMapValid) /* already got a subtree map to send up */
+           {
+               statemap_setmap (sgmt->Maps[clvl].OutputMap, cmRail->Levels[clvl].SubTreeMap);
+               sgmt->Maps[clvl].OutputMapValid = 1;
+
+               statemap_clearchanges (cmRail->Levels[clvl].SubTreeMap);
+           }
+       }
+       else                                    /* connection to subordinate */
+       {
+           ASSERT (sgmt->Sgmt != cmRail->Levels[sgmt->Level].MySgmt);
+
+           if (cmRail->Levels[clvl].GlobalMapValid) /* already got a global map to broadcast */
+           {
+               statemap_setmap (sgmt->Maps[clvl].OutputMap, cmRail->Levels[clvl].GlobalMap);
+               sgmt->Maps[clvl].OutputMapValid = 1;
+           }
+       }
+    }
+    
+    /* Initialise sequence counters */
+    sgmt->MsgSeq = sgmt->AckSeq = 0;
+    sgmt->MsgAcked = 1;                        /* ready to send a new sequenced message */
+   
+    sgmt->State      = CM_SGMT_PRESENT;
+    sgmt->NodeId     = NodeId;
+    sgmt->UpdateTick = lbolt;
+    sgmt->Timestamp  = Timestamp;
+}
+
+static void
+StartSubTreeDiscovery (CM_RAIL *cmRail, CM_SGMT *sgmt)
+{
+    sgmt->State = CM_SGMT_WAITING;
+    sgmt->UpdateTick = lbolt;
+    sgmt->WaitingTick = lbolt;
+
+    if (sgmt->Level > 0)
+       __Schedule_Discovery (cmRail);
+}
+
+void
+StartSubordinateDiscovery (CM_RAIL *cmRail)
+{
+    int       i;
+    int       lvl = cmRail->TopLevel - 1;
+    CM_LEVEL *level = &cmRail->Levels[lvl];
+
+    ASSERT (lvl >= 0 && lvl < cmRail->NumLevels);
+
+    for (i = 0; i < level->NumSegs; i++)
+    {
+        CM_SGMT *sgmt = &level->Sgmts[i];
+       
+       if (i != level->MySgmt)         /* No-one should connect here */
+           StartSubTreeDiscovery (cmRail, sgmt);
+    }
+}
+
+void
+StartLeaderDiscovery (CM_RAIL *cmRail)
+{
+    int       i;
+    int       clvl;
+    CM_LEVEL *level = &cmRail->Levels[cmRail->TopLevel];
+
+    ASSERT (cmRail->TopLevel < cmRail->NumLevels);
+
+    for (clvl = cmRail->TopLevel; clvl < cmRail->NumLevels; clvl++)
+    {
+        cmRail->Levels[clvl].GlobalMapValid = 0;
+       cmRail->Levels[clvl].SubTreeMapValid = 0;
+        level->SubordinateMapValid[clvl] = 0;
+    }
+
+    for (i = 0; i < level->NumSegs; i++)
+    {
+        CM_SGMT *sgmt = &level->Sgmts[i];
+       
+       sgmt->State = CM_SGMT_ABSENT;
+    }
+
+    cmRail->DiscoverStartTick = lbolt;
+    cmRail->Role = CM_ROLE_LEADER_CANDIDATE;
+   
+    __Schedule_Discovery (cmRail);
+}
+
+static void
+RaiseTopLevel (CM_RAIL *cmRail)
+{
+    ASSERT (cmRail->NumLevels != 0);
+    ASSERT (cmRail->TopLevel < cmRail->NumLevels);
+
+    CPRINTF2 (2, "%s: RaiseTopLevel %d\n", cmRail->Rail->Name, cmRail->TopLevel + 1);
+
+    if (++cmRail->TopLevel == cmRail->NumLevels)       /* whole machine leader? */
+       cmRail->Role = CM_ROLE_LEADER;
+    else
+       StartLeaderDiscovery (cmRail);          /* look for my leader */
+
+    StartSubordinateDiscovery (cmRail);                /* and any direct subordinates */
+}
+
+static void
+LowerTopLevel (CM_RAIL *cmRail, int lvl)
+{
+    ASSERT (cmRail->NumLevels != 0);
+    ASSERT (lvl < cmRail->NumLevels);
+
+    CPRINTF2 (2, "%s: LowerTopLevel %d\n", cmRail->Rail->Name, lvl);
+
+    if (lvl == 0)
+       cmRail->Timestamp = lbolt;
+
+    cmRail->TopLevel = lvl;
+
+    StartLeaderDiscovery (cmRail);             /* look for my leader */
+}
+
+static int
+IShouldLead (CM_RAIL *cmRail, CM_MSG *msg)
+{
+    /* NB, this function MUST be consistently calculated on any nodes, just
+     * from the info supplied in the message.  Otherwise leadership
+     * arbitration during concurrent discovery will fail.
+     */
+    return (cmRail->NodeId < msg->Hdr.NodeId);
+}
+
+static int
+SumCheck (CM_MSG *msg)
+{
+    CM_HDR   *hdr   = &msg->Hdr;
+    uint16_t  sum   = hdr->Checksum;
+    uint16_t  nmaps = hdr->NumMaps;
+
+    if (nmaps > CM_MSG_MAXMAPS) {
+       printk ("SumCheck: nmaps %d > CM_MSG_MAXMAPS\n", nmaps);
+       return 0;
+    }
+    
+    if ((hdr->Type != CM_MSG_TYPE_HEARTBEAT) && nmaps != 0) {
+       printk ("SumCheck: type(%d) not HEARTBEAT and nmaps(%d) != 0\n", hdr->Type, nmaps);
+       return 0;
+    }
+
+    hdr->Checksum = 0;
+    
+    if (CheckSum ((char *)msg + CM_MSG_BASE(nmaps), CM_MSG_SIZE(nmaps)) != sum) {
+       printk ("SumCheck: checksum failed %x %x\n", CheckSum ((char *)msg + CM_MSG_BASE(nmaps), CM_MSG_SIZE(nmaps)), sum);
+
+       return 0;
+    }
+       
+    return 1;
+}
+
+static void
+ProcessMessage (EP_RAIL *rail, void *arg, void *msgbuf)
+{
+    CM_RAIL       *cmRail = (CM_RAIL *) arg;
+    CM_MSG         *msg    = (CM_MSG *) msgbuf;
+    CM_HDR         *hdr    = &msg->Hdr;
+    int             lvl;
+    int             sidx;
+    CM_LEVEL       *level;
+    CM_SGMT        *sgmt;
+    bitmap_t        seg;
+    int             i;
+    int            delay;
+    static long    tlast;
+    static int     count;
+
+    /* Poll the message Version field until the message has completely
+     * arrived in main memory. */
+    for (delay = 1; hdr->Version == EP_SYSTEMQ_UNRECEIVED && delay < EP_SYSTEMQ_UNRECEIVED_TLIMIT; delay <<= 1)
+       DELAY (delay);
+
+    /* Display a message every 60 seconds if we see an "old" format message */
+    if (hdr->Version == EP_SYSTEMQ_UNRECEIVED && (((lbolt - tlast) > 60*HZ) ? (count = 0) : ++count) < 1)
+    {
+       printk ("%s: received old protocol message (type %d from node %d)\n", cmRail->Rail->Name, 
+               ((uint8_t *) msg)[20], ((uint16_t *) msg)[4]);
+
+       tlast = lbolt;
+       goto finished;
+    }
+
+    if (hdr->Version != CM_MSG_VERSION || hdr->ParamHash != cmRail->ParamHash || hdr->MachineId != MachineId)
+    {
+       CPRINTF8 (1, "%s: invalid message : Version %08x (%08x) ParamHash %08x (%08x) MachineId %04x (%04x) Nodeid %d\n", cmRail->Rail->Name,
+                 hdr->Version, CM_MSG_VERSION, hdr->ParamHash, cmRail->ParamHash, hdr->MachineId, MachineId, hdr->NodeId);
+       goto finished;
+    }
+
+    if (!SumCheck (msg))
+    {
+       printk ("%s: checksum failed on msg from %d?\n", cmRail->Rail->Name, hdr->NodeId);
+       goto finished;
+    }
+    
+    if (hdr->NodeId == cmRail->NodeId)         /* ignore my own broadcast */       
+    {
+       CPRINTF3 (6, "%s: node %d type %d: ignored (MESSAGE FROM ME)\n", 
+                 cmRail->Rail->Name, hdr->NodeId, hdr->Type);
+
+       if (hdr->Type != CM_MSG_TYPE_DISCOVER_LEADER && hdr->Type != CM_MSG_TYPE_RESOLVE_LEADER)
+           printk ("%s: node %d type %d: ignored (MESSAGE FROM ME)\n", 
+                   cmRail->Rail->Name, hdr->NodeId, hdr->Type);
+       goto finished;
+    }
+
+    lvl = hdr->Level;
+    level = &cmRail->Levels[lvl];
+
+    if (BrokenLevel != -1 && (lvl >= ((BrokenLevel >> (cmRail->Rail->Number*4)) & 0xf)))                       /* Simulate broken network? */
+       goto finished;
+    
+    if (lvl >= cmRail->NumLevels ||            /* from outer space  */
+       hdr->NodeId < level->MinNodeId ||       /* from outside this level's subtree */
+       hdr->NodeId >= level->MinNodeId + level->NumNodes)
+    {
+       printk ("%s: lvl %d node %d type %d: ignored (%s)\n", 
+               cmRail->Rail->Name, lvl, hdr->NodeId, hdr->Type, 
+               lvl >= cmRail->NumLevels ? "level too big for machine" : "outside subtree");
+       goto finished;
+    }
+
+    sidx = SegmentNo (cmRail, hdr->NodeId, lvl);
+    sgmt = &level->Sgmts[sidx];
+    
+    switch (hdr->Type)
+    {
+    case CM_MSG_TYPE_RESOLVE_LEADER:
+       if (lvl >= cmRail->TopLevel)
+       {
+           CPRINTF4 (6, "%s: lvl %d sidx %d node %d RESOLVE_LEADER: ignored (above my level)\n", 
+                     cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+           break;
+       }
+
+       /* someone else thinks they lead at the same level as me */
+       CPRINTF4 (1, "%s: lvl %d sidx %d node %d RESOLVE_LEADER: !REJOIN (putsch)\n", 
+                 cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+       
+       printk ("%s: lvl %d sidx %d node %d RESOLVE_LEADER: !REJOIN (putsch)\n", 
+               cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+       
+
+       SendMessage (cmRail, hdr->NodeId, lvl, CM_MSG_TYPE_REJOIN);
+       break;
+       
+    case CM_MSG_TYPE_DISCOVER_LEADER:
+       if (lvl > cmRail->TopLevel)
+       {
+           CPRINTF4 (6, "%s: lvl %d sidx %d node %d DISCOVER_LEADER: ignored (above my level)\n", 
+                     cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+           break;
+       }
+
+       if (sidx == level->MySgmt)              /* someone I led thinks they lead some of my subtrees */
+       {
+           CPRINTF4 (1, "%s: lvl %d sidx %d node %d DISCOVER_LEADER: !REJOIN (putsch)\n", 
+                     cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+
+           printk ("%s: lvl %d sidx %d node %d DISCOVER_LEADER: !REJOIN (putsch)\n", 
+                   cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+
+           SendMessage (cmRail, hdr->NodeId, hdr->Level, CM_MSG_TYPE_REJOIN);
+           break;
+       }       
+
+       if (lvl < cmRail->TopLevel)                     /* I'm the leader of this level */
+       {
+           if (sgmt->State == CM_SGMT_PRESENT &&       /* someone thinks someone I lead is dead */
+               sgmt->NodeId != hdr->NodeId)
+           {
+               /* My subordinate's death could be noticed by one of her peers
+                * before I do.  If she _is_ dead, I'll notice before long and
+                * NOTIFY this discover.  If this discover completes before I
+                * detect my subordinate's death, the discovering node will
+                * try to take over from me, and then I'll RESET her.
+                */
+               CPRINTF4 (6, "%s: lvl %d sidx %d node %d DISCOVER_LEADER: ignored (got established subordinate)\n", 
+                         cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+               return;
+           }
+
+           if (sgmt->State != CM_SGMT_PRESENT || /* New connection */
+               sgmt->Timestamp != hdr->Timestamp) /* new incarnation */
+               StartConnecting (cmRail, sgmt, hdr->NodeId, hdr->Timestamp);
+         
+           CPRINTF4 (2, "%s: lvl %d sidx %d node %d DISCOVER_LEADER: !NOTIFY)\n", 
+                     cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+         
+           SendToSgmt (cmRail, sgmt, CM_MSG_TYPE_NOTIFY);
+           break;
+       }
+
+       ASSERT (lvl == cmRail->TopLevel);
+
+       if (cmRail->Role == CM_ROLE_SUBORDINATE)
+       {
+           /* I think my leader is alive, in which case she'll NOTIFY this
+            * DISCOVER.  If she's dead, I'll start to become a leader
+            * candidate and handle this appropriately.
+            */
+           CPRINTF3 (6, "%s: lvl %d node %d DISCOVER: ignored (I'm a subordinate)\n", 
+                     cmRail->Rail->Name, lvl, hdr->NodeId);
+           break;
+       }
+       
+       ASSERT (cmRail->Role == CM_ROLE_LEADER_CANDIDATE);
+
+       /* A peer at this level is bidding for leadership along with me */
+       if (IShouldLead (cmRail, msg))
+       {
+           CPRINTF3 (6, "%s: lvl %d node %d DISCOVER: but I should lead\n", 
+                     cmRail->Rail->Name, lvl, hdr->NodeId);
+
+           /* So there _is_ someone there; She'll be seeing my DISCOVER
+            * messages and extending her discovery period, so that when I
+            * become leader, I'll NOTIFY her.  In the meantime I'll flag her
+            * activity, so she remains WAITING.
+            */
+           sgmt->UpdateTick = lbolt;
+           break;
+       }
+       
+       /* Defer to sender... */
+       CPRINTF3 (6, "%s: lvl %d node %d DISCOVER: delaying me becoming leader\n", 
+                 cmRail->Rail->Name, lvl, hdr->NodeId);
+       
+       StartLeaderDiscovery (cmRail);
+       break;
+
+    case CM_MSG_TYPE_DISCOVER_SUBORDINATE:
+       if (lvl <= cmRail->TopLevel)
+       {
+           CPRINTF3 (6, "%s: lvl %d node %d DISCOVER_SUBORDINATE: ignored (from my subtree)\n", 
+                     cmRail->Rail->Name, lvl, hdr->NodeId);
+           break;
+       }
+       
+       if (cmRail->Role != CM_ROLE_LEADER_CANDIDATE)
+       {
+           CPRINTF3 (6, "%s: lvl %d node %d DISCOVER_SUBORDINATE: ignored (I'm not looking for a leader)\n", 
+                     cmRail->Rail->Name, lvl, hdr->NodeId);
+           break;
+       }
+       
+       if (hdr->Level > cmRail->BroadcastLevel && AFTER (lbolt, cmRail->BroadcastLevelTick + EP_WITHDRAW_TIMEOUT))
+       {
+           CPRINTF3 (6, "%s: lvl %d node %d DISCOVER_SUBORDINATE: ignored (broadcast level too low)\n",
+                     cmRail->Rail->Name, lvl, hdr->NodeId);
+           break;
+       }
+
+       CPRINTF3 (2, "%s: lvl %d node %d DISCOVER_SUBORDINATE: !IMCOMING\n", 
+                 cmRail->Rail->Name, lvl, hdr->NodeId);
+       
+       SendMessage (cmRail, hdr->NodeId, hdr->Level, CM_MSG_TYPE_IMCOMING);
+       break;
+
+    case CM_MSG_TYPE_IMCOMING:
+       if (lvl > cmRail->TopLevel ||           /* from peer or node above me */
+           sgmt->State == CM_SGMT_PRESENT ||   /* already got a subtree */
+           sgmt->State == CM_SGMT_ABSENT)      /* already written off this subtree */
+       {
+           CPRINTF4 (2, "%s: lvl %d sidx %d node %d IMCOMING: ignored\n", cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+           break;
+       }
+
+       CPRINTF4 (2, "%s: lvl %d sidx %d node %d IMCOMING: waiting...\n", cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+
+       sgmt->State = CM_SGMT_COMING;
+       sgmt->UpdateTick = lbolt;
+       break;
+       
+    case CM_MSG_TYPE_NOTIFY:
+       if (cmRail->Role != CM_ROLE_LEADER_CANDIDATE || /* I'm not looking for a leader */
+           lvl != cmRail->TopLevel)            /* at this level */
+       {
+           /* If this person really should be my leader, my existing leader
+            * will time out, and I'll discover this one. */
+           CPRINTF4 (2, "%s: lvl %d node %d NOTIFY: ignored (%s)\n", 
+                     cmRail->Rail->Name, lvl, hdr->NodeId,
+                     lvl < cmRail->TopLevel ? "already leader" : 
+                     lvl > cmRail->TopLevel ? "lvl too high" : "already subordinate");
+           break;
+       }
+
+       CPRINTF3 (2, "%s: lvl %d node %d NOTIFY: becoming subordinate\n", 
+                 cmRail->Rail->Name, lvl, hdr->NodeId);
+       
+       cmRail->Role = CM_ROLE_SUBORDINATE;             /* Now I've found my level */
+       StartConnecting (cmRail, &level->Sgmts[0], hdr->NodeId, hdr->Timestamp);
+       break;
+
+    case CM_MSG_TYPE_HEARTBEAT:
+       if (lvl > cmRail->TopLevel)
+       {
+           CPRINTF3 (2, "%s: lvl %d node %d H/BEAT: ignored (lvl too high)\n", 
+                     cmRail->Rail->Name, lvl, hdr->NodeId);
+           break;
+       }
+
+       if (lvl == cmRail->TopLevel)                    /* heartbeat from my leader */
+       {
+           if (cmRail->Role == CM_ROLE_LEADER_CANDIDATE) /* but I've not got one */
+           {
+               /* I'm probably a new incarnation of myself; I'll keep doing
+                * discovery until my previous existence's leader NOTIFY's me.
+                * If I was this node's leader, she'll time me out (I'm not
+                * sending heartbeats to her) and we'll fight it out for
+                * leadership. */
+               CPRINTF3 (2, "%s: lvl %d node %d H/BEAT ignored (no leader)\n", 
+                         cmRail->Rail->Name, lvl, hdr->NodeId);
+               break;
+           }
+           sidx = 0;
+           sgmt = &level->Sgmts[0];
+       }
+      
+       if (sgmt->State != CM_SGMT_PRESENT ||   /* not fully connected with this guy */
+           sgmt->NodeId != hdr->NodeId ||      /* someone else impersonating my peer */
+           sgmt->Timestamp != hdr->Timestamp)  /* new incarnation of my peer */
+       {
+           CPRINTF4 (1, "%s: lvl %d sidx %d node %d H/BEAT: !REJOIN\n", 
+                     cmRail->Rail->Name, lvl, sidx, hdr->NodeId);
+         
+           printk ("%s: lvl %d sidx %d node %d H/BEAT: !REJOIN %s\n",
+                   cmRail->Rail->Name, lvl, sidx, hdr->NodeId,
+                   sgmt->State != CM_SGMT_PRESENT ? "not present" :
+                   sgmt->NodeId != hdr->NodeId ? "someone else" : "new incarnation");
+           
+           SendMessage (cmRail, hdr->NodeId, hdr->Level, CM_MSG_TYPE_REJOIN);
+           break;
+       }
+
+       if (!((hdr->Seq == sgmt->AckSeq) ||     /* NOT duplicate message or */
+             (hdr->Seq == (CM_SEQ)(sgmt->AckSeq + 1))) || /* expected message */
+           !((hdr->AckSeq == sgmt->MsgSeq) ||  /* NOT expected ack or */
+             (hdr->AckSeq == (CM_SEQ)(sgmt->MsgSeq - 1)))) /* duplicate ack */
+       {
+           CPRINTF9 (1, "%s: lvl %d sidx %d node %d type %d: H/BEAT !REJOIN (out-of-seq) M(%d,a%d) S%d,A%d\n", 
+                     cmRail->Rail->Name, lvl, sidx, hdr->NodeId, hdr->Type, 
+                     (int)hdr->Seq, (int)hdr->AckSeq, (int)sgmt->MsgSeq, (int)sgmt->AckSeq);
+        
+           printk ("%s: lvl %d sidx %d node %d type %d: H/BEAT !REJOIN (out-of-seq) M(%d,a%d) S%d,A%d\n", 
+                   cmRail->Rail->Name, lvl, sidx, hdr->NodeId, hdr->Type, 
+                   (int)hdr->Seq, (int)hdr->AckSeq, (int)sgmt->MsgSeq, (int)sgmt->AckSeq);
+        
+           SendMessage (cmRail, hdr->NodeId, hdr->Level, CM_MSG_TYPE_REJOIN);
+           break;
+       }
+
+       IncrStat (cmRail, HeartbeatsRcvd);
+
+       sgmt->UpdateTick = lbolt;
+       sgmt->SendMaps = 1;
+
+       if (sgmt->MsgSeq == hdr->AckSeq)                /* acking current message */
+           sgmt->MsgAcked = 1;                 /* can send the next one */
+
+       if (hdr->Seq == sgmt->AckSeq)           /* discard duplicate (or NULL heartbeat) */
+       {
+           CPRINTF6 (6, "%s: lvl %d sidx %d node %d type %d: %s H/BEAT\n", 
+                     cmRail->Rail->Name, lvl, sidx, hdr->NodeId, hdr->Type,
+                     hdr->NumMaps == 0 ? "null" : "duplicate");
+           break;
+       }
+
+       CPRINTF7 (6, "%s: lvl %d sidx %d node %d type %d: seq %d maps %d H/BEAT\n", 
+                 cmRail->Rail->Name, lvl, sidx, hdr->NodeId, hdr->Type, hdr->Seq, hdr->NumMaps);
+
+       sgmt->AckSeq = hdr->Seq;                        /* ready to receive next one */
+       
+       for (i = 0; i < hdr->NumMaps; i++)
+       {
+           CM_STATEMAP_ENTRY *map  = &msg->Payload.Statemaps[CM_MSG_MAP(i)];
+           int                clvl = map->level;
+           
+           if (clvl < 0)                       /* end of message */
+               break;
+
+           if (clvl < sgmt->Level)             /* bad level */
+           {
+               CPRINTF6 (1, "%s: lvl %d sidx %d node %d type %d: H/BEAT !REJOIN (bad clevel %d)\n", 
+                         cmRail->Rail->Name, lvl, sidx, hdr->NodeId, hdr->Type, clvl);
+
+               SendMessage (cmRail, hdr->NodeId, hdr->Level, CM_MSG_TYPE_REJOIN);
+               goto finished;
+           }
+
+           if (map->offset == STATEMAP_NOMORECHANGES) /* end of atomic changes */
+           {
+               if (!sgmt->Maps[clvl].InputMapValid || /* not set InputMap yet */
+                   statemap_changed (sgmt->Maps[clvl].CurrentInputMap)) /* previously applied changes */
+               {
+                   CPRINTF3 (4, "%s: received new clvl %d map from %d\n", cmRail->Rail->Name, clvl, sgmt->NodeId);
+
+                   statemap_setmap (sgmt->Maps[clvl].InputMap, sgmt->Maps[clvl].CurrentInputMap);
+                   sgmt->Maps[clvl].InputMapValid = 1;
+
+                   statemap_clearchanges (sgmt->Maps[clvl].CurrentInputMap);
+               }
+               continue;
+           }
+           
+           seg = ((bitmap_t)map->seg[0])
+               | (((bitmap_t)map->seg[1]) << 16)
+#if (BT_ULSHIFT == 6)
+               | (((bitmap_t)map->seg[2]) << 32)
+               | (((bitmap_t)map->seg[3]) << 48)
+#elif (BT_ULSHIFT != 5)
+#error "Bad value for BT_ULSHIFT"
+#endif
+               ;
+           statemap_setseg (sgmt->Maps[clvl].CurrentInputMap, map->offset, seg);
+       }
+       break;
+
+    case CM_MSG_TYPE_REJOIN:
+       CPRINTF5 (1, "%s: lvl %d sidx %d node %d type %d: REJOIN\n",
+                 cmRail->Rail->Name, lvl, sidx, hdr->NodeId, hdr->Type);
+       printk ("%s: lvl %d sidx %d node %d type %d: REJOIN\n", 
+               cmRail->Rail->Name, lvl, sidx, hdr->NodeId, hdr->Type);
+
+       LowerTopLevel (cmRail, 0);
+
+       IncrStat (cmRail, RejoinRequest);
+       break;
+
+    default:
+       printk ("%s: lvl=%d unknown message type %d\n", cmRail->Rail->Name, lvl, hdr->Type);
+       break;
+    }
+ finished:
+    hdr->Version = EP_SYSTEMQ_UNRECEIVED;
+}
+
+static void
+PollInputQueues (CM_RAIL *cmRail)
+{
+    ep_poll_inputq (cmRail->Rail, cmRail->IntrQueue, 0, ProcessMessage, cmRail);
+    ep_poll_inputq (cmRail->Rail, cmRail->PolledQueue, 0, ProcessMessage, cmRail);
+}
+
+static void
+IntrQueueCallback (EP_RAIL *rail, void *arg)
+{
+    CM_RAIL *cmRail = (CM_RAIL *) arg;
+    unsigned long flags;
+
+    /* If the lock is held, then don't bother spinning for it, 
+     * since the messages will be received at this, or the
+     * next heartbeat */
+    local_irq_save (flags);
+    if (spin_trylock (&cmRail->Lock))
+    {
+       if (AFTER (lbolt, cmRail->NextRunTime + MSEC2TICKS(CM_TIMER_SCHEDULE_TIMEOUT)))
+           printk ("%s: heartbeat timer stuck - scheduled\n", cmRail->Rail->Name);
+       else
+           ep_poll_inputq (rail, cmRail->IntrQueue, 0, ProcessMessage, cmRail);
+       spin_unlock (&cmRail->Lock);
+    }
+    local_irq_restore (flags);
+}
+
+char *
+sprintClPeers (char *str, CM_RAIL *cmRail, int clvl)
+{
+   int clLo     = cmRail->Levels[clvl].MinNodeId;
+   int clHi     = clLo + cmRail->Levels[clvl].NumNodes - 1;
+   int subClLo  = (clvl == 0) ? cmRail->NodeId : cmRail->Levels[clvl - 1].MinNodeId;
+   int subClHi  = subClLo + ((clvl == 0) ? 0 : cmRail->Levels[clvl - 1].NumNodes - 1);
+   
+   if (subClHi == clHi)
+      sprintf (str, "[%d-%d]", clLo, subClLo - 1);
+   else if (subClLo == clLo)
+      sprintf (str, "[%d-%d]", subClHi + 1, clHi);
+   else
+      sprintf (str, "[%d-%d][%d-%d]", clLo, subClLo - 1, subClHi + 1, clHi);
+
+   return (str);
+}
+
+static void
+RestartComms (CM_RAIL *cmRail, int clvl)
+{
+    int             base;
+    int             nodeId;
+    int             lstat;
+    int             numClNodes;
+    int             subClMin;
+    int             subClMax;
+    int             myClId;
+    int             thisClId;
+    
+    myClId     = ClusterIds (cmRail, clvl, &subClMin, &subClMax);
+    base       = myClId * CM_GSTATUS_BITS;
+    numClNodes = cmRail->Levels[clvl].NumNodes;
+
+    statemap_setbits (cmRail->Levels[clvl].LocalMap, base, 
+                     CM_GSTATUS_CLOSING | CM_GSTATUS_MAY_START | CM_GSTATUS_RESTART, CM_GSTATUS_BITS);
+    cmRail->Levels[clvl].Restarting = 1;
+
+    if (cmRail->Levels[clvl].Online)
+    {
+       cmRail->Levels[clvl].Online = 0;
+       
+       for (thisClId = 0; thisClId < numClNodes; thisClId++)
+       {
+           if (thisClId == subClMin)   /* skip sub-cluster; it's just someone in this cluster */
+           {                           /* that wants me to restart */
+               thisClId = subClMax;
+               continue;
+           }
+           
+           nodeId = cmRail->Levels[clvl].MinNodeId + thisClId;
+           base   = thisClId * CM_GSTATUS_BITS;
+           lstat  = statemap_getbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_BITS);
+           
+           if ((lstat & CM_GSTATUS_ACK_MASK) == CM_GSTATUS_MAY_RUN)
+           {
+               switch (ep_disconnect_node (cmRail->Rail, nodeId))
+               {
+               case EP_NODE_CONNECTING:
+                   /* gstat must == RUNNING */
+                   cmRail->Levels[clvl].Connected--;
+                   break;
+               case EP_NODE_DISCONNECTED:
+                   /* CLOSING || STARTING || (lstat & RESTART) */
+                   break;
+               }
+           }
+       }
+    }
+}
+
+static void
+UpdateGlobalStatus (CM_RAIL *cmRail)
+{
+    char            clNodeStr[32];                             /* [%d-%d][%d-%d] */
+    int             nodeId;
+    int             offset;
+    int             base;
+    bitmap_t        gstat;
+    bitmap_t        lgstat;
+    bitmap_t        lstat;
+    int             clvl;
+    int             numClNodes;
+    int             subClMin;
+    int             subClMax;
+    int             myClId;
+    int             thisClId;
+    int             lastClId;
+
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+    {
+       if (!cmRail->Levels[clvl].GlobalMapValid || /* not got the global map yet */
+           !statemap_changed (cmRail->Levels[clvl].GlobalMap)) /* no changes to respond to */
+       {
+           CPRINTF2 (6, "%s: Got invalid or unchanged clvl %d global map\n", cmRail->Rail->Name, clvl);
+           continue;
+       }
+       
+       CPRINTF2 (5, "%s: Got valid changed clvl %d global map\n", cmRail->Rail->Name, clvl);
+       
+       lastClId = -1;
+       myClId = ClusterIds (cmRail, clvl, &subClMin, &subClMax);
+       numClNodes = cmRail->Levels[clvl].NumNodes;
+       
+       while ((offset = statemap_findchange (cmRail->Levels[clvl].GlobalMap, &gstat, 1)) >= 0)
+       {
+           /*
+            * Check every node that this segment covers - however
+            * if the last node we checked in the previous segmemt
+            * is also the first node in this segment, then skip
+            * it.
+            */
+           if ((thisClId = (offset/CM_GSTATUS_BITS)) == lastClId)
+               thisClId++;
+           lastClId = (offset + BT_NBIPUL - 1)/CM_GSTATUS_BITS;
+           
+           /* check each node that might have changed */
+           for ( ; thisClId <= lastClId && thisClId < numClNodes; thisClId++)
+           {
+               base = thisClId * CM_GSTATUS_BITS;
+               nodeId = cmRail->Levels[clvl].MinNodeId + thisClId;
+
+               if (thisClId >= subClMin && thisClId <= subClMax) /* skip sub-cluster */
+                   continue;
+
+               /* This isn't me; I need to sense what this node is driving
+                * (just the starting and running bits) and respond
+                * appropriately...
+                */
+               lgstat = statemap_getbits (cmRail->Levels[clvl].LastGlobalMap, base, CM_GSTATUS_BITS) & CM_GSTATUS_STATUS_MASK;
+               gstat  = statemap_getbits (cmRail->Levels[clvl].GlobalMap,     base, CM_GSTATUS_BITS) & CM_GSTATUS_STATUS_MASK;
+
+               if (lgstat == gstat)            /* no change in peer state */
+                   continue;
+
+               CPRINTF5 (3, "%s: Node %d: lgstat %s, gstat %s, lstat %s\n", cmRail->Rail->Name, nodeId,
+                         GlobalStatusString (cmRail->Levels[clvl].LastGlobalMap, thisClId),
+                         GlobalStatusString (cmRail->Levels[clvl].GlobalMap, thisClId),
+                         GlobalStatusString (cmRail->Levels[clvl].LocalMap, thisClId));
+
+               /* What I'm currently driving as my acknowledgement */
+               lstat = statemap_getbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_BITS);
+
+               switch (gstat)
+               {
+               case CM_GSTATUS_STARTING:
+                   if ((lgstat == CM_GSTATUS_ABSENT || lgstat == CM_GSTATUS_CLOSING) && lstat == CM_GSTATUS_MAY_START)
+                   {
+                       CPRINTF2 (1, "%s: ===================node %d STARTING\n", cmRail->Rail->Name, nodeId);
+                       
+                       ASSERT (cmRail->Rail->Nodes[nodeId].State == EP_NODE_DISCONNECTED);
+
+                       statemap_setbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_MAY_RUN, CM_GSTATUS_BITS);
+                       continue;
+                   }
+                   break;
+                   
+               case CM_GSTATUS_RUNNING:
+                   if ((lgstat == CM_GSTATUS_ABSENT   && lstat == CM_GSTATUS_MAY_START) ||
+                       (lgstat == CM_GSTATUS_STARTING && lstat == CM_GSTATUS_MAY_RUN))
+                   {
+                       CPRINTF3 (1, "%s: ===================node %d%s RUNNING\n", cmRail->Rail->Name, nodeId,
+                                 lgstat == CM_GSTATUS_ABSENT ? " Already" : "");
+
+                       ASSERT (cmRail->Rail->Nodes[nodeId].State == EP_NODE_DISCONNECTED);
+
+                       if (cmRail->Levels[clvl].Online)
+                       {
+                           ep_connect_node (cmRail->Rail, nodeId);
+
+                           cmRail->Levels[clvl].Connected++;
+                       }
+
+                       statemap_setbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_MAY_RUN, CM_GSTATUS_BITS);
+                       continue;
+                   }
+                   break;
+                   
+               case CM_GSTATUS_CLOSING:
+                   CPRINTF4 (1, "%s: ===================node %d CLOSING%s%s\n", cmRail->Rail->Name, nodeId,
+                             (lstat & CM_GSTATUS_RESTART) ? " for Restart" : "",
+                             cmRail->Levels[clvl].Online ? "" : " (offline)");
+
+                   if ((lstat & CM_GSTATUS_ACK_MASK) == CM_GSTATUS_MAY_RUN)
+                   {
+                       switch (ep_disconnect_node (cmRail->Rail, nodeId))
+                       {
+                       case EP_NODE_CONNECTING:
+                           cmRail->Levels[clvl].Connected--;
+                           /* DROPTHROUGH */
+                       case EP_NODE_DISCONNECTED:
+                           lstat = CM_GSTATUS_MAY_START;
+                           break;
+                       }
+                   }
+
+                   if ((lstat & CM_GSTATUS_ACK_MASK) == CM_GSTATUS_MAY_START) /* clear restart if we've disconnected */
+                       statemap_setbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_MAY_START, CM_GSTATUS_BITS);
+                   continue;
+                   
+               default:
+                   break;
+               }
+
+               /* "unexpected" state change forces me to ask her to restart */
+               if (! (lstat & CM_GSTATUS_RESTART))             /* not requesting restart already */
+               {
+                   CPRINTF5 (1, "%s: ===================node %d %s, old %s new %s\n", cmRail->Rail->Name, nodeId,
+                             (gstat == CM_GSTATUS_ABSENT)  ? "ABSENT" : "REQUEST RESTART", 
+                             GlobalStatusString (cmRail->Levels[clvl].LastGlobalMap, thisClId),
+                             GlobalStatusString (cmRail->Levels[clvl].GlobalMap, thisClId));
+
+                   /* request restart */
+                   if (cmRail->Levels[clvl].Online && lstat == CM_GSTATUS_MAY_RUN)
+                   {
+                       switch (ep_disconnect_node (cmRail->Rail, nodeId))
+                       {
+                       case EP_NODE_CONNECTING:
+                           cmRail->Levels[clvl].Connected--;
+                           /* DROPTHROUGH */
+                       case EP_NODE_DISCONNECTED:
+                           lstat = CM_GSTATUS_MAY_START;
+                           break;
+                       }
+                   }
+
+                   statemap_setbits (cmRail->Levels[clvl].LocalMap, base, lstat | CM_GSTATUS_RESTART, CM_GSTATUS_BITS);
+                   continue;
+               }
+
+               continue;
+           }
+       }
+    
+       /* Now check myself - see what everyone else thinks I'm doing */
+       base   = myClId * CM_GSTATUS_BITS;
+       lstat  = statemap_getbits (cmRail->Levels[clvl].LocalMap,  base, CM_GSTATUS_BITS);
+       gstat  = statemap_getbits (cmRail->Levels[clvl].GlobalMap, base, CM_GSTATUS_BITS);
+       lgstat = statemap_getbits (cmRail->Levels[clvl].LastGlobalMap, base, CM_GSTATUS_BITS);
+
+       if (lgstat == gstat)                    /* my state in this cluster hasn't changed */
+       {
+           CPRINTF3 (6, "%s: my clvl %d global status unchanged from %s\n", cmRail->Rail->Name,
+                     clvl, GlobalStatusString (cmRail->Levels[clvl].GlobalMap, myClId));
+           goto all_done;
+       }
+
+       if ((gstat & CM_GSTATUS_RESTART) != 0)  /* someone wants me to restart */
+       {
+           if ((lstat & CM_GSTATUS_STATUS_MASK) == CM_GSTATUS_CLOSING) /* I'm already restarting */
+               goto all_done;
+           
+           CPRINTF2 (1, "%s: ===================RESTART REQUEST from %s\n", cmRail->Rail->Name,
+                     sprintClPeers (clNodeStr, cmRail, clvl));
+           
+           printk ("%s: Restart Request from %s\n", cmRail->Rail->Name,
+                   sprintClPeers (clNodeStr, cmRail, clvl));
+           
+           RestartComms (cmRail, clvl);
+           goto all_done;
+       }
+       
+       CPRINTF6 (5, "%s: clvl %d: lgstat %s gstat %s, lstat %s%s\n", cmRail->Rail->Name, clvl,
+                 GlobalStatusString (cmRail->Levels[clvl].LastGlobalMap, myClId),
+                 GlobalStatusString (cmRail->Levels[clvl].GlobalMap, myClId),
+                 GlobalStatusString (cmRail->Levels[clvl].LocalMap, myClId),
+                 (gstat != lstat) ? " (IGNORED)" : "");
+                       
+       if (gstat != lstat)                     /* not everyone agrees with me */
+           goto all_done;
+
+       switch (lstat)
+       {
+       default:
+           ASSERT (0);                         /* I never drive this */
+           
+       case CM_GSTATUS_CLOSING | CM_GSTATUS_MAY_START: /* I can restart now (have seen restart go away) */
+           ASSERT (!cmRail->Levels[clvl].Online);
+           
+           CPRINTF2 (1,"%s: ===================NODES %s AGREE I MAY START\n", cmRail->Rail->Name,
+                     sprintClPeers (clNodeStr, cmRail, clvl));
+           printk ("%s: ===================NODES %s AGREE I MAY START\n", cmRail->Rail->Name,
+                   sprintClPeers (clNodeStr, cmRail, clvl));
+           
+           statemap_setbits (cmRail->Levels[clvl].LocalMap, base, 
+                             CM_GSTATUS_STARTING | CM_GSTATUS_MAY_RUN, CM_GSTATUS_BITS);
+           goto all_done;
+           
+       case CM_GSTATUS_STARTING | CM_GSTATUS_MAY_RUN:
+           ASSERT (!cmRail->Levels[clvl].Online);
+           
+           CPRINTF2 (1, "%s: ===================NODES %s AGREE I MAY RUN\n", cmRail->Rail->Name,
+                     sprintClPeers (clNodeStr, cmRail, clvl));
+           printk ("%s: ===================NODES %s AGREE I MAY RUN\n", cmRail->Rail->Name,
+                   sprintClPeers (clNodeStr, cmRail, clvl));
+           
+           statemap_setbits (cmRail->Levels[clvl].LocalMap, base, 
+                             CM_GSTATUS_RUNNING | CM_GSTATUS_MAY_RUN, CM_GSTATUS_BITS);
+           goto all_done;
+           
+       case CM_GSTATUS_RUNNING | CM_GSTATUS_MAY_RUN:
+           if (! cmRail->Levels[clvl].Online)
+           {
+               CPRINTF2 (1, "%s: ===================NODES %s AGREE I'M RUNNING\n", cmRail->Rail->Name,
+                         sprintClPeers (clNodeStr, cmRail, clvl));
+               printk ("%s: ===================NODES %s AGREE I'M RUNNING\n", cmRail->Rail->Name,
+                       sprintClPeers (clNodeStr, cmRail, clvl));
+               
+               cmRail->Levels[clvl].Online = 1;
+               
+               for (thisClId = 0; thisClId < numClNodes; thisClId++)
+               {
+                   if (thisClId == subClMin)   /* skip sub-cluster */
+                   {
+                       thisClId = subClMax;
+                       continue;
+                   }
+                   
+                   nodeId = cmRail->Levels[clvl].MinNodeId + thisClId;
+                   
+                   base  = thisClId * CM_GSTATUS_BITS;
+                   lstat = statemap_getbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_BITS);
+                   gstat = statemap_getbits (cmRail->Levels[clvl].GlobalMap, base, CM_GSTATUS_BITS) & CM_GSTATUS_STATUS_MASK;
+                   
+                   /* Only connect to her if I see her as running and I'm not requesting her 
+                    * to restart - this means that I was offline when I saw her transition
+                    * to running and haven't seen her in a "bad" state since. */
+                   if (gstat == CM_GSTATUS_RUNNING && ! (lstat & CM_GSTATUS_RESTART))
+                   {
+                       CPRINTF5 (1, "%s: node %d lgstat %s gstat %s, lstat %s -> CONNECT\n", cmRail->Rail->Name, nodeId,
+                                 GlobalStatusString (cmRail->Levels[clvl].LastGlobalMap, thisClId),
+                                 GlobalStatusString (cmRail->Levels[clvl].GlobalMap, thisClId),
+                                 GlobalStatusString (cmRail->Levels[clvl].LocalMap, thisClId));
+                       
+                       if (lstat == CM_GSTATUS_MAY_START)
+                           statemap_setbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_MAY_RUN, CM_GSTATUS_BITS);
+
+                       ep_connect_node (cmRail->Rail, nodeId);
+
+                       cmRail->Levels[clvl].Connected++;
+                   }
+               }
+           }
+           goto all_done;
+       }
+
+    all_done:
+       statemap_setmap (cmRail->Levels[clvl].LastGlobalMap, cmRail->Levels[clvl].GlobalMap);
+    }
+}
+
+static void
+ReduceGlobalMap (CM_RAIL *cmRail, int clvl)
+{
+    int       lvl;
+    int       sidx;
+    int       recompute;
+    CM_LEVEL *level;
+    int       cTopLevel;
+    int       cRole;
+
+    if (clvl < cmRail->TopLevel)
+    {
+       cTopLevel = clvl + 1;
+       cRole = CM_ROLE_LEADER;
+    }
+    else
+    {
+       cTopLevel = cmRail->TopLevel;
+       cRole = cmRail->Role;
+    }
+    
+    /* Update cmRail->Levels[*].SubordinateMap[clvl] for all subordinate levels */
+    for (lvl = 0; lvl < cTopLevel; lvl++)
+    {
+       level = &cmRail->Levels[lvl];
+
+       /* We need to recompute this level's statemap if...
+        * . Previous level's statemap has changes to propagate OR
+        * . This level's statemap has not been computed yet OR
+        * . A subordinate at this level has sent me a change.
+        * Note that we can only do this if all subordinates from this
+        * level down are present with valid statemaps, or absent (i.e. not
+        * timing out).
+        */
+
+       ASSERT (lvl == 0 || cmRail->Levels[lvl - 1].SubordinateMapValid[clvl]);
+
+       recompute = !level->SubordinateMapValid[clvl] ||
+                   (lvl > 0 && statemap_changed (cmRail->Levels[lvl - 1].SubordinateMap[clvl]));
+         
+       for (sidx = 0; sidx < level->NumSegs; sidx++)
+       {
+           CM_SGMT *sgmt = &level->Sgmts[sidx];
+
+           if (!(sgmt->State == CM_SGMT_ABSENT || /* absent nodes contribute zeros */
+                 (sgmt->State == CM_SGMT_PRESENT && /* present nodes MUST have received a map to contribute */
+                  sgmt->Maps[clvl].InputMapValid)))
+           {
+               CPRINTF5 (5, "%s: waiting for clvl %d lvl %d seg %d node %d\n", cmRail->Rail->Name,
+                         clvl, lvl, sidx, sgmt->NodeId);
+
+               /* Gotta wait for this guy, so we can't compute this level,
+                * or any higher levels. */
+               return;
+           }
+
+           if (statemap_changed (sgmt->Maps[clvl].InputMap))
+           {
+               ASSERT (sgmt->Maps[clvl].InputMapValid);
+
+               recompute = 1;
+
+               CPRINTF7 (5, "%s: %s clvl %d map from @ %d %d (%d) - %s\n",
+                         cmRail->Rail->Name, sgmt->State == CM_SGMT_ABSENT ? "newly absent" : "got new",
+                         clvl, lvl, sidx, sgmt->NodeId, 
+                         MapString ("Input", sgmt->Maps[clvl].InputMap, cmRail->Levels[clvl].NumNodes, ""));
+           }
+       }
+
+       if (recompute)
+       {
+           if (lvl == 0)
+               statemap_reset (cmRail->Levels[clvl].TmpMap);
+           else
+           {
+               ASSERT (cmRail->Levels[lvl - 1].SubordinateMapValid[clvl]);
+               
+               statemap_copy (cmRail->Levels[clvl].TmpMap, cmRail->Levels[lvl - 1].SubordinateMap[clvl]);
+               statemap_clearchanges (cmRail->Levels[lvl - 1].SubordinateMap[clvl]);
+           }
+        
+           for (sidx = 0; sidx < level->NumSegs; sidx++)
+           {
+               CM_SGMT *sgmt = &level->Sgmts[sidx];
+               
+               if (sgmt->State != CM_SGMT_ABSENT)      /* absent nodes contribute zeroes */
+               {
+                   ASSERT (sgmt->State == CM_SGMT_PRESENT);
+                   ASSERT (sgmt->Maps[clvl].InputMapValid);
+                   statemap_ormap (cmRail->Levels[clvl].TmpMap, sgmt->Maps[clvl].InputMap);
+               }
+               statemap_clearchanges (sgmt->Maps[clvl].InputMap);
+           }
+
+           statemap_setmap (level->SubordinateMap[clvl], cmRail->Levels[clvl].TmpMap);
+           level->SubordinateMapValid[clvl] = 1;
+
+           CPRINTF4 (5, "%s: recompute clvl %d level %d statemap - %s\n", cmRail->Rail->Name, clvl, lvl,
+                     MapString ("level", level->SubordinateMap[clvl], cmRail->Levels[clvl].NumNodes, ""));
+       }
+    }
+
+    if (cRole == CM_ROLE_LEADER_CANDIDATE)     /* don't know this cluster's leader yet */
+       return;
+
+    ASSERT (cTopLevel == 0 || cmRail->Levels[cTopLevel - 1].SubordinateMapValid[clvl]);
+
+    /* Update SubTreeMap */
+    
+    if (!cmRail->Levels[clvl].SubTreeMapValid || 
+       statemap_changed (cmRail->Levels[clvl].LocalMap) ||
+       (cTopLevel > 0 && statemap_changed (cmRail->Levels[cTopLevel - 1].SubordinateMap[clvl])))
+    {
+       statemap_copy (cmRail->Levels[clvl].TmpMap, cmRail->Levels[clvl].LocalMap);
+       statemap_clearchanges (cmRail->Levels[clvl].LocalMap);
+
+       if (cTopLevel > 0)
+       {
+           statemap_ormap (cmRail->Levels[clvl].TmpMap, cmRail->Levels[cTopLevel - 1].SubordinateMap[clvl]);
+           statemap_clearchanges (cmRail->Levels[cTopLevel - 1].SubordinateMap[clvl]);
+       }
+
+       statemap_setmap (cmRail->Levels[clvl].SubTreeMap, cmRail->Levels[clvl].TmpMap);
+       cmRail->Levels[clvl].SubTreeMapValid = 1;
+
+       CPRINTF3 (5, "%s: recompute clvl %d subtree map - %s\n", cmRail->Rail->Name, clvl,
+                 MapString ("subtree", cmRail->Levels[clvl].SubTreeMap, cmRail->Levels[clvl].NumNodes, ""));
+    }
+
+    if (cRole == CM_ROLE_SUBORDINATE)          /* got a leader (Not me) */
+    {                                          /* => send SubTreeMap to her */
+       CM_SGMT *leader = &cmRail->Levels[cmRail->TopLevel].Sgmts[0];
+
+       ASSERT (leader->State == CM_SGMT_PRESENT);
+       ASSERT (cmRail->Levels[clvl].SubTreeMapValid);
+
+       if (!leader->Maps[clvl].OutputMapValid ||
+           statemap_changed (cmRail->Levels[clvl].SubTreeMap))
+       {
+           statemap_setmap (leader->Maps[clvl].OutputMap, cmRail->Levels[clvl].SubTreeMap);
+           leader->Maps[clvl].OutputMapValid = 1;
+
+           statemap_clearchanges (cmRail->Levels[clvl].SubTreeMap);
+
+           CPRINTF3 (5, "%s: sending clvl %d subtree map to leader (%d)\n", cmRail->Rail->Name, clvl, leader->NodeId);
+       }
+    }
+}
+
+void
+BroadcastGlobalMap (CM_RAIL *cmRail, int clvl)
+{
+    int       lvl;
+    int       sidx;
+    CM_LEVEL *level;
+    CM_SGMT  *leader;
+    int       cTopLevel;
+    int       cRole;
+
+    if (clvl < cmRail->TopLevel)
+    {
+       cTopLevel = clvl + 1;
+       cRole = CM_ROLE_LEADER;
+    }
+    else
+    {
+       cTopLevel = cmRail->TopLevel;
+       cRole = cmRail->Role;
+    }
+
+    switch (cRole)
+    {
+    default:
+       ASSERT (0);
+       
+    case CM_ROLE_LEADER_CANDIDATE:             /* don't know this cluster's leader yet */
+       return;
+
+    case CM_ROLE_LEADER:                       /* cluster leader: */
+       ASSERT (clvl < cmRail->TopLevel);               /* set GlobalMap from SubTreeMap */
+       
+       if (!cmRail->Levels[clvl].SubTreeMapValid)      /* can't set global map */
+           return;
+
+       if (cmRail->Levels[clvl].GlobalMapValid &&      /* already set global map */
+           !statemap_changed (cmRail->Levels[clvl].SubTreeMap)) /* no changes to propagate */
+           return;
+       
+       statemap_setmap (cmRail->Levels[clvl].GlobalMap, cmRail->Levels[clvl].SubTreeMap);
+       cmRail->Levels[clvl].GlobalMapValid = 1;
+       statemap_clearchanges (cmRail->Levels[clvl].SubTreeMap);
+
+       CPRINTF2 (5, "%s: whole cluster %d leader setting global map\n", cmRail->Rail->Name, clvl);
+
+       UpdateGlobalStatus (cmRail);
+       break;
+       
+    case CM_ROLE_SUBORDINATE:                  /* cluster subordinate: */
+       ASSERT (clvl >= cmRail->TopLevel);              /* receive GlobalMap from leader */
+       ASSERT (cmRail->TopLevel < cmRail->NumLevels);
+       
+       leader = &cmRail->Levels[cmRail->TopLevel].Sgmts[0];
+       ASSERT (leader->State == CM_SGMT_PRESENT);
+
+       if (!leader->Maps[clvl].InputMapValid)  /* can't set global map */
+           return;
+       
+       if (cmRail->Levels[clvl].GlobalMapValid &&      /* already set global map */
+           !statemap_changed (leader->Maps[clvl].InputMap)) /* no changes to propagate */
+           return;
+
+       statemap_setmap (cmRail->Levels[clvl].GlobalMap, leader->Maps[clvl].InputMap);
+       cmRail->Levels[clvl].GlobalMapValid = 1;
+       statemap_clearchanges (leader->Maps[clvl].InputMap);
+
+       CPRINTF3 (5, "%s: getting clvl %d global map from leader (%d)\n", cmRail->Rail->Name, clvl, leader->NodeId);
+
+       UpdateGlobalStatus (cmRail);
+       break;
+    }
+
+    CPRINTF3 (5, "%s: clvl %d %s\n", cmRail->Rail->Name, clvl,
+             MapString ("global", cmRail->Levels[clvl].GlobalMap, cmRail->Levels[clvl].NumNodes, ""));
+    
+    /* Broadcast global map to all subordinates */
+    for (lvl = 0; lvl < cTopLevel; lvl++)
+    {
+       level = &cmRail->Levels[lvl];
+       
+       for (sidx = 0; sidx < level->NumSegs; sidx++)
+       {
+           CM_SGMT *sgmt = &level->Sgmts[sidx];
+           
+           if (sgmt->State == CM_SGMT_PRESENT)
+           {
+               statemap_setmap (sgmt->Maps[clvl].OutputMap, cmRail->Levels[clvl].GlobalMap);
+               sgmt->Maps[clvl].OutputMapValid = 1;
+               
+               CPRINTF5 (5, "%s: sending clvl %d global map to subordinate %d %d (%d)\n", 
+                         cmRail->Rail->Name, clvl, lvl, sidx, sgmt->NodeId);
+           }
+       }
+    }
+}
+
+static void
+CheckPeerPulse (CM_RAIL *cmRail, CM_SGMT *sgmt)
+{
+    int clvl, sendRejoin;
+    
+    switch (sgmt->State)
+    {
+    case CM_SGMT_ABSENT:
+       break;
+
+    case CM_SGMT_WAITING:                      /* waiting for a subtree */
+       if (!AFTER (lbolt, sgmt->UpdateTick + MSEC2TICKS(CM_DISCOVER_TIMEOUT)))
+           break;
+      
+       CPRINTF3 (2, "%s: lvl %d subtree %d contains no live nodes\n", cmRail->Rail->Name, 
+                 sgmt->Level, (int) (sgmt - &cmRail->Levels[sgmt->Level].Sgmts[0]));
+
+       sgmt->State = CM_SGMT_ABSENT;
+       for (clvl = sgmt->Level; clvl < cmRail->NumLevels; clvl++)
+       {
+           statemap_zero (sgmt->Maps[clvl].InputMap);          /* need to start propagating zeros (flags change) */
+           sgmt->Maps[clvl].InputMapValid = 1;         /* and must indicate that the map is now valid */
+       }
+       break;
+
+    case CM_SGMT_COMING:                               /* lost/waiting subtree sent me IMCOMING */
+       ASSERT (sgmt->Level > 0);                       /* we only do subtree discovery below our own level */
+
+       if (AFTER (lbolt, sgmt->WaitingTick + MSEC2TICKS(CM_WAITING_TIMEOUT)))
+       {
+           CPRINTF3 (1, "%s: lvl %d subtree %d waiting too long\n", cmRail->Rail->Name,
+                     sgmt->Level, (int) (sgmt - &cmRail->Levels[sgmt->Level].Sgmts[0]));
+           printk ("%s: lvl %d subtree %d waiting too long\n", cmRail->Rail->Name,
+                   sgmt->Level, (int) (sgmt - &cmRail->Levels[sgmt->Level].Sgmts[0]));
+
+           sgmt->State = CM_SGMT_ABSENT;
+           for (clvl = sgmt->Level; clvl < cmRail->NumLevels; clvl++)
+           {
+               statemap_zero (sgmt->Maps[clvl].InputMap);              /* need to start propagating zeros (flags change) */
+               sgmt->Maps[clvl].InputMapValid = 1;             /* and must indicate that the map is now valid */
+           }
+           break;
+       }
+
+       if (!AFTER (lbolt, sgmt->UpdateTick + MSEC2TICKS(CM_DISCOVER_TIMEOUT)))
+           break;
+
+       CPRINTF3 (2, "%s: lvl %d subtree %d hasn't connected yet\n", cmRail->Rail->Name,
+                 sgmt->Level, (int) (sgmt - &cmRail->Levels[sgmt->Level].Sgmts[0]));
+
+       sgmt->State = CM_SGMT_WAITING;
+       sgmt->UpdateTick = lbolt;
+
+       if (sgmt->Level > 0)
+           __Schedule_Discovery (cmRail);
+       break;
+      
+    case CM_SGMT_PRESENT:
+       if (!AFTER (lbolt, sgmt->UpdateTick + MSEC2TICKS(CM_HEARTBEAT_TIMEOUT)))
+           break;
+
+       if (sgmt->Level == cmRail->TopLevel)            /* leader died */
+       {
+           sendRejoin = (sgmt->State == CM_SGMT_PRESENT && sgmt->AckSeq == 0);
+
+           CPRINTF4 (1, "%s: leader (%d) node %d JUST DIED%s\n", 
+                     cmRail->Rail->Name, sgmt->Level, sgmt->NodeId,
+                     sendRejoin ? ": !REJOIN" : "");
+           
+           printk ("%s: lvl %d leader (%d) JUST DIED%s\n", 
+                   cmRail->Rail->Name, sgmt->Level, sgmt->NodeId,
+                   sendRejoin ? ": !REJOIN" : "");
+       
+           if (sendRejoin)
+           {
+               /* she's not sent us any heartbeats even though she responded to a discover
+                * so tell her to rejoin the tree at the bottom, this will mean that she 
+                * has to run the heartbeat timer before being able to rejoin the tree. */
+               SendMessage (cmRail, sgmt->NodeId, sgmt->Level, CM_MSG_TYPE_REJOIN);
+           }
+
+           StartLeaderDiscovery (cmRail);
+           break;
+       }
+
+       sendRejoin = (sgmt->State == CM_SGMT_PRESENT && sgmt->AckSeq == 0);
+
+       CPRINTF5 (2, "%s: lvl %d subordinate %d (%d) JUST DIED%s\n", cmRail->Rail->Name, 
+                 sgmt->Level, (int) (sgmt - &cmRail->Levels[sgmt->Level].Sgmts[0]), sgmt->NodeId,
+                 sendRejoin ? ": !REJOIN" : "");
+       printk ("%s: lvl %d subordinate %d (%d) JUST DIED%s\n", cmRail->Rail->Name, 
+               sgmt->Level, (int) (sgmt - &cmRail->Levels[sgmt->Level].Sgmts[0]), sgmt->NodeId,
+               sendRejoin ? ": !REJOIN" : "");
+
+       if (sendRejoin)
+       {
+           /* she's not sent us any heartbeats even though she responded to a discover
+            * so tell her to rejoin the tree at the bottom, this will mean that she 
+            * has to run the heartbeat timer before being able to rejoin the tree. */
+           SendMessage (cmRail, sgmt->NodeId, sgmt->Level, CM_MSG_TYPE_REJOIN);
+       }
+
+       StartSubTreeDiscovery (cmRail, sgmt);
+       break;
+        
+    default:
+       ASSERT (0);
+    }
+}
+
+static void
+CheckPeerPulses (CM_RAIL *cmRail)
+{
+    int lvl;
+    int sidx;
+   
+    /* check children are alive */
+    for (lvl = 0; lvl < cmRail->TopLevel; lvl++)
+       for (sidx = 0; sidx < cmRail->Levels[lvl].NumSegs; sidx++)
+           CheckPeerPulse (cmRail, &cmRail->Levels[lvl].Sgmts[sidx]);
+
+    /* check leader is alive */
+    if (cmRail->Role == CM_ROLE_SUBORDINATE)
+    {
+       ASSERT (cmRail->TopLevel < cmRail->NumLevels);
+       ASSERT (cmRail->Levels[cmRail->TopLevel].Sgmts[0].State == CM_SGMT_PRESENT);
+      
+       CheckPeerPulse (cmRail, &cmRail->Levels[cmRail->TopLevel].Sgmts[0]);
+    }
+}
+
+static void
+SendHeartbeats (CM_RAIL *cmRail)
+{
+    int lvl;
+
+    /* Send heartbeats to my children */
+    for (lvl = 0; lvl < cmRail->TopLevel; lvl++)
+    {
+       CM_LEVEL *level = &cmRail->Levels[lvl];
+       int       sidx;
+       
+       for (sidx = 0; sidx < level->NumSegs; sidx++)
+       {
+           CM_SGMT *sgmt = &cmRail->Levels[lvl].Sgmts[sidx];
+
+           if (sgmt->State == CM_SGMT_PRESENT)
+               SendToSgmt (cmRail, sgmt, CM_MSG_TYPE_HEARTBEAT);
+       }
+    }
+
+    /* Send heartbeat to my leader */
+    if (cmRail->Role == CM_ROLE_SUBORDINATE)
+    {
+       ASSERT (cmRail->TopLevel < cmRail->NumLevels);
+       SendToSgmt (cmRail, &cmRail->Levels[cmRail->TopLevel].Sgmts[0], CM_MSG_TYPE_HEARTBEAT);
+    }
+}
+
+static int
+BroadcastDiscover (CM_RAIL *cmRail)
+{
+    int       sidx;
+    int              lvl;
+    int       msgType;
+    CM_LEVEL *level;
+    int       urgent;
+
+    ASSERT (cmRail->TopLevel <= cmRail->NumLevels);
+    ASSERT ((cmRail->Role == CM_ROLE_LEADER) ? (cmRail->TopLevel == cmRail->NumLevels) :
+           (cmRail->Role == CM_ROLE_SUBORDINATE) ? (cmRail->Levels[cmRail->TopLevel].Sgmts[0].State == CM_SGMT_PRESENT) :
+           (cmRail->Role == CM_ROLE_LEADER_CANDIDATE));
+
+    if (cmRail->Role != CM_ROLE_LEADER_CANDIDATE)      /* got a leader/lead whole machine */
+    {
+       urgent = 0;                             /* non-urgent leader discovery */
+       lvl = cmRail->TopLevel - 1;             /* on nodes I lead (resolves leader conflicts) */
+       msgType = CM_MSG_TYPE_RESOLVE_LEADER;
+    }
+    else
+    {
+       urgent = 1;                             /* urgent leader discovery */
+       lvl = cmRail->TopLevel;                 /* on nodes I'd like to lead */
+       msgType = CM_MSG_TYPE_DISCOVER_LEADER;
+    }
+
+    if (lvl >= 0)
+    {
+       if (lvl > cmRail->BroadcastLevel)
+       {
+           /* Unable to broadcast at this level in the spanning tree, so we 
+            * just continue doing discovery until we are able to broadcast */
+           CPRINTF4 (6, "%s: broadcast level %d too low to discover %d at level %d\n",
+                     cmRail->Rail->Name, cmRail->BroadcastLevel, msgType, lvl);
+
+           cmRail->DiscoverStartTick = lbolt;
+       }
+       else
+       {
+           level = &cmRail->Levels[lvl];
+           SendToSgmt (cmRail, &level->Sgmts[level->MySgmt], msgType);
+       }
+    }
+    
+    while (lvl > 0)
+    {
+       level = &cmRail->Levels[lvl];
+      
+       for (sidx = 0; sidx < level->NumSegs; sidx++)
+       {
+           CM_SGMT *sgmt = &level->Sgmts[sidx];
+        
+           if (sgmt->State == CM_SGMT_WAITING)
+           {
+               ASSERT (sidx != level->MySgmt);
+               /* Do subordinate discovery.  Existing subordinates will
+                * ignore it, but leader candidates will send IMCOMING.
+                * This is always urgent since we'll assume a subtree is
+                * absent if I don't get IMCOMING within the timeout.
+                */
+               SendToSgmt (cmRail, sgmt, CM_MSG_TYPE_DISCOVER_SUBORDINATE);
+               urgent = 1;
+           }
+       }
+       lvl--;
+    }
+   
+    return (urgent);
+}
+
+static void
+CheckBroadcast (CM_RAIL *cmRail)
+{
+    int  clvl;
+
+    for (clvl = cmRail->NumLevels-1; clvl >= 0 && cmRail->Rail->SwitchBroadcastLevel < cmRail->Levels[clvl].SwitchLevel; clvl--)
+       ;
+
+    if (cmRail->OfflineReasons || cmRail->Rail->System->Shutdown)
+       clvl = -1;
+
+    /* if the level at which we can broadcast drops, then we must rejoin the
+     * spanning tree at the highest level for which broadcast is good. */
+    if (cmRail->BroadcastLevel > clvl && clvl < (int)(cmRail->Role == CM_ROLE_LEADER ? cmRail->TopLevel - 1 : cmRail->TopLevel))
+    {
+       printk ("%s: REJOINING at level %d because %s\n", cmRail->Rail->Name, clvl+1, 
+               (cmRail->OfflineReasons & CM_OFFLINE_MANAGER) ? "of manager thread" :
+               (cmRail->OfflineReasons & CM_OFFLINE_PROCFS)  ? "force offline"  : 
+               cmRail->Rail->System->Shutdown ? "system shutdown" : "broadcast level changed");
+       LowerTopLevel (cmRail, clvl+1);
+    }
+    
+    if (cmRail->BroadcastLevel != clvl)
+    {
+       cmRail->BroadcastLevel     = clvl;
+       cmRail->BroadcastLevelTick = lbolt;
+    }
+
+    /* schedule the update thread, to withdraw from comms with 
+     * nodes "outside" of the valid broadcastable range. */
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+    {
+       if (cmRail->BroadcastLevel < clvl)
+       {
+           if (AFTER (lbolt, cmRail->BroadcastLevelTick + EP_WITHDRAW_TIMEOUT) && 
+               !(cmRail->Levels[clvl].OfflineReasons & CM_OFFLINE_BROADCAST))
+           {
+               printk ("%s: Withdraw at Level %d\n", cmRail->Rail->Name, clvl);
+               cmRail->Levels[clvl].OfflineReasons |= CM_OFFLINE_BROADCAST;
+           }
+       }
+       else
+       {
+           if (cmRail->Levels[clvl].OfflineReasons & CM_OFFLINE_BROADCAST)
+           {
+               printk ("%s: Rejoin at Level %d\n", cmRail->Rail->Name, clvl);
+               cmRail->Levels[clvl].OfflineReasons &= ~CM_OFFLINE_BROADCAST;
+           }
+       }
+    }
+       
+}
+
+static void
+CheckManager (CM_RAIL *cmRail)
+{
+    long time,  state = ep_kthread_state (&cmRail->Rail->System->ManagerThread, &time);
+
+    if (state == KT_STATE_RUNNING && BEFORE (lbolt, time + MSEC2TICKS(CM_THREAD_RUNNING_TIMEOUT)))
+       state = KT_STATE_SLEEPING;
+    if (state != KT_STATE_SLEEPING && BEFORE (lbolt, time + MSEC2TICKS(CM_THREAD_SCHEDULE_TIMEOUT)))
+       state = KT_STATE_SLEEPING;
+
+    if ((cmRail->OfflineReasons & CM_OFFLINE_MANAGER) && state == KT_STATE_SLEEPING)
+    {
+       printk ("%s: manager thread unstuck\n", cmRail->Rail->Name);
+
+       cmRail->OfflineReasons &= ~CM_OFFLINE_MANAGER;
+    }
+
+    if (!(cmRail->OfflineReasons & CM_OFFLINE_MANAGER) && state != KT_STATE_SLEEPING)
+    {
+       printk ("%s: manager thread stuck - %s\n", cmRail->Rail->Name,
+               state == KT_STATE_SCHEDULED ? "scheduled" : 
+               state == KT_STATE_RUNNING ? "running" : 
+               state == KT_STATE_STALLED ? "stalled" : "unknown");
+
+       cmRail->OfflineReasons |= CM_OFFLINE_MANAGER;
+    }
+}
+
+static void
+CheckOfflineReasons (CM_RAIL *cmRail, int clvl)
+{
+    int subClMin, subClMax, myClId;
+    char clNodeStr[32];                                /* [%d-%d][%d-%d] */
+
+    if (cmRail->Levels[clvl].OfflineReasons)
+    {
+       if (cmRail->Levels[clvl].Online)
+       {
+           printk ("%s: Withdraw from %s\n", cmRail->Rail->Name, sprintClPeers (clNodeStr, cmRail, clvl));
+           
+           RestartComms (cmRail, clvl);
+       }
+    }
+    else
+    {
+       if (cmRail->Levels[clvl].Restarting && cmRail->Levels[clvl].Connected == 0)
+       {
+           printk ("%s: Rejoin with %s\n", cmRail->Rail->Name, sprintClPeers (clNodeStr, cmRail, clvl));
+
+           myClId = ClusterIds (cmRail, clvl, &subClMin, &subClMax);
+           
+           ASSERT (statemap_getbits (cmRail->Levels[clvl].LocalMap, myClId * CM_GSTATUS_BITS, CM_GSTATUS_BITS) == 
+                   (CM_GSTATUS_CLOSING | CM_GSTATUS_MAY_START | CM_GSTATUS_RESTART));
+    
+           statemap_setbits (cmRail->Levels[clvl].LocalMap, myClId * CM_GSTATUS_BITS,
+                             CM_GSTATUS_CLOSING | CM_GSTATUS_MAY_START, CM_GSTATUS_BITS);
+
+           cmRail->Levels[clvl].Restarting = 0;
+       }
+    }
+}
+
+void
+DoHeartbeatWork (CM_RAIL *cmRail)
+{
+    long now = lbolt;
+    int  clvl;
+
+    if ((RejoinCheck || RejoinPanic) &&
+       AFTER (now, cmRail->NextRunTime + MSEC2TICKS (CM_TIMER_SCHEDULE_TIMEOUT))) /* If I've been unresponsive for too long */
+    {
+       /* I'd better reconnect to the network because I've not been playing the game */
+       CPRINTF4 (1, "%s: REJOINING because I was too slow (heartbeat) [%ld,%ld,(%ld)]\n", cmRail->Rail->Name, now,  cmRail->NextRunTime, (long int)MSEC2TICKS (CM_TIMER_SCHEDULE_TIMEOUT));
+       printk ("%s: REJOINING because I was too slow (heartbeat) [%ld,%ld,(%ld)]\n", cmRail->Rail->Name, now,  cmRail->NextRunTime, (long int)MSEC2TICKS (CM_TIMER_SCHEDULE_TIMEOUT));
+       
+       LowerTopLevel (cmRail, 0);
+       
+       IncrStat (cmRail, RejoinTooSlow);
+       
+       if (RejoinPanic)
+           panic ("ep: REJOINING because I was too slow (heartbeat)\n");
+    }
+    
+    PollInputQueues (cmRail);
+    
+    if (cmRail->NextDiscoverTime && ! BEFORE (now, cmRail->NextDiscoverTime))
+    {
+       if (BroadcastDiscover (cmRail))         /* urgent discovery required? */
+           cmRail->NextDiscoverTime = now + MSEC2TICKS (CM_URGENT_DISCOVER_INTERVAL);
+       else
+           cmRail->NextDiscoverTime = now + MSEC2TICKS (CM_PERIODIC_DISCOVER_INTERVAL);
+       
+       if (cmRail->Role == CM_ROLE_LEADER_CANDIDATE && AFTER (now, cmRail->DiscoverStartTick + MSEC2TICKS (CM_DISCOVER_TIMEOUT)))
+           RaiseTopLevel (cmRail);
+    }
+    
+    if (cmRail->NextHeartbeatTime && ! BEFORE (now, cmRail->NextHeartbeatTime))
+    {
+       CheckPosition (cmRail->Rail);
+       CheckPeerPulses (cmRail);
+       CheckBroadcast (cmRail);
+       CheckManager (cmRail);
+       
+       for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+       {
+           CheckOfflineReasons (cmRail, clvl);
+           ReduceGlobalMap (cmRail, clvl);
+           BroadcastGlobalMap (cmRail, clvl);
+       }
+       
+       SendHeartbeats (cmRail);
+       
+       /* Compute the next heartbeat time, but "drift" it towards the last
+        * periodic discovery time we saw from the whole machine leader */
+       cmRail->NextHeartbeatTime = now + MSEC2TICKS (CM_HEARTBEAT_INTERVAL);
+    }
+
+    if (cmRail->NextDiscoverTime && AFTER (cmRail->NextHeartbeatTime, cmRail->NextDiscoverTime))
+       cmRail->NextRunTime = cmRail->NextDiscoverTime;
+    else 
+       cmRail->NextRunTime = cmRail->NextHeartbeatTime;
+}
+
+#define CM_SVC_INDICATOR_OFFSET(CMRAIL,CLVL,IND,NODEID)     ( ( CMRAIL->Levels[CLVL].NumNodes * CM_GSTATUS_BITS ) \
+                                                               + ( CMRAIL->Levels[CLVL].NumNodes * IND ) \
+                                                               + ( NODEID - CMRAIL->Levels[CLVL].MinNodeId ) )
+int
+cm_svc_indicator_set (EP_RAIL *rail, int svc_indicator)
+{
+    CM_RAIL *cmRail = rail->ClusterRail;
+    unsigned long flags;
+    int           clvl;
+
+    EPRINTF2 (DBG_SVC,"cm_svc_indicator_set: rail %p ind %d\n", rail, svc_indicator);
+
+    if (svc_indicator < 0 || svc_indicator >= EP_SVC_NUM_INDICATORS)
+    {
+       EPRINTF1 (DBG_SVC,"cm_svc_indicator_set: service indicator %d not registered\n", svc_indicator);
+       return (-1);
+    }
+
+    if (rail->State == EP_RAIL_STATE_UNINITIALISED) 
+       return (-2);
+    
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)  {
+       statemap_setbits (cmRail->Levels[clvl].LocalMap, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, cmRail->NodeId), 1, 1); 
+       EPRINTF3 (DBG_SVC,"cm_svc_indicator_set: clvl %d nodeId %d offset %d\n", clvl, cmRail->NodeId, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, cmRail->NodeId));
+    }
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+
+    return (0);
+}
+
+int
+cm_svc_indicator_clear (EP_RAIL *rail, int svc_indicator)
+{
+    CM_RAIL *cmRail = rail->ClusterRail;
+    unsigned long flags;
+    int           clvl;
+
+    EPRINTF2 (DBG_SVC, "cm_svc_indicator_clear: rail %p ind %d\n", rail, svc_indicator);
+
+    if (svc_indicator < 0 || svc_indicator >= EP_SVC_NUM_INDICATORS)
+    {
+       EPRINTF1 (DBG_SVC, "cm_svc_indicator_clear: service indicator %d not registered\n", svc_indicator);
+       return (-1);
+    }
+
+    if (rail->State == EP_RAIL_STATE_UNINITIALISED) 
+       return (-2);
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)  {
+       statemap_setbits (cmRail->Levels[clvl].LocalMap, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, cmRail->NodeId), 0, 1); 
+       EPRINTF3 (DBG_SVC, "cm_svc_indicator_clear: clvl %d nodeId %d offset %d\n", clvl, cmRail->NodeId, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, cmRail->NodeId));
+    }
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+
+    return (0);
+}
+
+int
+cm_svc_indicator_is_set (EP_RAIL *rail, int svc_indicator, int nodeId)
+{
+    CM_RAIL *cmRail = rail->ClusterRail;
+    unsigned long flags;
+    int           clvl;
+    bitmap_t      bits;
+
+    EPRINTF4 (DBG_SVC, "cm_svc_indicator_is_set: rail %p ind %d nodeId %d (me=%d)\n", rail, svc_indicator, nodeId, cmRail->NodeId);
+
+    if (svc_indicator < 0 || svc_indicator > EP_SVC_NUM_INDICATORS)
+    {
+       EPRINTF1 (DBG_SVC, "cm_svc_indicator_is_set: service indicator %d not registered\n", svc_indicator);
+       return (0);
+    }
+
+    if (rail->State == EP_RAIL_STATE_UNINITIALISED) 
+       return (0);
+    
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+       if (nodeId >= cmRail->Levels[clvl].MinNodeId && nodeId < (cmRail->Levels[clvl].MinNodeId + cmRail->Levels[clvl].NumNodes))
+           break;
+
+    if ( clvl == cmRail->NumLevels) { 
+       EPRINTF1 (DBG_SVC, "cm_svc_indicator_is_set: node out of range %d \n", nodeId); 
+       spin_unlock_irqrestore (&cmRail->Lock, flags);
+       return (0);
+    }
+
+    if ( cmRail->NodeId == nodeId ) 
+       bits = statemap_getbits (cmRail->Levels[clvl].LocalMap, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, nodeId), 1);
+    else
+       bits = statemap_getbits (cmRail->Levels[clvl].GlobalMap, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, nodeId), 1);
+
+    EPRINTF4 (DBG_SVC, "cm_svc_indicator_is_set: clvl %d nodeId %d offset %d %x\n", clvl, nodeId, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, nodeId), bits);
+
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+
+    return  ( (bits == 0) ? (0) : (1) );
+}
+
+int
+cm_svc_indicator_bitmap (EP_RAIL *rail, int svc_indicator, bitmap_t * bitmap, int low, int nnodes)
+{
+    /* or in the bit map */  
+    CM_RAIL      *cmRail = rail->ClusterRail;
+    int           nodeId, clvl;
+    bitmap_t      bits;
+    unsigned long flags;
+    int           clip_out_low, clip_out_high;
+    int           curr_low, curr_high;
+    int           check_low, check_high;
+
+    EPRINTF4 (DBG_SVC, "cm_svc_indicator_bitmap: rail %p ind %d low %d high %d\n", rail, svc_indicator, low, (low + nnodes));
+
+    if (svc_indicator < 0 || svc_indicator >= EP_SVC_NUM_INDICATORS)
+    {
+       EPRINTF1 (DBG_SVC, "cm_svc_indicator_bitmap: service indicator %d not registered\n", svc_indicator);
+       return (-1);
+    }
+
+    if (rail->State != EP_RAIL_STATE_RUNNING) 
+       return (-2);
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    
+    clip_out_low = clip_out_high = -1; /* all in */
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++) {
+
+       /* curr_high/low is the range of the current lvl */
+       curr_low  = cmRail->Levels[clvl].MinNodeId;
+       curr_high = cmRail->Levels[clvl].MinNodeId + cmRail->Levels[clvl].NumNodes;
+
+       /* find out how much of low high is in this range and only check that part */
+       check_low  = ( low  < curr_low)  ? curr_low  : low; 
+       check_high = ( (low + nnodes) > curr_high) ? curr_high : (low + nnodes);
+
+       EPRINTF6 (DBG_SVC, "cm_svc_indicator_bitmap: curr(%d,%d) check(%d,%d) clip(%d,%d)\n", curr_low, curr_high, check_low, check_high, clip_out_low, clip_out_high);
+
+       for(nodeId = check_low; nodeId < check_high; nodeId++) {
+
+           if (  (clip_out_low <= nodeId) && (nodeId <= clip_out_high))
+               nodeId = clip_out_high; /* step over the cliped out section */
+           else {
+
+               if ( cmRail->NodeId == nodeId ) 
+                   bits = statemap_getbits (cmRail->Levels[clvl].LocalMap, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, nodeId), 1);
+               else
+                   bits = statemap_getbits (cmRail->Levels[clvl].GlobalMap, CM_SVC_INDICATOR_OFFSET (cmRail, clvl, svc_indicator, nodeId), 1);
+               
+               if ( bits ) {
+                   EPRINTF2 (DBG_SVC, "cm_svc_indicator_bitmap: its set nodeId %d (clvl %d)\n", nodeId, clvl);
+                   BT_SET ( bitmap , nodeId - low );
+               }
+           }
+       }
+
+       /* widen the clip out range */
+       clip_out_low  = curr_low;
+       clip_out_high = curr_high -1; 
+    }
+
+    spin_unlock_irqrestore (&cmRail->Lock, flags);      
+
+    return (0);
+}
+
+#if defined(PER_CPU_TIMEOUT)
+static void
+cm_percpu_timeout (void *arg)
+{
+    CM_RAIL          *cmRail = (CM_RAIL *) arg;
+    CM_TIMEOUT_DATA *hbd     = &cmRail->HeartbeatTimeoutsData[current_cpu()];
+    long             now     = lbolt;
+    unsigned         delay   = now - hbd->ScheduledAt;
+    unsigned long    flags;
+
+    if (delay > hbd->WorstDelay)
+       hbd->WorstDelay = delay;
+    if (hbd->BestDelay == 0 || delay < hbd->BestDelay)
+       hbd->BestDelay = delay;
+
+    if (cmRail->HeartbeatTimeoutsShouldStop)
+    {
+       spin_lock_irqsave (&cmRail->Lock, flags);
+       cmRail->HeartbeatTimeoutsStopped |= (1 << current_cpu());
+       kcondvar_wakeupall (&cmRail->HeartbeatTimeoutsWait, &cmRail->Lock);
+       spin_unlock_irqrestore (&cmRail->Lock, flags);
+       return;
+    }
+
+    if (cmRail->NextRunTime == 0 || AFTER (cmRail->NextRunTime, lbolt))
+       hbd->EarlyCount++;
+    else if (cmRail->HeartbeatTimeoutRunning)
+       hbd->MissedCount++;
+    else
+    {
+       local_irq_save (flags);
+       
+       if (! spin_trylock (&cmRail->HeartbeatTimeoutsLock))
+           hbd->WastedCount++;
+       else
+       {
+           cmRail->HeartbeatTimeoutRunning = 1;
+           hbd->WorkCount++;
+
+           spin_lock (&cmRail->Lock);
+
+           if ((delay = (lbolt - cmRail->NextRunTime)) > hbd->WorstHearbeatDelay)
+               hbd->WorstHearbeatDelay = delay;
+           if ((delay = (lbolt - now) > hbd->WorstLockDelay))
+               hbd->WorstLockDelay = delay;
+
+           DoHeartbeatWork (cmRail);
+
+           spin_unlock (&cmRail->Lock);
+           spin_unlock (&cmRail->HeartbeatTimeoutsLock);
+
+           cmRail->HeartbeatTimeoutRunning = 0;
+       }
+       local_irq_restore (flags);
+    }
+
+    hbd->ScheduledAt = lbolt + MSEC2TICKS (CM_PERCPU_TIMEOUT_INTERVAL);
+    timeout_cpu (cm_percpu_timeout, cmRail, MSECS2TICKS (CM_PERCPU_TIMEOUT_INTERVAL), CALLOUT_TYPE|CALLOUT_NOMALLOC);
+}
+
+static void
+StartPerCpuTimeouts (CM_RAIL *cmRail)
+{
+    register int c;
+
+    spin_lock_init (&cmRail->HeartbeatTimeoutsLock);
+
+    KMEM_ZALLOC (cmRail->HeartbeatTimeoutsData, CM_TIMEOUT_DATA *, ncpus * sizeof (CM_TIMEOUT_DATA), 1);
+
+    for (c = 0; c < cpus_in_box; c++)
+    {
+       if (cpu_to_processor (c))
+       {       
+           if (current_cpu() != c)
+           {
+               thread_bind (current_thread(), cpu_to_processor(c));
+               mpsleep (current_thread(), 0, "StartPerCpuTimeouts", 1, NULL, 0);
+
+               if (current_cpu() != c)
+                   panic ("ep: StartPerCpuTimeouts - failed to switch cpu\n");
+           }
+           
+           cmRail->HeartbeatTimeoutsStarted |= (1 << c);
+           cmRail->HeartbeatTimeoutsData[c].ScheduledAt = lbolt + c;
+
+           timeout_cpu (cm_percpu_timeout, cmRail, c, CALLOUT_TYPE|CALLOUT_NOMALLOC);
+       }
+    }
+
+    thread_bind(current_thread(), NULL);
+}
+
+static void
+StopPerCpuTimeouts (CM_RAIL *cmRail)
+{
+    register int c;
+    unsigned long flags;
+
+    cmRail->HeartbeatTimeoutsShouldStop = 1;
+
+    for (c = 0; c < cpus_in_box; c++)
+    {
+       if (cmRail->HeartbeatTimeoutsStarted & (1 << c))
+       {
+           printk ("%s: stopping cpu_timeout on cpu %d\n", cmRail->Rail->Name, c);
+
+           if (untimeout_cpu (cm_percpu_timeout, cmRail, c, CALLOUT_TYPE|CALLOUT_NOMALLOC, NULL))
+               cmRail->HeartbeatTimeoutsStopped |= (1 << c);
+       }
+    }
+    thread_bind(current_thread(), NULL);
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    while (cmRail->HeartbeatTimeoutsStopped != cmRail->HeartbeatTimeoutsStarted)
+       kcondvar_wait (&cmRail->HeartbeatTimeoutsWait, &cmRail->Lock, &flags);
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+
+    cmRail->HeartbeatTimeoutsStarted    = 0;
+    cmRail->HeartbeatTimeoutsStopped    = 0;
+    cmRail->HeartbeatTimeoutsShouldStop = 0;
+
+    KMEM_FREE (cmRail->HeartbeatTimeoutsData, ncpus * sizeof (CM_TIMEOUT_DATA));
+
+    spin_lock_destroy (&cmRail->HeartbeatTimeoutsLock);
+}
+
+#else
+
+static void
+cm_heartbeat_timer (unsigned long arg)
+{
+    CM_RAIL *cmRail = (CM_RAIL *) arg;
+    unsigned long flags;
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+
+    ASSERT (cmRail->Rail->State == EP_RAIL_STATE_RUNNING);
+
+    DoHeartbeatWork (cmRail);
+    
+    __Schedule_Timer (cmRail, cmRail->NextRunTime);
+
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+}
+
+#endif /* defined(PER_CPU_TIMEOUT) */
+
+
+
+void
+DisplayRailDo (DisplayInfo *di, EP_RAIL *rail)
+{
+    CM_RAIL *cmRail = rail->ClusterRail;
+    unsigned long flags;
+    int  i, j;
+
+    if (rail->State != EP_RAIL_STATE_RUNNING)
+       return;
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+
+    (di->func)(di->arg, "NodeId=%d NodeLevel=%d NumLevels=%d NumNodes=%d\n", 
+           cmRail->NodeId, cmRail->TopLevel, cmRail->NumLevels, cmRail->Rail->Position.pos_nodes);
+    
+    (di->func)(di->arg, "[");
+
+    for (i = 0; i < cmRail->NumLevels; i++)
+    {
+       if (i > 0)
+           (di->func)(di->arg, ",");
+       
+       if (i < cmRail->TopLevel)
+       {
+           (di->func)(di->arg, "L ");
+         
+           for (j = 0; j < cmRail->Levels[i].NumSegs; j++)
+               switch (cmRail->Levels[i].Sgmts[j].State)
+               {
+               case CM_SGMT_PRESENT: (di->func)(di->arg, "p%-4d", cmRail->Levels[i].Sgmts[j].NodeId); break;
+               case CM_SGMT_WAITING: (di->func)(di->arg, "w%4s", ""); break;
+               case CM_SGMT_COMING:  (di->func)(di->arg, "c%4s", ""); break;
+               case CM_SGMT_ABSENT:  (di->func)(di->arg, ".%4s", ""); break;
+               default:              (di->func)(di->arg, "?%4s", ""); break;
+               }
+       }
+       else
+           switch (cmRail->Role)
+           {
+           case CM_ROLE_LEADER_CANDIDATE:      
+               (di->func)(di->arg,"l "); 
+               for (j = 0; j < cmRail->Levels[i].NumSegs; j++)
+                   (di->func)(di->arg,"     ");
+               break;
+         
+           case CM_ROLE_SUBORDINATE:       
+               switch (cmRail->Levels[i].Sgmts[0].State)
+               {
+               case CM_SGMT_PRESENT: (di->func)(di->arg, "p%-4d", cmRail->Levels[i].Sgmts[0].NodeId); break;
+               case CM_SGMT_WAITING: (di->func)(di->arg, "w%4s", ""); break;
+               case CM_SGMT_COMING:  (di->func)(di->arg, "c%4s", ""); break;
+               case CM_SGMT_ABSENT:  (di->func)(di->arg, ".%4s", ""); break;
+               default:              (di->func)(di->arg, "?%4s", ""); break;
+               }
+               for (j = 1; j < cmRail->Levels[i].NumSegs; j++)
+                   (di->func)(di->arg, "     ");
+               break;
+         
+           default:
+               (di->func)(di->arg, "####");
+               break;
+           }
+    }
+    (di->func)(di->arg, "]\n");
+
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+}
+
+void
+DisplayRail (EP_RAIL *rail) 
+{
+    if (rail->State == EP_RAIL_STATE_RUNNING)
+       DisplayRailDo (&di_ep_debug, rail);
+}
+
+void
+DisplayStatus (EP_RAIL *rail)
+{
+    if (rail->State == EP_RAIL_STATE_RUNNING)
+    {
+       CM_RAIL *cmRail = rail->ClusterRail;
+       unsigned long flags;
+       
+       spin_lock_irqsave (&cmRail->Lock, flags);
+       
+       DisplayNodeMaps (&di_ep_debug, cmRail);
+       
+       spin_unlock_irqrestore (&cmRail->Lock, flags);
+    }
+}
+
+void
+DisplaySegs (EP_RAIL *rail)
+{
+    if (rail->State == EP_RAIL_STATE_RUNNING)
+    {
+       CM_RAIL *cmRail = rail->ClusterRail;
+       unsigned long flags;
+       
+       spin_lock_irqsave (&cmRail->Lock, flags);
+       
+       DisplayNodeSgmts (&di_ep_debug, cmRail);
+       
+       spin_unlock_irqrestore (&cmRail->Lock, flags);
+    }
+}
+
+static void
+LoadBroadcastRoute (CM_RAIL *cmRail, int lvl, int sidx)
+{
+    EP_RAIL *rail  = cmRail->Rail;
+    int      nsegs = cmRail->Levels[0].NumSegs;
+    int      vp    = EP_VP_BCAST(lvl, sidx);
+    int      nodes = 1;
+    int      baseNode;
+    int      i;
+
+    ASSERT (lvl > 0 && lvl <= cmRail->NumLevels);
+    ASSERT (sidx == 0 || lvl < cmRail->NumLevels);
+
+    ASSERT (vp >= EP_VP_BCAST_BASE && vp < EP_VP_BCAST_BASE + EP_VP_BCAST_COUNT);
+
+    for (i = 1; i <= lvl; i++)
+    {
+       nodes *= nsegs;
+       nsegs = (i == cmRail->NumLevels) ? 1 : cmRail->Levels[i].NumSegs;
+    }
+
+    baseNode = ((cmRail->NodeId / (nodes * nsegs)) * nsegs + sidx) * nodes;
+
+    CPRINTF5 (2, "%s: broadcast vp lvl %d sidx %d [%d,%d]\n", 
+             cmRail->Rail->Name, lvl, sidx, baseNode, baseNode + nodes - 1);
+    
+    rail->Operations.LoadSystemRoute (rail, vp, baseNode, baseNode + nodes - 1);
+}
+
+static void
+LoadRouteTable (CM_RAIL *cmRail)
+{
+    EP_RAIL *rail = cmRail->Rail;
+    int      i, j;
+   
+   if (cmRail->NumNodes > EP_MAX_NODES)
+   {
+       printk ("More nodes (%d) than point-to-point virtual process table entries (%d)\n", cmRail->NumNodes, EP_MAX_NODES);
+       panic ("LoadRouteTable\n");
+   }
+
+   for (i = 0; i < cmRail->NumNodes; i++)
+       rail->Operations.LoadSystemRoute (rail, EP_VP_NODE(i), i, i);
+
+   /* Generate broadcast routes for subtrees */
+   for (i = 1; i < cmRail->NumLevels; i++)
+      for (j = 0; j < cmRail->Levels[i].NumSegs; j++)
+         LoadBroadcastRoute (cmRail, i, j);
+
+   /* Generate broadcast route for whole machine */
+   LoadBroadcastRoute (cmRail, cmRail->NumLevels, 0);
+
+   /* Finally invalidate all the data routes */
+   for (i = 0; i < cmRail->NumNodes; i++)
+       rail->Operations.UnloadNodeRoute (cmRail->Rail, i);
+}
+
+void
+cm_node_disconnected (EP_RAIL *rail, unsigned nodeId)
+{
+    CM_RAIL *cmRail = rail->ClusterRail;
+    int      base, lstat, lgstat;
+    int             clvl, subClMin, subClMax;
+    int      thisClId, myClId;
+    unsigned long flags;
+
+    ASSERT (nodeId != cmRail->NodeId);
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+       if (nodeId >= cmRail->Levels[clvl].MinNodeId && nodeId < (cmRail->Levels[clvl].MinNodeId + cmRail->Levels[clvl].NumNodes))
+           break;
+
+    myClId   = ClusterIds (cmRail, clvl, &subClMin, &subClMax);
+    thisClId = nodeId - cmRail->Levels[clvl].MinNodeId;
+    base     = thisClId * CM_GSTATUS_BITS;
+    lstat    = statemap_getbits (cmRail->Levels[clvl].LocalMap,  base, CM_GSTATUS_BITS);
+    lgstat   = statemap_getbits (cmRail->Levels[clvl].LastGlobalMap, base, CM_GSTATUS_BITS) & CM_GSTATUS_STATUS_MASK;
+
+    ASSERT ((lstat & CM_GSTATUS_ACK_MASK) == CM_GSTATUS_MAY_RUN);
+
+    CPRINTF7 (2, "%s: cm_node_disconnected: Node %d: clvl %d, lgstat %s, gstat %s, lstat %s -> %sMAY_START\n",
+             cmRail->Rail->Name, nodeId, clvl,
+             GlobalStatusString (cmRail->Levels[clvl].LastGlobalMap, thisClId),
+             GlobalStatusString (cmRail->Levels[clvl].GlobalMap, thisClId),
+             GlobalStatusString (cmRail->Levels[clvl].LocalMap, thisClId),
+             ((lgstat != CM_GSTATUS_CLOSING) && (lstat & CM_GSTATUS_RESTART)) ? "RESTART|" : "");
+    
+    switch (lgstat)
+    {
+    case CM_GSTATUS_CLOSING:
+       /* delayed ack of closing - set MAY_START and clear RESTART */
+       statemap_setbits (cmRail->Levels[clvl].LocalMap, base, CM_GSTATUS_MAY_START, CM_GSTATUS_BITS);
+       break;
+    case CM_GSTATUS_STARTING:
+    case CM_GSTATUS_RUNNING:
+       IASSERT (! cmRail->Levels[clvl].Online || lstat & CM_GSTATUS_RESTART);
+       break;
+    case CM_GSTATUS_ABSENT:
+       IASSERT (lstat & CM_GSTATUS_RESTART);
+    }
+
+    cmRail->Levels[clvl].Connected--;
+
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+}
+
+void
+cm_restart_node (EP_RAIL *rail, unsigned nodeId)
+{
+    CM_RAIL *cmRail = rail->ClusterRail;
+    int      base, lstat, lgstat;
+    int             clvl, subClMin, subClMax;
+    int      thisClId, myClId;
+    unsigned long flags;
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    if (nodeId == rail->Position.pos_nodeid)
+    {
+       for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+           RestartComms (cmRail, clvl);
+    }
+    else
+    {
+       for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+           if (nodeId >= cmRail->Levels[clvl].MinNodeId && nodeId < (cmRail->Levels[clvl].MinNodeId + cmRail->Levels[clvl].NumNodes))
+               break;
+       
+       myClId   = ClusterIds (cmRail, clvl, &subClMin, &subClMax);
+       thisClId = nodeId - cmRail->Levels[clvl].MinNodeId;
+       base     = thisClId * CM_GSTATUS_BITS;
+       lstat    = statemap_getbits (cmRail->Levels[clvl].LocalMap,  base, CM_GSTATUS_BITS);
+       lgstat   = statemap_getbits (cmRail->Levels[clvl].LastGlobalMap,  base, CM_GSTATUS_BITS) & CM_GSTATUS_STATUS_MASK;
+
+       CPRINTF6 (2, "%s: cm_restart_node: Node %d: clvl %d, lgstat %s, gstat %s, lstat %s\n",
+                 cmRail->Rail->Name, nodeId, clvl,
+                 GlobalStatusString (cmRail->Levels[clvl].LastGlobalMap, thisClId),
+                 GlobalStatusString (cmRail->Levels[clvl].GlobalMap, thisClId),
+                 GlobalStatusString (cmRail->Levels[clvl].LocalMap, thisClId));
+       
+       if (lgstat != CM_GSTATUS_CLOSING)
+           statemap_setbits (cmRail->Levels[clvl].LocalMap, base, lstat | CM_GSTATUS_RESTART, CM_GSTATUS_BITS);
+    }
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+}
+
+void
+cm_force_offline (EP_RAIL *rail, int offline, unsigned int reason)
+{
+    CM_RAIL *cmRail = rail->ClusterRail;
+    unsigned long flags;
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+    if (offline)
+       cmRail->OfflineReasons |= reason;
+    else
+       cmRail->OfflineReasons &= ~reason;
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+}
+
+static void
+cm_remove_rail (EP_SUBSYS *subsys, EP_SYS *epsys, EP_RAIL *rail)
+{
+    CM_SUBSYS  *sys    = (CM_SUBSYS *) subsys;
+    CM_RAIL    *cmRail = sys->Rails[rail->Number];
+    int i, lvl, clvl;
+
+    cm_procfs_rail_fini (cmRail);
+
+    sys->Rails[rail->Number] = NULL;
+    rail->ClusterRail        = NULL;
+
+#if defined(PER_CPU_TIMEOUT)
+    StopPerCpuTimeouts (cmRail);
+#else
+    del_timer_sync (&cmRail->HeartbeatTimer);
+#endif
+    cmRail->NextRunTime      = 0;
+    cmRail->NextDiscoverTime = 0;
+    cmRail->NextHeartbeatTime = 0;
+    
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+    {
+       for (lvl = 0; lvl <= clvl; lvl++)
+       {
+           CM_LEVEL *level = &cmRail->Levels[lvl];
+           
+           statemap_destroy (level->SubordinateMap[clvl]);
+           
+           for (i = 0; i < level->NumSegs; i++)
+           {
+               statemap_destroy (level->Sgmts[i].Maps[clvl].CurrentInputMap);
+               statemap_destroy (level->Sgmts[i].Maps[clvl].InputMap);
+               statemap_destroy (level->Sgmts[i].Maps[clvl].OutputMap);
+           }
+       }
+       
+       cmRail->Levels[clvl].Online = 0;
+       
+       statemap_destroy (cmRail->Levels[clvl].TmpMap);
+       statemap_destroy (cmRail->Levels[clvl].GlobalMap);
+       statemap_destroy (cmRail->Levels[clvl].LastGlobalMap);
+       statemap_destroy (cmRail->Levels[clvl].SubTreeMap);
+       statemap_destroy (cmRail->Levels[clvl].LocalMap);
+    }
+    
+    spin_lock_destroy (&cmRail->Lock);
+
+    ep_free_inputq (cmRail->Rail, cmRail->PolledQueue);
+    ep_free_inputq (cmRail->Rail, cmRail->IntrQueue);
+    ep_free_outputq (cmRail->Rail, cmRail->MsgQueue);
+
+    KMEM_FREE (cmRail, sizeof (CM_RAIL));
+}
+
+static int
+cm_add_rail (EP_SUBSYS *subsys, EP_SYS *epsys, EP_RAIL *rail)
+{
+    CM_SUBSYS     *sys = (CM_SUBSYS *) subsys;
+    ELAN_POSITION *pos = &rail->Position;
+    CM_RAIL       *cmRail;
+    int lvl, n, nn, clvl, span, i;
+    unsigned long flags;
+
+    KMEM_ZALLOC (cmRail, CM_RAIL *, sizeof (CM_RAIL), 1);
+
+    if (cmRail == NULL)
+       return (ENOMEM);
+    
+    cmRail->Rail     = rail;
+    cmRail->NodeId   = pos->pos_nodeid;
+    cmRail->NumNodes = pos->pos_nodes;
+
+    spin_lock_init (&cmRail->Lock);
+
+    if ((cmRail->IntrQueue   = ep_alloc_inputq (rail, EP_SYSTEMQ_INTR,   sizeof (CM_MSG), CM_INPUTQ_ENTRIES, IntrQueueCallback, cmRail)) == NULL ||
+       (cmRail->PolledQueue = ep_alloc_inputq (rail, EP_SYSTEMQ_POLLED, sizeof (CM_MSG), CM_INPUTQ_ENTRIES, NULL, 0)) == NULL ||
+       (cmRail->MsgQueue    = ep_alloc_outputq (rail, sizeof (CM_MSG), CM_NUM_MSG_BUFFERS)) == NULL)
+    {
+       goto failed;
+    }
+
+    /* point to first "spare" message buffer */
+    cmRail->NextSpareMsg = 0;
+
+    /* Compute the branching ratios from the switcy arity */
+    for (lvl = 0; lvl < CM_MAX_LEVELS; lvl++)
+       BranchingRatios[lvl] = (lvl < pos->pos_levels) ? pos->pos_arity[pos->pos_levels - lvl - 1] : 4;
+    
+    /* now determine the number of levels of hierachy we have */
+    /* and how many nodes per level there are */
+    for (lvl = 0, nn = 1, n = pos->pos_nodes; 
+        n > 1; 
+        nn *= BranchingRatios[lvl], n = n / BranchingRatios[lvl], lvl++)
+    {
+       int       nSegs = (n > BranchingRatios[lvl]) ? BranchingRatios[lvl] : n;
+       int       nNodes = nn * nSegs;
+       CM_LEVEL *level = &cmRail->Levels[lvl];
+
+       for (clvl = 0, span = pos->pos_arity[pos->pos_levels - clvl - 1]; 
+            span < nNodes && clvl < pos->pos_levels - 1;
+            clvl++, span *= pos->pos_arity[pos->pos_levels - clvl - 1])
+           ;
+       
+       level->SwitchLevel = clvl;
+       level->MinNodeId = (pos->pos_nodeid / nNodes) * nNodes;
+       level->NumNodes = nNodes;
+       level->NumSegs = nSegs;
+    }
+    
+    cmRail->NumLevels      = lvl;
+    cmRail->BroadcastLevel = lvl-1;
+
+    CPRINTF4 (2, "%s: NodeId=%d NumNodes=%d NumLevels=%d\n", 
+             rail->Name, pos->pos_nodeid, pos->pos_nodes, cmRail->NumLevels);
+
+    LoadRouteTable (cmRail);
+    
+    /* Init SGMT constants */
+    for (lvl = 0; lvl < cmRail->NumLevels; lvl++)
+    {
+       CM_LEVEL *level = &cmRail->Levels[lvl];
+
+       level->MySgmt = SegmentNo (cmRail, cmRail->NodeId, lvl);
+       
+       for (i = 0; i < CM_SGMTS_PER_LEVEL; i++)
+       {
+           CM_SGMT *sgmt = &level->Sgmts[i];
+         
+           sgmt->MsgNumber = lvl * CM_SGMTS_PER_LEVEL + i;
+           sgmt->Level = lvl;
+           sgmt->Sgmt = i;
+       }
+    }
+
+    /* Init maps for each cluster level */
+    for (clvl = 0; clvl < cmRail->NumLevels; clvl++)
+    {
+       int nNodes = cmRail->Levels[clvl].NumNodes;
+       int mapBits = (nNodes * CM_GSTATUS_BITS) + (nNodes * EP_SVC_NUM_INDICATORS);
+       int clmin;
+       int clmax;
+       int clid = ClusterIds (cmRail, clvl, &clmin, &clmax);
+
+       for (lvl = 0; lvl <= clvl; lvl++)
+       {
+           CM_LEVEL *level = &cmRail->Levels[lvl];
+
+           level->SubordinateMap[clvl] = statemap_create (mapBits);
+
+           for (i = 0; i < level->NumSegs; i++)
+           {
+               level->Sgmts[i].Maps[clvl].CurrentInputMap = statemap_create (mapBits);
+               level->Sgmts[i].Maps[clvl].InputMap        = statemap_create (mapBits);
+               level->Sgmts[i].Maps[clvl].OutputMap       = statemap_create (mapBits);
+           }
+       }
+       
+       cmRail->Levels[clvl].Online = 0;
+
+       cmRail->Levels[clvl].TmpMap        = statemap_create (mapBits);
+       cmRail->Levels[clvl].GlobalMap     = statemap_create (mapBits);
+       cmRail->Levels[clvl].LastGlobalMap = statemap_create (mapBits);
+       cmRail->Levels[clvl].SubTreeMap    = statemap_create (mapBits);
+       cmRail->Levels[clvl].LocalMap      = statemap_create (mapBits);
+
+       /* Flag everyone outside my next lower cluster as sensed offline... */
+       for (i = 0; i < clmin; i++)
+           statemap_setbits (cmRail->Levels[clvl].LocalMap, i * CM_GSTATUS_BITS, CM_GSTATUS_MAY_START, CM_GSTATUS_BITS);
+       
+       for (i = clmax + 1; i < nNodes; i++)
+           statemap_setbits (cmRail->Levels[clvl].LocalMap, i * CM_GSTATUS_BITS, CM_GSTATUS_MAY_START, CM_GSTATUS_BITS);
+       
+       /* ...and set my own state */
+       statemap_setbits (cmRail->Levels[clvl].LocalMap, clid * CM_GSTATUS_BITS,
+                         CM_GSTATUS_CLOSING | CM_GSTATUS_MAY_START, CM_GSTATUS_BITS);
+    }
+    
+    /* compute parameter hash to add to messages */
+    cmRail->ParamHash = EP_PROTOCOL_VERSION;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_PERIODIC_DISCOVER_INTERVAL;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_URGENT_DISCOVER_INTERVAL;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_HEARTBEAT_INTERVAL;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_P2P_DMA_RETRIES;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_P2P_MSG_RETRIES;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_BCAST_MSG_RETRIES;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_TIMER_SCHEDULE_TIMEOUT;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_HEARTBEAT_TIMEOUT;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_DISCOVER_TIMEOUT;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + BT_NBIPUL;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + CM_GSTATUS_BITS;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + EP_SVC_NUM_INDICATORS;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + cmRail->NumLevels;
+    cmRail->ParamHash = cmRail->ParamHash * 127 + cmRail->NumNodes;
+    for (i = 0; i < cmRail->NumLevels; i++)
+       cmRail->ParamHash = cmRail->ParamHash * 127 + BranchingRatios[i];
+    
+#if defined(PER_CPU_TIMEOUT)
+    StartPerCpuTimeouts (cmRail);
+#endif
+
+    spin_lock_irqsave (&cmRail->Lock, flags);
+
+#if !defined(PER_CPU_TIMEOUT)
+    /* Initialise the timer, but don't add it yet, since
+     * __Schedule_Heartbeat() will do this. */
+
+    init_timer (&cmRail->HeartbeatTimer);
+    
+    cmRail->HeartbeatTimer.function = cm_heartbeat_timer;
+    cmRail->HeartbeatTimer.data     = (unsigned long) cmRail;
+    cmRail->HeartbeatTimer.expires  = lbolt + hz;
+#endif
+
+    /* start sending heartbeats */
+    __Schedule_Heartbeat (cmRail);
+
+    /* start discovering who else is out there */
+    LowerTopLevel (cmRail, 0);
+
+    /* connect to myself straight away - I know I'm here */
+    ep_connect_node (rail, cmRail->NodeId);
+    
+    /* add to all rails */
+    sys->Rails[rail->Number] = cmRail;
+    rail->ClusterRail = (void *) cmRail;
+
+    spin_unlock_irqrestore (&cmRail->Lock, flags);
+
+    /* Enable the input queues */
+    ep_enable_inputq (rail, cmRail->PolledQueue);
+    ep_enable_inputq (rail, cmRail->IntrQueue);
+
+    /* Create the procfs entries */
+    cm_procfs_rail_init (cmRail);
+
+    return 0;
+
+ failed:
+    cm_remove_rail (subsys, epsys, rail);
+    return -ENOMEM;
+}
+
+static void
+cm_fini (EP_SUBSYS *subsys, EP_SYS *epsys)
+{
+    CM_SUBSYS *sys = (CM_SUBSYS *) subsys;
+
+    cm_procfs_fini(sys);
+    
+    KMEM_FREE (sys, sizeof (CM_SUBSYS));
+}
+
+int
+cm_init (EP_SYS *sys)
+{
+    CM_SUBSYS *subsys;
+
+    KMEM_ZALLOC (subsys, CM_SUBSYS *, sizeof (CM_SUBSYS), 1);
+
+    if (subsys == NULL)
+       return (ENOMEM);
+
+    subsys->Subsys.Sys        = sys;
+    subsys->Subsys.Name              = "cm";
+    subsys->Subsys.Destroy    = cm_fini;
+    subsys->Subsys.AddRail    = cm_add_rail;
+    subsys->Subsys.RemoveRail = cm_remove_rail;
+
+    ep_subsys_add (sys, &subsys->Subsys);
+
+    cm_procfs_init (subsys);
+
+    /*
+     * Initialise the machineid if it wasn't specified by
+     * the modules.conf file - otherwise truncate it to 
+     * 16 bits.
+     */
+    if (MachineId != -1)
+       MachineId = (uint16_t) MachineId;
+    else
+    {
+#if defined(LINUX_ALPHA)
+       MachineId = (uint16_t)((5 << 12) | HZ);
+#elif defined(LINUX_SPARC)
+       MachineId = (uint16_t)((4 << 12) | HZ);
+#elif defined(LINUX_I386)
+       MachineId = (uint16_t)((3 << 12) | HZ);
+#elif defined( LINUX_IA64)
+       MachineId = (uint16_t)((2 << 12) | HZ);
+#elif defined(LINUX_X86_64)
+       MachineId = (uint16_t)((1 << 12) | HZ);
+#else
+       MachineId = (uint16_t)((0 << 12) | HZ);
+#endif
+    }
+
+    return (0);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/cm.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/cm.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/cm.h      2005-06-01 23:12:54.633433936 -0400
@@ -0,0 +1,412 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_CM_H
+#define __ELAN_CM_H
+
+#ident "@(#)$Id: cm.h,v 1.14.2.1 2004/11/12 10:54:50 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/cm.h,v $*/
+
+#include <elan/statemap.h>
+
+#if defined(DIGITAL_UNIX)
+/*
+ * On Tru64 - SMP doesn't mean Symmetric - cpu 0 is a master cpu and is responsible
+ * for handling all PCI interrupts and "funneled" operations.  When a kernel thread
+ * is made runnable, the scheduler will choose which cpu it will run on at that time,
+ * and will only execute a higher priority thread from another cpu's run queue when 
+ * it becomes totally idle (apparently also including user processes).  Also the 
+ * assert_wait_mesg_timo function uses a per-cpu timeout - these can only get executed
+ * at "preemptable" places - so again have no guarantee on when they will execute if
+ * they happen to be queued on a "hogged" cpu. The combination of these mean that the Tru64
+ * is incapable of scheduling a high priority kernel  thread within a deterministic time
+ * of when it should have become runnable - wonderfull.
+ *
+ * Hence the solution Compaq have proposed it to schedule a timeout onto all of the
+ * cpu's timeouts lists at the maximum frequency that we could want to execute code,
+ * then to handle the scheduling of work between these ourselves.  With a bit of luck
+ * ..... at least one cpu will be sufficiently unloaded to allow us to get a chance
+ * to do our important work.
+ *
+ * However ..... this still is not reliable, since timeouts under Tru64 are still 
+ * only run when the currently running kernel thread "co-operates" by calling one
+ * of a number of functions which is permitted to run the "lwc"s AND is not holding
+ * any spinlocks AND is running ai IPL 0.   However Compaq are unable to provide
+ * any upper limit on the time between the "lwc"'s being run and so it is possible
+ * for all 4 cpus to not run them for an unbounded time.
+ *
+ * The solution proposed is to use the RM_TEMP_BACKDOOR hook which was added to 
+ * hardclock() to "solve" this problem for Memory Channel.  However, since it
+ * is called within the clock interrupt it is not permissible to aquire any
+ * spinlocks, nor to run for "too long".  This means that it is not possible to
+ * call the heartbeat algorithm from this hook.  
+ *
+ * Our solution to these limitations is to use the hook to cause an elan interrupt 
+ * to be delivered, by issueing a mis-aligned SetEvent command - this causes the device 
+ * to trap and ep_cprocTrap() can then run the heartbeat code.  However there is a lock 
+ * order violation between the elan_dev::IntrLock and ep_dev::Lock, so we have to 
+ * use a trylock and if we fail, then hope that when the interrupt is delievered again
+ * some time later we will succeed.
+ *
+ * However this only works if the kernel is able to respond to the Elan interrupt,
+ * so we panic inside the RM_TEMP_BACKDOOR hook if the SetEvent's interrupt has
+ * not been taken for more than an CM_TIMER_SCHEDULE_TIMEOUT interval.
+ *
+ * In fact this is exactly the mechanism that other operating systems use to
+ * execute timeouts, since the hardclock interrupt posts a low priority 
+ * "soft interrupt" which "pre-eempts" the currently running thread and then
+ * executes the timeouts.To block timeouts you use splsoftclock() the same as 
+ * in Tru64.
+ */
+#define PER_CPU_TIMEOUT                        TRUE
+#endif
+
+
+#define CM_SGMTS_PER_LEVEL             8                       /* maximum nodes in each segment */
+#define CM_MAX_LEVELS                  6                       /* maximum depth of tree */
+
+/* message buffers/dmas/events etc */
+#define CM_NUM_NODE_MSG_BUFFERS                (CM_MAX_LEVELS * CM_SGMTS_PER_LEVEL) /* subordinates and leader */
+#define CM_NUM_SPARE_MSG_BUFFERS       8                       /* spare msg buffers for non-connected nodes */
+#define CM_NUM_MSG_BUFFERS             (CM_NUM_NODE_MSG_BUFFERS + CM_NUM_SPARE_MSG_BUFFERS)
+
+#define CM_INPUTQ_ENTRIES              128                     /* # entries in input queue */
+
+#define CM_PERIODIC_DISCOVER_INTERVAL  (5000)          /* 5s (infrequent resolution of established leader conflicts) */
+#define CM_URGENT_DISCOVER_INTERVAL    (50)            /* 0.05s (more frequently than heartbeats 'cause they don't retry) */
+#define CM_HEARTBEAT_INTERVAL          (125)           /* 0.125s */
+#define CM_TIMER_SCHEDULE_TIMEOUT      (4000)          /* 4s     Maximum time before a timer that's secheduled to run gets to run (eg blocked in interrupt handlers etc) */
+#define CM_THREAD_SCHEDULE_TIMEOUT     (30000)         /* 30s    Maximum time before a thread that's scheduled to run gets to run */
+#define CM_THREAD_RUNNING_TIMEOUT      (30000)         /* 30s    Don't expect the manager thread to be running longer than this */
+
+#ifdef PER_CPU_TIMEOUT
+#define CM_PERCPU_TIMEOUT_INTERVAL     (50)            /* 0.05s (must be less than all above intervals) */
+#define CM_PACEMAKER_INTERVAL          (500)           /* 0.05s */
+
+#define CM_HEARTBEAT_OVERDUE           (250)           /* 0.25s Maximum time a timeout can be overdue before taking extreme action */
+#endif
+
+#define CM_P2P_DMA_RETRIES             31
+
+/* We expect at least 1 point-to-point message in CM_P2P_MSG_RETRIES
+ * attempts to send one to be successfully received */
+#define CM_P2P_MSG_RETRIES             8
+
+/* We expect at least 1 broadcast message in CM_BCAST_MSG_RETRIES attempts
+ * to send one to be successfully received. */
+#define CM_BCAST_MSG_RETRIES           40
+
+/* Heartbeat timeout allows for a node stalling and still getting its
+ * heartbeat. The 2 is to allow for unsynchronised polling times. */
+#define CM_HEARTBEAT_TIMEOUT           (CM_TIMER_SCHEDULE_TIMEOUT + (2 + CM_P2P_MSG_RETRIES) * CM_HEARTBEAT_INTERVAL)
+
+/* Discover timeout must be > CM_HEARTBEAT_TIMEOUT to guarantee that people
+ * who don't see discovery are considered dead by their leader.  This
+ * ensures that by the time a node "discovers" it is a leader of a segment,
+ * the previous leader of that segment will have been deemed to be dead by
+ * its the parent segment's leader */
+#define CM_DISCOVER_TIMEOUT            (CM_TIMER_SCHEDULE_TIMEOUT + (2 + CM_BCAST_MSG_RETRIES) * CM_URGENT_DISCOVER_INTERVAL)
+
+#define CM_WAITING_TIMEOUT             (CM_DISCOVER_TIMEOUT * 100)
+
+/*
+ * Convert all timeouts specified in mS into "ticks"
+ */
+#define MSEC2TICKS(MSEC)               (((MSEC)*HZ)/1000)
+
+
+/* statemap entry */
+typedef struct cm_state_entry
+{
+    int16_t           level;                   /* cluster level to apply to */
+    int16_t          offset;                   /* from statemap_findchange() */
+    uint16_t          seg[BT_NBIPUL/16];       /* ditto */
+} CM_STATEMAP_ENTRY;
+
+/* offset is >= 0 for a change to apply and */
+#define STATEMAP_NOMORECHANGES (-1)            /* end of a set of updates */
+#define STATEMAP_RESET         (-2)            /* reset the target map */
+#define STATEMAP_NOOP          (-3)            /* null token */
+
+/* CM message format */
+typedef int8_t CM_SEQ;                         /* heartbeat sequence numbers; at least 2 bits, signed */
+
+/*
+ * The message header is received into the last 64 byte block of 
+ * the input queue and the Version *MUST* be the last word of the 
+ * block to ensure that we can see that the whole of the message
+ * has reached main memory after we've seen the input queue pointer
+ * have been updated.
+ */
+typedef struct ep_cm_hdr
+{
+    uint32_t          Pad0;
+    uint32_t          Pad1;
+
+    uint8_t           Type;
+    uint8_t           Level;
+    CM_SEQ            Seq;                     /* precision at least 2 bits each*/
+    CM_SEQ            AckSeq;
+    
+    uint16_t          NumMaps;
+    uint16_t          MachineId;
+
+    uint16_t          NodeId;
+    uint16_t          Checksum;
+
+    uint32_t           Timestamp;
+    uint32_t           ParamHash;
+    uint32_t          Version;
+} CM_HDR;
+
+#define CM_HDR_SIZE        sizeof (CM_HDR)
+
+typedef struct cm_msg
+{
+    union {
+       CM_STATEMAP_ENTRY   Statemaps[1];               /* piggy-backed statemap updates start here */
+       uint8_t             Space[EP_SYSTEMQ_MSG_MAX - CM_HDR_SIZE];
+    } Payload;
+    
+    CM_HDR                 Hdr;
+} CM_MSG;
+
+/* The maximum number of statemap entries that can fit within an EP_CM_MSG_BUFFER */
+#define CM_MSG_MAXMAPS         (offsetof (CM_MSG, Hdr) / sizeof (CM_STATEMAP_ENTRY))
+#define CM_MSG_MAP(mapno)      (CM_MSG_MAXMAPS - (mapno) - 1)
+
+/* The actual special message base & size, including 'nmaps' piggy-backed statemap entries */
+#define CM_MSG_BASE(nmaps)     (nmaps == 0 ? offsetof (CM_MSG, Hdr) : offsetof (CM_MSG, Payload.Statemaps[CM_MSG_MAXMAPS - nmaps]))
+#define CM_MSG_SIZE(nmaps)     (sizeof (CM_MSG) - CM_MSG_BASE(nmaps))
+
+#define CM_MSG_VERSION                         0xcad00005
+#define CM_MSG_TYPE_RESOLVE_LEADER             0
+#define CM_MSG_TYPE_DISCOVER_LEADER            1
+#define CM_MSG_TYPE_NOTIFY                     2
+#define CM_MSG_TYPE_DISCOVER_SUBORDINATE       3
+#define CM_MSG_TYPE_IMCOMING                   4
+#define CM_MSG_TYPE_HEARTBEAT                  5
+#define CM_MSG_TYPE_REJOIN                     6
+
+/* CM machine segment */
+typedef struct cm_sgmtMaps
+{
+    u_char       InputMapValid;                        /* Input map has been set */
+    u_char       OutputMapValid;               /* Output map has been set */
+    u_char       SentChanges;                  /* got an outstanding STATEMAP_NOMORECHANGES to send */
+    statemap_t  *OutputMap;                    /* state to send */
+    statemap_t  *InputMap;                     /* state received */
+    statemap_t  *CurrentInputMap;              /* state being received */
+} CM_SGMTMAPS;
+
+typedef struct cm_sgmt
+{
+   u_char       State;
+   u_char       SendMaps;
+   u_char       MsgAcked;
+   CM_SEQ      MsgSeq;
+   CM_SEQ      AckSeq;
+   u_int       NodeId;
+   long                UpdateTick;
+   long                WaitingTick;
+   uint32_t    Timestamp;
+   CM_SGMTMAPS  Maps[CM_MAX_LEVELS];           /* Maps[i] == state for cluster level i */
+   u_short      MsgNumber;                     /* msg buffer to use */
+   u_short     NumMaps;                        /* # maps in message buffer */
+   u_short      Level;
+   u_short      Sgmt;
+} CM_SGMT;
+
+#define CM_SGMT_ABSENT         0               /* no one there at all */
+#define CM_SGMT_WAITING                1               /* waiting for subtree to connect */
+#define CM_SGMT_COMING         2               /* expecting a subtree to reconnect */
+#define CM_SGMT_PRESENT                3               /* connected */
+
+typedef struct cm_level
+{
+    int               SwitchLevel;
+    u_int             MinNodeId;
+    u_int              NumNodes;
+    u_int              NumSegs;
+    u_int              MySgmt;
+   
+    /* SubordinateMap[i] == OR of all subordinate maps on this level and down for cluster level i */
+    u_char             SubordinateMapValid[CM_MAX_LEVELS];
+    statemap_t        *SubordinateMap[CM_MAX_LEVELS];
+
+    /* maps/flags for this cluster level */
+    u_int              Online:1;                               /* I've gone online (seen myself running) */
+    u_int             Restarting:1;                            /* driving my owm restart bit */
+    u_char            OfflineReasons;                          /* forced offline by broadcast */
+
+    u_char             GlobalMapValid;
+    u_char             SubTreeMapValid;
+    u_long            Connected;
+
+    statemap_t        *LocalMap;               /* state bits I drive */
+    statemap_t        *SubTreeMap;             /* OR of my and my subtree states */
+    statemap_t        *GlobalMap;              /* OR of all node states */
+    statemap_t        *LastGlobalMap;          /* last map I saw */
+    statemap_t        *TmpMap;                 /* scratchpad */
+
+    CM_SGMT           Sgmts[CM_SGMTS_PER_LEVEL];
+} CM_LEVEL;
+
+#define CM_ROLE_LEADER_CANDIDATE       0
+#define CM_ROLE_LEADER                 1
+#define CM_ROLE_SUBORDINATE            2
+
+/* global status bits */
+#define CM_GSTATUS_STATUS_MASK         0x03    /* bits nodes drive to broadcast their status */
+#define CM_GSTATUS_ABSENT              0x00    /* Off the network */
+#define CM_GSTATUS_STARTING            0x01    /* I'm waiting for everyone to see me online */
+#define CM_GSTATUS_RUNNING              0x03   /* up and running */
+#define CM_GSTATUS_CLOSING             0x02    /* I'm waiting for everyone to see me offline */
+
+#define CM_GSTATUS_ACK_MASK            0x0c    /* bits node drive to ack other status */
+#define CM_GSTATUS_MAY_START           0x04    /* Everyone thinks I may not start */
+#define CM_GSTATUS_MAY_RUN             0x08    /* Everyone thinks I may not run */
+
+#define CM_GSTATUS_RESTART             0x10    /* Someone thinks I should restart */
+#define CM_GSTATUS_BITS                        5
+
+#define CM_GSTATUS_BASE(node)          ((node) * CM_GSTATUS_BITS)
+
+#if defined(PER_CPU_TIMEOUT)
+typedef struct cm_timeout_data
+{
+    long               ScheduledAt;                            /* lbolt timeout was scheduled to run at */
+
+    unsigned long       EarlyCount;                            /* # times run early than NextRun */
+    unsigned long      MissedCount;                            /* # times run on time - but someone else was running it */
+    unsigned long       WastedCount;                           /* # times we failed to get the spinlock */
+    unsigned long      WorkCount;                              /* # times we're the one running */
+
+    unsigned long      WorstDelay;                             /* worst scheduling delay */
+    unsigned long      BestDelay;                              /* best scheduling delay */
+
+    unsigned long      WorstLockDelay;                         /* worst delay before getting rail->Lock */
+
+    unsigned long      WorstHearbeatDelay;                     /* worst delay before calling DoHeartbeatWork */
+} CM_TIMEOUT_DATA;
+#endif
+
+typedef struct cm_rail
+{
+    EP_RAIL          *Rail;                                    /* rail we're associated with */
+    struct list_head   Link;                                   /*   and linked on the CM_SUBSYS */
+
+    uint32_t          ParamHash;                               /* hash of critical parameters */
+    uint32_t           Timestamp;
+    long              DiscoverStartTick;                       /* when discovery start */
+
+    unsigned int       NodeId;                                 /* my node id */
+    unsigned int       NumNodes;                               /*   and number of nodes */
+    unsigned int       NumLevels;                              /* number of levels computed from machine size */
+    int                       BroadcastLevel;
+    long              BroadcastLevelTick;
+    unsigned int       TopLevel;                               /* level at which I'm not a leader */
+    unsigned char      Role;                                   /* state at TopLevel */
+
+    EP_INPUTQ        *PolledQueue;                             /* polled input queue */
+    EP_INPUTQ        *IntrQueue;                               /* intr input queue */
+    EP_OUTPUTQ       *MsgQueue;                                /* message  */
+    unsigned int       NextSpareMsg;                           /* next "spare" message buffer to use */
+
+    EP_CM_RAIL_STATS   Stats;                                  /* statistics */
+
+    kmutex_t          Mutex;
+    spinlock_t        Lock;
+    
+    long              NextHeartbeatTime;                       /* next time to check/send heartbeats */
+    long              NextDiscoverTime;                        /* next time to progress discovery  */
+    long              NextRunTime;                             /* the earlier of the above two or intr requires inputq poll*/
+
+    unsigned int       OfflineReasons;                         /* forced offline by procfs/manager thread stuck */
+
+#if defined(PER_CPU_TIMEOUT)
+    spinlock_t        HeartbeatTimeoutsLock;                   /* spinlock to sequentialise per-cpu timeouts */
+    long              HeartbeatTimeoutsStarted;                /* bitmap of which timeouts have started */
+    long              HeartbeatTimeoutsStopped;                /* bitmap of which timeouts have stopped */
+    long              HeartbeatTimeoutsShouldStop;             /* flag to indicate timeouts should stop */
+    kcondvar_t        HeartbeatTimeoutsWait;                   /* place to sleep waiting for timeouts to stop */
+    long              HeartbeatTimeoutRunning;                 /* someone is running the timeout - don't try for the lock */
+
+    long              HeartbeatTimeoutOverdue;                 /* heartbeat seen as overdue - interrupt requested */
+
+    CM_TIMEOUT_DATA   *HeartbeatTimeoutsData;                  /* per timeout data */
+#else
+    struct timer_list  HeartbeatTimer;                         /* timer for heartbeat/discovery */
+#endif
+
+    CM_LEVEL           Levels[CM_MAX_LEVELS];
+} CM_RAIL;
+
+/* OfflineReasons (both per-rail and  */
+#define CM_OFFLINE_BROADCAST           (1 << 0)
+#define CM_OFFLINE_PROCFS              (1 << 1)
+#define CM_OFFLINE_MANAGER             (1 << 2)
+
+typedef struct cm_subsys
+{
+    EP_SUBSYS          Subsys;
+    CM_RAIL            *Rails[EP_MAX_RAILS];
+} CM_SUBSYS;
+
+extern int  MachineId;
+
+extern void cm_node_disconnected (EP_RAIL *rail, unsigned nodeId);
+extern void cm_restart_node (EP_RAIL *rail, unsigned nodeId);
+extern void cm_restart_comms (CM_RAIL *cmRail);
+extern int  cm_init (EP_SYS *sys);
+
+extern void DisplayRail(EP_RAIL *rail);
+extern void DisplaySegs (EP_RAIL *rail);
+extern void DisplayStatus (EP_RAIL *rail);
+
+typedef struct proc_private
+{
+    struct nodeset_private *pr_next;
+    EP_RAIL                *pr_rail;
+    char                  *pr_data;
+    int                     pr_data_len;
+    unsigned               pr_off;
+    unsigned               pr_len;
+    DisplayInfo             pr_di;
+} PROC_PRIVATE;
+
+extern void    proc_character_fill (long mode, char *fmt, ...);
+extern int     proc_release (struct inode *inode, struct file *file);
+extern ssize_t proc_read (struct file *file, char *buf, size_t count, loff_t *ppos);
+
+
+extern void DisplayNodeMaps  (DisplayInfo *di, CM_RAIL *cmRail);
+extern void DisplayNodeSgmts (DisplayInfo *di, CM_RAIL *cmRail);
+extern void DisplayRailDo    (DisplayInfo *di, EP_RAIL *rail);
+
+extern int    cm_read_cluster(EP_RAIL *rail,char *page);
+extern void   cm_force_offline (EP_RAIL *rail, int offline, unsigned int reason);
+
+extern int    cm_svc_indicator_set      (EP_RAIL *rail, int svc_indicator);
+extern int    cm_svc_indicator_clear    (EP_RAIL *rail, int svc_indicator);
+extern int    cm_svc_indicator_is_set   (EP_RAIL *rail, int svc_indicator, int nodeId);
+extern int    cm_svc_indicator_bitmap   (EP_RAIL *rail, int svc_indicator, bitmap_t * bitmap, int low, int nnodes);
+
+/* cm_procfs.c */
+extern void   cm_procfs_init (CM_SUBSYS *subsys);
+extern void   cm_procfs_fini (CM_SUBSYS *subsys);
+extern void   cm_procfs_rail_init (CM_RAIL *rail);
+extern void   cm_procfs_rail_fini (CM_RAIL *rail);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN_CM_H */
+
Index: linux-2.4.21/drivers/net/qsnet/ep/cm_procfs.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/cm_procfs.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/cm_procfs.c       2005-06-01 23:12:54.633433936 -0400
@@ -0,0 +1,254 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2005 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: cm_procfs.c,v 1.5 2004/05/14 09:23:13 daniel Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/cm_procfs.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "debug.h"
+#include "cm.h"
+#include <elan/epsvc.h>
+
+#include <qsnet/procfs_linux.h>
+
+extern char *sprintClPeers (char *str, CM_RAIL *cmRail, int clvl);
+
+static int
+proc_read_cluster(char *page, char **start, off_t off,
+               int count, int *eof, void *data)
+{
+    CM_RAIL *cmRail = (CM_RAIL *) data;
+    char    *p      = page;
+
+    page[0] = 0;
+
+    if (cmRail->Rail->State != EP_RAIL_STATE_RUNNING)
+       p += sprintf(p, "<not running>\n");
+    else
+    {
+       CM_LEVEL *cmLevel;
+       unsigned long flags;
+       int  i, j;
+       char clNodeStr[32]; /* [%d-%d][%d-%d] */
+       char seperate_with;
+
+       struct { int val; char *name; } bitvals[] = {
+           {CM_OFFLINE_BROADCAST, "Broadcast"},
+           {CM_OFFLINE_PROCFS,    "Offline"},
+           {CM_OFFLINE_MANAGER,   "Manager"}};
+       
+       spin_lock_irqsave (&cmRail->Lock, flags);
+       
+       for (i = 0; i < cmRail->NumLevels; i++)
+       {
+           cmLevel = &cmRail->Levels[i];
+           
+           p += sprintf(p, "%23s %7s ", sprintClPeers (clNodeStr, cmRail, i), cmLevel->Online?"Online":"Offline");
+           
+           if ((cmLevel->Online ) | ( cmLevel->Connected > 0))
+               p += sprintf(p, "Connected=%lu ", cmLevel->Connected);
+           
+           seperate_with = '<';
+           
+           if ( cmLevel->Restarting ) {
+               p += sprintf(p, "%cRestarting", seperate_with);
+               seperate_with = ',';
+           }
+           
+           if ( ! (cmLevel->GlobalMapValid & cmLevel->SubTreeMapValid )) {
+               p += sprintf(p, "%cMap Not Valid", seperate_with);
+               seperate_with = ',';
+           }
+           
+           if ( cmLevel->OfflineReasons ) {
+               for (j = 0; j < sizeof (bitvals)/sizeof(bitvals[0]); j++)
+                   if (cmLevel->OfflineReasons & bitvals[j].val) {
+                       p += sprintf(p, "%c%s", seperate_with, bitvals[j].name);
+                       seperate_with = ',';
+                   }
+           }
+           if ( cmRail->OfflineReasons ) {
+               for (j = 0; j < sizeof (bitvals)/sizeof(bitvals[0]); j++)
+                   if (cmRail->OfflineReasons & bitvals[j].val) {
+                       p += sprintf(p, "%c%s", seperate_with, bitvals[j].name);
+                       seperate_with = ',';
+                   }
+           }
+           
+           if ( seperate_with != '<' ) 
+               p += sprintf(p,">\n");
+           else
+               p += sprintf(p,"\n");
+       }
+       
+       spin_unlock_irqrestore (&cmRail->Lock, flags);
+    }
+
+    return qsnet_proc_calc_metrics (page, start, off, count, eof, p - page);
+}
+
+static struct rail_info
+{
+    char *name;
+    int (*read_func) (char *page, char **start, off_t off, int count, int *eof, void *data);
+    int (*write_func) (struct file *file, const char *buf, unsigned long count, void *data);
+} rail_info[] = {
+    {"cluster", proc_read_cluster, NULL},
+};
+
+struct proc_dir_entry *svc_indicators_root;
+
+typedef struct svc_indicator_data
+{
+    int       svc_indicator;
+    EP_RAIL  *rail;
+} SVC_INDICATOR_DATA;
+
+static SVC_INDICATOR_DATA svc_indicator_data[EP_SVC_NUM_INDICATORS][EP_MAX_RAILS];
+static char              *svc_indicator_names[EP_SVC_NUM_INDICATORS] = EP_SVC_NAMES;
+
+static int
+proc_read_svc_indicator_rail_bitmap (char *page, char **start, off_t off,
+                                    int count, int *eof, void *data)
+{
+    SVC_INDICATOR_DATA  *svc_data = (SVC_INDICATOR_DATA  *)data;
+    unsigned int        nnodes   = ep_numnodes (ep_system());
+    bitmap_t           *bitmap;
+
+    KMEM_ZALLOC (bitmap, bitmap_t *, (BT_BITOUL(EP_MAX_NODES) * sizeof (bitmap_t)), 1);
+
+    cm_svc_indicator_bitmap (svc_data->rail, svc_data->svc_indicator, bitmap, 0, nnodes);
+
+    ep_sprintf_bitmap (page, PAGESIZE, bitmap, 0, 0, nnodes);
+    
+    KMEM_FREE (bitmap, (BT_BITOUL(EP_MAX_NODES) * sizeof (bitmap_t)));
+    
+    strcat (page, "\n");
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, strlen(page)));
+}
+
+static int
+proc_read_svc_indicator_bitmap(char *page, char **start, off_t off,
+                              int count, int *eof, void *data)
+{
+    unsigned int         num      = (unsigned long) data;
+    EP_SYS              *sys      = ep_system();
+    unsigned int         nnodes   = ep_numnodes (sys);
+    bitmap_t            *bitmap;
+
+    KMEM_ALLOC(bitmap, bitmap_t *, (BT_BITOUL(EP_MAX_NODES) * sizeof (bitmap_t)), 1);
+     
+    ep_svc_indicator_bitmap (sys, num, bitmap, 0, nnodes);
+
+    ep_sprintf_bitmap (page, PAGESIZE, bitmap, 0, 0, nnodes);
+    
+    KMEM_FREE (bitmap, (BT_BITOUL(EP_MAX_NODES) * sizeof (bitmap_t)));
+    
+    strcat (page, "\n");
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, strlen(page)));
+}
+
+void
+cm_procfs_rail_init (CM_RAIL *cmRail)
+{
+    EP_RAIL *rail = cmRail->Rail;
+    struct proc_dir_entry *p;
+    int i;
+    
+    for (i = 0; i < sizeof (rail_info)/sizeof (rail_info[0]); i++)
+    {
+       if ((p = create_proc_entry (rail_info[i].name, 0, cmRail->Rail->ProcDir)) != NULL)
+       {
+           p->read_proc  = rail_info[i].read_func;
+           p->write_proc = rail_info[i].write_func;
+           p->data       = cmRail;
+           p->owner      = THIS_MODULE;
+       }
+    }
+
+    if ((rail->SvcIndicatorDir = proc_mkdir ("svc_indicators", cmRail->Rail->ProcDir)) != NULL)
+    {
+       for (i = 0; i < EP_SVC_NUM_INDICATORS; i++)
+       {
+           if ((p = create_proc_entry (svc_indicator_names[i], 0, rail->SvcIndicatorDir)) != NULL)
+           {
+               svc_indicator_data[i][rail->Number].svc_indicator = i;
+               svc_indicator_data[i][rail->Number].rail          = rail; 
+               
+               p->write_proc = NULL;
+               p->read_proc  = proc_read_svc_indicator_rail_bitmap;
+               p->data       = (void *)&svc_indicator_data[i][rail->Number];
+               p->owner      = THIS_MODULE;
+           }
+       }
+    }
+}
+
+void
+cm_procfs_rail_fini (CM_RAIL *cmRail)
+{
+    EP_RAIL *rail = cmRail->Rail;
+    int i;
+
+    if (rail->SvcIndicatorDir)
+    {
+       for (i = 0; i < EP_SVC_NUM_INDICATORS; i++)
+           remove_proc_entry (svc_indicator_names[i], rail->SvcIndicatorDir);
+
+       remove_proc_entry ("svc_indicators", cmRail->Rail->ProcDir);
+    }
+
+    for (i = 0; i < sizeof (rail_info)/sizeof (rail_info[0]); i++)
+       remove_proc_entry (rail_info[i].name, cmRail->Rail->ProcDir);
+}
+
+void
+cm_procfs_init (CM_SUBSYS *subsys)
+{
+    struct proc_dir_entry *p;
+    int i;
+
+    qsnet_proc_register_hex (ep_config_root, "machine_id",      &MachineId,      0);
+
+    if ((svc_indicators_root = proc_mkdir("svc_indicators", ep_procfs_root)) != NULL)
+    {
+       for (i = 0; i < EP_SVC_NUM_INDICATORS; i++)
+       {
+           if ((p = create_proc_entry (svc_indicator_names[i], 0, svc_indicators_root)) != NULL)
+           {
+               p->write_proc = NULL;
+               p->read_proc  = proc_read_svc_indicator_bitmap;
+               p->data       = (void *)(long) i;
+               p->owner      = THIS_MODULE;
+           }
+       }
+       
+    }
+}
+
+void
+cm_procfs_fini (CM_SUBSYS *subsys)
+{
+    int i;
+
+    if (svc_indicators_root)
+    {
+       for (i = 0; i < EP_SVC_NUM_INDICATORS; i++)
+           remove_proc_entry (svc_indicator_names[i], svc_indicators_root);
+       
+       remove_proc_entry ("svc_indicators",   ep_procfs_root);
+    }
+
+    remove_proc_entry ("machine_id",      ep_config_root);
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/commands_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/commands_elan4.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/commands_elan4.c  2005-06-01 23:12:54.634433784 -0400
@@ -0,0 +1,173 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: commands_elan4.c,v 1.2 2003/10/23 15:07:53 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/commands_elan4.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "debug.h"
+
+#include <elan4/trtype.h>
+
+static __inline__ void
+elan4_command_write (ELAN4_CQ *cq, E4_uint64 val, unsigned off)
+{
+    writeq (val, cq->cq_mapping + offsetof (E4_CommandPort, Command[off]));
+}
+
+void
+elan4_nop_cmd (ELAN4_CQ *cq, E4_uint64 tag)
+{
+    elan4_command_write (cq, tag | NOP_CMD, 0);
+}
+
+void
+elan4_write_dword_cmd (ELAN4_CQ *cq, E4_Addr addr, E4_uint64 data)
+{
+    elan4_command_write (cq, addr | WRITE_DWORD_CMD, 0);
+    elan4_command_write (cq, data, 1);
+}
+
+void
+elan4_add_dword_cmd (ELAN4_CQ *cq, E4_Addr addr, E4_uint64 data)
+{
+    elan4_command_write (cq, addr | ADD_DWORD_CMD, 0);
+    elan4_command_write (cq, data,                 1);
+}
+
+void
+elan4_copy64_cmd (ELAN4_CQ *cq, E4_Addr from, E4_Addr to, E4_uint32 datatype)
+{
+    elan4_command_write (cq, from | (datatype << COPY64_DATA_TYPE_SHIFT) | COPY64_CMD, 0);
+    elan4_command_write (cq, to   | (datatype << COPY64_DATA_TYPE_SHIFT),              1);
+}
+
+void
+elan4_interrupt_cmd (ELAN4_CQ *cq, E4_uint64 cookie)
+{
+    elan4_command_write (cq, (cookie << E4_MAIN_INT_SHIFT) | INTERRUPT_CMD, 0);
+}
+
+
+void 
+elan4_run_thread_cmd (ELAN4_CQ *cq, E4_ThreadRegs *regs)
+{
+    elan4_command_write (cq, regs->Registers[0] | RUN_THREAD_CMD, 0);
+    elan4_command_write (cq, regs->Registers[1],                  1);
+    elan4_command_write (cq, regs->Registers[2],                  2);
+    elan4_command_write (cq, regs->Registers[3],                  3);
+    elan4_command_write (cq, regs->Registers[4],                  4);
+    elan4_command_write (cq, regs->Registers[5],                  5);
+    elan4_command_write (cq, regs->Registers[6],                  6);
+}
+
+void
+elan4_run_dma_cmd (ELAN4_CQ *cq, E4_DMA *dma)
+{
+    E4_uint64 *dmaptr = (E4_uint64 *) dma;
+
+    elan4_command_write (cq, dmaptr[0] | RUN_DMA_CMD, 0);
+    elan4_command_write (cq, dmaptr[1],               1);
+    elan4_command_write (cq, dmaptr[2],               2);
+    elan4_command_write (cq, dmaptr[3],               3);
+    elan4_command_write (cq, dmaptr[4],               4);
+    elan4_command_write (cq, dmaptr[5],               5);
+    elan4_command_write (cq, dmaptr[6],               6);
+}
+
+void
+elan4_set_event_cmd (ELAN4_CQ *cq, E4_Addr event)
+{
+    elan4_command_write (cq, event | SET_EVENT_CMD, 0);
+}
+
+void
+elan4_set_eventn_cmd (ELAN4_CQ *cq, E4_Addr event, E4_uint32 count)
+{
+    elan4_command_write (cq, SET_EVENTN_CMD,0);
+    elan4_command_write (cq, event | count, 1);
+}
+    
+void
+elan4_wait_event_cmd (ELAN4_CQ *cq, E4_Addr event, E4_uint64 candt, E4_uint64 param0, E4_uint64 param1)
+{
+    elan4_command_write (cq, event | WAIT_EVENT_CMD, 0);
+    elan4_command_write (cq, candt,                  1);
+    elan4_command_write (cq, param0,                 2);
+    elan4_command_write (cq, param1,                 3);
+}
+
+void
+elan4_open_packet (ELAN4_CQ *cq, E4_uint64 command)
+{
+    elan4_command_write (cq, command | OPEN_STEN_PKT_CMD, 0);
+}
+
+void
+elan4_guard (ELAN4_CQ *cq, E4_uint64 command)
+{
+    elan4_command_write (cq, command | GUARD_CMD, 0);
+}
+
+void
+elan4_sendtrans0 (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr)
+{
+    elan4_command_write (cq, (trtype << 16) | SEND_TRANS_CMD, 0);
+    elan4_command_write (cq, addr,                               1);
+}
+
+void
+elan4_sendtrans1 (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, E4_uint64 p0)
+{
+    elan4_command_write (cq, (trtype << 16) | SEND_TRANS_CMD, 0);
+    elan4_command_write (cq, addr,                               1);
+    elan4_command_write (cq, p0,                                 2);
+}
+
+void
+elan4_sendtrans2 (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, E4_uint64 p0, E4_uint64 p1)
+{
+    elan4_command_write (cq, (trtype << 16) | SEND_TRANS_CMD, 0);
+    elan4_command_write (cq, addr,                               1);
+    elan4_command_write (cq, p0,                                 2);
+    elan4_command_write (cq, p1,                                 3);
+}
+
+void
+elan4_sendtransn (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, ...)
+{
+    E4_uint32    ndword = ((trtype & TR_SIZE_MASK) >> TR_SIZE_SHIFT);
+    va_list      ap;
+    register int i;
+
+    elan4_command_write (cq, (trtype << 16) | SEND_TRANS_CMD, 0);
+    elan4_command_write (cq, addr,                               1);
+    
+    va_start (ap, addr);
+    for (i = 2; i < ndword+2; i++) 
+       elan4_command_write (cq, va_arg (ap, E4_uint64), i);
+    va_end (ap);
+}
+
+void
+elan4_sendtransp (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, E4_uint64 *ptr)
+{
+    E4_uint32    ndword = ((trtype &TR_SIZE_MASK) >> TR_SIZE_SHIFT);
+    register int i;
+
+    elan4_command_write (cq, (trtype << 16) | SEND_TRANS_CMD, 0);
+    elan4_command_write (cq, addr,                            1);
+    for (i = 2; i < ndword+2; i++)
+       elan4_command_write (cq, *ptr++, i);
+}
+
Index: linux-2.4.21/drivers/net/qsnet/ep/conf_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/conf_linux.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/conf_linux.c      2005-06-01 23:12:54.635433632 -0400
@@ -0,0 +1,309 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: conf_linux.c,v 1.37.2.3 2005/01/18 14:47:35 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/conf_linux.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <qsnet/autoconf.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "cm.h"
+
+#include "conf_linux.h"
+
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/reboot.h>
+#include <linux/notifier.h>
+
+/* Module parameters */
+unsigned int epdebug        = 0;
+unsigned int epdebug_console = 0;
+unsigned int epdebug_cmlevel = 0;
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+unsigned int epdebug_check_sum = 0;
+#endif
+int         disabled        = 0;
+int          sdram_assert    = 0;
+int          assfail_mode    = 0;
+int         txd_stabilise   = 7;
+int         portals_envelopes = 0;
+
+/* External module parameters */
+extern int     MaxSwitchLevels;
+extern int      RejoinCheck;
+extern int      RejoinPanic;
+extern int      PositionCheck;
+extern int      MachineId;
+
+/* Module globals */
+EP_SYS          epsys;
+
+#ifdef MODULE
+MODULE_AUTHOR("Quadrics Ltd");
+MODULE_DESCRIPTION("Elan Kernel Comms");
+
+MODULE_LICENSE("GPL");
+
+MODULE_PARM(epdebug,         "i");
+MODULE_PARM(epdebug_console, "i");
+MODULE_PARM(epdebug_cmlevel, "i");
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+MODULE_PARM(epdebug_check_sum, "i");
+#endif
+MODULE_PARM(disabled,        "i");
+
+MODULE_PARM(MachineId,       "i");
+MODULE_PARM(RejoinPanic,     "i");
+MODULE_PARM(RejoinCheck,     "i");
+MODULE_PARM(PositionCheck,   "i");
+MODULE_PARM(MaxSwitchLevels, "i");
+
+MODULE_PARM(sdram_assert,    "i");
+MODULE_PARM(assfail_mode,    "i");
+MODULE_PARM(txd_stabilise,   "i");
+MODULE_PARM(portals_envelopes,"i");
+
+/* epcomms.c large message service functions */
+EXPORT_SYMBOL(ep_alloc_xmtr);
+EXPORT_SYMBOL(ep_free_xmtr);
+EXPORT_SYMBOL(ep_transmit_message);
+EXPORT_SYMBOL(ep_multicast_message);
+EXPORT_SYMBOL(ep_transmit_rpc);
+
+EXPORT_SYMBOL(ep_alloc_rcvr);
+EXPORT_SYMBOL(ep_free_rcvr);
+EXPORT_SYMBOL(ep_queue_receive);
+EXPORT_SYMBOL(ep_requeue_receive);
+EXPORT_SYMBOL(ep_rpc_put);
+EXPORT_SYMBOL(ep_rpc_get);
+EXPORT_SYMBOL(ep_complete_rpc);
+EXPORT_SYMBOL(ep_complete_receive);
+
+EXPORT_SYMBOL(ep_poll_transmits);
+EXPORT_SYMBOL(ep_enable_txcallbacks);
+EXPORT_SYMBOL(ep_disable_txcallbacks);
+
+/* epcomms.c functions for accessing fields of rxds/txds */
+EXPORT_SYMBOL(ep_rxd_arg);
+EXPORT_SYMBOL(ep_rxd_len);
+EXPORT_SYMBOL(ep_rxd_isrpc);
+EXPORT_SYMBOL(ep_rxd_envelope);
+EXPORT_SYMBOL(ep_rxd_payload);
+EXPORT_SYMBOL(ep_rxd_node);
+EXPORT_SYMBOL(ep_rxd_status);
+EXPORT_SYMBOL(ep_rxd_statusblk);
+EXPORT_SYMBOL(ep_txd_node);
+EXPORT_SYMBOL(ep_txd_statusblk);
+
+/* kmap.c, nmh.c - handling mapping of pages into network memory */
+EXPORT_SYMBOL(ep_dvma_reserve);
+EXPORT_SYMBOL(ep_dvma_release);
+EXPORT_SYMBOL(ep_dvma_load);
+EXPORT_SYMBOL(ep_dvma_unload);
+EXPORT_SYMBOL(ep_nmd_subset);
+EXPORT_SYMBOL(ep_nmd_merge);
+
+EXPORT_SYMBOL(ep_system);
+
+/* kcomm.c */
+EXPORT_SYMBOL(ep_nodeid);
+EXPORT_SYMBOL(ep_numnodes);
+EXPORT_SYMBOL(ep_waitfor_nodeid);
+
+/* railhints.c */
+EXPORT_SYMBOL(ep_pickRail);
+EXPORT_SYMBOL(ep_xmtr_bcastrail);
+EXPORT_SYMBOL(ep_xmtr_prefrail);
+EXPORT_SYMBOL(ep_xmtr_availrails);
+EXPORT_SYMBOL(ep_xmtr_noderails);
+EXPORT_SYMBOL(ep_rcvr_prefrail);
+EXPORT_SYMBOL(ep_rcvr_availrails);
+EXPORT_SYMBOL(ep_rxd_railmask);
+
+EXPORT_SYMBOL(ep_svc_indicator_bitmap);
+EXPORT_SYMBOL(ep_svc_indicator_is_set);
+EXPORT_SYMBOL(ep_svc_indicator_clear);
+EXPORT_SYMBOL(ep_svc_indicator_set);
+
+/* cm.c */
+EXPORT_SYMBOL(cm_svc_indicator_clear);
+EXPORT_SYMBOL(cm_svc_indicator_set);
+EXPORT_SYMBOL(cm_svc_indicator_is_set);
+EXPORT_SYMBOL(cm_svc_indicator_bitmap);
+
+#endif
+
+EP_SYS *
+ep_system()
+{
+    return (&epsys);
+}
+
+void
+ep_mod_inc_usecount()
+{
+    MOD_INC_USE_COUNT;
+} 
+
+void
+ep_mod_dec_usecount()
+{
+    MOD_DEC_USE_COUNT;
+}
+
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+
+#include <linux/dump.h>
+
+static int
+ep_dump_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    if (event == DUMP_BEGIN)
+       ep_shutdown (&epsys);
+
+    return (NOTIFY_DONE);
+}
+static struct notifier_block ep_dump_notifier = 
+{
+    notifier_call:     ep_dump_event,
+    priority:          0,
+};
+
+#endif
+
+static int
+ep_reboot_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    if ((event == SYS_RESTART || event == SYS_HALT || event == SYS_POWER_OFF))
+       ep_shutdown (&epsys);
+
+    return (NOTIFY_DONE);
+}
+
+static struct notifier_block ep_reboot_notifier = 
+{
+    notifier_call:     ep_reboot_event,
+    priority:          0,
+};
+
+static int
+ep_panic_event (struct notifier_block *self, unsigned long event, void *buffer)
+{
+    ep_shutdown (&epsys);
+
+    return (NOTIFY_DONE);
+}
+
+static struct notifier_block ep_panic_notifier = 
+{
+    notifier_call:     ep_panic_event,
+    priority:          0,
+};
+
+/*
+ * Module configuration. 
+ */
+#ifdef MODULE
+static int __init ep_init(void)
+#else
+__initfunc(int ep_init(void))
+#endif
+{
+    register int rmask = 0;
+
+    ep_procfs_init ();
+
+    ep_sys_init (&epsys);
+
+#if defined(CONFIG_ELAN4) || defined(CONFIG_ELAN4_MODULE)
+    rmask = ep4_create_rails (&epsys, disabled);
+#endif
+    
+    /* If we've brought up an elan4 rail, then disable all elan3 rails. */
+    if ((rmask & ~disabled) != 0)
+       disabled = ~rmask;
+
+#if defined(CONFIG_ELAN3) || defined(CONFIG_ELAN3_MODULE)
+    rmask = ep3_create_rails (&epsys, disabled);
+#endif
+
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+    register_dump_notifier (&ep_dump_notifier);
+#endif
+    register_reboot_notifier (&ep_reboot_notifier);
+
+#if !defined(NO_PANIC_NOTIFIER)
+    notifier_chain_register (&panic_notifier_list, &ep_panic_notifier);
+#endif
+
+    return (0);
+}
+
+/*
+ * Module removal.
+ */
+#ifdef MODULE
+static void
+__exit ep_exit(void)
+{
+    register int i;
+
+#if defined(CONFIG_DUMP) || defined(CONFIG_DUMP_MODULE)
+    unregister_dump_notifier (&ep_dump_notifier);
+#endif
+    unregister_reboot_notifier (&ep_reboot_notifier);
+
+#if !defined(NO_PANIC_NOTIFIER)
+    notifier_chain_unregister (&panic_notifier_list, &ep_panic_notifier);
+#endif
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+    {
+       if (epsys.Rails[i])
+       {
+           switch (epsys.Rails[i]->State)
+           {
+           case EP_RAIL_STATE_UNINITIALISED:
+               break;
+
+           case EP_RAIL_STATE_STARTED:
+           case EP_RAIL_STATE_RUNNING:
+           case EP_RAIL_STATE_INCOMPATIBLE:
+               /* remove per-rail CM proc entries */
+               ep_stop_rail (epsys.Rails[i]);
+               break;
+           }
+
+           /* remove EP proc rail entries after per-rail CM entries */
+           ep_procfs_rail_fini (epsys.Rails[i]);
+           ep_destroy_rail (epsys.Rails[i]);
+       }
+    }
+
+    ep_sys_fini (&epsys);
+
+    ep_procfs_fini ();
+}
+
+/* Declare the module init and exit functions */
+module_init(ep_init);
+module_exit(ep_exit);
+
+#endif
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/conf_linux.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/conf_linux.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/conf_linux.h      2005-06-01 23:12:54.635433632 -0400
@@ -0,0 +1,29 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: conf_linux.h,v 1.6 2003/10/02 14:16:07 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/conf_linux.h,v $*/
+
+#ifndef __ELAN_CONF_LINUX_H
+#define __ELAN_CONF_LINUX_H
+
+extern void ep_procfs_init(void);
+extern void ep_procfs_fini(void);
+extern void ep_procfs_rail_init(EP_RAIL *rail);
+extern void ep_procfs_rail_fini(EP_RAIL *rail);
+
+extern void ep_procfs_svc_indicator_create(int svc_indicator, char *name);
+extern void ep_procfs_svc_indicator_remove(int svc_indicator, char *name);
+
+#endif /* __ELAN_CONF_LINUX_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/debug.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/debug.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/debug.c   2005-06-01 23:12:54.635433632 -0400
@@ -0,0 +1,145 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: debug.c,v 1.28.2.1 2004/11/12 10:54:50 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/debug.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "debug.h"
+
+DisplayInfo di_ep_debug = {ep_debugf, DBG_DEBUG};
+
+/*
+ * Generate a partial bitmap string, for the bitmap from offset "off" for "count" bits,
+ * to allow for displaying of subsets, treat entry 0 of the bitmap as having value "base".
+ */
+int
+ep_sprintf_bitmap (char *str, unsigned nbytes, bitmap_t *bitmap, int base, int off, int nbits)
+{
+    char entry[12];                                            /* space for N-N */
+    register int i, j, len;
+    register int notstart = off;
+    register int notfirst = 0;
+    char *p = str;
+    
+    for (i = off; i < nbits; i++)
+    {
+       if (BT_TEST (bitmap, i))
+       {
+           for (j = i+1; j < nbits; j++)
+               if (! BT_TEST (bitmap, j))
+                   break;
+
+           if (j == (i+1))
+               len = (int)sprintf (entry, "%d", base + i);
+           else
+               len = (int)sprintf (entry, "%d-%d", base + i, base + j-1);
+           
+           /* NOTE the 2 is for: one for comma, one for (possible) closing bracket */
+           if ((p - str) <= (nbytes - (len+3)))
+               p += (int)sprintf (p, "%c%s", notfirst++ ? ',' : notstart ? ' ' : '[', entry);
+           else
+           {
+               /* no more space on this line, so move onto next */
+               sprintf (p, "%c", notfirst++ ? ',' : '[');
+
+               return (i);
+           }
+
+           i = j;
+       }
+    }
+    
+    if (!notfirst)
+       sprintf (str, "<empty>");
+    else
+       strcpy (p, "]");
+
+    return (-1);
+}
+
+void
+ep_display_bitmap (char *prefix, char *tag, bitmap_t *bitmap, unsigned base, unsigned nbits)
+{
+    /* Tru64 kernel printf() truncates lines at 128 bytes - the man pages for printf (9)
+     * do not mention this restriction, nor that it does not terminate the line with a
+     * carriage return, this  is pretty naff. 
+     * Linux has a similar limit though is much more generous at 1024 - and you can just 
+     * look at the code to see why this has been done.
+     *
+     * Our nodeset information could well be longer than 128 characters,  so we're going to 
+     * have to split it into a number of lines. */
+
+#define LINEBUF_SIZE           128
+    char *p, linebuf[LINEBUF_SIZE+1];                          /* +1 for null termination */
+    int i, noff, off = 0;
+
+    do {
+       if (off == 0)
+           p = linebuf + (int)sprintf (linebuf, "%s: %s ", prefix, tag);
+       else
+       {
+           p = linebuf + (int)sprintf (linebuf, "%s:  ", prefix);
+           for (i = 0; tag[i] != '\0'; i++)
+               *p++ = ' ';
+       }
+
+       noff = ep_sprintf_bitmap (p, &linebuf[LINEBUF_SIZE-1]-p, bitmap, base, off, nbits);
+
+       printk ("%s\n", linebuf);
+
+    } while ((off = noff) != -1);
+
+#undef LINEBUF_SIZE
+}
+
+void
+ep_debugf (long mode, char *fmt, ...)
+{
+   va_list ap;
+   char prefix[32];
+   
+   va_start (ap, fmt);
+#if defined(LINUX)
+   sprintf (prefix, "[%08d.%04d] ", (int) lbolt, current->pid);
+#else
+   sprintf (prefix, "[%08d.----] ", (int) lbolt);
+#endif
+   qsnet_vdebugf ((mode & epdebug_console ? QSNET_DEBUG_CONSOLE: 0) | QSNET_DEBUG_BUFFER, prefix, fmt, ap);
+   va_end (ap);
+}
+
+int
+ep_assfail (EP_RAIL *rail, const char *ex, const char *func, const char *file, const int line)
+{
+    qsnet_debugf (QSNET_DEBUG_BUFFER, "ep: assertion failure: %s, function: %s, file %s, line: %d\n", ex, func, file, line);
+    
+    printk (KERN_EMERG "ep: assertion failure: %s, function: %s, file %s, line: %d\n", ex, func, file, line);
+    
+    if (panicstr)
+       return (0);
+    
+    if (assfail_mode & 1)                              /* return to BUG() */
+       return 1;
+    
+    if (assfail_mode & 2)
+       panic ("ep: assertion failure: %s, function: %s, file %s, line: %d\n", ex, func, file, line);
+    if (assfail_mode & 4)
+       epdebug = 0;
+    
+    return 0;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/debug_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/debug_elan4.c        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/debug_elan4.c     2005-06-01 23:12:54.636433480 -0400
@@ -0,0 +1,59 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: debug_elan4.c,v 1.1 2004/05/19 10:21:04 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/debug_elan4.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "conf_linux.h"
+#include "debug.h"
+
+static void
+ep4_display_ecqs (EP4_RAIL *rail)
+{
+    struct list_head *el;
+    unsigned long flags;
+    int i;
+
+    spin_lock_irqsave (&rail->r_ecq_lock, flags);
+    for (i = 0; i <EP4_NUM_ECQ; i++)
+    {
+       list_for_each (el, &rail->r_ecq_list[i]) {
+           EP4_ECQ *ecq = list_entry (el, EP4_ECQ, ecq_link);
+
+           ep_debugf (DBG_DEBUG, "ECQ: type %d: avail %d cqnum %d\n", i, ecq->ecq_avail, elan4_cq2num (ecq->ecq_cq));
+       }
+    }
+    spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+}
+
+void
+ep4_debug_rail (EP_RAIL *r)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+    EP_SYS   *sys  = rail->r_generic.System;
+
+    ep_debugf (DBG_DEBUG, "ep%d: is elan4 %d rev %c\n", rail->r_generic.Number,
+              rail->r_generic.Devinfo.dev_instance, 'a' + rail->r_generic.Devinfo.dev_revision_id);
+
+    ep4_display_ecqs (rail);
+
+    ep_display_alloc (&sys->Allocator);
+    ep_display_rmap (sys->Allocator.ResourceMap);
+
+    ep_display_alloc (&rail->r_generic.ElanAllocator);
+    ep_display_alloc (&rail->r_generic.MainAllocator);
+
+    ep_display_rmap (rail->r_generic.ElanAllocator.ResourceMap);
+}
+
Index: linux-2.4.21/drivers/net/qsnet/ep/debug.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/debug.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/debug.h   2005-06-01 23:12:54.636433480 -0400
@@ -0,0 +1,109 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_EPDEBUG_H
+#define _ELAN3_EPDEBUG_H
+
+#ident "$Id: debug.h,v 1.18.2.1 2004/11/12 10:54:50 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/debug.h,v $ */
+
+extern unsigned int epdebug;
+extern unsigned int epdebug_console;
+extern unsigned int epdebug_cmlevel;
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+extern unsigned int epdebug_check_sum;
+#endif
+#define DBG_CONFIG             0x00000001                      /* Module configuration */
+#define DBG_PROBE              0x00000002
+#define DBG_ROUTETABLE         0x00000004
+#define DBG_STATEMAP           0x00000008
+
+#define DBG_CM                 0x00000020
+#define DBG_XMTR               0x00000040
+#define DBG_RCVR               0x00000080
+#define DBG_FORWARD            0x00000100
+#define DBG_DISCON             0x00000200
+#define DBG_EPTRAP             0x00000400
+#define DBG_COMMAND            0x00000800
+#define DBG_RETRY              0x00001000
+#define DBG_DEBUG              0x00002000
+#define DBG_NETWORK_ERROR      0x00004000
+#define DBG_MSGSYS             0x00008000
+#define DBG_MANAGER            0x00010000
+#define DBG_KMAP               0x00020000
+#define DBG_FAILOVER           0x00040000
+#define DBG_MAPNMD             0x00080000
+#define DBG_KMSG               0x00100000
+#define DBG_SVC                 0x00200000
+#define DBG_STABILISE          0x00400000
+
+#if defined(DEBUG_PRINTF)
+
+#  define EPRINTF0(m,fmt)                      ((epdebug&(m)) ? ep_debugf(m,fmt)                     : (void)0)
+#  define EPRINTF1(m,fmt,a)                    ((epdebug&(m)) ? ep_debugf(m,fmt,a)                   : (void)0)
+#  define EPRINTF2(m,fmt,a,b)                  ((epdebug&(m)) ? ep_debugf(m,fmt,a,b)                 : (void)0)
+#  define EPRINTF3(m,fmt,a,b,c)                        ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c)               : (void)0)
+#  define EPRINTF4(m,fmt,a,b,c,d)              ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c,d)             : (void)0)
+#  define EPRINTF5(m,fmt,a,b,c,d,e)            ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c,d,e)           : (void)0)
+#  define EPRINTF6(m,fmt,a,b,c,d,e,f)          ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c,d,e,f)         : (void)0)
+#  define EPRINTF7(m,fmt,a,b,c,d,e,f,g)                ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c,d,e,f,g)       : (void)0)
+#  define EPRINTF8(m,fmt,a,b,c,d,e,f,g,h)      ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c,d,e,f,g,h)     : (void)0)
+#  define EPRINTF9(m,fmt,a,b,c,d,e,f,g,h,i)    ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c,d,e,f,g,h,i)   : (void)0)
+#  define EPRINTF10(m,fmt,a,b,c,d,e,f,g,h,i,j) ((epdebug&(m)) ? ep_debugf(m,fmt,a,b,c,d,e,f,g,h,i,j) : (void)0)
+
+#  define CPRINTF0(lvl,fmt)                    (((lvl) <= epdebug_cmlevel) ? EPRINTF0(DBG_CM,fmt)                   : (void)0)
+#  define CPRINTF1(lvl,fmt,a)                  (((lvl) <= epdebug_cmlevel) ? EPRINTF1(DBG_CM,fmt,a)                 : (void)0)
+#  define CPRINTF2(lvl,fmt,a,b)                        (((lvl) <= epdebug_cmlevel) ? EPRINTF2(DBG_CM,fmt,a,b)               : (void)0)
+#  define CPRINTF3(lvl,fmt,a,b,c)              (((lvl) <= epdebug_cmlevel) ? EPRINTF3(DBG_CM,fmt,a,b,c)             : (void)0)
+#  define CPRINTF4(lvl,fmt,a,b,c,d)            (((lvl) <= epdebug_cmlevel) ? EPRINTF4(DBG_CM,fmt,a,b,c,d)           : (void)0)
+#  define CPRINTF5(lvl,fmt,a,b,c,d,e)          (((lvl) <= epdebug_cmlevel) ? EPRINTF5(DBG_CM,fmt,a,b,c,d,e)         : (void)0)
+#  define CPRINTF6(lvl,fmt,a,b,c,d,e,f)                (((lvl) <= epdebug_cmlevel) ? EPRINTF6(DBG_CM,fmt,a,b,c,d,e,f)       : (void)0)
+#  define CPRINTF7(lvl,fmt,a,b,c,d,e,f,g)      (((lvl) <= epdebug_cmlevel) ? EPRINTF7(DBG_CM,fmt,a,b,c,d,e,f,g)     : (void)0)
+#  define CPRINTF8(lvl,fmt,a,b,c,d,e,f,g,h)    (((lvl) <= epdebug_cmlevel) ? EPRINTF8(DBG_CM,fmt,a,b,c,d,e,f,g,h)   : (void)0)
+#  define CPRINTF9(lvl,fmt,a,b,c,d,e,f,g,h,i)  (((lvl) <= epdebug_cmlevel) ? EPRINTF9(DBG_CM,fmt,a,b,c,d,e,f,g,h,i) : (void)0)
+
+#if defined __GNUC__
+extern void ep_debugf (long mode, char *fmt, ...) __attribute__ ((format (printf,2,3)));
+#else
+extern void ep_debugf (long mode, char *fmt, ...);
+#endif
+
+#else
+
+#  define EPRINTF0(m,fmt)                      (0)
+#  define EPRINTF1(m,fmt,a)                    (0)
+#  define EPRINTF2(m,fmt,a,b)                  (0)
+#  define EPRINTF3(m,fmt,a,b,c)                        (0)
+#  define EPRINTF4(m,fmt,a,b,c,d)              (0)
+#  define EPRINTF5(m,fmt,a,b,c,d,e)            (0)
+#  define EPRINTF6(m,fmt,a,b,c,d,e,f)          (0)
+#  define EPRINTF7(m,fmt,a,b,c,d,e,f,g)                (0)
+#  define EPRINTF8(m,fmt,a,b,c,d,e,f,g,h)      (0)
+#  define EPRINTF9(m,fmt,a,b,c,d,e,f,g,h,i)    (0)
+#  define EPRINTF9(m,fmt,a,b,c,d,e,f,g,h,i,j)  (0)
+
+#  define CPRINTF0(lvl,fmt)                    (0)
+#  define CPRINTF1(lvl,fmt,a)                  (0)
+#  define CPRINTF2(lvl,fmt,a,b)                        (0)
+#  define CPRINTF3(lvl,fmt,a,b,c)              (0)
+#  define CPRINTF4(lvl,fmt,a,b,c,d)            (0)
+#  define CPRINTF5(lvl,fmt,a,b,c,d,e)          (0)
+#  define CPRINTF6(lvl,fmt,a,b,c,d,e,f)                (0)
+#  define CPRINTF7(lvl,fmt,a,b,c,d,e,f,g)      (0)
+#  define CPRINTF8(lvl,fmt,a,b,c,d,e,f,g,h)    (0)
+#  define CPRINTF9(lvl,fmt,a,b,c,d,e,f,g,h,i)  (0)
+
+#endif /* DEBUG */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* _ELAN3_EPDEBUG_H */
+
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms_asm_elan4_thread.S
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms_asm_elan4_thread.S   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms_asm_elan4_thread.S        2005-06-01 23:12:54.637433328 -0400
@@ -0,0 +1,133 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcomms_asm_elan4_thread.S,v 1.5 2004/04/25 11:25:43 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms_asm_elan4_thread.S,v $*/
+
+#include <elan4/events.h>
+#include <elan4/commands.h>
+
+#include "assym_elan4.h"
+
+/* XXXXX - registers.h */
+#define E4_MAIN_INT_SHIFT              14
+
+/*
+ * c_waitevent_interrupt (E4_uint64 *commandport, E4_Event *event, E4_uint64 count, E4_uint64 intcookie)
+ */
+       .global c_waitevent_interrupt
+c_waitevent_interrupt:
+       add             %sp, -192, %sp
+       st64            %r16, [%sp + 64]                // preserve call preserved registers
+       st64            %r24, [%sp + 128]               // - see CALL_USED_REGISTERS.
+       mov             %r16,%r16                       // BUG FIX: E4 RevA
+       mov             %r24,%r24                       // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+
+       mov             %r7, %r18                       // (%r2) return pc
+1:     call            2f
+        mov            %sp, %r17                       // (%r1) SP
+2:     add             %r7, (3f-1b), %r16              // (%r0) PC
+       st32            %r16, [%sp]                     // event source block
+       mov             MAKE_EXT_CLEAN_CMD, %r23
+       st8             %r23, [%sp+56]                  // event source block
+       mov             %r16,%r16                       // BUG FIX: E4 RevA
+       mov             %r23,%r23                       // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       
+       or              %r9, WAIT_EVENT_CMD, %r16                                               ! WAIT_EVENT_CMD | event
+       sll8            %r10, 32, %r17
+       or              %r17, E4_EVENT_TYPE_VALUE(E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, 8), %r17  !   ev_CountAndType
+       mov             %sp, %r18                                                               !   ev_Source
+       mov             %r8, %r19                                                               !   ev_Dest
+       sll8            %r11, E4_MAIN_INT_SHIFT, %r20
+       or              %r20, INTERRUPT_CMD, %r20                                               ! INTERRUPT_CMD | (cookie << E4_MAIN_INT_SHIFT)
+       mov             NOP_CMD, %r21
+       mov             NOP_CMD, %r22
+       mov             NOP_CMD, %r23
+
+       st64suspend     %r16, [%r8]
+       
+3:     ld64            [%sp + 64], %r16                // restore call preserved register
+       ld64            [%sp + 128], %r24
+       jmpl            %r2+8, %r0                      // and return
+        add            %sp, 192, %sp
+
+
+#define EP4_RCVR_PENDING_STALLED               1               /* indicates thread has stalled for no descriptor (rcvr_pending_head) */
+
+#define RXD_DEBUG(VAL,RXD,TMP) \
+       mov     VAL, TMP; \
+       st8     TMP, [RXD + EP4_RXD_DEBUG]
+
+       
+       /*
+        * %r2  - rcvr elan
+        * %r3  - rxd elan
+        */
+       .global c_queue_rxd
+c_queue_rxd:
+       RXD_DEBUG(1, %r3, %r23)
+       
+       ld16    [%r2 + EP4_RCVR_PENDING_TAILP], %r18    /* r18 == tailp, r19 = head */
+       add     %r3, EP4_RXD_NEXT, %r4
+       
+       st8     %r0, [%r3 + EP4_RXD_NEXT]               /* rxd->rxd_next = NULL */
+       st8     %r4, [%r2 + EP4_RCVR_PENDING_TAILP]     /* tailp = &rxd->rxd_next */
+       st8     %r3, [%r18]                             /* *tailp = rxd */
+
+       cmp     %r19, EP4_RCVR_PENDING_STALLED          /* thread stalled ? */
+       beq     1f
+        mov    %r18, %r16                              /* must have used %r16, %r19, %r23 */
+       mov     %r3, %r23
+
+       RXD_DEBUG(2, %r3, %r23)
+       
+       st8suspend %r16, [%r3 + EP4_RXD_QUEUED]         /* no - mark as queued - all done */
+
+1:     st8     %r16, [%r3 + EP4_RXD_QUEUED]            /* mark as queued */
+
+       RXD_DEBUG(3, %r3, %r23)
+
+       mov     %r3, %r8                                /* return rxd from c_stall_thread */
+       ba      .epcomms_resume_thread                  /* resume the thread */
+        ld64   [%r2 + EP4_RCVR_THREAD_STALL], %r0
+
+       /*
+        *  c_stall_thread (EP4_RCVR_ELAN *rcvrElan)
+        */
+       .global c_stall_thread
+c_stall_thread:
+       add             %sp, -192, %sp
+       st64            %r16, [%sp + 64]                // preserve call preserved registers
+       st64            %r24, [%sp + 128]               // - see CALL_USED_REGISTERS.
+       mov             %r16,%r16                       // BUG FIX: E4 RevA
+       mov             %r24,%r24                       // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+       nop                                             // BUG FIX: E4 RevA
+
+       mov             EP4_RCVR_PENDING_STALLED, %r9   // Mark rcvr as stalled
+       st8             %r9, [%r8 + EP4_RCVR_PENDING_HEAD]
+
+       // XXXX _ TBD should generate interrupt
+
+       mov             %r1, %r17                       // SP 
+       mov             %r7, %r23                       // return pc
+
+       st64suspend     %r16, [%r8 + EP4_RCVR_THREAD_STALL]
+       
+.epcomms_resume_thread:
+       /* %r8 == rxdElan */
+       
+       ld64            [%sp + 64], %r16                // restore call preserved register
+       ld64            [%sp + 128], %r24
+       jmpl            %r7+8, %r0                      // and return
+        add            %sp, 192, %sp
+
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms.c 2005-06-01 23:12:54.637433328 -0400
@@ -0,0 +1,484 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcomms.c,v 1.71.2.6 2004/11/30 12:02:16 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+#include <qsnet/autoconf.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+#include "cm.h"
+#include "debug.h"
+
+static void
+ep_comms_thread (void *arg)
+{
+    EP_COMMS_SUBSYS  *subsys = (EP_COMMS_SUBSYS *) arg;
+    struct list_head *el;
+
+    kernel_thread_init ("ep_comms");
+
+    /* since ep_alloc_xmtr() has incremented the module use count,
+     * we would be preventing the module from being unloaded, so
+     * we decrement the use count since this thread must terminate
+     * during unload of the module.
+     */
+    ep_mod_dec_usecount();
+
+    for (;;)
+    {
+       long nextRunTime = 0;
+
+       /* NOTE - subsys->Lock serializes us against flush/relocations
+        *        caused by rail nodeset transitions.
+        */
+       kmutex_lock (&subsys->Lock);
+       list_for_each (el, &subsys->Transmitters) {
+           nextRunTime = ep_check_xmtr (list_entry (el, EP_XMTR, Link), nextRunTime);
+       }
+
+       list_for_each (el, &subsys->Receivers) {
+           nextRunTime = ep_check_rcvr (list_entry (el, EP_RCVR, Link), nextRunTime);
+       }
+       kmutex_unlock (&subsys->Lock);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+       ep_csum_rxds (subsys);  
+#endif
+       nextRunTime = ep_forward_rxds (subsys, nextRunTime);
+
+       if (ep_kthread_sleep (&subsys->Thread, nextRunTime) < 0)
+           break;
+    }
+
+    ep_mod_inc_usecount();
+
+    ep_kthread_stopped (&subsys->Thread);
+    kernel_thread_exit();
+}
+
+int
+ep_comms_add_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *rail)
+{
+    EP_COMMS_SUBSYS  *subsys = (EP_COMMS_SUBSYS *) s;
+    EP_COMMS_RAIL    *commsRail;
+    struct list_head *el;
+
+    printk ("%s: vendorid=%x deviceid=%x\n", rail->Name, rail->Devinfo.dev_vendor_id, rail->Devinfo.dev_device_id);
+
+    switch (rail->Devinfo.dev_device_id)
+    {
+#if defined(CONFIG_ELAN3) || defined(CONFIG_ELAN3_MODULE)
+    case PCI_DEVICE_ID_ELAN3:
+       commsRail = ep3comms_add_rail (s, sys, rail);
+       break;
+#endif
+#if defined(CONFIG_ELAN4) || defined(CONFIG_ELAN4_MODULE)
+    case PCI_DEVICE_ID_ELAN4:
+       commsRail = ep4comms_add_rail (s, sys, rail);
+       break;
+#endif
+    default:
+       return 0;
+    }
+
+    if (commsRail == NULL)
+       return 1;
+
+    commsRail->Rail   = rail;
+    commsRail->Subsys = subsys;
+
+    kmutex_lock (&subsys->Lock);
+    list_add_tail (&commsRail->Link, &subsys->Rails);
+    
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+
+       EP_RAIL_OP (commsRail, Rcvr.AddRail) (rcvr, commsRail);
+    }
+       
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+
+       EP_RAIL_OP (commsRail, Xmtr.AddRail) (xmtr, commsRail);
+    }
+
+    kmutex_unlock (&subsys->Lock);
+
+    return 0;
+}
+
+void
+ep_comms_del_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *rail)
+{
+    EP_COMMS_SUBSYS  *subsys    = (EP_COMMS_SUBSYS *) s;
+    EP_COMMS_RAIL    *commsRail = NULL;
+    struct list_head *el;
+
+    kmutex_lock (&subsys->Lock);
+    /* find out rail entry and remove from system list */
+    list_for_each (el, &subsys->Rails) {
+       if ((commsRail = list_entry (el, EP_COMMS_RAIL, Link))->Rail == rail)
+           break;
+    }
+
+    list_del (&commsRail->Link);
+    
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+
+       EP_RAIL_OP(commsRail, Rcvr.DelRail) (rcvr, commsRail);
+    }
+       
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+
+       EP_RAIL_OP(commsRail,Xmtr.DelRail) (xmtr, commsRail);
+    }
+
+    kmutex_unlock (&subsys->Lock);
+
+    EP_RAIL_OP (commsRail, DelRail) (commsRail);
+}
+
+void
+ep_comms_fini (EP_SUBSYS *s, EP_SYS *sys)
+{
+    EP_COMMS_SUBSYS *subsys = (EP_COMMS_SUBSYS *) s;
+
+    ep_kthread_stop (&subsys->Thread);
+    ep_kthread_destroy (&subsys->Thread);
+
+    if (subsys->ForwardXmtr)
+       ep_free_xmtr (subsys->ForwardXmtr);
+
+    spin_lock_destroy (&subsys->ForwardDescLock);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    spin_lock_destroy (&subsys->CheckSumDescLock);
+#endif
+
+    kmutex_destroy (&subsys->Lock);
+
+    KMEM_FREE (subsys, sizeof (EP_COMMS_SUBSYS));
+}
+
+int
+ep_comms_init (EP_SYS *sys)
+{
+    EP_COMMS_SUBSYS *subsys;
+
+    KMEM_ZALLOC (subsys, EP_COMMS_SUBSYS *, sizeof (EP_COMMS_SUBSYS), 1);
+
+    if (subsys == NULL)
+       return (ENOMEM);
+
+    INIT_LIST_HEAD (&subsys->Rails);
+    INIT_LIST_HEAD (&subsys->Receivers);
+    INIT_LIST_HEAD (&subsys->Transmitters);
+    INIT_LIST_HEAD (&subsys->ForwardDescList);
+
+    kmutex_init (&subsys->Lock);
+    spin_lock_init (&subsys->ForwardDescLock);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    INIT_LIST_HEAD (&subsys->CheckSumDescList);
+    spin_lock_init (&subsys->CheckSumDescLock);
+#endif
+
+    subsys->Subsys.Sys        = sys;
+    subsys->Subsys.Name              = "epcomms";
+    subsys->Subsys.Destroy    = ep_comms_fini;
+    subsys->Subsys.AddRail    = ep_comms_add_rail;
+    subsys->Subsys.RemoveRail = ep_comms_del_rail;
+
+    ep_subsys_add (sys, &subsys->Subsys);
+    ep_kthread_init (&subsys->Thread);
+
+    if ((subsys->ForwardXmtr = ep_alloc_xmtr (subsys->Subsys.Sys)) == NULL)
+       goto failed;
+
+    if (kernel_thread_create (ep_comms_thread, subsys) == NULL)
+       goto failed;
+    ep_kthread_started (&subsys->Thread);
+
+    return (0);
+
+ failed:
+    ep_subsys_del (sys, &subsys->Subsys);
+    ep_comms_fini (&subsys->Subsys, sys);
+
+    return (ENOMEM);
+}
+
+void
+ep_comms_display (EP_SYS *sys, char *how)
+{
+    EP_COMMS_SUBSYS  *subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (sys, EPCOMMS_SUBSYS_NAME);
+    struct list_head *el;
+
+    if (how == NULL || !strncmp (how, "rail", 4))
+    {
+       kmutex_lock (&subsys->Lock);
+       list_for_each (el, &subsys->Rails) {
+           EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+           EP_RAIL_OP(commsRail, DisplayRail) (commsRail);
+       }
+       kmutex_unlock (&subsys->Lock);
+    }
+           
+    if (how == NULL || !strncmp (how, "xmtr", 4))
+       list_for_each (el, &subsys->Transmitters)
+           ep_display_xmtr (&di_ep_debug, list_entry (el, EP_XMTR, Link));
+
+    if (how == NULL || !strncmp (how, "rcvr", 4)) 
+       list_for_each (el, &subsys->Receivers)
+           ep_display_rcvr (&di_ep_debug, list_entry (el, EP_RCVR, Link), (how && how[4] == ',') ? 1 : 0);
+}
+
+int
+ep_svc_indicator_set (EP_SYS *epsys, int svc_indicator) 
+{
+    EP_COMMS_SUBSYS  *subsys;
+    struct list_head *el;
+
+    EPRINTF1 (DBG_SVC,"ep_svc_indicator_set: %d \n",svc_indicator);
+
+    if (svc_indicator < 0 || svc_indicator > EP_SVC_NUM_INDICATORS)
+       return (EP_EINVAL);
+
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (epsys, "epcomms")) == NULL) {
+       EPRINTF0 (DBG_SVC,"ep_svc_indicator_set: ep_subsys_find failed\n");
+       return (EP_EINVAL);
+    }
+
+
+    kmutex_lock (&subsys->Lock); /* walking rails list and setting info on Rail */
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       cm_svc_indicator_set(commsRail->Rail, svc_indicator);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    EPRINTF1 (DBG_SVC,"ep_svc_indicator_set: %d success\n",svc_indicator);
+    return (EP_SUCCESS);
+}
+
+int
+ep_svc_indicator_clear (EP_SYS *epsys, int svc_indicator) 
+{
+    EP_COMMS_SUBSYS  *subsys;
+    struct list_head *el;
+
+    EPRINTF1 (DBG_SVC,"ep_svc_indicator_clear: %d \n",svc_indicator);
+
+    if (svc_indicator < 0 || svc_indicator >= EP_SVC_NUM_INDICATORS)
+       return (EP_EINVAL);
+
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (epsys, "epcomms")) == NULL) {
+       EPRINTF0 (DBG_SVC,"ep_svc_indicator_clear: ep_subsys_find failed\n");
+       return (EP_EINVAL);
+    }
+
+    kmutex_lock (&subsys->Lock); /* walking rails list and setting info on Rail */
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       cm_svc_indicator_clear(commsRail->Rail, svc_indicator);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    EPRINTF1 (DBG_SVC,"ep_svc_indicator_clear: %d success\n",svc_indicator);
+    return (EP_SUCCESS);
+}
+
+int 
+ep_svc_indicator_is_set (EP_SYS *epsys, int svc_indicator, int nodeId) 
+{
+    EP_COMMS_SUBSYS  *subsys;
+    struct list_head *el;
+    int               set = 0;
+
+    EPRINTF2 (DBG_SVC,"ep_svc_indicator_is_set: svc %d node %d \n", svc_indicator, nodeId);
+
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (epsys, "epcomms")) == NULL) {
+       EPRINTF0 (DBG_SVC,"ep_svc_indicator_is_set: ep_subsys_find failed\n");
+       return (0);
+    }
+
+    kmutex_lock (&subsys->Lock); /* walking rails list and setting info on Rail */
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       set |= cm_svc_indicator_is_set(commsRail->Rail, svc_indicator, nodeId);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    EPRINTF3 (DBG_SVC,"ep_svc_indicator_is_set: svc %d node %d returning %d\n", svc_indicator, nodeId, set);
+    return set;
+}
+
+int
+ep_svc_indicator_bitmap (EP_SYS *epsys, int svc_indicator, bitmap_t * bitmap, int low, int nnodes) 
+{
+    EP_COMMS_SUBSYS  *subsys;
+    struct list_head *el;
+
+    EPRINTF1 (DBG_SVC,"ep_svc_indicator_bitmap: svc %d\n", svc_indicator);
+
+    if (svc_indicator < 0 || svc_indicator >= EP_SVC_NUM_INDICATORS)
+       return (-1);
+
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (epsys, "epcomms")) == NULL) {
+       EPRINTF0 (DBG_SVC,"ep_svc_indicator_bitmap: ep_subsys_find failed\n");
+       return (-2);
+    }
+
+    /* clear bitmap */
+    bt_zero (bitmap, nnodes);
+
+    kmutex_lock (&subsys->Lock); /* walking rails list and setting info on Rail */
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       /* this will or in each bit map */
+       cm_svc_indicator_bitmap (commsRail->Rail, svc_indicator, bitmap, low, nnodes);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    return (0);
+}
+
+int
+ep_xmtr_svc_indicator_bitmap (EP_XMTR *xmtr, int svc_indicator, bitmap_t * bitmap, int low, int nnodes) 
+{
+    int i;
+
+    EPRINTF1 (DBG_SVC,"ep_xmtr_svc_indicator_bitmap: svc %d\n", svc_indicator);
+
+    if (svc_indicator < 0 || svc_indicator >= EP_SVC_NUM_INDICATORS)
+       return (-1);
+
+    /* clear bitmap */
+    bt_zero (bitmap, nnodes);
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+    {
+       if (xmtr->RailMask & (1 << i) )
+       {
+           /* this will or in each bit map */
+           cm_svc_indicator_bitmap (xmtr->Rails[i]->CommsRail->Rail, svc_indicator, bitmap, low, nnodes);
+       }
+    }
+
+    return (0);
+}
+
+EP_RAILMASK
+ep_svc_indicator_railmask (EP_SYS *epsys, int svc_indicator, int nodeId)
+{
+    EP_COMMS_SUBSYS  *subsys;
+    struct list_head *el;
+    EP_RAILMASK       rmask=0;
+
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (epsys, "epcomms")) == NULL)
+       return (rmask);
+
+    kmutex_lock (&subsys->Lock); /* walking rails list and reading info from Rail */
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       if ( cm_svc_indicator_is_set(commsRail->Rail, svc_indicator,nodeId))
+            rmask |= EP_RAIL2RAILMASK(commsRail->Rail->Number);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    return (rmask);
+}
+
+EP_RAILMASK
+ep_xmtr_svc_indicator_railmask (EP_XMTR *xmtr, int svc_indicator, int nodeId)
+{
+    EP_RAILMASK    rmask=0;
+    EP_COMMS_RAIL *commsRail;
+    int            i;
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+    {
+       if (xmtr->RailMask & (1 << i) )
+       {
+           commsRail = xmtr->Rails[i]->CommsRail;
+
+           if ( cm_svc_indicator_is_set(commsRail->Rail, svc_indicator,nodeId))
+               rmask |= EP_RAIL2RAILMASK(commsRail->Rail->Number);
+       }
+    }   
+    
+    EPRINTF3 (DBG_SVC, "ep_xmtr_svc_indicator_railmask: svc %d node %d mask 0x%x\n",  svc_indicator, nodeId, rmask);
+
+    return (rmask);
+}
+
+EP_RAILMASK
+ep_rcvr_railmask (EP_SYS *epsys, EP_SERVICE service)
+{
+    EP_COMMS_SUBSYS  *subsys;
+    EP_RAILMASK       rmask=0;
+    struct list_head *el;
+    
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (epsys, "epcomms")) == NULL)
+       return (rmask);
+    
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+ 
+       if (rcvr->Service == service)
+           rmask |= rcvr->RailMask; 
+    }
+    kmutex_unlock(&subsys->Lock);
+
+    return (rmask);
+}
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+uint32_t
+ep_calc_check_sum (EP_SYS *sys, EP_ENVELOPE *env, EP_NMD *nmd, int nFrags)
+{
+    EP_NMH   *nmh;
+    int       i;
+    uint16_t  check_data = 0;
+    uint16_t  check_env  = 0;
+
+    for (i = 0; i < nFrags; i++) {
+       /* find the nmh for this frag */
+       nmh = ep_nmh_find (&sys->MappingTable, &nmd[i]);
+
+       ASSERT( nmh != NULL);
+
+       /* add the next frag to the check sum */
+       check_data = nmh->nmh_ops->op_calc_check_sum (sys, nmh, &nmd[i], check_data);
+    }
+
+    check_env = rolling_check_sum ((char *) env, offsetof(EP_ENVELOPE, CheckSum), 0);
+
+    return (EP_ENVELOPE_CHECK_SUM | ( (check_env & 0x7FFF) << 16) | (check_data & 0xFFFF));
+}
+#endif
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms_elan3.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan3.c   2005-06-01 23:12:54.638433176 -0400
@@ -0,0 +1,191 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcomms_elan3.c,v 1.60 2004/08/03 11:34:34 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "kcomm_elan3.h"
+#include "epcomms_elan3.h"
+
+void
+ep3comms_flush_callback (void *arg, statemap_t *map)
+{
+    EP_COMMS_RAIL    *commsRail = (EP_COMMS_RAIL *) arg;
+    EP_COMMS_SUBSYS  *subsys    = commsRail->Subsys;
+    struct list_head *el;
+
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[commsRail->Rail->Number])
+           ep3xmtr_flush_callback (xmtr, (EP3_XMTR_RAIL *) xmtr->Rails[commsRail->Rail->Number]);
+    }
+
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[commsRail->Rail->Number])
+           ep3rcvr_flush_callback (rcvr, (EP3_RCVR_RAIL *) rcvr->Rails[commsRail->Rail->Number]);
+    }
+    kmutex_unlock (&subsys->Lock);
+}
+
+void
+ep3comms_failover_callback (void *arg, statemap_t *map)
+{
+    EP_COMMS_RAIL    *commsRail = (EP_COMMS_RAIL *) arg;
+    EP_COMMS_SUBSYS  *subsys    = commsRail->Subsys;
+    struct list_head *el;
+
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[commsRail->Rail->Number])
+           ep3xmtr_failover_callback (xmtr, (EP3_XMTR_RAIL *) xmtr->Rails[commsRail->Rail->Number]);
+    }
+
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[commsRail->Rail->Number])
+           ep3rcvr_failover_callback (rcvr, (EP3_RCVR_RAIL *) rcvr->Rails[commsRail->Rail->Number]);
+    }
+    kmutex_unlock (&subsys->Lock);
+}
+
+void
+ep3comms_disconnect_callback (void *arg, statemap_t *map)
+{
+    EP_COMMS_RAIL    *commsRail = (EP_COMMS_RAIL *) arg;
+    EP_COMMS_SUBSYS  *subsys    = commsRail->Subsys;
+    struct list_head *el;
+
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[commsRail->Rail->Number])
+           ep3xmtr_disconnect_callback (xmtr, (EP3_XMTR_RAIL *) xmtr->Rails[commsRail->Rail->Number]);
+    }
+
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[commsRail->Rail->Number])
+           ep3rcvr_disconnect_callback (rcvr, (EP3_RCVR_RAIL *) rcvr->Rails[commsRail->Rail->Number]);
+    }
+    kmutex_unlock (&subsys->Lock);
+}
+
+EP_COMMS_RAIL *
+ep3comms_add_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *r)
+{
+    EP3_RAIL         *rail   = (EP3_RAIL *) r;
+    ELAN3_DEV        *dev    = rail->Device;
+    EP3_COMMS_RAIL   *commsRail;
+    EP3_InputQueue    qdesc;
+    int i;
+
+    KMEM_ZALLOC (commsRail, EP3_COMMS_RAIL *, sizeof (EP3_COMMS_RAIL), TRUE);
+
+    if (commsRail == NULL)
+       return NULL;
+    
+    commsRail->Generic.Ops.DelRail          = ep3comms_del_rail;
+    commsRail->Generic.Ops.DisplayRail      = ep3comms_display_rail;
+    commsRail->Generic.Ops.Rcvr.AddRail     = ep3rcvr_add_rail;
+    commsRail->Generic.Ops.Rcvr.DelRail     = ep3rcvr_del_rail;
+    commsRail->Generic.Ops.Rcvr.Check       = ep3rcvr_check;
+    commsRail->Generic.Ops.Rcvr.QueueRxd    = ep3rcvr_queue_rxd;
+    commsRail->Generic.Ops.Rcvr.RpcPut      = ep3rcvr_rpc_put;
+    commsRail->Generic.Ops.Rcvr.RpcGet      = ep3rcvr_rpc_get;
+    commsRail->Generic.Ops.Rcvr.RpcComplete = ep3rcvr_rpc_complete;
+
+    commsRail->Generic.Ops.Rcvr.StealRxd    = ep3rcvr_steal_rxd;
+
+    commsRail->Generic.Ops.Rcvr.FillOutRailStats = ep3rcvr_fillout_rail_stats;
+
+    commsRail->Generic.Ops.Rcvr.DisplayRcvr = ep3rcvr_display_rcvr;
+    commsRail->Generic.Ops.Rcvr.DisplayRxd  = ep3rcvr_display_rxd;
+
+    commsRail->Generic.Ops.Xmtr.AddRail     = ep3xmtr_add_rail;
+    commsRail->Generic.Ops.Xmtr.DelRail     = ep3xmtr_del_rail;
+    commsRail->Generic.Ops.Xmtr.Check       = ep3xmtr_check;
+    commsRail->Generic.Ops.Xmtr.BindTxd     = ep3xmtr_bind_txd;
+    commsRail->Generic.Ops.Xmtr.UnbindTxd   = ep3xmtr_unbind_txd;
+    commsRail->Generic.Ops.Xmtr.PollTxd     = ep3xmtr_poll_txd;
+    commsRail->Generic.Ops.Xmtr.CheckTxdState = ep3xmtr_check_txd_state;
+
+    commsRail->Generic.Ops.Xmtr.DisplayXmtr = ep3xmtr_display_xmtr;
+    commsRail->Generic.Ops.Xmtr.DisplayTxd  = ep3xmtr_display_txd;
+
+    commsRail->Generic.Ops.Xmtr.FillOutRailStats = ep3xmtr_fillout_rail_stats;
+
+    /* Allocate the input queues at their fixed elan address */
+    if (! (commsRail->QueueDescs = ep_alloc_memory_elan (r, EP_EPCOMMS_QUEUE_BASE, roundup (EP_MSG_NSVC * sizeof (EP3_InputQueue), PAGESIZE), EP_PERM_ALL, 0)))
+    {
+       KMEM_FREE (commsRail, sizeof (EP3_COMMS_RAIL));
+       return NULL;
+    }
+
+    qdesc.q_state          = E3_QUEUE_FULL;
+    qdesc.q_base           = 0;
+    qdesc.q_top            = 0;
+    qdesc.q_fptr           = 0;
+    qdesc.q_bptr           = 0;
+    qdesc.q_size           = 0;
+    qdesc.q_event.ev_Count = 0;
+    qdesc.q_event.ev_Type  = 0;
+
+    /* Initialise all queue entries to be full */
+    for (i = 0; i < EP_MSG_NSVC; i++)
+       elan3_sdram_copyl_to_sdram (dev, &qdesc, commsRail->QueueDescs + (i * sizeof (EP3_InputQueue)), sizeof (EP3_InputQueue));
+
+    ep_register_callback (r, EP_CB_FLUSH_FILTERING, ep3comms_flush_callback,      commsRail);
+    ep_register_callback (r, EP_CB_FLUSH_FLUSHING,  ep3comms_flush_callback,      commsRail);
+    ep_register_callback (r, EP_CB_FAILOVER,        ep3comms_failover_callback,   commsRail);
+    ep_register_callback (r, EP_CB_DISCONNECTING,   ep3comms_disconnect_callback, commsRail);
+
+    return (EP_COMMS_RAIL *) commsRail;
+}
+
+void
+ep3comms_del_rail (EP_COMMS_RAIL *r)
+{
+    EP3_COMMS_RAIL *commsRail = (EP3_COMMS_RAIL *) r;
+    EP_RAIL        *rail      = commsRail->Generic.Rail;
+
+    ep_remove_callback (rail, EP_CB_FLUSH_FILTERING, ep3comms_flush_callback,      commsRail);
+    ep_remove_callback (rail, EP_CB_FLUSH_FLUSHING,  ep3comms_flush_callback,      commsRail);
+    ep_remove_callback (rail, EP_CB_FAILOVER,        ep3comms_failover_callback,   commsRail);
+    ep_remove_callback (rail, EP_CB_DISCONNECTING,   ep3comms_disconnect_callback, commsRail);
+
+    ep_free_memory_elan (rail, EP_EPCOMMS_QUEUE_BASE);
+
+    KMEM_FREE (commsRail, sizeof (EP3_COMMS_RAIL));
+}
+
+void
+ep3comms_display_rail (EP_COMMS_RAIL *r)
+{
+    
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan3.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms_elan3.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan3.h   2005-06-01 23:12:54.639433024 -0400
@@ -0,0 +1,330 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __EPCOMMS_ELAN3_H
+#define __EPCOMMS_ELAN3_H
+
+#ident "@(#)$Id: epcomms_elan3.h,v 1.27.2.1 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms_elan3.h,v $ */
+
+#define EP3_DMAFAILCOUNT               3
+
+
+/* Main/Elan spinlock */
+typedef struct ep3_spinlock_elan 
+{
+    volatile E3_uint32 sl_lock;                /* main wants a lock */
+    volatile E3_uint32 sl_seq;                 /* thread owns this word */
+    /* NOTE: The lock/seq words must be within the same 32-byte Elan cache-line */
+    E3_uint64          sl_pad[14];             /* pad to 64-bytes */
+} EP3_SPINLOCK_ELAN;
+
+/* Declare this as a main memory cache block for efficiency */
+typedef struct ep3_spinlock_main {
+    volatile E3_uint32 sl_seq;                 /* copy of seq number updated by Elan */
+    volatile E3_uint32 sl_pad[15];             /* pad to 64-bytes */
+} EP3_SPINLOCK_MAIN;
+
+#if defined (__ELAN3__)
+
+extern void ep3_spinblock (EP3_SPINLOCK_ELAN *, EP3_SPINLOCK_MAIN *);
+
+#define EP3_SPINENTER(SLE,SL) \
+do {\
+       (SLE)->sl_seq++; \
+       if ((SLE)->sl_lock) \
+               ep3_spinblock(SLE, SL);\
+} while (0)
+
+#define EP3_SPINEXIT(SLE,SL) \
+do {\
+       (SL)->sl_seq = (SLE)->sl_seq;\
+} while (0)
+
+#else
+
+#define EP3_SPINENTER(DEV,SLE,SL) do { \
+    E3_uint32 seq; \
+\
+    mb();\
+    elan3_sdram_writel (DEV, (SLE) + offsetof (EP3_SPINLOCK_ELAN, sl_lock), 1);\
+    mb();\
+    seq = elan3_sdram_readl (DEV, (SLE) + offsetof (EP3_SPINLOCK_ELAN, sl_seq));\
+    while (seq != (SL)->sl_seq)\
+    {\
+       while ((SL)->sl_seq == (seq - 1))\
+       {\
+           mb();\
+\
+           DELAY (1); \
+       }\
+       seq = elan3_sdram_readl (DEV, (SLE) + offsetof (EP3_SPINLOCK_ELAN, sl_seq));\
+    }\
+} while (0)
+
+#define EP3_SPINEXIT(DEV,SLE,SL) do { \
+       wmb(); \
+       elan3_sdram_writel (DEV, (SLE) + offsetof (EP3_SPINLOCK_ELAN, sl_lock), 0);\
+       mmiob(); \
+} while (0)
+
+#endif /* ! __ELAN3__ */
+
+/* per-rail elan memory portion receive descriptor */
+typedef struct ep3_rxd_rail_elan
+{
+    E3_DMA             Dmas[EP_MAXFRAG+1];                     /* Dma's for fetching data/putting data & status blk */
+    E3_Event           ChainEvent[EP_MAXFRAG];                 /* Events to chain dmas */
+    E3_BlockCopyEvent  DataEvent;                              /* message received block event */
+    E3_BlockCopyEvent  DoneEvent;                              /* RPC status block event */
+    
+    EP_NMD             Data;                                   /* Network mapping handle for receive data */
+
+    E3_Addr            RxdMain;                                /* pointer to main memory portion */
+
+    E3_Addr            Next;                                   /* linked list when on pending list (elan address) */
+
+    E3_uint64          MainAddr;                               /* kernel address of ep_rxd_main */
+} EP3_RXD_RAIL_ELAN;
+
+#define EP3_RXD_RAIL_ELAN_SIZE roundup (sizeof (EP3_RXD_RAIL_ELAN), E3_DMA_ALIGN)
+
+/* per-rail main memory portion of receive descriptor */
+typedef struct ep3_rxd_rail_main
+{
+    E3_uint32          DataEvent;                              /* dest for done event */
+    E3_uint32          DoneEvent;                              /* dest for done event */
+} EP3_RXD_RAIL_MAIN;
+
+#define EP3_RXD_RAIL_MAIN_SIZE roundup (sizeof(EP3_RXD_RAIL_MAIN), sizeof (E3_uint32))
+
+#if !defined(__ELAN3__)
+/* Kernel memory portion of per-rail receive descriptor */
+typedef struct ep3_rxd_rail
+{
+    EP_RXD_RAIL                Generic;                                /* generic rxd rail */
+
+    EP3_COOKIE         DataCookie;                             /* Event cookie */
+    EP3_COOKIE         DoneCookie;                             /* Event cookie */
+    EP3_COOKIE         ChainCookie[EP_MAXFRAG];                /* Event cookie */
+
+    sdramaddr_t                RxdElan;                                /* per-rail elan receive descriptor */
+    E3_Addr            RxdElanAddr;                            /*   and elan address */
+
+    EP3_RXD_RAIL_MAIN  *RxdMain;                               /* per-rail main receive descriptor */
+    E3_Addr            RxdMainAddr;                            /*   and elan address */
+
+    EP_BACKOFF         Backoff;                                /* dma backoff */
+} EP3_RXD_RAIL;
+
+#define EP3_NUM_RXD_PER_BLOCK  16
+
+typedef struct ep3_rxd_rail_block
+{
+    struct list_head   Link;
+    
+    EP3_RXD_RAIL        Rxd[EP3_NUM_RXD_PER_BLOCK];
+} EP3_RXD_RAIL_BLOCK;
+
+#endif /* ! __ELAN3__ */
+
+typedef struct ep3_rcvr_rail_elan                              /* Elan memory service structure */
+{
+    EP3_SPINLOCK_ELAN  ThreadLock;                             /* elan memory portion of spin lock */
+    EP3_SPINLOCK_ELAN  PendingLock;                            /* spin lock for pending rx list */
+
+    E3_Addr           PendingDescs;                            /* list of pending receive descriptors */
+    E3_uint32          ThreadShouldHalt;                        /* marks that the thread should halt */
+
+    E3_uint64         MainAddr;                                /* kernel address of ep_rcvr (for StallThreadForNoDescs)*/
+} EP3_RCVR_RAIL_ELAN;
+
+typedef struct ep3_rcvr_rail_main                              /* Main memory service strucure */
+{
+    EP3_SPINLOCK_MAIN  ThreadLock;                             /* main memory portion of spin lock */
+    EP3_SPINLOCK_MAIN  PendingLock;                            /* spinlock for pending rx list */
+
+    volatile unsigned   PendingDescsTailp;                     /* next pointer of last receive descriptor on pending list */
+} EP3_RCVR_RAIL_MAIN;
+
+#if !defined(__ELAN3__)
+
+typedef struct ep3_rcvr_rail_stats
+{
+    unsigned long some_stat;
+} EP3_RCVR_RAIL_STATS;
+
+typedef struct ep3_rcvr_rail
+{
+    EP_RCVR_RAIL       Generic;                                /* generic portion */
+    
+    EP3_RCVR_RAIL_MAIN *RcvrMain;
+    E3_Addr            RcvrMainAddr;
+    sdramaddr_t         RcvrElan;
+    E3_Addr             RcvrElanAddr;
+
+    sdramaddr_t                InputQueueBase;                         /* base of receive queue */
+    E3_Addr            InputQueueAddr;                         /* elan address of receive queue */
+
+    E3_Addr            ThreadStack;                            /* Thread processor stack */
+    E3_Addr            ThreadWaiting;                          /* Elan thread is waiting as no receive descriptors pending (sp stored here ) */
+    E3_Addr            ThreadHalted;                           /* Elan thread is waiting as it was requested to halt */
+
+    struct list_head   FreeDescList;                           /* freelist of per-rail receive descriptors */
+    unsigned int       FreeDescCount;                          /*   and number on free list */
+    unsigned int        TotalDescCount;                                /*   total number created */
+    spinlock_t         FreeDescLock;                           /*   and lock for free list */
+    struct list_head    DescBlockList;                         /* list of receive descriptor blocks */
+
+    unsigned int        FreeDescWaiting;                       /* waiting for descriptors to be freed */
+    kcondvar_t         FreeDescSleep;                          /*   and sleep here */
+
+    unsigned int       CleanupWaiting;                         /* waiting for cleanup */
+    kcondvar_t         CleanupSleep;                           /*   and sleep here */
+
+    EP3_RCVR_RAIL_STATS stats;                                  /* elan3 specific rcvr_rail stats */
+} EP3_RCVR_RAIL;
+
+#endif /* ! __ELAN3__ */
+
+/* per-rail portion of transmit descriptor */
+typedef struct ep3_txd_rail_elan
+{
+    EP_ENVELOPE               Envelope;                                /* message envelope */
+    EP_PAYLOAD        Payload;                                 /* message payload */
+
+    E3_BlockCopyEvent EnveEvent;                               /* envelope event */
+    E3_BlockCopyEvent DataEvent;                               /* data transfer event */
+    E3_BlockCopyEvent DoneEvent;                               /* rpc done event */
+} EP3_TXD_RAIL_ELAN;
+
+#define EP3_TXD_RAIL_ELAN_SIZE roundup (sizeof (EP3_TXD_RAIL_ELAN), E3_BLK_ALIGN)
+
+typedef struct ep3_txd_rail_main
+{
+    E3_uint32         EnveEvent;                               /* dest for envelope event */
+    E3_uint32         DataEvent;                               /* dest for data transfer event */
+    E3_uint32        DoneEvent;                                /* dest for rpc done event */
+} EP3_TXD_RAIL_MAIN;
+
+#define EP3_TXD_RAIL_MAIN_SIZE roundup (sizeof(EP3_TXD_RAIL_MAIN), E3_BLK_ALIGN)
+
+#if !defined(__ELAN3__)
+
+typedef struct ep3_txd_rail
+{
+    EP_TXD_RAIL               Generic;                                 /* generic txd rail */
+
+    EP3_COOKIE        EnveCookie;                              /* Event cookies */
+    EP3_COOKIE        DataCookie;
+    EP3_COOKIE        DoneCookie;
+
+    sdramaddr_t               TxdElan;                                 /* Elan TX descriptor */
+    E3_Addr           TxdElanAddr;                             /*  and elan address */
+
+    EP3_TXD_RAIL_MAIN *TxdMain;                                        /* Elan Main memory tx descriptor */
+    E3_Addr           TxdMainAddr;                             /*  and elan address */
+
+    EP_BACKOFF        Backoff;                                 /* dma backoff */
+} EP3_TXD_RAIL;
+
+
+#define EP3_NUM_TXD_PER_BLOCK  16
+
+typedef struct ep3_txd_rail_block
+{
+    struct list_head   Link;
+    
+    EP3_TXD_RAIL       Txd[EP3_NUM_TXD_PER_BLOCK];
+} EP3_TXD_RAIL_BLOCK;
+
+typedef struct ep3_xmtr_rail_stats
+{
+    unsigned long some_stat;
+} EP3_XMTR_RAIL_STATS;
+
+typedef struct ep3_xmtr_rail
+{
+    EP_XMTR_RAIL       Generic;                                /* generic portion */
+
+    struct list_head   FreeDescList;                           /* freelist of per-rail receive descriptors */
+    unsigned int       FreeDescCount;                          /*   and number on free list */
+    unsigned int        TotalDescCount;
+    spinlock_t         FreeDescLock;                           /*   and lock for free list */
+    struct list_head    DescBlockList;                         /* list of receive descriptor blocks */
+
+    unsigned int        FreeDescWaiting;                       /* waiting for descriptors to be freed */
+    kcondvar_t          FreeDescSleep;                         /*   and sleep here */
+
+    EP3_XMTR_RAIL_STATS stats;                                  /* elan3 specific xmtr rail stats */
+} EP3_XMTR_RAIL;
+
+typedef struct ep3_comms_rail
+{
+    EP_COMMS_RAIL      Generic;                                /* generic comms rail */
+    sdramaddr_t                QueueDescs;                             /* input queue descriptors */
+} EP3_COMMS_RAIL;
+
+/* epcommxTx_elan3.c */
+extern void           ep3xmtr_flush_callback (EP_XMTR *xmtr, EP3_XMTR_RAIL *xmtrRail);
+extern void           ep3xmtr_failover_callback (EP_XMTR *xmtr, EP3_XMTR_RAIL *xmtrRail);
+extern void           ep3xmtr_disconnect_callback (EP_XMTR *xmtr, EP3_XMTR_RAIL *xmtrRail);
+
+/* epcommsRx_elan3.c */
+extern void          CompleteEnvelope (EP3_RAIL *rail, E3_Addr rxdMainAddr, E3_uint32 PAckVal);
+extern void           StallThreadForNoDescs (EP3_RAIL *rail, E3_Addr rcvrElanAddr, E3_Addr sp);
+extern void           StallThreadForHalted  (EP3_RAIL *rail, E3_Addr rcvrElanAddr, E3_Addr sp);
+
+extern void           ep3rcvr_flush_callback (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail);
+extern void           ep3rcvr_failover_callback (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail);
+extern void           ep3rcvr_disconnect_callback (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail);
+
+/* epcomms_elan3.c */
+extern EP_COMMS_RAIL *ep3comms_add_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *r);
+extern void           ep3comms_del_rail (EP_COMMS_RAIL *r);
+extern void           ep3comms_display_rail (EP_COMMS_RAIL *r);
+
+/* epcommsTx_elan3.c */
+extern int            ep3xmtr_bind_txd (EP_TXD *txd, EP_XMTR_RAIL *xmtrRail, unsigned int phase);
+extern void           ep3xmtr_unbind_txd (EP_TXD *txd, unsigned int phase);
+extern int            ep3xmtr_poll_txd (EP_XMTR_RAIL *xmtrRail, EP_TXD_RAIL *txdRail, int how);
+extern long           ep3xmtr_check (EP_XMTR_RAIL *xmtrRail, long nextRunTime);
+extern void           ep3xmtr_add_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail);
+extern void           ep3xmtr_del_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail);
+extern int            ep3xmtr_check_txd_state(EP_TXD *txd);
+
+extern void           ep3xmtr_display_xmtr (DisplayInfo *di, EP_XMTR_RAIL *xmtrRail);
+extern void           ep3xmtr_display_txd  (DisplayInfo *di, EP_TXD_RAIL *txdRail);
+
+extern void           ep3xmtr_fillout_rail_stats (EP_XMTR_RAIL *xmtr_rail, char *str);
+
+/* epcommsRx_elan3.c */
+extern int           ep3rcvr_queue_rxd (EP_RXD *rxd, EP_RCVR_RAIL *rcvrRail);
+extern void          ep3rcvr_rpc_put (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+extern void          ep3rcvr_rpc_get (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+extern void          ep3rcvr_rpc_complete (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+
+extern EP_RXD       *ep3rcvr_steal_rxd (EP_RCVR_RAIL *rcvrRail);
+
+extern long          ep3rcvr_check (EP_RCVR_RAIL *rcvrRail, long nextRunTime);
+extern void           ep3rcvr_add_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *rail);
+extern void           ep3rcvr_del_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *rail);
+
+extern void           ep3rcvr_display_rcvr (DisplayInfo *di, EP_RCVR_RAIL *rcvrRail);
+extern void           ep3rcvr_display_rxd  (DisplayInfo *di, EP_RXD_RAIL *rxdRail);
+
+extern void           ep3rcvr_fillout_rail_stats (EP_RCVR_RAIL *rcvr_rail, char *str);
+
+#endif /* !defined(__ELAN3__) */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __EPCOMMS_ELAN3_H */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan3_thread.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms_elan3_thread.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan3_thread.c    2005-06-01 23:12:54.640432872 -0400
@@ -0,0 +1,296 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcomms_elan3_thread.c,v 1.4 2004/01/20 11:03:15 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms_elan3_thread.c,v $ */
+
+//#include <qsnet/types.h>
+
+typedef char               int8_t;
+typedef unsigned char      uint8_t;
+typedef short              int16_t;
+typedef unsigned short     uint16_t;
+typedef int                int32_t;
+typedef unsigned int       uint32_t;
+typedef long long          int64_t;
+typedef unsigned long long uint64_t;
+
+#include <elan3/e3types.h>
+#include <elan3/events.h>
+#include <elan3/elanregs.h>
+#include <elan3/intrinsics.h>
+
+#include <elan/nmh.h>
+#include <elan/kcomm.h>
+#include <elan/epcomms.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "epcomms_elan3.h"
+
+#ifndef offsetof
+#define offsetof(s, m)                 (unsigned long)(&(((s *)0)->m))
+#endif
+
+EP3_RAIL_ELAN *rail;
+EP3_RCVR_RAIL_ELAN *r;
+EP3_RCVR_RAIL_MAIN *rm;
+
+void
+ep3comms_rcvr (EP3_RAIL_ELAN *rail, EP3_RCVR_RAIL_ELAN *rcvrElan, EP3_RCVR_RAIL_MAIN *rcvrMain, 
+             EP3_InputQueue *q, unsigned int *cookies)
+{
+    int           count = 1;
+    E3_Addr       nfptr = q->q_fptr + q->q_size;
+    E3_uint32     tmp;
+    int           i;
+    E3_Addr       buffer;
+    int                  len;
+    E3_DMA       *dma;
+    E3_Event     *event;
+
+    /* clear the queue state to allow envelopes to arrive */
+    q->q_state = 0;
+
+    for (;;)
+    {
+       if (! rcvrElan->ThreadShouldHalt)
+           c_waitevent ((E3_Event *) &q->q_event, count);                                              /* HALT POINT */
+
+       if (rcvrElan->ThreadShouldHalt && nfptr == q->q_bptr)
+       {
+           asm volatile ("mov %0, %%g1" : /* no outputs */ : "r" (rcvrElan));
+           asm volatile ("ta %0"        : /* no outputs */ : "i" (EP3_UNIMP_THREAD_HALTED));           /* HALT POINT */
+           continue;
+       }
+
+       count = 0;
+       do {
+           /* Process the message at nfptr */
+           EP_ENVELOPE      *env = (EP_ENVELOPE *) nfptr;
+           EP3_RXD_RAIL_ELAN *rxd;
+           int ack;
+           
+           EP3_SPINENTER(&rcvrElan->ThreadLock, &rcvrMain->ThreadLock);                                        /* HALT POINT */
+           
+           while ((rxd = (EP3_RXD_RAIL_ELAN *)rcvrElan->PendingDescs) == 0)
+           {
+               /* no receive descriptors, so trap to the kernel to wait
+                * for receive descriptor to be queued, we pass the rcvr
+                * in %g1, so that the trap handler can restart us. */
+               EP3_SPINEXIT(&rcvrElan->ThreadLock, &rcvrMain->ThreadLock);
+               asm volatile ("mov %0, %%g1" : /* no outputs */ : "r" (rcvrElan));
+               asm volatile ("ta %0"        : /* no outputs */ : "i" (EP3_UNIMP_TRAP_NO_DESCS));       /* HALT POINT */
+               EP3_SPINENTER(&rcvrElan->ThreadLock, &rcvrMain->ThreadLock);                            /* HALT POINT */
+           }
+
+           if (env->Version != EP_ENVELOPE_VERSION)
+           {
+               /* This envelope has been cancelled - so just consume it */
+               EP3_SPINEXIT(&rcvrElan->ThreadLock, &rcvrMain->ThreadLock);
+               goto consume_envelope;
+           }
+
+           dma   = rxd->Dmas;
+           event = rxd->ChainEvent;
+
+           if (EP_IS_MULTICAST(env->Attr))
+           {
+               dma->dma_type            = E3_DMA_TYPE (DMA_BYTE, DMA_READ, DMA_NORMAL, EP3_DMAFAILCOUNT);
+               dma->dma_size            = BT_BITOUL(EP_MAX_NODES) * sizeof (bitmap_t);
+               dma->dma_source          = env->TxdMain.nmd_addr + offsetof (EP_TXD_MAIN, Bitmap);
+               dma->dma_dest            = (E3_Addr) &((EP_RXD_MAIN *) rxd->RxdMain)->Bitmap;
+               dma->dma_destEvent       = (E3_Addr) event;
+               dma->dma_destCookieVProc = DMA_COOKIE_THREAD | DMA_COOKIE (cookies[env->NodeId], EP_VP_DATA (rail->NodeId));
+               dma->dma_srcEvent        = env->TxdRail + offsetof (EP3_TXD_RAIL_ELAN, DataEvent);
+               dma->dma_srcCookieVProc  = DMA_COOKIE_THREAD | DMA_REMOTE_COOKIE (cookies[env->NodeId], EP_VP_DATA (env->NodeId));
+               
+               event->ev_Count = 1;
+
+               dma++; event++;
+           }
+
+           if (env->nFrags == 0)
+           {
+               /* Generate a "get" DMA to accept the envelope and fire the rx handler */
+               dma->dma_type            = E3_DMA_TYPE(DMA_BYTE, DMA_READ, DMA_NORMAL, EP3_DMAFAILCOUNT);
+               dma->dma_size            = 0;
+               dma->dma_destEvent       = (E3_Addr) &rxd->DataEvent;
+               dma->dma_destCookieVProc = DMA_COOKIE_THREAD | DMA_COOKIE (cookies[env->NodeId], EP_VP_DATA (rail->NodeId));
+               dma->dma_srcEvent        = env->TxdRail + offsetof (EP3_TXD_RAIL_ELAN, DataEvent);
+               dma->dma_srcCookieVProc  = DMA_COOKIE_THREAD | DMA_REMOTE_COOKIE (cookies[env->NodeId], EP_VP_DATA (env->NodeId));
+               len = 0;
+           }
+           else
+           {
+               /* Generate the DMA chain to fetch the data */
+               for (i = 0, buffer = rxd->Data.nmd_addr, len = 0; i < env->nFrags; i++, dma++, event++)
+               {
+                   dma->dma_type            = E3_DMA_TYPE(DMA_BYTE, DMA_READ, DMA_NORMAL, EP3_DMAFAILCOUNT);
+                   dma->dma_size            = env->Frags[i].nmd_len;
+                   dma->dma_source          = env->Frags[i].nmd_addr;
+                   dma->dma_dest            = buffer;
+                   dma->dma_destEvent       = (E3_Addr) event;
+                   dma->dma_destCookieVProc = DMA_COOKIE_THREAD | DMA_COOKIE (cookies[env->NodeId], EP_VP_DATA (rail->NodeId));
+                   dma->dma_srcEvent        = env->TxdRail + offsetof (EP3_TXD_RAIL_ELAN, DataEvent);
+                   dma->dma_srcCookieVProc  = DMA_COOKIE_THREAD | DMA_REMOTE_COOKIE (cookies[env->NodeId], EP_VP_DATA (env->NodeId));
+                   
+                   event->ev_Count = 1;
+                   
+                   buffer += dma->dma_size;
+                   len    += dma->dma_size;
+               }
+               
+               /* Point the last dma at the done event */
+               (--dma)->dma_destEvent = (E3_Addr) &rxd->DataEvent;
+               
+               if (rxd->Data.nmd_len < len)
+               {
+                   /* The receive descriptor was too small for the message */
+                   /* complete the message anyway,  but don't transfer any */
+                   /* data,  we set the length to EP_MSG_TOO_BIG */
+                   for (i = 0, dma = rxd->Dmas; i < env->nFrags; i++, dma++)
+                       dma->dma_size = 0;
+                   
+                   len = EP_MSG_TOO_BIG;
+               }
+           }
+           
+           /* Store the received message length in the rxdElan for CompleteEnvelope */
+           rxd->Data.nmd_len = len;
+
+           /* Initialise %g1 with the  "rxd" so the trap handler can
+            * complete the envelope processing if we trap while sending the
+            * packet */
+           asm volatile ("mov %0, %%g1" : /* no outputs */ : "r" (rxd));
+
+           /* Generate a packet to start the data transfer */
+           c_open (EP_VP_DATA (env->NodeId));
+           c_sendtrans2 (TR_THREADIDENTIFY, rxd->Dmas->dma_destCookieVProc, 0, 0);
+           c_sendmem (TR_SENDACK | TR_REMOTEDMA, 0, rxd->Dmas); 
+           ack = c_close();
+           
+           /*
+            * If we trapped for an output timeout, then the trap handler will have
+            * completed processing this envelope and cleared the spinlock, so we just
+            * need to update the queue descriptor.
+            */
+           if (ack == EP3_PAckStolen)
+               goto consume_envelope;
+           
+           if (ack != E3_PAckOk)
+           {
+               /* our packet got nacked, so trap into the kernel so that
+                * it can complete processing of this envelope.
+                */
+               asm volatile ("ta %0" : /* no outputs */ : "i" (EP3_UNIMP_TRAP_PACKET_NACKED));         /* HALT POINT */
+               goto consume_envelope;
+           }
+
+           /* remove the RXD from the pending list */
+           EP3_SPINENTER (&rcvrElan->PendingLock, &rcvrMain->PendingLock);
+           if ((rcvrElan->PendingDescs = rxd->Next) == 0)
+               rcvrMain->PendingDescsTailp = 0;
+           EP3_SPINEXIT (&rcvrElan->PendingLock, &rcvrMain->PendingLock);
+
+           /* Copy the envelope information - as 5 64 byte chunks.
+            * We force the parameters in g5, g6 so that they aren't
+            * trashed by the loadblk32 into the locals/ins
+            */
+           if (EP_HAS_PAYLOAD(env->Attr))
+           { 
+               register void *src asm ("g5") = (void *) env;
+               register void *dst asm ("g6") = (void *)  &((EP_RXD_MAIN *) rxd->RxdMain)->Envelope;
+
+               asm volatile (
+                   "and     %%sp,63,%%g7               ! Calculate stack alignment\n"
+                   "add     %%g7,64,%%g7               ! Space to save the registers\n"
+                   "sub     %%sp,%%g7,%%sp             ! align stack\n" 
+                   "stblock64 %%l0,[%%sp]              ! save the locals and ins\n"
+
+                   "ldblock64 [%0 + 0],%%l0            ! load 64-byte block into locals/ins\n"         /* copy envelope */
+                   "stblock64 %%l0,[%1 + 0]            ! store 64-byte block from local/ins\n"
+                   "ldblock64 [%0 + 64],%%l0           ! load 64-byte block into locals/ins\n"
+                   "stblock64 %%l0,[%1 + 64]           ! store 64-byte block from local/ins\n"
+
+                   "ldblock64 [%0 + 128],%%l0          ! load 64-byte block into locals/ins\n"         /* copy payload */
+                   "stblock64 %%l0,[%1 + 128]          ! store 64-byte block from local/ins\n"
+                   "ldblock64 [%0 + 192],%%l0          ! load 64-byte block into locals/ins\n"
+                   "stblock64 %%l0,[%1 + 192]          ! store 64-byte block from local/ins\n"
+
+                   "ldblock64 [%%sp],%%l0              ! restore locals and ins\n"
+                   "add     %%sp,%%g7,%%sp             ! restore stack pointer\n"
+                   : /* outputs */
+                   : /* inputs */ "r" (src), "r" (dst)
+                   : /* clobbered */ "g5", "g6", "g7" );
+           }
+           else
+           { 
+               register void *src asm ("g5") = (void *) env;
+               register void *dst asm ("g6") = (void *)  &((EP_RXD_MAIN *) rxd->RxdMain)->Envelope;
+
+               asm volatile (
+                   "and     %%sp,63,%%g7               ! Calculate stack alignment\n"
+                   "add     %%g7,64,%%g7               ! Space to save the registers\n"
+                   "sub     %%sp,%%g7,%%sp             ! align stack\n" 
+                   "stblock64 %%l0,[%%sp]              ! save the locals and ins\n"
+
+                   "ldblock64 [%0 + 0],%%l0            ! load 64-byte block into locals/ins\n"
+                   "stblock64 %%l0,[%1 + 0]            ! store 64-byte block from local/ins\n"
+                   "ldblock64 [%0 + 64],%%l0           ! load 64-byte block into locals/ins\n"
+                   "stblock64 %%l0,[%1 + 64]           ! store 64-byte block from local/ins\n"
+
+                   "ldblock64 [%%sp],%%l0              ! restore locals and ins\n"
+                   "add     %%sp,%%g7,%%sp             ! restore stack pointer\n"
+                   : /* outputs */
+                   : /* inputs */ "r" (src), "r" (dst)
+                   : /* clobbered */ "g5", "g6", "g7" );
+           }
+
+           /* Store the message length to indicate that I've finished */
+           ((EP_RXD_MAIN *) rxd->RxdMain)->Len = rxd->Data.nmd_len;                                    /* PCI write  */
+           
+           EP3_SPINEXIT(&rcvrElan->ThreadLock, &rcvrMain->ThreadLock);
+
+       consume_envelope:
+           /* Sample the queue full bit *BEFORE* moving the fptr.
+            * Then only clear it if it was full before, otherwise,
+            * as soon as the fptr is moved on the queue could fill 
+            * up, and so clearing it could mark a full queue as 
+            * empty.
+            *
+            * While the full bit is set, the queue is in a 'steady
+            * state', so it is safe to set the q_state
+            * 
+            */
+           if (((tmp = q->q_state) & E3_QUEUE_FULL) == 0)
+               q->q_fptr = nfptr;                              /* update queue */
+           else
+           {
+               q->q_fptr = nfptr;                              /* update queue */
+               q->q_state = tmp &~E3_QUEUE_FULL;               /* and clear full flag */
+           }
+
+           count++;                                            /* bump message count */
+           if (nfptr == q->q_top)                              /* queue wrap */
+               nfptr = q->q_base;
+           else
+               nfptr += q->q_size;
+
+           c_break_busywait();                                 /* be nice              HALT POINT */
+
+       } while (nfptr != q->q_bptr);                           /* loop until Fptr == Bptr */
+    }
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms_elan4.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan4.c   2005-06-01 23:12:54.640432872 -0400
@@ -0,0 +1,392 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcomms_elan4.c,v 1.11.2.1 2004/10/28 11:53:28 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms_elan4.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "debug.h"
+#include "kcomm_elan4.h"
+#include "epcomms_elan4.h"
+
+static void
+ep4comms_flush_interrupt (EP4_RAIL *rail, void *arg)
+{
+    EP4_COMMS_RAIL *commsRail = (EP4_COMMS_RAIL *) arg;
+    unsigned long  flags;
+
+    spin_lock_irqsave (&commsRail->r_flush_lock, flags);
+    commsRail->r_flush_count = 0;
+    kcondvar_wakeupall (&commsRail->r_flush_sleep, &commsRail->r_flush_lock);
+    spin_unlock_irqrestore  (&commsRail->r_flush_lock, flags);
+}
+
+void
+ep4comms_flush_start (EP4_COMMS_RAIL *commsRail)
+{
+    kmutex_lock (&commsRail->r_flush_mutex);
+}
+
+void
+ep4comms_flush_wait (EP4_COMMS_RAIL *commsRail)
+{
+    unsigned long flags;
+
+    ep4_wait_event_cmd (commsRail->r_flush_mcq, 
+                       commsRail->r_elan_addr + offsetof (EP4_COMMS_RAIL_ELAN, r_flush_event),
+                       E4_EVENT_INIT_VALUE (-32 * commsRail->r_flush_count, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0),
+                       commsRail->r_flush_ecq->ecq_addr, 
+                       INTERRUPT_CMD | (commsRail->r_flush_intcookie.int_val << E4_MAIN_INT_SHIFT));
+
+    spin_lock_irqsave (&commsRail->r_flush_lock, flags);
+    while (commsRail->r_flush_count != 0)
+       kcondvar_wait (&commsRail->r_flush_sleep, &commsRail->r_flush_lock, &flags);
+    spin_unlock_irqrestore (&commsRail->r_flush_lock, flags);
+    
+    kmutex_unlock (&commsRail->r_flush_mutex);
+}
+
+void
+ep4comms_flush_setevent (EP4_COMMS_RAIL *commsRail, ELAN4_CQ *cq)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&commsRail->r_flush_lock, flags);
+
+    elan4_set_event_cmd (cq, commsRail->r_elan_addr + offsetof (EP4_COMMS_RAIL_ELAN, r_flush_event));
+    
+    commsRail->r_flush_count++;
+    
+    spin_unlock_irqrestore (&commsRail->r_flush_lock, flags);
+}
+
+void
+ep4comms_flush_callback (void *arg, statemap_t *map)
+{
+    EP4_COMMS_RAIL   *commsRail = (EP4_COMMS_RAIL *) arg;
+    EP_COMMS_SUBSYS  *subsys    = commsRail->r_generic.Subsys;
+    EP4_RAIL        *rail      = (EP4_RAIL *) commsRail->r_generic.Rail;
+    unsigned int     rnum       = rail->r_generic.Number;
+    struct list_head *el;
+
+    /*
+     * We stall the retry thread from CB_FLUSH_FILTERING until
+     * we've finished CB_FLUSH_FLUSHING to ensure that sten 
+     * packets can not be being retried while we flush them
+     * through.
+     */
+    switch (rail->r_generic.CallbackStep)
+    {
+    case EP_CB_FLUSH_FILTERING:
+       ep_kthread_stall (&rail->r_retry_thread);
+
+       ep4comms_flush_start (commsRail);
+       break;
+
+    case EP_CB_FLUSH_FLUSHING:
+       break;
+    }
+
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[rnum])
+           ep4xmtr_flush_callback (xmtr, (EP4_XMTR_RAIL *) xmtr->Rails[rnum]);
+    }
+
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[rnum])
+           ep4rcvr_flush_callback (rcvr, (EP4_RCVR_RAIL *) rcvr->Rails[rnum]);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    switch (rail->r_generic.CallbackStep)
+    {
+    case EP_CB_FLUSH_FILTERING:
+       ep4comms_flush_wait (commsRail);
+       break;
+
+    case EP_CB_FLUSH_FLUSHING:
+       ep_kthread_resume (&rail->r_retry_thread);
+       break;
+    }
+}
+
+void
+ep4comms_failover_callback (void *arg, statemap_t *map)
+{
+    EP_COMMS_RAIL    *commsRail = (EP_COMMS_RAIL *) arg;
+    EP_COMMS_SUBSYS  *subsys    = commsRail->Subsys;
+    unsigned int     rnum       = commsRail->Rail->Number;
+    struct list_head *el;
+
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[rnum])
+           ep4xmtr_failover_callback (xmtr, (EP4_XMTR_RAIL *) xmtr->Rails[rnum]);
+    }
+
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[rnum])
+           ep4rcvr_failover_callback (rcvr, (EP4_RCVR_RAIL *) rcvr->Rails[rnum]);
+    }
+    kmutex_unlock (&subsys->Lock);
+}
+
+void
+ep4comms_disconnect_callback (void *arg, statemap_t *map)
+{
+    EP_COMMS_RAIL    *commsRail = (EP_COMMS_RAIL *) arg;
+    EP_COMMS_SUBSYS  *subsys    = commsRail->Subsys;
+    unsigned int     rnum       = commsRail->Rail->Number;
+    struct list_head *el;
+
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[rnum])
+           ep4xmtr_disconnect_callback (xmtr, (EP4_XMTR_RAIL *) xmtr->Rails[rnum]);
+    }
+
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[rnum])
+           ep4rcvr_disconnect_callback (rcvr, (EP4_RCVR_RAIL *) rcvr->Rails[rnum]);
+    }
+    kmutex_unlock (&subsys->Lock);
+}
+
+void
+ep4comms_neterr_callback (EP4_RAIL *rail, void *arg, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP_COMMS_RAIL    *commsRail = (EP_COMMS_RAIL *) arg;
+    EP_COMMS_SUBSYS  *subsys    = commsRail->Subsys;
+    unsigned int     rnum       = commsRail->Rail->Number;
+    struct list_head *el;
+    
+    /* First - stall the retry thread, so that it will no longer restart 
+     *         any sten packets from the retry lists */
+    ep_kthread_stall (&rail->r_retry_thread);
+
+    ep4comms_flush_start ((EP4_COMMS_RAIL *) commsRail);
+
+    /* Second - flush through all command queues for xmtrs and rcvrs */
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[rnum])
+           ep4xmtr_neterr_flush (xmtr, (EP4_XMTR_RAIL *) xmtr->Rails[rnum], nodeId, cookies);
+    }
+    
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[rnum])
+           ep4rcvr_neterr_flush (rcvr, (EP4_RCVR_RAIL *) rcvr->Rails[rnum], nodeId, cookies);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    /* Third - wait for flush to complete */
+    ep4comms_flush_wait ((EP4_COMMS_RAIL *) commsRail);
+    
+    /* Fourth - flush through all command queues */
+    ep4_flush_ecqs (rail);
+    
+    /* Fifth - search all the retry lists for the network error cookies */
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Transmitters) {
+       EP_XMTR *xmtr = list_entry (el, EP_XMTR, Link);
+       
+       if (xmtr->Rails[rnum])
+           ep4xmtr_neterr_check (xmtr, (EP4_XMTR_RAIL *) xmtr->Rails[rnum], nodeId, cookies);
+    }
+
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+       
+       if (rcvr->Rails[rnum])
+           ep4rcvr_neterr_check (rcvr, (EP4_RCVR_RAIL *) rcvr->Rails[rnum], nodeId, cookies);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    ep_kthread_resume (&rail->r_retry_thread);
+}
+
+
+EP_COMMS_RAIL *
+ep4comms_add_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *r)
+{
+    EP4_RAIL       *rail = (EP4_RAIL *)r;
+    ELAN4_DEV      *dev  = rail->r_ctxt.ctxt_dev;
+    EP4_COMMS_RAIL *commsRail;
+    E4_InputQueue   qdesc;
+    int i;
+
+    KMEM_ZALLOC (commsRail, EP4_COMMS_RAIL *,sizeof (EP4_COMMS_RAIL), 1);
+
+    if (commsRail == NULL)
+       return NULL;
+    
+    commsRail->r_generic.Ops.DelRail          = ep4comms_del_rail;
+    commsRail->r_generic.Ops.DisplayRail      = ep4comms_display_rail;
+    commsRail->r_generic.Ops.Rcvr.AddRail     = ep4rcvr_add_rail;
+    commsRail->r_generic.Ops.Rcvr.DelRail     = ep4rcvr_del_rail;
+    commsRail->r_generic.Ops.Rcvr.Check       = ep4rcvr_check;
+    commsRail->r_generic.Ops.Rcvr.QueueRxd    = ep4rcvr_queue_rxd;
+    commsRail->r_generic.Ops.Rcvr.RpcPut      = ep4rcvr_rpc_put;
+    commsRail->r_generic.Ops.Rcvr.RpcGet      = ep4rcvr_rpc_get;
+    commsRail->r_generic.Ops.Rcvr.RpcComplete = ep4rcvr_rpc_complete;
+
+    commsRail->r_generic.Ops.Rcvr.StealRxd    = ep4rcvr_steal_rxd;
+
+    commsRail->r_generic.Ops.Rcvr.DisplayRcvr = ep4rcvr_display_rcvr;
+    commsRail->r_generic.Ops.Rcvr.DisplayRxd  = ep4rcvr_display_rxd;
+
+    commsRail->r_generic.Ops.Rcvr.FillOutRailStats = ep4rcvr_fillout_rail_stats;
+
+    commsRail->r_generic.Ops.Xmtr.AddRail     = ep4xmtr_add_rail;
+    commsRail->r_generic.Ops.Xmtr.DelRail     = ep4xmtr_del_rail;
+    commsRail->r_generic.Ops.Xmtr.Check       = ep4xmtr_check;
+    commsRail->r_generic.Ops.Xmtr.BindTxd     = ep4xmtr_bind_txd;
+    commsRail->r_generic.Ops.Xmtr.UnbindTxd   = ep4xmtr_unbind_txd;
+    commsRail->r_generic.Ops.Xmtr.PollTxd     = ep4xmtr_poll_txd;
+    commsRail->r_generic.Ops.Xmtr.CheckTxdState = ep4xmtr_check_txd_state;
+
+    commsRail->r_generic.Ops.Xmtr.DisplayXmtr = ep4xmtr_display_xmtr;
+    commsRail->r_generic.Ops.Xmtr.DisplayTxd  = ep4xmtr_display_txd;
+
+    commsRail->r_generic.Ops.Xmtr.FillOutRailStats = ep4xmtr_fillout_rail_stats;
+
+    /* Allocate command queue space for flushing (1 dword for interrupt + 4 dwords for waitevent) */
+    if ((commsRail->r_flush_ecq = ep4_get_ecq (rail, EP4_ECQ_EVENT, 1)) == NULL)
+    {
+       KMEM_FREE (commsRail, sizeof (EP4_COMMS_RAIL));
+       return NULL;
+    }
+
+    if ((commsRail->r_flush_mcq = ep4_get_ecq (rail, EP4_ECQ_MAIN, 4)) == NULL)
+    {
+       ep4_put_ecq (rail, commsRail->r_flush_ecq, 1);
+       KMEM_FREE (commsRail, sizeof (EP4_COMMS_RAIL));
+       return NULL;
+    }
+
+    /* Allocate and initialise the elan memory part */
+    if ((commsRail->r_elan = ep_alloc_elan (r, EP4_COMMS_RAIL_ELAN_SIZE, 0, &commsRail->r_elan_addr)) == (sdramaddr_t) 0)
+    {
+       ep4_put_ecq (rail, commsRail->r_flush_mcq, 4);
+       ep4_put_ecq (rail, commsRail->r_flush_ecq, 1);
+       KMEM_FREE (commsRail, sizeof (EP4_COMMS_RAIL));
+       return NULL;
+    }
+
+    ep4_register_intcookie (rail, &commsRail->r_flush_intcookie, commsRail->r_elan_addr + offsetof (EP4_COMMS_RAIL_ELAN, r_flush_event),
+                           ep4comms_flush_interrupt, commsRail);
+
+    elan4_sdram_writeq (dev, commsRail->r_elan + offsetof (EP4_COMMS_RAIL_ELAN, r_flush_event.ev_CountAndType),
+                       E4_EVENT_INIT_VALUE (0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+
+
+    /* Allocate and initialise all the queue desriptors as "full" with no event */
+    if ((commsRail->r_descs = ep_alloc_memory_elan (r, EP_EPCOMMS_QUEUE_BASE, roundup (EP_MSG_NSVC * EP_QUEUE_DESC_SIZE, SDRAM_PAGE_SIZE), EP_PERM_ALL, 0)) == (sdramaddr_t) 0)
+    {
+       ep_free_elan (r, commsRail->r_elan_addr, EP4_COMMS_RAIL_ELAN_SIZE);
+       ep4_put_ecq (rail, commsRail->r_flush_mcq, 4);
+       ep4_put_ecq (rail, commsRail->r_flush_ecq, 1);
+       KMEM_FREE (commsRail, sizeof (EP4_COMMS_RAIL));
+       return NULL;
+    }
+
+    qdesc.q_bptr    = 0;
+    qdesc.q_fptr    = 8;
+    qdesc.q_control = E4_InputQueueControl (qdesc.q_bptr,qdesc.q_fptr, 8);
+    qdesc.q_event   = 0;
+
+    for (i = 0; i < EP_MSG_NSVC; i++)
+       elan4_sdram_copyq_to_sdram (rail->r_ctxt.ctxt_dev, &qdesc, commsRail->r_descs + (i * EP_QUEUE_DESC_SIZE),
+                                   sizeof (E4_InputQueue));
+
+    kmutex_init (&commsRail->r_flush_mutex);
+    spin_lock_init (&commsRail->r_flush_lock);
+    kcondvar_init (&commsRail->r_flush_sleep);
+
+    ep_register_callback (r, EP_CB_FLUSH_FILTERING, ep4comms_flush_callback,      commsRail);
+    ep_register_callback (r, EP_CB_FLUSH_FLUSHING,  ep4comms_flush_callback,      commsRail);
+    ep_register_callback (r, EP_CB_FAILOVER,        ep4comms_failover_callback,   commsRail);
+    ep_register_callback (r, EP_CB_DISCONNECTING,   ep4comms_disconnect_callback, commsRail);
+
+    commsRail->r_neterr_ops.op_func = ep4comms_neterr_callback;
+    commsRail->r_neterr_ops.op_arg  = commsRail;
+    
+    ep4_add_neterr_ops (rail, &commsRail->r_neterr_ops);
+
+    return (EP_COMMS_RAIL *) commsRail;
+}
+
+void
+ep4comms_del_rail (EP_COMMS_RAIL *r)
+{
+    EP4_COMMS_RAIL *commsRail = (EP4_COMMS_RAIL *) r;
+    EP4_RAIL       *rail      = (EP4_RAIL *) commsRail->r_generic.Rail;
+
+    ep_remove_callback (&rail->r_generic, EP_CB_FLUSH_FILTERING, ep4comms_flush_callback,      commsRail);
+    ep_remove_callback (&rail->r_generic, EP_CB_FLUSH_FLUSHING,  ep4comms_flush_callback,      commsRail);
+    ep_remove_callback (&rail->r_generic, EP_CB_FAILOVER,        ep4comms_failover_callback,   commsRail);
+    ep_remove_callback (&rail->r_generic, EP_CB_DISCONNECTING,   ep4comms_disconnect_callback, commsRail);
+
+    kcondvar_destroy (&commsRail->r_flush_sleep);
+    spin_lock_destroy (&commsRail->r_flush_lock);
+    kmutex_destroy (&commsRail->r_flush_mutex);
+
+    ep_free_memory_elan (&rail->r_generic, EP_EPCOMMS_QUEUE_BASE);
+    ep_free_elan (&rail->r_generic, commsRail->r_elan_addr, EP4_COMMS_RAIL_ELAN_SIZE);
+
+    ep4_deregister_intcookie (rail, &commsRail->r_flush_intcookie);
+
+    ep4_put_ecq (rail, commsRail->r_flush_mcq, 4);
+    ep4_put_ecq (rail, commsRail->r_flush_ecq, 1);
+
+    KMEM_FREE (commsRail, sizeof (EP4_COMMS_RAIL));
+}
+
+void
+ep4comms_display_rail (EP_COMMS_RAIL *r)
+{
+    EP4_COMMS_RAIL *commsRail = (EP4_COMMS_RAIL *) r;
+    EP4_RAIL       *rail      = (EP4_RAIL *) commsRail->r_generic.Rail;
+    ELAN4_DEV      *dev       = rail->r_ctxt.ctxt_dev;
+    
+    ep4_display_rail (rail);
+
+    ep_debugf (DBG_DEBUG, "   flush count=%d mcq=%p ecq=%p event %llx.%llx.%llx\n", 
+              commsRail->r_flush_count, commsRail->r_flush_mcq, commsRail->r_flush_ecq,
+              elan4_sdram_readq (dev, commsRail->r_elan + offsetof (EP4_COMMS_RAIL_ELAN, r_flush_event.ev_CountAndType)),
+              elan4_sdram_readq (dev, commsRail->r_elan + offsetof (EP4_COMMS_RAIL_ELAN, r_flush_event.ev_WritePtr)),
+              elan4_sdram_readq (dev, commsRail->r_elan + offsetof (EP4_COMMS_RAIL_ELAN, r_flush_event.ev_WriteValue)));
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan4.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms_elan4.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan4.h   2005-06-01 23:12:54.641432720 -0400
@@ -0,0 +1,470 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __EPCOMMS_ELAN4_H
+#define __EPCOMMS_ELAN4_H
+
+#ident "@(#)$Id: epcomms_elan4.h,v 1.13.2.1 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms_elan4.h,v $ */
+
+
+#include <elan4/types.h>
+
+/*
+ * Elan4 spinlocks are a pair of 64 bit words, one in elan sdram and one in main memory
+ * the sdram word holds the thread sequence number in the bottom 32 bits and the main
+ * lock in the top 32 bits.  The main memory word holds the sequence number only in
+ * it's bottom 32 bits */
+
+typedef volatile E4_uint64 EP4_SPINLOCK_MAIN;
+typedef volatile E4_uint64 EP4_SPINLOCK_ELAN;
+
+#define EP4_SPINLOCK_SEQ       0
+#define EP4_SPINLOCK_MLOCK     4
+
+#if defined(__elan4__)
+
+#define EP4_SPINENTER(CPORT,SLE,SLM) \
+do { \
+    register long tmp; \
+\
+    asm volatile ("ld4         [%1], %0\n" \
+                 "inc          %0\n" \
+                 "st4          %0, [%1]\n" \
+                 "ld4          [%1 + 4], %0\n" \
+                 "srl8,byte    %0, 4, %0\n" \
+                 : /* outputs */ "=r" (tmp)  \
+                 : /* inputs */ "r" (SLE), "r" (SLM)); \
+\
+    if (tmp) \
+       ep4_spinblock (CPORT,SLE, SLM); \
+} while (0)
+
+extern void ep4_spinblock(E4_uint64 *cport, EP4_SPINLOCK_ELAN *sle, EP4_SPINLOCK_MAIN *slm);
+
+#define EP4_SPINEXIT(CPORT,SLE,SLM) \
+do { \
+    register long tmp; \
+\
+    asm volatile ("ld4         [%1], %0\n" \
+                 "st4          %0, [%2]\n" \
+                 : /* outputs */ "=r" (tmp) \
+                 : /* inputs */ "r" (SLE), "r" (SLM)); \
+} while (0)
+
+#else
+
+#define EP4_SPINENTER(DEV,SLE,SLM) \
+do { \
+    uint32_t seq; \
+\
+    mb(); \
+    elan4_sdram_writel (DEV, (SLE) + EP4_SPINLOCK_MLOCK, 1); \
+    mb(); \
+    while ((seq = elan4_sdram_readl (DEV, (SLE) + EP4_SPINLOCK_SEQ)) != *((uint32_t *) (SLM))) \
+    { \
+       while (*((uint32_t *) (SLM)) == (seq - 1)) \
+       { \
+           mb(); \
+           DELAY(1); \
+       } \
+    } \
+} while (0)
+
+#define EP4_SPINEXIT(DEV,SLE,SLM) \
+do { \
+    wmb(); \
+    elan4_sdram_writel (DEV, (SLE) + EP4_SPINLOCK_MLOCK, 0); \
+} while (0)
+
+#endif /* !defined(__elan4__) */
+
+#define EP4_STEN_RETRYCOUNT    16
+#define EP4_DMA_RETRYCOUNT     16
+
+typedef struct ep4_intr_cmd
+{
+    E4_uint64          c_write_cmd;
+    E4_uint64          c_write_value;
+    E4_uint64          c_intr_cmd;
+} EP4_INTR_CMD;
+
+#define        EP4_INTR_CMD_NDWORDS    (sizeof (EP4_INTR_CMD) / 8)
+
+typedef struct ep4_rxd_sten_cmd
+{
+    E4_uint64          c_open;
+
+    E4_uint64          c_trans;
+    E4_uint64          c_cookie;
+    E4_uint64          c_dma_typeSize;
+    E4_uint64          c_dma_cookie;
+    E4_uint64          c_dma_vproc;
+    E4_uint64          c_dma_srcAddr;
+    E4_uint64          c_dma_dstAddr;
+    E4_uint64          c_dma_srcEvent;
+    E4_uint64          c_dma_dstEvent;
+
+    E4_uint64          c_ok_guard;
+    E4_uint64          c_ok_write_cmd;
+    E4_uint64          c_ok_write_value;
+    
+    E4_uint64          c_fail_guard;
+    E4_uint64          c_fail_setevent;
+
+    E4_uint64          c_nop_cmd;
+} EP4_RXD_STEN_CMD;
+
+#define EP4_RXD_STEN_CMD_NDWORDS       (sizeof (EP4_RXD_STEN_CMD) / 8)
+
+typedef struct ep4_rxd_dma_cmd
+{
+    E4_uint64          c_dma_typeSize;
+    E4_uint64          c_dma_cookie;
+    E4_uint64          c_dma_vproc;
+    E4_uint64          c_dma_srcAddr;
+    E4_uint64          c_dma_dstAddr;
+    E4_uint64          c_dma_srcEvent;
+    E4_uint64          c_dma_dstEvent;
+    E4_uint64          c_nop_cmd;
+} EP4_RXD_DMA_CMD;
+
+#define EP4_RXD_DMA_CMD_NDWORDS                (sizeof (EP4_RXD_DMA_CMD) / 8)
+#define EP4_RXD_START_CMD_NDWORDS      (sizeof (E4_ThreadRegs) / 8)
+
+typedef struct ep4_rxd_rail_elan
+{
+    EP4_RXD_STEN_CMD    rxd_sten[EP_MAXFRAG+1];
+
+    EP4_INTR_CMD       rxd_done_cmd;                           /* command stream issued by done event (aligned to 64 bytes) */
+    E4_Addr            rxd_next;                               /* linked list when on pending list (pad to 32 bytes)*/
+    E4_Event32         rxd_failed;                             /* event set when sten packet fails */
+
+    EP4_INTR_CMD        rxd_failed_cmd;                                /* command stream issued by fail event (aligned to 64 bytes) */
+    E4_uint64          rxd_queued;                             /* rxd queuing thread has executed (pad to 32 bytes)*/
+
+    E4_Event32         rxd_start;                              /* event to set to fire off and event chain (used as chain[0]) */
+    E4_Event32         rxd_chain[EP_MAXFRAG];                  /* chained events (aligned to 32 bytes) */
+    E4_Event32         rxd_done;                               /* event to fire done command stream causing interrupt (used as chain[EP_MAXFRAG]) */
+
+    E4_Addr            rxd_rxd;                                /* elan address of EP4_RXD_MAIN */
+    E4_Addr            rxd_main;                               /* elan address of EP4_RXD_RAIL_MAIN */
+    E4_uint64          rxd_debug;                              /* thread debug value */
+
+    EP_NMD             rxd_buffer;                             /* Network mapping descriptor for receive data */
+} EP4_RXD_RAIL_ELAN;
+
+#define EP4_RXD_RAIL_ELAN_SIZE roundup(sizeof (EP4_RXD_RAIL_ELAN), 64)
+
+typedef struct ep4_rxd_rail_main
+{
+    E4_uint64          rxd_sent[EP_MAXFRAG+1];                 /* sten packet sent */
+    E4_uint64          rxd_failed;                             /* sten packet failed */
+    E4_uint64          rxd_done;                               /* operation complete */
+
+    E4_Addr            rxd_scq;                                /* command port for scq */
+} EP4_RXD_RAIL_MAIN;
+
+#define EP4_RXD_RAIL_MAIN_SIZE roundup(sizeof (EP4_RXD_RAIL_MAIN), 8)
+
+#if !defined(__elan4__)
+typedef struct ep4_rxd_rail
+{
+    EP_RXD_RAIL                rxd_generic;
+
+    struct list_head    rxd_retry_link;
+    unsigned long       rxd_retry_time;
+
+    EP4_INTCOOKIE      rxd_intcookie;
+
+    sdramaddr_t                rxd_elan;
+    EP_ADDR            rxd_elan_addr;
+    
+    EP4_RXD_RAIL_MAIN  *rxd_main;
+    EP_ADDR            rxd_main_addr;
+
+    EP4_ECQ           *rxd_ecq;                                /* cq with 128 bytes targetted by event */
+    EP4_ECQ           *rxd_scq;                                /* cq with 8 bytes targetted by main/thread store */
+} EP4_RXD_RAIL;
+
+#define EP4_NUM_RXD_PER_BLOCK  16
+
+typedef struct ep4_rxd_rail_block
+{
+    struct list_head   blk_link;
+    EP4_RXD_RAIL       blk_rxds[EP4_NUM_RXD_PER_BLOCK];
+} EP4_RXD_RAIL_BLOCK;
+
+#endif /* !defined(__elan4__) */
+
+typedef struct ep4_rcvr_rail_elan
+{
+    E4_uint64          rcvr_thread_stall[8];                   /* place for thread to stall */
+    E4_Event32         rcvr_qevent;                            /* Input queue event */
+    E4_Event32         rcvr_thread_halt;                       /* place for thread to halt */
+
+    volatile E4_Addr    rcvr_pending_tailp;                    /* list of pending rxd's (elan addr) */
+    volatile E4_Addr   rcvr_pending_head;                      /*   -- this pair aligned to 16 bytes */
+
+    EP4_SPINLOCK_ELAN  rcvr_thread_lock;                       /* spinlock for thread processing loop */
+
+    E4_uint64          rcvr_stall_intcookie;                   /* interrupt cookie to use when requseted to halt */
+
+    E4_uint64          rcvr_qbase;                             /* base of input queue */
+    E4_uint64          rcvr_qlast;                             /* last item in input queue */
+
+    E4_uint64          rcvr_debug;                             /* thread debug value */
+} EP4_RCVR_RAIL_ELAN;
+
+typedef struct ep4_rcvr_rail_main
+{
+    EP4_SPINLOCK_MAIN   rcvr_thread_lock;                      /* spinlock for thread processing loop */
+} EP4_RCVR_RAIL_MAIN;
+
+#if !defined(__elan4__)
+
+typedef struct ep4_rcvr_rail_stats
+{
+    unsigned long some_stat;
+} EP4_RCVR_RAIL_STATS;
+
+typedef struct ep4_rcvr_rail
+{
+    EP_RCVR_RAIL       rcvr_generic;                           /* generic portion */
+    
+    sdramaddr_t                rcvr_elan;
+    EP_ADDR            rcvr_elan_addr;
+
+    EP4_RCVR_RAIL_MAIN *rcvr_main;
+    EP_ADDR            rcvr_main_addr;
+
+    sdramaddr_t                rcvr_slots;                             /* input queue slots */
+    EP_ADDR            rcvr_slots_addr;                        /*   and elan address */
+
+    EP_ADDR            rcvr_stack;                             /* stack for thread */
+
+    EP4_ECQ           *rcvr_ecq;                               /* command queue space for thread STEN packets */
+    EP4_ECQ           *rcvr_resched;                           /* command queue space to reschedule the thread */
+
+    struct list_head    rcvr_freelist;                         /* freelist of per-rail receive descriptors */
+    unsigned int        rcvr_freecount;                                /*   and number on free list */
+    unsigned int        rcvr_totalcount;                               /*   total number created */
+    spinlock_t          rcvr_freelock;                         /*   and lock for free list */
+    struct list_head    rcvr_blocklist;                                /* list of receive descriptor blocks */
+
+    unsigned int        rcvr_freewaiting;                      /* waiting for descriptors to be freed */
+    kcondvar_t         rcvr_freesleep;                         /*   and sleep here */
+
+    EP4_INTCOOKIE      rcvr_stall_intcookie;                   /* interrupt cookie for thread halt */
+    unsigned char      rcvr_thread_halted;                     /* thread has been halted */
+    unsigned char       rcvr_cleanup_waiting;                  /* waiting for cleanup */
+    kcondvar_t          rcvr_cleanup_sleep;                    /*   and sleep here */
+
+    EP4_RETRY_OPS      rcvr_retryops;
+
+    struct list_head    rcvr_retrylist;                                /* list of txd's to retry envelopes for */
+    struct list_head    rcvr_polllist;                         /* list of txd's to poll for completion */
+    spinlock_t          rcvr_retrylock;
+    
+    EP4_RCVR_RAIL_STATS rcvr_stats;                             /* elan4 specific rcvr_rail stats */
+
+} EP4_RCVR_RAIL;
+
+#endif /* !defined(__elan4__) */
+
+typedef struct ep4_txd_rail_elan
+{
+    EP4_INTR_CMD        txd_env_cmd;                           /* command stream for envelope event (64 byte aligned) */
+    E4_uint64          txd_pad0;                               /*  pad to 32 bytes */
+    E4_Event32         txd_env;                                /* event set when STEN packet fails */
+
+    EP4_INTR_CMD       txd_done_cmd;                           /* command stream for done event (64 byte aligned) */
+    E4_uint64          txd_pad1;                               /*  pad to 32 bytes */
+    E4_Event32         txd_done;                               /* event set when transmit complete */
+
+    E4_Event32         txd_data;                               /* event set when xmit completes (=> phase becomes passive) */
+} EP4_TXD_RAIL_ELAN;
+
+#define EP4_TXD_RAIL_ELAN_SIZE         roundup(sizeof(EP4_TXD_RAIL_ELAN), 64)
+
+typedef struct ep4_txd_rail_main
+{
+    E4_uint64          txd_env;
+    E4_uint64          txd_data;
+    E4_uint64          txd_done;
+} EP4_TXD_RAIL_MAIN;
+
+#define EP4_TXD_RAIL_MAIN_SIZE         roundup(sizeof(EP4_TXD_RAIL_MAIN), 8)
+
+#if !defined (__elan4__)
+typedef struct ep4_txd_rail
+{
+    EP_TXD_RAIL                txd_generic;
+
+    struct list_head    txd_retry_link;
+    unsigned long      txd_retry_time;
+
+    EP4_INTCOOKIE      txd_intcookie;
+
+    sdramaddr_t                txd_elan;
+    EP_ADDR            txd_elan_addr;
+    
+    EP4_TXD_RAIL_MAIN  *txd_main;
+    EP_ADDR            txd_main_addr;
+
+    EP4_ECQ           *txd_ecq;
+
+    E4_uint64          txd_cookie;
+} EP4_TXD_RAIL;
+
+#define EP4_NUM_TXD_PER_BLOCK  21
+
+typedef struct ep4_txd_rail_block
+{
+    struct list_head   blk_link;
+    EP4_TXD_RAIL       blk_txds[EP4_NUM_TXD_PER_BLOCK];
+} EP4_TXD_RAIL_BLOCK;
+
+typedef struct ep4_xmtr_rail_main
+{
+    E4_int64           xmtr_flowcnt;
+} EP4_XMTR_RAIL_MAIN;
+
+typedef struct ep4_xmtr_rail_stats
+{
+    unsigned long some_stat;
+} EP4_XMTR_RAIL_STATS;
+
+#define EP4_TXD_LIST_POLL      0
+#define EP4_TXD_LIST_STALLED   1
+#define EP4_TXD_LIST_RETRY     2
+#define EP4_TXD_NUM_LISTS      3
+typedef struct ep4_xmtr_rail
+{
+    EP_XMTR_RAIL       xmtr_generic;
+
+    EP4_XMTR_RAIL_MAIN *xmtr_main;
+    EP_ADDR            xmtr_main_addr;
+
+    struct list_head    xmtr_freelist;
+    unsigned int        xmtr_freecount;
+    unsigned int        xmtr_totalcount;
+    spinlock_t          xmtr_freelock;
+    struct list_head    xmtr_blocklist;
+    unsigned int        xmtr_freewaiting;
+    kcondvar_t         xmtr_freesleep;
+
+    EP4_INTCOOKIE      xmtr_intcookie;                         /* interrupt cookie for "polled" descriptors */
+
+    ELAN4_CQ           *xmtr_cq;
+    E4_int64           xmtr_flowcnt;
+
+    EP4_RETRY_OPS      xmtr_retryops;
+
+    struct list_head    xmtr_retrylist[EP4_TXD_NUM_LISTS];     /* list of txd's to retry envelopes for */
+    struct list_head    xmtr_polllist;                         /* list of txd's to poll for completion */
+    spinlock_t          xmtr_retrylock;
+
+    EP4_XMTR_RAIL_STATS stats;                                  /* elan4 specific xmtr rail stats */
+} EP4_XMTR_RAIL;
+
+#define EP4_XMTR_CQSIZE                CQ_Size64K                              /* size of command queue for xmtr */
+#define EP4_XMTR_FLOWCNT       (CQ_Size(EP4_XMTR_CQSIZE) / 512)        /* # of STEN packets which can fit in */
+
+typedef struct ep4_comms_rail_elan
+{
+    E4_Event32         r_flush_event;
+} EP4_COMMS_RAIL_ELAN;
+
+#define EP4_COMMS_RAIL_ELAN_SIZE       roundup(sizeof (EP4_COMMS_RAIL_ELAN), 32)
+
+typedef struct ep4_comms_rail
+{
+    EP_COMMS_RAIL      r_generic;                              /* generic comms rail */
+    sdramaddr_t                r_descs;                                /* input queue descriptors */
+
+    sdramaddr_t                r_elan;                                 /* elan portion */
+    EP_ADDR            r_elan_addr;
+
+    kmutex_t           r_flush_mutex;                          /* sequentialise flush usage */
+    EP4_INTCOOKIE       r_flush_intcookie;                     /* interrupt cookie to generate */
+
+    kcondvar_t         r_flush_sleep;                          /* place to sleep waiting */
+    spinlock_t         r_flush_lock;                           /*   and spinlock to use */
+
+    unsigned int       r_flush_count;                          /* # setevents issued */
+    EP4_ECQ           *r_flush_ecq;                            /* command queue for interrupt */
+    EP4_ECQ           *r_flush_mcq;                            /* command queeu to issue waitevent */
+
+    EP4_NETERR_OPS      r_neterr_ops;                          /* network error fixup ops */
+} EP4_COMMS_RAIL;
+
+/* epcommsTx_elan4.c */
+extern void           ep4xmtr_flush_callback (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail);
+extern void           ep4xmtr_failover_callback (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail);
+extern void           ep4xmtr_disconnect_callback (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail);
+
+extern void          ep4xmtr_neterr_flush (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+extern void          ep4xmtr_neterr_check (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+
+/* epcommsRx_elan4.c */
+extern void           ep4rcvr_flush_callback (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail);
+extern void           ep4rcvr_failover_callback (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail);
+extern void           ep4rcvr_disconnect_callback (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail);
+
+extern void          ep4rcvr_neterr_flush (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+extern void          ep4rcvr_neterr_check (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+
+/* epcomms_elan4.c */
+extern void           ep4comms_flush_start (EP4_COMMS_RAIL *commsRail);
+extern void           ep4comms_flush_wait (EP4_COMMS_RAIL *commsRail);
+extern void           ep4comms_flush_setevent (EP4_COMMS_RAIL *commsRail, ELAN4_CQ *cq);
+
+extern EP_COMMS_RAIL *ep4comms_add_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *r);
+extern void           ep4comms_del_rail (EP_COMMS_RAIL *r);
+extern void          ep4comms_display_rail (EP_COMMS_RAIL *r);
+
+/* epcommsTx_elan4.c */
+extern int            ep4xmtr_bind_txd (EP_TXD *txd, EP_XMTR_RAIL *xmtrRail, unsigned int phase);
+extern void           ep4xmtr_unbind_txd (EP_TXD *txd, unsigned int phase);
+extern int            ep4xmtr_poll_txd (EP_XMTR_RAIL *xmtrRail, EP_TXD_RAIL *txdRail, int how);
+extern long           ep4xmtr_check (EP_XMTR_RAIL *xmtrRail, long nextRunTime);
+extern void           ep4xmtr_add_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail);
+extern void           ep4xmtr_del_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail);
+extern int            ep4xmtr_check_txd_state(EP_TXD *txd);
+
+extern void           ep4xmtr_display_xmtr (DisplayInfo *di, EP_XMTR_RAIL *xmtrRail);
+extern void           ep4xmtr_display_txd  (DisplayInfo *di, EP_TXD_RAIL *txdRail);
+
+extern void           ep4xmtr_fillout_rail_stats (EP_XMTR_RAIL *xmtr_rail, char *str);
+
+/* epcommsRx_elan4.c */
+extern int           ep4rcvr_queue_rxd (EP_RXD *rxd, EP_RCVR_RAIL *rcvrRail);
+extern void          ep4rcvr_rpc_put (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+extern void          ep4rcvr_rpc_get (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+extern void          ep4rcvr_rpc_complete (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+
+extern EP_RXD       *ep4rcvr_steal_rxd (EP_RCVR_RAIL *rcvrRail);
+
+extern long          ep4rcvr_check (EP_RCVR_RAIL *rcvrRail, long nextRunTime);
+extern void           ep4rcvr_add_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *rail);
+extern void           ep4rcvr_del_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *rail);
+
+extern void           ep4rcvr_display_rcvr (DisplayInfo *di, EP_RCVR_RAIL *rcvrRail);
+extern void           ep4rcvr_display_rxd  (DisplayInfo *di, EP_RXD_RAIL *rxdRail);
+
+extern void           ep4rcvr_fillout_rail_stats (EP_RCVR_RAIL *rcvr_rail, char *str);
+
+#endif /* !defined(__elan4__) */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __EPCOMMS_ELAN4_H */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan4_thread.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcomms_elan4_thread.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcomms_elan4_thread.c    2005-06-01 23:12:54.642432568 -0400
@@ -0,0 +1,346 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcomms_elan4_thread.c,v 1.10.8.2 2004/09/28 10:36:51 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms_elan4_thread.c,v $*/
+
+//#include <qsnet/types.h>
+
+typedef char           int8_t;
+typedef unsigned char  uint8_t;
+typedef short          int16_t;
+typedef unsigned short uint16_t;
+typedef int            int32_t;
+typedef unsigned int   uint32_t;
+typedef long           int64_t;
+typedef unsigned long  uint64_t;
+
+#include <elan/nmh.h>
+#include <elan/kcomm.h>
+#include <elan/epcomms.h>
+
+#include <elan4/registers.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "epcomms_elan4.h"
+
+#include <elan4/trtype.h>
+
+/* assembler in epcomms_asm_elan4_thread.S */
+extern void               c_waitevent_interrupt (E4_uint64 *cport, E4_Event32 *event, E4_uint64 count, E4_uint64 intcookie);
+extern EP4_RXD_RAIL_ELAN *c_stall_thread (EP4_RCVR_RAIL_ELAN *rcvrRail);
+
+#define R32_to_R47             "%r32", "%r33", "%r34", "%r35", "%r36", "%r37", "%r38", "%r39", \
+                               "%r40", "%r41", "%r42", "%r43", "%r44", "%r45", "%r46", "%r47"
+#define R48_to_R63             "%r48", "%r49", "%r50", "%r51", "%r52", "%r53", "%r54", "%r55", \
+                               "%r56", "%r57", "%r58", "%r59", "%r60", "%r61", "%r62", "%r63"
+
+/* proto types for code in asm_elan4_thread.S */
+extern void c_waitevent (E4_uint64 *commandport, E4_Addr event, E4_uint64 count);
+extern void c_reschedule(E4_uint64 *commandport);
+
+static inline unsigned long
+c_load_u16(unsigned short *ptr)
+{
+    unsigned long value;
+
+    asm volatile ("ld2         [%1], %%r2\n"
+                 "srl8,byte    %%r2, %1, %0\n"
+                 "sll8         %0, 48, %0\n"
+                 "srl8         %0, 48, %0\n"
+                 : /* outputs */ "=r" (value) 
+                 : /* inputs */ "r" (ptr)
+                 : /* clobbered */ "%r2");
+    return value;
+}
+
+static inline unsigned long
+c_load_u32(unsigned int *ptr)
+{
+    unsigned long value;
+
+    asm volatile ("ld4         [%1], %%r2\n"
+                 "srl8,byte    %%r2, %1, %0\n"
+                 "sll8         %0, 32, %0\n"
+                 "srl8         %0, 32, %0\n"
+                 : /* outputs */ "=r" (value) 
+                 : /* inputs */ "r" (ptr)
+                 : /* clobbered */ "%r2");
+    return value;
+}
+
+static inline void
+c_store_u32(unsigned int *ptr, unsigned long value)
+{
+    asm volatile ("sll8,byte   %0, %1, %%r2\n"
+                 "st4          %%r2, [%1]\n"
+                 : /* no outputs */ 
+                 : /* inputs */ "r" (value), "r" (ptr)
+                 : /* clobbered */ "%r2");
+}
+
+/* Reschedule the current Elan thread to the back of the run queue 
+ * if there is another one ready to run */
+static inline void
+c_yield (E4_uint64 *commandport)
+{
+    unsigned long rval;
+
+    asm volatile ("breaktest %0" : /* outputs */ "=r" (rval) : /* inputs */);
+
+    if (rval  & ICC_SIGNED_BIT)
+       c_reschedule(commandport);
+}
+
+/* Reschedule the current thread if we're in danger of exceeding the 
+ * thread instruction count */
+static inline void
+c_insn_check(E4_uint64 *commandport)
+{
+    unsigned long rval;
+
+    asm volatile ("breaktest %0" : /* outputs */ "=r" (rval) : /* inputs */);
+
+    if (rval & ICC_ZERO_BIT)
+       c_reschedule(commandport);
+}
+
+void
+ep4_spinblock (E4_uint64 *cport, EP4_SPINLOCK_ELAN *sle, EP4_SPINLOCK_MAIN *slm)
+{
+    do {
+       unsigned long val = *sle & 0xfffffffff;
+
+       *slm = val;                                     /* Release my lock */
+       
+       while (*sle >> 32)                              /* Wait until the main */
+           c_yield(cport);                             /* releases the lock */
+       
+       c_store_u32 ((unsigned int *) sle, val + 1);    /* and try and relock */
+    } while (*sle >> 32);
+}
+
+#define RESCHED_AFTER_PKTS     ((CQ_Size(CQ_Size64K) / 128) - 1)
+
+void
+ep4comms_rcvr (EP4_RAIL_ELAN *rail, EP4_RCVR_RAIL_ELAN *rcvrElan, EP4_RCVR_RAIL_MAIN *rcvrMain,
+              E4_InputQueue *inputq, E4_uint64 *cport, E4_uint64 *resched)
+{
+    long count = 1;
+    long fptr  = inputq->q_fptr;
+
+    for (;;)
+    {
+       c_waitevent (cport, inputq->q_event, -count << 5);
+
+       count = 0;
+
+       while (fptr != inputq->q_bptr)
+       {
+           EP_ENVELOPE        *env      = (EP_ENVELOPE *) fptr;
+           unsigned long       nodeid   = c_load_u32 (&env->NodeId);
+           unsigned long       opencmd  = OPEN_STEN_PKT_CMD | OPEN_PACKET(0, PACK_OK | RESTART_COUNT_ZERO, EP_VP_DATA(nodeid));
+           unsigned long       vproc    = EP_VP_DATA(rail->r_nodeid);
+           EP_ATTRIBUTE        attr     = c_load_u32 (&env->Attr);
+           unsigned long       txdRail  = c_load_u32 (&env->TxdRail);
+           unsigned long       nFrags   = c_load_u32 (&env->nFrags);
+           E4_uint64           cookie   = rail->r_cookies[nodeid];
+           unsigned long       srcevent = (EP_IS_RPC(attr) ? txdRail + offsetof (EP4_TXD_RAIL_ELAN, txd_data) :
+                                           txdRail + offsetof (EP4_TXD_RAIL_ELAN, txd_done));
+           EP4_RXD_RAIL_ELAN  *rxdElan;
+           EP4_RXD_RAIL_MAIN  *rxdMain;
+           EP_RXD_MAIN        *rxd;
+           EP4_RXD_STEN_CMD   *sten;
+           E4_Event32         *event;
+           unsigned long       first;
+           unsigned long       buffer;
+           unsigned long       len;
+           unsigned long       i;
+
+           EP4_SPINENTER(resched, &rcvrElan->rcvr_thread_lock, &rcvrMain->rcvr_thread_lock);
+
+           if ((rxdElan = (EP4_RXD_RAIL_ELAN *) rcvrElan->rcvr_pending_head) == 0)
+           {
+               EP4_SPINEXIT (resched, &rcvrElan->rcvr_thread_lock, &rcvrMain->rcvr_thread_lock);
+
+               rxdElan = c_stall_thread (rcvrElan);
+
+               EP4_SPINENTER(resched, &rcvrElan->rcvr_thread_lock, &rcvrMain->rcvr_thread_lock);
+           }
+           
+           if (c_load_u32 (&env->Version) != EP_ENVELOPE_VERSION)              /* envelope has been cancelled */
+           {
+               EP4_SPINEXIT (resched, &rcvrElan->rcvr_thread_lock, &rcvrMain->rcvr_thread_lock);
+               goto consume_envelope;
+           }
+
+           rxd     = (EP_RXD_MAIN *) rxdElan->rxd_rxd;
+           rxdMain = (EP4_RXD_RAIL_MAIN *) rxdElan->rxd_main;
+           first   = (EP_MAXFRAG+1) - (( EP_IS_MULTICAST(attr) ? 1 : 0) + (nFrags == 0 ? 1 : nFrags));
+           sten    = &rxdElan->rxd_sten[first];
+           event   = &rxdElan->rxd_chain[first];
+
+           if (EP_IS_MULTICAST(attr))                          /* need to fetch broadcast bitmap */
+           {
+               sten->c_open          = opencmd;
+               sten->c_trans         = SEND_TRANS_CMD | ((TR_REMOTEDMA | TR_WAIT_FOR_EOP) << 16);
+               sten->c_cookie        = cookie | EP4_COOKIE_THREAD | EP4_COOKIE_STEN;
+               sten->c_dma_typeSize  = E4_DMA_TYPE_SIZE(BT_BITOUL(EP_MAX_NODES) * sizeof (bitmap_t), DMA_DataTypeWord, 0, EP4_DMA_RETRYCOUNT);
+               sten->c_dma_cookie    = cookie | EP4_COOKIE_THREAD | EP4_COOKIE_REMOTE | EP4_COOKIE_DMA | EP4_COOKIE_INC;
+               sten->c_dma_vproc     = vproc;
+               sten->c_dma_srcAddr   = c_load_u32 (&env->TxdMain.nmd_addr) + offsetof(EP_TXD_MAIN, Bitmap);
+               sten->c_dma_dstAddr   = (E4_Addr) &rxd->Bitmap;
+               sten->c_dma_srcEvent  = srcevent;
+               sten->c_dma_dstEvent  = (E4_Addr) event;
+
+               event->ev_CountAndType = E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_STEN_CMD_NDWORDS);
+
+               cookie += (EP4_COOKIE_INC << 1);
+
+               sten++; event++;
+           }
+
+           if (nFrags == 0)
+           {
+               /* Generate an empty "get" DMA to accept the envelope and fire the rx handler */
+               sten->c_open          = opencmd;
+               sten->c_trans         = SEND_TRANS_CMD | ((TR_REMOTEDMA | TR_WAIT_FOR_EOP) << 16);
+               sten->c_cookie        = cookie | EP4_COOKIE_THREAD | EP4_COOKIE_STEN;
+               sten->c_dma_typeSize  = E4_DMA_TYPE_SIZE(0, DMA_DataTypeByte, 0, EP4_DMA_RETRYCOUNT);
+               sten->c_dma_cookie    = cookie | EP4_COOKIE_THREAD | EP4_COOKIE_REMOTE | EP4_COOKIE_DMA | EP4_COOKIE_INC;
+               sten->c_dma_vproc     = vproc;
+               sten->c_dma_srcEvent  = srcevent;
+               sten->c_dma_dstEvent  = (E4_Addr) event;
+
+               event->ev_CountAndType = E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS);
+
+               len = 0;
+
+               cookie += (EP4_COOKIE_INC << 1);
+           }
+           else
+           {
+               /* Generate the DMA chain to fetch the data */
+               for (i = 0, buffer = c_load_u32 (&rxdElan->rxd_buffer.nmd_addr), len = 0; i < nFrags; i++)
+               {
+                   unsigned long fragLen = c_load_u32 (&env->Frags[i].nmd_len);
+
+                   sten->c_open          = opencmd;
+                   sten->c_trans         = SEND_TRANS_CMD | ((TR_REMOTEDMA | TR_WAIT_FOR_EOP) << 16);
+                   sten->c_cookie        = cookie | EP4_COOKIE_THREAD | EP4_COOKIE_STEN;
+                   sten->c_dma_typeSize  = E4_DMA_TYPE_SIZE(fragLen, DMA_DataTypeByte, 0, EP4_DMA_RETRYCOUNT);
+                   sten->c_dma_cookie    = cookie | EP4_COOKIE_THREAD | EP4_COOKIE_REMOTE | EP4_COOKIE_DMA | EP4_COOKIE_INC;
+                   sten->c_dma_vproc     = vproc;
+                   sten->c_dma_srcAddr   = c_load_u32 (&env->Frags[i].nmd_addr);
+                   sten->c_dma_dstAddr   = buffer;
+                   sten->c_dma_srcEvent  = srcevent;
+                   sten->c_dma_dstEvent  = (E4_Addr) event;
+                   
+                   event->ev_CountAndType = E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_STEN_CMD_NDWORDS);
+                   
+                   buffer += fragLen;
+                   len    += fragLen;
+
+                   cookie += (EP4_COOKIE_INC << 1);
+
+                   sten++; event++;
+               }
+               
+               (--event)->ev_CountAndType = E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS);
+
+               if (c_load_u32 (&rxdElan->rxd_buffer.nmd_len) < len)
+               {
+                   /* The receive descriptor was too small for the message */
+                   /* complete the message anyway,  but don't transfer any */
+                   /* data,  we set the length to EP_MSG_TOO_BIG */
+                   for (i = first, sten = &rxdElan->rxd_sten[first]; i <= EP_MAXFRAG; i++, sten++)
+                       sten->c_dma_typeSize = E4_DMA_TYPE_SIZE(0, DMA_DataTypeByte, 0, EP4_DMA_RETRYCOUNT);
+                   
+                   len = EP_MSG_TOO_BIG;
+               }
+           }
+           
+           /* Stuff the first STEN packet into the command queue, there's always enough space, 
+            * since we will insert a waitevent at least once for the queue size */
+           asm volatile ("ld64         [%0], %%r32\n"
+                         "ld64         [%0 + 64], %%r48\n"
+                         "st64         %%r32, [%1]\n"
+                         "st64         %%r48, [%1]\n"
+                         : /* no outputs */
+                         : /* inputs */ "r" (&rxdElan->rxd_sten[first]), "r" (cport)
+                         : /* clobbered */ R32_to_R47, R48_to_R63);
+
+           /* remove the RXD from the pending list */
+           if ((rcvrElan->rcvr_pending_head = rxdElan->rxd_next) == 0)
+               rcvrElan->rcvr_pending_tailp = (E4_Addr)&rcvrElan->rcvr_pending_head;
+
+           /* mark as not queued */
+           rxdElan->rxd_queued = 0;
+
+           /* copy down the envelope */
+           if (EP_HAS_PAYLOAD(attr))
+               asm volatile ("ld64     [%0],    %%r32\n"
+                             "ld64     [%0+64], %%r48\n"
+                             "st64     %%r32, [%1]\n"
+                             "ld64     [%0+128], %%r32\n"
+                             "st64     %%r48, [%1+64]\n"
+                             "ld64     [%0+192], %%r48\n"
+                             "st64     %%r32, [%1 + 128]\n"
+                             "st64     %%r48, [%1 + 192]\n"
+                             : /* no outputs */
+                             : /* inputs */    "r" (env), "r" (&rxd->Envelope)
+                             : /* clobbered */ R32_to_R47, R48_to_R63);
+ 
+           else
+               asm volatile ("ld64     [%0],    %%r32\n"
+                             "ld64     [%0+64], %%r48\n"
+                             "st64     %%r32, [%1]\n"
+                             "st64     %%r48, [%1+64]\n"
+                             : /* no outputs */
+                             : /* inputs */    "r" (env), "r" (&rxd->Envelope)
+                             : /* clobbered */ R32_to_R47, R48_to_R63);
+
+           /* Store the message length to indicate that I've finished */
+           c_store_u32 (&rxd->Len, len);
+           
+           /* Finally update the network error cookie */
+           rail->r_cookies[nodeid] = cookie;
+
+           EP4_SPINEXIT (resched, &rcvrElan->rcvr_thread_lock, &rcvrMain->rcvr_thread_lock);
+
+       consume_envelope:
+           if (fptr != rcvrElan->rcvr_qlast)
+               fptr += EP_INPUTQ_SIZE;
+           else
+               fptr = rcvrElan->rcvr_qbase;
+
+           if (! rcvrElan->rcvr_stall_intcookie)
+               inputq->q_fptr = fptr;
+
+           if (++count >= RESCHED_AFTER_PKTS)
+               break;
+
+           c_insn_check (cport);
+       }
+       
+       if (rcvrElan->rcvr_stall_intcookie)
+       {
+           c_waitevent_interrupt (cport, &rcvrElan->rcvr_thread_halt, -(1 << 5), rcvrElan->rcvr_stall_intcookie);
+           inputq->q_fptr = fptr;
+
+           count++;                                            /* one extra as we were given an extra set to wake us up */
+       }
+    }
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcommsFwd.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcommsFwd.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcommsFwd.c      2005-06-01 23:12:54.643432416 -0400
@@ -0,0 +1,310 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcommsFwd.c,v 1.12 2004/08/16 12:21:15 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/epcommsFwd.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "debug.h"
+
+unsigned int epcomms_forward_limit = 8;
+
+static void
+GenerateTree (unsigned nodeId, unsigned lowId, unsigned  highId, bitmap_t *bitmap, 
+             unsigned *parentp, unsigned *childrenp, int *nchildrenp)
+{
+    int i;
+    int count;
+    int branch;
+    int nSub;
+    int branchIndex;
+    int parent;
+    int nBranch;
+    int rem;
+    int self;
+    int branchRatio;
+    int node;
+    int x, y, z;
+
+
+#ifdef DEBUG_PRINTF
+    {
+#define OVERFLOW "...]"
+#define LINESZ  128
+       char space[LINESZ+1];
+
+       if (ep_sprintf_bitmap (space, LINESZ-strlen(OVERFLOW), bitmap, 0, 0, (highId - lowId)+1) != -1)
+           strcat (space, OVERFLOW);
+
+       EPRINTF3 (DBG_FORWARD, "GenerateTree; elan node low=%d node high=%d bitmap=%s\n", lowId, highId, space);
+#undef OVERFLOW
+#undef LINESZ
+    }
+#endif
+
+    /* Count the number of nodes in the partition */
+    /* and work out which one I am */
+    for (count = 0, self = ELAN_INVALID_NODE, i = lowId; i <= highId; i++)
+    {
+       if (BT_TEST (bitmap, i-lowId))
+       {
+           if (i == nodeId)
+               self = count;
+           count++;
+       }
+    }
+
+    EPRINTF2 (DBG_FORWARD, "GenerateTree: count=%d self=%d\n", count, self);
+
+    if (count == 0 || self == ELAN_INVALID_NODE)
+    {
+       *parentp    = ELAN_INVALID_NODE;
+       *nchildrenp = 0;
+       return;
+    }
+
+    /* search for position in tree */
+    branchRatio = EP_TREE_ARITY;               /* branching ratio */
+    branch      = 0;                           /* start with process 0 */
+    nSub        = count;                       /* and whole tree */
+    branchIndex = -1;                          /* my branch # in parent */
+    parent      = -1;                          /* my parent's group index # */
+    
+    while (branch != self)                     /* descend process tree */
+    {                                          /* until I find myself */
+       parent = branch;
+       branch++;                               /* parent + 1 = first born */
+       nSub--;                                 /* set # descendents */
+       
+       rem  = nSub % branchRatio;
+       nSub = nSub / branchRatio + 1;
+       x = rem * nSub;
+       y = self - branch;
+       
+       if (y < x)                              /* my first 'rem' branches have */
+       {                                       /* 1 more descendent... */
+           branchIndex = y / nSub;
+           branch += branchIndex * nSub;
+       }
+       else                                    /* than the rest of my branches */
+       {
+           nSub--;
+           z = (y - x) / nSub;
+           branchIndex = rem + z;
+           branch += x + z * nSub;
+       }
+    }
+
+    branch++;                                  /* my first born */
+    nSub--;                                    /* total # of my descendents */
+    /* leaves + their parents may have # children < branchRatio */
+    nBranch = (nSub < branchRatio) ? nSub : branchRatio;       
+
+    EPRINTF2 (DBG_FORWARD, "GenerateTree: parent=%d nBranch=%d\n", parent, nBranch);
+
+    /* Now calculate the real elan id's of the parent and my children */
+    if (parent == -1)
+       *parentp = ELAN_INVALID_NODE;
+    else
+    {
+       for (i = lowId, node = 0; i <= highId; i++)
+       {
+           if (BT_TEST(bitmap, i-lowId))
+               if (node++ == parent)
+                   break;
+       }
+       *parentp = i;
+    }
+
+    for (i = lowId, branchIndex = 0, node = 0; branchIndex < nBranch && i <= highId; i++)
+    {
+       if (BT_TEST(bitmap, i-lowId))
+       {
+           if (node == branch)
+           {
+               branch = branch + nSub / branchRatio + ((branchIndex < (nSub % branchRatio)) ? 1 : 0);
+
+               childrenp[branchIndex++] = i;
+           }
+           node++;
+       }
+    }
+
+    *nchildrenp = branchIndex;
+}
+
+static void
+ForwardTxDone (EP_TXD *txd, void *arg, EP_STATUS status)
+{
+    EP_FWD_DESC     *desc   = (EP_FWD_DESC *) arg;
+    EP_RXD          *rxd    = desc->Rxd;
+    EP_COMMS_SUBSYS *subsys = rxd->Rcvr->Subsys;
+    unsigned long    flags;
+
+    /* XXXX: if transmit fails, could step to next node in this subtree ? */
+
+    spin_lock_irqsave (&subsys->ForwardDescLock, flags);
+
+    if (--desc->NumChildren > 0)
+       spin_unlock_irqrestore (&subsys->ForwardDescLock, flags);
+    else
+    {
+       rxd->Rcvr->ForwardRxdCount--;
+
+       spin_unlock_irqrestore (&subsys->ForwardDescLock, flags);
+
+       KMEM_FREE (desc, sizeof (EP_FWD_DESC));
+
+       rxd->Handler (rxd);
+    }
+}
+
+long
+ep_forward_rxds (EP_COMMS_SUBSYS *subsys, long nextRunTime)
+{
+    unsigned long flags;
+    int i, res;
+
+    spin_lock_irqsave (&subsys->ForwardDescLock, flags);
+    while (! list_empty (&subsys->ForwardDescList)) 
+    {
+       EP_RXD      *rxd     = (EP_RXD *) list_entry (subsys->ForwardDescList.next, EP_RXD, Link);
+       EP_RXD_MAIN *rxdMain = rxd->RxdMain;
+       EP_ENVELOPE *env     = &rxdMain->Envelope;
+       EP_FWD_DESC *desc;
+
+       EPRINTF2 (DBG_FORWARD, "ep: forwarding rxd %p to range %x\n", rxd, env->Range);
+
+       list_del (&rxd->Link);
+
+       rxd->Rcvr->ForwardRxdCount++;
+
+       spin_unlock_irqrestore (&subsys->ForwardDescLock, flags);
+
+       KMEM_ALLOC (desc, EP_FWD_DESC *, sizeof (EP_FWD_DESC), 1);
+
+       if (desc == NULL)
+       {
+           spin_lock_irqsave (&subsys->ForwardDescLock, flags);
+           rxd->Rcvr->ForwardRxdCount--;
+           spin_unlock_irqrestore (&subsys->ForwardDescLock, flags);
+
+           rxd->Handler (rxd);
+       }
+       else
+       {
+           /* compute the spanning tree for this message */
+           unsigned int destLo = EP_RANGE_LOW (env->Range);
+           unsigned int destHi = EP_RANGE_HIGH (env->Range);
+           unsigned int parent;
+
+           GenerateTree (subsys->Subsys.Sys->Position.pos_nodeid, destLo, destHi, rxdMain->Bitmap, &parent, desc->Children, &desc->NumChildren);
+           
+           if (desc->NumChildren == 0 || (epcomms_forward_limit && (rxd->Rcvr->ForwardRxdCount >= epcomms_forward_limit)))
+           {
+               EPRINTF5 (DBG_FORWARD, "ep; don't forward rxd %p to /%d (%d children/ %d forwarding (%d))\n",
+                         rxd, rxd->Rcvr->Service, desc->NumChildren, rxd->Rcvr->ForwardRxdCount, epcomms_forward_limit);
+
+               spin_lock_irqsave (&subsys->ForwardDescLock, flags);
+               rxd->Rcvr->ForwardRxdCount--;
+               spin_unlock_irqrestore (&subsys->ForwardDescLock, flags);
+
+               KMEM_FREE (desc, sizeof (EP_FWD_DESC));
+               
+               rxd->Handler (rxd);
+           }
+           else
+           {
+               ep_nmd_subset (&desc->Data, &rxd->Data, 0, ep_rxd_len (rxd));
+               desc->Rxd = rxd;
+
+               /* NOTE - cannot access 'desc' after last call to multicast, since it could complete
+                *        and free the desc before we access it again.  Hence the reverse loop. */
+               for (i = desc->NumChildren-1; i >= 0; i--)
+               {
+                   ASSERT (desc->Children[i] < subsys->Subsys.Sys->Position.pos_nodes);
+
+                   EPRINTF3 (DBG_FORWARD, "ep: forwarding rxd %p to node %d/%d\n", rxd, desc->Children[i], rxd->Rcvr->Service);
+
+                   if ((res = ep_multicast_forward (subsys->ForwardXmtr, desc->Children[i], rxd->Rcvr->Service, 0, 
+                                                    ForwardTxDone, desc, env, EP_HAS_PAYLOAD(env->Attr) ? &rxdMain->Payload : NULL,  
+                                                    rxdMain->Bitmap, &desc->Data, 1)) != EP_SUCCESS)
+                   {
+                       ep_debugf (DBG_FORWARD, "ep: ep_multicast_forward failed\n");
+                       ForwardTxDone (NULL, desc, res);
+                   }
+               }
+               
+           }
+       }
+
+       spin_lock_irqsave (&subsys->ForwardDescLock, flags);
+    }
+    spin_unlock_irqrestore (&subsys->ForwardDescLock, flags);
+
+    return (nextRunTime);
+}
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+void
+ep_csum_rxds (EP_COMMS_SUBSYS *subsys)
+{
+    unsigned long flags;
+ 
+    spin_lock_irqsave (&subsys->CheckSumDescLock, flags);
+    while (! list_empty (&subsys->CheckSumDescList)) 
+    {
+       EP_RXD      *rxd = (EP_RXD *) list_entry (subsys->CheckSumDescList.next, EP_RXD, CheckSumLink);
+       EP_ENVELOPE *env = &rxd->RxdMain->Envelope;
+
+       list_del_init (&rxd->CheckSumLink);
+       spin_unlock_irqrestore (&subsys->CheckSumDescLock, flags);
+
+       if (env->CheckSum) {
+           EP_NMD nmd;
+           uint32_t csum; 
+
+           ep_nmd_subset ( &nmd, &rxd->Data, 0, ep_rxd_len (rxd));
+
+           csum = ep_calc_check_sum(subsys->Subsys.Sys, env, &nmd, 1);
+           if ( env->CheckSum  != csum ) {
+               int f;
+       
+
+               printk("Check Sum Error: env(0x%x,0x%x) data(0x%x,0x%x)\n", ((csum >> 16) & 0x7FFF), ((env->CheckSum >> 16) & 0x7FFF), 
+                      (csum & 0xFFFF),  (env->CheckSum & 0xFFFF));
+               printk("Check Sum Error: Sent : NodeId %u Range 0x%x Service %u Version 0x%x Attr 0x%x\n", env->NodeId, env->Range, rxd->Rcvr->Service, env->Version, env->Attr);
+               printk("Check Sum Error: Sent : Xid Generation 0x%x Handle 0x%x Unique 0x%llx\n", env->Xid.Generation, env->Xid.Handle, env->Xid.Unique);
+               printk("Check Sum Error: Sent : TxdRail 0x%x TxdMain nmd_addr 0x%x  nmd_len %u  nmd_attr 0x%x\n",  env->TxdRail, env->TxdMain.nmd_addr, env->TxdMain.nmd_len, env->TxdMain.nmd_attr ); 
+               printk("Check Sum Error: Sent : nFrags %d \n", env->nFrags);
+               for(f=0;f<env->nFrags;f++)
+                   printk("Check Sum Error: Sent (%d): nmd_addr 0x%x   nmd_len %u   nmd_attr 0x%x\n", f,
+                          env->Frags[f].nmd_addr, env->Frags[f].nmd_len, env->Frags[f].nmd_attr);
+               printk("Check Sum Error: Recv : nmd_addr 0x%x   nmd_len %u   nmd_attr 0x%x\n",
+                      nmd.nmd_addr, nmd.nmd_len, nmd.nmd_attr);
+
+           }
+       }
+       ep_rxd_received_now(rxd);
+
+       spin_lock_irqsave (&subsys->CheckSumDescLock, flags);
+    }
+    spin_unlock_irqrestore (&subsys->CheckSumDescLock, flags);
+}
+#endif
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcommsRx.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcommsRx.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcommsRx.c       2005-06-01 23:12:54.645432112 -0400
@@ -0,0 +1,1205 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcommsRx.c,v 1.27.2.5 2004/11/30 12:02:16 mike Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/epcommsRx.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "debug.h"
+
+unsigned int ep_rxd_lowat = 5;
+
+static int
+AllocateRxdBlock (EP_RCVR *rcvr, EP_ATTRIBUTE attr, EP_RXD **rxdp)
+{
+    EP_RXD_BLOCK *blk;
+    EP_RXD       *rxd;
+    EP_RXD_MAIN  *pRxdMain;
+    int                  i;
+    unsigned long flags;
+
+    KMEM_ZALLOC (blk, EP_RXD_BLOCK *, sizeof (EP_RXD_BLOCK), ! (attr & EP_NO_SLEEP));
+
+    if (blk == NULL)
+       return (ENOMEM);
+
+    if ((pRxdMain = ep_shared_alloc_main (rcvr->Subsys->Subsys.Sys, EP_RXD_MAIN_SIZE * EP_NUM_RXD_PER_BLOCK, attr, &blk->NmdMain)) == (sdramaddr_t) 0)
+    {
+       KMEM_FREE (blk, sizeof (EP_RXD_BLOCK));
+       return (ENOMEM);
+    }
+    
+    for (rxd = &blk->Rxd[0], i = 0; i < EP_NUM_RXD_PER_BLOCK; i++, rxd++)
+    {
+       rxd->Rcvr        = rcvr;
+       rxd->RxdMain     = pRxdMain;
+
+       ep_nmd_subset (&rxd->NmdMain, &blk->NmdMain, (i * EP_RXD_MAIN_SIZE), EP_RXD_MAIN_SIZE);
+
+       /* move onto next descriptor */
+       pRxdMain = (EP_RXD_MAIN *) ((unsigned long) pRxdMain + EP_RXD_MAIN_SIZE);
+    }
+
+    spin_lock_irqsave (&rcvr->FreeDescLock, flags);
+
+    list_add  (&blk->Link, &rcvr->DescBlockList);
+
+    rcvr->TotalDescCount += EP_NUM_RXD_PER_BLOCK;
+
+    for (i = rxdp ? 1 : 0; i < EP_NUM_RXD_PER_BLOCK; i++)
+    {
+       
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+       INIT_LIST_HEAD (&blk->Rxd[i].CheckSumLink);
+#endif
+
+       list_add (&blk->Rxd[i].Link, &rcvr->FreeDescList);
+       
+       rcvr->FreeDescCount++;
+
+       if (rcvr->FreeDescWanted)
+       {
+           rcvr->FreeDescWanted--;
+           kcondvar_wakeupone (&rcvr->FreeDescSleep, &rcvr->FreeDescLock);
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->FreeDescLock, flags);
+    
+    if (rxdp)
+    {
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+       INIT_LIST_HEAD (&blk->Rxd[0].CheckSumLink);
+#endif
+              
+       *rxdp = &blk->Rxd[0];
+    }
+    return (ESUCCESS);
+}
+
+static void
+FreeRxdBlock (EP_RCVR *rcvr, EP_RXD_BLOCK *blk)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rcvr->FreeDescLock, flags);
+
+    list_del (&blk->Link);
+
+    rcvr->TotalDescCount -= EP_NUM_RXD_PER_BLOCK;
+    rcvr->FreeDescCount -= EP_NUM_RXD_PER_BLOCK;
+
+    spin_unlock_irqrestore (&rcvr->FreeDescLock, flags);
+
+    ep_shared_free_main (rcvr->Subsys->Subsys.Sys, &blk->NmdMain);
+    KMEM_FREE (blk, sizeof (EP_RXD_BLOCK));
+}
+
+static EP_RXD *
+GetRxd (EP_RCVR *rcvr, EP_ATTRIBUTE attr)
+{
+    EP_RXD *rxd;
+    unsigned long flags;
+    int low_on_rxds;
+
+    spin_lock_irqsave (&rcvr->FreeDescLock, flags);
+
+    while (list_empty (&rcvr->FreeDescList))
+    {
+       if (! (attr & EP_NO_ALLOC))
+       {
+           spin_unlock_irqrestore (&rcvr->FreeDescLock, flags);
+
+           if (AllocateRxdBlock (rcvr, attr, &rxd) == ESUCCESS)
+               return (rxd);
+
+           spin_lock_irqsave (&rcvr->FreeDescLock, flags);
+       }
+
+       if (attr & EP_NO_SLEEP)
+       {
+           IncrStat (rcvr->Subsys, NoFreeRxds);
+           spin_unlock_irqrestore (&rcvr->FreeDescLock, flags);
+
+           ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+           return (NULL);
+       }
+
+       rcvr->FreeDescWanted++;
+       kcondvar_wait (&rcvr->FreeDescSleep, &rcvr->FreeDescLock, &flags);
+    }
+
+    rxd = list_entry (rcvr->FreeDescList.next, EP_RXD, Link);
+
+    list_del (&rxd->Link);
+
+    /* Wakeup the descriptor primer thread if there's not many left */
+    low_on_rxds = (--rcvr->FreeDescCount < ep_rxd_lowat);
+
+    spin_unlock_irqrestore (&rcvr->FreeDescLock, flags);
+
+    if (low_on_rxds)
+       ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+
+    return (rxd);
+}
+
+static void
+FreeRxd (EP_RCVR *rcvr, EP_RXD *rxd)
+{
+    unsigned long flags;
+
+    ASSERT (EP_XID_INVALID(rxd->MsgXid));
+
+    spin_lock_irqsave (&rcvr->FreeDescLock, flags);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    ASSERT(list_empty(&rxd->CheckSumLink));
+#endif
+   
+    list_add (&rxd->Link, &rcvr->FreeDescList);
+
+    rcvr->FreeDescCount++;
+
+    if (rcvr->FreeDescWanted)                                  /* someone waiting for a receive */
+    {                                                          /* descriptor, so wake them up */
+       rcvr->FreeDescWanted--;
+       kcondvar_wakeupone (&rcvr->FreeDescSleep, &rcvr->FreeDescLock);
+    }
+    
+    spin_unlock_irqrestore (&rcvr->FreeDescLock, flags);
+}
+
+int
+ep_queue_receive (EP_RCVR *rcvr, EP_RXH *handler, void *arg, EP_NMD *nmd, EP_ATTRIBUTE attr)
+{
+    EP_RCVR_RAIL *rcvrRail;
+    EP_RXD       *rxd;
+    int           rnum;
+    unsigned long flags;
+
+    if ((rxd = GetRxd (rcvr, attr)) == NULL)
+       return (ENOMEM);
+
+    rxd->Handler      = handler;
+    rxd->Arg          = arg;
+    rxd->Data         = *nmd;
+    rxd->RxdMain->Len = EP_RXD_PENDING;
+    
+    spin_lock_irqsave (&rcvr->Lock, flags);
+
+    list_add_tail (&rxd->Link, &rcvr->ActiveDescList);
+    
+    if (EP_IS_PREFRAIL_SET(attr))
+       rnum = EP_ATTR2PREFRAIL(attr);
+    else 
+       rnum = ep_rcvr_prefrail (rcvr, EP_NMD_RAILMASK(nmd));
+
+    if (rnum < 0 || !(EP_NMD_RAILMASK(nmd) & EP_RAIL2RAILMASK(rnum) & rcvr->RailMask))
+       rcvrRail = NULL;
+    else
+       rcvrRail = rcvr->Rails[rnum];
+
+    EPRINTF7 (DBG_RCVR,"ep_queue_receive: rxd=%p svc %d nmd=%08x,%d,%x rnum=%d rcvrRail=%p\n",
+             rxd, rcvr->Service, nmd->nmd_addr, nmd->nmd_len, nmd->nmd_attr, rnum, rcvrRail);
+
+    rxd->State = EP_RXD_RECEIVE_ACTIVE;
+
+    if (rcvrRail == NULL || !EP_RCVR_OP (rcvrRail, QueueRxd) (rxd, rcvrRail))
+    {
+       rxd->State = EP_RXD_RECEIVE_UNBOUND;
+
+       ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+    }
+
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    return (ESUCCESS);
+}
+
+void
+ep_requeue_receive (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_NMD *nmd, EP_ATTRIBUTE attr)
+{
+    EP_RCVR      *rcvr = rxd->Rcvr;
+    EP_SYS       *sys  = rcvr->Subsys->Subsys.Sys;
+    int           rnum = ep_pickRail(EP_NMD_RAILMASK(&rxd->Data));
+    EP_RCVR_RAIL *rcvrRail;
+    unsigned long flags;
+
+    ASSERT (rxd->RxdRail == NULL);
+
+    EPRINTF5 (DBG_RCVR,"ep_requeue_receive: rxd=%p svc %d nmd=%08x,%d,%x\n", 
+             rxd, rcvr->Service, nmd->nmd_addr, nmd->nmd_len, nmd->nmd_attr);
+
+    rxd->Handler      = handler;
+    rxd->Arg          = arg;
+    rxd->Data         = *nmd;
+    rxd->RxdMain->Len = EP_RXD_PENDING;
+    
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    
+    list_add_tail (&rxd->Link, &rcvr->ActiveDescList);
+
+    /*
+     * Rail selection: if they've asked for a particular rail, then use it, otherwise if
+     *                 the rail it was last received on is mapped for the nmd and is available
+     *                 then use that one, otherwise pick one that is mapped by the nmd.
+     */
+    if (EP_IS_PREFRAIL_SET(attr))
+       rnum = EP_ATTR2PREFRAIL(attr);
+    
+    if (rnum < 0 || ! (EP_RAIL2RAILMASK (rnum) & EP_NMD_RAILMASK(nmd) & ep_rcvr_availrails (rcvr)))
+       rnum = ep_rcvr_prefrail (rcvr, EP_NMD_RAILMASK(nmd));
+
+    if (rnum < 0)
+       rcvrRail = NULL;
+    else
+    {
+       rcvrRail = rcvr->Rails[rnum];
+
+       if (! (EP_NMD_RAILMASK(&rxd->Data) & EP_RAIL2RAILMASK(rnum)) && ep_nmd_map_rails (sys, &rxd->Data, EP_RAIL2RAILMASK(rnum)) < 0)
+           rcvrRail = NULL;
+    }
+
+    rxd->State = EP_RXD_RECEIVE_ACTIVE;
+
+    if (rcvrRail == NULL || !EP_RCVR_OP(rcvrRail, QueueRxd) (rxd, rcvrRail))
+    {
+       EPRINTF1 (DBG_RCVR, "ep_requeue_receive: rcvrRail=%p - setting unbound\n", rcvrRail);
+
+       rxd->State = EP_RXD_RECEIVE_UNBOUND;
+
+       ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+    }
+
+    if (rcvr->CleanupWaiting)
+       kcondvar_wakeupall (&rcvr->CleanupSleep, &rcvr->Lock);
+    rcvr->CleanupWaiting = 0;
+
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+void
+
+ep_complete_receive (EP_RXD *rxd)
+{
+    EP_RCVR *rcvr = rxd->Rcvr;
+    unsigned long flags;
+
+    ASSERT (rxd->RxdRail == NULL && rxd->State == EP_RXD_COMPLETED);
+
+    FreeRxd (rcvr, rxd);
+
+    /* if we're waiting for cleanup, then wake them up */
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    if (rcvr->CleanupWaiting)
+       kcondvar_wakeupall (&rcvr->CleanupSleep, &rcvr->Lock);
+    rcvr->CleanupWaiting = 0;
+    spin_unlock_irqrestore (&rcvr->Lock, flags);   
+}
+
+int
+ep_rpc_put (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_NMD *local, EP_NMD *remote, int nFrags)
+{
+    EP_RCVR      *rcvr = rxd->Rcvr;
+    EP_SYS       *sys  = rcvr->Subsys->Subsys.Sys;
+    EP_ENVELOPE  *env  = &rxd->RxdMain->Envelope;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    
+    if (rxd->State == EP_RXD_BEEN_ABORTED)
+    {
+       EPRINTF2 (DBG_RCVR, "ep_rpc_put: rcvr %p rxd %p completed because no rails available\n", rcvr, rxd);
+       
+       /* rxd no longer on active list - just free it */
+       /* off and return an error */
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+       
+       return EP_CONN_RESET;
+    }
+    else
+    {
+       EP_RXD_RAIL      *rxdRail   = rxd->RxdRail;
+       EP_RCVR_RAIL     *rcvrRail  = rxdRail->RcvrRail;
+       EP_COMMS_RAIL    *commsRail = rcvrRail->CommsRail;
+       EP_RAIL          *rail      = commsRail->Rail;
+       EP_NODE_RAIL     *nodeRail  = &rail->Nodes[env->NodeId];
+       int               i;
+       
+       /* Attempt to ensure that the local nmds are mapped */
+       for (i = 0; i < nFrags; i++)
+           if (! (EP_NMD_RAILMASK(&local[i]) & EP_RAIL2RAILMASK(rail->Number)))
+               ep_nmd_map_rails (sys, &local[i], EP_RAIL2RAILMASK(rail->Number));
+    
+       if (nodeRail->State == EP_NODE_CONNECTED &&                                                                     /* rail is connected */
+           (ep_nmd2railmask (local, nFrags) & ep_nmd2railmask (remote, nFrags) & EP_RAIL2RAILMASK (rail->Number)))     /* and NMDs valid for it */
+       {
+           rxd->State = EP_RXD_PUT_ACTIVE;
+
+           EP_RCVR_OP(rcvrRail, RpcPut) (rxd, local, remote, nFrags);
+       }
+       else
+       {
+           /* RPC completion cannot progress - either node is no longer connected on this 
+            * rail or some of the source/destination NMDs are not mapped on this rail.
+            * Save the NMDs into the RXD and schedule the thread to request mappings */
+           EPRINTF4 (DBG_RCVR, "%s: ep_rpc_put: rcvr %p rxd %p %s\n", rail->Name, rcvr, rxd,
+                     (nodeRail->State == EP_NODE_CONNECTED) ? "NMDs not valid on this rail" : "no longer connected on this rail");
+
+           rxd->State = EP_RXD_PUT_STALLED;
+
+           if (nodeRail->State == EP_NODE_CONNECTED)
+               ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+       }
+
+       /* install the handler */
+       rxd->Handler = handler;
+       rxd->Arg     = arg;
+       
+       /* store the arguements */
+       rxd->nFrags = nFrags;
+       for (i = 0; i < nFrags; i++)
+       {
+           rxd->Local[i]  = local[i];
+           rxd->Remote[i] = remote[i];
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    return EP_SUCCESS;
+}
+
+int
+ep_rpc_get (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_NMD *remote, EP_NMD *local, int nFrags)
+{
+    EP_RCVR      *rcvr = rxd->Rcvr;
+    EP_SYS       *sys  = rcvr->Subsys->Subsys.Sys;
+    EP_ENVELOPE  *env  = &rxd->RxdMain->Envelope;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    
+    if (rxd->State == EP_RXD_BEEN_ABORTED)
+    {
+       EPRINTF2 (DBG_RCVR, "ep_rpc_get: rcvr %p rxd %p completed because no rails available\n", rcvr, rxd);
+       
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+       return EP_CONN_RESET;
+    }
+    else
+    {
+       EP_RXD_RAIL      *rxdRail   = rxd->RxdRail;
+       EP_RCVR_RAIL     *rcvrRail  = rxdRail->RcvrRail;
+       EP_COMMS_RAIL    *commsRail = rcvrRail->CommsRail;
+       EP_RAIL          *rail      = commsRail->Rail;
+       EP_NODE_RAIL     *nodeRail  = &rail->Nodes[env->NodeId];
+       int               i;
+       
+       /* Attempt to ensure that the local nmds are mapped */
+       for (i = 0; i < nFrags; i++)
+           if (! (EP_NMD_RAILMASK(&local[i]) & EP_RAIL2RAILMASK(rail->Number)))
+               ep_nmd_map_rails (sys, &local[i], EP_RAIL2RAILMASK(rail->Number));
+
+       if (nodeRail->State == EP_NODE_CONNECTED &&                                                                     /* rail is connected */
+           (ep_nmd2railmask (local, nFrags) & ep_nmd2railmask (remote, nFrags) & EP_RAIL2RAILMASK (rail->Number)))     /* and NMDs valid for it */
+       {
+           rxd->State = EP_RXD_GET_ACTIVE;
+
+           EP_RCVR_OP (rcvrRail, RpcGet) (rxd, local, remote, nFrags);
+       }
+       else
+       {
+           /* RPC completion cannot progress - either node is no longer connected on this 
+            * node or some of the source/destination NMDs are not mapped on this rail.
+            * Save the NMDs into the RXD and schedule the thread to request mappings */
+           EPRINTF4 (DBG_RCVR, "%s: ep_rpc_get: rcvr %p rxd %p %s\n", rail->Name, rcvr, rxd, 
+                     (nodeRail->State == EP_NODE_CONNECTED) ? "NMDs not valid on this rail" : "no longer connected on this rail");
+           
+           rxd->State = EP_RXD_GET_STALLED;
+
+           if (nodeRail->State == EP_NODE_CONNECTED)
+               ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+       }
+
+       /* install the handler */
+       rxd->Handler = handler;
+       rxd->Arg     = arg;
+       
+       /* store the arguements */
+       rxd->nFrags = nFrags;
+       for (i = 0; i < nFrags; i++)
+       {
+           rxd->Local[i]  = local[i];
+           rxd->Remote[i] = remote[i];
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+    
+    return EP_SUCCESS;
+}
+
+int
+ep_complete_rpc (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_STATUSBLK *blk, EP_NMD *local, EP_NMD *remote, int nFrags)
+{
+    EP_RCVR      *rcvr = rxd->Rcvr;
+    EP_SYS       *sys  = rcvr->Subsys->Subsys.Sys;
+    EP_ENVELOPE  *env  = &rxd->RxdMain->Envelope;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+
+    if (rxd->State == EP_RXD_BEEN_ABORTED)
+    {
+       EPRINTF2 (DBG_RCVR, "ep_complete_rpc: rcvr %p rxd %p completed because no rails available\n", rcvr, rxd);
+       
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+       return EP_CONN_RESET;
+    }
+    else
+    {
+       EP_RXD_RAIL      *rxdRail   = rxd->RxdRail;
+       EP_RCVR_RAIL     *rcvrRail  = rxdRail->RcvrRail;
+       EP_COMMS_RAIL    *commsRail = rcvrRail->CommsRail;
+       EP_RAIL          *rail      = commsRail->Rail;
+       EP_NODE_RAIL     *nodeRail  = &rail->Nodes[env->NodeId];
+       int               i;
+
+       if (blk == NULL)
+           bzero (&rxd->RxdMain->StatusBlk, sizeof (EP_STATUSBLK));
+       else
+           bcopy (blk, &rxd->RxdMain->StatusBlk, sizeof (EP_STATUSBLK));
+
+       /* Attempt to ensure that the local nmds are mapped */
+       for (i = 0; i < nFrags; i++)
+           if (! (EP_NMD_RAILMASK(&local[i]) & EP_RAIL2RAILMASK(rail->Number)))
+               ep_nmd_map_rails (sys, &local[i], EP_RAIL2RAILMASK(rail->Number));
+
+       if (nodeRail->State == EP_NODE_CONNECTED &&                                                                     /* rail is connected */
+           (ep_nmd2railmask (local, nFrags) & ep_nmd2railmask (remote, nFrags) & EP_RAIL2RAILMASK (rail->Number)))     /* and NMDs valid for it */
+       {
+           rxd->State = EP_RXD_COMPLETE_ACTIVE;
+
+           EP_RCVR_OP (rcvrRail, RpcComplete) (rxd, local, remote, nFrags);
+       }
+       else
+       {
+           /* RPC completion cannot progress - either node is no longer connected on this 
+            * node or some of the source/destination NMDs are not mapped on this rail.
+            * Save the NMDs into the RXD and schedule the thread to request mappings */
+           EPRINTF4 (DBG_RCVR, "%s: ep_complete_rpc: rcvr %p rxd %p %s\n", rail->Name, rcvr, rxd, 
+                     (nodeRail->State == EP_NODE_CONNECTED) ? "NMDs not valid on this rail" : "no longer connected on this rail");
+
+           rxd->State = EP_RXD_COMPLETE_STALLED;
+
+           if (nodeRail->State == EP_NODE_CONNECTED)
+               ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+       }
+
+       /* install the handler */
+       rxd->Handler = handler;
+       rxd->Arg     = arg;
+       
+       /* store the arguements */
+       rxd->nFrags = nFrags;
+       for (i = 0; i < nFrags; i++)
+       {
+           rxd->Local[i]  = local[i];
+           rxd->Remote[i] = remote[i];
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+       
+    return (ESUCCESS);
+}
+
+/* functions for accessing fields of rxds */
+void            *ep_rxd_arg(EP_RXD *rxd)               { return (rxd->Arg); }
+int              ep_rxd_len(EP_RXD *rxd)               { return (rxd->RxdMain->Len); }
+EP_STATUS       ep_rxd_status(EP_RXD *rxd)             { return (rxd->RxdMain->Len < 0 ? rxd->RxdMain->Len : EP_SUCCESS); }
+int              ep_rxd_isrpc(EP_RXD *rxd)             { return (EP_IS_RPC(rxd->RxdMain->Envelope.Attr) != 0); }
+EP_ENVELOPE     *ep_rxd_envelope(EP_RXD *rxd)          { return (&rxd->RxdMain->Envelope); }
+EP_PAYLOAD      *ep_rxd_payload(EP_RXD *rxd)           { return (EP_HAS_PAYLOAD(rxd->RxdMain->Envelope.Attr) ? &rxd->RxdMain->Payload : NULL); }
+int              ep_rxd_node(EP_RXD *rxd)              { return (rxd->RxdMain->Envelope.NodeId); }
+EP_STATUSBLK    *ep_rxd_statusblk(EP_RXD *rxd)                 { return (&rxd->RxdMain->StatusBlk); }
+EP_RAILMASK      ep_rxd_railmask(EP_RXD *rxd)          { return (rxd->Data.nmd_attr); }
+
+static void
+ProcessNmdMapResponse (EP_RCVR *rcvr, EP_RXD *rxd, EP_MANAGER_MSG *msg)
+{
+    EP_RXD_RAIL  *rxdRail  = rxd->RxdRail;
+    EP_RCVR_RAIL *rcvrRail = rxdRail->RcvrRail;
+    EP_RAIL      *rail     = rcvrRail->CommsRail->Rail;
+    EP_NODE_RAIL *nodeRail = &rail->Nodes[rxd->RxdMain->Envelope.NodeId];
+    int           i;
+
+    ASSERT (msg->Body.MapNmd.nFrags == rxd->nFrags);
+    
+    for (i = 0; i < rxd->nFrags; i++)
+       rxd->Remote[i] = msg->Body.MapNmd.Nmd[i];
+    
+    if (nodeRail->State == EP_NODE_CONNECTED &&        /* node is still connected on this rail */
+       (ep_nmd2railmask (rxd->Local, rxd->nFrags) & ep_nmd2railmask (rxd->Remote, rxd->nFrags) & EP_RAIL2RAILMASK (rail->Number)))     /* NMDs are now valid for this rail */
+    {
+       switch (rxd->State)
+       {
+       case EP_RXD_PUT_STALLED:
+           rxd->State = EP_RXD_PUT_ACTIVE;
+
+           EP_RCVR_OP(rcvrRail, RpcPut) (rxd, rxd->Local, rxd->Remote, rxd->nFrags);
+           break;
+
+       case EP_RXD_GET_STALLED:
+           rxd->State = EP_RXD_GET_ACTIVE;
+
+           EP_RCVR_OP(rcvrRail, RpcGet) (rxd, rxd->Local, rxd->Remote, rxd->nFrags);
+           break;
+           
+       case EP_RXD_COMPLETE_STALLED:
+           rxd->State = EP_RXD_COMPLETE_ACTIVE;
+
+           EP_RCVR_OP(rcvrRail, RpcComplete) (rxd, rxd->Local, rxd->Remote, rxd->nFrags);
+           break;
+
+       default:
+           panic ("ProcessNmdMapResponse: XID match but rxd in invalid state\n");
+           break;
+       }
+
+       rxd->NextRunTime = 0;
+    }
+    else
+       ep_debugf (DBG_MANAGER, "%s: ep_rcvr_xid_msg_handler: rcvr=%p rxd=%p - still cannot proceed\n", rail->Name, rcvr, rxd);
+}
+
+static void
+ProcessFailoverResponse (EP_RCVR *rcvr, EP_RXD *rxd, EP_MANAGER_MSG *msg)
+{
+    /* XXXX - TBD */
+#ifdef NOTYET
+    EP_COMMS_SUBSYS *subsys   = rcvr->Subsys;
+    EP_RXD_RAIL     *rxdRail  = rxd->RxdRail;
+    EP_RCVR_RAIL    *rcvrRail = rxdRail->RcvrRail;
+    EP_RAIL         *rail     = rcvrRail->CommsRail->Rail;
+    EP_RCVR_RAIL    *nRcvrRail;
+    EP_RXD_RAIL     *nRxdRail;
+
+    ASSERT (rxd->RxdMain->Envelope.Attr & EP_RPC);
+
+    EPRINTF6 (DBG_RCVR, "ep_rcvr_xid_msg_handler: rcvr=%p rxd=%p Xid=%016llx state %x.%x - txd on rail %d\n", rcvr, rxd, 
+             rxd->MsgXid.Unique, rxdRail->RxdMain->DataEvent, rxdRail->RxdMain->DoneEvent, msg->Body.FailoverTxd.Rail);
+
+    if ((nRcvrRail = rcvr->Rails[msg->Body.FailoverTxd.Rail]) == NULL ||
+       (nRcvrRail->Rcvr->RailMask & EP_RAIL2RAILMASK (rail->Number)) == NULL)
+    {
+       ep_debugf (DBG_MANAGER, "%s: ep_rcvr_xid_msg_handler: rcvr=%p rxd=%p - still cannot proceed\n", rail->Name, rcvr,rxd);
+       return;
+    }
+
+
+    nRxdRail = EP_RCVR_OP (nrcvrRail, GetRxd) (rcvr, nRcvrRail);
+
+
+    /* If the RPC was in progress, then rollback and mark it as flagged, 
+     * this will then get treated as though the NMDs were not mapped
+     * for the rail when the user initiated the operation.
+     */
+    switch (rxdRail->RxdMain->DataEvent)
+    {
+    case EP_EVENT_ACTIVE|EP_RXD_PHASE_PUT:
+    case EP_EVENT_FLAGGED|EP_RXD_PHASE_PUT:
+       ASSERT (rxdRail->RxdMain->DoneEvent == EP_EVENT_PRIVATE ||
+               rxdRail->RxdMain->DoneEvent == EP_EVENT_PENDING);
+       
+       nRxdRail->RxdMain->DataEvent = EP_EVENT_FLAGGED|EP_RXD_PHASE_PUT;
+       nRxdRail->RxdMain->DoneEvent = EP_EVENT_PENDING;
+       break;
+
+    case EP_EVENT_ACTIVE|EP_RXD_PHASE_GET:
+    case EP_EVENT_FLAGGED|EP_RXD_PHASE_GET:
+       ASSERT (rxdRail->RxdMain->DoneEvent == EP_EVENT_PRIVATE ||
+               rxdRail->RxdMain->DoneEvent == EP_EVENT_PENDING);
+       
+       nRxdRail->RxdMain->DataEvent = EP_EVENT_FLAGGED|EP_RXD_PHASE_GET;
+       nRxdRail->RxdMain->DoneEvent = EP_EVENT_PENDING;
+       break;
+
+    case EP_EVENT_PRIVATE:
+       switch (rxdRail->RxdMain->DoneEvent)
+       {
+       case EP_EVENT_ACTIVE|EP_RXD_PHASE_COMPLETE:
+       case EP_EVENT_FLAGGED|EP_RXD_PHASE_COMPLETE:
+           nRxdRail->RxdMain->DataEvent = EP_EVENT_PRIVATE;
+           nRxdRail->RxdMain->DoneEvent = EP_EVENT_FLAGGED|EP_RXD_PHASE_COMPLETE;
+           break;
+
+       case EP_EVENT_PENDING:
+           break;
+
+       default:
+           panic ("ep_rcvr_xid_msg_handler: rxd in invalid state\n");
+       }
+       break;
+
+    default:
+       panic ("ep_rcvr_xid_msg_handler: rxd in invalid staten");
+    }
+    
+    UnbindRxdFromRail (rxd, rxdRail);
+
+    /* Mark rxdRail as no longer active */
+    rxdRail->RxdMain->DataEvent = EP_EVENT_PRIVATE;
+    rxdRail->RxdMain->DoneEvent = EP_EVENT_PRIVATE;
+
+    sdram_writel (rail->Device, rxdRail->RxdElan + offsetof (EP_RXD_RAIL_ELAN, DataEvent.ev_Count), 0);
+    sdram_writel (rail->Device, rxdRail->RxdElan + offsetof (EP_RXD_RAIL_ELAN, DoneEvent.ev_Count), 0);
+    
+    FreeRxdRail (rcvrRail, rxdRail);
+
+    BindRxdToRail (rxd, nRxdRail);
+
+    ep_kthread_schedule (&subsys->Thread, lbolt);
+#endif
+}
+
+void
+ep_rcvr_xid_msg_handler (void *arg, EP_MANAGER_MSG *msg)
+{
+    EP_RCVR          *rcvr = (EP_RCVR *) arg;
+    struct list_head *el;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    list_for_each (el, &rcvr->ActiveDescList) {
+       EP_RXD *rxd = list_entry (el,EP_RXD, Link);
+
+       if (EP_XIDS_MATCH (msg->Hdr.Xid, rxd->MsgXid))
+       {
+           EP_INVALIDATE_XID (rxd->MsgXid);
+
+           switch (msg->Hdr.Type)
+           {
+           case EP_MANAGER_MSG_TYPE_MAP_NMD_RESPONSE:
+               ProcessNmdMapResponse (rcvr, rxd, msg);
+               break;
+
+           case EP_MANAGER_MSG_TYPE_FAILOVER_RESPONSE:
+               ProcessFailoverResponse (rcvr, rxd, msg);
+               break;
+
+           default:
+               panic ("ep_rcvr_xid_msg_handler: XID match but invalid message type\n");
+           }
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+
+EP_RCVR *
+ep_alloc_rcvr (EP_SYS *sys, EP_SERVICE svc, unsigned int nenvs)
+{
+    EP_COMMS_SUBSYS  *subsys;
+    EP_RCVR          *rcvr;
+    struct list_head *el;
+    extern int portals_envelopes;
+
+    if (portals_envelopes && (svc == EP_MSG_SVC_PORTALS_SMALL || svc == EP_MSG_SVC_PORTALS_LARGE))
+    {
+       printk ("ep: use %d envelopes rather than %d for portals %s message service\n", sys->Position.pos_nodes * 16, nenvs,
+               svc == EP_MSG_SVC_PORTALS_SMALL ? "small" : "large");
+
+       nenvs = portals_envelopes;
+    }
+
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (sys, EPCOMMS_SUBSYS_NAME)) == NULL)
+       return (NULL);
+
+    KMEM_ZALLOC (rcvr, EP_RCVR *, sizeof (EP_RCVR), 1);
+
+    if (rcvr == NULL)
+       return (NULL);
+
+    rcvr->Subsys            = subsys;
+    rcvr->Service           = svc;
+    rcvr->InputQueueEntries = nenvs;
+    rcvr->FreeDescCount     = 0;
+    rcvr->TotalDescCount    = 0;
+    rcvr->ForwardRxdCount   = 0;
+
+    spin_lock_init (&rcvr->Lock);
+    INIT_LIST_HEAD (&rcvr->ActiveDescList);
+
+    kcondvar_init (&rcvr->CleanupSleep);
+    kcondvar_init (&rcvr->FreeDescSleep);
+    spin_lock_init (&rcvr->FreeDescLock);
+    INIT_LIST_HEAD (&rcvr->FreeDescList);
+    INIT_LIST_HEAD (&rcvr->DescBlockList);
+
+    ep_xid_cache_init (sys, &rcvr->XidCache);
+
+    rcvr->XidCache.MessageHandler = ep_rcvr_xid_msg_handler;
+    rcvr->XidCache.Arg            = rcvr;
+
+    kmutex_lock (&subsys->Lock);
+    /* See if this service is already in use */
+    list_for_each (el, &subsys->Receivers) {
+       EP_RCVR *rcvr = list_entry (el, EP_RCVR, Link);
+
+       if (rcvr->Service == svc)
+       {
+           KMEM_FREE (rcvr, sizeof (EP_RCVR));
+           kmutex_unlock (&subsys->Lock);   
+           return NULL;
+       }
+    }
+    
+    
+    list_add_tail (&rcvr->Link, &subsys->Receivers);
+
+    ep_procfs_rcvr_add(rcvr);
+
+    /* Now add all rails which are already started */
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       EP_RAIL_OP (commsRail, Rcvr.AddRail) (rcvr, commsRail);
+    }
+    kmutex_unlock (&subsys->Lock);   
+
+    ep_mod_inc_usecount();
+
+    return (rcvr);
+}
+
+void
+ep_free_rcvr (EP_RCVR *rcvr)
+{
+    EP_COMMS_SUBSYS  *subsys = rcvr->Subsys;
+    EP_SYS           *sys    = subsys->Subsys.Sys;
+    struct list_head  list;
+    struct list_head *el,*nel;
+    unsigned long flags;
+    
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+       
+       EP_RAIL_OP (commsRail, Rcvr.DelRail) (rcvr, commsRail);
+    }
+
+    ep_procfs_rcvr_del(rcvr);
+
+    list_del (&rcvr->Link);
+    kmutex_unlock (&subsys->Lock);
+
+    INIT_LIST_HEAD (&list);
+
+    /* abort all rxds - should not be bound to a rail */
+    spin_lock_irqsave (&rcvr->Lock, flags);   
+    for (;;)
+    {
+       if (! list_empty (&rcvr->ActiveDescList))
+       {
+           list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+               EP_RXD *rxd = list_entry (el, EP_RXD, Link);
+               
+               ASSERT (rxd->RxdRail == NULL);
+               ASSERT (rxd->RxdMain->Len == EP_RXD_PENDING);
+               
+               rxd->State = EP_RXD_COMPLETED;
+               rxd->RxdMain->Len = EP_SHUTDOWN;
+               
+               list_del (&rxd->Link);
+               list_add_tail (&rxd->Link, &list);
+           }
+           spin_unlock_irqrestore (&rcvr->Lock, flags);
+           
+           while (! list_empty (&list))
+           {
+               EP_RXD *rxd = list_entry (list.next, EP_RXD, Link);
+               
+               list_del (&rxd->Link);
+               
+               if (rxd->Handler) 
+                   rxd->Handler (rxd);
+           }
+           spin_lock_irqsave (&rcvr->Lock, flags);   
+           continue;
+       }
+
+       if (rcvr->FreeDescCount == rcvr->TotalDescCount)
+           break;
+
+       rcvr->CleanupWaiting++;
+       kcondvar_wait (&rcvr->CleanupSleep, &rcvr->Lock, &flags);
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    /* must all be in free list */
+    ASSERT( rcvr->FreeDescCount ==  rcvr->TotalDescCount);
+
+    while (! list_empty(& rcvr->DescBlockList) )
+       FreeRxdBlock (rcvr, list_entry (rcvr->DescBlockList.next, EP_RXD_BLOCK, Link));
+
+    /* had better be all gone now */
+    ASSERT((rcvr->FreeDescCount == 0) && (rcvr->TotalDescCount == 0));
+
+    ep_xid_cache_destroy (sys, &rcvr->XidCache);
+ 
+    spin_lock_destroy (&rcvr->Lock);
+    KMEM_FREE (rcvr, sizeof (EP_RCVR));
+
+    ep_mod_dec_usecount();
+}
+
+EP_RXD *
+StealRxdFromOtherRail (EP_RCVR *rcvr)
+{
+    EP_RXD          *rxd;
+    int               i;
+       
+    /* looking at the the rcvr railmask to find a rail to try to steal rxd from */
+    for (i = 0; i < EP_MAX_RAILS; i++) 
+       if (rcvr->RailMask & (1 << i) ) 
+           if ((rxd = EP_RCVR_OP (rcvr->Rails[i], StealRxd) (rcvr->Rails[i])) != NULL)
+               return rxd;
+
+    return NULL;
+}
+
+long
+CheckUnboundRxd (EP_RCVR *rcvr, EP_RXD *rxd, long nextRunTime)
+{
+    EP_SYS       *sys = rcvr->Subsys->Subsys.Sys;
+    EP_RCVR_RAIL *rcvrRail;
+    int           rnum;
+    
+    if ((rnum = ep_rcvr_prefrail (rcvr, EP_NMD_RAILMASK(&rxd->Data))) < 0)
+       rnum = ep_rcvr_prefrail (rcvr, ep_rcvr_availrails (rcvr));
+    
+    if ( rnum < 0 )    {
+       if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + RESOURCE_RETRY_TIME))
+           nextRunTime = lbolt + RESOURCE_RETRY_TIME;
+       
+       return (nextRunTime);
+    }
+
+    ASSERT ( rnum >= 0 );
+
+    rcvrRail = rcvr->Rails[rnum];
+
+    ASSERT ( rcvrRail != NULL);
+
+    rxd->State = EP_RXD_RECEIVE_ACTIVE;
+
+    if ((!(EP_NMD_RAILMASK (&rxd->Data) & EP_RAIL2RAILMASK(rnum)) &&                   /* not mapped already and */
+        ep_nmd_map_rails (sys, &rxd->Data, EP_RAIL2RAILMASK(rnum)) == 0) ||            /* failed mapping, or */
+       !EP_RCVR_OP (rcvrRail, QueueRxd) (rxd, rcvrRail))                               /* failed to queue */
+    {
+       ASSERT (rxd->RxdRail == NULL);
+
+       EPRINTF4 (DBG_RCVR,"CheckUnboundRxd: rcvr=%p rxd=%p -> rnum=%d rcvrRail=%p (failed)\n", rcvr, rxd, rnum, rcvrRail);
+
+       rxd->State = EP_RXD_RECEIVE_UNBOUND;
+       
+       if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + RESOURCE_RETRY_TIME))
+           nextRunTime = lbolt + RESOURCE_RETRY_TIME;
+    }
+
+    return (nextRunTime);
+}
+
+int
+CheckRxdNmdsMapped (EP_RCVR *rcvr, EP_RXD *rxd)
+{
+    EP_RXD_RAIL *rxdRail = rxd->RxdRail;
+    EP_RXD_MAIN *rxdMain = rxd->RxdMain;
+    EP_ENVELOPE *env     = &rxdMain->Envelope;
+    EP_SYS      *sys     = rcvr->Subsys->Subsys.Sys;
+    EP_RAIL     *rail    = rxdRail->RcvrRail->CommsRail->Rail;
+    int                 i;
+
+    /* Try and map the local NMDs before checking to see if we can proceed */
+    if (! (ep_nmd2railmask (rxd->Local, rxd->nFrags) & EP_RAIL2RAILMASK (rail->Number)))
+    {
+       EPRINTF3 (DBG_MAPNMD, "%s: rcvr=%p rxd=%p RPC Local NMDs not mapped\n", rail->Name, rcvr, rxd);
+       
+       for (i = 0; i < rxd->nFrags; i++)
+           if (! (EP_NMD_RAILMASK(&rxd->Local[i]) & EP_RAIL2RAILMASK(rail->Number)))
+               if (ep_nmd_map_rails (sys, &rxd->Local[i], EP_RAIL2RAILMASK(rail->Number)))
+                   rxd->NextRunTime = lbolt + RESOURCE_RETRY_TIME;
+    }
+    
+    /* Try and map remote NMDs if they are not valid for this rail */
+    if (! (ep_nmd2railmask (rxd->Remote, rxd->nFrags) & EP_RAIL2RAILMASK (rail->Number)))
+    {
+       EP_MANAGER_MSG_BODY msgBody;
+
+       EPRINTF3 (DBG_MAPNMD, "%s: rcvr=%p rxd=%p RPC Remote NMDs not mapped\n", rail->Name, rcvr, rxd);
+
+       if (EP_XID_INVALID(rxd->MsgXid))
+           rxd->MsgXid = ep_xid_cache_alloc (sys, &rcvr->XidCache);
+
+       msgBody.MapNmd.nFrags   = rxd->nFrags;
+       msgBody.MapNmd.Railmask = EP_RAIL2RAILMASK (rail->Number);
+       for (i = 0; i < rxd->nFrags; i++)
+           msgBody.MapNmd.Nmd[i] = rxd->Remote[i];
+
+       if (ep_send_message (rail, env->NodeId, EP_MANAGER_MSG_TYPE_MAP_NMD_REQUEST, rxd->MsgXid, &msgBody) == 0)
+           rxd->NextRunTime = lbolt + MESSAGE_RETRY_TIME;
+       else
+           rxd->NextRunTime = lbolt + MSGBUSY_RETRY_TIME;
+
+       return 0;
+    }
+
+    if ((ep_nmd2railmask (rxd->Local, rxd->nFrags) & ep_nmd2railmask (rxd->Remote, rxd->nFrags) & EP_RAIL2RAILMASK (rail->Number)) != 0)
+    {
+       rxd->NextRunTime = 0;
+       return 1;
+    }
+
+    return 0;
+}
+
+long
+ep_check_rcvr (EP_RCVR *rcvr, long nextRunTime)
+{
+    struct list_head *el, *nel;
+    unsigned long     flags;
+    int               i;
+
+    /* Check to see if we're low on rxds */
+    if (rcvr->FreeDescCount < ep_rxd_lowat)
+       AllocateRxdBlock (rcvr, 0, NULL);
+
+    for (i = 0; i < EP_MAX_RAILS; i++) 
+       if (rcvr->RailMask & (1 << i) )
+           nextRunTime = EP_RCVR_OP (rcvr->Rails[i], Check) (rcvr->Rails[i], nextRunTime);
+
+    /* See if we have any rxd's which need to be handled */
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD      *rxd     = list_entry (el, EP_RXD, Link);
+       EP_RXD_MAIN *rxdMain = rxd->RxdMain;
+       EP_ENVELOPE *env     = &rxdMain->Envelope;
+       EP_RXD_RAIL *rxdRail = rxd->RxdRail;
+
+       if (rxdRail == NULL)
+           nextRunTime = CheckUnboundRxd (rcvr, rxd, nextRunTime);
+       else
+       {
+           EP_RCVR_RAIL *rcvrRail = rxdRail->RcvrRail;
+           EP_RAIL      *rail     = rcvrRail->CommsRail->Rail;
+
+           if (rxd->RxdMain->Len == EP_RXD_PENDING ||                          /* envelope not received yet */
+               rail->Nodes[env->NodeId].State != EP_NODE_CONNECTED)            /* will be failing over */
+               continue;
+
+           switch (rxd->State)
+           {
+           case EP_RXD_PUT_STALLED:
+               if (CheckRxdNmdsMapped (rcvr, rxd))
+               {
+                   rxd->State = EP_RXD_PUT_ACTIVE;
+
+                   EP_RCVR_OP (rcvrRail, RpcPut) (rxd, rxd->Local, rxd->Remote, rxd->nFrags);
+               }
+               break;
+
+           case EP_RXD_GET_STALLED:
+               if (CheckRxdNmdsMapped (rcvr, rxd))
+               {
+                   rxd->State = EP_RXD_GET_ACTIVE;
+
+                   EP_RCVR_OP (rcvrRail, RpcGet) (rxd, rxd->Local, rxd->Remote, rxd->nFrags);
+               }
+               break;
+           
+           case EP_RXD_COMPLETE_STALLED:
+               if (CheckRxdNmdsMapped (rcvr, rxd))
+               {
+                   rxd->State = EP_RXD_COMPLETE_ACTIVE;
+
+                   EP_RCVR_OP (rcvrRail, RpcComplete)(rxd, rxd->Local, rxd->Remote, rxd->nFrags);
+               }
+               break;
+           }
+               
+           if (rxd->NextRunTime && (nextRunTime == 0 || AFTER (nextRunTime, rxd->NextRunTime)))
+               nextRunTime = rxd->NextRunTime;
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+    
+    return (nextRunTime);
+}
+
+void
+ep_display_rxd (DisplayInfo *di, EP_RXD *rxd)
+{
+    EP_RXD_MAIN *rxdMain  = rxd->RxdMain;
+    EP_ENVELOPE *env      = &rxdMain->Envelope;
+    EP_RXD_RAIL *rxdRail  = rxd->RxdRail;
+    
+    (di->func)(di->arg, "  RXD: %p State=%x RxdMain=%p(%x.%x.%x) Data=%x.%x.%x %s\n", rxd,
+              rxd->State, rxd->RxdMain, rxd->NmdMain.nmd_addr, rxd->NmdMain.nmd_len,
+              rxd->NmdMain.nmd_attr, rxd->Data.nmd_addr, rxd->Data.nmd_len, rxd->Data.nmd_attr,
+              rxd->RxdMain->Len == EP_RXD_PENDING ? "Pending" : "Active");
+    (di->func)(di->arg, "      NodeId=%d Range=%d.%d TxdRail=%x TxdMain=%x.%x.%x nFrags=%d XID=%08x.%08x.%016llx\n",
+              env->NodeId,  EP_RANGE_LOW(env->Range), EP_RANGE_HIGH(env->Range), env->TxdRail, env->TxdMain.nmd_addr,
+              env->TxdMain.nmd_len, env->TxdMain.nmd_attr, env->nFrags, env->Xid.Generation, env->Xid.Handle, env->Xid.Unique);;
+    (di->func)(di->arg, "      Frag[0] %08x.%08x.%08x\n", env->Frags[0].nmd_addr, env->Frags[0].nmd_len, env->Frags[0].nmd_attr);
+    (di->func)(di->arg, "      Frag[1] %08x.%08x.%08x\n", env->Frags[1].nmd_addr, env->Frags[1].nmd_len, env->Frags[1].nmd_attr);
+    (di->func)(di->arg, "      Frag[2] %08x.%08x.%08x\n", env->Frags[2].nmd_addr, env->Frags[2].nmd_len, env->Frags[2].nmd_attr);
+    (di->func)(di->arg, "      Frag[3] %08x.%08x.%08x\n", env->Frags[3].nmd_addr, env->Frags[3].nmd_len, env->Frags[3].nmd_attr);
+
+    if (rxdRail) EP_RCVR_OP (rxdRail->RcvrRail, DisplayRxd) (di, rxdRail);
+}
+
+void
+ep_display_rcvr (DisplayInfo *di, EP_RCVR *rcvr, int full)
+{
+    int               freeCount    = 0;
+    int                      activeCount  = 0;
+    int                      pendingCount = 0;
+    int                      railCounts[EP_MAX_RAILS];
+    struct list_head *el;
+    int               i;
+    unsigned long     flags;
+
+    for (i = 0; i <EP_MAX_RAILS; i++)
+       railCounts[i] = 0;
+
+    spin_lock_irqsave (&rcvr->FreeDescLock, flags);
+    list_for_each (el, &rcvr->FreeDescList)
+       freeCount++;
+    spin_unlock_irqrestore (&rcvr->FreeDescLock, flags);
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    list_for_each (el, &rcvr->ActiveDescList) {
+       EP_RXD      *rxd     = list_entry (el, EP_RXD, Link);
+       EP_RXD_RAIL *rxdRail = rxd->RxdRail;
+
+       if (rxd->RxdMain->Len == EP_RXD_PENDING)
+           pendingCount++;
+       else
+           activeCount++;
+
+       if (rxdRail)
+           railCounts[rxdRail->RcvrRail->CommsRail->Rail->Number]++;
+    }
+
+    (di->func)(di->arg, "RCVR: rcvr=%p number=%d\n", rcvr, rcvr->Service);
+    (di->func)(di->arg, "      RXDS Free=%d (%d) Pending=%d Active=%d Rails=%d.%d.%d.%d\n",
+              freeCount, rcvr->FreeDescCount, pendingCount, activeCount, railCounts[0], railCounts[1],
+              railCounts[2], railCounts[3]);
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+       if (rcvr->Rails[i] != NULL)
+           EP_RCVR_OP (rcvr->Rails[i], DisplayRcvr) (di, rcvr->Rails[i]);
+
+    list_for_each (el, &rcvr->ActiveDescList) {
+       EP_RXD *rxd = list_entry (el, EP_RXD, Link);
+
+       if (rxd->RxdMain->Len != EP_RXD_PENDING || full)
+           ep_display_rxd (di, rxd);
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+void
+ep_rxd_received_now(EP_RXD *rxd)
+{
+    EP_ENVELOPE   *env  = &rxd->RxdMain->Envelope;
+    EP_RCVR       *rcvr = rxd->Rcvr;
+    unsigned long  flags;
+
+    INC_STAT(rcvr->stats,rx);
+    ADD_STAT(rcvr->stats,rx_len, rxd->RxdMain->Len);
+
+    if (rxd->RxdMain->Len < 0 || !EP_IS_MULTICAST(env->Attr))
+    {
+       rxd->Handler (rxd);
+    }
+    else
+    {
+       EPRINTF5 (DBG_RCVR, "ep_rxd_received: forward rxd=%p Data=%08x.%08x.%08x len=%d\n", rxd, 
+                 rxd->Data.nmd_addr, rxd->Data.nmd_len, rxd->Data.nmd_attr, ep_rxd_len(rxd));
+
+       spin_lock_irqsave (&rcvr->Subsys->ForwardDescLock, flags);
+       list_add_tail (&rxd->Link, &rcvr->Subsys->ForwardDescList);
+       spin_unlock_irqrestore (&rcvr->Subsys->ForwardDescLock, flags);
+       
+       ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+    } 
+}
+
+#if defined(CONFIG_EP_NO_CHECK_SUM)
+void
+ep_rxd_received(EP_RXD *rxd) 
+{
+   ep_rxd_received_now(rxd);
+}
+
+#else
+
+void
+ep_rxd_received(EP_RXD *rxd) 
+{
+  EP_ENVELOPE   *env  = &rxd->RxdMain->Envelope;
+
+  if (env->CheckSum) 
+      ep_rxd_queue_csum(rxd);
+  else 
+      ep_rxd_received_now(rxd);
+}
+
+void
+ep_rxd_queue_csum(EP_RXD *rxd)
+{
+    EP_RCVR       *rcvr = rxd->Rcvr;
+    unsigned long flags;
+
+    EPRINTF5 (DBG_RCVR, "ep_rxd_queue_csum: rxd=%p Data=%08x.%08x.%08x len=%d\n", rxd, 
+             rxd->Data.nmd_addr, rxd->Data.nmd_len, rxd->Data.nmd_attr, ep_rxd_len(rxd));
+    
+    spin_lock_irqsave (&rcvr->Subsys->CheckSumDescLock, flags);
+    list_add_tail (&rxd->CheckSumLink, &rcvr->Subsys->CheckSumDescList);
+    spin_unlock_irqrestore (&rcvr->Subsys->CheckSumDescLock, flags);
+    
+    ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+}
+#endif
+
+void
+ep_rcvr_fillout_stats(EP_RCVR *rcvr, char *str)
+{
+    sprintf(str+strlen(str),"Rx     %lu  %lu /sec\n",       GET_STAT_TOTAL(rcvr->stats,rx),      GET_STAT_PER_SEC(rcvr->stats,rx) );
+    sprintf(str+strlen(str),"MBytes %lu  %lu Mbytes/sec\n", GET_STAT_TOTAL(rcvr->stats,rx_len) / (1024*1024),  GET_STAT_PER_SEC(rcvr->stats,rx_len) / (1024*1024));
+}
+
+void
+ep_rcvr_rail_fillout_stats(EP_RCVR_RAIL *rcvr_rail, char *str)
+{
+    sprintf(str+strlen(str),"Rx     %lu  %lu /sec\n",       GET_STAT_TOTAL(rcvr_rail->stats,rx),      GET_STAT_PER_SEC(rcvr_rail->stats,rx) );
+    sprintf(str+strlen(str),"MBytes %lu  %lu Mbytes/sec\n", GET_STAT_TOTAL(rcvr_rail->stats,rx_len) / (1024*1024),  GET_STAT_PER_SEC(rcvr_rail->stats,rx_len) / (1024*1024));
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcommsRx_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcommsRx_elan3.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcommsRx_elan3.c 2005-06-01 23:12:54.649431504 -0400
@@ -0,0 +1,1776 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcommsRx_elan3.c,v 1.19.2.3 2004/11/15 11:05:49 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcommsRx_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "epcomms_elan3.h"
+#include "debug.h"
+
+#define RCVR_TO_RAIL(rcvrRail)         ((EP3_RAIL *) ((EP_RCVR_RAIL *) rcvrRail)->CommsRail->Rail)
+#define RCVR_TO_DEV(rcvrRail)          (RCVR_TO_RAIL(rcvrRail)->Device)
+#define RCVR_TO_SUBSYS(rcvrRail)       (((EP_RCVR_RAIL *) rcvrRail)->Rcvr->Subsys)
+
+static void RxDataEvent (EP3_RAIL *rail, void *arg);
+static void RxDataRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status);
+static void RxDataVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma);
+
+static EP3_COOKIE_OPS RxDataCookieOps = 
+{
+    RxDataEvent,
+    RxDataRetry,
+    NULL, /* DmaCancelled */
+    RxDataVerify,
+};
+
+static void RxDoneEvent (EP3_RAIL *rail, void *arg);
+static void RxDoneRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status);
+static void RxDoneVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma);
+
+static EP3_COOKIE_OPS RxDoneCookieOps = 
+{
+    RxDoneEvent,
+    RxDoneRetry,
+    NULL, /* DmaCancelled */
+    RxDoneVerify,
+};
+
+static int
+AllocateRxdRailBlock (EP3_RCVR_RAIL *rcvrRail)
+{
+    EP3_RAIL         *rail = RCVR_TO_RAIL(rcvrRail);
+    ELAN3_DEV         *dev  = rail->Device;
+    EP3_RXD_RAIL_BLOCK *blk;
+    EP3_RXD_RAIL       *rxdRail;
+    sdramaddr_t        pRxdElan;
+    EP3_RXD_RAIL_MAIN  *pRxdMain;
+    E3_Addr           pRxdElanAddr;
+    E3_Addr           pRxdMainAddr;
+    E3_BlockCopyEvent  event;
+    int                       i, j;
+    unsigned long      flags;
+
+    KMEM_ZALLOC (blk, EP3_RXD_RAIL_BLOCK *, sizeof (EP3_RXD_RAIL_BLOCK), 1);
+    if (blk == NULL)
+       return 0;
+
+    if ((pRxdElan = ep_alloc_elan (&rail->Generic, EP3_RXD_RAIL_ELAN_SIZE * EP3_NUM_RXD_PER_BLOCK, 0, &pRxdElanAddr)) == (sdramaddr_t) 0)
+    {
+       KMEM_FREE (blk, sizeof (EP3_RXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    if ((pRxdMain = ep_alloc_main (&rail->Generic, EP3_RXD_RAIL_MAIN_SIZE * EP3_NUM_RXD_PER_BLOCK, 0, &pRxdMainAddr)) == (sdramaddr_t) 0)
+    {
+       ep_free_elan (&rail->Generic, pRxdElanAddr, EP3_RXD_RAIL_ELAN_SIZE * EP3_NUM_RXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP3_RXD_RAIL_BLOCK));
+       return 0;
+    }
+    
+    if (ReserveDmaRetries (rail, EP3_NUM_RXD_PER_BLOCK, 0) != ESUCCESS)
+    {
+       ep_free_main (&rail->Generic, pRxdMainAddr, EP3_RXD_RAIL_MAIN_SIZE * EP3_NUM_RXD_PER_BLOCK);
+       ep_free_elan (&rail->Generic, pRxdElanAddr, EP3_RXD_RAIL_ELAN_SIZE * EP3_NUM_RXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP3_RXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    for (rxdRail = &blk->Rxd[0], i = 0; i < EP3_NUM_RXD_PER_BLOCK; i++, rxdRail++)
+    {
+       rxdRail->Generic.RcvrRail = (EP_RCVR_RAIL *) rcvrRail;
+       rxdRail->RxdElan          = pRxdElan;
+       rxdRail->RxdElanAddr      = pRxdElanAddr;
+       rxdRail->RxdMain          = pRxdMain;
+       rxdRail->RxdMainAddr      = pRxdMainAddr;
+
+       elan3_sdram_writel (dev, pRxdElan + offsetof (EP3_RXD_RAIL_ELAN, RxdMain),  0);
+       elan3_sdram_writel (dev, pRxdElan + offsetof (EP3_RXD_RAIL_ELAN, Next),     0);
+       elan3_sdram_writeq (dev, pRxdElan + offsetof (EP3_RXD_RAIL_ELAN, MainAddr), (long) rxdRail);
+
+       for (j = 0; j < EP_MAXFRAG; j++)
+       {
+           RegisterCookie (&rail->CookieTable, &rxdRail->ChainCookie[j], pRxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[j]), &RxDataCookieOps, (void *) rxdRail);
+
+           event.ev_Type  = EV_TYPE_DMA | (pRxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, Dmas[j+1]));
+           event.ev_Count = 0;
+
+           elan3_sdram_copyl_to_sdram (dev, &event, pRxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[j]), sizeof (E3_BlockCopyEvent));
+       }
+       
+       RegisterCookie (&rail->CookieTable, &rxdRail->DataCookie, pRxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, DataEvent), &RxDataCookieOps, (void *) rxdRail);
+       RegisterCookie (&rail->CookieTable, &rxdRail->DoneCookie, pRxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent), &RxDoneCookieOps, (void *) rxdRail);
+
+       EP3_INIT_COPY_EVENT (event, rxdRail->DataCookie, pRxdMainAddr + offsetof (EP3_RXD_RAIL_MAIN, DataEvent), 1);
+       elan3_sdram_copyl_to_sdram (dev, &event, pRxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent), sizeof (E3_BlockCopyEvent));
+
+       EP3_INIT_COPY_EVENT (event, rxdRail->DoneCookie, pRxdMainAddr + offsetof (EP3_RXD_RAIL_MAIN, DoneEvent), 1);
+       elan3_sdram_copyl_to_sdram (dev, &event, pRxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent), sizeof (E3_BlockCopyEvent));
+       
+       pRxdMain->DataEvent = EP3_EVENT_FREE;
+       pRxdMain->DoneEvent = EP3_EVENT_FREE;
+
+       /* move onto next descriptor */
+       pRxdElan     += EP3_RXD_RAIL_ELAN_SIZE;
+       pRxdElanAddr += EP3_RXD_RAIL_ELAN_SIZE;
+       pRxdMain      = (EP3_RXD_RAIL_MAIN *) ((unsigned long) pRxdMain + EP3_RXD_RAIL_MAIN_SIZE);
+       pRxdMainAddr += EP3_RXD_RAIL_MAIN_SIZE;
+    }
+
+    spin_lock_irqsave (&rcvrRail->FreeDescLock, flags);
+
+    list_add  (&blk->Link, &rcvrRail->DescBlockList);
+    rcvrRail->TotalDescCount += EP3_NUM_RXD_PER_BLOCK;
+    rcvrRail->FreeDescCount  += EP3_NUM_RXD_PER_BLOCK;
+
+    for (i = 0; i < EP3_NUM_RXD_PER_BLOCK; i++)
+       list_add (&blk->Rxd[i].Generic.Link, &rcvrRail->FreeDescList);
+
+    spin_unlock_irqrestore (&rcvrRail->FreeDescLock, flags);
+    
+    return 1;
+}
+
+static void
+FreeRxdRailBlock (EP3_RCVR_RAIL *rcvrRail, EP3_RXD_RAIL_BLOCK *blk)
+{
+    EP3_RAIL     *rail = RCVR_TO_RAIL(rcvrRail);
+    EP3_RXD_RAIL *rxdRail;
+    unsigned long flags;
+    int i, j;
+
+    spin_lock_irqsave (&rcvrRail->FreeDescLock, flags);
+
+    list_del (&blk->Link);
+
+    rcvrRail->TotalDescCount -= EP3_NUM_RXD_PER_BLOCK;
+
+    for (rxdRail = &blk->Rxd[0], i = 0; i < EP3_NUM_RXD_PER_BLOCK; i++, rxdRail++)
+    {
+
+       rcvrRail->FreeDescCount--;
+
+       list_del (&rxdRail->Generic.Link);
+       
+       for (j = 0; j < EP_MAXFRAG; j++)
+           DeregisterCookie (&rail->CookieTable, &rxdRail->ChainCookie[j]);
+       
+       DeregisterCookie (&rail->CookieTable, &rxdRail->DataCookie);
+       DeregisterCookie (&rail->CookieTable, &rxdRail->DoneCookie);
+    }
+
+    spin_unlock_irqrestore (&rcvrRail->FreeDescLock, flags);
+
+    ReleaseDmaRetries (rail, EP3_NUM_RXD_PER_BLOCK);
+
+    ep_free_main (&rail->Generic, blk->Rxd[0].RxdMainAddr, EP3_RXD_RAIL_MAIN_SIZE * EP3_NUM_RXD_PER_BLOCK);
+    ep_free_elan (&rail->Generic, blk->Rxd[0].RxdElanAddr, EP3_RXD_RAIL_ELAN_SIZE * EP3_NUM_RXD_PER_BLOCK);
+
+    KMEM_FREE (blk, sizeof (EP3_RXD_RAIL_BLOCK));
+}
+
+static EP3_RXD_RAIL *
+GetRxdRail (EP3_RCVR_RAIL *rcvrRail)
+{
+    EP3_RXD_RAIL *rxdRail;
+    unsigned long flags;
+    int low_on_rxds;
+
+    spin_lock_irqsave (&rcvrRail->FreeDescLock, flags);
+
+    if (list_empty (&rcvrRail->FreeDescList))
+       rxdRail = NULL;
+    else
+    {
+       rxdRail = list_entry (rcvrRail->FreeDescList.next, EP3_RXD_RAIL, Generic.Link);
+
+       list_del (&rxdRail->Generic.Link);
+
+       rcvrRail->FreeDescCount--;
+    }
+
+    /* Wakeup the descriptor primer thread if there's not many left */
+    low_on_rxds = (rcvrRail->FreeDescCount < ep_rxd_lowat);
+
+    spin_unlock_irqrestore (&rcvrRail->FreeDescLock, flags);
+
+    if (low_on_rxds)
+       ep_kthread_schedule (&RCVR_TO_SUBSYS(rcvrRail)->Thread, lbolt);
+
+    return (rxdRail);
+}
+
+static void
+FreeRxdRail (EP3_RCVR_RAIL *rcvrRail, EP3_RXD_RAIL *rxdRail)
+{
+    unsigned long flags;
+
+#if defined(DEBUG_ASSERT)
+    {
+       EP_RAIL  *rail = (EP_RAIL *) RCVR_TO_RAIL(rcvrRail);
+       ELAN3_DEV *dev = RCVR_TO_DEV (rcvrRail);
+
+       EP_ASSERT (rail, rxdRail->Generic.RcvrRail == &rcvrRail->Generic);
+       
+       EP_ASSERT (rail, rxdRail->RxdMain->DataEvent == EP3_EVENT_PRIVATE);
+       EP_ASSERT (rail, rxdRail->RxdMain->DoneEvent == EP3_EVENT_PRIVATE);
+       EP_ASSERT (rail, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));
+       EP_ASSERT (rail, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)) == 0));
+
+       rxdRail->RxdMain->DataEvent = EP3_EVENT_FREE;
+       rxdRail->RxdMain->DoneEvent = EP3_EVENT_FREE;
+    }
+#endif
+
+    spin_lock_irqsave (&rcvrRail->FreeDescLock, flags);
+    
+    list_add (&rxdRail->Generic.Link, &rcvrRail->FreeDescList);
+
+    rcvrRail->FreeDescCount++;
+
+    if (rcvrRail->FreeDescWaiting)
+    {
+       rcvrRail->FreeDescWaiting--;
+       kcondvar_wakeupall (&rcvrRail->FreeDescSleep, &rcvrRail->FreeDescLock);
+    }
+
+    spin_unlock_irqrestore (&rcvrRail->FreeDescLock, flags);
+}
+
+static void
+BindRxdToRail (EP_RXD *rxd, EP3_RXD_RAIL *rxdRail)
+{
+    EP3_RAIL *rail = RCVR_TO_RAIL (rxdRail->Generic.RcvrRail);
+
+    ASSERT (SPINLOCK_HELD (&rxd->Rcvr->Lock));
+
+    EPRINTF3 (DBG_RCVR, "%s: BindRxdToRail: rxd=%p rxdRail=%p\n",  rail->Generic.Name, rxd, rxdRail);
+
+    elan3_sdram_writel (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, RxdMain), rxd->NmdMain.nmd_addr);                        /* PCI write */
+
+    rxd->RxdRail         = &rxdRail->Generic;
+    rxdRail->Generic.Rxd = rxd;
+}
+
+static void
+UnbindRxdFromRail (EP_RXD *rxd, EP3_RXD_RAIL *rxdRail)
+{
+    EP3_RCVR_RAIL *rcvrRail = (EP3_RCVR_RAIL *) rxdRail->Generic.RcvrRail;
+
+    ASSERT (SPINLOCK_HELD (&rxd->Rcvr->Lock));
+    ASSERT (rxd->RxdRail == &rxdRail->Generic && rxdRail->Generic.Rxd == rxd);
+    
+    EPRINTF3 (DBG_RCVR, "%s: UnbindRxdFromRail: rxd=%p rxdRail=%p\n",  RCVR_TO_RAIL(rxdRail->Generic.RcvrRail)->Generic.Name, rxd, rxdRail);
+
+    rxd->RxdRail         = NULL;
+    rxdRail->Generic.Rxd = NULL;
+
+    if (rcvrRail->CleanupWaiting)
+       kcondvar_wakeupall (&rcvrRail->CleanupSleep, &rxd->Rcvr->Lock);
+    rcvrRail->CleanupWaiting = 0;
+}
+
+static void
+LockRcvrThread (EP3_RCVR_RAIL *rcvrRail)
+{
+    EP_COMMS_RAIL     *commsRail   = rcvrRail->Generic.CommsRail;
+    EP3_RAIL          *rail        = RCVR_TO_RAIL(rcvrRail);
+    ELAN3_DEV        *dev         = rail->Device;
+    sdramaddr_t        sle         = rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, ThreadLock);
+    EP3_SPINLOCK_MAIN *sl          = &rcvrRail->RcvrMain->ThreadLock;
+    E3_uint32          RestartBits = 0;
+    int                delay       = 1;
+    E3_uint32          seq;
+    E3_uint32          reg;
+
+    ASSERT (SPINLOCK_HELD (&rcvrRail->Generic.Rcvr->Lock));
+
+    mb();
+    elan3_sdram_writel (dev, sle + offsetof (EP3_SPINLOCK_ELAN, sl_lock), 1);
+    mb();
+    seq = elan3_sdram_readl (dev, sle + offsetof (EP3_SPINLOCK_ELAN, sl_seq));
+    while (seq != sl->sl_seq)
+    {
+       while (sl->sl_seq == (seq - 1))
+       {
+           mb();
+
+           if ((read_reg32 (dev, Exts.InterruptReg) & (INT_TProc | INT_TProcHalted)) != 0 && spin_trylock (&dev->IntrLock))
+           {
+               reg=read_reg32 (dev, Exts.InterruptReg);
+               ELAN_REG_REC(reg);
+
+               if ((reg & (INT_TProc | INT_TProcHalted)) != 0&& 
+                   elan3_sdram_readl (dev, sle + offsetof (EP3_SPINLOCK_ELAN, sl_seq)) != sl->sl_seq)
+               {
+                   EPRINTF1 (DBG_RCVR, "%s: LockRcvrThread - thread trapped\n", rail->Generic.Name);
+
+                   /* The thread processor has *really* trapped, and the spinlock is still held.
+                    * thus is must have trapped due to a network error - we need to complete the
+                    * actions required for this envelope, since we may be spin-locking the receiver
+                    * to search the dma retry lists for a particular dma.  So must ensure that
+                    * if the thread had trapped then the dma has been queued onto the retry list
+                    * *before* we inspect them.
+                    */
+                   IncrStat (commsRail, LockRcvrTrapped);
+
+                   /* We're going to generate a spurious interrupt here - since we will
+                    * handle the thread processor trap directly */
+                   ELAN_REG_REC(reg);
+                   if (HandleTProcTrap (dev, &RestartBits))
+                   {
+                       /* NOTE - this is not an assert, since the "store" to unlock the lock could
+                        *        be held up on the PCI interface, whilst the thread processor has
+                        *        gone on and switched to a new thread, which has then trapped, and
+                        *        our read of the InterruptReg can overtake the unlock write.
+                        *
+                        * ASSERT (dev->ThreadTrap->Registers[REG_GLOBALS + (1^WordEndianFlip)] == 
+                        *         elan3_sdram_readl (dev, rcvr->RcvrElan + offsetof (EP_RCVR_ELAN, PendingRxDescsElan)));
+                        */
+
+                       PULSE_SCHED_STATUS (dev, RestartBits);
+
+                       DeliverTProcTrap (dev, dev->ThreadTrap, INT_TProc);
+                   }
+               }
+               spin_unlock (&dev->IntrLock);
+           }
+           
+           DELAY (delay); delay++;
+       }
+       seq = elan3_sdram_readl (dev, sle + offsetof (EP3_SPINLOCK_ELAN, sl_seq));
+    }
+}
+
+static void
+UnlockRcvrThread (EP3_RCVR_RAIL *rcvrRail)
+{
+    EP3_RAIL   *rail = RCVR_TO_RAIL(rcvrRail);
+    sdramaddr_t sle  = rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, ThreadLock);
+
+    mb();
+    elan3_sdram_writel (rail->Device, sle + offsetof (EP3_SPINLOCK_ELAN, sl_lock), 0);
+    mmiob(); 
+}
+
+void
+CompleteEnvelope (EP3_RAIL *rail, E3_Addr rxdElanAddr, E3_uint32 PAckVal)
+{
+    ELAN3_DEV         *dev       = rail->Device;
+    sdramaddr_t        rxdElan   = ep_elan2sdram (&rail->Generic, rxdElanAddr);
+    EP3_RXD_RAIL      *rxdRail   = (EP3_RXD_RAIL *) (unsigned long) elan3_sdram_readq (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, MainAddr));
+    EP_RXD_MAIN       *rxdMain   = rxdRail->Generic.Rxd->RxdMain;
+    EP_ENVELOPE       *env       = &rxdMain->Envelope;
+    EP3_RCVR_RAIL     *rcvrRail  = (EP3_RCVR_RAIL *) rxdRail->Generic.RcvrRail;
+    EP_COMMS_RAIL     *commsRail = rcvrRail->Generic.CommsRail;
+    EP_RCVR           *rcvr      = rcvrRail->Generic.Rcvr;
+    sdramaddr_t        queue     = ((EP3_COMMS_RAIL *) commsRail)->QueueDescs + rcvr->Service * sizeof (EP3_InputQueue);
+    sdramaddr_t        sle       = rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, ThreadLock);
+    EP3_SPINLOCK_MAIN *sl        = &rcvrRail->RcvrMain->ThreadLock;
+    int               nodeId;
+    EP_NODE_RAIL     *nodeRail;
+    E3_DMA_BE         dma;
+    E3_Addr           nfptr;
+    E3_Addr          next;
+
+    ASSERT (commsRail->Rail == &rail->Generic);
+    ASSERT (rxdElanAddr == elan3_sdram_readl (dev, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingDescs)));
+
+    IncrStat (commsRail, CompleteEnvelope);
+
+    /* We don't need to aquire the NodeLock here (however we might be holding it),
+     * since this can only get called while the node is connected, or disconnecting.
+     * If the node is disconnecting, then we can get called from FlushDisconnecting()
+     * while holding the NodeLock - after we cannot get called again until the node 
+     * has reconnected from scratch.
+     */
+    /* Copy the envelope information */
+    nfptr = elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_fptr));
+
+    if (nfptr == elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_top)))
+       nfptr = elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_base));
+    else
+       nfptr += elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_size));
+
+    /* Copy the envelope and payload (unconditionally) */
+    elan3_sdram_copyl_from_sdram (dev, rcvrRail->InputQueueBase + (nfptr - rcvrRail->InputQueueAddr), env, EP_ENVELOPE_SIZE + EP_PAYLOAD_SIZE);
+
+    ASSERT (env->Version == EP_ENVELOPE_VERSION);
+
+    /* Copy the received message length */
+    rxdMain->Len = elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Data.nmd_len));
+       
+    /* Remove the RXD from the pending desc list */
+    if ((next = elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Next))) == 0)
+       rcvrRail->RcvrMain->PendingDescsTailp = 0;
+    elan3_sdram_writel (dev, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingDescs), next);
+
+    /* Copy the DMA descriptor to queue on the approriate retry list */
+    elan3_sdram_copyq_from_sdram (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Dmas[0]), &dma, sizeof (E3_DMA));        /* PCI read block */
+    
+    EP_ASSERT (&rail->Generic, dma.s.dma_direction == DMA_READ);;
+
+#if defined(DEBUG_ASSERT) && defined(DEBUG_SDRAM_ASSERT)
+    /* NOTE: not an assertion, since the thread packet could have successfully
+     *       transferred the "put" dma to the far side - which could then have
+     *       completed - but the far side will see a network error which will
+     *       cause the virtual circuit to be dropped by the far side and this 
+     *       DMA will be removed */
+    if (rxdRail->RxdMain->DataEvent != EP3_EVENT_ACTIVE ||
+       elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) != 1)
+    {
+       printk ("CompleteEnvelope: suspicious dma : Node=%d DataBlock=%d Event=%d\n", 
+               env->NodeId, rxdRail->RxdMain->DataEvent, 
+               elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)));
+    }
+#endif
+    
+    EPRINTF6 (DBG_RCVR, "%s: CompleteEnvelope: rxd=%p NodeId=%d Xid=%llx Cookies=%08x,%08x\n", commsRail->Rail->Name,
+             rxdRail, env->NodeId, (long long) env->Xid.Unique, dma.s.dma_srcCookieVProc, dma.s.dma_destCookieVProc);
+
+    /* we MUST convert this into a DMA_READ_REQUEUE dma as if we don't the DMA descriptor will
+     * be read from the EP_RETRY_DMA rather than the original DMA - this can then get reused 
+     * and an incorrect DMA descriptor sent */
+    dma.s.dma_source    = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, Dmas[0]);
+    dma.s.dma_direction = (dma.s.dma_direction & ~DMA_READ) | DMA_READ_REQUEUE;
+    
+    nodeId   = EP_VP_TO_NODE(dma.s.dma_srcVProc);
+    nodeRail = &rail->Generic.Nodes[nodeId];
+
+    ASSERT (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE);
+
+    if (PAckVal != E3_PAckOk)
+    {
+       if (nodeRail->State == EP_NODE_CONNECTED)
+           QueueDmaForRetry (rail, &dma, EP_RETRY_LOW_PRI_RETRY);
+       else
+           QueueDmaOnStalledList (rail, &dma);
+    }
+
+    /* Finaly forcefully drop the spinlock for the thread */
+    sl->sl_seq = elan3_sdram_readl (dev, sle + offsetof (EP3_SPINLOCK_ELAN, sl_seq));
+
+    wmb();
+}
+
+void
+StallThreadForNoDescs (EP3_RAIL *rail, E3_Addr rcvrElanAddr, E3_Addr sp)
+{
+    ELAN3_DEV      *dev       = rail->Device;
+    sdramaddr_t    rcvrElan   = ep_elan2sdram (&rail->Generic, rcvrElanAddr);
+    EP3_RCVR_RAIL  *rcvrRail  = (EP3_RCVR_RAIL *) (unsigned long) elan3_sdram_readq (dev, rcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, MainAddr));
+    EP_RCVR        *rcvr      = rcvrRail->Generic.Rcvr;
+    EP_COMMS_RAIL  *commsRail = rcvrRail->Generic.CommsRail;
+
+    EPRINTF3 (DBG_RCVR, "%s: StallThreadForNoDescs - rcvrRail=%p sp=%x\n", commsRail->Rail->Name, rcvrRail, sp);
+    
+    IncrStat (commsRail, StallThread);
+
+    /* NOTE: spin lock not required as thread is trapped */
+    
+    if (rcvrRail->RcvrMain->PendingDescsTailp != 0)
+    {
+       EPRINTF1 (DBG_RCVR, "%s: StallThreadForNoDescs - pending descriptors, wakeup thread\n", commsRail->Rail->Name);
+       
+       /*
+        * A receive buffer was queued after the thread had decided to go to 
+        * sleep, but before the event interrupt occured.  Just restart the
+        * thread to consume the envelope.
+        */
+       IssueRunThread (rail, sp);
+    }
+    else
+    {
+       EPRINTF1 (DBG_RCVR, "%s: StallThreadForNoDescs - set ThreadWaiting\n", commsRail->Rail->Name);
+       
+       IncrStat (commsRail, ThrdWaiting);
+
+       /* Mark the rcvr as waiting for a rxd, and schedule a call of ep_check_rcvr
+        * to attempt to "steal" a descriptor from a different rail */
+       rcvrRail->ThreadWaiting = sp;
+
+       ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+    }
+}
+
+void
+StallThreadForHalted (EP3_RAIL *rail, E3_Addr rcvrElanAddr, E3_Addr sp)
+{
+    ELAN3_DEV     *dev       = rail->Device;
+    sdramaddr_t    rcvrElan  = ep_elan2sdram (&rail->Generic, rcvrElanAddr);
+    EP3_RCVR_RAIL *rcvrRail  = (EP3_RCVR_RAIL *) (unsigned long) elan3_sdram_readq (dev, rcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, MainAddr));
+    EP_RCVR       *rcvr      = rcvrRail->Generic.Rcvr;
+    unsigned long  flags     = 0;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+
+    rcvrRail->ThreadHalted = sp;
+
+    EPRINTF2 (DBG_EPTRAP, "%s: StallThreadForHalted: sp=%08x\n", rail->Generic.Name, sp);
+
+    if (rcvrRail->CleanupWaiting)
+       kcondvar_wakeupone (&rcvrRail->CleanupSleep, &rcvr->Lock);
+    rcvrRail->CleanupWaiting = 0;
+
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+/*
+ * RxDataEvent: arg == EP3_RXD_RAIL
+ *   Called on completion of receiving data.
+ */
+static void
+RxDataEvent (EP3_RAIL *rail, void *arg)
+{
+    EP3_RXD_RAIL  *rxdRail  = (EP3_RXD_RAIL *) arg;
+    EP3_RCVR_RAIL *rcvrRail = (EP3_RCVR_RAIL *) rxdRail->Generic.RcvrRail;
+    EP_RXD        *rxd      = rxdRail->Generic.Rxd;
+    EP_ENVELOPE   *env      = &rxd->RxdMain->Envelope;
+    EP_RCVR       *rcvr     = rxd->Rcvr;
+    ELAN3_DEV    *dev      = rail->Device;
+    unsigned long flags;
+    int delay = 1;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    for (;;)
+    {
+       if (EP3_EVENT_FIRED (rxdRail->DataCookie, rxdRail->RxdMain->DataEvent))
+           break;
+
+       if (EP3_EVENT_FIRING (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent), rxdRail->DataCookie, rxdRail->RxdMain->DataEvent))
+       {
+           if (delay > EP3_EVENT_FIRING_TLIMIT)
+               panic ("RxDataEvent: events set but block copy not completed\n");
+           DELAY(delay);
+           delay <<= 1;
+       }
+       else
+       {
+           printk ("%s: RxDataEvent: rxd %p not complete [%x,%x,%x]\n", rail->Generic.Name, rxd, rxdRail->RxdMain->DataEvent,
+                   elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)),
+                   elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Type)));
+           
+           spin_unlock_irqrestore (&rcvr->Lock, flags);
+           return;
+       }
+       mb();
+    }
+
+    /*
+     * Note, since the thread will have sent the "get" dma before copying the
+     * envelope, we must check that it has completed doing this,  if not then
+     * it might be that the thread trapped due to a network error, so we must
+     * spinlock against the thread 
+     */
+    if (rxd->RxdMain->Len == EP_RXD_PENDING)
+    {
+       LockRcvrThread (rcvrRail);
+       UnlockRcvrThread (rcvrRail);
+
+       ASSERT (env->Version == EP_ENVELOPE_VERSION && rxd->RxdMain->Len != EP_RXD_PENDING);
+    }
+
+    EPRINTF7 (DBG_RCVR, "%s: RxDataEvent: rxd=%p rxdRail=%p completed from elan node %d [XID=%llx] Length %d State %x\n", 
+             rail->Generic.Name, rxd, rxdRail, env->NodeId, (long long) env->Xid.Unique, rxd->RxdMain->Len, rxd->State);
+
+    EP_ASSERT (&rail->Generic, rxd->State == EP_RXD_RECEIVE_ACTIVE || rxd->State == EP_RXD_PUT_ACTIVE || rxd->State == EP_RXD_GET_ACTIVE);
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));      /* PCI read */
+    EP_ASSERT (&rail->Generic, rxdRail->RxdMain->DoneEvent == EP3_EVENT_PRIVATE);
+
+    rxdRail->RxdMain->DataEvent = EP3_EVENT_PRIVATE;
+    rxd->Data.nmd_attr          = EP_RAIL2RAILMASK (rail->Generic.Number);
+
+    if (rxd->RxdMain->Len >= 0 && EP_IS_RPC(env->Attr))
+       rxd->State = EP_RXD_RPC_IN_PROGRESS;
+    else
+    {
+       rxd->State = EP_RXD_COMPLETED;
+
+       /* remove from active list */
+       list_del (&rxd->Link);
+
+       UnbindRxdFromRail (rxd, rxdRail);
+       FreeRxdRail (rcvrRail, rxdRail);
+    }
+
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+    ep_rxd_received (rxd);
+
+}
+
+/* 
+ * RxDataRetry: arg == EP3_RXD_RAIL
+ *   Called on retry of "get" dma of large transmit data
+ *   and rpc_get/rpc_put and "put" of datavec of rpc completion.
+ */
+static void
+RxDataRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status)
+{
+    EP3_RXD_RAIL  *rxdRail  = (EP3_RXD_RAIL *) arg;
+    EP_COMMS_RAIL *commsRail = rxdRail->Generic.RcvrRail->CommsRail;
+    EP_RXD        *rxd       = rxdRail->Generic.Rxd;
+
+#if defined(DEBUG_ASSERT)
+    RxDataVerify (rail, arg, dma);
+#endif
+
+    IncrStat (commsRail, RxDataRetry);
+
+    EPRINTF4 (DBG_RCVR, "%s: RxDataRetry: rcvr %p rxd %p [XID=%llx]\n", rail->Generic.Name, rxd->Rcvr, rxd, (long long) rxd->RxdMain->Envelope.Xid.Unique);
+
+    QueueDmaForRetry (rail, dma, EP_RETRY_LOW_PRI_RETRY + ep_backoff (&rxdRail->Backoff, EP_BACKOFF_DATA));
+}
+
+static void
+RxDataVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma)
+{
+#if defined(DEBUG_ASSERT)
+    EP3_RXD_RAIL   *rxdRail  = (EP3_RXD_RAIL *) arg;
+    EP_RXD        *rxd       = rxdRail->Generic.Rxd;
+
+    if (dma->s.dma_direction == DMA_WRITE)
+    {
+       EP_ASSERT (&rail->Generic, 
+                  (rxd->State == EP_RXD_RECEIVE_ACTIVE  && rxdRail->RxdMain->DataEvent == EP3_EVENT_ACTIVE && rxdRail->RxdMain->DoneEvent == EP3_EVENT_PRIVATE) ||
+                  (rxd->State == EP_RXD_PUT_ACTIVE      && rxdRail->RxdMain->DataEvent == EP3_EVENT_ACTIVE && rxdRail->RxdMain->DoneEvent == EP3_EVENT_PRIVATE) ||
+                  (rxd->State == EP_RXD_COMPLETE_ACTIVE && rxdRail->RxdMain->DataEvent == EP3_EVENT_PRIVATE && rxdRail->RxdMain->DoneEvent == EP3_EVENT_ACTIVE));
+       EP_ASSERT (&rail->Generic, SDRAM_ASSERT (rxd->State == EP_RXD_COMPLETE_ACTIVE ?
+                                                elan3_sdram_readl (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)) == 1:            /* PCI read */
+                                                elan3_sdram_readl (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) == 1));          /* PCI read */
+    }
+    else
+    {
+       EP_ASSERT (&rail->Generic, dma->s.dma_direction == DMA_READ_REQUEUE);
+       
+#if defined(DEBUG_SDRAM_ASSERT)
+       /* NOTE: not an assertion, since the "get" DMA can still be running if
+        *       it's packet got a network error - and then the "put" from the
+        *       far side has completed - however the virtual circuit should
+        *       then be dropped by the far side and this DMA will be removed */
+       if (EP_VP_TO_NODE(dma->s.dma_srcVProc) != ep_rxd_node(rxd) || 
+           (rxd->State != EP_RXD_RECEIVE_ACTIVE && rxd->State != EP_RXD_GET_ACTIVE) ||
+           rxdRail->RxdMain->DataEvent != EP3_EVENT_ACTIVE ||
+           elan3_sdram_readl (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) != 1)
+       {
+           EPRINTF6 (DBG_RCVR, "%s: RxDataRetry: suspicious dma : VProc=%d NodeId=%d State=%d DataBlock=%x Event=%d\n",  
+                     rail->Generic.Name, EP_VP_TO_NODE(dma->s.dma_srcVProc), ep_rxd_node(rxd), rxd->State, rxdRail->RxdMain->DataEvent, 
+                     elan3_sdram_readl (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)));
+       }
+#endif /* defined(DEBUG_SDRAM_ASSERT) */
+    }
+#endif /* DEBUG_ASSERT */
+}
+
+/*
+ * RxDoneEvent: arg == EP_RXD
+ *   Called on completion of large receive.
+ */
+static void
+RxDoneEvent (EP3_RAIL *rail, void *arg)
+{
+    EP3_RXD_RAIL  *rxdRail   = (EP3_RXD_RAIL *) arg;
+    EP3_RCVR_RAIL *rcvrRail  = (EP3_RCVR_RAIL *) rxdRail->Generic.RcvrRail;
+    EP_COMMS_RAIL *commsRail = rcvrRail->Generic.CommsRail;
+    EP_RXD        *rxd       = rxdRail->Generic.Rxd;
+    EP_RCVR       *rcvr      = rxd->Rcvr;
+    ELAN3_DEV    *dev       = rail->Device;
+    int            delay     = 1;
+    unsigned long  flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    for (;;)
+    {
+       if (EP3_EVENT_FIRED (rxdRail->DoneCookie, rxdRail->RxdMain->DoneEvent))
+           break;
+       
+       if (EP3_EVENT_FIRING (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent), rxdRail->DoneCookie, rxdRail->RxdMain->DoneEvent))
+       {
+           if (delay > EP3_EVENT_FIRING_TLIMIT)
+               panic ("RxDoneEvent: events set but block copy not completed\n");
+           DELAY(delay);
+           delay <<= 1;
+       }
+       else
+       {
+           printk ("RxDoneEvent: rxd %p not complete [%x,%x.%x]\n", rxd, rxdRail->RxdMain->DoneEvent,
+                   elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)),
+                   elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Type)));
+           
+           spin_unlock_irqrestore (&rcvr->Lock, flags);
+           return;
+       }
+       mb();
+    }
+
+    EPRINTF4 (DBG_RCVR, "%s: RxDoneEvent: rxd %p completed from elan node %d [XID=%llx]\n", 
+             commsRail->Rail->Name, rxd, rxd->RxdMain->Envelope.NodeId, (long long) rxd->RxdMain->Envelope.Xid.Unique);
+    
+    IncrStat (commsRail, RxDoneEvent);
+
+    EP_ASSERT (&rail->Generic, rxdRail->RxdMain->DataEvent  == EP3_EVENT_PRIVATE);
+    EP_ASSERT (&rail->Generic, EP3_EVENT_FIRED (rxdRail->DoneCookie, rxdRail->RxdMain->DoneEvent));
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));      /* PCI read */
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)) == 0));      /* PCI read */
+
+    /* mark rxd as private  */
+    rxdRail->RxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+
+    /* remove from active list */
+    list_del (&rxd->Link);
+
+    UnbindRxdFromRail (rxd, rxdRail);
+    FreeRxdRail (rcvrRail, rxdRail);
+       
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    rxd->Handler (rxd);
+}
+
+/* 
+ * RxDoneRetry: arg == EP_RXD
+ *   Called on retry of "put" of RPC completion status block
+ */
+static void
+RxDoneRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status)
+{
+    EP3_RXD_RAIL  *rxdRail   = (EP3_RXD_RAIL *) arg;
+    EP_COMMS_RAIL *commsRail = rxdRail->Generic.RcvrRail->CommsRail;
+    EP_RXD        *rxd       = rxdRail->Generic.Rxd;
+
+#if defined(DEBUG_ASSERT)
+    RxDoneVerify (rail, arg, dma);
+#endif
+
+    IncrStat (commsRail, RxDoneRetry);
+
+    EPRINTF4 (DBG_RCVR, "%s: RxDoneRetry: rcvr %p rxd %p [XID=%llx]\n", commsRail->Rail->Name, rxd->Rcvr, rxd, (long long) rxd->RxdMain->Envelope.Xid.Unique);
+
+    QueueDmaForRetry (rail, dma, EP_RETRY_LOW_PRI_RETRY + ep_backoff (&rxdRail->Backoff, EP_BACKOFF_DONE));
+}
+
+static void
+RxDoneVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma)
+{
+#if defined(DEBUG_ASSERT)
+    EP3_RXD_RAIL *rxdRail = (EP3_RXD_RAIL *) arg;
+    EP_RXD       *rxd     = rxdRail->Generic.Rxd;
+
+    EP_ASSERT (&rail->Generic, dma->s.dma_direction == DMA_WRITE && EP_VP_TO_NODE(dma->s.dma_destVProc) == ep_rxd_node(rxd));
+    EP_ASSERT (&rail->Generic, rxd->State == EP_RXD_COMPLETE_ACTIVE && rxdRail->RxdMain->DoneEvent  == EP3_EVENT_ACTIVE);
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)) == 1));     /* PCI read */
+#endif /* defined(DEBUG_ASSERT) */
+}
+
+int
+ep3rcvr_queue_rxd (EP_RXD *rxd, EP_RCVR_RAIL *r)
+{
+    EP3_RCVR_RAIL *rcvrRail = (EP3_RCVR_RAIL *) r;
+    EP3_RAIL      *rail     = RCVR_TO_RAIL(rcvrRail);
+    ELAN3_DEV     *dev      = rail->Device;
+    EP3_RXD_RAIL  *rxdRail;
+
+    ASSERT ( SPINLOCK_HELD(&rxd->Rcvr->Lock));
+
+    if ((rxdRail = GetRxdRail (rcvrRail)) == NULL)
+       return 0;
+
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->Generic);
+
+    elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, Data.nmd_addr), rxd->Data.nmd_addr);              /* PCI write */
+    elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, Data.nmd_len),  rxd->Data.nmd_len);               /* PCI write */
+    elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, Data.nmd_attr), rxd->Data.nmd_attr);              /* PCI write */
+
+    /* Bind the rxdRail and rxd together */
+    BindRxdToRail (rxd, rxdRail);
+    
+    /* Mark as active */
+    elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 1);
+
+    rxdRail->RxdMain->DataEvent  = EP3_EVENT_ACTIVE;
+    rxdRail->RxdMain->DoneEvent  = EP3_EVENT_PRIVATE;
+
+    /* Interlock with StallThreadForNoDescs */
+    spin_lock (&dev->IntrLock);
+
+    EPRINTF4 (DBG_RCVR, "%s: ep3rcvr_queue_rxd: rcvr %p rxd %p rxdRail %p\n", rail->Generic.Name, rxd->Rcvr, rxd, rxdRail);
+
+    EP3_SPINENTER (dev, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingLock), &rcvrRail->RcvrMain->PendingLock);
+
+    elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, Next), 0);                                        /* PCI write */
+    if (rcvrRail->RcvrMain->PendingDescsTailp == 0)
+       elan3_sdram_writel (dev, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingDescs), rxdRail->RxdElanAddr);       /* PCI write */
+    else
+       elan3_sdram_writel (dev, rcvrRail->RcvrMain->PendingDescsTailp, rxdRail->RxdElanAddr);                          /* PCI write */
+    rcvrRail->RcvrMain->PendingDescsTailp = rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, Next);
+    
+    EP3_SPINEXIT (dev, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingLock), &rcvrRail->RcvrMain->PendingLock);
+
+    /* If the thread has paused because it was woken up with no receive buffer */
+    /* ready, then wake it up to process the one we've just added */
+    if (rcvrRail->ThreadWaiting)
+    {
+       EPRINTF1 (DBG_RCVR, "%s: DoReceive: ThreadWaiting - restart thread\n", rail->Generic.Name);
+
+       IssueRunThread (rail, rcvrRail->ThreadWaiting);
+
+       rcvrRail->ThreadWaiting = (E3_Addr) 0;
+    }
+
+    spin_unlock (&dev->IntrLock);
+
+    return 1;
+}
+
+void
+ep3rcvr_rpc_put (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags)
+{
+    EP3_RXD_RAIL      *rxdRail   = (EP3_RXD_RAIL *) rxd->RxdRail;
+    EP3_RCVR_RAIL     *rcvrRail  = (EP3_RCVR_RAIL *) rxdRail->Generic.RcvrRail;
+    EP3_RAIL          *rail      = RCVR_TO_RAIL (rcvrRail);
+    ELAN3_DEV         *dev       = rail->Device;
+
+    EP3_RXD_RAIL_MAIN *rxdMain   = rxdRail->RxdMain;
+    sdramaddr_t        rxdElan   = rxdRail->RxdElan;
+    EP_ENVELOPE       *env       = &rxd->RxdMain->Envelope;
+    E3_DMA_BE         dmabe;
+    int                       i, len;
+
+    EP_ASSERT (&rail->Generic, rxd->State == EP_RXD_PUT_ACTIVE);
+    EP_ASSERT (&rail->Generic, rxdMain->DataEvent == EP3_EVENT_PRIVATE && rxdMain->DoneEvent == EP3_EVENT_PRIVATE);
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));       /* PCI read */
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)) == 0));       /* PCI read */
+
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->Generic);
+    
+    /* Generate the DMA chain to put the data in two loops to burst
+     * the data across the PCI bus */
+    for (len = 0, i = (nFrags-1), local += (nFrags-1), remote += (nFrags-1); i >= 0;   len += local->nmd_len, i--, local--, remote--)
+    {
+       dmabe.s.dma_type            = E3_DMA_TYPE(DMA_BYTE, DMA_WRITE, DMA_NORMAL, EP3_DMAFAILCOUNT);
+       dmabe.s.dma_size            = local->nmd_len;
+       dmabe.s.dma_source          = local->nmd_addr;
+       dmabe.s.dma_dest            = remote->nmd_addr;
+       dmabe.s.dma_destEvent       = (E3_Addr) 0;
+       dmabe.s.dma_destCookieVProc = EP_VP_DATA (env->NodeId);
+       if (i == (nFrags-1))
+           dmabe.s.dma_srcEvent    = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, DataEvent);
+       else
+           dmabe.s.dma_srcEvent    = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[i]);
+       dmabe.s.dma_srcCookieVProc  = LocalCookie (rail, env->NodeId);
+       
+       EPRINTF9 (DBG_RCVR, "%s: ep3rcvr_rpc_put: rxd %p [XID=%llx] idx=%d Source=%08x Dest=%08x Len=%x Cookies=%x.%x\n", rail->Generic.Name, rxd,
+                 (long long) env->Xid.Unique, i, local->nmd_addr, remote->nmd_addr, local->nmd_len, dmabe.s.dma_destCookieVProc, dmabe.s.dma_srcCookieVProc);
+       
+       if (i != 0)
+           elan3_sdram_copyq_to_sdram (dev, &dmabe, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Dmas[i]), sizeof (E3_DMA)); /* PCI write block */
+    }
+    
+    for (i = 0; i < nFrags; i++)
+       elan3_sdram_writel (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[i].ev_Count), 1);                            /* PCI write */
+    
+    /* Initialise the data event */
+    elan3_sdram_writel (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 1);                   /* PCI write */
+    rxdMain->DataEvent = EP3_EVENT_ACTIVE;
+   
+    ASSERT (rail->Generic.Nodes[env->NodeId].State >= EP_NODE_CONNECTED && rail->Generic.Nodes[env->NodeId].State <= EP_NODE_LOCAL_PASSIVATE);
+
+    if (IssueDma (rail, &dmabe, EP_RETRY_LOW_PRI, FALSE) != ISSUE_COMMAND_OK)
+    {
+       /* Failed to issue the dma command, so copy the dma descriptor and queue it for retry */
+       EPRINTF2 (DBG_RCVR, "%s: ep3rcvr_rpc_put: queue rxd %p on retry thread\n", rail->Generic.Name, rxd);
+       
+       QueueDmaForRetry (rail, &dmabe, EP_RETRY_LOW_PRI);
+    }
+    
+    BucketStat (rxd->Rcvr->Subsys, RPCPut, len);
+}
+
+void
+ep3rcvr_rpc_get (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags)
+{
+    EP3_RXD_RAIL      *rxdRail   = (EP3_RXD_RAIL *) rxd->RxdRail;
+    EP3_RCVR_RAIL     *rcvrRail  = (EP3_RCVR_RAIL *) rxdRail->Generic.RcvrRail;
+    EP3_RAIL          *rail      = RCVR_TO_RAIL (rcvrRail);
+    ELAN3_DEV         *dev       = rail->Device;
+
+    EP3_RXD_RAIL_MAIN *rxdMain   = rxdRail->RxdMain;
+    sdramaddr_t        rxdElan   = rxdRail->RxdElan;
+    EP_ENVELOPE       *env       = &rxd->RxdMain->Envelope;
+    E3_DMA_BE         dmabe;
+    int                       i, len;
+
+    EP_ASSERT (&rail->Generic, rxd->State == EP_RXD_GET_ACTIVE);
+    EP_ASSERT (&rail->Generic, rxdMain->DataEvent == EP3_EVENT_PRIVATE && rxdMain->DoneEvent == EP3_EVENT_PRIVATE);
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));       /* PCI read */
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)) == 0));       /* PCI read */
+       
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->Generic);
+    
+    /* Generate the DMA chain to get the data in two loops to burst
+     * the data across the PCI bus */
+    for (len = 0, i = (nFrags-1), remote += (nFrags-1), local += (nFrags-1); i >= 0;   len += remote->nmd_len, i--, remote--, local--)
+    {
+       dmabe.s.dma_type            = E3_DMA_TYPE(DMA_BYTE, DMA_READ, DMA_NORMAL, EP3_DMAFAILCOUNT);
+       dmabe.s.dma_size            = remote->nmd_len;
+       dmabe.s.dma_source          = remote->nmd_addr;
+       dmabe.s.dma_dest            = local->nmd_addr;
+       if (i == (nFrags-1))
+           dmabe.s.dma_destEvent   = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, DataEvent);
+       else
+           dmabe.s.dma_destEvent   = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[i]);
+       dmabe.s.dma_destCookieVProc = LocalCookie (rail, env->NodeId);
+       dmabe.s.dma_srcEvent        = (E3_Addr) 0;
+       dmabe.s.dma_srcCookieVProc  = RemoteCookie (rail, env->NodeId);
+       
+       EPRINTF9 (DBG_RCVR, "%s: ep3rcvr_rpc_get rxd %p [XID=%llx] idx=%d Source=%08x Dest=%08x Len=%x Cookies=%x.%x\n", rail->Generic.Name, rxd,
+                 (long long) env->Xid.Unique, i, remote->nmd_addr, local->nmd_addr, remote->nmd_len, dmabe.s.dma_destCookieVProc, 
+                 dmabe.s.dma_srcCookieVProc);
+       
+       /* 
+        * Always copy down the dma descriptor, since we issue it as a READ_REQUEUE
+        * dma, and the elan will fetch the descriptor to send out of the link from
+        * the rxdElan->Dmas[i] location,  before issueing the DMA chain we modify
+        * the dma_source.
+        */
+       elan3_sdram_copyq_to_sdram (dev, &dmabe, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Dmas[i]), sizeof (E3_DMA)); /* PCI write block */
+    }
+    
+    for (i = 0; i < nFrags; i++)
+       elan3_sdram_writel (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[i].ev_Count), 1);                            /* PCI write */
+    
+    /* Initialise the data event */
+    elan3_sdram_writel (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 1);                   /* PCI write */
+    rxdMain->DataEvent  = EP3_EVENT_ACTIVE;
+    
+    ASSERT (rail->Generic.Nodes[env->NodeId].State >= EP_NODE_CONNECTED && rail->Generic.Nodes[env->NodeId].State <= EP_NODE_LOCAL_PASSIVATE);
+
+    /* we MUST convert this into a DMA_READ_REQUEUE dma as if we don't the DMA descriptor will
+     * be read from the EP_RETRY_DMA rather than the orignal DMA - this can then get reused 
+     * and an incorrect DMA descriptor sent */
+    dmabe.s.dma_source    = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, Dmas[0]);
+    dmabe.s.dma_direction = (dmabe.s.dma_direction & ~DMA_READ) | DMA_READ_REQUEUE;
+    
+    if (IssueDma (rail, &dmabe, EP_RETRY_LOW_PRI, FALSE) != ISSUE_COMMAND_OK)
+    {
+       /* Failed to issue the dma command, so copy the dma descriptor and queue it for retry */
+       EPRINTF2 (DBG_RCVR, "%s: ep3rcvr_rpc_get: queue rxd %p on retry thread\n", rail->Generic.Name, rxd);
+       
+       QueueDmaForRetry (rail, &dmabe, EP_RETRY_LOW_PRI);
+    }
+
+    BucketStat (rxd->Rcvr->Subsys, RPCGet, len);
+}
+       
+void
+ep3rcvr_rpc_complete (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags)
+{
+    EP3_RXD_RAIL      *rxdRail   = (EP3_RXD_RAIL *) rxd->RxdRail;
+    EP3_RCVR_RAIL     *rcvrRail  = (EP3_RCVR_RAIL *) rxdRail->Generic.RcvrRail;
+    EP3_RAIL          *rail      = RCVR_TO_RAIL (rcvrRail);
+    ELAN3_DEV         *dev       = rail->Device;
+
+    EP3_RXD_RAIL_MAIN *rxdMain   = rxdRail->RxdMain;
+    sdramaddr_t        rxdElan   = rxdRail->RxdElan;
+    EP_ENVELOPE       *env       = &rxd->RxdMain->Envelope;
+    E3_DMA_BE         dmabe;
+    int                       i, len;
+    
+    EP_ASSERT (&rail->Generic, rxd->State == EP_RXD_COMPLETE_ACTIVE);
+    EP_ASSERT (&rail->Generic, rxdMain->DataEvent == EP3_EVENT_PRIVATE && rxdMain->DoneEvent == EP3_EVENT_PRIVATE);
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));       /* PCI read */
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)) == 0));       /* PCI read */
+
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->Generic);
+    
+    /* Initialise the status block dma */
+    dmabe.s.dma_type            = E3_DMA_TYPE(DMA_BYTE, DMA_WRITE, DMA_NORMAL, EP3_DMAFAILCOUNT);
+    dmabe.s.dma_size            = sizeof (EP_STATUSBLK);
+    dmabe.s.dma_source          = rxd->NmdMain.nmd_addr + offsetof (EP_RXD_MAIN, StatusBlk);
+    dmabe.s.dma_dest            = env->TxdMain.nmd_addr + offsetof (EP_TXD_MAIN, StatusBlk);
+    dmabe.s.dma_destEvent       = env->TxdRail + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent);
+    dmabe.s.dma_destCookieVProc = EP_VP_DATA(env->NodeId);
+    dmabe.s.dma_srcEvent        = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent);
+    dmabe.s.dma_srcCookieVProc  = LocalCookie (rail, env->NodeId);
+    
+    EPRINTF8 (DBG_RCVR, "%s: ep3rcvr_rpc_complete: rxd %p [XID=%llx] statusblk source=%08x dest=%08x len=%x Cookies=%x.%x\n", rail->Generic.Name, rxd,
+             (long long) env->Xid.Unique, dmabe.s.dma_source, dmabe.s.dma_dest, dmabe.s.dma_size, dmabe.s.dma_destCookieVProc, 
+             dmabe.s.dma_srcCookieVProc);
+
+    for (len = 0, i = EP_MAXFRAG, remote += (nFrags-1), local += (nFrags-1); i > EP_MAXFRAG-nFrags; len += local->nmd_len, i--, local--, remote--)
+    {
+       /* copy down previous dma */
+       elan3_sdram_copyq_to_sdram (dev, &dmabe, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Dmas[i]),  sizeof (E3_DMA));    /* PCI write block */
+       
+       dmabe.s.dma_type            = E3_DMA_TYPE(DMA_BYTE, DMA_WRITE, DMA_NORMAL, EP3_DMAFAILCOUNT);
+       dmabe.s.dma_size            = local->nmd_len;
+       dmabe.s.dma_source          = local->nmd_addr;
+       dmabe.s.dma_dest            = remote->nmd_addr;
+       dmabe.s.dma_destEvent       = (E3_Addr) 0;
+       dmabe.s.dma_destCookieVProc = EP_VP_DATA (env->NodeId);
+       dmabe.s.dma_srcEvent        = rxdRail->RxdElanAddr + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[i-1]);
+       dmabe.s.dma_srcCookieVProc  = LocalCookie (rail, env->NodeId);
+       
+       EPRINTF9 (DBG_RCVR, "%s: ep3rcvr_rpc_complete: rxd %p [XID=%llx] idx=%d Source=%08x Dest=%08x Len=%x Cookies=%x.%x\n", rail->Generic.Name, rxd,
+                 (long long) env->Xid.Unique, i, local->nmd_addr, remote->nmd_addr, local->nmd_len, dmabe.s.dma_destCookieVProc, 
+                 dmabe.s.dma_srcCookieVProc);
+    }
+    
+    for (i = EP_MAXFRAG-nFrags; i < EP_MAXFRAG; i++)
+       elan3_sdram_writel (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[i].ev_Count), 1);                            /* PCI write */
+    
+    /* Initialise the done event */
+    elan3_sdram_writel (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count), 1);                                   /* PCI write */
+    rxdMain->DoneEvent  = EP3_EVENT_ACTIVE;
+
+    ASSERT (rail->Generic.Nodes[env->NodeId].State >= EP_NODE_CONNECTED && rail->Generic.Nodes[env->NodeId].State <= EP_NODE_LOCAL_PASSIVATE);
+
+    if (IssueDma (rail, &dmabe, EP_RETRY_LOW_PRI, FALSE) != ISSUE_COMMAND_OK)
+    {
+       /* Failed to issue the dma command, so copy the dma descriptor and queue it for retry */
+       EPRINTF2 (DBG_RCVR, "%s: ep3rcvr_rpc_complete: queue rxd %p on retry thread\n", rail->Generic.Name, rxd);
+       
+       QueueDmaForRetry (rail, &dmabe, EP_RETRY_LOW_PRI);
+    }
+
+    BucketStat (rxd->Rcvr->Subsys, CompleteRPC, len);
+}
+       
+void
+ep3rcvr_add_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *commsRail)
+{
+    EP3_RAIL          *rail   = (EP3_RAIL *) commsRail->Rail;
+    sdramaddr_t        qdescs = ((EP3_COMMS_RAIL *) commsRail)->QueueDescs;
+    EP3_RCVR_RAIL     *rcvrRail;
+    EP3_InputQueue     qdesc;
+    sdramaddr_t        stack;
+    unsigned long      flags;
+
+    KMEM_ZALLOC (rcvrRail, EP3_RCVR_RAIL *, sizeof (EP3_RCVR_RAIL), TRUE);
+
+    kcondvar_init (&rcvrRail->CleanupSleep);
+    spin_lock_init (&rcvrRail->FreeDescLock);
+    INIT_LIST_HEAD (&rcvrRail->FreeDescList);
+    INIT_LIST_HEAD (&rcvrRail->DescBlockList);
+
+    rcvrRail->Generic.CommsRail = commsRail;
+    rcvrRail->Generic.Rcvr      = rcvr;
+
+    rcvrRail->RcvrMain       = ep_alloc_main (&rail->Generic, sizeof (EP3_RCVR_RAIL_MAIN), 0, &rcvrRail->RcvrMainAddr);
+    rcvrRail->RcvrElan       = ep_alloc_elan (&rail->Generic, sizeof (EP3_RCVR_RAIL_ELAN), 0, &rcvrRail->RcvrElanAddr);
+    rcvrRail->InputQueueBase = ep_alloc_elan (&rail->Generic, EP_INPUTQ_SIZE * rcvr->InputQueueEntries, 0, &rcvrRail->InputQueueAddr);
+    stack                    = ep_alloc_elan (&rail->Generic, EP3_STACK_SIZE, 0, &rcvrRail->ThreadStack);
+
+    rcvrRail->TotalDescCount = 0;
+    rcvrRail->FreeDescCount  = 0;
+
+    /* Initialise the main/elan spin lock */
+    elan3_sdram_writel (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, ThreadLock.sl_lock), 0);
+    elan3_sdram_writel (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, ThreadLock.sl_seq),  0);
+
+    elan3_sdram_writel (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingLock.sl_lock), 0);
+    elan3_sdram_writel (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingLock.sl_seq), 0);
+    
+    /* Initialise the receive lists */
+    elan3_sdram_writel (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingDescs), 0);
+    
+    /* Initialise the ThreadShould Halt */
+    elan3_sdram_writel (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, ThreadShouldHalt), 0);
+
+    /* Initialise pointer to the ep_rcvr_rail */
+    elan3_sdram_writeq (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, MainAddr), (unsigned long) rcvrRail);
+
+    /* Initialise elan visible main memory */
+    rcvrRail->RcvrMain->ThreadLock.sl_seq  = 0;
+    rcvrRail->RcvrMain->PendingLock.sl_seq = 0;
+    rcvrRail->RcvrMain->PendingDescsTailp  = 0;
+
+    /* initialise and copy down the input queue descriptor */
+    qdesc.q_state          = E3_QUEUE_FULL;
+    qdesc.q_base           = rcvrRail->InputQueueAddr;
+    qdesc.q_top            = rcvrRail->InputQueueAddr + (rcvr->InputQueueEntries-1) * EP_INPUTQ_SIZE;
+    qdesc.q_fptr           = rcvrRail->InputQueueAddr;
+    qdesc.q_bptr           = rcvrRail->InputQueueAddr + EP_INPUTQ_SIZE;
+    qdesc.q_size           = EP_INPUTQ_SIZE;
+    qdesc.q_event.ev_Count = 0;
+    qdesc.q_event.ev_Type  = 0;
+
+    elan3_sdram_copyl_to_sdram (rail->Device, &qdesc, qdescs + rcvr->Service * sizeof (EP3_InputQueue), sizeof (EP3_InputQueue));
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    rcvr->Rails[rail->Generic.Number] = &rcvrRail->Generic;
+    rcvr->RailMask |= EP_RAIL2RAILMASK (rail->Generic.Number);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    /* initialise and run the Elan thread to process the queue */
+    IssueRunThread (rail, ep3_init_thread (rail->Device, ep_symbol (&rail->ThreadCode, "ep3comms_rcvr"),
+                                          rcvrRail->ThreadStack, stack, EP3_STACK_SIZE, 5,
+                                          rail->RailElanAddr, rcvrRail->RcvrElanAddr, rcvrRail->RcvrMainAddr,
+                                          EP_MSGQ_ADDR(rcvr->Service),
+                                          rail->ElanCookies));
+}
+
+void
+ep3rcvr_del_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *commsRail)
+{
+    EP3_RAIL         *rail     = (EP3_RAIL *) commsRail->Rail;
+    EP3_RCVR_RAIL    *rcvrRail = (EP3_RCVR_RAIL *) rcvr->Rails[rail->Generic.Number];  
+    unsigned long     flags;
+    struct list_head *el, *nel;
+
+    EPRINTF1 (DBG_RCVR, "%s: ep3rcvr_del_rail: removing rail\n", rail->Generic.Name);
+
+    /* flag the rail as no longer available */
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    rcvr->RailMask &= ~EP_RAIL2RAILMASK (rail->Generic.Number);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+    
+    /* mark the input queue descriptor as full */
+    SetQueueLocked(rail, ((EP3_COMMS_RAIL *)commsRail)->QueueDescs + rcvr->Service * sizeof (EP3_InputQueue));
+
+    /* need to halt the thread first         */
+    /*   set ThreadShouldHalt in elan memory */
+    /*   then trigger the event              */
+    /*   and wait on haltWait                */
+    elan3_sdram_writel  (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, ThreadShouldHalt), TRUE);
+
+    IssueSetevent (rail,  EP_MSGQ_ADDR(rcvr->Service) + offsetof(EP3_InputQueue, q_event));
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+
+    while (rcvrRail->ThreadHalted == 0)
+    {
+       rcvrRail->CleanupWaiting++;
+       kcondvar_wait (&rcvrRail->CleanupSleep, &rcvr->Lock, &flags);
+    }
+
+    /* at this point the thread is halted and it has no envelopes */
+ 
+    /* we need to wait until all the rxd's in the list that are 
+     * bound to the rail we are removing are not pending 
+     */
+    for (;;)
+    {
+       int mustWait = 0;
+       
+       list_for_each (el, &rcvr->ActiveDescList) {
+           EP_RXD       *rxd     = list_entry (el,EP_RXD, Link);
+           EP3_RXD_RAIL *rxdRail = (EP3_RXD_RAIL *) rxd->RxdRail;
+
+           if (rxdRail && RXD_BOUND2RAIL (rxdRail, rcvrRail) && rxd->RxdMain->Len != EP_RXD_PENDING)
+           {
+               mustWait++;
+               break;
+           }
+       }
+       
+       if (! mustWait)
+           break;
+
+       EPRINTF1 (DBG_RCVR, "%s: ep3rcvr_del_rail: waiting for active rxd's to be returned\n", rail->Generic.Name);
+
+       rcvrRail->CleanupWaiting++;
+       kcondvar_wait (&rcvrRail->CleanupSleep, &rcvr->Lock, &flags);
+    }
+
+    /* at this point all rxd's in the list that are bound to the deleting rail are not pending */
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD       *rxd     = list_entry (el, EP_RXD, Link);
+       EP3_RXD_RAIL *rxdRail = (EP3_RXD_RAIL *) rxd->RxdRail;
+       
+       if (rxdRail && RXD_BOUND2RAIL (rxdRail, rcvrRail))
+       {
+           /* here we need to unbind the remaining rxd's */
+           rxdRail->RxdMain->DataEvent = EP3_EVENT_PRIVATE;
+           rxdRail->RxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+ 
+           elan3_sdram_writel (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 0);  /* PCI write */
+           elan3_sdram_writel (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count), 0);  /* PCI write */
+
+           UnbindRxdFromRail (rxd, rxdRail);
+           FreeRxdRail(rcvrRail,  rxdRail );
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+    
+    /* wait for all rxd's for this rail to become free */
+    spin_lock_irqsave (&rcvrRail->FreeDescLock, flags);
+    while (rcvrRail->FreeDescCount != rcvrRail->TotalDescCount)
+    {
+       rcvrRail->FreeDescWaiting++;
+       kcondvar_wait (&rcvrRail->FreeDescSleep, &rcvrRail->FreeDescLock, &flags);
+    }
+    spin_unlock_irqrestore (&rcvrRail->FreeDescLock, flags);
+
+    /* can now remove the rail as it can no longer be used */
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    rcvr->Rails[rail->Generic.Number] = NULL;
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    /* all the rxd's accociated with DescBlocks must be in the FreeDescList */
+    ASSERT (rcvrRail->TotalDescCount == rcvrRail->FreeDescCount);
+
+    /* run through the DescBlockList deleting them */
+    while (!list_empty (&rcvrRail->DescBlockList))
+       FreeRxdRailBlock (rcvrRail, list_entry(rcvrRail->DescBlockList.next, EP3_RXD_RAIL_BLOCK , Link));
+
+    /* it had better be empty after that */
+    ASSERT ((rcvrRail->TotalDescCount == 0) && (rcvrRail->TotalDescCount == rcvrRail->FreeDescCount));
+    
+    ep_free_elan (&rail->Generic, rcvrRail->ThreadStack, EP3_STACK_SIZE);
+    ep_free_elan (&rail->Generic, rcvrRail->InputQueueAddr, EP_INPUTQ_SIZE * rcvr->InputQueueEntries);
+    ep_free_elan (&rail->Generic, rcvrRail->RcvrElanAddr, sizeof (EP3_RCVR_RAIL_ELAN));
+    ep_free_main (&rail->Generic, rcvrRail->RcvrMainAddr, sizeof (EP3_RCVR_RAIL_MAIN));
+
+    KMEM_FREE (rcvrRail, sizeof (EP3_RCVR_RAIL));
+}
+
+EP_RXD *
+ep3rcvr_steal_rxd (EP_RCVR_RAIL *r)
+{
+    EP3_RCVR_RAIL *rcvrRail = (EP3_RCVR_RAIL *) r;
+    EP3_RAIL      *rail     = RCVR_TO_RAIL (rcvrRail);
+    EP_RCVR       *rcvr     = rcvrRail->Generic.Rcvr;
+    E3_Addr        rxdElanAddr;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+
+    LockRcvrThread (rcvrRail);
+    if ((rxdElanAddr = elan3_sdram_readl (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingDescs))) != 0)
+    {
+       sdramaddr_t  rxdElan  = ep_elan2sdram (&rail->Generic, rxdElanAddr);
+       EP3_RXD_RAIL *rxdRail = (EP3_RXD_RAIL *) (unsigned long) elan3_sdram_readq (rail->Device, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, MainAddr));
+       EP_RXD      *rxd      = rxdRail->Generic.Rxd;
+       sdramaddr_t  next;
+       
+       EPRINTF2 (DBG_RCVR, "%s: StealRxdFromOtherRail stealing rxd %p\n", rail->Generic.Name, rail);
+       
+       /* Remove the RXD from the pending desc list */
+       if ((next = elan3_sdram_readl (rail->Device, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Next))) == 0)
+           rcvrRail->RcvrMain->PendingDescsTailp = 0;
+       elan3_sdram_writel (rail->Device, rcvrRail->RcvrElan + offsetof (EP3_RCVR_RAIL_ELAN, PendingDescs), next);
+       UnlockRcvrThread (rcvrRail);
+       
+       UnbindRxdFromRail (rxd, rxdRail);
+       
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+       
+       /* Mark rxdRail as no longer active */
+       rxdRail->RxdMain->DataEvent = EP3_EVENT_PRIVATE;
+       rxdRail->RxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+       elan3_sdram_writel (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 0);
+       elan3_sdram_writel (rail->Device, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count), 0);
+       
+       FreeRxdRail (rcvrRail, rxdRail);
+
+       return rxd;
+    }
+
+    UnlockRcvrThread (rcvrRail);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    return NULL;
+}
+
+long
+ep3rcvr_check (EP_RCVR_RAIL *r, long nextRunTime)
+{
+    EP3_RCVR_RAIL    *rcvrRail = (EP3_RCVR_RAIL *) r;
+    EP3_RAIL         *rail     = RCVR_TO_RAIL (rcvrRail);
+    EP_RCVR          *rcvr     = rcvrRail->Generic.Rcvr;
+    EP_COMMS_SUBSYS *subsys    = rcvr->Subsys;
+    EP_SYS           *sys       = subsys->Subsys.Sys;
+    EP_RXD           *rxd;
+    unsigned long     flags;
+
+    if (rcvrRail->FreeDescCount < ep_rxd_lowat && !AllocateRxdRailBlock (rcvrRail))
+    {
+       EPRINTF1 (DBG_RCVR,"%s: failed to grow rxd rail pool\n", rail->Generic.Name);
+               
+       if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + RESOURCE_RETRY_TIME))
+           nextRunTime = lbolt + RESOURCE_RETRY_TIME;
+    }
+    
+    if (rcvrRail->ThreadWaiting && (rxd = StealRxdFromOtherRail (rcvr)) != NULL)
+    {
+       /* Map the receive buffer into this rail as well */
+       EPRINTF4 (DBG_RCVR, "%s: mapping rxd->Data (%08x.%08x.%08x) into this rails\n",
+                 rail->Generic.Name, rxd->Data.nmd_addr,rxd->Data.nmd_len, rxd->Data.nmd_attr);
+
+       spin_lock_irqsave (&rcvr->Lock, flags);
+       if ((!(EP_NMD_RAILMASK (&rxd->Data) & EP_RAIL2RAILMASK(rail->Generic.Number)) &&                /* not already mapped and */
+            ep_nmd_map_rails (sys, &rxd->Data, EP_RAIL2RAILMASK(rail->Generic.Number)) == 0) ||        /* failed to map it */
+           ep3rcvr_queue_rxd (rxd, &rcvrRail->Generic))                                                /* or failed to queue it */
+       {
+           EPRINTF5 (DBG_RCVR,"%s: stolen rcvr=%p rxd=%p -> rnum=%d rcvrRail=%p (failed)\n", 
+                     rail->Generic.Name, rcvr, rxd, rail->Generic.Number, rcvrRail);
+               
+           if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + RESOURCE_RETRY_TIME))
+               nextRunTime = lbolt + RESOURCE_RETRY_TIME;
+       }
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+    }
+    
+    return nextRunTime;
+}
+
+static void
+ep3rcvr_flush_filtering (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail)
+{
+    EP3_COMMS_RAIL *commsRail = (EP3_COMMS_RAIL *) rcvrRail->Generic.CommsRail;
+    EP3_RAIL       *rail      = (EP3_RAIL *) commsRail->Generic.Rail;
+    ELAN3_DEV      *dev       = rail->Device;
+    sdramaddr_t    qdesc      = commsRail->QueueDescs + rcvr->Service*sizeof (EP3_InputQueue);
+    E3_Addr        qTop       = elan3_sdram_readl (dev, qdesc + offsetof (EP3_InputQueue, q_top));
+    E3_Addr        qBase      = elan3_sdram_readl (dev, qdesc + offsetof (EP3_InputQueue, q_base));
+    E3_Addr        qSize      = elan3_sdram_readl (dev,qdesc + offsetof (EP3_InputQueue, q_size));
+    E3_uint32      nfptr, qbptr;
+    unsigned long  flags;
+    
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    LockRcvrThread (rcvrRail);                                                                         /* PCI lock */
+
+    nfptr = elan3_sdram_readl (dev, qdesc + offsetof (EP3_InputQueue, q_fptr));
+    qbptr = elan3_sdram_readl (dev, qdesc + offsetof (EP3_InputQueue, q_bptr));
+    
+    if (nfptr == qTop)
+       nfptr = qBase;
+    else
+       nfptr += qSize;
+    
+    while (nfptr != qbptr)
+    {
+       unsigned nodeId = elan3_sdram_readl (dev, rcvrRail->InputQueueBase + (nfptr - rcvrRail->InputQueueAddr) + 
+                                      offsetof (EP_ENVELOPE, NodeId));
+       
+       EPRINTF3 (DBG_DISCON, "%s: ep3rcvr_flush_filtering: nodeId=%d State=%d\n", rail->Generic.Name, nodeId, rail->Generic.Nodes[nodeId].State);
+       
+       if (rail->Generic.Nodes[nodeId].State == EP_NODE_LOCAL_PASSIVATE)
+           elan3_sdram_writel (dev, rcvrRail->InputQueueBase + (nfptr - rcvrRail->InputQueueAddr) + 
+                         offsetof (EP_ENVELOPE, Version), 0);
+       
+       if (nfptr == qTop)
+           nfptr = qBase;
+       else
+           nfptr += qSize;
+    }
+    
+    UnlockRcvrThread (rcvrRail);                                                                               /* PCI unlock */
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+static void
+ep3rcvr_flush_flushing (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail)
+{
+    EP3_RAIL         *rail = RCVR_TO_RAIL (rcvrRail);
+    struct list_head *el, *nel;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    LockRcvrThread (rcvrRail);                                                                         /* PCI lock */
+    
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD       *rxd      = list_entry (el, EP_RXD, Link);
+       EP3_RXD_RAIL *rxdRail  = (EP3_RXD_RAIL *) rxd->RxdRail;
+       EP_ENVELOPE  *env      = &rxd->RxdMain->Envelope;
+       EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[env->NodeId];
+
+       if (rxd->RxdMain->Len == EP_RXD_PENDING || !RXD_BOUND2RAIL(rxdRail,rcvrRail) || nodeRail->State != EP_NODE_LOCAL_PASSIVATE)
+           continue;
+       
+       EPRINTF6 (DBG_DISCON, "%s: ep3rcvr_flush_flushing: rcvr %p rxd %p state %x.%x elan node %d\n", rail->Generic.Name,
+                 rcvr, rxd, rxdRail->RxdMain->DataEvent, rxdRail->RxdMain->DoneEvent, env->NodeId);
+       
+       switch (rxd->State)
+       {
+       case EP_RXD_FREE:
+           printk ("ep3rcvr_flush_flushing: rxd state is free but bound to a fail\n");
+           break;
+
+       case EP_RXD_RECEIVE_ACTIVE:
+           if (rxdRail->RxdMain->DataEvent == EP3_EVENT_ACTIVE)                /* incomplete message receive */
+           {
+               EPRINTF4 (DBG_RCVR, "%s: ep3rcvr_flush_flushing: rcvr %p rxd %p nodeId %d - passive\n", 
+                         rail->Generic.Name, rcvr, rxd, env->NodeId);
+               
+               nodeRail->MessageState |= EP_NODE_PASSIVE_MESSAGES;
+               continue;
+           }
+           break;
+           
+       default:
+           EP_ASSERT (&rail->Generic, EP_IS_RPC(env->Attr));
+
+           if (!EP3_EVENT_FIRED (rxdRail->DoneCookie, rxdRail->RxdMain->DoneEvent))    /* incomplete RPC */
+           {
+               EPRINTF4 (DBG_RCVR, "%s: ep3rcvr_flush_flushing: rcvr %p rxd %p nodeId %d - active\n", 
+                         rail->Generic.Name, rcvr, rxd, env->NodeId);
+               
+               EP_INVALIDATE_XID (rxd->MsgXid);                        /* Ignore any previous NMD map responses */
+               
+               nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+               continue;
+           }
+           break;
+
+       case EP_RXD_BEEN_ABORTED:
+           printk ("ep3rcvr_flush_flushing: rxd state is aborted but bound to a fail\n");
+           break;
+       }
+
+       EPRINTF4 (DBG_RCVR, "%s: ep3rcvr_flush_flushing: rcvr %p rxd %p nodeId %d - finished\n", 
+                 rail->Generic.Name, rcvr, rxd, env->NodeId);
+    }    
+
+    UnlockRcvrThread (rcvrRail);                                                                       /* PCI unlock */
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+void
+ep3rcvr_flush_callback (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail)
+{
+    EP3_RAIL *rail = RCVR_TO_RAIL(rcvrRail);
+
+    switch (rail->Generic.CallbackStep)
+    {
+    case EP_CB_FLUSH_FILTERING:
+       ep3rcvr_flush_filtering (rcvr, rcvrRail);
+       break;
+
+    case EP_CB_FLUSH_FLUSHING:
+       ep3rcvr_flush_flushing (rcvr, rcvrRail);
+       break;
+    }
+}
+
+void
+ep3rcvr_failover_callback (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail)
+{
+    EP_COMMS_SUBSYS  *subsys = rcvr->Subsys;
+    EP3_RAIL         *rail   = RCVR_TO_RAIL (rcvrRail);
+    ELAN3_DEV        *dev    = rail->Device;
+    struct list_head *el, *nel;
+    unsigned long     flags;
+#ifdef SUPPORT_RAIL_FAILOVER
+    EP_SYS           *sys    = subsys->Subsys.Sys;
+#endif
+   
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    LockRcvrThread (rcvrRail);                                                                         /* PCI lock */
+    
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD             *rxd      = list_entry (el, EP_RXD, Link);
+       EP3_RXD_RAIL       *rxdRail  = (EP3_RXD_RAIL *) rxd->RxdRail;
+       EP_ENVELOPE        *env      = &rxd->RxdMain->Envelope;
+       EP_NODE_RAIL       *nodeRail = &rail->Generic.Nodes[env->NodeId];
+#ifdef SUPPORT_RAIL_FAILOVER
+       EP_MANAGER_MSG_BODY msgBody;
+       EP_NODE            *node     = &sys->Nodes[env->NodeId];
+#endif
+       
+       if (rxd->RxdMain->Len == EP_RXD_PENDING || !RXD_BOUND2RAIL(rxdRail,rcvrRail) || nodeRail->State != EP_NODE_PASSIVATED)
+           continue;
+
+       EPRINTF6 (DBG_FAILOVER, "%s: ep3rcvr_failover_callback: rcvr %p rxd %p elan node %d state %x.%x\n", rail->Generic.Name, rcvr, rxd, env->NodeId,
+                 rxdRail->RxdMain->DataEvent, rxdRail->RxdMain->DoneEvent);
+
+       switch (rxd->State)
+       {
+       case EP_RXD_FREE:
+           printk ("ep4rcvr_failover_callback: rxd state is free but bound to a fail\n");
+           break;
+
+       case EP_RXD_RECEIVE_ACTIVE:
+           if (rxdRail->RxdMain->DataEvent == EP3_EVENT_ACTIVE)                /* incomplete message receive */
+           {
+               EPRINTF4 (DBG_FAILOVER, "%s: ep3rcvr_failover_callback: rcvr %p rxd %p nodeId %d - unbind\n", rail->Generic.Name, rcvr, rxd, env->NodeId);
+               
+               UnbindRxdFromRail (rxd, rxdRail);
+               
+               /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+               rxdRail->RxdMain->DataEvent = EP3_EVENT_PRIVATE;
+               rxdRail->RxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+               
+               /* clear the data event - the done event should already be zero */
+               elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 0);       /* PCI write */
+               
+               FreeRxdRail (rcvrRail, rxdRail);
+               
+               /* epcomms thread will requeue on different rail */
+               ep_kthread_schedule (&subsys->Thread, lbolt);
+               continue;
+           }
+           break;
+
+       default:
+           EP_ASSERT (&rail->Generic, EP_IS_RPC(env->Attr));
+
+#ifdef SUPPORT_RAIL_FAILOVER
+           if (!EP3_EVENT_FIRED (rxdRail->DoneCookie, rxdRail->RxdMain->DoneEvent) && !(EP_IS_NO_FAILOVER(env->Attr)))         /* incomplete RPC, which can be failed over  */
+           {
+               EPRINTF7 (DBG_FAILOVER, "%s: ep3rcvr_failover_callback: rxd %p State %x.%x Xid %llxx MsgXid %llxx nodeId %d - failover\n", 
+                         rail->Generic.Name, rxd, rxdRail->RxdMain->DataEvent, rxdRail->RxdMain->DoneEvent, 
+                         (long long) env->Xid.Unique, (long long) rxd->MsgXid.Unique, env->NodeId);
+               
+               if (EP_XID_INVALID(rxd->MsgXid))
+                   rxd->MsgXid = ep_xid_cache_alloc (sys, &rcvr->XidCache);
+               
+               /* XXXX maybe only send the message if the node failover retry is now ? */
+               msgBody.Failover.Xid      = env->Xid;
+               msgBody.Failover.Railmask = node->ConnectedRails;
+               
+               ep_send_message (&rail->Generic, env->NodeId, EP_MANAGER_MSG_TYPE_FAILOVER_REQUEST, rxd->MsgXid, &msgBody);
+               
+               nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+               continue;
+           }
+#endif
+           break;
+
+       case EP_RXD_BEEN_ABORTED:
+           printk ("ep4rcvr_failover_callback: rxd state is aborted but bound to a fail\n");
+           break;
+       }
+
+       EPRINTF3 (DBG_FAILOVER, "%s: ep3rcvr_failover_callback: rxd %p nodeId %d - finished\n", rail->Generic.Name, rxd, env->NodeId);
+    }
+    
+    UnlockRcvrThread (rcvrRail);                                                                       /* PCI unlock */
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+void
+ep3rcvr_disconnect_callback (EP_RCVR *rcvr, EP3_RCVR_RAIL *rcvrRail)
+{
+    EP3_RAIL         *rail = RCVR_TO_RAIL (rcvrRail);
+    ELAN3_DEV        *dev = rail->Device;
+    struct list_head *el, *nel;
+    struct list_head  rxdList;
+    unsigned long     flags;
+
+    INIT_LIST_HEAD (&rxdList);
+    
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    LockRcvrThread (rcvrRail);                                                                         /* PCI lock */
+    
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD       *rxd      = list_entry (el, EP_RXD, Link);
+       EP3_RXD_RAIL *rxdRail  = (EP3_RXD_RAIL *) rxd->RxdRail;
+       EP_ENVELOPE  *env      = &rxd->RxdMain->Envelope;
+       EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[env->NodeId];
+       
+       if (rxd->RxdMain->Len == EP_RXD_PENDING || !RXD_BOUND2RAIL(rxdRail,rcvrRail) || nodeRail->State != EP_NODE_DISCONNECTING)
+           continue;
+
+       EPRINTF4 (DBG_DISCON, "%s: ep3rcvr_disconnect_callback: rcvr %p rxd %p elan node %d\n", rail->Generic.Name, rcvr, rxd, env->NodeId);
+
+       switch (rxd->State)
+       {
+       case EP_RXD_FREE:
+           printk ("ep3rcvr_disconnect_callback: rxd state is free but bound to a fail\n");
+           break;
+
+       case EP_RXD_RECEIVE_ACTIVE:
+           if (rxdRail->RxdMain->DataEvent == EP3_EVENT_ACTIVE)                        /* incomplete message receive */
+           {
+               EPRINTF4 (DBG_RCVR, "%s: ep3rcvr_disconnect_callback: rcvr %p rxd %p nodeId %d - unbind\n", rail->Generic.Name, rcvr, rxd, env->NodeId);
+               
+               UnbindRxdFromRail (rxd, rxdRail);
+               
+               /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+               rxdRail->RxdMain->DataEvent = EP3_EVENT_PRIVATE;
+               rxdRail->RxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+               
+               /* clear the data event - the done event should already be zero */
+               elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 0);       /* PCI write */
+               
+               FreeRxdRail (rcvrRail, rxdRail);
+
+               /* remark it as pending if it was partially received */
+               rxd->RxdMain->Len = EP_RXD_PENDING;
+               
+               /* epcomms thread will requeue on different rail */
+               ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+               continue;
+           }
+           break;
+
+       default:
+           EP_ASSERT (&rail->Generic, EP_IS_RPC(env->Attr));
+
+           if (!EP3_EVENT_FIRED (rxdRail->DoneCookie, rxdRail->RxdMain->DoneEvent))    /* incomplete RPC */
+           {
+               EPRINTF4 (DBG_RCVR, "%s: ep3rcvr_disconnect_callback: rcvr %p rxd %p nodeId %d - not able to failover\n",
+                         rail->Generic.Name, rcvr, rxd, env->NodeId);
+           
+               /* Mark as no longer active */
+               rxdRail->RxdMain->DataEvent = EP3_EVENT_PRIVATE;
+               rxdRail->RxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+               
+               elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count), 0);       /* PCI write */
+               elan3_sdram_writel (dev, rxdRail->RxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count), 0);       /* PCI write */
+               
+               UnbindRxdFromRail (rxd, rxdRail);
+               FreeRxdRail (rcvrRail, rxdRail);
+
+               /* Ignore any previous NMD/failover responses */
+               EP_INVALIDATE_XID (rxd->MsgXid);
+               
+               /* Remove from active list */
+               list_del (&rxd->Link);
+               
+               if (rxd->State == EP_RXD_RPC_IN_PROGRESS)                               /* ownder by user .... */
+                   rxd->State = EP_RXD_BEEN_ABORTED;
+               else                                                                    /* queue for completion */
+               {
+                   rxd->RxdMain->Len = EP_CONN_RESET;                                  /* ensure ep_rxd_status() fails */
+                   list_add_tail (&rxd->Link, &rxdList);
+               }
+               continue;
+           }
+           break;
+
+       case EP_RXD_BEEN_ABORTED:
+           printk ("ep4rcvr_failover_callback: rxd state is aborted but bound to a fail\n");
+           break;
+       }
+           
+       EPRINTF4 (DBG_RCVR, "%s: ep3rcvr_disconnect_callback: rcvr %p rxd %p nodeId %d - finished\n", 
+                 rail->Generic.Name, rcvr, rxd, env->NodeId);
+    }
+    
+    UnlockRcvrThread (rcvrRail);                                                                       /* PCI unlock */
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    while (! list_empty (&rxdList)) 
+    {
+       EP_RXD *rxd = list_entry (rxdList.next, EP_RXD, Link);
+
+       list_del (&rxd->Link);
+
+       rxd->Handler (rxd);
+    }
+}
+
+void
+ep3rcvr_display_rxd (DisplayInfo *di, EP_RXD_RAIL *r)
+{
+    EP3_RXD_RAIL *rxdRail = (EP3_RXD_RAIL *) r;
+    sdramaddr_t   rxdElan = rxdRail->RxdElan;
+    EP3_RAIL     *rail    = RCVR_TO_RAIL (rxdRail->Generic.RcvrRail);
+    ELAN3_DEV    *dev     = rail->Device;
+
+    (di->func)(di->arg, "      ChainEvent=%x.%x %x.%x\n",
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[0].ev_Count)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[0].ev_Type)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[1].ev_Count)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[1].ev_Type)));
+    (di->func)(di->arg, "      ChainEvent=%x.%x %x.%x\n",
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[2].ev_Count)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[2].ev_Type)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[3].ev_Count)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, ChainEvent[3].ev_Type)));
+    (di->func)(di->arg, "      DataEvent=%x.%x DoneEvent=%x.%x\n",
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Count)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DataEvent.ev_Type)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Count)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, DoneEvent.ev_Type)));
+    (di->func)(di->arg, "      Data=%x Len=%x\n",
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Data.nmd_addr)),
+              elan3_sdram_readl (dev, rxdElan + offsetof (EP3_RXD_RAIL_ELAN, Data.nmd_len)));
+}
+
+void
+ep3rcvr_display_rcvr (DisplayInfo *di, EP_RCVR_RAIL *r)
+{
+    EP3_RCVR_RAIL  *rcvrRail  = (EP3_RCVR_RAIL *) r;
+    EP3_COMMS_RAIL *commsRail = (EP3_COMMS_RAIL *) rcvrRail->Generic.CommsRail;
+    EP3_RAIL       *rail      = RCVR_TO_RAIL (rcvrRail);
+    ELAN3_DEV      *dev       = rail->Device;
+    sdramaddr_t     queue     = commsRail->QueueDescs + rcvrRail->Generic.Rcvr->Service * sizeof (EP3_InputQueue);
+    E3_Addr         qbase      = elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_base));
+    E3_Addr         qtop       = elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_top));
+    E3_uint32       qsize      = elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_size));
+    int             freeCount  = 0;
+    int             blockCount = 0;
+    unsigned long   flags;
+    struct list_head *el;
+
+    spin_lock_irqsave (&rcvrRail->FreeDescLock, flags);
+    list_for_each (el, &rcvrRail->FreeDescList)
+       freeCount++;
+    list_for_each (el, &rcvrRail->DescBlockList)
+       blockCount++;
+    spin_unlock_irqrestore (&rcvrRail->FreeDescLock, flags);
+
+    (di->func)(di->arg, "                 Rail %d FreeDesc %d (%d) Total %d Blocks %d %s\n",
+              rail->Generic.Number, rcvrRail->FreeDescCount, freeCount, rcvrRail->TotalDescCount, blockCount, 
+              rcvrRail->ThreadWaiting ? "ThreadWaiting" : "");
+    
+    (di->func)(di->arg, "                 InputQueue state=%x bptr=%x size=%x top=%x base=%x fptr=%x\n",
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_state)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_bptr)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_size)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_top)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_base)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_fptr)));
+    (di->func)(di->arg, "                            event=%x.%x [%x.%x] wevent=%x.%x\n",
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_event.ev_Type)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_event.ev_Count)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_event.ev_Source)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_event.ev_Dest)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_wevent)),
+              elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_wcount)));
+    
+    LockRcvrThread (rcvrRail);
+    {
+       E3_Addr     nfptr = elan3_sdram_readl (dev, queue + offsetof (EP3_InputQueue, q_fptr));
+       EP_ENVELOPE env;
+       
+       if (nfptr == qtop)
+           nfptr = qbase;
+       else
+           nfptr += qsize;
+
+       while (nfptr != elan3_sdram_readl (dev, queue + offsetof (E3_Queue, q_bptr)))
+       {
+           elan3_sdram_copyl_from_sdram (dev, rcvrRail->InputQueueBase + (nfptr - rcvrRail->InputQueueAddr),
+                                         &env, sizeof (EP_ENVELOPE));
+           
+           (di->func)(di->arg, "                 ENVELOPE Version=%x Attr=%x Xid=%08x.%08x.%016llx\n",
+                      env.Version, env.Attr, env.Xid.Generation, env.Xid.Handle, (long long) env.Xid.Unique);
+           (di->func)(di->arg, "                          NodeId=%x Range=%x TxdRail=%x TxdMain=%x.%x.%x\n",
+                      env.NodeId, env.Range, env.TxdRail, env.TxdMain.nmd_addr,
+                      env.TxdMain.nmd_len, env.TxdMain.nmd_attr);
+           
+           
+           if (nfptr == qtop)
+               nfptr = qbase;
+           else
+               nfptr += qsize;
+       }
+    }
+    UnlockRcvrThread (rcvrRail);
+}
+
+void
+ep3rcvr_fillout_rail_stats(EP_RCVR_RAIL *rcvr_rail, char *str) {
+    /* no stats here yet */
+    /* EP3_RCVR_RAIL * ep4rcvr_rail = (EP3_RCVR_RAIL *) rcvr_rail; */
+}
+
Index: linux-2.4.21/drivers/net/qsnet/ep/epcommsRx_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcommsRx_elan4.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcommsRx_elan4.c 2005-06-01 23:12:54.653430896 -0400
@@ -0,0 +1,1758 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcommsRx_elan4.c,v 1.30.2.2 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcommsRx_elan4.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "debug.h"
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "epcomms_elan4.h"
+
+#include <elan4/trtype.h>
+
+#define RCVR_TO_COMMS(rcvrRail)                ((EP4_COMMS_RAIL *) ((EP_RCVR_RAIL *) rcvrRail)->CommsRail)
+#define RCVR_TO_RAIL(rcvrRail)         ((EP4_RAIL *) ((EP_RCVR_RAIL *) rcvrRail)->CommsRail->Rail)
+#define RCVR_TO_DEV(rcvrRail)          (RCVR_TO_RAIL(rcvrRail)->r_ctxt.ctxt_dev)
+#define RCVR_TO_SUBSYS(rcvrRail)       (((EP_RCVR_RAIL *) rcvrRail)->Rcvr->Subsys)
+
+#define RXD_TO_RCVR(txdRail)           ((EP4_RCVR_RAIL *) rxdRail->rxd_generic.RcvrRail)
+#define RXD_TO_RAIL(txdRail)           RCVR_TO_RAIL(RXD_TO_RCVR(rxdRail))
+
+static void rxd_interrupt (EP4_RAIL *rail, void *arg);
+
+static __inline__ void 
+__ep4_rxd_assert_free (EP4_RXD_RAIL *rxdRail, const char *file, const int line)
+{
+    EP4_RCVR_RAIL *rcvrRail = RXD_TO_RCVR(rxdRail);
+    ELAN4_DEV     *dev      = RCVR_TO_DEV(rcvrRail);
+    register int i, failed = 0;
+    
+    for (i = 0; i <= EP_MAXFRAG; i++)
+       if (((rxdRail)->rxd_main->rxd_sent[i] != EP4_STATE_FREE)) 
+           failed |= (1 << i);
+    
+    if (((rxdRail)->rxd_main->rxd_failed != EP4_STATE_FREE))
+       failed |= (1 << 5);
+    if (((rxdRail)->rxd_main->rxd_done   != EP4_STATE_FREE)) 
+       failed |= (1 << 6);
+    
+    if (sdram_assert)
+    {
+       if (((elan4_sdram_readq (RXD_TO_RAIL(rxdRail)->r_ctxt.ctxt_dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_CountAndType)) >> 32) != 0)) 
+           failed |= (1 << 7);
+       for (i = 0; i < EP_MAXFRAG; i++)
+           if (((elan4_sdram_readq (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[i].ev_CountAndType)) >> 32) != 0)) 
+               failed |= (1 << (8 + i));
+       if (((elan4_sdram_readq (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType)) >> 32) != 0)) 
+           failed |= (1 << 12);
+       if (((int)(elan4_sdram_readq (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CountAndType)) >> 32) != -32)) 
+           failed |= (1 << 13);
+    }
+
+    if (failed)
+    {
+       printk ("__ep4_rxd_assert_free: failed=%x rxdRail=%p %s - %d\n", failed, rxdRail, file, line);
+
+       ep_debugf (DBG_DEBUG, "__ep4_rxd_assert_free: failed=%x rxdRail=%p %s - %d\n", failed, rxdRail, file, line);
+       ep4rcvr_display_rxd (&di_ep_debug, &rxdRail->rxd_generic);
+
+       for (i = 0; i <= EP_MAXFRAG; i++)
+           (rxdRail)->rxd_main->rxd_sent[i] = EP4_STATE_FREE;
+
+       (rxdRail)->rxd_main->rxd_failed = EP4_STATE_FREE;
+       (rxdRail)->rxd_main->rxd_done   = EP4_STATE_FREE;
+
+       if (sdram_assert)
+       {
+           elan4_sdram_writew (RXD_TO_RAIL(rxdRail)->r_ctxt.ctxt_dev,
+                               (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_CountAndType) + 4, 0);
+
+           for (i = 0; i < EP_MAXFRAG; i++)
+               elan4_sdram_writew (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[i].ev_CountAndType) + 4, 0);
+           elan4_sdram_writew (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType) + 4, 0);
+           elan4_sdram_writew (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CountAndType) + 4, -32);
+       }
+       EP_ASSFAIL (RCVR_TO_RAIL(rcvrRail), "__ep4_rxd_assert_free");
+    }
+}
+
+static __inline__ void
+__ep4_rxd_assert_pending(EP4_RXD_RAIL *rxdRail, const char *file, const int line)
+{ 
+    EP4_RCVR_RAIL *rcvrRail = RXD_TO_RCVR(rcvrRail);
+    register int failed = 0;
+
+    failed |= ((rxdRail)->rxd_main->rxd_done != EP4_STATE_ACTIVE);
+
+    if (failed)
+    {
+       printk ("__ep4_rxd_assert_pending: %s - %d\n", file, line);
+
+       ep_debugf (DBG_DEBUG, "__ep4_rxd_assert_pending: %s - %d\n", file, line);
+       ep4rcvr_display_rxd (&di_ep_debug, &rxdRail->rxd_generic);
+
+       (rxdRail)->rxd_main->rxd_done = EP4_STATE_ACTIVE;
+
+       EP_ASSFAIL (RCVR_TO_RAIL(rcvrRail), "__ep4_rxd_assert_pending");
+    }
+}
+
+static __inline__ void
+__ep4_rxd_assert_private(EP4_RXD_RAIL *rxdRail, const char *file, const int line)
+{
+    EP4_RCVR_RAIL *rcvrRail = RXD_TO_RCVR(rxdRail);
+    ELAN4_DEV     *dev      = RCVR_TO_DEV(rcvrRail);
+    register int failed = 0;
+
+    if (((rxdRail)->rxd_main->rxd_failed != EP4_STATE_ACTIVE)) failed |= (1 << 0);
+    if (((rxdRail)->rxd_main->rxd_done != EP4_STATE_PRIVATE))  failed |= (1 << 1);
+    
+    if (sdram_assert)
+    {
+       if (((elan4_sdram_readq (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType)) >> 32) != 0))           failed |= (1 << 2);
+       if (((int) (elan4_sdram_readq (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CountAndType)) >> 32) != -32)) failed |= (1 << 3);
+    }
+
+    if (failed)
+    {
+       printk ("__ep4_rxd_assert_private: %s - %d\n", file, line);
+
+       ep_debugf (DBG_DEBUG, "__ep4_rxd_assert_private: %s - %d\n", file, line);
+       ep4rcvr_display_rxd (&di_ep_debug, &rxdRail->rxd_generic);
+
+       (rxdRail)->rxd_main->rxd_failed = EP4_STATE_ACTIVE;
+       (rxdRail)->rxd_main->rxd_done   = EP4_STATE_PRIVATE;
+
+       if (sdram_assert)
+       {
+           elan4_sdram_writew (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType) + 4, 0);
+           elan4_sdram_writew (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CountAndType) + 4, -32);
+       }
+
+       EP_ASSFAIL (RCVR_TO_RAIL(rcvrRail), "__ep4_rxd_assert_private");
+    }
+}
+
+static __inline__ void
+__ep4_rxd_private_to_free (EP4_RXD_RAIL *rxdRail)
+{
+    register int i;
+
+    for (i = 0; i <= EP_MAXFRAG; i++)
+       rxdRail->rxd_main->rxd_sent[i] = EP4_STATE_FREE;
+ 
+    rxdRail->rxd_main->rxd_failed = EP4_STATE_FREE;
+    rxdRail->rxd_main->rxd_done   = EP4_STATE_FREE;
+}
+
+static __inline__ void
+__ep4_rxd_force_private (EP4_RXD_RAIL *rxdRail)
+{
+    EP4_RAIL  *rail = RXD_TO_RAIL(rxdRail);
+    ELAN4_DEV *dev  = rail->r_ctxt.ctxt_dev;
+
+    (rxdRail)->rxd_main->rxd_failed = EP4_STATE_ACTIVE;
+    (rxdRail)->rxd_main->rxd_done = EP4_STATE_PRIVATE;
+
+    if (sdram_assert) 
+       elan4_sdram_writeq (dev, (rxdRail)->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE(0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+}
+
+#define EP4_RXD_ASSERT_FREE(rxdRail)           __ep4_rxd_assert_free(rxdRail, __FILE__, __LINE__)
+#define EP4_RXD_ASSERT_PENDING(rxdRail)                __ep4_rxd_assert_pending(rxdRail, __FILE__, __LINE__)
+#define EP4_RXD_ASSERT_PRIVATE(rxdRail)                __ep4_rxd_assert_private(rxdRail, __FILE__, __LINE__)
+#define EP4_RXD_PRIVATE_TO_FREE(rxdRail)       __ep4_rxd_private_to_free(rxdRail)
+#define EP4_RXD_FORCE_PRIVATE(rxdRail)         __ep4_rxd_force_private(rxdRail)
+
+static int
+alloc_rxd_block (EP4_RCVR_RAIL *rcvrRail)
+{
+    EP4_RAIL           *rail = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV          *dev  = rail->r_ctxt.ctxt_dev;
+    EP4_RXD_RAIL_BLOCK *blk;
+    EP4_RXD_RAIL_MAIN  *rxdMain;
+    EP_ADDR            rxdMainAddr;
+    sdramaddr_t                rxdElan;
+    EP_ADDR            rxdElanAddr;
+    EP4_RXD_RAIL       *rxdRail;
+    unsigned long       flags;
+    int                 i, j;
+
+    KMEM_ZALLOC (blk, EP4_RXD_RAIL_BLOCK *, sizeof (EP4_RXD_RAIL_BLOCK), 1);
+
+    if (blk == NULL)
+       return 0;
+
+    if ((rxdElan = ep_alloc_elan (&rail->r_generic, EP4_RXD_RAIL_ELAN_SIZE * EP4_NUM_RXD_PER_BLOCK, 0, &rxdElanAddr)) == (sdramaddr_t) 0)
+    {
+       KMEM_FREE (blk, sizeof (EP4_RXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    if ((rxdMain = ep_alloc_main (&rail->r_generic, EP4_RXD_RAIL_MAIN_SIZE * EP4_NUM_RXD_PER_BLOCK, 0, &rxdMainAddr)) == (EP4_RXD_RAIL_MAIN *) NULL)
+    {
+       ep_free_elan (&rail->r_generic, rxdElanAddr, EP4_RXD_RAIL_ELAN_SIZE * EP4_NUM_RXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP4_RXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    if (ep4_reserve_dma_retries (rail, EP4_NUM_RXD_PER_BLOCK, 0) != 0)
+    {
+       ep_free_main (&rail->r_generic, blk->blk_rxds[0].rxd_main_addr, EP4_RXD_RAIL_MAIN_SIZE * EP4_NUM_RXD_PER_BLOCK);
+       ep_free_elan (&rail->r_generic, rxdElanAddr, EP4_RXD_RAIL_ELAN_SIZE * EP4_NUM_RXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP4_RXD_RAIL_BLOCK));
+
+       return 0;
+    }
+
+    for (rxdRail = &blk->blk_rxds[0], i = 0; i < EP4_NUM_RXD_PER_BLOCK; i++, rxdRail++)
+    {
+       rxdRail->rxd_generic.RcvrRail = &rcvrRail->rcvr_generic;
+       rxdRail->rxd_elan             = rxdElan;
+       rxdRail->rxd_elan_addr        = rxdElanAddr;
+       rxdRail->rxd_main             = rxdMain;
+       rxdRail->rxd_main_addr        = rxdMainAddr;
+
+       /* reserve 128 bytes of "event" cq space for the chained STEN packets */
+       if ((rxdRail->rxd_ecq = ep4_get_ecq (rail, EP4_ECQ_EVENT, EP4_RXD_STEN_CMD_NDWORDS)) == NULL)
+           goto failed;
+
+       /* allocate a single word of "setevent" command space */
+       if ((rxdRail->rxd_scq = ep4_get_ecq (rail, EP4_ECQ_SINGLE, 1)) == NULL)
+       {
+           ep4_put_ecq (rail, rxdRail->rxd_ecq, EP4_RXD_STEN_CMD_NDWORDS);
+           goto failed;
+       }
+
+       /* initialise the completion events */
+       for (j = 0; j <= EP_MAXFRAG; j++)
+           rxdMain->rxd_sent[i] = EP4_STATE_FREE;
+
+       rxdMain->rxd_done   = EP4_STATE_FREE;
+       rxdMain->rxd_failed = EP4_STATE_FREE;
+
+       /* initialise the scq for the thread */
+       rxdMain->rxd_scq = rxdRail->rxd_scq->ecq_addr;
+
+       /* initialise the "start" event to copy the first STEN packet into the command queue */
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE(0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_START_CMD_NDWORDS));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_CopySource),
+                           rxdElanAddr + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[0]));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_CopyDest),
+                           rxdRail->rxd_ecq->ecq_addr);
+
+       /* initialise the "chain" events to copy the next STEN packet into the command queue */
+       for (j = 0; j < EP_MAXFRAG; j++)
+       {
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[j].ev_CountAndType),
+                               E4_EVENT_INIT_VALUE(0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_STEN_CMD_NDWORDS));
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[j].ev_CopySource),
+                               rxdElanAddr + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j+1]));
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[j].ev_CopyDest),
+                               rxdRail->rxd_ecq->ecq_addr);
+       }
+
+       /* initialise the portions of the sten packets which don't change */
+       for (j = 0; j < EP_MAXFRAG+1; j++)
+       {
+           if (j < EP_MAXFRAG)
+               elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_dma_dstEvent),
+                                   rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[j]));
+           else
+               elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_dma_dstEvent),
+                                   rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done));
+
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_ok_guard),
+                               GUARD_CMD | GUARD_CHANNEL (1) | GUARD_TEST(0, PACK_OK) | GUARD_RESET (EP4_STEN_RETRYCOUNT));
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_ok_write_cmd),
+                               WRITE_DWORD_CMD | (rxdMainAddr + offsetof (EP4_RXD_RAIL_MAIN, rxd_sent[j])));
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_ok_write_value),
+                               EP4_STATE_FINISHED);
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_fail_guard),
+                               GUARD_CMD | GUARD_CHANNEL (1) | GUARD_TEST(0, RESTART_COUNT_ZERO) | GUARD_RESET (EP4_STEN_RETRYCOUNT));
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_fail_setevent),
+                               SET_EVENT_CMD | (rxdElanAddr + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed)));
+           elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[j].c_nop_cmd),
+                               NOP_CMD);
+       }
+
+       /* register a main interrupt cookie */
+       ep4_register_intcookie (rail, &rxdRail->rxd_intcookie, rxdElanAddr + offsetof (EP4_RXD_RAIL_ELAN, rxd_done),
+                               rxd_interrupt, rxdRail);
+
+       /* initialise the command stream for the done event */
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done_cmd.c_write_cmd),
+                           WRITE_DWORD_CMD | (rxdMainAddr + offsetof (EP4_RXD_RAIL_MAIN, rxd_done)));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done_cmd.c_write_value),
+                           EP4_STATE_FINISHED);
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done_cmd.c_intr_cmd),
+                           INTERRUPT_CMD | (rxdRail->rxd_intcookie.int_val << E4_MAIN_INT_SHIFT));
+
+       /* initialise the command stream for the fail event */
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed_cmd.c_write_cmd),
+                           WRITE_DWORD_CMD | (rxdMainAddr + offsetof (EP4_RXD_RAIL_MAIN, rxd_failed)));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed_cmd.c_write_value),
+                           EP4_STATE_FAILED);
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed_cmd.c_intr_cmd),
+                           INTERRUPT_CMD | (rxdRail->rxd_intcookie.int_val << E4_MAIN_INT_SHIFT));
+
+       /* initialise the done and fail events */
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE(0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CopySource),
+                           rxdElanAddr + offsetof (EP4_RXD_RAIL_ELAN, rxd_done_cmd));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CopyDest),
+                           rxdRail->rxd_ecq->ecq_addr);
+
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CopySource),
+                           rxdElanAddr + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed_cmd));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CopyDest),
+                           rxdRail->rxd_ecq->ecq_addr);
+       
+       /* initialise the pointer to the main memory portion */
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_main), 
+                           rxdMainAddr);
+
+       /* move onto next descriptor */
+       rxdElan     += EP4_RXD_RAIL_ELAN_SIZE;
+       rxdElanAddr += EP4_RXD_RAIL_ELAN_SIZE;
+       rxdMain      = (EP4_RXD_RAIL_MAIN *) ((unsigned long) rxdMain + EP4_RXD_RAIL_MAIN_SIZE);
+       rxdMainAddr += EP4_RXD_RAIL_MAIN_SIZE;
+    }
+
+    spin_lock_irqsave (&rcvrRail->rcvr_freelock, flags);
+
+    list_add  (&blk->blk_link, &rcvrRail->rcvr_blocklist);
+
+    rcvrRail->rcvr_totalcount += EP4_NUM_RXD_PER_BLOCK;
+    rcvrRail->rcvr_freecount  += EP4_NUM_RXD_PER_BLOCK;
+
+    for (i = 0; i < EP4_NUM_RXD_PER_BLOCK; i++)
+       list_add (&blk->blk_rxds[i].rxd_generic.Link, &rcvrRail->rcvr_freelist);
+
+    spin_unlock_irqrestore (&rcvrRail->rcvr_freelock, flags);
+
+    return 1;
+
+ failed:
+    while (--i >= 0)
+    {
+       rxdRail--;
+
+       ep4_put_ecq (rail, rxdRail->rxd_ecq, EP4_RXD_STEN_CMD_NDWORDS);
+       ep4_put_ecq (rail, rxdRail->rxd_scq, 1);
+
+       ep4_deregister_intcookie (rail, &rxdRail->rxd_intcookie);
+    }
+
+    ep4_release_dma_retries (rail, EP4_NUM_RXD_PER_BLOCK);
+    
+    ep_free_main (&rail->r_generic, blk->blk_rxds[0].rxd_main_addr, EP4_RXD_RAIL_MAIN_SIZE * EP4_NUM_RXD_PER_BLOCK);
+    ep_free_elan (&rail->r_generic, rxdElanAddr, EP4_RXD_RAIL_ELAN_SIZE * EP4_NUM_RXD_PER_BLOCK);
+    KMEM_FREE (blk, sizeof (EP4_RXD_RAIL_BLOCK));
+
+    return 0;
+}
+
+
+static void
+free_rxd_block (EP4_RCVR_RAIL *rcvrRail, EP4_RXD_RAIL_BLOCK *blk)
+{
+    EP4_RAIL     *rail = RCVR_TO_RAIL (rcvrRail);
+    EP4_RXD_RAIL *rxdRail;
+    unsigned long flags;
+    int           i;
+
+    spin_lock_irqsave (&rcvrRail->rcvr_freelock, flags);
+
+    list_del (&blk->blk_link);
+
+    rcvrRail->rcvr_totalcount -= EP4_NUM_RXD_PER_BLOCK;
+
+    for (rxdRail = &blk->blk_rxds[0], i = 0; i < EP4_NUM_RXD_PER_BLOCK; i++, rxdRail++)
+    {
+       rcvrRail->rcvr_freecount--;
+
+       ep4_put_ecq (rail, rxdRail->rxd_ecq, EP4_RXD_STEN_CMD_NDWORDS);
+       ep4_put_ecq (rail, rxdRail->rxd_scq, 1);
+
+       ep4_deregister_intcookie (rail, &rxdRail->rxd_intcookie);
+
+       list_del (&rxdRail->rxd_generic.Link);
+    }
+    spin_unlock_irqrestore (&rcvrRail->rcvr_freelock, flags);
+
+    ep4_release_dma_retries (rail, EP4_NUM_RXD_PER_BLOCK);
+
+    ep_free_main (&rail->r_generic, blk->blk_rxds[0].rxd_main_addr, EP4_RXD_RAIL_MAIN_SIZE * EP4_NUM_RXD_PER_BLOCK);
+    ep_free_elan (&rail->r_generic, blk->blk_rxds[0].rxd_elan_addr, EP4_RXD_RAIL_ELAN_SIZE * EP4_NUM_RXD_PER_BLOCK);
+
+    KMEM_FREE (blk, sizeof (EP4_RXD_RAIL_BLOCK));
+}
+
+static EP4_RXD_RAIL *
+get_rxd_rail (EP4_RCVR_RAIL *rcvrRail)
+{
+    EP_COMMS_SUBSYS  *subsys = RCVR_TO_SUBSYS(rcvrRail);
+    EP4_RXD_RAIL     *rxdRail;
+    unsigned long flags;
+    int low_on_rxds;
+
+    spin_lock_irqsave (&rcvrRail->rcvr_freelock, flags);
+
+    if (list_empty (&rcvrRail->rcvr_freelist))
+       rxdRail = NULL;
+    else
+    {
+       rxdRail = list_entry (rcvrRail->rcvr_freelist.next, EP4_RXD_RAIL, rxd_generic.Link);
+
+       EP4_RXD_ASSERT_FREE(rxdRail);
+
+       list_del (&rxdRail->rxd_generic.Link);
+
+       rcvrRail->rcvr_freecount--;
+    }
+    /* Wakeup the descriptor primer thread if there's not many left */
+    low_on_rxds = (rcvrRail->rcvr_freecount < ep_rxd_lowat);
+
+    spin_unlock_irqrestore (&rcvrRail->rcvr_freelock, flags);
+
+    if (low_on_rxds)
+       ep_kthread_schedule (&subsys->Thread, lbolt);
+
+    return (rxdRail);
+}
+
+static void
+free_rxd_rail (EP4_RCVR_RAIL *rcvrRail, EP4_RXD_RAIL *rxdRail)
+{
+    unsigned long flags;
+
+    EP4_RXD_ASSERT_FREE(rxdRail);
+
+    spin_lock_irqsave (&rcvrRail->rcvr_freelock, flags);
+    
+    list_add (&rxdRail->rxd_generic.Link, &rcvrRail->rcvr_freelist);
+
+    rcvrRail->rcvr_freecount++;
+
+    if (rcvrRail->rcvr_freewaiting)
+    {
+       rcvrRail->rcvr_freewaiting--;
+       kcondvar_wakeupall (&rcvrRail->rcvr_freesleep, &rcvrRail->rcvr_freelock);
+    }
+
+    spin_unlock_irqrestore (&rcvrRail->rcvr_freelock, flags);
+}
+
+static void
+bind_rxd_rail (EP_RXD *rxd, EP4_RXD_RAIL *rxdRail)
+{
+    EP4_RAIL *rail = RCVR_TO_RAIL (rxdRail->rxd_generic.RcvrRail);
+
+    ASSERT (SPINLOCK_HELD (&rxd->Rcvr->Lock));
+
+    EPRINTF3 (DBG_RCVR, "%s: bind_rxd_rail: rxd=%p rxdRail=%p\n",  rail->r_generic.Name, rxd, rxdRail);
+
+    elan4_sdram_writeq (rail->r_ctxt.ctxt_dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_rxd), rxd->NmdMain.nmd_addr);                      /* PCI write */
+
+    rxd->RxdRail             = &rxdRail->rxd_generic;
+    rxdRail->rxd_generic.Rxd = rxd;
+}
+
+static void
+unbind_rxd_rail (EP_RXD *rxd, EP4_RXD_RAIL *rxdRail)
+{
+    EP4_RCVR_RAIL *rcvrRail = (EP4_RCVR_RAIL *) rxdRail->rxd_generic.RcvrRail;
+    
+    ASSERT (SPINLOCK_HELD (&rxd->Rcvr->Lock));
+    ASSERT (rxd->RxdRail == &rxdRail->rxd_generic && rxdRail->rxd_generic.Rxd == rxd);
+
+    EP4_RXD_ASSERT_PRIVATE (rxdRail);
+
+    EPRINTF3 (DBG_RCVR, "%s: unbind_rxd_rail: rxd=%p rxdRail=%p\n",  RCVR_TO_RAIL(rcvrRail)->r_generic.Name, rxd, rxdRail);
+
+    rxd->RxdRail             = NULL;
+    rxdRail->rxd_generic.Rxd = NULL;
+
+    if (rcvrRail->rcvr_cleanup_waiting)
+       kcondvar_wakeupall (&rcvrRail->rcvr_cleanup_sleep, &rxd->Rcvr->Lock);
+    rcvrRail->rcvr_cleanup_waiting = 0;
+
+    EP4_RXD_PRIVATE_TO_FREE (rxdRail);
+}
+
+
+static void
+rcvr_stall_interrupt (EP4_RAIL *rail, void *arg)
+{
+    EP4_RCVR_RAIL *rcvrRail = (EP4_RCVR_RAIL *) arg;
+    EP_RCVR       *rcvr     = rcvrRail->rcvr_generic.Rcvr;
+    unsigned long  flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    
+    EPRINTF1 (DBG_RCVR, "rcvr_stall_interrupt: rcvrRail %p thread halted\n", rcvrRail);
+
+    rcvrRail->rcvr_thread_halted = 1;
+
+    kcondvar_wakeupall (&rcvrRail->rcvr_cleanup_sleep, &rcvr->Lock);
+
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+static void
+rcvr_stall_haltop (ELAN4_DEV *dev, void *arg)
+{
+    EP4_RCVR_RAIL  *rcvrRail  = (EP4_RCVR_RAIL *) arg;
+    EP4_COMMS_RAIL *commsRail = RCVR_TO_COMMS(rcvrRail);
+    EP_RCVR        *rcvr      = rcvrRail->rcvr_generic.Rcvr;
+    sdramaddr_t     qdesc     = ((EP4_COMMS_RAIL *) commsRail)->r_descs + (rcvr->Service * EP_QUEUE_DESC_SIZE);
+    E4_uint64       qbptr     = elan4_sdram_readq (dev, qdesc + offsetof (E4_InputQueue, q_bptr));
+
+    /* Mark the queue as full by writing the fptr */
+    if (qbptr == (rcvrRail->rcvr_slots_addr + EP_INPUTQ_SIZE * (rcvr->InputQueueEntries-1)))
+       elan4_sdram_writeq (dev, qdesc + offsetof (E4_InputQueue, q_fptr), rcvrRail->rcvr_slots_addr);
+    else
+       elan4_sdram_writeq (dev, qdesc + offsetof (E4_InputQueue, q_fptr), qbptr + EP_INPUTQ_SIZE);
+
+    /* Notify the thread that it should stall after processing any outstanding envelopes */
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_stall_intcookie),
+                       rcvrRail->rcvr_stall_intcookie.int_val);
+
+    /* Issue a swtevent to the queue event to wake the thread up */
+    ep4_set_event_cmd (rcvrRail->rcvr_resched, rcvrRail->rcvr_elan_addr + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_qevent));
+}
+
+static void
+rxd_interrupt (EP4_RAIL *rail, void *arg)
+{
+    EP4_RXD_RAIL      *rxdRail  = (EP4_RXD_RAIL *) arg;
+    EP4_RCVR_RAIL     *rcvrRail = (EP4_RCVR_RAIL *) rxdRail->rxd_generic.RcvrRail;
+    EP_RCVR           *rcvr     = rcvrRail->rcvr_generic.Rcvr;
+    EP4_RXD_RAIL_MAIN *rxdMain  = rxdRail->rxd_main;
+    unsigned long      delay    = 1;
+    EP_RXD            *rxd;
+    EP_ENVELOPE       *env;
+    unsigned long      flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+
+    for (;;)
+    {
+       if (rxdMain->rxd_done == EP4_STATE_FINISHED || rxdMain->rxd_failed == EP4_STATE_FAILED)
+           break;
+
+       /* The write to rxd_done could be held up in the PCI bridge even though
+        * we've seen the interrupt cookie.  Unlike elan3, there is no possibility
+        * of spurious interrupts since we flush the command queues on node 
+        * disconnection and the txcallback mechanism */
+       mb();
+
+       if (delay > EP4_EVENT_FIRING_TLIMIT)
+       {
+           spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+           EP_ASSFAIL (RCVR_TO_RAIL(rcvrRail), "rxd_interrupt - not finished\n");
+           return;
+       }
+       DELAY(delay);
+       delay <<= 1;
+    }
+
+    if (rxdMain->rxd_done != EP4_STATE_FINISHED)
+    {
+       EPRINTF8 (DBG_RETRY, "%s: rxd_interrupt: rxdRail %p retry: done=%d failed=%d NodeId=%d XID=%08x.%08x.%016llx\n",
+                 rail->r_generic.Name, rxdRail, (int)rxdMain->rxd_done, (int)rxdMain->rxd_failed, rxdRail->rxd_generic.Rxd->RxdMain->Envelope.NodeId,
+                 rxdRail->rxd_generic.Rxd->RxdMain->Envelope.Xid.Generation, rxdRail->rxd_generic.Rxd->RxdMain->Envelope.Xid.Handle, 
+                 rxdRail->rxd_generic.Rxd->RxdMain->Envelope.Xid.Unique);
+    
+       spin_lock (&rcvrRail->rcvr_retrylock);
+
+       rxdRail->rxd_retry_time = lbolt + EP_RETRY_LOW_PRI_TIME;                        /* XXXX backoff ? */
+
+       list_add_tail (&rxdRail->rxd_retry_link, &rcvrRail->rcvr_retrylist);
+
+       ep_kthread_schedule (&rail->r_retry_thread, rxdRail->rxd_retry_time);
+       spin_unlock (&rcvrRail->rcvr_retrylock);
+
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+       return;
+    }
+    
+    rxd = rxdRail->rxd_generic.Rxd;
+    env = &rxd->RxdMain->Envelope;
+
+    /*
+     * Note, since the thread will have sent the remote dma packet before copying 
+     * the envelope, we must check that it has completed doing this,  we do this
+     * by acquiring the spinlock against the thread which it only drops once it's
+     * completed.
+     */
+    if (rxd->RxdMain->Len == EP_RXD_PENDING)
+    {
+       EP4_SPINENTER (rail->r_ctxt.ctxt_dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock),
+                      &rcvrRail->rcvr_main->rcvr_thread_lock);
+       
+       EP4_SPINEXIT (rail->r_ctxt.ctxt_dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock),
+                     &rcvrRail->rcvr_main->rcvr_thread_lock);
+       
+       ASSERT (env->Version == EP_ENVELOPE_VERSION && rxd->RxdMain->Len != EP_RXD_PENDING);
+    }
+
+    EPRINTF8 (DBG_RCVR, "%s: rxd_interrupt: rxd %p finished from %d XID %08x.%08x.%016llx len %d attr %x\n", rail->r_generic.Name, 
+             rxd, rxd->RxdMain->Envelope.NodeId, rxd->RxdMain->Envelope.Xid.Generation, rxd->RxdMain->Envelope.Xid.Handle, 
+             rxd->RxdMain->Envelope.Xid.Unique,  rxd->RxdMain->Len, rxd->RxdMain->Envelope.Attr);
+
+    rxdMain->rxd_done  = EP4_STATE_PRIVATE;
+    rxd->Data.nmd_attr = EP_RAIL2RAILMASK (rail->r_generic.Number);
+
+    switch (rxd->State)
+    {
+    case EP_RXD_RECEIVE_ACTIVE:
+       if (rxd->RxdMain->Len >= 0 && EP_IS_RPC(env->Attr))
+           rxd->State = EP_RXD_RPC_IN_PROGRESS;
+       else
+       {
+           rxd->State = EP_RXD_COMPLETED;
+
+           /* remove from active list */
+           list_del (&rxd->Link);
+
+           unbind_rxd_rail (rxd, rxdRail);
+           free_rxd_rail (rcvrRail, rxdRail);
+       }
+
+       if (rxd->RxdMain->Len >= 0) {
+           INC_STAT(rcvrRail->rcvr_generic.stats,rx);
+           ADD_STAT(rcvrRail->rcvr_generic.stats,rx_len,rxd->RxdMain->Len);
+           INC_STAT(rail->r_generic.Stats,rx);
+           ADD_STAT(rail->r_generic.Stats,rx_len,rxd->RxdMain->Len);
+       }
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+       ep_rxd_received (rxd);
+
+       break;
+
+    case EP_RXD_PUT_ACTIVE:
+    case EP_RXD_GET_ACTIVE:
+       rxd->State = EP_RXD_RPC_IN_PROGRESS;
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+       
+       rxd->Handler (rxd);
+       break;
+
+    case EP_RXD_COMPLETE_ACTIVE:
+       rxd->State = EP_RXD_COMPLETED;
+
+       /* remove from active list */
+       list_del (&rxd->Link);
+
+       unbind_rxd_rail (rxd, rxdRail);
+       free_rxd_rail (rcvrRail, rxdRail);
+
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+       rxd->Handler(rxd);
+       break;
+
+    default:
+       spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+       printk ("%s: rxd_interrupt: rxd %p in invalid state %d\n", rail->r_generic.Name, rxd, rxd->State);
+       /* NOTREACHED */
+    }
+}
+
+static void
+ep4rcvr_flush_filtering (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail)
+{
+    EP4_COMMS_RAIL *commsRail = RCVR_TO_COMMS(rcvrRail);
+    EP4_RAIL       *rail      = RCVR_TO_RAIL(rcvrRail);
+    ELAN4_DEV      *dev       = rail->r_ctxt.ctxt_dev;
+    sdramaddr_t    qdesc      = commsRail->r_descs + (rcvr->Service * EP_QUEUE_DESC_SIZE);
+    E4_Addr        qbase      = rcvrRail->rcvr_slots_addr;
+    E4_Addr        qlast      = qbase + EP_INPUTQ_SIZE * (rcvr->InputQueueEntries-1);
+    E4_uint64      qfptr, qbptr;
+    unsigned long  flags;
+    
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    EP4_SPINENTER (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    
+    /* zip down the input queue and invalidate any envelope we find to a node which is locally passivated */
+    qfptr = elan4_sdram_readq (dev, qdesc + offsetof (E4_InputQueue, q_fptr));
+    qbptr = elan4_sdram_readq (dev, qdesc + offsetof (E4_InputQueue, q_bptr));
+
+    while (qfptr != qbptr)
+    {
+       unsigned int nodeId = elan4_sdram_readl (dev, rcvrRail->rcvr_slots + (qfptr - qbase) + offsetof (EP_ENVELOPE, NodeId));
+
+       EPRINTF3 (DBG_DISCON, "%s: ep4rcvr_flush_filtering: nodeId=%d State=%d\n", rail->r_generic.Name, nodeId, rail->r_generic.Nodes[nodeId].State);
+       
+       if (rail->r_generic.Nodes[nodeId].State == EP_NODE_LOCAL_PASSIVATE)
+           elan4_sdram_writel (dev,  rcvrRail->rcvr_slots + (qfptr - qbase) + offsetof (EP_ENVELOPE, Version), 0);
+       
+       if (qfptr != qlast)
+           qfptr += EP_INPUTQ_SIZE;
+       else
+           qfptr = qbase;
+    }
+
+    /* Insert an setevent command into the thread's command queue
+     * to ensure that all sten packets have completed */
+    elan4_guard (rcvrRail->rcvr_ecq->ecq_cq, GUARD_ALL_CHANNELS);
+    ep4comms_flush_setevent (commsRail, rcvrRail->rcvr_ecq->ecq_cq);
+    
+    EP4_SPINEXIT (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+static void
+ep4rcvr_flush_flushing (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail)
+{
+    EP4_RAIL         *rail = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV       *dev  = rail->r_ctxt.ctxt_dev;
+    struct list_head *el, *nel;
+    struct list_head  rxdList;
+    unsigned long     flags;
+
+    INIT_LIST_HEAD (&rxdList);
+    
+    /* remove any sten packates which are retrying to nodes which are being passivated */
+    spin_lock_irqsave (&rcvrRail->rcvr_retrylock, flags);
+    list_for_each_safe (el, nel, &rcvrRail->rcvr_retrylist) {
+       EP4_RXD_RAIL *rxdRail  = list_entry (el, EP4_RXD_RAIL, rxd_retry_link);
+       EP_ENVELOPE  *env      = &rxdRail->rxd_generic.Rxd->RxdMain->Envelope;
+       EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[env->NodeId];
+
+       if (nodeRail->State == EP_NODE_LOCAL_PASSIVATE)
+       {
+           EPRINTF2 (DBG_XMTR, "%s; ep4rcvr_flush_flushing: removing rxdRail %p from retry list\n", rail->r_generic.Name, rxdRail);
+           
+           list_del (&rxdRail->rxd_retry_link);
+       }
+    }
+    spin_unlock_irqrestore (&rcvrRail->rcvr_retrylock, flags);
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    EP4_SPINENTER (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD       *rxd      = list_entry (el, EP_RXD, Link);
+       EP4_RXD_RAIL *rxdRail  = (EP4_RXD_RAIL *) rxd->RxdRail;
+       EP_ENVELOPE  *env      = &rxd->RxdMain->Envelope;
+       EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[env->NodeId];
+
+       if (rxd->RxdMain->Len == EP_RXD_PENDING || !RXD_BOUND2RAIL (rxdRail, rcvrRail) || nodeRail->State != EP_NODE_LOCAL_PASSIVATE)
+           continue;
+       
+       EPRINTF5 (DBG_DISCON, "%s: ep4rcvr_flush_flushing: rcvr %p rxd %p state %d elan node %d\n", 
+                 rail->r_generic.Name, rcvr, rxd, (int)rxdRail->rxd_main->rxd_done, env->NodeId);
+       
+       switch (rxd->State)
+       {
+       case EP_RXD_FREE:
+           printk ("ep4rcvr_flush_flushing: rxd state is free but bound to a fail\n");
+           break;
+
+       case EP_RXD_RECEIVE_ACTIVE:
+           if (rxdRail->rxd_main->rxd_done == EP4_STATE_ACTIVE)                /* incomplete message receive */
+           {
+               EPRINTF4 (DBG_RCVR, "%s: ep4rcvr_flush_flushing: rcvr %p rxd %p nodeId %d - passive\n", 
+                         rail->r_generic.Name, rcvr, rxd, env->NodeId);
+               
+               nodeRail->MessageState |= EP_NODE_PASSIVE_MESSAGES;
+               continue;
+           }
+           break;
+           
+       default:
+           EP4_ASSERT (rail, EP_IS_RPC(env->Attr));
+
+           if (rxdRail->rxd_main->rxd_done == EP4_STATE_ACTIVE)                /* incomplete RPC */
+           {
+               EPRINTF4 (DBG_RCVR, "%s: ep4rcvr_flush_flushing: rcvr %p rxd %p nodeId %d - active\n", 
+                         rail->r_generic.Name, rcvr, rxd, env->NodeId);
+               
+               EP_INVALIDATE_XID (rxd->MsgXid);                        /* Ignore any previous NMD map responses */
+               
+               nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+               continue;
+           }
+           break;
+
+       case EP_RXD_BEEN_ABORTED:
+           printk ("ep4rcvr_flush_flushing: rxd state is aborted but bound to a fail\n");
+           break;
+       }
+
+       EPRINTF4 (DBG_RCVR, "%s: ep4rcvr_flush_flushing: rcvr %p rxd %p nodeId %d - finished\n", 
+                 rail->r_generic.Name, rcvr, rxd, env->NodeId);
+    }    
+
+    EP4_SPINEXIT (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+void
+ep4rcvr_flush_callback (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail)
+{
+    EP4_RAIL *rail = RCVR_TO_RAIL(rcvrRail);
+
+    switch (rail->r_generic.CallbackStep)
+    {
+    case EP_CB_FLUSH_FILTERING:
+       ep4rcvr_flush_filtering (rcvr, rcvrRail);
+       break;
+
+    case EP_CB_FLUSH_FLUSHING:
+       ep4rcvr_flush_flushing (rcvr, rcvrRail);
+       break;
+    }
+}
+
+void
+ep4rcvr_failover_callback (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail)
+{
+    EP_COMMS_SUBSYS  *subsys = rcvr->Subsys;
+    EP4_RAIL         *rail   = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV       *dev    = rail->r_ctxt.ctxt_dev;
+    struct list_head *el, *nel;
+    unsigned long     flags;
+#if SUPPORT_RAIL_FAILOVER
+    EP_SYS           *sys    = subsys->Subsys.Sys;
+#endif
+    
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    EP4_SPINENTER (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD             *rxd      = list_entry (el, EP_RXD, Link);
+       EP4_RXD_RAIL       *rxdRail  = (EP4_RXD_RAIL *) rxd->RxdRail;
+       EP_ENVELOPE        *env      = &rxd->RxdMain->Envelope;
+       EP_NODE_RAIL       *nodeRail = &rail->r_generic.Nodes[env->NodeId];
+#if SUPPORT_RAIL_FAILOVER
+       EP_NODE            *node     = &sys->Nodes[env->NodeId];
+       EP_MANAGER_MSG_BODY msgBody;
+#endif
+       
+       if (rxd->RxdMain->Len == EP_RXD_PENDING || !RXD_BOUND2RAIL(rxdRail,rcvrRail) || nodeRail->State != EP_NODE_PASSIVATED)
+           continue;
+
+       EPRINTF5 (DBG_FAILOVER, "%s: ep4rcvr_failover_callback: rcvr %p rxd %p elan node %d state %d\n", 
+                 rail->r_generic.Name, rcvr, rxd, env->NodeId, (int)rxdRail->rxd_main->rxd_done);
+
+       switch (rxd->State)
+       {
+       case EP_RXD_FREE:
+           printk ("ep4rcvr_failover_callback: rxd state is free but bound to a fail\n");
+           break;
+
+       case EP_RXD_RECEIVE_ACTIVE:
+           if (rxdRail->rxd_main->rxd_done == EP4_STATE_ACTIVE)                        /* incomplete message receive */
+           {
+               EPRINTF4 (DBG_FAILOVER, "%s: ep4rcvr_failover_callback: rcvr %p rxd %p nodeId %d - unbind\n", rail->r_generic.Name, rcvr, rxd, env->NodeId);
+
+               EP4_RXD_FORCE_PRIVATE(rxdRail);
+               
+               unbind_rxd_rail (rxd, rxdRail);
+
+               free_rxd_rail (rcvrRail, rxdRail);
+           
+               /* epcomms thread will requeue on different rail */
+               ep_kthread_schedule (&subsys->Thread, lbolt);
+               continue;
+           }
+           break;
+
+       default:
+           EP4_ASSERT (rail, EP_IS_RPC(env->Attr));
+
+#if SUPPORT_RAIL_FAILOVER
+           /* XXXX - no rail failover for now .... */
+           if (rxdRail->rxd_main->rxd_done == EP4_STATE_ACTIVE && !EP_IS_NO_FAILOVER(env->Attr))       /* incomplete RPC, which can be failed over */
+           {
+               EPRINTF6 (DBG_FAILOVER, "%s: ep4rcvr_failover_callback: rxd %p State %d Xid %llxx MsgXid %llxx nodeId %d - failover\n", 
+                         rail->r_generic.Name, rxd, rxd->State, env->Xid.Unique, rxd->MsgXid.Unique, env->NodeId);
+               
+               if (EP_XID_INVALID(rxd->MsgXid))
+                   rxd->MsgXid = ep_xid_cache_alloc (sys, &rcvr->XidCache);
+               
+               /* XXXX maybe only send the message if the node failover retry is now ? */
+               msgBody.Failover.Xid      = env->Xid;
+               msgBody.Failover.Railmask = node->ConnectedRails;
+               
+               ep_send_message (&rail->r_generic, env->NodeId, EP_MANAGER_MSG_TYPE_FAILOVER_REQUEST, rxd->MsgXid, &msgBody);
+               
+               nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+               continue;
+           }
+#endif
+           break;
+
+       case EP_RXD_BEEN_ABORTED:
+           printk ("ep4rcvr_failover_callback: rxd state is aborted but bound to a fail\n");
+           break;
+       }
+       EPRINTF3 (DBG_FAILOVER, "%s: ep4rcvr_failover_callback: rxd %p nodeId %d - finished\n", rail->r_generic.Name, rxd, env->NodeId);
+    }
+    
+    EP4_SPINEXIT (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+void
+ep4rcvr_disconnect_callback (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail)
+{
+    EP4_RAIL         *rail = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV        *dev = rail->r_ctxt.ctxt_dev;
+    struct list_head *el, *nel;
+    struct list_head  rxdList;
+    unsigned long     flags;
+
+    INIT_LIST_HEAD (&rxdList);
+    
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    EP4_SPINENTER (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD       *rxd      = list_entry (el, EP_RXD, Link);
+       EP4_RXD_RAIL *rxdRail  = (EP4_RXD_RAIL *) rxd->RxdRail;
+       EP_ENVELOPE  *env      = &rxd->RxdMain->Envelope;
+       EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[env->NodeId];
+       
+       if (rxd->RxdMain->Len == EP_RXD_PENDING || !RXD_BOUND2RAIL(rxdRail,rcvrRail) || nodeRail->State != EP_NODE_DISCONNECTING)
+           continue;
+
+       EPRINTF5 (DBG_DISCON, "%s: ep4rcvr_disconnect_callback: rcvr %p rxd %p elan node %d state %x\n", rail->r_generic.Name, rcvr, rxd, env->NodeId, rxd->State);
+
+       switch (rxd->State)
+       {
+       case EP_RXD_FREE:
+           printk ("ep4rcvr_disconnect_callback: rxd state is free but bound to a rail\n");
+           break;
+
+       case EP_RXD_RECEIVE_ACTIVE:
+           if (rxdRail->rxd_main->rxd_done == EP4_STATE_ACTIVE)                /* incomplete message receive */
+           {
+               EPRINTF4 (DBG_RCVR, "%s: ep4rcvr_disconnect_callback: rcvr %p rxd %p nodeId %d - unbind\n", rail->r_generic.Name, rcvr, rxd, env->NodeId);
+
+               EP4_RXD_FORCE_PRIVATE (rxdRail);
+               
+               unbind_rxd_rail (rxd, rxdRail);
+               free_rxd_rail (rcvrRail, rxdRail);
+               
+               /* remark it as pending if it was partially received */
+               rxd->RxdMain->Len = EP_RXD_PENDING;
+               
+               /* epcomms thread will requeue on different rail */
+               ep_kthread_schedule (&rcvr->Subsys->Thread, lbolt);
+               continue;
+           }
+           break;
+
+       default:
+           if (rxdRail->rxd_main->rxd_done == EP4_STATE_ACTIVE || rxdRail->rxd_main->rxd_done == EP4_STATE_PRIVATE)            /* incomplete RPC */
+           {
+               EPRINTF5 (DBG_RCVR, "%s: ep4rcvr_disconnect_callback: rcvr %p rxd %p nodeId %d state %x - not able to failover\n",
+                         rail->r_generic.Name, rcvr, rxd, env->NodeId, rxd->State);
+           
+               EP4_RXD_FORCE_PRIVATE (rxdRail);
+
+               unbind_rxd_rail (rxd, rxdRail);
+               free_rxd_rail (rcvrRail, rxdRail);
+
+               /* Ignore any previous NMD/failover responses */
+               EP_INVALIDATE_XID (rxd->MsgXid);
+               
+               /* Remove from active list */
+               list_del (&rxd->Link);
+               
+               if (rxd->State == EP_RXD_RPC_IN_PROGRESS)                               /* ownder by user .... */
+                   rxd->State = EP_RXD_BEEN_ABORTED;
+               else                                                                    /* queue for completion */
+               {
+                   rxd->RxdMain->Len = EP_CONN_RESET;                                  /* ensure ep_rxd_status() fails */
+                   list_add_tail (&rxd->Link, &rxdList);
+               }
+               continue;
+           }
+           break;
+
+       case EP_RXD_BEEN_ABORTED:
+           printk ("ep4rcvr_disconnect_callback: rxd state is aborted but bound to a rail\n");
+           break;
+       }
+
+       printk ("%s: ep4rcvr_disconnect_callback: rcvr %p rxd %p nodeId %d - finished\n", 
+                 rail->r_generic.Name, rcvr, rxd, env->NodeId);
+       EPRINTF4 (DBG_RCVR, "%s: ep4rcvr_disconnect_callback: rcvr %p rxd %p nodeId %d - finished\n", 
+                 rail->r_generic.Name, rcvr, rxd, env->NodeId);
+       ep4rcvr_display_rxd (&di_ep_debug, &rxdRail->rxd_generic);
+    }
+    
+    EP4_SPINEXIT (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    while (! list_empty (&rxdList)) 
+    {
+       EP_RXD *rxd = list_entry (rxdList.next, EP_RXD, Link);
+
+       list_del (&rxd->Link);
+
+       rxd->Handler (rxd);
+    }
+}
+
+void
+ep4rcvr_neterr_flush (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP4_COMMS_RAIL *commsRail = RCVR_TO_COMMS(rcvrRail);
+    EP4_RAIL       *rail      = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV      *dev       = rail->r_ctxt.ctxt_dev;
+    unsigned long   flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    EP4_SPINENTER (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+
+    /* Insert an setevent command into the thread's command queue
+     * to ensure that all sten packets have completed */
+    elan4_guard (rcvrRail->rcvr_ecq->ecq_cq, GUARD_ALL_CHANNELS);
+    ep4comms_flush_setevent (commsRail, rcvrRail->rcvr_ecq->ecq_cq);
+    
+    EP4_SPINEXIT (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+void
+ep4rcvr_neterr_check (EP_RCVR *rcvr, EP4_RCVR_RAIL *rcvrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP4_RAIL         *rail = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV        *dev = rail->r_ctxt.ctxt_dev;
+    struct list_head *el;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    EP4_SPINENTER (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    
+    list_for_each (el, &rcvr->ActiveDescList) {
+       EP_RXD       *rxd      = list_entry (el, EP_RXD, Link);
+       EP4_RXD_RAIL *rxdRail  = (EP4_RXD_RAIL *) rxd->RxdRail;
+       EP_ENVELOPE  *env      = &rxd->RxdMain->Envelope;
+
+       if (rxd->RxdMain->Len == EP_RXD_PENDING || !RXD_BOUND2RAIL(rxdRail,rcvrRail) || env->NodeId != nodeId)
+           continue;
+
+       if (rxd->State == EP_RXD_RECEIVE_ACTIVE || rxd->State == EP_RXD_GET_ACTIVE)
+       {
+           EP_NETERR_COOKIE cookie;
+           unsigned int     first, this;
+
+           if (rxd->State == EP_RXD_RECEIVE_ACTIVE)
+               first = (EP_MAXFRAG+1) - (( EP_IS_MULTICAST(env->Attr) ? 1 : 0) + (env->nFrags == 0 ? 1 : env->nFrags));
+           else
+               first = (EP_MAXFRAG+1) - rxd->nFrags;
+
+           for (this = first; this < (EP_MAXFRAG+1); this++)
+               if (rxdRail->rxd_main->rxd_sent[this] == EP4_STATE_ACTIVE)
+                   break;
+           
+           if (this > first)
+           {
+               /* Look at the last completed STEN packet and if it's neterr cookie matches, then change
+                * the rxd to look the same as if the sten packet had failed and then schedule it for retry */
+               cookie = elan4_sdram_readq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[--this].c_cookie));
+               
+               if (cookie == cookies[0] || cookie == cookies[1])
+               {
+                   EPRINTF5 (DBG_NETWORK_ERROR, "%s: ep4rcvr_neterr_check: cookie <%lld%s%s%s%s> matches rxd %p rxdRail %p this %d\n",
+                             rail->r_generic.Name, EP4_COOKIE_STRING(cookie), rxd, rxdRail, this);
+                   
+                   printk ("%s: ep4rcvr_neterr_check: cookie <%lld%s%s%s%s> matches rxd %p rxdRail %p this %d : time %ld\n",
+                           rail->r_generic.Name, EP4_COOKIE_STRING(cookie), rxd, rxdRail, this, rxdRail->rxd_retry_time);
+                   
+                   rxdRail->rxd_main->rxd_sent[this] = EP4_STATE_ACTIVE;
+                   rxdRail->rxd_main->rxd_failed     = EP4_STATE_FAILED;
+                   
+                   spin_lock (&rcvrRail->rcvr_retrylock);
+                   
+                   ASSERT (rxdRail->rxd_retry_time == 0);
+
+                   rxdRail->rxd_retry_time = lbolt + EP_RETRY_LOW_PRI_TIME;
+                       
+                   list_add_tail (&rxdRail->rxd_retry_link, &rcvrRail->rcvr_retrylist);
+                       
+                   ep_kthread_schedule (&rail->r_retry_thread, rxdRail->rxd_retry_time);
+                   
+                   spin_unlock (&rcvrRail->rcvr_retrylock);
+               }
+           }
+       }
+    }
+    EP4_SPINEXIT (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), &rcvrRail->rcvr_main->rcvr_thread_lock);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+}
+
+int
+ep4rcvr_queue_rxd (EP_RXD *rxd, EP_RCVR_RAIL *r)
+{
+    EP4_RCVR_RAIL *rcvrRail = (EP4_RCVR_RAIL *) r;
+    EP4_RAIL      *rail     = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV     *dev      = rail->r_ctxt.ctxt_dev;
+    EP4_RXD_RAIL  *rxdRail;
+    register int   i;
+
+    ASSERT (SPINLOCK_HELD(&rxd->Rcvr->Lock));
+
+    if ((rxdRail = get_rxd_rail (rcvrRail)) == NULL)
+       return 0;
+    
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->r_generic);
+
+    EPRINTF6 (DBG_RCVR, "%s: ep4rcvr_queue_rxd: rcvr %p rxd %p rxdRail %p buffer %x len %x\n", 
+             rail->r_generic.Name, rxd->Rcvr, rxd, rxdRail, rxd->Data.nmd_addr, rxd->Data.nmd_len);
+
+    /* bind the rxdRail and rxd together */
+    bind_rxd_rail (rxd, rxdRail);
+
+    elan4_sdram_writel (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_buffer.nmd_addr), rxd->Data.nmd_addr);       /* PCI write */
+    elan4_sdram_writel (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_buffer.nmd_len),  rxd->Data.nmd_len);                /* PCI write */
+    elan4_sdram_writel (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_buffer.nmd_attr), rxd->Data.nmd_attr);       /* PCI write */
+
+    /* Mark as active */
+    elan4_sdram_writeq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType), 
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+    
+    for (i = 0; i <= EP_MAXFRAG; i++)
+       rxdRail->rxd_main->rxd_sent[i] = EP4_STATE_ACTIVE;
+
+    rxdRail->rxd_main->rxd_failed = EP4_STATE_ACTIVE;
+    rxdRail->rxd_main->rxd_done = EP4_STATE_ACTIVE;
+
+    elan4_sdram_writeq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[0]) + 0x00, /* %r0 */
+                       ep_symbol (&rail->r_threadcode, "c_queue_rxd"));
+    elan4_sdram_writeq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[0]) + 0x10, /* %r2 */
+                       rcvrRail->rcvr_elan_addr);
+    elan4_sdram_writeq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[0]) + 0x18, /* %r3 */
+                       rxdRail->rxd_elan_addr);
+
+    elan4_sdram_writeq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_CountAndType),
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_START_CMD_NDWORDS));
+
+    ep4_set_event_cmd (rxdRail->rxd_scq, rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_start));
+
+    return 1;
+}
+
+void
+ep4rcvr_rpc_put (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags)
+{
+    EP4_RXD_RAIL    *rxdRail  = (EP4_RXD_RAIL *) rxd->RxdRail;
+    EP4_RCVR_RAIL   *rcvrRail = (EP4_RCVR_RAIL *) rxdRail->rxd_generic.RcvrRail;
+    EP4_RAIL        *rail     = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV      *dev      = RCVR_TO_DEV (rcvrRail);
+    sdramaddr_t     rxdElan   = rxdRail->rxd_elan;
+    EP_ENVELOPE    *env       = &rxd->RxdMain->Envelope;
+    unsigned long   first     = (EP_MAXFRAG+1) - nFrags;
+    EP4_RXD_DMA_CMD cmd;
+    register int    i, len;
+
+    EP4_ASSERT (rail, rxd->State == EP_RXD_PUT_ACTIVE);
+    EP4_ASSERT (rail, rxdRail->rxd_main->rxd_done == EP4_STATE_PRIVATE);
+    EP4_SDRAM_ASSERT (rail, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType),
+                     E4_EVENT_INIT_VALUE (0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->r_generic);
+
+    /* Generate the DMA chain to put the data */
+    for (i = 0, len = 0; i < nFrags; i++, len += local->nmd_len, local++, remote++)
+    {
+       cmd.c_dma_typeSize     = RUN_DMA_CMD | E4_DMA_TYPE_SIZE(local->nmd_len, DMA_DataTypeByte, 0, EP4_DMA_RETRYCOUNT);
+       cmd.c_dma_cookie       = ep4_neterr_cookie (rail, env->NodeId) | EP4_COOKIE_DMA;
+       cmd.c_dma_vproc        = EP_VP_DATA(env->NodeId);
+       cmd.c_dma_srcAddr      = local->nmd_addr;
+       cmd.c_dma_dstAddr      = remote->nmd_addr;
+       if (i == (nFrags-1))
+           cmd.c_dma_srcEvent = rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_done);
+       else
+           cmd.c_dma_srcEvent = rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first + i]);
+       cmd.c_dma_dstEvent     = 0;
+       cmd.c_nop_cmd          = NOP_CMD;
+
+       EPRINTF7 (DBG_RCVR, "%s: ep4rcvr_rpc_put: rxd %p [XID=%llx] idx=%d Source=%08x Dest=%08x Len=%x\n", 
+                 rail->r_generic.Name, rxd, env->Xid.Unique, i, local->nmd_addr, remote->nmd_addr, local->nmd_len);
+       
+       elan4_sdram_copyq_to_sdram (dev, &cmd, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i]), sizeof (EP4_RXD_DMA_CMD));
+    }
+
+    /* Initialise the event chain */
+    for (i = 0; i < nFrags-1; i++)
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first + i]), 
+                           E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_DMA_CMD_NDWORDS));
+
+    elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done),
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+    for (i = 0; i <= EP_MAXFRAG; i++)
+       rxdRail->rxd_main->rxd_sent[i] = EP4_STATE_ACTIVE;
+
+    rxdRail->rxd_main->rxd_failed = EP4_STATE_ACTIVE;
+    rxdRail->rxd_main->rxd_done = EP4_STATE_ACTIVE;
+
+    /* Initialise the previous event to start the whole chain off */
+    elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first - 1]),
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_DMA_CMD_NDWORDS));
+
+    ASSERT (rail->r_generic.Nodes[env->NodeId].State >= EP_NODE_CONNECTED && rail->r_generic.Nodes[env->NodeId].State <= EP_NODE_LOCAL_PASSIVATE);
+
+    /* finally issue the setevent to start the whole chain */
+    ep4_set_event_cmd (rxdRail->rxd_scq, rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first - 1]));
+
+    BucketStat (rxd->Rcvr->Subsys, RPCPut, len);
+}    
+
+void
+ep4rcvr_rpc_get (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags)
+{
+    EP4_RXD_RAIL    *rxdRail  = (EP4_RXD_RAIL *) rxd->RxdRail;
+    EP4_RCVR_RAIL   *rcvrRail = (EP4_RCVR_RAIL *) rxdRail->rxd_generic.RcvrRail;
+    EP4_RAIL        *rail     = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV      *dev      = RCVR_TO_DEV (rcvrRail);
+    sdramaddr_t      rxdElan  = rxdRail->rxd_elan;
+    EP_ENVELOPE     *env      = &rxd->RxdMain->Envelope;
+    unsigned long    first    = (EP_MAXFRAG+1) - nFrags;
+    register int    i, len;
+
+    EP4_ASSERT (rail, rxd->State == EP_RXD_GET_ACTIVE);
+    EP4_ASSERT (rail, rxdRail->rxd_main->rxd_done == EP4_STATE_PRIVATE);
+    EP4_SDRAM_ASSERT (rail, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType),
+                     E4_EVENT_INIT_VALUE (0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->r_generic);
+
+    /* Generate the DMA chain to put the data */
+    for (i = 0, len = 0; i < nFrags; i++, len += local->nmd_len, local++, remote++)
+    {
+       EPRINTF7 (DBG_RCVR, "%s: ep4rcvr_rpc_get rxd %p [XID=%llx] idx=%d Source=%08x Dest=%08x Len=%x\n", 
+                 rail->r_generic.Name, rxd, env->Xid.Unique, i, remote->nmd_addr, local->nmd_addr, remote->nmd_len);
+       
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_open),
+                           OPEN_STEN_PKT_CMD | OPEN_PACKET(0, PACK_OK | RESTART_COUNT_ZERO, EP_VP_DATA(env->NodeId)));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_trans),
+                           SEND_TRANS_CMD | ((TR_REMOTEDMA | TR_WAIT_FOR_EOP) << 16));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_cookie),
+                           ep4_neterr_cookie (rail, env->NodeId) | EP4_COOKIE_STEN);
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_dma_typeSize),
+                           E4_DMA_TYPE_SIZE (local->nmd_len, DMA_DataTypeByte, 0, EP4_DMA_RETRYCOUNT));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_dma_cookie),
+                           ep4_neterr_cookie (rail, env->NodeId) | EP4_COOKIE_DMA);
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_dma_vproc),
+                           EP_VP_DATA (rail->r_generic.Position.pos_nodeid));
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_dma_srcAddr),
+                           remote->nmd_addr);
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_dma_dstAddr),
+                           local->nmd_addr);
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_dma_srcEvent),
+                           0);
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i].c_dma_dstEvent),
+                           i == (nFrags-1) ? rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_done) : 
+                                             rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first + i]));
+    }
+
+    /* Initialise the event chain */
+    for (i = 0; i < nFrags-1; i++)
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first + i]), 
+                           E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_STEN_CMD_NDWORDS));
+
+    elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done),
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+    for (i = 0; i <= EP_MAXFRAG; i++)
+       rxdRail->rxd_main->rxd_sent[i] = EP4_STATE_ACTIVE;
+
+    rxdRail->rxd_main->rxd_failed = EP4_STATE_ACTIVE;
+    rxdRail->rxd_main->rxd_done = EP4_STATE_ACTIVE;
+
+    /* Initialise the previous event to start the whole chain off */
+    elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first - 1]),
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_STEN_CMD_NDWORDS));
+
+    ASSERT (rail->r_generic.Nodes[env->NodeId].State >= EP_NODE_CONNECTED && rail->r_generic.Nodes[env->NodeId].State <= EP_NODE_LOCAL_PASSIVATE);
+
+    /* finally issue the setevent to start the whole chain */
+    ep4_set_event_cmd (rxdRail->rxd_scq, rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first - 1]));
+
+    BucketStat (rxd->Rcvr->Subsys, RPCPut, len);
+}
+
+void
+ep4rcvr_rpc_complete (EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags)
+{
+    EP4_RXD_RAIL    *rxdRail  = (EP4_RXD_RAIL *) rxd->RxdRail;
+    EP4_RCVR_RAIL   *rcvrRail = (EP4_RCVR_RAIL *) rxdRail->rxd_generic.RcvrRail;
+    EP4_RAIL        *rail     = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV      *dev      = RCVR_TO_DEV (rcvrRail);
+    sdramaddr_t     rxdElan   = rxdRail->rxd_elan;
+    EP_ENVELOPE    *env       = &rxd->RxdMain->Envelope;
+    unsigned long   first     = (EP_MAXFRAG+1) - nFrags - 1;
+    EP4_RXD_DMA_CMD cmd;
+    register int    i, len;
+
+    EP4_ASSERT (rail, rxd->State == EP_RXD_COMPLETE_ACTIVE);
+    EP4_ASSERT (rail, rxdRail->rxd_main->rxd_done == EP4_STATE_PRIVATE);
+    EP4_SDRAM_ASSERT (rail, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType),
+                     E4_EVENT_INIT_VALUE (0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->r_generic);
+
+    /* Generate the DMA chain to put the data */
+    for (i = 0, len = 0; i < nFrags; i++, len += local->nmd_len, local++, remote++)
+    {
+       cmd.c_dma_typeSize = RUN_DMA_CMD | E4_DMA_TYPE_SIZE(local->nmd_len, DMA_DataTypeByte, 0, EP4_DMA_RETRYCOUNT);
+       cmd.c_dma_cookie   = ep4_neterr_cookie (rail, env->NodeId) | EP4_COOKIE_DMA;
+       cmd.c_dma_vproc    = EP_VP_DATA(env->NodeId);
+       cmd.c_dma_srcAddr  = local->nmd_addr;
+       cmd.c_dma_dstAddr  = remote->nmd_addr;
+       cmd.c_dma_srcEvent = rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first + i]);
+       cmd.c_dma_dstEvent = 0;
+       cmd.c_nop_cmd      = NOP_CMD;
+
+       EPRINTF7 (DBG_RCVR, "%s: ep4rcvr_rpc_complete: rxd %p [XID=%llx] idx=%d Source=%08x Dest=%08x Len=%x\n", 
+                 rail->r_generic.Name, rxd, env->Xid.Unique, i, local->nmd_addr, remote->nmd_addr, local->nmd_len);
+
+       elan4_sdram_copyq_to_sdram (dev, &cmd, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[first + i]), sizeof (EP4_RXD_DMA_CMD));
+    }
+    
+    /* Initialise the status block dma */
+    cmd.c_dma_typeSize = RUN_DMA_CMD | E4_DMA_TYPE_SIZE(EP_STATUSBLK_SIZE, DMA_DataTypeByte, 0, EP4_DMA_RETRYCOUNT);
+    cmd.c_dma_cookie   = ep4_neterr_cookie (rail, env->NodeId) | EP4_COOKIE_DMA;
+    cmd.c_dma_vproc    = EP_VP_DATA(env->NodeId);
+    cmd.c_dma_srcAddr  = rxd->NmdMain.nmd_addr + offsetof (EP_RXD_MAIN, StatusBlk);
+    cmd.c_dma_dstAddr  = env->TxdMain.nmd_addr + offsetof (EP_TXD_MAIN, StatusBlk);
+    cmd.c_dma_srcEvent = rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_done);
+    cmd.c_dma_dstEvent = env->TxdRail + offsetof (EP4_TXD_RAIL_ELAN, txd_done);;
+    cmd.c_nop_cmd      = NOP_CMD;
+
+    EPRINTF6 (DBG_RCVR, "%s: ep4rcvr_rpc_complete: rxd %p [XID=%llx] statusblk source=%08x dest=%08x len=%x\n", 
+             rail->r_generic.Name, rxd, env->Xid.Unique, (int) cmd.c_dma_srcAddr, (int) cmd.c_dma_dstAddr, EP_STATUSBLK_SIZE);
+
+    elan4_sdram_copyq_to_sdram (dev, &cmd, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[EP_MAXFRAG]), sizeof (EP4_RXD_DMA_CMD));
+
+    /* Initialise the event chain */
+    for (i = 0; i < nFrags; i++)
+       elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first + i]), 
+                           E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_DMA_CMD_NDWORDS));
+
+    elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done),
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+    for (i = 0; i <= EP_MAXFRAG; i++)
+       rxdRail->rxd_main->rxd_sent[i] = EP4_STATE_ACTIVE;
+
+    rxdRail->rxd_main->rxd_failed = EP4_STATE_ACTIVE;
+    rxdRail->rxd_main->rxd_done = EP4_STATE_ACTIVE;
+
+    /* Initialise the previous event to start the whole chain off */
+    elan4_sdram_writeq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first - 1]),
+                       E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_DMA_CMD_NDWORDS));
+
+    ASSERT (rail->r_generic.Nodes[env->NodeId].State >= EP_NODE_CONNECTED && rail->r_generic.Nodes[env->NodeId].State <= EP_NODE_LOCAL_PASSIVATE);
+
+    /* finally issue the setevent to start the whole chain */
+    ep4_set_event_cmd (rxdRail->rxd_scq, rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first - 1]));
+
+    BucketStat (rxd->Rcvr->Subsys, CompleteRPC, len);
+}
+
+EP_RXD *
+ep4rcvr_steal_rxd (EP_RCVR_RAIL *r)
+{
+    /* XXXX - TBD */
+    return NULL;
+}
+
+long
+ep4rcvr_check (EP_RCVR_RAIL *r, long nextRunTime)
+{
+    EP4_RCVR_RAIL *rcvrRail = (EP4_RCVR_RAIL *) r;
+    EP4_RAIL      *rail     = RCVR_TO_RAIL (rcvrRail);
+
+    if (rcvrRail->rcvr_freecount < ep_rxd_lowat && !alloc_rxd_block (rcvrRail))
+    {
+       EPRINTF1 (DBG_RCVR,"%s: failed to grow rxd rail pool\n", rail->r_generic.Name);
+               
+       if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + RESOURCE_RETRY_TIME))
+           nextRunTime = lbolt + RESOURCE_RETRY_TIME;
+    }
+    
+    return nextRunTime;
+}
+
+unsigned long
+ep4rcvr_retry (EP4_RAIL *rail, void *arg, unsigned long nextRunTime)
+{
+    EP4_RCVR_RAIL *rcvrRail = (EP4_RCVR_RAIL *) arg;
+    ELAN4_DEV     *dev      = RCVR_TO_DEV(rcvrRail);
+    unsigned long  flags;
+
+    spin_lock_irqsave (&rcvrRail->rcvr_retrylock, flags);
+    while (! list_empty (&rcvrRail->rcvr_retrylist))
+    {
+       EP4_RXD_RAIL *rxdRail = list_entry (rcvrRail->rcvr_retrylist.next, EP4_RXD_RAIL, rxd_retry_link);
+       EP_ENVELOPE  *env     = &rxdRail->rxd_generic.Rxd->RxdMain->Envelope;
+       unsigned int  first   = (EP_MAXFRAG+1) - ((env->Attr & EP_MULTICAST ? 1 : 0) + (env->nFrags == 0 ? 1 : env->nFrags));
+       
+       if (BEFORE (lbolt, rxdRail->rxd_retry_time))
+       {
+           if (nextRunTime == 0 || AFTER (nextRunTime, rxdRail->rxd_retry_time))
+               nextRunTime = rxdRail->rxd_retry_time;
+
+           break;
+       }
+
+       list_del (&rxdRail->rxd_retry_link);
+       rxdRail->rxd_retry_time = 0;
+
+       /* determine which sten packet to resubmit */
+       for (; first < (EP_MAXFRAG+1); first++)
+           if (rxdRail->rxd_main->rxd_sent[first] == EP4_STATE_ACTIVE)
+               break;
+
+       EPRINTF3 (DBG_RETRY, "%s: ep4rcvr_retry: rxdRail %p, reissuing sten[%d]\n", rail->r_generic.Name, rxdRail, first);
+
+       /* re-initialise the fail event */
+       elan4_sdram_writeq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+       rxdRail->rxd_main->rxd_failed = EP4_STATE_ACTIVE;
+
+       /* re-initialise the chain event to resubmit this sten packet */
+       elan4_sdram_writeq (dev, rxdRail->rxd_elan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first-1].ev_CountAndType),
+                           E4_EVENT_INIT_VALUE(-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_RXD_STEN_CMD_NDWORDS));
+       
+       /* finally issue the setevent to start the chain again */
+       ep4_set_event_cmd (rxdRail->rxd_scq, rxdRail->rxd_elan_addr + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[first - 1]));
+    }
+    spin_unlock_irqrestore (&rcvrRail->rcvr_retrylock, flags);
+    
+    return nextRunTime;
+}
+
+void
+ep4rcvr_add_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *commsRail)
+{
+    EP4_RAIL          *rail   = (EP4_RAIL *) commsRail->Rail;
+    ELAN4_DEV         *dev    = rail->r_ctxt.ctxt_dev;
+    sdramaddr_t        qdescs = ((EP4_COMMS_RAIL *) commsRail)->r_descs;
+    EP4_RCVR_RAIL     *rcvrRail;
+    E4_InputQueue      qdesc;
+    E4_ThreadRegs      tregs;
+    sdramaddr_t        stack;
+    unsigned long      flags;
+
+    KMEM_ZALLOC (rcvrRail, EP4_RCVR_RAIL *, sizeof (EP4_RCVR_RAIL), 1);
+
+    spin_lock_init (&rcvrRail->rcvr_freelock);
+    INIT_LIST_HEAD (&rcvrRail->rcvr_freelist);
+    INIT_LIST_HEAD (&rcvrRail->rcvr_blocklist);
+
+    kcondvar_init (&rcvrRail->rcvr_cleanup_sleep);
+    kcondvar_init (&rcvrRail->rcvr_freesleep);
+
+    INIT_LIST_HEAD (&rcvrRail->rcvr_retrylist);
+    spin_lock_init (&rcvrRail->rcvr_retrylock);
+
+    rcvrRail->rcvr_generic.CommsRail = commsRail;
+    rcvrRail->rcvr_generic.Rcvr      = rcvr;
+
+    rcvrRail->rcvr_main  = ep_alloc_main (&rail->r_generic, sizeof (EP4_RCVR_RAIL_MAIN), 0, &rcvrRail->rcvr_main_addr);
+    rcvrRail->rcvr_elan  = ep_alloc_elan (&rail->r_generic, sizeof (EP4_RCVR_RAIL_ELAN), 0, &rcvrRail->rcvr_elan_addr);
+    rcvrRail->rcvr_slots = ep_alloc_elan (&rail->r_generic, EP_INPUTQ_SIZE * rcvr->InputQueueEntries, 0, &rcvrRail->rcvr_slots_addr);
+    stack                = ep_alloc_elan (&rail->r_generic, EP4_STACK_SIZE, 0, &rcvrRail->rcvr_stack);
+
+    /* allocate a command queue for the thread to use, plus space for it to wait/reschedule */
+    rcvrRail->rcvr_ecq     = ep4_alloc_ecq (rail, CQ_Size64K);
+    rcvrRail->rcvr_resched = ep4_get_ecq (rail, EP4_ECQ_ATOMIC, 8);
+
+    ep4_register_intcookie (rail, &rcvrRail->rcvr_stall_intcookie, rcvrRail->rcvr_elan_addr + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_stall_intcookie),
+                           rcvr_stall_interrupt, rcvrRail);
+
+    /* Initialise the elan portion */
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_qevent.ev_CountAndType), 0);
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_halt.ev_CountAndType), 0);
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock), 0);
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_pending_tailp),
+                       rcvrRail->rcvr_elan_addr + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_pending_head));
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_pending_head), 0);
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_stall_intcookie), 0);
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_qbase), rcvrRail->rcvr_slots_addr);
+    elan4_sdram_writeq (dev, rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_qlast), 
+                       rcvrRail->rcvr_slots_addr + EP_INPUTQ_SIZE * (rcvr->InputQueueEntries-1));
+
+    /* Initialise the main memory portion */
+    rcvrRail->rcvr_main->rcvr_thread_lock = 0;
+
+    /* Install our retry handler */
+    rcvrRail->rcvr_retryops.op_func = ep4rcvr_retry;
+    rcvrRail->rcvr_retryops.op_arg  = rcvrRail;
+
+    ep4_add_retry_ops (rail, &rcvrRail->rcvr_retryops);
+
+    /* Update the queue desriptor */
+    qdesc.q_bptr    = rcvrRail->rcvr_slots_addr;
+    qdesc.q_fptr    = rcvrRail->rcvr_slots_addr;
+    qdesc.q_control = E4_InputQueueControl (rcvrRail->rcvr_slots_addr, rcvrRail->rcvr_slots_addr + (EP_INPUTQ_SIZE * (rcvr->InputQueueEntries-1)), EP_INPUTQ_SIZE);
+    qdesc.q_event   = rcvrRail->rcvr_elan_addr + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_qevent);
+
+    ep4_write_qdesc (rail, qdescs + (rcvr->Service * EP_QUEUE_DESC_SIZE), &qdesc);
+
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    rcvr->Rails[rail->r_generic.Number] = &rcvrRail->rcvr_generic;
+    rcvr->RailMask |= EP_RAIL2RAILMASK (rail->r_generic.Number);
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    {
+       sdramaddr_t stackTop     = stack + EP4_STACK_SIZE;
+       E4_Addr     stackTopAddr = rcvrRail->rcvr_stack + EP4_STACK_SIZE;
+
+       ep4_init_thread (rail, &tregs, stackTop, stackTopAddr, ep_symbol (&rail->r_threadcode, "ep4comms_rcvr"), 6, 
+                        (E4_uint64) rail->r_elan_addr, (E4_uint64) rcvrRail->rcvr_elan_addr, (E4_uint64) rcvrRail->rcvr_main_addr,
+                        (E4_uint64) EP_MSGQ_ADDR(rcvr->Service), (E4_uint64) rcvrRail->rcvr_ecq->ecq_addr, (E4_uint64) rcvrRail->rcvr_resched->ecq_addr);
+    }
+    
+    /* Issue the command to the threads private command queue */
+    elan4_run_thread_cmd (rcvrRail->rcvr_ecq->ecq_cq, &tregs);
+
+    ep_procfs_rcvr_add_rail(&(rcvrRail->rcvr_generic));
+}
+
+void
+ep4rcvr_del_rail (EP_RCVR *rcvr, EP_COMMS_RAIL *commsRail)
+{
+    EP4_RAIL         *rail     = (EP4_RAIL *) commsRail->Rail;
+    EP4_RCVR_RAIL    *rcvrRail = (EP4_RCVR_RAIL *) rcvr->Rails[rail->r_generic.Number];  
+    ELAN4_HALTOP      haltop;
+    struct list_head *el, *nel;
+    unsigned long     flags;
+
+    ep_procfs_rcvr_del_rail(&(rcvrRail->rcvr_generic));
+
+    /* Run a halt operation to mark the input queue as full and
+     * request the thread to halt */
+    haltop.op_mask     = INT_DiscardingHighPri | INT_TProcHalted;
+    haltop.op_function = rcvr_stall_haltop;
+    haltop.op_arg      = rcvrRail;
+
+    elan4_queue_haltop (rail->r_ctxt.ctxt_dev, &haltop);
+
+    /* Wait for the thread to tell us it's processed the input queue */
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    while (! rcvrRail->rcvr_thread_halted)
+       kcondvar_wait (&rcvrRail->rcvr_cleanup_sleep, &rcvr->Lock, &flags);
+    rcvrRail->rcvr_thread_halted = 0;
+
+    /* flag the rail as no longer available */
+    rcvr->RailMask &= ~EP_RAIL2RAILMASK (rail->r_generic.Number);
+
+    /* wait for all active communications to terminate */
+    for (;;)
+    {
+       int mustWait = 0;
+
+       list_for_each (el, &rcvr->ActiveDescList) {
+           EP_RXD       *rxd     = list_entry (el, EP_RXD, Link);
+           EP4_RXD_RAIL *rxdRail = (EP4_RXD_RAIL *) rxd->RxdRail;
+           
+           if (rxdRail && RXD_BOUND2RAIL (rxdRail, rcvrRail) && rxd->RxdMain->Len != EP_RXD_PENDING)
+           {
+               mustWait++;
+               break;
+           }
+       }
+
+       if (! mustWait)
+           break;
+
+       rcvrRail->rcvr_cleanup_waiting++;
+       kcondvar_wait (&rcvrRail->rcvr_cleanup_sleep, &rcvr->Lock, &flags);
+    }
+
+    /* at this point all rxd's in the list that are bound to the deleting rail are pending */
+    list_for_each_safe (el, nel, &rcvr->ActiveDescList) {
+       EP_RXD       *rxd     = list_entry (el, EP_RXD, Link);
+       EP4_RXD_RAIL *rxdRail = (EP4_RXD_RAIL *) rxd->RxdRail;
+
+       if (rxdRail && RXD_BOUND2RAIL (rxdRail, rcvrRail))
+       {
+           EP4_RXD_ASSERT_PENDING (rxdRail);
+           EP4_RXD_FORCE_PRIVATE (rxdRail);
+
+           unbind_rxd_rail (rxd, rxdRail);
+           free_rxd_rail (rcvrRail, rxdRail);
+       }
+    }
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    /* wait for all rxd's for this rail to become free */
+    spin_lock_irqsave (&rcvrRail->rcvr_freelock, flags);
+    while (rcvrRail->rcvr_freecount != rcvrRail->rcvr_totalcount)
+    {
+       rcvrRail->rcvr_freewaiting++;
+       kcondvar_wait (&rcvrRail->rcvr_freesleep, &rcvrRail->rcvr_freelock, &flags);
+    }
+    spin_unlock_irqrestore (&rcvrRail->rcvr_freelock, flags);
+
+    /* can now remove the rail as it can no longer be used */
+    spin_lock_irqsave (&rcvr->Lock, flags);
+    rcvr->Rails[rail->r_generic.Number] = NULL;
+    spin_unlock_irqrestore (&rcvr->Lock, flags);
+
+    /* all the rxd's accociated with DescBlocks must be in the FreeDescList */
+    ASSERT (rcvrRail->rcvr_totalcount == rcvrRail->rcvr_freecount);
+
+    /* run through the DescBlockList deleting them */
+    while (!list_empty (&rcvrRail->rcvr_blocklist))
+       free_rxd_block (rcvrRail, list_entry(rcvrRail->rcvr_blocklist.next, EP4_RXD_RAIL_BLOCK , blk_link));
+
+    /* it had better be empty after that */
+    ASSERT ((rcvrRail->rcvr_totalcount == 0) && (rcvrRail->rcvr_totalcount == rcvrRail->rcvr_freecount));
+
+    ep4_remove_retry_ops (rail, &rcvrRail->rcvr_retryops);
+
+    ep4_deregister_intcookie (rail, &rcvrRail->rcvr_stall_intcookie);
+
+    ep4_put_ecq (rail, rcvrRail->rcvr_resched, 8);
+    ep4_free_ecq (rail, rcvrRail->rcvr_ecq);
+
+    ep_free_elan (&rail->r_generic, rcvrRail->rcvr_stack, EP4_STACK_SIZE);
+    ep_free_elan (&rail->r_generic, rcvrRail->rcvr_slots_addr, EP_INPUTQ_SIZE * rcvr->InputQueueEntries);
+    ep_free_elan (&rail->r_generic, rcvrRail->rcvr_elan_addr, sizeof (EP4_RCVR_RAIL_ELAN));
+    ep_free_main (&rail->r_generic, rcvrRail->rcvr_main_addr, sizeof (EP4_RCVR_RAIL_MAIN));
+
+    KMEM_FREE (rcvrRail, sizeof (EP4_RCVR_RAIL));
+}
+
+void
+ep4rcvr_display_rxd (DisplayInfo *di, EP_RXD_RAIL *r)
+{
+    EP4_RXD_RAIL *rxdRail = (EP4_RXD_RAIL *) r;
+    sdramaddr_t   rxdElan = rxdRail->rxd_elan;
+    EP4_RAIL     *rail    = RCVR_TO_RAIL (rxdRail->rxd_generic.RcvrRail);
+    ELAN4_DEV    *dev     = rail->r_ctxt.ctxt_dev;
+    int i;
+
+    (di->func)(di->arg, "    Rail %d rxd %p elan %lx(%x) main %p(%x) ecq %d scq %d debug %llx\n", rail->r_generic.Number,
+              rxdRail, rxdRail->rxd_elan, rxdRail->rxd_elan_addr, rxdRail->rxd_main, rxdRail->rxd_main_addr,
+              elan4_cq2num(rxdRail->rxd_ecq->ecq_cq), elan4_cq2num(rxdRail->rxd_scq->ecq_cq),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_debug)));
+    (di->func)(di->arg, "          start    %016llx %016llx %016llx [%016llx %016llx]\n",
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_CountAndType)),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_Params[0])),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_start.ev_Params[1])),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[0].c_cookie)),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[0].c_dma_cookie)));
+              
+    for (i = 0; i < EP_MAXFRAG; i++)
+       (di->func)(di->arg, "          chain[%d] %016llx %016llx %016llx [%016llx %016llx]\n", i,
+                  elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[i].ev_CountAndType)),
+                  elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[i].ev_Params[0])),
+                  elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_chain[i].ev_Params[1])),
+                  elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[i+1].c_cookie)),
+                  elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_sten[i+1].c_dma_cookie)));
+    (di->func)(di->arg, "          done    %016llx %016llx %016llx -> %016llx\n",
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_CountAndType)),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_Params[0])),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_done.ev_Params[1])),
+              rxdRail->rxd_main->rxd_done);
+    (di->func)(di->arg, "          fail    %016llx %016llx %016llx -> %016llx\n",
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_CountAndType)),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_Params[0])),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_failed.ev_Params[1])),
+              rxdRail->rxd_main->rxd_failed);
+    (di->func)(di->arg, "          next %016llx queued %016llx main %016llx\n",
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_next)),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_queued)),
+              elan4_sdram_readq (dev, rxdElan + offsetof (EP4_RXD_RAIL_ELAN, rxd_main)));
+    (di->func)(di->arg, "          sent %016llx %016llx %016llx %016llx %016llx\n",
+              rxdRail->rxd_main->rxd_sent[0], rxdRail->rxd_main->rxd_sent[1], rxdRail->rxd_main->rxd_sent[2],
+              rxdRail->rxd_main->rxd_sent[3], rxdRail->rxd_main->rxd_sent[4]);
+}
+
+void
+ep4rcvr_display_rcvr (DisplayInfo *di, EP_RCVR_RAIL *r)
+{
+    EP_RCVR          *rcvr       = r->Rcvr;
+    EP4_RCVR_RAIL    *rcvrRail   = (EP4_RCVR_RAIL *) r;
+    EP4_COMMS_RAIL   *commsRail  = RCVR_TO_COMMS(rcvrRail);
+    EP4_RAIL         *rail       = RCVR_TO_RAIL (rcvrRail);
+    ELAN4_DEV        *dev        = rail->r_ctxt.ctxt_dev;
+    sdramaddr_t       rcvrElan   = rcvrRail->rcvr_elan;
+    sdramaddr_t       qdesc      = commsRail->r_descs + (rcvr->Service * EP_QUEUE_DESC_SIZE);
+    sdramaddr_t       event      = rcvrRail->rcvr_elan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_qevent);
+    unsigned int      freeCount  = 0;
+    unsigned int      blockCount = 0;
+    struct list_head *el;
+    unsigned long     flags;
+    
+    spin_lock_irqsave (&rcvrRail->rcvr_freelock, flags);
+    list_for_each (el, &rcvrRail->rcvr_freelist)
+       freeCount++;
+    list_for_each (el, &rcvrRail->rcvr_blocklist)
+       blockCount++;
+    spin_unlock_irqrestore(&rcvrRail->rcvr_freelock, flags);
+
+    (di->func)(di->arg, "      Rail %d elan %lx(%x) main %p(%x) ecq %d resched %d debug %llx\n",
+              rail->r_generic.Number, rcvrRail->rcvr_elan, rcvrRail->rcvr_elan_addr,
+              rcvrRail->rcvr_main, rcvrRail->rcvr_main_addr, elan4_cq2num(rcvrRail->rcvr_ecq->ecq_cq),
+              elan4_cq2num (rcvrRail->rcvr_resched->ecq_cq),
+              elan4_sdram_readq (dev, rcvrElan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_debug)));
+    (di->func)(di->arg, "        free %d (%d) total %d blocks %d\n",
+              rcvrRail->rcvr_freecount, freeCount, rcvrRail->rcvr_totalcount, blockCount);
+    (di->func)(di->arg, "        spinlock %016llx %016llx\n", rcvrRail->rcvr_main->rcvr_thread_lock,
+              elan4_sdram_readq (dev, rcvrElan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_thread_lock)));
+    (di->func)(di->arg, "        queue: bptr %016llx fptr %016llx control %016llx (base %lx %x)\n",
+              elan4_sdram_readq (dev, qdesc + offsetof (E4_InputQueue, q_bptr)),
+              elan4_sdram_readq (dev, qdesc + offsetof (E4_InputQueue, q_fptr)),
+              elan4_sdram_readq (dev, qdesc + offsetof (E4_InputQueue, q_control)),
+              rcvrRail->rcvr_slots, rcvrRail->rcvr_slots_addr);
+    (di->func)(di->arg, "        event %016llx %016llx %016llx\n",
+              elan4_sdram_readq (dev, event + offsetof (E4_Event32, ev_CountAndType)),
+              elan4_sdram_readq (dev, event + offsetof (E4_Event32, ev_Params[0])),
+              elan4_sdram_readq (dev, event + offsetof (E4_Event32, ev_Params[1])));
+    (di->func)(di->arg, "        pending_tailp %016llx pending_head %016llx\n", 
+              elan4_sdram_readq (dev, rcvrElan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_pending_tailp)),
+              elan4_sdram_readq (dev, rcvrElan + offsetof (EP4_RCVR_RAIL_ELAN, rcvr_pending_head)));
+}
+
+void
+ep4rcvr_fillout_rail_stats(EP_RCVR_RAIL *rcvr_rail, char *str) {
+    /* no stats here yet */
+    /* EP4_RCVR_RAIL * ep4rcvr_rail = (EP4_RCVR_RAIL *) rcvr_rail; */
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcommsTx.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcommsTx.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcommsTx.c       2005-06-01 23:12:54.654430744 -0400
@@ -0,0 +1,919 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcommsTx.c,v 1.25.2.5 2004/12/09 10:02:42 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/epcommsTx.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "cm.h"
+#include "debug.h"
+
+unsigned int ep_txd_lowat = 5;
+
+static int
+AllocateTxdBlock (EP_XMTR *xmtr, EP_ATTRIBUTE attr, EP_TXD **txdp)
+{
+    EP_TXD_BLOCK *blk;
+    EP_TXD       *txd;
+    EP_TXD_MAIN  *pTxdMain;
+    int                  i;
+    unsigned long flags;
+
+    EPRINTF1 (DBG_XMTR, "AllocateTxdBlock: xmtr=%p\n", xmtr);
+
+    KMEM_ZALLOC (blk, EP_TXD_BLOCK *, sizeof (EP_TXD_BLOCK), ! (attr & EP_NO_SLEEP));
+
+    if (blk == NULL)
+       return -ENOMEM;
+
+    if ((pTxdMain = ep_shared_alloc_main (xmtr->Subsys->Subsys.Sys, EP_TXD_MAIN_SIZE * EP_NUM_TXD_PER_BLOCK, attr, &blk->NmdMain)) == (sdramaddr_t) 0)
+    {
+       KMEM_FREE (blk, sizeof (EP_TXD_BLOCK));
+       return -ENOMEM;
+    }
+
+    for (txd = &blk->Txd[0], i = 0; i < EP_NUM_TXD_PER_BLOCK; i++, txd++)
+    {
+       txd->Xmtr     = xmtr;
+       txd->TxdMain = pTxdMain;
+
+       ep_nmd_subset (&txd->NmdMain, &blk->NmdMain, (i * EP_TXD_MAIN_SIZE), EP_TXD_MAIN_SIZE);
+
+       /* move onto next descriptor */
+       pTxdMain = (EP_TXD_MAIN *) ((unsigned long) pTxdMain + EP_TXD_MAIN_SIZE);
+    }
+
+    spin_lock_irqsave (&xmtr->FreeDescLock, flags);
+
+    list_add  (&blk->Link, &xmtr->DescBlockList);
+    xmtr->TotalDescCount += EP_NUM_TXD_PER_BLOCK;
+
+    for (i = txdp ? 1 : 0; i < EP_NUM_TXD_PER_BLOCK; i++)
+    {
+       list_add (&blk->Txd[i].Link, &xmtr->FreeDescList);
+
+       xmtr->FreeDescCount++;
+
+       if (xmtr->FreeDescWanted)
+       {
+           xmtr->FreeDescWanted--;
+           kcondvar_wakeupone (&xmtr->FreeDescSleep, &xmtr->FreeDescLock);
+       }
+    }
+    spin_unlock_irqrestore (&xmtr->FreeDescLock, flags);
+    
+    if (txdp)
+       *txdp = &blk->Txd[0];
+
+    return 0;
+}
+
+static void
+FreeTxdBlock (EP_XMTR *xmtr, EP_TXD_BLOCK *blk)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&xmtr->FreeDescLock, flags);
+    list_del (&blk->Link);
+
+    xmtr->TotalDescCount -= EP_NUM_RXD_PER_BLOCK;
+    xmtr->FreeDescCount -= EP_NUM_RXD_PER_BLOCK;
+    spin_unlock_irqrestore (&xmtr->FreeDescLock, flags);
+
+    ep_shared_free_main (xmtr->Subsys->Subsys.Sys, &blk->NmdMain);
+    KMEM_FREE (blk, sizeof (EP_TXD_BLOCK));
+}
+
+static EP_TXD *
+GetTxd (EP_XMTR *xmtr, EP_ATTRIBUTE attr)
+{
+    EP_COMMS_SUBSYS *subsys = xmtr->Subsys;
+    EP_TXD          *txd;
+    int low_on_txds;
+    unsigned long flags;
+
+    spin_lock_irqsave (&xmtr->FreeDescLock, flags);
+
+    while (list_empty (&xmtr->FreeDescList))
+    {
+       if (! (attr & EP_NO_ALLOC))
+       {
+           spin_unlock_irqrestore (&xmtr->FreeDescLock, flags);
+
+           if (AllocateTxdBlock (xmtr, attr, &txd) == ESUCCESS)
+               return (txd);
+
+           spin_lock_irqsave (&xmtr->FreeDescLock, flags);
+       }
+
+       if (attr & EP_NO_SLEEP)
+       {
+           spin_unlock_irqrestore (&xmtr->FreeDescLock, flags);
+
+           return (NULL);
+       }
+
+       xmtr->FreeDescWanted++;
+       kcondvar_wait (&xmtr->FreeDescSleep, &xmtr->FreeDescLock, &flags);
+    }
+
+    txd = list_entry (xmtr->FreeDescList.next, EP_TXD, Link);
+
+    list_del (&txd->Link);
+
+    /* Wakeup the descriptor primer thread if there's not many left */
+    low_on_txds = (--xmtr->FreeDescCount < ep_txd_lowat);
+
+    spin_unlock_irqrestore (&xmtr->FreeDescLock, flags);
+
+    if (low_on_txds)
+       ep_kthread_schedule (&subsys->Thread, lbolt);
+
+    return (txd);
+}
+
+void
+FreeTxd (EP_XMTR *xmtr, EP_TXD *txd)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&xmtr->FreeDescLock, flags);
+    
+    list_add (&txd->Link, &xmtr->FreeDescList);
+
+    xmtr->FreeDescCount++;
+
+    if (xmtr->FreeDescWanted)                                  /* someone waiting for a receive */
+    {                                                          /* descriptor, so wake them up */
+       xmtr->FreeDescWanted--;
+       kcondvar_wakeupone (&xmtr->FreeDescSleep, &xmtr->FreeDescLock);
+    }
+    
+    spin_unlock_irqrestore (&xmtr->FreeDescLock, flags);
+}
+
+int
+TxdShouldStabalise (EP_TXD_RAIL *txdRail, EP_RAIL *rail)
+{
+    EP_TXD      *txd  = txdRail->Txd;
+    EP_XMTR     *xmtr = txd->Xmtr;
+    EP_ATTRIBUTE attr = txd->Envelope.Attr;
+    int                 stabilise;
+    extern int   txd_stabilise;
+
+    switch (EP_ATTR2TYPE (attr)) 
+    {
+    case EP_TYPE_SVC_INDICATOR:                                /* is the rail in the current service indicator rail mask */
+       if ((txd_stabilise & 4) == 0)
+           return 0;
+
+       stabilise = (ep_xmtr_svc_indicator_railmask (xmtr, EP_ATTR2DATA (attr), txd->NodeId) & EP_RAIL2RAILMASK (rail->Number)) == 0;
+       break;
+
+    case EP_TYPE_TIMEOUT:
+       if ((txd_stabilise & 2) == 0)
+           return 0;
+
+       stabilise = AFTER(lbolt, txdRail->Txd->TimeStamp + EP_ATTR2DATA(attr));
+       break;
+
+    default:
+       if ((txd_stabilise & 1) == 0)
+           return 0;
+
+       stabilise = AFTER(lbolt, txdRail->Txd->TimeStamp + EP_DEFAULT_TIMEOUT);
+       break;
+    }
+
+    if (stabilise)
+    {
+       txd->Envelope.Attr = EP_SET_TXD_STABALISING(txd->Envelope.Attr);
+       txd->RetryTime     = lbolt;
+
+       ep_kthread_schedule (&xmtr->Subsys->Thread, lbolt);    
+    }
+
+    return stabilise;
+}
+
+void ep_xmtr_txd_stat(EP_XMTR *xmtr, EP_TXD *txd) 
+{
+    int f;
+    unsigned long size;
+    EP_TXD_RAIL *txdRail = txd->TxdRail;
+
+    size = 0;
+    for (f=0; f < txd->Envelope.nFrags; f++)
+       size += txd->Envelope.Frags[f].nmd_len;
+
+    INC_STAT(xmtr->stats,tx);
+    ADD_STAT(xmtr->stats,tx_len, size);  
+    
+    if ((txdRail != NULL) && (txdRail->XmtrRail != NULL)){
+       INC_STAT(txdRail->XmtrRail->stats,tx);
+       ADD_STAT(txdRail->XmtrRail->stats,tx_len, size); 
+       
+       if ((txdRail->XmtrRail->CommsRail != NULL) && ( txdRail->XmtrRail->CommsRail->Rail != NULL)) {
+           INC_STAT(txdRail->XmtrRail->CommsRail->Rail->Stats,tx);
+           ADD_STAT(txdRail->XmtrRail->CommsRail->Rail->Stats,tx_len, size);
+       }
+    }
+}
+
+static int
+PollActiveTransmitList (EP_XMTR *xmtr, int flag)
+{
+    struct list_head *el, *nel;
+    struct list_head list;
+    unsigned long flags;
+    int count;
+
+    INIT_LIST_HEAD (&list);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    list_for_each_safe (el, nel, &xmtr->ActiveDescList) {
+       EP_TXD      *txd     = list_entry (el, EP_TXD, Link);
+       EP_TXD_RAIL *txdRail = txd->TxdRail;
+       
+       if (txdRail == NULL)
+           continue;
+
+       ASSERT (txdRail->Txd == txd);
+       
+       if (EP_XMTR_OP (txdRail->XmtrRail,PollTxd) (txdRail->XmtrRail, txdRail, flags))
+       {
+           list_del (&txd->Link);                              /* remove from active transmit list */
+           list_add_tail (&txd->Link, &list);                  /* and add to list to call handlers */
+       }
+    }
+    
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    for (count = 0; !list_empty (&list); count++)
+    {
+       EP_TXD *txd = list_entry (list.next, EP_TXD, Link);
+
+       list_del (&txd->Link);
+
+       txd->Handler (txd, txd->Arg, EP_SUCCESS);
+
+       FreeTxd (xmtr, txd);
+    }
+    return (count);
+}
+
+static inline void
+DoTransmit (EP_XMTR *xmtr, EP_TXD *txd)
+{
+    EP_RAILMASK   nmdRailMask = ep_nmd2railmask (txd->Envelope.Frags, txd->Envelope.nFrags);
+    EP_XMTR_RAIL *xmtrRail;
+    unsigned long flags;
+    int rnum;
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+
+    if (EP_IS_SVC_INDICATOR(txd->Envelope.Attr))
+       nmdRailMask = nmdRailMask & ep_xmtr_svc_indicator_railmask(xmtr, EP_ATTR2DATA(txd->Envelope.Attr), txd->NodeId);
+
+    if (EP_IS_PREFRAIL_SET(txd->Envelope.Attr))
+       rnum = EP_ATTR2PREFRAIL(txd->Envelope.Attr);
+    else 
+       rnum = ep_xmtr_prefrail (xmtr, nmdRailMask, txd->NodeId);
+    
+    if (rnum < 0 || !(nmdRailMask & EP_RAIL2RAILMASK(rnum)))
+       xmtrRail = NULL;
+    else
+       xmtrRail = xmtr->Rails[rnum];
+    
+    /* Allocate the XID while holding the xmtr->Lock from our XID cache */
+    txd->Envelope.Xid = ep_xid_cache_alloc (xmtr->Subsys->Subsys.Sys, &xmtr->XidCache);
+    
+    EPRINTF7 (DBG_XMTR, "ep: transmit txd %p to %d/%d: Xid %llx nFrags %d [%08x.%d]\n",
+             txd, txd->NodeId, txd->Service, (long long) txd->Envelope.Xid.Unique, 
+             txd->Envelope.nFrags, txd->Envelope.Frags[0].nmd_addr, txd->Envelope.Frags[0].nmd_len);
+
+    /* Store time transmit started to timeout if not received */
+    txd->TimeStamp = lbolt;
+    
+    /* Initialise the retry backoff */
+    txd->Backoff.type = EP_BACKOFF_FREE;
+
+    list_add_tail (&txd->Link, &xmtr->ActiveDescList);
+
+    if (xmtrRail == NULL || !EP_XMTR_OP(xmtrRail,BindTxd) (txd, xmtrRail, EP_TXD_PHASE_ACTIVE))
+       ep_kthread_schedule (&xmtr->Subsys->Thread, lbolt);
+    
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    if (EP_IS_NO_INTERRUPT(txd->Envelope.Attr))
+       PollActiveTransmitList (xmtr, POLL_TX_LIST);
+}
+
+EP_STATUS
+ep_transmit_message (EP_XMTR *xmtr, unsigned int dest, EP_SERVICE service, EP_ATTRIBUTE attr, 
+                    EP_TXH *handler, void *arg, EP_PAYLOAD *payload, EP_NMD *nmd, int nFrags)
+{
+    EP_TXD       *txd;
+    int           i, len;
+
+    if (nFrags > EP_MAXFRAG || service > EP_MSG_NSVC)
+       return (EP_EINVAL);
+
+    if ((txd = GetTxd (xmtr, attr)) == NULL)
+       return (EP_ENOMEM);
+
+    txd->Handler = handler;
+    txd->Arg     = arg;
+    txd->Service = service;
+    txd->NodeId  = (unsigned short) dest;
+
+    /* Initialise the envelope */
+    txd->Envelope.Version   = EP_ENVELOPE_VERSION;
+    txd->Envelope.Attr      = EP_CLEAR_LOCAL_ATTR(attr);
+    txd->Envelope.Range     = EP_RANGE (dest, dest);
+    txd->Envelope.TxdMain   = txd->NmdMain;
+    txd->Envelope.nFrags    = nFrags;
+
+    for (i = len = 0; i < nFrags; len += nmd[i].nmd_len, i++)
+       txd->Envelope.Frags[i] = nmd[i];
+
+    if (payload)
+    {
+       txd->Envelope.Attr = EP_SET_HAS_PAYLOAD(txd->Envelope.Attr);
+
+       bcopy (payload, &txd->Payload, sizeof (EP_PAYLOAD));
+    }
+
+    DoTransmit (xmtr, txd);
+
+    BucketStat (xmtr->Subsys, DataXmit, len);
+
+    return (EP_SUCCESS);
+}
+
+EP_STATUS
+ep_multicast_message (EP_XMTR *xmtr, unsigned int destLo, unsigned int destHi, bitmap_t *bitmap, EP_SERVICE service, 
+                    EP_ATTRIBUTE attr, EP_TXH *handler, void *arg, EP_PAYLOAD *payload, EP_NMD *nmd, int nFrags)
+{
+    EP_SYS       *sys = xmtr->Subsys->Subsys.Sys;
+    EP_TXD       *txd;
+    int           nnodes;
+    int           i, len;
+    unsigned long flags;    
+
+    if (nFrags > EP_MAXFRAG || service > EP_MSG_NSVC)
+       return (EP_EINVAL);
+
+    if (destLo == -1) 
+       destLo = sys->Position.pos_nodeid & ~(EP_MAX_NODES-1);
+
+    if (destHi == -1 && (destHi = ((sys->Position.pos_nodeid + EP_MAX_NODES) & ~(EP_MAX_NODES-1)) - 1) >= sys->Position.pos_nodes)
+       destHi = sys->Position.pos_nodes-1;
+
+    nnodes = (destHi-destLo+1);
+
+    if ((txd = GetTxd (xmtr, attr)) == NULL)
+       return (EP_ENOMEM);
+
+    txd->Handler = handler;
+    txd->Arg     = arg;
+    txd->Service = service;
+
+    /* Initialise the envelope */
+    txd->Envelope.Version   = EP_ENVELOPE_VERSION;
+    txd->Envelope.Attr      = EP_SET_MULTICAST(EP_CLEAR_LOCAL_ATTR(attr));
+    txd->Envelope.Range     = EP_RANGE (destLo, destHi);
+    txd->Envelope.TxdMain   = txd->NmdMain;
+    txd->Envelope.nFrags    = nFrags;
+
+    for (i = len = 0; i < nFrags; len += nmd[i].nmd_len, i++)
+       txd->Envelope.Frags[i] = nmd[i];
+
+    if (payload)
+    {
+       txd->Envelope.Attr = EP_SET_HAS_PAYLOAD(txd->Envelope.Attr);
+
+       bcopy (payload, &txd->Payload, sizeof (EP_PAYLOAD));
+    }
+
+    spin_lock_irqsave (&sys->NodeLock, flags);
+    if (EP_IS_SVC_INDICATOR(attr)) 
+       ep_xmtr_svc_indicator_bitmap(xmtr, EP_ATTR2DATA(attr), txd->TxdMain->Bitmap, destLo, nnodes);
+    else
+       bt_subset (statemap_tobitmap(sys->NodeSet), txd->TxdMain->Bitmap, destLo, nnodes);
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+
+    if (bitmap != NULL)                                                                        /* bitmap supplied, so intersect it with */
+       bt_intersect (txd->TxdMain->Bitmap, bitmap, nnodes);                            /* the current node set map */
+    
+    if ((attr & EP_NOT_MYSELF) && destLo <= sys->Position.pos_nodeid && sys->Position.pos_nodeid <= destHi)
+       BT_CLEAR (txd->TxdMain->Bitmap, (sys->Position.pos_nodeid-destLo));                     /* clear myself if not wanted */
+
+    if ((i = bt_lowbit (txd->TxdMain->Bitmap, nnodes)) < 0)
+    {
+       FreeTxd (xmtr, txd);
+       return (EP_NODE_DOWN);
+    }
+
+    txd->NodeId = (unsigned short) i;
+
+    DoTransmit (xmtr, txd);
+
+    BucketStat (xmtr->Subsys, McastXmit, len);
+
+    return (EP_SUCCESS);
+}
+
+EP_STATUS
+ep_transmit_rpc (EP_XMTR *xmtr, unsigned int dest, EP_SERVICE service, EP_ATTRIBUTE attr,
+                EP_TXH *handler, void *arg, EP_PAYLOAD *payload, EP_NMD *nmd, int nFrags)
+{
+    EP_TXD       *txd;
+    int           i, len;
+
+    if (nFrags > EP_MAXFRAG || service > EP_MSG_NSVC)
+       return (EP_EINVAL);
+
+    if ((txd = GetTxd (xmtr, attr)) == NULL)
+       return (EP_ENOMEM);
+
+    txd->Handler = handler;
+    txd->Arg     = arg;
+    txd->Service = service;
+    txd->NodeId  = dest;
+
+    /* Initialise the envelope */
+    txd->Envelope.Version   = EP_ENVELOPE_VERSION;
+    txd->Envelope.Attr      = EP_SET_RPC(EP_CLEAR_LOCAL_ATTR(attr));    
+    txd->Envelope.Range     = EP_RANGE (dest, dest);
+    txd->Envelope.TxdMain   = txd->NmdMain;
+    txd->Envelope.nFrags    = nFrags;
+     
+    for (i = len = 0; i < nFrags; len += nmd[i].nmd_len, i++)
+       txd->Envelope.Frags[i] = nmd[i];
+
+    if (payload)
+    {
+       txd->Envelope.Attr = EP_SET_HAS_PAYLOAD(txd->Envelope.Attr);
+
+       bcopy (payload, &txd->Payload, sizeof (EP_PAYLOAD));
+    }
+
+    DoTransmit (xmtr, txd);
+
+    BucketStat (xmtr->Subsys, RPCXmit, len);
+
+    return (EP_SUCCESS);
+}
+
+EP_STATUS
+ep_multicast_forward (EP_XMTR *xmtr, unsigned int dest, EP_SERVICE service, EP_ATTRIBUTE attr, EP_TXH *handler, void *arg,
+                     EP_ENVELOPE *env,  EP_PAYLOAD *payload, bitmap_t *bitmap, EP_NMD *nmd, int nFrags)
+{
+    EP_TXD       *txd;
+    int           i, len;
+
+    if (nFrags > EP_MAXFRAG || service > EP_MSG_NSVC)
+       return (EP_EINVAL);
+
+    if ((txd = GetTxd (xmtr, attr)) == NULL)
+       return (EP_ENOMEM);
+
+    txd->Handler = handler;
+    txd->Arg     = arg;
+    txd->Service = service;
+    txd->NodeId  = (unsigned short) dest;
+
+    /* Initialise the envelope */
+    txd->Envelope.Version   = EP_ENVELOPE_VERSION;
+    txd->Envelope.Attr      = EP_SET_MULTICAST(EP_CLEAR_LOCAL_ATTR(attr));
+    txd->Envelope.Range     = env->Range;
+    txd->Envelope.TxdMain   = txd->NmdMain;
+    txd->Envelope.nFrags    = nFrags;
+
+    for (i = len = 0; i < nFrags; len += nmd[i].nmd_len, i++)
+       txd->Envelope.Frags[i] = nmd[i];
+
+    bt_copy (bitmap, txd->TxdMain->Bitmap, EP_RANGE_HIGH(env->Range) - EP_RANGE_LOW(env->Range) + 1);
+
+    if (payload)
+    {
+       txd->Envelope.Attr = EP_SET_HAS_PAYLOAD(txd->Envelope.Attr);
+
+       bcopy (payload, &txd->Payload, sizeof (EP_PAYLOAD));
+    }
+
+    DoTransmit (xmtr, txd);
+
+    BucketStat (xmtr->Subsys, McastXmit, len);
+
+    return (EP_SUCCESS);
+}
+
+int
+ep_poll_transmits (EP_XMTR *xmtr)
+{
+    return (PollActiveTransmitList (xmtr, POLL_TX_LIST));
+}
+
+int
+ep_enable_txcallbacks (EP_XMTR *xmtr)
+{
+    return (PollActiveTransmitList (xmtr, ENABLE_TX_CALLBACK));
+}
+
+int
+ep_disable_txcallbacks (EP_XMTR *xmtr)
+{
+    return (PollActiveTransmitList (xmtr, DISABLE_TX_CALLBACK));
+}
+
+/* functions for accessing fields of txds */
+int              ep_txd_node(EP_TXD *txd)              { return (txd->NodeId); }
+EP_STATUSBLK    *ep_txd_statusblk(EP_TXD *txd)                 { return (&txd->TxdMain->StatusBlk); }
+
+void
+ep_xmtr_xid_msg_handler (void *arg, EP_MANAGER_MSG *msg)
+{
+    EP_XMTR          *xmtr = (EP_XMTR *) arg;
+    EP_SYS           *sys  = xmtr->Subsys->Subsys.Sys;
+    struct list_head *el,*nel;
+    unsigned long     flags;
+
+    switch (msg->Hdr.Type)
+    {
+    case EP_MANAGER_MSG_TYPE_FAILOVER_REQUEST:
+       spin_lock_irqsave (&xmtr->Lock, flags);
+       list_for_each (el, &xmtr->ActiveDescList) {
+           EP_TXD      *txd     = list_entry (el, EP_TXD, Link);
+           EP_TXD_RAIL *txdRail = txd->TxdRail;
+
+           if (txdRail != NULL && EP_XIDS_MATCH (msg->Body.Failover.Xid, txd->Envelope.Xid))
+           {
+               EP_XMTR_RAIL       *xmtrRail = txdRail->XmtrRail;
+               EP_RAIL            *rail     = xmtrRail->CommsRail->Rail;
+               EP_MANAGER_MSG_BODY msgBody;
+               int                 rnum;
+
+               if (! (msg->Body.Failover.Railmask & EP_RAIL2RAILMASK (rail->Number)))
+               {
+                   /* Need to failover this txd to a different rail, select a rail from
+                    * the set that she has asked us to use and which is connected to her
+                    * on this transmitter.   If there are no such rails, then in all probability
+                    * we're offline on all common rails and eventually she will see we have no
+                    * rails in common and abort the receive. */
+                   if ((rnum = ep_xmtr_prefrail (xmtr, msg->Body.Failover.Railmask, txd->NodeId)) < 0)
+                       ep_debugf (DBG_XMTR, "%s: ep_xmtr_xid_msg_handler: FAILOVER_REQUEST but can't determine rail (%04x,%04x,%d,%04x)\n",
+                                  rail->Name, msg->Body.Failover.Railmask, xmtr->RailMask, txd->NodeId, sys->Nodes[txd->NodeId].ConnectedRails);
+                   else
+                   {
+                       EP_XMTR_RAIL *nXmtrRail = xmtr->Rails[rnum];
+
+                       EPRINTF4 (DBG_XMTR, "%s: ep_xmtr_xid_msg_handler: FAILOVER_REQUEST txd=%p XID=%llx-> rail %d\n", rail->Name, txd, (long long) txd->Envelope.Xid.Unique, rnum);
+
+                       /* Bind the txd rail onto the new rail - it doesn't matter if we fail
+                        * as it will remain bound to the original rail */
+                       (void) EP_XMTR_OP (nXmtrRail, BindTxd) (txd, nXmtrRail, EP_TXD_PHASE_PASSIVE);
+                   }
+               }
+
+               /* Send a failover response including an envelope update */
+               msgBody.FailoverTxd.Rail     = rail->Number;
+               msgBody.FailoverTxd.Xid      = txd->Envelope.Xid;
+               msgBody.FailoverTxd.TxdRail  = txd->Envelope.TxdRail;
+
+               ep_send_message (rail, msg->Hdr.NodeId, EP_MANAGER_MSG_TYPE_FAILOVER_RESPONSE, msg->Hdr.Xid, &msgBody);
+           }
+       }
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_GET_NODE_STATE_RESPONSE: {
+       int         txd_has_not_sent_envelope = 0;
+       EP_TXD      *txd            = NULL;
+       EP_TXD_RAIL *txdRail        = NULL;
+
+       if (msg->Body.NodeState.NetworkErrorState != 0)
+           ep_kthread_schedule (&xmtr->Subsys->Thread, lbolt + MESSAGE_RETRY_TIME);
+       else
+       {
+           spin_lock_irqsave (&xmtr->Lock, flags);
+           list_for_each_safe (el, nel, &xmtr->ActiveDescList) {
+               
+               txd     = list_entry (el, EP_TXD, Link);
+               txdRail = txd->TxdRail;
+               
+               if (txdRail != NULL && EP_XIDS_MATCH (msg->Hdr.Xid, txd->Envelope.Xid)) {
+                   txd_has_not_sent_envelope = EP_XMTR_OP(txdRail->XmtrRail,CheckTxdState)(txd);
+                   break;
+               }
+           }
+           
+           if (txd_has_not_sent_envelope) {
+               EPRINTF2 (DBG_STABILISE, "ep_xmtr_xid_msg_handler: GET_NODE_STATE_RESPONSE txd=%p XID=%llx not sent envelope\n",
+                         txd, (long long) txd->Envelope.Xid.Unique);
+
+               /* at this point it has finished stabalising */
+               txd->Envelope.Attr = EP_CLEAR_TXD_STABALISING(txd->Envelope.Attr);
+
+               /* store railmask into txd if not a service indicator or timeout */
+               if (EP_IS_NO_TYPE(txd->Envelope.Attr))
+                   txd->Envelope.Attr = EP_SET_DATA(txd->Envelope.Attr, EP_TYPE_RAILMASK, msg->Body.NodeState.Railmask);
+
+               spin_unlock_irqrestore (&xmtr->Lock, flags);
+               
+               /* TXD is now no longer bound to a rail , so let ep_check_xmtr() handle it */
+               ep_kthread_schedule (&xmtr->Subsys->Thread, lbolt);
+           }
+           else
+               spin_unlock_irqrestore (&xmtr->Lock, flags);    
+       }
+       break;
+    }
+    default:
+       panic ("ep_xmtr_xid_msg_handler: XID match but invalid message type\n");
+    }
+}
+
+EP_XMTR *
+ep_alloc_xmtr (EP_SYS *sys)
+{
+    EP_COMMS_SUBSYS   *subsys;
+    EP_XMTR          *xmtr;
+    struct list_head *el;
+
+    if ((subsys = (EP_COMMS_SUBSYS *) ep_subsys_find (sys, EPCOMMS_SUBSYS_NAME)) == NULL)
+       return (NULL);
+
+    KMEM_ZALLOC (xmtr, EP_XMTR *, sizeof (EP_XMTR), 1);
+
+    if (xmtr == NULL)
+       return (NULL);
+    
+    xmtr->Subsys = subsys;
+
+    spin_lock_init (&xmtr->Lock);
+    INIT_LIST_HEAD (&xmtr->ActiveDescList);
+    
+    kcondvar_init (&xmtr->FreeDescSleep);
+    spin_lock_init (&xmtr->FreeDescLock);
+    INIT_LIST_HEAD (&xmtr->FreeDescList);
+    INIT_LIST_HEAD (&xmtr->DescBlockList);
+
+    ep_xid_cache_init (sys, &xmtr->XidCache);
+
+    xmtr->XidCache.MessageHandler = ep_xmtr_xid_msg_handler;
+    xmtr->XidCache.Arg            = xmtr;
+
+    kmutex_lock (&subsys->Lock);
+    list_add_tail (&xmtr->Link, &subsys->Transmitters);
+
+    ep_procfs_xmtr_add(xmtr);
+
+    /* Now add all rails which are already started */
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *commsRail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       EP_RAIL_OP(commsRail, Xmtr.AddRail) (xmtr, commsRail);
+    }
+    kmutex_unlock (&subsys->Lock);
+
+    ep_mod_inc_usecount();
+
+    return (xmtr);
+}
+
+void
+ep_free_xmtr (EP_XMTR *xmtr)
+{
+    EP_COMMS_SUBSYS  *subsys = xmtr->Subsys;
+    EP_SYS           *sys    = subsys->Subsys.Sys;
+    struct list_head *el;
+    
+    kmutex_lock (&subsys->Lock);
+    list_for_each (el, &subsys->Rails) { 
+       EP_COMMS_RAIL *rail = list_entry (el, EP_COMMS_RAIL, Link);
+
+       EP_RAIL_OP(rail,Xmtr.DelRail) (xmtr, rail);
+    }
+
+    list_del (&xmtr->Link);
+    kmutex_unlock (&subsys->Lock);
+
+    /* all the desc's must be free */
+    ASSERT(xmtr->FreeDescCount == xmtr->TotalDescCount);
+
+    /* delete the descs */
+    while (!list_empty (&xmtr->DescBlockList))
+       FreeTxdBlock( xmtr, list_entry(xmtr->DescBlockList.next, EP_TXD_BLOCK , Link));
+
+    /* they had better all be gone now */
+    ASSERT((xmtr->FreeDescCount == 0) && (xmtr->TotalDescCount == 0));
+
+    ep_procfs_xmtr_del(xmtr);
+
+    ep_xid_cache_destroy (sys, &xmtr->XidCache);
+
+    spin_lock_destroy (&xmtr->Lock);
+    KMEM_FREE (xmtr, sizeof (EP_XMTR));
+
+    ep_mod_dec_usecount();
+}
+
+long
+ep_check_xmtr (EP_XMTR *xmtr, long nextRunTime)
+{
+    EP_COMMS_SUBSYS  *subsys = xmtr->Subsys;
+    EP_SYS           *sys    = subsys->Subsys.Sys;
+    struct list_head *el, *nel;
+    struct list_head  txdList;
+    unsigned long       flags;
+    int                 timed_out=0;
+    int                 i;
+    EP_MANAGER_MSG_BODY body;
+
+    INIT_LIST_HEAD (&txdList);
+
+    /* See if we have any txd's which need to be bound to a rail */
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    list_for_each_safe (el, nel, &xmtr->ActiveDescList) {
+       EP_TXD      *txd      = list_entry (el, EP_TXD, Link);
+       EP_NODE     *node     = &sys->Nodes[txd->NodeId];
+       EP_RAILMASK nodeRails = node->ConnectedRails & xmtr->RailMask;
+       EP_ENVELOPE *env      = &txd->Envelope;
+
+       if (EP_IS_TXD_STABALISING(txd->Envelope.Attr)) 
+       {
+           ASSERT(txd->TxdRail != NULL);
+
+           if (AFTER (lbolt, txd->RetryTime))
+           {
+               EPRINTF6 (DBG_STABILISE, "ep_check_xmtr txd=%p txdRail=%p send get node state to %d Xid=%08x.%08x.%016llx\n",
+                         txd, txd->TxdRail, txd->NodeId, env->Xid.Generation, env->Xid.Handle, env->Xid.Unique);
+               
+               body.Service = txd->Service;
+               if (ep_send_message ( txd->TxdRail->XmtrRail->CommsRail->Rail, txd->NodeId, EP_MANAGER_MSG_TYPE_GET_NODE_STATE, env->Xid, &body) == 0)
+                   txd->RetryTime = lbolt + (MESSAGE_RETRY_TIME << ep_backoff (&txd->Backoff, EP_BACKOFF_STABILISE));
+               else
+                   txd->RetryTime = lbolt + MSGBUSY_RETRY_TIME;
+           }
+
+           ep_kthread_schedule (&subsys->Thread, txd->RetryTime);
+           continue;
+       }
+
+       if (txd->TxdRail != NULL)
+           continue;
+
+       switch (EP_ATTR2TYPE(txd->Envelope.Attr)) 
+       {
+       case EP_TYPE_SVC_INDICATOR: 
+       {
+           EP_RAILMASK       rmask=0;
+           struct list_head *tmp;
+
+           list_for_each (tmp, &subsys->Rails) { 
+               EP_COMMS_RAIL *commsRail = list_entry (tmp, EP_COMMS_RAIL, Link);
+               if ( cm_svc_indicator_is_set(commsRail->Rail, EP_ATTR2DATA(txd->Envelope.Attr), txd->NodeId))
+                   rmask |= EP_RAIL2RAILMASK(commsRail->Rail->Number);
+           } 
+           nodeRails &= rmask;
+           break;
+       }
+       case EP_TYPE_TIMEOUT:
+           timed_out = AFTER(lbolt, txd->TimeStamp + EP_ATTR2DATA(txd->Envelope.Attr)) ? (1) : (0);
+           break;
+       case EP_TYPE_RAILMASK:
+           nodeRails &= EP_ATTR2DATA(txd->Envelope.Attr);
+           break;
+       default:
+           timed_out = AFTER(lbolt, txd->TimeStamp +  EP_DEFAULT_TIMEOUT) ? (1) : (0);
+           break;
+       }
+
+       if (nodeRails == 0 || timed_out || (EP_IS_NO_FAILOVER(env->Attr) && EP_IS_PREFRAIL_SET(env->Attr) && 
+                                           (nodeRails & EP_RAIL2RAILMASK(EP_ATTR2PREFRAIL(env->Attr))) == 0))
+       {
+           EPRINTF5 (timed_out ? DBG_STABILISE : DBG_XMTR, "ep_check_xmtr: txd=%p XID=%llx to %d no rails connected or cannot failover (nodeRails=0x%x,timed_out=%d\n", 
+                     txd, (long long) env->Xid.Unique, txd->NodeId, nodeRails, timed_out);
+
+           list_del  (&txd->Link);
+           list_add_tail (&txd->Link, &txdList);
+       }
+       else
+       {
+           EP_XMTR_RAIL *xmtrRail;
+           int i, len, rnum;
+
+           if (EP_IS_PREFRAIL_SET(env->Attr) && (nodeRails & EP_RAIL2RAILMASK(EP_ATTR2PREFRAIL(env->Attr))))
+               rnum = EP_ATTR2PREFRAIL(env->Attr);
+           else
+               rnum = ep_pickRail (nodeRails);
+
+           EPRINTF3 (DBG_XMTR, "ep_check_xmtr: txd=%p XID=%llx mapping NMDs onto rail %d \n", txd, (long long) env->Xid.Unique, rnum);
+
+           for (i = len = 0; i < env->nFrags; i++, len += env->Frags[i].nmd_len)
+               ep_nmd_map_rails (sys, &env->Frags[i], nodeRails);
+
+           if ((xmtrRail = xmtr->Rails[rnum]) == NULL || 
+               !EP_XMTR_OP(xmtrRail,BindTxd) (txd, xmtrRail, EP_TXD_PHASE_ACTIVE))
+               ep_kthread_schedule (&subsys->Thread, lbolt + RESOURCE_RETRY_TIME);
+       }
+    }
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    while (! list_empty (&txdList))
+    {
+       EP_TXD *txd = list_entry (txdList.next, EP_TXD, Link);
+       list_del (&txd->Link);
+
+       txd->Handler (txd, txd->Arg, EP_NODE_DOWN);
+       FreeTxd (xmtr, txd);
+    }
+
+    /* Check to see if we're low on txds */
+    if (xmtr->FreeDescCount < ep_txd_lowat)
+       AllocateTxdBlock (xmtr, 0, NULL);
+    
+    /* Then check each rail */
+    for (i = 0; i < EP_MAX_RAILS; i++) 
+       if (xmtr->RailMask & (1 << i) ) 
+           nextRunTime = EP_XMTR_OP (xmtr->Rails[i],Check) (xmtr->Rails[i], nextRunTime);
+    return (nextRunTime);
+}
+
+void
+ep_display_txd (DisplayInfo *di, EP_TXD *txd)
+{
+    EP_ENVELOPE *env     = &txd->Envelope;
+    EP_TXD_RAIL *txdRail = txd->TxdRail;
+
+    (di->func)(di->arg, "TXD: %p Version=%x Attr=%x Xid=%08x.%08x.%016llx\n", txd, 
+              env->Version, env->Attr, env->Xid.Generation, env->Xid.Handle, (long long) env->Xid.Unique);
+    (di->func)(di->arg,  "     NodeId=%d Range=%d.%d TxdRail=%x TxdMain=%x.%x.%x nFrags=%d\n",
+              env->NodeId, EP_RANGE_LOW(env->Range), EP_RANGE_HIGH(env->Range), env->TxdRail,
+              env->TxdMain.nmd_addr, env->TxdMain.nmd_len, env->TxdMain.nmd_attr, env->nFrags);
+    (di->func)(di->arg,  "       Frag[0] %08x.%08x.%08x\n", env->Frags[0].nmd_addr, env->Frags[0].nmd_len, env->Frags[0].nmd_attr);
+    (di->func)(di->arg,  "       Frag[1] %08x.%08x.%08x\n", env->Frags[1].nmd_addr, env->Frags[1].nmd_len, env->Frags[1].nmd_attr);
+    (di->func)(di->arg,  "       Frag[2] %08x.%08x.%08x\n", env->Frags[2].nmd_addr, env->Frags[2].nmd_len, env->Frags[2].nmd_attr);
+    (di->func)(di->arg,  "       Frag[3] %08x.%08x.%08x\n", env->Frags[3].nmd_addr, env->Frags[3].nmd_len, env->Frags[3].nmd_attr);
+
+    if (txdRail != NULL) EP_XMTR_OP (txdRail->XmtrRail, DisplayTxd) (di, txdRail);
+}
+
+void
+ep_display_xmtr (DisplayInfo *di, EP_XMTR *xmtr)
+{
+    int               freeCount   = 0;
+    int               activeCount = 0;
+    struct list_head *el;
+    int               i;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&xmtr->FreeDescLock, flags);
+    list_for_each (el, &xmtr->FreeDescList)
+       freeCount++;
+    spin_unlock_irqrestore (&xmtr->FreeDescLock, flags);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    list_for_each (el, &xmtr->ActiveDescList)
+       activeCount++;
+    
+    (di->func)(di->arg, "ep_display_xmtr: xmtr=%p Free=%d Active=%d\n", xmtr, freeCount, activeCount);
+    for (i = 0; i < EP_MAX_RAILS; i++)
+       if (xmtr->Rails[i]) EP_XMTR_OP (xmtr->Rails[i], DisplayXmtr) (di, xmtr->Rails[i]);
+
+    list_for_each (el,&xmtr->ActiveDescList)
+       ep_display_txd (di, list_entry (el, EP_TXD, Link));
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+}
+
+void
+ep_xmtr_fillout_stats(EP_XMTR *xmtr, char *str)
+{
+    sprintf(str+strlen(str),"Tx     %lu  %lu /sec\n",       GET_STAT_TOTAL(xmtr->stats,tx),      GET_STAT_PER_SEC(xmtr->stats,tx) );
+    sprintf(str+strlen(str),"MBytes %lu  %lu Mbytes/sec\n", GET_STAT_TOTAL(xmtr->stats,tx_len) / (1024*1024),  GET_STAT_PER_SEC(xmtr->stats,tx_len) / (1024*1024));
+}
+
+void
+ep_xmtr_rail_fillout_stats(EP_XMTR_RAIL *xmtr_rail, char *str)
+{
+    sprintf(str+strlen(str),"Tx     %lu  %lu /sec\n",       GET_STAT_TOTAL(xmtr_rail->stats,tx),      GET_STAT_PER_SEC(xmtr_rail->stats,tx) );
+    sprintf(str+strlen(str),"MBytes %lu  %lu Mbytes/sec\n", GET_STAT_TOTAL(xmtr_rail->stats,tx_len) / (1024*1024),  GET_STAT_PER_SEC(xmtr_rail->stats,tx_len) / (1024*1024));
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcommsTx_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcommsTx_elan3.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcommsTx_elan3.c 2005-06-01 23:12:54.657430288 -0400
@@ -0,0 +1,1173 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcommsTx_elan3.c,v 1.17.2.2 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcommsTx_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "epcomms_elan3.h"
+#include "debug.h"
+
+#define XMTR_TO_RAIL(xmtrRail)         ((EP3_RAIL *) ((EP_XMTR_RAIL *) xmtrRail)->CommsRail->Rail)
+#define XMTR_TO_DEV(xmtrRail)          (XMTR_TO_RAIL(xmtrRail)->Device)
+#define XMTR_TO_SUBSYS(xmtrRail)       (((EP_XMTR_RAIL *) xmtrRail)->Xmtr->Subsys)
+
+static void TxEnveEvent (EP3_RAIL *rail, void *arg);
+static void TxEnveRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status);
+static void TxEnveVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma);
+
+static EP3_COOKIE_OPS EnveCookieOps =
+{
+    TxEnveEvent,
+    TxEnveRetry,
+    NULL, /* DmaCancelled */
+    TxEnveVerify
+};
+
+static void TxDataEvent (EP3_RAIL *rail, void *arg);
+static void TxDataRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status);
+static void TxDataVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma);
+
+static EP3_COOKIE_OPS DataCookieOps =
+{
+    TxDataEvent,
+    TxDataRetry,
+    NULL, /* DmaCancelled */
+    TxDataVerify
+};
+
+static void TxDoneEvent (EP3_RAIL *dev, void *arg);
+static void TxDoneRetry (EP3_RAIL *dev, void *arg, E3_DMA_BE *dma, int status);
+static void TxDoneVerify (EP3_RAIL  *dev, void *arg, E3_DMA_BE *dma);
+
+static EP3_COOKIE_OPS DoneCookieOps = 
+{
+    TxDoneEvent,
+    TxDoneRetry,
+    NULL, /* DmaCancelled */
+    TxDoneVerify,
+} ;
+
+static int
+AllocateTxdRailBlock (EP3_XMTR_RAIL *xmtrRail)
+{
+    EP3_RAIL          *rail = XMTR_TO_RAIL (xmtrRail);
+    ELAN3_DEV         *dev = rail->Device;
+    EP3_TXD_RAIL_BLOCK *blk;
+    EP3_TXD_RAIL       *txdRail;
+    sdramaddr_t        pTxdElan;
+    EP3_TXD_RAIL_MAIN  *pTxdMain;
+    E3_Addr           pTxdElanAddr;
+    E3_Addr           pTxdMainAddr;
+    E3_BlockCopyEvent  event;
+    int                       i;
+    unsigned long      flags;
+
+    KMEM_ZALLOC (blk, EP3_TXD_RAIL_BLOCK *, sizeof (EP3_TXD_RAIL_BLOCK), 1);
+
+    if (blk == NULL)
+       return 0;
+
+    if ((pTxdElan = ep_alloc_elan (&rail->Generic, EP3_TXD_RAIL_ELAN_SIZE * EP3_NUM_TXD_PER_BLOCK, 0, &pTxdElanAddr)) == (sdramaddr_t) 0)
+    {
+       KMEM_FREE (blk, sizeof (EP3_TXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    if ((pTxdMain = ep_alloc_main (&rail->Generic, EP3_TXD_RAIL_MAIN_SIZE * EP3_NUM_TXD_PER_BLOCK, 0, &pTxdMainAddr)) == (EP3_TXD_RAIL_MAIN *) NULL)
+    {
+       ep_free_elan (&rail->Generic, pTxdElanAddr, EP3_TXD_RAIL_ELAN_SIZE * EP3_NUM_TXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP3_TXD_RAIL_BLOCK));
+       return 0;
+    }
+    
+    if (ReserveDmaRetries (rail, EP3_NUM_TXD_PER_BLOCK, 0) != ESUCCESS)
+    {
+       ep_free_main (&rail->Generic, pTxdMainAddr, EP3_TXD_RAIL_MAIN_SIZE * EP3_NUM_TXD_PER_BLOCK);
+       ep_free_elan (&rail->Generic, pTxdElanAddr, EP3_TXD_RAIL_ELAN_SIZE * EP3_NUM_TXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP3_TXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    for (txdRail = &blk->Txd[0], i = 0; i < EP3_NUM_TXD_PER_BLOCK; i++, txdRail++)
+    {
+       txdRail->Generic.XmtrRail = &xmtrRail->Generic;
+       txdRail->TxdElan          = pTxdElan;
+       txdRail->TxdElanAddr      = pTxdElanAddr;
+       txdRail->TxdMain          = pTxdMain;
+       txdRail->TxdMainAddr      = pTxdMainAddr;
+
+       RegisterCookie (&rail->CookieTable, &txdRail->EnveCookie, pTxdElanAddr + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent), &EnveCookieOps, (void *) txdRail);
+       RegisterCookie (&rail->CookieTable, &txdRail->DataCookie, pTxdElanAddr + offsetof (EP3_TXD_RAIL_ELAN, DataEvent), &DataCookieOps, (void *) txdRail);
+       RegisterCookie (&rail->CookieTable, &txdRail->DoneCookie, pTxdElanAddr + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent), &DoneCookieOps, (void *) txdRail);
+
+       EP3_INIT_COPY_EVENT (event, txdRail->EnveCookie, pTxdMainAddr + offsetof (EP3_TXD_RAIL_MAIN, EnveEvent), 0);
+       elan3_sdram_copyl_to_sdram (dev, &event, pTxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent), sizeof (E3_BlockCopyEvent));
+
+       EP3_INIT_COPY_EVENT (event, txdRail->DataCookie, pTxdMainAddr + offsetof (EP3_TXD_RAIL_MAIN, DataEvent), 0);
+       elan3_sdram_copyl_to_sdram (dev, &event, pTxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent), sizeof (E3_BlockCopyEvent));
+
+       EP3_INIT_COPY_EVENT (event, txdRail->DoneCookie, pTxdMainAddr + offsetof (EP3_TXD_RAIL_MAIN, DoneEvent), 0);
+       elan3_sdram_copyl_to_sdram (dev, &event, pTxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent), sizeof (E3_BlockCopyEvent));
+       
+       pTxdMain->EnveEvent = EP3_EVENT_FREE;
+       pTxdMain->DataEvent = EP3_EVENT_FREE;
+       pTxdMain->DoneEvent = EP3_EVENT_FREE;
+
+       /* move onto next descriptor */
+       pTxdElan     += EP3_TXD_RAIL_ELAN_SIZE;
+       pTxdElanAddr += EP3_TXD_RAIL_ELAN_SIZE;
+       pTxdMain      = (EP3_TXD_RAIL_MAIN *) ((unsigned long) pTxdMain + EP3_TXD_RAIL_MAIN_SIZE);
+       pTxdMainAddr += EP3_TXD_RAIL_MAIN_SIZE;
+    }
+
+    spin_lock_irqsave (&xmtrRail->FreeDescLock, flags);
+
+    list_add  (&blk->Link, &xmtrRail->DescBlockList);
+    xmtrRail->TotalDescCount += EP3_NUM_TXD_PER_BLOCK;
+    xmtrRail->FreeDescCount  += EP3_NUM_TXD_PER_BLOCK;
+
+    for (i = 0; i < EP3_NUM_TXD_PER_BLOCK; i++)
+       list_add (&blk->Txd[i].Generic.Link, &xmtrRail->FreeDescList);
+
+    spin_unlock_irqrestore (&xmtrRail->FreeDescLock, flags);
+    
+    return 1;
+}
+
+static void
+FreeTxdRailBlock (EP3_XMTR_RAIL *xmtrRail, EP3_TXD_RAIL_BLOCK *blk)
+{
+    EP3_RAIL     *rail = XMTR_TO_RAIL(xmtrRail);
+    EP3_TXD_RAIL *txdRail;
+    unsigned long flags;
+    int i;
+
+    spin_lock_irqsave (&xmtrRail->FreeDescLock, flags);
+
+    list_del (&blk->Link);
+    
+    xmtrRail->TotalDescCount -= EP3_NUM_TXD_PER_BLOCK;
+    
+    for (txdRail = &blk->Txd[0], i = 0; i < EP3_NUM_TXD_PER_BLOCK; i++, txdRail++)
+    {
+       xmtrRail->FreeDescCount--;
+       
+       list_del (&txdRail->Generic.Link);
+       
+       DeregisterCookie (&rail->CookieTable, &txdRail->EnveCookie);
+       DeregisterCookie (&rail->CookieTable, &txdRail->DataCookie);
+       DeregisterCookie (&rail->CookieTable, &txdRail->DoneCookie);
+    }
+
+    spin_unlock_irqrestore (&xmtrRail->FreeDescLock, flags);
+
+    ReleaseDmaRetries (rail, EP3_NUM_TXD_PER_BLOCK);
+
+    ep_free_main (&rail->Generic, blk->Txd[0].TxdMainAddr, EP3_TXD_RAIL_MAIN_SIZE * EP3_NUM_TXD_PER_BLOCK);
+    ep_free_elan (&rail->Generic, blk->Txd[0].TxdElanAddr, EP3_TXD_RAIL_ELAN_SIZE * EP3_NUM_TXD_PER_BLOCK);
+    KMEM_FREE (blk, sizeof (EP3_TXD_RAIL_BLOCK));
+}
+
+static EP3_TXD_RAIL *
+GetTxdRail (EP3_XMTR_RAIL *xmtrRail)
+{
+    EP_COMMS_SUBSYS  *subsys = xmtrRail->Generic.Xmtr->Subsys;
+    EP3_TXD_RAIL     *txdRail;
+    int low_on_txds;
+    unsigned long flags;
+
+    spin_lock_irqsave (&xmtrRail->FreeDescLock, flags);
+
+    if (list_empty (&xmtrRail->FreeDescList))
+       txdRail = NULL;
+    else
+    {
+       txdRail = list_entry (xmtrRail->FreeDescList.next, EP3_TXD_RAIL, Generic.Link);
+
+#if defined(DEBUG)
+       {
+           EP_RAIL   *rail = xmtrRail->Generic.CommsRail->Rail;
+           ELAN3_DEV *dev  = ((EP3_RAIL *) rail)->Device;
+           
+           EP_ASSERT (rail, txdRail->TxdMain->EnveEvent == EP3_EVENT_FREE);
+           EP_ASSERT (rail, txdRail->TxdMain->DataEvent == EP3_EVENT_FREE);
+           EP_ASSERT (rail, txdRail->TxdMain->DoneEvent == EP3_EVENT_FREE);
+           EP_ASSERT (rail, SDRAM_ASSERT(elan3_sdram_readl (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count)) == 0));
+           EP_ASSERT (rail, SDRAM_ASSERT(elan3_sdram_readl (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));
+           EP_ASSERT (rail, SDRAM_ASSERT(elan3_sdram_readl (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count)) == 0));
+       }
+#endif
+       
+       list_del (&txdRail->Generic.Link);
+
+       xmtrRail->FreeDescCount--;
+    }
+    /* Wakeup the descriptor primer thread if there's not many left */
+    low_on_txds = (xmtrRail->FreeDescCount < ep_txd_lowat);
+
+    spin_unlock_irqrestore (&xmtrRail->FreeDescLock, flags);
+
+    if (low_on_txds)
+       ep_kthread_schedule (&subsys->Thread, lbolt);
+
+    return (txdRail);
+}
+
+static void
+FreeTxdRail (EP3_XMTR_RAIL *xmtrRail, EP3_TXD_RAIL *txdRail)
+{
+    unsigned long flags;
+
+#if defined(DEBUG_ASSERT)
+    {
+       EP_RAIL   *rail = xmtrRail->Generic.CommsRail->Rail;
+       ELAN3_DEV *dev  = ((EP3_RAIL *) rail)->Device;
+
+       EP_ASSERT (rail, txdRail->Generic.XmtrRail == &xmtrRail->Generic);
+       
+       EP_ASSERT (rail, txdRail->TxdMain->EnveEvent == EP3_EVENT_PRIVATE);
+       EP_ASSERT (rail, txdRail->TxdMain->DataEvent == EP3_EVENT_PRIVATE);
+       EP_ASSERT (rail, txdRail->TxdMain->DoneEvent == EP3_EVENT_PRIVATE);
+       EP_ASSERT (rail, SDRAM_ASSERT (elan3_sdram_readl (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count)) == 0));
+       EP_ASSERT (rail, SDRAM_ASSERT (elan3_sdram_readl (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));
+       EP_ASSERT (rail, SDRAM_ASSERT (elan3_sdram_readl (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count)) == 0));
+       
+       txdRail->TxdMain->EnveEvent = EP3_EVENT_FREE;
+       txdRail->TxdMain->DataEvent = EP3_EVENT_FREE;
+       txdRail->TxdMain->DoneEvent = EP3_EVENT_FREE;
+    }
+#endif
+
+    spin_lock_irqsave (&xmtrRail->FreeDescLock, flags);
+    
+    list_add (&txdRail->Generic.Link, &xmtrRail->FreeDescList);
+
+    xmtrRail->FreeDescCount++;
+
+    if (xmtrRail->FreeDescWaiting)
+    {
+       xmtrRail->FreeDescWaiting--;
+       kcondvar_wakeupall (&xmtrRail->FreeDescSleep, &xmtrRail->FreeDescLock);
+    }
+
+    spin_unlock_irqrestore (&xmtrRail->FreeDescLock, flags);
+}
+
+static void
+BindTxdToRail (EP_TXD *txd, EP3_TXD_RAIL *txdRail)
+{
+    ASSERT (SPINLOCK_HELD (&txd->Xmtr->Lock));
+
+    EPRINTF6 (DBG_XMTR, "%s: BindTxdToRail: txd=%p txdRail=%p XID=%08x.%08x.%016llx\n", 
+             XMTR_TO_RAIL(txdRail->Generic.XmtrRail)->Generic.Name, txd, txdRail, 
+             txd->Envelope.Xid.Generation, txd->Envelope.Xid.Handle, (long long) txd->Envelope.Xid.Unique);
+
+    txd->TxdRail = &txdRail->Generic;
+    txdRail->Generic.Txd = txd;
+}
+
+static void
+UnbindTxdFromRail (EP_TXD *txd, EP3_TXD_RAIL *txdRail)
+{
+    ASSERT (SPINLOCK_HELD (&txd->Xmtr->Lock));
+    ASSERT (txd->TxdRail == &txdRail->Generic && txdRail->Generic.Txd == txd);
+
+    EPRINTF6 (DBG_XMTR, "%s: UnbindTxdToRail: txd=%p txdRail=%p XID=%08x.%08x.%016llx\n", 
+             XMTR_TO_RAIL(txdRail->Generic.XmtrRail)->Generic.Name, txd, txdRail, 
+             txd->Envelope.Xid.Generation, txd->Envelope.Xid.Handle, (long long) txd->Envelope.Xid.Unique);
+    txd->TxdRail = NULL;
+    txdRail->Generic.Txd = NULL; 
+}
+
+/*
+ * TxEnveEvent: arg == EP_TXD
+ *    Called when envelope delivered
+ */
+static void
+TxEnveEvent (EP3_RAIL *rail, void *arg)
+{
+    panic ("TxEnveEvent");
+}
+
+/*
+ * TxEnveRetry: arg == EP3_TXD_RAIL
+ *    Called on retry of dma of large message envelope.
+ */
+static void
+TxEnveRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status)
+{
+    EP3_TXD_RAIL  *txdRail  = (EP3_TXD_RAIL *) arg;
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    
+    EPRINTF3 (DBG_XMTR, "%s: TxEnveRetry: xmtr %p txd %p\n",  rail->Generic.Name, xmtrRail, txdRail);
+    
+    EP_ASSERT (&rail->Generic, txdRail->TxdMain->EnveEvent == EP3_EVENT_ACTIVE);
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count)) == 1)); /* PCI read */
+    EP_ASSERT (&rail->Generic, dma->s.dma_direction == DMA_WRITE && EP_VP_TO_NODE(dma->s.dma_destVProc) == txdRail->Generic.Txd->NodeId);
+
+    if (! TxdShouldStabalise (&txdRail->Generic, &rail->Generic))
+       QueueDmaForRetry (rail, dma, EP_RETRY_LOW_PRI_RETRY + ep_backoff (&txdRail->Backoff, EP_BACKOFF_ENVELOPE));
+    else
+       QueueDmaForRetry (rail, dma, EP_RETRY_STABALISING);     /* place dma on stabilising list for neterr fixup */
+}
+
+static void
+TxEnveVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma)
+{
+    EP3_TXD_RAIL *txdRail = (EP3_TXD_RAIL *) arg;
+    
+    EP_ASSERT (&rail->Generic, txdRail->TxdMain->EnveEvent == EP3_EVENT_ACTIVE);
+    EP_ASSERT (&rail->Generic, SDRAM_ASSERT (elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count)) == 1)); /* PCI read */
+    EP_ASSERT (&rail->Generic, dma->s.dma_direction == DMA_WRITE && EP_VP_TO_NODE(dma->s.dma_destVProc) == txdRail->Generic.Txd->NodeId);
+}
+
+/*
+ * TxDataEvent: arg == EP3_TXD
+ *    Called on completion of a large transmit.
+ */
+static void
+TxDataEvent (EP3_RAIL *rail, void *arg)
+{
+    EP3_TXD_RAIL      *txdRail  = (EP3_TXD_RAIL *) arg;
+    EP3_XMTR_RAIL     *xmtrRail = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    EP_XMTR          *xmtr     = xmtrRail->Generic.Xmtr;
+    EP3_TXD_RAIL_MAIN *txdMain  = txdRail->TxdMain;
+    sdramaddr_t        txdElan  = txdRail->TxdElan;
+    int                delay    = 1;
+    EP_TXD            *txd;
+    unsigned long      flags;
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    for (;;)
+    {
+       if (EP3_EVENT_FIRED (txdRail->DataCookie, txdMain->DataEvent))
+           break;
+
+       if (EP3_EVENT_FIRING (rail->Device, txdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent), txdRail->DataCookie, txdMain->DataEvent))                /* PCI read */
+       {
+           if (delay > EP3_EVENT_FIRING_TLIMIT)
+               panic ("TxDataEvent: events set but block copy not completed\n");
+           DELAY(delay);
+           delay <<= 1;
+       }
+       else
+       {
+           EPRINTF3 (DBG_XMTR, "%s: TxDataEvent: xmtr %p txd %p previously collecting by polling\n", 
+                     rail->Generic.Name, xmtrRail, txdRail);
+           spin_unlock_irqrestore (&xmtr->Lock, flags);
+           return;
+       }
+       mb();
+    }
+
+    if ((txd = txdRail->Generic.Txd) == NULL ||                        /* If there is no txd, or if the descriptor is marked */
+       !(EP_IS_INTERRUPT_ENABLED(txd->Envelope.Attr)) ||       /* as no interrupt, or been reused as an RPC, */
+       (EP_IS_RPC(txd->Envelope.Attr)))                        /* then we were either called as a result of a previous */
+    {                                                          /* tx which was completed by polling or as a result */
+       spin_unlock_irqrestore (&xmtr->Lock, flags);            /* of a EnableTxCallBack/DisableTxCallback */
+
+       EPRINTF4 (DBG_XMTR, "%s: TxDataEvent: xmtr %p txd %p recyled (%x)\n", 
+                 rail->Generic.Name, xmtr, txd, txd ? txd->Envelope.Attr : 0);
+       return;
+    }
+
+    ASSERT (EP3_EVENT_FIRED (txdRail->EnveCookie, txdMain->EnveEvent));
+
+    EPRINTF5 (DBG_XMTR, "%s: TxDataEvent : xmtrRail=%p txdRail=%p tx=%p XID=%llx\n", 
+             rail->Generic.Name, xmtrRail, txdRail, txd, (long long) txd->Envelope.Xid.Unique);
+    
+    ep_xmtr_txd_stat(xmtr,txd);
+    
+    /* remove from active transmit lists */
+    list_del (&txd->Link);
+
+    UnbindTxdFromRail (txd, txdRail);
+    
+    /* clear the done flags for next time round */
+    txdMain->EnveEvent = EP3_EVENT_PRIVATE;
+    txdMain->DataEvent = EP3_EVENT_PRIVATE;
+    txdMain->DoneEvent = EP3_EVENT_PRIVATE;
+    
+    FreeTxdRail (xmtrRail, txdRail);
+
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+    
+    txd->Handler (txd, txd->Arg, EP_SUCCESS);
+    
+    FreeTxd (xmtr, txd);
+}
+
+/*
+ * TxDataRetry: arg == EP3_TXD
+ *    Called on retry of remote "put" dma of large transmit data.
+ */
+static void
+TxDataRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status)
+{
+    EP3_TXD_RAIL  *txdRail  = (EP3_TXD_RAIL *) arg;
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    EP_TXD        *txd      = txdRail->Generic.Txd;
+
+    EP_ASSERT (&rail->Generic, ((txdRail->TxdMain->DataEvent == EP3_EVENT_ACTIVE && 
+                                SDRAM_ASSERT (elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)) >= 1)) ||  /* PCI read */
+                               (EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent) &&
+                                SDRAM_ASSERT (elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)) == 0))));  /* PCI read */
+    EP_ASSERT (&rail->Generic, dma->s.dma_direction == DMA_WRITE && EP_VP_TO_NODE(dma->s.dma_destVProc) == txd->NodeId);
+
+    EPRINTF5 (DBG_XMTR, "%s: TxDataRetry: xmtrRail=%p txdRail=%p txd=%p XID=%llx\n", 
+             rail->Generic.Name, xmtrRail, txdRail, txd, (long long) txd->Envelope.Xid.Unique);
+    
+    QueueDmaForRetry (rail, dma, EP_RETRY_LOW_PRI_RETRY + ep_backoff (&txdRail->Backoff, EP_BACKOFF_DATA));
+}
+
+static void
+TxDataVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma)
+{
+    EP3_TXD_RAIL *txdRail = (EP3_TXD_RAIL *) arg;
+    EP_TXD       *txd     = txdRail->Generic.Txd;
+
+    EP_ASSERT (&rail->Generic, ((txdRail->TxdMain->DataEvent == EP3_EVENT_ACTIVE && 
+                                SDRAM_ASSERT (elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)) >= 1)) ||  /* PCI read */
+                               (EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent) &&
+                                SDRAM_ASSERT (elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)) == 0))));  /* PCI read */
+    EP_ASSERT (&rail->Generic, dma->s.dma_direction == DMA_WRITE && EP_VP_TO_NODE(dma->s.dma_destVProc) == txd->NodeId);
+}
+
+/*
+ * TxDoneEvent: arg == EP3_TXD
+ *    Called on completion of a RPC.
+ */
+static void
+TxDoneEvent (EP3_RAIL *rail, void *arg)
+{
+    EP3_TXD_RAIL      *txdRail  = (EP3_TXD_RAIL *) arg;
+    EP3_XMTR_RAIL     *xmtrRail = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    EP_XMTR          *xmtr     = xmtrRail->Generic.Xmtr;
+    int                delay   = 1;
+    EP_TXD           *txd;
+    unsigned long      flags;
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+
+    for (;;)
+    {
+       if (EP3_EVENT_FIRED (txdRail->DoneCookie, txdRail->TxdMain->DoneEvent) && 
+           EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent))
+           break;
+       
+       if (EP3_EVENT_FIRING (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent), txdRail->DoneCookie, txdRail->TxdMain->DoneEvent) && 
+           EP3_EVENT_FIRING (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent), txdRail->DataCookie, txdRail->TxdMain->DataEvent))
+       {
+           if (delay > EP3_EVENT_FIRING_TLIMIT)
+               panic ("TxDoneEvent: events set but block copy not completed\n");
+           DELAY(delay);
+           delay <<= 1;
+       }
+       else
+       {
+           EPRINTF3 (DBG_XMTR, "%s: TxDoneEvent: xmtr %p txdRail %p previously collecting by polling\n", 
+                     rail->Generic.Name, xmtr, txdRail);
+           spin_unlock_irqrestore (&xmtr->Lock, flags);
+           return;
+       }
+       mb();
+    }
+
+    if ((txd = txdRail->Generic.Txd) == NULL ||                                                 /* If there is no txd, or if the descriptor is marked */
+       !(EP_IS_INTERRUPT_ENABLED(txd->Envelope.Attr) || EP_IS_RPC(txd->Envelope.Attr))) /* marked as no interrupt, or been reused as an transmit, */
+    {                                                                                   /* then we were either called as a result of a previous */
+       spin_unlock_irqrestore (&xmtr->Lock, flags);                                     /* tx which was completed by polling or as a result */
+                                                                                        /* of a EnableTxCallBack/DisableTxCallback */
+
+       EPRINTF4 (DBG_XMTR, "%s: TxDoneEvent: xmtr %p txd %p recyled (%x)\n", 
+                 rail->Generic.Name, xmtr, txd, txd ? txd->Envelope.Attr : 0);
+       return; 
+    }
+
+    EPRINTF5 (DBG_XMTR, "%s: TxDoneEvent: xmtrRail=%p txdRail=%p txd=%p XID=%llx\n", 
+             rail->Generic.Name, xmtrRail, txdRail, txd, (long long) txd->Envelope.Xid.Unique);
+
+    ep_xmtr_txd_stat(xmtr,txd);
+
+    /* remove from active transmit list */
+    list_del (&txd->Link);
+    
+    UnbindTxdFromRail (txd, txdRail);
+    
+    /* clear the done flags for next time round */
+    txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+    txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+    txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+    
+    FreeTxdRail (xmtrRail, txdRail);
+
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+           
+    if (txd->Handler)
+       txd->Handler (txd, txd->Arg, EP_SUCCESS);
+       
+    FreeTxd (xmtr, txd);
+}
+
+/*
+ * TxDoneRetry: arg == EP3_TXD
+ */
+static void
+TxDoneRetry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int status)
+{
+    panic ("TxDoneRetry");
+}
+
+static void
+TxDoneVerify (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma)
+{
+    panic ("TxDoneVerify");
+}
+
+static void
+EnableTransmitCallback (EP_TXD *txd, EP3_TXD_RAIL *txdRail)
+{
+    ELAN3_DEV *dev = XMTR_TO_RAIL(txdRail->Generic.XmtrRail)->Device;
+
+    EPRINTF3 (DBG_XMTR, "%s: EnableTransmitCallback: txd %p txdRail %p\n", XMTR_TO_RAIL (txdRail->Generic.XmtrRail)->Generic.Name, txd, txdRail);
+
+    txd->Envelope.Attr = EP_SET_INTERRUPT_ENABLED(txd->Envelope.Attr);
+               
+    elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Type), EV_TYPE_BCOPY);
+       
+    if (EP_IS_RPC(txd->Envelope.Attr))
+    {
+       elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Type), EV_TYPE_BCOPY);
+       elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Type),  EV_TYPE_BCOPY | EV_TYPE_EVIRQ | txdRail->DoneCookie.Cookie);
+    }
+    else
+    {
+       elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Type), EV_TYPE_BCOPY | EV_TYPE_EVIRQ | txdRail->DataCookie.Cookie);
+       elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Type), EV_TYPE_BCOPY);
+    }
+}
+
+static void
+DisableTransmitCallback (EP_TXD *txd, EP3_TXD_RAIL *txdRail)
+{
+    ELAN3_DEV *dev = XMTR_TO_RAIL(txdRail->Generic.XmtrRail)->Device;
+
+    EPRINTF3 (DBG_XMTR, "%s: DisableTransmitCallback: txd %p txdRail %p\n", XMTR_TO_RAIL (txdRail->Generic.XmtrRail)->Generic.Name, txd, txdRail);
+
+    txd->Envelope.Attr = EP_CLEAR_INTERRUPT_ENABLED(txd->Envelope.Attr);
+
+    elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Type), EV_TYPE_BCOPY);
+    elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Type), EV_TYPE_BCOPY);
+    elan3_sdram_writel (dev, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Type), EV_TYPE_BCOPY);
+}
+
+static void
+InitialiseTxdRail (EP_TXD *txd, EP3_TXD_RAIL *txdRail, int phase)
+{
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    EP3_RAIL      *rail     = XMTR_TO_RAIL (xmtrRail);
+
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->Generic);
+
+    /* Initialise the per-rail fields in the envelope */
+    txd->Envelope.TxdRail = txdRail->TxdElanAddr;
+    txd->Envelope.NodeId  = rail->Generic.Position.pos_nodeid;
+
+    /* Initialise the dma backoff */
+    txdRail->Backoff.type = EP_BACKOFF_FREE;
+
+    /* Initialise the per-rail events */
+    switch (phase)
+    {
+    case EP_TXD_PHASE_ACTIVE:
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count), 1);
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count), 
+                           (txd->Envelope.nFrags ? txd->Envelope.nFrags : 1) + (EP_IS_MULTICAST(txd->Envelope.Attr) ? 1 : 0));
+       
+       txdRail->TxdMain->EnveEvent = EP3_EVENT_ACTIVE;
+       txdRail->TxdMain->DataEvent = EP3_EVENT_ACTIVE;
+       break;
+       
+    case EP_TXD_PHASE_PASSIVE:
+       ASSERT (EP_IS_RPC(txd->Envelope.Attr));
+
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count), 0);
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count), 0);
+
+       txdRail->TxdMain->EnveEvent = txdRail->EnveCookie.Cookie;
+       txdRail->TxdMain->DataEvent = txdRail->DataCookie.Cookie;
+       break;
+    }
+
+    if (! EP_IS_RPC(txd->Envelope.Attr))
+       txdRail->TxdMain->DoneEvent = txdRail->DoneCookie.Cookie;
+    else
+    {
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count), 1);
+       txdRail->TxdMain->DoneEvent = EP3_EVENT_ACTIVE;
+    }
+
+    if (EP_IS_NO_INTERRUPT(txd->Envelope.Attr))
+       DisableTransmitCallback (txd, txdRail);
+    else
+       EnableTransmitCallback (txd, txdRail);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    if ( epdebug_check_sum ) 
+       txd->Envelope.CheckSum = ep_calc_check_sum( txd->Xmtr->Subsys->Subsys.Sys, &txd->Envelope, txd->Envelope.Frags, txd->Envelope.nFrags);
+    else
+#endif
+       txd->Envelope.CheckSum = 0;  
+
+    /* copy the envelope and payload if present down to sdram */
+    elan3_sdram_copyl_to_sdram (rail->Device, &txd->Envelope, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, Envelope), EP_ENVELOPE_SIZE);
+    
+    if (EP_HAS_PAYLOAD(txd->Envelope.Attr))
+       elan3_sdram_copyl_to_sdram (rail->Device, &txd->Payload, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, Payload), EP_PAYLOAD_SIZE);
+}
+
+void
+ep3xmtr_flush_callback (EP_XMTR *xmtr, EP3_XMTR_RAIL *xmtrRail)
+{
+    EP3_RAIL *rail = XMTR_TO_RAIL (xmtrRail);
+    struct list_head *el;
+    unsigned long flags;
+
+    switch (rail->Generic.CallbackStep)
+    {
+    case EP_CB_FLUSH_FILTERING:
+       /* only need to acquire/release the Lock to ensure that
+        * the node state transition has been noticed. */
+       spin_lock_irqsave (&xmtr->Lock, flags);
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+       break;
+
+    case EP_CB_FLUSH_FLUSHING:
+       spin_lock_irqsave (&xmtr->Lock, flags);
+       
+       list_for_each (el, &xmtr->ActiveDescList) {
+           EP_TXD       *txd      = list_entry (el, EP_TXD, Link);
+           EP3_TXD_RAIL *txdRail  = (EP3_TXD_RAIL *) txd->TxdRail;
+           EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[txd->NodeId];
+           
+           if (!TXD_BOUND2RAIL(txdRail, xmtrRail) || nodeRail->State != EP_NODE_LOCAL_PASSIVATE)
+               continue;
+           
+           if (EP_IS_RPC(txd->Envelope.Attr))
+           {
+               if (! EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent))
+                   nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+               else if (! EP3_EVENT_FIRED (txdRail->DoneCookie, txdRail->TxdMain->DoneEvent))
+                   nodeRail->MessageState |= EP_NODE_PASSIVE_MESSAGES;
+           }
+           else
+           {
+               if (! EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent))
+                   nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+           }
+       }
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+       break;
+
+    default:
+       panic ("ep3xmtr_flush_callback: invalid callback step\n");
+       break;
+    }
+}
+
+void
+ep3xmtr_failover_callback (EP_XMTR *xmtr, EP3_XMTR_RAIL *xmtrRail)
+{
+    EP3_RAIL         *rail   = XMTR_TO_RAIL (xmtrRail);
+    struct list_head  txdList;
+    struct list_head *el, *nel;
+    unsigned long flags;
+#ifdef SUPPORT_RAIL_FAILOVER
+    EP_COMMS_SUBSYS  *subsys = xmtr->Subsys;
+#endif
+
+    INIT_LIST_HEAD (&txdList);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    list_for_each_safe (el, nel, &xmtr->ActiveDescList) {
+       EP_TXD       *txd       = list_entry (el, EP_TXD, Link);
+       EP3_TXD_RAIL *txdRail   = (EP3_TXD_RAIL *) txd->TxdRail;
+       EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[txd->NodeId];
+           
+       /* Only progress relocation of txd's bound to this rail */
+       if (!TXD_BOUND2RAIL(txdRail, xmtrRail) || nodeRail->State != EP_NODE_PASSIVATED)
+           continue;
+       
+#ifdef SUPPORT_RAIL_FAILOVER
+       /* Transmit data not been sent, so just restart on different rail */
+       if (! EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent))
+       {
+           EPRINTF4 (DBG_XMTR, "%s: ep3xmtr_failover_callback - xmtr %p txd %p node %d unbind an retry\n", rail->Generic.Name, xmtr, txd, txd->NodeId);
+           
+           UnbindTxdFromRail (txd, txdRail);
+           
+           /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+           txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+           txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+           txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+
+           /* reset all events, since non of them could have been set */
+           elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count), 0);                          /* PCI write */
+           elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count), 0);                          /* PCI write */
+           elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count), 0);                          /* PCI write */
+           
+           FreeTxdRail (xmtrRail, txdRail);
+           
+           /* epcomms thread will restart on different rail */
+           ep_kthread_schedule (&subsys->Thread, lbolt);
+           continue;
+       }
+
+       if (EP_IS_RPC(txd->Envelope.Attr) && !EP3_EVENT_FIRED (txdRail->DoneCookie, txdRail->TxdMain->DoneEvent))
+       {
+           if (EP_IS_NO_FAILOVER(txd->Envelope.Attr))
+           {
+               EPRINTF4 (DBG_XMTR, "%s: ep3xmtr_failover_callback - xmtr %p txd %p node %d - not able to failover\n",
+                         rail->Generic.Name, xmtr, txd, txd->NodeId);
+
+               list_del (&txd->Link);
+               UnbindTxdFromRail (txd, txdRail);
+               
+               /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+               txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+               txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+               txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+               
+               /* envelope and data events must have been set, so only clear the done event */
+               EP_ASSERT (&rail->Generic, SDRAM_ASSERT(elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count)) == 0));
+               EP_ASSERT (&rail->Generic, SDRAM_ASSERT(elan3_sdram_readl (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)) == 0));
+
+               elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count), 0);                              /* PCI write */
+               
+               FreeTxdRail (xmtrRail, txdRail);
+           
+               list_add_tail (&txd->Link, &txdList);
+               continue;
+           }
+           EPRINTF4 (DBG_XMTR, "%s: ep3xmtr_failover_callback - xmtr %p txd %p node %d passive\n", rail->Generic.Name, xmtr, txd, txd->NodeId);
+           
+           nodeRail->MessageState |= EP_NODE_PASSIVE_MESSAGES;
+           continue;
+       }
+
+       EPRINTF4 (DBG_XMTR, "%s: ep3xmtr_failover_callback - xmtr %p txd %p node %d completed\n", rail->Generic.Name, xmtr, txd, txd->NodeId);
+#endif
+
+    }
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    while (! list_empty (&txdList)) 
+    {
+       EP_TXD *txd = list_entry (txdList.next, EP_TXD, Link);
+
+       list_del (&txd->Link);
+
+       txd->Handler (txd, txd->Arg, EP_CONN_RESET);
+       
+       FreeTxd (xmtr, txd);
+    }
+}
+
+
+void
+ep3xmtr_disconnect_callback (EP_XMTR *xmtr, EP3_XMTR_RAIL *xmtrRail)
+{
+    EP3_RAIL         *rail = XMTR_TO_RAIL (xmtrRail);
+    struct list_head *el, *nel;
+    struct list_head  txdList;
+    unsigned long flags;
+    
+    INIT_LIST_HEAD (&txdList);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+
+    list_for_each_safe (el, nel, &xmtr->ActiveDescList) {
+       EP_TXD       *txd       = list_entry (el, EP_TXD, Link);
+       EP3_TXD_RAIL *txdRail   = (EP3_TXD_RAIL *) txd->TxdRail;
+       EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[txd->NodeId];
+           
+       if (!TXD_BOUND2RAIL(txdRail, xmtrRail) || nodeRail->State != EP_NODE_DISCONNECTING)
+           continue;
+       
+       if (EP3_EVENT_FIRED (txdRail->EnveCookie, txdRail->TxdMain->EnveEvent) &&
+           EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent) &&
+           EP3_EVENT_FIRED (txdRail->DoneCookie, txdRail->TxdMain->DoneEvent))
+       {
+           EPRINTF4 (DBG_XMTR, "%s: ep3xmtr_disconnect_callback - xmtr %p txd %p completed to node %d\n", rail->Generic.Name, xmtr, txd, txd->NodeId);
+           continue;
+       }
+
+       /* Remove from active list */
+       list_del (&txd->Link);
+       
+       UnbindTxdFromRail (txd, txdRail);
+       
+       /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+       txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+       txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+       txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+       
+       /* reset the envelope and data events, since only they could have been set */
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count), 0);                              /* PCI write */
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count), 0);                              /* PCI write */
+       elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count), 0);                              /* PCI write */
+       
+       FreeTxdRail (xmtrRail, txdRail);
+           
+       EPRINTF4 (DBG_XMTR, "%s: ep3xmtr_disconnect_callback - xmtr %p txd %p node %d not conected\n", rail->Generic.Name, xmtr, txd, txd->NodeId);
+
+       /* add to the list of txd's which are to be completed */
+       list_add_tail (&txd->Link, &txdList);
+    }
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+    
+    while (! list_empty (&txdList)) 
+    {
+       EP_TXD *txd = list_entry (txdList.next, EP_TXD, Link);
+
+       list_del (&txd->Link);
+
+       txd->Handler (txd, txd->Arg, EP_CONN_RESET);
+       
+       FreeTxd (xmtr, txd);
+    }
+}
+
+int
+ep3xmtr_poll_txd (EP_XMTR_RAIL *x, EP_TXD_RAIL *t, int how)
+{
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) x;
+    EP3_TXD_RAIL  *txdRail  = (EP3_TXD_RAIL *) t;
+    EP_TXD        *txd      = txdRail->Generic.Txd;
+
+    switch (how)
+    {
+    case ENABLE_TX_CALLBACK:
+       if (EP_IS_NO_INTERRUPT(txd->Envelope.Attr))
+           EnableTransmitCallback (txd, txdRail);
+       break;
+
+    case DISABLE_TX_CALLBACK:
+       if (EP_IS_NO_INTERRUPT(txd->Envelope.Attr))
+           DisableTransmitCallback (txd, txdRail);
+       break;
+    }
+
+    if (EP3_EVENT_FIRED (txdRail->EnveCookie, txdRail->TxdMain->EnveEvent) &&
+       EP3_EVENT_FIRED (txdRail->DataCookie, txdRail->TxdMain->DataEvent) &&
+       EP3_EVENT_FIRED (txdRail->DoneCookie, txdRail->TxdMain->DoneEvent))
+    {
+       EPRINTF3 (DBG_XMTR, "%s: ep3xmtr_poll_txd: txd=%p XID=%llx completed\n", 
+                 XMTR_TO_RAIL (xmtrRail)->Generic.Name, txd, (long long) txd->Envelope.Xid.Unique);
+
+       ep_xmtr_txd_stat(xmtrRail->Generic.Xmtr,txd);
+
+       UnbindTxdFromRail (txd, txdRail);
+       
+       /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+       txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+       txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+       txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+       
+       FreeTxdRail (xmtrRail, txdRail);
+
+       return 1;
+    }
+
+    return 0;
+}
+
+int
+ep3xmtr_bind_txd (EP_TXD *txd, EP_XMTR_RAIL *x, unsigned int phase)
+{
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) x;
+    EP3_RAIL      *rail     = XMTR_TO_RAIL (xmtrRail);
+    EP3_TXD_RAIL  *txdRail;
+    E3_DMA_BE      dmabe;
+
+    if ((txdRail = GetTxdRail (xmtrRail)) == NULL)
+       return 0;
+
+    switch (phase)
+    {
+    case EP_TXD_PHASE_ACTIVE:
+       if (rail->Generic.Nodes[txd->NodeId].State != EP_NODE_CONNECTED)
+       {
+           EPRINTF2 (DBG_XMTR, "%s: TransmitTxdOnRail: node %u not connected on this rail\n", rail->Generic.Name, txd->NodeId);
+
+           /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+           txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+           txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+           txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+
+           /* reset all events, since non of them could have been set */
+           elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count), 0);                          /* PCI write */
+           elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count), 0);                          /* PCI write */
+           elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count), 0);                          /* PCI write */
+
+           FreeTxdRail (xmtrRail, txdRail);
+           return 0;
+       }
+
+       InitialiseTxdRail (txd, txdRail, phase);
+
+       /* Initialise the dma descriptor */
+       dmabe.s.dma_type            = E3_DMA_TYPE (DMA_BYTE, DMA_WRITE, DMA_QUEUED, EP3_DMAFAILCOUNT);
+       dmabe.s.dma_size            = (EP_HAS_PAYLOAD(txd->Envelope.Attr) ? EP_INPUTQ_SIZE : EP_ENVELOPE_SIZE);
+       dmabe.s.dma_source          = txdRail->TxdElanAddr + offsetof (EP3_TXD_RAIL_ELAN, Envelope);
+       dmabe.s.dma_dest            = (E3_Addr) 0;
+       dmabe.s.dma_destEvent       = EP_MSGQ_ADDR(txd->Service);
+       dmabe.s.dma_destCookieVProc = EP_VP_DATA (txd->NodeId);
+       dmabe.s.dma_srcEvent        = txdRail->TxdElanAddr + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent);
+       dmabe.s.dma_srcCookieVProc  = LocalCookie (rail, txd->NodeId);
+
+       EPRINTF8 (DBG_XMTR, "%s: TransmitTxdOnRail: txd=%p txdRail=%p @ %x XID=%llx dest=%u srcEvent=%x srcCookie=%x\n", rail->Generic.Name, 
+                 txd, txdRail, txdRail->TxdElanAddr, (long long) txd->Envelope.Xid.Unique, txd->NodeId, dmabe.s.dma_srcEvent, dmabe.s.dma_srcCookieVProc);
+       
+       BindTxdToRail (txd, txdRail);
+       
+       if (IssueDma (rail, &dmabe, EP_RETRY_LOW_PRI, FALSE) != ISSUE_COMMAND_OK)
+           QueueDmaForRetry (rail, &dmabe, EP_RETRY_LOW_PRI);
+       break;
+
+    case EP_TXD_PHASE_PASSIVE:
+       InitialiseTxdRail (txd, txdRail, EP_TXD_PHASE_PASSIVE);                         /* initialise as passive (updated envelope) */
+       
+       EP_XMTR_OP (txd->TxdRail->XmtrRail, UnbindTxd) (txd, EP_TXD_PHASE_PASSIVE);     /* unbind from existing rail */
+
+       BindTxdToRail (txd, txdRail);                                                   /* and bind it to our new rail */
+       break;
+    }
+
+    return 1;
+}
+
+void
+ep3xmtr_unbind_txd (EP_TXD *txd, unsigned int phase)
+{
+    EP3_TXD_RAIL  *txdRail  = (EP3_TXD_RAIL *) txd->TxdRail;
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    EP3_RAIL      *rail     = XMTR_TO_RAIL (xmtrRail);
+
+    /* XXXX - TBD assertions on phase */
+
+    UnbindTxdFromRail (txd, txdRail);
+    
+    /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+    txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+    txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+    txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+    
+    /* reset the envelope and data events, since only they could have been set */
+    elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count), 0);                         /* PCI write */
+    elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count), 0);                         /* PCI write */
+    elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count), 0);                         /* PCI write */         
+    
+    FreeTxdRail (xmtrRail, txdRail);
+}
+
+long
+ep3xmtr_check (EP_XMTR_RAIL *x, long nextRunTime)
+{
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) x;
+
+    if (xmtrRail->FreeDescCount < ep_txd_lowat && !AllocateTxdRailBlock(xmtrRail))
+    {
+       EPRINTF1 (DBG_RCVR,"%s: failed to grow txd rail pool\n", XMTR_TO_RAIL(xmtrRail)->Generic.Name);
+               
+       if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + RESOURCE_RETRY_TIME))
+           nextRunTime = lbolt + RESOURCE_RETRY_TIME;
+    }
+    
+    return nextRunTime;
+}
+
+void
+ep3xmtr_add_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail)
+{
+    EP3_XMTR_RAIL *xmtrRail;
+    unsigned long  flags;
+
+    KMEM_ZALLOC (xmtrRail, EP3_XMTR_RAIL *, sizeof (EP3_XMTR_RAIL), 1);
+
+    spin_lock_init (&xmtrRail->FreeDescLock);
+    kcondvar_init  (&xmtrRail->FreeDescSleep);
+    INIT_LIST_HEAD (&xmtrRail->FreeDescList);
+    INIT_LIST_HEAD (&xmtrRail->DescBlockList);
+
+    xmtrRail->Generic.CommsRail = commsRail;
+    xmtrRail->Generic.Xmtr      = xmtr;
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+
+    xmtr->Rails[commsRail->Rail->Number] = &xmtrRail->Generic;
+    xmtr->RailMask |= EP_RAIL2RAILMASK(commsRail->Rail->Number);
+
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+}
+
+void
+ep3xmtr_del_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail)
+{
+    EP3_RAIL         *rail     = (EP3_RAIL *) commsRail->Rail;
+    EP3_XMTR_RAIL    *xmtrRail = (EP3_XMTR_RAIL *) xmtr->Rails[commsRail->Rail->Number];
+    unsigned long     flags;
+
+    /* rail mask set as not usable */
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    xmtr->RailMask &= ~EP_RAIL2RAILMASK (rail->Generic.Number);
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    /* wait for all txd's for this rail to become free */
+    spin_lock_irqsave (&xmtrRail->FreeDescLock, flags);
+    while (xmtrRail->FreeDescCount != xmtrRail->TotalDescCount)
+    {
+       xmtrRail->FreeDescWaiting++;
+       kcondvar_wait (&xmtrRail->FreeDescSleep, &xmtrRail->FreeDescLock, &flags);
+    }
+    spin_unlock_irqrestore (&xmtrRail->FreeDescLock, flags);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    xmtr->Rails[commsRail->Rail->Number] = NULL;
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    /* need to free up the txd's and blocks */
+    /* all the txd's accociated with DescBlocks must be in the FreeDescList */
+    ASSERT (xmtrRail->TotalDescCount == xmtrRail->FreeDescCount);
+
+    /* run through the DescBlockList deleting them */
+    while (!list_empty (&xmtrRail->DescBlockList))
+       FreeTxdRailBlock (xmtrRail, list_entry(xmtrRail->DescBlockList.next, EP3_TXD_RAIL_BLOCK , Link));
+    
+    /* it had better be empty after that */
+    ASSERT ((xmtrRail->FreeDescCount == 0) && (xmtrRail->TotalDescCount == 0));
+
+    spin_lock_destroy (&xmtrRail->FreeDescLock);
+    kcondvar_destroy (&xmtrRail->FreeDescSleep);
+
+    KMEM_FREE (xmtrRail, sizeof (EP3_XMTR_RAIL));
+}
+
+void
+ep3xmtr_display_xmtr (DisplayInfo *di, EP_XMTR_RAIL *x)
+{
+    EP3_XMTR_RAIL *xmtrRail = (EP3_XMTR_RAIL *) x;
+    EP3_RAIL      *rail     = XMTR_TO_RAIL (xmtrRail);
+    struct list_head *el;
+    unsigned long flags;
+    int freeCount = 0;
+
+    spin_lock_irqsave (&xmtrRail->FreeDescLock, flags);
+    list_for_each (el, &xmtrRail->FreeDescList)
+       freeCount++;
+    spin_unlock_irqrestore (&xmtrRail->FreeDescLock, flags);
+
+    (di->func)(di->arg, "                 Rail=%d Free=%d Total=%d (%d)\n",
+               rail->Generic.Number, xmtrRail->FreeDescCount, xmtrRail->TotalDescCount, freeCount);
+}
+
+void
+ep3xmtr_display_txd (DisplayInfo *di, EP_TXD_RAIL *t)
+{
+    EP3_TXD_RAIL      *txdRail   = (EP3_TXD_RAIL *) t;
+    EP3_XMTR_RAIL     *xmtrRail  = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    EP3_TXD_RAIL_MAIN *txdMain   = txdRail->TxdMain;
+    sdramaddr_t        txdElan   = txdRail->TxdElan;
+    EP3_RAIL          *rail      = (EP3_RAIL *) xmtrRail->Generic.CommsRail->Rail;
+    ELAN3_DEV         *dev       = rail->Device;
+    
+    (di->func)(di->arg, "     EnveEvent=%x DataEvent=%x DoneEvent=%x Rail=%s\n", 
+              txdMain->EnveEvent, txdMain->DataEvent, txdMain->DoneEvent, rail->Generic.Name);
+    (di->func)(di->arg, "     EnveEvent=%x.%x DataEvent=%x.%x DoneEvent=%x.%x\n",
+              elan3_sdram_readl (dev, txdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count)),
+              elan3_sdram_readl (dev, txdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Type)),
+              elan3_sdram_readl (dev, txdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count)),
+              elan3_sdram_readl (dev, txdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Type)),
+              elan3_sdram_readl (dev, txdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count)),
+              elan3_sdram_readl (dev, txdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Type)));
+}
+
+int
+ep3xmtr_check_txd_state (EP_TXD *txd)
+{
+    EP3_TXD_RAIL  *txdRail   = (EP3_TXD_RAIL *) txd->TxdRail;
+    EP3_XMTR_RAIL *xmtrRail  = (EP3_XMTR_RAIL *) txdRail->Generic.XmtrRail;
+    EP3_RAIL      *rail      = XMTR_TO_RAIL (xmtrRail);
+    E3_Addr        enveEvent = txdRail->TxdElanAddr + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent);
+    EP3_RETRY_DMA *retry = NULL;
+
+    struct list_head *el;
+    struct list_head *nel;
+    unsigned long     flags;
+
+    /*  is enevelope event is really not set */
+    if (EP3_EVENT_FIRED (txdRail->EnveCookie, txdRail->TxdMain->EnveEvent )) 
+       return (0);
+    
+    /* remove matching dma from stalled list */            
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    
+    list_for_each_safe(el, nel,  &rail->DmaRetries[EP_RETRY_STABALISING]) {
+       retry = list_entry (el, EP3_RETRY_DMA, Link);
+       
+       if ( retry->Dma.s.dma_srcEvent == enveEvent ) {
+           /* remove from retry list */
+           list_del (&retry->Link);
+           break; /* there can only be one */
+       } 
+    }
+    ASSERT ( retry != NULL); /* must find one in list */
+    ASSERT ( retry->Dma.s.dma_srcEvent == enveEvent ); /* better still be the right type then */    
+
+    /* add to free list */
+    list_add (&retry->Link, &rail->DmaRetryFreeList);
+
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);    
+                       
+    UnbindTxdFromRail (txd, txdRail);
+       
+    /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+    txdRail->TxdMain->EnveEvent = EP3_EVENT_PRIVATE;
+    txdRail->TxdMain->DataEvent = EP3_EVENT_PRIVATE;
+    txdRail->TxdMain->DoneEvent = EP3_EVENT_PRIVATE;
+    
+    /* reset the envelope and data events, since only they could have been set */
+    elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, EnveEvent.ev_Count), 0);                         /* PCI write */
+    elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DataEvent.ev_Count), 0);                         /* PCI write */
+    elan3_sdram_writel (rail->Device, txdRail->TxdElan + offsetof (EP3_TXD_RAIL_ELAN, DoneEvent.ev_Count), 0);                         /* PCI write */         
+    
+    FreeTxdRail (xmtrRail, txdRail);
+
+    return (1);
+}
+
+void
+ep3xmtr_fillout_rail_stats(EP_XMTR_RAIL *xmtr_rail, char *str) {
+    /* no stats here yet */
+    /* EP3_XMTR_RAIL * ep3xmtr_rail = (EP3_XMTR_RAIL *) xmtr_rail; */
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/epcommsTx_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/epcommsTx_elan4.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/epcommsTx_elan4.c 2005-06-01 23:12:54.659429984 -0400
@@ -0,0 +1,1389 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: epcommsTx_elan4.c,v 1.26.2.4 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcommsTx_elan4.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "debug.h"
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "epcomms_elan4.h"
+
+#include <elan4/trtype.h>
+
+#define XMTR_TO_COMMS(xmtrRail)                ((EP4_COMMS_RAIL *) ((EP_XMTR_RAIL *) xmtrRail)->CommsRail)
+#define XMTR_TO_RAIL(xmtrRail)         ((EP4_RAIL *) ((EP_XMTR_RAIL *) xmtrRail)->CommsRail->Rail)
+#define XMTR_TO_DEV(xmtrRail)          (XMTR_TO_RAIL(xmtrRail)->r_ctxt.ctxt_dev)
+#define XMTR_TO_SUBSYS(xmtrRail)       (((EP_XMTR_RAIL *) xmtrRail)->Xmtr->Subsys)
+
+#define TXD_TO_XMTR(txdRail)           ((EP4_XMTR_RAIL *) txdRail->txd_generic.XmtrRail)
+#define TXD_TO_RAIL(txdRail)           XMTR_TO_RAIL(TXD_TO_XMTR(txdRail))
+
+static void txd_interrupt (EP4_RAIL *rail, void *arg);
+static void poll_interrupt (EP4_RAIL *rail, void *arg);
+
+static __inline__ int
+on_list (struct list_head *ent, struct list_head *list)
+{
+    struct list_head *el;
+    unsigned int count = 0;
+    list_for_each (el, list) {
+       if (el == ent)
+           count++;
+    }
+    return count;
+}
+
+static __inline__ void
+__ep4_txd_assert_free (EP4_TXD_RAIL *txdRail, const char *file, const int line)
+{
+    EP4_XMTR_RAIL *xmtrRail = TXD_TO_XMTR (txdRail);
+    ELAN4_DEV     *dev      = XMTR_TO_DEV (xmtrRail);
+    register int   failed   = 0;
+    
+    if ((txdRail)->txd_retry_time     != 0)              failed |= (1 << 0);
+    if ((txdRail)->txd_main->txd_env  != EP4_STATE_FREE) failed |= (1 << 1);
+    if ((txdRail)->txd_main->txd_data != EP4_STATE_FREE) failed |= (1 << 2);
+    if ((txdRail)->txd_main->txd_done != EP4_STATE_FREE) failed |= (1 << 3);
+
+    if (sdram_assert)
+    {
+       if ((int)(elan4_sdram_readq (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType)) >> 32)  != -32) failed |= (1 << 4);
+       if ((int)(elan4_sdram_readq (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType)) >> 32) != 0)   failed |= (1 << 5);
+       if ((int)(elan4_sdram_readq (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType)) >> 32) != 0)   failed |= (1 << 6);
+    }
+
+    if (failed)
+    {
+       printk ("__ep4_txd_assert_free: failed=%x txdRail=%p at %s:%d\n", failed, txdRail, file, line);
+
+       ep_debugf (DBG_DEBUG, "__ep4_txd_assert_free: failed=%x txdRail=%p at %s:%d\n", failed, txdRail, file, line);
+       ep4xmtr_display_txd (&di_ep_debug, &txdRail->txd_generic);
+
+       (txdRail)->txd_retry_time     = 0;
+       (txdRail)->txd_main->txd_env  = EP4_STATE_FREE;
+       (txdRail)->txd_main->txd_data = EP4_STATE_FREE;
+       (txdRail)->txd_main->txd_done = EP4_STATE_FREE;
+
+       if (sdram_assert)
+       {
+           elan4_sdram_writel (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType)  + 4, -32);
+           elan4_sdram_writel (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType) + 4, 0);
+           elan4_sdram_writel (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType) + 4, 0);
+       }
+       EP_ASSFAIL (XMTR_TO_RAIL(xmtrRail), "__ep4_txd_assert_free");
+    }
+}
+
+static __inline__ void
+__ep4_txd_assert_finished (EP4_TXD_RAIL *txdRail, const char *file, const int line)
+{
+    EP4_XMTR_RAIL *xmtrRail = TXD_TO_XMTR (txdRail);
+    ELAN4_DEV     *dev      = XMTR_TO_DEV (xmtrRail);
+    register int   failed   = 0;
+    
+    if ((txdRail)->txd_retry_time     != 0)                  failed |= (1 << 0);
+    if ((txdRail)->txd_main->txd_env  != EP4_STATE_FINISHED) failed |= (1 << 1);
+    if ((txdRail)->txd_main->txd_data != EP4_STATE_FINISHED) failed |= (1 << 2);
+    if ((txdRail)->txd_main->txd_done != EP4_STATE_FINISHED) failed |= (1 << 3);
+    
+    if (sdram_assert)
+    {
+       if ((int)(elan4_sdram_readq (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType)) >> 32)  != -32) failed |= (1 << 4);
+       if ((int)(elan4_sdram_readq (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType)) >> 32) != 0)   failed |= (1 << 5);
+       if ((int)(elan4_sdram_readq (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType)) >> 32) != 0)   failed |= (1 << 6);
+    }
+
+    if (failed)
+    {
+       printk ("__ep4_txd_assert_finished: failed=%x txdRail=%p at %s:%d\n", failed, txdRail, file, line);
+
+       ep_debugf (DBG_DEBUG, "__ep4_txd_assert_finished: failed=%x txdRail=%p at %s:%d\n", failed, txdRail, file, line);
+       ep4xmtr_display_txd (&di_ep_debug, &txdRail->txd_generic);
+
+       (txdRail)->txd_retry_time     = 0;
+       (txdRail)->txd_main->txd_env  = EP4_STATE_FINISHED;
+       (txdRail)->txd_main->txd_data = EP4_STATE_FINISHED;
+       (txdRail)->txd_main->txd_done = EP4_STATE_FINISHED;
+
+       if (sdram_assert)
+       {
+           elan4_sdram_writel (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType)  + 4, -32);
+           elan4_sdram_writel (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType) + 4, 0);
+           elan4_sdram_writel (dev, (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType) + 4, 0);
+       }
+       EP_ASSFAIL (XMTR_TO_RAIL(xmtrRail), "__ep4_txd_assert_finished");
+    }
+}
+
+static __inline__ int
+__ep4_txd_assfail (EP4_TXD_RAIL *txdRail, const char *expr, const char *file, const int line)
+{
+    EP4_XMTR_RAIL *xmtrRail = TXD_TO_XMTR (txdRail);
+
+    printk ("__ep4_txd_assfail: %s:%d '%s'\n", file, line, expr);
+
+    ep_debugf (DBG_DEBUG, "__ep4_txd_assfail: %s:%d '%s'\n", file, line, expr);
+    ep4xmtr_display_txd (&di_ep_debug, &txdRail->txd_generic);
+
+    EP_ASSFAIL (XMTR_TO_RAIL (xmtrRail), "__ep4_txd_assfail");
+
+    return 0;
+}
+
+#define EP4_TXD_ASSERT(txdRail, EX)            ((void) ((EX) || (__ep4_txd_assfail(txdRail, #EX, __FILE__, __LINE__))))
+#define EP4_TXD_ASSERT_FREE(txdRail)           __ep4_txd_assert_free(txdRail, __FILE__, __LINE__)
+#define EP4_TXD_ASSERT_FINISHED(txdRail)       __ep4_txd_assert_finished(txdRail, __FILE__, __LINE__)
+
+static int
+alloc_txd_block (EP4_XMTR_RAIL *xmtrRail)
+{
+    EP4_RAIL           *rail = XMTR_TO_RAIL(xmtrRail);
+    ELAN4_DEV          *dev  = XMTR_TO_DEV(xmtrRail);
+    EP4_TXD_RAIL_BLOCK *blk;
+    EP4_TXD_RAIL_MAIN  *txdMain;
+    EP_ADDR            txdMainAddr;
+    sdramaddr_t                txdElan;
+    EP_ADDR            txdElanAddr;
+    EP4_TXD_RAIL       *txdRail;
+    unsigned long       flags;
+    int                 i;
+
+    KMEM_ZALLOC (blk, EP4_TXD_RAIL_BLOCK *, sizeof (EP4_TXD_RAIL_BLOCK), 1);
+
+    if (blk == NULL)
+       return 0;
+
+    if ((txdElan = ep_alloc_elan (&rail->r_generic, EP4_TXD_RAIL_ELAN_SIZE * EP4_NUM_TXD_PER_BLOCK, 0, &txdElanAddr)) == (sdramaddr_t) 0)
+    {
+       KMEM_FREE (blk, sizeof (EP4_TXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    if ((txdMain = ep_alloc_main (&rail->r_generic, EP4_TXD_RAIL_MAIN_SIZE * EP4_NUM_TXD_PER_BLOCK, 0, &txdMainAddr)) == (EP4_TXD_RAIL_MAIN *) NULL)
+    {
+       ep_free_elan (&rail->r_generic, txdElanAddr, EP4_TXD_RAIL_ELAN_SIZE * EP4_NUM_TXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP4_TXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    if (ep4_reserve_dma_retries (rail, EP4_NUM_TXD_PER_BLOCK, 0) != 0)
+    {
+       ep_free_main (&rail->r_generic, blk->blk_txds[0].txd_main_addr, EP4_TXD_RAIL_MAIN_SIZE * EP4_NUM_TXD_PER_BLOCK);
+       ep_free_elan (&rail->r_generic, txdElanAddr, EP4_TXD_RAIL_ELAN_SIZE * EP4_NUM_TXD_PER_BLOCK);
+       KMEM_FREE (blk, sizeof (EP4_TXD_RAIL_BLOCK));
+       return 0;
+    }
+
+    for (txdRail = &blk->blk_txds[0], i = 0; i < EP4_NUM_TXD_PER_BLOCK; i++, txdRail++)
+    {
+       txdRail->txd_generic.XmtrRail = &xmtrRail->xmtr_generic;
+       txdRail->txd_elan             = txdElan;
+       txdRail->txd_elan_addr        = txdElanAddr;
+       txdRail->txd_main             = txdMain;
+       txdRail->txd_main_addr        = txdMainAddr;
+
+       /* We only need to reserve space for one command stream, since the sten packet
+        * can only be retrying *before* the dma source event is set.
+        * reserve bytes of "event" cq space for the completion write + interrupt */
+       if ((txdRail->txd_ecq = ep4_get_ecq (rail, EP4_ECQ_EVENT, EP4_INTR_CMD_NDWORDS)) == NULL)
+           goto failed;
+
+       /* register the main interrupt cookies */
+       ep4_register_intcookie (rail, &txdRail->txd_intcookie, txdElanAddr + offsetof (EP4_TXD_RAIL_ELAN, txd_done), txd_interrupt, txdRail);
+
+       /* initialise the events */
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CopySource),
+                           txdElanAddr + offsetof (EP4_TXD_RAIL_ELAN, txd_env_cmd));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CopyDest),
+                           txdRail->txd_ecq->ecq_addr);
+
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_WritePtr),
+                           txdMainAddr + offsetof (EP4_TXD_RAIL_MAIN, txd_data));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_WriteValue),
+                           EP4_STATE_FINISHED);
+
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CopySource),
+                           txdElanAddr + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CopyDest),
+                           txdRail->txd_ecq->ecq_addr);
+
+       /* Initialise the command streams */
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env_cmd.c_write_cmd),
+                           WRITE_DWORD_CMD | (txdMainAddr + offsetof (EP4_TXD_RAIL_MAIN, txd_env)));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env_cmd.c_write_value),
+                           EP4_STATE_FAILED);
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env_cmd.c_intr_cmd),
+                           INTERRUPT_CMD | (txdRail->txd_intcookie.int_val << E4_MAIN_INT_SHIFT));
+
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd.c_write_cmd),
+                           WRITE_DWORD_CMD | (txdMainAddr + offsetof (EP4_TXD_RAIL_MAIN, txd_done)));
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd.c_write_value),
+                           EP4_STATE_FINISHED);
+       elan4_sdram_writeq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd.c_intr_cmd),
+                           INTERRUPT_CMD | (txdRail->txd_intcookie.int_val << E4_MAIN_INT_SHIFT));
+
+       txdMain->txd_env  = EP4_STATE_FREE;
+       txdMain->txd_data = EP4_STATE_FREE;
+       txdMain->txd_done = EP4_STATE_FREE;
+
+       /* move onto next descriptor */
+       txdElan     += EP4_TXD_RAIL_ELAN_SIZE;
+       txdElanAddr += EP4_TXD_RAIL_ELAN_SIZE;
+       txdMain      = (EP4_TXD_RAIL_MAIN *) ((unsigned long) txdMain + EP4_TXD_RAIL_MAIN_SIZE);
+       txdMainAddr += EP4_TXD_RAIL_MAIN_SIZE;
+    }
+
+    spin_lock_irqsave (&xmtrRail->xmtr_freelock, flags);
+
+    list_add  (&blk->blk_link, &xmtrRail->xmtr_blocklist);
+
+    xmtrRail->xmtr_totalcount += EP4_NUM_TXD_PER_BLOCK;
+    xmtrRail->xmtr_freecount  += EP4_NUM_TXD_PER_BLOCK;
+
+    for (i = 0; i < EP4_NUM_TXD_PER_BLOCK; i++)
+       list_add (&blk->blk_txds[i].txd_generic.Link, &xmtrRail->xmtr_freelist);
+
+    spin_unlock_irqrestore (&xmtrRail->xmtr_freelock, flags);
+
+    return 1;
+
+ failed:
+    while (--i >= 0)
+    {
+       ep4_put_ecq (rail, txdRail->txd_ecq, EP4_INTR_CMD_NDWORDS);
+       ep4_deregister_intcookie (rail, &txdRail->txd_intcookie);
+    }
+    ep4_release_dma_retries (rail, EP4_NUM_TXD_PER_BLOCK);
+
+    ep_free_main (&rail->r_generic, blk->blk_txds[0].txd_main_addr, EP4_TXD_RAIL_MAIN_SIZE * EP4_NUM_TXD_PER_BLOCK);
+    ep_free_elan (&rail->r_generic, blk->blk_txds[0].txd_elan_addr, EP4_TXD_RAIL_ELAN_SIZE * EP4_NUM_TXD_PER_BLOCK);
+
+    KMEM_FREE (blk, sizeof (EP4_TXD_RAIL_BLOCK));
+
+    return 0;
+}
+
+static void
+free_txd_block (EP4_XMTR_RAIL *xmtrRail, EP4_TXD_RAIL_BLOCK *blk)
+{
+    EP4_RAIL     *rail = XMTR_TO_RAIL (xmtrRail);
+    EP4_TXD_RAIL *txdRail;
+    unsigned long flags;
+    int           i;
+
+    spin_lock_irqsave (&xmtrRail->xmtr_freelock, flags);
+
+    list_del (&blk->blk_link);
+
+    xmtrRail->xmtr_totalcount -= EP4_NUM_TXD_PER_BLOCK;
+
+    for (txdRail = &blk->blk_txds[0], i = 0; i < EP4_NUM_TXD_PER_BLOCK; i++, txdRail++)
+    {
+       xmtrRail->xmtr_freecount--;
+
+       ep4_put_ecq (rail, txdRail->txd_ecq, EP4_INTR_CMD_NDWORDS);
+
+       ep4_deregister_intcookie (rail, &txdRail->txd_intcookie);
+
+       list_del (&txdRail->txd_generic.Link);
+    }
+    spin_unlock_irqrestore (&xmtrRail->xmtr_freelock, flags);
+
+    ep4_release_dma_retries (rail, EP4_NUM_TXD_PER_BLOCK);
+
+    ep_free_main (&rail->r_generic, blk->blk_txds[0].txd_main_addr, EP4_TXD_RAIL_MAIN_SIZE * EP4_NUM_TXD_PER_BLOCK);
+    ep_free_elan (&rail->r_generic, blk->blk_txds[0].txd_elan_addr, EP4_TXD_RAIL_ELAN_SIZE * EP4_NUM_TXD_PER_BLOCK);
+
+    KMEM_FREE (blk, sizeof (EP4_TXD_RAIL_BLOCK));
+}
+
+static EP4_TXD_RAIL *
+get_txd_rail (EP4_XMTR_RAIL *xmtrRail)
+{
+    EP_COMMS_SUBSYS  *subsys = XMTR_TO_SUBSYS(xmtrRail);
+    EP4_TXD_RAIL     *txdRail;
+    unsigned long flags;
+    int low_on_txds;
+
+    spin_lock_irqsave (&xmtrRail->xmtr_freelock, flags);
+
+    if (list_empty (&xmtrRail->xmtr_freelist))
+       txdRail = NULL;
+    else
+    {
+       txdRail = list_entry (xmtrRail->xmtr_freelist.next, EP4_TXD_RAIL, txd_generic.Link);
+
+       EP4_TXD_ASSERT_FREE(txdRail);
+
+       list_del (&txdRail->txd_generic.Link);
+
+       xmtrRail->xmtr_freecount--;
+    }
+    /* Wakeup the descriptor primer thread if there's not many left */
+    low_on_txds = (xmtrRail->xmtr_freecount < ep_txd_lowat);
+
+    spin_unlock_irqrestore (&xmtrRail->xmtr_freelock, flags);
+
+    if (low_on_txds)
+       ep_kthread_schedule (&subsys->Thread, lbolt);
+
+
+    return (txdRail);
+}
+
+static void
+free_txd_rail (EP4_XMTR_RAIL *xmtrRail, EP4_TXD_RAIL *txdRail)
+{
+    unsigned long flags;
+
+    EP4_TXD_ASSERT_FREE(txdRail);
+
+    spin_lock_irqsave (&xmtrRail->xmtr_freelock, flags);
+    
+    list_add (&txdRail->txd_generic.Link, &xmtrRail->xmtr_freelist);
+
+    xmtrRail->xmtr_freecount++;
+
+    if (xmtrRail->xmtr_freewaiting)
+    {
+       xmtrRail->xmtr_freewaiting--;
+       kcondvar_wakeupall (&xmtrRail->xmtr_freesleep, &xmtrRail->xmtr_freelock);
+    }
+
+    spin_unlock_irqrestore (&xmtrRail->xmtr_freelock, flags);
+}
+
+static void
+bind_txd_rail (EP_TXD *txd, EP4_TXD_RAIL *txdRail)
+{
+    EPRINTF6 (DBG_XMTR, "%s: bind_txd_rail: txd=%p txdRail=%p XID=%08x.%08x.%016llx\n", 
+             XMTR_TO_RAIL(txdRail->txd_generic.XmtrRail)->r_generic.Name, txd, txdRail, 
+             txd->Envelope.Xid.Generation, txd->Envelope.Xid.Handle, txd->Envelope.Xid.Unique);
+
+    txd->TxdRail = &txdRail->txd_generic;
+    txdRail->txd_generic.Txd = txd;
+}
+
+static void
+unbind_txd_rail (EP_TXD *txd, EP4_TXD_RAIL *txdRail)
+{
+    EP4_TXD_ASSERT (txdRail, txd->TxdRail == &txdRail->txd_generic && txdRail->txd_generic.Txd == txd);
+
+    EPRINTF6 (DBG_XMTR, "%s: unbind_txd_rail: txd=%p txdRail=%p XID=%08x.%08x.%016llx\n", 
+             XMTR_TO_RAIL(txdRail->txd_generic.XmtrRail)->r_generic.Name, txd, txdRail, 
+             txd->Envelope.Xid.Generation, txd->Envelope.Xid.Handle, txd->Envelope.Xid.Unique);
+
+
+    txdRail->txd_generic.Txd = NULL; 
+    txd->TxdRail = NULL;
+}
+
+static void
+initialise_txd (EP_TXD *txd, EP4_TXD_RAIL *txdRail, unsigned int phase)
+{
+    EP4_XMTR_RAIL *xmtrRail = (EP4_XMTR_RAIL *) txdRail->txd_generic.XmtrRail;
+    EP4_RAIL      *rail     = XMTR_TO_RAIL (xmtrRail);
+    ELAN4_DEV     *dev      = rail->r_ctxt.ctxt_dev;
+    
+    /* Flush the Elan TLB if mappings have changed */
+    ep_perrail_dvma_sync (&rail->r_generic);
+    
+    /* Initialise the per-rail fields in the envelope */
+    txd->Envelope.TxdRail = txdRail->txd_elan_addr;
+    txd->Envelope.NodeId  = rail->r_generic.Position.pos_nodeid;
+
+    /* Allocate a network error fixup cookie */
+    txdRail->txd_cookie = ep4_neterr_cookie (rail, txd->NodeId) | EP4_COOKIE_STEN;
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    if ( epdebug_check_sum ) 
+       txd->Envelope.CheckSum = ep_calc_check_sum( txd->Xmtr->Subsys->Subsys.Sys, &txd->Envelope, txd->Envelope.Frags, txd->Envelope.nFrags);
+    else
+#endif
+       txd->Envelope.CheckSum = 0;  
+
+    /* Initialise the per-rail events */
+    switch (phase)
+    {
+    case EP_TXD_PHASE_ACTIVE:
+    {
+       unsigned int nsets = (txd->Envelope.nFrags ? txd->Envelope.nFrags : 1) + ( EP_IS_MULTICAST(txd->Envelope.Attr) ? 1 : 0);
+
+       if (! EP_IS_RPC(txd->Envelope.Attr))
+       {
+           elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType),
+                               E4_EVENT_INIT_VALUE (-32 * nsets, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+           txdRail->txd_main->txd_data = EP4_STATE_FINISHED;
+       }
+       else
+       {
+           elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType),
+                               E4_EVENT_INIT_VALUE(-32 * nsets , E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+           elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType),
+                               E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+
+           txdRail->txd_main->txd_data = EP4_STATE_ACTIVE;
+       }
+                  
+       txdRail->txd_main->txd_env  = EP4_STATE_ACTIVE;
+       txdRail->txd_main->txd_done = EP4_STATE_ACTIVE;
+       break;
+    }
+
+    case EP_TXD_PHASE_PASSIVE:
+       EP4_TXD_ASSERT (txdRail, EP_IS_RPC(txd->Envelope.Attr));
+       
+       txdRail->txd_main->txd_env  = EP4_STATE_FINISHED;
+       txdRail->txd_main->txd_data = EP4_STATE_FINISHED;
+       txdRail->txd_main->txd_done = EP4_STATE_ACTIVE;
+
+       elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+       break;
+    }
+
+   if (EP_IS_NO_INTERRUPT(txd->Envelope.Attr))
+       elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd.c_intr_cmd), NOP_CMD);
+}
+
+static void
+terminate_txd_rail (EP4_XMTR_RAIL *xmtrRail, EP4_TXD_RAIL *txdRail)
+{
+    EP4_SDRAM_ASSERT (TXD_TO_RAIL(txdRail),\
+                     (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),\
+                     E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));\
+
+    /* clear the done flags - so that it will be ignored if an event interrupt is generated */
+    txdRail->txd_main->txd_env  = EP4_STATE_FREE;
+    txdRail->txd_main->txd_data = EP4_STATE_FREE;
+    txdRail->txd_main->txd_done = EP4_STATE_FREE;
+
+#if defined(DEBUG_ASSERT)
+    if (sdram_assert)
+    {
+       ELAN4_DEV *dev = XMTR_TO_RAIL (xmtrRail)->r_ctxt.ctxt_dev;
+
+       elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+       elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+    }
+#endif 
+}
+
+static void
+defer_txd_rail (EP4_TXD_RAIL *txdRail)
+{
+    EP4_XMTR_RAIL   *xmtrRail = TXD_TO_XMTR(txdRail);
+    EP4_RAIL        *rail     = XMTR_TO_RAIL(xmtrRail);
+    ELAN4_DEV       *dev      = rail->r_ctxt.ctxt_dev;
+    EP_COMMS_SUBSYS *subsys   = XMTR_TO_SUBSYS(xmtrRail);
+
+    EPRINTF5 (DBG_XMTR, "%s: defer_txd_rail: xmtrRail=%p txdRail=%p env/data (%d,%d) not finished\n",
+             rail->r_generic.Name, xmtrRail, txdRail, (int)txdRail->txd_main->txd_env, (int)txdRail->txd_main->txd_data);
+                   
+    /* transmit has completed, but the data dma has not completed
+     * (because of network error fixup), we queue the txdRail onto a list
+     * to be polled for completion later.
+     */
+    if (txdRail->txd_retry_time)
+    {
+       EP4_TXD_ASSERT (txdRail, (on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY]) == 1 ||
+                                 on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED]) == 1));
+
+       list_del (&txdRail->txd_retry_link);
+
+       txdRail->txd_main->txd_env = EP4_STATE_FINISHED;
+
+       /* re-initialise the envelope event */
+       elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+    }
+    
+    txdRail->txd_retry_time = lbolt;
+       
+    list_add_tail (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_POLL]);
+       
+    ep_kthread_schedule (&subsys->Thread, lbolt);
+}
+
+static void
+finalise_txd (EP_TXD *txd, EP4_TXD_RAIL *txdRail)
+{
+    EP4_XMTR_RAIL   *xmtrRail = TXD_TO_XMTR(txdRail);
+
+    EP4_TXD_ASSERT_FINISHED (txdRail);
+
+    unbind_txd_rail (txd, txdRail);
+    
+    terminate_txd_rail (xmtrRail, txdRail);
+    free_txd_rail (xmtrRail, txdRail);
+}
+
+static void
+txd_interrupt (EP4_RAIL *rail, void *arg)
+{
+    EP4_TXD_RAIL    *txdRail  = (EP4_TXD_RAIL *) arg;
+    EP4_XMTR_RAIL   *xmtrRail = TXD_TO_XMTR(txdRail);
+    EP_XMTR         *xmtr     = xmtrRail->xmtr_generic.Xmtr;
+    int              delay    = 1;
+    EP_TXD          *txd;
+    unsigned long    flags;
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    for (;;)
+    {
+       if (txdRail->txd_main->txd_done == EP4_STATE_FINISHED || txdRail->txd_main->txd_env == EP4_STATE_FAILED)
+           break;
+       
+       /* The write to txd_done could be held up in the PCI bridge even though
+        * we've seen the interrupt cookie.  Unlike elan3, there is no possibility
+        * of spurious interrupts since we flush the command queues on node 
+        * disconnection and the txcallback mechanism */
+       mb();
+
+       if (delay > EP4_EVENT_FIRING_TLIMIT)
+       {
+           spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+           EP_ASSFAIL (XMTR_TO_RAIL(xmtrRail), "txd_interrupt - not finished\n");
+           return;
+       }
+       DELAY (delay);
+       delay <<= 1;
+    }
+
+    txd = txdRail->txd_generic.Txd;
+
+    if (txdRail->txd_main->txd_env == EP4_STATE_FAILED)
+    {
+       spin_lock (&xmtrRail->xmtr_retrylock);
+
+       EP4_TXD_ASSERT (txdRail, txdRail->txd_retry_time == 0);                         /* cannot be on retry/poll list */
+       EP4_TXD_ASSERT (txdRail, txdRail->txd_main->txd_done != EP4_STATE_FINISHED);    /* data xfer cannot have finished */
+
+       if (TxdShouldStabalise (&txdRail->txd_generic, &rail->r_generic))
+       {
+           EPRINTF6 (DBG_STABILISE, "%s: txd_interrupt: stablise xmtrRail=%p txdRail=%p txd=%p XID=%llx dest=%u\n", rail->r_generic.Name,
+                     xmtrRail, txdRail, txd, txd->Envelope.Xid.Unique, txd->NodeId);
+
+           txdRail->txd_retry_time = lbolt;                    /* indicate on retry list */
+           
+           list_add_tail (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED]);
+       }
+       else
+       {
+           EPRINTF6 (DBG_RETRY, "%s: txd_interrupt: retry xmtrRail=%p txdRail=%p txd=%p XID=%llx dest=%u\n", rail->r_generic.Name,
+                     xmtrRail, txdRail, txd, txd->Envelope.Xid.Unique, txd->NodeId);
+
+           txdRail->txd_retry_time = lbolt + EP_RETRY_LOW_PRI_TIME;            /* XXXX: backoff ? */
+           
+           list_add_tail (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY]);
+           
+           ep_kthread_schedule (&rail->r_retry_thread, txdRail->txd_retry_time);
+       }
+       spin_unlock (&xmtrRail->xmtr_retrylock);
+
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+       return;
+    }
+
+    EP4_TXD_ASSERT (txdRail, txd != NULL && !(EP_IS_NO_INTERRUPT(txd->Envelope.Attr)));
+
+    EPRINTF6 (DBG_XMTR, "%s: txd_interrupt: xmtrRail=%p txdRail=%p txd=%p XID=%llx dest=%u\n", rail->r_generic.Name,
+             xmtrRail, txdRail, txd, txd->Envelope.Xid.Unique, txd->NodeId);
+            
+    if (txdRail->txd_main->txd_env != EP4_STATE_FINISHED || txdRail->txd_main->txd_data != EP4_STATE_FINISHED)
+    {
+       defer_txd_rail (txdRail);
+
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+    }
+    else
+    {
+       /* remove from active transmit list */
+       list_del (&txd->Link);
+
+       ep_xmtr_txd_stat(xmtr,txd);
+
+       finalise_txd (txd, txdRail);
+       
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+       
+       txd->Handler (txd, txd->Arg, EP_SUCCESS);
+       
+       FreeTxd (xmtr, txd);
+    }
+}
+
+static void
+poll_interrupt (EP4_RAIL *rail, void *arg)
+{
+    EP4_XMTR_RAIL   *xmtrRail = (EP4_XMTR_RAIL *) arg;
+
+    ep_poll_transmits (xmtrRail->xmtr_generic.Xmtr);
+}
+
+void
+issue_envelope_packet (EP4_XMTR_RAIL *xmtrRail, EP4_TXD_RAIL *txdRail)
+{
+    EP_TXD    *txd    = txdRail->txd_generic.Txd;
+    ELAN4_CQ  *cq     = xmtrRail->xmtr_cq;
+    E4_uint64 *blk0   = (E4_uint64 *) &txd->Envelope;
+    E4_uint64 *blk1   = EP_HAS_PAYLOAD(txd->Envelope.Attr) ? (E4_uint64 *) &txd->Payload : NULL;
+    E4_Addr    qaddr  = EP_MSGQ_ADDR(txd->Service);
+
+    EP4_SDRAM_ASSERT (TXD_TO_RAIL(txdRail),\
+                     (txdRail)->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),\
+                     E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));\
+
+    elan4_open_packet (cq, OPEN_PACKET (0, PACK_OK | RESTART_COUNT_ZERO, EP_VP_DATA(txd->NodeId)));
+    elan4_sendtrans0 (cq, TR_INPUT_Q_GETINDEX, EP_MSGQ_ADDR(txd->Service));
+           
+    /* send the payload if present */
+    if (blk0) elan4_sendtransp (cq, TR_WRITE(128 >> 3, 0, TR_DATATYPE_BYTE), 0,   blk0);
+    if (blk1) elan4_sendtransp (cq, TR_WRITE(128 >> 3, 0, TR_DATATYPE_BYTE), 128, blk1);
+
+    elan4_sendtrans1 (cq, TR_INPUT_Q_COMMIT, qaddr, txdRail->txd_cookie);
+
+    elan4_guard (cq, GUARD_CHANNEL (1) | GUARD_TEST(0, PACK_OK) | GUARD_RESET (EP4_STEN_RETRYCOUNT));
+    elan4_write_dword_cmd (cq, txdRail->txd_main_addr + offsetof (EP4_TXD_RAIL_MAIN, txd_env), EP4_STATE_FINISHED);
+           
+    elan4_guard (cq, GUARD_CHANNEL (1) | GUARD_TEST(0, RESTART_COUNT_ZERO) | GUARD_RESET (EP4_STEN_RETRYCOUNT));
+    elan4_set_event_cmd (cq, txdRail->txd_elan_addr + offsetof (EP4_TXD_RAIL_ELAN, txd_env));
+    
+    elan4_write_dword_cmd (cq, xmtrRail->xmtr_main_addr + offsetof (EP4_XMTR_RAIL_MAIN, xmtr_flowcnt), ++xmtrRail->xmtr_flowcnt);
+}
+
+void
+ep4xmtr_flush_callback (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail)
+{
+    EP4_RAIL       *rail      = XMTR_TO_RAIL (xmtrRail);
+    EP4_COMMS_RAIL *commsRail = XMTR_TO_COMMS (xmtrRail);
+    struct list_head *el, *nel;
+    unsigned long flags;
+
+    switch (rail->r_generic.CallbackStep)
+    {
+    case EP_CB_FLUSH_FILTERING:
+       /* need to acquire/release the Lock to ensure that the node state
+        * transition has been noticed and no new envelopes are queued to 
+        * nodes which are passivating. */
+       spin_lock_irqsave (&xmtr->Lock, flags);
+
+       /* Then we insert a "setevent" into the command queue to flush
+        * through the envelopes which have already been submitted */
+       ep4comms_flush_setevent (commsRail, xmtrRail->xmtr_cq);
+
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+       break;
+
+    case EP_CB_FLUSH_FLUSHING:
+       /* remove any envelopes which are retrying to nodes which are going down */
+       spin_lock_irqsave (&xmtrRail->xmtr_retrylock, flags);
+       list_for_each_safe (el, nel, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY]) {
+           EP4_TXD_RAIL *txdRail  = list_entry (el, EP4_TXD_RAIL, txd_retry_link);
+           EP_TXD       *txd      = txdRail->txd_generic.Txd;
+           EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[txd->NodeId];
+           
+           EP4_TXD_ASSERT (txdRail, txdRail->txd_main->txd_env == EP4_STATE_FAILED);
+           
+           if (nodeRail->State == EP_NODE_LOCAL_PASSIVATE)
+           {
+               EPRINTF2 (DBG_XMTR, "%s; ep4xmtr_flush_callback: removing txdRail %p from retry list\n", rail->r_generic.Name, txdRail);
+               
+               EP4_TXD_ASSERT (txdRail, txdRail->txd_retry_time != 0);
+
+               list_del (&txdRail->txd_retry_link);
+               list_add_tail (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED]);
+           }
+       }
+       spin_unlock_irqrestore (&xmtrRail->xmtr_retrylock, flags);
+
+       /* Determine whether we have active or passive messages to 
+        * any node which is passivating */
+       spin_lock_irqsave (&xmtr->Lock, flags);
+       list_for_each (el, &xmtr->ActiveDescList) {
+           EP_TXD       *txd      = list_entry (el, EP_TXD, Link);
+           EP4_TXD_RAIL *txdRail  = (EP4_TXD_RAIL *) txd->TxdRail;
+           EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[txd->NodeId];
+           
+           if (txdRail == NULL || txdRail->txd_generic.XmtrRail != &xmtrRail->xmtr_generic || nodeRail->State != EP_NODE_LOCAL_PASSIVATE)
+               continue;
+           
+           EPRINTF5 (DBG_XMTR, "%s: flush txd=%p txdRail=%p data=%llx done=%llx\n", rail->r_generic.Name,
+                     txd, txdRail, txdRail->txd_main->txd_data, txdRail->txd_main->txd_done);
+
+           if (EP_IS_RPC(txd->Envelope.Attr))
+           {
+               if (txdRail->txd_main->txd_data == EP4_STATE_ACTIVE)
+                   nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+               else if (txdRail->txd_main->txd_data == EP4_STATE_ACTIVE)
+                   nodeRail->MessageState |= EP_NODE_PASSIVE_MESSAGES;
+           }
+           else
+           {
+               if (txdRail->txd_main->txd_data == EP4_STATE_ACTIVE)
+                   nodeRail->MessageState |= EP_NODE_ACTIVE_MESSAGES;
+           }
+       }
+       spin_unlock_irqrestore (&xmtr->Lock, flags);
+       break;
+
+    default:
+       panic ("ep4xmtr_flush_callback: invalid callback step\n");
+       break;
+    }
+}
+
+void
+ep4xmtr_failover_callback (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail)
+{
+    EP4_RAIL         *rail   = XMTR_TO_RAIL (xmtrRail);
+    struct list_head  txdList;
+    struct list_head *el, *nel;
+    unsigned long flags;
+
+    INIT_LIST_HEAD (&txdList);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    list_for_each_safe (el, nel, &xmtr->ActiveDescList) {
+       EP_TXD       *txd       = list_entry (el, EP_TXD, Link);
+       EP4_TXD_RAIL *txdRail   = (EP4_TXD_RAIL *) txd->TxdRail;
+       EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[txd->NodeId];
+           
+       /* Only progress relocation of txd's bound to this rail */
+       if (! TXD_BOUND2RAIL (txdRail, xmtrRail) || nodeRail->State != EP_NODE_PASSIVATED)
+           continue;
+       
+       /* XXXX - no rail failover for now ....*/
+
+       EPRINTF4 (DBG_XMTR, "%s: ep4xmtr_failover_callback - xmtr %p txd %p node %d completed\n", rail->r_generic.Name, xmtr, txd, txd->NodeId);
+    }
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    while (! list_empty (&txdList)) 
+    {
+       EP_TXD *txd = list_entry (txdList.next, EP_TXD, Link);
+
+       list_del (&txd->Link);
+
+       txd->Handler (txd, txd->Arg, EP_CONN_RESET);
+       
+       FreeTxd (xmtr, txd);
+    }
+}
+
+
+void
+ep4xmtr_disconnect_callback (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail)
+{
+    EP4_RAIL         *rail = XMTR_TO_RAIL (xmtrRail);
+    ELAN4_DEV        *dev  = rail->r_ctxt.ctxt_dev;
+    struct list_head *el, *nel;
+    struct list_head  txdList;
+    unsigned long flags;
+    
+    INIT_LIST_HEAD (&txdList);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+
+    list_for_each_safe (el, nel, &xmtr->ActiveDescList) {
+       EP_TXD       *txd       = list_entry (el, EP_TXD, Link);
+       EP4_TXD_RAIL *txdRail   = (EP4_TXD_RAIL *) txd->TxdRail;
+       EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[txd->NodeId];
+           
+       if ( ! TXD_BOUND2RAIL (txdRail, xmtrRail) || nodeRail->State != EP_NODE_DISCONNECTING)
+           continue;
+       
+       if (txdRail->txd_main->txd_done == EP4_STATE_ACTIVE)
+       {
+
+           EPRINTF8 (DBG_DISCON, "ep4xmtr_disconnect_callback: txdRail=%p : events %llx,%llx,%llx done %llx,%llx,%llx retry %lx\n",txdRail,
+                     elan4_sdram_readq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType)),
+                     elan4_sdram_readq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType)),
+                     elan4_sdram_readq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType)),
+                     txdRail->txd_main->txd_env, txdRail->txd_main->txd_data, txdRail->txd_main->txd_done,
+                     txdRail->txd_retry_time);
+                      
+           if (txdRail->txd_retry_time)
+           {
+               /* re-initialise the envelope event */
+               elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),
+                                   E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+               
+               EP4_TXD_ASSERT (txdRail, on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED]) == 1);
+
+               txdRail->txd_retry_time  = 0;
+
+               list_del (&txdRail->txd_retry_link);
+           }
+
+           /* Remove from active list */
+           list_del (&txd->Link);
+       
+           unbind_txd_rail (txd, txdRail);
+
+           terminate_txd_rail (xmtrRail, txdRail);
+           free_txd_rail (xmtrRail, txdRail);
+           
+           EPRINTF4 (DBG_XMTR, "%s: ep4xmtr_disconnect_callback - xmtr %p txd %p node %d not conected\n", rail->r_generic.Name, xmtr, txd, txd->NodeId);
+
+           /* add to the list of txd's which are to be completed */
+           list_add_tail (&txd->Link, &txdList);
+       }
+    }
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+    
+    while (! list_empty (&txdList)) 
+    {
+       EP_TXD *txd = list_entry (txdList.next, EP_TXD, Link);
+
+       list_del (&txd->Link);
+
+       txd->Handler (txd, txd->Arg, EP_CONN_RESET);
+       
+       FreeTxd (xmtr, txd);
+    }
+}
+
+void
+ep4xmtr_neterr_flush (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP4_COMMS_RAIL   *commsRail = XMTR_TO_COMMS (xmtrRail);
+    unsigned long flags;
+    
+    spin_lock_irqsave (&xmtr->Lock, flags);
+
+    /* insert a "setevent" into the command queue to flush
+     * through the envelopes which have already been submitted */
+    ep4comms_flush_setevent (commsRail, xmtrRail->xmtr_cq);
+
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+}
+
+void
+ep4xmtr_neterr_check (EP_XMTR *xmtr, EP4_XMTR_RAIL *xmtrRail, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP4_RAIL *rail = XMTR_TO_RAIL (xmtrRail);
+    ELAN4_DEV *dev = rail->r_ctxt.ctxt_dev;
+    struct list_head *el;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    list_for_each (el, &xmtr->ActiveDescList) {
+       EP_TXD       *txd     = list_entry (el, EP_TXD, Link);
+       EP4_TXD_RAIL *txdRail = (EP4_TXD_RAIL *) txd->TxdRail;
+           
+       if ( ! TXD_BOUND2RAIL (txdRail, xmtrRail) || txd->NodeId != nodeId)
+           continue;
+       
+       /* The only non-dma associated with a txd is the initial sten packet, if it has been acked 
+        * and the neterr cookie matches, then change it to look like it's been acked since the
+        * INPUT_Q_COMMIT transaction has already been executed */
+       if (txdRail->txd_main->txd_env == EP4_STATE_FAILED && (txdRail->txd_cookie == cookies[0] || txdRail->txd_cookie == cookies[1]))
+       {
+           EPRINTF4 (DBG_NETWORK_ERROR, "%s: ep4xmtr_neterr_callback: cookie <%lld%s%s%s%s> matches txd %p txdRail %p\n", 
+                    rail->r_generic.Name, EP4_COOKIE_STRING(txdRail->txd_cookie), txd, txdRail);
+
+           EP4_TXD_ASSERT (txdRail, txdRail->txd_retry_time != 0);
+
+           txdRail->txd_main->txd_env = EP4_STATE_FINISHED;
+
+           /* re-initialise the envelope event */
+           elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),
+                               E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+           
+           spin_lock (&xmtrRail->xmtr_retrylock);
+
+           EP4_TXD_ASSERT (txdRail, (on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY]) == 1 ||
+                                     on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED]) == 1));
+
+           txdRail->txd_retry_time = 0;
+
+           list_del (&txdRail->txd_retry_link);
+
+           spin_unlock (&xmtrRail->xmtr_retrylock);
+       }
+    }
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+}
+
+int
+ep4xmtr_poll_txd (EP_XMTR_RAIL *x, EP_TXD_RAIL *t, int how)
+{
+    EP4_XMTR_RAIL *xmtrRail = (EP4_XMTR_RAIL *) x;
+    ELAN4_DEV     *dev      = XMTR_TO_DEV (xmtrRail);
+    EP4_TXD_RAIL  *txdRail  = (EP4_TXD_RAIL *) t;
+    EP_TXD        *txd      = txdRail->txd_generic.Txd;
+
+    if (! EP_IS_NO_INTERRUPT(txd->Envelope.Attr))
+       return 0;
+
+    switch (how)
+    {
+    case ENABLE_TX_CALLBACK:
+       if (!EP_IS_INTERRUPT_ENABLED(txd->Envelope.Attr))
+       {
+           elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd.c_intr_cmd),
+                               INTERRUPT_CMD | (xmtrRail->xmtr_intcookie.int_val << E4_MAIN_INT_SHIFT));
+
+           txd->Envelope.Attr |= EP_INTERRUPT_ENABLED;
+       }
+       break;
+
+    case DISABLE_TX_CALLBACK:
+       if (EP_IS_INTERRUPT_ENABLED(txd->Envelope.Attr & EP_INTERRUPT_ENABLED))
+       {
+           elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd.c_intr_cmd), NOP_CMD);
+
+           txd->Envelope.Attr &= ~EP_INTERRUPT_ENABLED;
+       }
+    }
+    
+    if (txdRail->txd_main->txd_env == EP4_STATE_FINISHED && txdRail->txd_main->txd_data == EP4_STATE_FINISHED && txdRail->txd_main->txd_done == EP4_STATE_FINISHED)
+    {
+       EPRINTF3 (DBG_XMTR, "%s: ep4xmtr_poll_txd: txd=%p XID=%llx completed\n",
+                 XMTR_TO_RAIL (xmtrRail)->r_generic.Name, txd, txd->Envelope.Xid.Unique);
+       
+       elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_done_cmd.c_intr_cmd),
+                           INTERRUPT_CMD | (txdRail->txd_intcookie.int_val << E4_MAIN_INT_SHIFT));
+
+
+       ep_xmtr_txd_stat(xmtrRail->xmtr_generic.Xmtr,txd);
+
+       finalise_txd (txd, txdRail);
+
+       return 1;
+    }
+
+    return 0;
+}
+
+int
+ep4xmtr_bind_txd (EP_TXD *txd, EP_XMTR_RAIL *x, unsigned int phase)
+{
+    EP4_XMTR_RAIL *xmtrRail = (EP4_XMTR_RAIL *) x;
+    EP4_RAIL      *rail     = XMTR_TO_RAIL (xmtrRail);
+    EP4_TXD_RAIL  *txdRail;
+    unsigned long  flags;
+
+    if ((txdRail = get_txd_rail (xmtrRail)) == NULL)
+       return 0;
+    
+    switch (phase)
+    {
+    case EP_TXD_PHASE_ACTIVE:
+       if (rail->r_generic.Nodes[txd->NodeId].State != EP_NODE_CONNECTED)
+       {
+           EPRINTF2 (DBG_XMTR, "%s: ep4xmtr_bind_txd: node %u not connected on this rail\n", rail->r_generic.Name, txd->NodeId);
+
+           free_txd_rail (xmtrRail, txdRail);
+           return 0;
+       }
+
+       initialise_txd (txd, txdRail, EP_TXD_PHASE_ACTIVE);
+
+       bind_txd_rail (txd, txdRail);
+       
+       /* generate the STEN packet to transfer the envelope */
+       spin_lock_irqsave (&xmtrRail->xmtr_retrylock, flags);
+       if (((int) (xmtrRail->xmtr_flowcnt - xmtrRail->xmtr_main->xmtr_flowcnt)) < EP4_XMTR_FLOWCNT)
+           issue_envelope_packet (xmtrRail, txdRail);
+       else
+       {
+           txdRail->txd_retry_time = lbolt;
+
+           list_add_tail (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY]);
+
+           ep_kthread_schedule (&rail->r_retry_thread, txdRail->txd_retry_time);
+       }
+       spin_unlock_irqrestore (&xmtrRail->xmtr_retrylock, flags);
+       break;
+
+    case EP_TXD_PHASE_PASSIVE:
+       initialise_txd (txd, txdRail, EP_TXD_PHASE_PASSIVE);
+       
+       EP_XMTR_OP (txd->TxdRail->XmtrRail, UnbindTxd) (txd, EP_TXD_PHASE_PASSIVE);     /* unbind from existing rail */
+
+       bind_txd_rail (txd, txdRail);                                                   /* and bind it to our new rail */
+       break;
+    }
+
+    return 1;
+}
+
+void
+ep4xmtr_unbind_txd (EP_TXD *txd, unsigned int phase)
+{
+    /* XXXX - TBD */
+}
+
+long
+ep4xmtr_check (EP_XMTR_RAIL *x, long nextRunTime)
+{
+    EP4_XMTR_RAIL *xmtrRail = (EP4_XMTR_RAIL *) x;
+    EP_XMTR       *xmtr     = xmtrRail->xmtr_generic.Xmtr;
+    struct list_head  txdList;
+    struct list_head *el, *nel;
+    unsigned long flags;
+
+    INIT_LIST_HEAD (&txdList);
+
+    if (xmtrRail->xmtr_freecount < ep_txd_lowat && !alloc_txd_block (xmtrRail))
+    {
+       EPRINTF1 (DBG_RCVR,"%s: failed to grow txd rail pool\n", XMTR_TO_RAIL(xmtrRail)->r_generic.Name);
+               
+       if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + RESOURCE_RETRY_TIME))
+           nextRunTime = lbolt + RESOURCE_RETRY_TIME;
+    }
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    list_for_each_safe (el, nel, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_POLL]) {
+       EP4_TXD_RAIL *txdRail = list_entry (el, EP4_TXD_RAIL, txd_retry_link);
+
+       if (txdRail->txd_main->txd_env != EP4_STATE_FINISHED || txdRail->txd_main->txd_data != EP4_STATE_FINISHED)
+       {
+           ep_debugf (DBG_XMTR, "%s: ep4xmtr_check: xmtrRail=%p txdRail=%p env/data (%d,%d) not finished\n",
+                      XMTR_TO_RAIL(xmtrRail)->r_generic.Name, xmtrRail, txdRail, (int)txdRail->txd_main->txd_env, (int)txdRail->txd_main->txd_data);
+                   
+           nextRunTime = lbolt + HZ;
+       }
+       else
+       {
+           EP_TXD *txd = txdRail->txd_generic.Txd;
+
+           ep_debugf (DBG_XMTR, "%s: ep4xmtr_check: xmtrRail=%p txdRail=%p env/data (%d,%d) finished\n",
+                      XMTR_TO_RAIL(xmtrRail)->r_generic.Name, xmtrRail, txdRail, (int)txdRail->txd_main->txd_env, (int)txdRail->txd_main->txd_data);
+
+           EPRINTF5 (DBG_XMTR, "%s: ep4xmtr_check: xmtrRail=%p txdRail=%p env/data (%d,%d) finished\n",
+                     XMTR_TO_RAIL(xmtrRail)->r_generic.Name, xmtrRail, txdRail, (int)txdRail->txd_main->txd_env, (int)txdRail->txd_main->txd_data);
+           EPRINTF3  (DBG_XMTR, "%s:    done %x data %x\n", XMTR_TO_RAIL(xmtrRail)->r_generic.Name,
+                      txdRail->txd_elan_addr + offsetof (EP4_TXD_RAIL_ELAN, txd_done),
+                      txdRail->txd_elan_addr + offsetof (EP4_TXD_RAIL_ELAN, txd_data));
+
+           EP4_TXD_ASSERT (txdRail, txdRail->txd_retry_time != 0);
+
+           /* remove txd from active list and add to list to call handlers */
+           list_del (&txd->Link);
+           list_add_tail (&txd->Link, &txdList);
+
+           /* remove and free of txdRail */
+           txdRail->txd_retry_time = 0;
+           list_del (&txdRail->txd_retry_link);
+
+           finalise_txd (txd, txdRail);
+
+       }
+    }
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    while (! list_empty (&txdList))
+    {
+       EP_TXD *txd = list_entry (txdList.next, EP_TXD, Link);
+
+       list_del (&txd->Link);
+
+       ep_xmtr_txd_stat (xmtr,txd);
+
+       txd->Handler (txd, txd->Arg, EP_SUCCESS);
+
+       FreeTxd (xmtr, txd);
+    }
+
+    return nextRunTime;
+}
+
+unsigned long
+ep4xmtr_retry (EP4_RAIL *rail, void *arg, unsigned long nextRunTime)
+{
+    EP4_XMTR_RAIL *xmtrRail = (EP4_XMTR_RAIL *) arg;
+    ELAN4_DEV     *dev      = XMTR_TO_DEV(xmtrRail);
+    unsigned long  flags;
+
+    spin_lock_irqsave (&xmtrRail->xmtr_retrylock, flags);
+    while (! list_empty (&xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY]))
+    {
+       EP4_TXD_RAIL *txdRail = list_entry (xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY].next, EP4_TXD_RAIL, txd_retry_link);
+
+       if (BEFORE (lbolt, txdRail->txd_retry_time))
+       {
+           if (nextRunTime == 0 || AFTER (nextRunTime, txdRail->txd_retry_time))
+               nextRunTime = txdRail->txd_retry_time;
+
+           break;
+       }
+
+       if (((int) (xmtrRail->xmtr_flowcnt - xmtrRail->xmtr_main->xmtr_flowcnt)) < EP4_XMTR_FLOWCNT)
+       {
+           txdRail->txd_retry_time = 0;
+
+           list_del (&txdRail->txd_retry_link);
+           
+           /* re-initialise the envelope event */
+           elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),
+                               E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+           
+           EPRINTF3 (DBG_RETRY, "%s: ep4xmtr_retry: re-issue envelope packet to %d for txdRail=%p\n", 
+                     rail->r_generic.Name, txdRail->txd_generic.Txd->Envelope.NodeId, txdRail);
+           
+           txdRail->txd_main->txd_env = EP4_STATE_ACTIVE;
+           
+           issue_envelope_packet (xmtrRail, txdRail);
+       }
+       else
+       {
+           EPRINTF2 (DBG_RETRY, "%s: ep4xmtr_retry: cannot re-issue envelope packet to %d\n", rail->r_generic.Name, txdRail->txd_generic.Txd->Envelope.NodeId);
+
+           if (nextRunTime == 0 || AFTER (nextRunTime, txdRail->txd_retry_time))
+               nextRunTime = txdRail->txd_retry_time;
+
+           break;
+       }
+    }
+    spin_unlock_irqrestore (&xmtrRail->xmtr_retrylock, flags);
+    
+    return nextRunTime;
+}
+
+void
+ep4xmtr_add_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail)
+{
+    EP4_RAIL         *rail   = (EP4_RAIL *) commsRail->Rail;
+    EP_COMMS_SUBSYS  *subsys = xmtr->Subsys;
+    EP4_XMTR_RAIL    *xmtrRail;
+    unsigned long     flags;
+    int                      i;
+
+    KMEM_ZALLOC (xmtrRail, EP4_XMTR_RAIL *, sizeof (EP4_XMTR_RAIL), 1);
+
+    spin_lock_init (&xmtrRail->xmtr_freelock);
+    kcondvar_init  (&xmtrRail->xmtr_freesleep);
+    INIT_LIST_HEAD (&xmtrRail->xmtr_freelist);
+    INIT_LIST_HEAD (&xmtrRail->xmtr_blocklist);
+
+    for (i = 0; i < EP4_TXD_NUM_LISTS; i++)
+       INIT_LIST_HEAD (&xmtrRail->xmtr_retrylist[i]);
+    spin_lock_init (&xmtrRail->xmtr_retrylock);
+
+    xmtrRail->xmtr_generic.CommsRail = commsRail;
+    xmtrRail->xmtr_generic.Xmtr      = xmtr;
+
+    xmtrRail->xmtr_main = ep_alloc_main (&rail->r_generic, sizeof (EP4_XMTR_RAIL_MAIN), 0, &xmtrRail->xmtr_main_addr);
+    xmtrRail->xmtr_cq   = elan4_alloccq (&rail->r_ctxt, EP4_XMTR_CQSIZE, CQ_EnableAllBits, CQ_Priority);
+
+    xmtrRail->xmtr_retryops.op_func = ep4xmtr_retry;
+    xmtrRail->xmtr_retryops.op_arg  = xmtrRail;
+
+    ep4_add_retry_ops (rail, &xmtrRail->xmtr_retryops);
+
+    ep4_register_intcookie (rail, &xmtrRail->xmtr_intcookie, xmtrRail->xmtr_main_addr,
+                           poll_interrupt, xmtrRail);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+
+    xmtr->Rails[commsRail->Rail->Number] = &xmtrRail->xmtr_generic;
+    xmtr->RailMask |= EP_RAIL2RAILMASK(commsRail->Rail->Number);
+
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    ep_kthread_schedule (&subsys->Thread, lbolt);
+
+    ep_procfs_xmtr_add_rail(&(xmtrRail->xmtr_generic));
+}
+
+void
+ep4xmtr_del_rail (EP_XMTR *xmtr, EP_COMMS_RAIL *commsRail)
+{
+    EP4_RAIL         *rail     = (EP4_RAIL *) commsRail->Rail;
+    EP4_XMTR_RAIL    *xmtrRail = (EP4_XMTR_RAIL *) xmtr->Rails[commsRail->Rail->Number];
+    unsigned long     flags;
+
+    /* rail mask set as not usable */
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    xmtr->RailMask &= ~EP_RAIL2RAILMASK (rail->r_generic.Number);
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    ep_procfs_xmtr_del_rail(&(xmtrRail->xmtr_generic));
+
+    /* wait for all txd's for this rail to become free */
+    spin_lock_irqsave (&xmtrRail->xmtr_freelock, flags);
+    while (xmtrRail->xmtr_freecount != xmtrRail->xmtr_totalcount)
+    {
+       xmtrRail->xmtr_freewaiting++;
+       kcondvar_wait (&xmtrRail->xmtr_freesleep, &xmtrRail->xmtr_freelock, &flags);
+    }
+    spin_unlock_irqrestore (&xmtrRail->xmtr_freelock, flags);
+
+    spin_lock_irqsave (&xmtr->Lock, flags);
+    xmtr->Rails[commsRail->Rail->Number] = NULL;
+    spin_unlock_irqrestore (&xmtr->Lock, flags);
+
+    /* all the txd's accociated with DescBlocks must be in the freelist */
+    ASSERT (xmtrRail->xmtr_totalcount == xmtrRail->xmtr_freecount);
+
+    /* run through the DescBlockList deleting them */
+    while (!list_empty (&xmtrRail->xmtr_blocklist))
+       free_txd_block (xmtrRail, list_entry(xmtrRail->xmtr_blocklist.next, EP4_TXD_RAIL_BLOCK , blk_link));
+    
+    /* it had better be empty after that */
+    ASSERT ((xmtrRail->xmtr_freecount == 0) && (xmtrRail->xmtr_totalcount == 0));
+
+    ep4_deregister_intcookie (rail, &xmtrRail->xmtr_intcookie);
+
+    ep4_remove_retry_ops (rail, &xmtrRail->xmtr_retryops);
+
+    elan4_freecq (&rail->r_ctxt, xmtrRail->xmtr_cq);
+    ep_free_main (&rail->r_generic, xmtrRail->xmtr_main_addr, sizeof (EP4_XMTR_RAIL_MAIN));
+
+    spin_lock_destroy (&xmtrRail->xmtr_retrylock);
+
+    spin_lock_destroy (&xmtrRail->xmtr_freelock);
+    kcondvar_destroy (&xmtrRail->xmtr_freesleep);
+
+    KMEM_FREE (xmtrRail, sizeof (EP4_XMTR_RAIL));
+}
+
+void
+ep4xmtr_display_xmtr (DisplayInfo *di, EP_XMTR_RAIL *x)
+{
+    EP4_XMTR_RAIL    *xmtrRail     = (EP4_XMTR_RAIL *) x;
+    EP4_RAIL         *rail         = XMTR_TO_RAIL (xmtrRail);
+    unsigned int      freeCount    = 0;
+    unsigned int      pollCount    = 0;
+    unsigned int      stalledCount = 0;
+    unsigned int      retryCount   = 0;
+    struct list_head *el;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&xmtrRail->xmtr_freelock, flags);
+    list_for_each (el, &xmtrRail->xmtr_freelist)
+       freeCount++;
+    spin_unlock_irqrestore (&xmtrRail->xmtr_freelock, flags);
+
+    spin_lock_irqsave (&xmtrRail->xmtr_retrylock, flags);
+    list_for_each (el, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_POLL])
+       pollCount++;
+    list_for_each (el, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED])
+       stalledCount++;
+    list_for_each (el, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY])
+       retryCount++;
+    spin_unlock_irqrestore (&xmtrRail->xmtr_retrylock, flags);
+
+    (di->func)(di->arg, "        rail=%d free=%d total=%d (%d) (retry %d,%d,%d)\n",
+              rail->r_generic.Number, xmtrRail->xmtr_freecount, xmtrRail->xmtr_totalcount, 
+              freeCount, pollCount, stalledCount, retryCount);
+    (di->func)(di->arg, "        cq %d flowcnt %lld,%lld\n", elan4_cq2num (xmtrRail->xmtr_cq), xmtrRail->xmtr_flowcnt, xmtrRail->xmtr_main->xmtr_flowcnt);
+}
+
+void
+ep4xmtr_display_txd (DisplayInfo *di, EP_TXD_RAIL *t)
+{
+    EP4_TXD_RAIL      *txdRail  = (EP4_TXD_RAIL *) t;
+    EP4_XMTR_RAIL     *xmtrRail = TXD_TO_XMTR(txdRail);
+    EP4_TXD_RAIL_MAIN *txdMain  = txdRail->txd_main;
+    sdramaddr_t        txdElan  = txdRail->txd_elan;
+    EP4_RAIL          *rail     = XMTR_TO_RAIL (xmtrRail);
+    ELAN4_DEV         *dev      = XMTR_TO_DEV (xmtrRail);
+    char             *list     = "";
+    unsigned long      flags;
+
+    spin_lock_irqsave (&xmtrRail->xmtr_retrylock, flags);
+    if (txdRail->txd_retry_time)
+    {
+       if (on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_POLL]))
+           list = " poll";
+       else if (on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED]))
+           list = " stalled";
+       else if (on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_RETRY]))
+           list = " retry";
+       else
+           list = " ERROR";
+    }
+    spin_unlock_irqrestore (&xmtrRail->xmtr_retrylock, flags);
+
+    (di->func)(di->arg, "      Rail %d txd %p elan %lx (%x) main %p (%x) cookie <%lld%s%s%s%s> ecq %d %s\n", rail->r_generic.Number,
+              txdRail, txdRail->txd_elan, txdRail->txd_elan_addr, txdRail->txd_main, txdRail->txd_main_addr, 
+              EP4_COOKIE_STRING(txdRail->txd_cookie), elan4_cq2num (txdRail->txd_ecq->ecq_cq), list);
+    
+    (di->func)(di->arg, "        env  %016llx %016llx %016llx -> %016llx\n",
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType)),
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_Params[0])),
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_Params[1])),
+              txdMain->txd_env);
+    (di->func)(di->arg, "        data %016llx %016llx %016llx -> %016llx\n",
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_CountAndType)),
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_Params[0])),
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_data.ev_Params[1])),
+              txdMain->txd_data);
+    (di->func)(di->arg, "        done %016llx %016llx %016llx -> %016llx\n",
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_CountAndType)),
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_Params[0])),
+              elan4_sdram_readq (dev, txdElan + offsetof (EP4_TXD_RAIL_ELAN, txd_done.ev_Params[1])),
+              txdMain->txd_done);
+}
+
+int
+ep4xmtr_check_txd_state (EP_TXD *txd) 
+{
+    EP4_TXD_RAIL  *txdRail  = (EP4_TXD_RAIL *) txd->TxdRail;
+    EP4_XMTR_RAIL *xmtrRail = (EP4_XMTR_RAIL *) txdRail->txd_generic.XmtrRail;
+    ELAN4_DEV     *dev      = XMTR_TO_DEV (xmtrRail);
+    unsigned long  flags;
+
+    if (txdRail->txd_main->txd_env == EP4_STATE_FINISHED)
+       return 0;
+
+    EP4_TXD_ASSERT (txdRail, txdRail->txd_retry_time != 0);
+
+    spin_lock_irqsave (&xmtrRail->xmtr_retrylock, flags);
+    EP4_TXD_ASSERT (txdRail, on_list (&txdRail->txd_retry_link, &xmtrRail->xmtr_retrylist[EP4_TXD_LIST_STALLED]) == 1);
+
+    list_del (&txdRail->txd_retry_link);
+    txdRail->txd_retry_time  = 0;
+    spin_unlock_irqrestore (&xmtrRail->xmtr_retrylock, flags);
+    
+    /* re-initialise the envelope event */
+    elan4_sdram_writeq (dev, txdRail->txd_elan + offsetof (EP4_TXD_RAIL_ELAN, txd_env.ev_CountAndType),
+                       E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_INTR_CMD_NDWORDS));
+           
+    unbind_txd_rail (txd, txdRail);
+
+    terminate_txd_rail (xmtrRail, txdRail);
+    free_txd_rail (xmtrRail, txdRail);
+
+    return 1;
+}
+
+void
+ep4xmtr_fillout_rail_stats(EP_XMTR_RAIL *xmtr_rail, char *str) {
+    /* no stats here yet */
+    /* EP4_XMTR_RAIL * ep4xmtr_rail = (EP4_XMTR_RAIL *) xmtr_rail; */
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/ep_procfs.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/ep_procfs.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/ep_procfs.c       2005-06-01 23:12:54.660429832 -0400
@@ -0,0 +1,331 @@
+
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: ep_procfs.c,v 1.5.6.3 2004/11/30 10:10:57 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/ep_procfs.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "cm.h"
+#include "debug.h"
+#include "conf_linux.h"
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "epcomms_elan4.h"
+
+#include <qsnet/procfs_linux.h>
+
+struct proc_dir_entry *ep_procfs_xmtr_root;
+struct proc_dir_entry *ep_procfs_rcvr_root;
+
+static int
+ep_proc_open (struct inode *inode, struct file *file)
+{
+    PROC_PRIVATE *pr;
+    int           pages = 4;
+
+    if ((pr = kmalloc (sizeof (PROC_PRIVATE), GFP_KERNEL)) == NULL)
+       return (-ENOMEM);
+    
+    do {       
+       pr->pr_data_len = PAGESIZE * pages;
+
+       KMEM_ZALLOC (pr->pr_data, char *, pr->pr_data_len, 1);
+       if (pr->pr_data == NULL) 
+       { 
+           pr->pr_len  = sprintf (pr->pr_data, "Out of Memory\n");
+           break;
+       } 
+       
+       pr->pr_off     = 0;
+       pr->pr_len     = 0;
+       pr->pr_data[0] = 0;
+       
+       pr->pr_di.func  = proc_character_fill;
+       pr->pr_di.arg   = (long)pr;
+       
+       if (!strcmp("debug_xmtr", file->f_dentry->d_iname)) 
+       {   
+           EP_XMTR *xmtr = (EP_XMTR *)(PDE(inode)->data);
+           ep_display_xmtr (&pr->pr_di, xmtr);
+       }
+       
+       if (!strcmp("debug_rcvr", file->f_dentry->d_iname)) 
+       {
+           EP_RCVR *rcvr = (EP_RCVR *)(PDE(inode)->data);
+           ep_display_rcvr (&pr->pr_di, rcvr, 0);
+       }
+       
+       if (!strcmp("debug_full", file->f_dentry->d_iname)) 
+       {
+           EP_RCVR *rcvr = (EP_RCVR *)(PDE(inode)->data);
+           ep_display_rcvr (&pr->pr_di, rcvr, 1);
+       }
+
+       if ( pr->pr_len < pr->pr_data_len) 
+           break; /* we managed to get all the output into the buffer */
+
+       pages++;
+       KMEM_FREE ( pr->pr_data,  pr->pr_data_len);
+    } while (1);
+       
+
+    file->private_data = (void *) pr;
+
+    MOD_INC_USE_COUNT;
+    return (0);
+}
+
+struct file_operations ep_proc_operations = 
+{
+    read:      proc_read,
+    open:      ep_proc_open,
+    release:   proc_release,
+};
+
+static int
+proc_read_rcvr_stats(char *page, char **start, off_t off,
+                    int count, int *eof, void *data)
+{
+    EP_RCVR *rcvr = (EP_RCVR *)data;
+    
+    if (rcvr == NULL) 
+       sprintf(page,"proc_read_rcvr_stats rcvr=NULL\n");
+    else {
+       page[0] = 0;
+       ep_rcvr_fillout_stats(rcvr,page);
+    }
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, strlen(page)));
+}
+
+static int
+proc_read_rcvr_rail_stats(char *page, char **start, off_t off,
+                    int count, int *eof, void *data)
+{
+    EP_RCVR_RAIL *rcvr_rail = (EP_RCVR_RAIL *)data;
+
+    if (rcvr_rail == NULL) {
+       strcpy(page,"proc_read_rcvr_rail_stats rcvr_rail=NULL");
+    } else {
+       page[0] = 0;
+       ep_rcvr_rail_fillout_stats(rcvr_rail, page);
+       EP_RCVR_OP(rcvr_rail,FillOutRailStats)(rcvr_rail,page);
+    }
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, strlen(page)));
+}
+
+void
+ep_procfs_rcvr_add(EP_RCVR *rcvr)
+{ 
+    /* ep/rcvr/service_number/stats       */
+    /* ep/rcvr/service_number/debug_rcvr  */
+    /* ep/rcvr/service_number/debug_full  */
+    struct proc_dir_entry *p;
+    char str[32];
+
+    sprintf(str,"%d", rcvr->Service);
+
+    rcvr->procfs_root = proc_mkdir (str, ep_procfs_rcvr_root);
+
+    if ((p = create_proc_entry ("stats", 0,  rcvr->procfs_root)) != NULL)
+    {
+       p->write_proc = NULL;
+       p->read_proc  = proc_read_rcvr_stats;
+       p->data       = rcvr;
+       p->owner      = THIS_MODULE;
+    }
+
+    if ((p = create_proc_entry ("debug_rcrv", 0, rcvr->procfs_root)) != NULL)
+    {
+       p->proc_fops = &ep_proc_operations;
+       p->owner     = THIS_MODULE;
+       p->data      = rcvr;
+    }
+
+    if ((p = create_proc_entry ("debug_full", 0, rcvr->procfs_root)) != NULL)
+    {
+       p->proc_fops = &ep_proc_operations;
+       p->owner     = THIS_MODULE;
+       p->data      = rcvr;
+    }
+}
+
+void
+ep_procfs_rcvr_del(EP_RCVR *rcvr)
+{  
+    char str[32];
+    sprintf(str,"%d", rcvr->Service);
+
+    remove_proc_entry ("stats",      rcvr->procfs_root);
+    remove_proc_entry ("debug_rcvr", rcvr->procfs_root);
+    remove_proc_entry ("debug_full", rcvr->procfs_root);
+
+    remove_proc_entry (str, ep_procfs_rcvr_root);
+}
+
+void 
+ep_procfs_rcvr_add_rail(EP_RCVR_RAIL *rcvrRail)
+{
+    /* ep/rcvr/service_number/railN/stats */
+
+    struct proc_dir_entry *p;
+    char str[32];
+    sprintf(str,"rail%d",rcvrRail->CommsRail->Rail->Number);
+
+    rcvrRail->procfs_root = proc_mkdir (str, rcvrRail->Rcvr->procfs_root);
+    
+    if ((p = create_proc_entry ("stats", 0,  rcvrRail->procfs_root)) != NULL)
+    {
+       p->write_proc = NULL;
+       p->read_proc  = proc_read_rcvr_rail_stats;
+       p->data       = rcvrRail;
+       p->owner      = THIS_MODULE;
+    } 
+}
+
+void 
+ep_procfs_rcvr_del_rail(EP_RCVR_RAIL *rcvrRail)
+{
+    char str[32];
+    sprintf(str,"rail%d",rcvrRail->CommsRail->Rail->Number);
+
+    remove_proc_entry ("stats", rcvrRail->procfs_root);
+
+    remove_proc_entry (str, rcvrRail->Rcvr->procfs_root);
+}
+
+
+
+
+static int
+proc_read_xmtr_stats(char *page, char **start, off_t off,
+                    int count, int *eof, void *data)
+{
+    EP_XMTR *xmtr = (EP_XMTR *)data;
+
+    if (xmtr == NULL) 
+       strcpy(page,"proc_read_xmtr_stats xmtr=NULL\n");
+    else {
+       page[0] = 0;
+       ep_xmtr_fillout_stats(xmtr, page);
+    }
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, strlen(page)));
+}
+
+static int
+proc_read_xmtr_rail_stats(char *page, char **start, off_t off,
+                    int count, int *eof, void *data)
+{
+    EP_XMTR_RAIL *xmtr_rail = (EP_XMTR_RAIL *)data;
+
+    if (xmtr_rail == NULL) 
+       strcpy(page,"proc_read_xmtr_rail_stats xmtr_rail=NULL\n");
+    else {
+       page[0] = 0;
+       ep_xmtr_rail_fillout_stats(xmtr_rail, page);
+       EP_XMTR_OP(xmtr_rail,FillOutRailStats)(xmtr_rail,page);
+    }
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, strlen(page)));
+}
+
+void
+ep_procfs_xmtr_add(EP_XMTR *xmtr)
+{ 
+    /* ep/xmtr/service_number/stats       */
+    /* ep/xmtr/service_number/debug_xmtr  */
+    struct proc_dir_entry *p;
+    char str[32];
+
+    sprintf(str,"%llx", (unsigned long long) (unsigned long)xmtr);
+
+    xmtr->procfs_root = proc_mkdir (str, ep_procfs_xmtr_root);
+
+    if ((p = create_proc_entry ("stats", 0,  xmtr->procfs_root)) != NULL)
+    {
+       p->write_proc = NULL;
+       p->read_proc  = proc_read_xmtr_stats;
+       p->data       = xmtr;
+       p->owner      = THIS_MODULE;
+    } 
+
+    if ((p = create_proc_entry ("debug_xmtr", 0, xmtr->procfs_root)) != NULL)
+    {
+       p->proc_fops = &ep_proc_operations;
+       p->owner     = THIS_MODULE;
+       p->data      = xmtr;
+    }
+}
+
+void
+ep_procfs_xmtr_del(EP_XMTR *xmtr)
+{  
+    char str[32];
+    sprintf(str,"%llx", (unsigned long long) (unsigned long)xmtr);
+
+    remove_proc_entry ("stats",      xmtr->procfs_root);
+    remove_proc_entry ("debug_xmtr", xmtr->procfs_root);
+
+    remove_proc_entry (str, ep_procfs_xmtr_root);
+}
+
+void 
+ep_procfs_xmtr_add_rail(EP_XMTR_RAIL *xmtrRail)
+{
+    /* ep/xmtr/service_number/railN/stats */
+    
+    struct proc_dir_entry *p;
+    char str[32];
+    sprintf(str,"rail%d",xmtrRail->CommsRail->Rail->Number);
+
+    xmtrRail->procfs_root = proc_mkdir (str, xmtrRail->Xmtr->procfs_root);
+
+    if ((p = create_proc_entry ("stats", 0,  xmtrRail->procfs_root)) != NULL)
+    {
+       p->write_proc = NULL;
+       p->read_proc  = proc_read_xmtr_rail_stats;
+       p->data       = xmtrRail;
+       p->owner      = THIS_MODULE;
+    } 
+}
+
+void 
+ep_procfs_xmtr_del_rail(EP_XMTR_RAIL *xmtrRail)
+{
+    char str[32];
+    sprintf(str,"rail%d",xmtrRail->CommsRail->Rail->Number);
+
+    remove_proc_entry ("stats", xmtrRail->procfs_root);
+
+    remove_proc_entry (str, xmtrRail->Xmtr->procfs_root);
+}
+
+void
+ep_procfs_rcvr_xmtr_init(void)
+{
+    ep_procfs_rcvr_root = proc_mkdir ("rcvr", ep_procfs_root);
+    ep_procfs_xmtr_root = proc_mkdir ("xmtr", ep_procfs_root); 
+}
+
+void
+ep_procfs_rcvr_xmtr_fini(void)
+{
+    remove_proc_entry ("rcvr", ep_procfs_root);
+    remove_proc_entry ("xmtr", ep_procfs_root);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/kalloc.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kalloc.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kalloc.c  2005-06-01 23:12:54.661429680 -0400
@@ -0,0 +1,677 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kalloc.c,v 1.17.8.2 2004/12/14 10:19:14 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kalloc.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "debug.h"
+
+static void
+HashInPool (EP_ALLOC *alloc, EP_POOL *pool)
+{
+    int idx0 = HASH (pool->Handle.nmh_nmd.nmd_addr);
+    int idx1 = HASH (pool->Handle.nmh_nmd.nmd_addr + pool->Handle.nmh_nmd.nmd_len);
+
+    list_add (&pool->HashBase, &alloc->HashBase[idx0]);
+    list_add (&pool->HashTop, &alloc->HashTop[idx1]);
+}
+
+static void
+HashOutPool (EP_ALLOC *alloc, EP_POOL *pool)
+{
+    list_del (&pool->HashBase);
+    list_del (&pool->HashTop);
+}
+
+static EP_POOL *
+LookupPool (EP_ALLOC *alloc, EP_ADDR addr)
+{
+    struct list_head *el;
+    
+    list_for_each (el, &alloc->HashBase[HASH(addr)]) {
+       EP_POOL *pool = list_entry (el, EP_POOL, HashBase);
+       
+       if (pool->Handle.nmh_nmd.nmd_addr <= addr && addr < (pool->Handle.nmh_nmd.nmd_addr + pool->Handle.nmh_nmd.nmd_len))
+           return (pool);
+    }
+    
+    list_for_each (el, &alloc->HashTop[HASH(addr)]) {
+       EP_POOL *pool = list_entry (el, EP_POOL, HashTop);
+       
+       if (pool->Handle.nmh_nmd.nmd_addr <= addr && addr < (pool->Handle.nmh_nmd.nmd_addr + pool->Handle.nmh_nmd.nmd_len))
+           return (pool);
+    }
+    
+    return (NULL);
+}
+
+static EP_POOL *
+AllocatePool (EP_ALLOC *alloc, EP_ADDR addr, unsigned size, unsigned int perm, EP_ATTRIBUTE attr)
+{
+    EP_ADDR base = 0;
+    EP_POOL *pool;
+    EP_RAIL *rail;
+    int i, railmask = 0;
+    struct list_head *el;
+
+    KMEM_ZALLOC (pool, EP_POOL *, sizeof (EP_POOL), !(attr & EP_NO_SLEEP));
+    
+    if (pool == NULL)
+       return (NULL);
+    
+    if (addr != 0)
+       base = addr;
+    else
+    {
+       for (i = LN2_MIN_SIZE; i <= LN2_MAX_SIZE; i ++)
+       {
+           KMEM_ZALLOC (pool->Bitmaps[i - LN2_MIN_SIZE], bitmap_t *, BT_BITOUL(1 << (LN2_MAX_SIZE-i)) * sizeof (bitmap_t), !(attr & EP_NO_SLEEP));
+           if (pool->Bitmaps[i - LN2_MIN_SIZE] == NULL)
+               goto failed;
+       }
+    
+       if ((base = ep_rmalloc (alloc->ResourceMap, size, !(attr & EP_NO_SLEEP))) == 0)
+           goto failed;
+    }
+
+    switch (alloc->Type)
+    {
+    case EP_ALLOC_TYPE_PRIVATE_SDRAM:
+       rail = alloc->Data.Private.Rail;
+
+       if ((pool->Buffer.Sdram = rail->Operations.SdramAlloc (rail, base, size)) == 0)
+           goto failed;
+
+       ep_perrail_sdram_map (rail, base, pool->Buffer.Sdram, size, perm, attr);
+
+       pool->Handle.nmh_nmd.nmd_addr = base;
+       pool->Handle.nmh_nmd.nmd_len  = size;
+       break;
+       
+    case EP_ALLOC_TYPE_PRIVATE_MAIN:
+        KMEM_GETPAGES(pool->Buffer.Ptr, unsigned long, btop (size), !(attr & EP_NO_SLEEP));
+       if (pool->Buffer.Ptr == 0)
+           goto failed;
+
+       ep_perrail_kaddr_map (alloc->Data.Private.Rail, base, pool->Buffer.Ptr, size, perm, attr);
+
+       pool->Handle.nmh_nmd.nmd_addr = base;
+       pool->Handle.nmh_nmd.nmd_len  = size;
+       break;
+
+    case EP_ALLOC_TYPE_SHARED_MAIN:
+        KMEM_GETPAGES(pool->Buffer.Ptr, unsigned long, btop (size), !(attr & EP_NO_SLEEP));
+       if (pool->Buffer.Ptr == 0)
+           goto failed;
+
+       list_for_each (el, &alloc->Data.Shared.Rails) {
+           EP_RAIL *rail = list_entry (el, EP_RAIL_ENTRY, Link)->Rail;
+
+           ep_perrail_kaddr_map (rail, base, pool->Buffer.Ptr, size, perm, attr);
+
+           railmask |= (1 << rail->Number);
+       }
+       pool->Handle.nmh_nmd.nmd_addr = base;
+       pool->Handle.nmh_nmd.nmd_len  = size;
+       pool->Handle.nmh_nmd.nmd_attr = EP_NMD_ATTR (alloc->Data.Shared.System->Position.pos_nodeid, railmask);
+
+       ep_nmh_insert (&alloc->Data.Shared.System->MappingTable, &pool->Handle);
+       break;
+
+    default:
+       goto failed;
+    }
+
+    return (pool);
+    
+ failed:
+    if (addr == 0 && base)
+       ep_rmfree (alloc->ResourceMap, size, base);
+
+    for (i = LN2_MIN_SIZE; i <= LN2_MAX_SIZE; i ++)
+       if (pool->Bitmaps[i - LN2_MIN_SIZE] != NULL)
+           KMEM_FREE (pool->Bitmaps[i - LN2_MIN_SIZE], BT_BITOUL(1 << (LN2_MAX_SIZE - i)) * sizeof (bitmap_t));
+    
+    KMEM_FREE (pool, sizeof (EP_POOL));
+    return (NULL);
+}
+
+static void
+FreePool (EP_ALLOC *alloc, EP_POOL *pool)
+{
+    struct list_head *el;
+    int i;
+
+    switch (alloc->Type)
+    {
+    case EP_ALLOC_TYPE_PRIVATE_SDRAM:
+       ep_perrail_unmap (alloc->Data.Private.Rail, pool->Handle.nmh_nmd.nmd_addr, pool->Handle.nmh_nmd.nmd_len);
+
+       alloc->Data.Private.Rail->Operations.SdramFree (alloc->Data.Private.Rail, pool->Buffer.Sdram, pool->Handle.nmh_nmd.nmd_len);
+       break;
+       
+    case EP_ALLOC_TYPE_PRIVATE_MAIN:
+       ep_perrail_unmap (alloc->Data.Private.Rail, pool->Handle.nmh_nmd.nmd_addr, pool->Handle.nmh_nmd.nmd_len);
+
+       KMEM_FREEPAGES (pool->Buffer.Ptr, btop (pool->Handle.nmh_nmd.nmd_len));
+       break;
+
+    case EP_ALLOC_TYPE_SHARED_MAIN:
+       ep_nmh_remove (&alloc->Data.Shared.System->MappingTable, &pool->Handle);
+
+       list_for_each (el, &alloc->Data.Shared.Rails) {
+           EP_RAIL *rail = list_entry (el, EP_RAIL_ENTRY, Link)->Rail;
+
+           ep_perrail_unmap (rail, pool->Handle.nmh_nmd.nmd_addr, pool->Handle.nmh_nmd.nmd_len);
+       }
+
+       KMEM_FREEPAGES (pool->Buffer.Ptr, btop (pool->Handle.nmh_nmd.nmd_len));
+       break;
+    }
+    
+    if (pool->Bitmaps[0])
+    {
+       ep_rmfree (alloc->ResourceMap, pool->Handle.nmh_nmd.nmd_len, pool->Handle.nmh_nmd.nmd_addr);
+    
+       for (i = LN2_MIN_SIZE; i <= LN2_MAX_SIZE; i ++)
+           KMEM_FREE (pool->Bitmaps[i - LN2_MIN_SIZE], BT_BITOUL(1 << (LN2_MAX_SIZE - i)) * sizeof (bitmap_t));
+    }
+    
+    KMEM_FREE (pool, sizeof (EP_POOL));
+}
+
+static int
+AddRail (EP_ALLOC *alloc, EP_RAIL *rail)
+{
+    struct list_head *el;
+    EP_RAIL_ENTRY *l;
+    unsigned long flags;
+    int i;
+
+    ASSERT (alloc->Type == EP_ALLOC_TYPE_SHARED_MAIN);
+
+    KMEM_ZALLOC (l, EP_RAIL_ENTRY *, sizeof (EP_RAIL_ENTRY), 1);
+
+    if (l == NULL)
+       return (ENOMEM);
+
+    l->Rail = rail;
+
+    spin_lock_irqsave (&alloc->Lock, flags);
+    for (i = 0; i < NHASH; i++)
+    {
+       list_for_each (el, &alloc->HashBase[i]) {
+           EP_POOL *pool = list_entry (el, EP_POOL, HashBase);
+
+           ep_perrail_kaddr_map (rail, pool->Handle.nmh_nmd.nmd_addr, pool->Buffer.Ptr, 
+                                 pool->Handle.nmh_nmd.nmd_len, EP_PERM_WRITE, EP_NO_SLEEP);
+
+           pool->Handle.nmh_nmd.nmd_attr |= EP_NMD_ATTR (0, 1 << rail->Number);
+       }
+    }
+
+    list_add (&l->Link, &alloc->Data.Shared.Rails);
+
+    spin_unlock_irqrestore (&alloc->Lock, flags); 
+    return (0);
+}
+
+static void
+RemoveRail (EP_ALLOC *alloc, EP_RAIL *rail)
+{
+    struct list_head *el;
+    unsigned long flags;
+    int i;
+
+    spin_lock_irqsave (&alloc->Lock, flags);
+    for (i = 0; i < NHASH; i++)
+    {
+       list_for_each (el, &alloc->HashBase[i]) {
+           EP_POOL *pool = list_entry (el, EP_POOL, HashBase);
+
+           ep_perrail_unmap (rail, pool->Handle.nmh_nmd.nmd_addr, pool->Handle.nmh_nmd.nmd_len);
+
+           pool->Handle.nmh_nmd.nmd_attr &= ~EP_NMD_ATTR (0, 1 << rail->Number);
+       }
+    }
+
+    list_for_each (el, &alloc->Data.Shared.Rails) {
+       EP_RAIL_ENTRY *tmp = list_entry (el, EP_RAIL_ENTRY, Link);
+       if (tmp->Rail == rail)
+       {
+           list_del (el);
+           KMEM_FREE(tmp, sizeof (EP_RAIL_ENTRY));
+           break;
+       }
+    }
+
+    spin_unlock_irqrestore (&alloc->Lock, flags);
+}
+
+static EP_POOL *
+AllocateBlock (EP_ALLOC *alloc, unsigned size, EP_ATTRIBUTE attr, int *offset)
+{
+    int block, j, k;
+    unsigned long flags;
+    EP_POOL *pool;
+
+
+    if (size > MAX_SIZE)
+    {
+       if ((attr & EP_NO_ALLOC) || (pool  = AllocatePool (alloc, 0, size, alloc->Perm, attr)) == NULL)
+           return (NULL);
+
+       spin_lock_irqsave (&alloc->Lock, flags);
+       HashInPool (alloc, pool);
+       spin_unlock_irqrestore (&alloc->Lock, flags);
+
+       *offset = 0;
+
+       return pool;
+    }
+
+    spin_lock_irqsave (&alloc->Lock, flags);
+
+    /* Round up size to next power of 2 */
+    for (k = LN2_MIN_SIZE; (1 << k) < size; k++)
+       ;
+    
+    /* k now has ln2 of the size to allocate. */
+    /* find the free list with the smallest block we can use*/
+    for (j = k; j <= LN2_MAX_SIZE && list_empty (&alloc->Freelists[j - LN2_MIN_SIZE]); j++)
+       ;
+    
+    /* j has ln2 of the smallest size block we can use */
+    if (j < LN2_MAX_SIZE)
+    {
+       int nbits = 1 << (LN2_MAX_SIZE-j);
+       
+       pool  = list_entry (alloc->Freelists[j - LN2_MIN_SIZE].next, EP_POOL, Link[j - LN2_MIN_SIZE]);
+       block = (bt_lowbit (pool->Bitmaps[j - LN2_MIN_SIZE], nbits) << j);
+       
+       BT_CLEAR (pool->Bitmaps[j - LN2_MIN_SIZE], block >> j);
+       
+       if (bt_lowbit (pool->Bitmaps[j - LN2_MIN_SIZE], nbits) == -1)
+           list_del (&pool->Link[j - LN2_MIN_SIZE]);
+    }
+    else
+    {
+       spin_unlock_irqrestore (&alloc->Lock, flags);
+       
+       if ((attr & EP_NO_ALLOC) || (pool  = AllocatePool (alloc, 0, MAX_SIZE, alloc->Perm, attr)) == NULL)
+           return (NULL);
+
+       block = 0;
+       j = LN2_MAX_SIZE;
+       
+       spin_lock_irqsave (&alloc->Lock, flags);
+       
+       HashInPool (alloc, pool);
+    }
+    
+    /* Split it until the buddies are the correct size, putting one
+     * buddy back on the free list and continuing to split the other */
+    while (--j >= k)
+    {
+       list_add (&pool->Link[j - LN2_MIN_SIZE], &alloc->Freelists[j - LN2_MIN_SIZE]);
+       
+       BT_SET (pool->Bitmaps[j - LN2_MIN_SIZE], block >> j);
+       
+       block += (1 << j);
+    }
+    spin_unlock_irqrestore (&alloc->Lock, flags);
+
+    *offset = block;
+
+    return (pool);
+}
+
+static void
+FreeBlock (EP_ALLOC *alloc, EP_ADDR addr, unsigned size)
+{
+    EP_POOL *pool;
+    int  k, block = 0;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&alloc->Lock, flags);
+    /* Round up size to next power of 2 */
+    for (k = LN2_MIN_SIZE; (1 << k) < size; k++)
+       ;
+
+    /* Find the pool containing this block */
+    pool = LookupPool (alloc, addr);
+
+    /* It must exist */
+    ASSERT (pool != NULL);
+
+    /* If we're freeing a subset of it, then update the bitmaps */
+    if (size <= MAX_SIZE)
+    {
+       ASSERT (BT_TEST (pool->Bitmaps[k - LN2_MIN_SIZE], (addr - pool->Handle.nmh_nmd.nmd_addr) >> k) == 0);
+       
+       block = addr - pool->Handle.nmh_nmd.nmd_addr;
+       
+       while (k < LN2_MAX_SIZE && BT_TEST (pool->Bitmaps[k - LN2_MIN_SIZE], (block >> k) ^ 1))
+       {
+           BT_CLEAR (pool->Bitmaps[k - LN2_MIN_SIZE], (block >> k) ^ 1);
+           
+           if (bt_lowbit (pool->Bitmaps[k - LN2_MIN_SIZE], (1 << (LN2_MAX_SIZE - k))) == -1)
+               list_del (&pool->Link[k - LN2_MIN_SIZE]);
+           
+           k++;
+       }
+    }
+
+    if (k >= LN2_MAX_SIZE)
+    {
+       HashOutPool (alloc, pool);
+       spin_unlock_irqrestore (&alloc->Lock, flags);
+
+       FreePool (alloc, pool);
+    }
+    else
+    {
+       if (bt_lowbit (pool->Bitmaps[k - LN2_MIN_SIZE], (1 << (LN2_MAX_SIZE - k))) == -1)
+           list_add (&pool->Link[k - LN2_MIN_SIZE], &alloc->Freelists[k - LN2_MIN_SIZE]);
+
+       BT_SET (pool->Bitmaps[k - LN2_MIN_SIZE], block >> k);
+
+       spin_unlock_irqrestore (&alloc->Lock, flags);
+    }
+}
+
+static void
+InitialiseAllocator (EP_ALLOC *alloc, EP_ALLOC_TYPE type, unsigned int perm, EP_RMAP *rmap)
+{
+    int i;
+
+    spin_lock_init (&alloc->Lock);
+
+    alloc->Type        = type;
+    alloc->ResourceMap = rmap;
+    alloc->Perm        = perm;
+
+    for (i = 0; i < NHASH; i++)
+    {
+       (&alloc->HashBase[i])->next = &alloc->HashBase[i];
+
+       INIT_LIST_HEAD (&alloc->HashBase[i]);
+       INIT_LIST_HEAD (&alloc->HashTop[i]);
+    }
+    
+    for (i = 0; i < NUM_FREELISTS; i++)
+       INIT_LIST_HEAD (&alloc->Freelists[i]);
+}
+
+static void
+DestroyAllocator (EP_ALLOC *alloc)
+{
+    struct list_head *el, *next;
+    int i;
+
+    for (i = 0; i < NHASH; i++)
+    {
+       list_for_each_safe (el, next, &alloc->HashBase[i]) { 
+           EP_POOL *pool = list_entry (el, EP_POOL, HashBase);
+
+           printk ("!!DestroyAllocator: pool=%p type=%d addr=%x len=%x\n", pool, alloc->Type,
+                   pool->Handle.nmh_nmd.nmd_addr, pool->Handle.nmh_nmd.nmd_len);
+
+           list_del (&pool->HashBase);
+           list_del (&pool->HashTop);
+
+           // XXXX: FreePool (alloc, pool);
+       }
+    }
+
+    spin_lock_destroy (&alloc->Lock);
+}
+
+void
+ep_display_alloc (EP_ALLOC *alloc)
+{
+    struct list_head *el;
+    int i;
+    int npools = 0;
+    int nbytes = 0;
+    int nfree = 0;
+    unsigned long flags;
+
+    spin_lock_irqsave (&alloc->Lock, flags);
+
+    ep_debugf (DBG_DEBUG, "Kernel comms memory allocator %p type %d\n", alloc, alloc->Type);
+    for (i = 0; i < NHASH; i++)
+    {
+       list_for_each (el, &alloc->HashBase[i]) {
+           EP_POOL *pool = list_entry (el, EP_POOL, HashBase);
+
+           ep_debugf (DBG_DEBUG, "  POOL %4x: %p -> %x.%x\n", i, pool, pool->Handle.nmh_nmd.nmd_addr,
+                      pool->Handle.nmh_nmd.nmd_addr + pool->Handle.nmh_nmd.nmd_len);
+
+           npools++;
+           nbytes += pool->Handle.nmh_nmd.nmd_len;
+       }
+    }
+    
+    for (i = LN2_MIN_SIZE; i <= LN2_MAX_SIZE; i++)
+    {
+       int n = 0;
+
+       list_for_each (el, &alloc->Freelists[i - LN2_MIN_SIZE]) {
+           EP_POOL *pool  = list_entry (el, EP_POOL, Link[i - LN2_MIN_SIZE]);
+           int      nbits = bt_nbits (pool->Bitmaps[i - LN2_MIN_SIZE], 1 << (LN2_MAX_SIZE - i));
+
+           n += nbits;
+           nfree += (nbits << i);
+       }
+       
+       if (n != 0)
+           ep_debugf (DBG_DEBUG, "  SIZE %5d : num %d\n", (1 << i), n);
+    }
+    ep_debugf (DBG_DEBUG, "%d pools with %d bytes and %d bytes free\n", npools, nbytes, nfree);
+
+    spin_unlock_irqrestore (&alloc->Lock, flags);
+}
+
+/* per-rail allocators */
+void
+ep_alloc_init (EP_RAIL *rail)
+{
+    EP_RMAP *rmap = ep_rmallocmap (EP_PRIVATE_RMAP_SIZE, "PrivateMap", 1);
+
+    ep_rmfree (rmap, EP_PRIVATE_TOP-EP_PRIVATE_BASE, EP_PRIVATE_BASE);
+
+    InitialiseAllocator (&rail->ElanAllocator, EP_ALLOC_TYPE_PRIVATE_SDRAM, EP_PERM_ALL, rmap);
+    InitialiseAllocator (&rail->MainAllocator, EP_ALLOC_TYPE_PRIVATE_MAIN, EP_PERM_WRITE, rmap);
+
+    rail->ElanAllocator.Data.Private.Rail = rail;
+    rail->MainAllocator.Data.Private.Rail = rail;
+}
+
+void
+ep_alloc_fini (EP_RAIL *rail)
+{
+    EP_RMAP *rmap = rail->ElanAllocator.ResourceMap;
+
+    DestroyAllocator (&rail->ElanAllocator);
+    DestroyAllocator (&rail->MainAllocator);
+    
+    ep_rmfreemap (rmap);
+}
+
+sdramaddr_t
+ep_alloc_memory_elan (EP_RAIL *rail, EP_ADDR addr, unsigned size, unsigned int perm, EP_ATTRIBUTE attr)
+{
+    EP_POOL *pool = AllocatePool (&rail->ElanAllocator, addr, size, perm, attr);
+    unsigned long flags;
+
+    if (pool == NULL)
+       return (0);
+
+    spin_lock_irqsave (&rail->ElanAllocator.Lock, flags);
+    HashInPool (&rail->ElanAllocator, pool);
+    spin_unlock_irqrestore (&rail->ElanAllocator.Lock, flags);
+
+    return (pool->Buffer.Sdram);
+}
+
+void
+ep_free_memory_elan (EP_RAIL *rail, EP_ADDR addr)
+{
+    EP_POOL *pool;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->ElanAllocator.Lock, flags);
+    pool = LookupPool (&rail->ElanAllocator, addr);
+    
+    HashOutPool (&rail->ElanAllocator, pool);
+    spin_unlock_irqrestore (&rail->ElanAllocator.Lock, flags);
+    
+    FreePool (&rail->ElanAllocator, pool);
+}
+
+sdramaddr_t
+ep_alloc_elan (EP_RAIL *rail, unsigned size, EP_ATTRIBUTE attr, EP_ADDR *addrp)
+{
+    int             offset;
+    EP_POOL *pool;
+
+    if ((pool = AllocateBlock (&rail->ElanAllocator, size, attr, &offset)) == NULL)
+       return (0);
+    
+    *addrp  = pool->Handle.nmh_nmd.nmd_addr + offset;
+
+    return (pool->Buffer.Sdram + offset);
+}
+
+void
+ep_free_elan (EP_RAIL *rail, EP_ADDR addr, unsigned size)
+{
+    FreeBlock (&rail->ElanAllocator, addr, size);
+}
+
+void *
+ep_alloc_main (EP_RAIL *rail, unsigned size, EP_ATTRIBUTE attr, EP_ADDR *addrp)
+{
+    int             offset;
+    EP_POOL *pool;
+
+    if ((pool = AllocateBlock (&rail->MainAllocator, size, attr, &offset)) == NULL)
+       return (NULL);
+    
+    *addrp  = pool->Handle.nmh_nmd.nmd_addr + offset;
+
+    return ((void *) ((unsigned long) pool->Buffer.Ptr + offset));
+}
+
+void
+ep_free_main (EP_RAIL *rail, EP_ADDR addr, unsigned size)
+{
+    FreeBlock (&rail->MainAllocator, addr, size);
+}
+
+sdramaddr_t
+ep_elan2sdram (EP_RAIL *rail, EP_ADDR addr)
+{
+    EP_POOL    *pool;
+    sdramaddr_t res;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->ElanAllocator.Lock, flags);
+    if ((pool = LookupPool (&rail->ElanAllocator, addr)) == NULL)
+       res = 0;
+    else
+       res = pool->Buffer.Sdram + (addr - pool->Handle.nmh_nmd.nmd_addr);
+    spin_unlock_irqrestore (&rail->ElanAllocator.Lock, flags);
+
+    return (res);
+}
+
+void *
+ep_elan2main (EP_RAIL *rail, EP_ADDR addr)
+{
+    EP_POOL *pool;
+    void *res;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->MainAllocator.Lock, flags);
+    if ((pool = LookupPool (&rail->MainAllocator, addr)) == NULL)
+       res = NULL;
+    else
+       res = (void *) ((unsigned long) pool->Buffer.Ptr + (addr - pool->Handle.nmh_nmd.nmd_addr));
+    spin_unlock_irqrestore (&rail->MainAllocator.Lock, flags);
+
+    return (res);
+}
+
+/* shared allocators */
+int
+ep_shared_alloc_add_rail (EP_SYS *sys, EP_RAIL *rail)
+{
+    return (AddRail (&sys->Allocator, rail));
+}
+
+void
+ep_shared_alloc_remove_rail (EP_SYS *sys, EP_RAIL *rail)
+{
+    RemoveRail (&sys->Allocator, rail);
+}
+
+void
+ep_shared_alloc_init (EP_SYS *sys)
+{
+    EP_RMAP *rmap = ep_rmallocmap (EP_SHARED_RMAP_SIZE, "shared_alloc_map", 1);
+
+    ep_rmfree (rmap, EP_SHARED_TOP - EP_SHARED_BASE, EP_SHARED_BASE);
+
+    InitialiseAllocator (&sys->Allocator, EP_ALLOC_TYPE_SHARED_MAIN, EP_PERM_WRITE, rmap);
+
+    INIT_LIST_HEAD (&sys->Allocator.Data.Shared.Rails);
+
+    sys->Allocator.Data.Shared.System = sys;
+}
+
+void
+ep_shared_alloc_fini (EP_SYS *sys)
+{
+    EP_RMAP *rmap = sys->Allocator.ResourceMap;
+
+    DestroyAllocator (&sys->Allocator);
+
+    ep_rmfreemap (rmap);
+}
+
+void *
+ep_shared_alloc_main (EP_SYS *sys, unsigned size, EP_ATTRIBUTE attr, EP_NMD *nmd)
+{
+    int offset;
+    EP_POOL *pool;
+
+    if ((pool = AllocateBlock (&sys->Allocator, size, attr, &offset)) == NULL)
+       return (NULL);
+
+    ep_nmd_subset (nmd, &pool->Handle.nmh_nmd, offset, size);
+
+    return ((void *) ((unsigned long) pool->Buffer.Ptr + offset));
+}
+
+void
+ep_shared_free_main (EP_SYS *sys, EP_NMD *nmd)
+{
+    FreeBlock (&sys->Allocator, nmd->nmd_addr, nmd->nmd_len);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/kcomm.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kcomm.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kcomm.c   2005-06-01 23:12:54.664429224 -0400
@@ -0,0 +1,1448 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kcomm.c,v 1.50.2.9 2004/12/09 10:02:42 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "cm.h"
+#include "debug.h"
+
+int MaxSwitchLevels = 5;                               /* Max 1024 sized machine */
+
+static char *NodeStateNames[EP_NODE_NUM_STATES] = 
+{
+    "Disconnected",
+    "Connecting",
+    "Connnected",
+    "LeavingConnected",
+    "LocalPassivate",
+    "RemotePassivate",
+    "Passivated",
+    "Disconnecting",
+};
+
+static void
+ep_xid_cache_fill (EP_SYS *sys, EP_XID_CACHE *cache)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&sys->XidLock, flags);
+
+    cache->Current = sys->XidNext;
+    cache->Last    = cache->Current + EP_XID_CACHE_CHUNKS-1;
+
+    sys->XidNext += EP_XID_CACHE_CHUNKS;
+
+    spin_unlock_irqrestore (&sys->XidLock, flags);
+}
+
+EP_XID
+ep_xid_cache_alloc (EP_SYS *sys, EP_XID_CACHE *cache)
+{
+    EP_XID xid;
+    
+    if (cache->Current == cache->Last)
+       ep_xid_cache_fill (sys, cache);
+
+    xid.Generation = sys->XidGeneration;
+    xid.Handle     = cache->Handle;
+    xid.Unique     = cache->Current++;
+
+    return (xid);
+}
+
+void
+ep_xid_cache_init (EP_SYS *sys, EP_XID_CACHE *cache)
+{
+    /* Stall manager thread - it doesn't lock the XidCacheList */
+    ep_kthread_stall (&sys->ManagerThread);
+
+    cache->Handle = ++sys->XidHandle;
+
+    list_add_tail (&cache->Link, &sys->XidCacheList);
+
+    ep_kthread_resume (&sys->ManagerThread);
+}
+
+void
+ep_xid_cache_destroy (EP_SYS *sys, EP_XID_CACHE *cache)
+{
+    /* Stall manager thread - it doesn't lock the XidCacheList */
+    ep_kthread_stall (&sys->ManagerThread);
+
+    list_del (&cache->Link);
+
+    ep_kthread_resume (&sys->ManagerThread);
+}
+
+EP_XID_CACHE *
+ep_xid_cache_find (EP_SYS *sys, EP_XID xid)
+{
+    struct list_head *el;
+
+    list_for_each (el, &sys->XidCacheList) {
+       EP_XID_CACHE *cache = list_entry (el, EP_XID_CACHE, Link);
+
+       if (sys->XidGeneration == xid.Generation && cache->Handle == xid.Handle)
+           return (cache);
+    }
+
+    return (NULL);
+}
+
+static int
+MsgBusy (EP_RAIL *rail, EP_OUTPUTQ *outputq, int slotNum)
+{
+    switch (rail->Operations.OutputQState (rail, outputq, slotNum))
+    {
+    case EP_OUTPUTQ_BUSY:                      /* still busy */
+       return 1;
+       
+    case EP_OUTPUTQ_FAILED:                    /* NACKed */
+    {
+#if defined(DEBUG_PRINTF)
+       EP_MANAGER_MSG *msg = rail->Operations.OutputQMsg (rail, outputq, slotNum);
+
+       EPRINTF4 (DBG_MANAGER, "%s: kcomm msg %d type %d to %d failed\n", rail->Name, slotNum, msg->Hdr.Type, msg->Hdr.DestId);
+#endif
+       break;
+    }
+    
+    case EP_OUTPUTQ_FINISHED:                  /* anything else is finished */
+       break;
+    }
+
+    return 0;
+}
+
+int
+ep_send_message (EP_RAIL *rail, int nodeId, int type, EP_XID xid, EP_MANAGER_MSG_BODY *body)
+{
+    EP_SYS         *sys  = rail->System;
+    EP_NODE        *node = &sys->Nodes[nodeId];
+    int             n    = EP_MANAGER_OUTPUTQ_SLOTS;
+    int             slotNum;
+    int             rnum;
+    EP_RAIL        *msgRail;
+    EP_MANAGER_MSG *msg;
+    unsigned long   flags;
+
+    ASSERT (! EP_XID_INVALID (xid));
+
+    if ((rnum = ep_pickRail (node->ConnectedRails)) >= 0)
+       msgRail = sys->Rails[rnum];
+    else
+    {
+       if (EP_MANAGER_MSG_TYPE_CONNECTED(type))
+       {
+           ep_debugf (DBG_MANAGER, "%s: no rails available, trying to send type %d to %d\n", rail->Name, type, nodeId);
+           return -EHOSTDOWN;
+       }
+
+       ep_debugf (DBG_MANAGER, "%s: no rails connected to %d - using receiving rail\n", rail->Name, nodeId);
+
+       msgRail = rail;
+    }
+    
+
+    spin_lock_irqsave (&msgRail->ManagerOutputQLock, flags);
+
+    slotNum = msgRail->ManagerOutputQNextSlot;
+
+    while (n-- > 0 && MsgBusy (msgRail, msgRail->ManagerOutputQ, slotNum))             /* search for idle message buffer */
+    {
+       if (++(msgRail->ManagerOutputQNextSlot) == EP_MANAGER_OUTPUTQ_SLOTS)
+           msgRail->ManagerOutputQNextSlot = 0;
+      
+       slotNum = msgRail->ManagerOutputQNextSlot;
+    }
+
+    if (n == 0)                                                        /* all message buffers busy */
+    {
+       spin_unlock_irqrestore (&msgRail->ManagerOutputQLock, flags);
+
+       ep_debugf (DBG_MANAGER, "%s: all message buffers busy: trying to send type %d to %d\n", msgRail->Name, type, nodeId);
+       return -EBUSY;
+    }
+
+    msg = msgRail->Operations.OutputQMsg (msgRail, msgRail->ManagerOutputQ, slotNum);
+    
+    EPRINTF7 (DBG_MANAGER, "%s: ep_send_message: type=%d nodeId=%d rail=%d xid=%08x.%08x.%016llx\n", 
+             msgRail->Name, type, nodeId, rail->Number, xid.Generation, xid.Handle, (long long) xid.Unique);
+
+    msg->Hdr.Version    = EP_MANAGER_MSG_VERSION;
+    msg->Hdr.Type       = type;
+    msg->Hdr.Rail       = rail->Number;
+    msg->Hdr.NodeId     = msgRail->Position.pos_nodeid;
+    msg->Hdr.DestId     = nodeId;
+    msg->Hdr.Xid        = xid;
+    msg->Hdr.Checksum   = 0;
+
+    if (body) bcopy (body, &msg->Body, sizeof (EP_MANAGER_MSG_BODY));
+
+    msg->Hdr.Checksum = CheckSum ((char *) msg, EP_MANAGER_MSG_SIZE);
+
+    if (rail->Operations.OutputQSend (msgRail, msgRail->ManagerOutputQ, slotNum, EP_MANAGER_MSG_SIZE,
+                                     nodeId, EP_SYSTEMQ_MANAGER, EP_MANAGER_OUTPUTQ_RETRIES) < 0)
+       IncrStat (msgRail, SendMessageFailed);
+    
+    if (++(msgRail->ManagerOutputQNextSlot) == EP_MANAGER_OUTPUTQ_SLOTS) /* check this one last next time */
+       msgRail->ManagerOutputQNextSlot = 0;
+
+    spin_unlock_irqrestore (&msgRail->ManagerOutputQLock, flags);
+
+    return 0;
+}
+
+void
+ep_panic_node (EP_SYS *sys, int nodeId, unsigned char *reason)
+{
+    EP_NODE            *node = &sys->Nodes[nodeId];
+    EP_MANAGER_MSG_BODY body;
+    EP_XID              xid;
+    kcondvar_t          sleep;
+    int                 rnum;
+    unsigned long       flags;
+
+    if (nodeId > sys->Position.pos_nodes)
+       return;
+
+    strncpy (body.PanicReason, reason, sizeof (body.PanicReason));
+
+    kcondvar_init (&sleep);
+    spin_lock_irqsave (&sys->NodeLock, flags);
+    for (;;)
+    {
+       if (node->ConnectedRails == 0)
+           break;
+
+       for (rnum = 0; rnum < EP_MAX_RAILS; rnum++)
+           if (node->ConnectedRails & (1 << rnum))
+               break;
+
+       xid = ep_xid_cache_alloc(sys, &sys->Rails[rnum]->XidCache);
+       
+       if (ep_send_message (sys->Rails[rnum], nodeId, EP_MANAGER_MSG_TYPE_REMOTE_PANIC, xid, &body) == 0)
+           break;
+
+       if (kcondvar_timedwaitsig (&sleep, &sys->NodeLock, &flags, lbolt + hz) == CV_RET_SIGPENDING)
+           break;
+    }
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+    kcondvar_destroy (&sleep);
+}
+
+static void
+ProcessNeterrRequest (EP_RAIL *msgRail, EP_RAIL *rail, EP_MANAGER_MSG *msg)
+{
+    EPRINTF4 (DBG_NETWORK_ERROR, "%s: process neterr request - node %d cookies %llx %llx\n", rail->Name, msg->Hdr.NodeId, msg->Body.Cookies[0], msg->Body.Cookies[1]);
+
+    rail->Operations.NeterrFixup (rail, msg->Hdr.NodeId, msg->Body.Cookies);
+    
+    ep_send_message (rail, msg->Hdr.NodeId, EP_MANAGER_MSG_TYPE_NETERR_RESPONSE, msg->Hdr.Xid, &msg->Body);
+}
+
+
+static void
+ProcessNeterrResponse (EP_RAIL *msgRail, EP_RAIL *rail, EP_MANAGER_MSG *msg)
+{
+    EP_SYS       *sys      = rail->System;
+    EP_NODE_RAIL *nodeRail = &rail->Nodes[msg->Hdr.NodeId];
+    unsigned long flags;
+
+    EPRINTF4 (DBG_NETWORK_ERROR, "%s: process neterr response - node %d cookies %llx %llx\n", rail->Name, msg->Hdr.NodeId, msg->Body.Cookies[0], msg->Body.Cookies[1]);
+
+    spin_lock_irqsave (&sys->NodeLock, flags);
+    if (EP_XIDS_MATCH (nodeRail->MsgXid, msg->Hdr.Xid))
+    {
+       EP_INVALIDATE_XID (nodeRail->MsgXid);
+
+       if (nodeRail->NetworkErrorCookies[0] != 0 && nodeRail->NetworkErrorCookies[0] == msg->Body.Cookies[0])
+           nodeRail->NetworkErrorCookies[0] = 0;
+
+       if (nodeRail->NetworkErrorCookies[1] != 0 && nodeRail->NetworkErrorCookies[1] == msg->Body.Cookies[1])
+           nodeRail->NetworkErrorCookies[1] = 0;
+       
+       if (nodeRail->NetworkErrorCookies[0] == 0 && nodeRail->NetworkErrorCookies[1] == 0)
+           nodeRail->NetworkErrorState &= ~EP_NODE_NETERR_ATOMIC_PACKET;
+    }
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+}
+
+
+static void
+ProcessGetNodeState (EP_RAIL *msgRail, EP_RAIL *rail, EP_MANAGER_MSG *msg)
+{
+    EP_NODE_RAIL *nodeRail = &rail->Nodes[msg->Hdr.NodeId];
+    unsigned int service = msg->Body.Service;
+
+    EPRINTF5 (DBG_MANAGER, "%s: ProcessGetNodeState: %s - %d %s%s\n",  msgRail->Name, rail->Name, msg->Hdr.NodeId,
+             NodeStateNames[nodeRail->State], nodeRail->NetworkErrorState ? " (NetworkError)" : "");
+
+    msg->Body.NodeState.State             = nodeRail->State;
+    msg->Body.NodeState.NetworkErrorState = nodeRail->NetworkErrorState;
+    msg->Body.NodeState.Railmask          = ep_rcvr_railmask (rail->System, service);
+
+    if (ep_send_message (rail, msg->Hdr.NodeId, EP_MANAGER_MSG_TYPE_GET_NODE_STATE_RESPONSE, msg->Hdr.Xid, &msg->Body) < 0)
+       printk ("%s: get node state for %s[%d] - failed to send response\n", msgRail->Name, rail->Name, msg->Hdr.NodeId);
+}
+
+static void
+ProcessFlushRequest (EP_RAIL *msgRail, EP_RAIL *rail, EP_MANAGER_MSG *msg)
+{
+    EP_NODE_RAIL *nodeRail = &rail->Nodes[msg->Hdr.NodeId];
+
+    EPRINTF5 (DBG_MANAGER, "%s: ProcessFlushRequest: %s - %d %s%s\n",  msgRail->Name, rail->Name, msg->Hdr.NodeId,
+             NodeStateNames[nodeRail->State], nodeRail->NetworkErrorState ? " (NetworkError)" : "");
+
+    switch (nodeRail->State)
+    {
+    case EP_NODE_REMOTE_PASSIVATE:
+       nodeRail->NextRunTime = lbolt + MSGBUSY_RETRY_TIME;     /* retransmit our flush request quickly */
+       EPRINTF3 (DBG_MANAGER, "%s: ProcessFlushRequest: NextRunTime -> %lx (%lx)\n", rail->Name, nodeRail->NextRunTime, lbolt);
+       /* DROPTHROUGH */
+
+    case EP_NODE_PASSIVATED:
+    case EP_NODE_DISCONNECTED:
+       if (nodeRail->NetworkErrorState != 0)
+           break;
+
+       if (ep_send_message (rail, msg->Hdr.NodeId, EP_MANAGER_MSG_TYPE_FLUSH_RESPONSE, msg->Hdr.Xid, NULL) < 0)
+           printk ("%s: flush request for %s[%d] - failed to send response\n", msgRail->Name, rail->Name, msg->Hdr.NodeId);
+       break;
+       
+    default:
+       EPRINTF4 (DBG_MANAGER, "%s: flush request for %s[%d] - node not in approriate state - %s\n", msgRail->Name, rail->Name, msg->Hdr.NodeId, NodeStateNames[nodeRail->State]);
+       break;
+    }
+}
+
+static void
+ProcessFlushResponse (EP_RAIL *msgRail, EP_RAIL *rail, EP_MANAGER_MSG *msg)
+{
+    EP_NODE_RAIL *nodeRail= &rail->Nodes[msg->Hdr.NodeId];
+
+    EPRINTF5 (DBG_MANAGER, "%s: ProcessFlushResponse: %s - %d %s%s\n",  msgRail->Name, rail->Name, msg->Hdr.NodeId,
+             NodeStateNames[nodeRail->State], EP_XIDS_MATCH (nodeRail->MsgXid, msg->Hdr.Xid) ? " (XIDS match)" : "");
+
+    if (nodeRail->State == EP_NODE_REMOTE_PASSIVATE && EP_XIDS_MATCH(nodeRail->MsgXid, msg->Hdr.Xid))
+    {
+       EP_INVALIDATE_XID (nodeRail->MsgXid);
+
+       printk ("%s: flush response from %d - move to passivated list\n", rail->Name, msg->Hdr.NodeId);
+       list_del (&nodeRail->Link);
+
+       /* Node is now passivated - attempt to failover  messages */
+       list_add_tail (&nodeRail->Link, &rail->PassivatedList);
+       nodeRail->State = EP_NODE_PASSIVATED;
+    }
+    else
+    {
+       printk ("%s: flush response from %d - not passivating (%s) or XIDs mismatch (%llx %llx)\n", rail->Name, 
+               msg->Hdr.NodeId, NodeStateNames[nodeRail->State], (long long) nodeRail->MsgXid.Unique, (long long) msg->Hdr.Xid.Unique);
+    }
+}
+
+static void
+ProcessMapNmdRequest (EP_RAIL *msgRail, EP_RAIL *rail, EP_MANAGER_MSG *msg)
+{
+    EP_SYS          *sys     = rail->System;
+    EP_MAP_NMD_BODY *msgBody = &msg->Body.MapNmd;
+    int              i;
+
+    EPRINTF4 (DBG_MANAGER, "%s: Map NMD request from %d for %d NMDs to railmask %x\n", rail->Name, msg->Hdr.NodeId, msgBody->nFrags, msgBody->Railmask);
+    
+    for (i = 0; i < msgBody->nFrags; i++)
+       ep_nmd_map_rails (sys, &msgBody->Nmd[i], msgBody->Railmask);
+    
+    /* Must flush TLBs before responding */
+    for (i = 0; i < EP_MAX_RAILS; i++)
+       if (sys->Rails[i] && sys->Rails[i]->TlbFlushRequired)
+           ep_perrail_dvma_sync (sys->Rails[i]);
+
+    if (ep_send_message (rail, msg->Hdr.NodeId, EP_MANAGER_MSG_TYPE_MAP_NMD_RESPONSE, msg->Hdr.Xid, &msg->Body) < 0)
+       printk ("%s: map nmd request for %s[%d] - failed to send response\n", msgRail->Name, rail->Name, msg->Hdr.NodeId);
+}
+
+static void
+ProcessXidMessage (EP_RAIL *msgRail, EP_MANAGER_MSG *msg, EP_XID xid)
+{
+    EP_XID_CACHE *xidCache = ep_xid_cache_find (msgRail->System, xid);
+
+    EPRINTF6 (DBG_MANAGER, "%s: ProcessXidMessage: XID=%08x.%0x8.%016llx -> %p(%p)\n",
+             msgRail->Name, xid.Generation, xid.Handle, (long long) xid.Unique,
+             xidCache  ? xidCache->MessageHandler : 0, xidCache  ? xidCache->Arg : 0);
+    
+    if (xidCache != NULL)
+       xidCache->MessageHandler (xidCache->Arg, msg);
+}
+
+static void
+ProcessMessage (EP_RAIL *msgRail, void *arg, void *msgbuf)
+{
+    EP_SYS         *sys  = msgRail->System;
+    EP_MANAGER_MSG *msg  = (EP_MANAGER_MSG *) msgbuf;
+    uint16_t        csum = msg->Hdr.Checksum;
+    EP_RAIL        *rail;
+
+    if (msg->Hdr.Version != EP_MANAGER_MSG_VERSION)
+       return;
+
+    msg->Hdr.Checksum= 0;
+    if (CheckSum ((char *) msg, EP_MANAGER_MSG_SIZE) != csum)
+    {
+       printk ("%s: checksum failed on msg from %d (%d) (%x != %x) ?\n", msgRail->Name, msg->Hdr.NodeId, msg->Hdr.Type, csum, CheckSum ((char *) msg, EP_MANAGER_MSG_SIZE));
+       return;
+    }
+
+    if ((rail = sys->Rails[msg->Hdr.Rail]) == NULL)
+    {
+       printk ("%s: rail no longer exists for msg from %d?\n", msgRail->Name, msg->Hdr.NodeId);
+       return;
+    }
+
+    EPRINTF7 (DBG_MANAGER, "%s: ProcessMessage (%s) type=%d node=%d XID=%08x.%0x8.%016llx\n", 
+             msgRail->Name, rail->Name, msg->Hdr.Type, msg->Hdr.NodeId,
+             msg->Hdr.Xid.Generation, msg->Hdr.Xid.Handle, msg->Hdr.Xid.Unique);
+
+    switch (msg->Hdr.Type)
+    {
+    case EP_MANAGER_MSG_TYPE_REMOTE_PANIC:
+       msg->Body.PanicReason[EP_PANIC_STRLEN] = '\0';          /* ensure string terminated */
+
+       printk ("%s: remote panic call from elan node %d - %s\n", msgRail->Name, msg->Hdr.NodeId, msg->Body.PanicReason);
+       panic ("ep: remote panic request\n");
+       break;
+
+    case EP_MANAGER_MSG_TYPE_NETERR_REQUEST:
+       ProcessNeterrRequest (msgRail, rail, msg);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_NETERR_RESPONSE:
+       ProcessNeterrResponse (msgRail, rail, msg);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_FLUSH_REQUEST:
+       ProcessFlushRequest (msgRail, rail, msg);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_FLUSH_RESPONSE:
+       ProcessFlushResponse (msgRail, rail, msg);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_MAP_NMD_REQUEST:
+       ProcessMapNmdRequest (msgRail, rail, msg);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_MAP_NMD_RESPONSE:
+       ProcessXidMessage (msgRail, msg, msg->Hdr.Xid);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_FAILOVER_REQUEST:
+       ProcessXidMessage (msgRail, msg, msg->Body.Failover.Xid);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_FAILOVER_RESPONSE:
+       ProcessXidMessage (msgRail, msg, msg->Hdr.Xid);
+       break;
+       
+    case EP_MANAGER_MSG_TYPE_GET_NODE_STATE:
+       ProcessGetNodeState (msgRail, rail, msg);
+       break;
+
+    case EP_MANAGER_MSG_TYPE_GET_NODE_STATE_RESPONSE: 
+       ProcessXidMessage (msgRail, msg, msg->Hdr.Xid);
+       break;
+
+    default:
+       printk ("%s: Unknown message type %d from %d\n", msgRail->Name, msg->Hdr.Type, msg->Hdr.NodeId);
+       break;
+    }
+}
+
+
+static void
+ManagerQueueEvent (EP_RAIL *rail, void *arg)
+{
+    ep_kthread_schedule ((EP_KTHREAD *) arg, lbolt);
+}
+
+void
+UpdateConnectionState (EP_RAIL *rail, statemap_t *map)
+{
+    EP_SYS *sys = rail->System;
+    bitmap_t seg;
+    int offset, nodeId;
+    unsigned long flags;
+    
+    while ((offset = statemap_findchange (map, &seg, 1)) >= 0)
+    {
+       for (nodeId = offset; nodeId < (offset + BT_NBIPUL) && nodeId < rail->Position.pos_nodes; nodeId++)
+       {
+           EP_NODE      *node     = &sys->Nodes[nodeId];
+           EP_NODE_RAIL *nodeRail = &rail->Nodes[nodeId];
+
+           if (statemap_getbits (map, nodeId, 1))
+           {
+               spin_lock_irqsave (&sys->NodeLock, flags);
+
+               switch (nodeRail->State)
+               {
+               case EP_NODE_DISCONNECTED:
+                   EPRINTF2 (DBG_MANAGER, "%s: Node %d -> Disconnected \n", rail->Name, nodeId);
+                   break;
+
+               case EP_NODE_CONNECTING:
+                   EPRINTF2 (DBG_MANAGER, "%s: Node %d -> Connect\n", rail->Name, nodeId);
+                   
+                   /* load the route table entry *before*  setting the state
+                    * to connected, since DMA's can be initiated as soon as
+                    * the node is marked as connected */
+                   rail->Operations.LoadNodeRoute (rail, nodeId);
+                   
+                   nodeRail->State = EP_NODE_CONNECTED;
+                   
+                   statemap_setbits (rail->NodeSet, nodeId, 1, 1);
+                   if (statemap_getbits (sys->NodeSet, nodeId, 1) == 0)
+                       statemap_setbits (sys->NodeSet, nodeId, 1, 1);
+
+                   /* Add to rails connected to this node */
+                   node->ConnectedRails |= (1 << rail->Number);
+
+                   /* Finally lower the per-node context filter */
+                   rail->Operations.LowerFilter (rail, nodeId);
+                   break;
+                   
+               case EP_NODE_LEAVING_CONNECTED:
+                   EPRINTF2 (DBG_MANAGER, "%s: Node %d -> Local Passivate\n", rail->Name, nodeId);
+                   
+                   /* Raise the per-node context filter */
+                   rail->Operations.RaiseFilter (rail, nodeId);
+
+                   /* If it's resolving network errors it will be on the NodeNeterrList,
+                    * remove if from this list before placing it on the LocalPassivateList
+                    * as we'll resolve the network error later in RemotePassivate */
+                   if (nodeRail->NetworkErrorState)
+                       list_del (&nodeRail->Link);
+
+                   list_add_tail (&nodeRail->Link, &rail->LocalPassivateList);
+                   nodeRail->State = EP_NODE_LOCAL_PASSIVATE;
+
+                   /* Remove from rails connected to this node */
+                   node->ConnectedRails &= ~(1 << rail->Number);
+                   break;
+
+               default:
+                   printk ("%s: Node %d - in NodeChangeMap with state %d\n", rail->Name, nodeId, nodeRail->State);
+                   panic ("Node in NodeChangeMap with invalid state\n");
+                   break;
+               }
+               spin_unlock_irqrestore (&sys->NodeLock, flags);
+           }
+       }
+    }
+}
+
+void
+ProgressNetworkError (EP_RAIL *rail, EP_NODE_RAIL *nodeRail)
+{
+    EP_SYS             *sys    = rail->System;
+    int                 nodeId = nodeRail - rail->Nodes;
+    EP_MANAGER_MSG_BODY msg;
+
+    ASSERT (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_REMOTE_PASSIVATE);
+
+    if (BEFORE (lbolt, nodeRail->NextRunTime))
+       return;
+
+    if (nodeRail->NetworkErrorState & EP_NODE_NETERR_DMA_PACKET)
+       nodeRail->NetworkErrorState &= ~EP_NODE_NETERR_DMA_PACKET;
+    
+    if (nodeRail->NetworkErrorState & EP_NODE_NETERR_ATOMIC_PACKET)
+    {
+       if (EP_XID_INVALID (nodeRail->MsgXid))
+           nodeRail->MsgXid = ep_xid_cache_alloc (sys, &rail->XidCache);
+
+       msg.Cookies[0] = nodeRail->NetworkErrorCookies[0];
+       msg.Cookies[1] = nodeRail->NetworkErrorCookies[1];
+       
+       EPRINTF4 (DBG_NETWORK_ERROR, "%s: progress neterr - node %d cookies %llx %llx\n", rail->Name, nodeId, msg.Cookies[0], msg.Cookies[1]);
+
+       if (ep_send_message (rail, nodeId, EP_MANAGER_MSG_TYPE_NETERR_REQUEST, nodeRail->MsgXid, &msg) == 0)
+           nodeRail->NextRunTime = lbolt + MESSAGE_RETRY_TIME;
+       else
+           nodeRail->NextRunTime = lbolt + MSGBUSY_RETRY_TIME;
+    }
+}
+
+long
+ProgressNodeLists (EP_RAIL *rail, long nextRunTime)
+{
+    EP_SYS           *sys = rail->System;
+    struct list_head *el, *nel;
+    unsigned long flags;
+
+    spin_lock_irqsave (&sys->NodeLock, flags);
+    list_for_each_safe (el, nel, &rail->NetworkErrorList) {
+       EP_NODE_RAIL *nodeRail = list_entry (el, EP_NODE_RAIL, Link);
+       int           nodeId   = nodeRail - rail->Nodes;
+
+       ProgressNetworkError (rail, nodeRail);
+
+       if (nodeRail->NetworkErrorState == 0)
+       {
+           EPRINTF2 (DBG_NETWORK_ERROR, "%s: lower context filter for node %d due to network error\n", rail->Name, nodeId);
+           printk ("%s: lower context filter for node %d due to network error\n", rail->Name, nodeId);
+
+           rail->Operations.LowerFilter (rail, nodeId);
+
+           list_del (&nodeRail->Link);
+           continue;
+       }
+       
+       if (nextRunTime == 0 || AFTER (nextRunTime, nodeRail->NextRunTime))
+           nextRunTime = nodeRail->NextRunTime;
+    }
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+
+    if (! list_empty (&rail->LocalPassivateList))
+    {
+       EPRINTF1 (DBG_MANAGER, "%s: Locally Passivating Nodes\n", rail->Name);
+       
+       /* We have disconnected from some nodes or have left ourselves
+        * flush through all communications and determine whether we
+        * need to perform rail failover */
+       rail->Operations.FlushFilters (rail);
+       
+       ep_call_callbacks (rail, EP_CB_FLUSH_FILTERING, rail->NodeSet);
+
+       rail->Operations.FlushQueues (rail);
+
+       ep_call_callbacks (rail, EP_CB_FLUSH_FLUSHING, rail->NodeSet);
+
+       while (! list_empty (&rail->LocalPassivateList))
+       {
+           EP_NODE_RAIL *nodeRail = list_entry (rail->LocalPassivateList.next, EP_NODE_RAIL, Link);
+           int           nodeId   = nodeRail - rail->Nodes;
+
+           list_del (&nodeRail->Link);
+
+           rail->Operations.UnloadNodeRoute (rail, nodeId);
+           
+           if (nodeRail->NetworkErrorState == 0 && nodeRail->MessageState == 0)
+           {
+               EPRINTF2 (DBG_MANAGER, "%s: Node %d -> Disconnecting\n", rail->Name, nodeId);
+
+               list_add_tail (&nodeRail->Link, &rail->DisconnectingList);
+               nodeRail->State = EP_NODE_DISCONNECTING;
+           }
+           else
+           {
+               EPRINTF2 (DBG_MANAGER, "%s: Node %d -> Remote Passivate\n", rail->Name, nodeId);
+
+               list_add_tail (&nodeRail->Link, &rail->RemotePassivateList);
+               nodeRail->State = EP_NODE_REMOTE_PASSIVATE;
+
+               if (nodeRail->NetworkErrorState == 0)
+                   nodeRail->NextRunTime = lbolt;
+           }
+       }
+
+       ep_call_callbacks (rail, EP_CB_PASSIVATED, rail->NodeSet);
+    }
+
+    list_for_each_safe (el, nel, &rail->RemotePassivateList) {
+       EP_NODE_RAIL *nodeRail = list_entry (el, EP_NODE_RAIL, Link);
+       int           nodeId   = nodeRail - rail->Nodes;
+       EP_NODE      *node     = &sys->Nodes[nodeId];
+
+       if (node->ConnectedRails == 0)                          /* no rails connected to this node (anymore) */
+       {
+           /*  Remove from this list */
+           list_del (&nodeRail->Link);
+
+           EPRINTF2 (DBG_MANAGER, "%s: Node %d, no rails, Remote Passivate -> Disconnecting\n", rail->Name, nodeId);
+
+           /* transition towards disconnected */
+           list_add_tail (&nodeRail->Link, &rail->DisconnectingList);
+           nodeRail->State = EP_NODE_DISCONNECTING;
+           continue;
+       }
+
+       EPRINTF6 (DBG_MANAGER, "%s: Node %d - %s NetworkErrorState=%x NextRunTime=%lx (%lx)\n",
+                 rail->Name, nodeId, NodeStateNames[nodeRail->State], nodeRail->NetworkErrorState,
+                 nodeRail->NextRunTime, nextRunTime);
+
+       if (nodeRail->NetworkErrorState)
+       {
+           ProgressNetworkError (rail, nodeRail);
+       }
+       else if (! BEFORE (lbolt, nodeRail->NextRunTime))
+       {
+           if (EP_XID_INVALID (nodeRail->MsgXid))
+               nodeRail->MsgXid = ep_xid_cache_alloc (sys, &rail->XidCache);
+
+           if (ep_send_message (rail, nodeId, EP_MANAGER_MSG_TYPE_FLUSH_REQUEST, nodeRail->MsgXid, NULL) == 0)
+               nodeRail->NextRunTime = lbolt + MESSAGE_RETRY_TIME;
+           else
+               nodeRail->NextRunTime = lbolt + MSGBUSY_RETRY_TIME;
+       }
+
+       if (nextRunTime == 0 || AFTER (nextRunTime, nodeRail->NextRunTime))
+           nextRunTime = nodeRail->NextRunTime;
+    }
+    
+    if (! list_empty (&rail->PassivatedList)) 
+    {
+       ep_call_callbacks (rail, EP_CB_FAILOVER, rail->NodeSet);
+
+       list_for_each_safe (el, nel, &rail->PassivatedList) {
+           EP_NODE_RAIL *nodeRail = list_entry (rail->PassivatedList.next, EP_NODE_RAIL, Link);
+           int           nodeId   = nodeRail - rail->Nodes;
+           EP_NODE      *node     = &sys->Nodes[nodeId];
+
+           ASSERT (nodeRail->NetworkErrorState == 0);
+
+           if (node->ConnectedRails == 0)
+           {
+               /*  Remove from this list */
+               list_del (&nodeRail->Link);
+
+               EPRINTF2 (DBG_MANAGER, "%s: Node %d, no rails, Passivated -> Disconnecting\n", rail->Name, nodeId);
+
+               /* transition towards disconnected */
+               list_add_tail (&nodeRail->Link, &rail->DisconnectingList);
+               nodeRail->State = EP_NODE_DISCONNECTING;
+               continue;
+           }
+           
+           EPRINTF6 (DBG_MANAGER, "%s: Node %d - %s NetworkErrorState=%x NextRunTime=%lx (%lx)\n",
+                     rail->Name, nodeId, NodeStateNames[nodeRail->State], nodeRail->NetworkErrorState,
+                     nodeRail->NextRunTime, nextRunTime);
+
+           if (nodeRail->MessageState == 0)
+           {
+               EPRINTF2 (DBG_MANAGER, "%s: Node %d, no messages, Passivated -> Disconnecting\n", rail->Name,nodeId);
+
+               list_del (&nodeRail->Link);
+               list_add_tail (&nodeRail->Link, &rail->DisconnectingList);
+               nodeRail->State = EP_NODE_DISCONNECTING;
+               continue;
+           }
+
+           nodeRail->MessageState = 0;
+           nodeRail->NextRunTime  = lbolt + FAILOVER_RETRY_TIME;
+
+           if (nextRunTime == 0 || AFTER (nextRunTime, nodeRail->NextRunTime))
+               nextRunTime = nodeRail->NextRunTime;
+       }
+    }
+
+    if (! list_empty (&rail->DisconnectingList))
+    {
+       ep_call_callbacks (rail, EP_CB_DISCONNECTING, rail->NodeSet);
+
+       while (! list_empty (&rail->DisconnectingList))
+       {
+           EP_NODE_RAIL *nodeRail = list_entry (rail->DisconnectingList.next, EP_NODE_RAIL, Link);
+           int           nodeId   = nodeRail - rail->Nodes;
+           EP_NODE      *node     = &sys->Nodes[nodeId];
+
+           EPRINTF2 (DBG_MANAGER, "%s: Node %d, Disconnecting -> Disconnected\n", rail->Name, nodeId);
+
+           list_del (&nodeRail->Link);
+
+           rail->Operations.NodeDisconnected (rail, nodeId);
+
+           /* Clear the network error state */
+           nodeRail->NextRunTime            = 0;
+           nodeRail->NetworkErrorState      = 0;
+           nodeRail->NetworkErrorCookies[0] = 0;
+           nodeRail->NetworkErrorCookies[1] = 0;
+
+           /* Clear the message state */
+           nodeRail->MessageState = 0;
+
+           cm_node_disconnected (rail, nodeId);
+
+           nodeRail->State = EP_NODE_DISCONNECTED;
+           
+           statemap_setbits (rail->NodeSet, nodeId, 0, 1);
+
+           if (node->ConnectedRails == 0)
+               statemap_setbits (sys->NodeSet, nodeId, 0, 1);
+       }
+
+       ep_call_callbacks (rail, EP_CB_DISCONNECTED, rail->NodeSet);
+    }
+
+    return (nextRunTime);
+}
+
+void
+DisplayNodes (EP_RAIL *rail)
+{
+    EP_SYS *sys = rail->System;
+    int i, state, count;
+    unsigned long flags;
+
+    spin_lock_irqsave (&sys->NodeLock, flags);
+
+    for (state = 0; state < EP_NODE_NUM_STATES; state++)
+    {
+       for (count = i = 0; i < rail->Position.pos_nodes; i++)
+       {
+           ASSERT (rail->Nodes[i].State < EP_NODE_NUM_STATES);
+
+           if (rail->Nodes[i].State == state)
+               if (state != EP_NODE_DISCONNECTED)
+                   printk ("%s %d", !count++ ? NodeStateNames[state] : "", i);
+       }
+       if (count)
+           printk ("%s (%d total)\n", state == EP_NODE_DISCONNECTED ? NodeStateNames[state] : "", count);
+    }
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+}
+
+static void
+PositionFound (EP_RAIL *rail, ELAN_POSITION *pos)
+{
+    EP_SYS           *sys = rail->System;
+    struct list_head *el;
+    int i;
+
+    /* only called from the ep_managage whilst rail->State == EP_RAIL_STATE_STARTED */
+    ASSERT ( rail->State == EP_RAIL_STATE_STARTED );
+
+#if defined(PER_CPU_TIMEOUT)
+    /*
+     * On Tru64 - if we're running in a "funnelled" thread, then we will be 
+     * unable to start the per-cpu timeouts, so if we return then eventually
+     * the ep_manager() thread will find the network position and we're
+     * in control of our own destiny.
+     */
+    if (THREAD_IS_FUNNELED(current_thread()))
+    {
+       ep_kthread_schedule (&sys->ManagerThread, lbolt);
+       return;
+    }
+#endif
+
+    sprintf (rail->Name, "ep%d[%d]", rail->Number, pos->pos_nodeid);
+
+    if (pos->pos_levels > MaxSwitchLevels)
+    {
+       for (i = 0; i < (pos->pos_levels - MaxSwitchLevels); i++)
+           pos->pos_nodes /= pos->pos_arity[i];
+
+       for (i = 0; i < MaxSwitchLevels; i++)
+           pos->pos_arity[i] = pos->pos_arity[i + (pos->pos_levels - MaxSwitchLevels)];
+
+       pos->pos_levels = MaxSwitchLevels;
+       pos->pos_nodeid = pos->pos_nodeid % pos->pos_nodes;
+                               
+       printk ("%s: limiting switch levels to %d\n", rail->Name, MaxSwitchLevels);
+       printk ("%s: nodeid=%d level=%d numnodes=%d\n", rail->Name, pos->pos_nodeid, pos->pos_levels, pos->pos_nodes);
+
+       sprintf (rail->Name, "ep%d[%d]", rail->Number, pos->pos_nodeid);
+    }
+
+    if (rail->Position.pos_mode != ELAN_POS_UNKNOWN && rail->Position.pos_nodeid != pos->pos_nodeid)
+    {
+       printk ("%s: NodeId has changed from %d to %d\n", rail->Name, rail->Position.pos_nodeid, pos->pos_nodeid);
+       panic ("ep: PositionFound: NodeId has changed\n");
+    }
+
+    if (sys->Position.pos_mode != ELAN_POS_UNKNOWN && (sys->Position.pos_nodeid != pos->pos_nodeid || sys->Position.pos_nodes != pos->pos_nodes))
+    {
+       printk ("%s: position incompatible - disabling rail\n", rail->Name);
+       rail->State = EP_RAIL_STATE_INCOMPATIBLE;
+       return;
+    }
+
+    if (sys->Position.pos_mode == ELAN_POS_UNKNOWN)
+    {
+       sys->Position = *pos;
+       sys->NodeSet  = statemap_create (pos->pos_nodes);
+       KMEM_ZALLOC (sys->Nodes, EP_NODE *, pos->pos_nodes * sizeof (EP_NODE), 1);
+    }
+
+    rail->Position             = *pos;
+    rail->SwitchBroadcastLevel = pos->pos_levels - 1;
+    rail->State                = EP_RAIL_STATE_RUNNING;
+
+    for (i = 0; i < pos->pos_levels; i++)
+    {
+       rail->SwitchProbeTick[i]   = lbolt;
+       rail->SwitchLast[i].uplink = 4;
+    }
+
+    rail->Operations.PositionFound (rail, pos);
+
+    INIT_LIST_HEAD (&rail->NetworkErrorList);
+    INIT_LIST_HEAD (&rail->LocalPassivateList);
+    INIT_LIST_HEAD (&rail->RemotePassivateList);
+    INIT_LIST_HEAD (&rail->PassivatedList);
+    INIT_LIST_HEAD (&rail->DisconnectingList);
+
+    rail->NodeSet       = statemap_create (rail->Position.pos_nodes);
+    rail->NodeChangeMap = statemap_create (rail->Position.pos_nodes);
+    rail->NodeChangeTmp = statemap_create (rail->Position.pos_nodes);
+
+    KMEM_ZALLOC (rail->Nodes, EP_NODE_RAIL *, rail->Position.pos_nodes * sizeof (EP_NODE_RAIL), 1);
+
+    for (i = 0; i < rail->Position.pos_nodes; i++)
+    {
+       spin_lock_init (&rail->Nodes[i].CookieLock);
+
+       INIT_LIST_HEAD (&rail->Nodes[i].StalledDmas);
+
+       rail->Nodes[i].State = EP_NODE_DISCONNECTED;
+    }
+
+    /* Notify all subsystems that a new rail has been enabled */
+    kmutex_lock (&sys->SubsysLock);
+    list_for_each (el, &sys->Subsystems) { 
+       EP_SUBSYS *subsys = list_entry (el, EP_SUBSYS, Link);
+
+       if (subsys->AddRail)
+           subsys->AddRail (subsys, sys, rail);
+
+       /* XXXX: what to do if the subsystem refused to add the rail ? */
+    }
+    kmutex_unlock (&sys->SubsysLock);
+
+    /* Now enable the manager input queue */
+    ep_enable_inputq (rail, rail->ManagerInputQ);
+}
+
+static void
+ep_manager (void *arg)
+{
+    EP_SYS            *sys = (EP_SYS *) arg;
+    struct list_head *el;
+    ELAN_POSITION     pos;
+    unsigned long     flags;
+
+    kernel_thread_init ("ep_manager");
+    kernel_thread_become_highpri();
+
+    for (;;)
+    {
+       long nextRunTime = lbolt + MSEC2TICKS(CM_THREAD_SCHEDULE_TIMEOUT);
+
+       list_for_each (el, &sys->ManagedRails) {
+           EP_RAIL *rail = list_entry (el, EP_RAIL, ManagerLink);
+
+           switch (rail->State)
+           {
+           case EP_RAIL_STATE_STARTED:
+               if (ProbeNetwork (rail, &pos) == 0)
+               {
+                   PositionFound (rail, &pos);
+                   break;
+               }
+
+               if (nextRunTime == 0 || AFTER (nextRunTime, lbolt + HZ))
+                   nextRunTime = lbolt + HZ;
+               break;
+
+           case EP_RAIL_STATE_RUNNING:
+               if (ep_poll_inputq (rail, rail->ManagerInputQ, 100, ProcessMessage, rail) >= 100)
+                   nextRunTime = lbolt;
+               
+               /* Handle any nodes which the cluster membership subsystem
+                * has indicated are to begin connecting or disconnecting */
+               spin_lock_irqsave (&sys->NodeLock, flags);
+               if (! statemap_changed (rail->NodeChangeMap))
+                   spin_unlock_irqrestore (&sys->NodeLock, flags);
+               else
+               {
+                   /*
+                    * Take a copy of the statemap, and zero all entries so
+                    * we only see new requests next time
+                    */
+                   statemap_copy (rail->NodeChangeTmp, rail->NodeChangeMap);
+                   statemap_zero (rail->NodeChangeMap);
+                   spin_unlock_irqrestore (&sys->NodeLock, flags);
+                   
+                   UpdateConnectionState (rail, rail->NodeChangeTmp);
+               }
+
+               nextRunTime = ProgressNodeLists (rail, nextRunTime);
+
+               if (statemap_changed (rail->NodeSet))
+               {
+                   ep_call_callbacks (rail, EP_CB_NODESET, rail->NodeSet);
+
+                   statemap_clearchanges (rail->NodeSet);
+               }
+               break;
+
+           case EP_RAIL_STATE_INCOMPATIBLE:
+               break;
+           }
+       }
+
+
+       EPRINTF5 (DBG_MANAGER, "ep_manager: sleep now=%lx nextRunTime=%lx (%ld) [%lx (%ld)]\n",
+                 lbolt, nextRunTime, nextRunTime ? nextRunTime - lbolt : 0, sys->ManagerThread.next_run,
+                 sys->ManagerThread.next_run ? sys->ManagerThread.next_run - lbolt : 0);
+
+       if (ep_kthread_sleep (&sys->ManagerThread, nextRunTime) < 0)
+           break;
+    }
+
+    ep_kthread_stopped (&sys->ManagerThread);
+    kernel_thread_exit();
+}
+
+void
+ep_connect_node (EP_RAIL *rail, int nodeId)
+{
+    EP_SYS       *sys  = rail->System;
+    EP_NODE_RAIL *node = &rail->Nodes[nodeId];
+    unsigned long flags;
+  
+    spin_lock_irqsave (&sys->NodeLock, flags);
+
+    EPRINTF2 (DBG_MANAGER, "%s: ep_connect_node: nodeId %d\n", rail->Name, nodeId);
+
+    ASSERT (node->State == EP_NODE_DISCONNECTED && statemap_getbits (rail->NodeChangeMap, nodeId, 1) == 0);
+    
+    node->State = EP_NODE_CONNECTING;
+
+    statemap_setbits (rail->NodeChangeMap, nodeId, 1, 1);
+
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+
+    ep_kthread_schedule (&sys->ManagerThread, lbolt);
+}
+
+int
+ep_disconnect_node (EP_RAIL *rail, int nodeId)
+{
+    EP_SYS       *sys  = rail->System;
+    EP_NODE_RAIL *node = &rail->Nodes[nodeId];
+    int                  state;
+    unsigned long flags;
+  
+    spin_lock_irqsave (&sys->NodeLock, flags);
+    
+    EPRINTF3 (DBG_MANAGER, "%s: ep_disconnect_node: nodeId %d - %s\n", rail->Name, nodeId, NodeStateNames[node->State]);
+
+    switch (state = node->State)
+    {
+    case EP_NODE_CONNECTING:
+       statemap_setbits (rail->NodeChangeMap, nodeId, 0, 1);
+
+       node->State = EP_NODE_DISCONNECTED;
+       break;
+       
+    case EP_NODE_CONNECTED:
+       statemap_setbits (rail->NodeChangeMap, nodeId, 1, 1);
+
+       node->State = EP_NODE_LEAVING_CONNECTED;
+       break;
+
+    case EP_NODE_LEAVING_CONNECTED:
+       /* no assert on NodeChangeMap as the map could have been taken but not acted on */
+       break;
+       
+    default:
+       ASSERT (statemap_getbits (rail->NodeChangeMap, nodeId, 1) == 0);
+       break;
+    }
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+
+    if (state == EP_NODE_CONNECTED)
+       ep_kthread_schedule (&sys->ManagerThread, lbolt);
+
+    return state;
+}
+
+int
+ep_manager_add_rail (EP_SYS *sys, EP_RAIL *rail)
+{
+    if ((rail->ManagerOutputQ = ep_alloc_outputq (rail, EP_MANAGER_MSG_SIZE, EP_MANAGER_OUTPUTQ_SLOTS)) == NULL)
+       return -ENOMEM;
+
+    if ((rail->ManagerInputQ = ep_alloc_inputq (rail, EP_SYSTEMQ_MANAGER, EP_MANAGER_MSG_SIZE, EP_MANAGER_INPUTQ_SLOTS,
+                                                ManagerQueueEvent, &sys->ManagerThread)) == NULL)
+    {
+       ep_free_outputq (rail, rail->ManagerOutputQ);
+       return -ENOMEM;
+    }
+
+    spin_lock_init (&rail->ManagerOutputQLock);
+
+    ep_xid_cache_init (sys, &rail->XidCache);
+
+    ep_kthread_stall (&sys->ManagerThread);
+    list_add_tail (&rail->ManagerLink, &sys->ManagedRails);
+    ep_kthread_resume (&sys->ManagerThread);
+
+    return (0);
+}
+
+void
+ep_manager_remove_rail (EP_SYS *sys, EP_RAIL *rail)
+{
+    if (rail->ManagerInputQ != NULL)
+    {
+       ep_kthread_stall (&sys->ManagerThread);
+       list_del (&rail->ManagerLink);
+       ep_kthread_resume (&sys->ManagerThread);
+
+       ep_xid_cache_destroy (sys, &rail->XidCache);
+
+       spin_lock_destroy (&rail->ManagerOutputQLock);
+
+       ep_disable_inputq (rail, rail->ManagerInputQ);
+       ep_free_inputq (rail, rail->ManagerInputQ);
+       ep_free_outputq (rail, rail->ManagerOutputQ);
+    }
+}
+
+int
+ep_manager_init (EP_SYS *sys)
+{
+    INIT_LIST_HEAD (&sys->ManagedRails);
+
+    ep_kthread_init (&sys->ManagerThread);
+
+    if (kernel_thread_create (ep_manager, (void *) sys) == 0)
+       return (ENOMEM);
+    
+    ep_kthread_started (&sys->ManagerThread);
+    
+    return (0);
+}
+
+void
+ep_manager_fini (EP_SYS *sys)
+{
+    ep_kthread_stop (&sys->ManagerThread);
+    ep_kthread_destroy (&sys->ManagerThread);
+}
+
+int
+ep_sys_init (EP_SYS *sys)
+{
+    kmutex_init (&sys->SubsysLock);   
+    kmutex_init (&sys->StartStopLock);
+    spin_lock_init (&sys->NodeLock);
+
+    INIT_LIST_HEAD (&sys->Subsystems);
+
+    /* initialise the xid allocators */
+    spin_lock_init (&sys->XidLock);
+    INIT_LIST_HEAD (&sys->XidCacheList);
+
+    /* initially don't know where we are in the network */
+    sys->Position.pos_mode = ELAN_POS_UNKNOWN;
+
+    /* initialise the network mapping descriptor hash tables */
+    ep_nmh_init (&sys->MappingTable);
+
+    /* intialise the shared allocators */
+    ep_shared_alloc_init (sys);
+
+    /* initialise the dvma space */
+    ep_dvma_init (sys);
+
+    /* intiialise the rail manager */
+    ep_manager_init (sys);
+
+    /* initialise all subsystems */
+    cm_init (sys);
+    ep_comms_init (sys);
+    //ep_msgsys_init (sys);
+
+    return (0);
+}
+
+void
+ep_sys_fini (EP_SYS *sys)
+{
+    /* Destroy the subsystems in the reverse order to their creation */
+    while (! list_empty (&sys->Subsystems))
+    {
+       EP_SUBSYS *subsys = list_entry (sys->Subsystems.prev, EP_SUBSYS, Link);
+
+       list_del (&subsys->Link);
+       
+       subsys->Destroy (subsys, sys);
+    }
+
+    ep_manager_fini(sys);
+    ep_dvma_fini (sys);
+    ep_shared_alloc_fini (sys);
+
+    ep_nmh_fini (&sys->MappingTable);
+
+    if (sys->Position.pos_mode != ELAN_POS_UNKNOWN) {
+       statemap_destroy (sys->NodeSet);
+       KMEM_FREE(sys->Nodes, sys->Position.pos_nodes * sizeof (EP_NODE));
+    }
+
+    spin_lock_destroy (&sys->XidLock);
+
+    spin_lock_destroy (&sys->NodeLock);
+    kmutex_destroy (&sys->SubsysLock);
+    kmutex_destroy (&sys->StartStopLock);
+}
+
+void
+ep_shutdown (EP_SYS *sys)
+{
+    sys->Shutdown = 1;
+}
+
+int
+ep_init_rail (EP_SYS *sys, EP_RAIL *rail)
+{
+    static int rnum;
+
+    rail->System              = sys;
+    rail->State               = EP_RAIL_STATE_UNINITIALISED;
+    rail->Number              = rnum++;
+    rail->Position.pos_mode   = ELAN_POS_UNKNOWN;
+    rail->Position.pos_nodeid = ELAN_INVALID_NODE;
+
+    rail->CallbackRegistered  = 0;
+
+    sprintf (rail->Name, "ep%d", rail->Number);
+
+    /* Initialise externally visible locks */
+    kmutex_init (&rail->CallbackLock);
+
+    ep_alloc_init (rail);
+
+    sys->Rails[rail->Number] = rail;
+
+    return 0;
+}
+
+void
+ep_destroy_rail (EP_RAIL *rail)
+{
+    ASSERT (rail->State == EP_RAIL_STATE_UNINITIALISED);
+
+    ep_alloc_fini (rail);
+
+    kmutex_destroy (&rail->CallbackLock);
+
+    rail->System->Rails[rail->Number] = NULL;
+
+    rail->Operations.DestroyRail (rail);
+}
+
+/* We need to traverse the Subsystems lists backwards
+ * but it's not defined in <linux/list.h> */
+#define list_for_each_backwards(pos,list) \
+       for (pos = (list)->prev; pos != (list); \
+            pos = (pos)->prev)
+
+void
+__ep_stop_rail (EP_RAIL *rail)
+{
+    /* called holding the sys->Lock */
+    EP_SYS           *sys = rail->System;
+    struct list_head *el;
+
+    rail->Operations.StallRail (rail);
+
+    /* Notify all subsystems that this rail is being stopped */
+    if (rail->State == EP_RAIL_STATE_RUNNING)
+    {
+       kmutex_lock (&sys->SubsysLock);
+       list_for_each_backwards (el, &sys->Subsystems) { 
+           EP_SUBSYS *subsys = list_entry (el, EP_SUBSYS, Link);
+           
+           if (subsys->RemoveRail)
+               subsys->RemoveRail (subsys, sys, rail);
+       }
+       kmutex_unlock (&sys->SubsysLock);
+
+       ep_manager_remove_rail (sys, rail);
+
+       KMEM_FREE (rail->Nodes, rail->Position.pos_nodes * sizeof (EP_NODE_RAIL));
+
+       statemap_destroy (rail->NodeChangeTmp);
+       statemap_destroy (rail->NodeChangeMap);
+       statemap_destroy (rail->NodeSet);
+    }
+
+    ep_dvma_remove_rail (sys, rail);
+    ep_shared_alloc_remove_rail (sys, rail);
+
+    rail->Operations.StopRail (rail);
+
+    rail->State = EP_RAIL_STATE_UNINITIALISED;
+}
+
+void
+ep_stop_rail (EP_RAIL *rail)
+{
+    EP_SYS *sys = rail->System;
+
+    /* stall ep_manager                      */
+    /* and remove the rail from the manaager */
+
+    ep_kthread_stall (&sys->ManagerThread);
+    if ( rail->State == EP_RAIL_STATE_STARTED ) 
+        ep_manager_remove_rail (sys, rail);
+    ep_kthread_resume (&sys->ManagerThread);
+
+    __ep_stop_rail (rail);
+}
+
+int
+ep_start_rail (EP_RAIL *rail)
+{
+    EP_SYS *sys = rail->System;
+
+    ASSERT (rail->State == EP_RAIL_STATE_UNINITIALISED);
+
+    if (rail->Operations.StartRail (rail) < 0)
+       return -ENXIO;
+    
+    kmutex_lock (&sys->StartStopLock);
+    /* Add this rail to the shared allocator */
+    if (ep_shared_alloc_add_rail (rail->System, rail))
+       goto failed;
+
+    /* Add this rail to dvma kmap */
+    if (ep_dvma_add_rail (rail->System, rail))
+       goto failed;
+
+    /* rail is now started */
+    rail->State = EP_RAIL_STATE_STARTED;
+
+    /* notify the rail manager of the new rail */
+    if (ep_manager_add_rail (rail->System, rail))
+       goto failed;
+
+    kmutex_unlock (&sys->StartStopLock);
+    return (ESUCCESS);
+
+ failed:
+    printk ("%s: start failed\n", rail->Name);
+    kmutex_unlock (&sys->StartStopLock);
+    __ep_stop_rail (rail);
+
+    return (ENOMEM);   
+}
+
+void
+ep_subsys_add (EP_SYS *sys, EP_SUBSYS *subsys)
+{
+    kmutex_lock (&sys->SubsysLock);
+    list_add_tail (&subsys->Link, &sys->Subsystems);
+    kmutex_unlock (&sys->SubsysLock);
+}
+
+void
+ep_subsys_del (EP_SYS *sys, EP_SUBSYS *subsys)
+{
+    kmutex_lock (&sys->SubsysLock);
+    list_del (&subsys->Link);
+    kmutex_unlock (&sys->SubsysLock);
+}
+
+EP_SUBSYS *
+ep_subsys_find (EP_SYS *sys, char *name)
+{
+    struct list_head *el;
+
+    ASSERT ( !in_interrupt());
+
+    kmutex_lock (&sys->SubsysLock); 
+    list_for_each (el, &sys->Subsystems) {
+       EP_SUBSYS *subsys = list_entry (el, EP_SUBSYS, Link);
+
+       if (! strcmp (subsys->Name, name))
+       {
+           kmutex_unlock (&sys->SubsysLock);
+           return (subsys);
+       }
+    }
+
+    kmutex_unlock (&sys->SubsysLock);
+    return (NULL);
+}
+
+int
+ep_waitfor_nodeid (EP_SYS *sys)
+{
+    int i, printed = 0;
+    kcondvar_t Sleep;
+    spinlock_t Lock;
+
+    kcondvar_init (&Sleep);
+    spin_lock_init (&Lock);
+
+#define TICKS_TO_WAIT  (10*hz)
+#define TICKS_PER_LOOP (hz/10)
+    for (i = 0; sys->Position.pos_mode == ELAN_POS_UNKNOWN && i < TICKS_TO_WAIT; i += TICKS_PER_LOOP)
+    {
+       if (! printed++)
+           printk ("ep: waiting for network position to be found\n");
+
+       spin_lock (&Lock);
+       kcondvar_timedwait (&Sleep, &Lock, NULL, lbolt + TICKS_PER_LOOP);
+       spin_unlock (&Lock);
+    }
+
+    if (sys->Position.pos_mode == ELAN_POS_UNKNOWN)
+       printk ("ep: network position not found after waiting\n");
+    else if (printed)
+       printk ("ep: network position found at nodeid %d\n", sys->Position.pos_nodeid);
+
+    spin_lock_destroy (&Lock);
+    kcondvar_destroy (&Sleep);
+
+    return (sys->Position.pos_mode == ELAN_POS_UNKNOWN ? ELAN_INVALID_NODE : sys->Position.pos_nodeid);
+}
+
+int
+ep_nodeid (EP_SYS *sys)
+{
+    return (sys->Position.pos_mode == ELAN_POS_UNKNOWN ? ELAN_INVALID_NODE : sys->Position.pos_nodeid);
+}
+
+int
+ep_numnodes (EP_SYS *sys)
+{
+    return (sys->Position.pos_nodes);
+}
+
+void
+ep_fillout_stats(EP_RAIL *r, char *str) 
+{
+    sprintf(str+strlen(str),"SendMessageFailed %lu NeterrAtomicPacket %lu NeterrDmaPacket %lu \n", r->Stats.SendMessageFailed, r->Stats.NeterrAtomicPacket, r->Stats.NeterrDmaPacket);
+    sprintf(str+strlen(str),"Rx %lu  %lu /sec\n",   GET_STAT_TOTAL(r->Stats,rx), GET_STAT_PER_SEC(r->Stats,rx) ); 
+    sprintf(str+strlen(str),"MBytes %lu  %lu MB/sec\n", GET_STAT_TOTAL(r->Stats,rx_len)/ (1024*1024), GET_STAT_PER_SEC(r->Stats,rx_len) / (1024*1024)); 
+    sprintf(str+strlen(str),"Tx %lu  %lu /sec\n",   GET_STAT_TOTAL(r->Stats,tx), GET_STAT_PER_SEC(r->Stats,tx) ); 
+    sprintf(str+strlen(str),"MBytes %lu  %lu MB/sec\n", GET_STAT_TOTAL(r->Stats,tx_len)/ (1024*1024), GET_STAT_PER_SEC(r->Stats,tx_len) / (1024*1024)); 
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kcomm_elan3.c        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan3.c     2005-06-01 23:12:54.665429072 -0400
@@ -0,0 +1,504 @@
+
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kcomm_elan3.c,v 1.31.8.3 2004/11/30 12:02:17 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "conf_linux.h"
+
+extern EP_CODE threadcode_elan3;
+
+unsigned int
+ep3_create_rails (EP_SYS *sys, unsigned int disabled)
+{
+    unsigned int rmask = 0;
+    ELAN3_DEV   *dev;
+    EP_RAIL     *rail;
+    int          i;
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+    {
+       if ((dev = elan3_device (i)) != NULL)
+       {
+           if ((rail = ep3_create_rail (sys, dev)) != NULL)
+           {
+               if (disabled & (1 << rail->Number))
+                   printk ("%s: auto-start of device disabled by configuration\n", rail->Name);
+               else
+                   ep_start_rail (rail);
+               
+               ep_procfs_rail_init(rail);
+
+               rmask |= (1 << rail->Number);
+           }
+       }
+    }
+
+    return rmask;
+}
+
+EP_RAIL *
+ep3_create_rail (EP_SYS *sys, ELAN3_DEV *dev)
+{
+    EP3_RAIL *rail;
+    int       res;
+
+    KMEM_ZALLOC (rail, EP3_RAIL *, sizeof (EP3_RAIL), TRUE);
+
+    if (rail == NULL)
+       return (EP_RAIL *) NULL;
+
+    if ((res = ep_init_rail (sys, &rail->Generic)) != 0)
+    {
+       KMEM_FREE (rail, sizeof (EP3_RAIL));
+       return (EP_RAIL *) NULL;
+    }
+
+    rail->Device = dev;
+
+    /* Install our rail operations */
+    rail->Generic.Operations.DestroyRail      = ep3_destroy_rail;
+    rail->Generic.Operations.StartRail        = ep3_start_rail;
+    rail->Generic.Operations.StallRail        = ep3_stall_rail;
+    rail->Generic.Operations.StopRail         = ep3_stop_rail;
+
+    rail->Generic.Operations.SdramAlloc       = ep3_sdram_alloc;
+    rail->Generic.Operations.SdramFree        = ep3_sdram_free;
+    rail->Generic.Operations.SdramWriteb      = ep3_sdram_writeb;
+
+    rail->Generic.Operations.KaddrMap         = ep3_kaddr_map;
+    rail->Generic.Operations.SdramMap         = ep3_sdram_map;
+    rail->Generic.Operations.Unmap            = ep3_unmap;
+
+    rail->Generic.Operations.DvmaReserve      = ep3_dvma_reserve;
+    rail->Generic.Operations.DvmaRelease      = ep3_dvma_release;
+    rail->Generic.Operations.DvmaSetPte       = ep3_dvma_set_pte;
+    rail->Generic.Operations.DvmaReadPte      = ep3_dvma_read_pte;
+    rail->Generic.Operations.DvmaUnload       = ep3_dvma_unload;
+    rail->Generic.Operations.FlushTlb         = ep3_flush_tlb;
+
+    rail->Generic.Operations.ProbeRoute       = ep3_probe_route;
+    rail->Generic.Operations.PositionFound    = ep3_position_found;
+    rail->Generic.Operations.CheckPosition    = ep3_check_position;
+    rail->Generic.Operations.NeterrFixup      = ep3_neterr_fixup;
+
+    rail->Generic.Operations.LoadSystemRoute  = ep3_load_system_route;
+
+    rail->Generic.Operations.LoadNodeRoute    = ep3_load_node_route;
+    rail->Generic.Operations.UnloadNodeRoute  = ep3_unload_node_route;
+    rail->Generic.Operations.LowerFilter      = ep3_lower_filter;
+    rail->Generic.Operations.RaiseFilter      = ep3_raise_filter;
+    rail->Generic.Operations.NodeDisconnected = ep3_node_disconnected;
+
+    rail->Generic.Operations.FlushFilters     = ep3_flush_filters;
+    rail->Generic.Operations.FlushQueues      = ep3_flush_queues;
+
+    rail->Generic.Operations.AllocInputQ      = ep3_alloc_inputq;
+    rail->Generic.Operations.FreeInputQ       = ep3_free_inputq;
+    rail->Generic.Operations.EnableInputQ     = ep3_enable_inputq;
+    rail->Generic.Operations.DisableInputQ    = ep3_disable_inputq;
+    rail->Generic.Operations.PollInputQ       = ep3_poll_inputq;
+
+    rail->Generic.Operations.AllocOutputQ     = ep3_alloc_outputq;
+    rail->Generic.Operations.FreeOutputQ      = ep3_free_outputq;
+    rail->Generic.Operations.OutputQMsg       = ep3_outputq_msg;
+    rail->Generic.Operations.OutputQState     = ep3_outputq_state;
+    rail->Generic.Operations.OutputQSend      = ep3_outputq_send;
+
+    rail->Generic.Operations.FillOutStats     = ep3_fillout_stats;
+
+    rail->Generic.Devinfo = dev->Devinfo;
+
+    printk ("%s: connected via elan3 rev%c device %d\n", rail->Generic.Name,
+           'a' + dev->Devinfo.dev_revision_id, dev->Instance);
+
+    return (EP_RAIL *) rail;
+}
+
+void
+ep3_destroy_rail (EP_RAIL *r)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+    
+    KMEM_FREE (rail, sizeof (EP3_RAIL));
+}
+
+static int
+ep3_attach_rail (EP3_RAIL *rail)
+{
+    ELAN3_DEV        *dev = rail->Device;
+    ELAN3_CTXT       *ctxt;
+    ELAN_CAPABILITY  *cap;
+    int               ctx;
+    unsigned long     flags;
+
+    if ((ctxt = elan3_alloc (dev, TRUE)) == (ELAN3_CTXT *) NULL)
+    {
+       printk ("%s: cannot allocate elan context\n", rail->Generic.Name);
+       return -ENXIO;
+    }
+    
+    ctxt->Operations = &ep3_elan3_ops;
+    ctxt->Private    = (void *) rail;
+    
+    /* Initialise a capability and attach to the elan*/
+    KMEM_ALLOC (cap, ELAN_CAPABILITY *, sizeof (ELAN_CAPABILITY), TRUE);
+    
+    elan_nullcap (cap);
+    
+    cap->cap_type        = ELAN_CAP_TYPE_KERNEL;
+    cap->cap_version     = ELAN_CAP_VERSION_NUMBER;
+    cap->cap_mycontext   = ELAN3_MRF_CONTEXT_NUM | SYS_CONTEXT_BIT;
+    cap->cap_lowcontext  = ELAN3_MRF_CONTEXT_NUM | SYS_CONTEXT_BIT;
+    cap->cap_highcontext = ELAN3_MRF_CONTEXT_NUM | SYS_CONTEXT_BIT;
+    cap->cap_railmask    = 1 << dev->Devinfo.dev_rail;
+    
+    /* Ensure the context filter is raised while we initialise */
+    elan3_block_inputter (ctxt, TRUE);
+
+    if (elan3_doattach (ctxt, cap) != 0)
+    {
+       printk ("%s: cannot attach to kernel context\n", rail->Generic.Name);
+
+       KMEM_FREE (cap, sizeof (ELAN_CAPABILITY));
+       elan3_free (ctxt);
+       return -ENXIO;
+    }
+    KMEM_FREE (cap, sizeof (ELAN_CAPABILITY));
+
+    /* now attach to all the kernel comms input/dmaring/data contexts */
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    for (ctx = ELAN3_DMARING_BASE_CONTEXT_NUM; ctx <= ELAN3_DMARING_TOP_CONTEXT_NUM; ctx++)
+    {
+       /* place it in the info table.  NOTE: don't call elan3mmu_set_info, as this */
+       /* will queue the info again on the devices info list */
+       dev->CtxtTable[ctx] = ctxt;
+       
+       elan3mmu_set_context_filter (dev, ctx|SYS_CONTEXT_BIT, TRUE, 0, NULL);
+       elan3mmu_attach (dev, ctx, ctxt->Elan3mmu, ctxt->RouteTable->Table, ctxt->RouteTable->Size-1);
+    }
+
+    for (ctx = ELAN3_KCOMM_BASE_CONTEXT_NUM; ctx <= ELAN3_KCOMM_TOP_CONTEXT_NUM; ctx++)
+    {
+       /* place it in the info table.  NOTE: don't call elan3mmu_set_info, as this */
+       /* will queue the info again on the devices info list */
+       dev->CtxtTable[ctx] = ctxt;
+       
+       elan3mmu_set_context_filter (dev, ctx|SYS_CONTEXT_BIT, TRUE, 0, NULL);
+       elan3mmu_attach (dev, ctx, ctxt->Elan3mmu, ctxt->RouteTable->Table, ctxt->RouteTable->Size-1);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    /* Stash the ctxt,commandport, mmu and route table */
+    rail->Ctxt        = ctxt;
+    rail->CommandPort = ctxt->CommandPort;
+    rail->Elan3mmu    = ctxt->Elan3mmu;
+    rail->RouteTable  = ctxt->RouteTable;
+
+    return 0;
+}
+
+static void
+ep3_detach_rail (EP3_RAIL *rail)
+{
+    ELAN3_DEV *dev = rail->Device;
+    unsigned long flags;
+    int ctx;
+
+    /* detach from the elan */
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    for (ctx = ELAN3_KCOMM_BASE_CONTEXT_NUM; ctx <= ELAN3_KCOMM_TOP_CONTEXT_NUM; ctx++)
+    {
+       dev->CtxtTable[ctx] = NULL;
+       elan3mmu_detach (dev, ctx);
+    }
+
+    for (ctx = ELAN3_DMARING_BASE_CONTEXT_NUM; ctx <= ELAN3_DMARING_TOP_CONTEXT_NUM; ctx++)
+    {
+       dev->CtxtTable[ctx] = NULL;
+       elan3mmu_detach (dev, ctx);
+    }
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    elan3_dodetach(rail->Ctxt);
+    elan3_free (rail->Ctxt);
+
+    rail->Ctxt        = NULL;
+    rail->CommandPort = 0;
+    rail->Elan3mmu    = NULL;
+    rail->RouteTable  = NULL;
+}
+
+int
+ep3_start_rail (EP_RAIL *r)
+{
+    EP3_RAIL     *rail = (EP3_RAIL *) r;
+    int           i, res;
+    unsigned long flags;
+
+    if ((res = ep3_attach_rail (rail)) != 0)
+       return res;
+
+    spin_lock_init (&rail->CookieLock);
+    kmutex_init (&rail->HaltOpMutex);
+    kcondvar_init (&rail->HaltOpSleep);
+
+    /* Initialise event interrupt cookie table */
+    InitialiseCookieTable (&rail->CookieTable);
+
+    /* Load and map the thread code */
+    rail->ThreadCode = threadcode_elan3;
+    if (ep_loadcode (&rail->Generic, &rail->ThreadCode) != ESUCCESS)
+       goto failed;
+
+    /* Map the command port to be visible to the Elan */
+    ep3_ioaddr_map (&rail->Generic, EP3_COMMANDPORT_ADDR, rail->Ctxt->CommandPage, PAGESIZE, EP_PERM_WRITE);
+    rail->CommandPortAddr = EP3_COMMANDPORT_ADDR + (rail->Ctxt->CommandPort - rail->Ctxt->CommandPage);
+
+    /* Allocate the elan visible sdram/main memory */
+    if ((rail->RailElan = ep_alloc_elan (&rail->Generic, sizeof (EP3_RAIL_ELAN), 0, &rail->RailElanAddr)) == 0 ||
+       (rail->RailMain = ep_alloc_main (&rail->Generic, sizeof (EP3_RAIL_MAIN), 0, &rail->RailMainAddr)) == 0)
+    {
+       goto failed;
+    }
+
+    /* Allocate the system input queues at their fixed elan address */
+    if (! (rail->QueueDescs = ep_alloc_memory_elan (&rail->Generic, EP_SYSTEM_QUEUE_BASE, PAGESIZE, EP_PERM_ALL, 0)))
+       goto failed;
+
+    /* Initialise all queue entries to be full */
+    for (i = 0; i < EP_NUM_SYSTEMQ; i++)
+       elan3_sdram_writel (rail->Device, EP_SYSTEMQ_DESC(rail->QueueDescs, i) + offsetof (EP3_InputQueue, q_state), E3_QUEUE_FULL);
+
+    /* initialise the dma rings */
+    if (DmaRingsCreate (rail))
+       goto failed;
+    
+    if (InitialiseDmaRetries (rail))
+       goto failed;
+
+    if (ep3_init_probenetwork (rail))
+       goto failed;
+
+    /* can now drop the context filter for the system context */
+    spin_lock_irqsave (&rail->Device->IntrLock, flags);
+    elan3mmu_set_context_filter (rail->Device, ELAN3_MRF_CONTEXT_NUM|SYS_CONTEXT_BIT, FALSE, 0, NULL);
+    spin_unlock_irqrestore (&rail->Device->IntrLock, flags);
+
+    return 0;
+
+ failed:
+    printk ("ep3_start_rail: failed for rail %d\n", rail->Generic.Number);
+    ep3_stop_rail (&rail->Generic);
+
+    return -ENOMEM;
+}
+
+void
+ep3_stall_rail (EP_RAIL *r)
+{
+    EP3_RAIL     *rail = (EP3_RAIL *) r;
+    int           ctx;
+    unsigned long flags;
+
+    /* raise all the context filters */
+    spin_lock_irqsave (&rail->Device->IntrLock, flags);
+
+    for (ctx = ELAN3_KCOMM_BASE_CONTEXT_NUM; ctx <= ELAN3_KCOMM_TOP_CONTEXT_NUM; ctx++)
+       elan3mmu_set_context_filter (rail->Device, ctx|SYS_CONTEXT_BIT, TRUE, 0, NULL);
+
+    for (ctx = ELAN3_DMARING_BASE_CONTEXT_NUM; ctx <= ELAN3_DMARING_TOP_CONTEXT_NUM; ctx++)
+       elan3mmu_set_context_filter (rail->Device, ctx|SYS_CONTEXT_BIT, TRUE, 0, NULL);
+
+    elan3mmu_set_context_filter (rail->Device, ELAN3_MRF_CONTEXT_NUM|SYS_CONTEXT_BIT, TRUE, 0, NULL);
+
+    spin_unlock_irqrestore (&rail->Device->IntrLock, flags);
+}
+
+void
+ep3_stop_rail (EP_RAIL *r)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+
+    ep3_destroy_probenetwork (rail);
+
+    if (rail->DmaRetryInitialised)
+       DestroyDmaRetries (rail);
+
+    DmaRingsRelease(rail);
+
+    if (rail->Generic.State == EP_RAIL_STATE_RUNNING)
+    {
+       KMEM_FREE (rail->MainCookies, rail->Generic.Position.pos_nodes * sizeof (E3_uint32));
+
+       ep_free_elan (&rail->Generic, rail->ElanCookies, rail->Generic.Position.pos_nodes * sizeof (E3_uint32));
+    }
+
+    if (rail->QueueDescs)
+       ep_free_memory_elan (&rail->Generic, EP_SYSTEM_QUEUE_BASE);
+    rail->QueueDescs = 0;
+
+    if (rail->RailMain)
+       ep_free_main (&rail->Generic, rail->RailMainAddr, sizeof (EP3_RAIL_MAIN));
+    rail->RailMain = 0;
+
+    if (rail->RailElan)
+       ep_free_elan (&rail->Generic, rail->RailElanAddr, sizeof (EP3_RAIL_ELAN));
+    rail->RailElan = 0;
+
+    ep_unloadcode (&rail->Generic, &rail->ThreadCode);
+
+    DestroyCookieTable (&rail->CookieTable);
+
+    ep_perrail_unmap (&rail->Generic, rail->Ctxt->CommandPage, PAGESIZE);
+
+    kcondvar_destroy (&rail->HaltOpSleep);
+    kmutex_destroy (&rail->HaltOpMutex);
+    spin_lock_destroy (&rail->CookieLock);
+
+    ep3_detach_rail (rail);
+}
+
+void
+ep3_position_found (EP_RAIL *r, ELAN_POSITION *pos)
+{
+    EP3_RAIL   *rail = (EP3_RAIL *) r;
+    sdramaddr_t addr;
+
+    rail->SwitchBroadcastLevelTick = lbolt;
+
+    elan3_sdram_writel (rail->Device, rail->RailElan + offsetof (EP3_RAIL_ELAN, NodeId), pos->pos_nodeid);
+
+    /* Allocate Network Identify cookie state */
+    KMEM_ZALLOC (rail->MainCookies, E3_uint32 *, pos->pos_nodes * sizeof (E3_uint32), 1);
+    
+    if (! (addr = ep_alloc_elan (&rail->Generic, pos->pos_nodes * sizeof (E3_uint32), 0, &rail->ElanCookies)))
+       panic ("ep: PositionFound: cannot allocate elan cookies array\n");
+
+    elan3_sdram_zeroq_sdram (rail->Device, addr, pos->pos_nodes * sizeof (E3_uint32));
+
+    ep3_probe_position_found (rail, pos);
+}
+
+sdramaddr_t
+ep3_sdram_alloc (EP_RAIL *r, EP_ADDR addr, unsigned size)
+{
+    return elan3_sdram_alloc (((EP3_RAIL *) r)->Device, size);
+}
+
+void
+ep3_sdram_free (EP_RAIL *r, sdramaddr_t addr, unsigned size)
+{
+    elan3_sdram_free (((EP3_RAIL *) r)->Device, addr, size);
+}
+
+void
+ep3_sdram_writeb (EP_RAIL *r, sdramaddr_t addr, unsigned char val)
+{
+    elan3_sdram_writeb (((EP3_RAIL *) r)->Device, addr, val);
+}
+
+void
+ep3_flush_tlb (EP_RAIL *r)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+    ELAN3_DEV *dev = rail->Device;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->TlbLock, flags);
+    
+    IncrStat (dev, TlbFlushes);
+    
+    write_reg32 (dev, Cache_Control_Reg.ContReg, dev->Cache_Control_Reg | MMU_FLUSH);
+    mmiob ();
+    spin_unlock_irqrestore (&dev->TlbLock, flags);
+
+    while (! (read_reg32 (dev, Cache_Control_Reg.ContReg) & MMU_FLUSHED))
+       mb();
+}
+
+void
+ep3_load_system_route (EP_RAIL *r, unsigned vp, unsigned lowNode, unsigned highNode)
+{
+    EP3_RAIL  *rail = (EP3_RAIL *) r;
+    E3_uint16  flits[MAX_FLITS];
+    int        nflits;
+    
+    nflits = GenerateRoute (&rail->Generic.Position, flits, lowNode, highNode, DEFAULT_ROUTE_TIMEOUT, HIGH_ROUTE_PRIORITY);
+       
+    if (LoadRoute (rail->Device, rail->RouteTable, vp, ELAN3_MRF_CONTEXT_NUM|SYS_CONTEXT_BIT, nflits, flits) != 0)
+    {
+       /* XXXX: whilst LoadRoute() can fail - it is not likely. */
+       panic ("ep3_load_system_route: cannot load p2p route entry\n");
+    }
+}
+
+void
+ep3_load_node_route (EP_RAIL *r, unsigned nodeId)
+{
+    EP3_RAIL     *rail = (EP3_RAIL *) r;
+    E3_uint16     flits[MAX_FLITS];
+    int           nflits;
+
+    nflits = GenerateRoute (&rail->Generic.Position, flits, nodeId, nodeId, DEFAULT_ROUTE_TIMEOUT, DEFAULT_ROUTE_PRIORITY);
+
+    if (LoadRoute (rail->Device, rail->RouteTable, EP_VP_DATA(nodeId), EP3_CONTEXT_NUM(rail->Generic.Position.pos_nodeid), nflits, flits) != 0)
+       panic ("ep3_load_node_route: cannot load p2p data route entry\n");
+}
+
+void
+ep3_unload_node_route (EP_RAIL *r, unsigned nodeId)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+
+    ClearRoute (rail->Device, rail->RouteTable, EP_VP_DATA(nodeId));
+}
+
+void
+ep3_lower_filter (EP_RAIL *r, unsigned nodeId)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->Device->IntrLock, flags);
+    elan3mmu_set_context_filter (rail->Device, EP3_CONTEXT_NUM(nodeId), 0, 0, NULL);
+    spin_unlock_irqrestore (&rail->Device->IntrLock, flags);
+}
+
+void
+ep3_raise_filter (EP_RAIL *r, unsigned nodeId)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->Device->IntrLock, flags);
+    elan3mmu_set_context_filter (rail->Device, EP3_CONTEXT_NUM(nodeId), 1, 0, NULL);
+    spin_unlock_irqrestore (&rail->Device->IntrLock, flags);
+}
+
+void
+ep3_node_disconnected (EP_RAIL *r, unsigned nodeId)
+{
+    FreeStalledDmas ((EP3_RAIL *) r, nodeId);
+}
+
+void
+ep3_fillout_stats(EP_RAIL *r, char *str) 
+{
+    /* no stats here yet */
+    /* EP3_RAIL *ep3rail = (EP3_RAIL *)r; */
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan3.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kcomm_elan3.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan3.h     2005-06-01 23:12:54.666428920 -0400
@@ -0,0 +1,431 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __EP_KCOMM_ELAN3_H
+#define __EP_KCOMM_ELAN3_H
+
+#ident "@(#)$Id: kcomm_elan3.h,v 1.50.8.3 2004/12/14 10:19:14 mike Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm_elan3.h,v $*/
+
+#if !defined(__ELAN3__)
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+#endif /* !defined(__ELAN3__) */
+
+#include <elan3/trtype.h>
+
+/* private address allocation */
+#define EP3_TEXT_BASE                  0xFF000000              /* base address for thread code (defined in makerules.elan3) */
+#define EP3_COMMANDPORT_ADDR           0xFFF00000              /* mapping address for elan command port */
+
+#define EP3_STACK_SIZE                 1024                    /* default thread code stack size */
+
+#define EP3_PACEMAKER_EVENTADDR                0xfeedbeef              /* mis-aligned address used by heartbeat pacemaker */
+
+/* context number allocation */
+#define EP3_CONTEXT_NUM(nodeId)                ((ELAN3_KCOMM_BASE_CONTEXT_NUM + (nodeId)) | SYS_CONTEXT_BIT)
+#define EP3_CONTEXT_ISDATA(ctx)                (((ctx) & MAX_ROOT_CONTEXT_MASK) >= ELAN3_KCOMM_BASE_CONTEXT_NUM && \
+                                        ((ctx) & MAX_ROOT_CONTEXT_MASK) <= ELAN3_KCOMM_TOP_CONTEXT_NUM)
+#define EP3_CONTEXT_TO_NODE(ctx)       (((ctx) & MAX_ROOT_CONTEXT_MASK) - ELAN3_KCOMM_BASE_CONTEXT_NUM)
+
+/* DMA issueing rings */
+#define EP3_RING_CRITICAL              0
+#define EP3_RING_CRITICAL_LEN          128
+#define EP3_RING_HIGH_PRI              1
+#define EP3_RING_HIGH_PRI_LEN          64
+#define EP3_RING_LOW_PRI               2
+#define EP3_RING_LOW_PRI_LEN           32
+#define EP3_NUM_RINGS                  3
+
+/* Value to "return" from c_close() when envelope handled  by the trap handler */
+#define EP3_PAckStolen                 4
+
+/* unimplemented instruction trap types for thread code */
+#define EP3_UNIMP_TRAP_NO_DESCS                0
+#define EP3_UNIMP_TRAP_PACKET_NACKED   1
+#define EP3_UNIMP_THREAD_HALTED                2
+#define EP3_NUM_UNIMP_TRAPS            3
+
+/* forward declarations */
+typedef struct ep3_rail        EP3_RAIL;
+
+/* block copy elan3 inputter queue - with waitvent0 */
+typedef struct ep3_inputqueue
+{
+    volatile E3_uint32 q_state;        /* queue is full=bit0, queue is locked=bit8 */
+    volatile E3_Addr   q_bptr;         /* block aligned ptr to current back item */
+    E3_uint32          q_size;         /* size of queue item; 0x1 <= size <= (0x40 * 5) */
+    E3_Addr            q_top;          /* block aligned ptr to last queue item */
+    E3_Addr            q_base;         /* block aligned ptr to first queue item */
+    volatile E3_Addr   q_fptr;         /* block aligned ptr to current front item */
+    E3_BlockCopyEvent  q_event;        /* queue block copy event */
+    E3_uint32          q_pad[4];       /* pad to 64 bytes */
+    E3_Addr            q_wevent;       /* WaitEvent0 struct */
+    E3_int32           q_wcount;
+} EP3_InputQueue;
+
+
+#if !defined(__ELAN3__)
+
+/* dma retries types and retry times */
+typedef struct ep3_retry_dma
+{
+    struct list_head    Link;                                  /* chained on free/retry list */
+    long               RetryTime;                              /* "lbolt" to retry at */
+    E3_DMA_BE          Dma;                                    /* DMA (in main memory) */
+} EP3_RETRY_DMA;
+
+typedef struct ep3_dma_ring
+{
+    sdramaddr_t                pEvent;  
+    E3_Addr            epEvent;
+    
+    sdramaddr_t                pDma;     
+    E3_Addr            epDma; 
+    
+    E3_uint32         *pDoneBlk; 
+    E3_Addr            epDoneBlk; 
+    
+    int                        Entries;                                /* number of slots in array  */
+    int                        Position;                               /* current position in array */
+
+    ioaddr_t            CommandPort;
+    ioaddr_t           CommandPage;
+    DeviceMappingHandle CommandPageHandle;
+} EP3_DMA_RING;
+
+#define DMA_RING_EVENT(ring,n)         ((ring)->pEvent + (n)*sizeof (E3_BlockCopyEvent))
+#define DMA_RING_EVENT_ELAN(ring,n)    ((ring)->epEvent + (n)*sizeof (E3_BlockCopyEvent))
+
+#define DMA_RING_DMA(ring,n)           ((ring)->pDma   + (n)*sizeof (E3_DMA))
+#define DMA_RING_DMA_ELAN(ring,n)      ((ring)->epDma   + (n)*sizeof (E3_DMA))
+
+#define DMA_RING_DONE_ELAN(ring,n)     ((ring)->epDoneBlk + (n)*sizeof (E3_uint32))
+
+/* Event interrupt cookie operations and lookup table */
+typedef struct ep3_cookie_ops
+{
+    void       (*Event)       (EP3_RAIL *rail, void *arg);                             /* called from the interrupt handler when an event is "set" */
+    void       (*DmaRetry)    (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int error);  /* called from the interrupt handler when a DMA is "nacked" */
+    void       (*DmaCancelled)(EP3_RAIL *rail, void *arg, E3_DMA_BE *dma);             /* called from the interrupt handler/flush disconnecting when cancelled. */
+    void       (*DmaVerify)   (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma);             /* called from multiple places, to check dma is consistent with state. */
+} EP3_COOKIE_OPS;
+
+typedef struct ep3_cookie
+{
+    struct ep3_cookie  *Next;                                  /* Cookies are chained in hash table. */
+    E3_uint32          Cookie;                                 /* Cooke store in ev_Type */
+    EP3_COOKIE_OPS     *Operations;                            /* Cookie operations */
+    void              *Arg;                                    /* Users arguement. */
+} EP3_COOKIE;
+
+#define EP3_COOKIE_HASH_SIZE           (256)
+#define EP3_HASH_COOKIE(a)             ((((a) >> 3) ^ ((a) >> 7) ^ ((a) >> 11)) & (EP3_COOKIE_HASH_SIZE-1))
+
+typedef struct ep3_cookie_table
+{
+    spinlock_t         Lock;
+    EP3_COOKIE        *Entries[EP3_COOKIE_HASH_SIZE];
+} EP3_COOKIE_TABLE;
+
+#endif /* !defined(__ELAN3__) */
+
+#define EP3_EVENT_FREE                 ((1 << 4) | EV_WCOPY)
+#define EP3_EVENT_ACTIVE               ((2 << 4) | EV_WCOPY)
+/* DONE == Cookie */
+#define EP3_EVENT_FAILED               ((3 << 4) | EV_WCOPY)
+#define EP3_EVENT_PRIVATE              ((4 << 4) | EV_WCOPY)
+
+/* The event cookie can get posted (and seen) before the write has */
+/* hit main memory - in this case the event count is <= 0 and the block */
+/* will be marked as ACTIVE - but could transition to DONE at any time */
+/* Also for a word copy event, the value written into the "done" word */
+/* can be the event interrupt cookie rather than the "source" value */
+/* this happens since the uCode does not wait for the write to have */
+/* occured before overwriting TMP_0 with the cookie */
+#define EP3_EVENT_FIRING(edev, event, cookie, done) \
+       (((((done) & ~(EV_TYPE_BCOPY | EV_TYPE_MASK_EVIRQ)) == (cookie).Cookie) || (done) == EP3_EVENT_ACTIVE) && \
+        (int) elan3_sdram_readl (edev, (event) + offsetof (E3_BlockCopyEvent, ev_Count)) <= 0)
+#define EP3_EVENT_FIRED(cookie, done) \
+       (((done) & ~(EV_TYPE_BCOPY | EV_TYPE_MASK_EVIRQ)) == (cookie).Cookie)
+
+
+/* Time limit to wait while event is firing and block write has not occured */
+#define EP3_EVENT_FIRING_TLIMIT                16384                   /* 1023 uS */
+
+#define EP3_INIT_COPY_EVENT(event, cookie, dest, intr)                                                 \
+{                                                                                                      \
+       (event).ev_Count  = 0;                                                                          \
+       (event).ev_Type   = (intr) ? EV_TYPE_BCOPY | EV_TYPE_EVIRQ | (cookie).Cookie : EV_TYPE_BCOPY;   \
+       (event).ev_Source = (cookie).Cookie | EV_WCOPY;                                                 \
+       (event).ev_Dest   = (dest) | EV_TYPE_BCOPY_WORD;                                                \
+}
+
+#if !defined(__ELAN3__)
+
+/* Generic input queues which can be polled */
+typedef struct ep3_inputq
+{
+    EP3_COOKIE         q_cookie;
+    unsigned int       q_slotSize;
+    unsigned int       q_slotCount;
+
+    void              *q_slots;
+    EP_ADDR            q_slotsAddr;
+
+    EP_INPUTQ_CALLBACK *q_callback;
+    void              *q_arg;
+
+    sdramaddr_t                q_desc;
+    E3_Addr            q_descAddr;
+
+    E3_Addr            q_base;
+    E3_Addr            q_top;
+    E3_Addr            q_fptr;
+
+    E3_uint32          q_waitCount;
+} EP3_INPUTQ;
+
+typedef struct ep3_outputq
+{
+    EP3_COOKIE         q_cookie;
+
+    unsigned int       q_slotCount;                            /* # slots allocated */
+    unsigned int       q_slotSize;                             /* size of each slot (rounded up) */
+
+    sdramaddr_t                q_elan;
+    E3_Addr            q_elanAddr;
+
+    void              *q_main;
+    E3_Addr            q_mainAddr;
+} EP3_OUTPUTQ;
+
+#endif /* !defined(__ELAN3__) */
+
+/* per-rail elan memory portion of device */
+typedef struct ep3_rail_elan
+{
+    E3_uint16           ProbeSource0[TR_TRACEROUTE_ENTRIES];   /* 32 byte aligned */
+    E3_uint16           ProbeSource1[TR_TRACEROUTE_ENTRIES];
+
+    E3_BlockCopyEvent   ProbeDone;                             /* 16 byte aligned */
+    E3_Event            ProbeStart;                            /* 8 byte aligned */
+    
+    E3_uint32           ProbeType;                             /* 4 byte aligned */
+    E3_uint32           ProbeLevel;
+
+    E3_uint32           NodeId;
+} EP3_RAIL_ELAN;
+
+/* values for ProbeType */
+#define PROBE_SINGLE                   0
+#define PROBE_MULTIPLE                 1
+/* number of attempts for each type */
+#define PROBE_SINGLE_ATTEMPTS          10
+#define PROBE_SINGLE_TIMEOUTS          5
+#define PROBE_MULTIPLE_ATTEMPTS                20
+#define PROBE_MULTIPLE_TIMEOUTS                10
+
+/* per-rail elan memory portsion of device */
+typedef struct ep3_rail_main
+{
+    E3_uint16          ProbeDest0[TR_TRACEROUTE_ENTRIES];      /* 32 byte aligned */
+    E3_uint16          ProbeDest1[TR_TRACEROUTE_ENTRIES];
+    
+    E3_uint32          ProbeDone;                              /* 4 byte aligned */
+    E3_uint32          ProbeResult;
+    E3_uint32          ProbeLevel;
+} EP3_RAIL_MAIN;
+
+#if !defined(__ELAN3__)
+
+struct ep3_rail
+{
+    EP_RAIL            Generic;                                /* Generic rail */
+
+    ELAN3_DEV          *Device;                                        /* Elan device we're using */
+    ELAN3_CTXT        *Ctxt;                                   /* Elan context struct */
+    ioaddr_t            CommandPort;                           /* commandport from context */
+    E3_Addr            CommandPortAddr;                        /*  and address mapped into elan */
+
+    ELAN3_ROUTE_TABLE  *RouteTable;                            /* routetable from context */
+    ELAN3MMU          *Elan3mmu;                               /* elanmmu from context */
+
+    EP3_COOKIE_TABLE    CookieTable;                           /* Event cookie table */
+
+    EP_CODE            ThreadCode;                             /* copy of thread code */
+    unsigned int        CommandPortEventTrap;                  /* flag to indicate command port eventint queue overflow trap */
+
+    sdramaddr_t         RailElan;                              /* Elan visible main/sdram portions of */
+    E3_Addr             RailElanAddr;                          /* device structure */
+    EP3_RAIL_MAIN      *RailMain;
+    E3_Addr            RailMainAddr;
+
+    /* small system message queues */
+    sdramaddr_t                QueueDescs;                             /* Input Queue descriptors */
+
+    /* Network position prober */
+    E3_Addr            ProbeStack;                             /* Network position thread command structure */
+    EP3_COOKIE         ProbeCookie;                            /*   event cookie for Done event */
+    kcondvar_t         ProbeWait;                              /*   place to wait on probe thread */
+    spinlock_t         ProbeLock;                              /*     and lock */
+    volatile int        ProbeDone;                             /*     and flag to indicate it's done */
+
+    E3_uint16          ProbeDest0[TR_TRACEROUTE_ENTRIES];      /* last result of CheckNetworkPosition */
+    E3_uint16          ProbeDest1[TR_TRACEROUTE_ENTRIES];
+    E3_uint32          ProbeResult;
+
+    long               ProbeLevelTick[ELAN_MAX_LEVELS];
+    long               SwitchBroadcastLevelTick;
+
+    /* rings for issueing dmas */
+    EP3_DMA_RING        DmaRings[EP3_NUM_RINGS];
+
+    /* retry lists for dmas */
+    struct list_head    DmaRetries[EP_NUM_RETRIES];            /* Dma retry lists */
+    struct list_head    DmaRetryFreeList;                      /*   and free list */
+    u_int              DmaRetryCount;                          /*   and total retry count */
+    u_int              DmaRetryReserved;                       /*   and number reserved */
+    u_int              DmaRetryThreadShouldStall;              /*   count of reasons to stall retries */
+    u_int              DmaRetryThreadStarted:1;                /*   dma retry thread running */
+    u_int              DmaRetryThreadShouldStop:1;             /*     but should stop */
+    u_int              DmaRetryThreadStopped:1;                /*     and now it's stopped */
+    u_int              DmaRetryInitialised:1;                  /* have initialise dma retries */
+
+    spinlock_t         DmaRetryLock;                           /*   spinlock protecting lists */
+    kcondvar_t         DmaRetryWait;                           /*   place retry thread sleeps */
+    long               DmaRetryTime;                           /*   and when it will next wakeup */
+    unsigned int        DmaRetrySleeping;                      /*   and it's sleeping there */
+
+    /* Network Identify Cookies */
+    E3_uint32         *MainCookies;                            /* One cookie allocator per-node for main*/
+    E3_Addr            ElanCookies;                            /*   and one for elan */
+    spinlock_t         CookieLock;                             /* spinlock to protect main cookies */
+
+    /* Halt operation flags for flushing. */
+    kmutex_t            HaltOpMutex;                           /* serialize access to halt operations */
+    unsigned int       HaltOpCompleted;                        /* flag to indicate halt operation completed */
+    kcondvar_t         HaltOpSleep;                            /*   place to wait for it to complete */
+
+    /* Network error state */
+    kcondvar_t         NetworkErrorSleep;                      /* place to sleep for network error halt operation */
+    u_int              NetworkErrorFlushed;                    /*   and flag to indicate flushed */
+
+
+    EP3_RAIL_STATS     Stats;                                  /* statistics */
+};
+
+/* support.c */
+
+extern ELAN3_OPS  ep3_elan3_ops;
+
+extern E3_uint32    LocalCookie (EP3_RAIL *rail, unsigned int remoteNode);
+extern E3_uint32    RemoteCookie (EP3_RAIL *rail, unsigned int remoteNode);
+
+extern void         InitialiseCookieTable (EP3_COOKIE_TABLE *table);
+extern void         DestroyCookieTable (EP3_COOKIE_TABLE *table);
+extern void         RegisterCookie (EP3_COOKIE_TABLE *table, EP3_COOKIE *cookie, 
+                                   E3_Addr event, EP3_COOKIE_OPS *ops, void *arg);
+extern void         DeregisterCookie (EP3_COOKIE_TABLE *table, EP3_COOKIE *cookie);
+extern EP3_COOKIE   *LookupCookie (EP3_COOKIE_TABLE *table, uint32_t cookie);
+extern EP3_COOKIE   *LookupEventCookie (EP3_RAIL *rail, EP3_COOKIE_TABLE *table, E3_Addr);
+
+extern int          DmaRingsCreate (EP3_RAIL *rail);
+extern void         DmaRingsRelease (EP3_RAIL *rail);
+extern int          IssueDma (EP3_RAIL *rail, E3_DMA_BE *dma, int type, int retryThread);
+
+extern int          IssueWaitevent (EP3_RAIL *rail, E3_Addr value);
+extern void         IssueSetevent (EP3_RAIL *rail, E3_Addr value);
+extern void         IssueRunThread (EP3_RAIL *rail, E3_Addr value);
+extern long         DmaRetryTime (int type);
+extern int          InitialiseDmaRetries (EP3_RAIL *rail);
+extern void         DestroyDmaRetries (EP3_RAIL *rail);
+extern int          ReserveDmaRetries (EP3_RAIL *rail, int count, EP_ATTRIBUTE attr);
+extern void         ReleaseDmaRetries (EP3_RAIL *rail, int count);
+extern void         StallDmaRetryThread (EP3_RAIL *rail);
+extern void         ResumeDmaRetryThread (EP3_RAIL *rail);
+extern void         QueueDmaForRetry (EP3_RAIL *rail, E3_DMA_BE *dma, int interval);
+extern void         QueueDmaOnStalledList (EP3_RAIL *rail, E3_DMA_BE *dma);
+extern void         FreeStalledDmas (EP3_RAIL *rail, unsigned int nodeId);
+
+extern void         SetQueueLocked(EP3_RAIL *rail, sdramaddr_t qaddr);
+
+/* threadcode_elan3.c */
+extern E3_Addr    ep3_init_thread (ELAN3_DEV *dev, E3_Addr fn, E3_Addr addr, sdramaddr_t stack,
+                                  int stackSize, int nargs, ...);
+
+/* probenetwork.c */
+extern int        ep3_init_probenetwork (EP3_RAIL *rail);
+extern void       ep3_destroy_probenetwork (EP3_RAIL *rail);
+extern void       ep3_probe_position_found (EP3_RAIL *rail, ELAN_POSITION *pos);
+extern int        ep3_probe_route (EP_RAIL *r, int level, int sw, int nodeid, int *linkup, int *linkdown, int attempts, EP_SWITCH *lsw);
+extern int        ep3_check_position (EP_RAIL *rail);
+
+/* neterr_elan3.c */
+extern void       ep3_neterr_fixup (EP_RAIL *r, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+
+/* kcomm_elan3.c */
+extern EP_RAIL    *ep3_create_rail (EP_SYS *sys, ELAN3_DEV *dev);
+extern void        ep3_destroy_rail (EP_RAIL *rail);
+
+extern int         ep3_start_rail (EP_RAIL *rail);
+extern void        ep3_stall_rail (EP_RAIL *rail);
+extern void        ep3_stop_rail (EP_RAIL *rail);
+
+extern void       ep3_position_found (EP_RAIL *rail, ELAN_POSITION *pos);
+
+extern sdramaddr_t ep3_sdram_alloc (EP_RAIL *rail, EP_ADDR addr, unsigned int size);
+extern void        ep3_sdram_free (EP_RAIL *rail, sdramaddr_t addr, unsigned int size);
+extern void        ep3_sdram_writeb (EP_RAIL *rail, sdramaddr_t addr, unsigned char val);
+
+extern void        ep3_flush_tlb (EP_RAIL *r);
+extern void        ep3_load_system_route (EP_RAIL *r, unsigned int vp, unsigned int lowNode, unsigned int highNode);
+extern void        ep3_load_node_route (EP_RAIL *r, unsigned int nodeId);
+extern void        ep3_unload_node_route (EP_RAIL *r, unsigned int nodeId);
+extern void        ep3_lower_filter (EP_RAIL *r, unsigned int nodeId);
+extern void        ep3_raise_filter (EP_RAIL *rail, unsigned int nodeId);
+extern void        ep3_node_disconnected (EP_RAIL *r, unsigned int nodeId);
+
+extern void        ep3_fillout_stats(EP_RAIL *rail, char *str);
+
+/* kmap_elan3.c */
+extern void        ep3_kaddr_map (EP_RAIL *r, EP_ADDR eaddr, virtaddr_t kaddr, unsigned int len, unsigned int perm, int ep_attr);
+extern void        ep3_sdram_map (EP_RAIL *r, EP_ADDR eaddr, sdramaddr_t saddr, unsigned int len, unsigned int perm, int ep_attr);
+extern void        ep3_ioaddr_map (EP_RAIL *r, EP_ADDR eaddr, ioaddr_t ioaddr, unsigned int len, unsigned int perm);
+extern void        ep3_unmap (EP_RAIL *r, EP_ADDR eaddr, unsigned int len);
+extern void       *ep3_dvma_reserve (EP_RAIL *r, EP_ADDR eaddr, unsigned int npages);
+extern void        ep3_dvma_release (EP_RAIL *r, EP_ADDR eaddr, unsigned int npages, void *private);
+extern void        ep3_dvma_set_pte (EP_RAIL *r, void *private, unsigned int index, physaddr_t paddr, unsigned int perm);
+extern physaddr_t  ep3_dvma_read_pte (EP_RAIL *r, void *private, unsigned int index);
+extern void        ep3_dvma_unload (EP_RAIL *r, void *private, unsigned int index, unsigned int npages);
+
+/* kmsg_elan3.c */
+extern EP_INPUTQ  *ep3_alloc_inputq (EP_RAIL *r, unsigned int qnum, unsigned int slotSize, unsigned int slotCount,
+                                    EP_INPUTQ_CALLBACK *callback, void *arg);
+extern void        ep3_free_inputq (EP_RAIL *r, EP_INPUTQ *q);
+extern void        ep3_enable_inputq (EP_RAIL *r, EP_INPUTQ *q);
+extern void        ep3_disable_inputq (EP_RAIL *r, EP_INPUTQ *q);
+extern int         ep3_poll_inputq (EP_RAIL *r, EP_INPUTQ *q, int maxCount, EP_INPUTQ_HANDLER *handler, void *arg);
+extern EP_OUTPUTQ *ep3_alloc_outputq (EP_RAIL *r, unsigned int slotSize, unsigned int slotCount);
+extern void        ep3_free_outputq (EP_RAIL *r, EP_OUTPUTQ *q);
+extern void       *ep3_outputq_msg (EP_RAIL *r, EP_OUTPUTQ *q, unsigned int slotNum);
+extern int         ep3_outputq_state (EP_RAIL *r, EP_OUTPUTQ *q, unsigned int slotNum);
+extern int         ep3_outputq_send (EP_RAIL *r, EP_OUTPUTQ *q, unsigned int slotNum, unsigned int size,
+                                    unsigned int nodeId, unsigned int qnum, unsigned int retries);
+
+/* support_elan3.c */
+extern void        ep3_flush_filters (EP_RAIL *r);
+extern void        ep3_flush_queues (EP_RAIL *r);
+
+#endif /* !defined(__ELAN3__) */
+
+#endif /* __EP_KCOMM_ELAN3_H */
Index: linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kcomm_elan4.c        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan4.c     2005-06-01 23:12:54.667428768 -0400
@@ -0,0 +1,526 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kcomm_elan4.c,v 1.16.2.3 2004/11/30 12:02:17 mike Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm_elan4.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "conf_linux.h"
+
+extern EP_CODE threadcode_elan4;
+
+unsigned int
+ep4_create_rails (EP_SYS *sys, unsigned int disabled)
+{
+    unsigned int rmask = 0;
+    ELAN4_DEV   *dev;
+    EP_RAIL     *rail;
+    int          i;
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+    {
+       if ((dev = elan4_reference_device (i, ELAN4_STATE_STARTED)) != NULL)
+       {
+           if ((rail = ep4_create_rail (sys, dev)) == NULL)
+               elan4_dereference_device (dev);
+           else
+           {
+               if (disabled & (1 << rail->Number))
+                   printk ("%s: auto-start of device disabled by configuration\n", rail->Name);
+               else
+                   ep_start_rail (rail);
+               
+               ep_procfs_rail_init(rail);
+
+               rmask |= (1 << rail->Number);
+           }
+       }
+    }
+
+    if (rmask)
+       qsnet_debug_alloc();
+
+    return rmask;
+}
+
+EP_RAIL *
+ep4_create_rail (EP_SYS *sys, ELAN4_DEV *dev)
+{
+    EP4_RAIL *rail;
+    int res;
+
+    KMEM_ZALLOC (rail, EP4_RAIL *, sizeof (EP4_RAIL), 1);
+
+    if (rail == NULL)
+       return (EP_RAIL *) NULL;
+    
+    if ((res = ep_init_rail (sys, &rail->r_generic)) != 0)
+    {
+       KMEM_FREE (rail, sizeof (EP4_RAIL));
+       return (EP_RAIL *) NULL;
+    }
+
+    rail->r_ctxt.ctxt_dev = dev;
+
+    /* install our rail operations */
+    rail->r_generic.Operations.DestroyRail      = ep4_destroy_rail;
+    rail->r_generic.Operations.StartRail        = ep4_start_rail;
+    rail->r_generic.Operations.StallRail        = ep4_stall_rail;
+    rail->r_generic.Operations.StopRail         = ep4_stop_rail;    
+
+    rail->r_generic.Operations.SdramAlloc       = ep4_sdram_alloc;
+    rail->r_generic.Operations.SdramFree        = ep4_sdram_free;
+    rail->r_generic.Operations.SdramWriteb      = ep4_sdram_writeb;
+
+    rail->r_generic.Operations.KaddrMap         = ep4_kaddr_map;
+    rail->r_generic.Operations.SdramMap         = ep4_sdram_map;
+    rail->r_generic.Operations.Unmap            = ep4_unmap;
+
+    rail->r_generic.Operations.DvmaReserve     = ep4_dvma_reserve;
+    rail->r_generic.Operations.DvmaRelease     = ep4_dvma_release;
+    rail->r_generic.Operations.DvmaSetPte      = ep4_dvma_set_pte;
+    rail->r_generic.Operations.DvmaReadPte     = ep4_dvma_read_pte;
+    rail->r_generic.Operations.DvmaUnload      = ep4_dvma_unload;
+    rail->r_generic.Operations.FlushTlb                = ep4_flush_tlb;
+
+    rail->r_generic.Operations.ProbeRoute       = ep4_probe_route;
+
+    rail->r_generic.Operations.PositionFound    = ep4_position_found;
+    rail->r_generic.Operations.CheckPosition    = ep4_check_position;
+    rail->r_generic.Operations.NeterrFixup      = ep4_neterr_fixup;
+
+    rail->r_generic.Operations.LoadSystemRoute  = ep4_load_system_route;
+
+    rail->r_generic.Operations.LoadNodeRoute    = ep4_load_node_route;
+    rail->r_generic.Operations.UnloadNodeRoute  = ep4_unload_node_route;
+    rail->r_generic.Operations.LowerFilter     = ep4_lower_filter;
+    rail->r_generic.Operations.RaiseFilter     = ep4_raise_filter;
+    rail->r_generic.Operations.NodeDisconnected = ep4_node_disconnected;
+
+    rail->r_generic.Operations.FlushFilters     = ep4_flush_filters;
+    rail->r_generic.Operations.FlushQueues     = ep4_flush_queues;
+
+    rail->r_generic.Operations.AllocInputQ     = ep4_alloc_inputq;
+    rail->r_generic.Operations.FreeInputQ      = ep4_free_inputq;
+    rail->r_generic.Operations.EnableInputQ     = ep4_enable_inputq;
+    rail->r_generic.Operations.DisableInputQ    = ep4_disable_inputq;
+    rail->r_generic.Operations.PollInputQ      = ep4_poll_inputq;
+
+    rail->r_generic.Operations.AllocOutputQ     = ep4_alloc_outputq;
+    rail->r_generic.Operations.FreeOutputQ     = ep4_free_outputq;
+    rail->r_generic.Operations.OutputQMsg      = ep4_outputq_msg;
+    rail->r_generic.Operations.OutputQState     = ep4_outputq_state;
+    rail->r_generic.Operations.OutputQSend     = ep4_outputq_send;
+
+    rail->r_generic.Operations.FillOutStats     = ep4_fillout_stats;
+    rail->r_generic.Operations.Debug           = ep4_debug_rail;
+
+    rail->r_generic.Devinfo = dev->dev_devinfo;
+
+    printk ("%s: connected via elan4 rev%c device %d\n", rail->r_generic.Name,
+           'a' + dev->dev_devinfo.dev_revision_id, dev->dev_instance);
+
+    return (EP_RAIL *) rail;
+}
+
+void
+ep4_destroy_rail (EP_RAIL *r)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+
+    elan4_dereference_device (rail->r_ctxt.ctxt_dev);
+
+    KMEM_FREE (rail, sizeof (EP4_RAIL));
+}
+
+static int
+ep4_attach_rail (EP4_RAIL *r)
+{
+    EP4_RAIL  *rail = (EP4_RAIL *) r;
+    ELAN4_DEV *dev  = rail->r_ctxt.ctxt_dev;
+    unsigned   ctx;
+
+    if (elan4_insertctxt (dev, &rail->r_ctxt, &ep4_trap_ops) != 0)
+       return -ENOMEM;
+    
+    if ((rail->r_routetable = elan4_alloc_routetable (dev, 4)) == NULL)        /* 512 << 4 == 8192 entries */
+    {
+       elan4_removectxt (dev, &rail->r_ctxt);
+       return -ENOMEM;
+    }
+    elan4_set_routetable (&rail->r_ctxt, rail->r_routetable);
+
+    /* Attach to the kernel comms nextwork context */
+    if (elan4_attach_filter (&rail->r_ctxt, ELAN4_KCOMM_CONTEXT_NUM) < 0)
+    {
+       elan4_free_routetable (dev, rail->r_routetable);
+       elan4_removectxt (dev, &rail->r_ctxt);
+
+       return -EBUSY;
+    }
+
+    for (ctx = ELAN4_KCOMM_BASE_CONTEXT_NUM; ctx <= ELAN4_KCOMM_TOP_CONTEXT_NUM; ctx++)
+       elan4_attach_filter (&rail->r_ctxt, ctx);
+
+    return 0;
+}
+
+static void
+ep4_detach_rail (EP4_RAIL *rail)
+{
+    ELAN4_DEV *dev = rail->r_ctxt.ctxt_dev;
+    unsigned   ctx;
+
+    elan4_detach_filter (&rail->r_ctxt, ELAN4_KCOMM_CONTEXT_NUM);
+
+    for (ctx = ELAN4_KCOMM_BASE_CONTEXT_NUM; ctx <= ELAN4_KCOMM_TOP_CONTEXT_NUM; ctx++)
+       elan4_detach_filter (&rail->r_ctxt, ctx);
+
+    if (rail->r_routetable)
+    {
+       elan4_set_routetable (&rail->r_ctxt, NULL);
+       elan4_free_routetable (dev, rail->r_routetable);
+    }
+
+    elan4_removectxt (dev, &rail->r_ctxt);
+}
+
+int
+ep4_start_rail (EP_RAIL *r)
+{
+    EP4_RAIL     *rail = (EP4_RAIL *) r;
+    ELAN4_DEV    *dev  = rail->r_ctxt.ctxt_dev;
+    E4_InputQueue qdesc;
+    int           i, res;
+
+    if ((res = ep4_attach_rail (rail)) < 0)
+       return res;
+
+    /* Initialise main interrupt cookie table */
+    spin_lock_init (&rail->r_intcookie_lock);
+    for (i = 0; i < EP4_INTCOOKIE_HASH_SIZE; i++)
+       INIT_LIST_HEAD (&rail->r_intcookie_hash[i]);
+
+    kmutex_init (&rail->r_haltop_mutex);
+    kcondvar_init (&rail->r_haltop_sleep);
+    spin_lock_init (&rail->r_haltop_lock);
+
+    spin_lock_init (&rail->r_cookie_lock);
+
+    INIT_LIST_HEAD (&rail->r_ecq_list[EP4_ECQ_EVENT]);
+    INIT_LIST_HEAD (&rail->r_ecq_list[EP4_ECQ_ATOMIC]);
+    INIT_LIST_HEAD (&rail->r_ecq_list[EP4_ECQ_SINGLE]);
+    INIT_LIST_HEAD (&rail->r_ecq_list[EP4_ECQ_MAIN]);
+    spin_lock_init (&rail->r_ecq_lock);
+
+    ep_kthread_init (&rail->r_retry_thread);
+    INIT_LIST_HEAD (&rail->r_retry_ops);
+
+    INIT_LIST_HEAD (&rail->r_neterr_ops);
+
+    kmutex_init (&rail->r_flush_mutex);
+    kcondvar_init (&rail->r_flush_sleep);
+
+    /* Allocate the elan visible sdram/main memory */
+    if ((rail->r_elan = ep_alloc_elan (&rail->r_generic, sizeof (EP4_RAIL_ELAN), 0, &rail->r_elan_addr)) == 0 ||
+       (rail->r_main = ep_alloc_main (&rail->r_generic, sizeof (EP4_RAIL_MAIN), 0, &rail->r_main_addr)) == 0)
+    {
+       goto failed;
+    }
+
+    for (i = 0; i < EP_NUM_SYSTEMQ; i++)
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_qevents[i].ev_CountAndType), 0);
+
+    elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_flush_event.ev_CountAndType), E4_EVENT_INIT_VALUE (0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+
+    /* Allocate the system input queues at their fixed elan address */
+    /*   avoid sdram address aliasing by allocating the min sdram pagesize */
+    if (! (rail->r_queuedescs= ep_alloc_memory_elan (&rail->r_generic, EP_SYSTEM_QUEUE_BASE, SDRAM_PAGE_SIZE, EP_PERM_ALL, 0)))
+       goto failed;
+
+    /* Initialise the input queue descriptor as "full" with no event */
+    qdesc.q_bptr    = 0;
+    qdesc.q_fptr    = 8;
+    qdesc.q_control = E4_InputQueueControl(qdesc.q_bptr, qdesc.q_fptr, 8);
+    qdesc.q_event   = 0;
+
+    for (i = 0; i < EP_NUM_SYSTEMQ; i++)
+       elan4_sdram_copyq_to_sdram (dev, &qdesc, EP_SYSTEMQ_DESC (rail->r_queuedescs, i), sizeof (E4_InputQueue));
+
+    /* Allocate the resource map for command queue mappings */
+    if ((rail->r_ecq_rmap = ep_rmallocmap (EP4_ECQ_RMAPSIZE, "r_ecq_rmap", 1)) == NULL)
+       goto failed;
+    
+    ep_rmfree (rail->r_ecq_rmap, EP4_ECQ_TOP - EP4_ECQ_BASE, EP4_ECQ_BASE);
+
+    /* register an interrupt cookie & allocate command queues for command queue flushing */
+    rail->r_flush_mcq = ep4_get_ecq (rail, EP4_ECQ_MAIN, 4);
+    rail->r_flush_ecq = ep4_get_ecq (rail, EP4_ECQ_EVENT, 1);
+
+    if (rail->r_flush_mcq == NULL || rail->r_flush_ecq == NULL)
+       goto failed;
+
+    ep4_register_intcookie (rail, &rail->r_flush_intcookie, rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_flush_event), ep4_flush_interrupt, rail);
+
+    /* startup the retry thread */
+    if (kernel_thread_create (ep4_retry_thread, (void *) rail) == 0)
+       goto failed;
+    ep_kthread_started (&rail->r_retry_thread);
+    
+    ep4_initialise_dma_retries (rail);
+
+    if ((rail->r_event_ecq = ep4_alloc_ecq (rail, CQ_Size1K)) == NULL)
+       goto failed;
+    
+    rail->r_threadcode = threadcode_elan4;
+    if (ep_loadcode (&rail->r_generic, &rail->r_threadcode))
+       goto failed;
+
+    elan4_flush_icache (&rail->r_ctxt);
+
+    if (ep4_probe_init (rail))
+       goto failed;
+
+    /* can now drop the context filter for the system context */
+    elan4_set_filter (&rail->r_ctxt, ELAN4_KCOMM_CONTEXT_NUM, E4_FILTER_HIGH_PRI);
+
+    return 0;
+
+ failed:
+    printk ("ep4_start_rail: failed for rail '%s'\n", rail->r_generic.Name);
+    ep4_stop_rail (&rail->r_generic);
+
+    return -ENOMEM;
+}
+
+void
+ep4_stall_rail (EP_RAIL *r)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+    unsigned  ctx;
+
+    /* Raise all the context filters */
+    elan4_set_filter (&rail->r_ctxt, ELAN4_KCOMM_CONTEXT_NUM, E4_FILTER_DISCARD_ALL);
+
+    for (ctx = ELAN4_KCOMM_BASE_CONTEXT_NUM; ctx <= ELAN4_KCOMM_TOP_CONTEXT_NUM; ctx++)
+       elan4_set_filter (&rail->r_ctxt, ctx, E4_FILTER_DISCARD_ALL);
+}
+
+void
+ep4_stop_rail (EP_RAIL *r)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+
+    if (rail->r_generic.State == EP_RAIL_STATE_RUNNING) /* undo ep4_position_found() */
+    {
+       ELAN_POSITION *pos  = &rail->r_generic.Position;
+       EP_ADDR        addr = elan4_sdram_readq (rail->r_ctxt.ctxt_dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_cookies));
+
+       ep_free_elan (&rail->r_generic, addr, pos->pos_nodes * sizeof (E4_uint64));
+
+       KMEM_FREE (rail->r_cookies, pos->pos_nodes * sizeof (E4_uint64));
+    }
+
+    ep4_probe_destroy (rail);
+
+    ep_unloadcode (&rail->r_generic, &rail->r_threadcode);
+
+    if (rail->r_event_ecq)
+       ep4_free_ecq (rail, rail->r_event_ecq);
+    rail->r_event_ecq = NULL;
+
+    ep4_finalise_dma_retries (rail);
+
+    ep_kthread_stop (&rail->r_retry_thread);
+    ep_kthread_destroy (&rail->r_retry_thread);
+
+    if (rail->r_flush_intcookie.int_arg)
+       ep4_deregister_intcookie (rail, &rail->r_flush_intcookie);
+    rail->r_flush_intcookie.int_arg = NULL;
+
+    if (rail->r_flush_mcq)
+       ep4_put_ecq (rail, rail->r_flush_mcq, 4);
+    rail->r_flush_mcq = NULL;
+
+    if (rail->r_flush_ecq)
+       ep4_put_ecq (rail, rail->r_flush_ecq, 1);
+    rail->r_flush_ecq = NULL;
+
+    if (rail->r_ecq_rmap)
+       ep_rmfreemap (rail->r_ecq_rmap);
+    
+    if (rail->r_queuedescs)
+       ep_free_memory_elan (&rail->r_generic, EP_SYSTEM_QUEUE_BASE);
+    rail->r_queuedescs = 0;
+
+    if (rail->r_elan)
+       ep_free_elan (&rail->r_generic, rail->r_elan_addr, sizeof (EP4_RAIL_ELAN));
+    rail->r_elan = 0;
+
+    if (rail->r_main)
+       ep_free_main (&rail->r_generic, rail->r_main_addr, sizeof (EP4_RAIL_MAIN));
+    rail->r_main = NULL;
+
+    kcondvar_destroy (&rail->r_flush_sleep);
+    kmutex_destroy (&rail->r_flush_mutex);
+
+    spin_lock_destroy (&rail->r_ecq_lock);
+    spin_lock_destroy (&rail->r_cookie_lock);
+
+    spin_lock_destroy (&rail->r_haltop_lock);
+    kcondvar_destroy(&rail->r_haltop_sleep);
+    kmutex_destroy (&rail->r_haltop_mutex);
+    spin_lock_destroy (&rail->r_intcookie_lock);
+
+    ep4_detach_rail (rail);
+}
+
+void
+ep4_position_found (EP_RAIL *r, ELAN_POSITION *pos)
+{
+    EP4_RAIL   *rail = (EP4_RAIL *) r;
+    sdramaddr_t cookies;
+    EP_ADDR     addr;
+    int         i;
+
+    KMEM_ZALLOC (rail->r_cookies, E4_uint64 *, pos->pos_nodes * sizeof (E4_uint64), 1);
+
+    if (! (cookies = ep_alloc_elan (&rail->r_generic, pos->pos_nodes * sizeof (E4_uint64), 0, &addr)))
+       panic ("ep4_position_found: cannot allocate elan cookies array\n");
+
+    for (i = 0; i < pos->pos_nodes; i++)
+       elan4_sdram_writeq (rail->r_ctxt.ctxt_dev, cookies + (i * sizeof (E4_uint64)), 0);
+    
+    for (i = 0; i < pos->pos_nodes; i++)
+       rail->r_cookies[i] = 0;
+
+    elan4_sdram_writeq (rail->r_ctxt.ctxt_dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_nodeid), pos->pos_nodeid);
+    elan4_sdram_writeq (rail->r_ctxt.ctxt_dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_cookies), addr);
+
+    ep4_probe_position_found (rail, pos);
+}
+
+sdramaddr_t
+ep4_sdram_alloc (EP_RAIL *r, EP_ADDR addr, unsigned size)
+{
+    ELAN4_DEV *dev = ((EP4_RAIL *) r)->r_ctxt.ctxt_dev;
+
+    if (size >= SDRAM_PAGE_SIZE)
+       return elan4_sdram_alloc (dev, size);
+    else
+    {
+       sdramaddr_t block = elan4_sdram_alloc (dev, SDRAM_PAGE_SIZE);
+       sdramaddr_t sdram = block + (addr & (SDRAM_PAGE_SIZE-1));
+
+       /* free of the portion before sdram */
+       if (sdram > block)
+           elan4_sdram_free (dev, block, sdram - block);
+
+       /* free of the portion after sdram */
+       if ((block + SDRAM_PAGE_SIZE) > (sdram + size))
+           elan4_sdram_free (dev, sdram + size, block + SDRAM_PAGE_SIZE - (sdram + size));
+
+       return sdram;
+    }
+}
+
+void
+ep4_sdram_free (EP_RAIL *r, sdramaddr_t addr, unsigned size)
+{
+    elan4_sdram_free (((EP4_RAIL *) r)->r_ctxt.ctxt_dev, addr, size);
+}
+
+void
+ep4_sdram_writeb (EP_RAIL *r, sdramaddr_t addr, unsigned char val)
+{
+    elan4_sdram_writeb (((EP4_RAIL *) r)->r_ctxt.ctxt_dev, addr, val);
+}
+
+void
+ep4_flush_tlb (EP_RAIL *r)
+{
+    elan4mmu_flush_tlb (((EP4_RAIL *) r)->r_ctxt.ctxt_dev);
+}
+
+void
+ep4_load_system_route (EP_RAIL *r, unsigned vp, unsigned lowNode, unsigned highNode)
+{
+    EP4_RAIL  *rail = (EP4_RAIL *) r;
+    ELAN4_DEV *dev  = rail->r_ctxt.ctxt_dev;
+    E4_VirtualProcessEntry route;
+
+    if (elan4_generate_route (&rail->r_generic.Position, &route, ELAN4_KCOMM_CONTEXT_NUM, 
+                             lowNode, highNode, FIRST_SYSTEM_PACKET | FIRST_HIGH_PRI | FIRST_TIMEOUT(3)) < 0)
+    {
+       panic ("ep4_load_system_route: generate route failed\n");
+       /* NOTREACHED */
+    }
+
+    elan4_write_route (dev, rail->r_routetable, vp, &route);
+}
+
+void
+ep4_load_node_route (EP_RAIL *r, unsigned nodeId)
+{
+    EP4_RAIL  *rail = (EP4_RAIL *) r;
+    ELAN4_DEV *dev  = rail->r_ctxt.ctxt_dev;
+    E4_VirtualProcessEntry route;
+
+    if (elan4_generate_route (&rail->r_generic.Position, &route, EP4_CONTEXT_NUM(rail->r_generic.Position.pos_nodeid),
+                             nodeId, nodeId, FIRST_SYSTEM_PACKET | FIRST_TIMEOUT(3)) < 0)
+    {
+       panic ("ep4_load_node_route: generate route failed\n");
+       /* NOTREACHED */
+    }
+
+    elan4_write_route (dev, rail->r_routetable, EP_VP_DATA(nodeId), &route);
+}
+
+void
+ep4_unload_node_route (EP_RAIL *r, unsigned nodeId)
+{
+    EP4_RAIL  *rail = (EP4_RAIL *) r;
+    ELAN4_DEV *dev  = rail->r_ctxt.ctxt_dev;
+    
+    elan4_invalidate_route (dev, rail->r_routetable, EP_VP_DATA(nodeId));
+}
+
+void
+ep4_lower_filter (EP_RAIL *r, unsigned nodeId)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+
+    elan4_set_filter (&rail->r_ctxt, EP4_CONTEXT_NUM(nodeId), E4_FILTER_HIGH_PRI);
+}
+
+void
+ep4_raise_filter (EP_RAIL *r, unsigned nodeId)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+
+    elan4_set_filter (&rail->r_ctxt, EP4_CONTEXT_NUM(nodeId), E4_FILTER_DISCARD_ALL);
+}
+
+void
+ep4_node_disconnected (EP_RAIL *r, unsigned nodeId)
+{
+    ep4_free_stalled_dmas ((EP4_RAIL *) r, nodeId);
+}
+
+void
+ep4_fillout_stats(EP_RAIL *r, char *str) 
+{
+    /* no stats here yet */
+    /* EP4_RAIL *ep4rail = (EP4_RAIL *)r; */
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan4.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kcomm_elan4.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kcomm_elan4.h     2005-06-01 23:12:54.668428616 -0400
@@ -0,0 +1,443 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __EP_KCOMM_ELAN4_H
+#define __EP_KCOMM_ELAN4_H
+
+#ident "@(#)$Id: kcomm_elan4.h,v 1.16.2.2 2004/12/14 10:19:14 mike Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm_elan4.h,v $*/
+
+#include <elan4/types.h>
+
+#include <elan4/dma.h>
+#include <elan4/events.h>
+#include <elan4/commands.h>
+
+#if !defined(__elan4__)
+#include <elan4/device.h>
+#endif /* ! defined(__elan4__) */
+
+/* private address allocation */
+#define EP4_TEXT_BASE                  0xF8000000              /* base address for thread code (defined in makerules.elan4) */
+#define EP4_ECQ_BASE                   0xFF000000              /* address space for mapping command queues */
+#define EP4_ECQ_TOP                    0xFF800000
+
+#define EP4_ECQ_RMAPSIZE               128
+#define EP4_STACK_SIZE                 1024                    /* default thread code stack size */
+#define EP4_MAX_LEVELS                 8                       /* same as ELAN_MAX_LEVELS */
+
+/* context number allocation */
+#define EP4_CONTEXT_NUM(nodeId)                (ELAN4_KCOMM_BASE_CONTEXT_NUM + (nodeId))
+#define EP4_CONTEXT_ISDATA(ctx)                ((ctx) >= ELAN4_KCOMM_BASE_CONTEXT_NUM && \
+                                        (ctx) <= ELAN4_KCOMM_TOP_CONTEXT_NUM)
+#define EP4_CONTEXT_TO_NODE(ctx)       ((ctx) - ELAN4_KCOMM_BASE_CONTEXT_NUM)
+
+/*
+ * network error cookie format:
+ *  -------------------------------------------------
+ *  | unique cookie value | Remote | DMA | Location |
+ *  -------------------------------------------------
+ * [63:4] Cookie   - unique cookie number
+ * [3]    Thread   - cookie generated by thread code
+ * [2]   Remote   - cookie generated by remote end
+ * [1]    STEN     - cookie is for a STEN packet
+ * [0]    DMA      - cookie is for a DMA
+ */
+#define EP4_COOKIE_DMA         (1    << 0)
+#define EP4_COOKIE_STEN                (1    << 1)
+#define EP4_COOKIE_REMOTE      (1    << 2)
+#define EP4_COOKIE_THREAD      (1    << 3)
+#define EP4_COOKIE_INC         (1ull << 4)
+
+#define EP4_COOKIE_STRING(val) ((val) & ~(EP4_COOKIE_INC-1)) >> 4, \
+                               ((val) & EP4_COOKIE_DMA)    ? ",dma" : "", \
+                               ((val) & EP4_COOKIE_REMOTE) ? ",remote" : "", \
+                               ((val) & EP4_COOKIE_THREAD) ? ",thread" : "", \
+                               ((val) & EP4_COOKIE_STEN)   ? ",sten" : ""
+/*
+ * Done "word" values 
+ */
+#define EP4_STATE_FREE         0
+#define EP4_STATE_ACTIVE       1
+#define EP4_STATE_FINISHED     2
+#define EP4_STATE_FAILED       3
+#define EP4_STATE_PRIVATE      4
+
+#define EP4_EVENT_FIRING_TLIMIT        16384                   /* 1023 uS */
+
+/* forward declarations */
+typedef struct ep4_rail        EP4_RAIL;
+
+#if !defined(__elan4__)
+
+typedef struct ep4_intcookie
+{
+    struct list_head            int_link;
+    E4_uint64                   int_val;
+    void                      (*int_callback)(EP4_RAIL *rail, void *arg);
+    void                       *int_arg;
+} EP4_INTCOOKIE;
+
+#define EP4_INTCOOKIE_HASH_SIZE        256
+#define EP4_INTCOOKIE_HASH(a)          ((((a) >> 3) ^ ((a) >> 7) ^ ((a) >> 11)) & (EP4_INTCOOKIE_HASH_SIZE-1))
+
+typedef struct ep4_ecq
+{
+    struct list_head   ecq_link;                               /* linked on r_ecq_list */
+    ELAN4_INTOP                ecq_intop;                              /* main interrupt op space */
+    ELAN4_CQ          *ecq_cq;                                 /* command queue */
+    E4_Addr            ecq_addr;                               /* address mapped into elan */
+    unsigned int       ecq_avail;                              /* # dwords still available */
+
+    spinlock_t         ecq_lock;                               /* spinlock for main accesses */
+    sdramaddr_t                ecq_event;                              /* event for flushing "event" queues */
+    EP_ADDR            ecq_event_addr;
+    struct ep4_ecq     *ecq_flushcq;                           /*  and command port to issue setevent to */
+} EP4_ECQ;
+
+#define EP4_ECQ_EVENT          0                               /* command queues targetted by multi-blocks events */
+#define EP4_ECQ_ATOMIC         1                               /* command queues targetted by atomic store operations */
+#define EP4_ECQ_SINGLE         2                               /* command queues targetted by single word commands from main */
+#define EP4_ECQ_MAIN           3                               /* command queues targetted by multi word commands from main */
+#define EP4_NUM_ECQ            4
+
+#define EP4_ECQ_Size(which)            ((which) == EP4_ECQ_EVENT  ? CQ_Size64K : \
+                                        (which) == EP4_ECQ_ATOMIC ? CQ_Size8K  : \
+                                        (which) == EP4_ECQ_SINGLE ? CQ_Size1K  : \
+                                        (which) == EP4_ECQ_MAIN   ? CQ_Size8K  : \
+                                        CQ_Size1K)
+
+typedef struct ep4_dma_retry
+{
+    struct list_head    retry_link;                            /* chained on free/retry list */
+    unsigned long      retry_time;                             /* "lbolt" to retry at */
+    E4_DMA             retry_dma;                              /* DMA (in main memory) */
+} EP4_DMA_RETRY;
+
+#define EP4_DMA_RETRY_CQSIZE           CQ_Size8K                               /* size of command queue for dma retry */
+#define EP4_DMA_RETRY_FLOWCNT          (CQ_Size(EP4_DMA_RETRY_CQSIZE)/72)      /* # of reissued DMA's which can fit in */
+
+typedef struct ep4_inputq
+{
+    EP4_INTCOOKIE      q_intcookie;
+    unsigned int       q_slotSize;
+    unsigned int       q_slotCount;
+
+    void              *q_slots;
+    EP_ADDR            q_slotsAddr;
+    
+    EP_INPUTQ_CALLBACK *q_callback;
+    void              *q_arg;
+
+    sdramaddr_t                q_desc;
+    EP_ADDR            q_descAddr;
+    EP_ADDR            q_eventAddr;
+    EP4_ECQ           *q_wcq;                                  /* command queue to issue waitevent to */
+    EP4_ECQ           *q_ecq;                                  /* command queue targetted by event to generate interrupt */
+
+    EP_ADDR            q_fptr;                                 /* cached current front pointer */
+    EP_ADDR            q_last;                                 /* elan addr for last queue slot  */
+
+    atomic_t           q_fired;                                /* atomic flag that interrupt received */
+    unsigned int       q_count;                                /* count of slots consumed */
+} EP4_INPUTQ;
+
+typedef struct ep4_outputq
+{
+    spinlock_t         q_lock;
+    unsigned int       q_slotCount;
+    unsigned int       q_slotSize;
+    unsigned int        q_dwords;
+    ELAN4_CQ          *q_cq;
+    void              *q_main;
+    EP_ADDR            q_mainAddr;
+    unsigned int       q_retries;
+} EP4_OUTPUTQ;
+
+#endif /* ! defined(__elan4__) */
+
+typedef struct ep4_check_sten
+{
+    E4_uint64          c_reset_event_cmd;                      /* WRITEDWORD to reset start event */
+    E4_uint64          c_reset_event_value;
+
+    E4_uint64          c_open;                                 /* OPEN VP_PROBE(lvl) */
+    E4_uint64          c_trans_traceroute0;                    /* SENDTRANS TR_TRACEROUTE 0s */
+    E4_uint64          c_addr_traceroute0;
+    E4_uint64          c_data_traceroute0[8];
+    E4_uint64          c_trans_traceroute1;                    /* SENDTRANS TR_TRACEROUTE 1s */
+    E4_uint64          c_addr_traceroute1;
+    E4_uint64          c_data_traceroute1[8];
+    E4_uint64          c_trans_sendack;                        /* SENDTRANS SENDACK */
+    E4_uint64          c_addr_sendack;
+    
+    E4_uint64          c_guard_ok;                             /* GUARD OK - write level */
+    E4_uint64          c_writedword_ok;
+    E4_uint64          c_value_ok;
+    
+    E4_uint64          c_guard_fail;                           /* GUARD FAIL - chain setevent/write fail */
+    E4_uint64          c_setevent_fail;
+    E4_uint64          c_setevent_nop;
+    E4_uint64          c_nop_pad;
+} EP4_CHECK_STEN;
+
+#define EP4_CHECK_STEN_NDWORDS (sizeof (EP4_CHECK_STEN) >> 3)
+
+typedef struct ep4_rail_elan
+{
+    EP4_CHECK_STEN     r_check_sten[EP4_MAX_LEVELS];
+    E4_Event32         r_check_fail;                                   /* Check failed (== r_check_start[-1]) */
+    E4_Event32          r_check_start[EP4_MAX_LEVELS];
+
+    E4_Event32         r_qevents[EP_NUM_SYSTEMQ];
+    E4_Event32         r_flush_event;
+
+    E4_uint64          r_nodeid;
+#ifdef __elan4__
+    E4_uint64         *r_cookies;
+#else
+    E4_Addr            r_cookies;
+#endif
+} EP4_RAIL_ELAN;
+
+#define TRACEROUTE_ENTRIES     16                      /* 2 * ELAN_MAX_LEVELS */
+#define TRACEROUTE_NDWORDS     (TRACEROUTE_ENTRIES/2)
+
+typedef struct ep4_rail_main
+{
+    E4_uint32          r_probe_dest0[TRACEROUTE_ENTRIES];
+    E4_uint32          r_probe_dest1[TRACEROUTE_ENTRIES];
+    E4_uint64          r_probe_result;
+    E4_uint64          r_probe_level;
+
+    E4_uint64           r_dma_flowcnt;                         /*  count of dma's queued */
+} EP4_RAIL_MAIN;
+
+#define EP4_PROBE_ACTIVE       (0xffff)
+#define EP4_PROBE_FAILED       (0xfffe)
+
+#if !defined(__elan4__)
+
+typedef struct ep4_retry_ops
+{
+    struct list_head   op_link;
+    unsigned long     (*op_func)(EP4_RAIL *rail, void *arg, unsigned long nextRunTime);
+    void              *op_arg;
+} EP4_RETRY_OPS;
+
+typedef struct ep4_neterr_ops
+{
+    struct list_head   op_link;
+    void             (*op_func) (EP4_RAIL *rail, void *arg, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+    void              *op_arg;
+} EP4_NETERR_OPS;
+
+struct ep4_rail
+{
+    EP_RAIL            r_generic;
+    ELAN4_CTXT         r_ctxt;
+    ELAN4_ROUTE_TABLE  *r_routetable;
+    
+    spinlock_t         r_intcookie_lock;
+    struct list_head    r_intcookie_hash[EP4_INTCOOKIE_HASH_SIZE];
+
+    sdramaddr_t                r_elan;
+    EP_ADDR            r_elan_addr;
+    EP4_RAIL_MAIN      *r_main;
+    EP_ADDR            r_main_addr;
+    
+    EP_CODE            r_threadcode;                           /* copy of thread code */
+
+    sdramaddr_t                r_queuedescs;                           /* systemq queue descriptors */
+
+    E4_uint64         *r_cookies;                              /* network error cookies */
+    spinlock_t          r_cookie_lock;                         /*    and spin lock */
+
+    kcondvar_t         r_probe_wait;                           /* network position probing */
+    spinlock_t         r_probe_lock;
+    volatile int       r_probe_done;
+    EP4_INTCOOKIE      r_probe_intcookie;
+    EP4_ECQ           *r_probe_cq;
+    E4_uint32          r_probe_source0[TRACEROUTE_ENTRIES];
+    E4_uint32          r_probe_source1[TRACEROUTE_ENTRIES];
+
+    kmutex_t           r_haltop_mutex;                         /* halt/flush operations */
+    ELAN4_HALTOP       r_haltop;
+    ELAN4_DMA_FLUSHOP   r_flushop;
+    kcondvar_t         r_haltop_sleep;
+    spinlock_t         r_haltop_lock;
+
+    struct list_head    r_ecq_list[EP4_NUM_ECQ];               /* list of statically allocated command queues */
+    EP_RMAP           *r_ecq_rmap;                             /* resource map for command queue mappings */
+    spinlock_t          r_ecq_lock;                            /* spinlock for list/space management */
+
+    kmutex_t           r_flush_mutex;                          /* serialize command queue flushing */
+    unsigned long      r_flush_count;                          /* # setevents issued for flushing */
+    EP4_ECQ           *r_flush_mcq;                            /*   and command queue for waitevent */
+    EP4_ECQ            *r_flush_ecq;                           /*   and command queue for interrupt */
+    EP4_INTCOOKIE       r_flush_intcookie;                     /*   and interrupt cookie */
+    kcondvar_t          r_flush_sleep;                         /*   and place to sleep ... */
+
+    EP_KTHREAD         r_retry_thread;                         /* retry thread */
+    struct list_head    r_retry_ops;                           /*  list of retry operations */
+
+    EP4_RETRY_OPS       r_dma_ops;                             /* dma retry operations */
+    EP4_ECQ           *r_dma_ecq;                              /*   command queue to reissue DMAs */
+    E4_uint64           r_dma_flowcnt;                         /*   count of dma's reissued */
+    struct list_head    r_dma_retrylist[EP_NUM_RETRIES];       /*   retry lists  */
+    struct list_head    r_dma_freelist;                                /*   and free list */
+    spinlock_t          r_dma_lock;                            /*   and spinlock to protect lists */
+    unsigned long       r_dma_allocated;                       /*   # retries allocated*/
+    unsigned long       r_dma_reserved;                                /*   # retries reserved */
+
+    EP4_ECQ           *r_event_ecq;                            /* command queue for occasional setevents */
+
+    struct list_head    r_neterr_ops;                          /* list of neterr fixup operations */
+
+    ELAN4_IPROC_TRAP    r_iproc_trap;
+    ELAN4_TPROC_TRAP    r_tproc_trap;
+} ;
+
+#define EP4_CTXT_TO_RAIL(ctxt) ((EP4_RAIL *) (((unsigned long) (ctxt)) - offsetof (EP4_RAIL, r_ctxt)))
+
+#if defined(DEBUG_ASSERT)
+#define EP4_ASSERT(rail,EXPR)                  EP_ASSERT(&((rail)->r_generic), EXPR)
+#define EP4_SDRAM_ASSERT(rail,off,value)       EP4_ASSERT(rail, (sdram_assert ? elan4_sdram_readq ((rail)->r_ctxt.ctxt_dev, (off)) == (value) : 1))
+#else
+#define EP4_ASSERT(rail,EXPR)
+#define EP4_SDRAM_ASSERT(rail,off,value)
+#endif
+
+/* kcomm_elan4.c */
+extern EP_RAIL    *ep4_create_rail (EP_SYS *sys, ELAN4_DEV *dev);
+extern void        ep4_destroy_rail (EP_RAIL *rail);
+
+extern int         ep4_start_rail (EP_RAIL *rail);
+extern void        ep4_stall_rail (EP_RAIL *rail);
+extern void        ep4_stop_rail (EP_RAIL *rail);
+
+extern void       ep4_debug_rail (EP_RAIL *rail);
+
+extern void        ep4_position_found (EP_RAIL *rail, ELAN_POSITION *pos);
+
+extern sdramaddr_t ep4_sdram_alloc (EP_RAIL *rail, EP_ADDR addr, unsigned int size);
+extern void        ep4_sdram_free (EP_RAIL *rail, sdramaddr_t addr, unsigned int size);
+extern void        ep4_sdram_writeb (EP_RAIL *rail, sdramaddr_t addr, unsigned char val);
+
+extern void        ep4_flush_tlb (EP_RAIL *r);
+extern void        ep4_load_system_route (EP_RAIL *r, unsigned int vp, unsigned int lowNode, unsigned int highNode);
+extern void        ep4_load_node_route (EP_RAIL *r, unsigned int nodeId);
+extern void        ep4_unload_node_route (EP_RAIL *r, unsigned int nodeId);
+extern void        ep4_lower_filter (EP_RAIL *r, unsigned int nodeId);
+extern void        ep4_raise_filter (EP_RAIL *rail, unsigned int nodeId);
+extern void        ep4_node_disconnected (EP_RAIL *r, unsigned int nodeId);
+
+/* kmap_elan4.c */
+extern void        ep4_kaddr_map (EP_RAIL *r, EP_ADDR eaddr, virtaddr_t kaddr, unsigned int len, unsigned int perm, int ep_attr);
+extern void        ep4_sdram_map (EP_RAIL *r, EP_ADDR eaddr, sdramaddr_t saddr, unsigned int len, unsigned int perm, int ep_attr);
+extern void        ep4_ioaddr_map (EP_RAIL *r, EP_ADDR eaddr, ioaddr_t ioaddr, unsigned int len, unsigned int perm);
+extern void        ep4_unmap (EP_RAIL *r, EP_ADDR eaddr, unsigned int len);
+extern void       *ep4_dvma_reserve (EP_RAIL *r, EP_ADDR eaddr, unsigned int npages);
+extern void        ep4_dvma_release (EP_RAIL *r, EP_ADDR eaddr, unsigned int npages, void *private);
+extern void        ep4_dvma_set_pte (EP_RAIL *r, void *private, unsigned int index, physaddr_t paddr, unsigned int perm);
+extern physaddr_t  ep4_dvma_read_pte (EP_RAIL *r, void *private, unsigned int index);
+extern void        ep4_dvma_unload (EP_RAIL *r, void *private, unsigned int index, unsigned int npages);
+
+/* kmsg_elan4.c */
+extern EP_INPUTQ  *ep4_alloc_inputq (EP_RAIL *r, unsigned int qnum, unsigned int slotSize, unsigned int slotCount,
+                                    EP_INPUTQ_CALLBACK *callback, void *arg);
+extern void        ep4_free_inputq (EP_RAIL *r, EP_INPUTQ *q);
+extern void        ep4_enable_inputq (EP_RAIL *r, EP_INPUTQ *q);
+extern void        ep4_disable_inputq (EP_RAIL *r, EP_INPUTQ *q);
+extern int         ep4_poll_inputq (EP_RAIL *r, EP_INPUTQ *q, int maxCount, EP_INPUTQ_HANDLER *handler, void *arg);
+extern EP_OUTPUTQ *ep4_alloc_outputq (EP_RAIL *r, unsigned int slotSize, unsigned int slotCount);
+extern void        ep4_free_outputq (EP_RAIL *r, EP_OUTPUTQ *q);
+extern void       *ep4_outputq_msg (EP_RAIL *r, EP_OUTPUTQ *q, unsigned int slotNum);
+extern int         ep4_outputq_state (EP_RAIL *r, EP_OUTPUTQ *q, unsigned int slotNum);
+extern int         ep4_outputq_send (EP_RAIL *r, EP_OUTPUTQ *q, unsigned int slotNum, unsigned int size,
+                                    unsigned int nodeId, unsigned int qnum, unsigned int retries);
+
+/* probenetwork_elan4.c */
+extern int         ep4_probe_init (EP4_RAIL *r);
+extern void        ep4_probe_destroy (EP4_RAIL *r);
+extern void        ep4_probe_position_found (EP4_RAIL *rail, ELAN_POSITION *pos);
+extern int         ep4_probe_route (EP_RAIL *r, int level, int sw, int nodeid, int *linkup, int *linkdown, int attempts, EP_SWITCH *lsw);
+extern int         ep4_check_position (EP_RAIL *rail);
+
+/* support_elan4.c */
+extern ELAN4_TRAP_OPS ep4_trap_ops;
+extern void           ep4_register_intcookie (EP4_RAIL *rail, EP4_INTCOOKIE *cp, E4_uint64 cookie, void (*callback)(EP4_RAIL *r, void *arg), void *arg);
+extern void           ep4_deregister_intcookie (EP4_RAIL *rail, EP4_INTCOOKIE *cp);
+extern EP4_INTCOOKIE *ep4_lookup_intcookie (EP4_RAIL *rail, E4_uint64 cookie);
+extern E4_uint64      ep4_neterr_cookie (EP4_RAIL *rail, unsigned int node);
+
+extern void           ep4_flush_filters (EP_RAIL *r);
+extern void           ep4_flush_queues (EP_RAIL *r);
+extern void          ep4_write_qdesc (EP4_RAIL *rail, sdramaddr_t qaddr, E4_InputQueue *qdesc);
+
+extern EP4_ECQ       *ep4_alloc_ecq (EP4_RAIL *rail, unsigned int cqsize);
+extern void           ep4_free_ecq (EP4_RAIL *rail, EP4_ECQ *ecq);
+extern EP4_ECQ      *ep4_get_ecq (EP4_RAIL *rail, unsigned int which, unsigned int ndwords);
+extern void           ep4_put_ecq (EP4_RAIL *rail, EP4_ECQ *ecq, unsigned int ndwords);
+
+extern void           ep4_nop_cmd (EP4_ECQ *ecq, E4_uint64 tag);
+extern void           ep4_set_event_cmd (EP4_ECQ *ecq, E4_Addr event);
+extern void           ep4_wait_event_cmd (EP4_ECQ *ecq, E4_Addr event, E4_uint64 candt, E4_uint64 param0, E4_uint64 param1);
+
+extern void           ep4_flush_interrupt (EP4_RAIL *rail, void *arg);
+extern void           ep4_flush_ecqs (EP4_RAIL *rail);
+
+extern void           ep4_init_thread (EP4_RAIL *rail, E4_ThreadRegs *regs, sdramaddr_t stackTop, 
+                                      EP_ADDR stackAddr, E4_Addr startpc, int nargs,...);
+
+extern void           ep4_initialise_dma_retries (EP4_RAIL *rail);
+extern void           ep4_finalise_dma_retries (EP4_RAIL *rail);
+extern int            ep4_reserve_dma_retries (EP4_RAIL *rail, unsigned int count, unsigned int attr);
+extern void          ep4_release_dma_retries(EP4_RAIL *rail, unsigned int count);
+extern void           ep4_queue_dma_retry (EP4_RAIL *rail, E4_DMA *dma, int interval);
+extern void           ep4_queue_dma_stalled (EP4_RAIL *rail, E4_DMA *dma);
+extern void           ep4_free_stalled_dmas (EP4_RAIL *rail, unsigned int nodeId);
+extern void           ep4_display_rail (EP4_RAIL *rail);
+
+extern void           ep4_add_retry_ops (EP4_RAIL *rail, EP4_RETRY_OPS *ops);
+extern void           ep4_remove_retry_ops (EP4_RAIL *rail, EP4_RETRY_OPS *ops);
+extern void           ep4_retry_thread (EP4_RAIL *rail);
+
+/* neterr_elan4.c */
+extern void           ep4_add_neterr_ops (EP4_RAIL *rail, EP4_NETERR_OPS *ops);
+extern void           ep4_remove_neterr_ops (EP4_RAIL *rail, EP4_NETERR_OPS *ops);
+extern void           ep4_neterr_fixup (EP_RAIL *r, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+
+/* commands_elan4.c */
+extern void           elan4_nop_cmd (ELAN4_CQ *cq, E4_uint64 tag);
+extern void           elan4_write_dword_cmd (ELAN4_CQ *cq, E4_Addr addr, E4_uint64 data);
+extern void           elan4_add_dword_cmd (ELAN4_CQ *cq, E4_Addr addr, E4_uint64 data);
+extern void           elan4_copy64_cmd (ELAN4_CQ *cq, E4_Addr from, E4_Addr to, E4_uint32 datatype);
+extern void           elan4_interrupt_cmd (ELAN4_CQ *cq, E4_uint64 cookie);
+extern void           elan4_run_thread_cmd (ELAN4_CQ *cq, E4_ThreadRegs *regs);
+extern void           elan4_run_dma_cmd (ELAN4_CQ *cq, E4_DMA *dma);
+extern void           elan4_set_event_cmd (ELAN4_CQ *cq, E4_Addr event);
+extern void           elan4_set_eventn_cmd (ELAN4_CQ *cq, E4_Addr event, E4_uint32 count);
+extern void           elan4_wait_event_cmd (ELAN4_CQ *cq, E4_Addr event, E4_uint64 candt, E4_uint64 param0, E4_uint64 param1);
+extern void           elan4_open_packet (ELAN4_CQ *cq, E4_uint64 command);
+extern void           elan4_guard (ELAN4_CQ *cq, E4_uint64 command);
+extern void           elan4_sendtrans0 (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr);
+extern void           elan4_sendtrans1 (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, E4_uint64 p0);
+extern void           elan4_sendtrans2 (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, E4_uint64 p0, E4_uint64 p1);
+extern void           elan4_sendtransn (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, ...);
+extern void           elan4_sendtransp (ELAN4_CQ *cq, E4_uint16 trtype, E4_uint64 addr, E4_uint64 *ptr);
+
+extern void           ep4_add_retry_ops (EP4_RAIL *rail, EP4_RETRY_OPS *ops);
+extern void          ep4_remove_retry_ops (EP4_RAIL *rail, EP4_RETRY_OPS *ops);
+extern void           ep4_retry_thread (EP4_RAIL *rail);
+
+extern void           ep4_fillout_stats(EP_RAIL *rail, char *str);
+
+#endif /* ! defined(__elan4__) */
+
+#endif /* __EP_KCOMM_ELAN4_H */
Index: linux-2.4.21/drivers/net/qsnet/ep/kcomm_vp.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kcomm_vp.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kcomm_vp.h        2005-06-01 23:12:54.668428616 -0400
@@ -0,0 +1,36 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __EP_KCOMM_VP_H
+#define __EP_KCOMM_VP_H
+
+#ident "@(#)$Id: kcomm_vp.h,v 1.2 2004/03/24 11:32:56 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm_vp.h,v $*/
+
+#define EP_MAX_NODES                   2048                    /* Max nodes we support */
+
+/* virtual process allocation */
+#define EP_VP_NODE_BASE                        (0)
+#define EP_VP_DATA_BASE                        (EP_VP_NODE_BASE + EP_MAX_NODES)
+#define EP_VP_PROBE_BASE               (EP_VP_DATA_BASE + EP_MAX_NODES)
+#define EP_VP_PROBE_COUNT              ELAN_MAX_LEVELS
+
+#define EP_VP_BCAST_BASE               (EP_VP_PROBE_BASE + EP_VP_PROBE_COUNT)
+#define EP_VP_BCAST_COUNT              (CM_SGMTS_PER_LEVEL * (CM_MAX_LEVELS - 1) + 1)
+
+#define EP_VP_NODE(nodeId)             (EP_VP_NODE_BASE + (nodeId))
+#define EP_VP_DATA(nodeId)             (EP_VP_DATA_BASE + (nodeId))
+#define EP_VP_PROBE(lvl)               (EP_VP_PROBE_BASE + (lvl))
+#define EP_VP_BCAST(lvl,sgmt)          (EP_VP_BCAST_BASE + ((lvl) - 1)*CM_SGMTS_PER_LEVEL + (sgmt))
+
+#define EP_VP_TO_NODE(vp)              ((vp) & (EP_MAX_NODES-1))
+#define EP_VP_ISDATA(vp)               ((vp) >= EP_VP_DATA_BASE && (vp) < (EP_VP_DATA_BASE + EP_MAX_NODES))
+
+#endif /* __EP_KCOMM_VP_H */
+
+
Index: linux-2.4.21/drivers/net/qsnet/ep/kmap.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kmap.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kmap.c    2005-06-01 23:12:54.669428464 -0400
@@ -0,0 +1,561 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kmap.c,v 1.10.6.2 2004/12/14 10:19:14 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kmap.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kpte.h>
+
+#include <elan/kcomm.h>
+
+#include "debug.h"
+
+#if defined(DIGITAL_UNIX)
+#  define kernel_map                   (first_task->map)
+#  define vaddr_to_phys(map, addr)     (pmap_extract (vm_map_pmap ((vm_map_t) map), (unsigned long) addr))
+#elif defined(LINUX)
+#  define kernel_map                   get_kern_mm()
+#  define vaddr_to_phys(map, addr)     (kmem_to_phys(addr))
+#elif defined(SOLARIS)
+#  define kernel_map                   &kas
+#  define vaddr_to_phys(map,addr)      ptob(hat_getpfnum (((struct as *) map)->a_hat, (caddr_t) addr))
+#endif
+
+void
+ep_perrail_kaddr_map (EP_RAIL *rail, EP_ADDR eaddr, virtaddr_t kaddr, unsigned long len,  unsigned int perm, int ep_attr)
+{
+    rail->Operations.KaddrMap (rail, eaddr, kaddr, len, perm, ep_attr);
+}
+
+void
+ep_perrail_sdram_map (EP_RAIL *rail, EP_ADDR eaddr, sdramaddr_t saddr, unsigned long len, unsigned int perm, int ep_attr)
+{
+    rail->Operations.SdramMap (rail, eaddr, saddr, len, perm, ep_attr);
+}
+
+void
+ep_perrail_unmap (EP_RAIL *rail, EP_ADDR eaddr, unsigned long len)
+{
+    rail->Operations.Unmap (rail, eaddr, len);
+}
+
+void
+ep_perrail_dvma_sync (EP_RAIL *rail)
+{
+    if (rail->TlbFlushRequired)
+    {
+       rail->TlbFlushRequired = 0;
+
+       rail->Operations.FlushTlb (rail);
+    }
+}
+
+
+static int ep_dvma_map_rails (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd, EP_RAILMASK mask);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+static uint16_t ep_dvma_calc_check_sum (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd, uint16_t check_sum);
+#endif
+
+EP_NMH_OPS ep_dvma_nmh_ops = 
+{
+    ep_dvma_map_rails,
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    ep_dvma_calc_check_sum
+#endif
+};
+
+extern void
+ep_dvma_init (EP_SYS *sys)
+{
+    EP_DVMA_STATE *d = &sys->DvmaState;
+
+    kmutex_init (&d->dvma_lock);
+    
+    INIT_LIST_HEAD (&d->dvma_handles);
+    INIT_LIST_HEAD (&d->dvma_rails);
+
+    d->dvma_rmap = ep_rmallocmap (EP_DVMA_RMAP_SIZE, "dvma_rmap", 1);
+
+    ep_rmfree (d->dvma_rmap, EP_DVMA_TOP - EP_DVMA_BASE, EP_DVMA_BASE);
+}
+
+extern void
+ep_dvma_fini (EP_SYS *sys)
+{
+    EP_DVMA_STATE *d = &sys->DvmaState;
+
+    ep_rmfreemap (d->dvma_rmap);
+
+    kmutex_destroy (&d->dvma_lock);
+}
+    
+extern int
+ep_dvma_add_rail (EP_SYS *sys, EP_RAIL *rail)
+{
+    EP_DVMA_STATE    *d = &sys->DvmaState;
+    EP_RAIL_ENTRY    *l;
+    struct list_head *el;
+
+    KMEM_ZALLOC (l, EP_RAIL_ENTRY *, sizeof (EP_RAIL_ENTRY), 1);
+
+    if (l == NULL)
+       return (ENOMEM);
+
+    kmutex_lock (&d->dvma_lock);
+
+    l->Rail = rail;
+
+    list_add_tail (&l->Link, &d->dvma_rails);
+
+    list_for_each (el, &d->dvma_handles) {
+       EP_DVMA_NMH *desc   = list_entry (el, EP_DVMA_NMH, dvma_link);
+       int          npages = desc->dvma_nmh.nmh_nmd.nmd_len >> PAGESHIFT;
+
+       desc->dvma_rails[rail->Number] = rail;
+       desc->dvma_railmask |= ( 1 << rail->Number);
+
+       desc->dvma_private[rail->Number] = rail->Operations.DvmaReserve (rail, desc->dvma_nmh.nmh_nmd.nmd_addr, npages);
+    }
+
+    kmutex_unlock (&d->dvma_lock);
+    return (0);
+}
+
+extern void
+ep_dvma_remove_rail (EP_SYS *sys, EP_RAIL *rail)
+{
+    EP_DVMA_STATE    *d = &sys->DvmaState;
+    struct list_head *el;
+
+    kmutex_lock (&d->dvma_lock);
+    
+    list_for_each (el, &d->dvma_handles) {
+       EP_DVMA_NMH *desc   = list_entry (el, EP_DVMA_NMH, dvma_link);
+       int          npages = desc->dvma_nmh.nmh_nmd.nmd_len >> PAGESHIFT;
+
+       desc->dvma_rails[rail->Number] = NULL;
+       desc->dvma_railmask &= ~(1 << rail->Number);
+
+       rail->Operations.DvmaRelease (rail, desc->dvma_nmh.nmh_nmd.nmd_addr, npages, desc->dvma_private[rail->Number]);
+    }
+
+    list_for_each (el, &d->dvma_rails) {
+       EP_RAIL_ENTRY *tmp = list_entry (el, EP_RAIL_ENTRY, Link);
+
+       if (tmp->Rail == rail)
+       {
+           list_del (el);
+
+           KMEM_FREE (tmp, sizeof (EP_RAIL_ENTRY));
+           break;
+       }
+    }
+    kmutex_unlock (&d->dvma_lock);
+}
+
+EP_NMH *
+ep_dvma_reserve (EP_SYS *sys, unsigned npages, unsigned perm)
+{
+    EP_DVMA_STATE    *d = &sys->DvmaState;
+    EP_DVMA_NMH      *desc;
+    EP_ADDR          addr;
+    struct list_head *el;
+    int               i;
+
+    KMEM_ZALLOC (desc, EP_DVMA_NMH *, offsetof (EP_DVMA_NMH, dvma_attrs[npages]), 1);
+    
+    if (desc == NULL)
+       return (NULL);
+
+    if ((addr = ep_rmalloc (d->dvma_rmap, npages << PAGESHIFT, 0)) == 0)
+    {
+
+       KMEM_FREE (desc, sizeof (EP_DVMA_NMH));
+       return (NULL);
+    }
+
+    spin_lock_init (&desc->dvma_lock);
+
+    desc->dvma_perm = perm;
+
+    kmutex_lock (&d->dvma_lock);
+    /* reserve the mapping resource */
+    list_for_each (el, &d->dvma_rails) {
+       EP_RAIL *rail = list_entry (el, EP_RAIL_ENTRY, Link)->Rail;
+
+       EPRINTF4 (DBG_KMAP, "%s: ep_dvma_reserve desc=%p npages=%d rail=%p\n", rail->Name, desc, npages, rail);
+
+       if ((desc->dvma_private[rail->Number] = rail->Operations.DvmaReserve (rail, addr, npages)) == NULL)
+       {
+           printk ("%s: !!ep_dvma_reserve - rail->DvmaReserve failed\n", rail->Name);
+           goto failed;
+       }
+
+       desc->dvma_rails[rail->Number] = rail;
+       desc->dvma_railmask |= (1 << rail->Number);
+    }
+
+    /* insert into the network mapping handle table */
+    desc->dvma_nmh.nmh_nmd.nmd_addr = addr;
+    desc->dvma_nmh.nmh_nmd.nmd_len  = npages << PAGESHIFT;
+    desc->dvma_nmh.nmh_nmd.nmd_attr = EP_NMD_ATTR (sys->Position.pos_nodeid, 0);
+    desc->dvma_nmh.nmh_ops          = &ep_dvma_nmh_ops;
+
+    ep_nmh_insert (&sys->MappingTable, &desc->dvma_nmh);
+
+    list_add (&desc->dvma_link, &d->dvma_handles);
+
+    kmutex_unlock (&d->dvma_lock);
+
+    return (&desc->dvma_nmh);
+
+ failed:
+
+    kmutex_unlock (&d->dvma_lock);
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+       if (desc->dvma_rails[i] != NULL)
+           desc->dvma_rails[i]->Operations.DvmaRelease (desc->dvma_rails[i], addr, npages, desc->dvma_private[i]);
+
+    ep_rmfree (d->dvma_rmap, npages << PAGESHIFT, addr);
+
+    KMEM_FREE (desc, sizeof (EP_DVMA_NMH));
+    return (NULL);
+}
+
+void
+ep_dvma_release (EP_SYS *sys, EP_NMH *nmh)
+{
+    EP_DVMA_STATE *d      = &sys->DvmaState;
+    EP_DVMA_NMH   *desc   = (EP_DVMA_NMH *) nmh;
+    EP_ADDR        addr   = nmh->nmh_nmd.nmd_addr;
+    int            npages = nmh->nmh_nmd.nmd_len >> PAGESHIFT;
+    EP_RAIL       *rail;
+    int            i;
+
+    kmutex_lock (&d->dvma_lock);
+
+    list_del (&desc->dvma_link);
+    
+    ep_nmh_remove (&sys->MappingTable, nmh);
+
+    for (i = 0; i < EP_MAX_RAILS; i++)
+       if ((rail = desc->dvma_rails[i]) != NULL)
+           rail->Operations.DvmaRelease (rail, addr, npages, desc->dvma_private[i]);
+
+    ep_rmfree (d->dvma_rmap, npages << PAGESHIFT, addr);
+
+    KMEM_FREE (desc, offsetof (EP_DVMA_NMH, dvma_attrs[npages]));
+
+    kmutex_unlock (&d->dvma_lock);
+}
+
+void
+ep_dvma_load (EP_SYS *sys, void *map, caddr_t vaddr, unsigned len, EP_NMH *nmh, unsigned index, EP_RAILMASK *hints, EP_NMD *subset)
+{
+    EP_DVMA_NMH *desc = (EP_DVMA_NMH *) nmh;
+    unsigned     offset = (unsigned long) vaddr & PAGEOFFSET;
+    unsigned     npages = btopr (len + offset);
+    EP_ADDR      addr   = nmh->nmh_nmd.nmd_addr + (index << PAGESHIFT);
+    int                 rmask  = *hints;
+    EP_RAIL     *rail;
+    register int i, rnum;
+    unsigned long flags;
+
+    EPRINTF7 (DBG_KMAP, "ep_dvma_load: map=%p vaddr=%p len=%x nmh=%p(%x,%x) index=%d\n",
+             map, vaddr, len, nmh, nmh->nmh_nmd.nmd_addr, nmh->nmh_nmd.nmd_len, index);
+
+    /* If no rail specified, then map into all rails */
+    if (rmask == 0)
+       rmask = desc->dvma_railmask;
+
+    ASSERT ((index + npages) <= (nmh->nmh_nmd.nmd_len >> PAGESHIFT));
+
+    /* If not map specified then use the kernel map */
+    if (map == NULL)
+       map = kernel_map;
+
+    spin_lock_irqsave (&desc->dvma_lock, flags);
+    /* Now map each of the specified pages (backwards) */
+
+    vaddr = (vaddr - offset) + (npages-1)*PAGESIZE;
+    for (i = npages-1; i >= 0; i--, vaddr -= PAGESIZE)
+    {
+       physaddr_t paddr = vaddr_to_phys (map, vaddr);
+       
+       for (rnum = 0; rnum < EP_MAX_RAILS; rnum++)
+       {
+           if (! (rmask & (1 << rnum)) || (rail = desc->dvma_rails[rnum]) == NULL)
+               rmask &= ~(1 << rnum);
+           else
+           {
+               rail->Operations.DvmaSetPte (rail, desc->dvma_private[rnum], index + i, paddr, desc->dvma_perm);
+
+               desc->dvma_attrs[index + i] |= (1 << rnum);
+           }
+       }
+    }
+
+    for (rnum = 0; rnum < EP_MAX_RAILS; rnum++)
+       if ((rmask & (1 << rnum)) && (rail = desc->dvma_rails[rnum]) != NULL)
+           rail->TlbFlushRequired = 1;
+
+    spin_unlock_irqrestore (&desc->dvma_lock, flags);
+
+    /* Construct the network mapping handle to be returned. */
+    subset->nmd_addr = addr + offset;
+    subset->nmd_len  = len;
+    subset->nmd_attr = EP_NMD_ATTR(sys->Position.pos_nodeid, rmask);
+}
+
+void
+ep_dvma_unload (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd)
+{
+    EP_DVMA_NMH *desc = (EP_DVMA_NMH *) nmh;
+    unsigned     offset = nmd->nmd_addr & PAGEOFFSET;
+    unsigned     npages = btopr (nmd->nmd_len + offset);
+    unsigned     index  = (nmd->nmd_addr - nmh->nmh_nmd.nmd_addr) >> PAGESHIFT;
+    EP_RAIL     *rail;
+    int          rnum;
+    int          rmask;
+    register int i;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&desc->dvma_lock, flags);
+
+    /* compute which rails we need to unload on */
+    for (rmask = 0, i = 0; i < npages; i++)
+    {
+       rmask |= desc->dvma_attrs[index + i];
+       
+       desc->dvma_attrs[index + i] = 0;
+    }
+    
+    for (rnum = 0; rnum < EP_MAX_RAILS; rnum++)
+       if ((rmask & (1 << rnum)) && (rail = desc->dvma_rails[rnum]) != NULL)
+           rail->Operations.DvmaUnload (rail, desc->dvma_private[rnum], index, npages);
+
+    spin_unlock_irqrestore (&desc->dvma_lock, flags);
+}
+
+int
+ep_dvma_map_rails (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd, EP_RAILMASK mask)
+{
+    EP_DVMA_NMH *desc = (EP_DVMA_NMH *) nmh;
+    unsigned     offset = nmd->nmd_addr & PAGEOFFSET;
+    unsigned     npages = btopr (nmd->nmd_len + offset);
+    unsigned     index  = (nmd->nmd_addr - nmh->nmh_nmd.nmd_addr) >> PAGESHIFT;
+    int          r, rnum;
+    register int i;
+    unsigned long flags;
+
+    spin_lock_irqsave (&desc->dvma_lock, flags);
+
+    EPRINTF4 (DBG_KMAP, "ep_dvma_map_rails: nmd=%08x.%08x.%08x mask=%04x\n", nmd->nmd_addr, nmd->nmd_len, nmd->nmd_attr, mask);
+
+    if ((mask &= desc->dvma_railmask) == 0)
+    {
+       printk ("ep_dvma_map_rails: no intersecting rails %04x.%04x\n", mask, desc->dvma_railmask);
+       spin_unlock_irqrestore (&desc->dvma_lock, flags);
+       return (-1);
+    }
+
+    for (i = npages-1; i >= 0; i--)
+    {
+       int pgidx = (index + i);
+
+       for (rnum = 0; rnum < EP_MAX_RAILS; rnum++)
+           if (desc->dvma_attrs[pgidx] & (1 << rnum))
+               break;
+       
+       if (rnum == EP_MAX_RAILS)
+       {
+           EPRINTF3 (DBG_KMAP, "ep_dvma_map_rails: nmh=%p idx=%x [%08x] not ptes valid\n", nmh, pgidx, 
+                     nmh->nmh_nmd.nmd_addr + ((pgidx) << PAGESHIFT));
+           mask = 0;
+       }
+       else
+       {
+           EP_RAIL   *rail  = desc->dvma_rails[rnum];
+           physaddr_t paddr = rail->Operations.DvmaReadPte (rail, desc->dvma_private[rnum], pgidx);
+           
+           EPRINTF5 (DBG_KMAP, "%s: ep_dvma_map_rails: nmh=%p idx=%x [%08x] paddr %llx\n", rail->Name, nmh, pgidx,
+                     nmh->nmh_nmd.nmd_addr + (pgidx << PAGESHIFT), (long long) paddr);
+           
+           for (r = 0; r < EP_MAX_RAILS; r++)
+           {
+               if ((mask & (1 << r)) == 0)
+                   continue;
+               
+               if ((desc->dvma_attrs[pgidx] & (1 << r)) == 0)
+               {
+                   EPRINTF5 (DBG_KMAP, "%s: ep_dvma_map_rails: nmh=%p idx=%x [%08x] paddr=%llx\n",
+                             desc->dvma_rails[rnum]->Name, nmh, pgidx, nmh->nmh_nmd.nmd_addr + (pgidx << PAGESHIFT), 
+                             (long long) paddr);
+                   
+                   rail->Operations.DvmaSetPte (rail, desc->dvma_private[rnum], pgidx, paddr, desc->dvma_perm);
+                   
+                   desc->dvma_attrs[pgidx] |= (1 << r);
+               }
+           }
+       }
+    }
+
+    for (rnum = 0; rnum < EP_MAX_RAILS; rnum++)
+       if ((mask & (1 << rnum)) != 0)
+           desc->dvma_rails[rnum]->TlbFlushRequired = 1;
+
+    EPRINTF4 (DBG_KMAP, "ep_dvma_map_rails: nmd=%08x.%08x.%08x|%04x\n", nmd->nmd_addr, nmd->nmd_len, nmd->nmd_attr, mask);
+
+    /* Finally update the network memory descriptor */
+    nmd->nmd_attr |= mask;
+
+    spin_unlock_irqrestore (&desc->dvma_lock, flags);
+
+    return (0);
+}
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+#include <linux/highmem.h>
+
+/* Generic rolling checksum algorithm */
+uint16_t
+rolling_check_sum (char *msg, int nob, uint16_t sum)
+{
+    while (nob-- > 0)
+       sum = sum * 13 + *msg++;
+
+    return (sum);
+}
+
+#if ! defined(NO_RMAP)
+void  
+unmap_phys_address(unsigned long phys_addr)
+{
+    unsigned long pfn = (phys_addr >> PAGE_SHIFT);
+    
+    if (pfn_valid(pfn)) 
+       kunmap(pfn_to_page(pfn));
+}
+
+void * 
+map_phys_address(unsigned long phys_addr)
+{
+    unsigned long pfn = (phys_addr >> PAGE_SHIFT);
+    
+    if (pfn_valid(pfn)) 
+       return  kmap(pfn_to_page(pfn));
+    
+    return NULL;
+}
+#else
+void  
+unmap_phys_address(unsigned long phys_addr)
+{
+    struct page *p = virt_to_page(__va(phys_addr));
+    
+    if (VALID_PAGE(p)) 
+       kunmap(p);
+}
+
+void * 
+map_phys_address(unsigned long phys_addr)
+{
+    struct page *p = virt_to_page(__va(phys_addr));
+                               
+    if (VALID_PAGE(p)) 
+       return  kmap(p);
+    
+    return NULL;
+}
+#endif
+
+uint16_t
+ep_dvma_calc_check_sum (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd, uint16_t check_sum)
+{
+    /* cant be called from an interupt */
+
+    EP_DVMA_NMH *desc = (EP_DVMA_NMH *) nmh;
+    unsigned     offset = nmd->nmd_addr & PAGEOFFSET;
+    unsigned     npages = btopr (nmd->nmd_len + offset);
+    unsigned     index  = (nmd->nmd_addr - nmh->nmh_nmd.nmd_addr) >> PAGESHIFT;
+    unsigned     start, len;
+    int          rnum;
+    register int i;
+    unsigned long flags;
+    EP_RAIL      *rail;
+
+
+    spin_lock_irqsave (&desc->dvma_lock, flags);
+
+    EPRINTF3 (DBG_KMAP, "ep_dvma_calc_check_sum: nmd=%08x.%08x.%08x \n", nmd->nmd_addr, nmd->nmd_len, nmd->nmd_attr);
+ 
+    /* find a rail */
+    for (rnum = 0; rnum < EP_MAX_RAILS; rnum++)
+       if (desc->dvma_attrs[index] & (1 << rnum))
+           break;
+       
+    ASSERT (rnum != EP_MAX_RAILS);
+ 
+    rail = desc->dvma_rails[rnum];
+
+    for (i = 0; i <= (npages-1); i++)
+    {
+       int        pgidx = (index + i);
+       physaddr_t paddr = rail->Operations.DvmaReadPte (rail, desc->dvma_private[rnum], pgidx);
+       void *     virt;
+
+       spin_unlock_irqrestore (&desc->dvma_lock, flags); /* unlock for check sum calc */
+
+       virt = map_phys_address(paddr);
+
+       if (!virt)
+           printk("ep_dvma_calc_check_sum: virt = NULL ! \n");
+       else {
+           if ( i == 0 ) {
+               /* last bit of the first page */
+               start =  (nmd->nmd_addr & (PAGESIZE - 1)) ;
+               len   =  PAGESIZE - start;
+               if ( len > nmd->nmd_len) /* less than the remaining page */ 
+                   len =  nmd->nmd_len;
+           } else {
+               if ( i != (npages-1)) {
+                   /* all of the middle pages    */
+                   start = 0;
+                   len   = PAGESIZE;
+               } else {
+                   /* first bit of the last page */
+                   start = 0;
+                   len   = ((nmd->nmd_addr + nmd->nmd_len -1) & (PAGESIZE -1)) +1;
+               }
+           }
+
+           check_sum = rolling_check_sum (((char *)virt)+start, len, check_sum);
+           unmap_phys_address(paddr);
+   
+           /* re aquire the lock */
+           spin_lock_irqsave (&desc->dvma_lock, flags);
+       }
+       
+       EPRINTF5 (DBG_KMAP, "%s: ep_dvma_calc_check_sum: nmh=%p idx=%x [%08x] paddr %llx\n", rail->Name, nmh, pgidx,
+                 nmh->nmh_nmd.nmd_addr + (pgidx << PAGESHIFT), (long long) paddr);     
+    }
+
+    EPRINTF4 (DBG_KMAP, "ep_dvma_calc_check_sum: nmd=%08x.%08x.%08x = %d\n", nmd->nmd_addr, nmd->nmd_len, nmd->nmd_attr, check_sum);
+
+    spin_unlock_irqrestore (&desc->dvma_lock, flags);
+
+    return (check_sum);
+}
+#endif
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/kmap_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kmap_elan3.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kmap_elan3.c      2005-06-01 23:12:54.670428312 -0400
@@ -0,0 +1,209 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kmap_elan3.c,v 1.3.8.1 2004/12/14 10:19:14 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kmap_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan3/elanregs.h>
+#include <elan3/elandev.h>
+#include <elan3/elanvp.h>
+#include <elan3/elan3mmu.h>
+#include <elan3/elanctxt.h>
+#include <elan3/elandebug.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_elan3.h"
+
+#if defined(DIGITAL_UNIX)
+#  define kernel_map                   (first_task->map)
+#  define vaddr_to_phys(map, addr)     (pmap_extract (vm_map_pmap ((vm_map_t) map), (unsigned long) addr))
+#elif defined(LINUX)
+#  define kernel_map                   get_kern_mm()
+#  define vaddr_to_phys(map, addr)     (kmem_to_phys(addr))
+#elif defined(SOLARIS)
+#  define kernel_map                   &kas
+#  define vaddr_to_phys(map,addr)      ptob(hat_getpfnum (((struct as *) map)->a_hat, (caddr_t) addr))
+#endif
+
+#define ELAN3_PTES_PER_PAGE            (PAGESIZE/ELAN3_PAGE_SIZE)
+
+#if defined(__LITTLE_ENDIAN__)
+#define PERM_ENDIAN    0
+#else
+#define PERM_ENDIAN    ELAN3_PTE_BIG_ENDIAN
+#endif
+
+static unsigned int main_permtable[] = 
+{
+    ELAN3_PERM_REMOTEALL,              /* EP_PERM_EXECUTE */
+    ELAN3_PERM_REMOTEREAD,             /* EP_PERM_READ */
+    ELAN3_PERM_REMOTEWRITE,            /* EP_PERM_WRITE */
+    ELAN3_PERM_REMOTEWRITE,            /* EP_PERM_ALL */
+};
+
+static unsigned int sdram_permtable[] = 
+{
+    ELAN3_PERM_REMOTEREAD,             /* EP_PERM_EXECUTE */
+    ELAN3_PERM_REMOTEREAD,             /* EP_PERM_READ */
+    ELAN3_PERM_REMOTEWRITE,            /* EP_PERM_WRITE */
+    ELAN3_PERM_REMOTEALL,              /* EP_PERM_ALL */
+};
+
+static unsigned int io_permtable[] = 
+{
+    ELAN3_PERM_LOCAL_READ,             /* EP_PERM_EXECUTE */
+    ELAN3_PERM_REMOTEREAD,             /* EP_PERM_READ */
+    ELAN3_PERM_REMOTEWRITE,            /* EP_PERM_WRITE */
+    ELAN3_PERM_REMOTEWRITE,            /* EP_PERM_ALL */
+};
+
+void
+ep3_kaddr_map (EP_RAIL *r, EP_ADDR eaddr, virtaddr_t kaddr, unsigned len, unsigned int perm, int ep_attr)
+{
+    EP3_RAIL    *rail   = (EP3_RAIL *) r;
+    unsigned     npages = len >> PAGESHIFT;
+    int          i;
+    unsigned int off;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (kaddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    for (i = 0; i < npages; i++)
+    {
+       physaddr_t paddr = vaddr_to_phys (kernel_map, (void *) kaddr);
+
+       for (off = 0; off < PAGESIZE; off += ELAN3_PAGE_SIZE)
+           elan3mmu_pteload (rail->Elan3mmu, PTBL_LEVEL_3, eaddr + off, paddr + off, 
+                             main_permtable[perm], PTE_LOAD_LOCK | PTE_LOAD_NOSYNC | ((ep_attr & EP_NO_SLEEP) ? PTE_NO_SLEEP : 0));
+
+       eaddr += PAGESIZE;
+       kaddr += PAGESIZE;
+    }
+}
+
+void
+ep3_sdram_map (EP_RAIL *r, EP_ADDR eaddr, sdramaddr_t saddr, unsigned len, unsigned int perm, int ep_attr)
+{
+    EP3_RAIL    *rail   = (EP3_RAIL *) r;
+    unsigned     npages = len >> PAGESHIFT;
+    int          i;
+    unsigned int off;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (saddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    for (i = 0; i < npages; i++)
+    {
+       physaddr_t paddr = elan3_sdram_to_phys (rail->Device, saddr);
+
+       for (off = 0; off < PAGESIZE; off += ELAN3_PAGE_SIZE)
+           elan3mmu_pteload (rail->Elan3mmu, PTBL_LEVEL_3, eaddr+off, paddr+off, 
+                             sdram_permtable[perm], PTE_LOAD_LOCK | PTE_LOAD_NOSYNC | ((ep_attr & EP_NO_SLEEP) ? PTE_NO_SLEEP : 0) );
+
+       eaddr += PAGESIZE;
+       saddr += PAGESIZE;
+    }
+}
+
+void
+ep3_ioaddr_map (EP_RAIL *r, EP_ADDR eaddr, ioaddr_t ioaddr, unsigned len, unsigned int perm)
+{
+    EP3_RAIL    *rail   = (EP3_RAIL *) r;
+    unsigned     npages = len >> PAGESHIFT;
+    int          i;
+    unsigned int off;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (ioaddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    for (i = 0; i < npages; i++)
+    {
+       physaddr_t paddr = vaddr_to_phys (kernel_map, (void *) ioaddr);
+
+       for (off = 0; off < PAGESIZE; off += ELAN3_PAGE_SIZE)
+           elan3mmu_pteload (rail->Elan3mmu, PTBL_LEVEL_3, eaddr + off, paddr + off, 
+                             io_permtable[perm], PTE_LOAD_LOCK | PTE_LOAD_NOSYNC);
+
+       eaddr += PAGESIZE;
+       ioaddr += PAGESIZE;
+    }
+}
+void
+ep3_unmap (EP_RAIL *r, EP_ADDR eaddr, unsigned len)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    elan3mmu_unload (rail->Elan3mmu, eaddr, len, PTE_UNLOAD_UNLOCK | PTE_UNLOAD_NOSYNC);
+}
+
+void *
+ep3_dvma_reserve (EP_RAIL *r, EP_ADDR eaddr, unsigned npages)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+    void     *private;
+
+    KMEM_ALLOC (private, void *, npages * ELAN3_PTES_PER_PAGE * sizeof (sdramaddr_t), 1);
+    
+    if (private == NULL)
+       return NULL;
+    
+    elan3mmu_reserve (rail->Elan3mmu, eaddr, npages * ELAN3_PTES_PER_PAGE, (sdramaddr_t *) private);
+
+    return private;
+}
+
+void
+ep3_dvma_release (EP_RAIL *r, EP_ADDR eaddr, unsigned npages, void *private)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+
+    elan3mmu_release (rail->Elan3mmu, eaddr, npages * ELAN3_PTES_PER_PAGE, (sdramaddr_t *) private);
+
+    KMEM_FREE (private, npages * ELAN3_PTES_PER_PAGE * sizeof (sdramaddr_t));
+}
+
+void
+ep3_dvma_set_pte (EP_RAIL *r, void *private, unsigned index, physaddr_t paddr, unsigned int perm)
+{
+    ELAN3_DEV    *dev  = ((EP3_RAIL *) r)->Device;
+    sdramaddr_t *ptep = &((sdramaddr_t *) private)[index * ELAN3_PTES_PER_PAGE];
+    int          off;
+
+    for (off =0 ; off < PAGESIZE; off += ELAN3_PAGE_SIZE)
+    {  
+       ELAN3_PTE newpte = elan3mmu_phys_to_pte (dev, paddr + off, main_permtable[perm]) | ELAN3_PTE_REF | ELAN3_PTE_MOD;
+
+       elan3_writepte (dev, *ptep, newpte);
+
+       ptep++;
+    }
+}
+
+physaddr_t
+ep3_dvma_read_pte (EP_RAIL *r, void *private, unsigned index)
+{
+    EP3_RAIL    *rail = (EP3_RAIL *) r;
+    sdramaddr_t *ptep = &((sdramaddr_t *) private)[index * ELAN3_PTES_PER_PAGE];
+    ELAN3_PTE     pte  = elan3_readpte (rail->Device, *ptep);
+
+    return pte & ELAN3_PTE_PFN_MASK;
+}
+
+void
+ep3_dvma_unload (EP_RAIL *r, void *private, unsigned index, unsigned npages)
+{
+    EP3_RAIL    *rail = (EP3_RAIL *) r;
+    sdramaddr_t *ptep = &((sdramaddr_t *) private)[index * ELAN3_PTES_PER_PAGE];
+    ELAN3_PTE     tpte = elan3mmu_kernel_invalid_pte (rail->Elan3mmu);
+    int i;
+
+    for (i = (npages * ELAN3_PTES_PER_PAGE) - 1; i >= 0; i--)
+       elan3_writepte (rail->Device, ptep[i], tpte);
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/kmap_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kmap_elan4.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kmap_elan4.c      2005-06-01 23:12:54.670428312 -0400
@@ -0,0 +1,226 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kmap_elan4.c,v 1.7.8.2 2004/12/14 10:19:14 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kmap_elan4.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "debug.h"
+#include "kcomm_elan4.h"
+
+#if defined(DIGITAL_UNIX)
+#  define kernel_map                   (first_task->map)
+#  define vaddr_to_phys(map, addr)     (pmap_extract (vm_map_pmap ((vm_map_t) map), (unsigned long) addr))
+#elif defined(LINUX)
+#  define kernel_map                   get_kern_mm()
+#  define vaddr_to_phys(map, addr)     (kmem_to_phys(addr))
+#elif defined(SOLARIS)
+#  define kernel_map                   &kas
+#  define vaddr_to_phys(map,addr)      ptob(hat_getpfnum (((struct as *) map)->a_hat, (caddr_t) addr))
+#endif
+
+static unsigned int main_permtable[] = 
+{
+    PERM_Unused,                       /* EP_PERM_EXECUTE */
+    PERM_RemoteReadOnly,               /* EP_PERM_READ */
+    PERM_DataReadWrite,                        /* EP_PERM_WRITE */
+    PERM_DataReadWrite,                        /* EP_PERM_ALL */
+};
+
+static unsigned int sdram_permtable[] = 
+{
+    PERM_LocExecute,                   /* EP_PERM_EXECUTE */
+    PERM_RemoteReadOnly,               /* EP_PERM_READ */
+    PERM_DataReadWrite,                        /* EP_PERM_WRITE */
+    PERM_RemoteAll,                    /* EP_PERM_ALL */
+};
+
+static unsigned int io_permtable[] = 
+{
+    PERM_Unused,                       /* EP_PERM_EXECUTE */
+    PERM_RemoteReadOnly,               /* EP_PERM_READ */
+    PERM_DataReadWrite,                        /* EP_PERM_WRITE */
+    PERM_Unused,                       /* EP_PERM_ALL */
+};
+
+void
+ep4_kaddr_map (EP_RAIL *r, EP_ADDR eaddr, virtaddr_t kaddr, unsigned int len, unsigned int perm, int ep_attr)
+{
+    EP4_RAIL    *rail   = (EP4_RAIL *) r;
+    ELAN4_DEV   *dev    = rail->r_ctxt.ctxt_dev;
+    unsigned int npages = len >> PAGESHIFT;
+    int          i;
+    unsigned int off;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (kaddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    for (i = 0; i < npages; i++)
+    {
+       physaddr_t paddr = vaddr_to_phys (kernel_map, (void *) kaddr);
+
+       for (off = 0; off < PAGESIZE; off += (1 << dev->dev_pageshift[0]))
+       {
+           E4_uint64 newpte = elan4mmu_phys2pte (dev, paddr + off, main_permtable[perm]);
+
+           elan4mmu_pteload (&rail->r_ctxt, 0, eaddr + off, newpte);
+       }
+
+       eaddr += PAGESIZE;
+       kaddr += PAGESIZE;
+    }
+}
+
+void
+ep4_sdram_map (EP_RAIL *r, EP_ADDR eaddr, sdramaddr_t saddr, unsigned int len, unsigned int perm, int ep_attr)
+{
+    EP4_RAIL    *rail   = (EP4_RAIL *) r;
+    ELAN4_DEV   *dev    = rail->r_ctxt.ctxt_dev;
+    unsigned int npages = len >> PAGESHIFT;
+    int          i;
+    unsigned int off;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (saddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    if ((eaddr & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT)) != (saddr & (SDRAM_PGOFF_OFFSET << PAGE_SHIFT)))
+       printk ("ep4_sdram_map: eaddr=%x saddr=%lx - incorrectly alised\n", eaddr, saddr);
+
+    for (i = 0; i < npages; i++)
+    {
+       for (off = 0; off < PAGESIZE; off += (1 << dev->dev_pageshift[0]))
+       {
+           E4_uint64 newpte = ((saddr + off) >> PTE_PADDR_SHIFT) | PTE_SetPerm (sdram_permtable[perm]);
+
+           elan4mmu_pteload (&rail->r_ctxt, 0, eaddr + off, newpte);
+       }
+
+       eaddr += PAGESIZE;
+       saddr += PAGESIZE;
+    }
+}
+
+void
+ep4_ioaddr_map (EP_RAIL *r, EP_ADDR eaddr, ioaddr_t ioaddr, unsigned int len, unsigned int perm)
+{
+    EP4_RAIL    *rail   = (EP4_RAIL *) r;
+    ELAN4_DEV   *dev    = rail->r_ctxt.ctxt_dev;
+    unsigned int npages = len >> PAGESHIFT;
+    int          i;
+    unsigned int off;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (ioaddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    for (i = 0; i < npages; i++)
+    {
+       physaddr_t paddr = vaddr_to_phys (kernel_map, (void *) ioaddr);
+
+       for (off = 0; off < PAGESIZE; off += (1 << dev->dev_pageshift[0]))
+       {
+           E4_uint64 newpte = elan4mmu_phys2pte (dev, paddr + off, io_permtable[perm]);
+
+           elan4mmu_pteload (&rail->r_ctxt, 0, eaddr + off, newpte);
+       }
+
+       eaddr += PAGESIZE;
+       ioaddr += PAGESIZE;
+    }
+}
+void
+ep4_unmap (EP_RAIL *r, EP_ADDR eaddr, unsigned int len)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+
+    ASSERT ((eaddr & PAGEOFFSET) == 0 && (len & PAGEOFFSET) == 0);
+
+    elan4mmu_unload_range (&rail->r_ctxt, 0, eaddr, len);
+}
+
+void *
+ep4_dvma_reserve (EP_RAIL *r, EP_ADDR eaddr, unsigned int npages)
+{
+    EP4_RAIL  *rail = (EP4_RAIL *) r;
+    ELAN4_DEV *dev  = rail->r_ctxt.ctxt_dev;
+
+    EPRINTF3 (DBG_KMAP, "ep4_dvma_reserve: eaddr=%x npages=%d (=> %d)\n", eaddr, npages, (npages << (PAGE_SHIFT - dev->dev_pageshift[0])));
+
+    return elan4mmu_reserve (&rail->r_ctxt, 0, (E4_Addr) eaddr, (npages << (PAGE_SHIFT - dev->dev_pageshift[0])), 0);
+}
+
+void
+ep4_dvma_release (EP_RAIL *r, EP_ADDR eaddr, unsigned int npages, void *private)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+
+    EPRINTF3 (DBG_KMAP, "ep4_dvma_release: eaddr=%x npages=%d private=%p\n", eaddr, npages, private);
+
+    elan4mmu_release (&rail->r_ctxt, (ELAN4_HASH_CACHE *) private);
+}
+
+void
+ep4_dvma_set_pte (EP_RAIL *r, void *private, unsigned int index, physaddr_t paddr, unsigned int perm)
+{
+    EP4_RAIL     *rail  = (EP4_RAIL *) r;
+    ELAN4_DEV    *dev   = rail->r_ctxt.ctxt_dev;
+    unsigned int  off;
+    unsigned long flags;
+
+    EPRINTF3 (DBG_KMAP, "ep4_dvma_set_pte: index %x -> eaddr %llx paddr %llx\n", 
+             index, ((ELAN4_HASH_CACHE *) private)->hc_start + (index * PAGE_SIZE), (long long) paddr);
+
+    local_irq_save (flags);
+    for (off = 0; off < PAGESIZE; off += (1 << dev->dev_pageshift[0]))
+    {
+       E4_uint64 newpte = elan4mmu_phys2pte (dev, paddr + off, main_permtable[perm]);
+
+       elan4mmu_set_pte (&rail->r_ctxt, (ELAN4_HASH_CACHE *) private, (index << (PAGE_SHIFT - dev->dev_pageshift[0])) +
+                         (off >> dev->dev_pageshift[0]), newpte);
+    }
+    local_irq_restore (flags);
+}
+
+physaddr_t
+ep4_dvma_read_pte (EP_RAIL *r, void *private, unsigned int index)
+{
+    EP4_RAIL     *rail  = (EP4_RAIL *) r;
+    ELAN4_DEV    *dev   = rail->r_ctxt.ctxt_dev;
+    E4_uint64     pte;
+    unsigned long flags;
+
+    local_irq_save (flags);
+    pte = elan4mmu_get_pte (&rail->r_ctxt, (ELAN4_HASH_CACHE *) private, index << (PAGE_SHIFT - dev->dev_pageshift[0]));
+    local_irq_restore (flags);
+
+    return elan4mmu_pte2phys (dev, pte);
+}
+
+void
+ep4_dvma_unload (EP_RAIL *r, void *private, unsigned int index, unsigned int npages)
+{
+    EP4_RAIL  *rail  = (EP4_RAIL *) r;
+    ELAN4_DEV *dev   = rail->r_ctxt.ctxt_dev;
+    EP_ADDR    eaddr = ((ELAN4_HASH_CACHE *) private)->hc_start + (index * PAGE_SIZE);
+    unsigned long idx = (index << (PAGE_SHIFT - dev->dev_pageshift[0]));
+    unsigned long lim = idx + (npages << (PAGE_SHIFT - dev->dev_pageshift[0]));
+    unsigned long flags;
+
+    EPRINTF5 (DBG_KMAP, "ep4_dvma_unload: eaddr %x -> %lx : index=%d idx=%ld lim=%ld\n", 
+             eaddr, (unsigned long)(eaddr + (npages * PAGE_SIZE)), index, idx, lim);
+
+    local_irq_save (flags);
+    for (; idx < lim; idx++)
+       elan4mmu_clear_pte (&rail->r_ctxt, (ELAN4_HASH_CACHE *) private, idx);
+    local_irq_restore (flags);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/kmsg_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kmsg_elan3.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kmsg_elan3.c      2005-06-01 23:12:54.671428160 -0400
@@ -0,0 +1,345 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kmsg_elan3.c,v 1.3.8.1 2004/09/30 09:52:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kmsg_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "debug.h"
+
+static void
+ep3_inputq_event (EP3_RAIL *rail, void *arg)
+{
+    EP3_INPUTQ *inputq = (EP3_INPUTQ *) arg;
+    
+    (*inputq->q_callback)((EP_RAIL *)rail, inputq->q_arg);
+}
+
+static EP3_COOKIE_OPS ep3_inputq_cookie_ops = 
+{
+    ep3_inputq_event,
+};
+
+EP_INPUTQ *
+ep3_alloc_inputq (EP_RAIL *r, unsigned qnum, unsigned slotSize, unsigned slotCount,
+                 EP_INPUTQ_CALLBACK *callback, void *arg)
+{
+    EP3_RAIL      *rail = (EP3_RAIL *) r;
+    EP3_INPUTQ    *inputq;
+    EP3_InputQueue qdesc;
+    void          *slots;
+    int            i;
+
+    ASSERT ((slotSize & (EP_SYSTEMQ_MSG_ALIGN-1)) == 0);
+
+    KMEM_ALLOC (inputq, EP3_INPUTQ *, sizeof (EP3_INPUTQ), TRUE);
+
+    if (inputq == NULL)
+       return (EP_INPUTQ *) NULL;
+    
+    if ((slots = ep_alloc_main (&rail->Generic, slotSize * slotCount, 0, &inputq->q_slotsAddr)) == NULL)
+    {
+       KMEM_FREE (inputq, sizeof (EP3_INPUTQ));
+       return (EP_INPUTQ *) NULL;
+    }
+
+    inputq->q_slotSize  = slotSize;
+    inputq->q_slotCount = slotCount;
+    inputq->q_callback  = callback;
+    inputq->q_arg       = arg;
+    inputq->q_slots     = slots;
+
+    /* Initialise all the slots to be "unreceived" */
+    for (i = 0; i < slotCount; i++)
+       ((uint32_t *) ((unsigned long) slots + (i+1) * slotSize))[-1] = EP_SYSTEMQ_UNRECEIVED;
+    
+    inputq->q_base     = inputq->q_slotsAddr;
+    inputq->q_top      = inputq->q_base + (slotCount-1) * slotSize;
+    inputq->q_fptr     = inputq->q_base;
+    inputq->q_desc     = EP_SYSTEMQ_DESC(rail->QueueDescs, qnum);
+    inputq->q_descAddr = EP_SYSTEMQ_ADDR (qnum);
+
+    if (callback)
+       RegisterCookie (&rail->CookieTable, &inputq->q_cookie, inputq->q_descAddr, &ep3_inputq_cookie_ops, inputq);
+
+    /* Initialise the input queue descriptor */
+    qdesc.q_state          = E3_QUEUE_FULL;
+    qdesc.q_bptr           = inputq->q_base + slotSize;
+    qdesc.q_fptr           = inputq->q_fptr;
+    qdesc.q_base           = inputq->q_base;
+    qdesc.q_top            = inputq->q_top;
+    qdesc.q_size           = slotSize;
+    qdesc.q_event.ev_Count = 1;
+    qdesc.q_event.ev_Type  = callback ? EV_TYPE_EVIRQ | inputq->q_cookie.Cookie : 0;
+    qdesc.q_wevent         = inputq->q_descAddr + offsetof (EP3_InputQueue, q_event);
+    qdesc.q_wcount         = 0;
+
+    /* copy the queue descriptor down to sdram */
+    elan3_sdram_copyl_to_sdram (rail->Device, &qdesc, inputq->q_desc, sizeof (EP3_InputQueue));
+
+    return (EP_INPUTQ *) inputq;
+}
+
+void
+ep3_free_inputq (EP_RAIL *r, EP_INPUTQ *q)
+{
+    EP3_RAIL   *rail   = (EP3_RAIL *) r;
+    EP3_INPUTQ *inputq = (EP3_INPUTQ *) q;
+
+    ep_free_main (&rail->Generic, inputq->q_slotsAddr, inputq->q_slotSize * inputq->q_slotCount);
+
+    if (inputq->q_callback)
+       DeregisterCookie (&rail->CookieTable, &inputq->q_cookie);
+
+    KMEM_FREE (inputq, sizeof (EP3_INPUTQ));
+}
+
+void
+ep3_enable_inputq (EP_RAIL *r, EP_INPUTQ *q)
+{
+    EP3_RAIL   *rail   = (EP3_RAIL *) r;
+    EP3_INPUTQ *inputq = (EP3_INPUTQ *) q;
+
+    elan3_sdram_writel (rail->Device, inputq->q_desc + offsetof (EP3_InputQueue, q_state), 0);
+}
+
+void
+ep3_disable_inputq (EP_RAIL *r, EP_INPUTQ *q)
+{
+    EP3_RAIL      *rail   = (EP3_RAIL *) r;
+    EP3_INPUTQ    *inputq = (EP3_INPUTQ *) q;
+    EP3_InputQueue qdesc;
+
+    /* mark the queue as locked */
+    SetQueueLocked (rail, inputq->q_desc);
+
+    /* re-initialise the queue as empty */
+    qdesc.q_state          = E3_QUEUE_FULL;
+    qdesc.q_bptr           = (E3_Addr) inputq->q_base + inputq->q_slotSize;
+    qdesc.q_fptr           = inputq->q_fptr;
+    qdesc.q_base           = inputq->q_base;
+    qdesc.q_top            = inputq->q_top;
+    qdesc.q_size           = inputq->q_slotSize;
+    qdesc.q_event.ev_Count = 1;
+    qdesc.q_event.ev_Type  = inputq->q_callback ? EV_TYPE_EVIRQ | inputq->q_cookie.Cookie : 0;
+    qdesc.q_wevent         = inputq->q_descAddr + offsetof (EP3_InputQueue, q_event);
+    qdesc.q_wcount         = 0;
+
+    /* copy the queue descriptor down to sdram */
+    elan3_sdram_copyl_to_sdram (rail->Device, &qdesc, inputq->q_desc, sizeof (EP3_InputQueue));
+}
+
+int
+ep3_poll_inputq (EP_RAIL *r, EP_INPUTQ *q, int maxCount, EP_INPUTQ_HANDLER *handler, void *arg)
+{
+    EP3_RAIL          *rail   = (EP3_RAIL *) r;
+    EP3_INPUTQ        *inputq = (EP3_INPUTQ *) q;
+    sdramaddr_t        qdesc  = inputq->q_desc;
+    E3_Addr            nfptr;
+    int                count = 0;
+    E3_uint32          state;
+    int                       delay;
+
+ run_again_because_of_eventqueue_overflow:
+    nfptr = inputq->q_fptr + inputq->q_slotSize;
+    if (nfptr > inputq->q_top)                                 
+       nfptr = inputq->q_base;
+
+    while (nfptr != elan3_sdram_readl (rail->Device, qdesc + offsetof (EP3_InputQueue, q_bptr)))                       /* PCI read */
+    {
+       unsigned long slot = (unsigned long) inputq->q_slots + (nfptr - inputq->q_base);
+
+       /* Poll the final word of the message until the message has completely
+        * arrived in main memory. */
+       for (delay = 1; ((uint32_t *) (slot + inputq->q_slotSize))[-1] == EP_SYSTEMQ_UNRECEIVED && delay < EP_SYSTEMQ_UNRECEIVED_TLIMIT; delay <<= 1)
+           DELAY (delay);
+
+       /* Call the message handler */
+       (*handler) (r, arg, (void *) slot);
+       
+       state = elan3_sdram_readl (rail->Device, qdesc + offsetof (EP3_InputQueue, q_state));                           /* PCI read */
+       if ((state & E3_QUEUE_FULL) == 0)
+           elan3_sdram_writel (rail->Device, qdesc + offsetof (EP3_InputQueue, q_fptr), nfptr);                        /* PCI write */
+       else
+       {
+           elan3_sdram_writel (rail->Device, qdesc + offsetof (EP3_InputQueue, q_fptr), nfptr);                        /* PCI write */
+           elan3_sdram_writel (rail->Device, qdesc + offsetof (EP3_InputQueue, q_state), (state & ~E3_QUEUE_FULL));    /* PCI write */
+       }
+       inputq->q_fptr = nfptr;
+       
+       nfptr += roundup (inputq->q_slotSize, E3_BLK_ALIGN);
+       if (nfptr > inputq->q_top)
+           nfptr = inputq->q_base;
+
+       if (++count >= maxCount && maxCount)
+           break;
+    }
+    
+    if (inputq->q_callback && count != 0)
+    {
+       if (count != inputq->q_waitCount)
+           elan3_sdram_writel (rail->Device, qdesc + offsetof (EP3_InputQueue, q_wcount), inputq->q_waitCount = count);
+
+       if (IssueWaitevent (rail, inputq->q_descAddr + offsetof (EP3_InputQueue, q_wevent)) == ISSUE_COMMAND_TRAPPED)
+           goto run_again_because_of_eventqueue_overflow;
+    }
+
+    return count;
+}
+
+#define Q_EVENT(q,slotNum)             ((q)->q_elan      + (slotNum) * sizeof (E3_BlockCopyEvent))
+#define Q_EVENT_ADDR(q,slotNum)                ((q)->q_elanAddr  + (slotNum) * sizeof (E3_BlockCopyEvent))
+#define Q_MSG(q,slotNum)       (void *)((q)->q_main      + (slotNum) * (q)->q_slotSize)
+#define Q_MSG_ADDR(q,slotNum)          ((q)->q_mainAddr  + (slotNum) * (q)->q_slotSize)
+#define Q_DONE(q,slotNum)     (*((int *)((q)->q_main      + (q)->q_slotCount * (q)->q_slotSize + (slotNum) * sizeof (E3_uint32))))
+#define Q_DONE_ADDR(q,slotNum)         ((q)->q_mainAddr  + (q)->q_slotCount * (q)->q_slotSize + (slotNum) * sizeof (E3_uint32))
+
+#define Q_ELAN_SIZE(q)                 ((q)->q_slotCount * sizeof (E3_BlockCopyEvent))
+#define Q_MAIN_SIZE(q)                 ((q)->q_slotCount * ((q)->q_slotSize + sizeof (E3_uint32)))
+
+static void
+ep3_outputq_retry (EP3_RAIL *rail, void *arg, E3_DMA_BE *dma, int error)
+{
+    E3_DMA_BE    *dmabe = (E3_DMA_BE *) dma;
+    sdramaddr_t   event = ep_elan2sdram (&rail->Generic, dmabe->s.dma_srcEvent);
+    E3_Addr       done  = elan3_sdram_readl (rail->Device, event + offsetof (E3_BlockCopyEvent, ev_Dest));
+    E3_uint32    *donep = ep_elan2main (&rail->Generic, done & ~EV_BCOPY_DTYPE_MASK);
+
+    EPRINTF1 (DBG_KMSG, "ep3_ouputq_retry: donep at %p -> FAILED\n", donep);
+    
+    *donep = EP3_EVENT_FAILED;
+}
+
+static EP3_COOKIE_OPS ep3_outputq_cookie_ops =
+{
+    NULL, /* Event */
+    ep3_outputq_retry,
+    NULL, /* DmaCancelled */
+    NULL, /* DmaVerify */
+};
+
+EP_OUTPUTQ *
+ep3_alloc_outputq (EP_RAIL *r, unsigned slotSize, unsigned slotCount)
+{
+    EP3_RAIL         *rail = (EP3_RAIL *) r;
+    EP3_OUTPUTQ      *outputq;
+    int               i;
+    E3_BlockCopyEvent event;
+
+    ASSERT ((slotSize & (EP_SYSTEMQ_MSG_ALIGN-1)) == 0);
+
+    KMEM_ALLOC (outputq, EP3_OUTPUTQ *, sizeof (EP3_OUTPUTQ), 1);
+
+    if (outputq == NULL)
+       return NULL;
+
+    outputq->q_slotCount = slotCount;
+    outputq->q_slotSize  = slotSize;
+
+    outputq->q_elan = ep_alloc_elan (r, Q_ELAN_SIZE(outputq), 0, &outputq->q_elanAddr);
+
+    if (outputq->q_elan == (sdramaddr_t) 0)
+    {
+       KMEM_FREE (outputq, sizeof (EP3_OUTPUTQ));
+       return NULL;
+    }
+
+    outputq->q_main = ep_alloc_main (r, Q_MAIN_SIZE(outputq), 0, &outputq->q_mainAddr);
+
+    if (outputq->q_main == (void *) NULL)
+    {
+       ep_free_elan (r, outputq->q_elanAddr, Q_ELAN_SIZE(outputq));
+       KMEM_FREE (outputq, sizeof (EP3_OUTPUTQ));
+       return NULL;
+    }
+
+    RegisterCookie (&rail->CookieTable, &outputq->q_cookie, outputq->q_elanAddr, &ep3_outputq_cookie_ops, outputq);
+
+    for (i = 0; i < slotCount; i++)
+    {
+       EP3_INIT_COPY_EVENT (event, outputq->q_cookie, Q_DONE_ADDR(outputq, i), 0);
+
+       Q_DONE(outputq, i) = outputq->q_cookie.Cookie;
+       
+       elan3_sdram_copyl_to_sdram (rail->Device, &event, Q_EVENT(outputq, i), sizeof (E3_BlockCopyEvent));
+    }
+
+    return (EP_OUTPUTQ *) outputq;
+}
+
+void
+ep3_free_outputq (EP_RAIL *r, EP_OUTPUTQ *q)
+{
+    EP3_RAIL    *rail    = (EP3_RAIL *) r;
+    EP3_OUTPUTQ *outputq = (EP3_OUTPUTQ *) q;
+
+    DeregisterCookie (&rail->CookieTable, &outputq->q_cookie);
+    
+    ep_free_main (r, outputq->q_mainAddr, Q_MAIN_SIZE(outputq));
+    ep_free_elan (r, outputq->q_elanAddr, Q_ELAN_SIZE(outputq));
+ 
+    KMEM_FREE (outputq, sizeof (EP3_OUTPUTQ));
+}
+
+void *
+ep3_outputq_msg (EP_RAIL *r, EP_OUTPUTQ *q, unsigned slotNum)
+{
+    return Q_MSG ((EP3_OUTPUTQ *) q, slotNum);
+}
+
+int
+ep3_outputq_state (EP_RAIL *r, EP_OUTPUTQ *q, unsigned slotNum)
+{
+    switch (Q_DONE((EP3_OUTPUTQ *) q, slotNum))
+    {
+    case EP3_EVENT_ACTIVE:
+       return EP_OUTPUTQ_BUSY;
+       
+    case EP3_EVENT_FAILED:
+       return EP_OUTPUTQ_FAILED;
+       
+    default:
+       return EP_OUTPUTQ_FINISHED;
+    }
+}
+
+int
+ep3_outputq_send (EP_RAIL *r, EP_OUTPUTQ *q, unsigned slotNum, unsigned size, 
+                 unsigned vp, unsigned qnum, unsigned retries)
+{
+    EP3_RAIL    *rail    = (EP3_RAIL *) r;
+    EP3_OUTPUTQ *outputq = (EP3_OUTPUTQ *) q;
+    unsigned     base    = outputq->q_slotSize - roundup (size, E3_BLK_ALIGN);
+    E3_DMA_BE    dmabe;
+
+    dmabe.s.dma_type           = E3_DMA_TYPE(DMA_BYTE, DMA_WRITE, DMA_QUEUED, retries);
+    dmabe.s.dma_size            = roundup (size, E3_BLK_ALIGN);
+    dmabe.s.dma_source          = Q_MSG_ADDR(outputq, slotNum) + base;
+    dmabe.s.dma_dest            = base;
+    dmabe.s.dma_destEvent       = EP_SYSTEMQ_ADDR(qnum);
+    dmabe.s.dma_destCookieVProc = vp;
+    dmabe.s.dma_srcEvent        = Q_EVENT_ADDR(outputq, slotNum);
+    dmabe.s.dma_srcCookieVProc  = 0;
+
+    Q_DONE(outputq, slotNum) = EP3_EVENT_ACTIVE;
+    
+    elan3_sdram_writel (rail->Device, Q_EVENT(outputq, slotNum), 1);
+
+    if (IssueDma (rail, &dmabe, EP_RETRY_CRITICAL, FALSE) != ISSUE_COMMAND_OK)
+    {
+       Q_DONE(outputq, slotNum) = EP3_EVENT_FAILED;
+       return FALSE;
+    }
+
+    return TRUE;
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/kmsg_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kmsg_elan4.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kmsg_elan4.c      2005-06-01 23:12:54.672428008 -0400
@@ -0,0 +1,416 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kmsg_elan4.c,v 1.8.6.1 2004/09/30 09:52:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kmsg_elan4.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "debug.h"
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+
+#include <elan4/trtype.h>
+
+static void
+ep4_inputq_interrupt (EP4_RAIL *rail, void *arg)
+{
+    EP4_INPUTQ *inputq = (EP4_INPUTQ *) arg;
+
+    /* mark the queue as "fired" to cause a single waitevent
+     * to be issued next time the queue is polled */
+    atomic_inc (&inputq->q_fired);
+    
+    (*inputq->q_callback)(&rail->r_generic, inputq->q_arg);
+}
+
+EP_INPUTQ *
+ep4_alloc_inputq (EP_RAIL *r, unsigned qnum, unsigned slotSize, unsigned slotCount,
+                 EP_INPUTQ_CALLBACK *callback, void *arg)
+{
+    EP4_RAIL     *rail = (EP4_RAIL *) r;
+    EP4_INPUTQ   *inputq;
+    E4_Event32    qevent;
+    void         *slots;
+    int           i;
+
+    ASSERT ((slotSize & (EP_SYSTEMQ_MSG_ALIGN-1)) == 0);
+
+    KMEM_ALLOC (inputq, EP4_INPUTQ *, sizeof (EP4_INPUTQ), 1);
+
+    if (inputq == NULL)
+       return (EP_INPUTQ *) NULL;
+    
+    if ((slots = ep_alloc_main (&rail->r_generic, slotSize * slotCount, 0, &inputq->q_slotsAddr)) == NULL)
+    {
+       KMEM_FREE (inputq, sizeof (EP4_INPUTQ));
+       return (EP_INPUTQ *) NULL;
+    }
+
+    inputq->q_slotSize  = slotSize;
+    inputq->q_slotCount = slotCount;
+    inputq->q_callback  = callback;
+    inputq->q_arg       = arg;
+    inputq->q_slots     = slots;
+
+    /* Initialise all the slots to be "unreceived" */
+    for (i = 0; i < slotCount; i++)
+       ((uint32_t *) ((unsigned long) slots + (i+1) * slotSize))[-1] = EP_SYSTEMQ_UNRECEIVED;
+    
+    inputq->q_last      = inputq->q_slotsAddr + (slotCount-1) * slotSize;
+    inputq->q_fptr      = inputq->q_slotsAddr;
+    inputq->q_desc      = EP_SYSTEMQ_DESC (rail->r_queuedescs, qnum);
+    inputq->q_descAddr  = EP_SYSTEMQ_ADDR (qnum);
+    inputq->q_eventAddr = rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_qevents[qnum]);
+
+    if (callback)
+    {
+       if ((inputq->q_ecq = ep4_get_ecq (rail, EP4_ECQ_EVENT, 1)) == 0)
+       {
+           ep_free_main (&rail->r_generic, inputq->q_slotsAddr, inputq->q_slotSize * inputq->q_slotCount);
+
+           KMEM_FREE (inputq, sizeof (EP4_INPUTQ));
+           return (EP_INPUTQ *) NULL;
+       }
+
+       if ((inputq->q_wcq = ep4_get_ecq (rail, EP4_ECQ_MAIN, 4)) == 0)
+       {
+           ep4_put_ecq (rail, inputq->q_ecq, 1);
+           ep_free_main (&rail->r_generic, inputq->q_slotsAddr, inputq->q_slotSize * inputq->q_slotCount);
+
+           KMEM_FREE (inputq, sizeof (EP4_INPUTQ));
+           return (EP_INPUTQ *) NULL;
+       }
+
+       ep4_register_intcookie (rail, &inputq->q_intcookie, inputq->q_descAddr, ep4_inputq_interrupt, inputq);
+
+       inputq->q_count = 0;
+
+       atomic_set (&inputq->q_fired, 0);
+
+       /* Initialise the queue event */
+       qevent.ev_CountAndType = E4_EVENT_INIT_VALUE (callback ? -32 : 0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0);
+       qevent.ev_WritePtr     = inputq->q_ecq->ecq_addr;
+       qevent.ev_WriteValue   = (inputq->q_intcookie.int_val << E4_MAIN_INT_SHIFT) | INTERRUPT_CMD;
+    }
+
+    /* copy the event down to sdram */
+    elan4_sdram_copyq_to_sdram (rail->r_ctxt.ctxt_dev, &qevent, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_qevents[qnum]), sizeof (E4_Event32));
+
+    return (EP_INPUTQ *) inputq;
+}
+
+void
+ep4_free_inputq (EP_RAIL *r, EP_INPUTQ *q)
+{
+    EP4_RAIL   *rail   = (EP4_RAIL *) r;
+    EP4_INPUTQ *inputq = (EP4_INPUTQ *) q;
+
+    ep_free_main (&rail->r_generic, inputq->q_slotsAddr, inputq->q_slotSize * inputq->q_slotCount);
+
+    if (inputq->q_callback)
+    {
+       ep4_deregister_intcookie (rail, &inputq->q_intcookie);
+       ep4_put_ecq (rail, inputq->q_ecq, 1);
+       ep4_put_ecq (rail, inputq->q_wcq, 4);
+    }
+
+    KMEM_FREE (inputq, sizeof (EP4_INPUTQ));
+}
+
+void
+ep4_enable_inputq (EP_RAIL *r, EP_INPUTQ *q)
+{
+    EP4_RAIL     *rail     = (EP4_RAIL *) r;
+    EP4_INPUTQ   *inputq   = (EP4_INPUTQ *) q;
+    EP_ADDR       lastSlot = inputq->q_slotsAddr + (inputq->q_slotCount-1) * inputq->q_slotSize;
+    E4_InputQueue qdesc;
+
+    qdesc.q_bptr    = inputq->q_slotsAddr;
+    qdesc.q_fptr    = inputq->q_slotsAddr;
+    qdesc.q_control = E4_InputQueueControl (inputq->q_slotsAddr, lastSlot, inputq->q_slotSize);
+    qdesc.q_event   = inputq->q_callback ? inputq->q_eventAddr : 0;
+
+    /* copy the queue descriptor down to sdram */
+    ep4_write_qdesc (rail, inputq->q_desc, &qdesc);
+
+    EPRINTF5 (DBG_KMSG,  "ep_enable_inputq: %x - %016llx %016llx %016llx %016llx\n", (int) inputq->q_descAddr,
+           elan4_sdram_readq (rail->r_ctxt.ctxt_dev, inputq->q_desc + 0),
+           elan4_sdram_readq (rail->r_ctxt.ctxt_dev, inputq->q_desc + 8),
+           elan4_sdram_readq (rail->r_ctxt.ctxt_dev, inputq->q_desc + 16),
+           elan4_sdram_readq (rail->r_ctxt.ctxt_dev, inputq->q_desc + 24));
+}
+
+void
+ep4_disable_inputq (EP_RAIL *r, EP_INPUTQ *q)
+{
+    EP4_RAIL     *rail   = (EP4_RAIL *) r;
+    EP4_INPUTQ   *inputq = (EP4_INPUTQ *) q;
+    E4_InputQueue qdesc;
+
+    /* Initialise the input queue descriptor as "full" with no event */
+    qdesc.q_bptr    = 0;
+    qdesc.q_fptr    = 8;
+    qdesc.q_control = E4_InputQueueControl(qdesc.q_bptr, qdesc.q_fptr, 8);
+    qdesc.q_event   = 0;
+
+    /* copy the queue descriptor down to sdram */
+    ep4_write_qdesc (rail, inputq->q_desc, &qdesc);
+}
+
+int
+ep4_poll_inputq (EP_RAIL *r, EP_INPUTQ *q, int maxCount, EP_INPUTQ_HANDLER *handler, void *arg)
+{
+    EP4_RAIL   *rail   = (EP4_RAIL *) r;
+    ELAN4_DEV  *dev    = rail->r_ctxt.ctxt_dev; 
+    EP4_INPUTQ *inputq = (EP4_INPUTQ *) q;
+    sdramaddr_t qdesc = inputq->q_desc;
+    E4_Addr     fptr  = inputq->q_fptr;
+    E4_Addr     bptr  = elan4_sdram_readl (dev, qdesc + offsetof (E4_InputQueue, q_bptr));
+    int                count = 0;
+    int         delay;
+
+    while (bptr != 0 && fptr != bptr)
+    {
+       while (fptr != bptr)
+       {
+           unsigned long slot = (unsigned long) inputq->q_slots + (fptr - inputq->q_slotsAddr);
+           
+           /* Poll the final word of the message until the message has completely
+            * arrived in main memory. */
+           for (delay = 1; ((uint32_t *) (slot + inputq->q_slotSize))[-1] == EP_SYSTEMQ_UNRECEIVED && delay < EP_SYSTEMQ_UNRECEIVED_TLIMIT; delay <<= 1)
+               DELAY (delay);
+           
+           EPRINTF4(DBG_KMSG, "ep4_poll_inputq: %x slot %d of %d [%08x]\n", (int)inputq->q_descAddr,
+                    ((int)(fptr - inputq->q_slotsAddr))/inputq->q_slotSize, 
+                    inputq->q_slotCount, ((uint32_t *) (slot + inputq->q_slotSize))[-1]);
+           
+           /* Call the message handler */
+           (*handler) (r, arg, (void *) slot);
+           
+           /* reset the last word of the slot to "unreceived" */
+           ((uint32_t *) (slot + inputq->q_slotSize))[-1] = EP_SYSTEMQ_UNRECEIVED;
+           
+           /* move on the front pointer */
+           fptr = (fptr == inputq->q_last) ? inputq->q_slotsAddr : fptr + inputq->q_slotSize;
+           
+           elan4_sdram_writel (dev, qdesc + offsetof (E4_InputQueue, q_fptr), fptr);
+           
+           inputq->q_count++;
+           
+           if (++count >= maxCount && maxCount)
+           {
+               inputq->q_fptr = fptr;
+
+               return count;
+           }
+       }
+
+       bptr = elan4_sdram_readl (dev, qdesc + offsetof (E4_InputQueue, q_bptr));
+    }
+
+    inputq->q_fptr = fptr;
+
+    /* Only insert a single wait event command if the callback has
+     * occured, otherwise just acrue the count as we've just periodically
+     * polled it.
+     */
+    if (inputq->q_callback && atomic_read (&inputq->q_fired))
+    {
+       atomic_dec (&inputq->q_fired);
+
+       ep4_wait_event_cmd (inputq->q_wcq, inputq->q_eventAddr,
+                           E4_EVENT_INIT_VALUE (-inputq->q_count << 5, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0),
+                           inputq->q_ecq->ecq_addr,
+                           (inputq->q_intcookie.int_val << E4_MAIN_INT_SHIFT) | INTERRUPT_CMD);
+
+       inputq->q_count = 0;
+    }
+
+    return count;
+}
+
+#define Q_MSG(q,slotNum)         (unsigned long)((q)->q_main      + (slotNum) * (q)->q_slotSize)
+#define Q_MSG_ADDR(q,slotNum)                  ((q)->q_mainAddr  + (slotNum) * (q)->q_slotSize)
+#define Q_DONE(q,slotNum)        *((E4_uint64 *)((q)->q_main      + (q)->q_slotCount * (q)->q_slotSize + (slotNum) * sizeof (E4_uint64)))
+#define Q_DONE_ADDR(q,slotNum)                 ((q)->q_mainAddr  + (q)->q_slotCount * (q)->q_slotSize + (slotNum) * sizeof (E4_uint64))
+
+#define Q_MAIN_SIZE(q)                 ((q)->q_slotCount * ((q)->q_slotSize + sizeof (E4_uint64)))
+
+#define Q_DONE_VAL(val,cnt)            ((cnt) << 16 | (val))
+#define Q_DONE_RET(done)               ((int) ((done) & 0xffff))
+#define Q_DONE_CNT(done)               ((int) ((done) >> 16))
+
+EP_OUTPUTQ *
+ep4_alloc_outputq (EP_RAIL *r, unsigned slotSize, unsigned slotCount)
+{
+    EP4_RAIL    *rail = (EP4_RAIL *) r;
+    EP4_OUTPUTQ *outputq;
+    int          i;
+
+    ASSERT ((slotSize & (EP_SYSTEMQ_MSG_ALIGN-1)) == 0);
+
+    KMEM_ALLOC (outputq, EP4_OUTPUTQ *, sizeof (EP4_OUTPUTQ), 1);
+
+    if (outputq == NULL)
+       return NULL;
+
+    spin_lock_init (&outputq->q_lock);
+
+    outputq->q_slotCount = slotCount;
+    outputq->q_slotSize  = slotSize;
+    outputq->q_main      = ep_alloc_main (r, Q_MAIN_SIZE(outputq), 0, &outputq->q_mainAddr);
+
+    if (outputq->q_main == (E4_uint64 *) NULL)
+    {
+       KMEM_FREE (outputq, sizeof (EP_OUTPUTQ));
+       return NULL;
+    }
+
+    outputq->q_cq = elan4_alloccq (&rail->r_ctxt, CQ_Size64K, CQ_STENEnableBit | CQ_WriteEnableBit, CQ_Priority);
+
+    if (outputq->q_cq == (ELAN4_CQ *) NULL)
+    {
+       ep_free_main (&rail->r_generic, outputq->q_mainAddr, Q_MAIN_SIZE(outputq));
+
+       KMEM_FREE (outputq, sizeof (EP_OUTPUTQ));
+    }
+
+    outputq->q_dwords = CQ_Size (outputq->q_cq->cq_size) >> 3;
+
+    /* mark all the queue slots as finished */
+    for (i = 0; i < slotCount; i++)
+       Q_DONE(outputq, i) = Q_DONE_VAL (EP_OUTPUTQ_FINISHED, 0);
+
+    return (EP_OUTPUTQ *) outputq;
+}
+
+void
+ep4_free_outputq (EP_RAIL *r, EP_OUTPUTQ *q)
+{
+    EP4_RAIL    *rail    = (EP4_RAIL *) r;
+    EP4_OUTPUTQ *outputq = (EP4_OUTPUTQ *) q;
+
+    elan4_freecq (&rail->r_ctxt, outputq->q_cq);
+
+    ep_free_main (&rail->r_generic, outputq->q_mainAddr, Q_MAIN_SIZE(outputq));
+
+    spin_lock_destroy (&outputq->q_lock);
+
+    KMEM_FREE (outputq, sizeof (EP4_OUTPUTQ));
+}
+
+void *
+ep4_outputq_msg (EP_RAIL *r, EP_OUTPUTQ *q, unsigned slotNum)
+{
+    return (void *) Q_MSG ((EP4_OUTPUTQ *) q, slotNum);
+}
+
+int
+ep4_outputq_state (EP_RAIL *r, EP_OUTPUTQ *q, unsigned slotNum)
+{
+    EPRINTF2 (DBG_KMSG, "ep4_outputq_state: slotNum %d state %x\n", slotNum, (int)Q_DONE((EP4_OUTPUTQ *) q, slotNum));
+
+    return Q_DONE_RET(Q_DONE((EP4_OUTPUTQ *)q, slotNum));
+}
+
+int
+ep4_outputq_send (EP_RAIL *r, EP_OUTPUTQ *q, unsigned slotNum, unsigned size, 
+                 unsigned vp, unsigned qnum, unsigned retries)
+{
+    EP4_OUTPUTQ *outputq = (EP4_OUTPUTQ *) q;
+    unsigned int nbytes  = roundup (size, 32);
+    unsigned int base    = outputq->q_slotSize - nbytes;
+    unsigned int i, dwords;
+    unsigned long flags;
+    E4_uint64 val;
+
+    spin_lock_irqsave (&outputq->q_lock, flags);
+
+    EPRINTF4 (DBG_KMSG, "ep4_outputq_send: slotNum=%d size=%d vp=%d qnum=%d\n", slotNum, size, vp, qnum);
+
+    /* compute command queue size as follows - each slot uses
+     *     overhead:   14 dwords +
+     *    data > 128 ? 36 dwords
+     *     data > 64  ? 18 dwords
+     *     data > 32  ? 10 dwords
+     *     else         6  dwords
+     */
+    dwords = 14 + (size > 128 ? 36 :
+                  size > 64  ? 18 :
+                  size  ? 10 : 6);
+
+    outputq->q_dwords += Q_DONE_CNT (Q_DONE(outputq, slotNum));
+
+    if (dwords > outputq->q_dwords)
+    {
+       /* attempt to reclaim command queue space from other slots */
+       i = slotNum;
+       do {
+           if (++i == outputq->q_slotCount)
+               i = 0;
+           
+           val = Q_DONE(outputq, i);
+
+           if ((Q_DONE_RET (val) == EP_OUTPUTQ_FINISHED || Q_DONE_RET (val) == EP_OUTPUTQ_FAILED) && Q_DONE_CNT(val) > 0)
+           {
+               outputq->q_dwords += Q_DONE_CNT (val);
+
+               Q_DONE(outputq, i) = Q_DONE_VAL(Q_DONE_RET(val), 0);
+           }
+       } while (i != slotNum && dwords > outputq->q_dwords);
+    }
+
+    if (dwords > outputq->q_dwords)
+    {
+       spin_unlock_irqrestore (&outputq->q_lock, flags);
+       
+       EPRINTF0 (DBG_KMSG, "ep4_outputq_state: no command queue space\n");
+       return 0;
+    }
+
+    outputq->q_dwords -= dwords;
+
+    Q_DONE(outputq, slotNum) = Q_DONE_VAL (EP_OUTPUTQ_BUSY, dwords);
+
+    if (outputq->q_retries != retries)
+    {
+       elan4_guard (outputq->q_cq, GUARD_CHANNEL(1) | GUARD_RESET(outputq->q_retries = retries));
+       elan4_nop_cmd (outputq->q_cq, 0);
+    }
+
+    /* transfer the top "size" bytes from message buffer to top of input queue */
+    elan4_open_packet (outputq->q_cq, OPEN_PACKET (0, PACK_OK | RESTART_COUNT_ZERO, vp));
+    elan4_sendtrans0 (outputq->q_cq, TR_INPUT_Q_GETINDEX, EP_SYSTEMQ_ADDR(qnum));
+
+    /* send upto EP_SYSTEMQ_MSG_MAX (256) bytes of message to the top of the slot */
+    if (size > 128)
+    {
+       elan4_sendtransp (outputq->q_cq, TR_WRITE (128 >> 3, 0, TR_DATATYPE_DWORD), base + 0,   (void *) (Q_MSG(outputq, slotNum) + base + 0));
+       elan4_sendtransp (outputq->q_cq, TR_WRITE (128 >> 3, 0, TR_DATATYPE_DWORD), base + 128, (void *) (Q_MSG(outputq, slotNum) + base + 128));
+    }
+    else if (size > 64)
+       elan4_sendtransp (outputq->q_cq, TR_WRITE (128 >> 3, 0, TR_DATATYPE_DWORD), base, (void *) (Q_MSG(outputq, slotNum) + base));
+    else if (size > 32)
+       elan4_sendtransp (outputq->q_cq, TR_WRITE (64 >> 3, 0, TR_DATATYPE_DWORD),  base, (void *) (Q_MSG(outputq, slotNum) + base));
+    else
+       elan4_sendtransp (outputq->q_cq, TR_WRITE (32 >> 3, 0, TR_DATATYPE_DWORD),  base, (void *) (Q_MSG(outputq, slotNum) + base));
+    elan4_sendtrans1 (outputq->q_cq, TR_INPUT_Q_COMMIT, EP_SYSTEMQ_ADDR(qnum), 0 /* no cookie */);
+
+    elan4_guard (outputq->q_cq, GUARD_CHANNEL (1) | GUARD_TEST(0, PACK_OK) | GUARD_RESET (outputq->q_retries));
+    elan4_write_dword_cmd (outputq->q_cq, Q_DONE_ADDR(outputq, slotNum), Q_DONE_VAL (EP_OUTPUTQ_FINISHED, dwords));
+
+    elan4_guard (outputq->q_cq, GUARD_CHANNEL (1) | GUARD_TEST(0, RESTART_COUNT_ZERO) | GUARD_RESET (outputq->q_retries));
+    elan4_write_dword_cmd (outputq->q_cq, Q_DONE_ADDR(outputq, slotNum), Q_DONE_VAL (EP_OUTPUTQ_FAILED, dwords));
+
+    spin_unlock_irqrestore (&outputq->q_lock, flags);
+
+    return 1;
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/kthread.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kthread.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kthread.c 2005-06-01 23:12:54.672428008 -0400
@@ -0,0 +1,186 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kthread.c,v 1.5 2004/05/19 08:54:57 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/kthread.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kthread.h>
+
+void
+ep_kthread_init (EP_KTHREAD *kt)
+{
+       spin_lock_init (&kt->lock);
+       kcondvar_init (&kt->wait);
+       
+       kt->next_run     = 0;
+       kt->should_stall = 0;
+       kt->started      = 0;
+       kt->should_stop  = 0;
+       kt->stopped      = 0;
+       kt->state        = KT_STATE_RUNNING;
+}
+
+void
+ep_kthread_destroy (EP_KTHREAD *kt)
+{
+       spin_lock_destroy (&kt->lock);
+       kcondvar_destroy (&kt->wait);
+}
+
+void
+ep_kthread_started (EP_KTHREAD *kt)
+{
+       unsigned long flags;
+       
+       spin_lock_irqsave (&kt->lock, flags);
+       kt->started = 1;
+       spin_unlock_irqrestore(&kt->lock, flags);
+}
+
+void
+ep_kthread_stopped (EP_KTHREAD *kt)
+{
+       unsigned long flags;
+       
+       spin_lock_irqsave (&kt->lock, flags);
+       kt->stopped = 1;
+       kcondvar_wakeupall (&kt->wait, &kt->lock);
+       spin_unlock_irqrestore(&kt->lock, flags);
+}
+
+int
+ep_kthread_should_stall (EP_KTHREAD *kth)
+{
+       return (kth->should_stall);
+}
+
+int
+ep_kthread_sleep (EP_KTHREAD *kt, long next_run)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave (&kt->lock, flags);
+       if (next_run && (kt->next_run == 0 || BEFORE (next_run, kt->next_run)))
+               kt->next_run = next_run;
+
+       if (kt->should_stop)
+       {
+               spin_unlock_irqrestore (&kt->lock, flags);
+               return (-1);
+       }
+       
+       do {
+               if (kt->should_stall)
+                       kcondvar_wakeupall (&kt->wait, &kt->lock);
+
+               kt->state = KT_STATE_SLEEPING;
+               kt->running = 0;
+               if (kt->should_stall || kt->next_run == 0)
+                       kcondvar_wait (&kt->wait, &kt->lock, &flags);
+               else
+                       kcondvar_timedwait (&kt->wait,&kt->lock, &flags, kt->next_run);
+               kt->state = KT_STATE_RUNNING;
+               kt->running = lbolt;
+       } while (kt->should_stall);
+       kt->next_run = 0;
+       spin_unlock_irqrestore (&kt->lock, flags);
+       
+       return (0);
+}
+
+void
+ep_kthread_schedule (EP_KTHREAD *kt, long tick)
+{
+       unsigned long flags;
+       
+       spin_lock_irqsave (&kt->lock, flags);
+       if (kt->next_run == 0 || BEFORE (tick, kt->next_run))
+       {
+               kt->next_run = tick;
+               if (!kt->should_stall && kt->state == KT_STATE_SLEEPING)
+               {
+                       kt->state = KT_STATE_SCHEDULED;
+                       kcondvar_wakeupone (&kt->wait, &kt->lock);
+               }
+       }
+       spin_unlock_irqrestore (&kt->lock, flags);
+}
+
+void
+ep_kthread_stall (EP_KTHREAD *kt)
+{
+       unsigned long flags;
+       
+       spin_lock_irqsave (&kt->lock, flags);
+       if (kt->should_stall++ == 0)
+               kcondvar_wakeupall (&kt->wait, &kt->lock);
+
+       while (kt->state != KT_STATE_SLEEPING)
+               kcondvar_wait (&kt->wait, &kt->lock, &flags);
+       spin_unlock_irqrestore (&kt->lock, flags);
+}
+
+void
+ep_kthread_resume (EP_KTHREAD *kt)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave (&kt->lock, flags);
+       if (--kt->should_stall == 0)
+       {
+               kt->state = KT_STATE_SCHEDULED;
+               kcondvar_wakeupone (&kt->wait, &kt->lock);
+       }
+       spin_unlock_irqrestore (&kt->lock, flags);
+}
+
+void
+ep_kthread_stop (EP_KTHREAD *kt)
+{
+       unsigned long flags;
+       
+       spin_lock_irqsave (&kt->lock, flags);
+       kt->should_stop = 1;
+       while (kt->started && !kt->stopped)
+       {
+               kcondvar_wakeupall (&kt->wait, &kt->lock);
+               kcondvar_wait (&kt->wait, &kt->lock, &flags);
+       }
+       spin_unlock_irqrestore (&kt->lock, flags);
+}
+
+int
+ep_kthread_state (EP_KTHREAD *kt, long *time)
+{
+       unsigned long flags;
+       int res = KT_STATE_SLEEPING;
+
+       spin_lock_irqsave (&kt->lock, flags);
+
+       if (kt->next_run) {
+               *time = kt->next_run;
+               res   = kt->should_stall ? KT_STATE_STALLED : KT_STATE_SCHEDULED;
+       }
+
+       if (kt->running) {
+               *time = kt->running;
+               res   = KT_STATE_RUNNING;
+       }
+
+       spin_unlock_irqrestore (&kt->lock, flags);
+       
+       return res;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/kthread.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/kthread.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/kthread.h 2005-06-01 23:12:54.673427856 -0400
@@ -0,0 +1,53 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_KTHREAD_H
+#define __ELAN3_KTHREAD_H
+
+#ident "@(#)$Id: kthread.h,v 1.4 2004/05/06 14:24:08 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/kthread.h,v $*/
+
+typedef struct ep_kthread
+{
+       kcondvar_t      wait;                                   /* place to sleep */
+       spinlock_t      lock;                                   /* and lock */
+       long            next_run;                               /* tick when thread should next run */
+       long            running;                                /* tick when thread started to run */
+       unsigned short  should_stall;
+       unsigned char   state;
+       unsigned int    started:1;
+       unsigned int    should_stop:1;
+       unsigned int    stopped:1;
+} EP_KTHREAD;
+
+#define KT_STATE_SLEEPING              0
+#define KT_STATE_SCHEDULED             1
+#define KT_STATE_RUNNING               2
+#define KT_STATE_STALLED               3
+
+#define AFTER(a, b)                    ((((long)(a)) - ((long)(b))) > 0)
+#define BEFORE(a,b)                    ((((long)(a)) - ((long)(b))) < 0)
+
+extern void ep_kthread_init (EP_KTHREAD *kt);
+extern void ep_kthread_destroy (EP_KTHREAD *kt);
+extern void ep_kthread_started (EP_KTHREAD *kt);
+extern void ep_kthread_stopped (EP_KTHREAD *kt);
+extern int  ep_kthread_should_stall (EP_KTHREAD *kth);
+extern int  ep_kthread_sleep (EP_KTHREAD *kth, long next_run);
+extern void ep_kthread_schedule (EP_KTHREAD *kt, long when);
+extern void ep_kthread_stall (EP_KTHREAD *kth);
+extern void ep_kthread_resume (EP_KTHREAD *kt);
+extern void ep_kthread_stop (EP_KTHREAD *kt);
+extern int  ep_kthread_state (EP_KTHREAD *kt, long *time);
+#endif /* __ELAN3_KTHREAD_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/Makefile     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/Makefile  2005-06-01 23:12:54.673427856 -0400
@@ -0,0 +1,33 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/ep/Makefile
+#
+
+
+ep3-$(CONFIG_ELAN3)    := kcomm_elan3.o kmsg_elan3.o kmap_elan3.o neterr_elan3.o probenetwork_elan3.o support_elan3.o threadcode_elan3.o threadcode_elan3_Linux.o epcomms_elan3.o epcommsTx_elan3.o epcommsRx_elan3.o
+ep4-$(CONFIG_ELAN4)    := kcomm_elan4.o kmsg_elan4.o kmap_elan4.o neterr_elan4.o probenetwork_elan4.o commands_elan4.o debug_elan4.o support_elan4.o threadcode_elan4_Linux.o epcomms_elan4.o epcommsTx_elan4.o epcommsRx_elan4.o
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/ep/Makefile
+#
+
+list-multi             := ep.o
+ep-objs        := cm.o debug.o kalloc.o kcomm.o kmap.o kthread.o neterr.o nmh.o probenetwork.o railhints.o rmap.o statemap.o support.o threadcode.o epcomms.o epcommsRx.o epcommsTx.o epcommsFwd.o conf_linux.o procfs_linux.o ep_procfs.o cm_procfs.o $(ep3-$(CONFIG_EP)) $(ep4-$(CONFIG_EP))
+export-objs            := conf_linux.o
+obj-$(CONFIG_EP)       := ep.o
+
+ep.o : $(ep-objs)
+       $(LD) -r -o $@ $(ep-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/ep/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/Makefile.conf        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/Makefile.conf     2005-06-01 23:12:54.673427856 -0400
@@ -0,0 +1,12 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       ep.o
+MODULENAME     =       ep
+KOBJFILES      =       cm.o debug.o kalloc.o kcomm.o kmap.o kthread.o neterr.o nmh.o probenetwork.o railhints.o rmap.o statemap.o support.o threadcode.o epcomms.o epcommsRx.o epcommsTx.o epcommsFwd.o conf_linux.o procfs_linux.o ep_procfs.o cm_procfs.o \$\(ep3-\$\(CONFIG_EP\)\) \$\(ep4-\$\(CONFIG_EP\)\)
+EXPORT_KOBJS   =       conf_linux.o
+CONFIG_NAME    =       CONFIG_EP
+SGALFC         =       
+# EXTRALINES START
+
+ep3-$(CONFIG_ELAN3)    := kcomm_elan3.o kmsg_elan3.o kmap_elan3.o neterr_elan3.o probenetwork_elan3.o support_elan3.o threadcode_elan3.o threadcode_elan3_Linux.o epcomms_elan3.o epcommsTx_elan3.o epcommsRx_elan3.o
+ep4-$(CONFIG_ELAN4)    := kcomm_elan4.o kmsg_elan4.o kmap_elan4.o neterr_elan4.o probenetwork_elan4.o commands_elan4.o debug_elan4.o support_elan4.o threadcode_elan4_Linux.o epcomms_elan4.o epcommsTx_elan4.o epcommsRx_elan4.o
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/ep/neterr.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/neterr.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/neterr.c  2005-06-01 23:12:54.674427704 -0400
@@ -0,0 +1,82 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: neterr.c,v 1.25.8.1 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/neterr.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <elan/kcomm.h>
+
+#include "debug.h"
+
+void
+ep_queue_network_error (EP_RAIL *rail, int nodeId, int what, int channel, EP_NETERR_COOKIE cookie)
+{
+    EP_SYS       *sys      = rail->System;
+    EP_NODE_RAIL *nodeRail = &rail->Nodes[nodeId];
+    unsigned long flags;
+
+    spin_lock_irqsave (&sys->NodeLock, flags);
+
+    ASSERT (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE);
+    
+    if (nodeRail->NetworkErrorState == 0)
+    {
+       EPRINTF2 (DBG_NETWORK_ERROR, "%s: raise context filter for node %d due to network error\n", rail->Name, nodeId);
+       printk ("%s: raise context filter for node %d due to network error\n", rail->Name, nodeId);
+       
+       rail->Operations.RaiseFilter (rail, nodeId);
+       
+       if (nodeRail->State == EP_NODE_LOCAL_PASSIVATE)
+           printk ("%s: node %d is flushing - deferring network error fixup\n", rail->Name, nodeId);
+       else
+           list_add_tail (&nodeRail->Link, &rail->NetworkErrorList);
+    }
+    
+    switch (what)
+    {
+    case EP_NODE_NETERR_ATOMIC_PACKET:
+       ASSERT (nodeRail->NetworkErrorCookies[channel] == 0);
+       
+       /* Need to raise the approriate context filter for this node,
+        * and periodically send a neterr fixup message to it until 
+        * we receive an ack from it
+        */
+       IncrStat (rail, NeterrAtomicPacket);
+       
+       nodeRail->NetworkErrorCookies[channel] = cookie;
+       
+       nodeRail->NetworkErrorState |= EP_NODE_NETERR_ATOMIC_PACKET;
+       nodeRail->MsgXid             = ep_xid_cache_alloc (sys, &rail->XidCache);
+       
+       EPRINTF3 (DBG_NETWORK_ERROR, "%s: atomic packet destroyed - node %d cookie %llx\n", rail->Name, nodeId, cookie);
+
+       printk ("%s: atomic packet destroyed - node %d cookie %llx\n", rail->Name, nodeId, cookie);
+       break;
+
+    case EP_NODE_NETERR_DMA_PACKET:
+       /* Must be an overlapped dma packet, raise the context filter,
+        * and hold it up for a NETWORK_ERROR_TIMEOUT */
+       IncrStat (rail, NeterrDmaPacket);
+       
+       nodeRail->NetworkErrorState |= EP_NODE_NETERR_DMA_PACKET;
+       break;
+    }
+
+    nodeRail->NextRunTime = lbolt + NETWORK_ERROR_TIMEOUT;
+    
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+
+    ep_kthread_schedule (&sys->ManagerThread, nodeRail->NextRunTime);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+
Index: linux-2.4.21/drivers/net/qsnet/ep/neterr_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/neterr_elan3.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/neterr_elan3.c    2005-06-01 23:12:54.674427704 -0400
@@ -0,0 +1,326 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: neterr_elan3.c,v 1.24 2003/11/17 13:26:45 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/neterr_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "debug.h"
+
+typedef struct neterr_halt_args
+{
+    EP3_RAIL        *Rail;
+    unsigned int      NodeId;
+    EP_NETERR_COOKIE *Cookies;
+} NETERR_HALT_ARGS;
+
+static int
+DmaMatchesCookie (EP3_RAIL *rail, E3_DMA_BE *dma, int nodeId, EP_NETERR_COOKIE *cookies, char *where)
+{
+    E3_uint32     cvproc;
+    E3_uint32     cookie;
+    
+    if (dma->s.dma_direction == DMA_WRITE)
+    {
+       cvproc = dma->s.dma_destCookieVProc;
+       cookie = dma->s.dma_srcCookieVProc;
+    }
+    else
+    {
+       cvproc = dma->s.dma_srcCookieVProc;
+       cookie = dma->s.dma_destCookieVProc;
+    }
+    
+    EPRINTF6 (DBG_NETWORK_ERROR, "%s: Neterr - %s: DMA %08x %08x %08x %08x\n", rail->Generic.Name, where,
+             dma->s.dma_type, dma->s.dma_size, dma->s.dma_source, dma->s.dma_dest);
+    EPRINTF5 (DBG_NETWORK_ERROR, "%s:                     %08x %08x %08x %08x\n", rail->Generic.Name,
+             dma->s.dma_destEvent, dma->s.dma_destCookieVProc, dma->s.dma_srcEvent, dma->s.dma_srcCookieVProc);
+
+    if (EP_VP_ISDATA((cvproc & DMA_PROCESS_MASK)) && EP_VP_TO_NODE(cvproc & DMA_PROCESS_MASK) == nodeId)
+    {
+       /*
+        * This is a DMA going to the node which has a network fixup
+        * request pending, so check if the cookie matches.
+        */
+       if ((cookie == cookies[0] || cookie == cookies[1]) /* && !WaitForEop */)
+       {
+           EPRINTF3 (DBG_NETWORK_ERROR, "%s: match cookie %08x on %s\n", rail->Generic.Name, cookie, where);
+           
+           return (TRUE);
+       }
+    }
+
+    return (FALSE);
+}
+
+
+static void
+NetworkErrorHaltOperation (ELAN3_DEV *dev, void *arg)
+{
+    NETERR_HALT_ARGS *args = (NETERR_HALT_ARGS *) arg;
+    EP3_RAIL         *rail = args->Rail;
+    EP_SYS           *sys  = rail->Generic.System;
+    sdramaddr_t       FPtr, BPtr;
+    sdramaddr_t       Base, Top;
+    E3_DMA_BE         dma;
+    unsigned long     flags;
+
+    spin_lock_irqsave (&sys->NodeLock, flags);
+
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProc.s.FSR)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData0.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData1.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData2.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData3.s.FSR.Status)) == 0);
+    
+    FPtr  = read_reg32 (dev, DProc_SysCntx_FPtr);
+    BPtr =  read_reg32 (dev, DProc_SysCntx_BPtr);
+    Base  = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[0]);
+    Top   = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[E3_SysCntxQueueSize-1]);
+    
+    while (FPtr != BPtr)
+    {
+       elan3_sdram_copyq_from_sdram (dev, FPtr, &dma, sizeof (E3_DMA_BE));
+       
+       if (DmaMatchesCookie (rail, &dma, args->NodeId, args->Cookies, "runq "))
+       {
+           /*
+            * Transfer the DMA to the node, it's source event will 
+            * get executed later.
+            */
+           QueueDmaOnStalledList (rail, &dma);
+           
+           /*
+            * Remove the DMA from the queue by replacing it with one with
+            * zero size and no events.
+            *
+            * NOTE: we must preserve the SYS_CONTEXT_BIT since the Elan uses this
+            * to mark the approriate run queue as empty.
+            */
+           dma.s.dma_type            = (SYS_CONTEXT_BIT << 16);
+           dma.s.dma_size            = 0;
+           dma.s.dma_source          = (E3_Addr) 0;
+           dma.s.dma_dest            = (E3_Addr) 0;
+           dma.s.dma_destEvent       = (E3_Addr) 0;
+           dma.s.dma_destCookieVProc = 0;
+           dma.s.dma_srcEvent        = (E3_Addr) 0;
+           dma.s.dma_srcCookieVProc  = 0;
+           
+           elan3_sdram_copyq_to_sdram (dev, &dma, FPtr, sizeof (E3_DMA_BE));
+       }
+
+       FPtr = (FPtr == Top) ? Base : FPtr + sizeof (E3_DMA);
+    }
+
+    rail->NetworkErrorFlushed = TRUE;
+    kcondvar_wakeupall (&rail->NetworkErrorSleep, &sys->NodeLock);
+
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+}
+
+void
+ep3_neterr_fixup (EP_RAIL *r, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP3_RAIL        *rail        = (EP3_RAIL *) r;
+    EP_SYS          *sys         = rail->Generic.System;
+    ELAN3_DEV       *dev         = rail->Device;
+    EP_NODE_RAIL    *nodeRail    = &rail->Generic.Nodes[nodeId];
+    E3_DMA_BE        dmabe;
+    EP3_COOKIE      *cp;
+    E3_uint32        vp;
+    NETERR_HALT_ARGS args;
+    struct list_head *el, *nel, matchedList;
+    int              i;
+    unsigned long    flags;
+
+    INIT_LIST_HEAD (&matchedList);
+
+    StallDmaRetryThread (rail);
+
+    args.Rail       = rail;
+    args.NodeId     = nodeId;
+    args.Cookies    = cookies;
+
+    spin_lock_irqsave (&rail->Device->IntrLock, flags);
+    QueueHaltOperation (rail->Device, 0, NULL, INT_TProcHalted | INT_DProcHalted, NetworkErrorHaltOperation, &args);
+    spin_unlock_irqrestore (&rail->Device->IntrLock, flags);
+    
+    spin_lock_irqsave (&sys->NodeLock, flags);
+    while (! rail->NetworkErrorFlushed)
+       kcondvar_wait (&rail->NetworkErrorSleep, &sys->NodeLock, &flags);
+    rail->NetworkErrorFlushed = FALSE;
+    
+    spin_lock (&rail->DmaRetryLock);
+    for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+    {
+       list_for_each_safe (el, nel, &rail->DmaRetries[i]) {
+           EP3_RETRY_DMA *retry = list_entry (el, EP3_RETRY_DMA, Link);
+
+           if (DmaMatchesCookie (rail, &retry->Dma, nodeId, cookies, "retry"))
+           {
+               /* remove from retry list */
+               list_del (&retry->Link);
+
+               /* add to list of dmas which matched */
+               list_add_tail (&retry->Link, &matchedList);
+           }
+       }
+    }
+    
+    list_for_each_safe (el, nel, &nodeRail->StalledDmas) {
+       EP3_RETRY_DMA *retry = list_entry (el, EP3_RETRY_DMA, Link);
+       
+       if (DmaMatchesCookie (rail, &retry->Dma, nodeId, cookies, "stalled"))
+       {
+           /* remove from retry list */
+           list_del (&retry->Link);
+           
+           /* add to list of dmas which matched */
+           list_add_tail (&retry->Link, &matchedList);
+       }
+    }
+    
+    spin_unlock (&rail->DmaRetryLock);
+    spin_unlock_irqrestore (&sys->NodeLock, flags);
+    
+    ResumeDmaRetryThread (rail);
+
+    /* Now "set" the source event of any write DMA's */
+    while (! list_empty (&matchedList))
+    {
+       EP3_RETRY_DMA *retry = list_entry (matchedList.next, EP3_RETRY_DMA, Link);
+       
+       list_del (&retry->Link);
+
+       if (retry->Dma.s.dma_direction == DMA_WRITE && retry->Dma.s.dma_srcEvent)
+       {
+           sdramaddr_t event = ep_elan2sdram (&rail->Generic, retry->Dma.s.dma_srcEvent);
+
+           /* Block local interrupts, since we need to atomically
+            * decrement the event count and perform the word write
+            */
+           local_irq_save (flags);
+           {
+               E3_uint32 type  = elan3_sdram_readl (dev, event + offsetof (E3_Event, ev_Type));
+               E3_uint32 count = elan3_sdram_readl (dev, event + offsetof (E3_Event, ev_Count));
+
+               elan3_sdram_writel (dev, event + offsetof (E3_Event, ev_Count), count - 1);
+
+               if (count == 1)
+               {
+                   if (type & EV_TYPE_MASK_BCOPY)
+                   {
+                       E3_Addr srcVal  = elan3_sdram_readl (dev, event + offsetof (E3_BlockCopyEvent, ev_Source));
+                       E3_Addr dstAddr = elan3_sdram_readl (dev, event + offsetof (E3_BlockCopyEvent, ev_Dest)) & ~EV_BCOPY_DTYPE_MASK;
+
+                       ASSERT ((srcVal & EV_WCOPY) != 0);
+                       
+                       EPRINTF3 (DBG_NETWORK_ERROR, "%s: neterr perform event word write at %08x with %08x\n", rail->Generic.Name, dstAddr, srcVal);
+
+                       ELAN3_OP_STORE32 (rail->Ctxt, dstAddr, srcVal);
+                   }
+
+                   if ((type & ~EV_TYPE_MASK_BCOPY) != 0)
+                   {
+                       if ((type & EV_TYPE_MASK_CHAIN) == EV_TYPE_CHAIN)
+                       {
+                           printk ("%s: event at %08x - chained event %x is invalid\n", rail->Generic.Name, retry->Dma.s.dma_srcEvent, type);
+                           panic ("ep: neterr invalid event type\n");
+                       }
+                       else if ((type & EV_TYPE_MASK_EVIRQ) == EV_TYPE_EVIRQ)
+                       {
+                           EPRINTF2 (DBG_NETWORK_ERROR, "%s: neterr event interrupt - cookie %08x\n", rail->Generic.Name, (type & ~(EV_TYPE_MASK_EVIRQ|EV_TYPE_MASK_BCOPY)));
+                           
+                           cp = LookupCookie (&rail->CookieTable, (type & ~(EV_TYPE_MASK_EVIRQ|EV_TYPE_MASK_BCOPY)));
+                           
+                           if (cp->Operations->Event)
+                               cp->Operations->Event(rail, cp->Arg);
+                       }
+                       else if ((type & EV_TYPE_MASK_DMA) == EV_TYPE_DMA)
+                       {
+                           sdramaddr_t dma = ep_elan2sdram (&rail->Generic, (type & ~EV_TYPE_MASK2));
+                           
+                           EPRINTF2 (DBG_NETWORK_ERROR, "%s: neterr chained dma - %08x\n", rail->Generic.Name, (type & ~EV_TYPE_MASK2));
+                           
+                           elan3_sdram_copyq_from_sdram (dev, dma, &dmabe, sizeof (E3_DMA));
+                           
+                           if (dmabe.s.dma_direction == DMA_WRITE)
+                           {
+                               vp = dmabe.s.dma_destVProc;
+                               cp = LookupEventCookie (rail, &rail->CookieTable, dmabe.s.dma_srcEvent);
+                           }
+                           else
+                           {
+                               vp = dmabe.s.dma_srcVProc;
+                               cp = LookupEventCookie (rail, &rail->CookieTable, dmabe.s.dma_destEvent);
+                               
+                               /* we MUST convert this into a DMA_READ_REQUEUE dma as if we don't the 
+                                * DMA descriptor will be read from the EP_RETRY_DMA rather than the 
+                                * original DMA - this can then get reused and an incorrect DMA 
+                                * descriptor sent 
+                                * eventp->ev_Type contains the dma address with type in the lower bits 
+                                */ 
+                           
+                               dmabe.s.dma_source    = (type & ~EV_TYPE_MASK2);
+                               dmabe.s.dma_direction = (dmabe.s.dma_direction & ~DMA_READ) | DMA_READ_REQUEUE;
+                           }
+                       
+                           ASSERT (EP_VP_ISDATA(vp));
+                       
+                           nodeRail = &rail->Generic.Nodes[EP_VP_TO_NODE(vp)];
+
+                           switch (nodeRail->State)
+                           {
+                           case EP_NODE_CONNECTED:
+                           case EP_NODE_LEAVING_CONNECTED:
+                               if (cp != NULL)
+                                   cp->Operations->DmaRetry (rail, cp->Arg, &dmabe, EAGAIN);
+                               else
+                               {
+                                   ASSERT (dmabe.s.dma_direction == DMA_WRITE && dmabe.s.dma_srcEvent == 0 && dmabe.s.dma_isRemote);
+                               
+                                   QueueDmaForRetry (rail, &dmabe, EP_RETRY_ANONYMOUS);
+                               }
+                               break;
+
+                           case EP_NODE_LOCAL_PASSIVATE:
+                               QueueDmaOnStalledList (rail, &dmabe);
+                               break;
+
+                           default:
+                               panic ("ep: neterr incorrect state for node\n");
+                           }
+                       }
+                       else if ((type & EV_TYPE_MASK_THREAD) == EV_TYPE_THREAD)
+                       {
+                           printk ("%s: event at %08x - thread waiting %x is invalid\n", rail->Generic.Name, retry->Dma.s.dma_srcEvent, type);
+                           panic ("ep: neterr invalid event type\n");
+                       }
+                   }
+               }
+           }
+           local_irq_restore(flags);
+       }
+       
+       /* add to free list */
+       spin_lock_irqsave (&rail->DmaRetryLock, flags);
+       list_add (&retry->Link, &rail->DmaRetryFreeList);
+       spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+    }
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+
Index: linux-2.4.21/drivers/net/qsnet/ep/neterr_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/neterr_elan4.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/neterr_elan4.c    2005-06-01 23:12:54.675427552 -0400
@@ -0,0 +1,251 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: neterr_elan4.c,v 1.2 2003/11/24 17:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/neterr_elan4.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "debug.h"
+
+struct neterr_desc
+{
+    EP4_RAIL         *rail;
+    unsigned int      nodeid;
+    EP_NETERR_COOKIE *cookies;
+    int                      done;
+} ;
+
+static int
+dma_matches_cookie (EP4_RAIL *rail, E4_uint64 vproc, E4_uint64 cookie, unsigned int nodeId, EP_NETERR_COOKIE *cookies, const char *where)
+{
+    if ((EP_VP_ISDATA (vproc) && EP_VP_TO_NODE (vproc) == nodeId) && (cookie == cookies[0] || cookie == cookies[1]))
+    {
+       EPRINTF3 (DBG_NETWORK_ERROR, "%s: match cookie %016llx on %s\n", rail->r_generic.Name, cookie, where);
+
+       return 1;
+    }
+    return 0;
+}
+
+static void
+ep4_neterr_dma_flushop (ELAN4_DEV *dev, void *arg, int qfull)
+{
+    struct neterr_desc *desc  = (struct neterr_desc *) arg;
+    EP4_RAIL           *rail  = desc->rail;
+    E4_uint64           qptrs = read_reg64 (dev, DProcHighPriPtrs);
+    E4_uint32           qsize = E4_QueueSize (E4_QueueSizeValue (qptrs));
+    E4_uint32           qfptr = E4_QueueFrontPointer (qptrs);
+    E4_uint32           qbptr = E4_QueueBackPointer (qptrs);
+    E4_DProcQueueEntry  qentry;
+    unsigned long       flags;
+
+    while ((qfptr != qbptr) || qfull)
+    {
+       E4_uint64 cookie = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_cookie));
+       E4_uint64 vproc  = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_vproc));
+
+       if (dma_matches_cookie (rail, vproc, cookie, desc->nodeid, desc->cookies, "runq "))
+       {
+           elan4_sdram_copyq_from_sdram (dev, qfptr, &qentry, sizeof (E4_DProcQueueEntry));
+
+           ep4_queue_dma_stalled (rail, &qentry.Desc);
+
+           /* Replace the dma with one which will "disappear" */
+           qentry.Desc.dma_typeSize = DMA_ShMemWrite | dev->dev_ctxt.ctxt_num;
+           qentry.Desc.dma_cookie   = 0;
+           qentry.Desc.dma_vproc    = 0;
+           qentry.Desc.dma_srcAddr  = 0;
+           qentry.Desc.dma_dstAddr  = 0;
+           qentry.Desc.dma_srcEvent = 0;
+           qentry.Desc.dma_dstEvent = 0;
+
+           elan4_sdram_copyq_to_sdram (dev, &qentry, qfptr, sizeof (E4_DProcQueueEntry));
+       }
+       
+       qfptr = (qfptr & ~(qsize-1)) | ((qfptr + sizeof (E4_DProcQueueEntry)) & (qsize-1));
+       qfull = 0;
+    }
+
+    spin_lock_irqsave (&rail->r_haltop_lock, flags);
+    desc->done = 1;
+    kcondvar_wakeupall (&rail->r_haltop_sleep, &rail->r_haltop_lock);
+    spin_unlock_irqrestore (&rail->r_haltop_lock, flags);
+}
+
+static void
+ep4_neterr_dma_haltop (ELAN4_DEV *dev, void *arg)
+{
+    struct neterr_desc *desc = (struct neterr_desc *) arg;
+
+    elan4_queue_dma_flushop (dev, &desc->rail->r_flushop, 1);
+}
+
+void
+ep4_neterr_fixup_dmas (EP4_RAIL *rail, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP_NODE_RAIL      *nodeRail = &rail->r_generic.Nodes[nodeId];
+    struct neterr_desc desc;
+    struct list_head   matchedList;
+    struct list_head  *el, *nel;
+    unsigned long      flags;
+    register int       i;
+
+    desc.rail    = rail;
+    desc.nodeid  = nodeId;
+    desc.cookies = cookies;
+    desc.done    = 0;
+
+    INIT_LIST_HEAD (&matchedList);
+
+    /* First -  stall the retry thread, so that it will no longer restart
+     *          any dma's from the retry list */
+    ep_kthread_stall (&rail->r_retry_thread);
+    
+    /* Second - flush through all command queues targetted by events, thread etc */
+    ep4_flush_ecqs (rail);
+    
+    /* Third - queue a halt operation to flush through all DMA's which are executing
+     *         or on the run queues */
+    kmutex_lock (&rail->r_haltop_mutex);
+    
+    rail->r_haltop.op_mask      = INT_DProcHalted;
+    rail->r_haltop.op_function  = ep4_neterr_dma_haltop;
+    rail->r_haltop.op_arg       = &desc;
+
+    rail->r_flushop.op_function = ep4_neterr_dma_flushop;
+    rail->r_flushop.op_arg      = &desc;
+    
+    elan4_queue_haltop (rail->r_ctxt.ctxt_dev, &rail->r_haltop);
+
+    spin_lock_irqsave (&rail->r_haltop_lock, flags);
+    while (! desc.done)
+       kcondvar_wait (&rail->r_haltop_sleep, &rail->r_haltop_lock, &flags);
+    spin_unlock_irqrestore (&rail->r_haltop_lock, flags);
+    kmutex_unlock (&rail->r_haltop_mutex);
+
+    /* Fourth - run down the dma retry lists and move all entries to the cancelled
+     *          list.  Any dma's which were on the run queues have already been
+     *          moved there */
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+    for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+    {
+       list_for_each_safe (el,nel, &rail->r_dma_retrylist[i]) {
+           EP4_DMA_RETRY *retry    = list_entry (el, EP4_DMA_RETRY, retry_link);
+           
+           if (dma_matches_cookie (rail, retry->retry_dma.dma_vproc, retry->retry_dma.dma_cookie, nodeId, cookies, "retry"))
+           {
+               /* remove from retry list */
+               list_del (&retry->retry_link);
+               
+               /* add to list of dmas which matched */
+               list_add_tail (&retry->retry_link, &matchedList);
+           }
+       }
+    }
+    
+    list_for_each_safe (el, nel, &nodeRail->StalledDmas) {
+       EP4_DMA_RETRY *retry = list_entry (el, EP4_DMA_RETRY, retry_link);
+       
+       if (dma_matches_cookie (rail, retry->retry_dma.dma_vproc, retry->retry_dma.dma_cookie, nodeId, cookies, "stalled"))
+       {
+           /* remove from retry list */
+           list_del (&retry->retry_link);
+           
+           /* add to list of dmas which matched */
+           list_add_tail (&retry->retry_link, &matchedList);
+       }
+    }
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+    
+    /* Now "set" the source event of any put DMA#'s we can use the dma 
+     * retry command queue as the retry thread is stalled */
+    while (! list_empty (&matchedList))
+    {
+       EP4_DMA_RETRY *retry = list_entry (matchedList.next, EP4_DMA_RETRY, retry_link);
+       
+       list_del (&retry->retry_link);
+
+       elan4_set_event_cmd (rail->r_dma_ecq->ecq_cq, retry->retry_dma.dma_srcEvent);
+
+       spin_lock_irqsave (&rail->r_dma_lock, flags);
+       list_add (&retry->retry_link, &rail->r_dma_freelist);
+       spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+    }
+
+    /* Flush through the command queues to ensure that all the setevents have executed */
+    ep4_flush_ecqs (rail);
+
+    /* Finally - allow the retry thread to run again */
+    ep_kthread_resume (&rail->r_retry_thread);
+}
+
+void
+ep4_add_neterr_ops (EP4_RAIL *rail, EP4_NETERR_OPS *ops)
+{
+    /* we're called from the ManagerThread, so no need to stall it */
+    list_add_tail (&ops->op_link, &rail->r_neterr_ops);
+}
+void
+ep4_remove_neterr_ops (EP4_RAIL *rail, EP4_NETERR_OPS *ops)
+{
+    EP_SYS *sys = rail->r_generic.System;
+
+    ep_kthread_stall (&sys->ManagerThread);
+    list_del (&ops->op_link);
+    ep_kthread_resume (&sys->ManagerThread);
+}
+
+void
+ep4_neterr_fixup_sten (EP4_RAIL *rail, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    struct list_head *el;
+
+    list_for_each (el, &rail->r_neterr_ops) {
+       EP4_NETERR_OPS *op = list_entry (el, EP4_NETERR_OPS, op_link);
+
+       (op->op_func) (rail, op->op_arg, nodeId, cookies);
+    }
+}
+
+void
+ep4_neterr_fixup (EP_RAIL *r, unsigned int nodeId, EP_NETERR_COOKIE *cookies)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+
+    /* network error cookies can come from the following :
+     *
+     *   DMA  engine
+     *     if a DMA matches a network error cookie, then we just need to 
+     *     execute the local setevent *before* returning.
+     *
+     *   STEN packet
+     *     if the STEN packet was generated with as a WAIT_FOR_EOP
+     *     and it's not present on the retry lists, then re-create
+     *     it.
+     *
+     */
+    EPRINTF4 (DBG_NETWORK_ERROR, "%s: ep4_neterr_fixup: node %d cookies <%lld%s%s%s%s> <%lld%s%s%s%s>\n",
+             rail->r_generic.Name, nodeId, EP4_COOKIE_STRING(cookies[0]), EP4_COOKIE_STRING(cookies[1]));
+
+    if ((cookies[0] & EP4_COOKIE_DMA) || (cookies[1] & EP4_COOKIE_DMA))
+       ep4_neterr_fixup_dmas (rail, nodeId, cookies);
+
+    if ((cookies[0] & EP4_COOKIE_STEN) || (cookies[1] & EP4_COOKIE_STEN))
+       ep4_neterr_fixup_sten (rail, nodeId, cookies);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+
Index: linux-2.4.21/drivers/net/qsnet/ep/nmh.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/nmh.c        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/nmh.c     2005-06-01 23:12:54.676427400 -0400
@@ -0,0 +1,181 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+#ident "@(#)$Id: nmh.c,v 1.6 2004/01/05 13:48:08 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/nmh.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#define EP_NMD_SPANS(nmd, base, top)   ((nmd)->nmd_addr <= (base) &&  \
+                                        ((nmd)->nmd_addr + (nmd)->nmd_len - 1) >= (top))
+
+#define EP_NMD_OVERLAPS(nmd, addr, len)        ((nmd)->nmd_addr <= ((addr) + (len)) && \
+                                        ((nmd)->nmd_addr + (nmd)->nmd_len - 1) >= (addr))
+
+#define EP_NMH_HASH(tbl,idx,addr)      ((addr) % (tbl)->tbl_size[idx])
+
+int
+ep_nmh_init (EP_NMH_TABLE *tbl)
+{
+    int i, idx, hsize = 1;
+
+    for (idx = EP_NMH_NUMHASH-1; idx >= 0; idx--, hsize <<= 1)
+    {
+       tbl->tbl_size[idx] = (hsize < EP_NMH_HASHSIZE) ? hsize : EP_NMH_HASHSIZE;
+
+       KMEM_ZALLOC (tbl->tbl_hash[idx], struct list_head *, sizeof (struct list_head) * tbl->tbl_size[idx], 1);
+       
+       if (tbl->tbl_hash == NULL)
+       {
+           while (++idx < EP_NMH_NUMHASH)
+               KMEM_FREE (tbl->tbl_hash[idx], sizeof (struct list_head) * tbl->tbl_size[idx]);
+           return (ENOMEM);
+       }
+
+       for (i = 0; i < tbl->tbl_size[idx]; i++)
+           INIT_LIST_HEAD (&tbl->tbl_hash[idx][i]);
+    }
+
+    return (0);
+}
+
+void
+ep_nmh_fini (EP_NMH_TABLE *tbl)
+{
+    int idx;
+
+    for (idx = 0; idx < EP_NMH_NUMHASH; idx++)
+       if (tbl->tbl_hash[idx])
+           KMEM_FREE (tbl->tbl_hash[idx], sizeof (struct list_head) * tbl->tbl_size[idx]);
+    
+    bzero (tbl, sizeof (EP_NMH_TABLE));
+}
+
+void
+ep_nmh_insert (EP_NMH_TABLE *tbl, EP_NMH *nmh)
+{
+    EP_ADDR base = nmh->nmh_nmd.nmd_addr;
+    EP_ADDR top  = base + nmh->nmh_nmd.nmd_len - 1;
+    int     idx;
+
+    for (idx = 0, base >>= 12, top >>= 12; base != top && idx < EP_NMH_NUMHASH; idx++, base >>= 1, top >>= 1)
+       ;
+
+    list_add_tail (&nmh->nmh_link, &tbl->tbl_hash[idx][EP_NMH_HASH(tbl, idx, base)]);
+}
+
+void
+ep_nmh_remove (EP_NMH_TABLE *tbl, EP_NMH *nmh)
+{
+    list_del (&nmh->nmh_link);
+}
+
+EP_NMH *
+ep_nmh_find (EP_NMH_TABLE *tbl, EP_NMD *nmd)
+{
+    EP_ADDR           base = nmd->nmd_addr;
+    EP_ADDR           top  = base + nmd->nmd_len - 1;
+    int               idx;
+    struct list_head *le;
+    
+    for (idx = 0, base >>= 12, top >>= 12; base != top && idx < EP_NMH_NUMHASH; idx++, base >>= 1, top >>= 1)
+       ;
+    
+    for (; idx < EP_NMH_NUMHASH; idx++, base >>= 1, top >>= 1) {
+
+       list_for_each (le, &tbl->tbl_hash[idx][EP_NMH_HASH(tbl, idx, base)]) {
+           EP_NMH *nmh = list_entry (le, EP_NMH, nmh_link);
+
+           if (EP_NMD_SPANS (&nmh->nmh_nmd, nmd->nmd_addr, nmd->nmd_addr + nmd->nmd_len - 1))
+               return (nmh);
+       }
+    }
+
+    return (0);
+}
+
+void
+ep_nmd_subset (EP_NMD *subset, EP_NMD *nmd, unsigned off, unsigned len)
+{
+    ASSERT ((off + len - 1) <= nmd->nmd_len);
+
+    subset->nmd_addr = nmd->nmd_addr + off;
+    subset->nmd_len  = len;
+    subset->nmd_attr = nmd->nmd_attr;
+}
+
+int
+ep_nmd_merge (EP_NMD *merged, EP_NMD *a, EP_NMD *b)
+{
+    if (EP_NMD_NODEID (a) != EP_NMD_NODEID (b))                        /* not generated on the same node */
+       return 0;
+    
+    if ((EP_NMD_RAILMASK (a) & EP_NMD_RAILMASK (b)) == 0)      /* no common rails */
+       return 0;
+    
+    if (b->nmd_addr == (a->nmd_addr + a->nmd_len))
+    {
+       if (merged != NULL)
+       {
+           merged->nmd_addr = a->nmd_addr;
+           merged->nmd_len  = a->nmd_len + b->nmd_len;
+           merged->nmd_attr = EP_NMD_ATTR(EP_NMD_NODEID(a), EP_NMD_RAILMASK(a) & EP_NMD_RAILMASK(b));
+       }
+       return 1;
+    }
+    
+    if (a->nmd_addr == (b->nmd_addr + b->nmd_len))
+    {
+       if (merged != NULL)
+       {
+           merged->nmd_addr = b->nmd_addr;
+           merged->nmd_len  = b->nmd_len + a->nmd_len;
+           merged->nmd_attr = EP_NMD_ATTR(EP_NMD_NODEID(b), EP_NMD_RAILMASK(a) & EP_NMD_RAILMASK(b));
+       }
+       
+       return 1;
+    }
+
+    return 0;
+}
+
+int
+ep_nmd_map_rails (EP_SYS *sys, EP_NMD *nmd, unsigned railmask)
+{
+    EP_NMH *nmh = ep_nmh_find (&sys->MappingTable, nmd);
+
+    if (nmh == NULL)
+    {
+       printk ("ep_nmd_map_rails: nmd=%08x.%08x.%08x cannot be found\n",
+               nmd->nmd_addr, nmd->nmd_len, nmd->nmd_attr);
+       return (-1);
+    }
+
+    return (nmh->nmh_ops->op_map_rails (sys, nmh, nmd, railmask));
+}
+
+EP_RAILMASK
+ep_nmd2railmask (EP_NMD *frags, int nFrags)
+{
+    EP_RAILMASK mask;
+
+    if (nFrags == 0)
+       return ((EP_RAILMASK)-1);
+    
+    for (mask = EP_NMD_RAILMASK(frags); --nFrags; )
+       mask &= EP_NMD_RAILMASK(++frags);
+
+    return (mask);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/probenetwork.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/probenetwork.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/probenetwork.c    2005-06-01 23:12:54.677427248 -0400
@@ -0,0 +1,446 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: probenetwork.c,v 1.43 2004/04/19 15:43:15 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/probenetwork.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include "debug.h"
+
+int PositionCheck = 1;
+
+#define NUM_DOWN_FROM_VAL(NumDownLinksVal, level)      (((NumDownLinksVal) >> ((level) << 2)) & 0xF)
+
+int
+ProbeNetwork (EP_RAIL *rail, ELAN_POSITION *pos)
+{
+    int               lvl, i;
+    int               level;
+    int               nodeid;
+    int               numnodes;
+    int                      randomRoutingDisabled;
+    int               sw;
+    int               nacks;
+    int               nowayup;
+    int                      nalias;
+    int                      upmask;
+    int                      partial;
+    int                      link;
+    int                      invalid;
+    int                      linkdown[ELAN_MAX_LEVELS];
+    int                      linkup[ELAN_MAX_LEVELS];
+    EP_SWITCH        *switches[ELAN_MAX_LEVELS];
+    int               switchCount[ELAN_MAX_LEVELS+1];
+    int               lowestBcast;
+    int               numUpLinks[ELAN_MAX_LEVELS];
+    int               routedown [ELAN_MAX_LEVELS];
+
+    EPRINTF1 (DBG_PROBE, "%s: ProbeNetwork started\n", rail->Name);
+
+    switchCount[0] = 1;
+    numUpLinks [0] = 4;
+
+    for (level = 0; level < ELAN_MAX_LEVELS; level++)
+    {
+       int ndown  = NUM_DOWN_FROM_VAL (rail->Devinfo.dev_num_down_links_value, level);
+
+       KMEM_ZALLOC (switches[level], EP_SWITCH *, sizeof (EP_SWITCH) * switchCount[level], 1);
+
+       for (sw = 0, nacks = 0, nowayup = 0, lowestBcast=7; sw < switchCount[level]; sw++)
+       {
+           EP_SWITCH *lsw  = &switches[level][sw];
+           int        good = 1;
+           int        tsw;
+
+           for (nodeid = 0,tsw = sw, lvl = level-1 ; lvl >= 0 ; lvl--)
+           {
+               EP_SWITCH *lsw;
+               int        link = (8-numUpLinks[lvl]) + (tsw % numUpLinks[lvl]);
+
+               tsw  = tsw / numUpLinks[lvl];
+               lsw  = &switches[lvl][tsw];
+
+               if (lsw->present == 0 || (lsw->lnr & (1 << link)))
+               {
+                   EPRINTF4 (DBG_PROBE, "lvl %d sw %d present=%d lnr=%x\n", lvl, sw, lsw->present, lsw->lnr);
+                   good = 0;
+               }
+               
+               linkup[lvl]   = link;
+               linkdown[lvl] = lsw->link;
+
+               if ( lvl ) nodeid = ((nodeid + linkdown[lvl]) * (8-numUpLinks[lvl-1]));
+               else       nodeid += linkdown[0];
+
+           }
+           
+           /* 
+            * don't bother probing routes which we we've already seen are unreachable 
+            * because a link upwards was in reset or the switch previously nacked us.
+            */
+           if (! good)
+           {
+               lsw->present = 0;
+
+               nacks++;
+               nowayup++;
+
+               continue;
+           }
+
+           lsw->present = rail->Operations.ProbeRoute (rail, level, sw, nodeid, linkup, linkdown, 5, lsw);
+
+           if (! lsw->present)
+           {
+               EPRINTF3 (DBG_PROBE, "%s: level %d switch %d - unexpected nack\n", rail->Name, level, sw);
+
+               nacks++;
+               nowayup++;
+           }
+           else
+           {
+               EPRINTF5 (DBG_PROBE, "%s: level %d switch %d - link %d bcast %d\n", rail->Name, level, sw, lsw->link, lsw->bcast);
+
+               if (level == 2 && rail->Devinfo.dev_device_id == PCI_DEVICE_ID_ELAN3)
+               {
+                   /* If we see broadcast top as 7, and we came in on a low link, then we can't
+                    * determine whether we're in a 128 way or a un-configured 64u64d switch, so
+                    * we treat it as a 64u64d and detect the 128 way case by "going over the top" 
+                    * below. Unless we've been told what it really is by NumDownLinksVal.
+                    */
+                   if (lsw->bcast == 7 && lsw->link < 4)
+                       lsw->bcast = ndown ? (ndown - 1) : 3;
+               }
+
+               if ( lowestBcast > lsw->bcast ) 
+                   lowestBcast = lsw->bcast;
+
+               if (lsw->link > (ndown ? (ndown-1) : (lowestBcast == 7 ? 3 : lowestBcast)))
+               {
+                   /* We've arrived on a "up-link" - this could be either
+                    * we're in the top half of a x8 top-switch - or we're
+                    * in the bottom half and have gone "over the top". We
+                    * differentiate these cases since the switches below
+                    * a x8 top-switch will have broadcast top set to 3, 
+                    * and the x8 topswitch have broadcast top set to 7.
+                    */
+                   if (lsw->bcast == 7)
+                       nowayup++;
+                   else
+                   {
+                       EPRINTF2 (DBG_PROBE, "%s: level %d - gone over the top\n",
+                                 rail->Name, level);
+
+                       if (level > 0)
+                       {
+                           KMEM_FREE (switches[level], sizeof (EP_SWITCH) * switchCount[level] );
+                           level--;
+                       }
+                       
+                       numUpLinks[level] = 0;
+                       goto finished;
+                   }
+               }
+
+           }
+       }
+
+       numUpLinks[level]    = ndown ? (8 - ndown) : (7 - lowestBcast);
+       switchCount[level+1] = switchCount[level] *  numUpLinks[level];
+       
+       /* Now we know which links are uplinks, we can see whether there is
+        * any possible ways up */
+       upmask = (ndown ? (0xFF << ndown) & 0xFF : (0xFF << (8 - numUpLinks[level])) & 0xFF);
+
+       for (sw = 0; sw < switchCount[level]; sw++)
+       {
+           EP_SWITCH *lsw  = &switches[level][sw];
+
+           if (lsw->present && lsw->link <= (ndown ? (ndown-1) : (lowestBcast == 7 ? 3 : lowestBcast)) && (switches[level][sw].lnr & upmask) == upmask)
+               nowayup++;
+       }
+
+       EPRINTF7 (DBG_PROBE, "%s: level %d - sw=%d nacks=%d nowayup=%d bcast=%d numup=%d\n", 
+                 rail->Name, level, sw, nacks, nowayup, lowestBcast, numUpLinks[level]);
+
+       if (nacks == sw)
+       {
+           static bitmap_t printed[BT_BITOUL(EP_MAX_RAILS)];
+
+           if (! BT_TEST (printed, rail->Number))
+               printk ("%s: cannot determine network position\n", rail->Name);
+           BT_SET (printed, rail->Number);
+           goto failed;
+       }
+
+       if (nowayup == sw)
+           goto finished;
+    }
+    
+    printk ("%s: exceeded number of levels\n", rail->Name);
+    level = ELAN_MAX_LEVELS - 1;
+
+ failed:
+    
+    for (lvl = 0; lvl <= level; lvl++)
+       KMEM_FREE (switches[lvl], sizeof (EP_SWITCH) * switchCount[lvl] );
+
+    return -EAGAIN;
+
+ finished:
+    /* we've successfully probed the network - now calculate our node 
+     * positon and what level of random routing is possible */
+    nalias = 1;
+    for (lvl = 0, invalid = 0, partial = 0, randomRoutingDisabled = 0; lvl <= level; lvl++)
+    {
+       int ndown  = NUM_DOWN_FROM_VAL (rail->Devinfo.dev_num_down_links_value, lvl);
+       int upmask = ndown ? (0xFF << ndown) & 0xFF : 0xF0;
+
+       for (sw = 0, nalias = 0; sw < switchCount[lvl]; sw++)
+       {
+           EP_SWITCH *lsw = &switches[lvl][sw];
+           
+           /* You can only use adaptive routing if links 4-7 are uplinks, and at least one of them is
+            * not in reset.   Otherwise you can randomly select an "uplink" if all the uplinks are not
+            * in reset. */
+           if (lsw->present && ((upmask == 0xF0) ? (lsw->lnr & upmask) == upmask : (lsw->lnr & upmask) != 0))
+               randomRoutingDisabled |= (1 << lvl);
+           
+           if (!lsw->present)
+               partial++;
+           else
+           {
+               if (lsw->invalid)
+               {
+                   printk ("%s: invalid switch detected (level %d switch %d)\n", rail->Name, lvl, sw);
+                   invalid++;
+               }
+               
+               for (i = 0; i < nalias; i++)
+                   if (linkdown[i] == lsw->link)
+                       break;
+               if (i == nalias)
+                   linkdown[nalias++] = lsw->link;
+           }
+       }
+       
+       link = linkdown[0];
+       for (i = 1; i < nalias; i++)
+           if (linkdown[i] < link)
+               link = linkdown[i];
+
+       if (nalias > 1 && lvl != level)
+       {
+           printk ("%s: switch aliased below top level (level %d)\n", rail->Name, lvl);
+           invalid++;
+       }
+       
+       routedown[lvl] = link;
+   }
+
+    for (lvl = 0; lvl <= level; lvl++) 
+       KMEM_FREE (switches[lvl], sizeof (EP_SWITCH) * switchCount[lvl] );
+
+    if (invalid)
+    {
+       printk ("%s: invalid switch configuration\n", rail->Name);
+       return (EINVAL);
+    }
+
+    /* Handle the aliasing case where a 16 way is used as multiple smaller switches */
+    if (nalias == 1)
+       level++;
+    else if (nalias == 2)                                      /* a 16 way as 2x8 ways */
+       numUpLinks[level++] = 6;                                /*   only 2 down links */
+    else if (nalias > 4)                                       /* a 16 way as 8x2 ways */
+       numUpLinks[level-1] = 6;
+    
+    /* 
+     * Compute my nodeid and number of nodes in the machine
+     * from the routedown and the number of downlinks at each level.
+     */
+    for(nodeid=0, lvl = level - 1; lvl >= 0; lvl--)
+    {
+       if (lvl) nodeid = ((nodeid + routedown[lvl]) * (8-numUpLinks[lvl-1]));  
+       else     nodeid += routedown[0];
+    }
+
+    for (numnodes = 1, lvl = 0; lvl < level; lvl++)
+       numnodes *= (8 - numUpLinks[lvl]);
+
+    sprintf (rail->Name, "ep%d[%d]", rail->Number, nodeid);
+
+    if (randomRoutingDisabled & ((1 << (level-1))-1))
+       printk ("%s: nodeid=%d level=%d numnodes=%d (random routing disabled 0x%x)\n", 
+               rail->Name, nodeid, level, numnodes, randomRoutingDisabled);
+    else if (partial)
+       printk ("%s: nodeid=%d level=%d numnodes=%d (random routing ok)\n",
+               rail->Name, nodeid, level, numnodes);
+    else
+       printk ("%s: nodeid=%d level=%d numnodes=%d\n",
+               rail->Name, nodeid, level, numnodes);
+
+    pos->pos_mode               = ELAN_POS_MODE_SWITCHED;
+    pos->pos_nodeid              = nodeid;
+    pos->pos_levels              = level;
+    pos->pos_nodes               = numnodes;
+    pos->pos_random_disabled     = randomRoutingDisabled;
+
+    for(lvl = 0; lvl < level; lvl++)
+       pos->pos_arity[level -lvl - 1] = (8-numUpLinks[lvl]);
+    pos->pos_arity[level] = 1;                         /* XXXX why does this need to be 1 ? */
+    
+    return 0;
+}
+
+/*
+ * broadcast top is invalid if it is not set to the number of downlinks-1,
+ * or at the topmost level it is less than ndown-1.
+ */
+#define BCAST_TOP_INVALID(lvl, bcast, ndown)   ((lvl) == 0 ? (bcast) < ((ndown)-1) : (bcast) != ((ndown) - 1))
+
+void
+CheckPosition (EP_RAIL *rail)
+{
+    ELAN_POSITION *pos     = &rail->Position;
+    unsigned int   nodeid  = pos->pos_nodeid;
+    unsigned int   invalid = 0;
+    unsigned int   changed = 0;
+    int lvl, slvl;
+
+    if (! PositionCheck)
+       return;
+
+    if (rail->Operations.CheckPosition(rail))          /* is update ready for this rail */
+    {
+       EPRINTF2 (DBG_ROUTETABLE, "%s: check position: SwitchProbeLevel=%d\n", rail->Name, rail->SwitchProbeLevel);
+
+       for (lvl = 0, slvl = pos->pos_levels-1; lvl <= rail->SwitchProbeLevel; lvl++, slvl--)
+       {
+           EP_SWITCHSTATE *state  = &rail->SwitchState[lvl];
+           EP_SWITCHSTATE *lstate = &rail->SwitchLast[lvl];
+           unsigned int    ndown  = pos->pos_arity[slvl];
+           unsigned int    upmask = (0xFF << ndown) & 0xFF;
+           unsigned int    mylink = nodeid % ndown;
+           unsigned int    error  = 0;
+           unsigned int    binval = 0;
+
+           nodeid /= ndown;
+
+           /*
+            * broadcast top is invalid if it is not set to the number of downlinks-1,
+            * or at the topmost level it is less than ndown-1.
+            */
+           if (BCAST_TOP_INVALID(lvl, state->bcast, ndown) || (state->LNR & upmask) == upmask)
+           {
+               /* no way up from here - we'd better be at the top */
+               if (lvl != (pos->pos_levels-1))
+               {
+                   if (state->bcast != (ndown-1))
+                       printk ("%s: invalid broadcast top %d at level %d\n", rail->Name, state->bcast, lvl);
+                   else if ((state->LNR & upmask) == upmask && (lstate->LNR & upmask) == upmask)
+                       printk ("%s: no way up to switch at level %d (turned off ?)\n", rail->Name, lvl+1);
+               }
+               else
+               {
+                   if (state->linkid != mylink)
+                       printk ("%s: moved at top level was connected to link %d now connected to %d\n", rail->Name, mylink, state->linkid);
+               }
+
+               if (state->linkid != mylink)
+                   error++;
+               
+               if (BCAST_TOP_INVALID (lvl, state->bcast, ndown))
+                   binval++;
+           }
+           else
+           {
+               if (state->linkid != mylink)
+               {
+                   if (state->linkid != rail->SwitchLast[lvl].linkid)
+                       printk ("%s: moved at lvl %d was connected to link %d now connected to %d\n", rail->Name, lvl, mylink, state->linkid);
+                       
+                   error++;
+               }
+           }
+
+           if (error == 0 && invalid == 0)
+               rail->SwitchProbeTick[lvl] = lbolt;
+           
+           EPRINTF10 (DBG_ROUTETABLE, "%s:   lvl=%d (slvl=%d) linkid=%d bcast=%d lnr=%02x uplink=%d : error=%d binval=%d invalid=%d\n", 
+                      rail->Name, lvl, slvl, state->linkid, state->bcast, state->LNR, state->uplink, error, binval, invalid);
+
+           invalid |= (error | binval);
+       }
+       
+       for (lvl = 0; lvl < rail->SwitchProbeLevel; lvl++)
+           if (rail->SwitchState[lvl].uplink != rail->SwitchLast[lvl].uplink)
+               changed++;
+
+       if (changed)
+       {
+           printk ("%s: broadcast tree has changed from", rail->Name);
+           for (lvl = 0; lvl < rail->SwitchProbeLevel; lvl++)
+               printk ("%c%d", lvl == 0 ? ' ' : ',', rail->SwitchLast[lvl].uplink);
+
+           for (lvl = 0; lvl < rail->SwitchProbeLevel; lvl++)
+               printk ("%s%d", lvl == 0 ? " to " : ",", rail->SwitchState[lvl].uplink);
+           printk ("\n");
+       }
+
+       if (rail->SwitchProbeLevel > 0)
+           bcopy (rail->SwitchState, rail->SwitchLast, rail->SwitchProbeLevel * sizeof (EP_SWITCHSTATE));
+    }
+
+    for (lvl = 0; lvl < pos->pos_levels; lvl++)
+    {
+       EPRINTF4 (DBG_ROUTETABLE, "%s: level %d lbolt=%lx ProbeLevelTick=%lx\n",
+                 rail->Name, lvl, lbolt, rail->SwitchProbeTick[lvl]);
+       
+       if (AFTER (lbolt, rail->SwitchProbeTick[lvl] + EP_POSITION_TIMEOUT))
+       {
+           if (lvl < rail->SwitchBroadcastLevel+1)
+           {
+               if (lvl == 0)
+                   printk ("%s: cable disconnected\n", rail->Name);
+               else
+                   printk ("%s: broadcast level has dropped to %d (should be %d)\n",
+                           rail->Name, lvl, rail->Position.pos_levels);
+           }
+           break;
+       }
+    }
+    
+    if (lvl > rail->SwitchBroadcastLevel+1)
+    {
+       if (rail->SwitchBroadcastLevel < 0)
+           printk ("%s: cable reconnected\n", rail->Name);
+       if (lvl == rail->Position.pos_levels)
+           printk ("%s: broadcast level has recovered\n", rail->Name);
+       else
+           printk ("%s: broadcast level has recovered to %d (should be %d)\n", 
+                   rail->Name, lvl, rail->Position.pos_levels);
+    }
+    
+    if (rail->SwitchBroadcastLevel != (lvl - 1))
+    {
+       EPRINTF2 (DBG_ROUTETABLE, "%s: setting SwitchBroadcastLevel to %d\n", rail->Name, lvl-1);
+       
+       rail->SwitchBroadcastLevel     = lvl - 1;
+       rail->SwitchBroadcastLevelTick = lbolt;
+    }
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/probenetwork_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/probenetwork_elan3.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/probenetwork_elan3.c      2005-06-01 23:12:54.677427248 -0400
@@ -0,0 +1,298 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: probenetwork_elan3.c,v 1.40 2004/04/15 12:30:08 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/probenetwork_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "debug.h"
+
+#include <elan3/intrinsics.h>
+
+static void ep3_probe_event (EP3_RAIL *rail, void *arg);
+static EP3_COOKIE_OPS ep3_probe_ops = 
+{
+    ep3_probe_event
+} ;
+
+int
+ep3_init_probenetwork (EP3_RAIL *rail)
+{
+    sdramaddr_t              stack;
+    E3_Addr           sp;
+    E3_BlockCopyEvent event;
+    int               i;
+
+    if (! (stack = ep_alloc_elan (&rail->Generic, EP3_STACK_SIZE, 0, &rail->ProbeStack)))
+       return -ENOMEM;
+
+    spin_lock_init (&rail->ProbeLock);
+    kcondvar_init (&rail->ProbeWait);
+
+    /* Initialise the probe command structure */
+    for (i = 0; i < TR_TRACEROUTE_ENTRIES; i++)
+       elan3_sdram_writew (rail->Device, rail->RailElan + offsetof (EP3_RAIL_ELAN, ProbeSource0[i]), 0);
+    for (i = 0; i < TR_TRACEROUTE_ENTRIES; i++)
+       elan3_sdram_writew (rail->Device, rail->RailElan + offsetof (EP3_RAIL_ELAN, ProbeSource1[i]), 1);
+    
+    RegisterCookie (&rail->CookieTable, &rail->ProbeCookie, rail->RailElanAddr + offsetof (EP3_RAIL_ELAN, ProbeDone), &ep3_probe_ops, rail);
+    
+    elan3_sdram_writel (rail->Device, rail->RailElan + offsetof (EP3_RAIL_ELAN, ProbeStart.ev_Type), 0);
+    elan3_sdram_writel (rail->Device, rail->RailElan + offsetof (EP3_RAIL_ELAN, ProbeStart.ev_Count), 0);
+
+    EP3_INIT_COPY_EVENT (event, rail->ProbeCookie, rail->RailMainAddr + offsetof (EP3_RAIL_MAIN, ProbeDone), 1);
+    elan3_sdram_copyl_to_sdram (rail->Device, &event, rail->RailElan + offsetof (EP3_RAIL_ELAN, ProbeDone), sizeof (E3_BlockCopyEvent));
+
+    rail->RailMain->ProbeDone = EP3_EVENT_FREE;
+
+    sp = ep3_init_thread (rail->Device, ep_symbol (&rail->ThreadCode, "kcomm_probe"),
+                         rail->ProbeStack, stack, EP3_STACK_SIZE,
+                         3, rail->CommandPortAddr, rail->RailElanAddr, rail->RailMainAddr);
+    
+    IssueRunThread (rail, sp);
+
+    return 0;
+}
+
+void
+ep3_destroy_probenetwork (EP3_RAIL *rail)
+{
+    if (rail->ProbeStack == (sdramaddr_t) 0)
+       return;
+
+    /* XXXX: ensure that the network probe thread is stopped */
+
+    DeregisterCookie (&rail->CookieTable, &rail->ProbeCookie);
+
+    kcondvar_destroy (&rail->ProbeWait);
+    spin_lock_destroy (&rail->ProbeLock);
+    
+    ep_free_elan (&rail->Generic, rail->ProbeStack, EP3_STACK_SIZE);
+}
+
+static void
+ep3_probe_event (EP3_RAIL *rail, void *arg)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->ProbeLock, flags);
+    rail->ProbeDone = 1;
+    kcondvar_wakeupone (&rail->ProbeWait, &rail->ProbeLock);
+    spin_unlock_irqrestore (&rail->ProbeLock, flags);
+}
+
+int
+ep3_probe_route (EP_RAIL *r, int level, int sw, int nodeid, int *linkup, int *linkdown, int attempts, EP_SWITCH *lsw)
+{
+    EP3_RAIL      *rail     = (EP3_RAIL *) r;
+    EP3_RAIL_MAIN *railMain = rail->RailMain;
+    sdramaddr_t    railElan = rail->RailElan;
+    E3_uint16      flits[MAX_FLITS];
+    E3_uint32      result;
+    int                   nflits;
+    unsigned long  flags;
+
+    spin_lock_irqsave (&rail->ProbeLock, flags);
+
+    nflits = GenerateProbeRoute ( flits, nodeid, level, linkup, linkdown, 0);
+           
+    if (LoadRoute (rail->Device, rail->RouteTable, EP_VP_PROBE(level), ELAN3_MRF_CONTEXT_NUM|SYS_CONTEXT_BIT, nflits, flits) != 0)
+    {
+       EPRINTF0 (DBG_ROUTETABLE, "ProbeRoute: cannot load route entry\n");
+       spin_unlock_irqrestore (&rail->ProbeLock, flags);
+       return (EINVAL);
+    }
+
+    do {
+       /* Initialise the probe source to include our partially computed nodeid */
+       elan3_sdram_writew (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeSource0[TR_TRACEROUTE_ENTRIES-1]), nodeid);
+       elan3_sdram_writew (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeSource1[TR_TRACEROUTE_ENTRIES-1]), nodeid);
+
+       /* Initialise the count result etc */
+       elan3_sdram_writel (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeType), PROBE_SINGLE);
+       elan3_sdram_writel (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeLevel), level);
+
+       railMain->ProbeResult  = -1;
+           
+       /* Clear the receive area */
+       bzero (railMain->ProbeDest0, sizeof (railMain->ProbeDest0));
+       bzero (railMain->ProbeDest1, sizeof (railMain->ProbeDest1));
+    
+       /* Re-arm the completion event */
+       elan3_sdram_writel (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeDone.ev_Count), 1);
+       railMain->ProbeDone = EP3_EVENT_ACTIVE;
+       rail->ProbeDone = 0;
+
+       /* And wakeup the thread to do the probe */
+       IssueSetevent (rail, rail->RailElanAddr + offsetof (EP3_RAIL_ELAN, ProbeStart));
+
+       /* Now wait for it to complete */
+       while (! rail->ProbeDone)
+           kcondvar_wait (&rail->ProbeWait, &rail->ProbeLock, &flags);
+
+       /* wait for block copy event to flush write buffers */
+       while (! EP3_EVENT_FIRED (rail->ProbeCookie, railMain->ProbeDone))
+           if (! EP3_EVENT_FIRING(rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeDone), rail->ProbeCookie, railMain->ProbeDone))
+               panic ("ProbeRoute: network probe event failure\n");
+
+       result = railMain->ProbeResult;
+
+       if (result == C_ACK_ERROR)
+           kcondvar_timedwait (&rail->ProbeWait, &rail->ProbeLock, &flags, lbolt + (hz/8));
+       
+       railMain->ProbeDone = EP3_EVENT_FREE;
+
+    } while (result != C_ACK_OK && --attempts);
+
+    if (result == C_ACK_OK)
+    {
+       if (railMain->ProbeDest0[TR_TRACEROUTE_ENTRIES - ((2*level)+1) - 1] != nodeid ||
+           railMain->ProbeDest1[TR_TRACEROUTE_ENTRIES - ((2*level)+1) - 1] != nodeid)
+       {
+           printk ("%s: lost nodeid at level %d switch %d - %d != %d\n", rail->Generic.Name, level, sw,
+                   railMain->ProbeDest0[TR_TRACEROUTE_ENTRIES - ((2*level)+1) - 1], nodeid);
+
+           result = C_ACK_ERROR;
+       }
+       else
+       {
+           E3_uint16 val0 = railMain->ProbeDest0[TR_TRACEROUTE_ENTRIES - level - 1];
+           E3_uint16 val1 = railMain->ProbeDest1[TR_TRACEROUTE_ENTRIES - level - 1];
+               
+           EPRINTF7 (DBG_PROBE, "%s: level %d switch %d - linkid=%d bcast=%d LNR=%02x%s\n", 
+                     rail->Generic.Name, level, sw, TR_TRACEROUTE0_LINKID(val0),
+                     TR_TRACEROUTE1_BCAST_TOP(val1), TR_TRACEROUTE0_LNR(val0),
+                     TR_TRACEROUTE0_REVID(val0) ? "" : " RevA Part");
+           
+           lsw->lnr     = TR_TRACEROUTE0_LNR(val0);
+           lsw->link    = TR_TRACEROUTE0_LINKID(val0);
+           lsw->bcast   = TR_TRACEROUTE1_BCAST_TOP(val1);
+           lsw->invalid = (TR_TRACEROUTE0_REVID(val0) == 0);
+       }
+    }
+    spin_unlock_irqrestore (&rail->ProbeLock, flags);
+    
+    return (result == C_ACK_OK);
+}
+
+void
+ep3_probe_position_found (EP3_RAIL *rail, ELAN_POSITION *pos)
+{
+    E3_uint16  flits[MAX_FLITS];
+    int        lvl, nflits;
+    
+    for (lvl = 0; lvl < pos->pos_levels; lvl++)
+    {
+       nflits = GenerateCheckRoute (pos, flits, pos->pos_levels - lvl - 1, 0);
+
+       if (LoadRoute (rail->Device, rail->Ctxt->RouteTable, EP_VP_PROBE(lvl), ELAN3_MRF_CONTEXT_NUM|SYS_CONTEXT_BIT, nflits, flits) != 0)
+           panic ("ep3_probe_position_found: cannot load probe route entry\n");
+    }
+    
+    /* Initialise the traceroute source data with our nodeid */
+    elan3_sdram_writew (rail->Device, rail->RailElan + offsetof (EP3_RAIL_ELAN, ProbeSource0[TR_TRACEROUTE_ENTRIES-1]), pos->pos_nodeid);
+    elan3_sdram_writew (rail->Device, rail->RailElan + offsetof (EP3_RAIL_ELAN, ProbeSource1[TR_TRACEROUTE_ENTRIES-1]), pos->pos_nodeid);
+}
+
+int
+ep3_check_position (EP_RAIL *r)
+{
+    EP3_RAIL      *rail     = (EP3_RAIL *) r;
+    EP3_RAIL_MAIN *railMain = rail->RailMain;
+    sdramaddr_t    railElan = rail->RailElan;
+    ELAN_POSITION *pos      = &rail->Generic.Position;
+    unsigned int   level    = rail->RailMain->ProbeLevel;
+    unsigned int   updated  = EP3_EVENT_FIRED (rail->ProbeCookie, railMain->ProbeDone);
+    unsigned int   lvl;
+
+    if (updated)
+    {
+       if (railMain->ProbeResult != C_ACK_OK)
+       {
+           EPRINTF2 (DBG_PROBE, "%s: CheckNetworkPosition: packet nacked result=%d\n", rail->Generic.Name, railMain->ProbeResult); 
+           
+           rail->Generic.SwitchProbeLevel = -1;
+       }
+       else
+       {
+           E3_uint16 val0 = railMain->ProbeDest0[TR_TRACEROUTE_ENTRIES - 2*(level+1)];
+           E3_uint16 val1 = railMain->ProbeDest1[TR_TRACEROUTE_ENTRIES - 2*(level+1)];
+
+           if (val0 != pos->pos_nodeid || val1 != pos->pos_nodeid)
+           {
+               static unsigned long printed = 0;
+
+               /* We've received a packet from another node - this probably means
+                * that we've moved */
+               if ((lbolt - printed) > (HZ*10))
+               {
+                   printk ("%s: ep3_check_position - level %d lost nodeid\n", rail->Generic.Name, level);
+                   printed = lbolt;
+               }
+
+               rail->Generic.SwitchProbeLevel = -1;
+           }
+           else
+           {
+               for (lvl = 0; lvl <= level; lvl++)
+               {
+                   E3_uint16 val0 = railMain->ProbeDest0[TR_TRACEROUTE_ENTRIES - ((2*level) - lvl + 1)];
+                   E3_uint16 val1 = railMain->ProbeDest1[TR_TRACEROUTE_ENTRIES - ((2*level) - lvl + 1)];
+
+                   rail->Generic.SwitchState[lvl].linkid = TR_TRACEROUTE0_LINKID(val0);
+                   rail->Generic.SwitchState[lvl].LNR    = TR_TRACEROUTE0_LNR(val0);
+                   rail->Generic.SwitchState[lvl].bcast  = TR_TRACEROUTE1_BCAST_TOP(val1);
+                   rail->Generic.SwitchState[lvl].uplink = 4;
+
+                   EPRINTF5 (DBG_PROBE, " --- lvl %d: linkid=%d LNR=%x bcast=%d uplink=%d\n", lvl, rail->Generic.SwitchState[lvl].linkid,
+                             rail->Generic.SwitchState[lvl].LNR, rail->Generic.SwitchState[lvl].bcast ,rail->Generic.SwitchState[lvl].uplink);
+               }
+               rail->Generic.SwitchProbeLevel = level;
+           }
+       }
+
+       railMain->ProbeDone = EP3_EVENT_FREE;
+    }
+
+    if (railMain->ProbeDone == EP3_EVENT_FREE)
+    {
+       if (rail->Generic.SwitchBroadcastLevel == rail->Generic.Position.pos_levels-1)
+           level = rail->Generic.Position.pos_levels - 1;
+       else
+           level = rail->Generic.SwitchBroadcastLevel + 1;
+
+       EPRINTF2 (DBG_PROBE, "%s: ep3_check_postiion: level %d\n", rail->Generic.Name, level);
+
+       /* Initialise the count result etc */
+       elan3_sdram_writel (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeType), PROBE_MULTIPLE);
+       elan3_sdram_writel (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeLevel), level);
+
+       railMain->ProbeResult = -1;
+       railMain->ProbeLevel  = -1;
+       
+       /* Clear the receive area */
+       bzero (railMain->ProbeDest0, sizeof (railMain->ProbeDest0));
+       bzero (railMain->ProbeDest1, sizeof (railMain->ProbeDest1));
+       
+       /* Re-arm the completion event */
+       elan3_sdram_writel (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeDone.ev_Type), EV_TYPE_BCOPY);
+       elan3_sdram_writel (rail->Device, railElan + offsetof (EP3_RAIL_ELAN, ProbeDone.ev_Count), 1);
+
+       railMain->ProbeDone = EP3_EVENT_ACTIVE;
+       
+       IssueSetevent (rail, rail->RailElanAddr + offsetof (EP3_RAIL_ELAN, ProbeStart));
+    }
+
+    return updated;
+}
+
Index: linux-2.4.21/drivers/net/qsnet/ep/probenetwork_elan3_thread.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/probenetwork_elan3_thread.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/probenetwork_elan3_thread.c       2005-06-01 23:12:54.678427096 -0400
@@ -0,0 +1,98 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: probenetwork_elan3_thread.c,v 1.19 2004/03/24 11:32:56 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/probenetwork_elan3_thread.c,v $*/
+
+#include <elan3/e3types.h>
+#include <elan3/events.h>
+#include <elan3/elanregs.h>
+#include <elan3/intrinsics.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+
+static int
+kcomm_probe_vp (EP3_RAIL_ELAN *railElan, EP3_RAIL_MAIN *railMain, int vp, int attempts, int timeouts)
+{
+    int rc;
+
+    /* Since we use %g1 to hold the "rxd" so the trap handler can
+     * complete the envelope processing - we pass zero to indicate we're
+     * not a receiver thread */
+    asm volatile ("mov %g0, %g1");
+
+    while (attempts && timeouts)
+    {
+       c_open (vp);
+       c_sendmem (TR_TRACEROUTE, &railMain->ProbeDest0, &railElan->ProbeSource0);
+       c_sendmem (TR_TRACEROUTE, &railMain->ProbeDest1, &railElan->ProbeSource1);
+       c_sendtrans0 (TR_SENDACK | TR_SETEVENT, (E3_Addr) 0);
+       
+       switch (rc = c_close())
+       {
+       case C_ACK_OK:
+           return (C_ACK_OK);
+           
+       case C_ACK_DISCARD:
+           attempts--;
+           break;
+
+       default:                                        /* output timeout */
+           timeouts--;
+       }
+
+       c_break_busywait();
+    }
+
+    return (timeouts == 0 ? C_ACK_ERROR : C_ACK_DISCARD);
+}
+
+void
+kcomm_probe (E3_CommandPort *cport, EP3_RAIL_ELAN *railElan, EP3_RAIL_MAIN *railMain)
+{
+    int level;
+
+    for (;;)
+    {
+       c_waitevent (&railElan->ProbeStart, 1);
+
+       switch (railElan->ProbeType)
+       {
+       case PROBE_SINGLE:
+           railMain->ProbeResult = kcomm_probe_vp (railElan, railMain, EP_VP_PROBE(railElan->ProbeLevel),
+                                                   PROBE_SINGLE_ATTEMPTS, PROBE_SINGLE_TIMEOUTS);
+
+           cport->SetEvent = (E3_Addr) &railElan->ProbeDone;
+           break;
+
+       case PROBE_MULTIPLE:
+           for (level = railElan->ProbeLevel; level >= 0; level--)
+           {
+               if (kcomm_probe_vp (railElan, railMain, EP_VP_PROBE(level),
+                                   PROBE_MULTIPLE_ATTEMPTS, PROBE_MULTIPLE_TIMEOUTS) == C_ACK_OK)
+               {
+                   railMain->ProbeLevel  = level;
+                   railMain->ProbeResult = C_ACK_OK;
+                   break;
+               }
+
+               c_break_busywait();
+           }
+           cport->SetEvent = (E3_Addr) &railElan->ProbeDone;
+           break;
+       }
+
+    }
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/probenetwork_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/probenetwork_elan4.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/probenetwork_elan4.c      2005-06-01 23:12:54.679426944 -0400
@@ -0,0 +1,396 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: probenetwork_elan4.c,v 1.9 2004/08/19 11:05:03 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/probenetwork_elan4.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "debug.h"
+
+#include <elan4/trtype.h>
+#include <elan4/commands.h>
+
+static void
+probe_interrupt (EP4_RAIL *rail, void *arg)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->r_probe_lock, flags);
+    rail->r_probe_done = 1;
+    kcondvar_wakeupone (&rail->r_probe_wait, &rail->r_probe_lock);
+    spin_unlock_irqrestore (&rail->r_probe_lock, flags);
+}
+
+int
+ep4_probe_init (EP4_RAIL *rail)
+{
+    spin_lock_init (&rail->r_probe_lock);
+    kcondvar_init (&rail->r_probe_wait);
+
+    rail->r_probe_cq = ep4_alloc_ecq (rail, CQ_Size1K);
+
+    if (rail->r_probe_cq == NULL)
+       return -ENOMEM;
+
+    ep4_register_intcookie (rail, &rail->r_probe_intcookie, rail->r_elan_addr, probe_interrupt, rail);
+
+    return 0;
+}
+
+void
+ep4_probe_destroy (EP4_RAIL *rail)
+{
+    if (rail->r_probe_cq)
+       ep4_free_ecq (rail, rail->r_probe_cq);
+
+    if (rail->r_probe_intcookie.int_arg == NULL)
+       return;
+    ep4_deregister_intcookie (rail, &rail->r_probe_intcookie);
+
+    kcondvar_destroy (&rail->r_probe_wait);
+    spin_lock_destroy (&rail->r_probe_lock);
+}
+
+#define LINKDOWN(nodeid, level)        ((nodeid >> (level << 1)) & 3)
+#define PROBE_PATTERN0(nodeid) (0xaddebabe ^ nodeid)
+#define PROBE_PATTERN1(nodeid)  (0xfeedbeef ^ nodeid)
+
+#define EP4_PROBE_RETRIES      4
+
+int
+ep4_probe_route (EP_RAIL *r, int level, int sw, int nodeid, int *linkup, int *linkdown, int attempts, EP_SWITCH *lsw)
+{
+    EP4_RAIL      *rail  = (EP4_RAIL *) r;
+    EP4_RAIL_MAIN *rmain = rail->r_main;
+    E4_uint16      first = 0;
+    int                   rb    = 0;
+
+    E4_uint8  packed[ROUTE_NUM_PACKED];
+    E4_VirtualProcessEntry route;
+    unsigned long flags;
+    int i;
+
+    for (i = 0; i < ROUTE_NUM_PACKED; i++)
+       packed[i] = 0;
+
+    /* Generate "up" routes */
+    for (i = 0; i < level; i++)
+       if (first == 0)
+           first = linkup ? FIRST_ROUTE(linkup[i]) : FIRST_ADAPTIVE;
+       else
+           packed[rb++] = linkup ? PACKED_ROUTE(linkup[i]) : PACKED_ADAPTIVE;
+    
+    /* Generate a "to-me" route down */
+    if (first == 0)
+       first = FIRST_MYLINK;
+    else
+       packed[rb++] = PACKED_MYLINK;
+    
+    /* Generate the "down" routes */
+    for (i = level-1; i >= 0; i--)
+       packed[rb++] = linkdown ? PACKED_ROUTE(linkdown[i]) : PACKED_ROUTE(LINKDOWN(nodeid, i));
+    
+    /* Pack up the routes into the virtual process entry */
+    route.Values[0] = first | FIRST_HIGH_PRI | FIRST_SYSTEM_PACKET | FIRST_TIMEOUT(3);
+    route.Values[1] = ROUTE_CTXT_VALUE(ELAN4_KCOMM_CONTEXT_NUM);
+
+    for (i = 0; i < (ROUTE_NUM_PACKED >> 1); i++)
+    {
+       route.Values[0] |= ((E4_uint64) packed[i]) << ((i << 2) + ROUTE_PACKED_OFFSET);
+       route.Values[1] |= ((E4_uint64) packed[i+(ROUTE_NUM_PACKED >> 1)]) << ((i << 2));
+    }
+
+    elan4_write_route (rail->r_ctxt.ctxt_dev, rail->r_routetable, EP_VP_PROBE(level), &route);
+    
+    while (attempts--)
+    {
+       rail->r_probe_done = 0;
+
+       /* generate the STEN packet - note we use a datatype of dword as we're copying to elan in dwords
+        *   NB - no flow control is required, since the max packet size is less than the command queue
+        *        size and it's dedicated for network probing.
+        */
+       
+       elan4_guard   (rail->r_probe_cq->ecq_cq, GUARD_CHANNEL(1) | GUARD_RESET(EP4_PROBE_RETRIES));
+       elan4_nop_cmd (rail->r_probe_cq->ecq_cq, 0);
+       
+       elan4_open_packet (rail->r_probe_cq->ecq_cq, OPEN_STEN_PKT_CMD | OPEN_PACKET(0, PACK_OK | RESTART_COUNT_ZERO, EP_VP_PROBE(level)));
+       elan4_sendtransn  (rail->r_probe_cq->ecq_cq, TR_TRACEROUTE(TRACEROUTE_NDWORDS),
+                          rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_dest0),
+                          0x0000000000000000ull, 0x0000000000000000ull, 0x0000000000000000ull, 0x0000000000000000ull, 
+                          0x0000000000000000ull, 0x0000000000000000ull, 0x0000000000000000ull, 0x0000000000000000ull | ((E4_uint64)PROBE_PATTERN0(nodeid) << 32));
+       elan4_sendtransn  (rail->r_probe_cq->ecq_cq, TR_TRACEROUTE(TRACEROUTE_NDWORDS),
+                          rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_dest1),
+                          0x0000000100000001ull, 0x0000000100000001ull, 0x0000000100000001ull, 0x0000000100000001ull, 
+                          0x0000000100000001ull, 0x0000000100000001ull, 0x0000000100000001ull, 0x0000000000000001ull | ((E4_uint64)PROBE_PATTERN1(nodeid) << 32));
+       elan4_sendtrans0  (rail->r_probe_cq->ecq_cq, TR_NOP_TRANS | TR_LAST_AND_SEND_ACK, 0);
+
+       elan4_guard           (rail->r_probe_cq->ecq_cq, GUARD_CHANNEL(1) | GUARD_TEST(0, PACK_OK) | GUARD_RESET(EP4_PROBE_RETRIES));
+       elan4_write_dword_cmd (rail->r_probe_cq->ecq_cq, rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_result), EP4_STATE_FINISHED);
+
+       elan4_guard            (rail->r_probe_cq->ecq_cq, GUARD_CHANNEL(1) | GUARD_TEST(0, RESTART_COUNT_ZERO) | GUARD_RESET(EP4_PROBE_RETRIES));
+       elan4_write_dword_cmd  (rail->r_probe_cq->ecq_cq, rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_result), EP4_STATE_FAILED);
+
+       elan4_interrupt_cmd   (rail->r_probe_cq->ecq_cq,  rail->r_probe_intcookie.int_val);
+
+       spin_lock_irqsave (&rail->r_probe_lock, flags);
+       while (! rail->r_probe_done)
+           kcondvar_wait (&rail->r_probe_wait, &rail->r_probe_lock, &flags);
+       spin_unlock_irqrestore (&rail->r_probe_lock, flags);
+
+       if (rmain->r_probe_result == EP4_STATE_FINISHED)
+       {
+           if (rmain->r_probe_dest0[TRACEROUTE_ENTRIES - ((2*level)+1) - 1] != PROBE_PATTERN0(nodeid) ||
+               rmain->r_probe_dest1[TRACEROUTE_ENTRIES - ((2*level)+1) - 1] != PROBE_PATTERN1(nodeid))
+           {
+               printk ("%s: lost nodeid at level %d switch %d - %d != %d\n", rail->r_generic.Name, level, sw,
+                       rmain->r_probe_dest0[TRACEROUTE_ENTRIES - ((2*level)+1) - 1], PROBE_PATTERN0(nodeid));
+           }
+           else
+           {
+               E4_uint32 val0 = rmain->r_probe_dest0[TRACEROUTE_ENTRIES - level - 1];
+               E4_uint32 val1 = rmain->r_probe_dest1[TRACEROUTE_ENTRIES - level - 1];
+               
+               lsw->lnr     = TR_TRACEROUTE0_LNR(val0);
+               lsw->link    = TR_TRACEROUTE0_LINKID(val0);
+               lsw->bcast   = TR_TRACEROUTE1_BCAST_TOP(val1);
+               lsw->invalid = 0;
+
+               return 1;
+           }
+       }
+
+       rmain->r_probe_result = EP4_STATE_FREE;
+    }
+
+    return 0;
+}
+
+
+void
+ep4_probe_position_found (EP4_RAIL *rail, ELAN_POSITION *pos)
+{
+    ELAN4_DEV  *dev  = rail->r_ctxt.ctxt_dev;
+    int         lvl;
+
+    for (lvl = 0; lvl < pos->pos_levels; lvl++)
+    {
+       /* Initialise the "probe" route to use the broadcast tree */
+       ELAN_POSITION *pos     = &rail->r_generic.Position;
+       unsigned char *arityp  = &pos->pos_arity[pos->pos_levels - 1];
+       unsigned int   spanned = *arityp;
+       E4_uint16      first   = 0;
+       int            rb      = 0;
+       
+       E4_uint8  packed[ROUTE_NUM_PACKED];
+       E4_VirtualProcessEntry route;
+       int i;
+       
+       for (i = 0; i < ROUTE_NUM_PACKED; i++)
+           packed[i] = 0;
+
+       /* Generate "up" routes */
+       for (i = 0; i < lvl; i++, spanned *= *(--arityp))
+       {
+           if (first == 0)
+               first = FIRST_BCAST_TREE;
+           else
+               packed[rb++] = PACKED_BCAST_TREE;
+       }
+
+       /* Generate a "to-me" route down */
+       if (first == 0)
+           first = FIRST_MYLINK;
+       else
+           packed[rb++] = PACKED_MYLINK;
+
+       spanned /= *arityp++;
+
+       /* Generate the "down" routes */
+       for (i = lvl-1; i >= 0; i--)
+       {
+           spanned /= *arityp;
+           packed[rb++] = PACKED_ROUTE((pos->pos_nodeid / spanned) % *arityp);
+           arityp++;
+       }
+
+    
+       /* Pack up the routes into the virtual process entry */
+       route.Values[0] = first | FIRST_HIGH_PRI | FIRST_SYSTEM_PACKET | FIRST_TIMEOUT(3);
+       route.Values[1] = ROUTE_CTXT_VALUE(ELAN4_KCOMM_CONTEXT_NUM);
+       
+       for (i = 0; i < (ROUTE_NUM_PACKED >> 1); i++)
+       {
+           route.Values[0] |= ((E4_uint64) packed[i]) << ((i << 2) + ROUTE_PACKED_OFFSET);
+           route.Values[1] |= ((E4_uint64) packed[i+(ROUTE_NUM_PACKED >> 1)]) << ((i << 2));
+       }
+       
+       elan4_write_route (rail->r_ctxt.ctxt_dev, rail->r_routetable, EP_VP_PROBE(lvl), &route);
+       
+       /* Initialise "start" event for this level */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_start[lvl].ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_CHECK_STEN_NDWORDS));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_start[lvl].ev_CopySource),
+                           rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl]));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_start[lvl].ev_CopyDest),
+                           rail->r_probe_cq->ecq_addr);
+
+       /* Initiailise command stream - reset the start event */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_reset_event_cmd),
+                           WRITE_DWORD_CMD | (rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_check_start[lvl])));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_reset_event_value),
+                           E4_EVENT_INIT_VALUE (-32, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, EP4_CHECK_STEN_NDWORDS));
+
+       /* Initiailise command stream - sten traceroute packet */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_open),
+                           OPEN_STEN_PKT_CMD | OPEN_PACKET (0, PACK_OK | RESTART_COUNT_ZERO, EP_VP_PROBE(lvl)));
+
+       /* Initiailise command stream - traceroute 0 */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_trans_traceroute0),
+                           SEND_TRANS_CMD | (TR_TRACEROUTE(TRACEROUTE_NDWORDS) << 16));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_addr_traceroute0),
+                           rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_dest0));
+       for (i = 0; i < (TRACEROUTE_NDWORDS-1); i++)
+           elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_data_traceroute0[i]),
+                               0x0000000000000000ull);
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_data_traceroute0[i]),
+                           0x0000000000000000ull | ((E4_uint64) PROBE_PATTERN0(pos->pos_nodeid) << 32));
+
+       /* Initiailise command stream - traceroute 1 */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_trans_traceroute1),
+                           SEND_TRANS_CMD | (TR_TRACEROUTE(TRACEROUTE_NDWORDS) << 16));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_addr_traceroute1),
+                           rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_dest1));
+       for (i = 0; i < (TRACEROUTE_NDWORDS-1); i++)
+           elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_data_traceroute1[i]),
+                               0x0000000100000001ull);
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_data_traceroute1[i]),
+                           0x0000000000000001ull | ((E4_uint64) PROBE_PATTERN1(pos->pos_nodeid) << 32));
+
+       /* Initiailise command stream - null sendack */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_trans_sendack),
+                           SEND_TRANS_CMD | ((TR_NOP_TRANS | TR_LAST_AND_SEND_ACK) << 16));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_addr_sendack),
+                           0);
+       
+       /* Initiailise command stream - guard ok, write done */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_guard_ok),
+                           GUARD_CMD | GUARD_CHANNEL(1) | GUARD_TEST(0, PACK_OK) | GUARD_RESET(EP4_PROBE_RETRIES));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_writedword_ok),
+                           WRITE_DWORD_CMD | (rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_level)));
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_value_ok),
+                           lvl);
+
+       /* Initiailise command stream - guard fail, chain to next or write done */
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_guard_fail),
+                           GUARD_CMD | GUARD_CHANNEL(1) | GUARD_TEST(0, RESTART_COUNT_ZERO) | GUARD_RESET(EP4_PROBE_RETRIES));
+
+       if (lvl > 0)
+       {
+           elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_setevent_fail),
+                               SET_EVENT_CMD | (rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_check_start[lvl-1])));
+           elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_setevent_nop),
+                               NOP_CMD);
+       }
+       else
+       {
+           elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_setevent_fail),
+                               WRITE_DWORD_CMD | (rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_probe_level)));
+           elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_setevent_nop),
+                               EP4_PROBE_FAILED);
+       }
+       elan4_sdram_writeq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_check_sten[lvl].c_nop_pad),
+                           NOP_CMD);
+    }
+
+    
+    rail->r_main->r_probe_level = EP4_PROBE_ACTIVE;
+
+    mb();
+    ep4_set_event_cmd (rail->r_probe_cq, rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_check_start[pos->pos_levels-1]));
+}
+
+int
+ep4_check_position (EP_RAIL *r)
+{
+    EP4_RAIL      *rail = (EP4_RAIL *) r;
+    ELAN_POSITION *pos  = &rail->r_generic.Position;
+    unsigned int level  = rail->r_main->r_probe_level;
+    unsigned int lvl;
+
+    EPRINTF2 (DBG_PROBE, "%s: ep4_check_position: level=%lld\n", rail->r_generic.Name, rail->r_main->r_probe_level);
+
+    if (rail->r_main->r_probe_level != EP4_PROBE_ACTIVE)
+    {
+       if (rail->r_main->r_probe_level == EP4_PROBE_FAILED)
+       {
+           EPRINTF1 (DBG_PROBE, "%s: ep4_check_position: packets all nacked\n", rail->r_generic.Name);
+
+           rail->r_generic.SwitchProbeLevel = -1;
+       }
+       else
+       {
+           E4_uint32 val0  = rail->r_main->r_probe_dest0[TRACEROUTE_ENTRIES - 2*(level+1)];
+           E4_uint32 val1  = rail->r_main->r_probe_dest1[TRACEROUTE_ENTRIES - 2*(level+1)];
+
+           if (val0 != PROBE_PATTERN0 (pos->pos_nodeid) || val1 != PROBE_PATTERN1 (pos->pos_nodeid))
+           {
+               static unsigned long printed = 0;
+
+               /* We've received a packet from another node - this probably means
+                * that we've moved */
+               if ((lbolt - printed) > (HZ*10))
+               {
+                   printk ("%s: ep4_check_position - level %d lost nodeid\n", rail->r_generic.Name, level);
+                   printed = lbolt;
+               }
+
+               rail->r_generic.SwitchProbeLevel = -1;
+           }
+           else
+           {
+               for (lvl = 0 ; lvl <= level; lvl++)
+               {
+                   E4_uint32 uval0  = rail->r_main->r_probe_dest0[TRACEROUTE_ENTRIES - lvl - 1];
+                   E4_uint32 dval0  = rail->r_main->r_probe_dest0[TRACEROUTE_ENTRIES - ((2*level) - lvl + 1)];
+                   E4_uint32 dval1  = rail->r_main->r_probe_dest1[TRACEROUTE_ENTRIES - ((2*level) - lvl + 1)];
+
+                   rail->r_generic.SwitchState[lvl].linkid = TR_TRACEROUTE0_LINKID (dval0);
+                   rail->r_generic.SwitchState[lvl].LNR    = TR_TRACEROUTE0_LNR(dval0);
+                   rail->r_generic.SwitchState[lvl].bcast  = TR_TRACEROUTE1_BCAST_TOP (dval1);
+                   rail->r_generic.SwitchState[lvl].uplink = TR_TRACEROUTE0_LINKID (uval0);
+
+                   EPRINTF5 (DBG_PROBE, " --- lvl %d: linkid=%d LNR=%x bcast=%d uplink=%d\n", lvl, rail->r_generic.SwitchState[lvl].linkid,
+                             rail->r_generic.SwitchState[lvl].LNR, rail->r_generic.SwitchState[lvl].bcast ,rail->r_generic.SwitchState[lvl].uplink);
+
+               }
+
+               rail->r_generic.SwitchProbeLevel = level;
+           }
+       }
+
+       rail->r_main->r_probe_level = EP4_PROBE_ACTIVE;
+       mb();
+
+       if (rail->r_generic.SwitchBroadcastLevel == rail->r_generic.Position.pos_levels-1)
+           level = rail->r_generic.Position.pos_levels - 1;
+       else
+           level = rail->r_generic.SwitchBroadcastLevel + 1;
+
+       ep4_set_event_cmd (rail->r_probe_cq, rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_check_start[level]));
+
+       return 1;
+    }
+
+    return 0;
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/procfs_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/procfs_linux.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/procfs_linux.c    2005-06-01 23:12:54.680426792 -0400
@@ -0,0 +1,693 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: procfs_linux.c,v 1.53.2.4 2005/01/18 14:18:42 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/procfs_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "cm.h"
+#include "debug.h"
+#include "conf_linux.h"
+#include <linux/module.h>
+#include <linux/wait.h>
+#include <linux/poll.h>
+
+#include <qsnet/procfs_linux.h>
+
+struct proc_dir_entry *ep_procfs_root;
+struct proc_dir_entry *ep_config_root;
+
+/*
+ * We provide a slightly "special" interface for /proc/elan/device%d/nodeset,
+ * so that it can be included in a "poll" system call.  On each "read" on the
+ * file, we generate a new nodeset if a) the previous one has been completely
+ * read and b) if it has changed since it was generated.
+ *
+ * Unfortunately ... this doesn't allow "tail -f" to work, since this uses
+ * fstat() on the fd, as we only hold the last nodeset string, we could not
+ * handle the case where two processes were reading a different rates.
+ * We could maybe have implemented this as a "sliding window", so that we 
+ * add a new nodeset string, when it has changed and someone reads past 
+ * end of the last one.   Then if someone read from before out "window"
+ * we would produce "padding" data.  The problem with this, is that a 
+ * simple "cat" on /proc/elan/device%d/nodeset will read the whole "file"
+ * which will be mostly padding !
+ *
+ * Just to not that the purpose of this interface is:
+ *    1) to allow cat /proc/elan/device%d/nodeset to show the current
+ *       nodeset.
+ *    2) to allow rms (or similar) to poll() on the file, and when the
+ *       nodeset changes read a new one.
+ *
+ * so ... we don't bother solving the troublesome "tail -f" problem.
+ */
+
+typedef struct nodeset_private
+{
+    struct nodeset_private *pr_next;
+    EP_RAIL                *pr_rail;
+    unsigned               pr_changed;
+    char                  *pr_page;
+    unsigned               pr_off;
+    unsigned               pr_len;
+} NODESET_PRIVATE;
+
+NODESET_PRIVATE   *ep_nodeset_list;
+wait_queue_head_t  ep_nodeset_wait;
+spinlock_t         ep_nodeset_lock;
+
+static int
+proc_write_state(struct file *file, const char *buffer,
+                unsigned long count, void *data)
+{
+    EP_RAIL *rail = (EP_RAIL *) data;
+    char    tmpbuf[128];
+    int     res;
+
+    if (count > sizeof (tmpbuf)-1)
+       return (-EINVAL);
+    
+    MOD_INC_USE_COUNT;
+    
+    if (copy_from_user (tmpbuf, buffer, count))
+       res = -EFAULT;
+    else 
+    {
+       tmpbuf[count] = '\0';   
+
+       if (tmpbuf[count-1] == '\n')
+           tmpbuf[count-1] = '\0';
+
+       if (! strcmp (tmpbuf, "start") && rail->State == EP_RAIL_STATE_UNINITIALISED)
+           ep_start_rail (rail);
+       
+       if (! strcmp (tmpbuf, "stop") && rail->State > EP_RAIL_STATE_UNINITIALISED)
+           ep_stop_rail (rail);
+       
+       if (! strcmp (tmpbuf, "offline") && rail->State > EP_RAIL_STATE_UNINITIALISED)
+           cm_force_offline (rail, 1, CM_OFFLINE_PROCFS);
+
+       if (! strcmp (tmpbuf, "online") && rail->State > EP_RAIL_STATE_UNINITIALISED)
+           cm_force_offline (rail, 0, CM_OFFLINE_PROCFS);
+
+       if (! strncmp (tmpbuf, "restart=", 8) && rail->State == EP_RAIL_STATE_RUNNING)
+           cm_restart_node (rail, simple_strtol (tmpbuf + 8, NULL, 0));
+
+       if (! strncmp (tmpbuf, "panic=", 6))
+           ep_panic_node (rail->System, simple_strtol(tmpbuf + 6, NULL, 0),
+                          strchr (tmpbuf, ',') ? strchr(tmpbuf, ',') + 1 : "remote panic request");
+
+       if (! strncmp (tmpbuf, "raise=", 6) && rail->State > EP_RAIL_STATE_UNINITIALISED)
+           rail->Operations.RaiseFilter (rail, simple_strtol (tmpbuf + 6, NULL, 0));
+
+       if (! strncmp (tmpbuf, "lower=", 6) && rail->State > EP_RAIL_STATE_UNINITIALISED)
+           rail->Operations.LowerFilter (rail, simple_strtol (tmpbuf + 6, NULL, 0));
+       
+       res = count;
+    }
+
+    MOD_DEC_USE_COUNT;
+
+    return (res);
+}
+
+static int
+proc_read_state(char *page, char **start, off_t off,
+               int count, int *eof, void *data)
+{
+    EP_RAIL *rail = (EP_RAIL *) data;
+    int     len;
+
+    switch (rail->State)
+    {
+    case EP_RAIL_STATE_UNINITIALISED:
+       len = sprintf (page, "uninitialised\n");
+       break;
+    case EP_RAIL_STATE_STARTED:
+       len = sprintf (page, "started\n");
+       break;
+    case EP_RAIL_STATE_RUNNING:
+       len = sprintf (page, "running NodeId=%d NumNodes=%d\n", rail->Position.pos_nodeid, rail->Position.pos_nodes);
+       break;
+    case EP_RAIL_STATE_INCOMPATIBLE:
+       len = sprintf (page, "incompatible NodeId=%d NumNodes=%d\n", rail->Position.pos_nodeid, rail->Position.pos_nodes);
+       break;
+    default:
+       len = sprintf (page, "<unknown>\n");
+       break;
+    }
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+static int
+proc_write_display(struct file *file, const char *buffer,
+                  unsigned long count, void *data)
+{
+    EP_RAIL *rail = (EP_RAIL *) data;
+    char    tmpbuf[128];
+    int     res;
+
+    if (count > sizeof (tmpbuf)-1)
+       return (-EINVAL);
+    
+    MOD_INC_USE_COUNT;
+    
+    if (copy_from_user (tmpbuf, buffer, count))
+       res = -EFAULT;
+    else 
+    {
+       tmpbuf[count] = '\0';   
+
+       if (tmpbuf[count-1] == '\n')
+           tmpbuf[count-1] = '\0';
+
+       if (! strcmp (tmpbuf, "rail"))
+           DisplayRail (rail);
+       if (! strcmp (tmpbuf, "segs"))
+           DisplaySegs (rail);
+       if (! strcmp (tmpbuf, "nodes"))
+           DisplayNodes (rail);
+       if (! strcmp (tmpbuf, "status"))
+           DisplayStatus (rail);
+       if (! strcmp (tmpbuf, "debug") && rail->Operations.Debug)
+           rail->Operations.Debug (rail);
+       if (! strncmp (tmpbuf, "epcomms", 7))
+           ep_comms_display (rail->System, tmpbuf[7] == '=' ? tmpbuf + 8 : NULL);
+       res = count;
+    }
+
+    MOD_DEC_USE_COUNT;
+
+    return (res);
+}
+
+static int
+proc_read_display(char *page, char **start, off_t off,
+                 int count, int *eof, void *data)
+{
+    int len = sprintf (page, "<unreadable>\n");
+    
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+
+static int
+proc_read_stats(char *page, char **start, off_t off,
+               int count, int *eof, void *data)
+{
+    EP_RAIL *rail = (EP_RAIL *) data;
+
+    if ( rail == NULL ) {
+       strcpy(page,"proc_read_stats rail=NULL\n");
+    } else {
+       page[0] = 0;
+       ep_fillout_stats(rail, page);
+       rail->Operations.FillOutStats (rail, page);
+    }
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, strlen(page)));
+}
+
+static int
+proc_read_devinfo(char *page, char **start, off_t off,
+                 int count, int *eof, void *data)
+{
+    EP_RAIL       *rail    = (EP_RAIL *) data;
+    ELAN_DEVINFO  *devinfo = &rail->Devinfo;
+    ELAN_POSITION *pos     = &rail->Position;
+    char          *p       = page;
+    
+    switch (devinfo->dev_device_id)
+    {
+    case PCI_DEVICE_ID_ELAN3:
+       p += sprintf (p, "ep%d is elan3 %d rev %c\n", rail->Number, 
+                     devinfo->dev_instance, 'a' + devinfo->dev_revision_id);
+       break;
+       
+    case PCI_DEVICE_ID_ELAN4:
+       p += sprintf (p, "ep%d is elan4 %d rev %c\n", rail->Number, 
+                     devinfo->dev_instance, 'a' + devinfo->dev_revision_id);
+       break;
+    default:
+       p += sprintf (p, "ep%d is unkown %x/%x\n", rail->Number, devinfo->dev_vendor_id, devinfo->dev_device_id);
+       break;
+    }
+
+    if (rail->State == EP_RAIL_STATE_RUNNING)
+       p += sprintf (p, "ep%d nodeid %d numnodes %d\n", rail->Number, pos->pos_nodeid, pos->pos_nodes);
+
+    return (qsnet_proc_calc_metrics (page, start, off, count, eof, p - page));
+}
+
+static struct rail_info
+{
+    char *name;
+    int (*read_func) (char *page, char **start, off_t off, int count, int *eof, void *data);
+    int (*write_func) (struct file *file, const char *buf, unsigned long count, void *data);
+} rail_info[] = {
+    {"state",   proc_read_state,   proc_write_state},
+    {"display", proc_read_display, proc_write_display},
+    {"stats",   proc_read_stats,   NULL},
+    {"devinfo", proc_read_devinfo, NULL},
+};
+
+static int
+nodeset_open (struct inode *inode, struct file *file)
+{
+    NODESET_PRIVATE *pr;
+
+    if ((pr = kmalloc (sizeof (NODESET_PRIVATE), GFP_KERNEL)) == NULL)
+       return (-ENOMEM);
+    
+    pr->pr_changed = 1;
+    pr->pr_off     = 0;
+    pr->pr_len     = 0;
+    pr->pr_page    = NULL;
+    pr->pr_rail    = (EP_RAIL *)( PDE(inode)->data );
+
+    spin_lock (&ep_nodeset_lock);
+    pr->pr_next = ep_nodeset_list;
+    ep_nodeset_list = pr;
+    spin_unlock (&ep_nodeset_lock);
+
+    file->private_data = (void *) pr;
+
+    MOD_INC_USE_COUNT;
+    return (0);
+}
+
+static int
+nodeset_release (struct inode *inode, struct file *file)
+{
+    NODESET_PRIVATE *pr = (NODESET_PRIVATE *) file->private_data;
+    NODESET_PRIVATE **ppr;
+
+    spin_lock (&ep_nodeset_lock);
+    for (ppr = &ep_nodeset_list; (*ppr) != pr; ppr = &(*ppr)->pr_next)
+       ;
+    (*ppr) = pr->pr_next;
+    spin_unlock (&ep_nodeset_lock);
+
+    if (pr->pr_page)
+       free_page ((unsigned long) pr->pr_page);
+    kfree (pr);
+    
+    MOD_DEC_USE_COUNT;
+    return (0);
+}
+
+static ssize_t
+nodeset_read (struct file *file, char *buf, size_t count, loff_t *ppos)
+{
+    NODESET_PRIVATE *pr  = (NODESET_PRIVATE *) file->private_data;
+    EP_RAIL          *rail = pr->pr_rail;
+    int              error;
+    unsigned long    flags;
+
+    if (!pr->pr_changed && pr->pr_off >= pr->pr_len)
+       return (0);
+
+    if ((error = verify_area (VERIFY_WRITE, buf, count)) != 0)
+       return (error);
+
+    if (pr->pr_page == NULL && (pr->pr_page = (char *) __get_free_page (GFP_KERNEL)) == NULL)
+       return (-ENOMEM);
+
+    if (pr->pr_off >= pr->pr_len)
+    {
+       kmutex_lock (&rail->CallbackLock);
+       if (rail->State == EP_RAIL_STATE_RUNNING)
+       {
+           spin_lock_irqsave (&rail->System->NodeLock, flags);
+           ep_sprintf_bitmap (pr->pr_page, PAGESIZE, statemap_tobitmap(rail->NodeSet), 0, 0, rail->Position.pos_nodes);
+           spin_unlock_irqrestore (&rail->System->NodeLock, flags);
+
+           if (rail->SwitchBroadcastLevel == -1)
+               strcat (pr->pr_page, "<disconnected>");
+           else if (rail->SwitchBroadcastLevel < (rail->Position.pos_levels-1))
+               sprintf (pr->pr_page + strlen (pr->pr_page), "<%d>", rail->SwitchBroadcastLevel);
+           strcat (pr->pr_page, "\n");
+       }
+       else
+           strcpy (pr->pr_page, "<not running>\n");
+       kmutex_unlock (&rail->CallbackLock);
+
+       pr->pr_len     = strlen (pr->pr_page);
+       pr->pr_off     = 0;
+       pr->pr_changed = 0;
+    }
+
+    if (count >= (pr->pr_len - pr->pr_off))
+       count = pr->pr_len - pr->pr_off;
+
+    copy_to_user (buf, pr->pr_page + pr->pr_off, count);
+
+    pr->pr_off += count;
+    *ppos      += count;
+
+    if (pr->pr_off >= pr->pr_len)
+    {
+       free_page ((unsigned long) pr->pr_page);
+       pr->pr_page = NULL;
+    }
+
+    return (count);
+}
+
+static unsigned int
+nodeset_poll (struct file *file, poll_table *wait)
+{
+    NODESET_PRIVATE *pr = (NODESET_PRIVATE *) file->private_data;
+
+    poll_wait (file, &ep_nodeset_wait, wait);
+    if (pr->pr_changed || pr->pr_off < pr->pr_len)
+       return (POLLIN | POLLRDNORM);
+    return (0);
+}
+
+static void 
+nodeset_callback (void *arg, statemap_t *map)
+{
+    EP_RAIL         *rail = (EP_RAIL *) arg;
+    NODESET_PRIVATE *pr;
+
+    ep_display_bitmap (rail->Name, "Nodeset", statemap_tobitmap(map), 0, ep_numnodes(rail->System));
+
+    spin_lock (&ep_nodeset_lock);
+    for (pr = ep_nodeset_list; pr; pr = pr->pr_next)
+       if (pr->pr_rail == rail)
+           pr->pr_changed = 1;
+    spin_unlock (&ep_nodeset_lock);
+
+    wake_up_interruptible (&ep_nodeset_wait);
+}
+
+void
+proc_character_fill (long mode, char *fmt, ...)
+{
+    int len;
+    va_list ap;
+    PROC_PRIVATE *private = (PROC_PRIVATE *)mode;
+    
+    /* is the buffer already full */
+    if (private->pr_len >= private->pr_data_len) 
+       return;
+    
+    /* attempt to fill up to the remaining space */
+    va_start (ap, fmt);
+    len = vsnprintf ( & private->pr_data[private->pr_len], (private->pr_data_len - private->pr_len), fmt, ap);
+    va_end (ap);
+    
+    if (len < 0 ) 
+    {
+       /* we have reached the end of buffer and need to fail all future writes
+        * the caller can check (pr_len >= pr_data_len) and recall with more space 
+        */
+       private->pr_len = private->pr_data_len;
+       return;
+    }
+    
+    /* move the length along */
+    private->pr_len += len;   
+}
+
+int
+proc_release (struct inode *inode, struct file *file)
+{
+    PROC_PRIVATE *pr = (PROC_PRIVATE *) file->private_data;
+    
+    if (pr->pr_data)
+       KMEM_FREE (pr->pr_data, pr->pr_data_len);
+    kfree (pr);
+    
+    MOD_DEC_USE_COUNT;
+    return (0);
+}
+
+ssize_t
+proc_read (struct file *file, char *buf, size_t count, loff_t *ppos)
+{
+    PROC_PRIVATE *pr  = (PROC_PRIVATE *) file->private_data;
+    int           error;
+
+    if (pr->pr_off >= pr->pr_len)
+       return (0);
+
+    if ((error = verify_area (VERIFY_WRITE, buf, count)) != 0)
+       return (error);
+
+    if (count >= (pr->pr_len - pr->pr_off))
+       count = pr->pr_len - pr->pr_off;
+
+    copy_to_user (buf, pr->pr_data + pr->pr_off, count);
+
+    pr->pr_off += count;
+    *ppos      += count;
+
+    return (count);
+}
+
+static int
+proc_open (struct inode *inode, struct file *file)
+{
+    PROC_PRIVATE *pr;
+    CM_RAIL      *cmRail;
+    int           pages = 4;
+    unsigned long flags;
+
+    if ((pr = kmalloc (sizeof (PROC_PRIVATE), GFP_KERNEL)) == NULL)
+       return (-ENOMEM);
+    
+    pr->pr_rail = (EP_RAIL *)(PDE(inode)->data);
+       
+    do {       
+       pr->pr_data_len = PAGESIZE * pages;
+
+       KMEM_ZALLOC (pr->pr_data, char *, pr->pr_data_len, 1);
+       if (pr->pr_data == NULL) 
+       { 
+           pr->pr_len  = sprintf (pr->pr_data, "Out of Memory\n");
+           break;
+       } 
+       
+       pr->pr_off     = 0;
+       pr->pr_len     = 0;
+       pr->pr_data[0] = 0;
+       
+       if (pr->pr_rail->State != EP_RAIL_STATE_RUNNING) 
+       { 
+           pr->pr_len  = sprintf (pr->pr_data, "Rail not Running\n");
+           break;
+       } 
+       else 
+       {
+           pr->pr_di.func  = proc_character_fill;
+           pr->pr_di.arg   = (long)pr;
+
+           if (!strcmp("maps", file->f_dentry->d_iname)) 
+           {
+               cmRail = pr->pr_rail->ClusterRail;
+
+               spin_lock_irqsave (&cmRail->Lock, flags);
+               DisplayNodeMaps (&pr->pr_di, cmRail);   
+               spin_unlock_irqrestore (&cmRail->Lock, flags);  
+           }
+
+           if (!strcmp("segs", file->f_dentry->d_iname)) 
+           {
+               cmRail = pr->pr_rail->ClusterRail;
+               
+               spin_lock_irqsave (&cmRail->Lock, flags);       
+               DisplayNodeSgmts (&pr->pr_di, cmRail);
+               spin_unlock_irqrestore (&cmRail->Lock, flags);
+           }
+
+           if (!strcmp("tree", file->f_dentry->d_iname)) 
+               DisplayRailDo (&pr->pr_di, pr->pr_rail);
+       }
+
+       if ( pr->pr_len < pr->pr_data_len) 
+           break; /* we managed to get all the output into the buffer */
+
+       pages++;
+       KMEM_FREE ( pr->pr_data,  pr->pr_data_len);
+    } while (1);
+       
+
+    file->private_data = (void *) pr;
+
+    MOD_INC_USE_COUNT;
+    return (0);
+}
+
+struct file_operations proc_nodeset_operations = 
+{
+    read:      nodeset_read,
+    poll:      nodeset_poll,
+    open:      nodeset_open,
+    release:   nodeset_release,
+};
+
+struct file_operations proc_operations = 
+{
+    read:      proc_read,
+    open:      proc_open,
+    release:   proc_release,
+};
+
+void
+ep_procfs_rail_init (EP_RAIL *rail)
+{
+    struct proc_dir_entry *dir;
+    struct proc_dir_entry *p;
+    char                   name[10];
+    int                    i;
+
+    sprintf (name, "rail%d", rail->Number);
+
+    if ((dir = rail->ProcDir = proc_mkdir (name, ep_procfs_root)) == NULL)
+       return;
+
+    for (i = 0; i < sizeof (rail_info)/sizeof (rail_info[0]); i++)
+    {
+       if ((p = create_proc_entry (rail_info[i].name, 0, dir)) != NULL)
+       {
+           p->read_proc  = rail_info[i].read_func;
+           p->write_proc = rail_info[i].write_func;
+           p->data       = rail;
+           p->owner      = THIS_MODULE;
+       }
+    }
+
+    if ((p = create_proc_entry ("nodeset", 0, dir)) != NULL)
+    {
+       p->proc_fops = &proc_nodeset_operations;
+       p->owner     = THIS_MODULE;
+       p->data      = rail;
+
+       rail->CallbackRegistered = 1;
+       ep_register_callback (rail, EP_CB_NODESET, nodeset_callback, rail);
+    }
+     
+    if ((p = create_proc_entry ("maps", 0, dir)) != NULL)
+    {
+       p->proc_fops = &proc_operations;
+       p->owner     = THIS_MODULE;
+       p->data      = rail;    
+    }
+    
+    if ((p = create_proc_entry ("segs", 0, dir)) != NULL)
+    {
+       p->proc_fops = &proc_operations;
+       p->owner     = THIS_MODULE;
+       p->data      = rail;
+    }
+    
+    if ((p = create_proc_entry ("tree", 0, dir)) != NULL)
+    {
+       p->proc_fops = &proc_operations;
+       p->owner     = THIS_MODULE;
+       p->data      = rail;
+    }
+
+}
+
+void
+ep_procfs_rail_fini (EP_RAIL *rail)
+{
+    struct proc_dir_entry *dir = rail->ProcDir;
+    char name[10];
+    int  i;
+
+    if (dir == NULL)
+       return;
+
+    if (rail->CallbackRegistered)
+    {
+       ep_remove_callback (rail, EP_CB_NODESET, nodeset_callback, rail);
+
+       remove_proc_entry ("nodeset", dir);
+    }
+
+    remove_proc_entry ("maps",    dir);
+    remove_proc_entry ("segs",    dir);
+    remove_proc_entry ("tree",    dir);
+
+    for (i = 0; i < sizeof (rail_info)/sizeof (rail_info[0]); i++)
+       remove_proc_entry (rail_info[i].name, dir);
+
+    sprintf (name, "rail%d", rail->Number);
+    remove_proc_entry (name, ep_procfs_root);
+}
+
+#include "quadrics_version.h"
+static char     quadrics_version[] = QUADRICS_VERSION;
+
+void
+ep_procfs_init()
+{
+    extern int txd_stabilise;
+    extern int MaxSwitchLevels;
+
+    spin_lock_init (&ep_nodeset_lock);
+    init_waitqueue_head (&ep_nodeset_wait);
+
+    ep_procfs_root = proc_mkdir ("ep", qsnet_procfs_root);
+    ep_config_root = proc_mkdir ("config", ep_procfs_root);
+
+    qsnet_proc_register_str (ep_procfs_root, "version", quadrics_version, 1);
+
+    qsnet_proc_register_hex (ep_config_root, "epdebug",               &epdebug,               0);
+    qsnet_proc_register_hex (ep_config_root, "epdebug_console",       &epdebug_console,       0);
+    qsnet_proc_register_hex (ep_config_root, "epdebug_cmlevel",       &epdebug_cmlevel,       0);
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    qsnet_proc_register_hex (ep_config_root, "epdebug_check_sum",     &epdebug_check_sum,     0);
+#endif
+    qsnet_proc_register_hex (ep_config_root, "epcomms_forward_limit", &epcomms_forward_limit, 0);
+    qsnet_proc_register_int (ep_config_root, "txd_stabilise",         &txd_stabilise,         0);
+    qsnet_proc_register_int (ep_config_root, "assfail_mode",          &assfail_mode,          0);
+    qsnet_proc_register_int (ep_config_root, "max_switch_levels",     &MaxSwitchLevels,       1);
+
+    ep_procfs_rcvr_xmtr_init();
+}
+
+void
+ep_procfs_fini(void)
+{
+    ep_procfs_rcvr_xmtr_fini();
+
+    remove_proc_entry ("max_switch_levels",     ep_config_root);
+    remove_proc_entry ("assfail_mode",          ep_config_root);
+    remove_proc_entry ("txd_stabilise",         ep_config_root);
+    remove_proc_entry ("epcomms_forward_limit", ep_config_root);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    remove_proc_entry ("epdebug_check_sum",     ep_config_root);
+#endif
+    remove_proc_entry ("epdebug_cmlevel",       ep_config_root);
+    remove_proc_entry ("epdebug_console",       ep_config_root);
+    remove_proc_entry ("epdebug",               ep_config_root);
+
+    remove_proc_entry ("version", ep_procfs_root);
+    
+    remove_proc_entry ("config", ep_procfs_root);
+    remove_proc_entry ("ep", qsnet_procfs_root);
+
+    spin_lock_destroy (&ep_nodeset_lock);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/quadrics_version.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/quadrics_version.h        2005-06-01 23:12:54.680426792 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/ep/railhints.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/railhints.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/railhints.c       2005-06-01 23:12:54.680426792 -0400
@@ -0,0 +1,103 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: railhints.c,v 1.5 2004/02/06 22:37:06 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/railhints.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "debug.h"
+
+int
+ep_pickRail(EP_RAILMASK railmask)
+{
+    static volatile int lastGlobal;
+    int i, rnum, last = lastGlobal;
+
+    /* Pick a single rail out of the railmask */
+    for (i = 0; i < EP_MAX_RAILS; i++)
+       if (railmask & (1 << ((last + i) % EP_MAX_RAILS)))
+           break;
+
+    if (i == EP_MAX_RAILS)
+       return (-1);
+
+    rnum = (last + i) % EP_MAX_RAILS;
+
+    lastGlobal = (rnum + 1) % EP_MAX_RAILS;
+
+    ASSERT (railmask & (1 << rnum));
+
+    return (rnum);
+}
+
+int
+ep_xmtr_bcastrail (EP_XMTR *xmtr, EP_RAILMASK allowedRails)
+{
+    /* Retrun a single rail out of allowed mask with the best connectivity for broadcast. */
+    return (ep_pickRail (allowedRails & xmtr->RailMask));
+}
+
+int
+ep_xmtr_prefrail (EP_XMTR *xmtr, EP_RAILMASK allowedRails, unsigned nodeId)
+{
+    EP_NODE *node = &xmtr->Subsys->Subsys.Sys->Nodes[nodeId];
+
+    EPRINTF5 (DBG_XMTR, "ep_xmtr_prefrail: xmtr=%p allowedRails=%x nodeId=%d xmtr->RailMaks=%x Connected=%x\n", 
+             xmtr, allowedRails, nodeId, xmtr->RailMask, node->ConnectedRails);
+
+    /* Return a single rail which is currently connected to nodeId (limited to rails
+     * in allowedmask) - if more than one rail is possible, then round-robin between 
+     * them */
+    return (ep_pickRail (allowedRails & xmtr->RailMask & node->ConnectedRails));
+}
+
+EP_RAILMASK
+ep_xmtr_availrails (EP_XMTR *xmtr)
+{
+    /* Return which rails can be used to transmit one. */
+
+    return (xmtr->RailMask);
+}
+
+EP_RAILMASK
+ep_xmtr_noderails (EP_XMTR *xmtr, unsigned nodeId)
+{
+    EP_NODE *node = &xmtr->Subsys->Subsys.Sys->Nodes[nodeId];
+
+    /* Return which rails can be used to transmit to this node. */
+
+    return (xmtr->RailMask & node->ConnectedRails);
+}
+
+int
+ep_rcvr_prefrail (EP_RCVR *rcvr, EP_RAILMASK allowedRails)
+{
+    /* Return the "best" rail for queueing a receive buffer out on - this will be a
+     * rail with ThreadWaiting set or the rail with the least descriptors queued
+     * on it. */
+    
+    return (ep_pickRail (allowedRails & rcvr->RailMask));
+}
+
+EP_RAILMASK
+ep_rcvr_availrails (EP_RCVR *rcvr)
+{
+    /* Return which rails can be used to queue receive buffers. */
+    return (rcvr->RailMask);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/rmap.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/rmap.c       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/rmap.c    2005-06-01 23:12:54.681426640 -0400
@@ -0,0 +1,365 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: rmap.c,v 1.15 2004/05/19 10:24:38 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/rmap.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <elan/rmap.h>
+
+#include "debug.h"
+
+void
+ep_display_rmap (EP_RMAP *mp)
+{
+    EP_RMAP_ENTRY *bp;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&mp->m_lock, flags);
+    ep_debugf (DBG_DEBUG, "map: %s size %d free %d\n", mp->m_name, mp->m_size, mp->m_free);
+    for (bp = &mp->m_map[0]; bp->m_size; bp++)
+       ep_debugf (DBG_DEBUG, "   [%lx - %lx]\n", bp->m_addr, bp->m_addr+bp->m_size-1);
+    spin_unlock_irqrestore (&mp->m_lock, flags);
+}
+
+void
+ep_mapinit (EP_RMAP *mp, char *name, u_int mapsize)
+{
+    spin_lock_init (&mp->m_lock);
+    kcondvar_init (&mp->m_wait);
+    
+    /* The final segment in the array has size 0 and acts as a delimiter
+     * we insure that we never use segments past the end of the array by
+     * maintaining a free segment count in m_free.  When excess segments
+     * occur we discard some resources */
+    
+    mp->m_size = mapsize;
+    mp->m_free = mapsize;
+    mp->m_name = name;
+    
+    bzero (mp->m_map, sizeof (EP_RMAP_ENTRY) * (mapsize+1));
+}
+
+EP_RMAP *
+ep_rmallocmap (size_t mapsize, char *name, int cansleep)
+{
+    EP_RMAP *mp;
+
+    KMEM_ZALLOC (mp, EP_RMAP *, sizeof (EP_RMAP) + mapsize*sizeof (EP_RMAP_ENTRY), cansleep);
+
+    if (mp != NULL)
+       ep_mapinit (mp, name, mapsize);
+
+    return (mp);
+}
+
+void
+ep_rmfreemap (EP_RMAP *mp)
+{
+    spin_lock_destroy (&mp->m_lock);
+    kcondvar_destroy (&mp->m_wait);
+    
+    KMEM_FREE (mp, sizeof (EP_RMAP) + mp->m_size * sizeof (EP_RMAP_ENTRY));
+}
+
+static u_long
+ep_rmalloc_locked (EP_RMAP *mp, size_t size)
+{
+    EP_RMAP_ENTRY *bp;
+    u_long            addr;
+    
+    ASSERT (size > 0);
+    ASSERT (SPINLOCK_HELD (&mp->m_lock));
+
+    for (bp = &mp->m_map[0]; bp->m_size; bp++)
+    {
+       if (bp->m_size >= size)
+       {
+           addr = bp->m_addr;
+           bp->m_addr += size;
+           
+           if ((bp->m_size -= size) == 0)
+           {
+               /* taken all of this slot - so shift the map down */
+               do {
+                   bp++;
+                   (bp-1)->m_addr = bp->m_addr;
+               } while (((bp-1)->m_size = bp->m_size) != 0);
+
+               mp->m_free++;
+           }
+           return (addr);
+       }
+    }
+
+    return (0);
+}
+
+u_long
+ep_rmalloc (EP_RMAP *mp, size_t size, int cansleep)
+{
+    unsigned long addr;
+    unsigned long flags;
+
+    spin_lock_irqsave (&mp->m_lock, flags);
+    while ((addr = ep_rmalloc_locked (mp, size)) == 0 && cansleep)
+    {
+       mp->m_want = 1;
+       kcondvar_wait (&mp->m_wait, &mp->m_lock, &flags);
+    }
+
+    spin_unlock_irqrestore (&mp->m_lock, flags);
+
+    return (addr);
+}
+
+
+
+u_long
+ep_rmalloc_constrained (EP_RMAP *mp, size_t size, u_long alo, u_long ahi, u_long align, int cansleep)
+{
+    EP_RMAP_ENTRY *bp, *bp2, *lbp;
+    unsigned long addr=0;
+    size_t        delta;
+    int           ok;
+    unsigned long flags;
+
+    spin_lock_irqsave (&mp->m_lock, flags);
+ again:
+    for (bp = &mp->m_map[0]; bp->m_size; bp++)
+    {
+       delta = 0;
+       
+       if (alo < bp->m_addr)
+       {
+           addr = bp->m_addr;
+           
+           if (addr & (align-1))
+               addr = (addr + (align-1)) & ~(align-1);
+           
+           delta = addr - bp->m_addr;
+           
+           if (ahi >= bp->m_addr + bp->m_size)
+               ok = (bp->m_size >= (size + delta));
+           else
+               ok = ((bp->m_addr + size + delta) <= ahi);
+       }
+       else
+       {
+           addr = alo;
+           if (addr & (align-1))
+               addr = (addr + (align-1)) & ~(align-1);
+           delta = addr - bp->m_addr;
+           
+           if (ahi >= bp->m_addr + bp->m_size)
+               ok = ((alo + size + delta) <= (bp->m_addr + bp->m_size));
+           else
+               ok = ((alo + size + delta) <= ahi);
+       }
+
+       if (ok)
+           break;
+    }  
+    
+    if (bp->m_size == 0)
+    {
+       if (cansleep)
+       {
+           mp->m_want = 1;
+           kcondvar_wait (&mp->m_wait, &mp->m_lock, &flags);
+           goto again;
+       }
+       spin_unlock_irqrestore (&mp->m_lock, flags);
+       return (0);
+    }
+
+    /* found an approriate map entry - so take the bit out which we want */
+    if (bp->m_addr == addr) 
+    {
+       if (bp->m_size == size) 
+       {
+           /* allocate entire segment and compress map */
+           bp2 = bp;
+           while (bp2->m_size) 
+           {
+               bp2++;
+               (bp2-1)->m_addr = bp2->m_addr;
+               (bp2-1)->m_size = bp2->m_size;
+           }
+           mp->m_free++;
+       }
+       else 
+       {
+           /* take from start of segment */
+           bp->m_addr += size;
+           bp->m_size -= size;
+       }
+    }
+    else 
+    {
+       if (bp->m_addr + bp->m_size == addr + size) 
+       {
+           /* take from end of segment */
+           bp->m_size -= size;
+       }
+       else 
+       {
+           /* split the segment loosing the last entry if there's no space */
+           if (mp->m_free == 0) 
+           {
+               /* find last map entry */
+               for (lbp = bp; lbp->m_size != 0; lbp++)
+                   ;
+               lbp--;
+               
+               if (lbp->m_size > (lbp-1)->m_size)
+                   lbp--;
+               
+               printk ("%s: lost resource map entry [%lx, %lx]\n",
+                       mp->m_name, lbp->m_addr, lbp->m_addr + lbp->m_size);
+               
+               *lbp = *(lbp+1);
+               (lbp+1)->m_size = 0;
+               
+               mp->m_free++;
+           }
+           
+           for (bp2 = bp; bp2->m_size != 0; bp2++)
+               continue;
+           
+           for (bp2--; bp2 > bp; bp2--)
+           {
+               (bp2+1)->m_addr = bp2->m_addr;
+               (bp2+1)->m_size = bp2->m_size;
+           }
+
+           mp->m_free--;
+           
+           (bp+1)->m_addr = addr + size;
+           (bp+1)->m_size = bp->m_addr + bp->m_size - (addr + size);
+           bp->m_size = addr - bp->m_addr;
+       }
+    }
+
+    spin_unlock_irqrestore (&mp->m_lock, flags);
+    return (addr);
+}
+
+void
+ep_rmfree (EP_RMAP *mp, size_t size, u_long addr)
+{
+    EP_RMAP_ENTRY *bp;
+    unsigned long t;
+    unsigned long flags;
+
+    spin_lock_irqsave (&mp->m_lock, flags);
+
+    ASSERT (addr != 0 && size > 0);
+       
+again:
+    /* find the piece of the map which starts after the returned space
+     * or the end of the map */
+    for (bp = &mp->m_map[0]; bp->m_addr <= addr && bp->m_size != 0; bp++)
+       ;
+
+    /* bp points to the piece to the right of where we want to go */
+    
+    if (bp > &mp->m_map[0] && (bp-1)->m_addr + (bp-1)->m_size >= addr) 
+    {
+       /* merge with piece on the left */
+       
+       ASSERT ((bp-1)->m_addr + (bp-1)->m_size <= addr);
+       
+       (bp-1)->m_size += size;
+       
+       ASSERT (bp->m_size == 0 || addr+size <= bp->m_addr);
+       
+       if (bp->m_size && (addr + size) == bp->m_addr)
+       {
+           /* merge witht he piece on the right by 
+            * growing the piece on the left and shifting
+            * the map down */
+           
+           ASSERT ((addr + size) <= bp->m_addr);
+           
+           (bp-1)->m_size += bp->m_size;
+           while (bp->m_size) 
+           {
+               bp++;
+               (bp-1)->m_addr = bp->m_addr;
+               (bp-1)->m_size = bp->m_size;
+           }
+           
+           mp->m_free++;
+       }
+    }
+    else if (addr + size >= bp->m_addr && bp->m_size)
+    {
+       /* merge with piece to the right */
+       
+       ASSERT ((addr + size) <= bp->m_addr);
+       
+       bp->m_addr -= size;
+       bp->m_size += size;
+    }
+    else
+    {
+       /* doesn't join with left or right - check for map
+          overflow and discard the smallest of the last or
+          next to last entries */
+
+       if (mp->m_free == 0)
+       {
+           EP_RMAP_ENTRY *lbp;
+           
+           /* find last map entry */
+           for (lbp = bp; lbp->m_size != 0; lbp++)
+               ;
+           lbp--;
+           
+           if (lbp->m_size > (lbp-1)->m_size)
+               lbp--;
+           
+           printk ("%s: lost resource map entry [%lx, %lx]\n", 
+                   mp->m_name, lbp->m_addr, lbp->m_addr + lbp->m_size);
+           
+           *lbp = *(lbp+1);
+           (lbp+1)->m_size = 0;
+
+           mp->m_free++;
+           goto again;
+       }
+
+       /* make a new entry and push the remaining ones up */
+       do {
+           t = bp->m_addr;
+           bp->m_addr = addr;
+           addr = t;
+           t = bp->m_size;
+           bp->m_size = size;
+           bp++;
+       } while ((size = t) != 0);
+
+       mp->m_free--;
+    }
+    
+    /* if anyone blocked on rmalloc failure, wake 'em up */
+    if (mp->m_want)
+    {
+       mp->m_want = 0;
+       kcondvar_wakeupall (&mp->m_wait, &mp->m_lock);
+    }
+
+    spin_unlock_irqrestore (&mp->m_lock, flags);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/spinlock_elan3_thread.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/spinlock_elan3_thread.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/spinlock_elan3_thread.c   2005-06-01 23:12:54.681426640 -0400
@@ -0,0 +1,44 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: spinlock_elan3_thread.c,v 1.9 2003/10/07 13:22:38 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/spinlock_elan3_thread.c,v $ */
+
+#include <qsnet/types.h>
+
+#include <elan3/e3types.h>
+#include <elan3/events.h>
+#include <elan3/elanregs.h>
+#include <elan3/intrinsics.h>
+
+#include <elan/nmh.h>
+#include <elan/kcomm.h>
+#include <elan/epcomms.h>
+
+#include "kcomm_elan3.h"
+#include "epcomms_elan3.h"
+
+void
+ep3_spinblock (EP3_SPINLOCK_ELAN *sle, EP3_SPINLOCK_MAIN *sl)
+{
+    do {
+       sl->sl_seq = sle->sl_seq;                       /* Release my lock */
+       
+       while (sle->sl_lock)                            /* Wait until the main */
+           c_break();                                  /* releases the lock */
+       
+       sle->sl_seq++;                                  /* and try and relock */
+    } while (sle->sl_lock);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/statemap.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/statemap.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/statemap.c        2005-06-01 23:12:54.682426488 -0400
@@ -0,0 +1,385 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: statemap.c,v 1.11.8.1 2004/11/18 12:05:00 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/statemap.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <elan/statemap.h>
+
+/******************************** global state bitmap stuff **********************************/
+static int
+statemap_setmapbit (bitmap_t *map, int offset, int bit)
+{
+   bitmap_t *e    = &map[offset >> BT_ULSHIFT];
+   bitmap_t  mask = ((bitmap_t)1) << (offset & BT_ULMASK);
+   int       rc = ((*e) & mask) != 0;
+   
+   if (bit)
+   {
+      *e |= mask;
+      return (!rc);
+   }
+
+   *e &= ~mask;
+   return (rc);
+}
+
+static int
+statemap_firstsegbit (bitmap_t seg)
+{
+   int            bit = 0;
+   
+   if (seg == 0)
+      return (-1);
+
+#if (BT_ULSHIFT == 6)
+   if ((seg & 0xffffffffL) == 0)
+   {
+      seg >>= 32;
+      bit += 32;
+   }
+#elif (BT_ULSHIFT != 5)
+# error "Unexpected value of BT_ULSHIFT"
+#endif
+
+   if ((seg & 0xffff) == 0)
+   {
+      seg >>= 16;
+      bit += 16;
+   }
+      
+   if ((seg & 0xff) == 0)
+   {
+      seg >>= 8;
+      bit += 8;
+   }
+      
+   if ((seg & 0xf) == 0)
+   {
+      seg >>= 4;
+      bit += 4;
+   }
+      
+   if ((seg & 0x3) == 0)
+   {
+      seg >>= 2;
+      bit += 2;
+   }
+
+   return (((seg & 0x1) == 0) ? bit + 1 : bit);
+}
+
+bitmap_t
+statemap_getseg (statemap_t *map, unsigned int offset)
+{
+   ASSERT (offset < map->size);
+   ASSERT ((offset & BT_ULMASK) == 0);
+
+   return (map->bitmap[offset >> BT_ULSHIFT]);
+}
+
+void
+statemap_setseg (statemap_t *map, unsigned int offset, bitmap_t seg)
+{
+   ASSERT (offset < map->size);
+   ASSERT ((offset & BT_ULMASK) == 0);
+
+   offset >>= BT_ULSHIFT;
+   if (map->bitmap[offset] == seg)
+      return;
+
+   map->bitmap[offset] = seg;
+
+   if (statemap_setmapbit (map->changemap2, offset,       1) &&
+       statemap_setmapbit (map->changemap1, offset >>= BT_ULSHIFT, 1))
+      statemap_setmapbit (map->changemap0, offset >>= BT_ULSHIFT, 1);
+}
+
+bitmap_t
+statemap_getbits (statemap_t *map, unsigned int offset, int nbits)
+{
+   int      index = offset >> BT_ULSHIFT;
+   bitmap_t mask  = (nbits == BT_NBIPUL) ? (bitmap_t) -1 : (((bitmap_t)1) << nbits) - 1;
+   
+   ASSERT (nbits <= BT_NBIPUL);
+   ASSERT (offset + nbits <= map->size);
+
+   offset &= BT_ULMASK;
+   if (offset + nbits <= BT_NBIPUL)
+      return ((map->bitmap[index] >> offset) & mask);
+   
+   return (((map->bitmap[index] >> offset) |
+           (map->bitmap[index + 1] << (BT_NBIPUL - offset))) & mask);
+}
+
+void
+statemap_setbits (statemap_t *map, unsigned int offset, bitmap_t bits, int nbits)
+{
+   int      index = offset >> BT_ULSHIFT;
+   bitmap_t mask;
+   bitmap_t seg;
+   bitmap_t newseg;
+
+   ASSERT (nbits <= BT_NBIPUL);
+   ASSERT (offset + nbits <= map->size);
+
+   offset &= BT_ULMASK;
+   if (offset + nbits <= BT_NBIPUL)
+   {
+      mask = ((nbits == BT_NBIPUL) ? -1 : ((((bitmap_t)1) << nbits) - 1)) << offset;
+      seg = map->bitmap[index];
+      newseg = ((bits << offset) & mask) | (seg & ~mask);
+      
+      if (seg == newseg)
+        return;
+   
+      map->bitmap[index] = newseg;
+      
+      if (statemap_setmapbit (map->changemap2, index,       1) &&
+         statemap_setmapbit (map->changemap1, index >>= BT_ULSHIFT, 1))
+        statemap_setmapbit (map->changemap0, index >>= BT_ULSHIFT, 1);
+      return;
+   }
+   
+   mask = ((bitmap_t)-1) << offset;
+   seg = map->bitmap[index];
+   newseg = ((bits << offset) & mask) | (seg & ~mask);
+
+   if (seg != newseg)
+   {
+      map->bitmap[index] = newseg;
+      
+      if (statemap_setmapbit (map->changemap2, index,       1) &&
+         statemap_setmapbit (map->changemap1, index >> BT_ULSHIFT, 1))
+        statemap_setmapbit (map->changemap0, index >> (2 * BT_ULSHIFT), 1);
+   }
+   
+   index++;
+   offset = BT_NBIPUL - offset;
+   mask = (((bitmap_t)1) << (nbits - offset)) - 1;
+   seg = map->bitmap[index];
+   newseg = ((bits >> offset) & mask) | (seg & ~mask);
+   
+   if (seg == newseg)
+      return;
+   
+   map->bitmap[index] = newseg;
+   
+   if (statemap_setmapbit (map->changemap2, index,       1) &&
+       statemap_setmapbit (map->changemap1, index >>= BT_ULSHIFT, 1))
+      statemap_setmapbit (map->changemap0, index >>= BT_ULSHIFT, 1);
+}
+
+void
+statemap_zero (statemap_t *dst)
+{
+   int       size       = dst->size;
+   int       offset     = 0;
+   bitmap_t *changemap0 = dst->changemap0;
+   bitmap_t *changemap1 = dst->changemap1;
+   bitmap_t *changemap2 = dst->changemap2;
+   bitmap_t *dstmap     = dst->bitmap;
+   bitmap_t  bit0;
+   bitmap_t  bit1;
+   bitmap_t  bit2;
+
+   for (bit0 = 1; offset < size; bit0 <<= 1, changemap1++)
+   {
+      for (bit1 = 1; bit1 != 0 && offset < size; bit1 <<= 1, changemap2++)
+      {
+        for (bit2 = 1; bit2 != 0 && offset < size; bit2 <<= 1, dstmap++, offset += BT_NBIPUL)
+        {
+            *dstmap = 0;
+            *changemap2 |= bit2;
+        }
+        *changemap1 |= bit1;
+      }
+      *changemap0 |= bit0;
+   }
+}
+   
+void
+statemap_setmap (statemap_t *dst, statemap_t *src)
+{
+   int       size       = dst->size;
+   int       offset     = 0;
+   bitmap_t *changemap0 = dst->changemap0;
+   bitmap_t *changemap1 = dst->changemap1;
+   bitmap_t *changemap2 = dst->changemap2;
+   bitmap_t *dstmap     = dst->bitmap;
+   bitmap_t *srcmap     = src->bitmap;
+   bitmap_t  bit0;
+   bitmap_t  bit1;
+   bitmap_t  bit2;
+
+   ASSERT (src->size == size);
+   
+   for (bit0 = 1; offset < size; bit0 <<= 1, changemap1++)
+   {
+      for (bit1 = 1; bit1 != 0 && offset < size; bit1 <<= 1, changemap2++)
+      {
+        for (bit2 = 1; bit2 != 0 && offset < size; bit2 <<= 1, dstmap++, srcmap++, offset += BT_NBIPUL)
+           if (*dstmap != *srcmap)
+           {
+              *dstmap = *srcmap;
+              *changemap2 |= bit2;
+           }
+        if (*changemap2 != 0)
+           *changemap1 |= bit1;
+      }
+      if (*changemap1 != 0)
+        *changemap0 |= bit0;
+   }
+}
+
+void
+statemap_ormap (statemap_t *dst, statemap_t *src)
+{
+   int       size       = dst->size;
+   int       offset     = 0;
+   bitmap_t *changemap0 = dst->changemap0;
+   bitmap_t *changemap1 = dst->changemap1;
+   bitmap_t *changemap2 = dst->changemap2;
+   bitmap_t *dstmap     = dst->bitmap;
+   bitmap_t *srcmap     = src->bitmap;
+   bitmap_t  bit0;
+   bitmap_t  bit1;
+   bitmap_t  bit2;
+   bitmap_t  seg;
+
+   ASSERT (src->size == size);
+   
+   for (bit0 = 1; offset < size; bit0 <<= 1, changemap1++)
+   {
+      for (bit1 = 1; bit1 != 0 && offset < size; bit1 <<= 1, changemap2++)
+      {
+        for (bit2 = 1; bit2 != 0 && offset < size; bit2 <<= 1, dstmap++, srcmap++, offset += BT_NBIPUL)
+        {
+           seg = *dstmap | *srcmap;
+           if (*dstmap != seg)
+           {
+              *dstmap = seg;
+              *changemap2 |= bit2;
+           }
+        }
+        if (*changemap2 != 0)
+           *changemap1 |= bit1;
+      }
+      if (*changemap1 != 0)
+        *changemap0 |= bit0;
+   }
+}
+
+int
+statemap_findchange (statemap_t *map, bitmap_t *newseg, int clearchange)
+{
+   int          bit0;
+   bitmap_t    *cm1;
+   int          bit1;
+   bitmap_t    *cm2;
+   int          bit2;
+   unsigned int offset;
+
+   bit0 = statemap_firstsegbit (*(map->changemap0));
+   if (bit0 < 0)
+      return (-1);
+
+   offset = bit0;
+   cm1 = map->changemap1 + offset;
+   bit1 = statemap_firstsegbit (*cm1);
+   ASSERT (bit1 >= 0);
+
+   offset = (offset << BT_ULSHIFT) + bit1;
+   cm2 = map->changemap2 + offset;
+   bit2 = statemap_firstsegbit (*cm2);
+   ASSERT (bit2 >= 0);
+   
+   offset = (offset << BT_ULSHIFT) + bit2;
+   *newseg = map->bitmap[offset];
+
+   if (clearchange &&
+       (*cm2 &= ~(((bitmap_t)1) << bit2)) == 0 &&
+       (*cm1 &= ~(((bitmap_t)1) << bit1)) == 0)
+      map->changemap0[0] &= ~(((bitmap_t)1) << bit0);
+
+   return (offset << BT_ULSHIFT);
+}
+
+int
+statemap_changed (statemap_t *map)
+{
+   return ((*(map->changemap0) != 0));
+}
+
+void
+statemap_reset (statemap_t *map)
+{
+   bzero (map->changemap0, map->changemap_nob + map->bitmap_nob);
+}
+
+void
+statemap_copy (statemap_t *dst, statemap_t *src)
+{
+   ASSERT (dst->size == src->size);
+   bcopy (src->changemap0, dst->changemap0, src->changemap_nob + src->bitmap_nob);
+}
+
+void
+statemap_clearchanges (statemap_t *map)
+{
+   if (statemap_changed (map))
+      bzero (map->changemap0, map->changemap_nob);
+}
+
+bitmap_t *
+statemap_tobitmap (statemap_t *map)
+{
+    return (map->bitmap);
+}
+
+statemap_t *
+statemap_create (int size)
+{
+   int   struct_entries     = (sizeof (statemap_t) * 8 + (BT_NBIPUL-1)) >> BT_ULSHIFT;
+   int   bitmap_entries     = (size + (BT_NBIPUL-1)) >> BT_ULSHIFT;
+   int   changemap2_entries = (bitmap_entries + (BT_NBIPUL-1)) >> BT_ULSHIFT;
+   int   changemap1_entries = (changemap2_entries + (BT_NBIPUL-1)) >> BT_ULSHIFT;
+   int   changemap0_entries = (changemap1_entries + (BT_NBIPUL-1)) >> BT_ULSHIFT;
+   int   changemap_entries  = changemap0_entries + changemap1_entries + changemap2_entries;
+   int   nob                = (struct_entries + bitmap_entries + changemap_entries) * sizeof (bitmap_t);
+   statemap_t *map;
+
+   ASSERT ((1 << BT_ULSHIFT) == BT_NBIPUL);
+   ASSERT (changemap0_entries == 1);
+
+   KMEM_ZALLOC (map, statemap_t *, nob, 1);
+
+   map->size = size;
+   map->nob  = nob;
+   map->changemap_nob = changemap_entries * sizeof (bitmap_t);
+   map->bitmap_nob = bitmap_entries * sizeof (bitmap_t);
+   map->changemap0 = ((bitmap_t *)map) + struct_entries;
+   map->changemap1 = map->changemap0 + changemap0_entries;
+   map->changemap2 = map->changemap1 + changemap1_entries;
+   map->bitmap     = map->changemap2 + changemap2_entries;
+
+   return (map);
+}
+
+void
+statemap_destroy (statemap_t *map)
+{
+   KMEM_FREE (map, map->nob);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/statusmon.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/statusmon.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/statusmon.h       2005-06-01 23:12:54.682426488 -0400
@@ -0,0 +1,44 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: statusmon.h,v 1.6 2003/10/07 13:22:38 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/statusmon.h,v $*/
+
+#ifndef __ELAN3_STATUSMON_H
+#define __ELAN3_STATUSMON_H
+
+typedef struct statusmon_node
+{
+    u_int      NodeId;
+    u_int      State;
+} STATUSMON_SGMT;
+
+typedef struct statusmon_level
+{
+    unsigned      Width;
+    STATUSMON_SGMT Nodes[CM_SGMTS_PER_LEVEL];
+} STATUSMON_LEVEL;
+
+typedef struct statusmon_msg
+{
+    unsigned       Type;
+    unsigned       NodeId;
+    unsigned       NumLevels;
+    unsigned       TopLevel;
+    unsigned        Role;
+    STATUSMON_LEVEL Levels[CM_MAX_LEVELS];
+} STATUSMON_MSG;
+
+
+#endif /* __ELAN3_STATUSMON_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/support.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/support.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/support.c 2005-06-01 23:12:54.683426336 -0400
@@ -0,0 +1,109 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: support.c,v 1.37.8.1 2004/09/30 15:01:53 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/support.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <elan/kcomm.h>
+
+/****************************************************************************************/
+/*
+ * Nodeset/flush callbacks.
+ */
+int
+ep_register_callback (EP_RAIL *rail, unsigned idx, void (*routine)(void *, statemap_t *), void *arg)
+{
+    EP_CALLBACK *cb;
+    
+    KMEM_ALLOC (cb, EP_CALLBACK *, sizeof (EP_CALLBACK), 1);
+    
+    cb->Routine = routine;
+    cb->Arg     = arg;
+
+    kmutex_lock (&rail->CallbackLock);
+    cb->Next = rail->CallbackList[idx];
+    rail->CallbackList[idx] = cb;
+    kmutex_unlock (&rail->CallbackLock);
+    
+    return (ESUCCESS);
+}
+
+void
+ep_remove_callback (EP_RAIL *rail, unsigned idx, void (*routine)(void *, statemap_t *), void *arg)
+{
+    EP_CALLBACK  *cb;
+    EP_CALLBACK **predp;
+
+    kmutex_lock (&rail->CallbackLock);
+    for (predp = &rail->CallbackList[idx]; (cb = *predp); predp = &cb->Next)
+       if (cb->Routine == routine && cb->Arg == arg)
+           break;
+
+    if (cb == NULL)
+       panic ("ep_remove_member_callback");
+    
+    *predp = cb->Next;
+    kmutex_unlock (&rail->CallbackLock);
+    
+    KMEM_FREE (cb, sizeof (EP_CALLBACK));
+}
+
+void
+ep_call_callbacks (EP_RAIL *rail, unsigned idx, statemap_t *map)
+{
+    EP_CALLBACK *cb;
+
+    kmutex_lock (&rail->CallbackLock);
+
+    rail->CallbackStep = idx;
+
+    for (cb = rail->CallbackList[idx]; cb; cb = cb->Next) {
+       (cb->Routine) (cb->Arg, map);
+    }
+    kmutex_unlock (&rail->CallbackLock);
+}
+
+unsigned int
+ep_backoff (EP_BACKOFF *backoff, int type)
+{
+    static int bcount[EP_NUM_BACKOFF] = {1, 16, 32, 64, 128, 256, 512, 1024};
+    
+    if (backoff->type != type)
+    {
+       backoff->type  = type;
+       backoff->indx  = 0;
+       backoff->count = 0;
+    }
+
+    if (++backoff->count > bcount[backoff->indx] && backoff->indx < (EP_NUM_BACKOFF-1))
+    {
+       backoff->indx++;
+       backoff->count = 0;
+    }
+
+    return (backoff->indx);
+}
+
+/* Generic checksum algorithm */
+uint16_t
+CheckSum (char *msg, int nob)
+{
+    uint16_t sum = 0;
+   
+    while (nob-- > 0)
+       sum = sum * 13 + *msg++;
+
+    return (sum);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/support_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/support_elan3.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/support_elan3.c   2005-06-01 23:12:54.687425728 -0400
@@ -0,0 +1,2111 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: support_elan3.c,v 1.42.8.3 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/support_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+
+#include <elan/kcomm.h>
+#include <elan/epsvc.h>
+#include <elan/epcomms.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan3.h"
+#include "epcomms_elan3.h"
+#include "debug.h"
+
+#include <elan3/thread.h>
+#include <elan3/urom_addrs.h>
+
+/****************************************************************************************/
+#define DMA_RING_NEXT_POS(ring)      ((ring)->Position+1 == ring->Entries ? 0 : ((ring)->Position+1))
+#define DMA_RING_PREV_POS(ring,pos)  ((pos) == 0 ? (ring)->Entries-1 : (pos) - 1)
+
+static int 
+DmaRingCreate (EP3_RAIL *rail, EP3_DMA_RING *ring, int ctxnum, int entries)
+{
+    unsigned long pgnum = (ctxnum * sizeof (E3_CommandPort)) / PAGE_SIZE;
+    unsigned long pgoff = (ctxnum * sizeof (E3_CommandPort)) & (PAGE_SIZE-1);
+    int           s;    
+        
+    /* set up the initial position */
+    ring->Entries  = entries;
+    ring->Position = 0;
+    
+    if (! (ring->pEvent = ep_alloc_elan (&rail->Generic, entries * sizeof (E3_BlockCopyEvent), 0, &ring->epEvent)))
+    {
+       ring->CommandPort = (ioaddr_t) NULL;
+       return (ENOMEM);
+    }
+    
+    if (! (ring->pDma = ep_alloc_elan (&rail->Generic, entries * sizeof (E3_DMA), 0, &ring->epDma)))
+    {
+       ep_free_elan (&rail->Generic, ring->epEvent, entries * sizeof (E3_BlockCopyEvent));
+
+       ring->CommandPort = (ioaddr_t) NULL;
+       return (ENOMEM);
+    }
+    
+    if (! (ring->pDoneBlk = ep_alloc_main (&rail->Generic, entries * sizeof (E3_uint32), 0, &ring->epDoneBlk)))
+    {
+       ep_free_elan (&rail->Generic, ring->epEvent, entries * sizeof (E3_BlockCopyEvent));
+       ep_free_elan (&rail->Generic, ring->epDma,   entries * sizeof (E3_DMA));
+
+       ring->CommandPort = (ioaddr_t) NULL;
+       return (ENOMEM);
+    }
+    
+    if (MapDeviceRegister (rail->Device, ELAN3_BAR_COMMAND_PORT, &ring->CommandPage, pgnum * PAGE_SIZE, PAGE_SIZE, &ring->CommandPageHandle) != ESUCCESS)
+    {
+       ep_free_elan (&rail->Generic, ring->epEvent,   entries * sizeof (E3_BlockCopyEvent));
+       ep_free_elan (&rail->Generic, ring->epDma,     entries * sizeof (E3_DMA));
+       ep_free_main (&rail->Generic, ring->epDoneBlk, entries * sizeof (E3_uint32));
+
+       ring->CommandPort = (ioaddr_t) NULL;
+       return (ENOMEM);
+    }
+    ring->CommandPort = ring->CommandPage + pgoff;
+       
+    for (s = 0; s < entries; s++)
+    {
+       /* setup the event */
+       elan3_sdram_writel(rail->Device, DMA_RING_EVENT(ring,s) + offsetof(E3_BlockCopyEvent,ev_Type),   
+                          EV_TYPE_BCOPY | EV_TYPE_DMA | DMA_RING_DMA_ELAN(ring, s));
+       elan3_sdram_writel(rail->Device, DMA_RING_EVENT(ring,s) + offsetof(E3_BlockCopyEvent,ev_Source), DMA_RING_DMA_ELAN(ring,s)  | EV_WCOPY);
+       elan3_sdram_writel(rail->Device, DMA_RING_EVENT(ring,s) + offsetof(E3_BlockCopyEvent,ev_Dest),   DMA_RING_DONE_ELAN(ring,s) | EV_TYPE_BCOPY_WORD );         
+
+       /* need to set all the doneBlks to appear that they have completed */
+       ring->pDoneBlk[s] = DMA_RING_DMA_ELAN(ring,s)  | EV_WCOPY;
+    }
+
+    return 0; /* success */
+}
+
+static void
+DmaRingRelease(EP3_RAIL *rail, EP3_DMA_RING *ring)
+{
+    if (ring->CommandPage != (ioaddr_t) 0)
+    {
+       UnmapDeviceRegister(rail->Device, &ring->CommandPageHandle);
+
+       ep_free_elan (&rail->Generic, ring->epEvent,   ring->Entries * sizeof (E3_BlockCopyEvent));
+       ep_free_elan (&rail->Generic, ring->epDma,     ring->Entries * sizeof (E3_DMA));
+       ep_free_main (&rail->Generic, ring->epDoneBlk, ring->Entries * sizeof (E3_uint32));
+    }
+    ring->CommandPage = (ioaddr_t) 0;
+}
+
+void 
+DmaRingsRelease (EP3_RAIL *rail)
+{
+    DmaRingRelease (rail, &rail->DmaRings[EP3_RING_CRITICAL]);
+    DmaRingRelease (rail, &rail->DmaRings[EP3_RING_HIGH_PRI]);
+    DmaRingRelease (rail, &rail->DmaRings[EP3_RING_LOW_PRI]);
+}
+
+int 
+DmaRingsCreate (EP3_RAIL *rail)
+{
+    if (DmaRingCreate (rail, &rail->DmaRings[EP3_RING_CRITICAL], ELAN3_DMARING_BASE_CONTEXT_NUM + EP3_RING_CRITICAL, EP3_RING_CRITICAL_LEN) ||
+       DmaRingCreate (rail, &rail->DmaRings[EP3_RING_HIGH_PRI], ELAN3_DMARING_BASE_CONTEXT_NUM + EP3_RING_HIGH_PRI, EP3_RING_HIGH_PRI_LEN) ||
+       DmaRingCreate (rail, &rail->DmaRings[EP3_RING_LOW_PRI],  ELAN3_DMARING_BASE_CONTEXT_NUM + EP3_RING_LOW_PRI,  EP3_RING_LOW_PRI_LEN))
+    {
+       DmaRingsRelease (rail);
+       return (ENOMEM);
+    }
+  
+    return 0;
+}
+
+static int 
+DmaRingNextSlot (EP3_DMA_RING *ring)
+{
+    int pos  = ring->Position;
+    int npos = DMA_RING_NEXT_POS(ring);
+
+    if (ring->pDoneBlk[npos] == EP3_EVENT_ACTIVE)
+       return (-1);
+    
+    ring->pDoneBlk[pos] = EP3_EVENT_ACTIVE;
+
+    ring->Position = npos; /* move on one */
+
+    return (pos);
+}
+
+
+/****************************************************************************************/
+/*
+ * Dma/event command issueing - these handle cproc queue overflow traps.
+ */
+static int
+DmaRunQueueSizeCheck (EP3_RAIL *rail, E3_uint32 len)
+{
+    E3_uint64  FandBPtr = read_reg64 (rail->Device, DProc_SysCntx_FPtr);
+    E3_uint32  FPtr, BPtr;
+    E3_uint32  qlen;
+
+#if (BYTE_ORDER == LITTLE_ENDIAN) || defined(__LITTLE_ENDIAN__)
+    FPtr = (FandBPtr & 0xFFFFFFFFull);
+    BPtr = (FandBPtr >> 32);
+#else
+    FPtr = (FandBPtr >> 32);
+    BPtr = (FandBPtr & 0xFFFFFFFFull);
+#endif
+    
+    qlen = (((BPtr - FPtr)/sizeof (E3_DMA)) & (E3_SysCntxQueueSize-1));
+    
+    if      (qlen < 4)   IncrStat (rail, DmaQueueLength[0]);
+    else if (qlen < 8)   IncrStat (rail, DmaQueueLength[1]);
+    else if (qlen < 16)  IncrStat (rail, DmaQueueLength[2]);
+    else if (qlen < 32)  IncrStat (rail, DmaQueueLength[3]);
+    else if (qlen < 64)  IncrStat (rail, DmaQueueLength[4]);
+    else if (qlen < 128) IncrStat (rail, DmaQueueLength[5]);
+    else if (qlen < 240) IncrStat (rail, DmaQueueLength[6]);
+    else                 IncrStat (rail, DmaQueueLength[7]);
+       
+    return (qlen < len);
+}
+
+int
+IssueDma (EP3_RAIL *rail, E3_DMA_BE * dmabe, int type, int retryThread)
+{
+    ELAN3_DEV     *dev = rail->Device;
+    EP3_RETRY_DMA *retry;
+    EP3_DMA_RING  *ring;
+    int           slot;
+    int           i, res;
+    unsigned long flags;
+
+    ASSERT (dmabe->s.dma_direction == DMA_WRITE || dmabe->s.dma_direction == DMA_READ_REQUEUE);
+
+    ASSERT (! EP_VP_ISDATA(dmabe->s.dma_destVProc) ||
+           (dmabe->s.dma_direction == DMA_WRITE ? 
+            EP_VP_TO_NODE(dmabe->s.dma_srcVProc) == rail->Generic.Position.pos_nodeid :
+            EP_VP_TO_NODE(dmabe->s.dma_destVProc) == rail->Generic.Position.pos_nodeid));
+    
+    /*
+     * If we're not the retry thread - then don't issue this DMA
+     * if there are any already queued on the retry lists with
+     * higher or equal priority than this one that are ready to
+     * retry.
+     */
+    if (! retryThread)
+    {
+       for (i = EP_RETRY_BASE; i < type; i++)
+       {
+           if (list_empty (&rail->DmaRetries[i]))
+               continue;
+
+           retry = list_entry (rail->DmaRetries[i].next, EP3_RETRY_DMA, Link);
+               
+           if (AFTER (lbolt, retry->RetryTime))
+           {
+               IncrStat (rail, IssueDmaFail[type]);
+               return (ISSUE_COMMAND_RETRY);
+           }
+       }
+    }
+
+    /*
+     * Depending on the type of DMA we're issuing - throttle back
+     * issueing of it if the DMA run queue is too full.  This then
+     * prioritises the "special" messages and completing data 
+     * transfers which have matched a receive buffer.
+     */
+
+    if (type >= EP_RETRY_LOW_PRI_RETRY)
+    {
+       if (! DmaRunQueueSizeCheck (rail, E3_SysCntxQueueSize / 2))
+       {
+           IncrStat (rail, IssueDmaFail[type]);
+           return (ISSUE_COMMAND_RETRY);
+       }
+       ring = &rail->DmaRings[EP3_RING_LOW_PRI];
+    } 
+    else if (type == EP_RETRY_LOW_PRI)
+    {
+       if (! DmaRunQueueSizeCheck (rail, E3_SysCntxQueueSize / 3))
+       {
+           IncrStat (rail, IssueDmaFail[type]);
+           return (ISSUE_COMMAND_RETRY);
+       }
+       ring = &rail->DmaRings[EP3_RING_LOW_PRI];
+    }
+    else if (type >= EP_RETRY_HIGH_PRI)
+       ring = &rail->DmaRings[EP3_RING_HIGH_PRI];
+    else
+       ring = &rail->DmaRings[EP3_RING_CRITICAL];
+
+    local_irq_save (flags);
+    if (! spin_trylock (&dev->CProcLock))
+    {
+       IncrStat (rail, IssueDmaFail[type]);
+
+       res = ISSUE_COMMAND_RETRY;
+    }
+    else
+    {
+       if ((slot = DmaRingNextSlot (ring)) == -1)
+       {
+           IncrStat (rail, IssueDmaFail[type]);
+           
+           res = ISSUE_COMMAND_RETRY;
+       }
+       else
+       {
+           EPRINTF4 (DBG_COMMAND, "IssueDma: type %08x size %08x Elan source %08x Elan dest %08x\n",
+                     dmabe->s.dma_type, dmabe->s.dma_size, dmabe->s.dma_source, dmabe->s.dma_dest);
+           EPRINTF2 (DBG_COMMAND, "          dst event %08x cookie/proc %08x\n",
+                     dmabe->s.dma_destEvent, dmabe->s.dma_destCookieVProc);
+           EPRINTF2 (DBG_COMMAND, "          src event %08x cookie/proc %08x\n",
+                     dmabe->s.dma_srcEvent, dmabe->s.dma_srcCookieVProc);
+
+           elan3_sdram_copyq_to_sdram (dev,  dmabe,  DMA_RING_DMA(ring, slot), sizeof (E3_DMA));                       /* PCI write block */
+           elan3_sdram_writel (dev, DMA_RING_EVENT(ring, slot) + offsetof (E3_BlockCopyEvent, ev_Count), 1);   /* PCI write */
+           
+           mb();                                                               /* ensure writes to main memory completed */
+           writel (DMA_RING_EVENT_ELAN(ring,slot), ring->CommandPort + offsetof (E3_CommandPort, SetEvent));
+           mmiob();                                                            /* and flush through IO writes */
+           
+           res = ISSUE_COMMAND_OK;
+       }
+       spin_unlock (&dev->CProcLock);
+    }
+    local_irq_restore (flags);
+
+    return (res);
+}
+
+int
+IssueWaitevent (EP3_RAIL *rail, E3_Addr value)
+{
+    ELAN3_DEV     *dev   = rail->Device;
+    int           res;
+    unsigned long flags;
+    
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    ASSERT (rail->CommandPortEventTrap == FALSE);
+
+    /*
+     * Disable the command processor interrupts, so that we don't see
+     * spurious interrupts appearing.
+     */
+    DISABLE_INT_MASK (dev, INT_CProc | INT_ComQueue);
+
+    EPRINTF1 (DBG_COMMAND, "IssueWaitevent: %08x\n", value);
+
+    mb();                                                              /* ensure writes to main memory completed */
+    writel (value, rail->CommandPort + offsetof (E3_CommandPort, WaitEvent0));
+    mmiob();                                                           /* and flush through IO writes */
+    
+    do {
+       res = CheckCommandQueueFlushed (rail->Ctxt, EventComQueueNotEmpty, ISSUE_COMMAND_CANT_WAIT, &flags);
+
+       EPRINTF1 (DBG_COMMAND, "IssueWaitevent: CheckCommandQueueFlushed -> %d\n", res);
+
+       if (res == ISSUE_COMMAND_WAIT)
+           HandleCProcTrap (dev, 0, NULL);
+    } while (res != ISSUE_COMMAND_OK);
+
+    if (! rail->CommandPortEventTrap)
+       res = ISSUE_COMMAND_OK;
+    else
+    {
+       rail->CommandPortEventTrap = FALSE;
+       res = ISSUE_COMMAND_TRAPPED;
+    }
+
+    EPRINTF1 (DBG_COMMAND, "IssueWaitevent: -> %d\n", res);
+
+    /*
+     * Re-enable the command processor interrupt as we've finished 
+     * polling it.
+     */
+    ENABLE_INT_MASK (dev, INT_CProc | INT_ComQueue);
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+
+    return (res);
+}
+
+void
+IssueSetevent (EP3_RAIL *rail, E3_Addr value)
+{
+    EPRINTF1 (DBG_COMMAND, "IssueSetevent: %08x\n", value);
+
+    mb();                                                              /* ensure writes to main memory completed */
+    writel (value, rail->CommandPort + offsetof (E3_CommandPort, SetEvent));
+    mmiob();                                                           /* and flush through IO writes */
+}
+
+void
+IssueRunThread (EP3_RAIL *rail, E3_Addr value)
+{
+    EPRINTF1 (DBG_COMMAND, "IssueRunThread: %08x\n", value);
+
+    mb();                                                              /* ensure writes to main memory completed */
+    writel (value, rail->CommandPort + offsetof (E3_CommandPort, RunThread));
+    mmiob();                                                           /* and flush through IO writes */
+}
+
+/****************************************************************************************/
+/*
+ * DMA retry list management
+ */
+static unsigned DmaRetryTimes[EP_NUM_RETRIES]; 
+
+static void
+ep3_dma_retry (EP3_RAIL *rail)
+{
+    EP3_COOKIE    *cp;
+    int            res;
+    int                   vp;
+    unsigned long  flags;
+    int            i;
+
+    kernel_thread_init("ep3_dma_retry");
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+
+    for (;;)
+    {
+       long yieldAt   = lbolt + (hz/10);
+       long retryTime = 0;
+
+       if (rail->DmaRetryThreadShouldStop)
+           break;
+       
+       for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+       {
+           while (! list_empty (&rail->DmaRetries[i]))
+           {
+               EP3_RETRY_DMA *retry = list_entry (rail->DmaRetries[i].next, EP3_RETRY_DMA, Link);
+
+               if (! AFTER (lbolt, retry->RetryTime))
+                   break;
+               
+               if (rail->DmaRetryThreadShouldStall || AFTER (lbolt, yieldAt))
+                   goto cant_do_more;
+
+               EPRINTF2 (DBG_RETRY, "%s: DmaRetryThread: retry %p\n", rail->Generic.Name, retry);
+               EPRINTF5 (DBG_RETRY, "%s:                 %08x %08x %08x %08x\n",
+                         rail->Generic.Name, retry->Dma.s.dma_type, retry->Dma.s.dma_size, retry->Dma.s.dma_source, retry->Dma.s.dma_dest);
+               EPRINTF5 (DBG_RETRY, "%s:                 %08x %08x %08x %08x\n",
+                         rail->Generic.Name, retry->Dma.s.dma_destEvent, retry->Dma.s.dma_destCookieVProc,
+                         retry->Dma.s.dma_srcEvent, retry->Dma.s.dma_srcCookieVProc);
+#if defined(DEBUG)
+               if (retry->Dma.s.dma_direction == DMA_WRITE)
+                   cp = LookupEventCookie (rail, &rail->CookieTable, retry->Dma.s.dma_srcEvent);
+               else
+                   cp = LookupEventCookie (rail, &rail->CookieTable, retry->Dma.s.dma_destEvent);
+
+               ASSERT (cp != NULL || (retry->Dma.s.dma_srcEvent == 0 && retry->Dma.s.dma_direction == DMA_WRITE && retry->Dma.s.dma_isRemote));
+               
+               if (cp && cp->Operations->DmaVerify)
+                   cp->Operations->DmaVerify (rail, cp->Arg, &retry->Dma);
+#endif
+
+#if defined(DEBUG_ASSERT)
+               if (retry->Dma.s.dma_direction == DMA_WRITE)
+                   vp = retry->Dma.s.dma_destVProc;
+               else
+                   vp = retry->Dma.s.dma_srcVProc;
+
+               ASSERT (!EP_VP_ISDATA(vp) || 
+                       (rail->Generic.Nodes[EP_VP_TO_NODE(vp)].State >= EP_NODE_CONNECTED &&
+                        rail->Generic.Nodes[EP_VP_TO_NODE(vp)].State <= EP_NODE_LOCAL_PASSIVATE));
+#endif
+               spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+               res = IssueDma (rail, &(retry->Dma), i, TRUE);
+               spin_lock_irqsave (&rail->DmaRetryLock, flags);
+               
+               if (res != ISSUE_COMMAND_OK)
+                   goto cant_do_more;
+               
+               /* Command issued, so remove from list, and add to free list */
+               list_del (&retry->Link);
+               list_add (&retry->Link, &rail->DmaRetryFreeList);
+           }
+       }
+    cant_do_more:
+       
+       for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+       {
+           if (!list_empty (&rail->DmaRetries[i]))
+           {
+               EP3_RETRY_DMA *retry = list_entry (rail->DmaRetries[i].next, EP3_RETRY_DMA, Link);
+
+               retryTime = retryTime ? MIN(retryTime, retry->RetryTime) : retry->RetryTime;
+           }
+       }
+
+       if (retryTime && !AFTER (retryTime, lbolt))
+           retryTime = lbolt + 1;
+
+       do {
+           EPRINTF3 (DBG_RETRY, "%s: ep_cm_retry: %s %lx\n", rail->Generic.Name, rail->DmaRetryThreadShouldStall ? "stalled" : "sleeping", retryTime);
+           
+           if (rail->DmaRetryTime == 0 || (retryTime != 0 && retryTime < rail->DmaRetryTime))
+               rail->DmaRetryTime = retryTime;
+           
+           rail->DmaRetrySleeping = TRUE;
+           
+           if (rail->DmaRetryThreadShouldStall)                                        /* wakeup threads waiting in StallDmaRetryThread */
+               kcondvar_wakeupall (&rail->DmaRetryWait, &rail->DmaRetryLock);  /* for us to really go to sleep for good. */
+
+           if (rail->DmaRetryTime == 0 || rail->DmaRetryThreadShouldStall)
+               kcondvar_wait (&rail->DmaRetryWait, &rail->DmaRetryLock, &flags);
+           else
+               kcondvar_timedwait (&rail->DmaRetryWait, &rail->DmaRetryLock, &flags, rail->DmaRetryTime);
+
+           rail->DmaRetrySleeping = FALSE;
+
+       } while (rail->DmaRetryThreadShouldStall);
+
+       rail->DmaRetryTime = 0;
+    }
+
+    rail->DmaRetryThreadStopped = 1;
+    kcondvar_wakeupall (&rail->DmaRetryWait, &rail->DmaRetryLock);
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+
+    kernel_thread_exit();
+}
+
+void
+StallDmaRetryThread (EP3_RAIL *rail)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    rail->DmaRetryThreadShouldStall++;
+
+    while (! rail->DmaRetrySleeping)
+       kcondvar_wait (&rail->DmaRetryWait, &rail->DmaRetryLock, &flags);
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+}
+
+void 
+ResumeDmaRetryThread (EP3_RAIL *rail)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+
+    ASSERT (rail->DmaRetrySleeping);
+
+    if (--rail->DmaRetryThreadShouldStall == 0)
+    {
+       rail->DmaRetrySleeping = 0;
+       kcondvar_wakeupone (&rail->DmaRetryWait, &rail->DmaRetryLock);
+    }
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+}
+
+int
+InitialiseDmaRetries (EP3_RAIL *rail)
+{
+    int i;
+
+    spin_lock_init (&rail->DmaRetryLock);
+    kcondvar_init (&rail->DmaRetryWait);
+
+    for (i = 0; i < EP_NUM_RETRIES; i++)
+       INIT_LIST_HEAD (&rail->DmaRetries[i]);
+
+    INIT_LIST_HEAD (&rail->DmaRetryFreeList);
+
+    DmaRetryTimes[EP_RETRY_HIGH_PRI]  = EP_RETRY_HIGH_PRI_TIME;
+
+    for (i =0 ; i < EP_NUM_BACKOFF; i++)
+       DmaRetryTimes[EP_RETRY_HIGH_PRI_RETRY+i] = EP_RETRY_HIGH_PRI_TIME << i;
+    
+    DmaRetryTimes[EP_RETRY_LOW_PRI] = EP_RETRY_LOW_PRI_TIME;
+
+    for (i =0 ; i < EP_NUM_BACKOFF; i++)
+       DmaRetryTimes[EP_RETRY_LOW_PRI_RETRY+i] = EP_RETRY_LOW_PRI_TIME << i;
+    
+    DmaRetryTimes[EP_RETRY_ANONYMOUS] = EP_RETRY_ANONYMOUS_TIME;
+    DmaRetryTimes[EP_RETRY_NETERR]    = EP_RETRY_NETERR_TIME;
+
+    rail->DmaRetryInitialised = 1;
+
+    if (kernel_thread_create (ep3_dma_retry, (void *) rail) == 0)
+    {
+       spin_lock_destroy (&rail->DmaRetryLock);
+       return (ENOMEM);
+    }
+
+    rail->DmaRetryThreadStarted = 1;
+
+    return (ESUCCESS);
+}
+
+void
+DestroyDmaRetries (EP3_RAIL *rail)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    rail->DmaRetryThreadShouldStop = 1;
+    while (rail->DmaRetryThreadStarted && !rail->DmaRetryThreadStopped)
+    {
+       kcondvar_wakeupall (&rail->DmaRetryWait, &rail->DmaRetryLock);
+       kcondvar_wait (&rail->DmaRetryWait, &rail->DmaRetryLock, &flags);
+    }
+    rail->DmaRetryThreadStarted = 0;
+    rail->DmaRetryThreadStopped = 0;
+    rail->DmaRetryThreadShouldStop = 0;
+    rail->DmaRetryInitialised = 0;
+
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+
+    /* Everyone should have given back their retry dma's by now */
+    ASSERT (rail->DmaRetryReserved == 0);
+
+    while (! list_empty (&rail->DmaRetryFreeList))
+    {
+       EP3_RETRY_DMA *retry = list_entry (rail->DmaRetryFreeList.next, EP3_RETRY_DMA, Link);
+       
+       list_del (&retry->Link);
+
+       KMEM_FREE (retry, sizeof (EP3_RETRY_DMA));
+    }
+
+    kcondvar_destroy (&rail->DmaRetryWait);
+    spin_lock_destroy (&rail->DmaRetryLock);
+}
+
+int
+ReserveDmaRetries (EP3_RAIL *rail, int count, EP_ATTRIBUTE attr)
+{
+    EP3_RETRY_DMA *retry;
+    int                  remaining = count;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    
+    if (remaining <= (rail->DmaRetryCount - rail->DmaRetryReserved))
+    {
+       rail->DmaRetryReserved += remaining;
+
+       spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+       return (ESUCCESS);
+    }
+
+    remaining -= (rail->DmaRetryCount - rail->DmaRetryReserved);
+
+    rail->DmaRetryReserved = rail->DmaRetryCount;
+
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+
+    while (remaining)
+    {
+       KMEM_ALLOC (retry, EP3_RETRY_DMA *, sizeof (EP3_RETRY_DMA), !(attr & EP_NO_SLEEP));
+       
+       if (retry == NULL)
+           goto failed;
+
+       /* clear E3_DMA */
+       bzero((char *)(&(retry->Dma.s)), sizeof(E3_DMA));
+
+       remaining--; 
+
+       spin_lock_irqsave (&rail->DmaRetryLock, flags);
+
+       list_add (&retry->Link, &rail->DmaRetryFreeList);
+
+       rail->DmaRetryCount++;
+       rail->DmaRetryReserved++;
+
+       spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+    }
+    return (ESUCCESS);
+
+ failed:
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    rail->DmaRetryReserved -= (count - remaining);
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+    return (ENOMEM);
+}
+
+void
+ReleaseDmaRetries (EP3_RAIL *rail, int count)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    rail->DmaRetryReserved -= count;
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+}
+
+void
+QueueDmaForRetry (EP3_RAIL *rail, E3_DMA_BE *dma, int interval)
+{
+    EP3_RETRY_DMA *retry;
+    unsigned long flags;
+
+    /*
+     * When requeueing DMAs they must never be "READ" dma's since
+     * these would fetch the DMA descriptor from the retryn descriptor
+     */
+    ASSERT (dma->s.dma_direction == DMA_WRITE || dma->s.dma_direction == DMA_READ_REQUEUE);
+    ASSERT (dma->s.dma_direction == DMA_WRITE ? 
+           EP_VP_TO_NODE(dma->s.dma_srcVProc) == rail->Generic.Position.pos_nodeid :
+           EP_VP_TO_NODE(dma->s.dma_destVProc) == rail->Generic.Position.pos_nodeid);
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    
+    EP_ASSERT (&rail->Generic, !list_empty (&rail->DmaRetryFreeList));
+
+    /* take an item of the free list */
+    retry = list_entry (rail->DmaRetryFreeList.next, EP3_RETRY_DMA, Link);
+
+    list_del (&retry->Link);
+    
+    EPRINTF5 (DBG_RETRY, "%s: QueueDmaForRetry: %08x %08x %08x %08x\n", rail->Generic.Name,
+             dma->s.dma_type, dma->s.dma_size, dma->s.dma_source, dma->s.dma_dest);
+    EPRINTF5 (DBG_RETRY, "%s:                   %08x %08x %08x %08x\n",rail->Generic.Name,
+            dma->s.dma_destEvent, dma->s.dma_destCookieVProc,
+            dma->s.dma_srcEvent, dma->s.dma_srcCookieVProc);
+
+    /* copy the DMA into the retry descriptor */
+    retry->Dma.s.dma_type            = dma->s.dma_type;
+    retry->Dma.s.dma_size            = dma->s.dma_size;
+    retry->Dma.s.dma_source          = dma->s.dma_source;
+    retry->Dma.s.dma_dest            = dma->s.dma_dest;
+    retry->Dma.s.dma_destEvent       = dma->s.dma_destEvent;
+    retry->Dma.s.dma_destCookieVProc = dma->s.dma_destCookieVProc;
+    retry->Dma.s.dma_srcEvent        = dma->s.dma_srcEvent;
+    retry->Dma.s.dma_srcCookieVProc  = dma->s.dma_srcCookieVProc;
+
+    retry->RetryTime = lbolt + DmaRetryTimes[interval];
+
+    /* chain onto the end of the approriate retry list */
+    list_add_tail (&retry->Link, &rail->DmaRetries[interval]);
+
+    /* now wakeup the retry thread */
+    if (rail->DmaRetryTime == 0 || retry->RetryTime < rail->DmaRetryTime)
+       rail->DmaRetryTime = retry->RetryTime;
+    
+    if (rail->DmaRetrySleeping && !rail->DmaRetryThreadShouldStall)
+    {
+       rail->DmaRetrySleeping = 0;
+       kcondvar_wakeupone (&rail->DmaRetryWait, &rail->DmaRetryLock);
+    }
+
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+}
+
+void
+QueueDmaOnStalledList (EP3_RAIL *rail, E3_DMA_BE *dma)
+{
+    EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[dma->s.dma_direction == DMA_WRITE ? 
+                                                 EP_VP_TO_NODE(dma->s.dma_srcVProc) :
+                                                 EP_VP_TO_NODE(dma->s.dma_destVProc)];
+    EP3_RETRY_DMA *retry;
+    unsigned long flags;
+
+    /*
+     * When requeueing DMAs they must never be "READ" dma's since
+     * these would fetch the DMA descriptor from the retryn descriptor
+     */
+    ASSERT (dma->s.dma_direction == DMA_WRITE || dma->s.dma_direction == DMA_READ_REQUEUE);
+    ASSERT (dma->s.dma_direction == DMA_WRITE ? 
+           EP_VP_TO_NODE(dma->s.dma_srcVProc) == rail->Generic.Position.pos_nodeid :
+           EP_VP_TO_NODE(dma->s.dma_destVProc) == rail->Generic.Position.pos_nodeid);
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    
+    EP_ASSERT (&rail->Generic, !list_empty (&rail->DmaRetryFreeList));
+
+    /* take an item of the free list */
+    retry = list_entry (rail->DmaRetryFreeList.next, EP3_RETRY_DMA, Link);
+
+    list_del (&retry->Link);
+    
+    EPRINTF5 (DBG_RETRY, "%s: QueueDmaOnStalledList: %08x %08x %08x %08x\n", rail->Generic.Name,
+             dma->s.dma_type, dma->s.dma_size, dma->s.dma_source, dma->s.dma_dest);
+    EPRINTF5 (DBG_RETRY, "%s:                        %08x %08x %08x %08x\n", rail->Generic.Name,
+             dma->s.dma_destEvent, dma->s.dma_destCookieVProc,
+             dma->s.dma_srcEvent, dma->s.dma_srcCookieVProc);
+
+    /* copy the DMA into the retry descriptor */
+    retry->Dma.s.dma_type            = dma->s.dma_type;
+    retry->Dma.s.dma_size            = dma->s.dma_size;
+    retry->Dma.s.dma_source          = dma->s.dma_source;
+    retry->Dma.s.dma_dest            = dma->s.dma_dest;
+    retry->Dma.s.dma_destEvent       = dma->s.dma_destEvent;
+    retry->Dma.s.dma_destCookieVProc = dma->s.dma_destCookieVProc;
+    retry->Dma.s.dma_srcEvent        = dma->s.dma_srcEvent;
+    retry->Dma.s.dma_srcCookieVProc  = dma->s.dma_srcCookieVProc;
+
+    /* chain onto the node cancelled dma list */
+    list_add_tail (&retry->Link, &nodeRail->StalledDmas);
+
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+}
+
+void
+FreeStalledDmas (EP3_RAIL *rail, unsigned int nodeId)
+{
+    EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[nodeId];
+    struct list_head *el, *nel;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    list_for_each_safe (el, nel, &nodeRail->StalledDmas) {
+       list_del (el);
+       list_add (el, &rail->DmaRetryFreeList);
+    }
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+}
+
+/****************************************************************************************/
+/*
+ * Connection management.
+ */
+static void
+DiscardingHaltOperation (ELAN3_DEV *dev, void *arg)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) arg;
+    unsigned long flags;
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    rail->HaltOpCompleted = 1;
+    kcondvar_wakeupall (&rail->HaltOpSleep, &dev->IntrLock);
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+ 
+typedef struct {
+     EP3_RAIL  *rail;
+    sdramaddr_t qaddr;
+} SetQueueFullData;
+ 
+static void
+SetQueueLockedOperation (ELAN3_DEV *dev, void *arg)
+{
+    SetQueueFullData *data =  (SetQueueFullData *) arg;
+    unsigned long     flags;     
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+
+    elan3_sdram_writel  (dev, data->qaddr, E3_QUEUE_LOCKED | elan3_sdram_readl(dev, data->qaddr));
+   
+    data->rail->HaltOpCompleted = 1;
+    kcondvar_wakeupall (&data->rail->HaltOpSleep, &dev->IntrLock);
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+static void
+FlushDmaQueuesHaltOperation (ELAN3_DEV *dev, void *arg)
+{
+    EP3_RAIL      *rail    = (EP3_RAIL *) arg;
+    sdramaddr_t    FPtr, BPtr;
+    sdramaddr_t           Base, Top;
+    E3_DMA_BE      dma;
+    EP_NODE_RAIL  *node;
+    int            vp;
+    unsigned long  flags;
+
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProc.s.FSR)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData0.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData1.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData2.s.FSR.Status)) == 0);
+    ASSERT (elan3_sdram_readl (dev, dev->TAndQBase + offsetof (E3_TrapAndQueue, DProcData3.s.FSR.Status)) == 0);
+    
+    FPtr  = read_reg32 (dev, DProc_SysCntx_FPtr);
+    BPtr =  read_reg32 (dev, DProc_SysCntx_BPtr);
+    Base  = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[0]);
+    Top   = dev->TAndQBase + offsetof (E3_TrapAndQueue, SysCntxDmaQueue[E3_SysCntxQueueSize-1]);
+    
+    while (FPtr != BPtr)
+    {
+       elan3_sdram_copyq_from_sdram (dev, FPtr, &dma, sizeof (E3_DMA_BE));
+       
+       EPRINTF5 (DBG_DISCON, "%s: FlushDmaQueuesHaltOperation: %08x %08x %08x %08x\n", rail->Generic.Name,
+                 dma.s.dma_type, dma.s.dma_size, dma.s.dma_source, dma.s.dma_dest);
+       EPRINTF5 (DBG_DISCON, "%s:                              %08x %08x %08x %08x\n", rail->Generic.Name,
+                 dma.s.dma_destEvent, dma.s.dma_destCookieVProc,
+                dma.s.dma_srcEvent, dma.s.dma_srcCookieVProc);
+       
+       ASSERT ((dma.s.dma_u.s.Context & SYS_CONTEXT_BIT) != 0);
+
+       if (dma.s.dma_direction == DMA_WRITE)
+           vp = dma.s.dma_destVProc;
+       else
+           vp = dma.s.dma_srcVProc;
+       
+       node = &rail->Generic.Nodes[EP_VP_TO_NODE(vp)];
+
+       ASSERT (!EP_VP_ISDATA(vp) || (node->State >= EP_NODE_CONNECTED && node->State <= EP_NODE_LOCAL_PASSIVATE));
+
+       if (EP_VP_ISDATA(vp) && node->State == EP_NODE_LOCAL_PASSIVATE)
+       {
+           /*
+            * This is a DMA going to the node which is being removed, 
+            * so move it onto the node dma list where it will get
+            * handled later.
+            */
+           EPRINTF1 (DBG_DISCON, "%s: FlushDmaQueuesHaltOperation: move dma to cancelled list\n", rail->Generic.Name);
+          
+           if (dma.s.dma_direction != DMA_WRITE)
+           {
+               /* for read dma's set the DMA_READ_REQUEUE bits as the dma_source has been 
+                * modified by the elan to point at the dma in the rxd where it was issued
+                * from */
+               dma.s.dma_direction = (dma.s.dma_direction & ~DMA_READ) | DMA_READ_REQUEUE;
+           }
+           
+           QueueDmaOnStalledList (rail, &dma);
+           
+           /*
+            * Remove the DMA from the queue by replacing it with one with
+            * zero size and no events.
+            *
+            * NOTE: we must preserve the SYS_CONTEXT_BIT since the Elan uses this
+            * to mark the approriate run queue as empty.
+            */
+           dma.s.dma_type            = (SYS_CONTEXT_BIT << 16);
+           dma.s.dma_size            = 0;
+           dma.s.dma_source          = (E3_Addr) 0;
+           dma.s.dma_dest            = (E3_Addr) 0;
+           dma.s.dma_destEvent       = (E3_Addr) 0;
+           dma.s.dma_destCookieVProc = 0;
+           dma.s.dma_srcEvent        = (E3_Addr) 0;
+           dma.s.dma_srcCookieVProc  = 0;
+           
+           elan3_sdram_copyq_to_sdram (dev, &dma, FPtr, sizeof (E3_DMA_BE));
+       }
+
+       FPtr = (FPtr == Top) ? Base : FPtr + sizeof (E3_DMA);
+    }
+
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    rail->HaltOpCompleted = 1;
+    kcondvar_wakeupall (&rail->HaltOpSleep, &dev->IntrLock);
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+}
+
+void
+SetQueueLocked (EP3_RAIL *rail, sdramaddr_t qaddr)
+{
+    ELAN3_DEV        *dev = rail->Device;
+    SetQueueFullData  data;
+    unsigned long     flags;
+    
+    /* Ensure that the context filter changes have been seen by halting
+     * then restarting the inputters - this also ensures that any setevent
+     * commands used to issue dma's have completed and any trap has been
+     * handled. */
+    data.rail  = rail;
+    data.qaddr = qaddr;
+
+    kmutex_lock (&rail->HaltOpMutex);
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    QueueHaltOperation (dev, 0, NULL, INT_DiscardingSysCntx | INT_TProcHalted, SetQueueLockedOperation, &data);
+
+    while (! rail->HaltOpCompleted)
+       kcondvar_wait (&rail->HaltOpSleep, &dev->IntrLock, &flags);
+    rail->HaltOpCompleted = 0;
+
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    kmutex_unlock (&rail->HaltOpMutex);
+}
+
+void
+ep3_flush_filters (EP_RAIL *r)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) r;
+    ELAN3_DEV *dev  = rail->Device;
+    unsigned long flags;
+
+    /* Ensure that the context filter changes have been seen by halting
+     * then restarting the inputters - this also ensures that any setevent
+     * commands used to issue dma's have completed and any trap has been
+     * handled. */
+    kmutex_lock (&rail->HaltOpMutex);
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    QueueHaltOperation (dev, 0, NULL, INT_DiscardingSysCntx, DiscardingHaltOperation, rail);
+    
+    while (! rail->HaltOpCompleted)
+       kcondvar_wait (&rail->HaltOpSleep, &dev->IntrLock, &flags);
+    rail->HaltOpCompleted = 0;
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    kmutex_unlock (&rail->HaltOpMutex);
+}
+
+void
+ep3_flush_queues (EP_RAIL *r)
+{
+    EP3_RAIL         *rail = (EP3_RAIL *) r;
+    ELAN3_DEV         *dev  = rail->Device;
+    struct list_head *el;
+    struct list_head *nel;
+    EP_NODE_RAIL     *node;
+    unsigned long flags;
+    int vp, i;
+
+    ASSERT (NO_LOCKS_HELD);
+    
+    /* First - stall the dma retry thread, so that it will no longer
+     *         restart any dma's from the rety lists. */
+    StallDmaRetryThread (rail);
+
+    /* Second - queue a halt operation to flush through all DMA's which are executing
+     *          or on the run queue. */
+    kmutex_lock (&rail->HaltOpMutex);
+    spin_lock_irqsave (&dev->IntrLock, flags);
+    QueueHaltOperation (dev, 0, NULL, INT_DProcHalted | INT_TProcHalted, FlushDmaQueuesHaltOperation, rail);
+    while (! rail->HaltOpCompleted)
+       kcondvar_wait (&rail->HaltOpSleep, &dev->IntrLock, &flags);
+    rail->HaltOpCompleted = 0;
+    spin_unlock_irqrestore (&dev->IntrLock, flags);
+    kmutex_unlock (&rail->HaltOpMutex);
+
+    /* Third - run down the dma retry lists and move all entries to the cancelled
+     *         list.  Any dma's which were on the run queues have already been
+     *         moved there */
+    spin_lock_irqsave (&rail->DmaRetryLock, flags);
+    for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+    {
+       list_for_each_safe (el, nel, &rail->DmaRetries[i]) {
+           EP3_RETRY_DMA *retry = list_entry (el, EP3_RETRY_DMA, Link);
+
+           if (retry->Dma.s.dma_direction == DMA_WRITE)
+               vp = retry->Dma.s.dma_destVProc;
+           else
+               vp = retry->Dma.s.dma_srcVProc;
+           
+           node = &rail->Generic.Nodes[EP_VP_TO_NODE(vp)];
+           
+           ASSERT (!EP_VP_ISDATA(vp) || (node->State >= EP_NODE_CONNECTED && node->State <= EP_NODE_LOCAL_PASSIVATE));
+
+           if (EP_VP_ISDATA(vp) && node->State == EP_NODE_LOCAL_PASSIVATE)
+           {
+               EPRINTF5 (DBG_DISCON, "%s: FlushDmaQueues: %08x %08x %08x %08x\n",rail->Generic.Name,
+                         retry->Dma.s.dma_type, retry->Dma.s.dma_size, retry->Dma.s.dma_source, retry->Dma.s.dma_dest);
+               EPRINTF5 (DBG_DISCON, "%s:                 %08x %08x %08x %08x\n", rail->Generic.Name,
+                         retry->Dma.s.dma_destEvent, retry->Dma.s.dma_destCookieVProc,
+                         retry->Dma.s.dma_srcEvent, retry->Dma.s.dma_srcCookieVProc);
+
+               list_del (&retry->Link);
+
+               list_add_tail (&retry->Link, &node->StalledDmas);
+           }
+       }
+    }
+    spin_unlock_irqrestore (&rail->DmaRetryLock, flags);
+
+    /* Finally - allow the dma retry thread to run again */
+    ResumeDmaRetryThread (rail);
+}
+
+/****************************************************************************************/
+/* NOTE - we require that all cookies are non-zero, which is 
+ *        achieved because EP_VP_DATA() is non-zero for all
+ *        nodes */
+E3_uint32
+LocalCookie (EP3_RAIL *rail, unsigned remoteNode)
+{
+    E3_uint32     cookie;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->CookieLock, flags);
+    cookie = DMA_COOKIE (rail->MainCookies[remoteNode], EP_VP_DATA(rail->Generic.Position.pos_nodeid));
+    spin_unlock_irqrestore (&rail->CookieLock, flags);
+
+    /* Main processor cookie for srcCookie - this is what is sent
+     * to the remote node along with the setevent from the put
+     * or the dma descriptor for a get */
+    return (cookie);
+}
+
+E3_uint32
+RemoteCookie (EP3_RAIL *rail, u_int remoteNode)
+{
+    uint32_t      cookie;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->CookieLock, flags);
+    cookie = DMA_REMOTE_COOKIE (rail->MainCookies[remoteNode], EP_VP_DATA(remoteNode));
+    spin_unlock_irqrestore (&rail->CookieLock, flags);
+
+    /* Main processor cookie for dstCookie - this is the cookie
+     * that the "remote put" dma uses for it's setevent packets for
+     * a get dma */
+    
+    return (cookie);
+}
+
+/****************************************************************************************/
+/*
+ * Event Cookie management.
+ *
+ *   We find the ep_cookie in one of two ways:
+ *     1) for block copy events
+ *          the cookie value is stored in the ev_Source - for EVIRQ events
+ *          it is also stored in the ev_Type
+ *     2) for normal events
+ *          we just use the event address.
+ */
+void 
+InitialiseCookieTable (EP3_COOKIE_TABLE *table)
+{
+    register int i;
+    
+    spin_lock_init (&table->Lock);
+    
+    for (i = 0; i < EP3_COOKIE_HASH_SIZE; i++)
+       table->Entries[i] = NULL;
+}
+
+void
+DestroyCookieTable (EP3_COOKIE_TABLE *table)
+{
+    register int i;
+
+    for (i = 0; i < EP3_COOKIE_HASH_SIZE; i++)
+       if (table->Entries[i])
+           printk ("DestroyCookieTable: entry %d not empty\n", i);
+
+    spin_lock_destroy (&table->Lock);
+}
+
+void
+RegisterCookie (EP3_COOKIE_TABLE *table, EP3_COOKIE *cp, E3_uint32 cookie, EP3_COOKIE_OPS *ops, void *arg)
+{
+    EP3_COOKIE *tcp;
+    int hashval = EP3_HASH_COOKIE(cookie);
+    unsigned long flags;
+
+    spin_lock_irqsave (&table->Lock, flags);
+    
+    cp->Operations = ops;
+    cp->Arg        = arg;
+    cp->Cookie     = cookie;
+    
+#if defined(DEBUG)
+    /* Check that the cookie is unique */
+    for (tcp = table->Entries[hashval]; tcp; tcp = tcp->Next)
+       if (tcp->Cookie == cookie)
+           panic ("RegisterEventCookie: non unique cookie\n");
+#endif
+    cp->Next = table->Entries[hashval];
+    
+    table->Entries[hashval] = cp;
+    
+    spin_unlock_irqrestore (&table->Lock, flags);
+}
+
+void
+DeregisterCookie (EP3_COOKIE_TABLE *table, EP3_COOKIE *cp)
+{
+    EP3_COOKIE **predCookiep;
+    unsigned long flags;
+
+    spin_lock_irqsave (&table->Lock, flags);
+    
+    for (predCookiep = &table->Entries[EP3_HASH_COOKIE (cp->Cookie)]; *predCookiep; predCookiep = &(*predCookiep)->Next)
+    {
+       if (*predCookiep == cp)
+       {
+           *predCookiep = cp->Next;
+           break;
+       }
+    }
+
+    spin_unlock_irqrestore (&table->Lock, flags);
+
+    cp->Operations = NULL;
+    cp->Arg        = NULL;
+    cp->Cookie     = 0;
+    cp->Next       = NULL;
+}
+
+EP3_COOKIE *
+LookupCookie (EP3_COOKIE_TABLE *table, E3_Addr cookie)
+{
+    EP3_COOKIE *cp;
+    unsigned long flags;
+
+    spin_lock_irqsave (&table->Lock, flags);
+    
+    for (cp = table->Entries[EP3_HASH_COOKIE(cookie)]; cp; cp = cp->Next)
+       if (cp->Cookie == cookie)
+           break;
+    
+    spin_unlock_irqrestore (&table->Lock, flags);
+    return (cp);
+}
+
+EP3_COOKIE *
+LookupEventCookie (EP3_RAIL *rail, EP3_COOKIE_TABLE *table, E3_Addr eaddr)
+{
+    sdramaddr_t event;
+    E3_uint32 type;
+
+    if ((event = ep_elan2sdram (&rail->Generic, eaddr)) != (sdramaddr_t) 0)
+    {
+       type = elan3_sdram_readl (rail->Device, event + offsetof (E3_BlockCopyEvent, ev_Type));
+
+       if (type & EV_TYPE_BCOPY)
+           return (LookupCookie (table, elan3_sdram_readl (rail->Device, event + offsetof (E3_BlockCopyEvent, ev_Source)) & ~EV_WCOPY));
+       else
+           return (LookupCookie (table, eaddr));
+    }
+
+    return (NULL);
+}
+
+/****************************************************************************************/
+/*
+ * Elan context operations - note only support interrupt ops.
+ */
+static int        ep3_event     (ELAN3_CTXT *ctxt, E3_uint32 cookie, int flag);
+static int        ep3_dprocTrap (ELAN3_CTXT *ctxt, DMA_TRAP *trap);
+static int        ep3_tprocTrap (ELAN3_CTXT *ctxt, THREAD_TRAP *trap);
+static int        ep3_iprocTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, int chan);
+static int        ep3_cprocTrap (ELAN3_CTXT *ctxt, COMMAND_TRAP *trap);
+static int        ep3_cprocReissue (ELAN3_CTXT *ctxt, CProcTrapBuf_BE *tbuf);
+
+static E3_uint8   ep3_load8 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void       ep3_store8 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint8 val);
+static E3_uint16  ep3_load16 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void       ep3_store16 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint16 val);
+static E3_uint32  ep3_load32 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void       ep3_store32 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint32 val);
+static E3_uint64  ep3_load64 (ELAN3_CTXT *ctxt, E3_Addr addr);
+static void       ep3_store64 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint64 val);
+
+ELAN3_OPS ep3_elan3_ops = 
+{
+    ELAN3_OPS_VERSION,         /* Version */
+    
+    NULL,                      /* Exception */
+    NULL,                      /* GetWordItem */
+    NULL,                      /* GetBlockItem */
+    NULL,                      /* PutWordItem */
+    NULL,                      /* PutBlockItem */
+    NULL,                      /* PutbackItem */
+    NULL,                      /* FreeWordItem */
+    NULL,                      /* FreeBlockItem */
+    NULL,                      /* CountItems */
+    ep3_event,                 /* Event */
+    NULL,                      /* SwapIn */
+    NULL,                      /* SwapOut */
+    NULL,                      /* FreePrivate */
+    NULL,                      /* FixupNetworkError */
+    ep3_dprocTrap,             /* DProcTrap */
+    ep3_tprocTrap,             /* TProcTrap */
+    ep3_iprocTrap,             /* IProcTrap */
+    ep3_cprocTrap,             /* CProcTrap */
+    ep3_cprocReissue,          /* CProcReissue */
+    NULL,                      /* StartFaultCheck */
+    NULL,                      /* EndFaulCheck */
+    ep3_load8,                 /* Load8 */
+    ep3_store8,                        /* Store8 */
+    ep3_load16,                        /* Load16 */
+    ep3_store16,               /* Store16 */
+    ep3_load32,                        /* Load32 */
+    ep3_store32,               /* Store32 */
+    ep3_load64,                        /* Load64 */
+    ep3_store64,               /* Store64 */
+};
+
+static int
+ep3_event (ELAN3_CTXT *ctxt, E3_uint32 cookie, int flag)
+{
+    EP3_RAIL  *rail = (EP3_RAIL *) ctxt->Private;
+    EP3_COOKIE *cp   = LookupCookie (&rail->CookieTable, cookie);
+    
+    if (cp == NULL)
+    {
+       printk ("ep3_event: cannot find event cookie for %x\n", cookie);
+       return (OP_HANDLED);
+    }
+    
+    if (cp->Operations->Event)
+       cp->Operations->Event(rail, cp->Arg);
+    
+    return (OP_HANDLED);
+}
+
+/* Trap interface */
+int
+ep3_dprocTrap (ELAN3_CTXT *ctxt, DMA_TRAP *trap)
+{
+    EP3_RAIL        *rail = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV        *dev = rail->Device;
+    EP3_COOKIE       *cp;
+    E3_FaultSave_BE *FaultArea;
+    E3_uint16        vp;
+    int                     validTrap;
+    int                     numFaults;
+    int                     i;
+    sdramaddr_t      event;
+    E3_uint32        type;
+    sdramaddr_t      dma;
+    E3_DMA_BE        dmabe;
+    int              status = EAGAIN;
+
+    EPRINTF4 (DBG_EPTRAP, "ep3_dprocTrap: WakeupFnt=%x Cntx=%x SuspAddr=%x TrapType=%s\n",
+             trap->Status.s.WakeupFunction, trap->Status.s.Context, 
+             trap->Status.s.SuspendAddr, MiToName (trap->Status.s.TrapType));
+    EPRINTF4 (DBG_EPTRAP, "              type %08x size %08x source %08x dest %08x\n",
+             trap->Desc.s.dma_type, trap->Desc.s.dma_size, trap->Desc.s.dma_source, trap->Desc.s.dma_dest);
+    EPRINTF2 (DBG_EPTRAP, "              Dest event %08x cookie/proc %08x\n",
+             trap->Desc.s.dma_destEvent, trap->Desc.s.dma_destCookieVProc);
+    EPRINTF2 (DBG_EPTRAP, "              Source event %08x cookie/proc %08x\n",
+             trap->Desc.s.dma_srcEvent, trap->Desc.s.dma_srcCookieVProc);
+
+    ASSERT (trap->Status.s.Context & SYS_CONTEXT_BIT);
+
+    switch (trap->Status.s.TrapType)
+    {
+    case MI_DmaPacketTimedOutOrPacketError:
+       if (trap->Desc.s.dma_direction == DMA_WRITE)
+           vp = trap->Desc.s.dma_destVProc;
+       else
+           vp = trap->Desc.s.dma_srcVProc;
+       
+       if (! trap->PacketInfo.s.PacketTimeout)
+           status = ETIMEDOUT;
+       else
+       {
+           status = EHOSTDOWN;
+
+           /* XXXX: dma timedout - might want to "restart" tree ? */
+       }
+       goto retry_dma;
+
+    case MI_DmaFailCountError:
+       goto retry_dma;
+
+    case MI_TimesliceDmaQueueOverflow:
+       IncrStat (rail, DprocDmaQueueOverflow);
+
+       goto retry_dma;
+
+    case MI_RemoteDmaCommand:
+    case MI_RunDmaCommand:
+    case MI_DequeueNonSysCntxDma:
+    case MI_DequeueSysCntxDma:
+       /*
+        * The DMA processor has trapped due to outstanding prefetches from the previous 
+        * dma.  The "current" dma has not been consumed, so we just ignore the trap
+        */
+       return (OP_HANDLED);
+       
+    case MI_EventQueueOverflow:
+       IncrStat (rail, DprocEventQueueOverflow);
+
+       if ((event = ep_elan2sdram (&rail->Generic, trap->Desc.s.dma_srcEvent)) != (sdramaddr_t) 0 &&
+           ((type  = elan3_sdram_readl (dev, event + offsetof(E3_Event,ev_Type))) & EV_TYPE_MASK_EVIRQ) == EV_TYPE_EVIRQ)
+       {
+           spin_unlock (&ctxt->Device->IntrLock);
+           ep3_event (ctxt, (type & ~(EV_TYPE_MASK_EVIRQ | EV_TYPE_MASK_BCOPY)), OP_LWP);
+           spin_lock (&ctxt->Device->IntrLock);
+       }
+       return (OP_HANDLED);
+       
+    case MI_DmaQueueOverflow:
+       IncrStat (rail, DprocDmaQueueOverflow);
+
+       if ((event = ep_elan2sdram (&rail->Generic, trap->Desc.s.dma_srcEvent)) != (sdramaddr_t) 0 &&
+           ((type = elan3_sdram_readl (dev, event + offsetof (E3_Event, ev_Type))) & EV_TYPE_MASK_DMA) == EV_TYPE_DMA &&
+           (dma  = ep_elan2sdram (&rail->Generic, (type & ~EV_TYPE_MASK2))) != (sdramaddr_t) 0)
+       {
+           elan3_sdram_copyq_from_sdram (dev, dma, &dmabe, sizeof (E3_DMA));
+           
+           /* We only chain together DMA's of the same direction, so since
+            * we took a DmaQueueOverflow trap - this means that DMA which
+            * trapped was a WRITE dma - hence the one we chain to must also
+            * be a WRITE dma.
+            */
+           ASSERT (dmabe.s.dma_direction == DMA_WRITE);
+           
+           cp = LookupEventCookie (rail, &rail->CookieTable, dmabe.s.dma_srcEvent);
+
+#ifdef DEBUG_ASSERT
+           {
+               E3_uint16     vp       = dmabe.s.dma_destVProc;
+               EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[EP_VP_TO_NODE(vp)];
+
+               ASSERT (cp != NULL && (!EP_VP_ISDATA(vp) || (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE)));
+           }
+#endif
+           cp->Operations->DmaRetry (rail, cp->Arg, &dmabe, EAGAIN);
+           
+           return (OP_HANDLED);
+       }
+
+       panic ("ep3_dprocTrap\n");
+       return (OP_HANDLED);
+
+    default:
+       break;
+    }
+
+    /* If it's a dma which traps past the end of the source, then */
+    /* just re-issue it */
+    numFaults = validTrap = (trap->FaultSave.s.FSR.Status != 0);
+    for (i = 0, FaultArea = &trap->Data0; i < 4; i++, FaultArea++)
+    {
+       if (FaultArea->s.FSR.Status != 0)
+       {
+           numFaults++;
+
+           /* XXXX: Rev B Elans can prefetch data past the end of the dma descriptor */
+           /*       if the fault relates to this, then just ignore it */
+           if (FaultArea->s.FaultAddress >= (trap->Desc.s.dma_source+trap->Desc.s.dma_size))
+           {
+               static int i;
+               if (i < 10 && i++ < 10)
+                   printk ("ep3_dprocTrap: Rev B prefetch trap error %08x %08x\n",
+                            FaultArea->s.FaultAddress, (trap->Desc.s.dma_source+trap->Desc.s.dma_size));
+               continue;
+           }
+
+           validTrap++;
+       }
+    }
+
+    /*
+     * NOTE: for physical errors (uncorrectable ECC/PCI parity errors) the FSR will
+     *       be zero - hence we will not see any faults - and none will be valid, 
+     *       so only ignore a Rev B prefetch trap if we've seen some faults. Otherwise
+     *       we can reissue a DMA which has already sent it's remote event !
+     */
+    if (numFaults != 0 && validTrap == 0)
+    {
+    retry_dma:
+       if (trap->Desc.s.dma_direction == DMA_WRITE)
+       {
+           vp = trap->Desc.s.dma_destVProc;
+           cp = LookupEventCookie (rail, &rail->CookieTable, trap->Desc.s.dma_srcEvent);
+       }
+       else
+       {
+           ASSERT (EP3_CONTEXT_ISDATA(trap->Desc.s.dma_queueContext) || trap->Desc.s.dma_direction == DMA_READ_REQUEUE);
+
+           vp = trap->Desc.s.dma_srcVProc;
+           cp = LookupEventCookie (rail, &rail->CookieTable, trap->Desc.s.dma_destEvent);
+
+           /* for read dma's set the DMA_READ_REQUEUE bits as the dma_source has been 
+            * modified by the elan to point at the dma in the rxd where it was issued
+            * from */
+           trap->Desc.s.dma_direction = (trap->Desc.s.dma_direction & ~DMA_READ) | DMA_READ_REQUEUE;
+       }
+
+#ifdef DEBUG_ASSERT
+       {
+           EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[EP_VP_TO_NODE(vp)];
+
+           ASSERT (!EP_VP_ISDATA(vp) || (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE));
+       }
+#endif
+       
+       if (cp != NULL)
+           cp->Operations->DmaRetry (rail, cp->Arg, &trap->Desc, status);
+       else
+       {
+           ASSERT (trap->Desc.s.dma_direction == DMA_WRITE && trap->Desc.s.dma_srcEvent == 0 && trap->Desc.s.dma_isRemote);
+
+           QueueDmaForRetry (rail, &trap->Desc, EP_RETRY_ANONYMOUS);
+       }
+
+       return (OP_HANDLED);
+    }
+    
+    printk ("ep3_dprocTrap: WakeupFnt=%x Cntx=%x SuspAddr=%x TrapType=%s\n",
+            trap->Status.s.WakeupFunction, trap->Status.s.Context, 
+            trap->Status.s.SuspendAddr, MiToName (trap->Status.s.TrapType));
+    printk ("                    FaultAddr=%x EventAddr=%x FSR=%x\n",
+            trap->FaultSave.s.FaultAddress, trap->FaultSave.s.EventAddress,
+            trap->FaultSave.s.FSR.Status);
+    for (i = 0, FaultArea = &trap->Data0; i < 4; i++, FaultArea++)
+       printk ("                  %d FaultAddr=%x EventAddr=%x FSR=%x\n", i,
+                FaultArea->s.FaultAddress, FaultArea->s.EventAddress, FaultArea->s.FSR.Status);
+    
+    printk ("                  type %08x size %08x source %08x dest %08x\n",
+            trap->Desc.s.dma_type, trap->Desc.s.dma_size, trap->Desc.s.dma_source, trap->Desc.s.dma_dest);
+    printk ("                  Dest event %08x cookie/proc %08x\n",
+            trap->Desc.s.dma_destEvent, trap->Desc.s.dma_destCookieVProc);
+    printk ("                  Source event %08x cookie/proc %08x\n",
+            trap->Desc.s.dma_srcEvent, trap->Desc.s.dma_srcCookieVProc);
+
+//    panic ("ep3_dprocTrap");
+
+    return (OP_HANDLED);
+}
+
+int
+ep3_tprocTrap (ELAN3_CTXT *ctxt, THREAD_TRAP *trap)
+{
+    EP3_RAIL *rail = (EP3_RAIL *) ctxt->Private;
+
+    EPRINTF6 (DBG_EPTRAP, "ep3_tprocTrap: SP=%08x PC=%08x NPC=%08x DIRTY=%08x TRAP=%08x MI=%s\n",
+             trap->sp, trap->pc, trap->npc, trap->DirtyBits.Bits, trap->TrapBits.Bits, MiToName (trap->mi));
+    EPRINTF4 (DBG_EPTRAP, "              g0=%08x g1=%08x g2=%08x g3=%08x\n", 
+             trap->Registers[REG_GLOBALS+(0^WordEndianFlip)], trap->Registers[REG_GLOBALS+(1^WordEndianFlip)], 
+             trap->Registers[REG_GLOBALS+(2^WordEndianFlip)], trap->Registers[REG_GLOBALS+(3^WordEndianFlip)]);
+    EPRINTF4 (DBG_EPTRAP, "              g4=%08x g5=%08x g6=%08x g7=%08x\n", 
+             trap->Registers[REG_GLOBALS+(4^WordEndianFlip)], trap->Registers[REG_GLOBALS+(5^WordEndianFlip)], 
+             trap->Registers[REG_GLOBALS+(6^WordEndianFlip)], trap->Registers[REG_GLOBALS+(7^WordEndianFlip)]);
+    EPRINTF4 (DBG_EPTRAP, "              o0=%08x o1=%08x o2=%08x o3=%08x\n", 
+             trap->Registers[REG_OUTS+(0^WordEndianFlip)], trap->Registers[REG_OUTS+(1^WordEndianFlip)], 
+             trap->Registers[REG_OUTS+(2^WordEndianFlip)], trap->Registers[REG_OUTS+(3^WordEndianFlip)]);
+    EPRINTF4 (DBG_EPTRAP, "              o4=%08x o5=%08x o6=%08x o7=%08x\n", 
+             trap->Registers[REG_OUTS+(4^WordEndianFlip)], trap->Registers[REG_OUTS+(5^WordEndianFlip)], 
+             trap->Registers[REG_OUTS+(6^WordEndianFlip)], trap->Registers[REG_OUTS+(7^WordEndianFlip)]);
+    EPRINTF4 (DBG_EPTRAP, "              l0=%08x l1=%08x l2=%08x l3=%08x\n", 
+             trap->Registers[REG_LOCALS+(0^WordEndianFlip)], trap->Registers[REG_LOCALS+(1^WordEndianFlip)], 
+             trap->Registers[REG_LOCALS+(2^WordEndianFlip)], trap->Registers[REG_LOCALS+(3^WordEndianFlip)]);
+    EPRINTF4 (DBG_EPTRAP, "              l4=%08x l5=%08x l6=%08x l7=%08x\n", 
+             trap->Registers[REG_LOCALS+(4^WordEndianFlip)], trap->Registers[REG_LOCALS+(5^WordEndianFlip)], 
+             trap->Registers[REG_LOCALS+(6^WordEndianFlip)], trap->Registers[REG_LOCALS+(7^WordEndianFlip)]);
+    EPRINTF4 (DBG_EPTRAP, "              i0=%08x i1=%08x i2=%08x i3=%08x\n", 
+             trap->Registers[REG_INS+(0^WordEndianFlip)], trap->Registers[REG_INS+(1^WordEndianFlip)], 
+             trap->Registers[REG_INS+(2^WordEndianFlip)], trap->Registers[REG_INS+(3^WordEndianFlip)]);
+    EPRINTF4 (DBG_EPTRAP, "              i4=%08x i5=%08x i6=%08x i7=%08x\n", 
+             trap->Registers[REG_INS+(4^WordEndianFlip)], trap->Registers[REG_INS+(5^WordEndianFlip)], 
+             trap->Registers[REG_INS+(6^WordEndianFlip)], trap->Registers[REG_INS+(7^WordEndianFlip)]);
+    
+    ASSERT (trap->Status.s.Context & SYS_CONTEXT_BIT);
+
+    switch (trap->mi)
+    {
+    case MI_UnimplementedError:
+       if (trap->TrapBits.s.ForcedTProcTrap)
+       {
+           ASSERT (trap->TrapBits.s.OutputWasOpen == 0);
+           
+           EPRINTF0 (DBG_EPTRAP, "ep3_tprocTrap: ForcedTProcTrap\n");
+
+           IssueRunThread (rail, SaveThreadToStack (ctxt, trap, FALSE));
+           return (OP_HANDLED);
+       }
+
+       if (trap->TrapBits.s.ThreadTimeout)
+       {
+           EPRINTF0 (DBG_EPTRAP, "ep3_tprocTrap: ThreadTimeout\n");
+
+           if (trap->Registers[REG_GLOBALS + (1^WordEndianFlip)] == 0)
+               RollThreadToClose (ctxt, trap, trap->TrapBits.s.PacketAckValue);
+           else
+           {
+               CompleteEnvelope (rail, trap->Registers[REG_GLOBALS + (1^WordEndianFlip)], trap->TrapBits.s.PacketAckValue);
+
+               RollThreadToClose (ctxt, trap, EP3_PAckStolen);
+           }
+               
+           IssueRunThread (rail, SaveThreadToStack (ctxt, trap, FALSE));
+           return (OP_HANDLED);
+       }
+
+       if (trap->TrapBits.s.Unimplemented)
+       {
+           E3_uint32 instr = ELAN3_OP_LOAD32 (ctxt, trap->pc & PC_MASK);
+
+           PRINTF1 (ctxt, DBG_EPTRAP, "ep3_tprocTrap: unimplemented instruction %08x\n", instr);
+
+           if ((instr & OPCODE_MASK) == OPCODE_Ticc &&
+               (instr & OPCODE_IMM)  == OPCODE_IMM &&
+               (Ticc_COND(instr)     == Ticc_TA))
+           {
+               switch (INSTR_IMM(instr))
+               {
+               case EP3_UNIMP_TRAP_NO_DESCS:
+                   StallThreadForNoDescs (rail, trap->Registers[REG_GLOBALS + (1^WordEndianFlip)], 
+                                          SaveThreadToStack (ctxt, trap, TRUE));
+                   return (OP_HANDLED);
+
+               case EP3_UNIMP_TRAP_PACKET_NACKED:
+                   CompleteEnvelope (rail, trap->Registers[REG_GLOBALS + (1^WordEndianFlip)], E3_PAckDiscard);
+
+                   IssueRunThread (rail, SaveThreadToStack (ctxt, trap, TRUE));
+                   return (OP_HANDLED);
+
+               case EP3_UNIMP_THREAD_HALTED: 
+                   StallThreadForHalted (rail, trap->Registers[REG_GLOBALS + (1^WordEndianFlip)], 
+                                         SaveThreadToStack (ctxt, trap, TRUE));
+                   return (OP_HANDLED);
+
+               default:
+                   break;
+                   
+               }
+           }
+       }
+       break;
+
+    default:
+       break;
+    }
+
+    /* All other traps should not happen for kernel comms */
+    printk ("ep3_tprocTrap: SP=%08x PC=%08x NPC=%08x DIRTY=%08x TRAP=%08x MI=%s\n",
+            trap->sp, trap->pc, trap->npc, trap->DirtyBits.Bits,
+            trap->TrapBits.Bits, MiToName (trap->mi));
+    printk ("              FaultSave : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->FaultSave.s.FaultAddress, trap->FaultSave.s.EventAddress, trap->FaultSave.s.FSR.Status);
+    printk ("              DataFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->DataFaultSave.s.FaultAddress, trap->DataFaultSave.s.EventAddress, trap->DataFaultSave.s.FSR.Status);
+    printk ("              InstFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->InstFaultSave.s.FaultAddress, trap->InstFaultSave.s.EventAddress, trap->InstFaultSave.s.FSR.Status);
+    printk ("              OpenFault : FaultAddress %08x EventAddress %08x FSR %08x\n",
+            trap->OpenFaultSave.s.FaultAddress, trap->OpenFaultSave.s.EventAddress, trap->OpenFaultSave.s.FSR.Status);
+
+    if (trap->DirtyBits.s.GlobalsDirty)
+    {
+       printk ("              g0=%08x g1=%08x g2=%08x g3=%08x\n", 
+                trap->Registers[REG_GLOBALS+(0^WordEndianFlip)], trap->Registers[REG_GLOBALS+(1^WordEndianFlip)], 
+                trap->Registers[REG_GLOBALS+(2^WordEndianFlip)], trap->Registers[REG_GLOBALS+(3^WordEndianFlip)]);
+       printk ("              g4=%08x g5=%08x g6=%08x g7=%08x\n", 
+                trap->Registers[REG_GLOBALS+(4^WordEndianFlip)], trap->Registers[REG_GLOBALS+(5^WordEndianFlip)], 
+                trap->Registers[REG_GLOBALS+(6^WordEndianFlip)], trap->Registers[REG_GLOBALS+(7^WordEndianFlip)]);
+    }
+    if (trap->DirtyBits.s.OutsDirty)
+    {
+       printk ("              o0=%08x o1=%08x o2=%08x o3=%08x\n", 
+                trap->Registers[REG_OUTS+(0^WordEndianFlip)], trap->Registers[REG_OUTS+(1^WordEndianFlip)], 
+                trap->Registers[REG_OUTS+(2^WordEndianFlip)], trap->Registers[REG_OUTS+(3^WordEndianFlip)]);
+       printk ("              o4=%08x o5=%08x o6=%08x o7=%08x\n", 
+                trap->Registers[REG_OUTS+(4^WordEndianFlip)], trap->Registers[REG_OUTS+(5^WordEndianFlip)], 
+                trap->Registers[REG_OUTS+(6^WordEndianFlip)], trap->Registers[REG_OUTS+(7^WordEndianFlip)]);
+    }
+    if (trap->DirtyBits.s.LocalsDirty)
+    {
+       printk ("              l0=%08x l1=%08x l2=%08x l3=%08x\n", 
+                trap->Registers[REG_LOCALS+(0^WordEndianFlip)], trap->Registers[REG_LOCALS+(1^WordEndianFlip)], 
+                trap->Registers[REG_LOCALS+(2^WordEndianFlip)], trap->Registers[REG_LOCALS+(3^WordEndianFlip)]);
+       printk ("              l4=%08x l5=%08x l6=%08x l7=%08x\n", 
+                trap->Registers[REG_LOCALS+(4^WordEndianFlip)], trap->Registers[REG_LOCALS+(5^WordEndianFlip)], 
+                trap->Registers[REG_LOCALS+(6^WordEndianFlip)], trap->Registers[REG_LOCALS+(7^WordEndianFlip)]);
+    }
+    if (trap->DirtyBits.s.InsDirty)
+    {
+       printk ("              i0=%08x i1=%08x i2=%08x i3=%08x\n", 
+                trap->Registers[REG_INS+(0^WordEndianFlip)], trap->Registers[REG_INS+(1^WordEndianFlip)], 
+                trap->Registers[REG_INS+(2^WordEndianFlip)], trap->Registers[REG_INS+(3^WordEndianFlip)]);
+       printk ("              i4=%08x i5=%08x i6=%08x i7=%08x\n", 
+                trap->Registers[REG_INS+(4^WordEndianFlip)], trap->Registers[REG_INS+(5^WordEndianFlip)], 
+                trap->Registers[REG_INS+(6^WordEndianFlip)], trap->Registers[REG_INS+(7^WordEndianFlip)]);
+    }
+    
+//    panic ("ep3_tprocTrap");
+
+    return (OP_HANDLED);
+}
+
+int
+ep3_iprocTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, int channel)
+{
+    EP3_RAIL      *rail = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV      *dev = ctxt->Device;
+    EP3_COOKIE    *cp;
+    sdramaddr_t    event;
+    E3_uint32      type;
+    sdramaddr_t    dma;
+    E3_DMA_BE      dmabe;
+
+    ASSERT (trap->Transactions[0].s.TrTypeCntx.s.Context & SYS_CONTEXT_BIT);
+
+    /*
+     * first process the trap to determine the cause
+     */
+    InspectIProcTrap (ctxt, trap);
+
+    if (! trap->AckSent && trap->LockQueuePointer)             /* Must be a network error in a queueing DMA */
+    {                                                          /* packet - unlock the queue */
+       IncrStat (rail, QueueingPacketTrap);
+
+       SimulateUnlockQueue (ctxt, trap->LockQueuePointer, FALSE);
+       return (OP_HANDLED);
+    }
+
+    if (trap->AckSent && trap->BadTransaction)
+    {
+       spin_unlock (&dev->IntrLock);
+
+       /* NOTE - no network error fixup is necessary for system context
+        *        messages since they are idempotent and are single packet 
+        *        dmas
+        */
+       if (EP3_CONTEXT_ISDATA (trap->Transactions[0].s.TrTypeCntx.s.Context))
+       {
+           int nodeId = EP3_CONTEXT_TO_NODE(trap->Transactions[0].s.TrTypeCntx.s.Context);
+           
+           if (trap->DmaIdentifyTransaction)
+               ep_queue_network_error (&rail->Generic, nodeId, EP_NODE_NETERR_ATOMIC_PACKET, channel, trap->DmaIdentifyTransaction->s.TrAddr);
+           else if (trap->ThreadIdentifyTransaction)
+               ep_queue_network_error (&rail->Generic, nodeId, EP_NODE_NETERR_ATOMIC_PACKET, channel, trap->ThreadIdentifyTransaction->s.TrAddr);
+           else
+               ep_queue_network_error (&rail->Generic, nodeId, EP_NODE_NETERR_DMA_PACKET, channel, 0);
+       }
+
+       spin_lock (&dev->IntrLock);
+       return (OP_HANDLED);
+    }
+    
+    if (trap->AckSent)
+    {
+       if (trap->TrappedTransaction == NULL)
+           return (OP_HANDLED);
+       
+       while (! trap->TrappedTransaction->s.TrTypeCntx.s.LastTrappedTrans)
+       {
+           E3_IprocTrapHeader_BE *hdrp  = trap->TrappedTransaction;
+           E3_IprocTrapData_BE   *datap = trap->TrappedDataBuffer;
+           
+           ASSERT (hdrp->s.TrTypeCntx.s.StatusRegValid != 0);
+           
+           if ((hdrp->s.TrTypeCntx.s.Type & TR_WRITEBLOCK_BIT) != 0)
+           {
+               printk ("ep3_iprocTrap: WRITEBLOCK : Addr %x\n", hdrp->s.TrAddr);
+//             panic ("ep3_iprocTrap\n");
+           }
+           else
+           {
+               switch (hdrp->s.TrTypeCntx.s.Type & TR_OPCODE_TYPE_MASK)
+               {
+               case TR_SETEVENT & TR_OPCODE_TYPE_MASK:
+                   switch (GET_STATUS_TRAPTYPE (hdrp->s.IProcTrapStatus))
+                   {
+                   case MI_DmaQueueOverflow:
+                       IncrStat (rail, IprocDmaQueueOverflow);
+
+                       if ((event = ep_elan2sdram (&rail->Generic, hdrp->s.TrAddr)) != (sdramaddr_t) 0 &&
+                           ((type = elan3_sdram_readl (dev, event + offsetof (E3_Event, ev_Type))) & EV_TYPE_MASK_DMA) == EV_TYPE_DMA &&
+                           (dma  = ep_elan2sdram (&rail->Generic, (type & ~EV_TYPE_MASK2))) != (sdramaddr_t) 0)
+                       {
+                           elan3_sdram_copyq_from_sdram (dev, dma, &dmabe, sizeof (E3_DMA));
+                           
+                           if (dmabe.s.dma_direction == DMA_WRITE)
+                               cp = LookupEventCookie (rail, &rail->CookieTable, dmabe.s.dma_srcEvent);
+                           else
+                           {
+                               cp = LookupEventCookie (rail, &rail->CookieTable, dmabe.s.dma_destEvent);
+                               
+                               /* we MUST convert this into a DMA_READ_REQUEUE dma as if we don't the 
+                                * DMA descriptor will be read from the EP3_RETRY_DMA rather than the 
+                                * original DMA - this can then get reused and an incorrect DMA 
+                                * descriptor sent 
+                                * eventp->ev_Type contains the dma address with type in the lower bits 
+                                */ 
+                               
+                               dmabe.s.dma_source    = (type & ~EV_TYPE_MASK2);
+                               dmabe.s.dma_direction = (dmabe.s.dma_direction & ~DMA_READ) | DMA_READ_REQUEUE;
+                           }
+
+#ifdef DEBUG_ASSERT
+                           {
+                               E3_uint16     vp       = (dmabe.s.dma_direction == DMA_WRITE ? dmabe.s.dma_destVProc : dmabe.s.dma_srcVProc);
+                               EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[EP_VP_TO_NODE(vp)];
+
+                               ASSERT (!EP_VP_ISDATA(vp) || (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE));
+                           }
+#endif
+
+                           if (cp != NULL)
+                               cp->Operations->DmaRetry (rail, cp->Arg, &dmabe, EAGAIN);
+                           else
+                           {
+                               ASSERT (dmabe.s.dma_direction == DMA_WRITE && dmabe.s.dma_srcEvent == 0 && dmabe.s.dma_isRemote);
+                               
+                               QueueDmaForRetry (rail, &dmabe, EP_RETRY_ANONYMOUS);
+                           }
+                           break;
+                       }
+
+                       printk ("ep3_iprocTrap: SETEVENT : %x - cannot find dma to restart\n", hdrp->s.TrAddr);
+//                     panic ("ep3_iprocTrap\n");
+                       break;
+
+                   case MI_EventQueueOverflow:
+                   {
+                       sdramaddr_t event;
+                       E3_uint32   type;
+
+                       IncrStat (rail, IprocEventQueueOverflow);
+
+                       if ((event = ep_elan2sdram (&rail->Generic, hdrp->s.TrAddr)) != (sdramaddr_t) 0 &&
+                           ((type = elan3_sdram_readl (dev, event + offsetof (E3_Event, ev_Type))) & EV_TYPE_MASK_EVIRQ) == EV_TYPE_EVIRQ)
+                       {
+                           spin_unlock (&dev->IntrLock);
+                           ep3_event (ctxt, (type & ~(EV_TYPE_MASK_EVIRQ|EV_TYPE_MASK_BCOPY)), OP_LWP);
+                           spin_lock (&dev->IntrLock);
+
+                           break;
+                       }
+
+                       printk ("ep3_iprocTrap: SETEVENT : %x - cannot find event\n", hdrp->s.TrAddr);
+//                     panic ("ep3_iprocTrap\n");
+                       break;
+                   }
+
+                   default:
+                       printk ("ep3_iprocTrap: SETEVENT : %x MI=%x\n", hdrp->s.TrAddr, GET_STATUS_TRAPTYPE(hdrp->s.IProcTrapStatus));
+//                     panic ("ep3_iprocTrap\n");
+                       break;
+                   }
+                   break;
+                   
+               case TR_SENDDISCARD & TR_OPCODE_TYPE_MASK:
+                   /* Just ignore send-discard transactions */
+                   break;
+                   
+               case TR_REMOTEDMA & TR_OPCODE_TYPE_MASK:
+               {
+                   E3_DMA_BE *dmap = (E3_DMA_BE *) datap;
+
+                   if (GET_STATUS_TRAPTYPE(hdrp->s.IProcTrapStatus) != MI_DmaQueueOverflow)
+                   {
+                       printk ("ep3_iprocTrap: MI=%x\n", GET_STATUS_TRAPTYPE(hdrp->s.IProcTrapStatus));
+                       break;
+                   }
+
+                   IncrStat (rail, IprocDmaQueueOverflow);
+
+                   cp = LookupEventCookie (rail, &rail->CookieTable, dmap->s.dma_srcEvent);
+
+                   /* modify the dma type since it will still be a "read" dma */
+                   dmap->s.dma_type = (dmap->s.dma_type & ~DMA_TYPE_READ) | DMA_TYPE_ISREMOTE;
+
+#ifdef DEBUG_ASSERT
+                   {
+                       E3_uint16     vp       = dmap->s.dma_destVProc;
+                       EP_NODE_RAIL *nodeRail = &rail->Generic.Nodes[EP_VP_TO_NODE(vp)];
+                       
+                       ASSERT (!EP_VP_ISDATA(vp) || (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE));
+                   }
+#endif
+                   if (cp != NULL)
+                       cp->Operations->DmaRetry (rail, cp->Arg, dmap, EAGAIN);
+                   else
+                   {
+                       ASSERT (dmap->s.dma_direction == DMA_WRITE && dmap->s.dma_srcEvent == 0 && dmap->s.dma_isRemote);
+                       
+                       QueueDmaForRetry (rail, dmap, EP_RETRY_ANONYMOUS);
+                   }
+                   break;
+               }   
+               default:
+                   printk ("ep3_iprocTrap: %s\n", IProcTrapString (hdrp, datap));
+                   break;
+               }
+           }
+           
+           /*
+            * We've successfully processed this transaction, so move onto the 
+            * next one.
+            */
+           trap->TrappedTransaction++;
+           trap->TrappedDataBuffer++;
+       }
+
+       return (OP_HANDLED);
+    }
+    
+    /* Workaround WRITEBLOCK transaction executed when LOCKQUEUE transaction missed */
+    if ((trap->TrappedTransaction->s.TrTypeCntx.s.Type & TR_WRITEBLOCK_BIT) &&         /* a DMA packet */
+       trap->LockQueuePointer == 0 && trap->UnlockQueuePointer &&              /* a queueing DMA */
+       trap->TrappedTransaction->s.TrAddr == trap->FaultSave.s.FaultAddress)   /* and missed lockqueue */
+    {
+       printk ("ep3_iprocTrap: missed lockqueue transaction for queue %x\n", trap->UnlockQueuePointer);
+       return (OP_HANDLED);
+    }
+
+    if (trap->FaultSave.s.FaultContext != 0)
+       printk ("ep3_iprocTrap: pagefault at %08x in context %x\n", 
+               trap->FaultSave.s.FaultAddress, trap->FaultSave.s.FaultContext);
+
+//    panic ("ep3_iprocTrap: unexpected inputter trap\n");
+    
+    return (OP_HANDLED);
+}
+
+/*
+ * Command processor trap
+ *   kernel comms should only be able to generate
+ *   queue overflow traps
+ */
+int
+ep3_cprocTrap (ELAN3_CTXT *ctxt, COMMAND_TRAP *trap)
+{
+    EP3_RAIL     *rail   = (EP3_RAIL *) ctxt->Private;
+    int           ctxnum = (trap->TrapBuf.r.Breg >> 16) & MAX_ROOT_CONTEXT_MASK;
+    ELAN3_DEV     *dev    = rail->Device;
+    EP3_DMA_RING  *ring;
+    EP3_COOKIE   *cp;
+    E3_DMA_BE     dmabe;
+    int           vp, slot;
+    unsigned long flags;
+
+    switch (trap->Status.s.TrapType)
+    {
+    case MI_DmaQueueOverflow:
+       IncrStat (rail, CprocDmaQueueOverflow);
+
+       /* Use the context number that the setevent was issued in,
+        * to find the appropriate dma ring, then since they are guaranteed
+        * to be issued in order, we just search backwards till we find the
+        * last one which has completed its word copy - this must be the
+        * one which had caused the DmaQueueOverflow trap ! */
+
+       ASSERT (ctxnum >= ELAN3_DMARING_BASE_CONTEXT_NUM && ctxnum < (ELAN3_DMARING_BASE_CONTEXT_NUM+EP3_NUM_RINGS));
+
+       spin_lock_irqsave (&dev->CProcLock, flags);
+
+       ring = &rail->DmaRings[ctxnum - ELAN3_DMARING_BASE_CONTEXT_NUM];
+       slot = DMA_RING_PREV_POS(ring, ring->Position);
+       
+       while (ring->pDoneBlk[slot] == EP3_EVENT_ACTIVE)
+           slot = DMA_RING_PREV_POS(ring, slot);
+       
+       elan3_sdram_copyq_from_sdram (rail->Device , DMA_RING_DMA(ring,slot), &dmabe, sizeof (E3_DMA));
+
+#if defined(DEBUG_ASSERT)
+       while (slot != DMA_RING_PREV_POS(ring, ring->Position))
+       {
+           ASSERT (ring->pDoneBlk[slot] != EP3_EVENT_ACTIVE);
+           
+           slot = DMA_RING_PREV_POS(ring, slot);
+       }
+#endif
+       spin_unlock_irqrestore (&dev->CProcLock, flags);
+
+       if (dmabe.s.dma_direction == DMA_WRITE)
+           cp = LookupEventCookie (rail, &rail->CookieTable, dmabe.s.dma_srcEvent);
+       else
+       {
+           ASSERT (dmabe.s.dma_direction = DMA_READ_REQUEUE);
+
+           cp = LookupEventCookie (rail, &rail->CookieTable, dmabe.s.dma_destEvent);
+       }
+
+#if defined(DEBUG_ASSERT)
+       if (dmabe.s.dma_direction == DMA_WRITE)
+           vp = dmabe.s.dma_destVProc;
+       else
+           vp = dmabe.s.dma_srcVProc;
+       
+       ASSERT (!EP_VP_ISDATA(vp) || (rail->Generic.Nodes[EP_VP_TO_NODE(vp)].State >= EP_NODE_CONNECTED &&
+                                     rail->Generic.Nodes[EP_VP_TO_NODE(vp)].State <= EP_NODE_LOCAL_PASSIVATE));
+#endif
+
+       if (cp != NULL)
+           cp->Operations->DmaRetry (rail, cp->Arg, &dmabe, EAGAIN);
+       else
+       {
+           ASSERT (dmabe.s.dma_direction == DMA_WRITE && dmabe.s.dma_srcEvent == 0 && dmabe.s.dma_isRemote);
+           
+           QueueDmaForRetry (rail, &dmabe, EP_RETRY_ANONYMOUS);
+       }
+       
+       return (OP_HANDLED);
+
+    case MI_EventQueueOverflow:
+       ASSERT (ctxnum == ELAN3_MRF_CONTEXT_NUM);
+
+       IncrStat (rail, CprocEventQueueOverflow);
+       
+       rail->CommandPortEventTrap = TRUE;
+       return (OP_HANDLED);
+       
+#if defined(PER_CPU_TIMEOUT)
+    case MI_SetEventReadWait:
+       if (ctxnum == ELAN3_MRF_CONTEXT_NUM && trap->FaultSave.s.EventAddress == EP_PACEMAKER_EVENTADDR)
+       {
+           HeartbeatPacemaker (rail);
+           return (OP_HANDLED);
+       }
+#endif
+
+    default:
+       printk ("ep3_cprocTrap : Context=%x Status=%x TrapType=%x\n", ctxnum, trap->Status.Status, trap->Status.s.TrapType);
+       printk ("               FaultAddr=%x EventAddr=%x FSR=%x\n",
+                trap->FaultSave.s.FaultAddress, trap->FaultSave.s.EventAddress,
+                trap->FaultSave.s.FSR.Status);
+       break;
+    }
+
+//    panic ("ep3_cprocTrap");
+
+    return (OP_HANDLED);
+}
+
+static int
+ep3_cprocReissue (ELAN3_CTXT *ctxt, CProcTrapBuf_BE *tbuf)
+{
+    EP3_RAIL   *rail    = (EP3_RAIL *) ctxt->Private;
+    unsigned  cmdoff = (tbuf->s.ContextType >> 5) & 0xFF;
+    int       ctxnum = (tbuf->s.ContextType >> 16) & MAX_ROOT_CONTEXT_MASK;
+    
+    if (ctxnum >= ELAN3_DMARING_BASE_CONTEXT_NUM && ctxnum < (ELAN3_DMARING_BASE_CONTEXT_NUM+EP3_NUM_RINGS))
+    {
+       EP3_DMA_RING *ring = &rail->DmaRings[ctxnum - ELAN3_DMARING_BASE_CONTEXT_NUM];
+
+       ASSERT ((cmdoff << 2) == offsetof (E3_CommandPort, SetEvent)); /* can only be setevent commands! */
+       ASSERT (tbuf->s.Addr >= DMA_RING_EVENT_ELAN(ring,0) && tbuf->s.Addr < DMA_RING_EVENT_ELAN(ring, ring->Entries));
+       
+       writel (tbuf->s.Addr, ring->CommandPort + (cmdoff << 2));
+    }
+    else
+    {
+       ASSERT (ctxnum == ELAN3_MRF_CONTEXT_NUM);
+
+       writel (tbuf->s.Addr, ctxt->CommandPort + (cmdoff << 2));
+    }
+    
+    return (OP_HANDLED);
+}
+
+static E3_uint8
+ep3_load8 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    EP3_RAIL    *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV    *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint8   *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       return (elan3_sdram_readb (dev, offset));
+    if ((ptr = ep_elan2main (&rail->Generic, addr)) != NULL)
+       return (*ptr);
+
+    printk ("ep3_load8: %08x\n", addr);
+    return (0);
+}
+
+static void
+ep3_store8 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint8 val)
+{
+    EP3_RAIL   *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV   *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint8   *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       elan3_sdram_writeb (dev, offset, val);
+    else if ((ptr = ep_elan2main (&rail->Generic, addr)) != 0)
+       *ptr = val;
+    else
+       printk ("ep3_store8 %08x\n", addr);
+}
+
+static E3_uint16
+ep3_load16 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    EP3_RAIL   *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV   *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint16  *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       return (elan3_sdram_readw (dev, offset));
+    if ((ptr = ep_elan2main (&rail->Generic, addr)) != 0)
+       return (*ptr);
+
+    printk ("ep3_load16 %08x\n", addr);
+    return (0);
+}
+
+static void
+ep3_store16 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint16 val)
+{
+    EP3_RAIL   *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV   *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint16  *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       elan3_sdram_writew (dev, offset, val);
+    else if ((ptr = ep_elan2main (&rail->Generic, addr)) != 0)
+       *ptr = val;
+    else
+       printk ("ep3_store16 %08x\n", addr);
+}
+
+static E3_uint32
+ep3_load32 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    EP3_RAIL   *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV   *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint32  *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       return (elan3_sdram_readl(dev, offset));
+    if ((ptr = ep_elan2main (&rail->Generic, addr)) != 0)
+       return (*ptr);
+    
+    printk ("ep3_load32 %08x\n", addr);
+    return (0);
+}
+
+static void
+ep3_store32 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint32 val)
+{
+    EP3_RAIL   *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV   *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint32  *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       elan3_sdram_writel (dev, offset, val);
+    else if ((ptr = ep_elan2main (&rail->Generic, addr)) != 0)
+       *ptr = val;
+    else
+       printk ("ep3_store32 %08x\n", addr);
+}
+
+static E3_uint64
+ep3_load64 (ELAN3_CTXT *ctxt, E3_Addr addr)
+{
+    EP3_RAIL   *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV   *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint64  *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       return (elan3_sdram_readq (dev, offset));
+    if ((ptr = ep_elan2main (&rail->Generic, addr)) != 0)
+       return (*ptr);
+
+    printk ("ep3_load64 %08x\n", addr);
+    return (0);
+}
+
+static void
+ep3_store64 (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint64 val)
+{
+    EP3_RAIL   *rail  = (EP3_RAIL *) ctxt->Private;
+    ELAN3_DEV   *dev = ctxt->Device;
+    sdramaddr_t offset;
+    E3_uint64  *ptr;
+
+    if ((offset = ep_elan2sdram (&rail->Generic, addr)) != 0)
+       elan3_sdram_writeq (dev, offset, val);
+    else if ((ptr = ep_elan2main (&rail->Generic, addr)) != 0)
+       *ptr = val;
+    else
+       printk ("ep3_store64 %08x\n", addr);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/support_elan4.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/support_elan4.c      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/support_elan4.c   2005-06-01 23:12:54.689425424 -0400
@@ -0,0 +1,1184 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: support_elan4.c,v 1.18.2.3 2004/11/18 12:05:00 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/support_elan4.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/kthread.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_vp.h"
+#include "kcomm_elan4.h"
+#include "debug.h"
+
+#include <elan4/trtype.h>
+#include <elan4/debug.h>
+
+void
+ep4_register_intcookie (EP4_RAIL *rail, EP4_INTCOOKIE *cp, E4_uint64 cookie, void (*callback)(EP4_RAIL *r, void *arg), void *arg)
+{
+    unsigned long flags;
+    
+    cp->int_val      = cookie;
+    cp->int_callback = callback;
+    cp->int_arg      = arg;
+       
+    spin_lock_irqsave (&rail->r_intcookie_lock, flags);
+    list_add_tail (&cp->int_link, &rail->r_intcookie_hash[EP4_INTCOOKIE_HASH(cookie)]);
+    spin_unlock_irqrestore (&rail->r_intcookie_lock, flags);
+}
+
+void
+ep4_deregister_intcookie (EP4_RAIL *rail, EP4_INTCOOKIE *cp)
+{
+    unsigned long flags;
+    
+    spin_lock_irqsave (&rail->r_intcookie_lock, flags);
+    list_del (&cp->int_link);
+    spin_unlock_irqrestore (&rail->r_intcookie_lock, flags);
+}
+
+
+EP4_INTCOOKIE *
+ep4_lookup_intcookie (EP4_RAIL *rail, E4_uint64 cookie)
+{
+    struct list_head *el;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->r_intcookie_lock, flags);
+    list_for_each (el, &rail->r_intcookie_hash[EP4_INTCOOKIE_HASH(cookie)]) {
+       EP4_INTCOOKIE *cp = list_entry (el, EP4_INTCOOKIE, int_link);
+
+       if (cp->int_val == cookie)
+       {
+           spin_unlock_irqrestore (&rail->r_intcookie_lock, flags);
+           return cp;
+       }
+    }
+    spin_unlock_irqrestore (&rail->r_intcookie_lock, flags);
+    return NULL;
+}
+
+E4_uint64
+ep4_neterr_cookie (EP4_RAIL *rail, unsigned int node)
+{
+    E4_uint64      cookie;
+    unsigned long  flags;
+
+    spin_lock_irqsave (&rail->r_cookie_lock, flags);
+    cookie = rail->r_cookies[node];
+
+    rail->r_cookies[node] += EP4_COOKIE_INC;
+    
+    spin_unlock_irqrestore (&rail->r_cookie_lock, flags);
+
+    return cookie;
+}
+
+void
+ep4_eproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status)
+{
+    EP4_RAIL        *rail = EP4_CTXT_TO_RAIL (ctxt);
+    ELAN4_EPROC_TRAP trap;
+
+    elan4_extract_eproc_trap (ctxt->ctxt_dev, status, &trap, 0);
+
+    if (epdebug & DBG_EPTRAP)
+       elan4_display_eproc_trap (DBG_BUFFER, 0, "ep4_eproc_trap", &trap);
+
+    switch (EPROC_TrapType (status))
+    {
+    case EventProcNoFault:
+       EPRINTF1 (DBG_EPTRAP, "%s: EventProcNoFault\n", rail->r_generic.Name);
+       return;
+
+    default:
+       printk ("%s: unhandled eproc trap %d\n", rail->r_generic.Name, EPROC_TrapType (status));
+       elan4_display_eproc_trap (DBG_CONSOLE, 0, "ep4_eproc_trap", &trap);
+    }
+}
+
+void
+ep4_cproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned cqnum)
+{
+    EP4_RAIL        *rail = EP4_CTXT_TO_RAIL (ctxt);
+    ELAN4_CPROC_TRAP trap;
+    struct list_head *el;
+    register int      i;
+
+    elan4_extract_cproc_trap (ctxt->ctxt_dev, status, &trap, cqnum);
+
+    if (epdebug & DBG_EPTRAP)
+       elan4_display_cproc_trap (DBG_BUFFER, 0, "ep4_cproc_trap", &trap);
+       
+    switch (CPROC_TrapType (status))
+    {
+    case CommandProcInterruptQueueOverflow:
+       /*
+        * Try and handle a bunch of elan main interrupts
+        */
+       for (i = 0; i <EP4_NUM_ECQ; i++) {
+           list_for_each (el, &rail->r_ecq_list[i]) {
+               EP4_ECQ *ecq = list_entry (el, EP4_ECQ, ecq_link);
+           
+               if (elan4_cq2num (ecq->ecq_cq) == cqnum)
+               {
+                   printk ("%s: defer command queue %d after trap %x\n",
+                           rail->r_generic.Name, cqnum, CPROC_TrapType (status));
+       
+                   elan4_queue_mainintop (ctxt->ctxt_dev, &ecq->ecq_intop);
+                   return;
+               }
+           }
+       }
+       break;
+
+    case CommandProcDmaQueueOverflow:
+    case CommandProcThreadQueueOverflow:
+       for (i = 0; i <EP4_NUM_ECQ; i++) {
+           list_for_each (el, &rail->r_ecq_list[i]) {
+               EP4_ECQ *ecq = list_entry (el, EP4_ECQ, ecq_link);
+           
+               if (elan4_cq2num (ecq->ecq_cq) == cqnum)
+               {
+                   printk ("%s: restart command queue %d after trap %x\n",
+                           rail->r_generic.Name, cqnum, CPROC_TrapType (status));
+
+                   elan4_restartcq (ctxt->ctxt_dev, ecq->ecq_cq);
+                   return;
+               }
+           }
+       }
+       break;
+    }
+
+    printk ("%s: unhandled cproc trap %d for cqnum %d\n", rail->r_generic.Name, CPROC_TrapType (status), cqnum);
+    elan4_display_cproc_trap (DBG_CONSOLE, 0, "ep4_cproc_trap", &trap);
+}
+
+void
+ep4_dproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit)
+{
+    EP4_RAIL        *rail = EP4_CTXT_TO_RAIL (ctxt);
+    ELAN4_DPROC_TRAP trap;
+
+    elan4_extract_dproc_trap (ctxt->ctxt_dev, status, &trap, unit);
+
+    if (epdebug & DBG_EPTRAP)
+       elan4_display_dproc_trap (DBG_BUFFER, 0, "ep4_dproc_trap", &trap);
+
+    if (! DPROC_PrefetcherFault (trap.tr_status))
+    {
+       switch (DPROC_TrapType (trap.tr_status))
+       {
+       case DmaProcFailCountError:
+           goto retry_this_dma;
+
+       case DmaProcPacketAckError:
+           goto retry_this_dma;
+
+       case DmaProcQueueOverflow:
+           goto retry_this_dma;
+       }
+    }
+
+    printk ("%s: unhandled dproc trap\n", rail->r_generic.Name);
+    elan4_display_dproc_trap (DBG_CONSOLE, 0, "ep4_dproc_trap", &trap);
+    return;
+
+ retry_this_dma:
+    /*XXXX implement backoff .... */
+
+    ep4_queue_dma_retry (rail, &trap.tr_desc, EP_RETRY_LOW_PRI);
+}
+
+void
+ep4_tproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status)
+{
+    EP4_RAIL         *rail = EP4_CTXT_TO_RAIL (ctxt);
+    ELAN4_TPROC_TRAP *trap = &rail->r_tproc_trap;
+
+    elan4_extract_tproc_trap (ctxt->ctxt_dev, status, trap);
+
+    if (epdebug & DBG_EPTRAP)
+       elan4_display_tproc_trap (DBG_BUFFER, 0, "ep4_tproc_trap", trap);
+       
+    printk ("%s: unhandled tproc trap\n", rail->r_generic.Name);
+    elan4_display_tproc_trap (DBG_CONSOLE, 0, "ep4_tproc_trap", trap);
+}
+
+void
+ep4_iproc_trap (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit)
+{
+    EP4_RAIL         *rail = EP4_CTXT_TO_RAIL (ctxt);
+    ELAN4_IPROC_TRAP *trap = &rail->r_iproc_trap;
+
+    elan4_extract_iproc_trap (ctxt->ctxt_dev, status, trap, unit);
+
+    if (epdebug & DBG_EPTRAP)
+       elan4_display_iproc_trap (DBG_BUFFER, 0, "ep4_iproc_trap", trap);
+       
+    elan4_inspect_iproc_trap (trap);
+
+    switch (IPROC_TrapValue (trap->tr_transactions[trap->tr_trappedTrans].IProcStatusCntxAndTrType))
+    {
+    case InputDmaQueueOverflow:
+       ep4_queue_dma_retry (rail, (E4_DMA *) &trap->tr_dataBuffers[trap->tr_trappedTrans], EP_RETRY_LOW_PRI);
+       return;
+
+    case InputEventEngineTrapped:
+    {
+       E4_IprocTrapHeader *hdrp = &trap->tr_transactions[trap->tr_trappedTrans];
+       sdramaddr_t         inputq;
+       E4_Addr             event;
+
+       /* XXXX: flow control on the command queue which we issue to is 
+        * rather difficult, we don't want to have space for an event 
+        * for each possible context, nor the mechanism to hold the 
+        * context filter up until the event has been executed.  Given
+        * that the event engine will be restarted by this same interrupt
+        * and we're using high priority command queues, then we just use
+        * a single small command queue for this.
+        */
+       switch (IPROC_TransactionType(hdrp->IProcStatusCntxAndTrType) & TR_OPCODE_MASK)
+       {
+       case TR_SETEVENT & TR_OPCODE_MASK:
+           if (hdrp->TrAddr != 0)
+               ep4_set_event_cmd (rail->r_event_ecq, hdrp->TrAddr);
+           return;
+
+       case TR_INPUT_Q_COMMIT & TR_OPCODE_MASK:
+           if ((inputq = ep_elan2sdram (&rail->r_generic, hdrp->TrAddr)) == 0)
+               printk ("%s: TR_INPUT_Q_COMMIT at %llx is not sdram\n", rail->r_generic.Name, hdrp->TrAddr);
+           else
+           {
+               if ((event = elan4_sdram_readq (rail->r_ctxt.ctxt_dev, inputq + offsetof (E4_InputQueue, q_event))) != 0)
+                   ep4_set_event_cmd (rail->r_event_ecq, event);
+               return;
+           }
+       }
+       break;
+    }
+
+    case InputEopErrorOnWaitForEop:
+    case InputEopErrorTrap:
+    case InputCrcErrorAfterPAckOk:
+       if (! (trap->tr_flags & TR_FLAG_ACK_SENT) || (trap->tr_flags & TR_FLAG_EOP_BAD))
+           return;
+       
+       if (EP4_CONTEXT_ISDATA (IPROC_NetworkContext (status)))
+       {
+           unsigned int nodeId = EP4_CONTEXT_TO_NODE (IPROC_NetworkContext (status));
+
+           if ((trap->tr_flags & (TR_FLAG_DMA_PACKET | TR_FLAG_BAD_TRANS)) || 
+               ((trap->tr_flags & TR_FLAG_EOP_ERROR) && (trap->tr_identifyTrans == TR_TRANS_INVALID)))
+           {
+               printk ("%s: network error on dma packet from node %d\n", rail->r_generic.Name, nodeId);
+
+               ep_queue_network_error (&rail->r_generic, EP4_CONTEXT_TO_NODE(IPROC_NetworkContext (status)), EP_NODE_NETERR_DMA_PACKET, unit & 1, 0);
+               return;
+           }
+           
+           if (trap->tr_flags & TR_FLAG_EOP_ERROR)
+           {
+               E4_uint64        status = trap->tr_transactions[trap->tr_identifyTrans].IProcStatusCntxAndTrType;
+               EP_NETERR_COOKIE cookie = 0;
+
+               switch (IPROC_TransactionType (status) & TR_OPCODE_MASK)
+               {
+               case TR_SETEVENT_IDENTIFY & TR_OPCODE_MASK:
+                   if (IPROC_TrapValue(status) == InputNoFault)
+                       cookie = trap->tr_transactions[trap->tr_identifyTrans].TrAddr;
+                   else
+                       cookie = trap->tr_dataBuffers[trap->tr_identifyTrans].Data[0];
+                   printk ("%s: network error on setevent <%lld%s%s%s%s> from node %d\n", rail->r_generic.Name, EP4_COOKIE_STRING(cookie), nodeId);
+                   break;
+
+               case TR_INPUT_Q_COMMIT & TR_OPCODE_MASK:
+                   if (IPROC_TrapValue(status) == InputNoFault)
+                       cookie = trap->tr_transactions[trap->tr_identifyTrans].TrAddr;
+                   else
+                       cookie = trap->tr_dataBuffers[trap->tr_identifyTrans].Data[0];
+                   printk ("%s: network error on queue commit <%lld%s%s%s%s> from node %d\n", rail->r_generic.Name, EP4_COOKIE_STRING(cookie), nodeId);
+                   break;
+                   
+               case TR_REMOTEDMA & TR_OPCODE_MASK:
+                   cookie = trap->tr_transactions[trap->tr_identifyTrans].TrAddr;
+                   printk ("%s: network error on remote dma <%lld%s%s%s%s> from node %d\n", rail->r_generic.Name, EP4_COOKIE_STRING(cookie), nodeId);
+                   break;
+
+               case TR_IDENTIFY & TR_OPCODE_MASK:
+                   cookie = trap->tr_transactions[trap->tr_identifyTrans].TrAddr;
+                   printk ("%s: network error on identify <%lld%s%s%s%s> from node %d\n", rail->r_generic.Name, EP4_COOKIE_STRING(cookie), nodeId);
+                   break;
+
+               default:
+                   panic ("%s: unknown identify transaction type %x for eop error from node %d\n", rail->r_generic.Name,
+                           IPROC_TransactionType (trap->tr_transactions[trap->tr_identifyTrans].IProcStatusCntxAndTrType), nodeId);
+                   break;
+               }
+
+               ep_queue_network_error (&rail->r_generic, nodeId, EP_NODE_NETERR_ATOMIC_PACKET, unit & 1, cookie);
+           }
+       }
+       return;
+    }
+
+    printk ("%s: unhandled iproc trap\n", rail->r_generic.Name);
+    elan4_display_iproc_trap (DBG_CONSOLE, 0, "ep4_iproc_trap", trap);
+}
+
+void
+ep4_interrupt (ELAN4_CTXT *ctxt, E4_uint64 cookie)
+{
+    EP4_RAIL      *rail = EP4_CTXT_TO_RAIL (ctxt);
+    EP4_INTCOOKIE *cp  = ep4_lookup_intcookie (rail, cookie);
+
+    if (cp == NULL)
+    {
+       printk ("ep4_interrupt: cannot find event cookie for %016llx\n", (long long) cookie);
+       return;
+    }
+
+    cp->int_callback (rail, cp->int_arg);
+}
+
+ELAN4_TRAP_OPS ep4_trap_ops = 
+{
+    ep4_eproc_trap,
+    ep4_cproc_trap,
+    ep4_dproc_trap,
+    ep4_tproc_trap,
+    ep4_iproc_trap,
+    ep4_interrupt,
+};
+
+void
+ep4_flush_filters (EP_RAIL *r)
+{
+    /* nothing to do here as elan4_set_filter() flushes the context filter */
+}
+
+struct flush_queues_desc
+{
+    EP4_RAIL      *rail;
+    volatile int   done;
+} ;
+
+static void
+ep4_flush_queues_flushop (ELAN4_DEV *dev, void *arg, int qfull)
+{
+    struct flush_queues_desc *desc  = (struct flush_queues_desc *) arg;
+    EP4_RAIL                 *rail  = desc->rail;
+    E4_uint64                qptrs = read_reg64 (dev, DProcHighPriPtrs);
+    E4_uint32                 qsize = E4_QueueSize (E4_QueueSizeValue (qptrs));
+    E4_uint32                 qfptr = E4_QueueFrontPointer (qptrs);
+    E4_uint32                 qbptr = E4_QueueBackPointer (qptrs);
+    E4_DProcQueueEntry        qentry;
+    unsigned long             flags;
+
+    while ((qfptr != qbptr) || qfull)
+    {
+       E4_uint64 typeSize = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_typeSize));
+       
+       if (DMA_Context (qentry.Desc.dma_typeSize) == rail->r_ctxt.ctxt_num)
+       {
+           E4_uint64     vp       = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_vproc));
+           EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[EP_VP_TO_NODE(vp)];
+           
+           EP4_ASSERT (rail, !EP_VP_ISDATA(vp) || (nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE));
+           
+           if (EP_VP_ISDATA(vp) && nodeRail->State == EP_NODE_LOCAL_PASSIVATE)
+           {
+               /*
+                * This is a DMA going to the node which is being removed, 
+                * so move it onto the node dma list where it will get
+                * handled later.
+                */
+               qentry.Desc.dma_typeSize = typeSize;
+               qentry.Desc.dma_cookie   = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_cookie));
+               qentry.Desc.dma_vproc    = vp;
+               qentry.Desc.dma_srcAddr  = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_srcAddr));
+               qentry.Desc.dma_dstAddr  = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_dstAddr));
+               qentry.Desc.dma_srcEvent = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_srcEvent));
+               qentry.Desc.dma_dstEvent = elan4_sdram_readq (dev, qfptr + offsetof (E4_DProcQueueEntry, Desc.dma_dstEvent));
+               
+               EPRINTF4 (DBG_RETRY, "ep4_flush_dmas: %016llx %016llx %016llx %016llx\n", qentry.Desc.dma_typeSize, 
+                         qentry.Desc.dma_cookie, qentry.Desc.dma_vproc, qentry.Desc.dma_srcAddr);
+               EPRINTF3 (DBG_RETRY, "                %016llx %016llx %016llx\n", qentry.Desc.dma_dstAddr, 
+                         qentry.Desc.dma_srcEvent, qentry.Desc.dma_dstEvent);
+               
+               ep4_queue_dma_stalled (rail, &qentry.Desc);
+               
+               qentry.Desc.dma_typeSize = DMA_ShMemWrite | dev->dev_ctxt.ctxt_num;
+               qentry.Desc.dma_cookie   = 0;
+               qentry.Desc.dma_vproc    = 0;
+               qentry.Desc.dma_srcAddr  = 0;
+               qentry.Desc.dma_dstAddr  = 0;
+               qentry.Desc.dma_srcEvent = 0;
+               qentry.Desc.dma_dstEvent = 0;
+               
+               elan4_sdram_copyq_to_sdram (dev, &qentry, qfptr, sizeof (E4_DProcQueueEntry));
+           }
+       }
+
+       qfptr = (qfptr & ~(qsize-1)) | ((qfptr + sizeof (E4_DProcQueueEntry)) & (qsize-1));
+       qfull = 0;
+    }
+
+    spin_lock_irqsave (&rail->r_haltop_lock, flags);
+    desc->done = 1;
+    kcondvar_wakeupall (&rail->r_haltop_sleep, &rail->r_haltop_lock);
+    spin_unlock_irqrestore (&rail->r_haltop_lock, flags);
+}
+
+static void
+ep4_flush_queues_haltop (ELAN4_DEV *dev, void *arg)
+{
+    struct flush_queues_desc *desc = (struct flush_queues_desc *) arg;
+
+    elan4_queue_dma_flushop (dev, &desc->rail->r_flushop, 1);
+}
+
+void
+ep4_flush_queues (EP_RAIL *r)
+{
+    EP4_RAIL *rail = (EP4_RAIL *) r;
+    struct flush_queues_desc desc;
+    struct list_head *el, *nel;
+    unsigned long flags;
+    int i;
+
+    /* initialise descriptor */
+    desc.rail  = rail;
+    desc.done  = 0;
+
+    /* First -  stall the dma retry thread, so that it will no longer restart
+     *          any dma's from the retry list */
+    ep_kthread_stall (&rail->r_retry_thread);
+
+    /* Second - flush through all command queues targetted by events, thread etc */
+    ep4_flush_ecqs (rail);
+
+    /* Third - queue a halt operation to flush through all DMA's which are executing
+     *         or on the run queues */
+    kmutex_lock (&rail->r_haltop_mutex);
+
+    rail->r_haltop.op_mask      = INT_DProcHalted;
+    rail->r_haltop.op_function  = ep4_flush_queues_haltop;
+    rail->r_haltop.op_arg       = &desc;
+
+    rail->r_flushop.op_function = ep4_flush_queues_flushop;
+    rail->r_flushop.op_arg      = &desc;
+    
+    elan4_queue_haltop (rail->r_ctxt.ctxt_dev, &rail->r_haltop);
+
+    spin_lock_irqsave (&rail->r_haltop_lock, flags);
+    while (! desc.done)
+       kcondvar_wait (&rail->r_haltop_sleep, &rail->r_haltop_lock, &flags);
+    spin_unlock_irqrestore (&rail->r_haltop_lock, flags);
+    kmutex_unlock (&rail->r_haltop_mutex);
+
+    /* Fourth - run down the dma retry lists and move all entries to the cancelled
+     *          list.  Any dma's which were on the run queues have already been
+     *          moved there */
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+    for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+    {
+       list_for_each_safe (el,nel, &rail->r_dma_retrylist[i]) {
+           EP4_DMA_RETRY *retry    = list_entry (el, EP4_DMA_RETRY, retry_link);
+           EP_NODE_RAIL  *nodeRail = &rail->r_generic.Nodes[EP_VP_TO_NODE(retry->retry_dma.dma_vproc)];
+
+           EP4_ASSERT (rail, nodeRail->State >= EP_NODE_CONNECTED && nodeRail->State <= EP_NODE_LOCAL_PASSIVATE);
+
+           if (nodeRail->State == EP_NODE_LOCAL_PASSIVATE)
+           {
+               list_del (&retry->retry_link);
+               list_add_tail (&retry->retry_link, &nodeRail->StalledDmas);
+           }
+       }
+    }
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+    
+    /* Finally - allow the retry thread to run again */
+    ep_kthread_resume (&rail->r_retry_thread);
+}
+
+struct write_qdesc_desc
+{
+    EP4_RAIL      *rail;
+    sdramaddr_t    qaddr;
+    E4_InputQueue *qdesc;
+    volatile int   done;
+} ;
+
+static void
+ep4_write_qdesc_haltop (ELAN4_DEV *dev, void *arg)
+{
+    struct write_qdesc_desc *desc = (struct write_qdesc_desc *) arg;
+    EP4_RAIL                *rail = desc->rail;
+    unsigned long            flags;
+
+    elan4_sdram_copyq_to_sdram (dev, desc->qdesc, desc->qaddr, sizeof (E4_InputQueue));
+
+    spin_lock_irqsave (&rail->r_haltop_lock, flags);
+    desc->done = 1;
+    kcondvar_wakeupall (&rail->r_haltop_sleep, &rail->r_haltop_lock);
+    spin_unlock_irqrestore (&rail->r_haltop_lock, flags);
+}
+
+void
+ep4_write_qdesc (EP4_RAIL *rail, sdramaddr_t qaddr, E4_InputQueue *qdesc)
+{
+    struct write_qdesc_desc desc;
+    unsigned long flags;
+
+    /* initialise descriptor */
+    desc.rail  = rail;
+    desc.qaddr = qaddr;
+    desc.qdesc = qdesc;
+    desc.done  = 0;
+
+    kmutex_lock (&rail->r_haltop_mutex);
+
+    rail->r_haltop.op_mask     = INT_DiscardingHighPri;
+    rail->r_haltop.op_function = ep4_write_qdesc_haltop;
+    rail->r_haltop.op_arg      = &desc;
+    
+    elan4_queue_haltop (rail->r_ctxt.ctxt_dev, &rail->r_haltop);
+
+    spin_lock_irqsave (&rail->r_haltop_lock, flags);
+    while (! desc.done)
+       kcondvar_wait (&rail->r_haltop_sleep, &rail->r_haltop_lock, &flags);
+    spin_unlock_irqrestore (&rail->r_haltop_lock, flags);
+    
+    kmutex_unlock (&rail->r_haltop_mutex);
+}
+#define CQ_SIZE_NWORDS ((CQ_Size (ecq->ecq_cq->cq_size) >> 3) - 8)     /* available number of dwords (less enough to flush) */
+EP4_ECQ *
+ep4_alloc_ecq (EP4_RAIL *rail, unsigned cqsize)
+{
+    EP4_ECQ *ecq;
+    unsigned long pgoff;
+
+    /* no space available, so allocate a new entry */
+    KMEM_ZALLOC (ecq, EP4_ECQ *, sizeof (EP4_ECQ), 1);
+
+    if (ecq == NULL)
+       return 0;
+
+    if ((ecq->ecq_cq = elan4_alloccq (&rail->r_ctxt, cqsize, CQ_EnableAllBits, CQ_Priority)) == NULL)
+    {
+       KMEM_FREE (ecq, sizeof (EP4_ECQ));
+       return 0;
+    }
+
+    pgoff = (ecq->ecq_cq->cq_mapping & (PAGE_SIZE-1));
+
+    ecq->ecq_addr  = ep_rmalloc (rail->r_ecq_rmap, PAGESIZE, 0) + pgoff;
+    ecq->ecq_avail = CQ_SIZE_NWORDS;                   /* available number of dwords (less enough to flush) */
+
+    ecq->ecq_intop.op_function = (ELAN4_HALTFN *) elan4_restartcq;
+    ecq->ecq_intop.op_arg      = ecq->ecq_cq;
+
+    ep4_ioaddr_map (&rail->r_generic, ecq->ecq_addr - pgoff, ecq->ecq_cq->cq_mapping - pgoff, PAGESIZE, EP_PERM_WRITE);
+
+    spin_lock_init (&ecq->ecq_lock);
+
+    return ecq;
+}
+
+void
+ep4_free_ecq (EP4_RAIL *rail, EP4_ECQ *ecq)
+{
+    unsigned long pgoff = (ecq->ecq_cq->cq_mapping & (PAGE_SIZE-1));
+
+    spin_lock_destroy (&ecq->ecq_lock);
+
+    ep4_unmap (&rail->r_generic, ecq->ecq_addr - pgoff, PAGESIZE);
+    ep_rmfree (rail->r_ecq_rmap, PAGESIZE, ecq->ecq_addr - pgoff);
+
+    elan4_freecq (&rail->r_ctxt, ecq->ecq_cq);
+    
+    KMEM_FREE (ecq, sizeof (EP4_ECQ));
+}
+
+EP4_ECQ *
+ep4_get_ecq (EP4_RAIL *rail, unsigned which, unsigned ndwords)
+{
+    ELAN4_DEV *dev = rail->r_ctxt.ctxt_dev;
+    struct list_head *el;
+    unsigned long flags;
+    EP4_ECQ *ecq;
+    
+    spin_lock_irqsave (&rail->r_ecq_lock, flags);
+    list_for_each (el, &rail->r_ecq_list[which]) {
+       EP4_ECQ *ecq = list_entry (el, EP4_ECQ, ecq_link);
+
+       if (ecq->ecq_avail >= ndwords)
+       {
+           ecq->ecq_avail -= ndwords;
+
+           spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+
+           return ecq;
+       }
+    }
+    spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+
+    if ((ecq = ep4_alloc_ecq (rail, EP4_ECQ_Size (which))) == NULL)
+       return NULL;
+
+    if (which == EP4_ECQ_EVENT)
+    {
+       if ((ecq->ecq_event = ep_alloc_elan (&rail->r_generic, sizeof (E4_Event32), 0, &ecq->ecq_event_addr)) == 0)
+       {
+           ep4_free_ecq (rail, ecq);
+           return NULL;
+       }
+       
+       elan4_sdram_writeq (dev, ecq->ecq_event + offsetof (E4_Event32, ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+       elan4_sdram_writeq (dev, ecq->ecq_event + offsetof (E4_Event32, ev_WritePtr),
+                           ecq->ecq_addr);
+       elan4_sdram_writeq (dev, ecq->ecq_event + offsetof (E4_Event32, ev_WriteValue),
+                           SET_EVENT_CMD | (rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_flush_event)));
+       
+       if ((ecq->ecq_flushcq = ep4_get_ecq (rail, EP4_ECQ_SINGLE, 1)) == NULL)
+       {
+           ep_free_elan (&rail->r_generic, ecq->ecq_event_addr, sizeof (E4_Event32));
+           ep4_free_ecq (rail, ecq);
+           return NULL;
+       }
+    }
+
+    spin_lock_irqsave (&rail->r_ecq_lock, flags);
+    list_add (&ecq->ecq_link, &rail->r_ecq_list[which]);
+
+    ecq->ecq_avail -= ndwords;
+    spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+
+    return ecq;
+}
+
+void
+ep4_put_ecq (EP4_RAIL *rail, EP4_ECQ *ecq, unsigned ndwords)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->r_ecq_lock, flags);
+
+    ecq->ecq_avail += ndwords;
+    
+    if (ecq->ecq_avail !=  CQ_SIZE_NWORDS) 
+       spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+    else
+    {
+       list_del (&ecq->ecq_link);
+       spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+       
+       if (ecq->ecq_flushcq)
+           ep4_put_ecq (rail, ecq->ecq_flushcq, 1);
+       if (ecq->ecq_event_addr)
+           ep_free_elan (&rail->r_generic, ecq->ecq_event_addr, sizeof (E4_Event32));
+
+       ep4_free_ecq (rail, ecq);
+    }
+}
+
+void
+ep4_nop_cmd (EP4_ECQ *ecq, E4_uint64 tag)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&ecq->ecq_lock, flags);
+    elan4_nop_cmd (ecq->ecq_cq, tag);
+    spin_unlock_irqrestore (&ecq->ecq_lock, flags);
+    
+}
+
+void
+ep4_set_event_cmd (EP4_ECQ *ecq, E4_Addr event)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&ecq->ecq_lock, flags);
+    elan4_set_event_cmd (ecq->ecq_cq, event);
+    spin_unlock_irqrestore (&ecq->ecq_lock, flags);
+}
+
+void
+ep4_wait_event_cmd (EP4_ECQ *ecq, E4_Addr event, E4_uint64 candt, E4_uint64 param0, E4_uint64 param1)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&ecq->ecq_lock, flags);
+    elan4_wait_event_cmd (ecq->ecq_cq, event, candt, param0, param1);
+    spin_unlock_irqrestore (&ecq->ecq_lock, flags);
+}
+
+void
+ep4_flush_interrupt (EP4_RAIL *rail, void *arg)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->r_ecq_lock, flags);
+    rail->r_flush_count = 0;
+    kcondvar_wakeupone (&rail->r_flush_sleep, &rail->r_ecq_lock);
+    spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+}
+
+void
+ep4_flush_ecqs (EP4_RAIL *rail)
+{
+    ELAN4_DEV *dev = rail->r_ctxt.ctxt_dev;
+    struct list_head *el;
+    unsigned long flags;
+    int i;
+
+    kmutex_lock (&rail->r_flush_mutex);
+
+    EP4_SDRAM_ASSERT (rail, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_flush_event), E4_EVENT_INIT_VALUE (0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG,0));
+
+    spin_lock_irqsave (&rail->r_ecq_lock, flags);
+    /* first flush all the "event" queues */
+    list_for_each (el, &rail->r_ecq_list[EP4_ECQ_EVENT]) {
+       EP4_ECQ *ecq = list_entry (el, EP4_ECQ, ecq_link);
+
+       elan4_sdram_writeq (dev, ecq->ecq_event + offsetof (E4_Event32, ev_CountAndType),
+                           E4_EVENT_INIT_VALUE (-32, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0));
+
+       ep4_set_event_cmd (ecq->ecq_flushcq, ecq->ecq_event_addr);
+
+       rail->r_flush_count++;
+    }
+
+    /* next issue the setevents to all the other queues */
+    for (i = EP4_ECQ_ATOMIC; i <EP4_NUM_ECQ; i++)
+    {
+       list_for_each (el,&rail->r_ecq_list[i]) {
+           EP4_ECQ *ecq = list_entry (el, EP4_ECQ, ecq_link);
+
+           ep4_set_event_cmd (ecq, rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_flush_event));
+
+           rail->r_flush_count++;
+       }
+    }
+
+    /* issue the waitevent command */
+    ep4_wait_event_cmd (rail->r_flush_mcq,  rail->r_elan_addr + offsetof (EP4_RAIL_ELAN, r_flush_event),
+                       E4_EVENT_INIT_VALUE (-32 * rail->r_flush_count, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG,0),
+                       rail->r_flush_ecq->ecq_addr,
+                       INTERRUPT_CMD | (rail->r_flush_intcookie.int_val << E4_MAIN_INT_SHIFT));
+    
+    while (rail->r_flush_count)
+       kcondvar_wait (&rail->r_flush_sleep, &rail->r_ecq_lock, &flags);
+    
+    spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+
+    EP4_SDRAM_ASSERT (rail, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_flush_event), E4_EVENT_INIT_VALUE (0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG,0));
+
+    kmutex_unlock (&rail->r_flush_mutex);
+}
+
+void
+ep4_init_thread (EP4_RAIL *rail, E4_ThreadRegs *regs, sdramaddr_t stackTop, 
+                EP_ADDR stackAddr, E4_Addr startpc, int nargs,...)
+{
+    sdramaddr_t   sp = stackTop - roundup (nargs * sizeof (E4_uint64), E4_STACK_ALIGN);
+    int           i;
+    va_list       ap;
+    
+    /*
+     * the thread start code expects the following :
+     *   %r1 = stack pointer
+     *   %r6 = frame pointer
+     *   %r2 = function to call
+     *
+     *   function args are store on stack above %sp
+     */
+
+    va_start(ap, nargs);
+    for (i = 0; i < nargs; i++)
+       elan4_sdram_writeq (rail->r_ctxt.ctxt_dev, sp + (i * sizeof (E4_uint64)), va_arg (ap, E4_uint64));
+    va_end (ap);
+    
+    regs->Registers[0] = ep_symbol (&rail->r_threadcode, ".thread_start");             /* %r0 - PC */
+    regs->Registers[1] = stackAddr - (stackTop - sp);                                  /* %r1 - stack pointer */
+    regs->Registers[2] = startpc;                                                      /* %r2 - start pc */
+    regs->Registers[3] = 0;
+    regs->Registers[4] = 0;
+    regs->Registers[5] = 0;
+    regs->Registers[6] = stackTop;                                                     /* %r6 - frame pointer */ 
+}
+
+/* retransmission thread */
+
+void
+ep4_add_retry_ops (EP4_RAIL *rail, EP4_RETRY_OPS *ops)
+{
+    ep_kthread_stall (&rail->r_retry_thread);
+    list_add_tail (&ops->op_link, &rail->r_retry_ops);
+    ep_kthread_resume (&rail->r_retry_thread);
+}
+
+void
+ep4_remove_retry_ops (EP4_RAIL *rail, EP4_RETRY_OPS *ops)
+{
+    ep_kthread_stall (&rail->r_retry_thread);
+    list_del (&ops->op_link);
+    ep_kthread_resume (&rail->r_retry_thread);
+}
+
+void
+ep4_retry_thread (EP4_RAIL *rail)
+{
+    struct list_head *el;
+
+    kernel_thread_init ("ep4_retry");
+    
+    for (;;)
+    {
+       long nextRunTime = 0;
+
+       list_for_each (el, &rail->r_retry_ops) {
+           EP4_RETRY_OPS *ops = list_entry (el, EP4_RETRY_OPS, op_link);
+
+           nextRunTime = ops->op_func (rail, ops->op_arg, nextRunTime);
+       }
+
+       if (ep_kthread_sleep (&rail->r_retry_thread, nextRunTime) < 0)
+           break;
+    }
+
+    ep_kthread_stopped (&rail->r_retry_thread);
+
+    kernel_thread_exit();
+}
+
+/* DMA retransmission */
+static unsigned ep4_dma_retry_times[EP_NUM_RETRIES];
+
+static unsigned long
+ep4_retry_dmas (EP4_RAIL *rail, void *arg, unsigned long nextRunTime)
+{
+    unsigned long yieldAt = lbolt + (hz/10);
+    unsigned long flags;
+    int           i;
+
+    for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+    {
+       while (! list_empty (&rail->r_dma_retrylist[i]))
+       {
+           EP4_DMA_RETRY *retry = list_entry (rail->r_dma_retrylist[i].next, EP4_DMA_RETRY, retry_link);
+           
+           if (! AFTER(lbolt, retry->retry_time))
+               break;
+
+           if (ep_kthread_should_stall (&rail->r_retry_thread) || AFTER (lbolt, yieldAt))
+               goto cant_do_more;
+           
+           EPRINTF3 (DBG_RETRY, "%s: ep4_retry_dmas: flowcnt %llx %llx\n", rail->r_generic.Name, rail->r_dma_flowcnt, rail->r_main->r_dma_flowcnt);
+
+           if ((rail->r_dma_flowcnt - rail->r_main->r_dma_flowcnt) > EP4_DMA_RETRY_FLOWCNT)
+           {
+               printk ("ep4_retry_dmas: flowcnt %llx %llx\n", rail->r_dma_flowcnt, rail->r_main->r_dma_flowcnt);
+
+               goto cant_do_more;
+           }
+
+           EPRINTF4 (DBG_RETRY, "%s: ep4_retry_dmas: %016llx %016llx %016llx\n", rail->r_generic.Name,
+                     retry->retry_dma.dma_typeSize, retry->retry_dma.dma_cookie, retry->retry_dma.dma_vproc);
+           EPRINTF5 (DBG_RETRY, "%s:                  %016llx %016llx %016llx %016llx\n", rail->r_generic.Name,
+                     retry->retry_dma.dma_srcAddr, retry->retry_dma.dma_dstAddr, retry->retry_dma.dma_srcEvent, 
+                     retry->retry_dma.dma_dstEvent);
+
+           elan4_run_dma_cmd (rail->r_dma_ecq->ecq_cq, &retry->retry_dma);
+           elan4_write_dword_cmd (rail->r_dma_ecq->ecq_cq, rail->r_main_addr + offsetof (EP4_RAIL_MAIN, r_dma_flowcnt), ++rail->r_dma_flowcnt);
+
+           spin_lock_irqsave (&rail->r_dma_lock, flags);
+           list_del (&retry->retry_link);
+           list_add (&retry->retry_link, &rail->r_dma_freelist);
+           spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+       }
+    }
+ cant_do_more:
+
+    /* re-compute the next retry time */
+    for (i = EP_RETRY_BASE; i < EP_NUM_RETRIES; i++)
+    {
+       if (! list_empty (&rail->r_dma_retrylist[i]))
+       {
+           EP4_DMA_RETRY *retry = list_entry (rail->r_dma_retrylist[i].next, EP4_DMA_RETRY, retry_link);
+
+           SET_NEXT_RUN_TIME (nextRunTime, retry->retry_time);
+       }
+    }
+
+    return nextRunTime;
+}
+
+void
+ep4_initialise_dma_retries (EP4_RAIL *rail)
+{
+    int i;
+
+    spin_lock_init (&rail->r_dma_lock);
+    
+    for (i = 0; i < EP_NUM_RETRIES; i++)
+       INIT_LIST_HEAD (&rail->r_dma_retrylist[i]);
+    
+    INIT_LIST_HEAD (&rail->r_dma_freelist);
+    
+    rail->r_dma_ecq = ep4_alloc_ecq (rail, EP4_DMA_RETRY_CQSIZE);
+    
+    rail->r_dma_allocated = 0;
+    rail->r_dma_reserved  = 0;
+
+    ep4_dma_retry_times[EP_RETRY_HIGH_PRI] = EP_RETRY_HIGH_PRI_TIME;
+
+    for (i =0 ; i < EP_NUM_BACKOFF; i++)
+       ep4_dma_retry_times[EP_RETRY_HIGH_PRI_RETRY+i] = EP_RETRY_HIGH_PRI_TIME << i;
+    
+    ep4_dma_retry_times[EP_RETRY_LOW_PRI] = EP_RETRY_LOW_PRI_TIME;
+
+    for (i =0 ; i < EP_NUM_BACKOFF; i++)
+       ep4_dma_retry_times[EP_RETRY_LOW_PRI_RETRY+i] = EP_RETRY_LOW_PRI_TIME << i;
+    
+    ep4_dma_retry_times[EP_RETRY_ANONYMOUS] = EP_RETRY_ANONYMOUS_TIME;
+    ep4_dma_retry_times[EP_RETRY_NETERR]    = EP_RETRY_NETERR_TIME;
+
+    rail->r_dma_ops.op_func = ep4_retry_dmas;
+    rail->r_dma_ops.op_arg  = NULL;
+
+    ep4_add_retry_ops (rail, &rail->r_dma_ops);
+}
+
+void
+ep4_finalise_dma_retries (EP4_RAIL *rail)
+{
+    ep4_remove_retry_ops (rail, &rail->r_dma_ops);
+
+    /* Everyone should have given back their retry dma's by now */
+    EP4_ASSERT (rail, rail->r_dma_reserved == 0);
+
+    while (! list_empty (&rail->r_dma_freelist))
+    {
+       EP4_DMA_RETRY *retry = list_entry (rail->r_dma_freelist.next, EP4_DMA_RETRY, retry_link);
+
+       list_del (&retry->retry_link);
+
+       KMEM_FREE (retry, sizeof (EP4_DMA_RETRY));
+    }
+
+    ep4_free_ecq (rail, rail->r_dma_ecq);
+
+    spin_lock_destroy (&rail->r_dma_lock);
+}
+
+int
+ep4_reserve_dma_retries (EP4_RAIL *rail, unsigned int count, EP_ATTRIBUTE attr)
+{
+    EP4_DMA_RETRY *retry;
+    unsigned int   remaining = count;
+    unsigned long  flags;
+
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+
+    if (remaining <= (rail->r_dma_allocated - rail->r_dma_reserved))
+    {
+       rail->r_dma_reserved += remaining;
+
+       spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+
+       return 0;
+    }
+    
+    remaining -= (rail->r_dma_allocated - rail->r_dma_reserved);
+
+    rail->r_dma_reserved = rail->r_dma_allocated;
+
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+
+    while (remaining > 0)
+    {
+       KMEM_ALLOC (retry, EP4_DMA_RETRY *, sizeof (EP4_DMA_RETRY), !(attr & EP_NO_SLEEP));
+
+       if (retry == NULL)
+           goto failed;
+       
+       remaining--;
+
+       spin_lock_irqsave (&rail->r_dma_lock, flags);
+       list_add (&retry->retry_link, &rail->r_dma_freelist);
+
+       rail->r_dma_allocated++;
+       rail->r_dma_reserved++;
+       spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+    }
+
+    return 0;
+
+ failed:
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+    rail->r_dma_reserved -= (count - remaining);
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+
+    return 1;
+}
+
+void
+ep4_release_dma_retries (EP4_RAIL *rail, unsigned int count)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+    rail->r_dma_reserved -= count;
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+}
+
+void
+ep4_queue_dma_retry (EP4_RAIL *rail, E4_DMA *dma, int interval)
+{
+    EP4_DMA_RETRY *retry;
+    unsigned long  flags;
+    
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+
+    EP4_ASSERT (rail, !list_empty (&rail->r_dma_freelist));
+    
+    /* take an item of the free list */
+    retry = list_entry (rail->r_dma_freelist.next, EP4_DMA_RETRY, retry_link);
+
+    list_del (&retry->retry_link);
+    
+    EPRINTF5 (DBG_RETRY, "%s: ep4_queue_dma_retry: %016llx %016llx %016llx %016llx\n", rail->r_generic.Name,
+             dma->dma_typeSize, dma->dma_cookie, dma->dma_vproc, dma->dma_srcAddr);
+    EPRINTF5 (DBG_RETRY, "%s:                      %016llx %016llx %016llx (%d)\n", rail->r_generic.Name,
+             dma->dma_dstAddr, dma->dma_srcEvent, dma->dma_dstEvent, interval);
+
+    retry->retry_dma.dma_typeSize = dma->dma_typeSize;
+    retry->retry_dma.dma_cookie   = dma->dma_cookie;
+    retry->retry_dma.dma_vproc    = dma->dma_vproc;
+    retry->retry_dma.dma_srcAddr  = dma->dma_srcAddr;
+    retry->retry_dma.dma_dstAddr  = dma->dma_dstAddr;
+    retry->retry_dma.dma_srcEvent = dma->dma_srcEvent;
+    retry->retry_dma.dma_dstEvent = dma->dma_dstEvent;
+
+    retry->retry_time             = lbolt + ep4_dma_retry_times[interval];
+
+    /* chain onto the end of the approriate retry list */
+    list_add_tail (&retry->retry_link, &rail->r_dma_retrylist[interval]);
+
+    ep_kthread_schedule (&rail->r_retry_thread, retry->retry_time);
+
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+}
+
+void
+ep4_queue_dma_stalled (EP4_RAIL *rail, E4_DMA *dma)
+{
+    EP_NODE_RAIL  *nodeRail = &rail->r_generic.Nodes[EP_VP_TO_NODE(dma->dma_vproc)];
+    EP4_DMA_RETRY *retry;
+    unsigned long  flags;
+    
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+
+    EP4_ASSERT (rail, !list_empty (&rail->r_dma_freelist));
+    
+    /* take an item of the free list */
+    retry = list_entry (rail->r_dma_freelist.next, EP4_DMA_RETRY, retry_link);
+
+    list_del (&retry->retry_link);
+    
+    EPRINTF5 (DBG_RETRY, "%s: ep4_queue_dma_stalled: %016llx %016llx %016llx %016llx\n", rail->r_generic.Name,
+             dma->dma_typeSize, dma->dma_cookie, dma->dma_vproc, dma->dma_srcAddr);
+    EPRINTF4 (DBG_RETRY, "%s:                        %016llx %016llx %016llx\n", rail->r_generic.Name,
+             dma->dma_dstAddr, dma->dma_srcEvent, dma->dma_dstEvent);
+
+    retry->retry_dma.dma_typeSize = dma->dma_typeSize;
+    retry->retry_dma.dma_cookie   = dma->dma_cookie;
+    retry->retry_dma.dma_vproc    = dma->dma_vproc;
+    retry->retry_dma.dma_srcAddr  = dma->dma_srcAddr;
+    retry->retry_dma.dma_dstAddr  = dma->dma_dstAddr;
+    retry->retry_dma.dma_srcEvent = dma->dma_srcEvent;
+    retry->retry_dma.dma_dstEvent = dma->dma_dstEvent;
+
+    /* chain onto the node cancelled dma list */
+    list_add_tail (&retry->retry_link, &nodeRail->StalledDmas);
+
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+}
+
+void
+ep4_free_stalled_dmas (EP4_RAIL *rail, unsigned int nodeId)
+{
+    EP_NODE_RAIL *nodeRail = &rail->r_generic.Nodes[nodeId];
+    struct list_head *el, *nel;
+    unsigned long flags;
+
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+    list_for_each_safe (el, nel, &nodeRail->StalledDmas) {
+       list_del (el);
+       list_add (el, &rail->r_dma_freelist);
+    }
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+}
+
+void
+ep4_display_rail (EP4_RAIL *rail)
+{
+    ELAN4_DEV        *dev = rail->r_ctxt.ctxt_dev;
+    struct list_head *el;
+    register int      i;
+    unsigned long     flags;
+
+    ep_debugf (DBG_DEBUG, "%s: vendorid=%x deviceid=%x\n", rail->r_generic.Name, 
+              rail->r_generic.Devinfo.dev_vendor_id, rail->r_generic.Devinfo.dev_device_id);
+
+    spin_lock_irqsave (&rail->r_ecq_lock, flags);
+    for (i = 0; i < EP4_NUM_ECQ; i++)
+    {
+       list_for_each (el, &rail->r_ecq_list[i]) {
+           EP4_ECQ *ecq = list_entry (el, EP4_ECQ, ecq_link);
+           
+           if (i == EP4_ECQ_EVENT)
+               ep_debugf (DBG_DEBUG, "   ECQ[%d] ecq=%p cqnum=%d addr=%llx avail=%d event=%llx,%llx,%llx\n",
+                          i, ecq, elan4_cq2num (ecq->ecq_cq), ecq->ecq_addr, ecq->ecq_avail,
+                          elan4_sdram_readq (dev, ecq->ecq_event + offsetof (E4_Event32, ev_CountAndType)),
+                          elan4_sdram_readq (dev, ecq->ecq_event + offsetof (E4_Event32, ev_WriteValue)),
+                          elan4_sdram_readq (dev, ecq->ecq_event + offsetof (E4_Event32, ev_WritePtr)));
+
+           else
+               ep_debugf (DBG_DEBUG, "   ECQ[%d] ecq=%p cqnum=%d addr=%llx avail=%d\n",
+                          i, ecq, elan4_cq2num (ecq->ecq_cq), ecq->ecq_addr, ecq->ecq_avail);
+       }
+    }
+    spin_unlock_irqrestore (&rail->r_ecq_lock, flags);
+
+    ep_debugf (DBG_DEBUG, "   flush count=%ld mcq=%p ecq=%p event %llx.%llx.%llx\n", 
+              rail->r_flush_count, rail->r_flush_mcq, rail->r_flush_ecq,
+              elan4_sdram_readq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_flush_event.ev_CountAndType)),
+              elan4_sdram_readq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_flush_event.ev_WritePtr)),
+              elan4_sdram_readq (dev, rail->r_elan + offsetof (EP4_RAIL_ELAN, r_flush_event.ev_WriteValue)));
+    
+    spin_lock_irqsave (&rail->r_dma_lock, flags);
+    for (i = 0; i < EP_NUM_RETRIES; i++)
+    {
+       list_for_each (el, &rail->r_dma_retrylist[i]) {
+           EP4_DMA_RETRY *retry = list_entry (el, EP4_DMA_RETRY, retry_link);
+           
+           ep_debugf (DBG_DEBUG, "    RETRY[%d] typeSize %llx cookie %llx vproc %llx events %llx %llx\n",
+                      i, retry->retry_dma.dma_typeSize, retry->retry_dma.dma_cookie,
+                      retry->retry_dma.dma_vproc, retry->retry_dma.dma_srcEvent, retry->retry_dma.dma_dstEvent);
+       }
+    }
+    spin_unlock_irqrestore (&rail->r_dma_lock, flags);
+}
Index: linux-2.4.21/drivers/net/qsnet/ep/threadcode.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/threadcode.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/threadcode.c      2005-06-01 23:12:54.689425424 -0400
@@ -0,0 +1,146 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: threadcode.c,v 1.11 2003/10/07 13:22:38 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/threadcode.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+EP_ADDR
+ep_symbol (EP_CODE *code, char *name)
+{
+    EP_SYMBOL *s = code->symbols;
+    
+    while (s->name && strcmp (s->name, name))
+       s++;
+    
+    return (s->name ? s->value : (EP_ADDR) 0);
+}
+
+int
+ep_loadcode (EP_RAIL *rail, EP_CODE *code)
+{
+    register int i;
+
+    EP_ADDR  _stext  = ep_symbol (code, "_stext");
+    EP_ADDR  _etext  = ep_symbol (code, "_etext");
+    EP_ADDR  _sdata  = ep_symbol (code, "_sdata");
+    EP_ADDR  _edata  = ep_symbol (code, "_edata");
+    EP_ADDR  _end    = ep_symbol (code, "_end");
+    EP_ADDR  _rodata = roundup (_etext, sizeof (uint64_t));
+
+    if (_stext == (EP_ADDR) 0 || _etext == (EP_ADDR) 0 ||
+       _sdata == (EP_ADDR) 0 || _edata == (EP_ADDR) 0 ||
+       _end == (EP_ADDR) 0)
+    {
+       printk ("ep_loadcode: symbols not defined correctly for code at %p\n", code);
+       return (EINVAL);
+    }
+
+    /*
+     * Include the rodata in the text segment
+     */
+    _etext = _rodata + code->rodata_size;
+
+    /*
+     * If _etext is in the same page as _sdata,  then allocate a contiguous
+     * chunk of memory and map it as read/write. otherwise allocate two chunks
+     * and map the code in as read-only.
+     */
+    if ((_etext & PAGEMASK) == (_sdata & PAGEMASK))
+    {
+       code->ntext  = btopr (_end - (_stext & PAGEMASK));
+       code->pptext = ep_alloc_memory_elan (rail, _stext & PAGEMASK, ptob (code->ntext), EP_PERM_EXECUTE, 0);
+
+       if (code->pptext == (sdramaddr_t) 0)
+           return (ENOMEM);
+       
+       code->_stext  = code->pptext + (_stext & PAGEOFFSET);
+       code->_rodata = code->_stext + (_rodata - _stext);
+       code->_sdata  = code->_stext + (_sdata - _stext);
+    }
+    else
+    {
+       code->ntext  = btopr (_etext - (_stext & PAGEMASK));
+       code->ndata  = btopr (_end - (_sdata & PAGEMASK));
+
+       if (code->ntext)
+       {
+           code->pptext = ep_alloc_memory_elan (rail, _stext & PAGEMASK, ptob (code->ntext), EP_PERM_EXECUTE, 0);
+
+           if (code->pptext == (sdramaddr_t) 0)
+               return (ENOMEM);
+
+           code->_stext  = code->pptext + (_stext & PAGEOFFSET);
+           code->_rodata = code->_stext + (_rodata - _stext);
+       }
+       
+       if (code->ndata)
+       {
+           code->ppdata = ep_alloc_memory_elan (rail, _sdata & PAGEMASK, ptob (code->ndata), EP_PERM_WRITE, 0);
+
+           if (code->ppdata == (sdramaddr_t) 0)
+           {
+               if (code->ntext) ep_free_memory_elan (rail, _sdata & PAGEMASK);
+               code->ntext = 0;
+
+               return (ENOMEM);
+           }
+           
+           code->_sdata = code->ppdata + (_sdata & PAGEOFFSET);
+       }
+    }
+    
+#ifdef __LITTLE_ENDIAN__
+#  define Flip 3
+#else
+#  define Flip  0
+#endif
+
+    /*
+     * Now copy the text and rodata into the SDRAM
+     * this is linked into the module to be byte 
+     * copied to the SDRAM, since we want to copy
+     * with word accesses we have to do the byte
+     * assembly correctly.
+     */
+    for (i = 0; i < code->text_size; i++)
+       rail->Operations.SdramWriteb (rail, code->_stext + i, code->text[i^Flip]);
+
+    for (i = 0; i < code->rodata_size; i++)
+       rail->Operations.SdramWriteb (rail, code->_rodata + i, code->rodata[i^Flip]);
+    
+    /*
+     * And the initialised data segment.
+     */
+    for (i = 0; i < code->data_size; i++)
+       rail->Operations.SdramWriteb (rail, code->_sdata + i, code->data[i^Flip]);
+
+    return (ESUCCESS);
+}
+
+void
+ep_unloadcode (EP_RAIL *rail, EP_CODE *code)
+{
+    EP_ADDR  _stext = ep_symbol (code, "_stext");
+    EP_ADDR  _sdata = ep_symbol (code, "_sdata");
+
+    if (code->pptext)
+       ep_free_memory_elan (rail, _stext & PAGEMASK);
+    if (code->ppdata)
+       ep_free_memory_elan (rail, _sdata & PAGEMASK);
+    code->pptext = code->ppdata = 0;
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/threadcode_elan3.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/threadcode_elan3.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/threadcode_elan3.c        2005-06-01 23:12:54.690425272 -0400
@@ -0,0 +1,85 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: threadcode_elan3.c,v 1.11 2003/10/07 13:22:38 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/threadcode_elan3.c,v $ */
+
+#include <qsnet/kernel.h>
+
+#include <elan/kcomm.h>
+
+#include "kcomm_elan3.h"
+#include "debug.h"
+
+#include <elan3/thread.h>
+
+E3_Addr
+ep3_init_thread (ELAN3_DEV  *dev,
+                E3_Addr     fn,                                /* Elan address of function */
+                E3_Addr     addr,                              /* Elan address of stack */
+                sdramaddr_t stack,                             /* sdram address of stack */
+                int           stackSize,                       /* stack size (in bytes) */
+                int           nargs,
+                ...)
+{
+    sdramaddr_t  frame;
+    sdramaddr_t  regs;
+    sdramaddr_t  argsp;
+    int                 i;
+    va_list      ap;
+
+    /*
+     * Align the stack pointer at the top of the stack and leave space for a stack frame
+     */
+    stack = ((stack + stackSize) & ~(E3_STACK_ALIGN-1)) - sizeof (E3_Frame);
+    addr  = ((addr  + stackSize) & ~(E3_STACK_ALIGN-1)) - sizeof (E3_Frame);
+
+    va_start (ap, nargs);
+
+    if (nargs > 6)
+    {
+       stack -= (((nargs*sizeof (E3_uint32))+E3_STACK_ALIGN-1) & ~(E3_STACK_ALIGN-1));
+       addr  -= (((nargs*sizeof (E3_uint32))+E3_STACK_ALIGN-1) & ~(E3_STACK_ALIGN-1));
+    }
+    
+    frame  = stack;
+    regs   = stack - sizeof (E3_OutsRegs);
+
+    /*
+     * Initialise the registers, and stack frame.
+     */
+    elan3_sdram_writel (dev, regs + offsetof (E3_OutsRegs, o[6]), fn);
+    elan3_sdram_writel (dev, regs + offsetof (E3_OutsRegs, o[7]), 0);
+    
+    if (nargs <= 6)
+    {
+       for (i = 0; i < nargs; i++)
+           elan3_sdram_writel (dev, regs + offsetof (E3_OutsRegs, o[i]), va_arg (ap, E3_uint32));
+    }
+    else
+    {
+       for (i = 0; i < 6; i++)
+           elan3_sdram_writel (dev, regs + offsetof (E3_OutsRegs, o[i]), va_arg (ap, E3_uint32));
+       
+       for (argsp = frame + offsetof (E3_Frame, fr_argx[0]); i < nargs; i++, argsp += sizeof (E3_uint32))
+           elan3_sdram_writel (dev, argsp, va_arg (ap, int));
+    }
+
+    elan3_sdram_writel (dev, frame + offsetof (E3_Frame, fr_savefp), 0);
+    elan3_sdram_writel (dev, frame + offsetof (E3_Frame, fr_savepc), 0);
+
+    va_end (ap);
+
+    return (addr);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/ep/threadcode_elan3_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/threadcode_elan3_Linux.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/threadcode_elan3_Linux.c  2005-06-01 23:12:54.690425272 -0400
@@ -0,0 +1,112 @@
+/* --------------------------------------------------------*/
+/* MACHINE GENERATED ELAN CODE                             */
+#include <qsnet/kernel.h>
+#include <elan/kcomm.h>
+#include "kcomm_elan3.h"
+static uint32_t threadcode_elan3_text[] = {
+0x80a0239c, 0x00001082, 0x00e0a280, 0x47008002, 0x0020a380, 0x20600288, 0x20200286, 0x43008002, 
+0x00000001, 0x0a006081, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 
+0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 
+0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 
+0x00000001, 0x00000001, 0xa800c613, 0xa300c609, 0x0020108a, 0x0080900b, 0x00006885, 0x0580a080, 
+0x06008002, 0x02a0a080, 0x06008022, 0xffff0296, 0x04008010, 0xff3f0398, 0x1f008010, 0x00201090, 
+0x00007081, 0x1600801c, 0x00000001, 0x60a0239c, 0x00a0a3c0, 0x20a0a3f0, 0x40a0a3e0, 0x00c03f3f, 
+0xf8e017be, 0x04e08f80, 0x06008012, 0x00000001, 0x00c01ffc, 0x0000a081, 0x06008010, 0x40a083e0, 
+0x14e007be, 0x00c01ffc, 0x0000a081, 0x40a083e0, 0x20a083f0, 0x00a083c0, 0x60a0039c, 0x00e0a280, 
+0xbfffbf12, 0x0020a380, 0x03008012, 0x02201090, 0x03201090, 0x08e0c381, 0x80a0039c, 0xe0a0239c, 
+0x60a023de, 0x80a0a3e0, 0xa0a0a3f0, 0x080010b8, 0x090010b0, 0x0a0010b2, 0x04000037, 0x402006b4, 
+0x50200690, 0x01201092, 0x20a0239c, 0x00a0a3f0, 0x00c03f3f, 0x8ce117be, 0x04e08f80, 0x06008012, 
+0x00000001, 0x00c01ff8, 0x0000b081, 0x06008010, 0x00a083f0, 0x14e007be, 0x00c01ff8, 0x0000b081, 
+0x00a083f0, 0x20a0039c, 0x582006d0, 0x0020a280, 0x05008002, 0x0900a280, 0x10008002, 0x50200690, 
+0xeaffbf30, 0x5c2006d4, 0x18001090, 0x19001092, 0x1b800294, 0x0a201096, 0x8affff7f, 0x05201098, 
+0x446026d0, 0x302027f4, 0xdfffbf10, 0x50200690, 0xfdffbf10, 0x446026c0, 0x5c2006e0, 0x0020a480, 
+0xf9ffbf06, 0x18001090, 0x19001092, 0x1b000494, 0x14201096, 0x7bffff7f, 0x0a201098, 0x0020a280, 
+0xf4ffbf22, 0x486026e0, 0x00007081, 0x1600801c, 0x00000001, 0x60a0239c, 0x00a0a3c0, 0x20a0a3f0, 
+0x40a0a3e0, 0x00c03f3f, 0x60e217be, 0x04e08f80, 0x06008012, 0x00000001, 0x00c01ffc, 0x0000a081, 
+0x06008010, 0x40a083e0, 0x14e007be, 0x00c01ffc, 0x0000a081, 0x40a083e0, 0x20a083f0, 0x00a083c0, 
+0x60a0039c, 0xff3f84a0, 0xe0ffbf1c, 0x18001090, 0xd5ffbf30, 0x60a003de, 0x80a083e0, 0xa0a083f0, 
+0x08e0c381, 0xe0a0039c, 0x00a1239c, 0x60a023de, 0x80a0a3e0, 0xa0a0a3f0, 0x44a123d0, 0x090010b0, 
+0x0a0010b6, 0x0b0010b8, 0x0c0010b4, 0x012010ba, 0xdca023fa, 0x142007d2, 0x082007d0, 0x084002b2, 
+0x000027c0, 0xf42006d0, 0x0020a280, 0x15008032, 0xf42006d0, 0x18200790, 0xdca003d2, 0x20a0239c, 
+0x00a0a3f0, 0x00c03f3f, 0x20e317be, 0x04e08f80, 0x06008012, 0x00000001, 0x00c01ff8, 0x0000b081, 
+0x06008010, 0x00a083f0, 0x14e007be, 0x00c01ff8, 0x0000b081, 0x00a083f0, 0x20a0039c, 0xf42006d0, 
+0x0020a280, 0x0a008022, 0xdca023c0, 0x042007d0, 0x0840a680, 0x06008032, 0xdca023c0, 0x18001082, 
+0x0220d091, 0xe1ffbf10, 0xf42006d0, 0x06008010, 0x190010a2, 0x042006d0, 0x00c026d0, 0x18001082, 
+0x0020d091, 0x042006d0, 0x01200290, 0x042026d0, 0x000006d0, 0x0020a280, 0x04008002, 0x18001090, 
+0x4f010040, 0x1b001092, 0xf02006e0, 0x0020a480, 0xf1ffbf02, 0x40b03611, 0x004004d2, 0x01201290, 
+0x0840a280, 0x0e018012, 0x10001096, 0x046004d0, 0x01208a80, 0x33008002, 0xa0200484, 0x0c2610ba, 
+0x000024fa, 0x00211090, 0x042024d0, 0x246004d0, 0x80200290, 0x082024d0, 0xec2004d0, 0x00210290, 
+0x0c2024d0, 0x102024c4, 0x186004d2, 0x02602a93, 0x098006d0, 0x0001003b, 0x1d000290, 0x098026d0, 
+0xc0ff3f3b, 0x1d000a90, 0x44a103fa, 0x606007d2, 0x00680292, 0x09001290, 0x4000003b, 0x1d001290, 
+0x142024d0, 0x206004d0, 0x10210290, 0x182024d0, 0x186004d0, 0x02202a91, 0x088006d2, 0x0001003b, 
+0x1d400292, 0x088026d2, 0xc0ff3f3b, 0x1d400a92, 0x186004d0, 0x00280290, 0x80000015, 0x0a001290, 
+0x08401292, 0x4000003b, 0x1d401292, 0x1c2024d2, 0x01201090, 0xa02024d0, 0x20200496, 0xa8200484, 
+0x306004d0, 0x0020a280, 0x2b008012, 0x00201098, 0x0c2610ba, 0x00c022fa, 0x04e022c0, 0xc0200490, 
+0x10e022d0, 0x186004d2, 0x02602a93, 0x098006d0, 0x0001003b, 0x1d000290, 0x098026d0, 0xc0ff3f3b, 
+0x1d000a90, 0x44a103fa, 0x606007d2, 0x00680292, 0x09001290, 0x4000003b, 0x1d001290, 0x14e022d0, 
+0x206004d0, 0x10210290, 0x18e022d0, 0x186004d0, 0x02202a91, 0x088006d2, 0x0001003b, 0x1d400292, 
+0x088026d2, 0xc0ff3f3b, 0x1d400a92, 0x186004d0, 0x00280290, 0x80000015, 0x0a001290, 0x08401292, 
+0x4000003b, 0x1d401292, 0x1ce022d2, 0x4f008010, 0x0020109a, 0x0c00109a, 0x306004d0, 0x0840a380, 
+0x3b00801a, 0xe02004c6, 0x0c2610ba, 0x00c022fa, 0x01202b91, 0x0c000290, 0x02202a91, 0x08400490, 
+0x382002d2, 0x04e022d2, 0x342002d0, 0x08e022d0, 0x0ce022c6, 0x10e022c4, 0x186004d0, 0x02202a91, 
+0x088006d2, 0x0001003b, 0x1d400292, 0x088026d2, 0xc0ff3f3b, 0x1d400a92, 0x44a103fa, 0x606007d0, 
+0x00280290, 0x08401292, 0x4000003b, 0x1d401292, 0x14e022d2, 0x206004d0, 0x10210290, 0x18e022d0, 
+0x186004d0, 0x02202a91, 0x088006d4, 0x0001003b, 0x1d800294, 0x088026d4, 0xc0ff3f3b, 0x1d800a94, 
+0x186004d0, 0x00280290, 0x80000013, 0x09001290, 0x08801294, 0x4000003b, 0x1d801294, 0x1ce022d4, 
+0x01201090, 0x008020d0, 0x04e002d0, 0x08c00086, 0x0840039a, 0x01200398, 0x20e00296, 0x306004d0, 
+0x0800a380, 0xc9ffbf0a, 0x08a00084, 0xc0200490, 0xf0ff22d0, 0xe42004d0, 0x0d00a280, 0x0b00801a, 
+0x00201098, 0x04008010, 0x10001096, 0x01200398, 0x20e00296, 0x306004d0, 0x0800a380, 0xfcffbf2a, 
+0x04e022c0, 0xfc3f109a, 0xe42024da, 0x10001082, 0x186004d0, 0x00280290, 0x08006081, 0x00000001, 
+0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 
+0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 
+0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00201098, 
+0x0c00109a, 0x142004fa, 0xec00823b, 0x3080d61b, 0x00006891, 0x0420a280, 0x3b008002, 0x0c00a280, 
+0x04008002, 0x00000001, 0x0120d091, 0x36008030, 0x7c2006d0, 0x01200290, 0x7c2026d0, 0x782006d0, 
+0x0020a280, 0x04008002, 0x78200690, 0x64000040, 0x40e00692, 0xf02004d0, 0x0020a280, 0x03008012, 
+0xf02026d0, 0x80e026c0, 0x7c2006d0, 0x40e026d0, 0x046004d0, 0x04208a80, 0x13008002, 0x1100108a, 
+0xec2004cc, 0x3fa00b8e, 0x40e0018e, 0x0780239c, 0x0080bbe0, 0x006099e0, 0x00a0b9e0, 0x406099e0, 
+0x40a0b9e0, 0x806099e0, 0x80a0b9e0, 0xc06099e0, 0xc0a0b9e0, 0x00809be0, 0x0780039c, 0x0e008010, 
+0xec2004d2, 0xec2004cc, 0x3fa00b8e, 0x40e0018e, 0x0780239c, 0x0080bbe0, 0x006099e0, 0x00a0b9e0, 
+0x406099e0, 0x40a0b9e0, 0x00809be0, 0x0780039c, 0xec2004d2, 0xe42004d0, 0x886222d0, 0x042006d0, 
+0x00c026d0, 0x000007d0, 0x01208a80, 0x05008012, 0x00000001, 0x142027f2, 0x06008010, 0xdca003fa, 
+0x142027f2, 0xfe3f0a90, 0x000027d0, 0xdca003fa, 0x016007ba, 0xdca023fa, 0x0c2007d0, 0x0840a680, 
+0x04008032, 0x082007d0, 0x03008010, 0x102007f2, 0x084006b2, 0x00007081, 0x1600801c, 0x00000001, 
+0x60a0239c, 0x00a0a3c0, 0x20a0a3f0, 0x40a0a3e0, 0x02c03f3f, 0x8ce017be, 0x04e08f80, 0x06008012, 
+0x00000001, 0x00c01ffc, 0x0000a081, 0x06008010, 0x40a083e0, 0x14e007be, 0x00c01ffc, 0x0000a081, 
+0x40a083e0, 0x20a083f0, 0x00a083c0, 0x60a0039c, 0x042007d0, 0x0840a680, 0xb3febf12, 0x190010a2, 
+0x8afebf10, 0xf42006d0, 0x60a003de, 0x80a083e0, 0xa0a083f0, 0x08e0c381, 0x00a1039c, 0x80a0239c, 
+0x042002c4, 0x004022c4, 0x18008030, 0x00007081, 0x16008012, 0x00000001, 0x60a0239c, 0x00a0a3c0, 
+0x20a0a3f0, 0x40a0a3e0, 0x02c03f3f, 0x24e117be, 0x04e08f80, 0x06008012, 0x00000001, 0x00c01ffc, 
+0x0000a081, 0x06008010, 0x40a083e0, 0x14e007be, 0x00c01ffc, 0x0000a081, 0x40a083e0, 0x20a083f0, 
+0x00a083c0, 0x60a0039c, 0x000002c4, 0x00a0a080, 0xe7ffbf12, 0x00000001, 0x042002c4, 0x01a00084, 
+0x042022c4, 0x000002c4, 0x00a0a080, 0xddffbf12, 0x00000001, 0x08e0c381, 0x80a0039c, };
+#define threadcode_elan3_text_size 0x97c
+static uint32_t threadcode_elan3_data[] = {
+0};
+#define threadcode_elan3_data_size 0x0
+static uint32_t threadcode_elan3_rodata[] = {
+0};
+#define threadcode_elan3_rodata_size 0x0
+static EP_SYMBOL threadcode_elan3_symbols[] = {
+    {"__bss_start", 0xff00297c},
+    {"_edata", 0xff00297c},
+    {"_end", 0xff002988},
+    {"_etext", 0xff00097c},
+    {"_sdata", 0xff00297c},
+    {"_stext", 0xff000000},
+    {"ep3_spinblock", 0xff0008dc},
+    {"ep3comms_rcvr", 0xff0002a8},
+    {"kcomm_probe", 0xff00013c},
+    {"r", 0xff00297c},
+    {"rail", 0xff002984},
+    {"rm", 0xff002980},
+    {0, 0}};
+EP_CODE threadcode_elan3 = {
+   (unsigned char *) threadcode_elan3_text,
+   threadcode_elan3_text_size,
+   (unsigned char *) threadcode_elan3_data,
+   threadcode_elan3_data_size,
+   (unsigned char *) threadcode_elan3_rodata,
+   threadcode_elan3_rodata_size,
+   threadcode_elan3_symbols,
+};
Index: linux-2.4.21/drivers/net/qsnet/ep/threadcode_elan4_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/ep/threadcode_elan4_Linux.c     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/ep/threadcode_elan4_Linux.c  2005-06-01 23:12:54.691425120 -0400
@@ -0,0 +1,112 @@
+/* --------------------------------------------------------*/
+/* MACHINE GENERATED ELAN CODE                             */
+#include <qsnet/kernel.h>
+#include <elan/kcomm.h>
+#include "kcomm_elan4.h"
+static uint32_t threadcode_elan4_text[] = {
+0x00a00087, 0xc04060cb, 0x00003080, 0x80001080, 0x02606180, 0x02004032, 0x807f60cb, 0x04606180, 
+0x02004032, 0x407f60d3, 0x08606180, 0x02004032, 0x007f60db, 0x10606180, 0x02004032, 0xc07e60e3, 
+0x20606180, 0x02004032, 0x807e60eb, 0x40606180, 0x02004032, 0x407e60f3, 0x80606180, 0x02004032, 
+0x007e60fb, 0x40001180, 0xc3801080, 0xc07f60c3, 0x20002000, 0x20002000, 0x20002000, 0x20002000, 
+0x407f8001, 0x4060c0c7, 0x4860c0d0, 0x5060c0d1, 0x5860c0d2, 0x6060c0d3, 0x6860c0d4, 0x00208292, 
+0x00608291, 0x00a08294, 0xff3f8088, 0x1c381293, 0xc04044c8, 0x13004290, 0xc000c5d0, 0x08004030, 
+0x00001088, 0x04204288, 0x0020b200, 0x04004003, 0x00208080, 0x9c010040, 0x00a08488, 0xc04044c8, 
+0x20381288, 0x0020b200, 0xf6ff7f13, 0x01208408, 0x11161282, 0x804094c2, 0xc04044c8, 0x20381288, 
+0x0020b200, 0xebff7f13, 0x00208080, 0x406040c7, 0x486040d0, 0x506040d1, 0x586040d2, 0x606040d3, 
+0x686040d4, 0x08e00180, 0xc0608001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 
+0x807e8001, 0x4060c0c7, 0x4860c0d0, 0x5060c0d1, 0x5860c0d2, 0x6060c0d3, 0x6860c0d4, 0x7060c0d5, 
+0x7860c0d6, 0x8060c0d7, 0x8860c0d8, 0x9060c0d9, 0x9860c0da, 0xa060c0db, 0xa860c0dc, 0xb060c0dd, 
+0xb860c0de, 0xc060c0df, 0x8061c0c8, 0x00608296, 0x00a0829a, 0x9861c0cb, 0xa061c0cc, 0xa861c0cd, 
+0x01208088, 0x3861c0c8, 0x08e042d2, 0x386140c9, 0x0900900a, 0xa06140c8, 0x986140cb, 0x18e042c9, 
+0x72010040, 0x05b4128a, 0x0020808c, 0x3861c0cc, 0x986140c9, 0xc04042c8, 0x0880b400, 0x39014003, 
+0xffff3f08, 0x90a0851c, 0xe023829f, 0x20f4179f, 0x10e3879f, 0xffff3f08, 0xe023829e, 0x20b4179e, 
+0x03a3879e, 0xffff3f08, 0xe023829d, 0x2074179d, 0x0363879d, 0x00a08495, 0x18a08408, 0x800012c2, 
+0x089a109b, 0x20f4169b, 0x20f8169b, 0x00e88609, 0x20741289, 0x01120008, 0x0a381288, 0x08408297, 
+0x45208088, 0x06341288, 0x806140ca, 0xc88042c8, 0x00288218, 0x04a08408, 0x800012c2, 0x089a1088, 
+0x20341288, 0x20381288, 0x00281299, 0x20a08408, 0x800012c2, 0x089a108a, 0x20b4128a, 0x20b8128a, 
+0x30a08408, 0x800012c2, 0x089a1093, 0x20f41493, 0x20f81493, 0x03f41689, 0x806140cb, 0x2922808c, 
+0x0334138c, 0xccc042c8, 0xc90042d1, 0x02604688, 0x0020b200, 0x03004002, 0x60a08214, 0x80a08214, 
+0x90a08509, 0x804012c8, 0x01208208, 0x804092c8, 0x046012c8, 0x043a1288, 0x0020b200, 0x04004003, 
+0xa86140c8, 0x67ffff7f, 0x00a0868a, 0x88a045d0, 0x0020b400, 0x12004013, 0x00208080, 0x800017c8, 
+0x808096c8, 0x72010040, 0x00a08588, 0x00208290, 0x90a08509, 0x804012c8, 0x01208208, 0x804092c8, 
+0x046012c8, 0x043a1288, 0x0020b200, 0x04004003, 0xa86140c8, 0x53ffff7f, 0x00a0868a, 0x804015c2, 
+0x159a1089, 0x20741289, 0x20781289, 0x40b03608, 0x01208288, 0x0840b200, 0x06004023, 0xa02344c4, 
+0x800017c8, 0x808096c8, 0xbb004010, 0xa8a045c8, 0x01604688, 0x00281288, 0x08009008, 0x00e0b400, 
+0x05004003, 0x3f381289, 0x13408209, 0x03004010, 0x05208088, 0x04208088, 0x09009220, 0x07341889, 
+0x0900840b, 0x05341888, 0x0023820a, 0x01604688, 0x0020b200, 0x1d004002, 0x0a00840c, 0xc900c4d7, 
+0x40c40f08, 0x09208288, 0x08e0c2c8, 0x0a608488, 0x10e0c2c8, 0x81001008, 0x0a341288, 0x18e0c2c8, 
+0x1d608488, 0x20e0c2c8, 0x28e0c2d8, 0x24608508, 0x800012c2, 0x089a1088, 0x20341288, 0x20381288, 
+0x80208208, 0x30e0c2c8, 0x00218108, 0x38e0c2c8, 0x40e0c2d4, 0x48e0c2cc, 0xca00c4df, 0x20608411, 
+0x80e0820b, 0x2020830c, 0x00e0b400, 0x13004013, 0x0020808e, 0xc0c0c2d7, 0x40c40f09, 0x09608289, 
+0x08e0c2c9, 0x0a608488, 0x10e0c2c8, 0x00040008, 0x18e0c2c8, 0x1d608488, 0x20e0c2c8, 0x28e0c2d8, 
+0x40e0c2d4, 0x48e0c2cc, 0xc000c3de, 0x00208083, 0x4c004010, 0x20608411, 0xb8238408, 0x800012c2, 
+0x089a108f, 0x20f4138f, 0x20f8138f, 0x00208083, 0x13c0b000, 0x2e00401b, 0x40c40f08, 0x092082a2, 
+0x00040021, 0xffff3f08, 0xe023828d, 0x2074138d, 0x1063838d, 0x0e808309, 0x0e408209, 0x02741289, 
+0x1540820a, 0x38a0820a, 0x808012c2, 0x0a9a108a, 0x20b4128a, 0x20b8128a, 0xc0c0c2d7, 0x08e0c2e2, 
+0x0a608488, 0x10e0c2c8, 0x20b41288, 0x21008288, 0x18e0c2c8, 0x1d608488, 0x20e0c2c8, 0x28e0c2d8, 
+0x15408209, 0x34608209, 0x804012c2, 0x099a1089, 0x20741289, 0x20781289, 0x30e0c2c9, 0x38e0c2cf, 
+0x40e0c2d4, 0x48e0c2cc, 0xc000c3cd, 0x0ac0830f, 0x0ac08003, 0x20608411, 0x80e0820b, 0x01a0830e, 
+0x1380b300, 0xdcff7f0b, 0x2020830c, 0xe03f830c, 0xc000c3dd, 0xbc238408, 0x800012c2, 0x089a1088, 
+0x20341288, 0x20381288, 0x0300b200, 0x0d00401b, 0x07341888, 0x0020888e, 0x0420b800, 0x08004019, 
+0x0800840b, 0x00040008, 0x18e0c2c8, 0x01a0830e, 0x04a0b300, 0xfdff7f09, 0x80e0820b, 0xfc3f8083, 
+0x07341888, 0x08008408, 0xa06140ca, 0xc00062e3, 0x402062f3, 0xc080e2e3, 0xc080e2f3, 0x982244c8, 
+0x88a0c5c8, 0x88a045c8, 0x0020b200, 0x05004013, 0x04604688, 0x88a08508, 0x80a0c5c8, 0x04604688, 
+0x0020b200, 0x0c004002, 0xd822c4c0, 0xc04065e3, 0x406065f3, 0xc000e1e3, 0x806065e3, 0x4020e1f3, 
+0xc06065f3, 0x8020e1e3, 0xc020e1f3, 0x07004010, 0x88228108, 0xc04065e3, 0x406065f3, 0xc000e1e3, 
+0x4020e1f3, 0x88228108, 0x08d61082, 0x800092c2, 0x03f41689, 0x806140cb, 0x2922808c, 0x0334138c, 
+0xccc042c8, 0xc900c2d1, 0x800017c8, 0x808096c8, 0xa8a045c8, 0x0880b400, 0x03004013, 0x00a18412, 
+0xa0a045d2, 0x98a045c8, 0x0020b200, 0x05004013, 0x386140c9, 0x986140c8, 0x0820c2d2, 0x386140c9, 
+0x01608209, 0xfe61b200, 0x0e004015, 0x3861c0c9, 0x00001088, 0x02204288, 0x0020b200, 0x05004003, 
+0x986140ca, 0x28000040, 0xa06140c8, 0x986140ca, 0xc08042c8, 0x0880b400, 0xd8fe7f13, 0x00a08495, 
+0x98a045cb, 0x00e0b200, 0xbafe7f03, 0x386140c9, 0xa06140c8, 0x60a08509, 0x48000040, 0xe03f808a, 
+0x986140cb, 0x08e0c2d2, 0x386140cc, 0x0120830c, 0xaffe7f10, 0x3861c0cc, 0x406040c7, 0x486040d0, 
+0x506040d1, 0x586040d2, 0x606040d3, 0x686040d4, 0x706040d5, 0x786040d6, 0x806040d7, 0x886040d8, 
+0x906040d9, 0x986040da, 0xa06040db, 0xa86040dc, 0xb06040dd, 0xb86040de, 0xc06040df, 0x08e00180, 
+0x80618001, 0x807f8001, 0xc040e0d3, 0x4060e0db, 0x00208490, 0x00208698, 0x00208080, 0x00208080, 
+0x00e08192, 0x02000040, 0x00608091, 0x14e08110, 0x17208097, 0xc000f2d3, 0xc04060d3, 0x406060db, 
+0x08a00080, 0x80608001, 0x407f8001, 0x4060e0d3, 0x8060e0db, 0x00208490, 0x00208698, 0x00208080, 
+0x00208080, 0x00e08192, 0x02000040, 0x00608091, 0x40e08110, 0xc040e0d1, 0x37208097, 0x3860c0d7, 
+0x00208490, 0x00e08597, 0x00208080, 0x00208080, 0x1f608290, 0x20b41291, 0x08638491, 0x00608092, 
+0x00208293, 0xc000f2d1, 0x406060d3, 0x806060db, 0x08a00080, 0xc0608001, 0x407f8001, 0x4060e0d3, 
+0x8060e0db, 0x00208490, 0x00208698, 0x00208080, 0x00208080, 0x00e08192, 0x02000040, 0x00608091, 
+0x54e08110, 0xc040e0d1, 0x37208097, 0x3860c0d7, 0x00208490, 0x00e08597, 0x00208080, 0x00208080, 
+0x1f608290, 0x20b41291, 0x08638491, 0x00608092, 0x00208293, 0x0ef41294, 0x0d208594, 0x17208095, 
+0x17208096, 0x17208097, 0xc000f2d3, 0x406060d3, 0x806060db, 0x08a00080, 0xc0608001, 0x01208097, 
+0xb0e3c0d7, 0x80a060d2, 0x98e28004, 0x98e2c0c0, 0x80a0c0c4, 0xc080c4c3, 0x01e0b400, 0x06004002, 
+0x00a08490, 0x00e08097, 0x02208097, 0xb0e3c0d7, 0xd8e2d0d0, 0xd8e2c0d0, 0x03208097, 0xb0e3c0d7, 
+0x00e08088, 0x0e004010, 0x00a060c3, 0x407f8001, 0x4060e0d3, 0x8060e0db, 0x00208490, 0x00208698, 
+0x00208080, 0x00208080, 0x01208089, 0x8820c2c9, 0x00608091, 0x00e08197, 0x0020f2d3, 0x406060d3, 
+0x806060db, 0x08e00180, 0xc0608001, };
+#define threadcode_elan4_text_size 0x90c
+static uint32_t threadcode_elan4_data[] = {
+0};
+#define threadcode_elan4_data_size 0x0
+static uint32_t threadcode_elan4_rodata[] = {
+0};
+#define threadcode_elan4_rodata_size 0x0
+static EP_SYMBOL threadcode_elan4_symbols[] = {
+    {".thread_restart", 0x00000000f800000c},
+    {".thread_start", 0x00000000f8000000},
+    {"__bss_start", 0x00000000f810090c},
+    {"_edata", 0x00000000f810090c},
+    {"_end", 0x00000000f8100910},
+    {"_etext", 0x00000000f800090c},
+    {"_sdata", 0x00000000f810090c},
+    {"_stext", 0x00000000f8000000},
+    {"c_queue_rxd", 0x00000000f800087c},
+    {"c_reschedule", 0x00000000f8000744},
+    {"c_stall_thread", 0x00000000f80008cc},
+    {"c_waitevent", 0x00000000f8000788},
+    {"c_waitevent_interrupt", 0x00000000f80007f8},
+    {"ep4_spinblock", 0x00000000f8000080},
+    {"ep4comms_rcvr", 0x00000000f8000140},
+    {0, 0}};
+EP_CODE threadcode_elan4 = {
+   (unsigned char *) threadcode_elan4_text,
+   threadcode_elan4_text_size,
+   (unsigned char *) threadcode_elan4_data,
+   threadcode_elan4_data_size,
+   (unsigned char *) threadcode_elan4_rodata,
+   threadcode_elan4_rodata_size,
+   threadcode_elan4_symbols,
+};
Index: linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/jtag/jtagdrv.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv.c       2005-06-01 23:12:54.692424968 -0400
@@ -0,0 +1,451 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: jtagdrv.c,v 1.12 2003/06/07 16:02:35 david Exp $"
+/*      $Source: /cvs/master/quadrics/jtagmod/jtagdrv.c,v $*/
+
+#include <qsnet/types.h>
+
+#include "jtagdrv.h"
+#include <jtag/jtagio.h>
+
+int
+jtagdrv_strobe_data (JTAG_DEV *dev, u_char data)
+{
+    u_char dsr;
+
+    PRINTF (DBG_ECPP, ("jtagdrv_strobe_data: %s %s %s -> ", (data & LPT_DATA_TRST) ? "TRST" : "trst", 
+                      (data & LPT_DATA_TDI) ? "TDI" : "tdi", (data & LPT_DATA_TMS) ? "TMS" : "tms"));
+
+
+    LPT_WRITE_DATA (dev, data); DELAY(5);                      /* Drive NEW values on data wires */
+    LPT_WRITE_CTRL (dev, LPT_CTRL_TCLK); DELAY(5);             /* Drive strobe low */
+    LPT_READ_STAT  (dev, dsr); DELAY(5);                       /* Sample TDI from ring */
+    LPT_WRITE_CTRL (dev, 0); DELAY(5);                         /* Drive strobe high */
+
+    PRINTF (DBG_ECPP, ("%s\n", (dsr & LPT_STAT_PE) ? "TDO" : "tdo"));
+
+    return ((dsr & LPT_STAT_PE) ? 1 : 0);
+}
+
+void
+jtagdrv_select_ring (JTAG_DEV *dev, u_int ring)
+{
+    PRINTF (DBG_ECPP, ("jtagdrv_select_ring: ring=0x%x\n", ring));
+
+    LPT_WRITE_CTRL (dev, 0); DELAY(5);                         /* Drive strobe and TCLK high */
+    LPT_WRITE_DATA (dev, ring);        DELAY(5);                       /* Drive ring address */
+    LPT_WRITE_CTRL (dev, LPT_CTRL_RCLK); DELAY(5);             /* Drive strobe low */
+    LPT_WRITE_CTRL (dev, 0); DELAY(5);                         /* Drive strobe high */
+}
+
+void
+jtagdrv_reset (JTAG_DEV *dev)
+{
+    register int i;
+
+    for (i = 0; i < 5; i++)
+       jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS);                /* 5 clocks to Reset from any state */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST);                                  /* to Run-Test/Idle */
+}
+
+void
+jtagdrv_shift_ir (JTAG_DEV *dev, u_char *value, int nbits)
+{
+    register int i;
+    register int bit;
+
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS);                   /* to Select DR-Scan */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS);                   /* to Select IR-Scan */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST);                                  /* to Capture-IR */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST);                                  /* to Shift-IR */
+    
+    for (i = 0; i < nbits; i++)
+    {
+       /* strobe through the instruction bits,  asserting TMS on the last bit */
+
+       if (i == (nbits-1))
+           bit = jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS | (JTAG_BIT(value, i) ? LPT_DATA_TDI : 0));
+       else
+           bit = jtagdrv_strobe_data (dev, LPT_DATA_TRST | (JTAG_BIT(value, i) ? LPT_DATA_TDI : 0));
+       
+       if (bit)
+           JTAG_SET_BIT(value, i);
+       else
+           JTAG_CLR_BIT(value, i);
+    }
+    
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS);                   /* to Update-IR */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST);                                  /* to Run-Test/Idle */
+}
+
+
+void
+jtagdrv_shift_dr (JTAG_DEV *dev, u_char *value, int nbits)
+{
+    register int i;
+    register int bit;
+
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS);                   /* to Select DR-Scan */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST);                                  /* to Capture-DR */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST);                                  /* to Shift-DR */
+    
+    for (i = 0; i < nbits; i++)
+    {
+       /* strobe through the data bits,  asserting TMS on the last bit */
+
+       if (i == (nbits-1))
+           bit = jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS | (JTAG_BIT(value, i) ? LPT_DATA_TDI : 0));
+       else
+           bit = jtagdrv_strobe_data (dev, LPT_DATA_TRST | (JTAG_BIT(value, i) ? LPT_DATA_TDI : 0));
+       
+       if (bit)
+           JTAG_SET_BIT(value, i);
+       else
+           JTAG_CLR_BIT(value, i);
+    }
+    
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST | LPT_DATA_TMS);                   /* to Update-DR */
+    jtagdrv_strobe_data (dev, LPT_DATA_TRST);                                  /* to Run-Test/Idle */
+}
+
+static int
+jtagdrv_i2c_start (JTAG_DEV *dev)
+{
+    u_char dsr;
+    int i;
+
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_start\n"));
+    
+    /* Issue a stop sequence */
+    LPT_WRITE_CTRL (dev,  LPT_CTRL_SCLK); DELAY(1);            /* SCLK low */
+    LPT_WRITE_DATA (dev, 0); DELAY(5);                         /* SDA low */
+    LPT_WRITE_CTRL (dev, 0); DELAY(5);                         /* SCLK high */
+    LPT_WRITE_DATA (dev, LPT_DATA_SDA); DELAY(5);              /* SDA high */
+    
+    /* sample the line to see if we're idle */
+    LPT_READ_STAT (dev, dsr);                                  /* sample SDA */
+    if ((dsr & LPT_STAT_SDA) == 0)                             /* Cannot start if SDA already driven */
+    {
+       PRINTF (DBG_ECPP, ("jtagdrv_i2c_start: cannot start - sda driven low\n"));
+
+       for (i = 0; i < 16 ; i++)
+       {
+           LPT_WRITE_CTRL (dev, LPT_CTRL_SCLK); DELAY(5);      /* SCLK low */
+           LPT_WRITE_CTRL (dev, 0); DELAY(5);                  /* SCLK high */
+           LPT_READ_STAT  (dev, dsr);
+           
+           if (dsr & LPT_STAT_SDA)
+           {
+               PRINTF (DBG_ECPP, ("jtagdrv_i2c_start - stopped after %d clocks\n", i));
+               break;
+           }
+       }
+
+       if ((dsr & LPT_STAT_SDA) == 0)
+       {
+           PRINTF (DBG_ECPP, ("jtagdrv_i2c_start - cannot start - not idle\n"));
+           return (0);
+       }
+
+       /* seen SDA float high, so issue a stop sequence */
+       LPT_WRITE_CTRL (dev, LPT_CTRL_SCLK); DELAY(1);          /* SCLK low */
+       LPT_WRITE_DATA (dev, 0); DELAY(5);                      /* SDA low */
+       LPT_WRITE_CTRL (dev, 0); DELAY(5);                      /* SCLK high */
+       LPT_WRITE_DATA (dev, LPT_DATA_SDA); DELAY(5);           /* SDA high */
+    }
+
+    LPT_WRITE_DATA (dev, 0); DELAY(4);                         /* drive SDA low */
+    return (1);
+}
+
+static void
+jtagdrv_i2c_stop (JTAG_DEV *dev)
+{
+    u_char dsr;
+    int    i;
+
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_stop\n"));
+
+    LPT_WRITE_CTRL (dev, LPT_CTRL_SCLK); DELAY(1);             /* SCLK low */
+    LPT_WRITE_DATA (dev, 0); DELAY(5);                         /* SDA low */
+    LPT_WRITE_CTRL (dev, 0); DELAY(5);                         /* SCLK high */
+    LPT_WRITE_DATA (dev, LPT_DATA_SDA); DELAY(5);              /* SDA high */
+
+    /* 
+     * bug fix for temperature sensor chip
+     * if it's still driving SDA, then clock
+     * it until it stops driving it 
+     */
+    LPT_READ_STAT (dev, dsr);
+    if ((dsr & LPT_STAT_SDA) == 0)
+    {
+       PRINTF (DBG_ECPP, ("jtagdrv_i2c_stop - slave not stodeved\n"));
+       for (i = 0; i < 16 ; i++)
+       {
+           LPT_WRITE_CTRL (dev, LPT_CTRL_SCLK); DELAY(5);      /* SCLK low */
+           LPT_WRITE_CTRL (dev, 0); DELAY(5);                  /* SCLK high */
+           LPT_READ_STAT  (dev, dsr);
+           
+           if (dsr & LPT_STAT_SDA)
+               break;
+       }
+       PRINTF (DBG_ECPP, ("jtagdrv_i2c_stop - stodeved after %d clocks\n", i));
+    }
+}
+
+static int
+jtagdrv_i2c_strobe (JTAG_DEV *dev, u_char data)
+{
+    u_char dsr;
+    
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_strobe : %s", (data & LPT_DATA_SDA) ? "SDA" : "sda"));
+
+    LPT_WRITE_CTRL (dev, LPT_CTRL_SCLK); DELAY(1);             /* SCLK low */
+    LPT_WRITE_DATA (dev, data);        DELAY(5);                       /* write data */
+    LPT_WRITE_CTRL (dev, 0);                                   /* SCLK high */
+    LPT_READ_STAT  (dev, dsr); DELAY(4);                       /* Sample SDA */
+
+    PRINTF (DBG_ECPP, (" -> %s\n", (dsr & LPT_STAT_SDA) ? "SDA" : "sda"));
+
+    return ((dsr & LPT_STAT_SDA) ? 1 : 0);
+}
+
+static int
+jtagdrv_i2c_get_ack (JTAG_DEV *dev)
+{
+    u_char dsr;
+
+    LPT_WRITE_CTRL (dev, LPT_CTRL_SCLK); DELAY(1);             /* SCLK low */
+    LPT_WRITE_DATA (dev, LPT_DATA_SDA);        DELAY(5);               /* SDA high */
+    LPT_WRITE_CTRL (dev, 0);                                   /* SCLK high */
+    LPT_READ_STAT  (dev, dsr); DELAY(4);                       /* Sample SDA */
+
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_get_ack -> %s\n", (dsr & LPT_STAT_SDA) ? "no ack" : "ack"));
+    
+    return ((dsr & LPT_STAT_SDA) ? 0 : 1);
+}
+
+static int
+jtagdrv_i2c_drive_ack (JTAG_DEV *dev, int nack)
+{
+    u_char dsr;
+
+    LPT_WRITE_CTRL (dev, LPT_CTRL_SCLK); DELAY(1);             /* SCLK low */
+    LPT_WRITE_DATA (dev, nack ? LPT_DATA_SDA : 0); DELAY(5);   /* SDA low for ack, high for nack */
+    LPT_WRITE_CTRL (dev, 0);                                   /* SCLK high */
+    LPT_READ_STAT  (dev, dsr); DELAY(4);                       /* Sample SDA for ack */
+
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_drive_ack %d -> %s\n", nack, (dsr & LPT_STAT_SDA) ? "done" : "more"));
+    
+    return ((dsr & LPT_STAT_SDA) ? 1 : 0);
+}
+
+static void
+jtagdrv_i2c_shift_addr (JTAG_DEV *dev, u_int address, int readNotWrite)
+{
+    register int i;
+
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_shift_addr: %x\n", address));
+
+    for (i = I2C_ADDR_LEN-1; i >= 0; i--)
+       jtagdrv_i2c_strobe (dev, (address & (1 << i)) ? LPT_DATA_SDA : 0);
+    
+    jtagdrv_i2c_strobe (dev, readNotWrite ? LPT_DATA_SDA : 0);
+}
+
+static u_char
+jtagdrv_i2c_shift_data (JTAG_DEV *dev, u_char data)
+{
+    register int i;
+    u_char val = 0;
+    
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_shift_data : %02x\n", data));
+
+    for (i = I2C_DATA_LEN-1; i >= 0; i--)
+       if (jtagdrv_i2c_strobe (dev, data & (1 << i) ? LPT_DATA_SDA : 0))
+           val |= (1 << i);
+
+    PRINTF (DBG_ECPP, ("jtagdrv_i2c_shift_data : -> %02x\n", val));
+
+    return (val);
+}
+
+int
+jtagdrv_i2c_write (JTAG_DEV *dev, u_int address, u_int count, u_char *data)
+{
+    register int i;
+
+    PRINTF (DBG_FN, ("jtagdrv_i2c_write: address=%x count=%d data=%02x\n", address, count, data[0]));
+
+    if (! jtagdrv_i2c_start (dev))
+       return (I2C_OP_NOT_IDLE);
+
+    jtagdrv_i2c_shift_addr (dev, address, 0);
+    
+    if (! jtagdrv_i2c_get_ack (dev))
+    {
+       PRINTF (DBG_FN, ("jtagdrv_i2c_write: no ack on address phase\n"));
+
+       jtagdrv_i2c_stop (dev);
+       return (I2C_OP_NO_DEVICE);
+    }
+    
+    for (i = 0; i < count; i++)
+    {
+       jtagdrv_i2c_shift_data (dev, data[i]);
+       
+       if (! jtagdrv_i2c_get_ack (dev))
+       {
+           PRINTF (DBG_FN, ("jtagdrv_i2c_write: no ack on data phase %d\n", i));
+
+           jtagdrv_i2c_stop (dev);
+           return (I2C_OP_WRITE_TO_BIG);
+       }
+    }
+
+    jtagdrv_i2c_stop (dev);
+    return (I2C_OP_SUCCESS);
+}
+
+int
+jtagdrv_i2c_read (JTAG_DEV *dev, u_int address, u_int count, u_char *data)
+{
+    register int i;
+
+    PRINTF (DBG_FN, ("jtagdrv_i2c_read: address=%x count=%d\n", address, count));
+
+    if (! jtagdrv_i2c_start (dev))
+       return (I2C_OP_NOT_IDLE);
+
+    jtagdrv_i2c_shift_addr (dev, address, 1);
+    
+    if (! jtagdrv_i2c_get_ack (dev))
+    {
+       PRINTF (DBG_FN, ("jtagdrv_i2c_read: no ack on address phase\n"));
+
+       jtagdrv_i2c_stop (dev);
+       return (I2C_OP_NO_DEVICE);
+    }
+    
+    for (i = 0; i < count; i++)
+    {
+       data[i] = jtagdrv_i2c_shift_data (dev, 0xff);
+
+       jtagdrv_i2c_drive_ack (dev, (i == (count-1) ? 1 : 0));
+    }
+
+    jtagdrv_i2c_stop (dev);
+    
+    return (I2C_OP_SUCCESS);
+}
+
+int
+jtagdrv_i2c_writereg (JTAG_DEV *dev, u_int address, u_int intaddress, u_int count, u_char *data)
+{
+    register int i;
+
+    PRINTF (DBG_FN, ("jtagdrv_i2c_writereg: address=%x count=%d\n", address, count));
+
+    if (! jtagdrv_i2c_start (dev))
+       return (I2C_OP_NOT_IDLE);
+
+    jtagdrv_i2c_shift_addr (dev, address, 0);
+    
+    if (! jtagdrv_i2c_get_ack (dev))
+    {
+       PRINTF (DBG_FN, ("jtagdrv_i2c_writereg: no ack on address phase\n"));
+
+       jtagdrv_i2c_stop (dev);
+       return (I2C_OP_NO_DEVICE);
+    }
+    
+    jtagdrv_i2c_shift_data (dev, intaddress);
+    
+    if (! jtagdrv_i2c_get_ack (dev))
+    {
+       PRINTF (DBG_FN, ("jtagdrv_i2c_writereg: no ack on intaddress phase\n"));
+       jtagdrv_i2c_stop (dev);
+       return (I2C_OP_NO_DEVICE);
+    }
+    
+    for (i = 0; i < count; i++)
+    {
+       jtagdrv_i2c_shift_data (dev, data[i]);
+       if (! jtagdrv_i2c_get_ack (dev))
+       {
+           PRINTF (DBG_FN, ("jtagdrv_i2c_writedate: no ack on byte %d\n", i));
+           jtagdrv_i2c_stop (dev);
+           return (I2C_OP_WRITE_TO_BIG);
+       }
+    }
+    
+    jtagdrv_i2c_stop (dev);
+    return (I2C_OP_SUCCESS);
+}
+
+int
+jtagdrv_i2c_readreg (JTAG_DEV *dev, u_int address, u_int intaddress, u_int count, u_char *data)
+{
+    PRINTF (DBG_FN, ("jtagdrv_i2c_readreg: address=%x count=%d\n", address, count));
+
+    if (! jtagdrv_i2c_start (dev))
+       return (I2C_OP_NOT_IDLE);
+
+    jtagdrv_i2c_shift_addr (dev, address, 0);
+    
+    if (! jtagdrv_i2c_get_ack (dev))
+    {
+       PRINTF (DBG_FN, ("jtagdrv_i2c_readreg: no ack on address phase\n"));
+
+       jtagdrv_i2c_stop (dev);
+       return (I2C_OP_NO_DEVICE);
+    }
+    
+    jtagdrv_i2c_shift_data (dev, intaddress);
+    
+    if (! jtagdrv_i2c_get_ack (dev))
+    {
+       PRINTF (DBG_FN, ("jtagdrv_i2c_readreg: no ack on intaddress phase\n"));
+       jtagdrv_i2c_stop (dev);
+       return (I2C_OP_NO_DEVICE);
+    }
+
+    jtagdrv_i2c_stop (dev);
+
+    return (jtagdrv_i2c_read (dev, address, count, data));
+}
+
+void
+jtagdrv_i2c_clock_shift (JTAG_DEV *dev, u_int t, u_int n, u_int m)
+{
+    int i;
+
+    for (i = 2; i >= 0; i--)
+    {
+       LPT_WRITE_DATA (dev, ((t & (1 << i)) ? LPT_DATA_TDI : 0)); DELAY(1);                    /* clock low  | data */
+       LPT_WRITE_DATA (dev, ((t & (1 << i)) ? LPT_DATA_TDI : 0) | LPT_DATA_TMS); DELAY(1);     /* clock high | data */
+    }
+
+    for (i = 1; i >= 0; i--)
+    {
+       LPT_WRITE_DATA (dev, ((n & (1 << i)) ? LPT_DATA_TDI : 0)); DELAY(1);                    /* clock low  | data */
+       LPT_WRITE_DATA (dev, ((n & (1 << i)) ? LPT_DATA_TDI : 0)| LPT_DATA_TMS); DELAY(1);      /* clock high | data */
+    }    
+
+    for (i = 6; i >= 0; i--)
+    {
+       LPT_WRITE_DATA (dev, ((m & (1 << i)) ? LPT_DATA_TDI : 0)); DELAY(1);                    /* clock low  | data */
+       LPT_WRITE_DATA (dev, ((m & (1 << i)) ? LPT_DATA_TDI : 0) | LPT_DATA_TMS); DELAY(1);     /* clock high | data */
+    }    
+
+    LPT_WRITE_DATA (dev, 0); DELAY(1);                                                         /* clock low  | 0 */
+
+    LPT_WRITE_CTRL (dev, LPT_CTRL_TCLK); DELAY(1);                                             /* strobe low */
+    LPT_WRITE_CTRL (dev, 0); DELAY(1);                                                         /* strobe low */
+}
+
Index: linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/jtag/jtagdrv.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv.h       2005-06-01 23:12:54.692424968 -0400
@@ -0,0 +1,57 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __JTAGDRV_COMMON_H
+#define __JTAGDRV_COMMON_H
+
+#ident "@(#)$Id: jtagdrv.h,v 1.5 2002/08/09 11:18:37 addy Exp $"
+/*      $Source: /cvs/master/quadrics/jtagmod/jtagdrv.h,v $*/
+
+#include <qsnet/config.h>
+
+/* include OS specific header file */
+#if defined(LINUX)
+#  include "jtagdrv_Linux.h"
+#elif defined(DIGITAL_UNIX)
+#  include "jtagdrv_OSF1.h"
+#elif defined(QNX)
+#  include "jtagdrv_QNX.h"
+#else
+#  error cannot determint os type
+#endif
+
+extern int jtagdebug;
+
+#define DBG_CFG                (1 << 0)
+#define DBG_OPEN       (1 << 1)
+#define DBG_IOCTL      (1 << 2)
+#define DBG_ECPP       (1 << 3)
+#define DBG_FN         (1 << 4)
+
+#define DRIVER_NAME    "jtag"
+
+#if defined(LINUX)
+#define PRINTF(n,X)    ((n) & jtagdebug ? (void) printk X : (void) 0)
+#define PRINTMSG(fmt, arg...) printk(KERN_INFO DRIVER_NAME ": " fmt, ##arg)
+#else
+#define PRINTF(n,X)    ((n) & jtagdebug ? (void) printf X : (void) 0)
+#define PRINTMSG(M, A) printf ("jtag: " M, A)
+#endif
+
+extern void jtagdrv_select_ring (JTAG_DEV *pp, u_int ring);
+extern void jtagdrv_reset (JTAG_DEV *pp);
+extern void jtagdrv_shift_ir (JTAG_DEV *pp, u_char *value, int nbits);
+extern void jtagdrv_shift_dr (JTAG_DEV *pp, u_char *value, int nbits);
+
+extern int  jtagdrv_i2c_write (JTAG_DEV *pp, u_int address, u_int count, u_char *data);
+extern int  jtagdrv_i2c_read (JTAG_DEV *pp, u_int address, u_int count, u_char *data);
+extern int  jtagdrv_i2c_writereg (JTAG_DEV *pp, u_int address, u_int intaddress, u_int count, u_char *data);
+extern int  jtagdrv_i2c_readreg (JTAG_DEV *pp, u_int address, u_int intaddress, u_int count, u_char *data);
+extern void jtagdrv_i2c_clock_shift (JTAG_DEV *pp, u_int t, u_int n, u_int m);
+
+
+#endif /* __JTAGDRV_COMMON_H */
Index: linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/jtag/jtagdrv_Linux.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv_Linux.c 2005-06-01 23:12:54.693424816 -0400
@@ -0,0 +1,319 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+/*
+ * $Id: jtagdrv_Linux.c,v 1.18 2004/01/06 11:15:46 fabien Exp $
+ * $Source: /cvs/master/quadrics/jtagmod/jtagdrv_Linux.c,v $
+ */
+
+#include "jtagdrv.h"
+#include <jtag/jtagio.h>
+
+#include <linux/module.h>
+#include <linux/ioport.h>
+
+MODULE_AUTHOR("Quadrics Ltd.");
+MODULE_DESCRIPTION("JTAG Parallel port QsNet switch interface");
+
+MODULE_LICENSE("GPL");
+
+#define MAJOR_INSTANCE 0       /* 0 is dynamic assign of device major  */ 
+#define MAX_JTAG_DEV   4
+
+int jtag_major = MAJOR_INSTANCE;
+int jtagdebug  = 0;
+MODULE_PARM(jtag_major, "i");
+MODULE_PARM(jtagdebug, "i");
+
+JTAG_DEV       jtag_devs[MAX_JTAG_DEV];
+
+int io[MAX_JTAG_DEV]= { 0, };
+MODULE_PARM(io, "1-4i");
+
+
+/* The fops functions */
+int jtag_open(struct inode *, struct file *);
+int jtag_close(struct inode *, struct file *);
+int jtag_ioctl(struct inode *, struct file *, unsigned int, unsigned long );
+
+struct file_operations jtag_fops = {
+    ioctl:   jtag_ioctl,
+    open:    jtag_open,
+    release: jtag_close,
+};
+
+int
+jtag_probe(void)
+{
+       int i=0;        
+       int default_io = 1;
+       JTAG_DEV *dev;
+       unsigned char value=0xff;
+
+       /* see if there are any user supplied io addr */
+       for ( i = 0; i < MAX_JTAG_DEV; i++) {
+               if ( io[i] != 0x00)
+                       default_io = 0;
+               jtag_devs[i].base = io[i];
+       }
+       
+       if ( default_io ) {
+               jtag_devs[0].base = 0x3bc;
+               jtag_devs[1].base = 0x378;
+               jtag_devs[2].base = 0x278;
+               jtag_devs[3].base = 0x268;
+       }
+
+       for ( i = 0 ; i < MAX_JTAG_DEV; i++) {
+               if ( jtag_devs[i].base == 0x3bc ) 
+                       jtag_devs[i].region = 3;
+               else
+                       jtag_devs[i].region = 8;
+               jtag_devs[i].present = 0;
+       }       
+
+
+       if( default_io )
+       {
+               for( i = 0 ; i < MAX_JTAG_DEV; i++) {
+                       dev=&(jtag_devs[i]);
+                       if(dev->base && request_region(dev->base, dev->region, "jtag")) {
+                               LPT_WRITE(dev, 0,0);
+                               LPT_READ(dev, 0,value);
+                               if ( value != 0xff) {
+                                       PRINTMSG("(%d , %d) present, io=0x%04lx\n",jtag_major,i,dev->base);
+                       
+                                       dev->present=1; 
+                               }
+                               else
+                                       release_region(dev->base, dev->region);
+                       }
+               }
+               return 0;
+       }     
+       else /* Force the region to be present, this makes the PCI parallel cards work */
+       {
+               for( i = 0 ; i < MAX_JTAG_DEV; i++) 
+               {
+                        dev=&(jtag_devs[i]);
+                        if(dev->base && request_region(dev->base, dev->region, "jtag") && (dev->base != 0)) 
+                       {
+                                PRINTMSG("(%d , %d) forced by user, io=0x%04lx\n",jtag_major,i,dev->base);
+                                        dev->present=1;
+                       }       
+                        else   
+                       {
+                                if( dev->base != 0)
+                                       release_region(dev->base, dev->region);
+                       }
+               }
+                return 0;
+       }
+}
+
+int init_module(void)
+{
+       int result,i;
+       result = register_chrdev(jtag_major, DRIVER_NAME, &jtag_fops);
+       if (result < 0) {
+               PRINTMSG("Couldn't register char device err == %d\n",jtag_major);
+               return -1;
+       }
+
+       if ( jtag_major == 0 ) 
+               jtag_major = result;
+
+       for ( i = 0; i < MAX_JTAG_DEV; i++) {
+               jtag_devs[i].base=io[i];        
+       }
+
+       jtag_probe();
+
+       PRINTMSG("Registered character device, major == %d\n",jtag_major);      
+       return 0;
+}      
+
+void cleanup_module(void)
+{
+       int i=0;
+
+       for( i = 0; i < MAX_JTAG_DEV; i++) {
+               if( jtag_devs[i].present)       
+                       release_region(jtag_devs[i].base, jtag_devs[i].region);
+       }
+                       
+       unregister_chrdev(jtag_major, DRIVER_NAME);
+       PRINTMSG("Unloaded char device\n");
+}
+
+
+int
+jtag_open (struct inode *inode, struct file *filp)
+{
+    int unit = MINOR(inode->i_rdev);
+    JTAG_DEV *dev = &jtag_devs[unit];
+
+    if (unit < 0 || unit > MAX_JTAG_DEV || !dev->present)
+       return (-ENXIO);
+
+    /*
+     * Only allow a single open at a time 
+     */
+    if (dev->open)
+       return (-EBUSY);
+    dev->open = 1;
+    
+    /*
+     * Initialise the hardware registers
+     */
+   
+    LPT_WRITE (dev, LPT_CTRL, 0);
+    DELAY(50);
+    LPT_WRITE (dev, LPT_CTRL, LPT_CTRL_INIT);
+
+    MOD_INC_USE_COUNT;
+
+    return (0);
+}
+
+int
+jtag_close(struct inode *inode, struct file *filp)
+{
+  
+    int unit = MINOR(inode->i_rdev);
+    JTAG_DEV *dev = &jtag_devs[unit];
+    
+    if (unit < 0 || unit > MAX_JTAG_DEV || !dev->present)
+       return (-ENXIO);
+    
+    dev->open = 0;
+
+    MOD_DEC_USE_COUNT;
+
+    return (0);
+}
+
+int
+jtag_ioctl (struct inode *inode, struct file *filp, unsigned int io_cmd, unsigned long io_data)
+{
+    int                  unit = MINOR(inode->i_rdev);
+    JTAG_DEV             *dev = &jtag_devs[unit];
+    JTAG_RESET_ARGS      *resetargs;
+    JTAG_SHIFT_ARGS      *shiftargs;
+    I2C_ARGS            *i2cargs;
+    I2C_CLOCK_SHIFT_ARGS *clockargs;
+    u_char              *buf;
+    int                          freq;
+
+    if (unit < 0 || unit > MAX_JTAG_DEV || !dev->present)
+       return (-ENXIO);
+    
+    PRINTF (DBG_IOCTL, ("jtag_ioctl: device %d cmd=%x\n", unit, io_cmd));
+
+    switch (io_cmd)
+    {
+    case JTAG_RESET:
+       resetargs = (JTAG_RESET_ARGS *) io_data;
+
+       if (! VALID_JTAG_RING (resetargs->ring))
+           return (-EINVAL);
+       
+       jtagdrv_select_ring (dev, resetargs->ring);
+       jtagdrv_reset (dev);
+       return (0);
+       
+    case JTAG_SHIFT_IR:
+    case JTAG_SHIFT_DR:
+       shiftargs = (JTAG_SHIFT_ARGS *) io_data;
+       
+       if (! VALID_JTAG_RING (shiftargs->ring) || shiftargs->nbits > (JTAG_MAX_DATA_LEN*JTAG_MAX_CHIPS)) {
+           return (-EFAULT);
+               }
+
+       buf = (u_char *) kmalloc (JTAG_NBYTES(shiftargs->nbits), GFP_KERNEL);
+
+       if (buf == (u_char *) NULL)
+           return (-ENOMEM);
+       
+       if (copy_from_user (buf, shiftargs->value, JTAG_NBYTES(shiftargs->nbits)))
+       {
+           kfree(buf);
+           return (-EFAULT);
+       }
+
+
+       jtagdrv_select_ring (dev, shiftargs->ring);
+
+       if (io_cmd == JTAG_SHIFT_IR)
+           jtagdrv_shift_ir (dev, buf, shiftargs->nbits);
+       else
+           jtagdrv_shift_dr (dev, buf, shiftargs->nbits);
+       
+       if (copy_to_user (shiftargs->value, buf, JTAG_NBYTES (shiftargs->nbits)))
+       {
+           kfree (buf);
+           return (-EFAULT);
+       }
+
+       kfree (buf);
+       return (0);
+
+    case I2C_WRITE:
+    case I2C_READ:
+    case I2C_WRITEREG:
+    case I2C_READREG:
+       i2cargs = (I2C_ARGS *) io_data;
+
+       if (! VALID_I2C_RING(i2cargs->ring) || i2cargs->count > I2C_MAX_DATA_LEN)
+           return (-EFAULT);
+
+       jtagdrv_select_ring (dev, RING_I2C_BIT | i2cargs->ring);
+       switch (io_cmd)
+       {
+       case I2C_WRITE:
+           i2cargs->ok = jtagdrv_i2c_write (dev, i2cargs->device, i2cargs->count, i2cargs->data);
+           break;
+
+       case I2C_READ:
+           i2cargs->ok = jtagdrv_i2c_read (dev, i2cargs->device, i2cargs->count, i2cargs->data);
+           break;
+
+       case I2C_WRITEREG:
+           i2cargs->ok = jtagdrv_i2c_writereg (dev, i2cargs->device, i2cargs->reg, i2cargs->count, i2cargs->data);
+           break;
+
+       case I2C_READREG:
+           i2cargs->ok = jtagdrv_i2c_readreg (dev, i2cargs->device, i2cargs->reg, i2cargs->count, i2cargs->data);
+           break;
+       }
+       return (0);
+
+    case I2C_CLOCK_SHIFT:
+       clockargs = (I2C_CLOCK_SHIFT_ARGS *) io_data;
+
+       freq = (10 * clockargs->m / (1 << (((clockargs->n + 1) & 3))));
+       
+       /* validate the value, and initialise the ring */
+       if (clockargs->t != 0 || clockargs->n > 3 || clockargs->m > 127)
+           return (-EINVAL);
+       
+       jtagdrv_select_ring (dev, RING_I2C_BIT | RING_CLOCK_SHIFT);
+       jtagdrv_i2c_clock_shift (dev, clockargs->t, clockargs->n, clockargs->m);
+       jtagdrv_select_ring (dev, 0);
+       return (0);
+
+    default:
+       return (-EINVAL);
+    }
+    return (-EINVAL);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv_Linux.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/jtag/jtagdrv_Linux.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/jtag/jtagdrv_Linux.h 2005-06-01 23:12:54.693424816 -0400
@@ -0,0 +1,174 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: jtagdrv_Linux.h,v 1.3 2002/08/09 11:18:37 addy Exp $"
+/*      $Source: /cvs/master/quadrics/jtagmod/jtagdrv_Linux.h,v $*/
+
+#ifndef __JTAGDRV_LINUX_H
+#define __JTAGDRV_LINUX_H
+
+#include <qsnet/kernel.h>
+#include <asm/io.h>
+
+typedef struct jtag_dev
+{
+    unsigned long      base;
+    int                 region;
+
+    u_int              present:1;
+    u_int              open:1;
+} JTAG_DEV;
+
+/*
+**
+**                     Hardware Defines
+**
+*/
+
+/*
+ * Assume that bit 4 of the Control Register is set to 1 (by default) 
+ * to enable the printer port (CS3).
+ *
+ * The default base address is 3BC-3BF. 
+ */
+
+#define LPT0   0x3BC                   /* CSR Base Address - note this can
+                                        * change depending on the setting
+                                        * in the Control Register 0.
+                                        *
+                                        * LPT1 0x378
+                                        * LPT2 0x278
+                                        * LPT3 0x268
+                                       */
+
+/*
+ *     Register offsets from the port base address
+ */
+
+#define LPT_REGISTER_0 0
+#define LPT_REGISTER_1 1
+#define LPT_REGISTER_2 2
+#define LPT_REGISTER_3 0x400
+#define LPT_REGISTER_4 0x401
+#define LPT_REGISTER_5 0x402
+
+/*
+ *     Chip control registers
+ */
+                                       /* Base address for Super I/O National*/
+
+#define SIO_BASE_ADDR  0x26e           /* Semiconductor PC87332VLJ combo-chip*/
+#define CR4_REG                0x04            /* index 4, printer control reg 4 */
+
+#define LPT_EPP                0x01            /* Enable bit for epp */
+#define LPT_ECP                0x04            /* Enable bit for ecp */
+
+/*
+ *     Registers for use with centronics, nibble and byte modes.
+ */
+
+#define LPT_DATA       LPT_REGISTER_0          /* line printer port data */
+#define LPT_STAT       LPT_REGISTER_1          /* LPT port status        */
+#define LPT_CTRL       LPT_REGISTER_2          /* LPT port control       */
+
+/*
+ *     Registers for use with ECP mode.
+ */ 
+
+#define LPT_DFIFO      LPT_REGISTER_3          /* r/w fifo register    */
+#define LPT_CFGB       LPT_REGISTER_4          /* Configuration B      */
+#define LPT_ECR                LPT_REGISTER_5          /* Exteded control      */
+
+/*
+ * Bit assignments for ECR register.
+ */
+
+       /* Bits 0-4 */
+
+#define LPT_ECR_EMPTY  0x01            /* FIFO is empty */
+#define LPT_ECR_FULL   0x02            /* FIFO is full */
+#define LPT_ECR_SERV   0x04            /* Service bit */
+#define LPT_ECR_DMA    0x08            /* DMA enable */
+#define LPT_ECR_nINTR  0x10            /* Interrupt disable */
+
+       /*
+        * Bits 5-7 are ECR modes.
+        */
+
+#define LPT_ECR_PAR    0x20            /* Parallel port FIFO mode */
+#define LPT_ECR_ECP    0x60            /* ECP mode */
+#define LPT_ECR_CFG    0xE0            /* Configuration mode */
+#define LPT_ECR_CLEAR  ~0xE0           /* Cear mode bits */
+
+/*
+ * Bit assignments for the parallel port STATUS register:
+ */
+
+#define LPT_STAT_BIT0  0X1     /* Reserved. Bit always set.            */
+#define LPT_STAT_BIT1  0X2     /* Reserved. Bit always set.            */
+#define LPT_STAT_IRQ   0x4     /* interrupt status bit                 */
+#define LPT_STAT_ERROR 0x8     /* set to 0 to indicate error           */
+#define LPT_STAT_SLCT  0x10    /* status of SLCT lead from printer     */
+#define LPT_STAT_PE    0x20    /* set to 1 when out of paper           */
+#define LPT_STAT_ACK   0x40    /* acknowledge - set to 0 when ready    */
+#define LPT_STAT_nBUSY 0x80    /* busy status bit, 0=busy, 1=ready     */
+
+/*
+ * Bit assignments for the parallel port CONTROL register:
+ */
+
+#define LPT_CTRL_nSTROBE       0x1     /* Printer Strobe Control       */
+#define LPT_CTRL_nAUTOFD       0x2     /* Auto Feed Control            */
+#define LPT_CTRL_INIT          0x4     /* Initialize Printer Control   */
+#define LPT_CTRL_nSLCTIN       0x8     /* 0=select printer, 1=not selected */
+#define LPT_CTRL_IRQ           0x10    /* Interrupt Request Enable Control */
+#define LPT_CTRL_DIR           0x20    /* Direction control            */
+#define LPT_CTRL_BIT6          0X40    /* Reserved. Bit always set.    */
+#define LPT_CTRL_BIT7          0X80    /* Reserved. Bit always set.    */
+
+
+#define LPT_WRITE(dev, regname, value) do { outb(value, (dev)->base + regname); } while (0)
+#define LPT_READ(dev, regname,value)   do { value = inb((dev)->base + regname); } while (0)
+
+
+
+/* Standard register access macros */
+#define LPT_WRITE_CTRL(dev, value)     LPT_WRITE(dev, LPT_CTRL, LPT_CTRL_INIT | value)
+#define LPT_WRITE_DATA(dev, value)     LPT_WRITE(dev, LPT_DATA, value)
+#define LPT_READ_STAT(dev, value)      LPT_READ(dev, LPT_STAT, value)
+
+/*
+ * The jtag signals are connected to the parallel port as follows :
+ *
+ *  TRST       bit 0
+ *  TDI                bit 1
+ *  TMS                bit 2
+ *  TCLK       AFX
+ *  TDO                PE
+ */
+#define LPT_DATA_TRST  1
+#define LPT_DATA_TDI   2
+#define LPT_DATA_TMS   4
+#define LPT_CTRL_TCLK  LPT_CTRL_nAUTOFD
+#define LPT_STAT_TDO   LPT_STAT_PE
+
+/*
+ * The I2C signals are connected as follows :
+ */
+#define LPT_DATA_SDA   2
+#define LPT_CTRL_SCLK  LPT_CTRL_nAUTOFD
+#define LPT_STAT_SDA   LPT_STAT_PE
+
+/*
+ * The ring selection signals are as follows :
+ *  addr       bit 0-7
+ *  clock      nSLCTIN
+ */
+#define LPT_CTRL_RCLK  LPT_CTRL_nSLCTIN
+
+
+#endif /* __JTAGDRV_LINUX_H */
Index: linux-2.4.21/drivers/net/qsnet/jtag/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/jtag/Makefile   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/jtag/Makefile        2005-06-01 23:12:54.694424664 -0400
@@ -0,0 +1,31 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/jtag/Makefile
+#
+
+
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/jtag/Makefile
+#
+
+list-multi             := jtag.o
+jtag-objs      := jtagdrv_Linux.o jtagdrv.o
+export-objs            := 
+obj-$(CONFIG_JTAG)     := jtag.o
+
+jtag.o : $(jtag-objs)
+       $(LD) -r -o $@ $(jtag-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/jtag/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/jtag/Makefile.conf      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/jtag/Makefile.conf   2005-06-01 23:12:54.694424664 -0400
@@ -0,0 +1,10 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       jtag.o
+MODULENAME     =       jtag
+KOBJFILES      =       jtagdrv_Linux.o jtagdrv.o
+EXPORT_KOBJS   =       
+CONFIG_NAME    =       CONFIG_JTAG
+SGALFC         =       
+# EXTRALINES START
+
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/jtag/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/jtag/quadrics_version.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/jtag/quadrics_version.h      2005-06-01 23:12:54.694424664 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/Makefile        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/Makefile     2005-06-01 23:12:54.695424512 -0400
@@ -0,0 +1,17 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2003 Quadrics Ltd.
+#
+# File: driver/net/qsnet/Makefile
+#
+
+subdir-$(CONFIG_QSNET)  += qsnet elan
+subdir-$(CONFIG_ELAN3)  += elan3
+subdir-$(CONFIG_ELAN4)  += elan4
+subdir-$(CONFIG_EP)     += ep
+subdir-$(CONFIG_EIP)    += eip
+subdir-$(CONFIG_RMS)    += rms
+subdir-$(CONFIG_JTAG)   += jtag
+
+include $(TOPDIR)/Rules.make
Index: linux-2.4.21/drivers/net/qsnet/qsnet/debug.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/qsnet/debug.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/qsnet/debug.c        2005-06-01 23:12:54.696424360 -0400
@@ -0,0 +1,583 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: debug.c,v 1.21 2004/08/19 08:09:57 david Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/debug.c,v $ */
+
+#include <qsnet/kernel.h>
+#include <qsnet/debug.h>
+#include <qsnet/procfs_linux.h>
+
+caddr_t        qsnet_debug_buffer_ptr = NULL;
+int           qsnet_debug_front      = 0;
+int           qsnet_debug_back       = 0;
+int            qsnet_debug_lost_lines = 0;
+int           qsnet_debug_disabled   = 0;
+
+int           qsnet_debug_line_size  = 256;
+int           qsnet_debug_num_lines  = 8192;
+
+int           qsnet_assfail_mode     = 1;                      /* default to BUG() */
+
+int            qsnet_debug_running    = 0;
+int            kqsnet_debug_running   = 0;
+
+static spinlock_t qsnet_debug_lock;
+static kcondvar_t qsnet_debug_wait;
+static char       qsnet_debug_buffer_space[8192];
+
+#define QSNET_DEBUG_PREFIX_MAX_SIZE    32
+#define QSNET_DEBUG_MAX_WORDWRAP       15
+
+/* must be larger than  QSNET_DEBUG_PREFIX_MAX_SIZE +  QSNET_DEBUG_MAX_WORDWRAP + 2 */
+#if defined(DIGITAL_UNIX) 
+#define QSNET_DEBUG_CONSOLE_WIDTH 80
+#elif defined(LINUX)
+#define QSNET_DEBUG_CONSOLE_WIDTH 128
+#endif
+
+#define isspace(CH)    ((CH==' ') | (CH=='\t') | (CH=='\n'))
+
+#ifdef LINUX
+#define ALLOC_DEBUG_BUFFER(ptr)                do { (ptr) = (void *)__get_free_pages (GFP_KERNEL, get_order (qsnet_debug_num_lines * qsnet_debug_line_size)); } while (0)
+#define FREE_DEBUG_BUFFER(ptr)         free_pages ((unsigned long) ptr, get_order (qsnet_debug_num_lines * qsnet_debug_line_size))
+#else
+#define ALLOC_DEBUG_BUFFER(ptr)                KMEM_ALLOC (ptr, caddr_t, qsnet_debug_num_lines * qsnet_debug_line_size, 1)
+#define FREE_DEBUG_BUFFER(ptr)         KMEM_FREE (ptr, qsnet_debug_num_lines * qsnet_debug_line_size)
+#endif
+
+void
+qsnet_debug_init ()
+{
+       spin_lock_init (&qsnet_debug_lock);
+       kcondvar_init  (&qsnet_debug_wait);
+
+       qsnet_debug_front      = 0;
+       qsnet_debug_back       = 0;
+       qsnet_debug_lost_lines = 0;
+
+       if (qsnet_debug_line_size < (QSNET_DEBUG_PREFIX_MAX_SIZE + QSNET_DEBUG_MAX_WORDWRAP + 2))
+               qsnet_debug_line_size = 256;
+
+       qsnet_debug_running    = 1;
+
+       qsnet_proc_register_int (qsnet_procfs_config, "assfail_mode", &qsnet_assfail_mode, 0);
+}
+
+void
+qsnet_debug_fini()
+{
+       if (!qsnet_debug_running) return;
+
+       remove_proc_entry ("assfail_mode", qsnet_procfs_config);
+
+       spin_lock_destroy (&qsnet_debug_lock);
+       kcondvar_destroy  (&qsnet_debug_wait);
+       
+       if (qsnet_debug_buffer_ptr)
+               FREE_DEBUG_BUFFER (qsnet_debug_buffer_ptr);
+
+       qsnet_debug_buffer_ptr     = NULL;
+       qsnet_debug_lost_lines = 0;     
+       qsnet_debug_running    = 0;     
+}
+
+void
+qsnet_debug_disable(int val)
+{
+       qsnet_debug_disabled = val;
+}
+
+void
+qsnet_debug_alloc()
+{
+       caddr_t ptr;
+       unsigned long flags;
+
+       if (!qsnet_debug_running) return;
+
+       if (qsnet_debug_buffer_ptr == NULL)
+       {
+               ALLOC_DEBUG_BUFFER (ptr);
+
+               if (ptr != NULL)
+               {
+                       spin_lock_irqsave (&qsnet_debug_lock, flags);
+                       if (qsnet_debug_buffer_ptr == NULL)
+                       {
+                               qsnet_debug_buffer_ptr = ptr;
+                               spin_unlock_irqrestore (&qsnet_debug_lock, flags);
+                       }
+                       else
+                       {
+                               spin_unlock_irqrestore (&qsnet_debug_lock, flags);
+
+                               FREE_DEBUG_BUFFER (ptr);
+                       }
+               }
+       }
+       
+}
+
+static void 
+qsnet_prefix_debug(unsigned int mode, char *prefix, char *buffer) 
+{
+       /* assumes caller has lock */
+
+       int  prefixlen = strlen(prefix);
+       char pref[QSNET_DEBUG_PREFIX_MAX_SIZE];
+       int  prefix_done = 0;
+
+       if (!qsnet_debug_running) return;
+
+       if (qsnet_debug_disabled)
+               return;
+
+       if (prefixlen >= QSNET_DEBUG_PREFIX_MAX_SIZE) 
+       {
+               strncpy(pref,prefix,QSNET_DEBUG_PREFIX_MAX_SIZE -2);
+               strcpy (&pref[QSNET_DEBUG_PREFIX_MAX_SIZE-5],"... ");
+
+               prefix = pref;
+                prefixlen = strlen(prefix);
+       }
+
+#ifdef CONFIG_MPSAS
+       {
+               char *p;
+#define TRAP_PUTCHAR_B                 (0x17a - 256)
+#define SAS_PUTCHAR(c)                 do {\
+                       register int o0 asm ("o0") = (c);\
+\
+                       asm volatile ("ta %0; nop" \
+                                     : /* no outputs */\
+                                     : /* inputs */ "i" (TRAP_PUTCHAR_B), "r" (o0)\
+                                     : /* clobbered */ "o0");\
+\
+                       if (o0 == '\n') {\
+                               o0 = '\r';\
+\
+                               asm volatile ("ta %0; nop" \
+                                             : /* no outputs */\
+                                             : /* inputs */ "i" (TRAP_PUTCHAR_B), "r" (o0)\
+                                             : /* clobbered */ "o0");\
+                       }\
+               } while(0)
+
+               for (p = prefix; *p; p++)
+                       SAS_PUTCHAR (*p);
+
+               for (p = buffer; *p; p++)
+                       SAS_PUTCHAR (*p);
+       }
+#else
+       if (mode & QSNET_DEBUG_BUFFER)
+       {
+               if (qsnet_debug_buffer_ptr == NULL)
+                       qsnet_debug_lost_lines++;
+               else
+               {                   
+                       caddr_t base = &qsnet_debug_buffer_ptr[qsnet_debug_line_size * qsnet_debug_back];
+                       caddr_t lim  = base + qsnet_debug_line_size - 2;
+                       caddr_t p;
+               
+                       p = buffer; 
+                       prefix_done = 0;
+                       while (*p) 
+                       {
+                               /* sort out prefix */
+                               if ( prefix_done++ ) 
+                               {
+                                       int i;
+                                       for(i=0;i<prefixlen;i++)
+                                               base[i] = ' ';
+                                       /* memset(base,' ',prefixlen); */
+                               }
+                               else
+                                       strcpy(base,prefix);
+                               base += prefixlen; /* move the base on */
+
+                               /* copy data */
+                               for ( ; *p && (base < lim); )
+                                       *base++ = *p++;
+
+                               /* if line split then add \n */
+                               if ((base == lim) && (*base != '\n'))
+                               {
+                                       char *ptr;
+                                       int   count;
+
+                                       *base = '\n';
+                                       /* we added a \n cos it was end of line put next char was \n */
+                                       if (*p == '\n') 
+                                               p++;
+                                       else
+                                       {
+                                               /* lets see if we can back track and find a white space to break on */
+                                               ptr = base-1;
+                                               count = 1;
+                                               while ( ( !isspace(*ptr) ) && ( count < QSNET_DEBUG_MAX_WORDWRAP ))
+                                               {
+                                                       count++;
+                                                       ptr--;
+                                               }
+
+                                               if ( isspace(*ptr) ) 
+                                               {
+                                                       /* found somewhere to wrap to */
+                                                       p -= (count-1); /* need to loose the white space */
+                                                       base = ptr;
+                                                       *base = '\n';
+                                               }
+                                       }
+                                       base++;
+                               }
+                               *base = '\0';
+
+                               /* move on pointers */
+                               qsnet_debug_back = (++qsnet_debug_back == qsnet_debug_num_lines) ? 0 : qsnet_debug_back;            
+                               if (qsnet_debug_back == qsnet_debug_front)
+                               {
+                                       qsnet_debug_lost_lines++;
+                                       qsnet_debug_front = (++qsnet_debug_front == qsnet_debug_num_lines) ? 0 : qsnet_debug_front;
+                               }
+                               base  = &qsnet_debug_buffer_ptr[qsnet_debug_line_size * qsnet_debug_back];
+                               lim  =  base + qsnet_debug_line_size - 2;
+                       }
+                       kcondvar_wakeupone (&qsnet_debug_wait, &qsnet_debug_lock);
+               }
+       }
+
+       if (mode & QSNET_DEBUG_CONSOLE)
+       {
+               int     remaining = QSNET_DEBUG_CONSOLE_WIDTH - prefixlen;
+               caddr_t p;
+               char    line[QSNET_DEBUG_CONSOLE_WIDTH +2];
+               int     len;
+           
+               strcpy (pref,prefix);
+               prefix_done = 0;
+
+               p = buffer;
+               while ( *p )
+               {
+                       /* use the prefix only once */
+                       if  ( prefix_done++ > 0 ) 
+                               {
+                                       int i;
+                                       for(i=0;i<prefixlen;i++)
+                                               pref[i] = ' ';
+                                       /* memset(perf,' ',prefixlen); */
+                               }       
+
+                       len=strlen(p);
+                       if (len > remaining) len = remaining;
+                 
+                       strncpy(line, p, len);
+                       line[len] = 0;
+                       p += len;
+                   
+                       /* word wrap */
+                       if ((len == remaining) && *p && !isspace(*p))
+                       {
+                               /* lets see if we can back track and find a white space to break on */
+                               char * ptr = &line[len-1];
+                               int    count = 1;
+
+                               while ( ( !isspace(*ptr) ) && ( count < QSNET_DEBUG_MAX_WORDWRAP ))
+                               {
+                                       count++;
+                                       ptr--;
+                               }
+
+                               if ( isspace(*ptr) ) 
+                               {
+                                       /* found somewhere to wrap to */
+                                       p -= (count-1); /* need to loose the white space */
+                                       len -= count;
+                               }               
+                       }
+
+                       if (line[len-1] != '\n' ) 
+                       {
+                               line[len] = '\n';
+                               line[len+1] = 0;
+                       }
+
+                       /* we put a \n in so dont need another one next */
+                       if ( *p == '\n')
+                               p++;
+
+#if defined(DIGITAL_UNIX)
+                       {
+                               char *pr;
+
+                               for (pr = pref; *pr; pr++)
+                                       cnputc (*pr);
+
+                               for (pr = line; *pr; pr++)
+                                       cnputc (*pr); 
+                       }
+#elif defined(LINUX)
+                       printk("%s%s",pref,line);
+#endif
+               }
+       }
+#endif /* CONFIG_MPSAS */
+}
+
+void
+qsnet_vdebugf (unsigned int mode, char *prefix, char *fmt, va_list ap)
+{
+       unsigned long flags;
+
+       if (!qsnet_debug_running) return;
+
+       spin_lock_irqsave (&qsnet_debug_lock, flags);
+
+       qsnet_debug_buffer_space[0] = '\0';
+
+#if defined(DIGITAL_UNIX)
+       prf (qsnet_debug_buffer_space+strlen(qsnet_debug_buffer_space), NULL, fmt, ap);
+#elif defined(LINUX)
+       vsprintf (qsnet_debug_buffer_space+strlen(qsnet_debug_buffer_space), fmt, ap);
+#endif
+
+       if (prefix == NULL)
+               printk ("qsnet_vdebugf: prefix==NULL\n");
+       else
+               qsnet_prefix_debug(mode, prefix, qsnet_debug_buffer_space);
+
+       spin_unlock_irqrestore (&qsnet_debug_lock, flags);
+}
+
+void kqsnet_debugf(char *fmt,...)
+{
+       if ( kqsnet_debug_running ) {
+               va_list ap;
+               char string[20];
+               
+               sprintf (string, "mm=%p:", current->mm);
+               va_start(ap, fmt);
+               qsnet_vdebugf(QSNET_DEBUG_BUFFER, string, fmt, ap);
+               va_end(ap);
+       }       
+}
+void 
+qsnet_debugf(unsigned int mode, char *fmt,...)
+{
+       va_list       ap;
+       unsigned long flags;
+
+       if (!qsnet_debug_running) return;
+
+       spin_lock_irqsave (&qsnet_debug_lock, flags);
+
+       qsnet_debug_buffer_space[0] = '\0';
+
+       va_start (ap, fmt);
+#if defined(DIGITAL_UNIX)
+       prf (qsnet_debug_buffer_space+strlen(qsnet_debug_buffer_space), NULL, fmt, ap);
+#elif defined(LINUX)
+       vsprintf (qsnet_debug_buffer_space+strlen(qsnet_debug_buffer_space), fmt, ap);
+#endif
+       va_end (ap);
+
+       qsnet_prefix_debug(mode, "", qsnet_debug_buffer_space); 
+
+       spin_unlock_irqrestore (&qsnet_debug_lock, flags);
+}
+
+int
+qsnet_debug_buffer (caddr_t ubuffer, int len)
+{
+       caddr_t buffer, ptr, base;
+       int     remain, len1;
+       unsigned long flags;
+       static  char qsnet_space[65536];
+
+       if (!qsnet_debug_running) return (0);
+
+       if (len < qsnet_debug_line_size)
+               return (-1);
+
+       if (len > (qsnet_debug_line_size * qsnet_debug_num_lines))
+               len = qsnet_debug_line_size * qsnet_debug_num_lines;
+    
+       if ( len > 65536 ) {
+               KMEM_ZALLOC (buffer, caddr_t, len, 1);
+       } else 
+               buffer = qsnet_space;
+
+       if (buffer == NULL)
+               return (-1);
+
+       if (qsnet_debug_buffer_ptr == NULL)
+               qsnet_debug_alloc();
+
+       if (qsnet_debug_buffer_ptr == NULL)
+       {
+               if ( len > 65536 )
+                       KMEM_FREE (buffer, len);
+               return (-1);
+       }
+
+       spin_lock_irqsave (&qsnet_debug_lock, flags);
+    
+       while (!qsnet_debug_lost_lines && (qsnet_debug_back == qsnet_debug_front))
+               if (kcondvar_waitsig (&qsnet_debug_wait, &qsnet_debug_lock, &flags) == 0)
+                       break;
+    
+       ptr    = buffer;
+       remain = len;
+
+       if (qsnet_debug_lost_lines)
+       {
+               qsnet_debug_lost_lines = 0;
+               strcpy (ptr, "Debug Buffer has overflowed!!\n");
+               len1 = strlen (ptr);
+
+               remain -= len1;
+               ptr    += len1;
+       }
+
+       while (qsnet_debug_front != qsnet_debug_back)
+       {
+               /* copy the line from DebugFront */
+               base = &qsnet_debug_buffer_ptr[qsnet_debug_front*qsnet_debug_line_size];
+
+               len1 = strlen (base);
+
+               if (len1 > remain)
+                       break;
+       
+               bcopy (base, ptr, len1);
+       
+               ptr += len1;
+               remain -= len1;
+
+               qsnet_debug_front = (++qsnet_debug_front == qsnet_debug_num_lines) ? 0 : qsnet_debug_front;
+       }
+
+       spin_unlock_irqrestore (&qsnet_debug_lock, flags);
+
+       len1 = ptr - buffer;
+
+       if (len1 != 0 && copyout (buffer, ubuffer, len1))
+               len1 = -1;
+
+       if ( len > 65536 )
+               KMEM_FREE (buffer, len);
+   
+       return (len1);
+}
+
+void
+qsnet_debug_buffer_on() 
+{
+       if (qsnet_debug_buffer_ptr == NULL)
+               qsnet_debug_alloc();
+}
+
+void 
+qsnet_debug_buffer_clear()
+{
+       unsigned long flags;
+
+       qsnet_debug_buffer_on();
+       
+       if (qsnet_debug_buffer_ptr != NULL){
+               spin_lock_irqsave (&qsnet_debug_lock, flags);
+               qsnet_debug_front      = 0;
+               qsnet_debug_back       = 0;
+               qsnet_prefix_debug(QSNET_DEBUG_BUFFER,"Clear","");
+               spin_unlock_irqrestore (&qsnet_debug_lock, flags);      
+       }
+}
+
+void 
+qsnet_debug_buffer_mark(char *str)
+{
+       unsigned long flags;    
+
+       qsnet_debug_buffer_on();
+
+       if (qsnet_debug_buffer_ptr != NULL) {
+               spin_lock_irqsave (&qsnet_debug_lock, flags);
+               qsnet_prefix_debug(QSNET_DEBUG_BUFFER,"Mark",str);
+               spin_unlock_irqrestore (&qsnet_debug_lock, flags);
+       }
+}
+int
+qsnet_debug_dump ()
+{
+       unsigned long flags;
+
+       if (!qsnet_debug_running) return (0);
+
+       if (qsnet_debug_buffer_ptr == NULL)
+               qsnet_debug_alloc();
+
+       if (qsnet_debug_buffer_ptr == NULL)
+               return (-1);
+
+       spin_lock_irqsave (&qsnet_debug_lock, flags);
+
+       while (qsnet_debug_front != qsnet_debug_back)
+       {
+               printk ("%s", &qsnet_debug_buffer_ptr[qsnet_debug_front*qsnet_debug_line_size]);
+
+               qsnet_debug_front = (++qsnet_debug_front == qsnet_debug_num_lines) ? 0 : qsnet_debug_front;
+       }
+
+       if (qsnet_debug_lost_lines)
+               printk ("\n**** Debug buffer has lost %d lines\n****\n",qsnet_debug_lost_lines);
+
+       spin_unlock_irqrestore (&qsnet_debug_lock, flags);
+
+       return (0);
+}
+
+int
+qsnet_debug_kmem (void *handle)
+{
+       if (!qsnet_debug_running) return (0);
+
+#ifdef KMEM_DEBUG
+       qsnet_kmem_display(handle);
+#endif
+       return (0);
+}
+
+int
+qsnet_assfail (char *ex, const char *func, char *file, int line)
+{
+       qsnet_debugf (QSNET_DEBUG_BUFFER, "qsnet: assertion failure: %s, function: %s, file %s, line: %d\n", ex, func, file, line);
+
+       printk (KERN_EMERG "qsnet: assertion failure: %s, function: %s, file %s, line: %d\n", ex, func, file, line);
+
+       if (panicstr)
+               return (0);
+
+       if (qsnet_assfail_mode & 1)                             /* return to BUG() */
+               return 1;
+
+       if (qsnet_assfail_mode & 2)
+               panic ("qsnet: assertion failure: %s, function: %s, file %s, line: %d\n", ex, func, file, line);
+       if (qsnet_assfail_mode & 4)
+               qsnet_debug_disable (1);
+
+       return 0;
+
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/qsnet/i686_mmx.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/qsnet/i686_mmx.c        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/qsnet/i686_mmx.c     2005-06-01 23:12:54.696424360 -0400
@@ -0,0 +1,99 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: i686_mmx.c,v 1.11 2004/01/05 12:08:25 mike Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/i686_mmx.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#if defined(LINUX_I386)
+
+#include <linux/config.h>
+#include <linux/sched.h>
+#include <asm/processor.h>
+#include <asm/i387.h>
+
+int mmx_disabled = 0;
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 0)
+/* These functions are lifted from arch/i386/kernel/i387.c
+ * and MUST be kept in step with the kernel (currently 2.4.17)
+ * alternatively we should export the kernel_fpu_begin() function
+ */
+static inline void __save_init_fpu( struct task_struct *tsk )
+{
+       if ( cpu_has_fxsr ) {
+               asm volatile( "fxsave %0 ; fnclex"
+                             : "=m" (tsk->thread.i387.fxsave) );
+       } else {
+               asm volatile( "fnsave %0 ; fwait"
+                             : "=m" (tsk->thread.i387.fsave) );
+       }
+       tsk->flags &= ~PF_USEDFPU;
+}
+#if defined(MODULE)
+void kernel_fpu_begin(void)
+{
+       struct task_struct *tsk = current;
+
+       if (tsk->flags & PF_USEDFPU) {
+               __save_init_fpu(tsk);
+               return;
+       }
+       clts();
+}
+#endif
+#endif
+
+extern inline int
+mmx_preamble(void)
+{
+    if (mmx_disabled || in_interrupt())
+       return (0);
+
+    kernel_fpu_begin();
+
+    return (1);
+}
+
+extern inline void
+mmx_postamble(void)
+{
+    kernel_fpu_end();
+}
+
+extern u64
+qsnet_readq (volatile u64 *ptr)
+{
+    u64 value;
+
+    if (! mmx_preamble())
+       value = *ptr;
+    else
+    {
+       asm volatile ("movq (%0), %%mm0\n"
+                     "movq %%mm0, (%1)\n"
+                     : : "r" (ptr), "r" (&value) : "memory");
+       mmx_postamble();
+    }
+    return (value);
+}
+
+void
+qsnet_writeq(u64 value, volatile u64 *ptr)
+{
+    if (! mmx_preamble())
+       *ptr = value;
+    else
+    {
+       asm volatile ("movq (%0), %%mm0\n"
+                     "movq %%mm0, (%1)\n"
+                     : : "r" (&value), "r" (ptr) : "memory");
+       mmx_postamble();
+    }
+}
+#endif
Index: linux-2.4.21/drivers/net/qsnet/qsnet/kernel_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/qsnet/kernel_linux.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/qsnet/kernel_linux.c 2005-06-01 23:12:54.697424208 -0400
@@ -0,0 +1,856 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: kernel_linux.c,v 1.71.2.3 2004/11/04 11:03:47 david Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/kernel_linux.c,v $*/
+
+#include <qsnet/kernel.h>
+#include <qsnet/ctrl_linux.h>
+#include <qsnet/kpte.h>
+
+#include <linux/sysctl.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/vmalloc.h>
+
+#include <qsnet/procfs_linux.h>
+
+#include <linux/smp.h>         /* for smp_call_function() prototype */
+#include <linux/smp_lock.h>
+#include <linux/mm.h>
+
+#include <linux/highmem.h>
+
+extern int mmx_disabled;
+extern int qsnet_debug_line_size;
+extern int qsnet_debug_num_lines;
+
+gid_t                 qsnet_procfs_gid;
+struct proc_dir_entry *qsnet_procfs_root;
+struct proc_dir_entry *qsnet_procfs_config;
+
+MODULE_AUTHOR("Quadrics Ltd.");
+MODULE_DESCRIPTION("QsNet Kernel support code");
+
+MODULE_LICENSE("GPL");
+
+#if defined(LINUX_I386)
+MODULE_PARM(mmx_disabled, "i");
+#endif
+
+MODULE_PARM(qsnet_debug_line_size, "i");
+MODULE_PARM(qsnet_debug_num_lines, "i");
+
+MODULE_PARM(qsnet_procfs_gid, "i");
+
+#ifdef KMEM_DEBUG
+EXPORT_SYMBOL(qsnet_kmem_alloc_debug);
+EXPORT_SYMBOL(qsnet_kmem_free_debug);
+#else
+EXPORT_SYMBOL(qsnet_kmem_alloc);
+EXPORT_SYMBOL(qsnet_kmem_free);
+#endif
+
+EXPORT_SYMBOL(qsnet_kmem_display);
+EXPORT_SYMBOL(kmem_to_phys);
+
+EXPORT_SYMBOL(cpu_hold_all);
+EXPORT_SYMBOL(cpu_release_all);
+
+#if defined(LINUX_I386)
+EXPORT_SYMBOL(qsnet_readq);
+EXPORT_SYMBOL(qsnet_writeq);
+#endif
+
+/* debug.c */
+EXPORT_SYMBOL(qsnet_debugf);
+EXPORT_SYMBOL(kqsnet_debugf);
+EXPORT_SYMBOL(qsnet_vdebugf);
+EXPORT_SYMBOL(qsnet_debug_buffer);
+EXPORT_SYMBOL(qsnet_debug_alloc);
+EXPORT_SYMBOL(qsnet_debug_dump);
+EXPORT_SYMBOL(qsnet_debug_kmem);
+EXPORT_SYMBOL(qsnet_debug_disable);
+
+EXPORT_SYMBOL(qsnet_assfail);
+
+EXPORT_SYMBOL(qsnet_procfs_gid);
+EXPORT_SYMBOL(qsnet_procfs_root);
+
+static int qsnet_open    (struct inode *ino, struct file *fp);
+static int qsnet_release (struct inode *ino, struct file *fp);
+static int qsnet_ioctl   (struct inode *ino, struct file *fp, unsigned int cmd, unsigned long arg);
+
+static struct file_operations qsnet_ioctl_fops = 
+{
+       ioctl:   qsnet_ioctl,
+       open:    qsnet_open,
+       release: qsnet_release,
+};
+
+static int
+qsnet_open (struct inode *inode, struct file *fp)
+{
+       MOD_INC_USE_COUNT;
+       fp->private_data = NULL;
+       return (0);
+}
+
+static int
+qsnet_release (struct inode *inode, struct file *fp)
+{
+       MOD_DEC_USE_COUNT;
+       return (0);
+}
+
+static int 
+qsnet_ioctl(struct inode *inode, struct file *fp, unsigned int cmd, unsigned long arg)
+{
+       int res=0;
+
+       switch (cmd) 
+       {
+       case QSNETIO_DEBUG_KMEM:
+       {
+               QSNETIO_DEBUG_KMEM_STRUCT args;
+
+               if (copy_from_user (&args, (void *) arg, sizeof (QSNETIO_DEBUG_KMEM_STRUCT)))
+                       return (-EFAULT);
+
+               /* doesnt use handle as a pointer */
+               qsnet_kmem_display(args.handle);
+               break;
+       }
+
+       case QSNETIO_DEBUG_DUMP : 
+       {
+               res = qsnet_debug_dump();
+               break;
+       }
+
+       case QSNETIO_DEBUG_BUFFER :
+       {
+               QSNETIO_DEBUG_BUFFER_STRUCT args;
+
+               if (copy_from_user (&args, (void *) arg, sizeof (QSNETIO_DEBUG_BUFFER_STRUCT)))
+                       return (-EFAULT);
+
+               /* qsnet_debug_buffer uses copyout */
+               if ((res = qsnet_debug_buffer (args.addr, args.len)) != -1)
+               {
+                       args.len = res;
+                       if (copy_to_user ((void *) arg, &args, sizeof (QSNETIO_DEBUG_BUFFER_STRUCT)))
+                               return (-EFAULT);
+                       res = 0;
+               }
+               break;
+       }
+       default:
+               res = EINVAL;
+               break;
+       }
+
+       return ((res == 0) ? 0 : -res);
+}
+
+#ifdef KMEM_DEBUG
+static int qsnet_kmem_open    (struct inode *ino, struct file *fp);
+static int qsnet_kmem_release (struct inode *ino, struct file *fp);
+static ssize_t qsnet_kmem_read (struct file *file, char *buf, size_t count, loff_t *ppos);
+
+static struct file_operations qsnet_kmem_fops = 
+{
+       open:    qsnet_kmem_open,
+       release: qsnet_kmem_release,
+       read:    qsnet_kmem_read,
+};
+
+typedef struct qsnet_private_space
+{
+       char * space;
+       int    size;
+       struct qsnet_private_space *next;
+} QSNET_PRIVATE_SPACE;
+
+typedef struct qsnet_private  
+{
+       QSNET_PRIVATE_SPACE *space_chain;
+        QSNET_PRIVATE_SPACE *current_space;
+       int                  current_pos;
+
+} QSNET_PRIVATE;
+
+#define QSNET_KMEM_DEBUG_LINE_SIZE ((int)512)
+#define QSNET_PRIVATE_PAGE_SIZE    ((int)(4*1024))
+
+static int qsnet_kmem_fill(QSNET_PRIVATE *pd);
+
+void
+destroy_chain(QSNET_PRIVATE * pd)
+{
+       QSNET_PRIVATE_SPACE *mem, *next;
+       
+       if (pd == NULL) return;
+
+       for(mem = pd->space_chain ; mem != NULL; )
+       {
+               next = mem->next; 
+               if ( mem->space ) 
+                       kfree ( mem->space);
+               kfree(mem);
+               mem = next;
+       }
+       kfree (pd);
+}
+
+QSNET_PRIVATE *
+make_chain(int len)
+{
+       QSNET_PRIVATE       * pd;
+       QSNET_PRIVATE_SPACE * mem;
+       int                   i;
+
+       /* make the private data block */
+       if ((pd = kmalloc (sizeof (QSNET_PRIVATE), GFP_KERNEL)) == NULL)
+               return NULL;
+       pd->space_chain = NULL;
+
+       /* first make the holders */
+       for(i=0;i<len;i++)
+       {
+               if ((mem = kmalloc (sizeof (QSNET_PRIVATE_SPACE), GFP_KERNEL)) == NULL)
+               {
+                       destroy_chain(pd);
+                       return (NULL);
+               }
+               mem->next  = pd->space_chain;
+               mem->size  = 0;
+               mem->space = 0;
+               pd->space_chain = mem;
+
+               /* now add the space */
+               if ((mem->space = kmalloc (QSNET_PRIVATE_PAGE_SIZE, GFP_KERNEL)) == NULL)
+               {
+                       destroy_chain(pd);
+                       return (NULL);
+               }                       
+
+               mem->space[0] = 0;
+
+       }
+
+       pd->current_space = pd->space_chain;
+       pd->current_pos   = 0;
+
+       return pd;
+}
+
+static int
+qsnet_kmem_open (struct inode *inode, struct file *fp)
+{
+       MOD_INC_USE_COUNT;
+       fp->private_data = NULL;
+       return (0);
+}
+
+static int
+qsnet_kmem_release (struct inode *inode, struct file *fp)
+{
+       if ( fp->private_data )
+       {
+               QSNET_PRIVATE * pd = (QSNET_PRIVATE *) fp->private_data;
+
+               /* free the space */
+               if (pd->space_chain)
+                       kfree (pd->space_chain);        
+
+               /* free struct */
+               kfree (pd);
+       }
+       MOD_DEC_USE_COUNT;
+       return (0);
+}
+
+static ssize_t
+qsnet_kmem_read (struct file *file, char *buf, size_t count, loff_t *ppos)
+{
+       QSNET_PRIVATE * pd = (QSNET_PRIVATE *) file->private_data;
+       int             error;
+       int             output_count;
+       int             num_of_links=10;
+
+       /* make a buffer to output count bytes in */
+       if ((error = verify_area (VERIFY_WRITE, buf, count)) != 0)
+               return (error);
+
+       if ( pd == NULL) 
+       {
+               /* first time */
+
+               /* ok we have to guess at how much space we are going to need  */
+               /* if it fails we up the space and carry try again             */
+               /* we have to do it this way as we cant get more memory whilst */
+               /* holding the lock                                            */
+               if ((pd = make_chain(num_of_links)) == NULL)
+                       return (-ENOMEM);       
+
+               while ( qsnet_kmem_fill(pd) ) 
+               {
+                       destroy_chain(pd);
+                       num_of_links += 10;
+                       if ((pd = make_chain(num_of_links)) == NULL)
+                               return (-ENOMEM);       
+               }
+
+               /* we have the space and filled it */
+               file->private_data = (void *)pd;        
+       }
+               
+       /* output buffer */
+       if ( pd->current_pos >= pd->current_space->size )
+               return (0); /* finished */
+
+       output_count = pd->current_space->size - pd->current_pos;
+       if ( output_count > count ) 
+               output_count = count;
+
+       copy_to_user(buf, (pd->current_space->space + pd->current_pos), output_count);
+
+       pd->current_pos += output_count;
+       ppos            += output_count;
+
+       /* just check to see if we have finished the current space */
+       if ( pd->current_pos >= pd->current_space->size )
+       {
+               if ( pd->current_space->next )
+               {
+                       pd->current_space = pd->current_space->next;
+                       pd->current_pos   = 0;
+               }
+       }       
+
+       return (output_count);
+}
+#endif /* KMEM_DEBUG */
+
+static int
+proc_write_qsnetdebug(struct file *file, const char *buffer,
+                     unsigned long count, void *data)
+{
+       char    tmpbuf[128];
+       int     res;
+       
+       if (count > sizeof (tmpbuf)-1)
+               return (-EINVAL);
+       
+       MOD_INC_USE_COUNT;
+       
+       if (copy_from_user (tmpbuf, buffer, count))
+               res = -EFAULT;
+       else 
+       {
+               tmpbuf[count] = '\0';   
+               
+               if (tmpbuf[count-1] == '\n')
+                       tmpbuf[count-1] = '\0';
+               
+               if (! strcmp (tmpbuf, "on"))
+                       qsnet_debug_buffer_on();
+
+               if (! strcmp (tmpbuf, "clear"))
+                       qsnet_debug_buffer_clear();
+
+               if (! strncmp (tmpbuf, "mark",4))
+                       qsnet_debug_buffer_mark( &tmpbuf[4] );
+               
+               res = count;
+       }
+       
+       MOD_DEC_USE_COUNT;
+       
+       return (res);
+}
+
+static int
+proc_read_qsnetdebug(char *page, char **start, off_t off,
+                    int count, int *eof, void *data)
+{
+       int len = sprintf (page, "echo command > /proc/qsnet/config/qsnetdebug\ncommand = on | off | clear | mark text\n");
+       return (qsnet_proc_calc_metrics (page, start, off, count, eof, len));
+}
+
+#include "quadrics_version.h"
+extern int kqsnet_debug_running;
+static char       quadrics_version[] = QUADRICS_VERSION;
+
+static int __init qsnet_init(void)
+{
+       struct proc_dir_entry *p;
+
+       if ((qsnet_procfs_root = proc_mkdir ("qsnet", 0)) == NULL)
+       {
+               printk ("qsnet: failed to create /proc/qsnet \n");
+               return (-ENXIO);
+       }
+       
+       if ((p = create_proc_entry ("ioctl", S_IRUGO|S_IWUSR|S_IWGRP, qsnet_procfs_root)) == NULL)
+       {
+               printk ("qsnet: failed to register /proc/qsnet/ioctl\n");
+               return (-ENXIO);
+       }
+       p->proc_fops = &qsnet_ioctl_fops;
+       p->owner     = THIS_MODULE;
+       p->data      = NULL;
+       p->gid       = qsnet_procfs_gid;
+
+       qsnet_proc_register_str (qsnet_procfs_root, "version", quadrics_version, S_IRUGO);
+
+       if ((qsnet_procfs_config = proc_mkdir ("config", qsnet_procfs_root)) == NULL)
+       {
+               printk ("qsnet: failed to create /proc/qsnet/config \n");
+               return (-ENXIO);
+       }
+
+#ifdef KMEM_DEBUG
+       if ((p = create_proc_entry ("kmem_debug", S_IRUGO|S_IWUSR|S_IWGRP, qsnet_procfs_config)) == NULL)
+       {
+               printk ("qsnet: failed to register /proc/qsnet/config/kmem_debug\n");
+               return (-ENXIO);
+       }
+       p->proc_fops = &qsnet_kmem_fops;
+       p->owner     = THIS_MODULE;
+       p->data      = NULL;
+       p->gid       = qsnet_procfs_gid;
+#endif         
+      
+       qsnet_debug_init(); 
+
+       qsnet_proc_register_int (qsnet_procfs_config, "kqsnet_debug_running", &kqsnet_debug_running, 0);
+
+       if ((p = create_proc_entry ("qsnetdebug", S_IRUGO|S_IWUSR|S_IWGRP, qsnet_procfs_config)) == NULL)
+       {
+               printk ("qsnet: failed to register /proc/qsnet/config/qsnetdebug\n");
+               return (-ENXIO);
+       }
+       p->read_proc  = proc_read_qsnetdebug;
+       p->write_proc = proc_write_qsnetdebug;
+       p->owner      = THIS_MODULE;
+       p->data       = NULL;
+       p->gid        = qsnet_procfs_gid;
+       
+       return (0);
+}
+
+static void __exit qsnet_exit(void)
+{
+#ifdef KMEM_DEBUG
+       qsnet_kmem_display(0);
+#endif
+       qsnet_debug_fini();
+
+       remove_proc_entry ("qsnetdebug",           qsnet_procfs_config);
+       remove_proc_entry ("kqsnet_debug_running", qsnet_procfs_config);
+#ifdef KMEM_DEBUG
+       remove_proc_entry ("kmem_debug",           qsnet_procfs_config);
+#endif
+       remove_proc_entry ("config",               qsnet_procfs_root);
+
+       remove_proc_entry ("version", qsnet_procfs_root);
+       remove_proc_entry ("ioctl",   qsnet_procfs_root);
+
+       remove_proc_entry ("qsnet", 0);
+}
+
+/* Declare the module init and exit functions */
+module_init(qsnet_init);
+module_exit(qsnet_exit);
+
+#ifdef KMEM_DEBUG
+/*
+ * Kernel memory allocation.  We maintain our own list of allocated mem
+ * segments so we can free them on module cleanup.
+ * 
+ * We use kmalloc for allocations less than one page in size; vmalloc for
+ * larger sizes.
+ */
+
+typedef struct {
+       struct list_head list;
+       void            *ptr;
+       int             len;
+       int             used_vmalloc;
+       void            *owner;
+       void            *caller;
+       unsigned int     time;
+       int              line;
+       char             filename[20];
+} kmalloc_t;
+
+static LIST_HEAD(kmalloc_head);
+
+static spinlock_t      kmalloc_lock = SPIN_LOCK_UNLOCKED;
+
+/*
+ * Kernel memory allocation.  We use kmalloc for allocations less 
+ * than one page in size; vmalloc for larger sizes.
+ */
+
+static int
+qsnet_kmem_fill(QSNET_PRIVATE *pd)
+{
+       kmalloc_t *kp;
+       struct list_head *lp;
+       unsigned long flags;
+       char str[QSNET_KMEM_DEBUG_LINE_SIZE];
+       QSNET_PRIVATE_SPACE * current_space;
+       int                   current_pos;
+       int                   len;
+       current_space = pd->space_chain;
+       current_pos   = 0;
+       
+       
+       current_space->space[0] = 0;    
+       spin_lock_irqsave(&kmalloc_lock, flags);
+       for (lp = kmalloc_head.next; lp != &kmalloc_head;  lp = lp->next) {
+               kp = list_entry(lp, kmalloc_t, list);
+               
+               /* make the next line */
+               sprintf(str,"%p %d %d %p %p %u %d %s\n",
+                       kp->ptr, kp->len, kp->used_vmalloc, kp->caller, kp->owner, kp->time, kp->line, kp->filename);
+               len = strlen(str);
+               
+               /* does it fit on the current page */
+               if ( (current_pos + len + 1) >=  QSNET_PRIVATE_PAGE_SIZE)
+               {
+                       /* move onto next page */
+                       if ((current_space = current_space->next) == NULL)
+                       {
+                               /* run out of space !!!! */
+                               spin_unlock_irqrestore(&kmalloc_lock, flags);
+                               return (1);
+                       }
+                       current_space->space[0] = 0;    
+                       current_pos = 0;
+               }
+               strcat( current_space->space + current_pos, str);
+               current_pos += len;
+
+               /* remember how much we wrote to this page */
+               current_space->size = current_pos;
+
+       }
+       spin_unlock_irqrestore(&kmalloc_lock, flags);
+
+       return (0);
+}
+
+void * 
+qsnet_kmem_alloc_debug(int len, int cansleep, int zerofill, char *file, int line)
+{
+       void *new;
+       unsigned long flags;
+       kmalloc_t *kp;
+
+       if (len < PAGE_SIZE || !cansleep)
+               new = kmalloc(len, cansleep ? GFP_KERNEL : GFP_ATOMIC);
+       else
+               new = vmalloc(len);
+
+       if (len >= PAGE_SIZE)
+               ASSERT(PAGE_ALIGNED((uintptr_t) new));
+
+       if (new && zerofill)
+               memset(new,0,len);
+
+       /* record allocation */
+       kp = kmalloc(sizeof(kmalloc_t), cansleep ? GFP_KERNEL : GFP_ATOMIC);
+       ASSERT(kp != NULL);
+       kp->len = len;
+       kp->ptr = new;
+       kp->used_vmalloc = (len >= PAGE_SIZE || cansleep);
+       kp->owner  = current;
+       kp->caller = __builtin_return_address(0);
+       kp->time = lbolt;
+       kp->line = line;
+       len = strlen(file);
+
+       if (len > 18) 
+               strcpy(kp->filename,&file[len-18]);
+       else
+               strcpy(kp->filename,file);
+
+       spin_lock_irqsave(&kmalloc_lock, flags);
+       list_add(&kp->list, &kmalloc_head);
+       spin_unlock_irqrestore(&kmalloc_lock, flags);
+
+       return new;
+}
+
+void 
+qsnet_kmem_free_debug(void *ptr, int len, char *file, int line)
+{
+       unsigned long flags;
+       kmalloc_t *kp;
+       struct list_head *lp;
+
+       spin_lock_irqsave(&kmalloc_lock, flags);
+       for (lp = kmalloc_head.next; lp != &kmalloc_head; lp = lp->next) {
+               kp = list_entry(lp, kmalloc_t, list);
+               if (kp->ptr == ptr) {
+                       if (kp->len != len)
+                               printk("qsnet_kmem_free_debug(%p) ptr %p len %d mismatch: expected %d caller %p owner %p (%s:%d)\n",
+                                      current, ptr, len, kp->len, __builtin_return_address(0), kp->caller, file, line);
+                       list_del(lp);
+                       kfree(kp); /* free off descriptor */
+                       break;
+               }
+       }
+       spin_unlock_irqrestore(&kmalloc_lock, flags);
+
+       if (lp == &kmalloc_head) /* segment must be found */
+       {
+               printk( "qsnet_kmem_free_debug(%p) ptr %p len %d not found: caller %p (%s:%d)\n",
+                       current, ptr, len, __builtin_return_address(0), file, line);
+       }
+
+       if ((((unsigned long) ptr) >= VMALLOC_START && ((unsigned long) ptr) < VMALLOC_END)) 
+               vfree (ptr);
+       else
+               kfree (ptr);
+}
+
+#else /* !KMEM_DEBUG */
+
+void * 
+qsnet_kmem_alloc(int len, int cansleep, int zerofill)
+{
+       void *new;
+
+       if (len < PAGE_SIZE || !cansleep)
+               new = kmalloc(len, cansleep ? GFP_KERNEL : GFP_ATOMIC);
+       else
+               new = vmalloc(len);
+
+       if (len >= PAGE_SIZE)
+               ASSERT(PAGE_ALIGNED((unsigned long) new));
+
+       if (new && zerofill)
+               memset(new,0,len);
+
+       return new;
+}
+
+void 
+qsnet_kmem_free(void *ptr, int len)
+{
+       if ((((unsigned long) ptr) >= VMALLOC_START && ((unsigned long) ptr) < VMALLOC_END)) 
+               vfree (ptr);
+       else
+               kfree (ptr);
+}
+#endif /* !KMEM_DEBUG */
+
+void
+qsnet_kmem_display(void *handle)
+{
+#ifdef KMEM_DEBUG
+       kmalloc_t *kp;
+       struct list_head *lp;
+       unsigned long flags;
+       int count = 0, totsize = 0;
+
+       spin_lock_irqsave(&kmalloc_lock, flags);
+       for (lp = kmalloc_head.next; lp != &kmalloc_head;  lp = lp->next) {
+               kp = list_entry(lp, kmalloc_t, list);
+
+               if (!handle || handle == kp->owner)
+               {
+                       printk("qsnet_kmem_display(%p): mem %p len %d unfreed caller %p (%p) \n",
+                              handle, kp->ptr, kp->len, kp->caller, kp->owner);
+                   
+                       count++;
+                       totsize += kp->len;
+               }
+       }
+       spin_unlock_irqrestore(&kmalloc_lock, flags);
+
+       printk("qsnet_kmem_display(%p): %d bytes left in %d objects\n", handle, totsize, count);
+#endif
+}
+
+physaddr_t
+kmem_to_phys(void *ptr)
+{
+       virtaddr_t virt = (virtaddr_t) ptr;
+       physaddr_t phys;
+       pte_t     *pte;
+
+       if ((virt >= VMALLOC_START && virt < VMALLOC_END))
+       {
+               pte = find_pte_kernel(virt);
+               ASSERT(pte && !pte_none(*pte));
+               phys = pte_phys(*pte) + (virt & (PAGE_SIZE-1));
+       }
+#if defined(PKMAP_BASE)
+       else if (virt >= PKMAP_BASE && virt < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE))
+       {
+               pte = find_pte_kernel(virt);
+               ASSERT(pte && !pte_none(*pte));
+               phys = pte_phys(*pte) + (virt & (PAGE_SIZE-1));
+       }
+#endif
+#if defined(__ia64)
+       else if (virt >= __IA64_UNCACHED_OFFSET && virt < PAGE_OFFSET)
+       {
+               /* ia64 non-cached KSEG */
+               phys = ((physaddr_t) ptr - __IA64_UNCACHED_OFFSET);
+       }
+#endif
+       else /* otherwise it's KSEG */
+       {
+               phys = __pa(virt);
+       }
+           
+#if defined(CONFIG_ALPHA_GENERIC) || (defined(CONFIG_ALPHA_EV6) && !defined(USE_48_BIT_KSEG))
+       /* 
+        * with TS_BIAS as bit 40 - the tsunami pci space is mapped into
+        * the kernel at 0xfffff500.00000000 however we need to convert
+        * this to the true physical address 0x00000800.00000000.
+        *
+        * there is no need for PHYS_TWIDDLE since we knew we'd get a kernel
+        * virtual address already and handled this with __pa().
+        */
+       if (phys & (1ul << 40)) {
+               phys &= ~(1ul << 40);   /*   clear bit 40 (kseg I/O select) */
+               phys |= (1ul << 43);    /*   set   bit 43 (phys I/O select) */
+       }
+#endif
+       return phys;
+}
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,4,0)
+
+EXPORT_SYMBOL(pci_resource_size);
+EXPORT_SYMBOL(pci_get_base_address);
+EXPORT_SYMBOL(pci_base_to_kseg);
+
+
+/*
+ * PCI stuff.  
+ *
+ * XXX pci_base_to_kseg() and pci_kseg_to_phys() are problematic
+ * in that they may not work on non-Tsunami (DS20, ES40, etc) 
+ * architectures, and may not work in non-zero PCI bus numbers.
+ */
+
+unsigned long 
+pci_get_base_address(struct pci_dev *pdev, int index)
+{
+       unsigned long base;
+
+       ASSERT(index >= 0 && index <= 5);
+       /* borrowed in part from drivers/scsi/sym53c8xx.c */
+       base = pdev->base_address[index++];
+
+#if BITS_PER_LONG > 32
+       if ((base & 0x7) == 0x4)
+               base |= (((unsigned long)pdev->base_address[index]) << 32);
+#endif
+       return base;
+}
+
+unsigned long 
+pci_resource_size(struct pci_dev *pdev, int index)
+{
+       u32 addr, mask, size;
+
+       static u32 bar_addr[] = {
+               PCI_BASE_ADDRESS_0, 
+               PCI_BASE_ADDRESS_1, 
+               PCI_BASE_ADDRESS_2,
+               PCI_BASE_ADDRESS_3, 
+               PCI_BASE_ADDRESS_4, 
+               PCI_BASE_ADDRESS_5, 
+       };
+       ASSERT(index >= 0 && index <= 5);
+
+       /* algorithm from Rubini book */
+       pci_read_config_dword (pdev,    bar_addr[index], &addr);
+       pci_write_config_dword(pdev,    bar_addr[index], ~0);
+       pci_read_config_dword (pdev,    bar_addr[index], &mask);
+       pci_write_config_dword(pdev,    bar_addr[index], addr);
+
+       mask &= PCI_BASE_ADDRESS_MEM_MASK;
+       size = ~mask + 1;
+       return size;
+}
+
+/*
+ * Convert BAR register value to KSEG address.
+ */
+void *
+pci_base_to_kseg(u64 baddr, int bus)
+{
+       u64 kseg;
+
+       /* XXX tsunami specific */
+       baddr &= ~(u64)0x100000000;  /* mask out hose bit */
+       kseg = TSUNAMI_MEM(bus) + baddr;
+       return (void *)kseg; 
+}
+
+#endif /* LINUX_VERSION_CODE <= KERNEL_VERSION(2,4,0) */
+
+/*
+ * Spin the other CPU's in an SMP system.
+ * smp_call_function() needed to be exported to modules.  It will be
+ * papered over in <linux/smp.h> if running on a non-SMP box.
+ */
+static spinlock_t hold_lock = SPIN_LOCK_UNLOCKED;
+
+#if 0
+static void cpu_hold(void *unused)
+{
+       spin_lock(&hold_lock);
+       spin_unlock(&hold_lock);
+}
+#endif
+
+void cpu_hold_all(void)
+{
+       spin_lock(&hold_lock);
+
+#if 0
+       {
+               int res;
+               int retries = 10; 
+           
+               /* XXXXX: cannot call smp_call_function() from interrupt context */
+           
+               do {
+                       /* only request blocking retry if not in interrupt context */
+                       res = smp_call_function(cpu_hold, NULL, !in_interrupt(), 0);
+                       if (res)
+                               mdelay(5);
+               } while (res && retries--);
+           
+               if (res)
+                       printk("cpu_hold_all: IPI timeout\n");
+       }
+#endif
+}
+
+void cpu_release_all(void)
+{
+       spin_unlock(&hold_lock);
+}
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/qsnet/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/qsnet/Makefile  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/qsnet/Makefile       2005-06-01 23:12:54.697424208 -0400
@@ -0,0 +1,31 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/qsnet/Makefile
+#
+
+
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/qsnet/Makefile
+#
+
+list-multi             := qsnet.o
+qsnet-objs     := debug.o kernel_linux.o i686_mmx.o
+export-objs            := kernel_linux.o
+obj-$(CONFIG_QSNET)    := qsnet.o
+
+qsnet.o : $(qsnet-objs)
+       $(LD) -r -o $@ $(qsnet-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/qsnet/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/qsnet/Makefile.conf     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/qsnet/Makefile.conf  2005-06-01 23:12:54.698424056 -0400
@@ -0,0 +1,10 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       qsnet.o
+MODULENAME     =       qsnet
+KOBJFILES      =       debug.o kernel_linux.o i686_mmx.o
+EXPORT_KOBJS   =       kernel_linux.o
+CONFIG_NAME    =       CONFIG_QSNET
+SGALFC         =       
+# EXTRALINES START
+
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/qsnet/qsnetkmem_linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/qsnet/qsnetkmem_linux.c 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/qsnet/qsnetkmem_linux.c      2005-06-01 23:12:54.698424056 -0400
@@ -0,0 +1,325 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: qsnetkmem_linux.c,v 1.3 2003/08/13 10:03:27 fabien Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/qsnetkmem_linux.c,v $*/
+
+/* macro macros */
+#define MACRO_BEGIN     do {
+#define MACRO_END       } while (0)
+#define offsetof(T,F) ((int )&(((T *)0)->F))
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <ctype.h>
+#include <sys/types.h>
+#include <errno.h>
+#include <unistd.h>
+#include <string.h>
+#include <qsnet/config.h>
+#include <qsnet/list.h>
+#include <qsnet/procfs_linux.h>
+#include <signal.h>
+#include <sys/wait.h>
+
+#define LIST_HEAD_INIT(name) { &(name), &(name) }
+
+#define LIST_HEAD(name) \
+       struct list_head name = LIST_HEAD_INIT(name)
+
+typedef struct {
+       struct list_head list;
+       void            *ptr;
+       int             len;
+       int             used_vmalloc;
+       void            *owner;
+       void            *caller;
+       unsigned int     time;
+       int              mark;
+       int              line;
+       char             file[256];
+       
+} kmalloc_t;
+
+
+static LIST_HEAD(current_kmem);
+static LIST_HEAD(stored_kmem);
+
+void
+count_kmem(struct list_head * list, long * count, long * size )
+{
+       long              c,s;
+       struct list_head *tmp;
+       kmalloc_t        *kmem_ptr = NULL;
+
+
+       c = s = 0L;
+
+       list_for_each(tmp, list) {
+               kmem_ptr = list_entry(tmp, kmalloc_t , list);
+               c++;
+               s += kmem_ptr->len;
+       }       
+
+       *count = c;
+       *size  = s;
+}
+
+void
+clear_kmem(struct list_head * list)
+{
+       struct list_head *tmp,*tmp2;
+       kmalloc_t        *kmem_ptr = NULL;
+
+       list_for_each_safe(tmp, tmp2, list) {
+               kmem_ptr = list_entry(tmp, kmalloc_t , list);
+               list_del_init(&kmem_ptr->list);
+               free( kmem_ptr );
+       }
+}
+
+void 
+move_kmem(struct list_head * dest, struct list_head *src)
+{
+       struct list_head *tmp,*tmp2;
+       kmalloc_t        *kp= NULL;
+
+       list_for_each_safe(tmp, tmp2, src) {
+               kp = list_entry(tmp, kmalloc_t , list);
+               list_del_init(&kp->list);
+
+/*
+               printf("mem %p len %d (vm=%d)  caller %p owner %p (%s:%d)\n",
+                      kp->ptr, kp->len, kp->used_vmalloc, kp->caller, kp->owner, kp->file, kp->line);
+*/                 
+
+               list_add_tail(&kp->list, dest);
+       }
+}
+
+void
+read_kmem(struct list_head * list)
+{
+       FILE      * fd;
+       char        line[1024];
+       int         line_size = 100;
+       char      * rep;
+       kmalloc_t * kp;
+
+       clear_kmem(list);
+
+       fd = fopen(QSNET_PROCFS_KMEM_DEBUG,"r");
+       if ( fd == NULL) 
+       {
+               printf("No Kmem Debug\n");
+               return;
+       }
+
+       rep = fgets(line,line_size, fd);
+
+       while ( rep != NULL ) 
+       {
+               kp = malloc(sizeof(kmalloc_t));
+
+               sscanf(line,"%p %d %d %p %p %u %d %s\n",
+                      &kp->ptr, &kp->len, &kp->used_vmalloc, &kp->caller, &kp->owner, &kp->time, &kp->line, &kp->file[0]);
+
+/*
+               printf(">>%s<<\n",line);
+               printf("%p %d %d %p %p %u %d %s\n",
+                      kp->ptr, kp->len, kp->used_vmalloc, kp->caller, kp->owner, kp->time, kp->line, kp->file);
+*/
+
+               list_add_tail(&kp->list, list);
+
+               rep = fgets(line,line_size, fd);
+       }
+       fclose(fd);
+}
+
+void
+mark_kmem(struct list_head * list, int mark)
+{
+       struct list_head *tmp;
+       kmalloc_t        *kp = NULL;
+
+       list_for_each(tmp, list) {
+               kp = list_entry(tmp, kmalloc_t , list);
+
+               kp->mark = mark;
+       }
+}
+
+kmalloc_t *
+find_kmem(kmalloc_t * value, struct list_head * list)
+{
+       struct list_head *tmp;
+       kmalloc_t        *kp = NULL;
+
+       
+       list_for_each(tmp, list) {
+               kp = list_entry(tmp, kmalloc_t , list);
+               if ( (kp->ptr == value->ptr)
+                    && (kp->len == value->len)
+                    && (kp->used_vmalloc  == value->used_vmalloc )
+                    && (kp->owner  == value->owner )
+                    && (kp->caller  == value->caller )
+                    && (kp->time  == value->time )
+                    && (kp->line  == value->line )
+                    && !(strcmp(kp->file,value->file) ))
+                       return kp;
+       }       
+       return NULL;
+}
+
+void 
+diff_kmem(struct list_head *curr, struct list_head *stored)
+{
+       struct list_head *tmp;
+       kmalloc_t        *kp = NULL;
+       long              c,s;
+
+       mark_kmem(stored,  0);
+       mark_kmem(curr,    0);
+       
+       list_for_each(tmp, stored) {
+               kp = list_entry(tmp, kmalloc_t , list);
+               if (find_kmem( kp, curr) != NULL) 
+                       kp->mark = 1;
+       }
+       
+       list_for_each(tmp, curr) {
+               kp = list_entry(tmp, kmalloc_t , list);
+               if (find_kmem( kp, stored) != NULL) 
+                       kp->mark = 1;
+       }               
+
+       c=s=0L;
+       list_for_each(tmp, stored) {
+               kp = list_entry(tmp, kmalloc_t , list);
+               if (kp->mark != 1)
+               {
+                       printf("-- mem %p len %d (vm=%d)  caller %p owner %p (%s:%d)\n",
+                              kp->ptr, kp->len, kp->used_vmalloc, kp->caller, kp->owner, kp->file, kp->line);
+                       c++;
+                       s+= kp->len;
+               }
+       }
+       printf("-- %4ld %10ld \n",c,s);
+       
+       c=s=0L;
+       list_for_each(tmp, curr) {
+               kp = list_entry(tmp, kmalloc_t , list);
+               if (kp->mark != 1)
+               {
+                       printf("++ mem %p len %d (vm=%d)  caller %p owner %p (%s:%d)\n",
+                              kp->ptr, kp->len, kp->used_vmalloc, kp->caller, kp->owner, kp->file, kp->line);
+                       c++;
+                       s+= kp->len;
+               }
+       }               
+       printf("++ %4ld %10ld \n",c,s);
+}
+
+
+void
+print_kmem(struct list_head * list)
+{
+       struct list_head *tmp;
+       kmalloc_t        *kp = NULL;
+
+       list_for_each(tmp, list) {
+               kp = list_entry(tmp, kmalloc_t , list);
+
+               printf("mem %p len %d (vm=%d)  caller %p owner %p (%s:%d)\n",
+                      kp->ptr, kp->len, kp->used_vmalloc, kp->caller, kp->owner, kp->file, kp->line);
+                   
+       }
+}
+
+void 
+print_cmds()
+{
+       long c,s;
+
+       printf("q : quits \n");
+       printf("r : read\n");
+       printf("c : print current\n");
+       printf("o : print stored\n");
+       printf("s : store\n");
+
+       count_kmem(&current_kmem, &c, &s );
+       printf("\ncurrent : %4ld %10ld\n", c , s);
+ 
+       count_kmem(&stored_kmem, &c, &s );
+       printf("store   : %4ld %10ld\n", c , s);
+ 
+}
+
+int
+main()
+{
+       char            line[128];
+       int             line_size=127;
+       int             len;
+
+
+       while (1)
+       {
+               
+               printf(">> ");
+               fgets(line,line_size, stdin);
+       
+               
+               len = strlen( line ) -1;
+               if ( len ) 
+               {
+                       switch ( tolower(line[0]) ) 
+                       {
+                       case 'q':
+                               exit(0);
+
+                       case 'r' :
+                               read_kmem(&current_kmem);
+                               break;
+
+                       case 'c' :
+                               print_kmem(&current_kmem);
+                               break;
+
+                       case 'o' :
+                               print_kmem(&stored_kmem);
+                               break;
+
+                       case 's' :
+                               clear_kmem(&stored_kmem);
+                               move_kmem(&stored_kmem, &current_kmem);
+                               break;
+
+                       case 'd' :
+                               diff_kmem(&current_kmem, &stored_kmem);
+                               break;
+
+                       default:
+                               print_cmds();   
+                       }
+
+               
+                       
+               }
+               else
+                       print_cmds();
+       }
+
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/drivers/net/qsnet/qsnet/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/qsnet/quadrics_version.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/qsnet/quadrics_version.h     2005-06-01 23:12:54.699423904 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/rms/Makefile
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/rms/Makefile    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/rms/Makefile 2005-06-01 23:12:54.699423904 -0400
@@ -0,0 +1,31 @@
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2002-2004 Quadrics Ltd
+#
+# File: drivers/net/qsnet/rms/Makefile
+#
+
+
+#
+
+#
+# Makefile for Quadrics QsNet
+#
+# Copyright (c) 2004 Quadrics Ltd.
+#
+# File: driver/net/qsnet/rms/Makefile
+#
+
+list-multi             := rms.o
+rms-objs       := rms_kern.o rms_kern_Linux.o
+export-objs            := 
+obj-$(CONFIG_RMS)      := rms.o
+
+rms.o : $(rms-objs)
+       $(LD) -r -o $@ $(rms-objs)
+
+EXTRA_CFLAGS           +=  -DDEBUG -DDEBUG_PRINTF -DDEBUG_ASSERT
+
+include $(TOPDIR)/Rules.make
+
Index: linux-2.4.21/drivers/net/qsnet/rms/Makefile.conf
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/rms/Makefile.conf       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/rms/Makefile.conf    2005-06-01 23:12:54.699423904 -0400
@@ -0,0 +1,10 @@
+# Flags for generating QsNet Linux Kernel Makefiles
+MODNAME                =       rms.o
+MODULENAME     =       rms
+KOBJFILES      =       rms_kern.o rms_kern_Linux.o
+EXPORT_KOBJS   =       
+CONFIG_NAME    =       CONFIG_RMS
+SGALFC         =       
+# EXTRALINES START
+
+# EXTRALINES END
Index: linux-2.4.21/drivers/net/qsnet/rms/quadrics_version.h
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/rms/quadrics_version.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/rms/quadrics_version.h       2005-06-01 23:12:54.700423752 -0400
@@ -0,0 +1 @@
+#define QUADRICS_VERSION "4.30qsnet"
Index: linux-2.4.21/drivers/net/qsnet/rms/rms_kern.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/rms/rms_kern.c  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/rms/rms_kern.c       2005-06-01 23:12:54.702423448 -0400
@@ -0,0 +1,1757 @@
+/*
+ * Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ * For licensing information please see the supplied COPYING file
+ *
+ * rms_kern.c:    RMS kernel module
+ *
+ * $Source: /cvs/master/quadrics/rmsmod/rms_kern.c,v $
+ */
+
+#ident "@(#)$Id: rms_kern.c,v 1.62.2.4 2005/01/18 11:05:45 duncan Exp $"
+
+#include <stddef.h>
+#include <qsnet/kernel.h>
+#include <qsnet/autoconf.h>
+#include <rms/rmscall.h>
+
+/*
+ * extend stats added in version 5
+ */
+#define RMS_MODVERSION 5
+
+#if defined(SOLARIS)
+
+#define CURUID() CURPROC()->p_cred->cr_uid
+#define RMS_NCPUS() 4
+#define PROC_STRUCT proc
+
+#include <sys/time.h>
+
+#elif defined(LINUX)
+
+#ifdef PROCESS_ACCT 
+#define TIMEVAL_TO_MSEC(tv) ((tv)->tv_sec * 1000 + (tv)->tv_usec / 1000)
+#define TIMEVAL_TO_CT(tv)   ((tv)->tv_sec * HZ + (tv)->tv_usec / (1000000L / HZ))
+#endif
+
+#ifdef RSS_ATOMIC
+#define PROC_RSS(proc) ((proc)->mm ? atomic_read(&(proc)->mm->rss) : 0)
+#else
+#define PROC_RSS(proc) ((proc)->mm ? (proc)->mm->rss : 0)
+#endif
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 0)
+#      define  RMS_NCPUS()     smp_num_cpus
+#else
+#      define RMS_NCPUS()      num_online_cpus()
+#endif
+
+#define CURUID()       CURPROC()->uid
+#define p_pid          pid
+#define PROC_STRUCT    task_struct
+
+/* care needed with conversion to millisecs on 32-bit Linux */
+#ifdef LINUX
+#ifdef LINUX_I386
+#define CT_TO_MSEC(x)  ct_to_msec(x)
+
+uint64_t ct_to_msec(clock_t t)
+{
+    uint64_t msecs;
+    if (t < 2000000)
+    {
+       t = (1000 * t)/HZ;
+       msecs = t;
+    }
+    else
+    {
+       t = t / HZ;
+       msecs = t * 1000;
+    }
+    return(msecs);
+}
+
+#else
+#define CT_TO_MSEC(x)  (((x) * 1000)/HZ)
+#endif
+#endif
+
+#ifndef FALSE
+#define FALSE          (0)
+#define TRUE           (!FALSE)
+#endif
+
+#include <linux/time.h>
+#include <linux/proc_fs.h>
+#include <linux/ptrack.h>
+
+#include <linux/module.h>
+
+#elif defined(DIGITAL_UNIX)
+
+#define CURUID() CURPROC()->p_ruid
+extern  int ncpus;
+#define RMS_NCPUS() ncpus
+#define PROC_STRUCT proc
+#define TIMEVAL_TO_MSEC(tv) ((tv)->tv_sec * 1000 + (tv)->tv_usec / 1000)
+
+#include <sys/time.h>
+
+#else
+#error cannot determine operating system
+#endif
+
+int shm_cleanup(void);
+
+struct cap_desc {
+
+    struct cap_desc *next;
+    int              index;    /* index of capability in program */
+    ELAN_CAPABILITY  cap;      /* elan capability */
+
+};
+
+struct proc_desc {
+    
+    struct proc_desc    *next;
+    struct PROC_STRUCT  *proc;
+    struct prg_desc     *program;      /* controlling program         */
+    int                  mycap;                /* index of my capability      */
+    int                  myctx;                /* context number for process  */
+    int                  flags;
+    int                  vp;           /* elan virtual process number */
+};
+
+struct prg_desc {
+    
+    struct prg_desc  *next;            
+    int               id;      /* program id                          */
+    int               flags;   /* program status flags                */
+    uid_t             uid;     /* user id                             */
+    int               ncpus;   /* number of cpus allocated to program */
+    int               nprocs;  /* number of processes in program      */
+    struct proc_desc *pdescs;  /* processes in this program           */
+    int               ncaps;   /* number of capabilities              */
+    struct cap_desc  *caps;    /* elan capabilities                   */
+    char             *corepath;        /* core path for parallel program      */
+    int               psid;    /* processor set id                    */
+
+    uint64_t       cutime;     /* user time accumulated by children   */
+    uint64_t       cstime;     /* system time accumulated by children */
+    uint64_t       start_time; /* time program created                */
+    uint64_t       end_time;   /* time last process exited            */
+    uint64_t       sched_time; /* last time job was scheduled         */
+    uint64_t       accum_atime;        /* allocated time last deschedule      */
+    uint64_t       memint;     /* accumulated memory integral         */
+    uint64_t       ebytes;     /* data transferred by the Elan(s)     */
+    uint64_t       exfers;     /* number of Elan data transfers       */
+    long           maxrss;     /* maximum size to date                */
+    long           majflt;
+    
+#ifdef LINUX
+    struct proc_dir_entry *proc_entry;
+#endif
+
+};
+
+#if defined(LINUX)
+static int rms_ptrack_callback (void *arg, int phase, struct task_struct *child);
+#else
+static void rms_xd_callback(void *arg, int phase, void *ctask);
+static void rms_xa_callback (void *arg, int phase, void *ctask);
+#endif
+
+static void prgsignal(struct prg_desc *program, int signo);
+static uint64_t gettime(void);
+static void freeProgram(struct prg_desc *program);
+
+static struct prg_desc *programs = 0;
+
+kmutex_t rms_lock;
+
+int rms_init(void)
+{
+    kmutex_init (&rms_lock);
+
+    DBG(printk("rms: initialising\n"));
+
+    return(ESUCCESS);
+}
+
+int rms_reconfigure(void)
+{
+    return(ESUCCESS);
+}
+
+int rms_programs_registered(void)
+{
+    /*
+    ** Called when trying to unload rms.mod will not succeed
+    ** if programs registered
+    */
+ 
+   struct prg_desc *program, **pp;
+
+   kmutex_lock(&rms_lock);
+
+   for (program = programs; program; program = program->next)
+   {
+       if (program->nprocs != 0)
+       {
+            kmutex_unlock(&rms_lock);
+            return(EBUSY);
+       }
+   }
+
+   /*
+   ** We have traversed the programs list and no processes registered
+   ** Now free the memory
+   */
+       
+    pp = &programs;
+    while ((program = *pp) != NULL)
+    {
+        *pp = program->next;
+        freeProgram(program);
+    }
+    kmutex_unlock(&rms_lock);
+   
+    return(ESUCCESS);
+
+}
+
+int rms_fini(void)
+{
+    /*
+     * don't allow an unload if there are programs registered
+     */
+    if (rms_programs_registered())
+        return(EBUSY);
+
+    kmutex_destroy (&rms_lock);
+
+    DBG(printk("rms: removed\n"));
+
+    return(ESUCCESS);
+}
+
+#ifdef LINUX
+
+extern struct proc_dir_entry *rms_procfs_programs;
+
+/*
+ * display one pid per line if there isn't enough space 
+ * for another pid then add "...\n" and stop 
+ */
+int pids_callback(char* page, char** start, off_t off, int count, int* eof, void* data)
+{
+    struct prg_desc *program = (struct prg_desc *)data;
+    struct proc_desc *pdesc;
+    char *ptr = page;
+    int bytes = 0, nb;
+
+    kmutex_lock(&rms_lock);
+    
+    for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+    {
+       if (bytes > count - 15)
+       {
+           bytes += sprintf(ptr,"...\n");
+           break;
+       }
+        nb = sprintf(ptr, "%d %d\n", pdesc->proc->p_pid, pdesc->vp);
+       bytes += nb;
+       ptr += nb;
+    }
+    kmutex_unlock(&rms_lock);
+    
+    return(bytes);
+}
+
+int status_callback(char* page, char** start, off_t off, int count, int* eof, void* data)
+{
+    struct prg_desc *program = (struct prg_desc *)data;
+    int bytes;
+    if (program->flags & PRG_KILLED)
+       bytes = sprintf(page, "killed\n");
+    else
+       bytes = sprintf(page, "running\n");
+    return(bytes);
+}
+
+void rms_create_proc_entry(struct prg_desc *program)
+{
+    struct proc_dir_entry *p;
+    char name[32];
+
+    if (rms_procfs_programs)
+    {
+       sprintf(name,"%d", program->id);
+       if ((program->proc_entry = proc_mkdir(name, rms_procfs_programs)) != NULL)
+       {
+           if ((p = create_proc_entry ("pids", S_IRUGO, program->proc_entry)) != NULL)
+           {
+               p->owner = THIS_MODULE;
+               p->data = program;
+               p->read_proc = pids_callback;
+           }
+           if ((p = create_proc_entry ("status", S_IRUGO, program->proc_entry)) != NULL)
+           {
+               p->owner = THIS_MODULE;
+               p->data = program;
+               p->read_proc = status_callback;
+           }
+       }
+    }
+}
+
+void rms_remove_proc_entry(struct prg_desc *program)
+{
+    char name[32];
+    if (rms_procfs_programs)
+    {
+       if (program->proc_entry)
+       {
+           remove_proc_entry ("pids", program->proc_entry);
+           remove_proc_entry ("status", program->proc_entry);
+       }
+       sprintf(name,"%d", program->id);
+       remove_proc_entry (name, rms_procfs_programs);
+    }
+}
+
+#endif
+
+/*
+ * find a program from its index/pid
+ *
+ * Duncan:  make the lookup more efficient for large numbers of programs/processes
+ */
+static struct prg_desc *findProgram(const int id)
+{
+    struct prg_desc *program;
+    for (program = programs; program; program = program->next)
+       if (program->id == id)
+           return(program);
+    return(0);
+}
+
+static struct proc_desc *findProcess(const int pid)
+{
+    struct prg_desc *program;
+    struct proc_desc *pdesc;
+    for (program = programs; program; program = program->next)
+       for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+           if (pdesc->proc->p_pid == pid)
+               return(pdesc);
+    return(0);
+}
+
+static void freeProgram(struct prg_desc *program)
+{
+    struct proc_desc *pdesc;
+    struct cap_desc *cdesc;
+
+#ifdef LINUX
+    rms_remove_proc_entry(program);
+#endif
+
+    while ((pdesc = program->pdescs) != NULL)
+    {
+       program->pdescs = pdesc->next;
+       KMEM_FREE(pdesc, sizeof(struct proc_desc));
+    }
+
+    while ((cdesc = program->caps) != NULL)
+    {
+       program->caps = cdesc->next;
+       KMEM_FREE(cdesc, sizeof(struct cap_desc));
+    }
+
+    if (program->corepath)
+       KMEM_FREE(program->corepath, MAXCOREPATHLEN + 1);
+
+    KMEM_FREE(program, sizeof(struct prg_desc));
+
+#ifdef LINUX
+    MOD_DEC_USE_COUNT;
+#endif
+}
+
+/*
+ * rms_prgcreate
+ *
+ * create a new program description
+ */
+int rms_prgcreate(int id, uid_t uid, int cpus)
+{
+    struct prg_desc *program;
+    struct proc_desc *pdesc;
+    
+    DBG(printk("rms_prgcreate :: program %d pid %d uid %d cpus %d\n", id, CURPROC()->p_pid, uid, cpus));
+    
+    /*
+     * parallel programs are created as root by the rmsd as it forks the loader
+     */
+    if (CURUID())
+       return(EACCES);
+    
+    /*
+     * program ids must be unique
+     */
+    kmutex_lock(&rms_lock);
+    program = findProgram(id);
+    kmutex_unlock(&rms_lock);
+    if (program)
+       return(EINVAL);
+
+    /*
+     * create a new program description
+     */
+    KMEM_ALLOC(program, struct prg_desc *, sizeof(struct prg_desc), TRUE);
+    if (!program)
+       return(ENOMEM);
+
+    program->id = id;
+    program->flags = PRG_RUNNING;
+    program->ncpus = cpus;
+    program->nprocs = 1;
+    program->uid = uid;
+    program->ncaps = 0;
+    program->caps = 0;
+    program->corepath = 0;
+    program->psid = 0;
+    program->start_time = program->sched_time = gettime();
+    program->end_time = 0;
+    program->accum_atime = 0;
+    program->cutime = 0;
+    program->cstime = 0;
+    program->maxrss = 0;
+    program->memint = 0;
+    program->majflt = 0;
+    program->ebytes = 0;
+    program->exfers = 0;
+
+    KMEM_ALLOC(pdesc, struct proc_desc *, sizeof(struct proc_desc), TRUE);
+    if (!pdesc)
+       return(ENOMEM);
+
+    pdesc->proc = CURPROC();
+    pdesc->next = 0;
+    pdesc->mycap = ELAN_CAP_UNINITIALISED;
+    pdesc->myctx = ELAN_CAP_UNINITIALISED;
+    pdesc->vp = -1;            /* rmsloader */
+    pdesc->program = program;
+    program->pdescs = pdesc;
+    
+#ifdef LINUX
+    rms_create_proc_entry(program);
+#endif
+    
+    kmutex_lock(&rms_lock);
+
+#if defined(LINUX)
+    if (ptrack_register (rms_ptrack_callback, NULL) != 0)
+    {
+       kmutex_unlock(&rms_lock);
+        KMEM_FREE(pdesc,sizeof(struct proc_desc));
+        KMEM_FREE(program,sizeof(struct prg_desc));
+       return(ENOMEM);
+    }
+#else
+    /*
+     * install a fork handler
+     */
+    if (HANDLER_REGISTER((void *)(unsigned long)rms_xa_callback, NULL, XA_FORK | XA_EXIT | XA_IOF | XA_KOF | XA_KOE) == NULL)
+    {
+       kmutex_unlock(&rms_lock);
+        KMEM_FREE(pdesc,sizeof(struct proc_desc));
+        KMEM_FREE(program,sizeof(struct prg_desc));
+       return(ENOMEM);
+    }
+#endif
+
+    program->next = programs;
+    programs = program;
+    
+#ifdef LINUX
+    MOD_INC_USE_COUNT;
+#endif
+    
+    kmutex_unlock(&rms_lock);
+    return(ESUCCESS);
+}
+
+
+/*
+ * rms_prgdestroy
+ *
+ * destroy a program description
+ */
+int rms_prgdestroy(int id)
+{
+    struct prg_desc *program, **pp;
+    int status = ESRCH;
+
+    /*
+     * parallel programs are created and destroyed by the rmsd
+     */
+    if (CURUID())
+       return(EACCES);
+
+    kmutex_lock(&rms_lock);
+    
+    pp = &programs;
+    while ((program = *pp) != NULL)
+    {
+       if (program->id == id)
+       {
+           if (program->nprocs == 0)
+           {
+               DBG(printk("rms_prgdestro :: removing program %d\n", program->id));
+               *pp = program->next;
+               freeProgram(program);
+               status = ESUCCESS;
+           }
+           else
+           {
+               DBG(printk("rms_prgdestro :: failed to remove program %d: %d\n", program->id, program->nprocs));
+               status = ECHILD;
+               pp = &program->next;
+           }
+       }
+       else
+           pp = &program->next;
+    }
+    
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+/*
+ * rms_prgids
+ */
+int rms_prgids(int maxids, int *prgids, int *nprgs)
+{
+    struct prg_desc *program;
+    int count = 0, *buf, *bufp;
+    int status = ESUCCESS;
+
+    if (maxids < 1)
+        return(EINVAL);
+
+    kmutex_lock(&rms_lock);
+
+    for (program = programs; program; program = program->next)
+        count++;
+    count = MIN(count, maxids);
+
+    if (count > 0)
+    {
+        KMEM_ALLOC(buf, int *, count * sizeof(int), TRUE);
+       if (buf)
+       {                  
+           for (program = programs, bufp=buf; bufp < buf + count; 
+                program = program->next)
+               *bufp++ = program->id;
+       
+           if (copyout(buf, prgids, sizeof(int) * count))
+               status = EFAULT;
+
+           KMEM_FREE(buf, count * sizeof(int));
+       }
+       else
+           status = ENOMEM;
+    }
+    
+    if (copyout(&count, nprgs, sizeof(int)))
+       status = EFAULT;
+
+    kmutex_unlock(&rms_lock);
+    
+    return(status);
+}
+
+/*
+ * rms_prginfo
+ */
+int rms_prginfo(int id, int maxpids, pid_t *pids, int *nprocs)
+{
+    struct prg_desc *program;
+    struct proc_desc *pdesc;
+    pid_t *pidp, *buf;
+    int status = ESUCCESS;
+
+    kmutex_lock(&rms_lock);
+
+    if ((program = findProgram(id)) != NULL)
+    {
+       if (program->nprocs > 0)
+       {
+           KMEM_ALLOC(buf, pid_t *, program->nprocs * sizeof(pid_t), TRUE);
+           if (buf)
+           {
+               for (pidp = buf, pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+                   *pidp++ = pdesc->proc->p_pid;
+               
+               if (copyout(buf, pids, sizeof(pid_t) * MIN(program->nprocs, maxpids)))
+                   status = EFAULT;
+               
+               KMEM_FREE(buf, program->nprocs * sizeof(pid_t));
+           }
+           else
+               status = ENOMEM;
+       }
+       
+       if (copyout(&program->nprocs, nprocs, sizeof(int)))
+           status = EFAULT;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    
+    return(status);
+}
+
+/*
+ * rmsmod always used to use psignal but this doesn't work
+ * on Linux 2.6.7 so we have changed to kill_proc
+ */
+static void prgsignal(struct prg_desc *program, int signo)
+{
+    struct proc_desc *pdesc;
+    for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+       kill_proc(pdesc->proc->p_pid, signo, 1);
+}
+
+
+int rms_prgsignal(int id, int signo)
+{
+    struct prg_desc *program;
+    int status = ESUCCESS;
+    
+    kmutex_lock(&rms_lock);
+    
+    if ((program = findProgram(id)) != NULL)
+    {
+       if (CURUID() == 0 || CURUID() == program->uid)
+       {
+           prgsignal(program, signo);
+           if (signo == SIGKILL)
+               program->flags |= PRG_KILLED;
+       }
+       else
+           status = EACCES;
+    }
+    else
+       status = ESRCH;
+    
+    kmutex_unlock(&rms_lock);
+    
+    return(status);
+}
+
+int rms_prgaddcap(int id, int index, ELAN_CAPABILITY *cap)
+{
+    struct prg_desc *program;
+    struct cap_desc *cdesc;
+    int status = ESUCCESS;
+
+    if (cap == NULL)
+        return(EINVAL);
+
+    kmutex_lock(&rms_lock);
+    if ((program = findProgram(id)) != NULL)
+    {
+       KMEM_ALLOC(cdesc, struct cap_desc *, sizeof(struct cap_desc), TRUE);
+       if (cdesc)
+       {
+           cdesc->index = index;
+           if (copyin(cap, &cdesc->cap, sizeof(ELAN_CAPABILITY)))
+           {
+               KMEM_FREE(cdesc, sizeof(struct cap_desc));
+               status = EFAULT;
+           }
+           else
+           {
+               DBG(printk("rms_prgaddcap :: program %d index %d context %d<-->%d\n",
+                          program->id, index, cdesc->cap.cap_lowcontext, cdesc->cap.cap_highcontext));
+               cdesc->next = program->caps;
+               program->caps = cdesc;
+               program->ncaps++;
+           }
+       }
+       else
+           status = ENOMEM;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+static uint64_t gettime(void)
+{
+    uint64_t now;
+
+#if defined(SOLARIS)
+    timespec_t tv;
+    gethrestime(&tv);
+    now = tv.tv_sec * 1000 + tv.tv_nsec / 1000000;
+#elif defined(LINUX)
+    struct timeval tv;
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,4,17)
+    get_fast_time(&tv);
+#else
+    do_gettimeofday(&tv);
+#endif
+    now = tv.tv_sec * 1000 + tv.tv_usec / 1000;
+#elif defined(DIGITAL_UNIX)
+    struct timeval tv;
+    microtime(&tv);
+    now = tv.tv_sec * 1000 + tv.tv_usec / 1000;
+#endif
+
+    return(now);
+}
+
+#ifdef DIGITAL_UNIX
+
+int rms_getrusage(struct proc_desc *pdesc, struct rusage *ru)
+{
+    task_t   task;
+    thread_t thread;
+    
+    if (!pdesc->proc)
+       return(-1);
+    
+    /*
+     * locking required unless called from the current proc
+     */
+    if (pdesc->proc != CURPROC())
+    {
+       if (!P_REF(pdesc->proc))
+           return(-1);
+       
+       task = proc_to_task(pdesc->proc);
+       if (!task) 
+       {
+           P_UNREF(pdesc->proc);
+           DBG(printk("rms_getrusage :: process (%d) has no task\n", pdesc->proc->p_pid));
+           return(-1);
+       }
+
+       task_reference(task);
+       task_lock(task);
+       
+       if (!queue_empty(&task->thread_list))
+           thread = (thread_t) queue_first(&task->thread_list);
+       else 
+       {
+           task_unlock(task);
+           task_deallocate(task);
+           P_UNREF(pdesc->proc);
+           return(-1);
+       }
+       
+       thread_reference(thread);
+       task_unlock(task);
+    }
+    
+    *ru = proc_to_utask(pdesc->proc)->uu_ru;
+    task_get_rusage(ru, proc_to_task(pdesc->proc));
+    
+    if (pdesc->proc != CURPROC())
+    {
+       task_deallocate(task);
+       thread_deallocate(thread);
+       P_UNREF(pdesc->proc);
+    }
+    return(0);
+}
+
+#endif
+
+/*
+ * new stats collection interface, 64-bit with addition of Elan stats
+ */
+int rms_prggetstats(int id, prgstats_t *stats)
+{
+#ifdef DIGITAL_UNIX
+    long ruixrss, ruidrss, ruisrss, rumaxrss, rumajflt;
+#endif
+    struct prg_desc *program = 0;
+    struct proc_desc *pdesc;
+    int status = ESUCCESS;
+    prgstats_t totals;
+    uint64_t now = gettime();
+#if defined(SOLARIS)
+    clock_t utime, stime;
+#elif defined(LINUX)
+    uint64_t utime, stime;
+#endif
+
+    long maxrss;
+
+    kmutex_lock(&rms_lock);
+    
+    if (id < 0)
+    {
+       if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+           program = pdesc->program;
+    }
+    else
+       program = findProgram(id);
+
+    if (program)
+    {
+       if (CURUID() == 0 || CURUID() == program->uid)
+       {
+           totals.flags = program->flags;
+           totals.ncpus = program->ncpus;
+           maxrss = 0;
+
+           if (program->nprocs > 0)
+               totals.etime = now - program->start_time;
+           else
+               totals.etime = program->end_time - program->start_time;
+           
+           totals.atime = program->accum_atime;
+           if (program->flags & PRG_RUNNING)
+               totals.atime += program->ncpus * (now - program->sched_time);
+           
+#if defined(SOLARIS)
+           utime = stime = 0;
+           for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+           {
+               utime += pdesc->proc->p_utime;
+               stime += pdesc->proc->p_stime;
+           }
+           totals.utime = TICK_TO_MSEC(utime);
+           totals.stime = TICK_TO_MSEC(stime);
+
+#elif defined(LINUX)
+           utime = stime = 0;
+           totals.memint = program->memint;
+           totals.pageflts = program->majflt;
+
+           for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+           {
+#ifdef  PROCESS_ACCT
+       DBG(printk("rms_prggetsta :: process %d utime %ld clks stime %ld clks\n", 
+                               pdesc->proc->p_pid, TIMEVAL_TO_CT(&pdesc->proc->utime),
+                               TIMEVAL_TO_CT(&pdesc->proc->stime)));                 
+       utime += TIMEVAL_TO_CT(&pdesc->proc->utime);                  
+       stime += TIMEVAL_TO_CT(&pdesc->proc->stime);                  
+#elif LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 0)
+               DBG(printk("rms_prggetsta :: process %d utime %ld clks stime %ld clks\n", 
+                          pdesc->proc->p_pid, pdesc->proc->times.tms_utime, 
+                          pdesc->proc->times.tms_stime));
+               utime += pdesc->proc->times.tms_utime;
+               stime += pdesc->proc->times.tms_stime;
+#else
+               DBG(printk("rms_prggetsta :: process %d utime %ld clks stime %ld clks\n", 
+                          pdesc->proc->p_pid, pdesc->proc->utime, pdesc->proc->stime));
+               utime += pdesc->proc->utime;
+               stime += pdesc->proc->stime;
+#endif
+
+               totals.pageflts += pdesc->proc->maj_flt; 
+
+               maxrss += PROC_RSS(pdesc->proc) >> (20 - PAGE_SHIFT);
+           }
+
+           /* convert user and system times to millisecs */
+           totals.utime = CT_TO_MSEC(utime);
+           totals.stime = CT_TO_MSEC(stime);
+           
+#elif defined(DIGITAL_UNIX)
+           totals.utime = totals.stime = 0;
+           totals.memint = program->memint;
+           totals.pageflts = program->majflt;
+
+           for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+           {
+               struct rusage ru;
+               if (rms_getrusage(pdesc, &ru) < 0)
+                   continue;
+               
+               totals.utime += TIMEVAL_TO_MSEC(&ru.ru_utime);
+               totals.stime += TIMEVAL_TO_MSEC(&ru.ru_stime);
+               
+               /* convert maxrss to megabytes */
+               rumaxrss = ru.ru_maxrss >> 10;
+               rumajflt = ru.ru_majflt;
+               totals.pageflts += rumajflt;
+               
+               /*
+                * memory intergals are still broken in 5.1
+                */
+               
+#ifdef FIXED_MEMINIT
+               
+               /* convert from pages * clock ticks to Mbytes * secs */
+               ruixrss = (ru.ru_ixrss >> (20 - PAGE_SHIFT)) / hz;
+               ruidrss = (ru.ru_idrss >> (20 - PAGE_SHIFT)) / hz;
+               ruisrss = (ru.ru_isrss >> (20 - PAGE_SHIFT)) / hz;
+               
+               DBG(printk("rms_prggetsta :: process %d mem %d int %d %d %d flt %d\n", pdesc->proc->p_pid, 
+                          rumaxrss, ruixrss, ruidrss, ruisrss, rumajflt));
+               
+               totals.memint += ruixrss + ruidrss + ruisrss;
+#else
+               DBG(printk("rms_prggetsta :: process %d mem %d flt %d\n", pdesc->proc->p_pid, rumaxrss, rumajflt));
+               totals.memint = 0;
+#endif
+               maxrss += rumaxrss;
+           }
+#endif /* DIGITAL_UNIX */
+
+           if (maxrss > program->maxrss)
+               program->maxrss = maxrss;
+           
+           totals.utime += program->cutime;
+           totals.stime += program->cstime;
+           totals.mem = program->maxrss;
+           totals.ebytes = program->ebytes;
+           totals.exfers = program->exfers;
+
+           DBG(printk("rms_prggetsta :: program %d mem %d flt %d\n", program->id, totals.mem, totals.pageflts));
+           
+           if (copyout(&totals, stats, sizeof(prgstats_t)))
+               status = EFAULT;
+       }
+       else
+           status = EACCES;
+    }
+    else
+       status = ESRCH;
+    
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+/*
+ * preserve the old stats stats collection interface
+ */
+
+int rms_prggetoldstats(int id, prgstats_old_t *stats)
+{
+#ifdef DIGITAL_UNIX
+    long ruixrss, ruidrss, ruisrss, rumaxrss, rumajflt;
+#endif
+    struct prg_desc *program = 0;
+    struct proc_desc *pdesc;
+    int status = ESUCCESS;
+    prgstats_old_t totals;
+    uint64_t now = gettime();
+#if defined(SOLARIS) || defined(LINUX)
+    clock_t utime, stime;
+#endif
+    long maxrss;
+
+    kmutex_lock(&rms_lock);
+    
+    if (id < 0)
+    {
+       if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+           program = pdesc->program;
+    }
+    else
+       program = findProgram(id);
+
+    if (program)
+    {
+       if (CURUID() == 0 || CURUID() == program->uid)
+       {
+           totals.flags = program->flags;
+           totals.ncpus = program->ncpus;
+           maxrss = 0;
+
+           if (program->nprocs > 0)
+               totals.etime = now - program->start_time;
+           else
+               totals.etime = program->end_time - program->start_time;
+           
+           totals.atime = program->accum_atime;
+           if (program->flags & PRG_RUNNING)
+               totals.atime += program->ncpus * (now - program->sched_time);
+           
+#if defined(SOLARIS)
+           utime = stime = 0;
+           for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+           {
+               utime += pdesc->proc->p_utime;
+               stime += pdesc->proc->p_stime;
+           }
+           totals.utime = TICK_TO_MSEC(utime);
+           totals.stime = TICK_TO_MSEC(stime);
+
+#elif defined(LINUX)
+           utime = stime = 0;
+           totals.memint = program->memint;
+           totals.pageflts = program->majflt;
+
+           for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+           {
+#ifdef  PROCESS_ACCT
+       DBG(printk("rms_getoldsta :: process %d utime %ld clks stime %ld clks\n", 
+                               pdesc->proc->p_pid, TIMEVAL_TO_CT(&pdesc->proc->utime),
+                               TIMEVAL_TO_CT(&pdesc->proc->stime)));                 
+       utime += TIMEVAL_TO_CT(&pdesc->proc->utime);                  
+       stime += TIMEVAL_TO_CT(&pdesc->proc->stime);                  
+#elif LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 0)
+               DBG(printk("rms_getoldsta :: process %d utime %ld clks stime %ld clks\n", 
+                          pdesc->proc->p_pid, pdesc->proc->times.tms_utime, 
+                          pdesc->proc->times.tms_stime));
+               utime += pdesc->proc->times.tms_utime;
+               stime += pdesc->proc->times.tms_stime;
+#else
+               DBG(printk("rms_getoldsta :: process %d utime %ld clks stime %ld clks\n", 
+                          pdesc->proc->p_pid, pdesc->proc->utime, pdesc->proc->stime));
+               utime += pdesc->proc->utime;
+               stime += pdesc->proc->stime;
+#endif
+
+               totals.pageflts += pdesc->proc->maj_flt; 
+               maxrss += PROC_RSS(pdesc->proc) >> (20 - PAGE_SHIFT);
+           }
+
+           /* convert user and system times to millisecs */
+           totals.utime = CT_TO_MSEC(utime);
+           totals.stime = CT_TO_MSEC(stime);
+           
+#elif defined(DIGITAL_UNIX)
+           totals.utime = totals.stime = 0;
+           totals.memint = program->memint;
+           totals.pageflts = program->majflt;
+
+           for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+           {
+               struct rusage ru;
+               if (rms_getrusage(pdesc, &ru) < 0)
+                   continue;
+               
+               totals.utime += TIMEVAL_TO_MSEC(&ru.ru_utime);
+               totals.stime += TIMEVAL_TO_MSEC(&ru.ru_stime);
+               
+               /* convert maxrss to megabytes */
+               rumaxrss = ru.ru_maxrss >> 10;
+               rumajflt = ru.ru_majflt;
+               totals.pageflts += rumajflt;
+               
+               /*
+                * memory intergals are still broken in 5.1
+                */
+               
+#ifdef FIXED_MEMINIT
+               
+               /* convert from pages * clock ticks to Mbytes * secs */
+               ruixrss = (ru.ru_ixrss >> (20 - PAGE_SHIFT)) / hz;
+               ruidrss = (ru.ru_idrss >> (20 - PAGE_SHIFT)) / hz;
+               ruisrss = (ru.ru_isrss >> (20 - PAGE_SHIFT)) / hz;
+               
+               DBG(printk("rms_getoldsta :: process %d mem %d int %d %d %d flt %d\n", pdesc->proc->p_pid, 
+                          rumaxrss, ruixrss, ruidrss, ruisrss, rumajflt));
+               
+               totals.memint += ruixrss + ruidrss + ruisrss;
+#else
+               DBG(printk("rms_getoldsta :: process %d mem %d flt %d\n", pdesc->proc->p_pid, rumaxrss, rumajflt));
+               totals.memint = 0;
+#endif
+               maxrss += rumaxrss;
+           }
+#endif /* DIGITAL_UNIX */
+
+           if (maxrss > program->maxrss)
+               program->maxrss = maxrss;
+           
+           totals.utime += program->cutime;
+           totals.stime += program->cstime;
+           totals.mem = program->maxrss;
+           
+           DBG(printk("rms_getoldsta :: program %d mem %d flt %d\n", program->id, totals.mem, totals.pageflts));
+           
+           if (copyout(&totals, stats, sizeof(prgstats_old_t)))
+               status = EFAULT;
+       }
+       else
+           status = EACCES;
+    }
+    else
+       status = ESRCH;
+    
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+
+int rms_prgsuspend(int id)
+{
+    struct prg_desc *program;
+    int status = ESUCCESS;
+
+    kmutex_lock(&rms_lock);
+    
+    if ((program = findProgram(id)) != NULL)
+    {
+       if (CURUID() == 0 || CURUID() == program->uid)
+       {
+           program->flags &= ~PRG_RUNNING;
+           program->flags |=  PRG_SUSPEND;
+           program->accum_atime += program->ncpus * (gettime() - program->sched_time);
+
+           /* suspend/resume just use signals for now */
+           prgsignal(program, SIGSTOP);
+       }
+       else
+           status = EACCES;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+int rms_prgresume(int id)
+{
+    struct prg_desc *program;
+    int status = ESUCCESS;
+
+    kmutex_lock(&rms_lock);
+    
+    if ((program = findProgram(id)) != NULL)
+    {
+       if (CURUID() == 0 || CURUID() == program->uid)
+       {
+           program->flags &= ~PRG_SUSPEND;
+           program->flags |=  PRG_RUNNING;
+           program->sched_time = gettime();
+           prgsignal(program, SIGCONT);
+       }
+       else
+           status = EACCES;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+
+int rms_ncaps(int *ncaps)
+{
+    struct proc_desc *pdesc;
+    int status = ESUCCESS;
+    
+    kmutex_lock(&rms_lock);
+    if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+    {
+       if (copyout(&pdesc->program->ncaps, ncaps, sizeof(int)))
+           status = EFAULT;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+int rms_getprgid(pid_t pid, int *id)
+{
+    struct proc_desc *pdesc;
+    int status = ESUCCESS;
+    
+    if (pid == 0)
+       pid = CURPROC()->p_pid;
+    
+    kmutex_lock(&rms_lock);
+    if ((pdesc = findProcess(pid)) != NULL)
+    {
+       if (copyout(&pdesc->program->id, id, sizeof(int)))
+           status = EFAULT;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+int rms_setcap(int index, int ctx)
+{
+    struct proc_desc *pdesc;
+    struct cap_desc *cdesc;
+    int status = EINVAL;
+    
+    DBG(printk("rms_setcap    :: process %d cap %d ctx %d\n",CURPROC()->p_pid,index,ctx));
+
+    kmutex_lock(&rms_lock);
+    if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+    {
+       for (cdesc = pdesc->program->caps; cdesc; cdesc = cdesc->next)
+           if (cdesc->index == index && 0 <= ctx && ctx <= (cdesc->cap.cap_highcontext - cdesc->cap.cap_lowcontext + 1))
+           {
+               pdesc->mycap = index;
+               pdesc->myctx = cdesc->cap.cap_lowcontext + ctx;
+               status = ESUCCESS;
+           }
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+
+int rms_mycap(int *index)
+{
+    struct proc_desc *pdesc;
+    int status = ESUCCESS;
+    
+    DBG(printk("rms_mycap :: process %d\n", CURPROC()->p_pid));
+    
+    kmutex_lock(&rms_lock);
+    if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+    {
+       DBG(printk("rms_mycap :: found process %d mycap = %d\n", CURPROC()->p_pid, pdesc->mycap));
+       if (copyout(&pdesc->mycap, index, sizeof(int)))
+           status = EFAULT;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+int rms_getcap(int index, ELAN_CAPABILITY *cap)
+{
+    struct proc_desc *pdesc;
+    struct cap_desc *cdesc;
+    int status = ESUCCESS;
+    
+    kmutex_lock(&rms_lock);
+    if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+    {
+       for (cdesc = pdesc->program->caps; cdesc; cdesc = cdesc->next)
+           if (cdesc->index == index)
+               break;
+       
+       if (cdesc)
+       {
+           /* tell each process about its own context */
+           cdesc->cap.cap_mycontext = pdesc->myctx;
+           
+           if (copyout(&cdesc->cap, cap, ELAN_CAP_SIZE(&cdesc->cap)))
+               status = EFAULT;
+           
+           DBG(printk("rms_getcap    :: program %d index %d context %d<-->%d\n", pdesc->program->id, 
+                      cdesc->index, cdesc->cap.cap_lowcontext, cdesc->cap.cap_highcontext));
+       }
+       else
+           status = EINVAL;
+    }
+    else
+       status = ESRCH;
+    
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+
+static int
+rms_fork_callback (struct PROC_STRUCT *curproc, struct PROC_STRUCT *child)
+{
+    struct prg_desc *program;
+    struct proc_desc *parent;
+    struct proc_desc *pdesc = NULL;
+
+    kmutex_lock(&rms_lock);
+    
+    DBG(printk("rms_fork_func :: phase is fork pid %d child %d\n", curproc->p_pid, child->p_pid));
+
+    /*
+     * find the process that forked
+     */
+    if ((parent = findProcess(curproc->p_pid)) != NULL)
+    {
+       program = parent->program;
+       
+       DBG(printk("rms_fork_func :: program is %d flags %d\n", program->id, program->flags));
+       
+       /*
+        * processes can be blocked in fork while prgsignal is in progress
+        * so check to see if the PRG_KILLED flag is set
+        */
+       if (program->flags & PRG_KILLED)
+           DBG(printk("rms_fork_func :: fork handler called after program killed\n"));
+       else
+       {
+           /*
+            * create a new process description and add to program
+            */
+           KMEM_ALLOC(pdesc, struct proc_desc *, sizeof(struct proc_desc), TRUE);
+           if (pdesc)
+           {
+               pdesc->next = program->pdescs;
+               program->pdescs = pdesc;
+               pdesc->proc = child;
+               pdesc->mycap = parent->mycap;
+               pdesc->myctx = parent->myctx;
+               pdesc->program = program;
+               pdesc->vp = -1;              /* assigned by elaninitdone */
+               program->nprocs++;
+           }
+           else
+               printk("rms_fork_func :: memory allocation failed\n");
+       }
+    }
+    else
+       DBG(printk("rms_fork_func :: no program\n"));
+    
+    kmutex_unlock (&rms_lock);
+
+    return pdesc == NULL;
+}
+
+static void
+rms_exit_callback (struct PROC_STRUCT *curproc)
+{
+    struct prg_desc *program;
+    struct proc_desc *pdesc, **pdescp, *p;
+#ifdef DIGITAL_UNIX
+    struct rusage ru;
+#endif
+    long maxrss;
+
+    kmutex_lock(&rms_lock);
+    
+    DBG(printk("rms_exit_func :: process %d exiting\n", curproc->p_pid));
+
+    /*
+     * find the process that exited and accumulate 
+     * resource usage in its parent program
+     */
+    for (program = programs, pdesc = 0; program && !pdesc; program = program->next)
+    {
+       pdescp = &program->pdescs;
+       while ((pdesc = *pdescp) != NULL)
+       {
+           if (pdesc->proc->p_pid == curproc->p_pid)
+           {
+               /*
+                * keep track of the resources used
+                */
+#if defined(SOLARIS)
+               program->cutime += TICK_TO_MSEC(pdesc->proc->p_utime);
+               program->cstime += TICK_TO_MSEC(pdesc->proc->p_stime);
+               
+#elif defined(LINUX)
+#ifdef PROCESS_ACCT
+       DBG(printk("rms_exit_func :: process %d exit utime %ld clks stime %ld clks\n",
+                               pdesc->proc->p_pid,
+                               TIMEVAL_TO_CT(&pdesc->proc->utime),                
+                               TIMEVAL_TO_CT(&pdesc->proc->stime)));              
+       program->cutime += TIMEVAL_TO_MSEC(&pdesc->proc->utime);      
+       program->cstime += TIMEVAL_TO_MSEC(&pdesc->proc->stime);      
+#elif LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 0)                           
+               DBG(printk("rms_exit_func :: process %d exit utime %ld clks stime %ld clks\n", 
+                          pdesc->proc->p_pid, pdesc->proc->times.tms_utime, 
+                          pdesc->proc->times.tms_stime));
+               
+               program->cutime += CT_TO_MSEC(pdesc->proc->times.tms_utime);
+               program->cstime += CT_TO_MSEC(pdesc->proc->times.tms_stime);
+#else
+               DBG(printk("rms_exit_func :: process %d exit utime %ld clks stime %ld clks\n", 
+                          pdesc->proc->p_pid, pdesc->proc->utime, pdesc->proc->stime));
+               
+               program->cutime += CT_TO_MSEC(pdesc->proc->utime);
+               program->cstime += CT_TO_MSEC(pdesc->proc->stime);
+#endif
+               program->majflt += pdesc->proc->maj_flt;
+               maxrss = PROC_RSS(pdesc->proc) >> (20 - PAGE_SHIFT);
+               
+#elif defined(DIGITAL_UNIX)
+               if (rms_getrusage(pdesc, &ru) == 0)
+               {
+                   program->cutime += TIMEVAL_TO_MSEC(&ru.ru_utime);
+                   program->cstime += TIMEVAL_TO_MSEC(&ru.ru_stime);
+                   program->majflt += ru.ru_majflt;
+                   
+                   /* convert maxrss to megabytes */
+                   maxrss = ru.ru_maxrss >> 10;
+               }
+#endif
+               
+               /*
+                * shared memory segment cleanup
+                */
+#if defined(DIGITAL_UNIX)
+               rms_shmcleanup(-1);
+#elif defined(LINUX)
+               shm_cleanup();
+#endif 
+               
+               /* 
+                * remove process from program 
+                */
+               *pdescp = pdesc->next;
+               KMEM_FREE(pdesc, sizeof(struct proc_desc));
+               program->nprocs--;
+               
+               /*
+                * update the memory high water mark for the program
+                */
+               for (p = program->pdescs; p; p = p->next)
+               {
+#if defined(DIGITAL_UNIX)
+                   if (rms_getrusage(p, &ru) < 0)
+                       continue;
+                   
+                   /* convert maxrss to megabytes */
+                   maxrss += ru.ru_maxrss >> 10;
+                   
+#elif defined(LINUX)                   
+                   maxrss += PROC_RSS(p->proc) >> (20 - PAGE_SHIFT);
+#endif
+               }
+               if (maxrss > program->maxrss)
+                   program->maxrss = maxrss;
+               
+               DBG(printk("rms_exit_func :: program %d procs %d mem %ld\n", program->id, program->nprocs, program->maxrss));
+               
+               /*
+                * final update to the program if this is the last process
+                */
+               if (program->nprocs == 0)
+               {
+                   program->end_time = gettime();
+                   program->flags &= ~PRG_RUNNING;
+                   program->accum_atime += program->ncpus * (program->end_time - program->sched_time);
+                   DBG(printk("rms_exit_func :: last process has gone\n"));
+               }
+               break;
+           }
+           else
+               pdescp = &pdesc->next;
+       }
+    }
+    kmutex_unlock  (&rms_lock);
+}
+
+#if defined(LINUX)
+static int
+rms_ptrack_callback (void *arg, int phase, struct task_struct *child)
+{
+    switch (phase)
+    {
+    case PTRACK_PHASE_CLONE:
+       if (rms_fork_callback (current, child))
+           return PTRACK_DENIED;
+       else
+           return PTRACK_INNHERIT;
+
+    case PTRACK_PHASE_CLONE_FAIL:
+       DBG(printk("rms_fork_func :: fork failed pid %d child %d\n", current->p_pid, child->p_pid));
+       rms_exit_callback(child);
+       break;
+
+    case PTRACK_PHASE_EXIT:
+       rms_exit_callback(current);
+       break;
+    }
+    return PTRACK_FINISHED;
+}
+
+#else
+
+static void
+rms_xa_callback (void *arg, int phase, void *ctask)
+{
+    switch (phase)
+    {
+    case XA_FORK:
+       if (rms_fork_callback (CURPROC(), (struct PROC_STRUCT *)task_to_proc(ctask)))
+           psignal(task_to_proc(ctask), SIGKILL);
+       break;
+    case XA_EXIT:
+       rms_exit_callback (CURPROC());
+       break;
+    }
+}
+
+#endif
+
+#ifdef DIGITAL_UNIX
+
+/*
+ * NB: These functions will only work on steelos.
+ */
+
+/*
+ * rms_setcorepath
+ *
+ * set a path at which to dump core if the task aborts  
+ *
+ * enhanced core file names must be enabled for this to work
+ */
+int rms_setcorepath(char *corepath)
+{
+    int    length;
+    char  *path;
+    int    status; 
+    struct proc_desc *pdesc;
+    
+    /* 
+     * access restricted - we don't want users moving
+     * their corepath and generating a huge I/O load
+     */
+    if (CURUID())
+       return(EACCES);
+    
+    if (!(pdesc = findProcess(CURPROC()->p_pid)))
+       return(ESRCH);
+    
+    if (pdesc->program->corepath)
+       return(EEXIST);
+    
+    KMEM_ALLOC(path, char *, MAXCOREPATHLEN + 1, TRUE);
+    if (path == 0)
+       return(ENOMEM);
+    
+    if (copyinstr(corepath, path, MAXCOREPATHLEN, &length))
+       return(EFAULT);
+    
+    path[length] = 0;
+    status = add_corepath(path);
+    
+    DBG(printk("rms_setcorepa :: id %d corepath %s status %d\n", pdesc->program->id, path, status));
+    
+    if (status == ESUCCESS)
+       pdesc->program->corepath = path;
+    else
+       KMEM_FREE(path, MAXCOREPATHLEN + 1);
+    
+    return(status);
+}
+
+static int find_corepath(pid_t pid, char *path, int len)
+{
+    struct proc *procp;
+    struct utask *utask;
+    int status = ESUCCESS;
+
+    procp = pfind(pid);
+    if (procp == NULL)
+        return(ENOENT);
+    
+    utask = proc_to_utask(procp);
+    
+    if (utask->uu_coredir)
+        bcopy(utask->uu_coredir,path,len);
+    else
+        status = ENOENT;
+    
+    /* pfind takes out a reference */
+    P_UNREF(procp);
+
+    return(status);
+}
+
+int rms_getcorepath(pid_t pid, char *corepath, int maxlen)
+{
+    char src[MAXCOREPATHLEN];
+    int len;
+    int status;
+    
+    if (maxlen < 2)
+       return(EINVAL);
+    
+    len = MIN(maxlen, MAXCOREPATHLEN);
+    
+    status = find_corepath(pid, src, len);
+    
+    if (status == ESUCCESS)
+        len = strlen(src)+1;
+    else if (status == ENOENT) 
+    {
+       len = 2;
+       src[0] = '.';
+        src[1] = '\0';
+        status = ESUCCESS;
+    }
+    
+    if (copyout(src, corepath, len))
+       return(EFAULT);
+    
+    return(status);
+}
+
+#endif
+
+/*
+ * rms_elaninitdone - mark a process as having successfully completed elan initialisation
+ */
+int rms_elaninitdone(int vp)
+{
+    int status = ESUCCESS;
+    struct proc_desc *pdesc;
+    
+    DBG(printk("rms_elaninit  :: process %d vp %d\n", CURPROC()->p_pid, vp));
+    
+    kmutex_lock(&rms_lock);
+    if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+       pdesc->vp = vp;
+    else
+       status = ESRCH;
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+
+/*
+ * rms_prgelanpids - return the ids of processes that have completed elan initialisation
+ */
+int rms_prgelanpids(int id, int maxpids, int *vps, pid_t *pids, int *npids)
+{
+    struct prg_desc *program;
+    struct proc_desc *pdesc;
+    pid_t *pidbuf;
+    int status = ESUCCESS, count = 0, *vpbuf;
+    
+    DBG(printk("rms_elanpids  :: process %d id %d\n", CURPROC()->p_pid, id));
+    
+    kmutex_lock(&rms_lock);
+    
+    if ((program = findProgram(id)) != NULL)
+    {
+       if (program->nprocs > 0)
+       {
+           KMEM_ALLOC(pidbuf, pid_t *, program->nprocs * sizeof(pid_t), TRUE);
+           KMEM_ALLOC(vpbuf, int *, program->nprocs * sizeof(int), TRUE);
+           if (pidbuf && vpbuf)
+           {
+               for (pdesc = program->pdescs; pdesc; pdesc = pdesc->next)
+                   if (pdesc->vp >= 0)
+                   {
+                       pidbuf[count] = pdesc->proc->p_pid;
+                       vpbuf[count] = pdesc->vp;
+                       count++;
+                   }
+           
+               if (count > 0 && (copyout(pidbuf, pids, sizeof(pid_t) * MIN(count, maxpids)) ||
+                                 copyout(vpbuf, vps, sizeof(int) * MIN(count, maxpids))))
+                   status = EFAULT;
+               
+               KMEM_FREE(pidbuf, program->nprocs * sizeof(pid_t));
+               KMEM_FREE(vpbuf, program->nprocs * sizeof(int));
+           }
+           else
+               status = ENOMEM;
+       }
+
+       if (copyout(&count, npids, sizeof(int)))
+           status = EFAULT;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    
+    return(status);
+
+}
+
+int rms_setpset(int psid)
+{
+    struct prg_desc *program;
+    struct proc_desc *pdesc;
+    int status = ESUCCESS;
+
+    if (CURUID())
+       return(EACCES);
+
+    kmutex_lock(&rms_lock);
+    
+    if ((pdesc = findProcess(CURPROC()->p_pid)) != NULL)
+    {
+       program = pdesc->program;
+       program->psid = psid;
+    }
+    else
+       status = ESRCH;
+
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+
+int rms_getpset(int id, int *psid)
+{
+    struct prg_desc *program;
+    int status = ESUCCESS;
+    
+    kmutex_lock(&rms_lock);
+    if ((program = findProgram(id)) != NULL)
+    {
+       if (copyout(&program->psid, psid, sizeof(int)))
+           status = EFAULT;
+    }
+    else
+       status = ESRCH;
+    
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+int
+rms_setelanstats(int id, uint64_t ebytes, uint64_t exfers)
+{
+    struct prg_desc *program;
+    int status = ESUCCESS;
+    
+    DBG(printk("rms_setelanst :: process %d id %d\n", CURPROC()->p_pid, id));
+
+    kmutex_lock(&rms_lock);
+    if ((program = findProgram(id)) != NULL)
+    {
+       if (CURUID() == 0 || CURUID() == program->uid)
+       {
+           program->ebytes = ebytes;
+           program->exfers = exfers;
+       }
+       else
+           status = EACCES;
+    }
+    else
+       status = ESRCH;
+    
+    kmutex_unlock(&rms_lock);
+    return(status);
+}
+
+rms_modversion()
+{
+    return(RMS_MODVERSION);
+}
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+
+
+
+
+
+
+
Index: linux-2.4.21/drivers/net/qsnet/rms/rms_kern_Linux.c
===================================================================
--- linux-2.4.21.orig/drivers/net/qsnet/rms/rms_kern_Linux.c    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/drivers/net/qsnet/rms/rms_kern_Linux.c 2005-06-01 23:12:54.703423296 -0400
@@ -0,0 +1,430 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: rms_kern_Linux.c,v 1.20 2004/05/14 08:55:57 duncan Exp $"
+/*      $Source: /cvs/master/quadrics/rmsmod/rms_kern_Linux.c,v $*/
+
+#include <qsnet/kernel.h>
+
+#include <linux/sysctl.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/proc_fs.h>
+
+#include <rms/rmscall.h>
+#include <rms/rmsio.h>
+
+MODULE_AUTHOR("Quadrics Ltd");
+MODULE_DESCRIPTION("RMS support module");
+MODULE_LICENSE("GPL");
+
+int rms_debug = 0;
+
+ctl_table rms_table[] = {
+    {
+       .ctl_name = 1,
+       .procname = "rms_debug",
+       .data     = &rms_debug,
+       .maxlen   = sizeof(int),
+       .mode     = 0644,
+       .child    = NULL,
+       .proc_handler = &proc_dointvec,
+    },
+    {0}
+};
+
+ctl_table rms_root_table[] = {
+    {
+       .ctl_name = CTL_DEBUG,
+       .procname = "rms",
+       .data     = NULL,
+       .maxlen   = 0,
+       .mode     = 0555,
+       .child    = rms_table,
+    },
+    {0}
+};
+
+static struct ctl_table_header *rms_sysctl_header;
+
+static int rms_open (struct inode *ino, struct file *fp);
+static int rms_release (struct inode *ino, struct file *fp);
+static int rms_ioctl (struct inode *inode, struct file *fp, unsigned int cmd, unsigned long arg);
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+static int
+rms_ioctl32_cmds[] =
+{
+    RMSIO_GETPRGID32,
+    RMSIO_GETCAP32
+};
+
+static int      rms_ioctl32 (unsigned int fd, unsigned int cmd, 
+                            unsigned long arg, struct file *file);
+#endif
+
+static struct file_operations rms_fops =
+{
+    .owner   = THIS_MODULE,
+    .ioctl   = rms_ioctl,
+    .open    = rms_open,
+    .release = rms_release,
+};
+
+struct proc_dir_entry *rms_procfs_programs;
+static struct proc_dir_entry *rms_procfs_root;
+
+int version_callback(char* page, char** start, off_t off, int count, int* eof, void* data)
+{
+    return(sprintf(page, "$Id: rms_kern_Linux.c,v 1.20 2004/05/14 08:55:57 duncan Exp $\n"));
+}
+
+static int __init rms_start(void)
+{
+    struct proc_dir_entry *p;
+    int res;
+
+    if ((rms_sysctl_header = register_sysctl_table(rms_root_table, 1)) == 0)
+    {
+       printk ("rmsmod: failed to register sysctl table\n");
+       return (-ENXIO);
+    }
+    
+    if ((rms_procfs_root = proc_mkdir("rms",  NULL)) == NULL ||
+       (rms_procfs_programs = proc_mkdir("programs",  rms_procfs_root)) == NULL ||
+       (p = create_proc_entry ("control", S_IRUGO, rms_procfs_root)) == NULL)
+    {
+       unregister_sysctl_table (rms_sysctl_header);
+       printk ("rmsmod: failed to register /proc/rms\n");
+       return (-ENXIO);
+    }
+    p->proc_fops = &rms_fops;
+    p->owner     = THIS_MODULE;
+    p->data      = NULL;
+
+    if ((p = create_proc_entry ("version", S_IRUGO, rms_procfs_root)) != NULL)
+    {
+       p->owner = THIS_MODULE;
+       p->data = NULL;
+       p->read_proc = version_callback;
+    }
+
+    if ((res = rms_init()) != ESUCCESS)
+    {
+       remove_proc_entry ("programs", rms_procfs_root);
+       remove_proc_entry ("control", rms_procfs_root);
+       remove_proc_entry ("rms", NULL);
+       unregister_sysctl_table (rms_sysctl_header);
+       return (-res);
+    }
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+    lock_kernel();
+    {
+       extern int register_ioctl32_conversion(unsigned int cmd, int (*handler)(unsigned int, unsigned int, unsigned long, struct file *));
+       register int i;
+       for (i = 0; i < sizeof (rms_ioctl32_cmds)/sizeof(rms_ioctl32_cmds[0]); i++)
+           register_ioctl32_conversion (rms_ioctl32_cmds[i], rms_ioctl32);
+    }
+    unlock_kernel();
+#endif
+    return (0);
+}
+
+static void __exit rms_exit(void)
+{
+    rms_fini();
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+    lock_kernel();
+    {
+       extern void unregister_ioctl32_conversion(unsigned int cmd);
+       register int i;
+
+       for (i = 0; i < sizeof (rms_ioctl32_cmds)/sizeof(rms_ioctl32_cmds[0]); i++)
+           unregister_ioctl32_conversion (rms_ioctl32_cmds[i]);
+    }
+    unlock_kernel();
+#endif
+
+    remove_proc_entry ("version", rms_procfs_root);
+    remove_proc_entry ("programs", rms_procfs_root);
+    remove_proc_entry ("control", rms_procfs_root);
+    remove_proc_entry ("rms", NULL);
+    unregister_sysctl_table(rms_sysctl_header);
+}
+
+/* Declare the module init and exit functions */
+module_init(rms_start);
+module_exit(rms_exit);
+
+static int
+rms_open (struct inode *inode, struct file *fp)
+{
+    MOD_INC_USE_COUNT;
+    fp->private_data = NULL;
+
+    return (0);
+}
+
+static int
+rms_release (struct inode *inode, struct file *fp)
+{
+    MOD_DEC_USE_COUNT;
+    return (0);
+}
+
+static int 
+rms_ioctl(struct inode *inode, struct file *fp, unsigned int cmd, unsigned long arg)
+{
+    int res;
+
+    switch (cmd) 
+    {
+/* no corepath support in Linux yet */
+#if 0
+    case RMSIO_SETCOREPATH:
+       res = rms_setcorepath((caddr_t)arg);
+       break;
+       
+    case RMSIO_GETCOREPATH:
+    {
+       RMSIO_GETCOREPATH_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_getcorepath(args.pid, args.corepath, args.maxlen);
+       break;
+    }
+#endif
+       
+    case RMSIO_PRGCREATE:
+    {
+       RMSIO_PRGCREATE_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prgcreate(args.id, args.uid, args.cpus);
+       break;
+    }
+
+    case RMSIO_PRGDESTROY:
+       res = rms_prgdestroy(arg);
+       break;
+       
+    case RMSIO_PRGIDS:
+    {
+       RMSIO_PRGIDS_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prgids(args.maxids, args.prgids, args.nprgs);
+       break;
+    }
+
+    case RMSIO_PRGINFO:
+    {
+       RMSIO_PRGINFO_STRUCT args;
+       
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prginfo(args.id, args.maxpids, args.pids, args.nprocs);
+       break;
+    }
+       
+    case RMSIO_PRGSIGNAL:
+    {
+       RMSIO_PRGSIGNAL_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prgsignal(args.id, args.signo);
+       break;
+    }
+       
+    case RMSIO_PRGADDCAP:
+    {
+       RMSIO_PRGADDCAP_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prgaddcap(args.id, args.index, args.cap);
+       break;
+    }
+
+    case RMSIO_SETCAP:
+    {
+       RMSIO_SETCAP_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_setcap(args.index, args.ctx);
+       break;
+    }
+       
+    case RMSIO_NCAPS:
+       res = rms_ncaps((int *)arg);
+       break;
+       
+    case RMSIO_GETPRGID:
+    {
+       RMSIO_GETPRGID_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_getprgid(args.pid, args.id);
+       break;
+    }
+
+    case RMSIO_GETMYCAP:
+       res = rms_mycap((int *)arg);
+       break;
+       
+    case RMSIO_GETCAP:
+    {
+       RMSIO_GETCAP_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_getcap(args.index, args.cap);
+       break;
+    }
+
+    case RMSIO_PRGGETSTATS:
+    {
+       RMSIO_PRGGETSTATS_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prggetoldstats(args.id, args.stats);
+       break;
+    }
+
+    case RMSIO_PRGGETSTATS2:
+    {
+       RMSIO_PRGGETSTATS2_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prggetstats(args.id, args.stats);
+       break;
+    }
+
+    case RMSIO_PRGSUSPEND:
+       res = rms_prgsuspend(arg);
+       break;
+       
+    case RMSIO_PRGRESUME:
+       res = rms_prgresume(arg);
+       break;
+
+    case RMSIO_ELANINITDONE:
+       res = rms_elaninitdone(arg);
+       break;
+
+    case RMSIO_PRGELANPIDS:
+    {
+       RMSIO_PRGELANPIDS_STRUCT args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_prgelanpids(args.id, args.maxpids, args.vps, args.pids, args.npids);
+       break;
+    }
+
+    case RMSIO_SETELANSTATS:
+    {
+       RMSIO_SETELANSTATS_STRUCT args;
+       elanstats_t estats;
+
+       if (copy_from_user(&args, (void *)arg, sizeof(args)) ||
+           copy_from_user(&estats, (void *)args.estats, sizeof(estats)))
+           return(-EFAULT);
+       
+       res = rms_setelanstats(args.id, estats.ebytes, estats.exfers);
+       break;
+    }
+
+    case RMSIO_MODVERSION:
+    {
+       RMSIO_MODVERSION_STRUCT args;
+       int version = rms_modversion();
+       
+       if (copy_from_user (&args, (void *)arg, sizeof (args)))
+           return (-EFAULT);
+       
+       if (copyout(&version, args.version, sizeof(int)))
+           res = EFAULT;
+       else
+           res = ESUCCESS;
+
+       break;
+    }
+
+    default:
+       res = EINVAL;
+       break;
+    }
+
+    return ((res == 0) ? 0 : -res);
+}
+
+#if defined(CONFIG_PPC64) || defined(CONFIG_SPARC64) || defined(CONFIG_X86_64)
+static int
+rms_ioctl32 (unsigned int fd, unsigned int cmd, unsigned long arg, struct file *file)
+{
+    int res;
+
+    switch (cmd)
+    {
+    case RMSIO_GETPRGID32:
+    {
+       RMSIO_GETPRGID_STRUCT32 args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_getprgid(args.pid, (int *)(unsigned long) args.idptr);
+       break;
+    }
+       
+    case RMSIO_GETCAP32:
+    {
+       RMSIO_GETCAP_STRUCT32 args;
+
+       if (copy_from_user (&args, (void *) arg, sizeof (args)))
+           return (-EFAULT);
+
+       res = rms_getcap(args.index, (ELAN_CAPABILITY *)(unsigned long) args.capptr);
+       break;
+    }
+
+    default:
+       return (sys_ioctl (fd, cmd, arg));
+    }
+
+    return ((res == 0) ? 0 : -res);
+}
+#endif
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/fs/exec.c
===================================================================
--- linux-2.4.21.orig/fs/exec.c 2005-06-01 22:58:09.044063984 -0400
+++ linux-2.4.21/fs/exec.c      2005-06-01 23:12:54.704423144 -0400
@@ -51,6 +51,7 @@
 #ifdef CONFIG_KMOD
 #include <linux/kmod.h>
 #endif
+#include <linux/ptrack.h>
 
 int core_uses_pid;
 char core_pattern[65] = "core";
@@ -1125,6 +1126,10 @@
        if (retval < 0) 
                goto out; 
 
+
+       /* Notify any ptrack callbacks of the process exec */
+       ptrack_call_callbacks (PTRACK_PHASE_EXEC, NULL);
+       
        retval = search_binary_handler(&bprm,regs);
        if (retval >= 0)
                /* execve success */
Index: linux-2.4.21/include/elan/bitmap.h
===================================================================
--- linux-2.4.21.orig/include/elan/bitmap.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/bitmap.h  2005-06-01 23:12:54.704423144 -0400
@@ -0,0 +1,74 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __QSNET_BITMAP_H
+#define __QSNET_BITMAP_H
+
+#ident "$Id: bitmap.h,v 1.5 2004/01/20 17:32:15 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/bitmap.h,v $ */
+
+typedef unsigned int                   bitmap_t;
+
+#define BT_NBIPUL                      32                      /* n bits per bitmap_t */
+#define BT_ULSHIFT                     5                       /* log 2 BT_NBIPUL to extract word index */
+#define BT_ULMASK                      0x1f                    /* to extract bit index */
+
+#define BT_WIM(bitmap,bitindex)                ((bitmap)[(bitindex) >> BT_ULSHIFT])            /* word in map */
+#define BT_BIW(bitindex)               (1 << ((bitindex) & BT_ULMASK))         /* bit in word */
+
+/* BT_BITOUL -- n bits to n words */
+#define BT_BITOUL(nbits)               (((nbits) + BT_NBIPUL -1) / BT_NBIPUL)
+
+#define BT_TEST(bitmap,bitindex)       ((BT_WIM((bitmap), (bitindex)) & BT_BIW(bitindex)) ? 1 : 0)
+#define BT_SET(bitmap,bitindex)                do { BT_WIM((bitmap), (bitindex)) |= BT_BIW(bitindex); } while (0)
+#define BT_CLEAR(bitmap,bitindex)      do { BT_WIM((bitmap), (bitindex)) &= ~BT_BIW(bitindex); } while (0)
+
+/* return first free bit in the bitmap, or -1 for failure */
+extern int  bt_freebit (bitmap_t *bitmap, int nbits);
+
+/* return the index of the lowest set bit in the bitmap or -1 for failure */
+extern int bt_lowbit (bitmap_t *bitmap, int nbits);
+
+/* return the index of the next set/clear bit in the bitmap or -1 for failure */
+extern int bt_nextbit (bitmap_t *bitmap, int nbits, int last, int isset);
+
+/* copy/zero/fill/compare a bit map */
+extern void bt_copy (bitmap_t *a, bitmap_t *b, int nbits);
+extern void bt_zero (bitmap_t *a, int nbits);
+extern void bt_fill (bitmap_t *a, int nbits);
+extern int  bt_cmp (bitmap_t *a, bitmap_t *b, int nbits);
+
+/* intersect bitmap 'a' with bitmap 'b' and return in 'a' */
+extern void bt_intersect (bitmap_t *a, bitmap_t *b, int nbits);
+
+/* remove/add bitmap 'b' from bitmap 'a' */
+extern void bt_remove (bitmap_t *a, bitmap_t *b, int nbits);
+extern void bt_add (bitmap_t *a, bitmap_t *b, int nbits);
+
+/* check whether bitmap 'a' spans bitmap 'b' */
+extern int  bt_spans (bitmap_t *a, bitmap_t *b, int nbits);
+
+/* copy [base,base+nbits-1] from 'a' to 'b' */
+extern void bt_subset (bitmap_t *a, bitmap_t *b, int base, int nbits);
+
+/* find bits clear in 'a' and set in 'b', put result in 'c' */
+extern void bt_up (bitmap_t *a, bitmap_t *b, bitmap_t *c, int nbits);
+
+/* find bits set in 'a' and clear in 'b', put result in 'c' */
+extern void bt_down (bitmap_t *a, bitmap_t *b, bitmap_t *c, int nbits);
+
+/* return number of bits set in bitmap */
+extern int  bt_nbits (bitmap_t *a, int nbits);
+
+
+#endif /* __QSNET_BITMAP_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/capability.h
===================================================================
--- linux-2.4.21.orig/include/elan/capability.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/capability.h      2005-06-01 23:12:54.705422992 -0400
@@ -0,0 +1,197 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: capability.h,v 1.16 2004/07/20 10:15:33 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/capability.h,v $*/
+
+#ifndef __ELAN_CAPABILITY_H
+#define __ELAN_CAPABILITY_H
+
+#include <elan/bitmap.h>
+
+/* Maximum number of rails */
+#define ELAN_MAX_RAILS          (31)
+/* Maximum number of virtual processes we support */
+#define ELAN_MAX_VPS           (16384)
+
+/* Number of words in a bitmap capability */
+#define ELAN_BITMAPSIZE                BT_BITOUL(ELAN_MAX_VPS)
+
+/* Guaranteed invalid values */
+#define ELAN_INVALID_PROCESS   (0x7fffffff)            /* A GUARANTEED invalid process # */
+#define ELAN_INVALID_NODE      (0xFFFF)
+#define ELAN_INVALID_CONTEXT   (0xFFFF)
+
+/* Number of values in a user key */
+#define ELAN_USERKEY_ENTRIES   4
+
+typedef void * ELAN_CAP_OWNER;
+
+/* 
+ * When used in userspace this is relative to the base of
+ * the capabality but is an absolute location for kernel space.
+ */
+typedef struct elan_location
+{
+       unsigned short loc_node;
+       unsigned short loc_context;
+} ELAN_LOCATION;
+
+typedef struct elan_userkey
+{
+       unsigned        key_values[ELAN_USERKEY_ENTRIES];
+} ELAN_USERKEY;
+
+typedef struct elan_capability
+{
+       ELAN_USERKEY    cap_userkey;                            /* User defined protection */
+
+       int             cap_version;                            /* Version number */
+       unsigned short  cap_type;                               /* Capability Type */
+       unsigned short  cap_spare;                              /* spare was cap_elan_type */
+
+       int             cap_lowcontext;                         /* low context number in block */
+       int             cap_highcontext;                        /* high context number in block */
+       int             cap_mycontext;                          /* my context number */
+    
+       int             cap_lownode;                            /* low elan id of group */
+       int             cap_highnode;                           /* high elan id of group */
+
+       unsigned int    cap_railmask;                           /* which rails this capability is valid for */
+       
+       bitmap_t        cap_bitmap[ELAN_BITMAPSIZE];            /* Bitmap of process to processor translation */
+} ELAN_CAPABILITY;
+
+#define ELAN_CAP_UNINITIALISED         (-1)
+
+#define ELAN_CAP_VERSION_NUMBER                (0x00010002)
+
+#define ELAN_CAP_NUM_NODES(cap)                ((cap)->cap_highnode - (cap)->cap_lownode + 1)
+#define ELAN_CAP_NUM_CONTEXTS(cap)     ((cap)->cap_highcontext - (cap)->cap_lowcontext + 1)
+
+/* using or defining our own MIN/MAX had confilicts with dunix so we define ELAN_ ones */
+#define ELAN_MIN(a,b)  ((a) > (b) ? (b) : (a))
+#define ELAN_MAX(a,b)  ((a) > (b) ? (a) : (b))
+#define ELAN_CAP_BITMAPSIZE(cap)       (ELAN_MAX (ELAN_MIN (ELAN_CAP_NUM_NODES(cap) * ELAN_CAP_NUM_CONTEXTS(cap), ELAN_MAX_VPS), 0))
+
+#define ELAN_CAP_SIZE(cap)             (offsetof (ELAN_CAPABILITY, cap_bitmap[BT_BITOUL(ELAN_CAP_BITMAPSIZE(cap))]))
+#define ELAN_CAP_ENTRIES(cap)           (((cap)->cap_type & ELAN_CAP_TYPE_NO_BITMAP) ? ELAN_CAP_BITMAPSIZE((cap)) : bt_nbits((cap)->cap_bitmap, ELAN_CAP_BITMAPSIZE((cap))))
+
+#define ELAN_CAP_IS_RAIL_SET(cap,rail)  ((cap)->cap_railmask & (1<<rail))
+
+#define ELAN_CAP_KEY_MATCH(cap1,cap2)  ((cap1)->cap_userkey.key_values[0] == (cap2)->cap_userkey.key_values[0] && \
+                                        (cap1)->cap_userkey.key_values[1] == (cap2)->cap_userkey.key_values[1] && \
+                                        (cap1)->cap_userkey.key_values[2] == (cap2)->cap_userkey.key_values[2] && \
+                                        (cap1)->cap_userkey.key_values[3] == (cap2)->cap_userkey.key_values[3])
+
+#define ELAN_CAP_TYPE_MATCH(cap1,cap2)  ((cap1)->cap_version           == (cap2)->cap_version           && \
+                                        (cap1)->cap_type              == (cap2)->cap_type)
+
+#define ELAN_CAP_GEOM_MATCH(cap1,cap2) ((cap1)->cap_lowcontext        == (cap2)->cap_lowcontext        && \
+                                        (cap1)->cap_highcontext       == (cap2)->cap_highcontext       && \
+                                        (cap1)->cap_lownode           == (cap2)->cap_lownode           && \
+                                        (cap1)->cap_highnode          == (cap2)->cap_highnode          && \
+                                         (cap1)->cap_railmask          == (cap2)->cap_railmask          && \
+                                        !bcmp (&(cap1)->cap_bitmap[0], &(cap2)->cap_bitmap[0],            \
+                                               BT_BITOUL(ELAN_CAP_BITMAPSIZE(cap1)*sizeof(bitmap_t))))
+
+#define ELAN_CAP_MATCH(cap1,cap2)      (ELAN_CAP_KEY_MATCH (cap1, cap2)  && \
+                                        ELAN_CAP_TYPE_MATCH (cap1, cap2) && \
+                                        ELAN_CAP_GEOM_MATCH (cap1, cap2))
+
+#define ELAN_CAP_VALID_MYCONTEXT(cap)   (    ((cap)->cap_lowcontext  != ELAN_CAP_UNINITIALISED)     \
+                                         && ((cap)->cap_mycontext   != ELAN_CAP_UNINITIALISED)     \
+                                         && ((cap)->cap_highcontext != ELAN_CAP_UNINITIALISED)     \
+                                         && ((cap)->cap_lowcontext <= (cap)->cap_mycontext)        \
+                                         && ((cap)->cap_mycontext <= (cap)->cap_highcontext)) 
+
+/*
+ * Definitions for type 
+ */
+#define ELAN_CAP_TYPE_BLOCK            1               /* Block distribution */
+#define ELAN_CAP_TYPE_CYCLIC           2               /* Cyclic distribution */
+#define ELAN_CAP_TYPE_KERNEL           3               /* Kernel capability */
+
+#define ELAN_CAP_TYPE_MASK             (0xFFF)         /* Mask for type */
+
+/* OR these bits in for extra features */
+#define ELAN_CAP_TYPE_HWTEST           (1 << 12)       /* Hardware test capability type */
+#define ELAN_CAP_TYPE_MULTI_RAIL       (1 << 13)       /* "new" multi rail capability */
+#define ELAN_CAP_TYPE_NO_BITMAP                (1 << 14)       /* don't use bit map */
+#define ELAN_CAP_TYPE_BROADCASTABLE    (1 << 15)       /* broadcastable */
+
+
+extern void          elan_nullcap     (ELAN_CAPABILITY *cap);
+extern char         *elan_capability_string (ELAN_CAPABILITY *cap, char *str);
+extern ELAN_LOCATION elan_vp2location (unsigned process, ELAN_CAPABILITY *cap);
+extern int           elan_location2vp (ELAN_LOCATION location, ELAN_CAPABILITY *cap);
+extern int           elan_nvps        (ELAN_CAPABILITY *cap);
+extern int           elan_nlocal      (int node, ELAN_CAPABILITY *cap);
+extern int           elan_maxlocal    (ELAN_CAPABILITY *cap);
+extern int           elan_localvps    (int node, ELAN_CAPABILITY *cap, int *vps, int size);
+extern int           elan_nrails      (ELAN_CAPABILITY *cap);
+extern int           elan_rails       (ELAN_CAPABILITY *cap, int *rails);
+extern int           elan_cap_overlap (ELAN_CAPABILITY *cap1, ELAN_CAPABILITY *cap2);
+
+/*
+ * capability creation/access fns provide for running
+ * new libelan code on old OS releases
+ */
+extern int elan_lowcontext(ELAN_CAPABILITY *cap);
+extern int elan_mycontext(ELAN_CAPABILITY *cap);
+extern int elan_highcontext(ELAN_CAPABILITY *cap);
+extern int elan_lownode(ELAN_CAPABILITY *cap);
+extern int elan_highnode(ELAN_CAPABILITY *cap);
+extern int elan_captype(ELAN_CAPABILITY *cap);
+extern int elan_railmask(ELAN_CAPABILITY *cap);
+
+extern int elan_getenvCap (ELAN_CAPABILITY *cap, int index);
+extern ELAN_CAPABILITY *elan_createCapability(void);
+extern ELAN_CAPABILITY *elan_copyCapability(ELAN_CAPABILITY *from, int ctxShift);
+extern int elan_generateCapability(char *string);
+
+typedef struct elan_cap_struct
+{
+       ELAN_CAP_OWNER   owner;
+       ELAN_CAPABILITY  cap;
+
+       unsigned int     attached; /* count of people attached */
+       unsigned int     active;   /* ie not being destroyed   */
+} ELAN_CAP_STRUCT;
+
+#if ! defined(__KERNEL__)
+extern void          elan_get_random_key(ELAN_USERKEY *key);
+extern int           elan_prefrails(ELAN_CAPABILITY *cap, int *pref, int nvp);
+#endif
+
+#if defined(__KERNEL__)
+/* capability.c */
+extern int elan_validate_cap  (ELAN_CAPABILITY *cap);
+extern int elan_validate_map  (ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map);
+
+extern int elan_create_cap  (ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap);
+extern int elan_destroy_cap (ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap);
+extern int elan_create_vp   (ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map);
+extern int elan_destroy_vp  (ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map);
+
+typedef        void (*ELAN_DESTROY_CB)(void *args, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map);
+
+extern int elan_attach_cap  (ELAN_CAPABILITY *cap, unsigned int rail, void *args, ELAN_DESTROY_CB callback);
+extern int elan_detach_cap  (ELAN_CAPABILITY *cap, unsigned int rail);
+
+extern int elan_get_caps    (uint *number_of_results, uint array_size, ELAN_CAP_STRUCT *caps);
+extern int elan_cap_dump    (void);
+#endif /* __KERNEL__ */
+
+
+#endif /* __ELAN_CAPABILITY_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/cm.h
===================================================================
--- linux-2.4.21.orig/include/elan/cm.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/cm.h      2005-06-01 23:12:54.706422840 -0400
@@ -0,0 +1,412 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_CM_H
+#define __ELAN_CM_H
+
+#ident "@(#)$Id: cm.h,v 1.14.2.1 2004/11/12 10:54:50 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/cm.h,v $*/
+
+#include <elan/statemap.h>
+
+#if defined(DIGITAL_UNIX)
+/*
+ * On Tru64 - SMP doesn't mean Symmetric - cpu 0 is a master cpu and is responsible
+ * for handling all PCI interrupts and "funneled" operations.  When a kernel thread
+ * is made runnable, the scheduler will choose which cpu it will run on at that time,
+ * and will only execute a higher priority thread from another cpu's run queue when 
+ * it becomes totally idle (apparently also including user processes).  Also the 
+ * assert_wait_mesg_timo function uses a per-cpu timeout - these can only get executed
+ * at "preemptable" places - so again have no guarantee on when they will execute if
+ * they happen to be queued on a "hogged" cpu. The combination of these mean that the Tru64
+ * is incapable of scheduling a high priority kernel  thread within a deterministic time
+ * of when it should have become runnable - wonderfull.
+ *
+ * Hence the solution Compaq have proposed it to schedule a timeout onto all of the
+ * cpu's timeouts lists at the maximum frequency that we could want to execute code,
+ * then to handle the scheduling of work between these ourselves.  With a bit of luck
+ * ..... at least one cpu will be sufficiently unloaded to allow us to get a chance
+ * to do our important work.
+ *
+ * However ..... this still is not reliable, since timeouts under Tru64 are still 
+ * only run when the currently running kernel thread "co-operates" by calling one
+ * of a number of functions which is permitted to run the "lwc"s AND is not holding
+ * any spinlocks AND is running ai IPL 0.   However Compaq are unable to provide
+ * any upper limit on the time between the "lwc"'s being run and so it is possible
+ * for all 4 cpus to not run them for an unbounded time.
+ *
+ * The solution proposed is to use the RM_TEMP_BACKDOOR hook which was added to 
+ * hardclock() to "solve" this problem for Memory Channel.  However, since it
+ * is called within the clock interrupt it is not permissible to aquire any
+ * spinlocks, nor to run for "too long".  This means that it is not possible to
+ * call the heartbeat algorithm from this hook.  
+ *
+ * Our solution to these limitations is to use the hook to cause an elan interrupt 
+ * to be delivered, by issueing a mis-aligned SetEvent command - this causes the device 
+ * to trap and ep_cprocTrap() can then run the heartbeat code.  However there is a lock 
+ * order violation between the elan_dev::IntrLock and ep_dev::Lock, so we have to 
+ * use a trylock and if we fail, then hope that when the interrupt is delievered again
+ * some time later we will succeed.
+ *
+ * However this only works if the kernel is able to respond to the Elan interrupt,
+ * so we panic inside the RM_TEMP_BACKDOOR hook if the SetEvent's interrupt has
+ * not been taken for more than an CM_TIMER_SCHEDULE_TIMEOUT interval.
+ *
+ * In fact this is exactly the mechanism that other operating systems use to
+ * execute timeouts, since the hardclock interrupt posts a low priority 
+ * "soft interrupt" which "pre-eempts" the currently running thread and then
+ * executes the timeouts.To block timeouts you use splsoftclock() the same as 
+ * in Tru64.
+ */
+#define PER_CPU_TIMEOUT                        TRUE
+#endif
+
+
+#define CM_SGMTS_PER_LEVEL             8                       /* maximum nodes in each segment */
+#define CM_MAX_LEVELS                  6                       /* maximum depth of tree */
+
+/* message buffers/dmas/events etc */
+#define CM_NUM_NODE_MSG_BUFFERS                (CM_MAX_LEVELS * CM_SGMTS_PER_LEVEL) /* subordinates and leader */
+#define CM_NUM_SPARE_MSG_BUFFERS       8                       /* spare msg buffers for non-connected nodes */
+#define CM_NUM_MSG_BUFFERS             (CM_NUM_NODE_MSG_BUFFERS + CM_NUM_SPARE_MSG_BUFFERS)
+
+#define CM_INPUTQ_ENTRIES              128                     /* # entries in input queue */
+
+#define CM_PERIODIC_DISCOVER_INTERVAL  (5000)          /* 5s (infrequent resolution of established leader conflicts) */
+#define CM_URGENT_DISCOVER_INTERVAL    (50)            /* 0.05s (more frequently than heartbeats 'cause they don't retry) */
+#define CM_HEARTBEAT_INTERVAL          (125)           /* 0.125s */
+#define CM_TIMER_SCHEDULE_TIMEOUT      (4000)          /* 4s     Maximum time before a timer that's secheduled to run gets to run (eg blocked in interrupt handlers etc) */
+#define CM_THREAD_SCHEDULE_TIMEOUT     (30000)         /* 30s    Maximum time before a thread that's scheduled to run gets to run */
+#define CM_THREAD_RUNNING_TIMEOUT      (30000)         /* 30s    Don't expect the manager thread to be running longer than this */
+
+#ifdef PER_CPU_TIMEOUT
+#define CM_PERCPU_TIMEOUT_INTERVAL     (50)            /* 0.05s (must be less than all above intervals) */
+#define CM_PACEMAKER_INTERVAL          (500)           /* 0.05s */
+
+#define CM_HEARTBEAT_OVERDUE           (250)           /* 0.25s Maximum time a timeout can be overdue before taking extreme action */
+#endif
+
+#define CM_P2P_DMA_RETRIES             31
+
+/* We expect at least 1 point-to-point message in CM_P2P_MSG_RETRIES
+ * attempts to send one to be successfully received */
+#define CM_P2P_MSG_RETRIES             8
+
+/* We expect at least 1 broadcast message in CM_BCAST_MSG_RETRIES attempts
+ * to send one to be successfully received. */
+#define CM_BCAST_MSG_RETRIES           40
+
+/* Heartbeat timeout allows for a node stalling and still getting its
+ * heartbeat. The 2 is to allow for unsynchronised polling times. */
+#define CM_HEARTBEAT_TIMEOUT           (CM_TIMER_SCHEDULE_TIMEOUT + (2 + CM_P2P_MSG_RETRIES) * CM_HEARTBEAT_INTERVAL)
+
+/* Discover timeout must be > CM_HEARTBEAT_TIMEOUT to guarantee that people
+ * who don't see discovery are considered dead by their leader.  This
+ * ensures that by the time a node "discovers" it is a leader of a segment,
+ * the previous leader of that segment will have been deemed to be dead by
+ * its the parent segment's leader */
+#define CM_DISCOVER_TIMEOUT            (CM_TIMER_SCHEDULE_TIMEOUT + (2 + CM_BCAST_MSG_RETRIES) * CM_URGENT_DISCOVER_INTERVAL)
+
+#define CM_WAITING_TIMEOUT             (CM_DISCOVER_TIMEOUT * 100)
+
+/*
+ * Convert all timeouts specified in mS into "ticks"
+ */
+#define MSEC2TICKS(MSEC)               (((MSEC)*HZ)/1000)
+
+
+/* statemap entry */
+typedef struct cm_state_entry
+{
+    int16_t           level;                   /* cluster level to apply to */
+    int16_t          offset;                   /* from statemap_findchange() */
+    uint16_t          seg[BT_NBIPUL/16];       /* ditto */
+} CM_STATEMAP_ENTRY;
+
+/* offset is >= 0 for a change to apply and */
+#define STATEMAP_NOMORECHANGES (-1)            /* end of a set of updates */
+#define STATEMAP_RESET         (-2)            /* reset the target map */
+#define STATEMAP_NOOP          (-3)            /* null token */
+
+/* CM message format */
+typedef int8_t CM_SEQ;                         /* heartbeat sequence numbers; at least 2 bits, signed */
+
+/*
+ * The message header is received into the last 64 byte block of 
+ * the input queue and the Version *MUST* be the last word of the 
+ * block to ensure that we can see that the whole of the message
+ * has reached main memory after we've seen the input queue pointer
+ * have been updated.
+ */
+typedef struct ep_cm_hdr
+{
+    uint32_t          Pad0;
+    uint32_t          Pad1;
+
+    uint8_t           Type;
+    uint8_t           Level;
+    CM_SEQ            Seq;                     /* precision at least 2 bits each*/
+    CM_SEQ            AckSeq;
+    
+    uint16_t          NumMaps;
+    uint16_t          MachineId;
+
+    uint16_t          NodeId;
+    uint16_t          Checksum;
+
+    uint32_t           Timestamp;
+    uint32_t           ParamHash;
+    uint32_t          Version;
+} CM_HDR;
+
+#define CM_HDR_SIZE        sizeof (CM_HDR)
+
+typedef struct cm_msg
+{
+    union {
+       CM_STATEMAP_ENTRY   Statemaps[1];               /* piggy-backed statemap updates start here */
+       uint8_t             Space[EP_SYSTEMQ_MSG_MAX - CM_HDR_SIZE];
+    } Payload;
+    
+    CM_HDR                 Hdr;
+} CM_MSG;
+
+/* The maximum number of statemap entries that can fit within an EP_CM_MSG_BUFFER */
+#define CM_MSG_MAXMAPS         (offsetof (CM_MSG, Hdr) / sizeof (CM_STATEMAP_ENTRY))
+#define CM_MSG_MAP(mapno)      (CM_MSG_MAXMAPS - (mapno) - 1)
+
+/* The actual special message base & size, including 'nmaps' piggy-backed statemap entries */
+#define CM_MSG_BASE(nmaps)     (nmaps == 0 ? offsetof (CM_MSG, Hdr) : offsetof (CM_MSG, Payload.Statemaps[CM_MSG_MAXMAPS - nmaps]))
+#define CM_MSG_SIZE(nmaps)     (sizeof (CM_MSG) - CM_MSG_BASE(nmaps))
+
+#define CM_MSG_VERSION                         0xcad00005
+#define CM_MSG_TYPE_RESOLVE_LEADER             0
+#define CM_MSG_TYPE_DISCOVER_LEADER            1
+#define CM_MSG_TYPE_NOTIFY                     2
+#define CM_MSG_TYPE_DISCOVER_SUBORDINATE       3
+#define CM_MSG_TYPE_IMCOMING                   4
+#define CM_MSG_TYPE_HEARTBEAT                  5
+#define CM_MSG_TYPE_REJOIN                     6
+
+/* CM machine segment */
+typedef struct cm_sgmtMaps
+{
+    u_char       InputMapValid;                        /* Input map has been set */
+    u_char       OutputMapValid;               /* Output map has been set */
+    u_char       SentChanges;                  /* got an outstanding STATEMAP_NOMORECHANGES to send */
+    statemap_t  *OutputMap;                    /* state to send */
+    statemap_t  *InputMap;                     /* state received */
+    statemap_t  *CurrentInputMap;              /* state being received */
+} CM_SGMTMAPS;
+
+typedef struct cm_sgmt
+{
+   u_char       State;
+   u_char       SendMaps;
+   u_char       MsgAcked;
+   CM_SEQ      MsgSeq;
+   CM_SEQ      AckSeq;
+   u_int       NodeId;
+   long                UpdateTick;
+   long                WaitingTick;
+   uint32_t    Timestamp;
+   CM_SGMTMAPS  Maps[CM_MAX_LEVELS];           /* Maps[i] == state for cluster level i */
+   u_short      MsgNumber;                     /* msg buffer to use */
+   u_short     NumMaps;                        /* # maps in message buffer */
+   u_short      Level;
+   u_short      Sgmt;
+} CM_SGMT;
+
+#define CM_SGMT_ABSENT         0               /* no one there at all */
+#define CM_SGMT_WAITING                1               /* waiting for subtree to connect */
+#define CM_SGMT_COMING         2               /* expecting a subtree to reconnect */
+#define CM_SGMT_PRESENT                3               /* connected */
+
+typedef struct cm_level
+{
+    int               SwitchLevel;
+    u_int             MinNodeId;
+    u_int              NumNodes;
+    u_int              NumSegs;
+    u_int              MySgmt;
+   
+    /* SubordinateMap[i] == OR of all subordinate maps on this level and down for cluster level i */
+    u_char             SubordinateMapValid[CM_MAX_LEVELS];
+    statemap_t        *SubordinateMap[CM_MAX_LEVELS];
+
+    /* maps/flags for this cluster level */
+    u_int              Online:1;                               /* I've gone online (seen myself running) */
+    u_int             Restarting:1;                            /* driving my owm restart bit */
+    u_char            OfflineReasons;                          /* forced offline by broadcast */
+
+    u_char             GlobalMapValid;
+    u_char             SubTreeMapValid;
+    u_long            Connected;
+
+    statemap_t        *LocalMap;               /* state bits I drive */
+    statemap_t        *SubTreeMap;             /* OR of my and my subtree states */
+    statemap_t        *GlobalMap;              /* OR of all node states */
+    statemap_t        *LastGlobalMap;          /* last map I saw */
+    statemap_t        *TmpMap;                 /* scratchpad */
+
+    CM_SGMT           Sgmts[CM_SGMTS_PER_LEVEL];
+} CM_LEVEL;
+
+#define CM_ROLE_LEADER_CANDIDATE       0
+#define CM_ROLE_LEADER                 1
+#define CM_ROLE_SUBORDINATE            2
+
+/* global status bits */
+#define CM_GSTATUS_STATUS_MASK         0x03    /* bits nodes drive to broadcast their status */
+#define CM_GSTATUS_ABSENT              0x00    /* Off the network */
+#define CM_GSTATUS_STARTING            0x01    /* I'm waiting for everyone to see me online */
+#define CM_GSTATUS_RUNNING              0x03   /* up and running */
+#define CM_GSTATUS_CLOSING             0x02    /* I'm waiting for everyone to see me offline */
+
+#define CM_GSTATUS_ACK_MASK            0x0c    /* bits node drive to ack other status */
+#define CM_GSTATUS_MAY_START           0x04    /* Everyone thinks I may not start */
+#define CM_GSTATUS_MAY_RUN             0x08    /* Everyone thinks I may not run */
+
+#define CM_GSTATUS_RESTART             0x10    /* Someone thinks I should restart */
+#define CM_GSTATUS_BITS                        5
+
+#define CM_GSTATUS_BASE(node)          ((node) * CM_GSTATUS_BITS)
+
+#if defined(PER_CPU_TIMEOUT)
+typedef struct cm_timeout_data
+{
+    long               ScheduledAt;                            /* lbolt timeout was scheduled to run at */
+
+    unsigned long       EarlyCount;                            /* # times run early than NextRun */
+    unsigned long      MissedCount;                            /* # times run on time - but someone else was running it */
+    unsigned long       WastedCount;                           /* # times we failed to get the spinlock */
+    unsigned long      WorkCount;                              /* # times we're the one running */
+
+    unsigned long      WorstDelay;                             /* worst scheduling delay */
+    unsigned long      BestDelay;                              /* best scheduling delay */
+
+    unsigned long      WorstLockDelay;                         /* worst delay before getting rail->Lock */
+
+    unsigned long      WorstHearbeatDelay;                     /* worst delay before calling DoHeartbeatWork */
+} CM_TIMEOUT_DATA;
+#endif
+
+typedef struct cm_rail
+{
+    EP_RAIL          *Rail;                                    /* rail we're associated with */
+    struct list_head   Link;                                   /*   and linked on the CM_SUBSYS */
+
+    uint32_t          ParamHash;                               /* hash of critical parameters */
+    uint32_t           Timestamp;
+    long              DiscoverStartTick;                       /* when discovery start */
+
+    unsigned int       NodeId;                                 /* my node id */
+    unsigned int       NumNodes;                               /*   and number of nodes */
+    unsigned int       NumLevels;                              /* number of levels computed from machine size */
+    int                       BroadcastLevel;
+    long              BroadcastLevelTick;
+    unsigned int       TopLevel;                               /* level at which I'm not a leader */
+    unsigned char      Role;                                   /* state at TopLevel */
+
+    EP_INPUTQ        *PolledQueue;                             /* polled input queue */
+    EP_INPUTQ        *IntrQueue;                               /* intr input queue */
+    EP_OUTPUTQ       *MsgQueue;                                /* message  */
+    unsigned int       NextSpareMsg;                           /* next "spare" message buffer to use */
+
+    EP_CM_RAIL_STATS   Stats;                                  /* statistics */
+
+    kmutex_t          Mutex;
+    spinlock_t        Lock;
+    
+    long              NextHeartbeatTime;                       /* next time to check/send heartbeats */
+    long              NextDiscoverTime;                        /* next time to progress discovery  */
+    long              NextRunTime;                             /* the earlier of the above two or intr requires inputq poll*/
+
+    unsigned int       OfflineReasons;                         /* forced offline by procfs/manager thread stuck */
+
+#if defined(PER_CPU_TIMEOUT)
+    spinlock_t        HeartbeatTimeoutsLock;                   /* spinlock to sequentialise per-cpu timeouts */
+    long              HeartbeatTimeoutsStarted;                /* bitmap of which timeouts have started */
+    long              HeartbeatTimeoutsStopped;                /* bitmap of which timeouts have stopped */
+    long              HeartbeatTimeoutsShouldStop;             /* flag to indicate timeouts should stop */
+    kcondvar_t        HeartbeatTimeoutsWait;                   /* place to sleep waiting for timeouts to stop */
+    long              HeartbeatTimeoutRunning;                 /* someone is running the timeout - don't try for the lock */
+
+    long              HeartbeatTimeoutOverdue;                 /* heartbeat seen as overdue - interrupt requested */
+
+    CM_TIMEOUT_DATA   *HeartbeatTimeoutsData;                  /* per timeout data */
+#else
+    struct timer_list  HeartbeatTimer;                         /* timer for heartbeat/discovery */
+#endif
+
+    CM_LEVEL           Levels[CM_MAX_LEVELS];
+} CM_RAIL;
+
+/* OfflineReasons (both per-rail and  */
+#define CM_OFFLINE_BROADCAST           (1 << 0)
+#define CM_OFFLINE_PROCFS              (1 << 1)
+#define CM_OFFLINE_MANAGER             (1 << 2)
+
+typedef struct cm_subsys
+{
+    EP_SUBSYS          Subsys;
+    CM_RAIL            *Rails[EP_MAX_RAILS];
+} CM_SUBSYS;
+
+extern int  MachineId;
+
+extern void cm_node_disconnected (EP_RAIL *rail, unsigned nodeId);
+extern void cm_restart_node (EP_RAIL *rail, unsigned nodeId);
+extern void cm_restart_comms (CM_RAIL *cmRail);
+extern int  cm_init (EP_SYS *sys);
+
+extern void DisplayRail(EP_RAIL *rail);
+extern void DisplaySegs (EP_RAIL *rail);
+extern void DisplayStatus (EP_RAIL *rail);
+
+typedef struct proc_private
+{
+    struct nodeset_private *pr_next;
+    EP_RAIL                *pr_rail;
+    char                  *pr_data;
+    int                     pr_data_len;
+    unsigned               pr_off;
+    unsigned               pr_len;
+    DisplayInfo             pr_di;
+} PROC_PRIVATE;
+
+extern void    proc_character_fill (long mode, char *fmt, ...);
+extern int     proc_release (struct inode *inode, struct file *file);
+extern ssize_t proc_read (struct file *file, char *buf, size_t count, loff_t *ppos);
+
+
+extern void DisplayNodeMaps  (DisplayInfo *di, CM_RAIL *cmRail);
+extern void DisplayNodeSgmts (DisplayInfo *di, CM_RAIL *cmRail);
+extern void DisplayRailDo    (DisplayInfo *di, EP_RAIL *rail);
+
+extern int    cm_read_cluster(EP_RAIL *rail,char *page);
+extern void   cm_force_offline (EP_RAIL *rail, int offline, unsigned int reason);
+
+extern int    cm_svc_indicator_set      (EP_RAIL *rail, int svc_indicator);
+extern int    cm_svc_indicator_clear    (EP_RAIL *rail, int svc_indicator);
+extern int    cm_svc_indicator_is_set   (EP_RAIL *rail, int svc_indicator, int nodeId);
+extern int    cm_svc_indicator_bitmap   (EP_RAIL *rail, int svc_indicator, bitmap_t * bitmap, int low, int nnodes);
+
+/* cm_procfs.c */
+extern void   cm_procfs_init (CM_SUBSYS *subsys);
+extern void   cm_procfs_fini (CM_SUBSYS *subsys);
+extern void   cm_procfs_rail_init (CM_RAIL *rail);
+extern void   cm_procfs_rail_fini (CM_RAIL *rail);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN_CM_H */
+
Index: linux-2.4.21/include/elan/compat.h
===================================================================
--- linux-2.4.21.orig/include/elan/compat.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/compat.h  2005-06-01 23:12:54.706422840 -0400
@@ -0,0 +1,23 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: compat.h,v 1.1 2003/12/03 13:18:48 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/compat.h,v $*/
+
+#ifndef __ELAN_COMPAT_H
+#define __ELAN_COMPAT_H
+
+#define ELANMOD_STATS_MAP      ELAN_STATS_MAP
+
+#endif  /* __ELAN_COMPAT_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/device.h
===================================================================
--- linux-2.4.21.orig/include/elan/device.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/device.h  2005-06-01 23:12:54.707422688 -0400
@@ -0,0 +1,62 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: device.h,v 1.5 2003/09/24 13:55:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/device.h,v $*/
+
+#ifndef __ELAN_DEVICE_H
+#define __ELAN_DEVICE_H
+
+/* non-kernel headings */
+typedef unsigned int ELAN_DEV_IDX;
+
+#if defined(__KERNEL__)
+
+/* device callbacks */
+#define ELAN_DEV_OPS_VERSION ((u_int)1)
+
+typedef struct elan_dev_ops
+{
+       /* dev info */
+       int (*get_position)          (void *user_data, ELAN_POSITION *position);
+       int (*set_position)          (void *user_data, unsigned short nodeId, unsigned short numNodes);
+
+       /* cap */
+
+       u_int  ops_version;
+} ELAN_DEV_OPS;
+
+typedef struct elan_dev_struct
+{
+       struct list_head node;
+
+       ELAN_DEV_IDX     devidx;
+       ELAN_DEVINFO    *devinfo;
+       void            *user_data;
+       ELAN_DEV_OPS *ops;
+} ELAN_DEV_STRUCT;
+
+/* device.c */
+extern ELAN_DEV_IDX         elan_dev_register   (ELAN_DEVINFO    *devinfo, 
+                                                   ELAN_DEV_OPS *ops,
+                                                   void            *userdata);
+extern int                  elan_dev_deregister (ELAN_DEVINFO *devinfo);
+
+extern ELAN_DEV_STRUCT * elan_dev_find       (ELAN_DEV_IDX devidx);
+
+extern ELAN_DEV_STRUCT * elan_dev_find_byrail(unsigned short deviceid, unsigned rail);
+extern int                  elan_dev_dump       (void);
+
+#endif /* __KERNEL__ */
+
+#endif /* __ELAN_DEVICE_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/devinfo.h
===================================================================
--- linux-2.4.21.orig/include/elan/devinfo.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/devinfo.h 2005-06-01 23:12:54.707422688 -0400
@@ -0,0 +1,81 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: devinfo.h,v 1.11 2004/03/12 14:27:39 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/devinfo.h,v $*/
+
+#ifndef __ELAN_DEVINFO_H
+#define __ELAN_DEVINFO_H
+
+#define ELAN_MAX_LEVELS                        8                       /* maximum number of levels in switch network */
+
+typedef struct elan_position
+{
+       unsigned        pos_mode;                               /* mode we're operating in */
+       unsigned        pos_nodeid;                             /* port this device connected to */
+       unsigned        pos_levels;                             /* number of levels to top switch */
+       unsigned        pos_nodes;                              /* number of nodes in the machine */
+       unsigned        pos_random_disabled;                    /* levels at which "random" routing is not possible */
+       unsigned char   pos_arity[ELAN_MAX_LEVELS];             /* number of downlinks per switch level */
+} ELAN_POSITION;
+
+#define ELAN4_PARAM_PCI_PADDING_FLAGS          0               /* A bit field, representing good places to burst across the pci                      */
+#define ELAN4_PARAM_EVENT_COPY_WIN             1               /* The num of cmds when it becomes quicker to send via event copy than write directly */
+#define ELAN4_PARAM_WRITE_COMBINING            2               /* If set the device supports bursts accesses across the pci bus                      */
+#define ELAN4_PARAM_COUNT                      12
+
+typedef struct elan_params
+{
+       unsigned        values[ELAN4_PARAM_COUNT];
+} ELAN_PARAMS;
+
+/* values for pos_mode */
+#define ELAN_POS_UNKNOWN               0                       /* network position unknown */
+#define ELAN_POS_MODE_SWITCHED         1                       /* connected to a switch */
+#define ELAN_POS_MODE_LOOPBACK         2                       /* loopback connector */
+#define ELAN_POS_MODE_BACKTOBACK       3                       /* cabled back-to-back to another node */
+
+typedef struct elan_devinfo
+{
+       unsigned short  dev_vendor_id;                          /* pci vendor id */
+       unsigned short  dev_device_id;                          /* pci device id */
+       unsigned char   dev_revision_id;                        /* pci revision id */
+       unsigned char   dev_instance;                           /* device instance number */
+       unsigned char   dev_rail;                               /* device rail number */
+
+       unsigned short  dev_driver_version;                     /* device driver version */
+       unsigned short  dev_params_mask;                        /* mask for valid entries in dev_params array */
+       ELAN_PARAMS     dev_params;                             /* device parametization */
+
+       unsigned        dev_num_down_links_value;               /* MRH hint as to machine size NEEDS coding XXXXX */
+} ELAN_DEVINFO;
+
+#define PCI_VENDOR_ID_QUADRICS         0x14fc
+#define PCI_DEVICE_ID_ELAN3            0x0000
+#define   PCI_REVISION_ID_ELAN3_REVA   0x0000
+#define   PCI_REVISION_ID_ELAN3_REVB   0x0001
+#define PCI_DEVICE_ID_ELAN4            0x0001
+#define   PCI_REVISION_ID_ELAN4_REVA   0x0000
+#define   PCI_REVISION_ID_ELAN4_REVB   0x0001
+
+#if defined(__KERNEL__)
+/* devinfo.c */
+#include <elan/capability.h>
+#include <elan/device.h>
+extern int elan_get_devinfo  (ELAN_DEV_IDX devidx, ELAN_DEVINFO  *devinfo);
+extern int elan_get_position (ELAN_DEV_IDX devidx, ELAN_POSITION *position);
+extern int elan_set_position (ELAN_DEV_IDX devidx, unsigned short nodeId, unsigned short numNodes);
+#endif /* __KERNEL__ */
+
+
+#endif /* __ELAN_DEVINFO_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/elanmoddebug.h
===================================================================
--- linux-2.4.21.orig/include/elan/elanmoddebug.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/elanmoddebug.h    2005-06-01 23:12:54.707422688 -0400
@@ -0,0 +1,63 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN_DEBUG_H
+#define _ELAN_DEBUG_H
+
+
+#ident "$Id: elanmoddebug.h,v 1.5 2003/09/24 13:55:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/elanmoddebug.h,v $ */
+
+#if defined(__KERNEL__)
+
+/* 0 | QSNET_DEBUG_BUFFER | QSNET_DEBUG_CONSOLE */
+extern int elan_debug_mode; 
+extern int elan_debug_mask;
+
+#define ELAN_DBG_VP            0x00000001
+#define ELAN_DBG_CAP            0x00000002
+#define ELAN_DBG_CTRL           0x00000004
+#define ELAN_DBG_SYS_FN         0x00000008
+#define ELAN_DBG_ALL           0xffffffff
+
+
+#if defined(DEBUG_PRINTF)
+#  define ELAN_DEBUG0(m,fmt)                   ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode,fmt)             : (void)0)
+#  define ELAN_DEBUG1(m,fmt,a)                 ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode,fmt,a)           : (void)0)
+#  define ELAN_DEBUG2(m,fmt,a,b)               ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode,fmt,a,b)         : (void)0)
+#  define ELAN_DEBUG3(m,fmt,a,b,c)             ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode,fmt,a,b,c)       : (void)0)
+#  define ELAN_DEBUG4(m,fmt,a,b,c,d)           ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode,fmt,a,b,c,d)     : (void)0)
+#  define ELAN_DEBUG5(m,fmt,a,b,c,d,e)         ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode,fmt,a,b,c,d,e)   : (void)0)
+#  define ELAN_DEBUG6(m,fmt,a,b,c,d,e,f)       ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode,fmt,a,b,c,d,e,f) : (void)0)
+#ifdef __GNUC__
+#  define ELAN_DEBUG(m,args...)                        ((elan_debug_mask&(m)) ? qsnet_debugf(elan_debug_mode, ##args)         : (void)0)
+#endif
+
+#else
+
+#  define ELAN_DEBUG0(m,fmt)                   (0)
+#  define ELAN_DEBUG1(m,fmt,a)                 (0)
+#  define ELAN_DEBUG2(m,fmt,a,b)               (0)
+#  define ELAN_DEBUG3(m,fmt,a,b,c)             (0)
+#  define ELAN_DEBUG4(m,fmt,a,b,c,d)           (0)
+#  define ELAN_DEBUG5(m,fmt,a,b,c,d,e)         (0)
+#  define ELAN_DEBUG6(m,fmt,a,b,c,d,e,f)       (0)
+#ifdef __GNUC__
+#  define ELAN_DEBUG(m,args...)
+#endif
+
+#endif /* DEBUG_PRINTF */
+
+
+#endif /* __KERNEL__ */
+#endif /* _ELAN_DEBUG_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/elanmod.h
===================================================================
--- linux-2.4.21.orig/include/elan/elanmod.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/elanmod.h 2005-06-01 23:12:54.708422536 -0400
@@ -0,0 +1,59 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elanmod.h,v 1.10 2004/06/18 09:28:16 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/elanmod.h,v $*/
+
+#ifndef __ELAN_MOD_H
+#define __ELAN_MOD_H
+
+#include <elan/devinfo.h>
+#include <elan/device.h>
+#include <elan/capability.h>
+#include <elan/stats.h>
+
+#if defined(__KERNEL__)
+
+#include <elan/elanmoddebug.h>
+
+extern kmutex_t elan_mutex;
+
+/* elan_general.c */
+extern int elan_init(void);
+extern int elan_fini(void);
+
+/* return codes, -ve => errno, +ve => success */
+#define ELAN_CAP_OK  (0)
+#define ELAN_CAP_RMS (1)
+
+#define ELAN_USER_ATTACH    (1)
+#define ELAN_USER_DETACH    (2)
+#define ELAN_USER_P2P       (3)
+#define ELAN_USER_BROADCAST (4)
+
+extern int elanmod_classify_cap (ELAN_POSITION *position, ELAN_CAPABILITY *cap, unsigned use);
+
+#define ELAN_USER_BASE_CONTEXT_NUM     0x000                   /* first user allowable context */
+#define ELAN_USER_TOP_CONTEXT_NUM      0x7FF                   /* last user allowable context */
+
+#define ELAN_RMS_BASE_CONTEXT_NUM      0x400                   /* reserved for RMS allocation */
+#define ELAN_RMS_TOP_CONTEXT_NUM       0x7FF
+
+#define ELAN_USER_CONTEXT(ctx)         ((ctx) >= ELAN_USER_BASE_CONTEXT_NUM && \
+                                        (ctx) <= ELAN_USER_TOP_CONTEXT_NUM)    
+
+#define ELAN_RMS_CONTEXT(ctx)          ((ctx) >= ELAN_RMS_BASE_CONTEXT_NUM && \
+                                        (ctx) <= ELAN_RMS_TOP_CONTEXT_NUM)    
+#endif /* __KERNEL__ */
+
+#endif /* __ELAN_MOD_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/elanmod_linux.h
===================================================================
--- linux-2.4.21.orig/include/elan/elanmod_linux.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/elanmod_linux.h   2005-06-01 23:12:54.708422536 -0400
@@ -0,0 +1,140 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: elanmod_linux.h,v 1.6 2003/09/29 15:36:20 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/elanmod_linux.h,v $*/
+
+#ifndef __ELAN_MOD_LINUX_H
+#define __ELAN_MOD_LINUX_H
+
+#define ELANCRTL_USER_BASE             0x40
+
+/* stats */
+typedef struct elanctrl_stats_get_next_struct
+{
+       ELAN_STATS_IDX   statidx; 
+       ELAN_STATS_IDX  *next_statidx; /* return value */
+} ELANCTRL_STATS_GET_NEXT_STRUCT;
+#define ELANCTRL_STATS_GET_NEXT   _IOR   ('e', ELANCRTL_USER_BASE + 0,  ELANCTRL_STATS_GET_NEXT_STRUCT)
+
+typedef struct elanctrl_stats_find_index_struct
+{
+       caddr_t          block_name;
+       ELAN_STATS_IDX  *statidx; /* return value */
+       uint        *num_entries; /* return value */
+} ELANCTRL_STATS_FIND_INDEX_STRUCT;
+#define ELANCTRL_STATS_FIND_INDEX   _IOR   ('e', ELANCRTL_USER_BASE + 1,  ELANCTRL_STATS_FIND_INDEX_STRUCT)
+
+typedef struct elanctrl_stats_get_block_info_struct
+{
+       ELAN_STATS_IDX  statidx; 
+       caddr_t       block_name; /* return value */
+       uint        *num_entries; /* return value */
+} ELANCTRL_STATS_GET_BLOCK_INFO_STRUCT;
+#define ELANCTRL_STATS_GET_BLOCK_INFO   _IOR   ('e', ELANCRTL_USER_BASE + 2, ELANCTRL_STATS_GET_BLOCK_INFO_STRUCT)
+
+typedef struct elanctrl_stats_get_index_name_struct
+{
+       ELAN_STATS_IDX statidx; 
+       uint           index;
+       caddr_t        name; /* return value */
+} ELANCTRL_STATS_GET_INDEX_NAME_STRUCT;
+#define ELANCTRL_STATS_GET_INDEX_NAME   _IOR   ('e', ELANCRTL_USER_BASE + 3, ELANCTRL_STATS_GET_INDEX_NAME_STRUCT)
+
+typedef struct elanctrl_stats_clear_block_struct
+{
+       ELAN_STATS_IDX statidx; 
+} ELANCTRL_STATS_CLEAR_BLOCK_STRUCT;
+#define ELANCTRL_STATS_CLEAR_BLOCK   _IOR   ('e', ELANCRTL_USER_BASE + 4, ELANCTRL_STATS_CLEAR_BLOCK_STRUCT)
+
+typedef struct elanctrl_stats_get_block_struct
+{
+       ELAN_STATS_IDX statidx; 
+       uint           entries;  
+       ulong         *values; /* return values */
+} ELANCTRL_STATS_GET_BLOCK_STRUCT;
+#define ELANCTRL_STATS_GET_BLOCK        _IOR   ('e', ELANCRTL_USER_BASE + 5, ELANCTRL_STATS_GET_BLOCK_STRUCT)
+
+
+typedef struct elanctrl_get_devinfo_struct
+{
+       ELAN_DEV_IDX  devidx; 
+       ELAN_DEVINFO *devinfo; /* return values */
+} ELANCTRL_GET_DEVINFO_STRUCT;
+#define ELANCTRL_GET_DEVINFO        _IOR   ('e', ELANCRTL_USER_BASE + 6, ELANCTRL_GET_DEVINFO_STRUCT)
+
+typedef struct elanctrl_get_position_struct
+{
+       ELAN_DEV_IDX   devidx; 
+       ELAN_POSITION *position; /* return values */
+} ELANCTRL_GET_POSITION_STRUCT;
+#define ELANCTRL_GET_POSITION        _IOR   ('e', ELANCRTL_USER_BASE + 7, ELANCTRL_GET_POSITION_STRUCT)
+
+typedef struct elanctrl_set_position_struct
+{
+       ELAN_DEV_IDX   devidx; 
+       unsigned short nodeId;
+       unsigned short numNodes;
+} ELANCTRL_SET_POSITION_STRUCT;
+#define ELANCTRL_SET_POSITION        _IOR   ('e', ELANCRTL_USER_BASE + 8, ELANCTRL_SET_POSITION_STRUCT)
+
+typedef struct elanctrl_create_cap_struct
+{
+       ELAN_CAPABILITY cap;
+} ELANCTRL_CREATE_CAP_STRUCT;
+#define ELANCTRL_CREATE_CAP             _IOW   ('e', ELANCRTL_USER_BASE + 9, ELANCTRL_CREATE_CAP_STRUCT)
+
+typedef struct elanctrl_destroy_cap_struct
+{
+       ELAN_CAPABILITY cap;
+} ELANCTRL_DESTROY_CAP_STRUCT;
+#define ELANCTRL_DESTROY_CAP             _IOW   ('e', ELANCRTL_USER_BASE + 10, ELANCTRL_DESTROY_CAP_STRUCT)
+
+typedef struct elanctrl_create_vp_struct
+{
+       ELAN_CAPABILITY cap;
+       ELAN_CAPABILITY map;
+} ELANCTRL_CREATE_VP_STRUCT;
+#define ELANCTRL_CREATE_VP             _IOW   ('e', ELANCRTL_USER_BASE + 11, ELANCTRL_CREATE_VP_STRUCT)
+
+typedef struct elanctrl_destroy_vp_struct
+{
+       ELAN_CAPABILITY cap;
+       ELAN_CAPABILITY map;
+} ELANCTRL_DESTROY_VP_STRUCT;
+#define ELANCTRL_DESTROY_VP          _IOW   ('e', ELANCRTL_USER_BASE + 12, ELANCTRL_DESTROY_VP_STRUCT)
+
+#define ELANCTRL_DEBUG_DUMP          _IO    ('e', ELANCRTL_USER_BASE + 13)
+
+typedef struct elanctrl_get_caps_struct
+{
+       uint            *number_of_results;
+       uint             array_size;
+       ELAN_CAP_STRUCT *caps;
+} ELANCTRL_GET_CAPS_STRUCT;
+#define ELANCTRL_GET_CAPS          _IOW   ('e', ELANCRTL_USER_BASE + 14, ELANCTRL_GET_CAPS_STRUCT)
+
+
+typedef struct elanctrl_debug_buffer_struct
+{
+       caddr_t buffer;
+       int     size;
+} ELANCTRL_DEBUG_BUFFER_STRUCT;
+#define ELANCTRL_DEBUG_BUFFER _IOW ('e', ELANCRTL_USER_BASE + 15, ELANCTRL_DEBUG_BUFFER_STRUCT)
+
+#define ELANMOD_PROCFS_IOCTL      "/proc/qsnet/elan/ioctl"
+#define ELANMOD_PROCFS_VERSION    "/proc/qsnet/elan/version"
+#define ELANMOD_PROCFS_DEBUG_MASK "/proc/qsnet/elan/debug_mask"
+#define ELANMOD_PROCFS_DEBUG_MODE "/proc/qsnet/elan/debug_mode"
+
+#endif /* __ELAN_MOD_LINUX_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/elanmod_subsystem.h
===================================================================
--- linux-2.4.21.orig/include/elan/elanmod_subsystem.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/elanmod_subsystem.h       2005-06-01 23:12:54.708422536 -0400
@@ -0,0 +1,138 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_SUBSYSTEM_H
+#define __ELAN_SUBSYSTEM_H
+
+#include <sys/types.h>
+#include <sys/param.h>
+
+#if defined( __KERNEL__) 
+int elan_configure(
+    cfg_op_t op,
+    caddr_t  indata,
+    ulong    indata_size,
+    caddr_t  outdata,
+    ulong    outdata_size);
+#endif
+
+#define ELAN_KMOD_CODE(x)      ((x)+CFG_OP_SUBSYS_MIN)
+#define ELAN_MAX_KMOD_CODES 100
+
+#define ELAN_SUBSYS "elan"
+
+#define ELAN_STATS_GET_NEXT    0x01
+typedef struct {
+       ELAN_STATS_IDX statidx;
+       ELAN_STATS_IDX *next_statidx;   
+} elan_stats_get_next_struct;
+
+
+#define ELAN_STATS_FIND_INDEX   0x02
+typedef struct {
+       caddr_t          block_name;
+       ELAN_STATS_IDX  *statidx; /* return value */
+       uint        *num_entries; /* return value */
+} elan_stats_find_index_struct;
+
+#define ELAN_STATS_GET_BLOCK_INFO  0x03
+typedef struct {
+       ELAN_STATS_IDX  statidx; 
+       caddr_t       block_name; /* return value */
+       uint        *num_entries; /* return value */
+} elan_stats_get_block_info_struct;
+
+#define ELAN_STATS_GET_INDEX_NAME  0x04
+typedef struct {
+       ELAN_STATS_IDX statidx; 
+       uint           index;
+       caddr_t        name; /* return value */
+} elan_stats_get_index_name_struct;
+
+#define ELAN_STATS_CLEAR_BLOCK  0x05
+typedef struct {
+       ELAN_STATS_IDX statidx; 
+} elan_stats_clear_block_struct;
+
+#define ELAN_STATS_GET_BLOCK     0x06
+typedef struct 
+{
+       ELAN_STATS_IDX statidx; 
+       uint           entries;  
+       ulong         *values; /* return values */
+} elan_stats_get_block_struct;
+
+#define ELAN_GET_DEVINFO     0x07
+typedef struct 
+{
+       ELAN_DEV_IDX  devidx; 
+       ELAN_DEVINFO *devinfo; /* return values */
+} elan_get_devinfo_struct;
+
+#define ELAN_GET_POSITION  0x08
+typedef struct {
+       ELAN_DEV_IDX   devidx; 
+       ELAN_POSITION *position; /* return values */
+} elan_get_position_struct;
+
+#define ELAN_SET_POSITION   0x09
+typedef struct {
+       ELAN_DEV_IDX   devidx; 
+       unsigned short nodeId;
+       unsigned short numNodes;
+} elan_set_position_struct;
+
+#define ELAN_CREATE_CAP  0x0a
+typedef struct {
+       ELAN_CAPABILITY cap;
+} elan_create_cap_struct;
+
+#define ELAN_DESTROY_CAP    0x0b
+typedef struct {
+       ELAN_CAPABILITY cap;
+} elan_destroy_cap_struct;
+
+#define ELAN_CREATE_VP   0x0c
+typedef struct {
+       ELAN_CAPABILITY cap;
+       ELAN_CAPABILITY map;
+} elan_create_vp_struct;
+
+#define ELAN_DESTROY_VP    0x0d
+typedef struct {
+       ELAN_CAPABILITY cap;
+       ELAN_CAPABILITY map;
+} elan_destroy_vp_struct;
+
+
+#define ELAN_DEBUG_DUMP   0x0e
+
+#define ELAN_GET_CAPS    0x0f
+typedef struct {
+       uint            *number_of_results;
+       uint             array_size;
+       ELAN_CAP_STRUCT *caps;
+} elan_get_caps_struct;
+
+#define ELAN_DEBUG_BUFFER 0x10
+typedef struct {
+       caddr_t addr;
+       int     len;
+} elan_debug_buffer_struct;
+
+#define ELANMOD_PROCFS_IOCTL      "/proc/qsnet/elan/ioctl"
+#define ELANMOD_PROCFS_VERSION    "/proc/qsnet/elan/version"
+#define ELANMOD_PROCFS_DEBUG_MASK "/proc/qsnet/elan/debug_mask"
+#define ELANMOD_PROCFS_DEBUG_MODE "/proc/qsnet/elan/debug_mode"
+
+#endif /* __ELAN_SUBSYSTEM_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/epcomms.h
===================================================================
--- linux-2.4.21.orig/include/elan/epcomms.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/epcomms.h 2005-06-01 23:12:54.710422232 -0400
@@ -0,0 +1,635 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_EPCOMMS_H
+#define __ELAN_EPCOMMS_H
+
+#ident "$Id: epcomms.h,v 1.44.2.2 2004/11/12 10:54:50 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epcomms.h,v $ */
+
+#include <elan/kcomm.h>
+#include <elan/bitmap.h>
+
+#define EPCOMMS_SUBSYS_NAME    "epcomms"
+
+/* message service numbers */
+#define EP_MSG_SVC_EIP512              0x00                            /* Quadrics EIP services */
+#define EP_MSG_SVC_EIP1K               0x01
+#define EP_MSG_SVC_EIP2K               0x02
+#define EP_MSG_SVC_EIP4K               0x03
+#define EP_MSG_SVC_EIP8K               0x04
+#define EP_MSG_SVC_EIP16K              0x05
+#define EP_MSG_SVC_EIP32K              0x06
+#define EP_MSG_SVC_EIP64K              0x07
+#define EP_MSG_SVC_EIP128K             0x08
+
+#define EP_MSG_SVC_PFS                 0x09                            /* Quadrics PFS rpc service */
+
+#define EP_MSG_SVC_PORTALS_SMALL       0x10                            /* Lustre Portals */
+#define EP_MSG_SVC_PORTALS_LARGE       0x11
+
+#define EP_MSG_NSVC                    0x40                            /* Max number of services */
+
+#define EP_MSGQ_ADDR(qnum)             (EP_EPCOMMS_QUEUE_BASE + (qnum) * EP_QUEUE_DESC_SIZE)
+
+/*
+ * EP_ENVELOPE
+ *   Messages are sent by sending an envelope to the destination
+ *   describing the source buffers to transfer.  The receiving thread
+ *   then allocates a receive buffer and fetches the data by issuing
+ *   "get" dmas.
+ *
+ * NOTE:  envelopes are not explicitly converted to network byte order
+ *        since they are always transferred little endian as they are
+ *        copied to/from elan memory using word operations.
+ */
+typedef struct ep_envelope
+{
+    uint32_t     Version;                                      /* Protocol version field */
+
+    EP_ATTRIBUTE  Attr;                                        /* Attributes */
+
+    EP_XID       Xid;                                          /* transaction id */
+
+    uint32_t     NodeId;                                       /* Source processor */
+    uint32_t     Range;                                        /* range we're sending to (high << 16 | low) */
+
+    EP_ADDR      TxdRail;                                      /* address of per-rail txd */
+    EP_NMD       TxdMain;                                      /* address of main memory portion of txd */
+
+    uint32_t      nFrags;                                      /* # fragments */
+    EP_NMD       Frags[EP_MAXFRAG];                            /* network mapping handles of source data */
+
+    uint32_t      CheckSum;                                     /* holds the check sum value when active 
+                                                                * must be after all members to be checksum'd
+                                                                */
+
+    uint32_t     Pad[6];                                       /* Pad to 128 bytes */
+} EP_ENVELOPE;
+
+#define EP_ENVELOPE_VERSION            0xdac10001
+#define EP_ENVELOPE_SIZE               roundup (sizeof (EP_ENVELOPE), EP_BLK_SIZE)
+
+/*
+ * RPC payload - this small amount of data is transfered in
+ * the envelope for RPCs
+ */
+typedef struct ep_payload
+{
+    uint32_t   Data[128/sizeof(uint32_t)];
+} EP_PAYLOAD;
+
+#define EP_PAYLOAD_SIZE                        roundup (sizeof (EP_PAYLOAD), EP_BLK_SIZE)
+
+#define EP_INPUTQ_SIZE                 (EP_ENVELOPE_SIZE + EP_PAYLOAD_SIZE)
+
+/*
+ * EP_STATUSBLK
+ *   RPC completion transfers a status block to the client.
+ */
+typedef struct ep_statusblk
+{
+    uint32_t   Data[128/sizeof(uint32_t)];
+} EP_STATUSBLK;
+
+#define EP_STATUSBLK_SIZE              roundup (sizeof(EP_STATUSBLK), EP_BLK_SIZE)
+
+#define EP_RANGE(low,high)             ((high) << 16 | (low))
+#define EP_RANGE_LOW(range)            ((range) & 0xFFFF)
+#define EP_RANGE_HIGH(range)           (((range) >> 16) & 0xFFFF)
+
+/* return codes from functions, + 'res' parameter to txd callback, ep_rxd_status() */
+typedef enum
+{
+    EP_SUCCESS         = 0,                                    /* message sent/received successfully */
+    EP_RXD_PENDING     = -1,                                   /* rxd not completed by thread */
+    EP_CONN_RESET      = -2,                                   /* virtual circuit reset */
+    EP_NODE_DOWN       = -3,                                   /* node down - transmit not attempted */
+    EP_MSG_TOO_BIG      = -4,                                  /* received message larger than buffer */
+    EP_ENOMEM          = -5,                                   /* memory alloc failed */
+    EP_EINVAL          = -6,                                   /* invalid parameters */
+    EP_SHUTDOWN                = -7,                                   /* receiver is being shut down */
+} EP_STATUS;
+
+/* forward declarations */
+typedef struct ep_rxd          EP_RXD;
+typedef struct ep_txd          EP_TXD;
+typedef struct ep_rcvr_rail    EP_RCVR_RAIL;
+typedef struct ep_rcvr         EP_RCVR;
+typedef struct ep_xmtr_rail    EP_XMTR_RAIL;
+typedef struct ep_xmtr         EP_XMTR;
+typedef struct ep_comms_rail    EP_COMMS_RAIL;
+typedef struct ep_comms_subsys  EP_COMMS_SUBSYS;
+
+typedef struct ep_rcvr_stats           EP_RCVR_STATS;
+typedef struct ep_xmtr_stats           EP_XMTR_STATS;
+typedef struct ep_rcvr_rail_stats      EP_RCVR_RAIL_STATS;
+typedef struct ep_xmtr_rail_stats      EP_XMTR_RAIL_STATS;
+
+typedef void (EP_RXH)(EP_RXD *rxd);                            /* callback function from receive completion */
+typedef void (EP_TXH)(EP_TXD *txd, void *arg, EP_STATUS res);  /* callback function from transmit completion  */
+
+/* Main memory portion shared descriptor */
+typedef struct ep_rxd_main
+{
+    EP_ENVELOPE                Envelope;                               /* 128 byte aligned envelope */
+    EP_PAYLOAD         Payload;                                /* 128 byte aligned payload */
+    bitmap_t           Bitmap[BT_BITOUL(EP_MAX_NODES)];        /* broadcast bitmap */
+    EP_STATUSBLK       StatusBlk;                              /* RPC status block to return */
+    uint64_t           Next;                                   /* linked list when on active list (main address) */
+    int32_t            Len;                                    /* Length of message received */
+} EP_RXD_MAIN;
+
+#define EP_RXD_MAIN_SIZE       roundup (sizeof (EP_RXD_MAIN), EP_BLK_SIZE)
+
+/* Phases for message/rpc */
+#ifndef __ELAN__
+
+/* Kernel memory portion of per-rail receive descriptor */
+typedef struct ep_rxd_rail
+{
+    struct list_head    Link;                                  /* linked on freelist */
+    EP_RCVR_RAIL       *RcvrRail;                              /* rvcr we're associated with */
+    
+    EP_RXD            *Rxd;                                    /* receive descriptor we're bound to */
+} EP_RXD_RAIL;
+
+#define RXD_BOUND2RAIL(rxdRail,rcvrRail)       ((rxdRail) != NULL && ((EP_RXD_RAIL *) (rxdRail))->RcvrRail == (EP_RCVR_RAIL *) rcvrRail)
+
+struct ep_rxd
+{
+    struct list_head   Link;                                   /* linked on free/active list */
+    EP_RCVR           *Rcvr;                                   /* owning receiver */
+
+    EP_RXD_MAIN               *RxdMain;                                /* shared main memory portion. */
+    EP_NMD             NmdMain;                                /*  and network mapping descriptor */
+
+    EP_RXD_RAIL               *RxdRail;                                /* per-rail rxd we're bound to */
+    
+    EP_RXH            *Handler;                                /* completion function */
+    void              *Arg;                                    /*    and arguement */
+
+    unsigned int       State;                                  /* RXD status (active,stalled,failed) */
+
+    EP_NMD             Data;                                   /* network mapping descriptor for user buffer */
+
+    int                        nFrags;                                 /* network mapping descriptor for put/get/complete */
+    EP_NMD             Local[EP_MAXFRAG];
+    EP_NMD             Remote[EP_MAXFRAG];
+
+    long               NextRunTime;                            /* time to resend failover/map requests */
+    EP_XID             MsgXid;                                 /*   and transaction id */
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    struct list_head   CheckSumLink;                           /* linked on check sum list */
+#endif
+};
+
+#define EP_NUM_RXD_PER_BLOCK   16
+
+/* rxd->State */
+#define EP_RXD_FREE            0
+
+#define EP_RXD_RECEIVE_UNBOUND 1
+#define EP_RXD_RECEIVE_ACTIVE  2
+
+#define EP_RXD_PUT_ACTIVE      3
+#define EP_RXD_PUT_STALLED     4
+#define EP_RXD_GET_ACTIVE      5
+#define EP_RXD_GET_STALLED     6
+
+#define EP_RXD_COMPLETE_ACTIVE 7
+#define EP_RXD_COMPLETE_STALLED        8
+
+#define EP_RXD_RPC_IN_PROGRESS 9
+#define EP_RXD_COMPLETED       10      
+
+#define EP_RXD_BEEN_ABORTED    11                              /* rxd was aborted while in a private state */
+
+typedef struct ep_rxd_block
+{
+    struct list_head   Link;
+
+    EP_NMD             NmdMain;
+
+    EP_RXD             Rxd[EP_NUM_RXD_PER_BLOCK];
+} EP_RXD_BLOCK;
+
+struct ep_rcvr_rail_stats 
+{
+    EP_STATS_COUNT rx;
+    EP_STATS_COUNT rx_len;
+};
+
+struct ep_rcvr_rail
+{
+    EP_RCVR           *Rcvr;                                   /* associated receiver */
+    EP_COMMS_RAIL      *CommsRail;                             /* comms rail */
+
+    struct proc_dir_entry *procfs_root;                         /* root of this rcvr_rail's procfs entry */
+    EP_RCVR_RAIL_STATS     stats;                               /* generic rcvr_rail stats */
+};
+
+struct ep_rcvr_stats
+{
+    EP_STATS_COUNT rx;
+    EP_STATS_COUNT rx_len;
+};
+
+struct ep_rcvr
+{
+    struct list_head  Link;                                    /* queued on subsystem */
+    EP_COMMS_SUBSYS  *Subsys;                                  /* kernel comms subsystem */
+    EP_SERVICE        Service;                                 /* service number */
+
+    unsigned int      InputQueueEntries;                       /* # entries on receive queue */
+
+    EP_RAILMASK              RailMask;                                 /* bitmap of which rails are available */
+    EP_RCVR_RAIL     *Rails[EP_MAX_RAILS];
+
+    spinlock_t       Lock;                                     /* spinlock for rails/receive lists */
+
+    struct list_head  ActiveDescList;                          /* List of pending/active receive descriptors */
+
+    EP_XID_CACHE      XidCache;                                        /* XID cache (protected by Lock) */
+
+    struct list_head  FreeDescList;                            /* List of free receive descriptors */
+    unsigned int      FreeDescCount;                           /*   and number on free list */
+    unsigned int      TotalDescCount;                           /*   total number created */
+    spinlock_t       FreeDescLock;                             /*   and lock for free list */
+    kcondvar_t       FreeDescSleep;                            /*   with place to sleep for rx desc */
+    int                      FreeDescWanted;                           /*   and flag */
+    struct list_head  DescBlockList;
+
+    unsigned int      ForwardRxdCount;                         /* count of rxd's being forwarded */
+    unsigned int      CleanupWaiting;                          /* waiting for cleanup */
+    kcondvar_t       CleanupSleep;                             /*   and place to sleep */
+
+    struct proc_dir_entry *procfs_root;                         /* place where this rcvr's proc entry is */
+    EP_RCVR_STATS          stats;                                    
+};
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+#define EP_ENVELOPE_CHECK_SUM      (1<<31)
+extern uint32_t ep_calc_check_sum (EP_SYS *sys, EP_ENVELOPE *env, EP_NMD *nmd, int nFrags);
+#endif
+
+#endif /* ! __ELAN__ */
+
+typedef struct ep_txd_main
+{
+    EP_STATUSBLK      StatusBlk;                               /* RPC status block */
+    bitmap_t          Bitmap[BT_BITOUL(EP_MAX_NODES)];         /* broadcast bitmap */
+} EP_TXD_MAIN;
+
+#define EP_TXD_MAIN_SIZE       roundup (sizeof (EP_TXD_MAIN), EP_BLK_SIZE)
+
+#ifndef __ELAN__
+typedef struct ep_txd_rail
+{
+    struct list_head  Link;                                    /* linked on freelist */
+    EP_XMTR_RAIL     *XmtrRail;                                        /* xmtr we're associated with */
+
+    EP_TXD          *Txd;                                      /* txd we're bound to */
+} EP_TXD_RAIL;
+
+#define TXD_BOUND2RAIL(rxdRail,xmtrRail)       ((txdRail) != NULL && ((EP_TXD_RAIL *) (txdRail))->XmtrRail == (EP_XMTR_RAIL *) xmtrRail)
+
+struct ep_txd
+{
+    struct list_head  Link;                                    /* linked on free/active list */
+    EP_XMTR         *Xmtr;                                     /* service we're associated with */
+
+    EP_TXD_MAIN             *TxdMain;                                  /* shared main memory portion */
+    EP_NMD           NmdMain;                                  /*   and network mapping descriptor */
+
+    EP_TXD_RAIL      *TxdRail;                                 /* per-rail txd for this phase */
+
+    EP_TXH          *Handler;                                  /* completion function */
+    void            *Arg;                                      /*    and arguement */
+    
+    unsigned short    NodeId;                                  /* node transmit is to. */
+    EP_SERVICE        Service;                                 /*    and seervice */
+
+    long              TimeStamp;                                 /* time we where created at, to find sends taking too long */
+    long             RetryTime;
+    EP_BACKOFF       Backoff;
+
+    EP_ENVELOPE              Envelope;                                 /* envelope for transmit */
+    EP_PAYLOAD       Payload;                                  /* payload for transmit */
+};
+
+#define EP_NUM_TXD_PER_BLOCK   16
+
+/* "phase" parameter to BindTxd */
+#define EP_TXD_PHASE_ACTIVE            1
+#define EP_TXD_PHASE_PASSIVE           2
+
+typedef struct ep_txd_block
+{
+    struct list_head   Link;
+    EP_NMD             NmdMain;
+    EP_TXD             Txd[EP_NUM_TXD_PER_BLOCK];              /* transmit descriptors */
+} EP_TXD_BLOCK;
+
+struct ep_xmtr_rail_stats
+{
+    EP_STATS_COUNT tx;
+    EP_STATS_COUNT tx_len;
+};
+
+struct ep_xmtr_rail
+{
+    EP_COMMS_RAIL      *CommsRail;                             /* associated comms rail */
+    EP_XMTR           *Xmtr;                                   /* associated transmitter */
+
+    struct proc_dir_entry *procfs_root;                         /* place where this xmtr's proc entry is */
+
+    EP_XMTR_RAIL_STATS     stats;
+};
+
+struct ep_xmtr_stats
+{
+    EP_STATS_COUNT tx;
+    EP_STATS_COUNT tx_len;
+};
+
+struct ep_xmtr
+{
+    struct list_head  Link;                                    /* Linked on subsys */
+    EP_COMMS_SUBSYS  *Subsys;                                  /* kernel comms subsystem */
+
+    EP_RAILMASK              RailMask;                                 /* bitmap of which rails are available */
+    EP_XMTR_RAIL     *Rails[EP_MAX_RAILS];                     /* per-rail state */
+
+    spinlock_t       Lock;                                     /* lock for active descriptor list */
+
+    struct list_head  ActiveDescList;                          /* list of active transmit descriptors */
+
+    EP_XID_CACHE      XidCache;                                        /* XID cache (protected by Lock) */
+
+    struct list_head  FreeDescList;                            /* List of free receive descriptors */
+    unsigned int      FreeDescCount;                           /*   and number on free list */
+    unsigned int      TotalDescCount;
+    spinlock_t       FreeDescLock;                             /*   and lock for free list */
+    kcondvar_t       FreeDescSleep;                            /*   with place to sleep for rx desc */
+    int                      FreeDescWanted;                           /*   and flag */
+    struct list_head  DescBlockList;
+
+    struct proc_dir_entry *procfs_root;                         /* place where this rcvr's proc entry is */
+    EP_XMTR_STATS          stats;   
+};
+
+/* forward descriptor */
+#define EP_TREE_ARITY          3
+
+typedef struct ep_fwd_desc
+{
+    struct list_head    Link;                                  /* linked on forward/free lists */
+    EP_RXD            *Rxd;                                    /* rxd to forward */
+    EP_NMD             Data;                                   /* nmd of subset of receive buffer */
+    unsigned           NumChildren;                            /*   number of places we're forwarding */
+    unsigned           Children[EP_TREE_ARITY];
+} EP_FWD_DESC;
+
+typedef struct ep_comms_ops
+{
+    void            (*DelRail) (EP_COMMS_RAIL *rail);
+    void            (*DisplayRail) (EP_COMMS_RAIL *rail);
+
+    struct {
+       void         (*AddRail) (EP_RCVR *rcvr, EP_COMMS_RAIL *rail);
+       void         (*DelRail) (EP_RCVR *rcvr, EP_COMMS_RAIL *rail);
+
+       long         (*Check) (EP_RCVR_RAIL *rcvrRail, long nextRunTime);
+
+       int          (*QueueRxd) (EP_RXD *rxd, EP_RCVR_RAIL *rcvrRail);
+       void         (*RpcPut)(EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+       void         (*RpcGet)(EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+       void         (*RpcComplete)(EP_RXD *rxd, EP_NMD *local, EP_NMD *remote, unsigned nFrags);
+
+       EP_RXD      *(*StealRxd)(EP_RCVR_RAIL *rcvrRail);
+
+       void         (*DisplayRcvr) (DisplayInfo *di, EP_RCVR_RAIL *rcvrRail);
+       void         (*DisplayRxd)  (DisplayInfo *di, EP_RXD_RAIL *rxdRail);
+
+       void         (*FillOutRailStats) (EP_RCVR_RAIL *rcvr_rail, char *str);
+
+    } Rcvr;
+
+    struct {
+       void         (*AddRail) (EP_XMTR *xmtr, EP_COMMS_RAIL *rail);
+       void         (*DelRail) (EP_XMTR *xmtr, EP_COMMS_RAIL *rail);
+
+       long         (*Check) (EP_XMTR_RAIL *xmtrRail, long nextRunTime);
+       
+       int          (*BindTxd) (EP_TXD *txd, EP_XMTR_RAIL *xmtrRail, unsigned int phase);
+       void         (*UnbindTxd) (EP_TXD *txd, unsigned int phase);
+       int          (*PollTxd) (EP_XMTR_RAIL *xmtrRail, EP_TXD_RAIL *txdRail, int how);
+
+       void         (*DisplayXmtr) (DisplayInfo *di, EP_XMTR_RAIL *xmtrRail);
+       void         (*DisplayTxd)  (DisplayInfo *di, EP_TXD_RAIL *txdRail);
+
+       int          (*CheckTxdState) (EP_TXD *txd);
+
+       void         (*FillOutRailStats) (EP_XMTR_RAIL *xmtr_rail, char *str);
+
+    } Xmtr;
+} EP_COMMS_OPS;
+
+#define EP_RAIL_OP(commsRail, Which)   (commsRail)->Ops.Which
+#define EP_RCVR_OP(rcvrRail, Which)    (rcvrRail)->CommsRail->Ops.Rcvr.Which
+#define EP_XMTR_OP(xmtrRail, Which)    (xmtrRail)->CommsRail->Ops.Xmtr.Which
+
+/* "how" parameter to PollTxd */
+#define POLL_TX_LIST           0
+#define ENABLE_TX_CALLBACK     1
+#define DISABLE_TX_CALLBACK    2
+
+struct ep_comms_rail
+{
+    struct list_head   Link;                                   /* Linked on subsys */
+    EP_RAIL           *Rail;                                   /* kernel comms rail */
+    EP_COMMS_SUBSYS    *Subsys;
+    EP_COMMS_OPS        Ops;
+
+    EP_COMMS_RAIL_STATS Stats;                                 /* statistics */
+};
+
+struct ep_comms_subsys
+{
+    EP_SUBSYS          Subsys;                                 /* is a kernel comms subsystem */
+
+    kmutex_t           Lock;                                   /* global lock */
+
+    EP_COMMS_STATS     Stats;                                  /* statistics */
+
+    struct list_head   Rails;                                  /* list of all rails */
+
+    struct list_head    Receivers;                             /* list of receivers */
+    struct list_head   Transmitters;                           /* and transmitters */
+
+    /* forward/allocator thread */
+    EP_KTHREAD         Thread;                                 /* place thread sleeps */
+
+    /* message passing "broadcast" forward lists */
+    spinlock_t         ForwardDescLock;                        /* Lock for broadcast forwarding */
+    struct list_head    ForwardDescList;                       /* List of rxd's to forward */
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    spinlock_t         CheckSumDescLock;                       /* Lock for CheckSums */
+    struct list_head    CheckSumDescList;                      /* List of rxd's to be CheckSumed */
+#endif
+
+    EP_XMTR           *ForwardXmtr;                            /* and transmitter to forward with */
+};
+
+/* epcomms.c subsystem initialisation */
+extern unsigned int   epcomms_forward_limit;
+
+extern int           ep_comms_init (EP_SYS *sys);
+extern void           ep_comms_display (EP_SYS *sys, char *how);
+extern EP_RAILMASK    ep_rcvr_railmask (EP_SYS *epsys, EP_SERVICE service);
+
+/* epcomms_elan3.c */
+extern EP_COMMS_RAIL *ep3comms_add_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *rail);
+
+/* epcomms_elan4.c */
+extern EP_COMMS_RAIL *ep4comms_add_rail (EP_SUBSYS *s, EP_SYS *sys, EP_RAIL *rail);
+
+/* epcommsTx.c */
+extern int            TxdShouldStabalise (EP_TXD_RAIL *txdRail, EP_RAIL *rail);
+extern void          FreeTxd (EP_XMTR *xmtr, EP_TXD *txd);
+
+extern unsigned int   ep_txd_lowat;
+extern long           ep_check_xmtr (EP_XMTR *xmtr, long nextRunTime);
+extern void           ep_display_xmtr (DisplayInfo *di, EP_XMTR *xmtr);
+extern void           ep_xmtr_flush_callback (EP_XMTR *xmtr, EP_XMTR_RAIL *xmtrRail);
+extern void           ep_xmtr_reloc_callback (EP_XMTR *xmtr, EP_XMTR_RAIL *xmtrRail);
+
+extern void           ep_xmtr_fillout_stats      (EP_XMTR      *xmtr,      char *str);
+extern void           ep_xmtr_rail_fillout_stats (EP_XMTR_RAIL *xmtr_rail, char *str);
+
+extern void           ep_xmtr_txd_stat (EP_XMTR *xmtr, EP_TXD *txd);
+
+/* epcommsRx.c */
+extern EP_RXD        *StealRxdFromOtherRail (EP_RCVR *rcvr);
+
+extern unsigned int   ep_rxd_lowat;
+extern long          ep_check_rcvr (EP_RCVR *rcvr, long nextRunTime);
+extern void           ep_rcvr_flush_callback (EP_RCVR *rcvr, EP_RCVR_RAIL *rcvrRail);
+extern void           ep_rcvr_reloc_callback (EP_RCVR *rcvr, EP_RCVR_RAIL *rcvrRail);
+extern void           ep_display_rcvr (DisplayInfo *di, EP_RCVR *rcvr, int full);
+
+extern long           ep_forward_rxds (EP_COMMS_SUBSYS *subsys, long nextRunTime);
+
+extern void           ep_rcvr_fillout_stats      (EP_RCVR      *rcvr,      char *str);
+extern void           ep_rcvr_rail_fillout_stats (EP_RCVR_RAIL *rcvr_rail, char *str);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+extern void           ep_csum_rxds    (EP_COMMS_SUBSYS *subsys);
+extern void           ep_rxd_queue_csum (EP_RXD *rxd);
+#endif
+
+extern void           ep_rxd_received     (EP_RXD *rxd);
+extern void           ep_rxd_received_now (EP_RXD *rxd);
+
+/* ep_procfs.c */
+extern struct proc_dir_entry *ep_procfs_root;
+
+extern void ep_procfs_rcvr_xmtr_init(void);
+extern void ep_procfs_rcvr_xmtr_fini(void);
+
+extern void ep_procfs_rcvr_add(EP_RCVR *rcvr);
+extern void ep_procfs_rcvr_del(EP_RCVR *rcvr);
+
+extern void ep_procfs_rcvr_add_rail(EP_RCVR_RAIL *rcvrRail);
+extern void ep_procfs_rcvr_del_rail(EP_RCVR_RAIL *rcvrRail);
+
+extern void ep_procfs_xmtr_add(EP_XMTR *xmtr);
+extern void ep_procfs_xmtr_del(EP_XMTR *xmtr);
+
+extern void ep_procfs_xmtr_add_rail(EP_XMTR_RAIL *xmtrRail);
+extern void ep_procfs_xmtr_del_rail(EP_XMTR_RAIL *xmtrRail);
+
+
+/* Public Interface */
+
+
+/* epcomms.c message xmtr functions */
+extern EP_XMTR       *ep_alloc_xmtr (EP_SYS *sys);
+extern void           ep_free_xmtr (EP_XMTR *xmtr);
+
+extern EP_STATUS      ep_transmit_message (EP_XMTR *xmtr, unsigned int dest, EP_SERVICE service, EP_ATTRIBUTE attr, 
+                                          EP_TXH *handler, void *arg, EP_PAYLOAD *payload,
+                                          EP_NMD *nmd, int nFrag);
+extern EP_STATUS      ep_multicast_message (EP_XMTR *xmtr, unsigned int destLo, unsigned int destHi, bitmap_t *bitmap, 
+                                           EP_SERVICE service, EP_ATTRIBUTE attr, EP_TXH *handler, void *arg, 
+                                           EP_PAYLOAD *payload, EP_NMD *nmd, int nFrag);
+extern EP_STATUS      ep_transmit_rpc (EP_XMTR *xmtr, unsigned int dest, EP_SERVICE service, EP_ATTRIBUTE attr, 
+                                      EP_TXH *handler, void *arg, EP_PAYLOAD *payload,
+                                      EP_NMD *nmd, int nFrag);
+extern EP_STATUS      ep_multicast_forward (EP_XMTR *xmtr, unsigned int dest, EP_SERVICE service, EP_ATTRIBUTE attr, 
+                                           EP_TXH *handler, void *arg, EP_ENVELOPE *env, EP_PAYLOAD *payload, 
+                                           bitmap_t *bitmap, EP_NMD *nmd, int nFrags);
+
+/* epcomms.c functions for use with polled transmits */
+extern int            ep_poll_transmits (EP_XMTR *xmtr);
+extern int            ep_enable_txcallbacks (EP_XMTR *xmtr);
+extern int            ep_disable_txcallbacks (EP_XMTR *xmtr);
+
+/* epcomms.c message rcvr functions */
+extern EP_RCVR       *ep_alloc_rcvr (EP_SYS *sys, EP_SERVICE svc, unsigned int nenvelopes);
+extern void          ep_free_rcvr (EP_RCVR *rcvr);
+
+extern EP_STATUS      ep_queue_receive (EP_RCVR *rcvr, EP_RXH *handler, void *arg, EP_NMD *nmd, EP_ATTRIBUTE attr);
+extern void          ep_requeue_receive (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_NMD *nmd, EP_ATTRIBUTE attr);
+extern EP_STATUS      ep_rpc_put (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_NMD *from, EP_NMD *to, int nFrags);
+extern EP_STATUS      ep_rpc_get (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_NMD *from, EP_NMD *to, int nFrags);
+extern EP_STATUS      ep_complete_rpc (EP_RXD *rxd, EP_RXH *handler, void *arg, EP_STATUSBLK *blk, 
+                                      EP_NMD *from, EP_NMD *to, int nFrags);
+extern void          ep_complete_receive (EP_RXD *rxd);
+
+/* railhints.c */
+extern int            ep_xmtr_bcastrail (EP_XMTR *xmtr, EP_RAILMASK allowedRails);
+extern int            ep_xmtr_prefrail (EP_XMTR *xmtr, EP_RAILMASK allowedRails, unsigned nodeId);
+extern EP_RAILMASK    ep_xmtr_availrails (EP_XMTR *xmtr);
+extern EP_RAILMASK    ep_xmtr_noderails (EP_XMTR *xmtr, unsigned nodeId);
+extern int            ep_rcvr_prefrail (EP_RCVR *rcvr, EP_RAILMASK allowedRails);
+extern EP_RAILMASK    ep_rcvr_availrails (EP_RCVR *rcvr);
+extern EP_RAILMASK    ep_rxd_railmask (EP_RXD *rxd);
+
+/* epcomms.c functions for accessing fields of rxds */
+extern void          *ep_rxd_arg(EP_RXD *rxd);
+extern int            ep_rxd_len(EP_RXD *rxd);
+extern EP_STATUS      ep_rxd_status(EP_RXD *rxd);
+extern int            ep_rxd_isrpc(EP_RXD *rxd);
+extern EP_ENVELOPE   *ep_rxd_envelope(EP_RXD *rxd);
+extern EP_PAYLOAD    *ep_rxd_payload(EP_RXD *rxd);
+extern int            ep_rxd_node(EP_RXD *rxd);
+extern EP_STATUSBLK  *ep_rxd_statusblk(EP_RXD *rxd);
+
+/* functions for accessing fields of txds */
+extern int            ep_txd_node(EP_TXD *txd);
+extern EP_STATUSBLK  *ep_txd_statusblk(EP_TXD *txd);
+
+/* functions for controlling how many processes are using module */
+extern void              ep_mod_dec_usecount (void);
+extern void              ep_mod_inc_usecount (void);
+
+extern EP_RAILMASK ep_xmtr_svc_indicator_railmask (EP_XMTR *xmtr, int svc_indicator, int nodeId);
+extern int ep_xmtr_svc_indicator_bitmap (EP_XMTR *xmtr, int svc_indicator, bitmap_t * bitmap, int low, int nnodes);
+
+#endif /* ! __ELAN__ */
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN_EPCOMMS_H */
+
Index: linux-2.4.21/include/elan/epsvc.h
===================================================================
--- linux-2.4.21.orig/include/elan/epsvc.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/epsvc.h   2005-06-01 23:12:54.710422232 -0400
@@ -0,0 +1,36 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_EPSVC_H
+#define __ELAN_EPSVC_H
+
+#ident "@(#)$Id: epsvc.h,v 1.9 2004/02/13 10:03:27 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/epsvc.h,v $ */
+
+
+#define EP_SVC_NUM_INDICATORS       8
+#define EP_SVC_INDICATOR_MAX_NAME  32
+
+#define EP_SVC_EIP     0
+#define EP_SVC_NAMES   {"eip", "1", "2", "3", "4", "5", "6", "7"};
+
+#if defined(__KERNEL__)
+extern int         ep_svc_indicator_set      (EP_SYS *epsys, int svc_indicator);
+extern int         ep_svc_indicator_clear    (EP_SYS *epsys, int svc_indicator);
+extern int         ep_svc_indicator_is_set   (EP_SYS *epsys, int svc_indicator, int nodeId);
+extern int         ep_svc_indicator_bitmap   (EP_SYS *epsys, int svc_indicator, bitmap_t * bitmap, int low, int nnodes);
+extern EP_RAILMASK ep_svc_indicator_railmask (EP_SYS *epsys, int svc_indicator, int nodeId);
+#endif
+
+#endif /* __ELAN_EPSVC_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/kalloc.h
===================================================================
--- linux-2.4.21.orig/include/elan/kalloc.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/kalloc.h  2005-06-01 23:12:54.710422232 -0400
@@ -0,0 +1,108 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_KALLOC_H
+#define __ELAN3_KALLOC_H
+
+#ident "$Id: kalloc.h,v 1.11 2004/05/19 10:23:59 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kalloc.h,v $ */
+
+#include <elan/rmap.h>
+
+/*
+ * Memory allocator
+ */
+#define LN2_MIN_SIZE   6                                       /* 64 bytes */
+#define LN2_MAX_SIZE   16                                      /* 64k bytes */
+#define NUM_FREELISTS  (LN2_MAX_SIZE-LN2_MIN_SIZE + 1)
+#define MIN_SIZE       (1 << LN2_MIN_SIZE)
+#define MAX_SIZE       (1 << LN2_MAX_SIZE)
+
+#define HASHSHIFT      LN2_MAX_SIZE
+#define NHASH          32
+#define HASH(addr)     (((addr) >> HASHSHIFT) & (NHASH-1))
+
+typedef enum
+{
+    EP_ALLOC_TYPE_PRIVATE_SDRAM,
+    EP_ALLOC_TYPE_PRIVATE_MAIN,
+    EP_ALLOC_TYPE_SHARED_MAIN,
+} EP_ALLOC_TYPE;
+
+typedef struct ep_pool
+{
+    EP_NMH               Handle;                               /* network mapping handle */
+
+    struct list_head     HashBase;                             /* linked on hash lists */
+    struct list_head     HashTop;                              /* linked on hash lists */
+
+    struct list_head     Link[NUM_FREELISTS];                  /* linked on free lists */
+    bitmap_t            *Bitmaps[NUM_FREELISTS];               /* bitmaps for each size */
+
+    union {
+       sdramaddr_t     Sdram;
+       unsigned long   Ptr;
+    } Buffer;
+} EP_POOL;
+
+typedef struct ep_alloc
+{
+    spinlock_t      Lock;
+    
+    EP_ALLOC_TYPE    Type;
+    unsigned int     Perm;
+
+    EP_RMAP         *ResourceMap;
+
+    struct list_head HashBase[NHASH];
+    struct list_head HashTop[NHASH];
+    struct list_head Freelists[NUM_FREELISTS];
+
+    union {
+       struct {
+           EP_SYS             *System;
+           struct list_head    Rails;
+       } Shared;
+       
+       struct {
+           EP_RAIL            *Rail;
+       } Private;
+    } Data;
+} EP_ALLOC;
+
+extern void            ep_display_alloc (EP_ALLOC *alloc);
+
+extern void            ep_alloc_init (EP_RAIL *rail);
+extern void            ep_alloc_fini (EP_RAIL *rail);
+
+extern sdramaddr_t     ep_alloc_memory_elan (EP_RAIL *rail, EP_ADDR addr, unsigned size, unsigned int perm, EP_ATTRIBUTE attr);
+extern void            ep_free_memory_elan (EP_RAIL *rail, EP_ADDR addr);
+
+extern sdramaddr_t     ep_alloc_elan (EP_RAIL *rail, unsigned size, EP_ATTRIBUTE attr, EP_ADDR *addrp);
+extern void            ep_free_elan (EP_RAIL *rail, EP_ADDR addr, unsigned size);
+extern void           *ep_alloc_main (EP_RAIL *rail, unsigned size, EP_ATTRIBUTE attr, EP_ADDR *addr);
+extern void            ep_free_main (EP_RAIL *rail, EP_ADDR addr, unsigned size);
+
+extern sdramaddr_t     ep_elan2sdram (EP_RAIL *rail, EP_ADDR addr);
+extern void            *ep_elan2main (EP_RAIL *rail, EP_ADDR addr);
+
+extern void            ep_shared_alloc_init (EP_SYS *sys);
+extern void            ep_shared_alloc_fini (EP_SYS *sys);
+extern int             ep_shared_alloc_add_rail (EP_SYS *sys, EP_RAIL *rail);
+extern void            ep_shared_alloc_remove_rail (EP_SYS *sys, EP_RAIL *rail);
+
+extern void           *ep_shared_alloc_main (EP_SYS *sys, unsigned size, EP_ATTRIBUTE attr, EP_NMD *nmd);
+extern void            ep_shared_free_main (EP_SYS *sys, EP_NMD *nmd);
+
+#endif /* __ELAN_KALLOC_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/kcomm.h
===================================================================
--- linux-2.4.21.orig/include/elan/kcomm.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/kcomm.h   2005-06-01 23:12:54.712421928 -0400
@@ -0,0 +1,839 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_KCOMM_H
+#define __ELAN_KCOMM_H
+
+#ident "$Id: kcomm.h,v 1.71.2.8 2004/12/14 10:19:14 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm.h,v $*/
+#define EP_KCOMM_MAJOR_VERSION         3
+#define EP_KCOMM_MINOR_VERSION         1
+
+#define EP_PROTOCOL_VERSION            1                       /* CM/KCOMM protocol revision */
+
+#define EP_MAX_NODES                   2048                    /* Max nodes we support */
+#define EP_MAX_RAILS                   16                      /* max number of rails (we use an unsigned short for bitmaps !) */
+#define EP_MAXFRAG                     4                       /* max number of fragments */
+
+#define EP_BLK_SIZE                    64                      /* align objects for elan access */
+
+/* Elan virtual address address space */
+#define EP_SYSTEM_QUEUE_BASE           0x00010000              /* Base address for system queues */
+#define EP_MSGSYS_QUEUE_BASE           0x00020000              /* Base address for msgsys queues */
+#define EP_EPCOMMS_QUEUE_BASE          0x00030000              /* Base address for message queues */
+#define EP_DVMA_BASE                   0x10000000              /* elan address range for dvma mapping. */
+#define EP_DVMA_TOP                    0xE0000000
+
+#define EP_SHARED_BASE                 0xE0000000              /* shared main/elan allocators */
+#define EP_SHARED_TOP                  0xF0000000
+
+#define EP_PRIVATE_BASE                        0xF0000000              /* private main/elan allocators */
+#define EP_PRIVATE_TOP                 0xF8000000
+
+#define EP_DVMA_RMAP_SIZE              1024                    /* size of resource map for dvma address space */
+#define EP_SHARED_RMAP_SIZE            1024                    /* size of resource map for shared address space */
+#define EP_PRIVATE_RMAP_SIZE           1024                    /* size of resource map for private address space */
+
+/* Input queue descriptors fit into 64 bytes */
+#define EP_QUEUE_DESC_SIZE             64
+
+/* Timeouts for checking network position */
+#define EP_POSITION_TIMEOUT            (4*HZ)          /* 1s   time to notice CheckNetworkPosition changes */
+#define EP_WITHDRAW_TIMEOUT            (2*HZ)          /* 2s   time before withdrawing from unreachable nodes */
+
+/* Time to try again due to resource failue (eg malloc etc) */
+#define RESOURCE_RETRY_TIME            (HZ/20)
+
+/* Time to retransmit message when send failed */
+#define MSGBUSY_RETRY_TIME             (HZ/20)
+
+/* Time between retransmits of messages network flush requests */
+#define MESSAGE_RETRY_TIME             (HZ/5)
+
+/* time to hold the context filter up to ensure that the
+ * next packet of a dma is guaranteed to get nacked (8mS) */
+#define NETWORK_ERROR_TIMEOUT          (1 + roundup (HZ * 8 / 1000, 1))
+
+/* Time between retransmits of message failover requests */
+#define FAILOVER_RETRY_TIME            (HZ/5)
+
+/* compute earliest time */
+#define SET_NEXT_RUN_TIME(nextRunTime, time) \
+do { \
+    if ((nextRunTime) == 0 || AFTER(nextRunTime, (time)))\
+       (nextRunTime) = (time);\
+} while (0)
+
+/* DMA retry backoff/priorities/issue rings */
+#define EP_NUM_BACKOFF                 8
+#define EP_RETRY_STABALISING            0
+#define EP_RETRY_BASE                  1
+
+#define EP_RETRY_CRITICAL              EP_RETRY_BASE
+#define EP_RETRY_HIGH_PRI              (EP_RETRY_CRITICAL + 1)
+#define EP_RETRY_HIGH_PRI_TIME         (1)
+#define EP_RETRY_HIGH_PRI_RETRY                (EP_RETRY_HIGH_PRI + 1)
+#define EP_RETRY_HIGH_PRI_RETRY_TIME   (2)
+#define EP_RETRY_LOW_PRI               (EP_RETRY_HIGH_PRI_RETRY + EP_NUM_BACKOFF)
+#define EP_RETRY_LOW_PRI_TIME          (2)
+#define EP_RETRY_LOW_PRI_RETRY         (EP_RETRY_LOW_PRI + 1)
+#define EP_RETRY_LOW_PRI_RETRY_TIME    (4)
+#define EP_RETRY_ANONYMOUS             (EP_RETRY_LOW_PRI_RETRY + EP_NUM_BACKOFF)
+#define EP_RETRY_ANONYMOUS_TIME                (10)
+#define EP_RETRY_NETERR                        (EP_RETRY_ANONYMOUS + EP_NUM_BACKOFF)
+#define EP_RETRY_NETERR_TIME           (10)
+#define EP_NUM_RETRIES                 (EP_RETRY_NETERR + 1)
+
+typedef unsigned short EP_SERVICE;
+
+/* EP_ATTRIBUTE 32 bits 
+ *
+ * 0-2
+ *   for initial call :-
+ *     0 (0x1) EP_NO_ALLOC                             used once
+ *     1 (0x2) EP_NO_SLEEP                             used once
+ *     2 (0x4) EP_NOT_MYSELF                           used once
+ *
+ *   when stored and transmited :-
+ *     0 (0x0) EP_MULTICAST                            envelope
+ *     1 (0x2) EP_RPC                                  envelope
+ *     2 (0x4) EP_HAS_PAYLOAD                          envelope
+ *
+ * 3-11
+ *     3   (0x08) EP_PREFRAIL_SET                      preserved
+ *     4-7 (0xf0) Pref Rail
+ *     8  (0x100) EP_NO_INTERUPT
+ *     9  (0x200) EP_NO_FAILOVER
+ *
+ *    10 (0x400) EP_INTERRUPT_ENABLED                  internal
+ *    11 (0x800) EP_TXD_STABALISING                    internal
+ *
+ * 12-13 Not Used.
+ * 
+ * 14-15 (0xC000) Data Type.                           passed in
+ *    00 none. 
+ *    01 Service Indicator.
+ *    10 TimeOut.
+ *    11 RailMask
+ *          
+ * 16-31 (0x10000)  Data.  Service Indicator, TimeOut, RailMask, Pref Rail.
+ *         
+*/
+
+typedef uint32_t EP_ATTRIBUTE;
+
+#define EP_LOCAL_ATTR_MASK 0x07
+#define EP_CLEAR_LOCAL_ATTR(ATTR)  ( (ATTR) & ~EP_LOCAL_ATTR_MASK )
+
+#define EP_NO_ALLOC      0x01  /* Don't call allocators if no free descriptors */
+#define EP_NO_SLEEP      0x02  /* Don't sleep if no free descriptors */
+#define EP_NOT_MYSELF    0x04  /* Don't send multicast to me */
+
+#define EP_MULTICAST         0x01      /* Message is a multicast */
+#define EP_RPC               0x02      /* Wait for RPC reply */
+#define EP_HAS_PAYLOAD_BIT   0x04      /* transfer payload */
+
+
+#define EP_PREFRAIL_SET  0x08  /* preferred rail is set (otherwise pick one from the NMDs) */
+
+#define EP_PREFRAIL_SHIFT  (4)
+#define EP_PREFRAIL_MASK   0xf0
+#define EP_IS_PREFRAIL_SET(ATTR)      (((ATTR) &  EP_PREFRAIL_SET) != 0)
+#define EP_CLEAR_PREFRAIL(ATTR)       (((ATTR) & ~EP_PREFRAIL_SET) & ~EP_PREFRAIL_MASK) 
+#define EP_SET_PREFRAIL(ATTR,RAIL)    (EP_CLEAR_PREFRAIL(ATTR) | (((RAIL) <<  EP_PREFRAIL_SHIFT ) &  EP_PREFRAIL_MASK ) |  EP_PREFRAIL_SET)
+
+
+#define EP_ATTR2PREFRAIL(ATTR)            (((ATTR) & EP_PREFRAIL_MASK) >> EP_PREFRAIL_SHIFT)
+
+
+#define EP_INTERRUPT_ENABLED 0x400     /* event interrupt enabled on EP_NO_INTERRUPT */
+#define EP_TXD_STABALISING   0x800      /* flag to indicate this is attempting to stabalise */
+
+#define EP_IS_MULTICAST(ATTR)                 (((ATTR) &  EP_MULTICAST) != 0)
+#define EP_SET_MULTICAST(ATTR)                ( (ATTR) |  EP_MULTICAST)
+#define EP_CLEAR_MULTICAST(ATTR)              ( (ATTR) & ~EP_MULTICAST)
+
+#define EP_IS_RPC(ATTR)                       (((ATTR) &  EP_RPC) != 0)
+#define EP_SET_RPC(ATTR)                      ( (ATTR) |  EP_RPC)
+#define EP_CLEAR_RPC(ATTR)                    ( (ATTR) & ~EP_RPC)
+
+#define EP_HAS_PAYLOAD(ATTR)                  (((ATTR) &  EP_HAS_PAYLOAD_BIT) != 0)
+#define EP_SET_HAS_PAYLOAD(ATTR)              ( (ATTR) |  EP_HAS_PAYLOAD_BIT)
+#define EP_CLEAR_HAS_PAYLOAD(ATTR)            ( (ATTR) & ~EP_HAS_PAYLOAD_BIT)
+
+#define EP_IS_INTERRUPT_ENABLED(ATTR)         (((ATTR) &  EP_INTERRUPT_ENABLED) != 0)
+#define EP_SET_INTERRUPT_ENABLED(ATTR)        ( (ATTR) |  EP_INTERRUPT_ENABLED)
+#define EP_CLEAR_INTERRUPT_ENABLED(ATTR)      ( (ATTR) & ~EP_INTERRUPT_ENABLED)
+
+#define EP_IS_TXD_STABALISING(ATTR)           (((ATTR) &  EP_TXD_STABALISING) != 0)
+#define EP_SET_TXD_STABALISING(ATTR)          ( (ATTR) |  EP_TXD_STABALISING)
+#define EP_CLEAR_TXD_STABALISING(ATTR)        ( (ATTR) & ~EP_TXD_STABALISING)
+
+#define EP_NO_INTERRUPT      0x100     /* Don't generate completion interrupt (tx) */
+#define EP_NO_FAILOVER       0x200     /* don't attempt rail failover, just abort */
+
+#define EP_IS_NO_INTERRUPT(ATTR)    (((ATTR) &  EP_NO_INTERRUPT) != 0)
+#define EP_SET_NO_INTERRUPT(ATTR)   ( (ATTR) |  EP_NO_INTERRUPT)
+#define EP_CLEAR_NO_INTERRUPT(ATTR) ( (ATTR) & ~EP_NO_INTERRUPT)
+
+#define EP_IS_NO_FAILOVER(ATTR)    (((ATTR) &  EP_NO_FAILOVER) != 0)
+#define EP_SET_NO_FAILOVER(ATTR)   ( (ATTR) |  EP_NO_FAILOVER)
+#define EP_CLEAR_NO_FAILOVER(ATTR) ( (ATTR) & ~EP_NO_FAILOVER)
+
+#define EP_TYPE_MASK           0xC000
+#define EP_TYPE_SVC_INDICATOR  0x4000
+#define EP_TYPE_TIMEOUT        0x8000
+#define EP_TYPE_RAILMASK       0xC000
+
+#define EP_ATTR2TYPE(ATTR)        ( (ATTR) & EP_TYPE_MASK )
+
+#define EP_IS_SVC_INDICATOR(ATTR) (EP_ATTR2TYPE(ATTR) == EP_TYPE_SVC_INDICATOR)
+#define EP_IS_TIMEOUT(ATTR)       (EP_ATTR2TYPE(ATTR) == EP_TYPE_TIMEOUT)
+#define EP_IS_RAILMASK(ATTR)      (EP_ATTR2TYPE(ATTR) == EP_TYPE_RAILMASK)
+#define EP_IS_NO_TYPE(ATTR)       (EP_ATTR2TYPE(ATTR) == 0)
+
+#define EP_DATA_SHIFT          (16)
+#define EP_DATA_MASK           0xffff0000
+
+#define EP_ATTR2DATA(ATTR)     (((ATTR) & EP_DATA_MASK) >> EP_DATA_SHIFT)
+#define EP_DATA2ATTR(DATA)     (((DATA) <<  EP_DATA_SHIFT) & EP_DATA_MASK)
+
+#define EP_CLEAR_DATA(ATTR)    (((ATTR) & ~EP_TYPE_MASK) & ~EP_DATA_MASK)
+#define EP_SET_DATA(ATTR,TYPE,DATA)   (EP_CLEAR_DATA(ATTR) | ((TYPE) & EP_TYPE_MASK) | (((DATA) <<  EP_DATA_SHIFT) & EP_DATA_MASK))
+
+#define EP_DEFAULT_TIMEOUT     (HZ*30)
+
+#if !defined(offsetof)
+#define offsetof(s, m)         (unsigned long)(&(((s *)0)->m))
+#endif
+#if !defined(roundup)
+#define roundup(x, y)          ((((x)+((y)-1))/(y))*(y))
+#endif
+
+/* 
+ * Message transaction ID's - these are unique 64 bts 
+ * numbers which include the initial rail number.
+ */
+typedef struct ep_xid
+{
+    uint32_t   Generation;
+    uint32_t   Handle;
+    uint64_t   Unique;
+} EP_XID;
+
+#define EP_INVALIDATE_XID(xid) ((xid).Generation = (xid).Handle = (xid).Unique = 0)
+
+#define EP_XID_INVALID(xid)    ((xid).Generation == 0 && (xid).Handle == 0 && (xid).Unique == 0)
+#define EP_XIDS_MATCH(a,b)     ((a).Generation == (b).Generation && (a).Handle == (b).Handle && (a).Unique == (b).Unique)
+
+typedef struct ep_backoff
+{
+    unsigned char      type;
+    unsigned char      indx;
+    unsigned short     count;
+} EP_BACKOFF;
+
+/* values for "type" */
+#define EP_BACKOFF_FREE                0
+#define EP_BACKOFF_ENVELOPE    1
+#define EP_BACKOFF_FETCH       2
+#define EP_BACKOFF_DATA                3
+#define EP_BACKOFF_DONE                4
+#define EP_BACKOFF_STABILISE   5
+
+#ifndef __ELAN__
+
+/* forward declaration of types */
+typedef struct ep_rail EP_RAIL;
+typedef struct ep_sys  EP_SYS;
+
+#include <elan/nmh.h>
+#include <elan/kmap.h>
+#include <elan/statemap.h>
+#include <elan/kalloc.h>
+#include <elan/kthread.h>
+#include <elan/kcomm_stats.h>
+#include <elan/devinfo.h>
+
+typedef struct ep_callback
+{
+    struct ep_callback *Next;
+    void              (*Routine)(void *, statemap_t *);
+    void              *Arg;
+} EP_CALLBACK;
+
+#define EP_CB_FLUSH_FILTERING          0
+#define EP_CB_FLUSH_FLUSHING           1
+#define EP_CB_PASSIVATED               2
+#define EP_CB_FAILOVER                 3
+#define EP_CB_DISCONNECTING            4
+#define EP_CB_DISCONNECTED             5
+#define EP_CB_NODESET                  6
+#define EP_CB_COUNT                    7
+
+#endif /* !defined(__ELAN__) */
+
+/* Small unreliable system message queues */
+#define EP_SYSTEMQ_INTR                        0                       /* input queue for cluster membership generating an interrupt */
+#define EP_SYSTEMQ_POLLED              1                       /* input queue for cluster membership polled on clock tick */
+#define EP_SYSTEMQ_MANAGER             2                       /* input queue for manager messages */
+#define EP_NUM_SYSTEMQ                 64
+
+#define EP_SYSTEMQ_ADDR(qnum)          (EP_SYSTEM_QUEUE_BASE + (qnum) * EP_QUEUE_DESC_SIZE)
+#define EP_SYSTEMQ_DESC(base,qnum)     ((base) + (qnum) * EP_QUEUE_DESC_SIZE)
+
+#define EP_SYSTEMQ_MSG_ALIGN           64                      /* message sizes aligned to 64 byte boundaries */
+#define EP_SYSTEMQ_MSG_MAX             (4*64)                  /* max message size */
+
+/* Special flag for Version field to indicate message not
+ * seen in main memory yet and time limit to poll for it */
+#define EP_SYSTEMQ_UNRECEIVED                  0xdeadbabe
+#define EP_SYSTEMQ_UNRECEIVED_TLIMIT           16384                   /* 1023 uS */
+
+#ifndef __ELAN__
+
+typedef void (EP_INPUTQ_HANDLER) (EP_RAIL *rail, void *arg, void *msg);
+typedef void (EP_INPUTQ_CALLBACK) (EP_RAIL *rail, void *arg);
+
+typedef struct ep_inputq
+{
+    unsigned long      q_hidden;                               /* implementation hidden as ep3 or ep4 */
+} EP_INPUTQ;
+
+typedef struct ep_outputq
+{
+    unsigned long      q_hidden;                               /* implementation hidden as ep3 or ep4 */
+} EP_OUTPUTQ;
+
+/* returned values for ep_outputq_state */
+#define EP_OUTPUTQ_BUSY                0
+#define EP_OUTPUTQ_FAILED      1
+#define EP_OUTPUTQ_FINISHED    2
+
+typedef struct ep_switch
+{
+    unsigned    present:1;
+    unsigned   invalid:1;
+    unsigned   link:3;
+    unsigned   bcast:3;
+    unsigned   lnr;
+} EP_SWITCH;
+
+/*
+ * Network error fixup, flush, relocation messges
+ */
+typedef struct ep_map_nmd_body
+{
+    uint32_t           nFrags;
+    EP_RAILMASK                Railmask;
+    EP_NMD             Nmd[EP_MAXFRAG];
+} EP_MAP_NMD_BODY;
+
+typedef struct ep_failover_body
+{
+    EP_XID             Xid;
+    EP_RAILMASK                Railmask;
+} EP_FAILOVER_BODY;
+
+typedef struct ep_failover_txd
+{
+    EP_XID             Xid;
+    uint32_t           Rail;
+    EP_ADDR            TxdRail;
+} EP_FAILOVER_TXD;
+
+typedef uint64_t EP_NETERR_COOKIE;
+
+#define EP_PANIC_STRLEN                31
+
+typedef struct ep_node_state
+{
+    unsigned char       State;
+    unsigned char       NetworkErrorState;
+    EP_RAILMASK         Railmask;
+} EP_NODE_STATE;
+
+#define EP_MANAGER_MSG_SIZE            (2 * EP_SYSTEMQ_MSG_ALIGN)
+
+typedef struct ep_manager_msg_hdr
+{
+    EP_XID             Xid;                                    /* Message transaction id */
+
+    uint16_t           NodeId;                                 /* Originating node number */
+    uint16_t           DestId;                                 /* destination node id */
+
+    uint16_t           Checksum;                               /* Message checksum */
+    uint8_t            Rail;                                   /* Rail message associated with */
+    uint8_t            Type;                                   /* Message type */
+
+    uint32_t           Pad;                                    /* pad to 32 bytes */
+
+    uint32_t           Version;                                /* Message Version */
+} EP_MANAGER_MSG_HDR;
+
+typedef union ep_manager_msg_body
+{
+    unsigned char       Space[EP_MANAGER_MSG_SIZE - sizeof (EP_MANAGER_MSG_HDR)];
+
+    EP_NETERR_COOKIE   Cookies[2];                             /* EP_MSG_TYPE_NETERR */
+    EP_MAP_NMD_BODY    MapNmd;                                 /* EP_MSG_TYPE_MAP_NMD */
+    EP_FAILOVER_BODY   Failover;                               /* EP_MSG_TYPE_FAILOVER_REQUEST */
+    EP_FAILOVER_TXD    FailoverTxd;                            /* EP_MSG_TYPE_FAILOVER_RESPONSE */
+    unsigned char       PanicReason[EP_PANIC_STRLEN+1];                /* EP_MSG_TYPE_REMOTE_PANIC */
+    EP_NODE_STATE       NodeState;                              /* EP_MSG_TYPE_GET_NODE_STATE_RESPONSE */   
+    EP_SERVICE          Service;                                /* EP_MSG_TYPE_GET_NODE_STATE */
+} EP_MANAGER_MSG_BODY;
+
+typedef struct ep_manager_msg
+{
+    EP_MANAGER_MSG_BODY Body;
+    EP_MANAGER_MSG_HDR  Hdr;
+} EP_MANAGER_MSG;
+
+#define EP_MANAGER_MSG_VERSION                         0xcad01000
+#define EP_MANAGER_MSG_TYPE_REMOTE_PANIC               0x00
+#define EP_MANAGER_MSG_TYPE_NETERR_REQUEST             0x01
+#define EP_MANAGER_MSG_TYPE_NETERR_RESPONSE            0x02
+#define EP_MANAGER_MSG_TYPE_FLUSH_REQUEST              0x03
+#define EP_MANAGER_MSG_TYPE_FLUSH_RESPONSE             0x04
+#define EP_MANAGER_MSG_TYPE_MAP_NMD_REQUEST            0x05
+#define EP_MANAGER_MSG_TYPE_MAP_NMD_RESPONSE           0x06
+#define EP_MANAGER_MSG_TYPE_FAILOVER_REQUEST           0x07
+#define EP_MANAGER_MSG_TYPE_FAILOVER_RESPONSE          0x08
+#define EP_MANAGER_MSG_TYPE_GET_NODE_STATE              0x09
+#define EP_MANAGER_MSG_TYPE_GET_NODE_STATE_RESPONSE     0x0a
+
+/* Message types which should only be sent when a rail is connected */
+#define EP_MANAGER_MSG_TYPE_CONNECTED(type)            (((type) & 1) == 1)
+
+#define EP_MANAGER_OUTPUTQ_SLOTS       128                     /* # entries in outputq */
+#define EP_MANAGER_INPUTQ_SLOTS                128                     /* # entries in inputq */
+#define EP_MANAGER_OUTPUTQ_RETRIES     31                      /* # retries for manager messages */
+
+/* XID's are allocated from a cache, which doesn't
+ * require locking since it relies on the caller to
+ * manage the locking for us.
+ */
+typedef struct ep_xid_cache
+{
+    struct list_head   Link;
+
+    uint32_t           Handle;                                 /* my XID cache handle */
+    uint64_t           Current;                                /* range of XID.Unique we can allocate from */
+    uint64_t           Last;
+
+    void             (*MessageHandler)(void *arg, EP_MANAGER_MSG *);
+    void              *Arg;
+} EP_XID_CACHE;
+
+#define EP_XID_CACHE_CHUNKS    (10000)
+
+typedef struct ep_node_rail
+{
+    struct list_head    Link;                                  /* can be linked on work lists */
+
+    unsigned char       State;                                 /* node connection state */
+    unsigned char      NetworkErrorState;                      /* reasons for keeping the context filter up */
+    unsigned char      MessageState;                           /* state of messages during passivate/relocate */
+
+    EP_XID             MsgXid;                                 /* neterr/flush transaction id */
+    long               NextRunTime;                            /* time to drop context filter for destroyed dma packet, or to send next request */
+    EP_NETERR_COOKIE   NetworkErrorCookies[2];                 /* identify cookie for destroyed atomic packet */
+
+    uint32_t           Cookie;                                 /* per-node network error cookie */
+    spinlock_t         CookieLock;                             /* and spinlock for it. */
+
+    struct list_head    StalledDmas;                           /* list of stalled DMAs */
+} EP_NODE_RAIL;
+
+#define EP_NODE_DISCONNECTED           0                       /* node is disconnected */
+#define EP_NODE_CONNECTING             1                       /* awaiting connection */
+#define EP_NODE_CONNECTED              2                       /* node is connected */
+#define EP_NODE_LEAVING_CONNECTED      3                       /* node is starting to disconnect */
+#define EP_NODE_LOCAL_PASSIVATE        4                       /* flushing context filter/run queues */
+#define EP_NODE_REMOTE_PASSIVATE       5                       /* stalling for neterr flush */
+#define EP_NODE_PASSIVATED             6                       /* relocating active/passive messages */
+#define EP_NODE_DISCONNECTING          7                       /* entering disconncted - abort remaining comms */
+#define EP_NODE_NUM_STATES             8
+
+#define EP_NODE_NETERR_ATOMIC_PACKET   (1 << 0)
+#define EP_NODE_NETERR_DMA_PACKET      (1 << 1)
+
+#define EP_NODE_PASSIVE_MESSAGES       (1 << 0)
+#define EP_NODE_ACTIVE_MESSAGES                (1 << 1)
+
+/*
+ * Kernel thread code is loaded as a table.
+ */
+typedef struct ep_symbol
+{
+    char   *name;
+    EP_ADDR value;
+} EP_SYMBOL;
+
+typedef struct ep_code
+{
+    u_char        *text;
+    u_int         text_size;
+    u_char        *data;
+    u_int         data_size;
+    u_char        *rodata;
+    u_int         rodata_size;
+    EP_SYMBOL     *symbols;
+    
+    int                   ntext;
+    sdramaddr_t    pptext;
+    EP_ADDR       etext;
+    sdramaddr_t   _stext;
+    sdramaddr_t          _rodata;
+
+    int                   ndata;
+    sdramaddr_t    ppdata;
+    EP_ADDR       edata;
+    sdramaddr_t   _sdata;
+} EP_CODE;
+
+typedef struct ep_switchstate
+{
+    unsigned char       linkid;
+    unsigned char       LNR;
+    unsigned char       bcast;
+    unsigned char       uplink;
+} EP_SWITCHSTATE;
+
+typedef struct ep_rail_ops
+{
+    void       (*DestroyRail) (EP_RAIL *rail);
+
+    int        (*StartRail) (EP_RAIL *rail);
+    void       (*StallRail) (EP_RAIL *rail);
+    void       (*StopRail) (EP_RAIL *rail);
+
+    sdramaddr_t (*SdramAlloc) (EP_RAIL *rail, EP_ADDR addr, unsigned size);
+    void        (*SdramFree) (EP_RAIL *rail, sdramaddr_t addr, unsigned size);
+    void        (*SdramWriteb) (EP_RAIL *rail, sdramaddr_t addr, unsigned char val);
+    
+    void       (*KaddrMap) (EP_RAIL *rail, EP_ADDR eaddr, virtaddr_t kaddr, unsigned len, unsigned int perm, int ep_attr);
+    void       (*SdramMap) (EP_RAIL *rail, EP_ADDR eaddr, sdramaddr_t saddr, unsigned len, unsigned int perm, int ep_attr);
+    void       (*Unmap) (EP_RAIL *rail, EP_ADDR eaddr, unsigned len);
+
+    void       *(*DvmaReserve) (EP_RAIL *rail, EP_ADDR eaddr, unsigned npages);
+    void       (*DvmaRelease) (EP_RAIL *rail, EP_ADDR eaddr, unsigned npages, void *private);
+    void       (*DvmaSetPte) (EP_RAIL *rail, void *private, unsigned index, physaddr_t phys, unsigned int perm);
+    physaddr_t (*DvmaReadPte) (EP_RAIL *rail, void *private, unsigned index);
+    void       (*DvmaUnload)(EP_RAIL *rail, void *private, unsigned index, unsigned npages);
+    void       (*FlushTlb) (EP_RAIL *rail);
+
+    int        (*ProbeRoute) (EP_RAIL *r, int level, int sw, int nodeid, int *linkup, 
+                              int *linkdown, int attempts, EP_SWITCH *lsw);
+    void       (*PositionFound) (EP_RAIL *rail, ELAN_POSITION *pos);
+    int                (*CheckPosition) (EP_RAIL *rail);
+    void       (*NeterrFixup) (EP_RAIL *rail, unsigned int nodeId, EP_NETERR_COOKIE *cookies);
+
+    void       (*LoadSystemRoute) (EP_RAIL *rail, unsigned int vp, unsigned int lowNode, unsigned int highNode);
+
+    void       (*LoadNodeRoute) (EP_RAIL *rail, unsigned nodeId);
+    void       (*UnloadNodeRoute) (EP_RAIL *rail, unsigned nodeId);
+    void       (*LowerFilter) (EP_RAIL *rail, unsigned nodeId);
+    void       (*RaiseFilter) (EP_RAIL *rail, unsigned nodeId);
+    void       (*NodeDisconnected) (EP_RAIL *rail, unsigned nodeId);
+
+    void       (*FlushFilters) (EP_RAIL *rail);
+    void       (*FlushQueues) (EP_RAIL *rail);
+
+
+    EP_INPUTQ  *(*AllocInputQ) (EP_RAIL *rail, unsigned qnum, unsigned slotSize, unsigned slotCount,
+                               void (*callback)(EP_RAIL *rail, void *arg), void *arg);
+    void       (*FreeInputQ) (EP_RAIL *rail, EP_INPUTQ *q);
+    void       (*EnableInputQ) (EP_RAIL *rail, EP_INPUTQ *q);
+    void       (*DisableInputQ) (EP_RAIL *rail, EP_INPUTQ *q);
+    int                (*PollInputQ) (EP_RAIL *rail, EP_INPUTQ *q, int maxCount, EP_INPUTQ_HANDLER *handler, void *arg);
+
+    EP_OUTPUTQ *(*AllocOutputQ) (EP_RAIL *rail, unsigned slotSize, unsigned slotCount);
+    void       (*FreeOutputQ) (EP_RAIL *rail, EP_OUTPUTQ *outputq);
+    void       *(*OutputQMsg) (EP_RAIL *rail, EP_OUTPUTQ *outputq, unsigned slotNum);
+    int         (*OutputQState) (EP_RAIL *rail, EP_OUTPUTQ *outputq, unsigned slotNum);
+    int                (*OutputQSend) (EP_RAIL *rail, EP_OUTPUTQ *outputq, unsigned slotNum, unsigned size,
+                               unsigned vp, unsigned qnum, unsigned retries);
+
+    void        (*FillOutStats) (EP_RAIL *rail, char *str);
+    void       (*Debug) (EP_RAIL *rail);
+
+} EP_RAIL_OPS;
+
+#define ep_alloc_inputq(rail,qnum,slotSize,slotCount,callback,arg) \
+       (rail)->Operations.AllocInputQ(rail,qnum,slotSize,slotCount,callback,arg)
+#define ep_free_inputq(rail,inputq) \
+       (rail)->Operations.FreeInputQ(rail,inputq)
+#define ep_enable_inputq(rail,inputq) \
+       (rail)->Operations.EnableInputQ(rail,inputq)
+#define ep_disable_inputq(rail,inputq) \
+       (rail)->Operations.DisableInputQ(rail,inputq)
+#define ep_poll_inputq(rail,inputq,maxCount,handler,arg) \
+       (rail)->Operations.PollInputQ(rail,inputq,maxCount,handler,arg)
+#define ep_alloc_outputq(rail,slotSize,slotCount)\
+       (rail)->Operations.AllocOutputQ(rail,slotSize,slotCount)
+#define ep_free_outputq(rail,outputq)\
+       (rail)->Operations.FreeOutputQ(rail,outputq)
+#define ep_outputq_msg(rail,outputq,slotNum)\
+       (rail)->Operations.OutputQMsg(rail,outputq,slotNum)
+#define ep_outputq_state(rail,outputq,slotNum)\
+       (rail)->Operations.OutputQState(rail,outputq,slotNum)
+#define ep_outputq_send(rail,outputq,slotNum,size,vp,qnum,retries)\
+       (rail)->Operations.OutputQSend(rail,outputq,slotNum,size,vp,qnum,retries)
+
+struct ep_rail
+{
+    EP_SYS            *System;                                 /* "system" we've attached to */
+
+    unsigned char      Number;                                 /* Rail number */
+    unsigned char       State;                                 /* Rail state */
+    char               Name[32];                               /* Rail name */
+
+    struct list_head    ManagerLink;                           /* linked on ManagedRails list */
+
+    ELAN_DEVINFO       Devinfo;                                /* Device information for this rail */
+    ELAN_POSITION       Position;                              /* Position on switch device is connected to */
+
+    EP_RAIL_OPS                Operations;                             /* device specific operations */
+    EP_RAIL_STATS      Stats;                                  /* statistics */
+
+    EP_ALLOC            ElanAllocator;                         /* per-rail elan memory allocator */
+    EP_ALLOC            MainAllocator;                         /* per-rail main memory allocator */
+
+    unsigned           TlbFlushRequired;                       /* lazy TLB flushing */
+
+    int                SwitchBroadcastLevel;                   /* current switch level ok for broadcast */
+    unsigned long       SwitchBroadcastLevelTick;
+
+    int                        SwitchProbeLevel;                       /* result of last switch probe */
+    EP_SWITCHSTATE      SwitchState[ELAN_MAX_LEVELS];
+    EP_SWITCHSTATE      SwitchLast[ELAN_MAX_LEVELS];
+    unsigned long       SwitchProbeTick[ELAN_MAX_LEVELS];
+    
+    /* Node disconnecting/connecting state */
+    EP_CALLBACK        *CallbackList[EP_CB_COUNT];             /* List of callbacks */
+    kmutex_t           CallbackLock;                           /*   and lock for it. */
+    unsigned           CallbackStep;                           /*  step through UpdateConnectionState. */
+
+    /* back pointer for cluster membership */
+    void              *ClusterRail;
+
+    /* Per node state for message passing */
+    EP_NODE_RAIL       *Nodes;                                 /* array of per-node state */
+    statemap_t         *NodeSet;                               /* per-rail statemap of connected nodes */
+    statemap_t        *NodeChangeMap;                          /* statemap of nodes to being connected/disconnected */
+    statemap_t        *NodeChangeTmp;                          /*   and temporary copies */
+
+    struct list_head    NetworkErrorList;                      /* list of nodes resolving network errors */
+    struct list_head    LocalPassivateList;                    /* list of nodes in state LOCAL_PASSIVATE */
+    struct list_head    RemotePassivateList;                   /* list of nodes waiting for remote network error flush */
+    struct list_head    PassivatedList;                                /* list of nodes performing message relocation */
+    struct list_head    DisconnectingList;                     /* list of nodes transitioning to disconnected */
+
+    EP_XID_CACHE       XidCache;                               /* XID cache for node messages (single threaded access) */
+
+    /* Manager messages */
+    EP_INPUTQ         *ManagerInputQ;
+    EP_OUTPUTQ        *ManagerOutputQ;
+    unsigned           ManagerOutputQNextSlot;
+    spinlock_t         ManagerOutputQLock;
+
+    /* /proc entries */
+    struct proc_dir_entry *ProcDir;
+    struct proc_dir_entry *SvcIndicatorDir;
+    int                    CallbackRegistered;
+};
+
+/* values for State */
+#define EP_RAIL_STATE_UNINITIALISED    0                       /* device uninitialised */
+#define EP_RAIL_STATE_STARTED          1                       /* device started but network position unknown */
+#define EP_RAIL_STATE_RUNNING          2                       /* device started and position known */
+#define EP_RAIL_STATE_INCOMPATIBLE     3                       /* device started, but position incompatible */
+
+typedef struct ep_rail_entry
+{
+    struct list_head   Link;
+    EP_RAIL           *Rail;
+} EP_RAIL_ENTRY;
+
+typedef struct ep_subsys
+{
+    EP_SYS            *Sys;
+
+    struct list_head   Link;                                   /* Linked on sys->Subsystems */
+    char              *Name;                                   /* Name to lookup */
+    
+    void              (*Destroy)    (struct ep_subsys *subsys, EP_SYS *sys);
+
+    int                       (*AddRail)    (struct ep_subsys *subsys, EP_SYS *sys, EP_RAIL *rail);
+    void              (*RemoveRail) (struct ep_subsys *subsys, EP_SYS *sys, EP_RAIL *rail);
+} EP_SUBSYS;
+
+typedef struct ep_node
+{
+    EP_RAILMASK                ConnectedRails;
+} EP_NODE;
+
+struct ep_sys
+{
+    EP_RAIL         *Rails[EP_MAX_RAILS];                      /* array of all available devices */
+
+    kmutex_t        StartStopLock;                             /* lock for starting stopping rails */
+
+    ELAN_POSITION    Position;                                 /* primary node position */
+
+    EP_NMH_TABLE     MappingTable;                             /* Network mapping handle table */
+
+    EP_ALLOC        Allocator;                                 /* shared main memory allocator */
+
+    EP_DVMA_STATE    DvmaState;                                        /* dvma state */
+
+    kmutex_t        SubsysLock;                               /* lock on the Subsytems list */
+    struct list_head Subsystems;                               /* list of subsystems */
+
+    /* device manager state */
+    struct list_head ManagedRails;                             /* list of managed devices */
+    EP_KTHREAD       ManagerThread;                            /* place for manager thread to sleep */
+
+    /* global node state */
+    spinlock_t      NodeLock;                                  /* spinlock for node state (including per-device node state) */
+    EP_NODE        *Nodes;                                     /* system wide node state */
+    statemap_t      *NodeSet;                                  /* system wide nodeset */
+    struct list_head NodesetCallbackList;                      /* list of "callbacks" */
+
+    /* Transaction Id */
+    struct list_head XidCacheList;                             /* list of XID caches */
+    uint32_t        XidGeneration;                             /* XID generation number (distinguishes reboots) */
+    uint32_t        XidHandle;                                 /* XID handles (distinguishes XID caches) */
+    uint64_t        XidNext;                                   /* next XID to prime cache */
+    spinlock_t      XidLock;                                   /*   and it's spinlock  */
+
+    /* Shutdown/Panic */
+    unsigned int     Shutdown;                                 /* node has shutdown/panic'd */
+};
+
+#if defined(DEBUG_ASSERT)
+extern int ep_assfail (EP_RAIL *rail, const char *string, const char *func, const char *file, const int line);
+extern int sdram_assert;
+extern int assfail_mode;
+
+#define EP_ASSERT(rail, EX)    do { \
+    if (!(EX) && ep_assfail ((EP_RAIL *) (rail), #EX, __FUNCTION__, __FILE__, __LINE__)) { \
+       BUG(); \
+    } \
+} while (0)
+#define EP_ASSFAIL(rail,EX)    do { \
+   if (ep_assfail ((EP_RAIL *) (rail), EX, __FUNCTION__, __FILE__, __LINE__)) { \
+       BUG(); \
+    } \
+} while (0)
+#define SDRAM_ASSERT(EX)       (sdram_assert ? (EX) : 1)
+#else
+#define EP_ASSERT(rail, EX)    ((void) 0)
+#define EP_ASSFAIL(rail,str)   ((void) 0)
+#define SDRAM_ASSERT(EX)       (1)
+#endif
+
+/* conf_osdep.c */
+extern EP_SYS    *ep_system(void);
+extern void       ep_mod_dec_usecount (void);
+extern void       ep_mod_inc_usecount (void);
+
+/* procfs_osdep.c */
+extern struct proc_dir_entry *ep_procfs_root;
+extern struct proc_dir_entry *ep_config_root;
+
+/* kcomm.c */
+extern int        ep_sys_init (EP_SYS *sys);
+extern void       ep_sys_fini (EP_SYS *sys);
+extern void      ep_shutdown (EP_SYS *sys);
+extern int        ep_init_rail (EP_SYS *sys, EP_RAIL *rail);
+extern void       ep_destroy_rail (EP_RAIL *rail);
+extern int        ep_start_rail (EP_RAIL *rail);
+extern void       ep_stop_rail (EP_RAIL *rail);
+
+extern void       ep_connect_node (EP_RAIL *rail, int nodeId);
+extern int        ep_disconnect_node (EP_RAIL *rail, int nodeId);
+
+extern EP_XID     ep_xid_cache_alloc (EP_SYS *sys, EP_XID_CACHE *cache);
+extern void       ep_xid_cache_init (EP_SYS *sys, EP_XID_CACHE *cache);
+extern void       ep_xid_cache_destroy (EP_SYS *sys, EP_XID_CACHE *cache);
+
+extern int        ep_send_message (EP_RAIL *rail, int nodeId, int type, EP_XID xid, EP_MANAGER_MSG_BODY *body);
+
+extern void       ep_panic_node (EP_SYS *sys, int nodeId, unsigned char *reason);
+
+extern void      ep_subsys_add (EP_SYS *sys, EP_SUBSYS *subsys);
+extern void      ep_subsys_del (EP_SYS *sys, EP_SUBSYS *subsys);
+extern EP_SUBSYS *ep_subsys_find (EP_SYS *sys, char *name);
+
+extern void       DisplayNodes (EP_RAIL *rail);
+
+extern void       ep_fillout_stats(EP_RAIL *rail, char *str);
+
+/* neterr.c */
+extern void       ep_queue_network_error (EP_RAIL *rail, int nodeId, int what, int channel, EP_NETERR_COOKIE cookie);
+
+/* kcomm_elan3.c */
+extern unsigned int ep3_create_rails (EP_SYS *sys, unsigned int disabled);
+
+/* kcomm_elan4.c */
+extern unsigned int ep4_create_rails (EP_SYS *sys, unsigned int disabled);
+
+/* probenetwork.c */
+extern int       ProbeNetwork (EP_RAIL *rail, ELAN_POSITION *pos);
+extern void      CheckPosition (EP_RAIL *rail);
+
+extern uint16_t   CheckSum (char *msg, int nob);
+
+/* threadcode.c */
+extern EP_ADDR    ep_symbol (EP_CODE *code, char *name);
+extern int        ep_loadcode (EP_RAIL *rail, EP_CODE *code);
+extern void       ep_unloadcode (EP_RAIL *rail, EP_CODE *code);
+
+/* Public interface */
+/* debug.c */
+extern int              ep_sprintf_bitmap (char *str, unsigned nbytes, bitmap_t *bitmap, int base, int count, int off);
+extern void             ep_display_bitmap (char *prefix, char *tag, bitmap_t *bitmap, unsigned base, unsigned nbits);
+
+/* epcomms.c */
+extern int              ep_waitfor_nodeid (EP_SYS *sys);
+extern int              ep_nodeid (EP_SYS *sys);
+extern int              ep_numnodes (EP_SYS *sys);
+
+/* railhints.c */
+extern int              ep_pickRail(EP_RAILMASK railmask);
+
+/* support.c */
+extern int              ep_register_nodeset_callback (EP_SYS *sys, void (*routine)(void *, statemap_t *), void *arg);
+extern void             ep_remove_nodeset_callback (EP_SYS *sys, void (*routine)(void *, statemap_t *), void *arg);
+extern void             ep_call_nodeset_callbacks (EP_SYS *sys, statemap_t *map);
+
+extern int              ep_register_callback (EP_RAIL *rail, unsigned idx, void (*routine)(void *, statemap_t *), void *arg);
+extern void             ep_remove_callback (EP_RAIL *rail, unsigned idx, void (*routine)(void *, statemap_t *), void *arg);
+extern void             ep_call_callbacks (EP_RAIL *rail, unsigned idx, statemap_t *);
+extern unsigned int     ep_backoff (EP_BACKOFF *backoff, int type);
+
+#endif /* !__ELAN__ */
+
+typedef struct display_info {
+    void (*func)(long, char *, ...);
+    long arg;
+} DisplayInfo;
+
+extern DisplayInfo di_ep_debug;
+
+
+#endif /* __ELAN_KCOMM_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/kcomm_stats.h
===================================================================
--- linux-2.4.21.orig/include/elan/kcomm_stats.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/kcomm_stats.h     2005-06-01 23:12:54.712421928 -0400
@@ -0,0 +1,153 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __EP_EPSTATS_H
+#define __EP_EPSTATS_H
+
+#ident "$Id: kcomm_stats.h,v 1.4.8.1 2004/11/12 10:54:51 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kcomm_stats.h,v $ */
+
+#define EP_BUCKET_SLOTS                        8
+
+#define BucketStat(obj,stat,size)      ((size) < 128   ? (obj)->Stats.stat[0]++ : \
+                                        (size) < 512   ? (obj)->Stats.stat[1]++ : \
+                                        (size) < 1024  ? (obj)->Stats.stat[2]++ : \
+                                        (size) < 8192  ? (obj)->Stats.stat[3]++ : \
+                                        (size) < 16384 ? (obj)->Stats.stat[4]++ : \
+                                        (size) < 32768 ? (obj)->Stats.stat[5]++ : \
+                                        (size) < 65536 ? (obj)->Stats.stat[6]++ : \
+                                        (obj)->Stats.stat[7]++)
+#define IncrStat(obj,stat)             ((obj)->Stats.stat++)
+
+
+#define EP3_NUM_DMA_FAIL               11      /* NOTE - the same as EP_NUM_RETRIES */
+
+#define ADD_STAT(STATS,STAT,VALUE) { unsigned long now = lbolt;\
+   STATS.STAT.total += VALUE; \
+   if  ( ( now - STATS.STAT.last_time ) > HZ ) { \
+ STATS.STAT.last_per_sec = ( STATS.STAT.total - STATS.STAT.last_count)/ ( (( now - STATS.STAT.last_time ) + (HZ/2)) / HZ);\
+ STATS.STAT.last_time = now; \
+ STATS.STAT.last_count = STATS.STAT.total; \
+   }} \
+
+#define INC_STAT(STATS,STAT) ADD_STAT(STATS,STAT,1)
+
+#define GET_STAT_PER_SEC(STATS, STAT) (  (( lbolt - STATS.STAT.last_time ) < (HZ * 5)) ? STATS.STAT.last_per_sec : 0 )
+#define GET_STAT_TOTAL(STATS, STAT) ( STATS.STAT.total )
+
+struct ep_stats_count 
+{
+    unsigned long total;
+    unsigned long last_time;
+    unsigned long last_count;
+    unsigned long last_per_sec;
+};
+
+typedef struct ep_stats_count          EP_STATS_COUNT;
+
+typedef struct ep3_rail_stats
+{
+    unsigned long      IssueDmaFail[EP3_NUM_DMA_FAIL];
+
+    unsigned long      DmaQueueLength[EP_BUCKET_SLOTS];
+    unsigned long      CprocDmaQueueOverflow;
+    unsigned long      DprocDmaQueueOverflow;
+    unsigned long      IprocDmaQueueOverflow;
+    unsigned long      CprocEventQueueOverflow;
+    unsigned long      DprocEventQueueOverflow;
+    unsigned long      IprocEventQueueOverflow;
+
+    unsigned long      QueueingPacketTrap;
+    unsigned long      DmaIdentifyTrap;
+    unsigned long      ThreadIdentifyTrap;
+    unsigned long      DmaPacketTrap;
+} EP3_RAIL_STATS;
+
+typedef struct ep4_rail_stats
+{
+    unsigned long       somestatsgohere;
+} EP4_RAIL_STATS;
+
+typedef struct ep_rail_stats
+{
+    unsigned long      SendMessageFailed;
+    unsigned long      NeterrAtomicPacket;
+    unsigned long       NeterrDmaPacket;
+
+    EP_STATS_COUNT      rx;
+    EP_STATS_COUNT      rx_len;
+
+    EP_STATS_COUNT      tx;
+    EP_STATS_COUNT      tx_len;
+
+} EP_RAIL_STATS;
+
+typedef struct ep_cm_rail_stats
+{
+    /* cluster membership statistics */
+    unsigned long      HeartbeatsSent;
+    unsigned long      HeartbeatsRcvd;
+    
+    unsigned long      RetryHeartbeat;
+    unsigned long      RejoinRequest;
+    unsigned long      RejoinTooSlow;
+    unsigned long      LaunchMessageFail;
+    unsigned long      MapChangesSent;
+
+    /* Heartbeat scheduling stats */
+    unsigned long      HeartbeatOverdue;
+} EP_CM_RAIL_STATS;
+
+typedef struct ep_comms_rail_stats
+{
+    /* kernel comms large message statistics */
+    unsigned long      TxEnveEvent;
+    unsigned long      TxDataEvent;
+    unsigned long      TxDoneEvent;
+    unsigned long      RxDoneEvent;
+    unsigned long      MulticastTxDone;
+    unsigned long      QueueReceive;
+
+    unsigned long      TxEnveRetry;
+    unsigned long      TxDataRetry;
+    unsigned long      TxDoneRetry;
+    unsigned long      RxThrdEvent;
+    unsigned long      RxDataRetry;
+    unsigned long      RxDoneRetry;
+    unsigned long      StallThread;
+    unsigned long      ThrdWaiting;
+    unsigned long      CompleteEnvelope;
+
+    unsigned long      NoFreeTxds;
+    unsigned long      NoFreeRxds;
+
+    unsigned long      LockRcvrTrapped;
+} EP_COMMS_RAIL_STATS;
+
+typedef struct ep_comms_stats
+{
+    unsigned long      DataXmit[8];
+    unsigned long      McastXmit[8];
+    unsigned long      RPCXmit[8];
+    unsigned long      RPCPut[8];
+    unsigned long      RPCGet[8];
+    unsigned long      CompleteRPC[8];
+    unsigned long      RxData[8];
+    unsigned long      RxMcast[8];
+
+    unsigned long      NoFreeTxds;
+    unsigned long      NoFreeRxds;
+} EP_COMMS_STATS;
+
+#endif /* __EP_EPSTATS_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/kmap.h
===================================================================
--- linux-2.4.21.orig/include/elan/kmap.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/kmap.h    2005-06-01 23:12:54.713421776 -0400
@@ -0,0 +1,68 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_KMAP_H
+#define __ELAN_KMAP_H
+
+#ident "$Id: kmap.h,v 1.3.8.1 2004/12/14 10:19:14 mike Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kmap.h,v $ */
+
+#include <elan/rmap.h>
+
+extern void ep_perrail_kaddr_map (EP_RAIL *rail, EP_ADDR eaddr, virtaddr_t vaddr, unsigned long len, unsigned int perm, int ep_attr);
+extern void ep_perrail_sdram_map (EP_RAIL *rail, EP_ADDR eaddr, sdramaddr_t saddr, unsigned long len, unsigned int perm, int ep_attr);
+extern void ep_perrail_unmap (EP_RAIL *rail, EP_ADDR eaddr, unsigned long len);
+extern void ep_perrail_dvma_sync (EP_RAIL *rail);
+
+typedef struct ep_dvma_nmh
+{
+    EP_NMH             dvma_nmh;
+    
+    struct list_head   dvma_link;                              /* chained on ep_dvma_state */
+    unsigned           dvma_perm;                              /* permissions for region */
+
+    spinlock_t         dvma_lock;
+    EP_RAILMASK                dvma_railmask;                          /* bitmap of rails */
+    EP_RAIL           *dvma_rails[EP_MAX_RAILS];               /* assoicated rails */
+    void              *dvma_private[EP_MAX_RAILS];             /* pointers to rail private data */
+    unsigned int        dvma_attrs[1];                         /* bitmap of which rails pages are loaded NOTE - max 32 rails */
+} EP_DVMA_NMH;
+
+/* values for dvma_perm */
+#define EP_PERM_EXECUTE                0
+#define EP_PERM_READ           1
+#define EP_PERM_WRITE          2
+#define EP_PERM_ALL            3
+
+typedef struct ep_dvma_state
+{
+    kmutex_t           dvma_lock;
+    struct list_head    dvma_handles;
+    struct list_head    dvma_rails;
+    EP_RMAP           *dvma_rmap;
+} EP_DVMA_STATE;
+
+extern void    ep_dvma_init (EP_SYS *sys);
+extern void    ep_dvma_fini (EP_SYS *sys);
+extern EP_NMH *ep_dvma_reserve (EP_SYS *sys, unsigned npages, unsigned perm);
+extern void    ep_dvma_release (EP_SYS *sys, EP_NMH *nmh);
+extern void    ep_dvma_load (EP_SYS *sys, void *map, caddr_t vaddr, unsigned len, 
+                            EP_NMH *nmh, unsigned index, EP_RAILMASK *hints, EP_NMD *subset);
+extern void    ep_dvma_unload (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd);
+  
+extern void    ep_dvma_remove_rail (EP_SYS *sys, EP_RAIL *rail);
+extern int     ep_dvma_add_rail (EP_SYS *sys, EP_RAIL *rail);
+
+extern uint16_t rolling_check_sum (char *msg, int nob, uint16_t sum);
+
+#endif /* __ELAN_KMAP_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/kmsg.h
===================================================================
--- linux-2.4.21.orig/include/elan/kmsg.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/kmsg.h    2005-06-01 23:12:54.713421776 -0400
@@ -0,0 +1,14 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_KMSG_H
+#define __ELAN_KMSG_H
+
+#ident "@(#)$Id: kmsg.h,v 1.1 2003/09/23 13:55:12 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/kmsg.h,v $ */
+
+#endif /* __ELAN_KMSG_H */
Index: linux-2.4.21/include/elan/kthread.h
===================================================================
--- linux-2.4.21.orig/include/elan/kthread.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/kthread.h 2005-06-01 23:12:54.713421776 -0400
@@ -0,0 +1,53 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_KTHREAD_H
+#define __ELAN3_KTHREAD_H
+
+#ident "@(#)$Id: kthread.h,v 1.4 2004/05/06 14:24:08 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/epmod/kthread.h,v $*/
+
+typedef struct ep_kthread
+{
+       kcondvar_t      wait;                                   /* place to sleep */
+       spinlock_t      lock;                                   /* and lock */
+       long            next_run;                               /* tick when thread should next run */
+       long            running;                                /* tick when thread started to run */
+       unsigned short  should_stall;
+       unsigned char   state;
+       unsigned int    started:1;
+       unsigned int    should_stop:1;
+       unsigned int    stopped:1;
+} EP_KTHREAD;
+
+#define KT_STATE_SLEEPING              0
+#define KT_STATE_SCHEDULED             1
+#define KT_STATE_RUNNING               2
+#define KT_STATE_STALLED               3
+
+#define AFTER(a, b)                    ((((long)(a)) - ((long)(b))) > 0)
+#define BEFORE(a,b)                    ((((long)(a)) - ((long)(b))) < 0)
+
+extern void ep_kthread_init (EP_KTHREAD *kt);
+extern void ep_kthread_destroy (EP_KTHREAD *kt);
+extern void ep_kthread_started (EP_KTHREAD *kt);
+extern void ep_kthread_stopped (EP_KTHREAD *kt);
+extern int  ep_kthread_should_stall (EP_KTHREAD *kth);
+extern int  ep_kthread_sleep (EP_KTHREAD *kth, long next_run);
+extern void ep_kthread_schedule (EP_KTHREAD *kt, long when);
+extern void ep_kthread_stall (EP_KTHREAD *kth);
+extern void ep_kthread_resume (EP_KTHREAD *kt);
+extern void ep_kthread_stop (EP_KTHREAD *kt);
+extern int  ep_kthread_state (EP_KTHREAD *kt, long *time);
+#endif /* __ELAN3_KTHREAD_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan/nmh.h
===================================================================
--- linux-2.4.21.orig/include/elan/nmh.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/nmh.h     2005-06-01 23:12:54.714421624 -0400
@@ -0,0 +1,95 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_NMH_H
+#define __ELAN3_NMH_H
+
+#ident "@(#)$Id: nmh.h,v 1.7 2004/01/06 10:29:55 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/nmh.h,v $*/
+
+
+/* Forward declarations */
+typedef struct ep_nmd          EP_NMD;
+typedef struct ep_nmh_ops      EP_NMH_OPS;
+typedef struct ep_nmh          EP_NMH;
+
+/* Railmask held in 16 bit field (packs with nodeId into NMD */
+typedef uint16_t               EP_RAILMASK;
+
+#define EP_RAIL2RAILMASK(rnum) (1 << (rnum))
+#define EP_RAILMASK_ALL                0xffff
+
+/* kernel comms elan network address */
+typedef uint32_t               EP_ADDR;
+
+/* network mapping descriptor - this is returned to the user from a map operation,
+ * and is what is passed to all communication functions */
+struct ep_nmd
+{
+    EP_ADDR    nmd_addr;                                       /* base address */
+    uint32_t   nmd_len;                                        /* size in bytes */
+    uint32_t   nmd_attr;                                       /* nodeid << 16 | railmask */
+};
+
+#define EP_NMD_ATTR(nodeid,railmask)   (((nodeid) << 16) | (railmask))
+#define EP_NMD_NODEID(nmd)             ((nmd)->nmd_attr >> 16)
+#define EP_NMD_RAILMASK(nmd)           ((nmd)->nmd_attr & EP_RAILMASK_ALL)
+
+#if !defined(__ELAN__)
+
+struct ep_nmh_ops
+{
+    int           (*op_map_rails) (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd, EP_RAILMASK mask);   /* add mappings to different rail(s) */
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+    uint16_t (*op_calc_check_sum) (EP_SYS *sys, EP_NMH *nmh, EP_NMD *nmd, uint16_t check_sum); /* calculates check sum              */
+#endif
+};
+
+struct ep_nmh
+{
+    EP_NMD          nmh_nmd;                                   /* public field */
+    struct list_head nmh_link;                                 /* linked on hash table */
+    EP_NMH_OPS     *nmh_ops;                                   /* operations to perform on object */
+};
+
+#define EP_NMH_NUMHASH                 (32 - 11 + 1)           /* one hash table for each power of 2 above pagesize */
+#define EP_NMH_HASHSIZE                        (64)                    /* max size of each hash table */
+
+typedef struct ep_nmh_table
+{
+    struct list_head *tbl_hash[EP_NMH_NUMHASH];
+    unsigned         tbl_size[EP_NMH_NUMHASH];
+} EP_NMH_TABLE;
+
+extern int         ep_nmh_init (EP_NMH_TABLE *tbl);
+extern void        ep_nmh_fini (EP_NMH_TABLE *tbl);
+
+extern void        ep_nmh_insert (EP_NMH_TABLE *tbl, EP_NMH *nmd);
+extern void        ep_nmh_remove (EP_NMH_TABLE *tbl, EP_NMH *nmd);
+extern EP_NMH     *ep_nmh_find (EP_NMH_TABLE *tbl, EP_NMD *nmh);
+
+#if ! defined(CONFIG_EP_NO_CHECK_SUM)
+extern uint32_t    ep_nmd_calc_data_check_sum(EP_SYS *sys, EP_NMD *nmd, int nFrags);
+#endif
+
+/* Public interface */
+extern EP_RAILMASK ep_nmd2railmask (EP_NMD *frags, int nFrags);
+extern void        ep_nmd_subset (EP_NMD *subset, EP_NMD *nmd, unsigned off, unsigned len);
+extern int        ep_nmd_merge (EP_NMD *merged, EP_NMD *a, EP_NMD *b);
+extern int         ep_nmd_map_rails (EP_SYS *sys, EP_NMD *nmd, unsigned railmask);
+
+#endif /* __ELAN__ */
+
+#endif /* __ELAN3_NMH_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/rmap.h
===================================================================
--- linux-2.4.21.orig/include/elan/rmap.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/rmap.h    2005-06-01 23:12:54.714421624 -0400
@@ -0,0 +1,49 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_RMAP_H
+#define __ELAN_RMAP_H
+
+#ident "$Id: rmap.h,v 1.8 2004/05/19 10:24:40 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/rmap.h,v $ */
+
+
+typedef struct ep_rmap_entry 
+{
+    size_t     m_size;
+    u_long     m_addr;
+} EP_RMAP_ENTRY;
+
+typedef struct ep_rmap 
+{
+    spinlock_t      m_lock;
+    kcondvar_t       m_wait;
+    u_int           m_size;
+    u_int           m_free;
+    u_int           m_want;
+    char            *m_name;
+    EP_RMAP_ENTRY    m_map[1];
+} EP_RMAP;
+
+extern void         ep_display_rmap (EP_RMAP *map);
+
+extern void          ep_rmapinit (EP_RMAP *rmap, char *name, u_int mapsize);
+extern unsigned long ep_rmalloc (EP_RMAP *rmap, size_t size, int cansleep);
+extern unsigned long ep_rmalloc_constrained (EP_RMAP *mp, size_t size, unsigned long alo, unsigned long ahi, unsigned long align, int cansleep);
+extern void          ep_rmfree (EP_RMAP *rmap, size_t size, unsigned long addr);
+extern unsigned long ep_rmget (EP_RMAP *rmap, size_t size, unsigned long addr);
+extern EP_RMAP      *ep_rmallocmap (size_t size, char *name, int cansleep);
+extern void          ep_rmfreemap (EP_RMAP *map);
+
+#endif /* __ELAN3_RMAP_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/statemap.h
===================================================================
--- linux-2.4.21.orig/include/elan/statemap.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/statemap.h        2005-06-01 23:12:54.714421624 -0400
@@ -0,0 +1,52 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN_STATEMAP_H
+#define __ELAN_STATEMAP_H
+
+#ident "$Id: statemap.h,v 1.8 2003/10/07 13:22:38 david Exp $"
+/*      $Source: /cvs/master/quadrics/epmod/statemap.h,v $ */
+
+#include <elan/bitmap.h>
+
+/******************************** global state bitmap stuff **********************************/
+typedef struct
+{
+   unsigned int size;
+   unsigned int nob;
+   unsigned int changemap_nob;
+   unsigned int bitmap_nob;
+   bitmap_t    *changemap0;
+   bitmap_t    *changemap1;
+   bitmap_t    *changemap2;
+   bitmap_t    *bitmap;
+} statemap_t;
+
+extern bitmap_t              statemap_getseg (statemap_t *map, unsigned int offset);
+extern void           statemap_setseg (statemap_t *map, unsigned int offset, bitmap_t seg);
+extern bitmap_t       statemap_getbits (statemap_t *map, unsigned int offset, int nbits);
+extern void           statemap_setbits (statemap_t *map, unsigned int offset, bitmap_t bits, int nbits);
+extern void           statemap_zero (statemap_t *map);
+extern void           statemap_setmap (statemap_t *dst, statemap_t *src);
+extern void           statemap_ormap (statemap_t *dst, statemap_t *src);
+extern int           statemap_findchange (statemap_t *map, bitmap_t *newseg, int clearchange);
+extern int            statemap_changed (statemap_t *map);
+extern void           statemap_reset (statemap_t *map);
+extern void           statemap_copy (statemap_t *dst, statemap_t *src);
+extern void           statemap_clearchanges (statemap_t *map);
+extern bitmap_t      *statemap_tobitmap (statemap_t *map);
+extern statemap_t    *statemap_create (int size);
+extern void           statemap_destroy (statemap_t *map);
+
+#endif /* __ELAN_STATEMAP_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan/stats.h
===================================================================
--- linux-2.4.21.orig/include/elan/stats.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan/stats.h   2005-06-01 23:12:54.715421472 -0400
@@ -0,0 +1,85 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: stats.h,v 1.5 2003/09/24 13:55:37 david Exp $"
+/*      $Source: /cvs/master/quadrics/elanmod/modsrc/stats.h,v $*/
+
+#ifndef __ELAN_STATS_H
+#define __ELAN_STATS_H
+
+
+/* non-kernel headings */
+#define ELAN_STATS_NAME_MAX_LEN ((uint)64)
+typedef unsigned int ELAN_STATS_IDX;
+
+typedef struct elan_stats_map
+{
+       char entry_name[ELAN_STATS_NAME_MAX_LEN];
+       int  index;
+} ELAN_STATS_MAP;
+
+#if defined(__KERNEL__)
+
+/* stats callbacks */
+#define ELAN_STATS_OPS_VERSION ((u_int)1)
+typedef struct elan_stats_ops
+{
+       u_int  ops_version;
+
+       int (*elan_stats_get_name)    (void * arg, uint index,   caddr_t  name);
+       int (*elan_stats_get_block)   (void * arg, uint entries, ulong   *values);
+       int (*elan_stats_clear_block) (void * arg);
+
+} ELAN_STATS_OPS;
+
+typedef struct elan_stats_struct
+{
+       struct list_head   node;
+
+       ELAN_STATS_IDX     statidx;
+       char               block_name[ELAN_STATS_NAME_MAX_LEN];
+       uint               num_entries;
+       ELAN_STATS_OPS *ops;
+       void              *arg;
+
+} ELAN_STATS_STRUCT;
+
+/* stats.c */
+extern int                   elan_stats_register   (ELAN_STATS_IDX    *statidx, 
+                                                      char              *block_name, 
+                                                      uint               num_entries,
+                                                      ELAN_STATS_OPS *ops,
+                                                      void              *arg);
+
+extern int                   elan_stats_deregister  (ELAN_STATS_IDX  statidx);
+extern ELAN_STATS_STRUCT *elan_stats_find        (ELAN_STATS_IDX  statidx);
+extern ELAN_STATS_STRUCT *elan_stats_find_by_name(caddr_t         block_name);
+extern ELAN_STATS_STRUCT *elan_stats_find_next   (ELAN_STATS_IDX  statidx);
+
+
+/* elan_stats.c */
+extern int elan_stats_get_next_index (ELAN_STATS_IDX statidx, ELAN_STATS_IDX *next_statidx);
+
+extern int elan_stats_find_index     (caddr_t  block_name, ELAN_STATS_IDX *statidx, uint *num_entries);
+
+extern int elan_stats_get_block_info (ELAN_STATS_IDX statidx, caddr_t block_name, uint *num_entries);
+
+extern int elan_stats_get_index_name (ELAN_STATS_IDX statidx, uint index, caddr_t name);
+
+extern int elan_stats_get_block      (ELAN_STATS_IDX statidx, uint entries, ulong   *values);
+
+extern int elan_stats_clear_block    (ELAN_STATS_IDX statidx);
+
+#endif /* __KERNEL__ */
+
+#endif /* __ELAN_STATS_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/compat.h
===================================================================
--- linux-2.4.21.orig/include/elan3/compat.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/compat.h 2005-06-01 23:12:54.715421472 -0400
@@ -0,0 +1,177 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: compat.h,v 1.4 2004/06/09 09:07:03 mike Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/compat.h,v $*/
+
+#ifndef __ELAN3_COMPAT_H
+#define __ELAN3_COMPAT_H
+
+/* compatibility header to allow Eagle branch QSNETLIBS 
+ * to compile against head kernel */
+
+#define ELAN_EAGLE_COMPAT
+
+/* vmseg.h */
+#define ELAN_FLAGSTATS                         ELAN3_FLAGSTATS
+
+/* uregs.h */
+#define ELAN_STATS_NAME                                ELAN3_STATS_NAME
+#define elan3_stats_names                      elan_stats_names
+
+/* spinlock.h */
+#define ELAN_SPINLOCK                          ELAN3_SPINLOCK
+#define ELAN_SPINLOCK_MAIN                     ELAN3_SPINLOCK_MAIN
+#define ELAN_SPINLOCK_ELAN                     ELAN3_SPINLOCK_ELAN
+#define ELAN_ME_SPINENTER                      ELAN3_ME_SPINENTER
+#define ELAN_ME_FORCEENTER                     ELAN3_ME_FORCEENTER
+#define ELAN_ME_SPINEXIT                       ELAN3_ME_SPINEXIT
+#define ELAN_SPINENTER                         ELAN3_SPINENTER
+#define ELAN_SPINEXIT                          ELAN3_SPINEXIT
+#define elan3_me_spinblock                     elan_me_spinblock
+#define elan3_spinenter                                elan_spinenter
+
+/* elanio.h */
+#define ELANIO_CONTROL_PATHNAME                        ELAN3IO_CONTROL_PATHNAME
+#define ELANIO_USER_PATHNAME                   ELAN3IO_USER_PATHNAME
+#define ELANIO_SDRAM_PATHNAME                  ELAN3IO_SDRAM_PATHNAME
+#define ELANIO_MAX_PATHNAMELEN                 ELAN3IO_MAX_PATHNAMELEN
+
+#define ELANIO_SET_BOUNDARY_SCAN               ELAN3IO_SET_BOUNDARY_SCAN
+#define ELANIO_CLEAR_BOUNDARY_SCAN             ELAN3IO_CLEAR_BOUNDARY_SCAN
+#define ELANIO_READ_LINKVAL                    ELAN3IO_READ_LINKVAL
+#define ELANIO_WRITE_LINKVAL                   ELAN3IO_WRITE_LINKVAL
+#define ELANIO_SET_DEBUG_STRUCT                        ELAN3IO_SET_DEBUG_STRUCT
+#define ELANIO_SET_DEBUG                       ELAN3IO_SET_DEBUG
+#define ELANIO_DEBUG_BUFFER_STRUCT             ELAN3IO_DEBUG_BUFFER_STRUCT
+#define ELANIO_DEBUG_BUFFER                    ELAN3IO_DEBUG_BUFFER
+#define ELANIO_NETERR_SERVER_STRUCT            ELAN3IO_NETERR_SERVER_STRUCT
+#define ELANIO_NETERR_SERVER                   ELAN3IO_NETERR_SERVER
+#define ELANIO_NETERR_FIXUP                    ELAN3IO_NETERR_FIXUP
+
+#define ELANIO_FREE                            ELAN3IO_FREE
+#define ELANIO_ATTACH                          ELAN3IO_ATTACH
+#define ELANIO_DETACH                          ELAN3IO_DETACH
+#define ELANIO_ADDVP_STRUCT                    ELAN3IO_ADDVP_STRUCT
+#define ELANIO_ADDVP                           ELAN3IO_ADDVP
+#define ELANIO_REMOVEVP                                ELAN3IO_REMOVEVP
+#define ELANIO_BCASTVP_STRUCT                  ELAN3IO_BCASTVP_STRUCT
+#define ELANIO_BCASTVP                         ELAN3IO_BCASTVP
+#define ELANIO_LOAD_ROUTE_STRUCT               ELAN3IO_LOAD_ROUTE_STRUCT
+#define ELANIO_LOAD_ROUTE                      ELAN3IO_LOAD_ROUTE
+#define ELANIO_PROCESS                         ELAN3IO_PROCESS
+#define ELANIO_SETPERM_STRUCT                  ELAN3IO_SETPERM_STRUCT
+#define ELANIO_SETPERM                         ELAN3IO_SETPERM
+#define ELANIO_CLEARPERM_STRUCT                        ELAN3IO_CLEARPERM_STRUCT
+#define ELANIO_CLEARPERM                       ELAN3IO_CLEARPERM
+#define ELANIO_CHANGEPERM_STRUCT               ELAN3IO_CHANGEPERM_STRUCT
+#define ELANIO_CHANGEPERM                      ELAN3IO_CHANGEPERM
+#define ELANIO_HELPER_THREAD                   ELAN3IO_HELPER_THREAD
+#define ELANIO_WAITCOMMAND                     ELAN3IO_WAITCOMMAND
+#define ELANIO_BLOCK_INPUTTER                  ELAN3IO_BLOCK_INPUTTER
+#define ELANIO_SET_FLAGS                       ELAN3IO_SET_FLAGS
+#define ELANIO_WAITEVENT                       ELAN3IO_WAITEVENT
+#define ELANIO_ALLOC_EVENTCOOKIE               ELAN3IO_ALLOC_EVENTCOOKIE
+#define ELANIO_FREE_EVENTCOOKIE                        ELAN3IO_FREE_EVENTCOOKIE
+#define ELANIO_ARM_EVENTCOOKIE                 ELAN3IO_ARM_EVENTCOOKIE
+#define ELANIO_WAIT_EVENTCOOKIE                        ELAN3IO_WAIT_EVENTCOOKIE
+#define ELANIO_SWAPSPACE                       ELAN3IO_SWAPSPACE
+#define ELANIO_EXCEPTION_SPACE                 ELAN3IO_EXCEPTION_SPACE
+#define ELANIO_GET_EXCEPTION                   ELAN3IO_GET_EXCEPTION
+#define ELANIO_UNLOAD_STRUCT                   ELAN3IO_UNLOAD_STRUCT
+#define ELANIO_UNLOAD                          ELAN3IO_UNLOAD
+#define ELANIO_GET_ROUTE_STRUCT                        ELAN3IO_GET_ROUTE_STRUCT
+#define ELANIO_GET_ROUTE                       ELAN3IO_GET_ROUTE
+#define ELANIO_RESET_ROUTE_STRUCT              ELAN3IO_RESET_ROUTE_STRUCT
+#define ELANIO_RESET_ROUTE                     ELAN3IO_RESET_ROUTE
+#define ELANIO_CHECK_ROUTE_STRUCT              ELAN3IO_CHECK_ROUTE_STRUCT
+#define ELANIO_CHECK_ROUTE                     ELAN3IO_CHECK_ROUTE
+#define ELANIO_VP2NODEID_STRUCT                        ELAN3IO_VP2NODEID_STRUCT
+#define ELANIO_VP2NODEID                       ELAN3IO_VP2NODEID
+#define ELANIO_SET_SIGNAL                      ELAN3IO_SET_SIGNAL
+#define ELANIO_PROCESS_2_LOCATION_STRUCT       ELAN3IO_PROCESS_2_LOCATION_STRUCT
+#define ELANIO_PROCESS_2_LOCATION              ELAN3IO_PROCESS_2_LOCATION
+#define ELANIO_GET_DEVINFO_STRUCT              ELAN3IO_GET_DEVINFO_STRUCT
+#define ELANIO_GET_DEVINFO                     ELAN3IO_GET_DEVINFO
+#define ELANIO_GET_POSITION_STRUCT             ELAN3IO_GET_POSITION_STRUCT
+#define ELANIO_GET_POSITION                    ELAN3IO_GET_POSITION
+#define ELANIO_STATS_STRUCT                    ELAN3IO_STATS_STRUCT
+#define ELANIO_STATS                           ELAN3IO_STATS
+#  define ELAN_SYS_STATS_DEVICE                        ELAN3_SYS_STATS_DEVICE
+#  define ELAN_SYS_STATS_ELAN3MMU              ELAN3_SYS_STATS_MMU
+
+#define ELANIO_OFF_FLAG_PAGE                   ELAN3IO_OFF_FLAG_PAGE
+#define ELANIO_OFF_UREG_PAGE                   ELAN3IO_OFF_UREG_PAGE
+#define ELANIO_OFF_COMMAND_PAGE                        ELAN3IO_OFF_COMMAND_PAGE
+
+
+/* elanvp.h */
+#define ELAN_ROUTE_SUCCESS                     ELAN3_ROUTE_SUCCESS
+#define ELAN_ROUTE_SYSCALL_FAILED              ELAN3_ROUTE_SYSCALL_FAILED
+#define ELAN_ROUTE_INVALID                     ELAN3_ROUTE_INVALID
+#define ELAN_ROUTE_TOO_LONG                    ELAN3_ROUTE_TOO_LONG
+#define ELAN_ROUTE_LOAD_FAILED                 ELAN3_ROUTE_LOAD_FAILED
+#define ELAN_ROUTE_PROC_RANGE                  ELAN3_ROUTE_PROC_RANGE
+#define ELAN_ROUTE_INVALID_LEVEL               ELAN3_ROUTE_INVALID_LEVEL
+#define ELAN_ROUTE_OCILATES                    ELAN3_ROUTE_OCILATES
+#define ELAN_ROUTE_WRONG_DEST                  ELAN3_ROUTE_WRONG_DEST
+#define ELAN_ROUTE_TURN_LEVEL                  ELAN3_ROUTE_TURN_LEVEL
+#define ELAN_ROUTE_NODEID_UNKNOWN              ELAN3_ROUTE_NODEID_UNKNOWN
+
+/* elandev.h */
+#define ELAN_STATS                             ELAN3_STATS
+#define ELAN_STATS_VERSION                     ELAN3_STATS_VERSION
+
+/* perm.h */
+#define ELAN_PERM_NOREMOTE                     ELAN3_PERM_NOREMOTE
+#define ELAN_PERM_LOCAL_READ                   ELAN3_PERM_LOCAL_READ
+#define ELAN_PERM_REMOTEALL                    ELAN3_PERM_REMOTEALL
+
+/* threadsyscall.h */
+#define ELAN_ABORT_TRAPNUM                     ELAN3_ABORT_TRAPNUM
+#define ELAN_ELANCALL_TRAPNUM                  ELAN3_ELANCALL_TRAPNUM
+#define ELAN_SYSCALL_TRAPNUM                   ELAN3_SYSCALL_TRAPNUM
+#define ELAN_SYS_close                         ELAN3_SYS_close
+#define ELAN_SYS_getpid                                ELAN3_SYS_getpid
+#define ELAN_SYS_ioctl                         ELAN3_SYS_ioctl
+#define ELAN_SYS_kill                          ELAN3_SYS_kill
+#define ELAN_SYS_lseek                         ELAN3_SYS_lseek
+#define ELAN_SYS_mmap                          ELAN3_SYS_mmap
+#define ELAN_SYS_munmap                                ELAN3_SYS_munmap
+#define ELAN_SYS_open                          ELAN3_SYS_open
+#define ELAN_SYS_poll                          ELAN3_SYS_poll
+#define ELAN_SYS_read                          ELAN3_SYS_read
+#define ELAN_SYS_write                         ELAN3_SYS_write
+#define ELAN_T_SYSCALL_CODE                    ELAN3_T_SYSCALL_CODE
+#define ELAN_T_SYSCALL_ERRNO                   ELAN3_T_SYSCALL_ERRNO
+
+/* elansyscall.h */
+#define ELAN_SYS_FLAG_DMA_BADVP                        ELAN3_SYS_FLAG_DMA_BADVP
+#define ELAN_SYS_FLAG_THREAD_BADVP             ELAN3_SYS_FLAG_THREAD_BADVP
+#define ELAN_SYS_FLAG_DMAFAIL                  ELAN3_SYS_FLAG_DMAFAIL
+#define ELAN_SYS_FLAG_NETERR                   ELAN3_SYS_FLAG_NETERR
+
+/* intrinsics.h */
+#define elan_copy64w                           elan3_copy64w
+#define elan_read64dw                          elan3_read64dw
+#define elan_write64dw                         elan3_write64dw
+
+#ifndef ELAN_POLL_EVENT
+#define ELAN_POLL_EVENT                                ELAN3_POLL_EVENT
+#endif
+#ifndef ELAN_WAIT_EVENT
+#define ELAN_WAIT_EVENT                                ELAN3_WAIT_EVENT
+#endif
+
+#endif /* __ELAN3_COMPAT_H */
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+
Index: linux-2.4.21/include/elan3/dma.h
===================================================================
--- linux-2.4.21.orig/include/elan3/dma.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/dma.h    2005-06-01 23:12:54.716421320 -0400
@@ -0,0 +1,213 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_DMA_H
+#define __ELAN3_DMA_H
+
+#ident "$Id: dma.h,v 1.38 2002/08/21 12:43:27 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/dma.h,v $ */
+
+#include <elan3/e3types.h>
+#include <elan3/events.h>
+
+/* Alignment for a DMA descriptor */
+#define E3_DMA_ALIGN           (32)
+
+/* The maximum size a DMA can be (i.e. < 2GB) */
+#define E3_MAX_DMA_SIZE                0x7fffffff
+
+/* This macro returns TRUE if a fixup for the ELAN_REVB_BUG_2 problem is required 
+ * i.e. if the DMA begins in the last 64-bytes of a page and its size causes it to enter the
+ * next page, hence causing the Elan to issue 2 (64-byte) block reads to different pages.
+ * See GNAT hw-elan3/3263
+ */
+#define E3_DMA_REVB_BUG_2(SIZE, ADDR, PAGESIZE)        \
+       ( (((int) (ADDR) & (PAGESIZE-64)) == (PAGESIZE-64)) && (-(((int) (ADDR) | ~(PAGESIZE-1))) < (SIZE)) )
+
+/* There is a point where a dma runs quicker from main memory than
+ * when running from sdram and having to copy all the data down
+ * first.
+ */
+#define E3_DMA_SDRAM_CUTOFF    128
+
+typedef union _e3_DmaType
+{
+    E3_uint32 type;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 dataType:2;   /* Bits 0 to 1   */
+       E3_uint32 direction:3;  /* Bit  4 to 2   */
+       E3_uint32 opCode:4;     /* Bits 5 to 8   */
+       E3_uint32 failCount:6;  /* Bits 9 to 14  */
+       E3_uint32 isRemote:1;   /* Bit  15       */
+       E3_uint32 Context:13;   /* Bits 16 to 28 */
+       E3_uint32 :3;           /* Bits 29 to 31 */
+#else
+       E3_uint32 :3;           /* Bits 29 to 31 */
+       E3_uint32 Context:13;   /* Bits 16 to 28 */
+       E3_uint32 isRemote:1;   /* Bit  15       */
+       E3_uint32 failCount:6;  /* Bits 9 to 14  */
+       E3_uint32 opCode:4;     /* Bits 5 to 8   */
+       E3_uint32 direction:3;  /* Bit  4 to 2   */
+       E3_uint32 dataType:2;   /* Bits 0 to 1   */
+#endif
+    } s;
+} E3_DmaType;
+
+#define E3_DMA_CONTEXT_MASK    (ALL_CONTEXT_BITS << 16)
+
+#define E3_DMA_CONTEXT(type)   (((type) >> 16) & ALL_CONTEXT_BITS)
+#define E3_DMA_ISREMOTE(type)  (((type) >> 15) & 1)
+#define E3_DMA_FAILCOUNT(type) (((type) >> 9) & 0x3F)
+#define E3_DMA_OPCODE(type)    (((type) >> 5) & 0xF)
+#define E3_DMA_DIRECTION(type) (((type) >> 2) & 0x7)
+#define EP_DMA_DATATYPE(type)  (((type) >> 0) & 0x3)
+
+#define E3_DMA_TYPE(dataType, direction, opCode, failCount) \
+    (((dataType) & 0x3) | (((direction) & 7) << 2) | (((opCode) & 0xF) << 5) | (((failCount) & 0x3F) << 9))
+
+
+typedef union _e3_CookieVProc
+{
+    E3_uint32 cookie_vproc;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 vproc:16;                     /* Bit  15 to 0  */
+       E3_uint32 cookie:16;                    /* Bits 31 to 16 */
+#else
+       E3_uint32 cookie:16;                    /* Bits 31 to 16 */
+       E3_uint32 vproc:16;                     /* Bit  15 to 0  */
+#endif
+    } s;
+} E3_CookieVProc;
+
+#define E3_DMA_COOKIE_PROC(Cookie, VProc)  (((VProc) & 0xffff) | (((Cookie) << 16)))
+
+#define DMA_COOKIE_MASK                (0xffff0000)
+#define DMA_PROCESS_MASK       (0x0000ffff)
+
+/* We use the bottom bit of the cookie to
+ * distinguish main/thread generated cookies
+ */
+#define DMA_COOKIE_THREAD      (0x01 << 16)
+
+/* We use the next bit of the cookie to
+ * distinguish locally/remotely generated cookies 
+ */
+#define DMA_COOKIE_REMOTE      (0x02 << 16)
+
+/* Assign and increment cookie (NB: we have reserved the bottom two bits)
+ */
+#define DMA_COOKIE(COOKIE, VPROC)       ((((COOKIE) += (0x4 << 16)) & DMA_COOKIE_MASK) | VPROC)
+#define DMA_REMOTE_COOKIE(COOKIE, VPROC) ((((COOKIE) += (0x4 << 16)) & DMA_COOKIE_MASK) | DMA_COOKIE_REMOTE | VPROC)
+
+#define DMA_COOKIE_REFRESH(COOKIEVP, COOKIE)                           \
+do {                                                                   \
+       COOKIEVP &= ~DMA_COOKIE_MASK;           /* Clear cookie */      \
+       COOKIEVP |= DMA_COOKIE(COOKIE,0);       /* Assign new cookie */ \
+} while (0)
+
+typedef struct e3_dma
+{
+    E3_DmaType         dma_u;
+    E3_uint32          dma_size;
+    E3_Addr            dma_source;
+    E3_Addr            dma_dest;
+    E3_Addr            dma_destEvent;
+    E3_CookieVProc     dma_destCookieProc;
+    E3_Addr            dma_srcEvent;
+    E3_CookieVProc     dma_srcCookieProc;
+} E3_DMA;
+
+
+/*
+ * Word-swapped version of DMA descriptor.
+ * This is used by the UltraSPARC code to format the descriptor
+ * in main memory before block-copying it down to Elan SDRAM.
+ * In the process it does a dword (64-bit) conversion and so swaps
+ * the word order on a double-word pair basis
+ */
+typedef struct e3_dma_swapped
+{
+    E3_uint32          dma_size;
+    E3_DmaType         dma_u;
+    E3_Addr            dma_dest;
+    E3_Addr            dma_source;
+    E3_CookieVProc     dma_destCookieProc;
+    E3_Addr            dma_destEvent;
+    E3_CookieVProc     dma_srcCookieProc;
+    E3_Addr            dma_srcEvent;
+} E3_DMA_SWAPPED;
+
+/* Define a Main memory structure for DMA desc based on Endianess of machine */
+#if defined(__LITTLE_ENDIAN__)
+#define E3_DMA_MAIN E3_DMA
+#else
+#define E3_DMA_MAIN E3_DMA_SWAPPED;
+#endif
+
+#define dma_type        dma_u.type
+#define dma_failCount    dma_u.s.failCount
+#define dma_isRemote     dma_u.s.isRemote
+#define dma_opCode       dma_u.s.opCode
+#define dma_direction    dma_u.s.direction
+#define dma_dataType     dma_u.s.dataType
+#define dma_queueContext dma_u.s.Context
+
+#define dma_destCookieVProc   dma_destCookieProc.cookie_vproc
+#define dma_destVProc        dma_destCookieProc.s.vproc
+#define dma_destCookie       dma_destCookieProc.s.cookie
+#define dma_srcCookieVProc    dma_srcCookieProc.cookie_vproc
+#define dma_srcVProc         dma_srcCookieProc.s.vproc
+#define dma_srcCookie        dma_srcCookieProc.s.cookie
+
+/*
+ * Values for dma_opCode
+ */
+#define DMA_NORMAL                     0
+#define DMA_QUEUED                     1
+#define DMA_NORMAL_BROADCAST           2
+#define DMA_QUEUED_BROADCAST           3
+#define DMA_NORMAL_UNSAFE              4
+#define DMA_QUEUED_UNSAFE              5
+#define DMA_NORMAL_BROADCAST_UNSAFE    6
+#define DMA_QUEUED_BROADCAST_UNSAFE    7
+
+/*
+ * Values for dma_direction
+ */
+#define DMA_WRITE              0
+#define DMA_READ_REQUEUE       1
+#define DMA_READ               3
+#define DMA_READ_BROADCAST     7
+
+/*
+ * Values for dma_dataType
+ */
+#define DMA_BYTE               0
+#define DMA_HALFWORD           1
+#define DMA_WORD               2
+#define DMA_DOUBLE             3
+
+/* OUT OF DATE ?
+  #define DMA_OPCODE_SHIFT     3
+  #define DMA_FAILCOUNT_SHIFT  9
+*/
+#define DMA_TYPE_ISREMOTE      (1 << 15)
+#define DMA_TYPE_READ          (3 << 2)
+#define DMA_TYPE_READ_REQUEUE  (1 << 2)
+#define DMA_TYPE_DIRECTION_MASK        (3 << 2)
+
+#endif /* __ELAN3_DMA_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/e3types.h
===================================================================
--- linux-2.4.21.orig/include/elan3/e3types.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/e3types.h        2005-06-01 23:12:54.716421320 -0400
@@ -0,0 +1,82 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_E3TYPES_H
+#define __ELAN3_E3TYPES_H
+
+#ident "$Id: e3types.h,v 1.18 2002/08/09 11:23:33 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/e3types.h,v $ */
+
+#include <qsnet/config.h>
+/*
+ * "flip" values for correctly indexing into
+ * block data which was copied from the Elan
+ * using 64 bit accesses.
+ */
+#if defined(__LITTLE_ENDIAN__)
+#  define ByteEndianFlip  0
+#  define ShortEndianFlip 0
+#  define WordEndianFlip  0
+#else
+#  define ByteEndianFlip  7
+#  define ShortEndianFlip 3
+#  define WordEndianFlip  1
+#endif
+
+
+#ifndef _ASM
+
+typedef signed int        E3_int;
+typedef unsigned int              E3_uint;
+
+typedef signed char       E3_int8;
+typedef unsigned char     E3_uint8;
+
+typedef signed short      E3_int16;
+typedef unsigned short            E3_uint16;
+
+typedef signed int        E3_int32;
+typedef unsigned int              E3_uint32;
+
+#ifdef __ELAN3__
+typedef signed long long   E3_int64;
+typedef unsigned long long E3_uint64;
+#ifdef _MAIN_LP64
+/* NOTE: If the Main is 64-bit we declare the Elan thread's
+ * E3_uintptr to be 64-bits too
+ */
+typedef unsigned long long E3_uintptr;
+#else
+typedef unsigned long      E3_uintptr;
+#endif
+
+#else
+
+#ifdef _LP64
+typedef signed long        E3_int64;
+typedef unsigned long      E3_uint64;
+typedef unsigned long      E3_uintptr;
+#else /* _ILP32 */
+typedef signed long long   E3_int64;
+typedef unsigned long long E3_uint64;
+typedef unsigned long      E3_uintptr;
+#endif
+
+#endif /* __ELAN3__ */
+
+/* 32-bit Elan3 address */
+typedef E3_uint32         E3_Addr;
+
+#endif /* _ASM */
+
+#endif /* __ELAN3_E3TYPES_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elan3mmu.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elan3mmu.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elan3mmu.h       2005-06-01 23:12:54.717421168 -0400
@@ -0,0 +1,346 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_ELAN3MMU_H
+#define __ELAN3_ELAN3MMU_H
+
+#ident "$Id: elan3mmu.h,v 1.40.2.1 2004/12/14 10:19:48 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elan3mmu.h,v $*/
+
+
+#include <elan3/pte.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct elan3mmu_global_stats
+{
+    int                version;
+    int                pteload;
+    int                pteunload;
+    int                ptereload;
+
+    int                streamable_alloc;
+    int                streamable_free;
+    int                streamable_alloc_failed;
+
+    int                num_ptbl_level[4]; /* number of level N  ptbls */
+
+    int                create_ptbl_failed;                     /* count of ptbl creation failure */
+
+    int         lX_alloc_l3;                           /* count of l3 ptbls used as lX */
+    int         lX_freed_l3;                           /* count of lX ptbls freed as l3 */
+
+    int                l2_alloc_l3;                            /* count of l3 ptbls used as l2 */
+    int                l2_freed_l3;                            /* count of l2 ptbls freed as l3 */
+
+    int                stolen_ptbls;                           /* count of l3 ptbls stolen */
+} ELAN3MMU_GLOBAL_STATS;
+
+#define ELAN3MMU_STATS_VERSION         1
+
+#define ELAN3MMU_STAT(what)            (elan3mmu_global_stats.what++)
+#define ELAN3MMU_SET_STAT(what,count)  (elan3mmu_global_stats.what = count)
+
+#ifdef __KERNEL__
+
+#define ELAN3_PT_SHIFT (ELAN3_L2_SHIFT + 2)
+
+typedef struct elan3_ptbl
+{
+    struct elan3_ptbl   *ptbl_parent;                          /* Parent page table, or next on freelist */
+    struct elan3mmu     *ptbl_elan3mmu;                        /* elan3mmu we're allocated for */
+    E3_Addr              ptbl_base;                            /* Virtual address we're mapping */
+    u_char               ptbl_index;                           /* Index in ptbl group */
+    u_char               ptbl_valid;                           /* Number of valid entries */
+    u_char               ptbl_flags;                           /* Flags, defined below. */
+    u_char               ptbl_spare;
+} ELAN3_PTBL;
+
+#define ptbl_next      ptbl_parent                             /* Parent pointer is next pointer when on free list */
+
+#define PTBL_LEVEL_X            0x00
+#define PTBL_LEVEL_1           0x01
+#define PTBL_LEVEL_2           0x02
+#define PTBL_LEVEL_3           0x03
+#define PTBL_LEVEL_MASK                0x03
+#define PTBL_LOCKED            0x04                            /* Page table is locked,  protects all fields */
+#define PTBL_KEEP              0x08                            /* This ptbl is not to be stolen */
+#define PTBL_ALLOCED           0x10                            /* This ptbl has been allocated, and is not free */
+#define PTBL_GROUPED           0x20                            /* This ptbl is a member of a group of ptbls */
+#define PTBL_KERNEL            0x80                            /* This ptbl is allocated for the kernel */
+
+#define PTBL_LEVEL(flags)      ((flags) & PTBL_LEVEL_MASK)
+#define PTBL_IS_LOCKED(flags)  (((flags) & (PTBL_LOCKED|PTBL_ALLOCED)) == (PTBL_LOCKED|PTBL_ALLOCED))
+
+#if ELAN3_PAGE_SHIFT == 13
+#  define PTBL_GROUP_SIZE      8192                            /* page table groups are 8k bytes */
+#  define PTBLS_PER_GROUP_L1   8                               /* Number of level 1 tables in a group */
+#  define PTBLS_PER_GROUP_L2   32                              /*   ... level 2 */
+#  define PTBLS_PER_GROUP_L3   32                              /*   ... level 3 */
+#  define PTBLS_PER_GROUP_LX   32                              /*   ... level X */
+#  define PTBLS_PER_GROUP_MAX  32                              /*  max of l1,l2,l3,lX */
+#else
+#  define PTBL_GROUP_SIZE      4096                            /* page table groups are 4k bytes */
+#  define PTBLS_PER_GROUP_L1   4                               /* Number of level 1 tables in a group */
+#  define PTBLS_PER_GROUP_L2   16                              /*   ... level 2 */
+#  define PTBLS_PER_GROUP_L3   8                               /*   ... level 3 */
+#  define PTBLS_PER_GROUP_LX   16                              /*   ... level X */
+#  define PTBLS_PER_GROUP_MAX  16                              /*  max of l1,l2,l3,lX */
+#endif
+
+#define HMES_PER_GROUP         (PTBLS_PER_GROUP_L3*ELAN3_L3_ENTRIES)
+
+#if ELAN3_PAGE_SHIFT == 13
+#  define PTBLS_PER_PTBL_L1    4                               /* 256 PTPs */
+#  define PTBLS_PER_PTBL_L2    1                               /* 64 PTPs */
+#  define PTBLS_PER_PTBL_L3    1                               /* 32 PTEs */
+#else
+#  define PTBLS_PER_PTBL_L1    4                               /* 256 PTPs */
+#  define PTBLS_PER_PTBL_L2    1                               /* 64 PTPs */
+#  define PTBLS_PER_PTBL_L3    2                               /* 64 PTEs */
+#endif
+
+#define ELAN3_LX_ENTRIES     (32) 
+#define PTBLS_PER_PTBL_LX   (1)        
+
+#define L1_VA_PER_PTBL (ELAN3_L1_SIZE*(ELAN3_L1_ENTRIES/PTBLS_PER_PTBL_L1))    /* 4 ptbl for L1 */
+#define L2_VA_PER_PTBL (ELAN3_L2_SIZE*(ELAN3_L2_ENTRIES/PTBLS_PER_PTBL_L2))    /* 1 ptbl for L2 */
+#define L3_VA_PER_PTBL (ELAN3_L3_SIZE*(ELAN3_L3_ENTRIES/PTBLS_PER_PTBL_L3))    /* 1 ptbl for L3 */
+
+typedef struct elan3_ptbl_gr
+{
+    struct elan3_ptbl_gr *pg_next;                             /* Next in list. */
+    int                         pg_level;                              /* Level PG allocated for */
+    sdramaddr_t                 pg_addr;                               /* sdram offset of ptes/ptps */    
+    ELAN3_PTBL          pg_ptbls[PTBLS_PER_GROUP_MAX];         /* The actual page tables */
+} ELAN3_PTBL_GR;
+
+
+/*
+ * The elan3mmu structure is the mmu dependant hardware address translation
+ * structure linked to the address space structure to show the translatioms
+ * provided by the elan for an address sapce.
+ *
+ * We also have a doubly linked list of 'regions' which allow the 
+ * elan3mmu code to determine the access permissions for the elan 
+ * dependant on the virtual address that the translation is being
+ * loaded at.
+ */
+
+typedef struct elan3mmu_rgn
+{
+    struct elan3mmu_rgn *rgn_mnext;                            /* Doubly linked list of regions */
+    struct elan3mmu_rgn *rgn_mprev;                            /*   sorted on main address */ 
+    caddr_t             rgn_mbase;                             /* main address of base of region */
+
+    struct elan3mmu_rgn *rgn_enext;                            /* Doubly linked list of regions */
+    struct elan3mmu_rgn *rgn_eprev;                            /*   sorted on elan address */
+    E3_Addr             rgn_ebase;                             /* elan address of base of region */
+
+    u_int               rgn_len;                               /* length of region */
+    u_int               rgn_perm;                              /* elan access permission */
+} ELAN3MMU_RGN;
+
+typedef struct elan3mmu
+{
+    spinlock_t             elan3mmu_lock;                      /* spinlock lock for regions */
+    ELAN3MMU_RGN           *elan3mmu_mrgns;                    /* Doubly linked list of memory regions */
+    ELAN3MMU_RGN          *elan3mmu_mtail;                     /* Last memory region on list */
+    ELAN3MMU_RGN          *elan3mmu_mrgnlast;                  /* Last region 'hit' */
+
+    ELAN3MMU_RGN           *elan3mmu_ergns;                    /* Doubly linked list of memory regions */
+    ELAN3MMU_RGN          *elan3mmu_etail;                     /* Last memory region on list */
+    ELAN3MMU_RGN          *elan3mmu_ergnlast;                  /* Last region 'hit' */
+
+    struct elan3_dev        *elan3mmu_dev;                     /* Elan device we're using. */
+    struct elan3_ctxt     *elan3mmu_ctxt;                      /* Elan ctxt we're associated with */
+
+    sdramaddr_t             elan3mmu_ctp;                      /* Context table entry for our context */
+    ELAN3_PTBL            *elan3mmu_l1ptbl;                    /* Level 1 Page table (first of 4) */
+
+    spinlock_t             elan3mmu_lXptbl_lock;               /* spinlock for level X table list */
+    ELAN3_PTBL              *elan3mmu_lXptbl;                    /* Level X Page table list         */
+
+#ifdef LINUX
+    struct mm_struct       *elan3mmu_coproc_mm;                        /* Linux mm we're mapping */
+#endif
+} ELAN3MMU;
+
+_NOTE(LOCK_ORDER(elan3mmu::elan3mmu_lock elan3_dev::IntrLock))
+
+_NOTE(MUTEX_PROTECTS_DATA(elan3mmu::elan3mmu_lock,
+                         elan3mmu::elan3mmu_mrgns elan3mmu::elan3mmu_mtail
+                         elan3mmu::elan3mmu_ergns elan3mmu::elan3mmu_etail))
+/* protected by dev->IntrLock for read by device driver */
+_NOTE(DATA_READABLE_WITHOUT_LOCK(elan3mmu::elan3mmu_mrgns elan3mmu::elan3mmu_mtail
+                                elan3mmu::elan3mmu_ergns elan3mmu::elan3mmu_etail))
+
+_NOTE(SCHEME_PROTECTS_DATA("only set to valid region", 
+                          elan3mmu::elan3mmu_ergnlast elan3mmu::elan3mmu_mrgnlast))
+
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::IntrLock,
+                         elan3mmu::elan3mmu_l1ptbl 
+                         elan3mmu::elan3mmu_ctp 
+                         elan3mmu::elan3mmu_dev))
+
+_NOTE(DATA_READABLE_WITHOUT_LOCK(elan3mmu::elan3mmu_l1ptbl
+                                elan3mmu::elan3mmu_ctp 
+                                elan3mmu::elan3mmu_dev))
+
+/*
+ * Macros for accessing ptes/ptbls/ptbl_grs
+ */
+
+#define OFFSETOF(object,member)        /* calculate offset of structure member */ \
+       ((size_t) (&(((object *)0)->member)))
+#define PTBL_TO_GR(ptbl)       /* convert ptbl to ptbl group */ \
+       ((ELAN3_PTBL_GR *) ((caddr_t) ((ptbl) - (ptbl)->ptbl_index) - OFFSETOF(ELAN3_PTBL_GR,pg_ptbls[0])))
+#define PTBL_TO_PTADDR(ptbl)   /* convert ptbl to a ptp pointing at it */ \
+        (PTBL_TO_GR(ptbl)->pg_addr + ((ptbl)->ptbl_index<<ELAN3_PT_SHIFT))
+#define PTE_TO_HME(ptbl,pte)   /* convert pte to corresponding hme */ \
+        (PTBL_TO_GR(ptbl)->pg_hmes + ((pte) - (ELAN3_PTE *) PTBL_TO_GR(ptbl)->pg_vaddr))
+#define HME_TO_PTE(ptebl,hme)  /* convert hme to corresponding pte */ \
+        ((ELAN3_PTE *) PTBL_TO_GR(ptbl)->pg_vaddr + ((hme) - (PTBL_TO_GR(ptbl)->pg_hmes)))
+
+
+/* Flags for lock_ptbl */
+#define LK_PTBL_NOWAIT         0x1
+#define LK_PTBL_FAILOK         0x2
+
+/* Return values for lock_ptbl */
+#define LK_PTBL_OK             0x0
+#define LK_PTBL_MISMATCH       0x1
+#define LK_PTBL_FAILED         0x2
+
+/* Flags for elan3mmu_ptesync */
+#define        NO_MLIST_LOCK   0
+#define        MLIST_LOCKED    1
+
+/* Flags for elan3mmu_pteload */
+#define PTE_LOAD               0x00
+#define PTE_LOAD_LOCK          0x01                            /* translation should be locked */
+#define PTE_LOAD_NOSYNC                0x02                            /* ref/mod bits should not be sync'ed to page */
+#define PTE_NO_SLEEP            0x04                            /* true if we cant sleep */
+#define PTE_NO_STEAL           0x08                            /* true if we don't want to steal ptbls */
+
+#define PTE_LOAD_ENDIAN_MASK   0x10                            /* mask for endian-ness */
+#define PTE_LOAD_LITTLE_ENDIAN 0x00                            /* translation is to little-endian memory */
+#define PTE_LOAD_BIG_ENDIAN    0x10                            /* translation is to big-endian memory */
+
+
+/* Flags for elan3mmu_unload */
+#define PTE_UNLOAD             0x00
+#define PTE_UNLOAD_UNLOCK      0x01
+#define PTE_UNLOAD_NOFLUSH     0x02
+#define PTE_UNLOAD_NOSYNC      0x04
+
+extern int         elan3mmu_debug;
+#ifdef DEBUG_PRINTF
+#  define HAT_PRINTF0(n,msg)            ((elan3mmu_debug & n) ? (void) elan3_debugf (NULL, DBG_HAT, msg)             : (void) 0)
+#  define HAT_PRINTF1(n,msg,a)          ((elan3mmu_debug & n) ? (void) elan3_debugf (NULL, DBG_HAT, msg,a)           : (void) 0)
+#  define HAT_PRINTF2(n,msg,a,b)        ((elan3mmu_debug & n) ? (void) elan3_debugf (NULL, DBG_HAT, msg,a,b)         : (void) 0)
+#  define HAT_PRINTF3(n,msg,a,b,c)      ((elan3mmu_debug & n) ? (void) elan3_debugf (NULL, DBG_HAT, msg,a,b,c)       : (void) 0)
+#  define HAT_PRINTF4(n,msg,a,b,c,d)    ((elan3mmu_debug & n) ? (void) elan3_debugf (NULL, DBG_HAT, msg,a,b,c,d)     : (void) 0)
+#  define HAT_PRINTF5(n,msg,a,b,c,d,e)  ((elan3mmu_debug & n) ? (void) elan3_debugf (NULL, DBG_HAT, msg,a,b,c,d,e)   : (void) 0)
+#  define HAT_PRINTF6(n,msg,a,b,c,d,e,f) ((elan3mmu_debug & n) ? (void) elan3_debugf (NULL, DBG_HAT, msg,a,b,c,d,e,f) : (void) 0)
+#  ifdef LINUX
+#    define HAT_PRINTF(n,args...)        ((elan3mmu_debug & n) ? (void) elan3_debugf(NULL, DBG_HAT, ##args) : (void) 0)
+#  endif
+#else
+#  define HAT_PRINTF0(n,msg)
+#  define HAT_PRINTF1(n,msg,a)
+#  define HAT_PRINTF2(n,msg,a,b)
+#  define HAT_PRINTF3(n,msg,a,b,c)
+#  define HAT_PRINTF4(n,msg,a,b,c,d)
+#  define HAT_PRINTF5(n,msg,a,b,c,d,e)
+#  define HAT_PRINTF6(n,msg,a,b,c,d,e,f)
+#  ifdef LINUX
+#    define HAT_PRINTF(n,args...)
+#  endif
+#endif
+
+/* elan3mmu_generic.c */
+extern ELAN3MMU_GLOBAL_STATS elan3mmu_global_stats;
+
+extern void         elan3mmu_init (void);
+extern void         elan3mmu_fini (void);
+
+extern ELAN3MMU           *elan3mmu_alloc (struct elan3_ctxt *ctxt);
+extern void        elan3mmu_free (ELAN3MMU *elan3mmu);
+
+extern void          elan3mmu_set_context_filter (ELAN3_DEV *dev, int ctx, int disabled, E3_uint32 Pend, E3_uint32 *Maskp);
+extern int          elan3mmu_attach (ELAN3_DEV *dev, int ctx, ELAN3MMU *elan3mmu, sdramaddr_t routeTable, E3_uint32 routeMask);
+extern void         elan3mmu_detach (ELAN3_DEV *dev, int ctx);
+
+extern ELAN3MMU_RGN *elan3mmu_findrgn_elan (ELAN3MMU *elan3mmu, E3_Addr addr, int tail);
+extern int           elan3mmu_addrgn_elan (ELAN3MMU *elan3mmu, ELAN3MMU_RGN *nrgn);
+extern ELAN3MMU_RGN *elan3mmu_removergn_elan (ELAN3MMU *elan3mmu, E3_Addr addr);
+extern ELAN3MMU_RGN *elan3mmu_rgnat_elan (ELAN3MMU *elan3mmu, E3_Addr addr);
+extern ELAN3MMU_RGN *elan3mmu_findrgn_main (ELAN3MMU *elan3mmu, caddr_t addr, int tail);
+extern int           elan3mmu_addrgn_main (ELAN3MMU *elan3mmu, ELAN3MMU_RGN *nrgn);
+extern ELAN3MMU_RGN *elan3mmu_removergn_main (ELAN3MMU *elan3mmu, caddr_t addr);
+extern ELAN3MMU_RGN *elan3mmu_rgnat_main (ELAN3MMU *elan3mmu, caddr_t addr);
+
+extern int          elan3mmu_setperm (ELAN3MMU *elan3mmu, caddr_t maddr, E3_Addr eaddr, u_int len, u_int perm);
+extern void         elan3mmu_clrperm (ELAN3MMU *elan3mmu, E3_Addr addr, u_int len);
+extern int          elan3mmu_checkperm (ELAN3MMU *elan3mmu, E3_Addr addr, u_int len, u_int access);
+extern caddr_t      elan3mmu_mainaddr (ELAN3MMU *elan3mmu, E3_Addr addr);
+extern E3_Addr      elan3mmu_elanaddr (ELAN3MMU *elan3mmu, caddr_t addr);
+
+extern void          elan3mmu_expand (ELAN3MMU *elan3mmu, E3_Addr addr, int len, int level, int attr);
+extern void          elan3mmu_reserve (ELAN3MMU *elan3mmu, E3_Addr addr, u_int npages, sdramaddr_t *);
+extern void          elan3mmu_release (ELAN3MMU *elan3mmu, E3_Addr addr, u_int npages, sdramaddr_t *);
+
+extern void          elan3mmu_pteload (ELAN3MMU *elan3mmu, int level, E3_Addr addr, physaddr_t paddr, int perm, int attr);
+extern void         elan3mmu_unload (ELAN3MMU *elan3mmu, E3_Addr addr, u_int len, int flags);
+extern void         elan3mmu_sync (ELAN3MMU *elan3mmu, E3_Addr addr, u_int len, u_int clearflag);
+extern void         elan3mmu_pteunload (ELAN3_PTBL *ptbl, sdramaddr_t pte, int flags, int got_mlist_lock);
+extern void         elan3mmu_ptesync (ELAN3_PTBL *ptbl, sdramaddr_t pte, int flags, int got_mlist_lock);
+extern sdramaddr_t   elan3mmu_ptp2pte (ELAN3MMU *elan3mmu, sdramaddr_t ptp, int level);
+extern sdramaddr_t   elan3mmu_ptefind (ELAN3MMU *elan3mmu, E3_Addr, int *level, ELAN3_PTBL **pptbl, spinlock_t **plock, unsigned long *flags);
+extern sdramaddr_t   elan3mmu_ptealloc (ELAN3MMU *elan3mmu, E3_Addr, int level, ELAN3_PTBL **pptbl, spinlock_t **plock, int attr, unsigned long *flags);
+extern void         elan3mmu_l1inval (ELAN3MMU *elan3mmu, ELAN3_PTBL *l1ptbl, int flags);
+extern int           elan3mmu_l2inval (ELAN3MMU *elan3mmu, ELAN3_PTBL *l2ptbl, int flags, E3_Addr addr, spinlock_t **pl2lock, unsigned long *lock_flags);
+extern int           elan3mmu_l3inval (ELAN3MMU *elan3mmu, ELAN3_PTBL *l3ptbl, int flags, E3_Addr addr, spinlock_t **pl3lock, unsigned long *lock_flags);
+
+extern void          elan3mmu_free_l1ptbl (ELAN3_DEV *dev, ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags);
+extern void          elan3mmu_free_l2ptbl (ELAN3_DEV *dev, ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags);
+extern void          elan3mmu_free_l3ptbl (ELAN3_DEV *dev, ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags);
+
+extern int          elan3mmu_lock_this_ptbl (ELAN3_PTBL *ptbl, int flag, spinlock_t **plock, unsigned long *flags);
+extern int           elan3mmu_lock_ptbl (ELAN3_PTBL *ptbl, u_int flag, ELAN3MMU *elan3mmu, E3_Addr va, int level, spinlock_t **plock, unsigned long *flags);
+extern void         elan3mmu_unlock_ptbl (ELAN3_PTBL *ptbl, spinlock_t *lock, unsigned long flags);
+
+/* elan3mmu_osdep.c */
+extern void         elan3mmu_init_osdep (void);
+extern void         elan3mmu_fini_osdep (void);
+extern void         elan3mmu_alloc_osdep (ELAN3MMU *elan3mmu);
+extern void         elan3mmu_free_osdep (ELAN3MMU *elan3mmu);
+extern ELAN3_PTE     elan3mmu_phys_to_pte (ELAN3_DEV *dev, physaddr_t paddr, int perm);
+extern ELAN3_PTE     elan3mmu_kernel_invalid_pte (ELAN3MMU *elan3mmu);
+
+#if defined (DIGITAL_UNIX)
+#  include <elan3/elan3mmu_dunix.h>
+#elif defined (LINUX)
+#  include <elan3/elan3mmu_linux.h>
+#endif
+
+#endif /* __KERNEL__ */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __ELAN3_ELAN3MMU_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elan3mmu_linux.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elan3mmu_linux.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elan3mmu_linux.h 2005-06-01 23:12:54.717421168 -0400
@@ -0,0 +1,39 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_MMU_LINUX_H
+#define __ELAN3_MMU_LINUX_H
+
+#ident "$Id: elan3mmu_linux.h,v 1.12 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elan3mmu_linux.h,v $*/
+
+/* XXX copy of elan3mmu_dunix.h */
+
+#define ALLOC_ELAN3MMU(ptr,cansleep)           KMEM_ALLOC(ptr, ELAN3MMU *, sizeof (ELAN3MMU), cansleep)
+#define ALLOC_PTBL_GR(ptr,cansleep)            KMEM_ALLOC(ptr, ELAN3_PTBL_GR *, sizeof (ELAN3_PTBL_GR), cansleep)
+#define ALLOC_ELAN3MMU_RGN(ptr,cansleep)       KMEM_ALLOC(ptr, ELAN3MMU_RGN *, sizeof (ELAN3MMU_RGN), cansleep)
+#define ALLOC_HMENTS(ptr,cansleep)             KMEM_ALLOC((ptr,ELAN3_HMENT *, sizeof (ELAN3_HMENT), cansleep)
+
+#define FREE_ELAN3MMU(ptr)                     KMEM_FREE(ptr,sizeof (ELAN3MMU))
+#define FREE_PTBL_GR(ptr)                      KMEM_FREE(ptr,sizeof (ELAN3_PTBL_GR))
+#define FREE_ELAN3MMU_RGN(ptr)                 KMEM_FREE(ptr,sizeof (ELAN3MMU_RGN))
+#define FREE_HMENTS(ptr)                       KMEM_FREE(ptr,sizeof (ELAN3_HMENT))
+
+extern void         elan3mmu_init_osdep(void);
+extern void         elan3mmu_fini_osdep(void);
+
+extern void          elan3mmu_pte_range_unload (ELAN3MMU *elan3mmu, struct mm_struct *mm, caddr_t addr, unsigned long len);
+extern void          elan3mmu_pte_range_update (ELAN3MMU *elan3mmu, struct mm_struct *mm, caddr_t addr, unsigned long len);
+extern void          elan3mmu_pte_ctxt_unload(ELAN3MMU *elan3mmu);
+
+#endif
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elan3ops.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elan3ops.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elan3ops.h       2005-06-01 23:12:54.718421016 -0400
@@ -0,0 +1,42 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+/* $Id: elan3ops.h,v 1.3 2003/09/24 13:57:24 david Exp $ */
+/* $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elan3ops.h,v $ */
+
+#ifndef _ELAN3_OPS_H
+#define _ELAN3_OPS_H
+
+int get_position          (void *arg, ELAN_POSITION *position);
+int set_position          (void *arg, unsigned short nodeId, unsigned short numNodes);
+
+int elan3mod_create_cap   (void *arg, ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap);
+int elan3mod_destroy_cap  (void *arg, ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap);
+
+int elan3mod_create_vp    (void *arg, ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map);
+int elan3mod_destroy_vp   (void *arg, ELAN_CAP_OWNER owner, ELAN_CAPABILITY *cap, ELAN_CAPABILITY *map);
+
+int elan3mod_attach_cap   (void *arg_ctxt, ELAN_CAPABILITY *cap);
+int elan3mod_detach_cap   (void *arg_ctxt);
+
+extern ELAN_DEV_OPS elan3_dev_ops;
+
+int stats_get_index_name  (void *arg, uint index, caddr_t name);
+int stats_get_block       (void *arg, uint entries, ulong *value);
+int stats_clear_block     (void *arg);
+
+int elan3_register_dev_stats   (ELAN3_DEV * dev);
+void elan3_deregister_dev_stats (ELAN3_DEV * dev);
+
+
+#endif /* __ELAN3_OPS_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elanctxt.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elanctxt.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elanctxt.h       2005-06-01 23:12:54.719420864 -0400
@@ -0,0 +1,856 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_ELANCTXT_H
+#define _ELAN3_ELANCTXT_H
+
+#ident "$Id: elanctxt.h,v 1.81 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elanctxt.h,v $*/
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <elan3/elanregs.h>
+#include <elan3/vmseg.h>
+
+#define BumpUserStat(ctxt, stat)       ((ctxt)->FlagPage->stat++)
+
+#if defined(__LITTLE_ENDIAN__)
+
+typedef union _CProcTrapBuf
+{
+    E3_uint64 Align64;
+    struct
+    {
+       E3_uint32 Areg;
+       E3_uint32 Breg;
+    } r;
+    struct
+    {
+       E3_uint32 Addr;
+       E3_uint32 ContextType;
+    } s;
+} CProcTrapBuf_BE;
+
+typedef E3_EventInt        E3_EventInt_BE;
+typedef E3_IprocTrapHeader E3_IprocTrapHeader_BE;
+typedef E3_IprocTrapData   E3_IprocTrapData_BE;
+typedef E3_FaultSave      E3_FaultSave_BE;
+
+typedef union
+{
+    E3_uint64  Align64;
+    E3_DMA      s;
+} E3_DMA_BE;
+
+typedef E3_ThreadQueue     E3_ThreadQueue_BE;
+
+#else
+
+/* "Big-Endian" data structures copied by 64 bit loads, these are 32 bit word flipped */
+/* from the corresponding data structure. */
+
+typedef union _CProcTrapBuf
+{
+    E3_uint64 Align64;
+    struct
+    {
+       E3_uint32 Breg;
+       E3_uint32 Areg;
+    } r;
+    struct
+    {
+       E3_uint32 ContextType;
+       E3_uint32 Addr;
+    } s;
+} CProcTrapBuf_BE;
+
+typedef union _E3_EventInt_BE
+{
+    E3_uint64    Align64;
+    struct {
+       E3_uint32 EventContext; /* Bits 16 to 28 */
+       E3_uint32 IntCookie;
+    } s;
+} E3_EventInt_BE;
+
+typedef union _E3_IprocTrapHeader_BE
+{
+   E3_uint64            Align64;
+
+   struct
+   {
+      E3_uint32                 TrAddr;
+      E3_TrTypeCntx     TrTypeCntx;
+      union
+      {
+        E3_IProcStatus_Reg u_IProcStatus;
+        E3_uint32          u_TrData1;
+      } ipsotd;
+      E3_uint32                 TrData0;
+   } s;
+} E3_IprocTrapHeader_BE;
+
+typedef E3_IprocTrapData E3_IprocTrapData_BE;
+
+typedef union _E3_FaultSave_be
+{
+    E3_uint64                  Align64;
+    struct {
+       volatile E3_uint32      FaultContext;
+       E3_FaultStatusReg       FSR;
+       volatile E3_uint32      EventAddress;
+       volatile E3_uint32      FaultAddress;
+    } s;
+} E3_FaultSave_BE;
+
+typedef union _e3_dma_be
+{
+    E3_uint64          Align64;
+    struct {
+       E3_uint32       dma_size;
+       E3_DmaType      dma_u;
+       E3_Addr         dma_dest;
+       E3_Addr         dma_source;
+       E3_CookieVProc  dma_destCookieProc;
+       E3_Addr         dma_destEvent;
+       E3_CookieVProc  dma_srcCookieProc;
+       E3_Addr         dma_srcEvent;
+    } s;
+} E3_DMA_BE;
+
+typedef union _E3_ThreadQueue_BE
+{
+   E3_uint64   Align64;
+   struct
+   {
+       /* copied by 64 bit copy from elan to main */
+       E3_uint32 :3;           /* Bits 29 to 31 */
+       E3_uint32 Context:13;   /* Bits 16 to 28 */
+       E3_uint32 :16;          /* Bits 0  to 15 */
+       E3_Addr  Thread;        /* Bits 32 to 63 */
+   } s;
+} E3_ThreadQueue_BE;
+
+#endif /* defined(LITTLE_ENDIAN) || defined(__LITTLE_ENDIAN__) */
+
+typedef struct neterr_msg
+{
+    E3_uint32          Rail;                                   /* Rail error received on */
+    ELAN_CAPABILITY    SrcCapability;                          /* Capability of source of packet */
+    ELAN_CAPABILITY    DstCapability;                          /* Capability of dest of packet */
+
+    E3_uint32          DstProcess;                             /* Virtual Process of dest of packet */
+    E3_Addr            CookieAddr;                             /* Cookie Address (or NULL for DMA) */
+    E3_uint32          CookieVProc;                            /* Cookie and VP (identifies DMA) */
+    E3_uint32          NextCookie;                             /* Next Cookie value (for thread) */
+    E3_uint32          WaitForEop;                             /* Wait for EOP transaction */
+} NETERR_MSG;
+
+#ifdef __KERNEL__
+
+/*
+ * Associated with each input channel can be a network error
+ * resolver structure, which can be queued on the network 
+ * error resolver threads to perform RPCs to the other kernels
+ * when a network error occurs with an identify transaction
+ * included
+ */
+typedef struct neterr_resolver
+{
+    struct neterr_resolver *Next;
+
+    spinlock_t             Lock;
+
+    struct elan3_ctxt       *Ctxt;
+    ELAN_LOCATION          Location;
+
+    int                            Completed;
+    int                            Status;
+    long                   Timestamp;
+
+    NETERR_MSG             Message;
+} NETERR_RESOLVER;
+
+
+typedef struct neterr_fixup
+{
+    struct neterr_fixup           *Next;
+
+    kcondvar_t             Wait;
+    int                            Completed;
+    int                            Status;
+
+    NETERR_MSG             Message;
+} NETERR_FIXUP;
+
+#endif /* __KERNEL__ */
+
+/* Each of the following structures must be padded to a whole */
+/* number of 64 bit words since the kernel uses 64 bit load/stores */
+/* to transfer the elan register state. */
+typedef struct command_trap
+{
+    E3_Status_Reg      Status;                                 /* 4  bytes */
+    E3_uint32          Pad;                                    /* 4  bytes */
+    E3_FaultSave_BE    FaultSave;                              /* 16 bytes */
+    CProcTrapBuf_BE            TrapBuf;                                /* 8  bytes */
+} COMMAND_TRAP;
+
+typedef struct thread_trap
+{
+    E3_uint32          Registers[32];                          /* 128 bytes */
+#define REG_GLOBALS    0
+#define REG_OUTS       8
+#define REG_LOCALS     16
+#define REG_INS                24
+
+    E3_FaultSave_BE    FaultSave;                              /* 16 bytes */
+    E3_FaultSave_BE     DataFaultSave;                         /* 16 bytes */
+    E3_FaultSave_BE     InstFaultSave;                         /* 16 bytes */
+    E3_FaultSave_BE     OpenFaultSave;                         /* 16 bytes */
+    
+    E3_Status_Reg      Status;                                 /* 4 bytes */
+
+    E3_Addr            pc;                                     /* 4 bytes */
+    E3_Addr            npc;                                    /* 4 bytes */
+    E3_Addr            StartPC;                                /* 4 bytes */
+    E3_Addr            sp;                                     /* 4 bytes */
+    E3_uint32          mi;                                     /* 4 bytes */
+    E3_TrapBits                TrapBits;                               /* 4 bytes */
+    E3_DirtyBits       DirtyBits;                              /* 4 bytes */
+} THREAD_TRAP;
+
+typedef struct dma_trap
+{
+    E3_DMA_BE          Desc;                                   /* 32 bytes */
+    E3_FaultSave_BE    FaultSave;                              /* 16 bytes */
+    E3_FaultSave_BE    Data0;                                  /* 16 bytes */
+    E3_FaultSave_BE    Data1;                                  /* 16 bytes */
+    E3_FaultSave_BE    Data2;                                  /* 16 bytes */
+    E3_FaultSave_BE    Data3;                                  /* 16 bytes */
+    E3_Status_Reg      Status;                                 /* 4 bytes */
+    E3_DmaInfo         PacketInfo;                             /* 4 bytes */
+} DMA_TRAP;
+
+typedef struct input_trap
+{
+    E3_uint32             State;                               /* 4 bytes */
+    E3_Status_Reg         Status;                              /* 4 bytes */
+    E3_FaultSave_BE       FaultSave;                           /* 16 bytes */
+    
+    u_int                 NumTransactions;                     /* 4 bytes */
+    u_int                 Overflow;                            /* 4 bytes */
+    u_int                 AckSent;                             /* 4 bytes */
+    u_int                 BadTransaction;                      /* 4 bytes */
+
+    E3_IprocTrapHeader_BE *TrappedTransaction;                 /* 4 bytes */
+    E3_IprocTrapData_BE   *TrappedDataBuffer;                  /* 4 bytes */
+    E3_IprocTrapHeader_BE *WaitForEopTransaction;              /* 4 bytes */
+    E3_IprocTrapData_BE   *WaitForEopDataBuffer;               /* 4 bytes */
+    E3_IprocTrapHeader_BE *DmaIdentifyTransaction;             /* 4 bytes */
+    E3_IprocTrapHeader_BE *ThreadIdentifyTransaction;          /* 4 bytes */
+    E3_Addr               LockQueuePointer;                    /* 4 bytes */
+    E3_Addr               UnlockQueuePointer;                  /* 4 bytes */
+
+    E3_IprocTrapHeader_BE  Transactions[MAX_TRAPPED_TRANS];    /* n * 8 bytes */
+    E3_IprocTrapData_BE           DataBuffers[MAX_TRAPPED_TRANS];      /* n * 64 bytes */
+} INPUT_TRAP;
+
+typedef struct input_fault_save
+{
+    struct input_fault_save *Next;
+    E3_Addr                 Addr;
+    E3_uint32               Count;
+} INPUT_FAULT_SAVE;
+
+#define NUM_INPUT_FAULT_SAVE   32
+#define MIN_INPUT_FAULT_PAGES  8
+#define MAX_INPUT_FAULT_PAGES  128
+
+typedef E3_uint32 EVENT_COOKIE;
+
+#ifdef __KERNEL__
+
+typedef struct event_cookie_entry
+{
+    struct event_cookie_entry *ent_next;
+    struct event_cookie_entry *ent_prev;
+
+    spinlock_t                ent_lock;
+    unsigned                  ent_ref;
+
+    EVENT_COOKIE              ent_cookie;
+    EVENT_COOKIE              ent_fired;
+    kcondvar_t                ent_wait;
+} EVENT_COOKIE_ENTRY;
+
+typedef struct event_cookie_table
+{
+    struct event_cookie_table *tbl_next;
+    struct event_cookie_table *tbl_prev;
+
+    unsigned long              tbl_task;
+    unsigned long              tbl_handle;
+
+    spinlock_t                tbl_lock;
+    unsigned                  tbl_ref;
+    EVENT_COOKIE_ENTRY        *tbl_entries;
+} EVENT_COOKIE_TABLE;
+
+#define NBYTES_PER_SMALL_ROUTE 8
+#define NBYTES_PER_LARGE_ROUTE 16
+
+#define ROUTE_BLOCK_SIZE       ELAN3_PAGE_SIZE
+#define NROUTES_PER_BLOCK      (ROUTE_BLOCK_SIZE/NBYTES_PER_LARGE_ROUTE)
+
+typedef struct elan3_routes
+{
+    struct elan3_routes                *Next;                                  /* Can be chained together */
+
+    sdramaddr_t                         Routes;                                /* sdram offset of route entries */
+    bitmap_t                    Bitmap[BT_BITOUL(NROUTES_PER_BLOCK)];  /* Bitmap of which entries are used */
+} ELAN3_ROUTES; 
+
+
+typedef struct elan3_route_table
+{
+    spinlock_t          Lock;                          /* Route lock */
+    sdramaddr_t                 Table;                         /* Kernel address for route table */
+    u_int               Size;                          /* # entries in route table */
+
+    ELAN3_ROUTES       *LargeRoutes;                   /* Large routes */
+} ELAN3_ROUTE_TABLE;
+
+typedef struct elan3_vpseg
+{
+    struct elan3_vpseg         *Next;
+    int                                 Process;                       /* Virtual process */
+    int                                 Entries;                       /*  and # processes */
+    int                                 Type;                          /* Type of cookie */
+
+    union
+    {
+       
+       ELAN_CAPABILITY Capability;                     /* Capability of remote segment */
+#  define SegCapability                SegUnion.Capability
+       struct {
+           u_short             LowProc;                        /* Base process number */
+           u_short             HighProc;                       /*   and high process number */
+#  define SegLowProc           SegUnion.BROADCAST.LowProc
+#  define SegHighProc          SegUnion.BROADCAST.HighProc
+       } BROADCAST;
+    } SegUnion;
+} ELAN3_VPSEG;
+
+#define ELAN3_VPSEG_UNINT      0                               /* Unitialised */
+#define ELAN3_VPSEG_P2P                1                               /* Point to Point */
+#define ELAN3_VPSEG_BROADCAST  2                               /* Broadcast */
+
+#define NUM_LISTS      7                                       /* Number of "swap" lists */
+
+typedef struct elan3_ctxt
+{
+    struct elan3_ctxt    *Next;                                        /* can be queued on a task */
+    struct elan3_ctxt    *Prev;
+
+    CtxtHandle          Handle;                                /* user handle */
+    int                         RefCnt;                                /* reference count */
+
+    ELAN3MMU           *Elan3mmu;                              /* elan3mmu allocated for Elan translations */
+
+    struct elan3_ops     *Operations;                          /* User supplied helper functions */
+    void               *Private;                               /* Users private pointer */
+
+    int                         Status;                                /* Status (guarded by dev_mutex) */
+    int                         OthersState;                           /* State of halt queueing for dma/thread */
+    int                         LwpCount;                              /* Number of lwp's running */
+
+    ELAN3_DEV          *Device;                                /* Elan device */
+
+    ELAN_CAPABILITY     Capability;                            /* Capability I've attached as */
+    ELAN_POSITION       Position;                              /* Position when I was created */
+    
+    ELAN3_VPSEG                *VpSegs;                                /* List of virtual process segments */
+    ELAN3_ROUTE_TABLE    *RouteTable;
+
+    krwlock_t           VpLock;                                /* Reader/writer lock for vp list */
+    kmutex_t            SwapListsLock;                         /* mutex to lock swap lists */
+    kmutex_t            CmdLock;                               /* mutex to lock trapped dma command */
+    kmutex_t            CmdPortLock;                           /* mutex to load/unload commandport xlation */
+
+    kcondvar_t          Wait;                                  /* Condition variable to sleep on */
+    kcondvar_t          CommandPortWait;                       /* Condition variable to wait for commandport */
+    kcondvar_t          LwpWait;                               /* Condition variable to wait for lwps to stop */
+    kcondvar_t          HaltWait;                              /* Condition variable to wait for halt */
+    int                         Halted;                                /*  and flag for halt cv */
+
+    caddr_t             CommandPageMapping;                    /* user virtual address for command page mapping */
+    ioaddr_t             CommandPage;                          /* Elan command port mapping page */
+    DeviceMappingHandle  CommandPageHandle;                    /* DDI Handle */
+    ioaddr_t            CommandPort;                           /* Elan command port */
+    void               *CommandPortItem;                       /* Item we're re-issuing to commandport */
+
+    ELAN3_FLAGSTATS      *FlagPage;                            /* Page visible to user process */
+
+    COMMAND_TRAP       *CommandTraps;                          /* Command port traps */
+    ELAN3_SPLIT_QUEUE     CommandTrapQ;
+                                                                  
+    CProcTrapBuf_BE    *Commands;                              /* Overflowed commands */
+    ELAN3_QUEUE           CommandQ;
+
+    THREAD_TRAP                *ThreadTraps;                           /* Thread processor traps */
+    ELAN3_QUEUE                 ThreadTrapQ;
+    
+    DMA_TRAP           *DmaTraps;                              /* Dma processor tra[ed */
+    ELAN3_QUEUE                 DmaTrapQ;
+
+    INPUT_TRAP          Input0Trap;                            /* Inputter channel 0 trap */
+    INPUT_TRAP          Input1Trap;                            /* Inputter channel 1 trap */
+    NETERR_RESOLVER    *Input0Resolver;                        /* Inputter channel 0 network error resolver */
+    NETERR_RESOLVER    *Input1Resolver;                        /* Inputter channel 1 network error resolver */
+
+    INPUT_FAULT_SAVE    InputFaults[NUM_INPUT_FAULT_SAVE];     /* stored writeblock addresses */
+    INPUT_FAULT_SAVE    *InputFaultList;                       /* organized in list for LRU */
+    spinlock_t          InputFaultLock;                        /* and lock for list */
+
+    kmutex_t            NetworkErrorLock;
+    NETERR_FIXUP       *NetworkErrorFixups;
+
+    EVENT_COOKIE        *EventCookies;                         /* Event cookies. */
+    ELAN3_QUEUE                 EventCookieQ;
+
+    E3_Addr            *SwapThreads;                           /* Swapped Thread Queue */
+    ELAN3_QUEUE                 SwapThreadQ;
+
+    E3_DMA_BE          *SwapDmas;                              /* Swapped Dmas Queue */
+    ELAN3_QUEUE                 SwapDmaQ;
+
+    int                         ItemCount[NUM_LISTS];                  /* Count of items on each swap list */
+    int                         inhibit;                               /* if set lwp not to reload translations */
+
+    int                  Disabled;
+} ELAN3_CTXT;
+
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::IntrLock, 
+                         elan3_ctxt::Status elan3_ctxt::OthersState
+                         elan3_ctxt::CommandTrapQ elan3_ctxt::CommandQ elan3_ctxt::ThreadTrapQ elan3_ctxt::DmaTrapQ 
+                         elan3_ctxt::Input0Trap elan3_ctxt::Input1Trap elan3_ctxt::EventCookieQ elan3_ctxt::SwapThreadQ 
+                         elan3_ctxt::SwapDmaQ elan3_ctxt::CommandPortItem elan3_ctxt::LwpCount))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_ctxt::SwapListsLock, 
+                         elan3_ctxt::ItemCount))
+_NOTE(RWLOCK_PROTECTS_DATA(elan3_ctxt::VpLock, 
+                          elan3_ctxt::VpSegs elan3_vpseg::Next elan3_vpseg::Process 
+                          elan3_vpseg::Entries elan3_vpseg::Type))
+
+_NOTE(DATA_READABLE_WITHOUT_LOCK(elan3_ctxt::ItemCount elan3_ctxt::Status elan3_ctxt::CommandPortItem))
+
+_NOTE(LOCK_ORDER(elan3_ctxt::SwapListsLock elan3_ctxt::CmdLock elan3_dev::IntrLock))
+_NOTE(LOCK_ORDER(elan3_ctxt::SwapListsLock as::a_lock))                        /* implicit by pagefault */
+
+#define CTXT_DETACHED                          (1 << 0)                /* Context is detached. */
+#define CTXT_NO_LWPS                           (1 << 1)                /* No lwp's to handle faults */
+#define CTXT_EXITING                           (1 << 2)                /* User process is exiting */
+
+#define CTXT_SWAPPING_OUT                      (1 << 3)                /* Context is swapping out */
+#define CTXT_SWAPPED_OUT                       (1 << 4)                /* Context is swapped out */
+
+#define CTXT_SWAP_FREE                         (1 << 5)                /* Swap buffer is free */
+#define CTXT_SWAP_VALID                                (1 << 6)                /* Swap buffer has queue entries in it */
+
+#define CTXT_DMA_QUEUE_FULL                    (1 << 7)                /* Dma trap queue is full */
+#define CTXT_THREAD_QUEUE_FULL                 (1 << 8)                /* Thread trap queue is full */
+#define CTXT_EVENT_QUEUE_FULL                  (1 << 9)                /* Event interrupt queue is full */
+#define CTXT_COMMAND_OVERFLOW_ERROR            (1 << 10)               /* Trap queue overflow */
+
+#define CTXT_SWAP_WANTED                       (1 << 11)               /* Some one wanted to swap */
+#define CTXT_WAITING_SWAPIN                    (1 << 12)               /* Someone waiting on swapin */
+
+#define CTXT_WAITING_COMMAND                   (1 << 13)               /* swgelan waiting on command port */
+#define CTXT_COMMAND_MAPPED_MAIN               (1 << 14)               /* segelan has mapped command port */
+
+#define CTXT_QUEUES_EMPTY                      (1 << 15)               /* dma/thread run queues are empty */
+#define CTXT_QUEUES_EMPTYING                   (1 << 16)               /* dma/thread run queues are being emptied */
+
+#define CTXT_USER_FILTERING                    (1 << 17)               /* user requested context filter */
+
+#define CTXT_KERNEL                            (1 << 18)               /* context is a kernel context */
+#define CTXT_COMMAND_MAPPED_ELAN               (1 << 19)               /* command port is mapped for elan */
+#define CTXT_FIXUP_NETERR                      (1 << 20)               /* fixing up a network error */
+
+
+#define CTXT_SWAPPED_REASONS           (CTXT_NO_LWPS   |               \
+                                        CTXT_DETACHED  |               \
+                                        CTXT_EXITING   |               \
+                                        CTXT_FIXUP_NETERR)
+
+#define CTXT_OTHERS_REASONS            (CTXT_EVENT_QUEUE_FULL  |       \
+                                        CTXT_DMA_QUEUE_FULL    |       \
+                                        CTXT_THREAD_QUEUE_FULL |       \
+                                        CTXT_COMMAND_OVERFLOW_ERROR |  \
+                                        CTXT_SWAPPED_REASONS)
+
+#define CTXT_INPUTTER_REASONS          (CTXT_USER_FILTERING |          \
+                                        CTXT_OTHERS_REASONS)
+
+#define CTXT_COMMAND_MAPPED            (CTXT_COMMAND_MAPPED_MAIN |     \
+                                        CTXT_COMMAND_MAPPED_ELAN)
+
+#define CTXT_IS_KERNEL(ctxt)           ((ctxt)->Status & CTXT_KERNEL)
+
+/*
+ * State values for ctxt_inputterState/ctxt_commandportStats
+ */
+#define CTXT_STATE_OK                  0
+#define CTXT_STATE_TRAPPED             1               /* Inputter channel 0 trapped */
+#define CTXT_STATE_RESOLVING           2               /* An LWP is resolving the trap */
+#define CTXT_STATE_NEEDS_RESTART       3               /* Th trapped packet needs to be executed */
+#define CTXT_STATE_NETWORK_ERROR       4               /* We're waiting on an RPC for the identify transaction */
+#define CTXT_STATE_EXECUTING           5               /* An LWP is executing the trapped packet */
+
+/*
+ * State values for OthersState.
+ */
+#define CTXT_OTHERS_RUNNING            0
+#define CTXT_OTHERS_HALTING            1
+#define CTXT_OTHERS_SWAPPING           2
+#define CTXT_OTHERS_HALTING_MORE       3
+#define CTXT_OTHERS_SWAPPING_MORE      4
+#define CTXT_OTHERS_SWAPPED            5
+
+typedef struct elan3_ops
+{
+    u_int  Version;
+
+    int         (*Exception)   (ELAN3_CTXT *ctxt, int type, int proc, void *trap, va_list ap);
+
+    /* swap item list functions */
+    int  (*GetWordItem)                (ELAN3_CTXT *ctxt, int list, void **itemp, E3_uint32 *valuep);
+    int  (*GetBlockItem)       (ELAN3_CTXT *ctxt, int list, void **itemp, E3_Addr *valuep);
+    void (*PutWordItem)                (ELAN3_CTXT *ctxt, int list, E3_Addr value);
+    void (*PutBlockItem)       (ELAN3_CTXT *ctxt, int list, E3_uint32 *ptr);
+    void (*PutbackItem)                (ELAN3_CTXT *ctxt, int list, void *item);
+    void (*FreeWordItem)       (ELAN3_CTXT *ctxt, void *item);
+    void (*FreeBlockItem)      (ELAN3_CTXT *ctxt, void *item);
+    int  (*CountItems)         (ELAN3_CTXT *ctxt, int list);
+
+    /* event interrupt cookie */
+    int  (*Event)              (ELAN3_CTXT *ctxt, E3_uint32 cookie, int flag);
+
+    /* swapin/swapout functions. */
+    void (*Swapin)             (ELAN3_CTXT *ctxt);
+    void (*Swapout)            (ELAN3_CTXT *ctxt);
+
+    /* Free of private data */
+    void (*FreePrivate)                (ELAN3_CTXT *ctxt);
+
+    /* Fixup a network error */
+    int  (*FixupNetworkError)  (ELAN3_CTXT *ctxt, NETERR_FIXUP *nef);
+
+    /* Interrupt handler trap interface */
+    int  (*DProcTrap)          (ELAN3_CTXT *ctxt, DMA_TRAP *trap);
+    int  (*TProcTrap)          (ELAN3_CTXT *ctxt, THREAD_TRAP *trap);
+    int         (*IProcTrap)           (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, int chan);
+    int         (*CProcTrap)           (ELAN3_CTXT *ctxt, COMMAND_TRAP *trap);
+    int  (*CProcReissue)        (ELAN3_CTXT *ctxt, CProcTrapBuf_BE *TrapBuf);
+
+    /* User memory access functions */
+    int              (*StartFaultCheck)(ELAN3_CTXT *ctxt);
+    void      (*EndFaultCheck)  (ELAN3_CTXT *ctxt);
+
+    E3_uint8  (*Load8)         (ELAN3_CTXT *ctxt, E3_Addr addr);
+    void      (*Store8)                (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint8 val);
+    E3_uint16 (*Load16)                (ELAN3_CTXT *ctxt, E3_Addr addr);
+    void      (*Store16)       (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint16 val);
+    E3_uint32 (*Load32)                (ELAN3_CTXT *ctxt, E3_Addr addr);
+    void      (*Store32)       (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint32 val);
+    E3_uint64 (*Load64)                (ELAN3_CTXT *ctxt, E3_Addr addr);
+    void      (*Store64)       (ELAN3_CTXT *ctxt, E3_Addr addr, E3_uint64 val);
+
+} ELAN3_OPS;
+
+#define ELAN3_OPS_VERSION      0xdeef0001
+
+/*
+ * Flags for ops_event.
+ */
+#define OP_INTR                        0               /* Called from interrupt handler */
+#define OP_LWP                 1               /* Called from "lwp" */
+
+/*
+ * Return codes for "ops" functions.
+ */
+#define OP_DEFER               0               /* Defer to next lower interrupt */
+#define OP_IGNORE              1               /* No event hander, so ignore it */
+#define OP_HANDLED             2               /* Handled event (resume thread) */
+#define OP_FAILED              3               /* Failed */
+
+#define ELAN3_CALL_OP(ctxt,fn)                         ((ctxt)->Operations && (ctxt)->Operations->fn) ? (ctxt)->Operations->fn 
+
+#define ELAN3_OP_EXCEPTION(ctxt,type,proc,trap,ap)     (ELAN3_CALL_OP(ctxt,Exception)    (ctxt,type,proc,trap,ap)      : OP_IGNORE)
+#define ELAN3_OP_GET_WORD_ITEM(ctxt,list,itemp,valuep) (ELAN3_CALL_OP(ctxt,GetWordItem)  (ctxt,list,itemp,valuep)      : 0)
+#define ELAN3_OP_GET_BLOCK_ITEM(ctxt,list,itemp,valuep)        (ELAN3_CALL_OP(ctxt,GetBlockItem) (ctxt,list,itemp,valuep)      : 0)
+#define ELAN3_OP_PUT_WORD_ITEM(ctxt,list,value)                (ELAN3_CALL_OP(ctxt,PutWordItem)  (ctxt,list,value)             : (void)0)
+#define ELAN3_OP_PUT_BLOCK_ITEM(ctxt,list,ptr)         (ELAN3_CALL_OP(ctxt,PutBlockItem) (ctxt,list,ptr)               : (void)0)
+#define ELAN3_OP_PUTBACK_ITEM(ctxt,list,item)          (ELAN3_CALL_OP(ctxt,PutbackItem)  (ctxt,list,item)              : (void)0)
+#define ELAN3_OP_FREE_WORD_ITEM(ctxt,item)             (ELAN3_CALL_OP(ctxt,FreeWordItem) (ctxt,item)                   : (void)0)
+#define ELAN3_OP_FREE_BLOCK_ITEM(ctxt,item)            (ELAN3_CALL_OP(ctxt,FreeBlockItem)(ctxt,item)                   : (void)0)
+#define ELAN3_OP_COUNT_ITEMS(ctxt,list)                        (ELAN3_CALL_OP(ctxt,CountItems)(ctxt,list)                      : 0)
+#define ELAN3_OP_EVENT(ctxt,cookie,flag)               (ELAN3_CALL_OP(ctxt,Event)(ctxt,cookie,flag)                    : OP_IGNORE)
+#define ELAN3_OP_SWAPIN(ctxt)                          (ELAN3_CALL_OP(ctxt,Swapin)(ctxt)                               : (void)0)
+#define ELAN3_OP_SWAPOUT(ctxt)                         (ELAN3_CALL_OP(ctxt,Swapout)(ctxt)                              : (void)0)
+#define ELAN3_OP_FREE_PRIVATE(ctxt)                    (ELAN3_CALL_OP(ctxt,FreePrivate)(ctxt)                          : (void)0)
+#define ELAN3_OP_FIXUP_NETWORK_ERROR(ctxt, nef)                (ELAN3_CALL_OP(ctxt,FixupNetworkError)(ctxt,nef)                        : OP_FAILED)
+
+#define ELAN3_OP_DPROC_TRAP(ctxt, trap)                        (ELAN3_CALL_OP(ctxt,DProcTrap)(ctxt,trap)                       : OP_DEFER)
+#define ELAN3_OP_TPROC_TRAP(ctxt, trap)                        (ELAN3_CALL_OP(ctxt,TProcTrap)(ctxt,trap)                       : OP_DEFER)
+#define ELAN3_OP_IPROC_TRAP(ctxt, trap, chan)          (ELAN3_CALL_OP(ctxt,IProcTrap)(ctxt,trap,chan)                  : OP_DEFER)
+#define ELAN3_OP_CPROC_TRAP(ctxt, trap)                        (ELAN3_CALL_OP(ctxt,CProcTrap)(ctxt,trap)                       : OP_DEFER)
+#define ELAN3_OP_CPROC_REISSUE(ctxt,tbuf)              (ELAN3_CALL_OP(ctxt,CProcReissue)(ctxt, tbuf)                   : OP_DEFER)
+
+#define ELAN3_OP_START_FAULT_CHECK(ctxt)               (ELAN3_CALL_OP(ctxt,StartFaultCheck)(ctxt)                      : 0)
+#define ELAN3_OP_END_FAULT_CHECK(ctxt)                 (ELAN3_CALL_OP(ctxt,EndFaultCheck)(ctxt)                                : (void)0)
+#define ELAN3_OP_LOAD8(ctxt,addr)                      (ELAN3_CALL_OP(ctxt,Load8)(ctxt,addr)                           : 0)
+#define ELAN3_OP_STORE8(ctxt,addr,val)                 (ELAN3_CALL_OP(ctxt,Store8)(ctxt,addr,val)                      : (void)0)
+#define ELAN3_OP_LOAD16(ctxt,addr)                     (ELAN3_CALL_OP(ctxt,Load16)(ctxt,addr)                          : 0)
+#define ELAN3_OP_STORE16(ctxt,addr,val)                        (ELAN3_CALL_OP(ctxt,Store16)(ctxt,addr,val)                     : (void)0)
+#define ELAN3_OP_LOAD32(ctxt,addr)                     (ELAN3_CALL_OP(ctxt,Load32)(ctxt,addr)                          : 0)
+#define ELAN3_OP_STORE32(ctxt,addr,val)                        (ELAN3_CALL_OP(ctxt,Store32)(ctxt,addr,val)                     : (void)0)
+#define ELAN3_OP_LOAD64(ctxt,addr)                     (ELAN3_CALL_OP(ctxt,Load64)(ctxt,addr)                          : 0)
+#define ELAN3_OP_STORE64(ctxt,addr,val)                        (ELAN3_CALL_OP(ctxt,Store64)(ctxt,addr,val)                     : (void)0)
+
+#endif /* __KERNEL__ */
+
+/* "list" arguement to ops functions */
+#define LIST_DMA_PTR           0
+#define LIST_DMA_DESC          1
+#define LIST_THREAD                    2
+#define LIST_COMMAND           3
+#define LIST_SETEVENT          4
+#define LIST_FREE_WORD         5
+#define LIST_FREE_BLOCK                6
+
+#define MAX_LISTS              7
+
+#if defined(__KERNEL__) && MAX_LISTS != NUM_LISTS
+#  error Check NUM_LISTS == MAX_LISTS
+#endif
+
+/*
+ * Values for the 'type' field to PostException().
+ */
+#define EXCEPTION_INVALID_ADDR         1               /* FaultArea, res */
+#define EXCEPTION_UNIMP_INSTR          2               /* instr */
+#define EXCEPTION_INVALID_PROCESS      3               /* proc, res */
+#define EXCEPTION_SIMULATION_FAILED    4               /* */
+#define EXCEPTION_UNIMPLEMENTED                5               /* */
+#define EXCEPTION_SWAP_FAULT           6               /* */
+#define EXCEPTION_SWAP_FAILED          7               /* */
+#define EXCEPTION_BAD_PACKET           8               /* */
+#define EXCEPTION_FAULTED              9               /* addr */
+#define EXCEPTION_QUEUE_OVERFLOW       10              /* FaultArea, TrapType */
+#define EXCEPTION_COMMAND_OVERFLOW     11              /* count */
+#define EXCEPTION_DMA_RETRY_FAIL       12              /* */
+#define EXCEPTION_CHAINED_EVENT                13              /* EventAddr */
+#define EXCEPTION_THREAD_KILLED                14              /* */
+#define EXCEPTION_CANNOT_SAVE_THREAD   15
+#define EXCEPTION_BAD_SYSCALL          16              /* */
+#define EXCEPTION_DEBUG                        17
+#define EXCEPTION_BAD_EVENT            18              /* */
+#define EXCEPTION_NETWORK_ERROR                19              /* rvp */
+#define EXCEPTION_BUS_ERROR            20
+#define EXCEPTION_COOKIE_ERROR         21
+#define EXCEPTION_PACKET_TIMEOUT       22
+#define EXCEPTION_BAD_DMA              23              /* */
+#define EXCEPTION_ENOMEM               24
+
+/*
+ * Values for the 'proc' field to ElanException().
+ */
+#define COMMAND_PROC                   1
+#define THREAD_PROC                    2
+#define DMA_PROC                       3
+#define INPUT_PROC                     4
+#define EVENT_PROC                     5
+
+/* Flags to IssueDmaCommand */
+#define ISSUE_COMMAND_FOR_CPROC                1
+#define ISSUE_COMMAND_CANT_WAIT                2
+
+/* Return code from IssueDmaCommand.*/
+#define ISSUE_COMMAND_OK               0
+#define ISSUE_COMMAND_TRAPPED          1
+#define ISSUE_COMMAND_RETRY            2
+#define ISSUE_COMMAND_WAIT             3
+
+#ifdef __KERNEL__
+
+extern ELAN3_CTXT *elan3_alloc(ELAN3_DEV *dev, int kernel);
+extern void       elan3_free      (ELAN3_CTXT *ctxt);
+
+extern int        elan3_attach    (ELAN3_CTXT *ctxt, ELAN_CAPABILITY *cap);
+extern int         elan3_doattach  (ELAN3_CTXT *ctxt, ELAN_CAPABILITY *cap);
+extern void       elan3_detach    (ELAN3_CTXT *ctxt);
+extern void        elan3_dodetach  (ELAN3_CTXT *ctxt);
+
+extern int        elan3_addvp     (ELAN3_CTXT *ctxt, int process, ELAN_CAPABILITY *cap);
+extern int        elan3_removevp  (ELAN3_CTXT *ctxt, int process);
+extern int        elan3_addbcastvp(ELAN3_CTXT *ctxt, int process, int base, int count);
+
+extern int         elan3_process   (ELAN3_CTXT *ctxt);
+
+extern int        elan3_load_route (ELAN3_CTXT *ctxt, int process, E3_uint16 *flits);
+extern int        elan3_check_route(ELAN3_CTXT *ctxt, int process, E3_uint16 *flits, E3_uint32 *routeError);
+
+extern int        elan3_lwp       (ELAN3_CTXT *ctxt);
+
+extern void       elan3_swapin (ELAN3_CTXT *ctxt, int reason);
+extern void       elan3_swapout (ELAN3_CTXT *ctxt, int reason);
+extern int         elan3_pagefault (ELAN3_CTXT *ctxt, E3_FaultSave_BE *FaultSave, int npages);
+extern void        elan3_block_inputter (ELAN3_CTXT *ctxt, int block);
+
+
+extern E3_Addr     elan3_init_thread (ELAN3_DEV *dev, E3_Addr fn, E3_Addr addr, sdramaddr_t stack, int stackSize, int nargs, ...);
+
+extern void       SetInputterState (ELAN3_CTXT *ctxt, E3_uint32 Pend, E3_uint32 *Maskp);
+extern void       SetInputterStateForContext (ELAN3_CTXT *ctxt, E3_uint32 Pend, E3_uint32 *Maskp);
+extern void        UnloadCommandPageMapping (ELAN3_CTXT *ctxt);
+extern void       StartSwapoutContext (ELAN3_CTXT *ctxt, E3_uint32 Pend, E3_uint32 *Maskp);
+
+extern int        HandleExceptions (ELAN3_CTXT *ctxt, unsigned long *flags);
+extern int        RestartContext (ELAN3_CTXT *ctxt, unsigned long *flags);
+extern int         CheckCommandQueueFlushed (ELAN3_CTXT *ctxt, E3_uint32 cflags, int how, unsigned long *flags);
+extern int        IssueCommand (ELAN3_CTXT *ctxt, unsigned cmdoff, E3_Addr value, int flags);
+extern int        IssueDmaCommand (ELAN3_CTXT *ctxt, E3_Addr value, void *item, int flags);
+extern int         WaitForDmaCommand (ELAN3_CTXT *ctxt, void *item, int flags);
+extern void       FixupEventTrap (ELAN3_CTXT *ctxt, int proc, void *trap, E3_uint32 TrapType, 
+                                  E3_FaultSave_BE *FaultSaveArea, int flags);
+extern int        SimulateBlockCopy (ELAN3_CTXT *ctxt, E3_Addr EventAddress);
+extern void       ReissueEvent (ELAN3_CTXT *ctxt, E3_Addr addr,int flags);
+extern int         SetEventsNeedRestart (ELAN3_CTXT *ctxt);
+extern void        RestartSetEvents (ELAN3_CTXT *ctxt);
+extern int        RunEventType (ELAN3_CTXT *ctxt, E3_FaultSave_BE *FaultSaveArea, E3_uint32 EventType);
+extern void        WakeupLwp (ELAN3_DEV *dev, void *arg);
+extern void       QueueEventInterrupt (ELAN3_CTXT *ctxt, E3_uint32 cookie);
+extern int         WaitForCommandPort (ELAN3_CTXT *ctxt);
+
+extern int        ElanException (ELAN3_CTXT *ctxt, int type, int proc, void *trap, ...);
+
+/* context_osdep.c */
+extern int        LoadElanTranslation (ELAN3_CTXT *ctxt, E3_Addr elanAddr, int len, int protFault, int writeable);
+extern void       LoadCommandPortTranslation (ELAN3_CTXT *ctxt);
+
+#if defined(DIGITAL_UNIX)
+/* seg_elan.c */
+extern caddr_t    elan3_segelan3_create (ELAN3_CTXT *ctxt);
+extern void       elan3_segelan3_destroy (ELAN3_CTXT *ctxt);
+extern int         elan3_segelan3_map (ELAN3_CTXT *ctxt);
+extern void        elan3_segelan3_unmap (ELAN3_CTXT *ctxt);
+
+/* seg_elanmem.c */
+extern int        elan3_segelanmem_create (ELAN3_DEV *dev, unsigned object, unsigned off, vm_offset_t *addrp, int len);
+#endif /* defined(DIGITAL_UNIX) */
+
+/* route_table.c */
+extern ELAN3_ROUTE_TABLE *AllocateRouteTable (ELAN3_DEV *dev, int size);
+extern void              FreeRouteTable  (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl);
+extern int               LoadRoute       (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int vp, int ctxnum, int nflits, E3_uint16 *flits);
+extern int               GetRoute        (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int process, E3_uint16 *flits);
+extern void             InvalidateRoute (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int vp);
+extern void             ValidateRoute   (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int vp);
+extern void             ClearRoute      (ELAN3_DEV *dev, ELAN3_ROUTE_TABLE *tbl, int vp);
+
+extern int               GenerateRoute (ELAN_POSITION *pos, E3_uint16 *flits, int lowid, int highid, int timeout, int highPri);
+extern int               GenerateProbeRoute (E3_uint16 *flits, int nodeid, int level, int *linkup, int *linkdown, int adaptive);
+extern int               GenerateCheckRoute (ELAN_POSITION *pos, E3_uint16 *flits, int level, int adaptive);
+
+/* virtual_process.c */
+extern ELAN_LOCATION  ProcessToLocation     (ELAN3_CTXT *ctxt, ELAN3_VPSEG *seg, int process, ELAN_CAPABILITY *cap);
+extern int           ResolveVirtualProcess (ELAN3_CTXT *ctxt, int process);
+extern caddr_t        CapabilityString      (ELAN_CAPABILITY *cap);
+extern void           UnloadVirtualProcess  (ELAN3_CTXT *ctxt, ELAN_CAPABILITY *cap);
+
+extern int           elan3_get_route   (ELAN3_CTXT *ctxt, int process, E3_uint16 *flits);
+extern int           elan3_reset_route (ELAN3_CTXT *ctxt, int process);
+
+/* cproc.c */
+extern int       NextCProcTrap (ELAN3_CTXT *ctxt, COMMAND_TRAP *trap);
+extern void      ResolveCProcTrap (ELAN3_CTXT *ctxt);
+extern int       RestartCProcTrap (ELAN3_CTXT *ctxt);
+
+/* iproc.c */
+extern void       InspectIProcTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap);
+extern void      ResolveIProcTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, NETERR_RESOLVER **rvp);
+extern int       RestartIProcTrap (ELAN3_CTXT *ctxt, INPUT_TRAP *trap);
+extern char      *IProcTrapString (E3_IprocTrapHeader_BE *hdrp, E3_IprocTrapData *datap);
+extern void       SimulateUnlockQueue (ELAN3_CTXT *ctxt, E3_Addr QueuePointer, int SentAck);
+
+/* tproc.c */
+extern int       NextTProcTrap (ELAN3_CTXT *ctxt, THREAD_TRAP *trap);
+extern void      ResolveTProcTrap (ELAN3_CTXT *ctxt, THREAD_TRAP *trap);
+extern int       TProcNeedsRestart (ELAN3_CTXT *ctxt);
+extern void      RestartTProcItems (ELAN3_CTXT *ctxt);
+extern E3_Addr    SaveThreadToStack (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, int SkipInstruction);
+extern void       ReissueStackPointer (ELAN3_CTXT *ctxt, E3_Addr StackPointer);
+
+/* tprocinsts.c */
+extern int        RollThreadToClose (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, E3_uint32 PAckVal);
+
+/* tproc_osdep.c */
+extern int        ThreadSyscall (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, int *skip);
+extern int       ThreadElancall (ELAN3_CTXT *ctxt, THREAD_TRAP *trap, int *skip);
+
+/* dproc.c */
+extern int       NextDProcTrap (ELAN3_CTXT *ctxt, DMA_TRAP *trap);
+extern void      ResolveDProcTrap (ELAN3_CTXT *ctxt, DMA_TRAP *trap);
+extern int       DProcNeedsRestart (ELAN3_CTXT *ctxt);
+extern void      RestartDProcItems (ELAN3_CTXT *ctxt);
+extern void       RestartDmaDesc (ELAN3_CTXT *ctxt, E3_DMA_BE *desc);
+extern void       RestartDmaTrap (ELAN3_CTXT *ctxt, DMA_TRAP *trap);
+extern void      RestartDmaPtr (ELAN3_CTXT *ctxt, E3_Addr ptr);
+
+/* network_error.c */
+extern void       InitialiseNetworkErrorResolver (void);
+extern void       FinaliseNetworkErrorResolver (void);
+extern int        QueueNetworkErrorResolver (ELAN3_CTXT *ctxt, INPUT_TRAP *trap, NETERR_RESOLVER **rvpp);
+extern void      FreeNetworkErrorResolver (NETERR_RESOLVER *rvp);
+extern void       CancelNetworkErrorResolver (NETERR_RESOLVER *rvp);
+extern int       ExecuteNetworkErrorFixup (NETERR_MSG *msg);
+extern void      CompleteNetworkErrorFixup (ELAN3_CTXT *ctxt, NETERR_FIXUP *nef, int status);
+
+extern int        AddNeterrServerSyscall (int elanId, void *configp, void *addrp, char *namep);
+
+/* eventcookie.c */
+extern void                cookie_init(void);
+extern void                cookie_fini(void);
+extern EVENT_COOKIE_TABLE *cookie_alloc_table (unsigned long task, unsigned long handle);
+extern void                cookie_free_table (EVENT_COOKIE_TABLE *tbl);
+extern int                 cookie_alloc_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie);
+extern int                 cookie_free_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie);
+extern int                 cookie_fire_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie);
+extern int                 cookie_wait_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie);
+extern int                 cookie_arm_cookie (EVENT_COOKIE_TABLE *tbl, EVENT_COOKIE cookie);
+
+/* routecheck.c */
+extern int elan3_route_check          (ELAN3_CTXT *ctxt, E3_uint16 *flits, int destNode);
+extern int elan3_route_broadcast_check(ELAN3_CTXT *ctxt, E3_uint16 *flitsA, int lowNode, int highNode);
+
+
+#endif /* __KERNEL__ */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _ELAN3_ELANCTXT_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elandebug.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elandebug.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elandebug.h      2005-06-01 23:12:54.720420712 -0400
@@ -0,0 +1,106 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_ELANDEBUG_H
+#define _ELAN3_ELANDEBUG_H
+
+#ident "$Id: elandebug.h,v 1.38 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elandebug.h,v $ */
+
+#if defined(__KERNEL__)
+
+extern u_int elan3_debug;
+extern u_int elan3_debug_console;
+extern u_int elan3_debug_buffer;
+extern u_int elan3_debug_ignore_dev;
+extern u_int elan3_debug_ignore_kcomm;
+extern u_int elan3_debug_ignore_ctxt;
+extern u_int elan3_debug_display_ctxt;
+
+#define DBG_CONFIG     0x00000001                      /* Module configuration */
+#define DBG_HAT                0x00000002
+#define DBG_FN         0x00000004
+#define DBG_SEG                0x00000008
+#define DBG_INTR       0x00000010
+#define DBG_LWP                0x00000020
+#define DBG_FAULT      0x00000040
+#define DBG_EVENT      0x00000080
+#define DBG_CPROC      0x00000100
+#define DBG_TPROC      0x00000200
+#define DBG_DPROC      0x00000400
+#define DBG_IPROC      0x00000800
+#define DBG_SWAP       0x00001000
+#define DBG_CMD                0x00002000
+#define DBG_VP         0x00004000
+#define DBG_SYSCALL    0x00008000
+#define DBG_BSCAN      0x00010000
+#define DBG_LINKERR    0x00020000
+#define DBG_NETERR     0x00040000
+#define DBG_NETRPC     0x00080000
+#define DBG_EVENTCOOKIE 0x00100000
+#define DBG_SDRAM      0x00200000
+
+#define DBG_EP         0x10000000
+#define DBG_EPCONSOLE  0x20000000
+
+#define DBG_EIP                0x40000000
+#define DBG_EIPFAIL    0x80000000
+
+#define DBG_ALL                0xffffffff
+
+/* values to pass as "ctxt" rather than a "ctxt" pointer */
+#define DBG_DEVICE     ((void *) 0)
+#define DBG_KCOMM      ((void *) 1)
+#define DBG_ICS                ((void *) 2)
+#define DBG_USER       ((void *) 3)
+#define DBG_NTYPES     64
+
+#if defined(DEBUG_PRINTF)
+#  define DBG(m,fn)                            ((elan3_debug&(m)) ? (void)(fn) : (void)0)
+#  define PRINTF0(ctxt,m,fmt)                  ((elan3_debug&(m)) ? elan3_debugf(ctxt,m,fmt)             : (void)0)
+#  define PRINTF1(ctxt,m,fmt,a)                        ((elan3_debug&(m)) ? elan3_debugf(ctxt,m,fmt,a)           : (void)0)
+#  define PRINTF2(ctxt,m,fmt,a,b)              ((elan3_debug&(m)) ? elan3_debugf(ctxt,m,fmt,a,b)         : (void)0)
+#  define PRINTF3(ctxt,m,fmt,a,b,c)            ((elan3_debug&(m)) ? elan3_debugf(ctxt,m,fmt,a,b,c)       : (void)0)
+#  define PRINTF4(ctxt,m,fmt,a,b,c,d)          ((elan3_debug&(m)) ? elan3_debugf(ctxt,m,fmt,a,b,c,d)     : (void)0)
+#  define PRINTF5(ctxt,m,fmt,a,b,c,d,e)                ((elan3_debug&(m)) ? elan3_debugf(ctxt,m,fmt,a,b,c,d,e)   : (void)0)
+#  define PRINTF6(ctxt,m,fmt,a,b,c,d,e,f)      ((elan3_debug&(m)) ? elan3_debugf(ctxt,m,fmt,a,b,c,d,e,f) : (void)0)
+#ifdef __GNUC__
+#  define PRINTF(ctxt,m,args...)               ((elan3_debug&(m)) ? elan3_debugf(ctxt,m, ##args)         : (void)0)
+#endif
+
+#else
+
+#  define DBG(m, fn)                           do { ; } while (0)
+#  define PRINTF0(ctxt,m,fmt)                  do { ; } while (0)
+#  define PRINTF1(ctxt,m,fmt,a)                        do { ; } while (0)
+#  define PRINTF2(ctxt,m,fmt,a,b)              do { ; } while (0)
+#  define PRINTF3(ctxt,m,fmt,a,b,c)            do { ; } while (0)
+#  define PRINTF4(ctxt,m,fmt,a,b,c,d)          do { ; } while (0)
+#  define PRINTF5(ctxt,m,fmt,a,b,c,d,e)                do { ; } while (0)
+#  define PRINTF6(ctxt,m,fmt,a,b,c,d,e,f)      do { ; } while (0)
+#ifdef __GNUC__
+#  define PRINTF(ctxt,m,args...)               do { ; } while (0)
+#endif
+
+#endif /* DEBUG_PRINTF */
+
+#ifdef __GNUC__
+extern void       elan3_debugf (void *ctxt, unsigned int mode, char *fmt, ...)
+    __attribute__ ((format (printf,3,4)));
+#else
+extern void       elan3_debugf (void *ctxt, unsigned int mode, char *fmt, ...);
+#endif
+
+
+#endif /* __KERNEL__ */
+#endif /* _ELAN3_ELANDEBUG_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elandev.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elandev.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elandev.h        2005-06-01 23:12:54.721420560 -0400
@@ -0,0 +1,581 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_ELANDEV_H
+#define __ELAN3_ELANDEV_H
+
+#ident "$Id: elandev.h,v 1.74.2.2 2004/12/10 11:10:19 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elandev.h,v $ */
+
+#include <elan/bitmap.h>
+#include <elan/devinfo.h>
+#include <elan/stats.h>
+
+#if defined(DIGITAL_UNIX)
+#  include <elan3/elandev_dunix.h>
+#elif defined(LINUX)
+#  include <elan3/elandev_linux.h>
+#elif defined(SOLARIS)
+#  include <elan3/elandev_solaris.h>
+#endif
+
+#ifndef TRUE
+#  define TRUE 1
+#endif
+#ifndef FALSE
+#  define FALSE 0
+#endif
+
+/*
+ * Elan base address registers defined as follows :
+ */
+#define ELAN3_BAR_SDRAM                0
+#define ELAN3_BAR_COMMAND_PORT 1
+#define ELAN3_BAR_REGISTERS    2
+#define ELAN3_BAR_EBUS         3
+
+/* Macro to generate 'offset' to mmap "mem" device */
+#define OFF_TO_SPACE(off)      ((off) >> 28)
+#define OFF_TO_OFFSET(off)     ((off) & 0x0FFFFFFF)
+#define GEN_OFF(space,off)     (((space) << 28) | ((off) & 0x0FFFFFFF))
+
+#ifdef __KERNEL__
+
+/*
+ * Elan EBUS is configured as follows :
+ */
+#define ELAN3_EBUS_ROM_OFFSET          0x000000                /* rom */
+#define ELAN3_EBUS_INTPAL_OFFSET       0x180000                /* interrupt pal (write only) */
+
+#define ELAN3_EBUS_ROM_SIZE            0x100000
+
+/*
+ * Elan SDRAM is arranged as follows :
+ */
+#define ELAN3_TANDQ_SIZE               0x0020000               /* Trap And Queue Size */
+#define ELAN3_CONTEXT_SIZE             0x0010000               /* Context Table Size */
+#define ELAN3_COMMAND_TRAP_SIZE                0x0010000               /* Command Port Trap Size */
+
+#ifdef MPSAS
+#define ELAN3_LN2_NUM_CONTEXTS 8                               /* Support 256 contexts */
+#else
+#define ELAN3_LN2_NUM_CONTEXTS 12                              /* Support 4096 contexts */
+#endif
+#define ELAN3_NUM_CONTEXTS     (1 << ELAN3_LN2_NUM_CONTEXTS)   /* Entries in context table */
+
+#define ELAN3_SDRAM_NUM_BANKS  4                               /* Elan supports 4 Banks of Sdram */
+#define ELAN3_SDRAM_BANK_SHIFT 26                              /* each of which can be 64 mbytes ? */
+#define ELAN3_SDRAM_BANK_SIZE  (1 << ELAN3_SDRAM_BANK_SHIFT)
+
+#define ELAN3_MAX_CACHE_SIZE   (64 * 1024)                     /* Maximum cache size */
+#define ELAN3_CACHE_SIZE       (64 * 4 * E3_CACHELINE_SIZE)    /* Elan3 has 8K cache */
+
+#ifndef offsetof
+#define offsetof(s, m)         (size_t)(&(((s *)0)->m))
+#endif
+
+/*
+ * circular queue and macros to access members.
+ */
+typedef struct
+{
+    u_int      q_back;                 /* Next free space */
+    u_int      q_front;                /* First object to remove */
+    u_int      q_size;                 /* Size of queue */
+    u_int      q_count;                /* Current number of entries */
+    u_int      q_slop;                 /* FULL <=> (count+slop) == size */
+} ELAN3_QUEUE;
+
+typedef struct 
+{
+    u_int      q_back;                 /* Next free space */
+    u_int      q_middle;               /* Middle pointer */
+    u_int      q_front;                /* First object to remove */
+    u_int      q_size;                 /* Size of queue */
+    u_int      q_count;                /* Current number of entries */
+    u_int      q_slop;                 /* FULL <=> (count+slop) == size */
+} ELAN3_SPLIT_QUEUE;
+
+#define ELAN3_QUEUE_INIT(q,num,slop)   ((q).q_size = (num), (q).q_slop = (slop)+1, (q).q_front = (q).q_back = 0, (q).q_count = 0)
+#define ELAN3_QUEUE_FULL(q)            ((q).q_count == ((q).q_size - (q).q_slop))
+#define ELAN3_QUEUE_REALLY_FULL(q)     ((q).q_count == (q).q_size - 1)
+#define ELAN3_QUEUE_EMPTY(q)           ((q).q_count == 0)
+#define ELAN3_QUEUE_FRONT_EMPTY(q)     ((q).q_front == (q).q_middle)
+#define ELAN3_QUEUE_BACK_EMPTY(q)      ((q).q_middle == (q).q_back)
+#define ELAN3_QUEUE_ADD(q)             ((q).q_back = ((q).q_back+1) % (q).q_size, (q).q_count++)
+#define ELAN3_QUEUE_REMOVE(q)          ((q).q_front = ((q).q_front+1) % (q).q_size, (q).q_count--)
+#define ELAN3_QUEUE_ADD_FRONT(q)               ((q).q_front = ((q).q_front-1) % (q).q_size, (q).q_count++)
+#define ELAN3_QUEUE_CONSUME(q)         ((q).q_middle = ((q).q_middle+1) % (q).q_size)
+#define ELAN3_QUEUE_FRONT(q,qArea)     (&(qArea)[(q).q_front])
+#define ELAN3_QUEUE_MIDDLE(q,qArea)    (&(qArea)[(q).q_middle])
+#define ELAN3_QUEUE_BACK(q,qArea)      (&(qArea)[(q).q_back])
+
+#define SDRAM_MIN_BLOCK_SHIFT  10
+#define SDRAM_NUM_FREE_LISTS   17                              /* allows max 64Mb block */
+#define SDRAM_MIN_BLOCK_SIZE   (1 << SDRAM_MIN_BLOCK_SHIFT)
+#define SDRAM_MAX_BLOCK_SIZE   (SDRAM_MIN_BLOCK_SIZE << (SDRAM_NUM_FREE_LISTS-1))
+#define SDRAM_FREELIST_TRIGGER 32
+
+typedef struct elan3_sdram_bank
+{
+    u_int              Size;                                   /* Size of bank of memory */
+
+    ioaddr_t           Mapping;                                /* Where mapped in the kernel */
+    DeviceMappingHandle Handle;                                        /* and mapping handle */
+
+    struct elan3_ptbl_gr **PtblGroups;
+    
+    bitmap_t          *Bitmaps[SDRAM_NUM_FREE_LISTS];
+} ELAN3_SDRAM_BANK;
+
+typedef struct elan3_haltop
+{
+    struct elan3_haltop          *Next;                                /* Chain to next in list. */
+    E3_uint32             Mask;                                /* Interrupt mask to see before calling function */
+    
+    void                (*Function)(void *, void *);           /* Function to call */
+    void                 *Arguement;                           /* Arguement to pass to function */
+} ELAN3_HALTOP;
+
+#define HALTOP_BATCH   32
+
+#endif /* __KERNEL__ */
+
+typedef struct elan3_stats
+{
+    u_long     Version;                                        /* version field */
+    u_long     Interrupts;                                     /* count of elan interrupts */
+    u_long     TlbFlushes;                                     /* count of tlb flushes */
+    u_long     InvalidContext;                                 /* count of traps with invalid context */
+    u_long     ComQueueHalfFull;                               /* count of interrupts due to com queue being half full */
+
+    u_long     CProcTraps;                                     /* count of cproc traps */
+    u_long     DProcTraps;                                     /* count of dproc traps */
+    u_long     TProcTraps;                                     /* cound of tproc traps */
+    u_long     IProcTraps;                                     /* count of iproc traps */
+    u_long     EventInterrupts;                                /* count of event interrupts */
+
+    u_long     PageFaults;                                     /* count of elan page faults */
+
+    /* inputter related */
+    u_long     EopBadAcks;                                     /* count of EOP_BAD_ACKs */
+    u_long     EopResets;                                      /* count of EOP_ERROR_RESET */
+    u_long      InputterBadLength;                             /* count of BadLength */
+    u_long      InputterCRCDiscards;                           /* count of CRC_STATUS_DISCARD */
+    u_long      InputterCRCErrors;                             /* count of CRC_STATUS_ERROR */
+    u_long      InputterCRCBad;                                        /* count of CRC_STATUS_BAD */
+    u_long     DmaNetworkErrors;                               /* count of errors in dma data */
+    u_long     DmaIdentifyNetworkErrors;                       /* count of errors after dma identify */
+    u_long     ThreadIdentifyNetworkErrors;                    /* count of errors after thread identify */
+
+    /* dma related */
+    u_long     DmaRetries;                                     /* count of dma retries (due to retry fail count) */    
+    u_long     DmaOutputTimeouts;                              /* count of dma output timeouts */
+    u_long     DmaPacketAckErrors;                             /* count of dma packet ack errors */
+
+    /* thread related */
+    u_long     ForcedTProcTraps;                               /* count of forced tproc traps */
+    u_long     TrapForTooManyInsts;                            /* count of too many instruction traps */
+    u_long     ThreadOutputTimeouts;                           /* count of thread output timeouts */
+    u_long       ThreadPacketAckErrors;                                /* count of thread packet ack errors */
+
+    /* link related */
+    u_long     LockError;                                      /* count of RegPtr->Exts.LinkErrorTypes:LS_LockError */
+    u_long     DeskewError;                                    /* count of RegPtr->Exts.LinkErrorTypes:LS_DeskewError */
+    u_long     PhaseError;                                     /* count of RegPtr->Exts.LinkErrorTypes:LS_PhaseError */
+    u_long     DataError;                                      /* count of RegPtr->Exts.LinkErrorTypes:LS_DataError */
+    u_long     FifoOvFlow0;                                    /* count of RegPtr->Exts.LinkErrorTypes:LS_FifoOvFlow0 */
+    u_long     FifoOvFlow1;                                    /* count of RegPtr->Exts.LinkErrorTypes:LS_FifoOvFlow1 */
+    u_long     LinkErrorValue;                                 /* link error value on data error */
+
+    /* memory related */
+    u_long     CorrectableErrors;                              /* count of correctable ecc errors */
+    u_long     UncorrectableErrors;                            /* count of uncorrectable ecc errors */
+    u_long       MultipleErrors;                                       /* count of multiple ecc errors */
+    u_long     SdramBytesFree;                                 /* count of sdram bytes free */
+    
+    /* Interrupt related */
+    u_long     LongestInterrupt;                               /* length of longest interrupt in ticks */
+
+    u_long     EventPunts;                                     /* count of punts of event interrupts to thread */
+    u_long     EventRescheds;                                  /* count of reschedules of event interrupt thread */
+} ELAN3_STATS;
+
+#define ELAN3_STATS_VERSION    (ulong)2
+#define ELAN3_NUM_STATS                (sizeof (ELAN3_STATS)/sizeof (u_long))
+
+#define ELAN3_STATS_DEV_FMT   "elan3_stats_dev_%d"
+
+#ifdef __KERNEL__
+
+#define BumpStat(dev,stat)     ((dev)->Stats.stat++)
+
+typedef struct elan3_level_ptbl_block
+{
+    spinlock_t             PtblLock;                           /* Page table freelist lock */
+    int                            PtblTotal;                          /* Count of level N page tables allocated */
+    int                            PtblFreeCount;                      /* Count of free level N page tables */
+    struct elan3_ptbl     *PtblFreeList;                       /* Free level N page tables */
+    struct elan3_ptbl_gr          *PtblGroupList;                      /* List of Groups of level N page tables */
+} ELAN3_LEVEL_PTBL_BLOCK;
+ 
+typedef struct elan3_dev
+{
+    ELAN3_DEV_OSDEP        Osdep;                              /* OS specific entries */
+    int                            Instance;                           /* Device number */
+    ELAN_DEVINFO            Devinfo;                           
+    ELAN_POSITION          Position;                           /* position in switch network (for user code) */
+    ELAN_DEV_IDX           DeviceIdx;                          /* device index registered with elanmod */
+
+    int                            ThreadsShouldStop;                  /* flag that kernel threads should stop */
+
+    spinlock_t             IntrLock;
+    spinlock_t             TlbLock;
+    spinlock_t             CProcLock;
+    kcondvar_t             IntrWait;                           /* place event interrupt thread sleeps */
+    unsigned               EventInterruptThreadStarted:1;      /* event interrupt thread started */
+    unsigned               EventInterruptThreadStopped:1;      /* event interrupt thread stopped */
+    
+    DeviceMappingHandle            RegHandle;                          /* DDI Handle */
+    ioaddr_t               RegPtr;                             /* Elan Registers */
+
+    volatile E3_uint32     InterruptMask;                      /* copy of RegPtr->InterruptMask */
+    volatile E3_uint32     Event_Int_Queue_FPtr;               /* copy of RegPtr->Event_Int_Queue_FPtr */
+    volatile E3_uint32      SchCntReg;                         /* copy of RegPtr->SchCntReg */
+    volatile E3_uint32      Cache_Control_Reg;                 /* true value for RegPtr->Cache_Control_Reg */
+    
+    ELAN3_SDRAM_BANK       SdramBanks[ELAN3_SDRAM_NUM_BANKS];  /* Elan sdram banks */
+    spinlock_t             SdramLock;                          /* Sdram allocator */
+    sdramaddr_t                    SdramFreeLists[SDRAM_NUM_FREE_LISTS];
+    unsigned               SdramFreeCounts[SDRAM_NUM_FREE_LISTS];
+               
+    sdramaddr_t                    TAndQBase;                          /* Trap and Queue area */
+    sdramaddr_t                    ContextTable;                       /* Elan Context Table */
+    u_int                  ContextTableSize;                   /* # entries in context table */
+
+    struct elan3_ctxt      **CtxtTable;                         /* array of ctxt pointers or nulls */
+
+    sdramaddr_t                    CommandPortTraps[2];                /* Command port trap overflow */
+    int                            CurrentCommandPortTrap;             /* Which overflow queue we're using */
+    
+    u_int                  HaltAllCount;                       /* Count of reasons to halt context 0 queues */
+    u_int                  HaltNonContext0Count;               /* Count of reasons to halt non-context 0 queues */
+    u_int                  HaltDmaDequeueCount;                /* Count of reasons to halt dma from dequeuing */
+    u_int                  HaltThreadCount;                    /* Count of reasons to halt the thread processor */
+    u_int                  FlushCommandCount;                  /* Count of reasons to flush command queues */
+    u_int                  DiscardAllCount;                    /* Count of reasons to discard context 0 */
+    u_int                  DiscardNonContext0Count;            /* Count of reasons to discard non context 0 */
+
+    struct thread_trap    *ThreadTrap;                         /* Thread Processor trap space */
+    struct dma_trap       *DmaTrap;                            /* DMA Processor trap space */
+
+    spinlock_t             FreeHaltLock;                       /* Lock for haltop free list */
+    ELAN3_HALTOP                  *FreeHaltOperations;                 /* Free list of haltops */
+    u_int                  NumHaltOperations;                  /* Number of haltops allocated */
+    u_int                  ReservedHaltOperations;             /* Number of haltops reserved */
+
+    ELAN3_HALTOP                  *HaltOperations;                     /* List of operations to call */
+    ELAN3_HALTOP                 **HaltOperationsTailpp;               /* Pointer to last "next" pointer in list */
+    E3_uint32              HaltOperationsMask;                 /* Or of all bits in list of operations */
+
+    physaddr_t             SdramPhysBase;                      /* Physical address of SDRAM */
+    physaddr_t             SdramPhysMask;                      /* and mask of significant bits */ 
+    
+    physaddr_t             PciPhysBase;                        /* physical address of local PCI segment */
+    physaddr_t             PciPhysMask;                        /* and mask of significant bits */
+
+    long                   ErrorTime;                          /* lbolt at last error (link,ecc etc) */
+    long                   ErrorsPerTick;                      /* count of errors for this tick */
+    timer_fn_t             ErrorTimeoutId;                     /* id of timeout when errors masked out */
+    timer_fn_t             DmaPollTimeoutId;                   /* id of timeout to poll for "bad" dmas */
+    int                            FilterHaltQueued;
+
+    /*
+     * HAT layer specific entries.
+     */
+    ELAN3_LEVEL_PTBL_BLOCK   Level[4];
+    spinlock_t             PtblGroupLock;                      /* Lock for Page Table group lists */
+    struct elan3_ptbl_gr    *Level3PtblGroupHand;              /* Hand for ptbl stealing */
+
+    /*
+     * Per-Context Information structures.
+     */
+    struct elan3_info     *Infos;                              /* List of "infos" for this device */
+
+    char                    LinkShutdown;                       /* link forced into reset by panic/shutdown/dump */
+
+    /*
+     * Device statistics.
+     */
+    ELAN3_STATS                    Stats;
+    ELAN_STATS_IDX          StatsIndex;
+
+    struct {
+       E3_Regs            *RegPtr;
+       char               *Sdram[ELAN3_SDRAM_NUM_BANKS];
+    } PanicState;
+} ELAN3_DEV;
+
+#define ELAN3_DEV_CTX_TABLE(dev,ctxtn) ( (dev)->CtxtTable[ (ctxtn) &  MAX_ROOT_CONTEXT_MASK] )
+
+/* macros for accessing dev->RegPtr.Tags/Sets. */
+#define write_cache_tag(dev,what,val)  writeq (val, dev->RegPtr + offsetof (E3_Regs, Tags.what))
+#define read_cache_tag(dev,what)       readq (dev->RegPtr + offsetof (E3_Regs, Tags.what))
+#define write_cache_set(dev,what,val)  writeq (val, dev->RegPtr + offsetof (E3_Regs, Sets.what))
+#define read_cache_set(dev,what)       readq (dev->RegPtr + offsetof (E3_Regs, Sets.what))
+
+/* macros for accessing dev->RegPtr.Regs. */
+#define write_reg64(dev,what,val)      writeq (val, dev->RegPtr + offsetof (E3_Regs, Regs.what))
+#define write_reg32(dev,what,val)      writel (val, dev->RegPtr + offsetof (E3_Regs, Regs.what))
+#define read_reg64(dev,what)           readq (dev->RegPtr + offsetof (E3_Regs, Regs.what))
+#define read_reg32(dev,what)           readl (dev->RegPtr + offsetof (E3_Regs, Regs.what))
+
+/* macros for accessing dev->RegPtr.uRegs. */
+#define write_ureg64(dev,what,val)     writeq (val, dev->RegPtr + offsetof (E3_Regs, URegs.what))
+#define write_ureg32(dev,what,val)     writel (val, dev->RegPtr + offsetof (E3_Regs, URegs.what))
+#define read_ureg64(dev,what)          readq (dev->RegPtr + offsetof (E3_Regs, URegs.what))
+#define read_ureg32(dev,what)          readl (dev->RegPtr + offsetof (E3_Regs, URegs.what))
+
+/* macros for accessing dma descriptor/thread regs */
+#define copy_dma_regs(dev, desc) \
+MACRO_BEGIN \
+    register int i;  \
+    for (i = 0; i < sizeof (E3_DMA)/sizeof(E3_uint64); i++) \
+       ((E3_uint64 *) desc)[i] = readq (dev->RegPtr + offsetof (E3_Regs, Regs.Dma_Desc) + i*sizeof (E3_uint64)); \
+MACRO_END
+
+#define copy_thread_regs(dev, regs) \
+MACRO_BEGIN \
+    register int i;  \
+    for (i = 0; i < (32*sizeof (E3_uint32))/sizeof(E3_uint64); i++) \
+       ((E3_uint64 *) regs)[i] = readq (dev->RegPtr + offsetof (E3_Regs, Regs.Globals[0]) + i*sizeof (E3_uint64)); \
+MACRO_END
+
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::IntrLock, 
+                         _E3_DataBusMap::Exts _E3_DataBusMap::Input_Context_Fil_Flush
+                         elan3_dev::CurrentCommandPortTrap elan3_dev::HaltAllCount elan3_dev::HaltDmaDequeueCount
+                         elan3_dev::FlushCommandCount elan3_dev::DiscardAllCount elan3_dev::DiscardNonContext0Count
+                         elan3_dev::HaltOperations elan3_dev::HaltOperationsMask))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::TlbLock, 
+                         _E3_DataBusMap::Cache_Control_Reg))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::InfoLock, 
+                         elan3_dev::Infos elan3_dev::InfoTable))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::FreeHaltLock, 
+                         elan3_dev::FreeHaltOperations elan3_dev::NumHaltOperations elan3_dev::ReservedHaltOperations))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::PageFreeListLock, 
+                         elan3_dev::PageFreeList elan3_dev::PageFreeListSize))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::Level1PtblLock,
+                         elan3_dev::Level1PtblTotal elan3_dev::Level1PtblFreeCount elan3_dev::Level1PtblFreeList))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::Level2PtblLock,
+                         elan3_dev::Level2PtblTotal elan3_dev::Level2PtblFreeCount elan3_dev::Level2PtblFreeList))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::Level3PtblLock,
+                         elan3_dev::Level3PtblTotal elan3_dev::Level3PtblFreeCount elan3_dev::Level3PtblFreeList))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::PtblGroupLock,
+                         elan3_dev::Level1PtblGroupList elan3_dev::Level2PtblGroupList elan3_dev::Level3PtblGroupList))
+
+_NOTE(DATA_READABLE_WITHOUT_LOCK(elan3_dev::InfoTable elan3_dev::Level1PtblFreeList
+                                elan3_dev::Level2PtblFreeList elan3_dev::Level3PtblFreeList))
+
+_NOTE(LOCK_ORDER(elan3_dev::InfoLock elan3_dev::IntrLock))
+_NOTE(LOCK_ORDER(as::a_lock elan3_dev::InfoLock))
+_NOTE(LOCK_ORDER(as::a_lock elan3_dev::IntrLock))
+
+#define SET_INT_MASK(dev,Mask)         MACRO_BEGIN write_reg32 (dev, Exts.InterruptMask, ((dev)->InterruptMask = (Mask)));  mmiob(); MACRO_END
+#define ENABLE_INT_MASK(dev, bits)     MACRO_BEGIN write_reg32 (dev, Exts.InterruptMask, ((dev->InterruptMask |= (bits)))); mmiob(); MACRO_END
+#define DISABLE_INT_MASK(dev, bits)    MACRO_BEGIN write_reg32 (dev, Exts.InterruptMask, ((dev->InterruptMask &= ~(bits)))); mmiob(); MACRO_END
+
+#define INIT_SCHED_STATUS(dev, val) \
+MACRO_BEGIN \
+       (dev)->SchCntReg = (val); \
+       write_reg32 (dev, Exts.SchCntReg, (dev)->SchCntReg); \
+       mmiob(); \
+MACRO_END
+
+#define SET_SCHED_STATUS(dev, val) \
+MACRO_BEGIN \
+       ASSERT (((val) & HaltStopAndExtTestMask) == (val)); \
+       (dev)->SchCntReg |= (val); \
+       write_reg32 (dev, Exts.SchCntReg, (dev)->SchCntReg); \
+        mmiob (); \
+MACRO_END
+
+#define CLEAR_SCHED_STATUS(dev, val) \
+MACRO_BEGIN \
+       ASSERT (((val) & HaltStopAndExtTestMask) == (val)); \
+       (dev)->SchCntReg &= ~(val); \
+       write_reg32 (dev, Exts.SchCntReg, (dev)->SchCntReg); \
+        mmiob(); \
+MACRO_END
+
+#define MODIFY_SCHED_STATUS(dev, SetBits, ClearBits) \
+MACRO_BEGIN \
+       ASSERT ((((SetBits)|(ClearBits)) & HaltStopAndExtTestMask) == ((SetBits)|(ClearBits))); \
+       (dev)->SchCntReg = (((dev)->SchCntReg | (SetBits)) & ~(ClearBits)); \
+       write_reg32 (dev, Exts.SchCntReg, (dev)->SchCntReg); \
+       mmiob(); \
+MACRO_END
+
+#define PULSE_SCHED_STATUS(dev, RestartBits) \
+MACRO_BEGIN \
+       ASSERT (((RestartBits) & HaltStopAndExtTestMask) == 0); \
+       write_reg32 (dev, Exts.SchCntReg, (dev)->SchCntReg | (RestartBits)); \
+       mmiob(); \
+MACRO_END
+
+#define SET_SCHED_LINK_VALUE(dev, enabled, val) \
+MACRO_BEGIN \
+       (dev)->SchCntReg = (((dev)->SchCntReg & HaltAndStopMask) | ((enabled) ? LinkBoundaryScan : 0) | LinkSetValue(val, 0)); \
+       write_reg32 (dev, Exts.SchCntReg, (dev)->SchCntReg); \
+       mmiob(); \
+MACRO_END
+
+#ifdef DEBUG_ASSERT
+#  define ELAN3_ASSERT(dev, EX)        ((void)((EX) || elan3_assfail(dev, #EX, __FILE__, __LINE__)))
+#else
+#  define ELAN3_ASSERT(dev, EX)
+#endif
+
+/* elandev_generic.c */
+extern int        InitialiseElan (ELAN3_DEV *dev, ioaddr_t CmdPort);
+extern void       FinaliseElan (ELAN3_DEV *dev);
+extern int        InterruptHandler (ELAN3_DEV *dev);
+extern void       PollForDmaHungup (void *arg);
+
+extern int        SetLinkBoundaryScan (ELAN3_DEV *dev);
+extern void       ClearLinkBoundaryScan (ELAN3_DEV *dev);
+extern int        WriteBoundaryScanValue (ELAN3_DEV *dev, int value);
+extern int        ReadBoundaryScanValue(ELAN3_DEV *dev, int link);
+
+extern int        ReadVitalProductData (ELAN3_DEV *dev, int *CasLatency);
+
+extern struct elan3_ptbl_gr *ElanGetPtblGr (ELAN3_DEV *dev, sdramaddr_t offset);
+extern void       ElanSetPtblGr (ELAN3_DEV *dev, sdramaddr_t offset, struct elan3_ptbl_gr *ptg);
+
+extern void       ElanFlushTlb (ELAN3_DEV *dev);
+
+extern void       SetSchedStatusRegister (ELAN3_DEV *dev, E3_uint32 Pend, volatile E3_uint32 *Maskp);
+extern void      FreeHaltOperation (ELAN3_DEV *dev, ELAN3_HALTOP *op);
+extern int       ReserveHaltOperations (ELAN3_DEV *dev, int count, int cansleep);
+extern void      ReleaseHaltOperations (ELAN3_DEV *dev, int count);
+extern void      ProcessHaltOperations (ELAN3_DEV *dev, E3_uint32 Pend);
+extern void      QueueHaltOperation (ELAN3_DEV *dev, E3_uint32 Pend, volatile E3_uint32 *Maskp,
+                                     E3_uint32 ReqMask, void (*Function)(ELAN3_DEV *, void *), void *Arguement);
+
+extern int        ComputePosition (ELAN_POSITION *pos, unsigned NodeId, unsigned NumNodes, unsigned numDownLinksVal);
+
+extern caddr_t   MiToName (int mi);
+extern void      ElanBusError (ELAN3_DEV *dev);
+
+extern void      TriggerLsa (ELAN3_DEV *dev);
+
+extern ELAN3_DEV  *elan3_device (int instance);
+extern int       DeviceRegisterSize (ELAN3_DEV *dev, int rnumber, int *sizep);
+extern int       MapDeviceRegister (ELAN3_DEV *dev, int rnumber, ioaddr_t *addrp, int offset, 
+                                    int len, DeviceMappingHandle *handlep);
+extern void       UnmapDeviceRegister (ELAN3_DEV *dev, DeviceMappingHandle *handlep);
+
+
+/* sdram.c */
+/* sdram accessing functions - define 4 different types for 8,16,32,64 bit accesses */
+extern unsigned char      elan3_sdram_readb (ELAN3_DEV *dev, sdramaddr_t ptr);
+extern unsigned short     elan3_sdram_readw (ELAN3_DEV *dev, sdramaddr_t ptr);
+extern unsigned int       elan3_sdram_readl (ELAN3_DEV *dev, sdramaddr_t ptr);
+extern unsigned long long elan3_sdram_readq (ELAN3_DEV *dev, sdramaddr_t ptr);
+extern void               elan3_sdram_writeb (ELAN3_DEV *dev, sdramaddr_t ptr, unsigned char val);
+extern void               elan3_sdram_writew (ELAN3_DEV *dev, sdramaddr_t ptr, unsigned short val);
+extern void               elan3_sdram_writel (ELAN3_DEV *dev, sdramaddr_t ptr, unsigned int val);
+extern void               elan3_sdram_writeq (ELAN3_DEV *dev, sdramaddr_t ptr, unsigned long long val);
+
+extern void              elan3_sdram_zerob_sdram (ELAN3_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern void              elan3_sdram_zerow_sdram (ELAN3_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern void              elan3_sdram_zerol_sdram (ELAN3_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern void              elan3_sdram_zeroq_sdram (ELAN3_DEV *dev, sdramaddr_t ptr, int nbytes);
+
+extern void               elan3_sdram_copyb_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan3_sdram_copyw_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan3_sdram_copyl_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan3_sdram_copyq_from_sdram (ELAN3_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan3_sdram_copyb_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+extern void               elan3_sdram_copyw_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+extern void               elan3_sdram_copyl_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+extern void               elan3_sdram_copyq_to_sdram (ELAN3_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+
+extern void              elan3_sdram_init (ELAN3_DEV *dev);
+extern void               elan3_sdram_fini (ELAN3_DEV *dev);
+extern void              elan3_sdram_add (ELAN3_DEV *dev, sdramaddr_t base, sdramaddr_t top);
+extern sdramaddr_t        elan3_sdram_alloc (ELAN3_DEV *dev, int nbytes);
+extern void               elan3_sdram_free (ELAN3_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern physaddr_t         elan3_sdram_to_phys (ELAN3_DEV *dev, sdramaddr_t addr);
+
+/* cproc.c */
+extern void      HandleCProcTrap (ELAN3_DEV *dev, E3_uint32 Pend, E3_uint32 *Mask);
+
+/* iproc.c */
+extern void      HandleIProcTrap (ELAN3_DEV *dev, int Channel, E3_uint32 Pend, sdramaddr_t FaultSaveOff, 
+                                  sdramaddr_t TransactionsOff, sdramaddr_t DataOff);
+
+/* tproc.c */
+extern int       HandleTProcTrap (ELAN3_DEV *dev, E3_uint32 *RestartBits);
+extern void      DeliverTProcTrap (ELAN3_DEV *dev, struct thread_trap *threadTrap, E3_uint32 Pend);
+
+/* dproc.c */
+extern int       HandleDProcTrap (ELAN3_DEV *dev, E3_uint32 *RestartBits);
+extern void      DeliverDProcTrap (ELAN3_DEV *dev, struct dma_trap *dmaTrap, E3_uint32 Pend);
+
+#if defined(LINUX)
+/* procfs_linux.h */
+extern struct proc_dir_entry *elan3_procfs_root;
+extern struct proc_dir_entry *elan3_config_root;
+
+extern void elan3_procfs_init(void);
+extern void elan3_procfs_fini(void);
+extern void elan3_procfs_device_init (ELAN3_DEV *dev);
+extern void elan3_procfs_device_fini (ELAN3_DEV *dev);
+#endif /* defined(LINUX) */
+
+/* elan3_osdep.c */
+extern int        BackToBackMaster;
+extern int        BackToBackSlave;
+
+#define ELAN_REG_REC_MAX (100)
+#define ELAN_REG_REC(REG)  {                                         \
+elan_reg_rec_file [elan_reg_rec_index] = __FILE__;                   \
+elan_reg_rec_line [elan_reg_rec_index] = __LINE__;                   \
+elan_reg_rec_reg  [elan_reg_rec_index] = REG;                        \
+elan_reg_rec_cpu  [elan_reg_rec_index] = smp_processor_id();         \
+elan_reg_rec_lbolt[elan_reg_rec_index] = lbolt;                      \
+elan_reg_rec_index = ((elan_reg_rec_index+1) % ELAN_REG_REC_MAX);}
+
+extern char *    elan_reg_rec_file [ELAN_REG_REC_MAX];
+extern int       elan_reg_rec_line [ELAN_REG_REC_MAX];
+extern long      elan_reg_rec_lbolt[ELAN_REG_REC_MAX];
+extern int       elan_reg_rec_cpu  [ELAN_REG_REC_MAX];
+extern E3_uint32 elan_reg_rec_reg  [ELAN_REG_REC_MAX];
+extern int       elan_reg_rec_index;
+ 
+#endif /* __KERNEL__ */
+
+
+#define ELAN3_PROCFS_ROOT          "/proc/qsnet/elan3"
+#define ELAN3_PROCFS_VERSION       "/proc/qsnet/elan3/version"
+#define ELAN3_PROCFS_DEBUG         "/proc/qsnet/elan3/config/elandebug"
+#define ELAN3_PROCFS_DEBUG_CONSOLE "/proc/qsnet/elan3/config/elandebug_console"
+#define ELAN3_PROCFS_DEBUG_BUFFER  "/proc/qsnet/elan3/config/elandebug_buffer"
+#define ELAN3_PROCFS_MMU_DEBUG     "/proc/qsnet/elan3/config/elan3mmu_debug"
+#define ELAN3_PROCFS_PUNT_LOOPS    "/proc/qsnet/elan3/config/eventint_punt_loops"
+
+#define ELAN3_PROCFS_DEVICE_STATS_FMT    "/proc/qsnet/elan3/device%d/stats"
+#define ELAN3_PROCFS_DEVICE_POSITION_FMT "/proc/qsnet/elan3/device%d/position"
+#define ELAN3_PROCFS_DEVICE_NODESET_FMT  "/proc/qsnet/elan3/device%d/nodeset"
+
+#endif /* __ELAN3_ELANDEV_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elandev_linux.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elandev_linux.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elandev_linux.h  2005-06-01 23:12:54.721420560 -0400
@@ -0,0 +1,56 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELANDEV_LINUX_H
+#define __ELANDEV_LINUX_H
+
+#ident "$Id: elandev_linux.h,v 1.11 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elandev_linux.h,v $*/
+
+#ifdef __KERNEL__
+#include <linux/mm.h>
+#include <linux/sched.h>
+#include <linux/pci.h>
+#endif
+
+#define ELAN3_MAJOR              60
+#define ELAN3_NAME               "elan3"
+#define ELAN3_MAX_CONTROLLER     16                      /* limited to 4 bits */
+ 
+#define ELAN3_MINOR_DEVNUM(m)    ((m) & 0x0f)            /* card number */
+#define ELAN3_MINOR_DEVFUN(m)    (((m) >> 4) & 0x0f)     /* function */
+#define ELAN3_MINOR_CONTROL      0                       /* function values */
+#define ELAN3_MINOR_MEM          1
+#define ELAN3_MINOR_USER               2
+ 
+typedef void                   *DeviceMappingHandle;
+
+/* task and ctxt handle types */
+typedef struct mm_struct       *TaskHandle;
+typedef int                    CtxtHandle;
+ 
+#define ELAN3_MY_TASK_HANDLE() (current->mm)
+#define KERNEL_TASK_HANDLE()   (get_kern_mm())
+ 
+/*
+ * OS-dependent component of ELAN3_DEV struct.
+ */
+typedef struct elan3_dev_osdep
+{
+       struct pci_dev  *pci;                   /* PCI config data */
+       int             ControlDeviceOpen;      /* flag to indicate control */
+                                               /*   device open */
+       struct proc_dir_entry *procdir;
+} ELAN3_DEV_OSDEP;
+
+#endif /* __ELANDEV_LINUX_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elanio.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elanio.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elanio.h 2005-06-01 23:12:54.722420408 -0400
@@ -0,0 +1,226 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_ELAN3IO_H
+#define __ELAN3_ELAN3IO_H
+
+#ident "$Id: elanio.h,v 1.19 2003/12/08 15:40:26 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elanio.h,v $*/
+
+#define ELAN3IO_CONTROL_PATHNAME       "/dev/elan3/control%d"
+#define ELAN3IO_MEM_PATHNAME   "/dev/elan3/mem%d"
+#define ELAN3IO_USER_PATHNAME  "/dev/elan3/user%d"
+#define ELAN3IO_SDRAM_PATHNAME   "/dev/elan3/sdram%d"
+#define ELAN3IO_MAX_PATHNAMELEN        32
+
+/* ioctls on /dev/elan3/control */
+#define ELAN3IO_CONTROL_BASE           0
+
+#define ELAN3IO_SET_BOUNDARY_SCAN      _IO   ('e', ELAN3IO_CONTROL_BASE + 0)
+#define ELAN3IO_CLEAR_BOUNDARY_SCAN    _IO   ('e', ELAN3IO_CONTROL_BASE + 1)
+#define ELAN3IO_READ_LINKVAL           _IOWR ('e', ELAN3IO_CONTROL_BASE + 2, E3_uint32)
+#define ELAN3IO_WRITE_LINKVAL          _IOWR ('e', ELAN3IO_CONTROL_BASE + 3, E3_uint32)
+
+typedef struct elanio_set_debug_struct
+{
+    char       what[32];
+    u_long     value;
+} ELAN3IO_SET_DEBUG_STRUCT;
+#define ELAN3IO_SET_DEBUG              _IOW  ('e', ELAN3IO_CONTROL_BASE + 4, ELAN3IO_SET_DEBUG_STRUCT)
+
+typedef struct elanio_debug_buffer_struct
+{
+    caddr_t    addr;
+    size_t     len;
+} ELAN3IO_DEBUG_BUFFER_STRUCT;
+#define ELAN3IO_DEBUG_BUFFER           _IOWR ('e', ELAN3IO_CONTROL_BASE + 5, ELAN3IO_DEBUG_BUFFER_STRUCT)
+
+typedef struct elanio_neterr_server_struct
+{
+    u_int              elanid;
+    void              *addr;
+    char              *name;
+} ELAN3IO_NETERR_SERVER_STRUCT;
+#define ELAN3IO_NETERR_SERVER          _IOW  ('e', ELAN3IO_CONTROL_BASE + 6, ELAN3IO_NETERR_SERVER_STRUCT)
+#define ELAN3IO_NETERR_FIXUP           _IOWR ('e', ELAN3IO_CONTROL_BASE + 7, NETERR_MSG)
+
+typedef struct elanio_set_position_struct
+{
+    u_int              device;
+    unsigned short      nodeId;
+    unsigned short      numNodes;
+} ELAN3IO_SET_POSITION_STRUCT;
+#define ELAN3IO_SET_POSITION           _IOW ('e', ELAN3IO_CONTROL_BASE + 8, ELAN3IO_SET_POSITION_STRUCT)
+
+#if defined(LINUX)
+
+/* ioctls on /dev/elan3/sdram */
+#define ELAN3IO_SDRAM_BASE             20
+
+/* ioctls on /dev/elan3/user */
+#define ELAN3IO_USER_BASE              30
+
+#define ELAN3IO_FREE                   _IO  ('e', ELAN3IO_USER_BASE + 0)
+
+#define ELAN3IO_ATTACH                 _IOWR('e', ELAN3IO_USER_BASE + 1, ELAN_CAPABILITY)
+#define ELAN3IO_DETACH                 _IO  ('e', ELAN3IO_USER_BASE + 2)
+
+typedef struct elanio_addvp_struct
+{
+    u_int              process;
+    ELAN_CAPABILITY     capability;
+} ELAN3IO_ADDVP_STRUCT;
+#define ELAN3IO_ADDVP                  _IOWR('e', ELAN3IO_USER_BASE + 3, ELAN3IO_ADDVP_STRUCT)
+#define ELAN3IO_REMOVEVP                       _IOW ('e', ELAN3IO_USER_BASE + 4, int)
+
+typedef struct elanio_bcastvp_struct
+{
+    u_int              process;
+    u_int              lowvp;
+    u_int              highvp;
+} ELAN3IO_BCASTVP_STRUCT;
+#define ELAN3IO_BCASTVP                        _IOW ('e', ELAN3IO_USER_BASE + 5, ELAN3IO_BCASTVP_STRUCT)
+
+typedef struct elanio_loadroute_struct
+{
+    u_int              process;
+    E3_uint16          flits[MAX_FLITS];
+} ELAN3IO_LOAD_ROUTE_STRUCT;
+#define ELAN3IO_LOAD_ROUTE             _IOW ('e', ELAN3IO_USER_BASE + 6, ELAN3IO_LOAD_ROUTE_STRUCT)
+
+#define ELAN3IO_PROCESS                        _IO  ('e', ELAN3IO_USER_BASE + 7)
+
+typedef struct elanio_setperm_struct
+{
+    caddr_t            maddr;
+    E3_Addr            eaddr;
+    size_t             len;
+    int                        perm;
+} ELAN3IO_SETPERM_STRUCT;
+#define ELAN3IO_SETPERM                        _IOW ('e', ELAN3IO_USER_BASE + 8, ELAN3IO_SETPERM_STRUCT)
+
+typedef struct elanio_clearperm_struct
+{
+    E3_Addr            eaddr;
+    size_t             len;
+} ELAN3IO_CLEARPERM_STRUCT;
+#define ELAN3IO_CLEARPERM              _IOW ('e', ELAN3IO_USER_BASE + 9, ELAN3IO_CLEARPERM_STRUCT)
+
+typedef struct elanio_changeperm_struct
+{
+    E3_Addr            eaddr;
+    size_t             len;
+    int                        perm;
+} ELAN3IO_CHANGEPERM_STRUCT;
+#define ELAN3IO_CHANGEPERM             _IOW ('e', ELAN3IO_USER_BASE + 10, ELAN3IO_CHANGEPERM_STRUCT)
+
+
+#define ELAN3IO_HELPER_THREAD          _IO  ('e', ELAN3IO_USER_BASE + 11)
+#define ELAN3IO_WAITCOMMAND            _IO  ('e', ELAN3IO_USER_BASE + 12)
+#define ELAN3IO_BLOCK_INPUTTER         _IOW ('e', ELAN3IO_USER_BASE + 13, int)
+#define ELAN3IO_SET_FLAGS              _IOW ('e', ELAN3IO_USER_BASE + 14, int)
+
+#define ELAN3IO_WAITEVENT              _IOW ('e', ELAN3IO_USER_BASE + 15, E3_Event)
+#define ELAN3IO_ALLOC_EVENTCOOKIE      _IOW ('e', ELAN3IO_USER_BASE + 16, EVENT_COOKIE)
+#define ELAN3IO_FREE_EVENTCOOKIE               _IOW ('e', ELAN3IO_USER_BASE + 17, EVENT_COOKIE)
+#define ELAN3IO_ARM_EVENTCOOKIE                _IOW ('e', ELAN3IO_USER_BASE + 18, EVENT_COOKIE)
+#define ELAN3IO_WAIT_EVENTCOOKIE               _IOW ('e', ELAN3IO_USER_BASE + 19, EVENT_COOKIE)
+
+#define ELAN3IO_SWAPSPACE              _IOW ('e', ELAN3IO_USER_BASE + 20, SYS_SWAP_SPACE)
+#define ELAN3IO_EXCEPTION_SPACE                _IOW ('e', ELAN3IO_USER_BASE + 21, SYS_EXCEPTION_SPACE)
+#define ELAN3IO_GET_EXCEPTION          _IOR ('e', ELAN3IO_USER_BASE + 22, SYS_EXCEPTION)
+
+typedef struct elanio_unload_struct
+{
+    void       *addr;
+    size_t      len;
+} ELAN3IO_UNLOAD_STRUCT;
+#define ELAN3IO_UNLOAD                 _IOW ('e', ELAN3IO_USER_BASE + 23, ELAN3IO_UNLOAD_STRUCT)
+
+
+
+typedef struct elanio_getroute_struct
+{
+    u_int              process;
+    E3_uint16          flits[MAX_FLITS];
+} ELAN3IO_GET_ROUTE_STRUCT;
+#define ELAN3IO_GET_ROUTE              _IOW ('e', ELAN3IO_USER_BASE + 24, ELAN3IO_GET_ROUTE_STRUCT)
+
+typedef struct elanio_resetroute_struct
+{
+    u_int              process;
+} ELAN3IO_RESET_ROUTE_STRUCT;
+#define ELAN3IO_RESET_ROUTE            _IOW ('e', ELAN3IO_USER_BASE + 25, ELAN3IO_RESET_ROUTE_STRUCT)
+
+typedef struct elanio_checkroute_struct
+{
+    u_int              process;
+    E3_uint32           routeError;
+    E3_uint16          flits[MAX_FLITS];
+} ELAN3IO_CHECK_ROUTE_STRUCT;
+#define ELAN3IO_CHECK_ROUTE            _IOW ('e', ELAN3IO_USER_BASE + 26, ELAN3IO_CHECK_ROUTE_STRUCT)
+
+typedef struct elanio_vp2nodeId_struct
+{
+    u_int              process;
+    unsigned short      nodeId;
+    ELAN_CAPABILITY    cap;
+} ELAN3IO_VP2NODEID_STRUCT;
+#define ELAN3IO_VP2NODEID      _IOWR('e', ELAN3IO_USER_BASE + 27, ELAN3IO_VP2NODEID_STRUCT)
+
+#define ELAN3IO_SET_SIGNAL     _IOW ('e', ELAN3IO_USER_BASE + 28, int)
+
+typedef struct elanio_process_2_location_struct
+{
+    u_int              process;
+    ELAN_LOCATION       loc;
+} ELAN3IO_PROCESS_2_LOCATION_STRUCT;
+#define ELAN3IO_PROCESS_2_LOCATION     _IOW ('e', ELAN3IO_USER_BASE + 29, ELAN3IO_PROCESS_2_LOCATION_STRUCT)
+
+
+
+/* ioctls on all device */
+#define ELAN3IO_GENERIC_BASE           100
+typedef struct elanio_get_devinfo_struct
+{
+    ELAN_DEVINFO *devinfo;
+} ELAN3IO_GET_DEVINFO_STRUCT;
+#define ELAN3IO_GET_DEVINFO            _IOR ('e', ELAN3IO_GENERIC_BASE + 0, ELAN_DEVINFO)
+
+typedef struct elanio_get_position_struct
+{
+    ELAN_POSITION *position;
+} ELAN3IO_GET_POSITION_STRUCT;
+#define ELAN3IO_GET_POSITION             _IOR ('e', ELAN3IO_GENERIC_BASE + 1, ELAN_POSITION)
+
+typedef struct elanio_stats_struct
+{
+    int                which;
+    void       *ptr;
+} ELAN3IO_STATS_STRUCT;
+#define ELAN3IO_STATS                  _IOR ('e', ELAN3IO_GENERIC_BASE + 2, ELAN3IO_STATS_STRUCT)
+#  define ELAN3_SYS_STATS_DEVICE       0
+#  define ELAN3_SYS_STATS_MMU          1
+
+/* offsets on /dev/elan3/control */
+
+/* offsets on /dev/elan3/mem */
+
+/* page numbers on /dev/elan3/user */
+#define ELAN3IO_OFF_COMMAND_PAGE               0
+#define ELAN3IO_OFF_FLAG_PAGE          1
+#define ELAN3IO_OFF_UREG_PAGE          2
+
+#endif /* LINUX */
+
+#endif /* __ELAN3_ELAN3IO_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elanregs.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elanregs.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elanregs.h       2005-06-01 23:12:54.724420104 -0400
@@ -0,0 +1,1063 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+/*
+ * Header file for internal slave mapping of the ELAN3 registers
+ */
+
+#ifndef _ELAN3_ELANREGS_H
+#define _ELAN3_ELANREGS_H
+
+#ident "$Id: elanregs.h,v 1.87 2004/04/22 12:27:21 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elanregs.h,v $*/
+
+#include <elan3/e3types.h>
+#include <elan3/dma.h>
+#include <elan3/elanuregs.h>
+
+#define MAX_ROOT_CONTEXT_MASK  0xfff
+#define SYS_CONTEXT_BIT                0x1000
+#define ALL_CONTEXT_BITS       (MAX_ROOT_CONTEXT_MASK | SYS_CONTEXT_BIT)
+#define ROOT_TAB_OFFSET(Cntxt) (((Cntxt) & MAX_ROOT_CONTEXT_MASK) << 4)
+#define CLEAR_SYS_BIT(Cntxt)   ((Cntxt) & ~SYS_CONTEXT_BIT)
+
+#define E3_CACHELINE_SIZE      (32)
+#define E3_CACHE_SIZE          (8192)
+
+typedef volatile struct _E3_CacheSets
+{
+   E3_uint64   Set0[256];      /* 2k bytes per set */
+   E3_uint64   Set1[256];      /* 2k bytes per set */
+   E3_uint64   Set2[256];      /* 2k bytes per set */
+   E3_uint64   Set3[256];      /* 2k bytes per set */
+} E3_CacheSets;
+
+typedef union e3_cache_tag
+{
+   E3_uint64   Value;
+   struct {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 pad2:8;               /* Undefined value when read */
+       E3_uint32 LineError:1;          /* A line error has occured */
+       E3_uint32 Modified:1;           /* Cache data is modified */
+       E3_uint32 FillPending:1;                /* Pipelined fill occuring*/
+       E3_uint32 AddrTag27to11:17;     /* Tag address bits 27 to 11 */
+       E3_uint32 pad1:4;               /* Undefined value when read */
+       E3_uint32 pad0;                 /* Undefined value when read */
+#else
+       E3_uint32 pad0;                 /* Undefined value when read */
+       E3_uint32 pad1:4;               /* Undefined value when read */
+       E3_uint32 AddrTag27to11:17;     /* Tag address bits 27 to 11 */
+       E3_uint32 FillPending:1;                /* Pipelined fill occuring*/
+       E3_uint32 Modified:1;           /* Cache data is modified */
+       E3_uint32 LineError:1;          /* A line error has occured */
+       E3_uint32 pad2:8;               /* Undefined value when read */
+#endif
+   } s;
+} E3_CacheTag;
+
+#define E3_NumCacheLines       64
+#define E3_NumCacheSets                4
+
+typedef volatile struct _E3_CacheTags
+{
+   E3_CacheTag Tags[E3_NumCacheLines][E3_NumCacheSets];        /* 2k bytes per set */
+} E3_CacheTags;
+
+typedef union E3_IProcStatus_Reg
+{
+    E3_uint32 Status;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 TrapType:8;           /* iprocs trap ucode address */
+       E3_uint32 SuspendAddr:8;        /* iprocs suspend address */
+       E3_uint32 EopType:2;            /* Type of Eop Received */
+       E3_uint32 QueueingPacket:1;     /* receiving a queueing packet */
+       E3_uint32 AckSent:1;            /* a packet ack has been sent */
+       E3_uint32 Reject:1;             /* a packet nack has been sent */
+       E3_uint32 CrcStatus:2;          /* Crc Status value */
+       E3_uint32 BadLength:1;          /* Eop was received in a bad place */
+       E3_uint32 Chan1:1;              /* This packet received on v chan1 */
+       E3_uint32 First:1;              /* This is the first transaction in the packet */
+       E3_uint32 Last:1;               /* This is the last transaction in the packet */
+       E3_uint32 Unused:2;
+       E3_uint32 WakeupFunction:3;     /* iprocs wakeup function */
+#else
+       E3_uint32 WakeupFunction:3;     /* iprocs wakeup function */
+       E3_uint32 Unused:2;
+       E3_uint32 Last:1;               /* This is the last transaction in the packet */
+       E3_uint32 First:1;              /* This is the first transaction in the packet */
+       E3_uint32 Chan1:1;              /* This packet received on v chan1 */
+       E3_uint32 BadLength:1;          /* Eop was received in a bad place */
+       E3_uint32 CrcStatus:2;          /* Crc Status value */
+       E3_uint32 Reject:1;             /* a packet nack has been sent */
+       E3_uint32 AckSent:1;            /* a packet ack has been sent */
+       E3_uint32 QueueingPacket:1;     /* receiving a queueing packet */
+       E3_uint32 EopType:2;            /* Type of Eop Received */
+       E3_uint32 SuspendAddr:8;        /* iprocs suspend address */
+       E3_uint32 TrapType:8;           /* iprocs trap ucode address */
+#endif
+    } s;
+} E3_IProcStatus_Reg;
+
+#define CRC_STATUS_GOOD    (0 << 21)
+#define CRC_STATUS_DISCARD (1 << 21)
+#define CRC_STATUS_ERROR   (2 << 21)
+#define CRC_STATUS_BAD     (3 << 21)
+
+#define CRC_MASK          (3 << 21)
+
+#define EOP_GOOD          (1 << 16)
+#define EOP_BADACK        (2 << 16)
+#define EOP_ERROR_RESET           (3 << 16)
+
+#define E3_IPS_LastTrans       (1 << 26)
+#define E3_IPS_FirstTrans      (1 << 25)
+#define E3_IPS_VChan1          (1 << 24)
+#define E3_IPS_BadLength       (1 << 23)
+#define E3_IPS_CrcMask         (3 << 21)
+#define E3_IPS_Rejected                (1 << 20)
+#define E3_IPS_AckSent         (1 << 19)
+#define E3_IPS_QueueingPacket  (1 << 18)
+#define E3_IPS_EopType         (3 << 16)
+
+typedef union E3_Status_Reg
+{
+    E3_uint32 Status;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 TrapType:8;           /* procs trap ucode address */
+       E3_uint32 SuspendAddr:8;        /* procs suspend address */
+       E3_uint32 Context:13;           /* procs current context */
+       E3_uint32 WakeupFunction:3;     /* procs wakeup function */
+#else
+       E3_uint32 WakeupFunction:3;     /* procs wakeup function */
+       E3_uint32 Context:13;           /* procs current context */
+       E3_uint32 SuspendAddr:8;        /* procs suspend address */
+       E3_uint32 TrapType:8;           /* procs trap ucode address */
+#endif
+    } s;
+} E3_Status_Reg;
+
+/* values for WakeupFunction */
+#define SleepOneTick                   0
+#define WakeupToSendTransOrEop         1
+#define SleepOneTickThenRunnable       2
+#define WakeupNever                    4
+/* extra dma wakeup functions */
+#define WakupeToSendTransOrEop         1
+#define WakeupForPacketAck             3
+#define WakeupToSendTrans              5
+/* extra thread wakup function */
+#define WakeupStopped                  3
+/* extra cproc wakup function */
+#define WakeupSetEvent                 3
+
+#define GET_STATUS_CONTEXT(Ptr)      ((Ptr.Status >> 16) & 0x1fff)
+#define GET_STATUS_SUSPEND_ADDR(Ptr) ((Ptr.Status >> 8) & 0xff)
+#define GET_STATUS_TRAPTYPE(Ptr)     ((E3_uint32)(Ptr.Status & 0xff))
+
+/*
+ * Interrupt register bits
+ */
+#define INT_PciMemErr                  (1<<15) /* Pci memory access error */
+#define INT_SDRamInt                   (1<<14) /* SDRam ECC interrupt */
+#define INT_EventInterrupt             (1<<13) /* Event Interrupt */
+#define INT_LinkError                  (1<<12) /* Link Error */
+#define INT_ComQueue                   (1<<11) /* a comm queue half full */
+#define INT_TProcHalted                        (1<<10) /* Tproc Halted */
+#define INT_DProcHalted                        (1<<9) /* Dmas Halted */
+#define INT_DiscardingNonSysCntx       (1<<8) /* Inputters Discarding Non-SysCntx */
+#define INT_DiscardingSysCntx          (1<<7) /* Inputters Discarding SysCntx */
+#define INT_TProc                      (1<<6) /* tproc interrupt */
+#define INT_CProc                      (1<<5) /* cproc interrupt */
+#define INT_DProc                      (1<<4) /* dproc interrupt */
+#define INT_IProcCh1NonSysCntx         (1<<3) /* iproc non-SysCntx interrupt */
+#define INT_IProcCh1SysCntx            (1<<2) /* iproc SysCntx interrupt */
+#define INT_IProcCh0NonSysCntx         (1<<1) /* iproc non-SysCntx interrupt */
+#define INT_IProcCh0SysCntx            (1<<0) /* iproc SysCntx interrupt */
+
+#define INT_Inputters          (INT_IProcCh0SysCntx | INT_IProcCh0NonSysCntx | INT_IProcCh1SysCntx | INT_IProcCh1NonSysCntx)
+#define INT_Discarding         (INT_DiscardingSysCntx | INT_DiscardingNonSysCntx)
+#define INT_Halted             (INT_DProcHalted | INT_TProcHalted)
+#define INT_ErrorInterrupts    (INT_PciMemErr | INT_SDRamInt | INT_LinkError)
+
+/*
+ * Link state bits.
+ */
+#define LS_LinkNotReady        (1 << 0) /* Link is in reset or recovering from an error */
+#define LS_Locked      (1 << 1) /* Linkinput PLL is locked */
+#define LS_LockError   (1 << 2) /* Linkinput PLL was unable to lock onto the input clock. */
+#define LS_DeskewError (1 << 3) /* Linkinput was unable to Deskew all the inputs. (Broken wire?) */
+#define LS_PhaseError  (1 << 4) /* Linkinput Phase alignment error. */
+#define LS_DataError   (1 << 5) /* Received value was neither good data or a token. */
+#define LS_FifoOvFlow0 (1 << 6) /* Channel 0 input fifo overflowed. */
+#define LS_FifoOvFlow1 (1 << 7) /* Channel 1 input fifo overflowed. */
+
+/*
+ * Link State Constant defines, used for writing to LinkSetValue
+ */
+
+#define LRS_DataDel0           0x0
+#define LRS_DataDel1           0x1
+#define LRS_DataDel2           0x2
+#define LRS_DataDel3           0x3
+#define LRS_DataDel4           0x4
+#define LRS_DataDel5           0x5
+#define LRS_DataDel6           0x6
+#define LRS_DataDel7           0x7
+#define LRS_DataDel8           0x8
+#define LRS_PllDelValue                0x9
+#define LRS_ClockEven          0xA
+#define LRS_ClockOdd           0xB
+#define LRS_ErrorLSW           0xC
+#define LRS_ErrorMSW           0xD
+#define LRS_FinCoarseDeskew    0xE
+#define LRS_LinkInValue                0xF
+#define LRS_NumLinkDels         0x10
+
+#define LRS_Pllfast             0x40
+ 
+union Sched_Status
+{
+    E3_uint32 Status;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 StopNonSysCntxs:1;
+       E3_uint32 FlushCommandQueues:1;
+       E3_uint32 HaltDmas:1;
+       E3_uint32 HaltDmaDequeue:1;
+       E3_uint32 HaltThread:1;
+       E3_uint32 CProcStop:1;
+       E3_uint32 DiscardSysCntxIn:1;
+       E3_uint32 DiscardNonSysCntxIn:1;
+       E3_uint32 RestartCh0SysCntx:1;
+       E3_uint32 RestartCh0NonSysCntx:1;
+       E3_uint32 RestartCh1SysCntx:1;
+       E3_uint32 RestartCh1NonSysCntx:1;
+       E3_uint32 RestartDProc:1;
+       E3_uint32 RestartTProc:1;
+       E3_uint32 RestartCProc:1;
+       E3_uint32 ClearLinkErrorInt:1;
+       E3_uint32 :3;
+       E3_uint32 LinkSetValue:10; 
+       E3_uint32 FixLinkDelays:1;
+       E3_uint32 LinkBoundaryScan:1;
+#else
+       E3_uint32 LinkBoundaryScan:1;
+       E3_uint32 FixLinkDelays:1;
+       E3_uint32 LinkSetValue:10; 
+       E3_uint32 :3;
+       E3_uint32 ClearLinkErrorInt:1;
+       E3_uint32 RestartCProc:1;
+       E3_uint32 RestartTProc:1;
+       E3_uint32 RestartDProc:1;
+       E3_uint32 RestartCh1NonSysCntx:1;
+       E3_uint32 RestartCh1SysCntx:1;
+       E3_uint32 RestartCh0NonSysCntx:1;
+       E3_uint32 RestartCh0SysCntx:1;
+       E3_uint32 DiscardNonSysCntxIn:1;
+       E3_uint32 DiscardSysCntxIn:1;
+       E3_uint32 CProcStop:1;
+       E3_uint32 HaltThread:1;
+       E3_uint32 HaltDmaDequeue:1;
+       E3_uint32 HaltDmas:1;
+       E3_uint32 FlushCommandQueues:1;
+       E3_uint32 StopNonSysCntxs:1;
+#endif
+    } s;
+};
+
+#define LinkBoundaryScan       ((E3_uint32) 1<<31) /* Clears the link error interrupt */
+#define FixLinkDelays          ((E3_uint32) 1<<30) /* Clears the link error interrupt */
+#define LinkSetValue(Val, OldVal) ((E3_uint32) (((Val) & 0x3ff) << 20) | ((OldVal) & ((~0x3ff) << 20)))
+
+#define ClearLinkErrorInt      ((E3_uint32) 1<<16) /* Clears the link error interrupt */
+#define RestartCProc           ((E3_uint32) 1<<15) /* Clears command proc interrupt */
+#define RestartTProc           ((E3_uint32) 1<<14) /* Clears thread interrupt */
+#define RestartDProc           ((E3_uint32) 1<<13) /* Clears dma0 interrupt */
+#define RestartCh1NonSysCntx   ((E3_uint32) 1<<12) /* Clears interrupt */
+#define RestartCh1SysCntx      ((E3_uint32) 1<<11) /* Clears interrupt */
+#define RestartCh0NonSysCntx   ((E3_uint32) 1<<10) /* Clears interrupt */
+#define RestartCh0SysCntx      ((E3_uint32) 1<<9) /* Clears interrupt */
+#define CProcStopped           ((E3_uint32) 1<<9) /* Read value only */
+
+#define TraceSetEvents         ((E3_uint32) 1<<8)
+#define DiscardNonSysCntxIn    ((E3_uint32) 1<<7)
+#define DiscardSysCntxIn       ((E3_uint32) 1<<6)
+#define CProcStop              ((E3_uint32) 1<<5) /* Will empty all the command port queues. */
+#define HaltThread             ((E3_uint32) 1<<4) /* Will stop the thread proc and clear the tproc command queue */
+#define HaltDmaDequeue         ((E3_uint32) 1<<3) /* Will stop the dmaers starting new dma's. */
+#define HaltDmas               ((E3_uint32) 1<<2) /* Will stop the dmaers and clear the dma command queues */
+#define FlushCommandQueues     ((E3_uint32) 1<<1) /* Causes the command ports to be flushed. */
+#define StopNonSysCntxs                ((E3_uint32) 1<<0) /* Prevents a non-SysCntx from starting. */
+
+/* Initial value of schedule status register */
+#define LinkResetToken         0x00F
+
+#define Sched_Initial_Value    (LinkBoundaryScan | (LinkResetToken << 20) | \
+                                DiscardSysCntxIn | DiscardNonSysCntxIn | HaltThread | HaltDmas)
+
+#define StopDmaQueues       (HaltDmaDequeue | HaltDmas | \
+                             DiscardNonSysCntxIn | DiscardSysCntxIn)
+#define CheckDmaQueueStopped (INT_DiscardingNonSysCntx | INT_DiscardingSysCntx | INT_DProcHalted)
+
+#define HaltStopAndExtTestMask 0xfff001ff
+#define HaltAndStopMask                0x000001ff
+
+
+#define DmaComQueueNotEmpty    (1<<0)
+#define ThreadComQueueNotEmpty (1<<1)
+#define EventComQueueNotEmpty  (1<<2)
+#define DmaComQueueHalfFull    (1<<3)
+#define ThreadComQueueHalfFull (1<<4)
+#define EventComQueueHalfFull  (1<<5)
+#define DmaComQueueError       (1<<6)
+#define ThreadComQueueError    (1<<7)
+#define EventComQueueError     (1<<8)
+
+#define ComQueueNotEmpty       (DmaComQueueNotEmpty | ThreadComQueueNotEmpty | EventComQueueNotEmpty)
+#define ComQueueError          (DmaComQueueError | ThreadComQueueError | EventComQueueError)
+
+typedef union _E3_DmaInfo
+{
+    E3_uint32  Value;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 DmaOutputOpen:1;      /* The packet is currently open */
+       E3_uint32 :7;
+       E3_uint32 TimeSliceCount:2;     /* Time left to timeslice */
+       E3_uint32 UseRemotePriv:1;      /* Set for remote read dmas */
+       E3_uint32 DmaLastPacket:1;      /* Set for the last packet of a dma */
+       E3_uint32 PacketAckValue:2;     /* Packet ack type. Valid if AckBufferValid set. */
+       E3_uint32 PacketTimeout:1;      /* Packet timeout. Sent an EopError. Valid if AckBufferValid set. */
+       E3_uint32 AckBufferValid:1;     /* Packet ack is valid. */
+       E3_uint32 :16;                  /* read as Zero */
+#else
+       E3_uint32 :16;                  /* read as Zero */
+       E3_uint32 AckBufferValid:1;     /* Packet ack is valid. */
+       E3_uint32 PacketTimeout:1;      /* Packet timeout. Sent an EopError. Valid if AckBufferValid set. */
+       E3_uint32 PacketAckValue:2;     /* Packet ack type. Valid if AckBufferValid set. */
+       E3_uint32 DmaLastPacket:1;      /* Set for the last packet of a dma */
+       E3_uint32 UseRemotePriv:1;      /* Set for remote read dmas */
+       E3_uint32 TimeSliceCount:2;     /* Time left to timeslice */
+       E3_uint32 :7;
+       E3_uint32 DmaOutputOpen:1;      /* The packet is currently open */
+#endif
+    } s;
+} E3_DmaInfo;
+
+typedef volatile struct _E3_DmaRds
+{
+   E3_uint32           DMA_Source4to0AndTwoReads;
+   E3_uint32           pad13;
+   E3_uint32           DMA_BytesToRead;
+   E3_uint32           pad14;
+   E3_uint32           DMA_MinusPacketSize;
+   E3_uint32           pad15;
+   E3_uint32           DMA_MaxMinusPacketSize;
+   E3_uint32           pad16;
+   E3_uint32           DMA_DmaOutputOpen;
+   E3_uint32           pad16a;
+   E3_DmaInfo          DMA_PacketInfo;
+   E3_uint32           pad17[7];
+   E3_uint32           IProcTrapBase;
+   E3_uint32           pad18;
+   E3_uint32           IProcBlockTrapBase;
+   E3_uint32           pad19[11];
+} E3_DmaRds;
+   
+typedef volatile struct _E3_DmaWrs
+{
+   E3_uint64           pad0;
+   E3_uint64           LdAlignment;
+   E3_uint64           ResetAckNLdBytesToWr;
+   E3_uint64           SetAckNLdBytesToWr;
+   E3_uint64           LdBytesToRd;
+   E3_uint64           LdDmaType;
+   E3_uint64           SendRoutes;
+   E3_uint64           SendEop;
+   E3_uint64           pad1[8];
+} E3_DmaWrs;
+
+typedef volatile struct _E3_Exts
+{
+   E3_uint32           CurrContext;                            /* 0x12a00 */
+   E3_uint32           pad0;
+   E3_Status_Reg       DProcStatus;                            /* 0x12a08 */
+   E3_uint32           pad1;
+   E3_Status_Reg       CProcStatus;                            /* 0x12a10 */
+   E3_uint32           pad2;
+   E3_Status_Reg       TProcStatus;                            /* 0x12a18 */
+   E3_uint32           pad3;
+   E3_IProcStatus_Reg  IProcStatus;                            /* 0x12a20 */
+   E3_uint32           pad4[3];
+
+   E3_uint32           IProcTypeContext;                       /* 0x12a30 */
+   E3_uint32           pad5;
+   E3_uint32           IProcTransAddr;                         /* 0x12a38 */
+   E3_uint32           pad6;
+   E3_uint32           IProcCurrTransData0;                    /* 0x12a40 */
+   E3_uint32           pad7;
+   E3_uint32           IProcCurrTransData1;                    /* 0x12a48 */
+   E3_uint32           pad8;
+
+   E3_uint32           SchCntReg;                              /* 0x12a50 */
+   E3_uint32           pad9;
+   E3_uint32           InterruptReg;                           /* 0x12a58 */
+   E3_uint32           pad10;
+   E3_uint32           InterruptMask;                          /* 0x12a60 */
+   E3_uint32           pad11;
+   E3_uint32           LinkErrorTypes;                         /* 0x12a68 */
+   E3_uint32           pad12[3];
+   E3_uint32           LinkState;      /* a read here returens the DataDel value for the */
+                                       /* link that has just been defined by a write to */
+                                       /* Regs.Exts.SchCntReg.LinkSetValue */
+   E3_uint32           pad13;
+
+   union                                                       /* 0x12a80 */
+   {
+      E3_DmaWrs                DmaWrs;
+      E3_DmaRds                DmaRds;
+   } Dmas;
+} E3_Exts;
+
+typedef union com_port_entry
+{
+    E3_uint64  type;
+    struct
+    {
+       E3_uint32 Address;              /* Command VAddr */
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 Context0Issue:1;      /* Issue was for context 0 */
+       E3_uint32 EventNotCommand:1;    /* Issue address bit 3 */
+       E3_uint32 RemoteDesc:1;         /* Issue address bit 5 */
+       E3_uint32 :13;                  /* read as Zero */
+       E3_uint32 Context:12;           /* Command Context */
+       E3_uint32 :4;                   /* read as Zero */
+#else
+       E3_uint32 :4;                   /* read as Zero */
+       E3_uint32 Context:12;           /* Command Context */
+       E3_uint32 :13;                  /* read as Zero */
+       E3_uint32 RemoteDesc:1;         /* Issue address bit 5 */
+       E3_uint32 EventNotCommand:1;    /* Issue address bit 3 */
+       E3_uint32 Context0Issue:1;      /* Issue was for context 0 */
+#endif
+    } s;
+} E3_ComPortEntry;
+
+/* control reg bits */
+#define CONT_MMU_ENABLE                (1 << 0) /* bit 0 enables mmu */
+#define CONT_ENABLE_8K_PAGES   (1 << 1) /* When set smallest page is 8k instead of 4k. */
+#define CONT_EN_ALL_SETS       (1 << 2) /* enable cache */
+#define CONT_CACHE_LEVEL0      (1 << 3) /* cache context table */
+#define CONT_CACHE_LEVEL1      (1 << 4) /* cache up level 1 PTD/PTE */
+#define CONT_CACHE_LEVEL2      (1 << 5) /* cache up level 2 PTD/PTE */
+#define CONT_CACHE_LEVEL3      (1 << 6) /* cache up level 3 PTD/PTE */
+#define CONT_CACHE_TRAPS       (1 << 7) /* cache up traps */
+#define CONT_CACHE_LEV0_ROUTES (1 << 8) /* cache up small routes */
+#define CONT_CACHE_LEV1_ROUTES (1 << 9) /* cache up large routes */
+#define CONT_CACHE_ALL         (CONT_CACHE_LEVEL0 | CONT_CACHE_LEVEL1 | CONT_CACHE_LEVEL2 | \
+                                CONT_CACHE_LEVEL3 | CONT_CACHE_TRAPS | \
+                                CONT_CACHE_LEV0_ROUTES | CONT_CACHE_LEV1_ROUTES)
+
+#define CONT_SYNCHRONOUS       (1 << 10) /* PCI running sync */
+#define CONT_SER               (1 << 11) /* Single bit output (Elan1 SER bit) */
+#define CONT_SIR               (1 << 12) /* Writing 1 resets elan. */
+
+#define CONT_PSYCHO_MODE       (1 << 13) /* Enables all the perversion required by psycho */
+#define CONT_ENABLE_ECC                (1 << 14) /* Enables error detecting on the ECC */
+#define CONT_SDRAM_TESTING     (1 << 15) /* Switches to test mode for checking EEC data bits */
+
+/* defines SDRam CasLatency. Once set will not change again unless reset is reasserted. */
+/* 1 = Cas Latency is 3, 0 = Cas Latency is 2 */
+#define CAS_LATENCY_2          (0 << 16)
+#define CAS_LATENCY_3          (1 << 16)
+#define REFRESH_RATE_2US       (0 << 17) /* defines 2us SDRam Refresh rate. */
+#define REFRESH_RATE_4US       (1 << 17) /* defines 4us SDRam Refresh rate. */
+#define REFRESH_RATE_8US       (2 << 17) /* defines 8us SDRam Refresh rate. */
+#define REFRESH_RATE_16US      (3 << 17) /* defines 16us SDRam Refresh rate. */
+
+#define CONT_PCI_ERR           (1 << 19) /* Read 1 if PCI Error */
+#define CONT_CLEAR_PCI_ERROR   (1 << 19) /* Clears an PCI error. */
+
+/* Will cause the PCI error bit to become set. This is used to force the threads proc
+   and the uProc to start to stall. */
+#define CONT_SET_PCI_ERROR     (1 << 20)
+
+/* Writes SDram control reg when set. Also starts SDram memory system refreshing. */
+#define SETUP_SDRAM            (1 << 21)
+
+/* Flushes the tlb */
+#define MMU_FLUSH              (1 << 22)
+/* and read back when it's finished */
+#define MMU_FLUSHED            (1 << 0)
+
+/* Clears any ECC error detected by SDRam interface */
+#define CLEAR_SDRAM_ERROR      (1 << 23)
+
+#define ECC_ADDR_MASK          0x0ffffff8
+#define ECC_UE_MASK            0x1 
+#define ECC_CE_MASK            0x2
+#define ECC_ME_MASK            0x4 
+#define ECC_SYN_MASK           0xff
+
+/* define page table entry bit fields */
+#define TLB_PageSizeBits       (3 << 0)
+#define TLB_ACCBits            (7 << 2)
+#define TLB_LocalBit           (1 << 5)
+#define TLB_PCI64BitTargetBit  (1 << 6)
+#define TLB_PCIBigEndianBit    (1 << 7)
+
+#define TLB_ModifiedBit                (1 << 55)
+#define TLB_ReferencedBit      (1 << 63)
+
+/* Used to read values from the tlb. */
+#define TLB_TlbReadCntBitsSh   56
+#define TLB_UseSelAddrSh       (1ULL << 60)
+#define TLB_WriteTlbLine       (1ULL << 61)
+
+#define TLB_SEL_LINE(LineNo) (TLB_UseSelAddrSh | \
+                             ((E3_uint64)((LineNo) & 0xf) << TLB_TlbReadCntBitsSh))
+
+typedef union _E3_CacheContReg
+{
+    E3_uint32 ContReg;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 MMU_Enable:1;         /* wr 1 to enable the MMU */
+       E3_uint32 Set8kPages:1;         /* wr 1 smallest page is 8k. */
+       E3_uint32 EnableAllSets:1;      /* wr 1 All the cache sets are enabled */
+       E3_uint32 Cache_Level0:1;       /* wr 1 lev0 page tabs will be cached */
+       E3_uint32 Cache_Level1:1;       /* wr 1 lev1 page tabs will be cached */
+       E3_uint32 Cache_Level2:1;       /* wr 1 lev2 page tabs will be cached */
+       E3_uint32 Cache_Level3:1;       /* wr 1 lev3 page tabs will be cached */
+       E3_uint32 Cache_Traps:1;        /* wr 1 trap info will be cached */
+       E3_uint32 Cache_Lev0_Routes:1;  /* wr 1 small routes will be cached */
+       E3_uint32 Cache_Lev1_Routes:1;  /* wr 1 big routes will be cached */
+       E3_uint32 PCI_Synchronous:1;    /* Pci and sys clocks are running synchronously*/
+       E3_uint32 SER:1;                /* 1 bit output port */
+       E3_uint32 SIR:1;                /* write 1 will reset elan */
+       E3_uint32 PsychoMode:1;         /* Enables psycho perversion mode. */
+       E3_uint32 CasLatency:1;         /* 1=cas latency=3, 1=cas latency=2 */
+       E3_uint32 RefreshRate:2;        /* 0=2us, 1=4us, 2=8us, 3=16us */
+       E3_uint32 Pci_Err:1;            /* pci error. Write 1 clears err */
+       E3_uint32 Set_Pci_Error:1;      /* Will simulate an Pci error */
+       E3_uint32 StartSDRam:1;         /* Starts the sdram subsystem */
+       E3_uint32 FlushTlb:1;           /* Flush the contence of the tlb */
+       E3_uint32 :11;
+#else
+       E3_uint32 :11;
+       E3_uint32 FlushTlb:1;           /* Flush the contence of the tlb */
+       E3_uint32 StartSDRam:1;         /* Starts the sdram subsystem */
+       E3_uint32 Set_Pci_Error:1;      /* Will simulate an Pci error */
+       E3_uint32 Pci_Err:1;            /* pci error. Write 1 clears err */
+       E3_uint32 RefreshRate:2;        /* 0=2us, 1=4us, 2=8us, 3=16us */
+       E3_uint32 CasLatency:1;         /* 1=cas latency=3, 1=cas latency=2 */
+       E3_uint32 PsychoMode:1;         /* Enables psycho perversion mode. */
+       E3_uint32 SIR:1;                /* write 1 will reset elan */
+       E3_uint32 SER:1;                /* 1 bit output port */
+       E3_uint32 PCI_Synchronous:1;    /* Pci and sys clocks are running synchronously*/
+       E3_uint32 Cache_Lev1_Routes:1;  /* wr 1 big routes will be cached */
+       E3_uint32 Cache_Lev0_Routes:1;  /* wr 1 small routes will be cached */
+       E3_uint32 Cache_Traps:1;        /* wr 1 trap info will be cached */
+       E3_uint32 Cache_Level3:1;       /* wr 1 lev3 page tabs will be cached */
+       E3_uint32 Cache_Level2:1;       /* wr 1 lev2 page tabs will be cached */
+       E3_uint32 Cache_Level1:1;       /* wr 1 lev1 page tabs will be cached */
+       E3_uint32 Cache_Level0:1;       /* wr 1 lev0 page tabs will be cached */
+       E3_uint32 EnableAllSets:1;      /* wr 1 All the cache sets are enabled */
+       E3_uint32 Set8kPages:1;         /* wr 1 smallest page is 8k. */
+       E3_uint32 MMU_Enable:1;         /* wr 1 to enable the MMU */
+#endif
+    } s;
+} E3_CacheContReg;
+
+typedef union _E3_TrapBits
+{
+    volatile E3_uint32 Bits;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 ForcedTProcTrap:1;     /* The theads proc has been halted */
+       E3_uint32 InstAccessException:1; /* An instruction access exception */
+       E3_uint32 Unimplemented:1;       /* Unimplemented instruction executed */
+       E3_uint32 DataAccessException:1; /* A data access exception */  
+
+       E3_uint32 ThreadTimeout:1;       /* The threads outputer has timed out */
+       E3_uint32 OpenException:1;       /* Invalid sequence of open, sendtr or close */
+       E3_uint32 OpenRouteFetch:1;      /* Fault while fetching routes for previous open*/
+       E3_uint32 TrapForTooManyInsts:1; /* Thread has been executing for too long */
+       
+       E3_uint32 PacketAckValue:2;      /* Packet ack type. Valid if AckBufferValid set. */
+       E3_uint32 PacketTimeout:1;       /* Packet timeout. Sent an EopError. Valid if AckBufferValid set. */
+
+       E3_uint32 AckBufferValid:1;      /* The PacketAckValue bits are valid */
+       E3_uint32 OutputWasOpen:1;       /* The output was open when tproc trapped */
+       E3_uint32 TProcDeschedule:2;     /* The reason the tproc stopped running. */
+       E3_uint32 :17;
+#else
+       E3_uint32 :17;
+       E3_uint32 TProcDeschedule:2;     /* The reason the tproc stopped running. */
+       E3_uint32 OutputWasOpen:1;       /* The output was open when tproc trapped */
+       E3_uint32 AckBufferValid:1;      /* The PacketAckValue bits are valid */
+       
+       E3_uint32 PacketTimeout:1;       /* Packet timeout. Sent an EopError. Valid if AckBufferValid set. */
+       E3_uint32 PacketAckValue:2;      /* Packet ack type. Valid if AckBufferValid set. */
+       
+       E3_uint32 TrapForTooManyInsts:1; /* Thread has been executing for too long */
+       E3_uint32 OpenRouteFetch:1;      /* Fault while fetching routes for previous open*/
+       E3_uint32 OpenException:1;       /* Invalid sequence of open, sendtr or close */
+       E3_uint32 ThreadTimeout:1;       /* The threads outputer has timed out */
+
+       E3_uint32 DataAccessException:1; /* A data access exception */
+       E3_uint32 Unimplemented:1;       /* Unimplemented instruction executed */
+       E3_uint32 InstAccessException:1; /* An instruction access exception */
+       E3_uint32 ForcedTProcTrap:1;     /* The theads proc has been halted */
+#endif
+    } s;
+} E3_TrapBits;
+
+typedef union _E3_DirtyBits
+{
+    volatile E3_uint32 Bits;
+    struct
+    {
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 GlobalsDirty:8;
+       E3_uint32 OutsDirty:8;          /* will always read as dirty. */
+       E3_uint32 LocalsDirty:8;
+       E3_uint32 InsDirty:8;
+#else
+       E3_uint32 InsDirty:8;
+       E3_uint32 LocalsDirty:8;
+       E3_uint32 OutsDirty:8;          /* will always read as dirty. */
+       E3_uint32 GlobalsDirty:8;
+#endif
+    } s;
+} E3_DirtyBits;
+
+#define E3_TProcDescheduleMask    0x6000
+#define E3_TProcDescheduleWait    0x2000
+#define E3_TProcDescheduleSuspend 0x4000
+#define E3_TProcDescheduleBreak   0x6000
+
+#define E3_TrapBitsMask          0x7fff
+
+#define ThreadRestartFromTrapBit       1
+#define ThreadReloadAllRegs            2
+
+#define E3_PAckOk      0
+#define E3_PAckTestFail        1
+#define E3_PAckDiscard 2
+#define E3_PAckError   3
+
+typedef volatile struct _E3_DataBusMap
+{
+   E3_uint64            Dma_Alignment_Port[8];         /* 0x00002800 */
+   E3_uint32            pad0[0x30];                    /* 0x00002840 */
+
+   E3_uint32            Input_Trans0_Data[0x10];       /* 0x00002900 */
+   E3_uint32            Input_Trans1_Data[0x10];
+   E3_uint32            Input_Trans2_Data[0x10];
+   E3_uint32            Input_Trans3_Data[0x10];
+
+/* this is the start of the exts directly addressable from the ucode. */
+   E3_Exts              Exts;                          /* 0x00002a00 */
+
+/* this is the start of the registers directly addressable from the ucode. */
+   E3_DMA               Dma_Desc;                      /* 0x00002b00 */
+
+   E3_uint32            Dma_Last_Packet_Size;          /* 0x00002b20 */
+   E3_uint32            Dma_This_Packet_Size;          /* 0x00002b24 */
+   E3_uint32            Dma_Tmp_Source;                /* 0x00002b28 */
+   E3_uint32            Dma_Tmp_Dest;                  /* 0x00002b2c */
+
+   E3_Addr              Thread_SP_Save_Ptr;    /* points to the thread desched save word. */
+   E3_uint32            Dma_Desc_Size_InProg;          /* 0x00002b34 */
+
+   E3_uint32            Thread_Desc_SP;                /* 0x00002b38 */
+   E3_uint32            Thread_Desc_Context;           /* 0x00002b3c */
+
+   E3_uint32            uCode_TMP[0x10];               /* 0x00002b40 */
+
+   E3_uint32            TProc_NonSysCntx_FPtr;         /* 0x00002b80 */
+   E3_uint32            TProc_NonSysCntx_BPtr;         /* 0x00002b84 */
+   E3_uint32            TProc_SysCntx_FPtr;            /* 0x00002b88 */
+   E3_uint32            TProc_SysCntx_BPtr;            /* 0x00002b8c */
+   E3_uint32            DProc_NonSysCntx_FPtr;         /* 0x00002b90 */
+   E3_uint32            DProc_NonSysCntx_BPtr;         /* 0x00002b94 */
+   E3_uint32            DProc_SysCntx_FPtr;            /* 0x00002b98 */
+   E3_uint32            DProc_SysCntx_BPtr;            /* 0x00002b9c */
+
+   E3_uint32            Input_Trap_Base;               /* 0x00002ba0 */
+   E3_uint32            Input_Queue_Offset;            /* 0x00002ba4 */
+   E3_uint32            CProc_TrapSave_Addr;           /* 0x00002ba8 */
+   E3_uint32            Input_Queue_Addr;              /* 0x00002bac */
+   E3_uint32            uCode_TMP10;                   /* 0x00002bb0 */
+   E3_uint32            uCode_TMP11;                   /* 0x00002bb4 */
+   E3_uint32            Event_Trace_Ptr;               /* 0x00002bb8 */
+   E3_uint32            Event_Trace_Mask;              /* 0x00002bbc */
+
+   E3_ComPortEntry      DmaComQueue[3];                /* 0x00002bc0 */
+
+   E3_uint32            Event_Int_Queue_FPtr;          /* 0x00002bd8 */
+   E3_uint32            Event_Int_Queue_BPtr;          /* 0x00002bdc */
+
+   E3_ComPortEntry      ThreadComQueue[2];             /* 0x00002be0 */
+   E3_ComPortEntry      SetEventComQueue[2];           /* 0x00002bf0 */
+
+   E3_uint32            pad1[96];                      /* 0x00002c00 */
+   E3_uint32            ComQueueStatus;                /* 0x00002d80 */
+   E3_uint32            pad2[31];                      /* 0x00002d84 */
+
+/* These are the internal registers of the threads proc. */
+   E3_uint32            Globals[8];                    /* 0x00002e00 */
+   E3_uint32            Outs[8];
+   E3_uint32            Locals[8];
+   E3_uint32            Ins[8];
+
+   E3_uint32            pad3[16];
+
+   E3_uint32            IBufferReg[4];
+
+   E3_uint32            ExecuteNPC;
+   E3_uint32            ExecutePC;
+
+   E3_uint32            StartPC;
+   E3_uint32            pad4;
+
+   E3_uint32            StartnPC;
+   E3_uint32            pad5;
+
+   E3_TrapBits          TrapBits;
+   E3_DirtyBits                 DirtyBits;
+   E3_uint64            LoadDataReg;
+   E3_uint64            StoreDataReg;
+
+   E3_uint32            ECC_STATUS0;
+   E3_uint32            ECC_STATUS1;
+   E3_uint32            pad6[0xe];
+
+/* Pci slave port regs */
+   E3_uint32            PciSlaveReadCache[0x10];
+
+   E3_uint32            Fault_Base_Ptr;
+   E3_uint32            pad7;
+   E3_uint32            Context_Ptr;
+   E3_uint32            pad8;
+   E3_uint32            Input_Context_Filter;      /* write only, No data */
+   E3_uint32            Input_Context_Fil_Flush;   /* write only, No data */
+   E3_CacheContReg      Cache_Control_Reg;
+   E3_uint32            pad9;
+
+   E3_uint64            Tlb_Line_Value;
+   
+   E3_uint32            Walk_Datareg1;
+   E3_uint32            Walk_VAddr_Tab_Base;
+   E3_uint32            Walk_Datareg;
+   E3_uint32            Walk_ContextReg;
+   E3_uint32            Walk_FaultAddr;
+   E3_uint32            Walk_EventAddr;
+
+/* outputers output cont ext registers. */
+   E3_uint64            Dma_Route_012345_Context;
+   E3_uint64            pad10;
+   E3_uint64            Dma_Route_01234567;
+   E3_uint64            Dma_Route_89ABCDEF;
+
+   E3_uint64            Thread_Route_012345_Context;
+   E3_uint64            pad11;
+   E3_uint64            Thread_Route_01234567;
+   E3_uint64            Thread_Route_89ABCDEF;
+} E3_DataBusMap;
+
+typedef volatile struct _E3_Regs
+{
+   E3_CacheSets                  Sets;                         /* 0x00000000 */
+   E3_CacheTags                  Tags;                         /* 0x00002000 */
+   E3_DataBusMap         Regs;                         /* 0x00002800 */
+   E3_uint32             pad1[0x400];
+   E3_User_Regs          URegs;
+} E3_Regs;
+
+#define MAX_TRAPPED_TRANS      16
+#define TRANS_DATA_WORDS       16
+#define TRANS_DATA_BYTES       64
+
+/*
+ * Event interrupt
+ */
+typedef volatile union _E3_EventInt
+{
+   E3_uint64    ForceAlign;
+   struct {
+       E3_uint32 IntCookie;
+       E3_uint32 EventContext; /* Bits 16 to 28 */
+    } s;
+} E3_EventInt;
+
+#define GET_EVENT_CONTEXT(Ptr) ((Ptr->s.EventContext >> 16) & MAX_ROOT_CONTEXT_MASK)
+
+typedef volatile union _E3_ThreadQueue
+{
+   E3_uint64   ForceAlign;
+   struct
+   {
+       E3_Addr  Thread;
+#if defined(__LITTLE_ENDIAN__)
+       E3_uint32 :16;          /* Bits 0  to 15 */
+       E3_uint32 Context:13;   /* Bits 16 to 28 */
+       E3_uint32 :3;           /* Bits 29 to 31 */
+#else
+       E3_uint32 :3;           /* Bits 29 to 31 */
+       E3_uint32 Context:13;   /* Bits 16 to 28 */
+       E3_uint32 :16;          /* Bits 0  to 15 */
+#endif
+   } s;
+} E3_ThreadQueue;
+
+typedef volatile union _E3_FaultStatusReg
+{
+   E3_uint32 Status;
+   struct
+   {
+#if defined(__LITTLE_ENDIAN__)
+      E3_uint32 AccTypePerm:3; /* Access permission. See below. Bits 0 to 2 */
+      E3_uint32 AccSize:4;     /* Access size. See below for different types. Bits 3 to 6 */
+      E3_uint32 WrAcc:1;       /* Access was a write. Bit 7 */
+      E3_uint32 NonAllocAcc:1; /* Access was a cache non allocate type. Bit 8 */
+      E3_uint32 BlkDataType:2; /* Data size used for endian flips. Bits 9 to 10 */
+      E3_uint32 RdLine:1;      /* Access was a dma read line. Bit 11 */
+      E3_uint32 RdMult:1;      /* Access was a dma read multiple. Bit 12 */
+      E3_uint32 Walking:1;     /* The fault occued when walking. Bit 13 */
+      E3_uint32 Level:2;       /* Page table level when the fault occued. Bits 14 to 15 */
+      E3_uint32 ProtFault:1;   /* A protection fault occured. Bit 16 */
+      E3_uint32 FaultPte:2;    /* Page table type when the fault occured. Bit 17 */
+      E3_uint32 AlignmentErr:1;        /* Address alignment did not match the access size. Bit 19 */
+      E3_uint32 VProcSizeErr:1;        /* VProc number is out of range. Bit 20 */
+      E3_uint32 WalkBadData:1; /* Memory CRC error during a walk. Bit 21 */
+      E3_uint32 :10;           /* Bits 22 to 31 */
+#else
+      E3_uint32 :10;           /* Bits 22 to 31 */
+      E3_uint32 WalkBadData:1; /* Memory CRC error during a walk. Bit 21 */
+      E3_uint32 VProcSizeErr:1;        /* VProc number is out of range. Bit 20 */
+      E3_uint32 AlignmentErr:1;        /* Address alignment did not match the access size. Bit 19 */
+      E3_uint32 FaultPte:2;    /* Page table type when the fault occured. Bit 17 */
+      E3_uint32 ProtFault:1;   /* A protection fault occured. Bit 16 */
+      E3_uint32 Level:2;       /* Page table level when the fault occued. Bits 14 to 15 */
+      E3_uint32 Walking:1;     /* The fault occued when walking. Bit 13 */
+      E3_uint32 RdMult:1;      /* Access was a dma read multiple. Bit 12 */
+      E3_uint32 RdLine:1;      /* Access was a dma read line. Bit 11 */
+      E3_uint32 BlkDataType:2; /* Data size used for endian flips. Bits 9 to 10 */
+      E3_uint32 NonAllocAcc:1; /* Access was a cache non allocate type. Bit 8 */
+      E3_uint32 WrAcc:1;       /* Access was a write. Bit 7 */
+      E3_uint32 AccSize:4;     /* Access size. See below for different types. Bits 3 to 6 */
+      E3_uint32 AccTypePerm:3; /* Access permission. See below. Bits 0 to 2 */
+#endif
+   } s;
+} E3_FaultStatusReg;
+
+typedef union _E3_FaultSave
+{
+   E3_uint64            ForceAlign;
+   struct {
+      E3_FaultStatusReg         FSR;
+      volatile E3_uint32 FaultContext;
+      volatile E3_uint32 FaultAddress;
+      volatile E3_uint32 EventAddress;
+   } s;
+} E3_FaultSave;
+
+/* MMU fault status reg bit positions. */
+#define FSR_WritePermBit       0       /* 1=Write access perm, 0=Read access perm */
+#define FSR_RemotePermBit      1       /* 1=Remote access perm, 0=local access perm */
+#define FSR_EventPermBit       2       /* 1=Event access perm, 0=data access perm */
+#define FSR_Size0Bit           3
+#define FSR_Size1Bit           4
+#define FSR_Size2Bit           5
+#define FSR_Size3Bit           6
+#define FSR_WriteAccBit                7       /* 1=Write access, 0=Read access. */
+#define FSR_NonAllocBit                8       /* 1=Do not fill cache with this data */
+#define FSR_BlkDataTy0Bit      9
+#define FSR_BlkDataTy1Bit      10
+#define FSR_ReadLineBit                11
+#define FSR_ReadMultipleBit    12
+
+#define FSR_PermMask           (0xf << FSR_WritePermBit)
+#define FSR_SizeMask           (0xf << FSR_Size0Bit)
+#define FSR_AccTypeMask                (3 << FSR_WriteAccBit)
+#define FSR_BlkDataTyMask      (3 << FSR_BlkDataTy0Bit)
+#define FSR_PciAccTyMask       (3 << FSR_ReadLineBit)
+#define FSR_Walking            (0x1 << 13)
+#define FSR_Level_Mask         (0x3 << 14)
+#define FSR_ProtFault          (0x1 << 16)
+#define FSR_FaultPTEType       (0x2 << 17)
+#define FSR_AddrSizeError      (0x1 << 19)
+#define FSR_VProcSizeError     (0x1 << 20)
+#define FSR_WalkBadData                (0x1 << 21)
+
+#define FSR_PermRead           0
+#define FSR_PermWrite          1
+#define FSR_PermRemoteRead     2
+#define FSR_PermRemoteWrite    3
+#define FSR_PermEventRd                4
+#define FSR_PermEventWr                5
+#define FSR_PermRemoteEventRd  6
+#define FSR_PermRemoteEventWr  7
+
+/* AT size values for each access type */
+#define FSR_Word               (0x0 << FSR_Size0Bit)
+#define FSR_DWord              (0x1 << FSR_Size0Bit)
+#define FSR_QWord              (0x2 << FSR_Size0Bit)
+#define FSR_Block32            (0x3 << FSR_Size0Bit)
+#define FSR_ReservedBlock      (0x6 << FSR_Size0Bit)
+#define FSR_Block64            (0x7 << FSR_Size0Bit)
+#define FSR_GetCntxFilter      (0x8 << FSR_Size0Bit)
+#define FSR_QueueDWord         (0x9 << FSR_Size0Bit)
+#define FSR_RouteFetch         (0xa << FSR_Size0Bit)
+#define FSR_QueueBlock         (0xb << FSR_Size0Bit)
+#define FSR_Block32PartWrite   (0xe << FSR_Size0Bit)
+#define FSR_Block64PartWrite   (0xf << FSR_Size0Bit)
+
+#define FSR_AllocRead          (0 << FSR_WriteAccBit)
+#define FSR_AllocWrite         (1 << FSR_WriteAccBit)
+#define FSR_NonAllocRd         (2 << FSR_WriteAccBit)
+#define FSR_NonAllocWr         (3 << FSR_WriteAccBit)
+
+#define FSR_TypeByte           (0 << FSR_BlkDataTy0Bit)
+#define FSR_TypeHWord          (1 << FSR_BlkDataTy0Bit)
+#define FSR_TypeWord           (2 << FSR_BlkDataTy0Bit)
+#define FSR_TypeDWord          (3 << FSR_BlkDataTy0Bit)
+
+typedef union E3_TrTypeCntx
+{
+   E3_uint32 TypeContext;
+   struct
+   {
+#if defined(__LITTLE_ENDIAN__)
+      E3_uint32 Type:16;               /* Transaction type field */
+      E3_uint32 Context:13;            /* Transaction context */
+      E3_uint32 TypeCntxInvalid:1;     /* Bit  29 */
+      E3_uint32 StatusRegValid:1;      /* Bit  30 */
+      E3_uint32 LastTrappedTrans:1;    /* Bit  31 */
+#else
+      E3_uint32 LastTrappedTrans:1;    /* Bit  31 */
+      E3_uint32 StatusRegValid:1;      /* Bit  30 */
+      E3_uint32 TypeCntxInvalid:1;     /* Bit  29 */
+      E3_uint32 Context:13;            /* Transaction context */
+      E3_uint32 Type:16;               /* Transaction type field */
+#endif
+   } s;
+} E3_TrTypeCntx;
+
+#define GET_TRAP_TYPE(Ptr)    (Ptr.TypeContext & 0xfff)
+#define GET_TRAP_CONTEXT(Ptr) ((Ptr.TypeContext >> 16) & 0x1fff)
+
+/* Words have been swapped for big endian access when fetched with dword access from elan.*/
+typedef union _E3_IprocTrapHeader
+{
+   E3_uint64   forceAlign;
+
+   struct
+   {
+      E3_TrTypeCntx     TrTypeCntx;
+      E3_uint32                 TrAddr;
+      E3_uint32                 TrData0;
+      union
+      {
+        E3_IProcStatus_Reg u_IProcStatus;
+        E3_uint32          u_TrData1;
+      } ipsotd;
+   } s;
+} E3_IprocTrapHeader;
+
+#define IProcTrapStatus ipsotd.u_IProcStatus
+#define TrData1                ipsotd.u_TrData1
+
+typedef struct E3_IprocTrapData
+{
+   E3_uint32 TrData[TRANS_DATA_WORDS];
+} E3_IprocTrapData;
+
+/*
+ * 64 kbytes of elan local memory. Must be aligned on a 64k boundary
+ */
+#define E3_NonSysCntxQueueSize 0x400
+#define E3_SysCntxQueueSize    0x100
+
+typedef struct _E3_TrapAndQueue
+{
+   E3_DMA              NonSysCntxDmaQueue[E3_NonSysCntxQueueSize];                     /* 0x000000 */
+   E3_DMA              SysCntxDmaQueue[E3_SysCntxQueueSize];                           /* 0x008000 */
+   E3_EventInt         EventIntQueue[E3_NonSysCntxQueueSize];                          /* 0x00A000 */
+   E3_ThreadQueue      NonSysCntxThreadQueue[E3_NonSysCntxQueueSize];                  /* 0x00C000 */  
+   E3_ThreadQueue      SysCntxThreadQueue[E3_SysCntxQueueSize];                        /* 0x00E000 */
+   E3_FaultSave                IProcSysCntx;                                                   /* 0x00E800 */
+   E3_Addr             Thread_SP_Save;                                                 /* 0x00E810 */
+   E3_uint32           dummy0[3];                                                      /* 0x00E814 */
+   E3_FaultSave                ThreadProcData;                                                 /* 0x00E820 */
+   E3_FaultSave                ThreadProcInst;                                                 /* 0x00E830 */
+   E3_FaultSave                dummy1[2];                                                      /* 0x00E840 */  
+   E3_FaultSave                ThreadProcOpen;                                                 /* 0x00E860 */
+   E3_FaultSave                dummy2;                                                         /* 0x00E870 */
+   E3_FaultSave                IProcNonSysCntx;                                                /* 0x00E880 */
+   E3_FaultSave                DProc;                                                          /* 0x00E890 */
+   E3_FaultSave                CProc;                                                          /* 0x00E8A0 */
+   E3_FaultSave                TProc;                                                          /* 0x00E8B0 */
+   E3_FaultSave                DProcData0;                                                     /* 0x00E8C0 */
+   E3_FaultSave                DProcData1;                                                     /* 0x00E8D0 */
+   E3_FaultSave                DProcData2;                                                     /* 0x00E8E0 */
+   E3_FaultSave                DProcData3;                                                     /* 0x00E8F0 */
+   E3_uint32           dummy3[0xc0];                                                   /* 0x00E900 */
+   E3_IprocTrapHeader  VCh0_C0_TrHead[MAX_TRAPPED_TRANS];
+   E3_IprocTrapHeader  VCh0_NonC0_TrHead[MAX_TRAPPED_TRANS];
+   E3_IprocTrapHeader  VCh1_C0_TrHead[MAX_TRAPPED_TRANS];
+   E3_IprocTrapHeader  VCh1_NonC0_TrHead[MAX_TRAPPED_TRANS];
+   E3_IprocTrapData    VCh0_C0_TrData[MAX_TRAPPED_TRANS];
+   E3_IprocTrapData    VCh0_NonC0_TrData[MAX_TRAPPED_TRANS];
+   E3_IprocTrapData    VCh1_C0_TrData[MAX_TRAPPED_TRANS];
+   E3_IprocTrapData    VCh1_NonC0_TrData[MAX_TRAPPED_TRANS];
+   E3_uint64           DmaOverflowQueueSpace[0x1000];
+   E3_uint64           ThreadOverflowQueueSpace[0x800];
+   E3_uint64           EventOverflowQueueSpace[0x800];
+} E3_TrapAndQueue;
+
+
+typedef struct _E3_ContextControlBlock 
+{
+   E3_uint32   rootPTP;
+   E3_uint32   filter;
+   E3_uint32   VPT_ptr;
+   E3_uint32   VPT_mask;
+} E3_ContextControlBlock;
+
+#define E3_CCB_CNTX0           (0x20000000)
+#define E3_CCB_DISCARD_ALL     (0x40000000)
+#define E3_CCB_ACKOK_ALL       (0x80000000)
+#define E3_CCB_MASK            (0xc0000000)
+
+#define E3_NUM_CONTEXT_0       (0x20)
+
+/* Macros to manipulate event queue pointers */
+/*     generate index in EventIntQueue */
+#define E3_EVENT_INTQ_INDEX(fptr)      (((fptr) & 0x1fff) >> 3)
+/*     generate next fptr */
+#define E3_EVENT_INTQ_NEXT(fptr)       ((((fptr) + 8) & ~0x4000) | 0x2000)
+
+
+#endif /* notdef _ELAN3_ELANREGS_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elansyscall.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elansyscall.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elansyscall.h    2005-06-01 23:12:54.724420104 -0400
@@ -0,0 +1,124 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_ELANSYSCALL_H
+#define __ELAN3_ELANSYSCALL_H
+
+#ident "$Id: elansyscall.h,v 1.34 2004/06/07 13:50:06 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elansyscall.h,v $*/
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef _ASM
+
+typedef struct sys_word_item
+{
+    struct sys_word_item *Next;
+    E3_uint32            Value;
+} SYS_WORD_ITEM;
+
+typedef struct sys_block_item
+{
+    struct sys_block_item *Next;
+    E3_uint32            *Pointer;
+} SYS_BLOCK_ITEM;
+
+typedef struct sys_swap_space
+{
+    int                 Magic;
+    void       *ItemListsHead[MAX_LISTS];
+    void       **ItemListsTailp[MAX_LISTS];
+} SYS_SWAP_SPACE;
+
+typedef struct sys_exception
+{
+    int                        Type;
+    int                        Proc;
+    u_long             Res;
+    u_long             Value;
+    E3_FaultSave_BE    FaultArea;
+    
+    union
+    {
+       DMA_TRAP        Dma;
+       THREAD_TRAP     Thread;
+       COMMAND_TRAP    Command;
+       INPUT_TRAP      Input;
+    }                  Union;
+} SYS_EXCEPTION;
+
+typedef struct sys_exception_space
+{
+    struct sys_exception_space *Next;
+    int                                Magic;
+    int                                Front;
+    int                                Back;
+    int                                Count;
+    int                                Overflow;
+    SYS_EXCEPTION              Exceptions[1];
+} SYS_EXCEPTION_SPACE;
+
+#ifdef __KERNEL__
+
+typedef struct sys_ctxt
+{
+    SYS_SWAP_SPACE      *Swap;
+    SYS_EXCEPTION_SPACE *Exceptions;
+    kmutex_t            Lock;
+
+    spinlock_t          WaitLock;
+    kcondvar_t          NetworkErrorWait;
+
+    int                         Armed;
+    int                         Backoff;
+    long                Time;
+
+    u_long              Flags;
+    int                  signal;
+
+    EVENT_COOKIE_TABLE  *Table;
+} SYS_CTXT;
+
+extern SYS_CTXT *sys_init (ELAN3_CTXT *ctxt);
+extern int       sys_waitevent (ELAN3_CTXT *ctxt, E3_Event *event);
+extern void      sys_addException (SYS_CTXT *sctx, int type, int proc, caddr_t ptr, int size, 
+                                  E3_FaultSave_BE *, u_long res, u_long value);
+extern int       sys_getException (SYS_CTXT *sctx, SYS_EXCEPTION *ex);
+
+/* returns -ve error or ELAN_CAP_OK or ELAN_CAP_RMS */
+/* use = ELAN_USER_ATTACH, ELAN_USER_P2P, ELAN_USER_BROADCAST */
+extern int  elan3_validate_cap (ELAN3_DEV *dev, ELAN_CAPABILITY *cap ,int use);
+
+#endif /* __KERNEL__ */
+
+#endif /* _ASM */
+
+/* values for "Flags" */
+#define ELAN3_SYS_FLAG_DMA_BADVP               1
+#define ELAN3_SYS_FLAG_THREAD_BADVP    2
+#define ELAN3_SYS_FLAG_DMAFAIL         4
+#define ELAN3_SYS_FLAG_NETERR          8
+
+#define SYS_SWAP_MAGIC         0xB23C52DF
+#define SYS_EXCEPTION_MAGIC    0xC34D63E0
+
+#define EXCEPTION_GLOBAL_STRING        "elan3_exceptions"
+#define EXCEPTION_ABORT_STRING  "elan3_abortstring"
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __ELAN3_ELANSYSCALL_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elanuregs.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elanuregs.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elanuregs.h      2005-06-01 23:12:54.725419952 -0400
@@ -0,0 +1,295 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_ELANUREGS_H
+#define __ELAN3_ELANUREGS_H
+
+#ident "$Id: elanuregs.h,v 1.10 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elanuregs.h,v $*/
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Statistic control reg values
+ * Each 4-bit nibble of the control word specifies what statistic
+ * is to be recorded in each of the 8 statistic counters
+ */
+
+/* Count reg 0 */
+#define STC_INPUT_TRANSACTIONS         0
+#define STP_DMA_EOP_WAIT_ACK           1
+#define STP_THREAD_RUNNING             2
+#define STP_UCODE_WAIT_MEM             3
+#define STC_CACHE_WRITE_BACKS          4
+#define STC_PCI_SLAVE_READS            5
+#define STC_REG0_UNUSED6               6
+#define STP_REG0_UNUSED7               7
+
+#define STATS_REG0_NAMES {             \
+         "STC_INPUT_TRANSACTIONS",     \
+         "STP_DMA_EOP_WAIT_ACK",       \
+         "STP_THREAD_RUNNING",         \
+         "STP_UCODE_WAIT_MEM",         \
+         "STC_CACHE_WRITE_BACKS",      \
+         "STC_PCI_SLAVE_READS",        \
+         "STC_REG0_UNUSED6",           \
+         "STP_REG0_UNUSED7"            \
+}
+
+/* Count reg 1 */
+#define STC_INPUT_WRITE_BLOCKS         (0 << 4)
+#define STP_DMA_DATA_TRANSMITTING      (1 << 4)
+#define STP_THEAD_WAITING_INST         (2 << 4)
+#define STC_REG1_UNUSED3               (3 << 4)
+#define STP_FETCHING_ROUTES            (4 << 4)
+#define STC_REG1_UNUSED5               (5 << 4)
+#define STC_PCI_SLAVE_WRITES           (6 << 4)
+#define STP_PCI_SLAVE_READ_WAITING     (7 << 4)
+
+#define STATS_REG1_NAMES {             \
+      "STC_INPUT_WRITE_BLOCKS",                \
+         "STP_DMA_DATA_TRANSMITTING",  \
+         "STP_THEAD_WAITING_INST",     \
+         "STC_REG1_UNUSED3",           \
+         "STP_FETCHING_ROUTES",        \
+         "STC_REG1_UNUSED5",           \
+         "STC_PCI_SLAVE_WRITES",       \
+         "STP_PCI_SLAVE_READ_WAITING"  \
+}
+
+/* Count reg 2 */
+#define STC_INPUT_PKTS                 (0 << 8)
+#define STP_DMA_WAITING_MEM            (1 << 8)
+#define STP_THREAD_WAIT_OPEN_PKT       (2 << 8)
+#define STC_REG2_UNUSED3               (3 << 8)
+#define STC_ROUTE_FETCHES              (4 << 8)
+#define STC_CACHE_NON_ALLOC_MISSES     (5 << 8)
+#define STC_REG2_UNUSED6               (6 << 8)
+#define STP_PCI_SLAVE_WRITE_WAITING    (7 << 8)
+
+#define STATS_REG2_NAMES {             \
+      "STC_INPUT_PKTS",                        \
+         "STP_DMA_WAITING_MEM",        \
+         "STP_THREAD_WAIT_OPEN_PKT",   \
+         "STC_REG2_UNUSED3",           \
+         "STC_ROUTE_FETCHES",          \
+         "STC_CACHE_NON_ALLOC_MISSES", \
+         "STC_REG2_UNUSED6",           \
+         "STP_PCI_SLAVE_WRITE_WAITING" \
+}
+
+/* Count reg 3 */
+#define STC_INPUT_PKTS_REJECTED                (0 << 12)
+#define STP_DMA_WAIT_NETWORK_BUSY      (1 << 12)
+#define STP_THREAD_WAIT_PACK           (2 << 12)
+#define STP_UCODE_BLOCKED_UCODE                (3 << 12)
+#define STC_TLB_HITS                   (4 << 12)
+#define STC_REG3_UNUSED5               (5 << 12)
+#define STC_PCI_MASTER_READS           (6 << 12)
+#define STP_PCI_MASTER_WRITE_WAITING   (7 << 12)
+
+#define STATS_REG3_NAMES {             \
+      "STC_INPUT_PKTS_REJECTED",       \
+         "STP_DMA_WAIT_NETWORK_BUSY",  \
+         "STP_THREAD_WAIT_PACK",       \
+         "STP_UCODE_BLOCKED_UCODE",    \
+         "STC_TLB_HITS",               \
+         "STC_REG3_UNUSED5",           \
+         "STC_PCI_MASTER_READS",       \
+         "STP_PCI_MASTER_WRITE_WAITING"\
+}
+
+/* Count reg 4 */
+#define STP_INPUT_DATA_TRANSMITTING    (0 << 16)
+#define STC_DMA_NON_CTX0_PKTS          (1 << 16)
+#define STP_THREAD_EOP_WAIT_ACK                (2 << 16)
+#define STP_UCODE_DPROC_RUNNING                (3 << 16)
+#define STC_TLB_MEM_WALKS              (4 << 16)
+#define STC_REG4_UNUSED5               (5 << 16)
+#define STC_PCI_MASTER_WRITES          (6 << 16)
+#define STP_PCI_MASTER_READ_WAITING    (7 << 16)
+
+#define STATS_REG4_NAMES {             \
+      "STP_INPUT_DATA_TRANSMITTING",   \
+         "STC_DMA_NON_CTX0_PKTS",      \
+         "STP_THREAD_EOP_WAIT_ACK",    \
+         "STP_UCODE_DPROC_RUNNING",    \
+         "STC_TLB_MEM_WALKS",          \
+         "STC_REG4_UNUSED5",           \
+         "STC_PCI_MASTER_WRITES",      \
+         "STP_PCI_MASTER_READ_WAITING" \
+}
+
+/* Count reg 5 */
+#define STP_INPUT_WAITING_NETWORK_DATA (0 << 20)
+#define STC_DMA_NON_CTX0_PKTS_REJECTED (1 << 20)
+#define STP_THREAD_WAITING_DATA                (2 << 20)
+#define STP_UCODE_CPROC_RUNNING                (3 << 20)
+#define STP_THREAD_TRANSMITTING_DATA   (4 << 20)
+#define STP_PCI_WAITING_MAIN           (5 << 20)
+#define STC_REG5_UNUSED6               (6 << 20)
+#define STC_REG5_UNUSED7               (7 << 20)
+
+#define STATS_REG5_NAMES {                     \
+      "STP_INPUT_WAITING_NETWORK_DATA",                \
+         "STC_DMA_NON_CTX0_PKTS_REJECTED",     \
+         "STP_THREAD_WAITING_DATA",            \
+         "STP_UCODE_CPROC_RUNNING",            \
+         "STP_THREAD_TRANSMITTING_DATA",       \
+         "STP_PCI_WAITING_MAIN",               \
+         "STC_REG5_UNUSED6",                   \
+         "STC_REG5_UNUSED7"                    \
+}
+
+/* Count reg 6 */
+#define STP_INPUT_WAITING_MEMORY       (0 << 24)
+#define STC_DMA_CTX0_PKTS              (1 << 24)
+#define STP_THREAD_WAITING_MEMORY      (2 << 24)
+#define STP_UCODE_TPROC_RUNNING                (3 << 24)
+#define STC_CACHE_HITS                 (4 << 24)
+#define STP_PCI_WAITING_ELAN           (5 << 24)
+#define STC_REG6_UNUSED4               (6 << 24)
+#define STC_REG6_UNUSED7               (7 << 24)
+
+#define STATS_REG6_NAMES {             \
+      "STP_INPUT_WAITING_MEMORY",      \
+         "STC_DMA_CTX0_PKTS",          \
+         "STP_THREAD_WAITING_MEMORY",  \
+         "STP_UCODE_TPROC_RUNNING",    \
+         "STC_CACHE_HITS",             \
+         "STP_PCI_WAITING_ELAN",       \
+         "STC_REG6_UNUSED4",           \
+         "STC_REG6_UNUSED7"            \
+}
+
+/* Count reg 7 */
+#define STC_INPUT_CTX_FILTER_FILL      (0 << 28)       
+#define STC_DMA_CTX0_PKTS_REJECTED     (1 << 28)
+#define STP_THREAD_WAIT_NETWORK_BUSY   (2 << 28)
+#define STP_UCODE_IPROC_RUNNING                (3 << 28)
+#define STP_TLB_MEM_WALKING            (4 << 28)
+#define STC_CACHE_ALLOC_MISSES         (5 << 28)
+#define STP_PCI_DATA_TRANSFER          (6 << 28)
+#define STC_REG7_UNUSED7               (7 << 28)
+
+#define STATS_REG7_NAMES {             \
+      "STC_INPUT_CTX_FILTER_FILL",     \
+         "STC_DMA_CTX0_PKTS_REJECTED", \
+         "STP_THREAD_WAIT_NETWORK_BUSY",\
+         "STP_UCODE_IPROC_RUNNING",    \
+         "STP_TLB_MEM_WALKING",        \
+         "STC_CACHE_ALLOC_MISSES",     \
+         "STP_PCI_DATA_TRANSFER",      \
+         "STC_REG7_UNUSED7"            \
+}
+
+#define STATS_REG_NAMES { \
+    STATS_REG0_NAMES, \
+    STATS_REG1_NAMES, \
+    STATS_REG2_NAMES, \
+    STATS_REG3_NAMES, \
+    STATS_REG4_NAMES, \
+    STATS_REG5_NAMES, \
+    STATS_REG6_NAMES, \
+    STATS_REG7_NAMES, \
+}
+
+extern const char *elan3_stats_names[8][8];
+
+#define ELAN3_STATS_NAME(COUNT, CONTROL) (elan3_stats_names[(COUNT)][(CONTROL) & 7])
+
+typedef volatile union e3_StatsControl
+{
+   E3_uint32 StatsControl;
+   struct
+   {
+#if defined(__LITTLE_ENDIAN__)
+      E3_uint32 StatCont0:4;
+      E3_uint32 StatCont1:4;
+      E3_uint32 StatCont2:4;
+      E3_uint32 StatCont3:4;
+      E3_uint32 StatCont4:4;
+      E3_uint32 StatCont5:4;
+      E3_uint32 StatCont6:4;
+      E3_uint32 StatCont7:4;
+#else
+      E3_uint32 StatCont7:4;
+      E3_uint32 StatCont6:4;
+      E3_uint32 StatCont5:4;
+      E3_uint32 StatCont4:4;
+      E3_uint32 StatCont3:4;
+      E3_uint32 StatCont2:4;
+      E3_uint32 StatCont1:4;
+      E3_uint32 StatCont0:4;
+#endif
+   } s;
+} E3_StatsControl;
+
+typedef volatile union e3_StatsCount
+{
+   E3_uint64    ClockStat; 
+   struct
+   {
+       E3_uint32 ClockLSW;     /* read only */
+       E3_uint32 StatsCount;
+   } s;
+} E3_StatsCount;
+
+typedef volatile union e3_clock
+{
+   E3_uint64 NanoSecClock;
+   struct
+   {
+      E3_uint32 ClockLSW;
+      E3_uint32 ClockMSW;
+   } s;
+} E3_Clock;
+#define E3_TIME( X ) ((X).NanoSecClock)
+
+typedef volatile struct _E3_User_Regs
+{
+   E3_StatsCount       StatCounts[8];
+   E3_StatsCount       InstCount;
+   E3_uint32           pad0;
+   E3_StatsControl     StatCont;
+   E3_Clock            Clock;
+   E3_uint32           pad1[0x7ea];
+} E3_User_Regs;
+
+typedef volatile struct _E3_CommandPort 
+{
+   E3_Addr             PutDma;         /* 0x000 */
+   E3_uint32           Pad1;
+   E3_Addr             GetDma;         /* 0x008 */
+   E3_uint32           Pad2;
+   E3_Addr             RunThread;      /* 0x010 */
+   E3_uint32           Pad3[3];
+   E3_Addr             WaitEvent0;     /* 0x020 */
+   E3_uint32           Pad4;
+   E3_Addr             WaitEvent1;     /* 0x028 */
+   E3_uint32           Pad5;
+   E3_Addr             SetEvent;       /* 0x030 */
+   E3_uint32           Pad6[3];
+   E3_uint32           Pad7[0x7f0];    /* Fill out to an 8K page */
+} E3_CommandPort;
+/* Should have the new structures for the top four pages of the elan3 space */
+
+#define E3_COMMANDPORT_SIZE    (sizeof (E3_CommandPort))
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __ELAN3_ELANUREGS_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/elanvp.h
===================================================================
--- linux-2.4.21.orig/include/elan3/elanvp.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/elanvp.h 2005-06-01 23:12:54.726419800 -0400
@@ -0,0 +1,165 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_ELANVP_H
+#define _ELAN3_ELANVP_H
+
+#ident "$Id: elanvp.h,v 1.45 2004/06/18 09:28:06 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/elanvp.h,v $ */
+
+#include <elan3/e3types.h>
+#include <elan/bitmap.h>
+#include <elan/capability.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Context number allocation.
+ * [0-31]      system contexts
+ * [32-63]     hardware test
+ * [64-1023]   available
+ * [1024-2047] RMS allocatable
+ * [2048-4095] kernel comms data contexts
+ */
+#define ELAN3_KCOMM_CONTEXT_NUM                0x001                   /* old kernel comms context (system) */
+#define ELAN3_CM_CONTEXT_NUM           0x002                   /* new cluster member ship comms context (system) */
+#define ELAN3_MRF_CONTEXT_NUM          0x003                   /* multi-rail kernel comms context */
+#define ELAN3_DMARING_BASE_CONTEXT_NUM 0x010                   /* 16 contexts for dma ring issue (system) */
+#define ELAN3_DMARING_TOP_CONTEXT_NUM  0x01f
+
+#define ELAN3_HWTEST_BASE_CONTEXT_NUM  0x020                   /* reserved for hardware test */
+#define ELAN3_HWTEST_TOP_CONTEXT_NUM   0x03f
+
+#define ELAN3_KCOMM_BASE_CONTEXT_NUM   0x800                   /* kernel comms data transfer contexts */
+#define ELAN3_KCOMM_TOP_CONTEXT_NUM    0xfff
+
+#define ELAN3_HWTEST_CONTEXT(ctx)      ((ctx) >= ELAN3_HWTEST_BASE_CONTEXT_NUM && \
+                                        (ctx) <= ELAN3_HWTEST_TOP_CONTEXT_NUM)    
+
+#define ELAN3_SYSTEM_CONTEXT(ctx)      (((ctx) & SYS_CONTEXT_BIT) != 0 || \
+                                        (ctx) < E3_NUM_CONTEXT_0 || \
+                                        (ctx) >= ELAN3_KCOMM_BASE_CONTEXT_NUM)
+
+/* Maximum number of virtual processes */
+#define ELAN3_MAX_VPS          (16384)
+
+#define ELAN3_INVALID_PROCESS  (0x7fffffff)            /* A GUARANTEED invalid process # */
+#define ELAN3_INVALID_NODE     (0xFFFF)
+#define ELAN3_INVALID_CONTEXT  (0xFFFF)
+
+
+
+#if defined(__KERNEL__) && !defined(__ELAN3__)
+
+/*
+ * Contexts are accessible via Elan capabilities,
+ * for each context that can be "attached" to there
+ * is a ELAN3_CTXT_INFO structure created by its
+ * "owner".  This also "remembers" all remote 
+ * segments that have "blazed" a trail to it.
+ *
+ * If the "owner" goes away the soft info is 
+ * destroyed when it is no longer "attached" or 
+ * "referenced" by a remote segment.
+ *
+ * If the owner changes the capability, then 
+ * the soft info must be not "referenced" or 
+ * "attached" before a new process can "attach"
+ * to it.
+ */
+
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::InfoLock,
+                         elan3_info::Next elan3_info::Prev elan3_info::Device elan3_info::Owner
+                         elan3_info::Capability elan3_info::AttachedCapability elan3_info::Context))
+_NOTE(MUTEX_PROTECTS_DATA(elan3_dev::IntrLock,
+                         elan3_info::Nacking elan3_info::Disabled))
+_NOTE(DATA_READABLE_WITHOUT_LOCK(elan3_info::Context elan3_info::Device elan3_info::Capability))
+
+#endif /* __KERNEL__ */
+
+#define LOW_ROUTE_PRIORITY     0
+#define HIGH_ROUTE_PRIORITY    1
+
+#define DEFAULT_ROUTE_TIMEOUT  3
+#define DEFAULT_ROUTE_PRIORITY LOW_ROUTE_PRIORITY
+
+
+/* a small route is 4 flits (8 bytes), a big route  */
+/* is 8 flits (16 bytes) - each packed route is 4 bits */
+/* so giving us a maximum of 28 as flit0 does not contain */
+/* packed routes */
+#define MAX_FLITS              8
+#define MAX_PACKED             28
+
+/* bit definitions for 64 bit route pointer */
+#define ROUTE_VALID            (1ULL << 63)
+#define ROUTE_PTR              (1ULL << 62)
+#define ROUTE_CTXT_SHIFT       48
+#define ROUTE_PTR_MASK         ((1ull << ROUTE_CTXT_SHIFT)-1)
+#define ROUTE_GET_CTXT          ((VAL >> ROUTE_CTXT_SHIFT) & 0x3fff )
+
+#define SMALL_ROUTE(flits, context)    (((E3_uint64) (flits)[0] <<  0) | ((E3_uint64) (flits)[1] << 16) | \
+                                        ((E3_uint64) (flits)[2] << 32) | ((E3_uint64) (context) << ROUTE_CTXT_SHIFT) | \
+                                        ROUTE_VALID)
+
+#define BIG_ROUTE_PTR(paddr, context)  ((E3_uint64) (paddr) | ((E3_uint64) context << ROUTE_CTXT_SHIFT) | ROUTE_VALID | ROUTE_PTR)
+
+#define BIG_ROUTE0(flits)              (((E3_uint64) (flits)[0] <<  0) | ((E3_uint64) (flits)[1] << 16) | \
+                                        ((E3_uint64) (flits)[2] << 32) | ((E3_uint64) (flits)[3] << 48))
+#define BIG_ROUTE1(flits)              (((E3_uint64) (flits)[4] <<  0) | ((E3_uint64) (flits)[5] << 16) | \
+                                        ((E3_uint64) (flits)[6] << 32) | ((E3_uint64) (flits)[7] << 48))
+
+
+/* defines for first flit of a route */
+#define FIRST_HIGH_PRI         (1 << 15)
+#define FIRST_AGE(Val)         ((Val) << 11)
+#define FIRST_TIMEOUT(Val)     ((Val) << 9)
+#define FIRST_PACKED(X)                ((X) << 7)
+#define FIRST_ROUTE(Val)       (Val)
+#define FIRST_ADAPTIVE         (0x30)
+#define FIRST_BCAST_TREE       (0x20)
+#define FIRST_MYLINK           (0x10)
+#define FIRST_BCAST(Top, Bot)  (0x40 | ((Top) << 3) | (Bot))
+
+/* defines for 3 bit packed entries for subsequent flits */
+#define PACKED_ROUTE(Val)      (8 | (Val))
+#define PACKED_ADAPTIVE                (3)
+#define PACKED_BCAST_TREE      (2)
+#define PACKED_MYLINK          (1)
+#define PACKED_BCAST0(Top,Bot) (4 | (Bot & 3))
+#define PACKED_BCAST1(Top,Bot) ((Top << 1) | (Bot >> 2))
+
+/* ---------------------------------------------------------- 
+ * elan3_route functions 
+ * return ELAN3_ROUTE_xxx codes
+ * ---------------------------------------------------------- */
+
+#define ELAN3_ROUTE_SUCCESS        (0x00)
+#define ELAN3_ROUTE_SYSCALL_FAILED (0x01)
+#define ELAN3_ROUTE_INVALID        (0x02)
+#define ELAN3_ROUTE_TOO_LONG       (0x04)
+#define ELAN3_ROUTE_LOAD_FAILED    (0x08)
+#define ELAN3_ROUTE_PROC_RANGE     (0x0f)
+#define ELAN3_ROUTE_INVALID_LEVEL  (0x10)
+#define ELAN3_ROUTE_OCILATES       (0x20)
+#define ELAN3_ROUTE_WRONG_DEST     (0x40)
+#define ELAN3_ROUTE_TURN_LEVEL     (0x80)
+#define ELAN3_ROUTE_NODEID_UNKNOWN (0xf0)
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _ELAN3_ELANVP_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/events.h
===================================================================
--- linux-2.4.21.orig/include/elan3/events.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/events.h 2005-06-01 23:12:54.726419800 -0400
@@ -0,0 +1,183 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_EVENTS_H
+#define _ELAN3_EVENTS_H
+
+#ident "$Id: events.h,v 1.45 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/events.h,v $*/
+
+/*
+ * Alignments for events, event queues and blockcopy blocks.
+ */
+#define E3_EVENT_ALIGN                 (8)
+#define E3_QUEUE_ALIGN         (32)
+#define E3_BLK_ALIGN           (64)
+#define E3_BLK_SIZE            (64)
+#define E3_BLK_PATTERN                 (0xfeedface)
+
+#define E3_EVENT_FREE          ((0 << 4) | EV_WCOPY)
+#define E3_EVENT_PENDING       ((1 << 4) | EV_WCOPY)
+#define E3_EVENT_ACTIVE                ((2 << 4) | EV_WCOPY)
+#define E3_EVENT_FIRED         ((3 << 4) | EV_WCOPY)
+#define E3_EVENT_FAILED                ((4 << 4) | EV_WCOPY)
+#define E3_EVENT_DONE          ((5 << 4) | EV_WCOPY)
+#define E3_EVENT_PRIVATE       ((6 << 4) | EV_WCOPY)
+
+/*
+ * Event values and masks
+ *
+ * Block Copy event    xxxxxxxxxxxxxxxx1
+ * Chained event       30 bit ptr ....0x
+ * Event interrupt     29 bit cookie 01x
+ * Dma event           28 bit ptr   011x
+ * thread event                28 bit ptr   111x
+ */
+#define EV_CLEAR               (0x00000000)
+#define EV_TYPE_BCOPY          (0x00000001)
+#define EV_TYPE_CHAIN          (0x00000000)
+#define EV_TYPE_EVIRQ          (0x00000002)
+#define EV_TYPE_DMA            (0x00000006)
+#define EV_TYPE_THREAD         (0x0000000e)
+
+#define EV_TYPE_BCOPY_BYTE     (0)
+#define EV_TYPE_BCOPY_HWORD    (1)
+#define EV_TYPE_BCOPY_WORD     (2)
+#define EV_TYPE_BCOPY_DWORD    (3)
+
+/*
+ * Data type is in the lowest two bits of the Dest pointer.
+ */
+#define EV_BCOPY_DTYPE_MASK    (3)
+#define EV_WCOPY               (1)     /* [DestWord] = Source */
+#define EV_BCOPY               (0)     /* [DestBlock] = [SourceBlock] */
+
+#define EV_TYPE_MASK           (0x0000000e)
+#define EV_TYPE_MASK_BCOPY     (0x00000001)
+#define EV_TYPE_MASK_CHAIN     (0x00000002)
+#define EV_TYPE_MASK_EVIRQ     (0x00000006)
+#define EV_TYPE_MASK_DMA       (0x0000000e)
+#define EV_TYPE_MASK_THREAD    (0x0000000e)
+#define EV_TYPE_MASK2          (0x0000000f)
+
+/*
+ * Min/Max size for Elan queue entries 
+ */
+#define E3_QUEUE_MIN   E3_BLK_SIZE
+#define E3_QUEUE_MAX   (E3_BLK_SIZE * 5)
+
+/*
+ * Elan queue state bits
+ */
+#define E3_QUEUE_FULL  (1<<0)
+#define E3_QUEUE_LOCKED        (1<<8)
+
+#ifndef _ASM
+
+typedef union _E3_Event
+{
+   E3_uint64   ev_Int64;
+   struct {
+      volatile E3_int32        u_Count;
+      E3_uint32                u_Type;
+   } ev_u;
+} E3_Event;
+
+typedef union _E3_BlockCopyEvent
+{
+   E3_uint64 ev_ForceAlign;
+   struct E3_BlockCopyEvent_u {
+      volatile E3_int32        u_Count;
+      E3_uint32                u_Type;
+      E3_Addr          u_Source;
+      E3_Addr          u_Dest;   /* lowest bits are the data type for endian conversion */
+   } ev_u;
+} E3_BlockCopyEvent;
+
+#define ev_Type   ev_u.u_Type
+#define ev_Count  ev_u.u_Count
+#define ev_Source ev_u.u_Source
+#define ev_Dest   ev_u.u_Dest
+
+typedef union _E3_WaitEvent0
+{
+   E3_uint64            we_ForceAlign;
+   struct {
+       E3_Addr         u_EventLoc;
+       E3_int32        u_WaitCount;
+   } we_u;
+} E3_WaitEvent0;
+#define we_EventLoc we_u.u_EventLoc
+#define we_WaitCount we_u.u_WaitCount
+
+typedef union _E3_Event_Blk
+{
+    E3_uint8  eb_Bytes[E3_BLK_SIZE];
+    E3_uint32 eb_Int32[E3_BLK_SIZE/sizeof (E3_uint32)];
+    E3_uint64 eb_Int64[E3_BLK_SIZE/sizeof (E3_uint64)];
+} E3_Event_Blk;
+
+/* We make eb_done the last word of the blk
+ * so that we can guarantee the rest of the blk is
+ * correct when this value is set.
+ * However, when the TPORT code copies the envelope
+ * info into the blk, it uses a dword endian type.
+ * Thus we must correct for this when initialising
+ * the pattern in the Elan SDRAM blk (eeb_done)
+ */
+#define eb_done eb_Int32[15]
+#define eeb_done eb_Int32[15^WordEndianFlip]
+
+#define EVENT_WORD_READY(WORD) (*((volatile E3_uint32 *) WORD) != 0)
+#define EVENT_BLK_READY(BLK) (((volatile E3_Event_Blk *) (BLK))->eb_done != 0)
+#define EVENT_READY(EVENT)   (((volatile E3_Event *) (EVENT))->ev_Count <= 0)
+
+#define ELAN3_WAIT_EVENT (0)
+#define ELAN3_POLL_EVENT (-1)
+
+#define SETUP_EVENT_TYPE(ptr,typeval) (((unsigned long)(ptr)) | (typeval))
+
+#define E3_RESET_BCOPY_BLOCK(BLK)                                                      \
+       do {                                                                            \
+               (BLK)->eb_done = 0;                                                     \
+       } while (0)
+
+typedef struct e3_queue
+{
+   volatile E3_uint32  q_state;        /* queue is full=bit0, queue is locked=bit8 */
+   volatile E3_Addr    q_bptr;         /* block aligned ptr to current back item */
+   E3_uint32           q_size;         /* size of queue item; 0x1 <= size <= (0x40 * 5) */
+   E3_Addr             q_top;          /* block aligned ptr to last queue item */
+   E3_Addr             q_base;         /* block aligned ptr to first queue item */
+   volatile E3_Addr    q_fptr;         /* block aligned ptr to current front item */
+   E3_Event            q_event;        /* queue event */
+} E3_Queue;
+
+typedef struct e3_blockcopy_queue
+{
+   volatile E3_uint32  q_state;        /* queue is full=bit0, queue is locked=bit8 */
+   volatile E3_Addr    q_bptr;         /* block aligned ptr to current back item */
+   E3_uint32           q_size;         /* size of queue item; 0x1 <= size <= (0x40 * 5) */
+   E3_Addr             q_top;          /* block aligned ptr to last queue item */
+   E3_Addr             q_base;         /* block aligned ptr to first queue item */
+   volatile E3_Addr    q_fptr;         /* block aligned ptr to current front item */
+   E3_BlockCopyEvent   q_event;        /* queue event */
+   E3_uint32           q_pad[6];
+} E3_BlockCopyQueue;
+
+#define E3_QUEUE_EVENT_OFFSET  24
+#define QUEUE_FULL(Q)          ((Q)->q_state & E3_QUEUE_FULL)          
+
+#endif /* ! _ASM */
+
+#endif /* _ELAN3_EVENTS_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/intrinsics.h
===================================================================
--- linux-2.4.21.orig/include/elan3/intrinsics.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/intrinsics.h     2005-06-01 23:12:54.727419648 -0400
@@ -0,0 +1,320 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_INTRINSICS_H
+#define _ELAN3_INTRINSICS_H
+
+#ident "$Id: intrinsics.h,v 1.35 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/intrinsics.h,v $ */
+
+#include <elan3/e3types.h>
+#include <elan3/events.h>
+
+/* 
+ * This file contains definitions of the macros for accessing the QSW
+ * specific instructions, as if they were functions.
+ * The results from the function 
+ */
+
+#define C_ACK_OK       0                       /* return from c_close() */
+#define C_ACK_TESTFAIL 1                       /* return from c_close() */
+#define C_ACK_DISCARD  2                       /* return from c_close() */
+#define C_ACK_ERROR    3                       /* return from c_close() */
+
+/*
+ * Elan asi's for tproc block accesses
+ */
+#define EASI_BYTE      0
+#define EASI_HALF      1
+#define EASI_WORD      2
+#define EASI_DOUBLE    3
+
+#if defined(__ELAN3__) && !defined (_ASM)
+
+extern inline void c_abort(void) 
+{
+    asm volatile (".word 0x0000                ! die you thread you " : : );
+}
+
+extern inline void c_suspend(void) 
+{
+    asm volatile (
+       "set 1f, %%i7                   ! RevB bug fix. get address of the wakeup inst\n"
+       "andcc %%i7,0x4,%%g0            ! RevB bug fix. check alignment\n"
+       "bne 1f                         ! RevB bug fix. jump to other alignment\n"
+       "nop                            ! RevB bug fix. delay slot\n"
+       "ldd [%%i7],%%i6                ! RevB bug fix. data fetch of instructions\n"
+       "suspend                        ! do the real suspend\n"
+       "1: add %%i7,5*4,%%i7           ! RevB bug fix. Point i7 to first ldblock\n"
+       "ldd [%%i7],%%i6                ! RevB bug fix. data fetch of instructions\n"
+       "suspend                        ! do the real suspend\n" : : );
+}
+
+extern inline int c_close(void) 
+{
+    register int rc asm("o0");
+
+    asm volatile ("close %0" : "=r" (rc) : );
+
+    return (rc);
+}
+
+extern inline int c_close_cookie(volatile E3_uint32 *cookiep, E3_uint32 next)
+{
+    register int rc asm("o0");
+
+    asm volatile ("close       %0              ! close the packet\n"
+                 "bz,a         1f              ! ack received\n"
+                 "st           %1, [%2]        ! update cookie on ack\n"
+                 "1:                           ! label for not-ack\n"
+                 : "=r" (rc) : "r" (next), "r" (cookiep));
+
+    return (rc);
+}
+
+extern inline void c_break_busywait(void)
+{
+    asm volatile (
+       "breaktest                      ! test to see if break necessary\n"
+       "bpos 1f                        ! no other thread ready\n"
+       "nop                            ! delay slot\n"
+       "sub     %%sp,3*8*4,%%sp        ! Space to save the registers\n"
+       "stblock %%g0,[%%sp+0]          ! save the globals\n"
+       "stblock %%i0,[%%sp+8*4]        ! save the ins\n"
+       "stblock %%l0,[%%sp+16*4]       ! save the locals\n"
+       "set 2f, %%i7                   ! RevB bug fix. get address of the wakeup inst\n"
+       "andcc %%i7,0x4,%%g0            ! RevB bug fix. check alignment\n"
+       "bne 3f                         ! RevB bug fix. jump to other alignment\n"
+       "nop                            ! RevB bug fix. delay slot\n"
+       "ldd [%%i7],%%i6                ! RevB bug fix. data fetch of instructions\n"
+       "break                          ! do the real break\n"
+       "2: b 4f                        ! RevB bug fix. Branch over other alignment case\n"
+       " ldblock [%%sp+16*4],%%l0      ! RevB bug fix. restore locals in delay slot\n"
+       "3: add %%i7,5*4,%%i7           ! RevB bug fix. Point i7 to first ldblock\n"
+       "ldd [%%i7],%%i6                ! RevB bug fix. data fetch of instructions\n"
+       "break                          ! do the real break\n"
+       "ldblock [%%sp+16*4],%%l0       ! restore locals\n"
+       "4: ldblock [%%sp+8*4], %%i0    ! restore ins\n"
+       "ldblock [%%sp+0],%%g0          ! restore globals\n"
+       "add     %%sp,3*8*4,%%sp        ! restore stack pointer\n"
+       "1: " : : );
+}
+
+extern inline void c_break(void)
+{
+    asm volatile (
+       "breaktest                      ! test to see if break necessary\n"
+       "bne 1f                         ! haven't exceeded our inst count yet\n"
+       "nop                            ! delay slot\n"
+       "sub     %%sp,3*8*4,%%sp        ! Space to save the registers\n"
+       "stblock %%g0,[%%sp+0]          ! save the globals\n"
+       "stblock %%i0,[%%sp+8*4]        ! save the ins\n"
+       "stblock %%l0,[%%sp+16*4]       ! save the locals\n"
+       "set 2f, %%i7                   ! RevB bug fix. get address of the wakeup inst\n"
+       "andcc %%i7,0x4,%%g0            ! RevB bug fix. check alignment\n"
+       "bne 3f                         ! RevB bug fix. jump to other alignment\n"
+       "nop                            ! RevB bug fix. delay slot\n"
+       "ldd [%%i7],%%i6                ! RevB bug fix. data fetch of instructions\n"
+       "break                          ! do the real break\n"
+       "2: b 4f                        ! RevB bug fix. Branch over other alignment case\n"
+       " ldblock [%%sp+16*4],%%l0      ! RevB bug fix. restore locals in delay slot\n"
+       "3: add %%i7,5*4,%%i7           ! RevB bug fix. Point i7 to first ldblock\n"
+       "ldd [%%i7],%%i6                ! RevB bug fix. data fetch of instructions\n"
+       "break                          ! do the real break\n"
+       "ldblock [%%sp+16*4],%%l0       ! restore locals\n"
+       "4: ldblock [%%sp+8*4], %%i0    ! restore ins\n"
+       "ldblock [%%sp+0],%%g0          ! restore globals\n"
+       "add     %%sp,3*8*4,%%sp        ! restore stack pointer\n"
+       "1: " : : );
+}
+
+extern inline void c_open( const int arg ) 
+{
+    asm volatile ("open %0" : : "r" (arg) );
+    asm volatile ("nop; nop; nop; nop");
+    asm volatile ("nop; nop; nop; nop");
+    asm volatile ("nop; nop; nop; nop");
+    asm volatile ("nop; nop; nop; nop");
+    asm volatile ("nop; nop; nop; nop");
+    asm volatile ("nop; nop; nop; nop");
+}
+
+extern inline void c_waitevent( volatile E3_Event *const ptr,
+                               const int count) 
+{
+    register volatile E3_Event *a_unlikely asm("o0") = ptr;
+    register int a_very_unlikely asm("o1") = count;
+
+    asm volatile (
+        "sub     %%sp,1*8*4,%%sp       ! Space to save the registers\n"
+        "stblock %%i0,[%%sp+0]         ! save the ins\n"
+       "set    2f, %%i7                ! RevB bug fix. get address of the wakeup inst\n"
+       "andcc %%i7,0x4,%%g0            ! RevB bug fix. check alignment\n"
+       "bne 3f                         ! RevB bug fix. jump to other alignment\n"
+       "nop                            ! RevB bug fix. delay slot\n"
+       "ldd [%%i7],%%i4                ! RevB bug fix. data fetch of instructions\n"
+        "waitevent                     ! do the business\n"
+       "2: b 4f                        ! RevB bug fix. Branch over other alignment case\n"
+        "  ldblock [%%sp+0],%%i0       ! RevB bug fix. restore ins in delay slot\n"
+       "3: add %%i7,5*4,%%i7           ! RevB bug fix. Point i7 to first ldblock\n"
+       "ldd [%%i7],%%i4                ! RevB bug fix. data fetch of instructions\n"
+        "waitevent                     ! do the business\n"
+        "ldblock [%%sp+0],%%i0         ! restore ins\n"
+        "4: add     %%sp,1*8*4,%%sp    ! restore stack pointer\n"
+        : /* no outputs */
+        : /* inputs */ "r" (a_unlikely), "r" (a_very_unlikely)
+        : /* clobbered */ "g0", "g1", "g2", "g3", "g4", "g5", "g6", "g7",
+                         "l0", "l1", "l2", "l3", "l4", "l5", "l6", "l7" );
+
+}
+
+#define c_sendtrans0(type,dest)                        \
+       asm volatile ("sendtrans %0, %%g0, %1" : : "i" (type), "r" (dest))
+
+#define c_sendtrans1(type,dest,arg)            \
+       asm volatile ("sendtrans %0, %2, %1" : : "i" (type), "r" (dest), "r" (arg))
+
+#define c_sendtrans2(type,dest,arg1,arg2)      \
+       do {                                    \
+            register const unsigned long a_unlikely_1 asm("o4") = arg1;                        \
+            register const unsigned long a_unlikely_2 asm("o5") = arg2;                        \
+            asm volatile ("sendtrans %0, %2, %1"                                       \
+                : : "i" (type), "r" (dest), "r" (a_unlikely_1), "r" (a_unlikely_2));   \
+       } while(0)
+
+#define c_sendmem(type,dest,ptr)               \
+       asm volatile ("sendtrans %0, [%2], %1" : : "i" (type), "r" (dest), "r" (ptr))
+
+/* Copy a single 64-byte block (src blk is read using a BYTE endian type) */
+extern inline void elan3_copy64b(void *src, void *dst)
+{
+    /* Copy 64 bytes using ldblock/stblock
+     * We save and restore the locals/ins because if we don't gcc
+     * really makes a bad job of optimisising the rest of the thread code!
+     *
+     * We force the parameters in g5, g6 so that they aren't
+     * trashed by the loadblk32 into the locals/ins
+     */
+    register void *tmp1 asm("g5") = src;
+    register void *tmp2 asm("g6") = dst;
+
+    asm volatile (
+       "and     %%sp,63,%%g7           ! Calculate stack alignment\n"
+       "sub     %%sp,2*8*4,%%sp        ! Space to save the registers\n"
+       "sub     %%sp,%%g7,%%sp         ! align stack\n" 
+       "stblock64 %%l0,[%%sp]          ! save the locals and ins\n"
+       "ldblock64a [%0]%2,%%l0         ! load 64-byte block into locals/ins\n"
+       "stblock64a %%l0,[%1]%2         ! store 64-byte block from local/ins\n"
+       "ldblock64 [%%sp],%%l0          ! restore locals and ins\n"
+       "add     %%sp,%%g7, %%sp        ! undo alignment\n"
+       "add     %%sp,2*8*4,%%sp        ! restore stack pointer\n"
+       : /* outputs */
+       : /* inputs */ "r" (tmp1), "r" (tmp2), "n" (EASI_BYTE)
+       : /* clobbered */ "g5", "g6", "g7" );
+}
+
+/* Copy a single 64-byte block (src blk is read using a WORD endian type) */
+extern inline void elan3_copy64w(void *src, void *dst)
+{
+    /* Copy 64 bytes using ldblock/stblock
+     * We save and restore the locals/ins because if we don't gcc
+     * really makes a bad job of optimisising the rest of the thread code!
+     *
+     * We force the parameters in g5, g6 so that they aren't
+     * trashed by the loadblk32 into the locals/ins
+     */
+    register void *tmp1 asm("g5") = src;
+    register void *tmp2 asm("g6") = dst;
+
+    asm volatile (
+       "and     %%sp,63,%%g7           ! Calculate stack alignment\n"
+       "sub     %%sp,2*8*4,%%sp        ! Space to save the registers\n"
+       "sub     %%sp,%%g7,%%sp         ! align stack\n" 
+       "stblock64 %%l0,[%%sp]          ! save the locals and ins\n"
+       "ldblock64a [%0]%2,%%l0         ! load 64-byte block into locals/ins\n"
+       "stblock64a %%l0,[%1]%2         ! store 64-byte block from local/ins\n"
+       "ldblock64 [%%sp],%%l0          ! restore locals and ins\n"
+       "add     %%sp,%%g7, %%sp        ! undo alignment\n"
+       "add     %%sp,2*8*4,%%sp        ! restore stack pointer\n"
+       : /* outputs */
+       : /* inputs */ "r" (tmp1), "r" (tmp2), "n" (EASI_WORD)
+       : /* clobbered */ "g5", "g6", "g7" );
+}
+
+/* Read a 64-bit value with a WORD (32-bit) endian type */
+extern inline E3_uint64 elan3_read64w( volatile E3_uint64 *const ptr )
+{
+    E3_uint64 result;
+
+    asm volatile (
+       "ldblock8a [%1]%2, %0\n"
+       : /* outputs */ "=r" (result)
+       : /* inputs */ "r" (ptr), "n" (EASI_WORD) );
+
+    return( result );
+}
+
+/* Read a 64-bit value with a DOUBLEWORD (64-bit) endian type */
+extern inline E3_uint64 elan3_read64dw( volatile E3_uint64 *const ptr )
+{
+    E3_uint64 result;
+
+    asm volatile (
+       "ldblock8a [%1]%2, %0\n"
+       : /* outputs */ "=r" (result)
+       : /* inputs */ "r" (ptr), "n" (EASI_DOUBLE) );
+
+    return( result );
+}
+
+/* Write a 32-bit value with a WORD (32-bit) endian type */
+extern inline void elan3_write64w( volatile E3_uint64 *const ptr, E3_uint64 value )
+{
+    asm volatile (
+       "stblock8a %1, [%0]%2\n"
+       : /* no outputs */
+       : /* inputs */ "r" (ptr), "r" (value), "n" (EASI_WORD) );
+}
+
+/* Write a 64-bit value with a DOUBLEWORD (64-bit) endian type */
+extern inline void elan3_write64dw( volatile E3_uint64 *const ptr, E3_uint64 value )
+{
+    asm volatile (
+       "stblock8a %1, [%0]%2\n"
+       : /* no outputs */
+       : /* inputs */ "r" (ptr), "r" (value), "n" (EASI_DOUBLE) );
+}
+
+extern inline E3_uint32 c_swap(volatile E3_uint32 *source, E3_uint32 result)
+{
+   asm volatile("swap [%1],%0\n"
+               : "=r" (result)
+               : "r" (source) ,"0" (result)
+               : "memory");
+   return result;
+}
+
+extern inline E3_uint32 c_swap_save(volatile E3_uint32 *source, const E3_uint32 result)
+{
+   register E3_uint32 a_unlikely;
+   asm volatile("" : "=r" (a_unlikely) : );
+
+   asm volatile("mov %2,%0; swap [%1],%0\n"
+               : "=r" (a_unlikely)
+               : "r" (source) ,"r" (result), "0" (a_unlikely)
+               : "memory");
+   return a_unlikely;
+}
+#endif /* (__ELAN3__) && !(_ASM) */
+
+#endif /* _ELAN3_INTRINSICS_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/minames.h
===================================================================
--- linux-2.4.21.orig/include/elan3/minames.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/minames.h        2005-06-01 23:12:54.728419496 -0400
@@ -0,0 +1,256 @@
+{MI_WaitForRemoteDescRead,     "MI_WaitForRemoteDescRead"},
+{MI_WaitForRemoteDescRead2,    "MI_WaitForRemoteDescRead2"},
+{MI_WaitForRemoteDescRead2_seq1,       "MI_WaitForRemoteDescRead2_seq1"},
+{MI_SendRemoteDmaRoutes,       "MI_SendRemoteDmaRoutes"},
+{MI_IProcTrapped,      "MI_IProcTrapped"},
+{MI_DProcTrapped,      "MI_DProcTrapped"},
+{MI_CProcTrapped,      "MI_CProcTrapped"},
+{MI_TProcTrapped,      "MI_TProcTrapped"},
+{MI_TestWhichDmaQueue, "MI_TestWhichDmaQueue"},
+{MI_TestWhichDmaQueue_seq1,    "MI_TestWhichDmaQueue_seq1"},
+{MI_InputRemoteDmaUpdateBPtr,  "MI_InputRemoteDmaUpdateBPtr"},
+{MI_FixupQueueContextAndRemoteBit,     "MI_FixupQueueContextAndRemoteBit"},
+{MI_FixupQueueContextAndRemoteBit_seq1,        "MI_FixupQueueContextAndRemoteBit_seq1"},
+{MI_FixupQueueContextAndRemoteBit_seq2,        "MI_FixupQueueContextAndRemoteBit_seq2"},
+{MI_FixupQueueContextAndRemoteBit_seq3,        "MI_FixupQueueContextAndRemoteBit_seq3"},
+{MI_FixupQueueContextAndRemoteBit_seq4,        "MI_FixupQueueContextAndRemoteBit_seq4"},
+{MI_RunDmaCommand,     "MI_RunDmaCommand"},
+{MI_DoSendRemoteDmaDesc,       "MI_DoSendRemoteDmaDesc"},
+{MI_DequeueNonSysCntxDma,      "MI_DequeueNonSysCntxDma"},
+{MI_WaitForRemoteDescRead1,    "MI_WaitForRemoteDescRead1"},
+{MI_RemoteDmaCommand,  "MI_RemoteDmaCommand"},
+{MI_WaitForRemoteRoutes,       "MI_WaitForRemoteRoutes"},
+{MI_DequeueSysCntxDma, "MI_DequeueSysCntxDma"},
+{MI_ExecuteDmaDescriptorForQueue,      "MI_ExecuteDmaDescriptorForQueue"},
+{MI_ExecuteDmaDescriptor1,     "MI_ExecuteDmaDescriptor1"},
+{MI_ExecuteDmaDescriptor1_seq1,        "MI_ExecuteDmaDescriptor1_seq1"},
+{MI_ExecuteDmaDescriptor1_seq2,        "MI_ExecuteDmaDescriptor1_seq2"},
+{MI_ExecuteDmaDescriptor1_seq3,        "MI_ExecuteDmaDescriptor1_seq3"},
+{MI_GetNewSizeInProg,  "MI_GetNewSizeInProg"},
+{MI_GetNewSizeInProg_seq1,     "MI_GetNewSizeInProg_seq1"},
+{MI_FirstBlockRead,    "MI_FirstBlockRead"},
+{MI_ExtraFirstBlockRead,       "MI_ExtraFirstBlockRead"},
+{MI_UnimplementedError,        "MI_UnimplementedError"},
+{MI_UpdateDescriptor,  "MI_UpdateDescriptor"},
+{MI_UpdateDescriptor_seq1,     "MI_UpdateDescriptor_seq1"},
+{MI_UpdateDescriptor_seq2,     "MI_UpdateDescriptor_seq2"},
+{MI_UpdateDescriptor_seq3,     "MI_UpdateDescriptor_seq3"},
+{MI_UpdateDescriptor_seq4,     "MI_UpdateDescriptor_seq4"},
+{MI_UpdateDescriptor_seq5,     "MI_UpdateDescriptor_seq5"},
+{MI_GetNextSizeInProg, "MI_GetNextSizeInProg"},
+{MI_DoStopThisDma,     "MI_DoStopThisDma"},
+{MI_DoStopThisDma_seq1,        "MI_DoStopThisDma_seq1"},
+{MI_GenNewBytesToRead, "MI_GenNewBytesToRead"},
+{MI_WaitForEventReadTy1,       "MI_WaitForEventReadTy1"},
+{MI_WaitUpdateEvent,   "MI_WaitUpdateEvent"},
+{MI_WaitUpdateEvent_seq1,      "MI_WaitUpdateEvent_seq1"},
+{MI_DoSleepOneTickThenRunable, "MI_DoSleepOneTickThenRunable"},
+{MI_RunEvent,  "MI_RunEvent"},
+{MI_EnqueueThread,     "MI_EnqueueThread"},
+{MI_CheckContext0,     "MI_CheckContext0"},
+{MI_EnqueueDma,        "MI_EnqueueDma"},
+{MI_CprocTrapping,     "MI_CprocTrapping"},
+{MI_CprocTrapping_seq1,        "MI_CprocTrapping_seq1"},
+{MI_WaitForRemoteRoutes1,      "MI_WaitForRemoteRoutes1"},
+{MI_SetEventCommand,   "MI_SetEventCommand"},
+{MI_DoSetEvent,        "MI_DoSetEvent"},
+{MI_DoRemoteSetEventNowOrTrapQueueingDma,      "MI_DoRemoteSetEventNowOrTrapQueueingDma"},
+{MI_DoRemoteSetEventNowOrTrapQueueingDma_seq1, "MI_DoRemoteSetEventNowOrTrapQueueingDma_seq1"},
+{MI_SendRemoteDmaRoutes2,      "MI_SendRemoteDmaRoutes2"},
+{MI_WaitForRemoteRoutes2,      "MI_WaitForRemoteRoutes2"},
+{MI_WaitEventCommandTy0,       "MI_WaitEventCommandTy0"},
+{MI_DequeueNonSysCntxDma2,     "MI_DequeueNonSysCntxDma2"},
+{MI_WaitEventCommandTy1,       "MI_WaitEventCommandTy1"},
+{MI_WaitEventCommandTy1_seq1,  "MI_WaitEventCommandTy1_seq1"},
+{MI_DequeueNonSysCntxThread,   "MI_DequeueNonSysCntxThread"},
+{MI_DequeueSysCntxDma1,        "MI_DequeueSysCntxDma1"},
+{MI_DequeueSysCntxThread,      "MI_DequeueSysCntxThread"},
+{MI_TestNonSysCntxDmaQueueEmpty,       "MI_TestNonSysCntxDmaQueueEmpty"},
+{MI_TestNonSysCntxDmaQueueEmpty_seq1,  "MI_TestNonSysCntxDmaQueueEmpty_seq1"},
+{MI_TestNonSysCntxDmaQueueEmpty_seq2,  "MI_TestNonSysCntxDmaQueueEmpty_seq2"},
+{MI_RunThreadCommand,  "MI_RunThreadCommand"},
+{MI_SetEventWaitForLastAcess,  "MI_SetEventWaitForLastAcess"},
+{MI_SetEventReadWait,  "MI_SetEventReadWait"},
+{MI_SetEventReadWait_seq1,     "MI_SetEventReadWait_seq1"},
+{MI_TestEventType,     "MI_TestEventType"},
+{MI_TestEventType_seq1,        "MI_TestEventType_seq1"},
+{MI_TestEventBit2,     "MI_TestEventBit2"},
+{MI_DmaDescOrBlockCopyOrChainedEvent,  "MI_DmaDescOrBlockCopyOrChainedEvent"},
+{MI_RunThread, "MI_RunThread"},
+{MI_RunThread1,        "MI_RunThread1"},
+{MI_RunThread1_seq1,   "MI_RunThread1_seq1"},
+{MI_IncDmaSysCntxBPtr, "MI_IncDmaSysCntxBPtr"},
+{MI_IncDmaSysCntxBPtr_seq1,    "MI_IncDmaSysCntxBPtr_seq1"},
+{MI_IncDmaSysCntxBPtr_seq2,    "MI_IncDmaSysCntxBPtr_seq2"},
+{MI_WaitForCntxDmaDescRead,    "MI_WaitForCntxDmaDescRead"},
+{MI_FillInContext,     "MI_FillInContext"},
+{MI_FillInContext_seq1,        "MI_FillInContext_seq1"},
+{MI_WriteNewDescToQueue,       "MI_WriteNewDescToQueue"},
+{MI_WriteNewDescToQueue_seq1,  "MI_WriteNewDescToQueue_seq1"},
+{MI_TestForQueueWrap,  "MI_TestForQueueWrap"},
+{MI_TestForQueueWrap_seq1,     "MI_TestForQueueWrap_seq1"},
+{MI_TestQueueIsFull,   "MI_TestQueueIsFull"},
+{MI_TestQueueIsFull_seq1,      "MI_TestQueueIsFull_seq1"},
+{MI_TestQueueIsFull_seq2,      "MI_TestQueueIsFull_seq2"},
+{MI_CheckPsychoShitFixup,      "MI_CheckPsychoShitFixup"},
+{MI_PsychoShitFixupForcedRead, "MI_PsychoShitFixupForcedRead"},
+{MI_PrepareDMATimeSlice,       "MI_PrepareDMATimeSlice"},
+{MI_PrepareDMATimeSlice_seq1,  "MI_PrepareDMATimeSlice_seq1"},
+{MI_TProcRestartFromTrapOrTestEventBit2,       "MI_TProcRestartFromTrapOrTestEventBit2"},
+{MI_TProcRestartFromTrapOrTestEventBit2_seq1,  "MI_TProcRestartFromTrapOrTestEventBit2_seq1"},
+{MI_WaitForGlobalsRead,        "MI_WaitForGlobalsRead"},
+{MI_WaitForNPCRead,    "MI_WaitForNPCRead"},
+{MI_EventInterrupt,    "MI_EventInterrupt"},
+{MI_EventInterrupt_seq1,       "MI_EventInterrupt_seq1"},
+{MI_EventInterrupt_seq2,       "MI_EventInterrupt_seq2"},
+{MI_EventInterrupt_seq3,       "MI_EventInterrupt_seq3"},
+{MI_TestSysCntxDmaQueueEmpty,  "MI_TestSysCntxDmaQueueEmpty"},
+{MI_TestSysCntxDmaQueueEmpty_seq1,     "MI_TestSysCntxDmaQueueEmpty_seq1"},
+{MI_TestIfRemoteDesc,  "MI_TestIfRemoteDesc"},
+{MI_DoDmaLocalSetEvent,        "MI_DoDmaLocalSetEvent"},
+{MI_DoDmaLocalSetEvent_seq1,   "MI_DoDmaLocalSetEvent_seq1"},
+{MI_DoDmaLocalSetEvent_seq2,   "MI_DoDmaLocalSetEvent_seq2"},
+{MI_DmaLoop1,  "MI_DmaLoop1"},
+{MI_ExitDmaLoop,       "MI_ExitDmaLoop"},
+{MI_ExitDmaLoop_seq1,  "MI_ExitDmaLoop_seq1"},
+{MI_RemoteDmaTestPAckType,     "MI_RemoteDmaTestPAckType"},
+{MI_PacketDiscardOrTestFailRecIfCCis0, "MI_PacketDiscardOrTestFailRecIfCCis0"},
+{MI_PacketDiscardOrTestFailRecIfCCis0_seq1,    "MI_PacketDiscardOrTestFailRecIfCCis0_seq1"},
+{MI_TestNackFailIsZero2,       "MI_TestNackFailIsZero2"},
+{MI_TestNackFailIsZero3,       "MI_TestNackFailIsZero3"},
+{MI_DmaFailCountError, "MI_DmaFailCountError"},
+{MI_TestDmaForSysCntx, "MI_TestDmaForSysCntx"},
+{MI_TestDmaForSysCntx_seq1,    "MI_TestDmaForSysCntx_seq1"},
+{MI_TestDmaForSysCntx_seq2,    "MI_TestDmaForSysCntx_seq2"},
+{MI_TestAeqB2, "MI_TestAeqB2"},
+{MI_TestAeqB2_seq1,    "MI_TestAeqB2_seq1"},
+{MI_GetNextDmaDescriptor,      "MI_GetNextDmaDescriptor"},
+{MI_DequeueSysCntxDma2,        "MI_DequeueSysCntxDma2"},
+{MI_InputSetEvent,     "MI_InputSetEvent"},
+{MI_PutBackSysCntxDma, "MI_PutBackSysCntxDma"},
+{MI_PutBackSysCntxDma_seq1,    "MI_PutBackSysCntxDma_seq1"},
+{MI_PutBackSysCntxDma_seq2,    "MI_PutBackSysCntxDma_seq2"},
+{MI_InputRemoteDma,    "MI_InputRemoteDma"},
+{MI_InputRemoteDma_seq1,       "MI_InputRemoteDma_seq1"},
+{MI_WaitOneTickForWakeup1,     "MI_WaitOneTickForWakeup1"},
+{MI_SendRemoteDmaDesc, "MI_SendRemoteDmaDesc"},
+{MI_InputLockQueue,    "MI_InputLockQueue"},
+{MI_CloseTheTrappedPacketIfCCis1,      "MI_CloseTheTrappedPacketIfCCis1"},
+{MI_CloseTheTrappedPacketIfCCis1_seq1, "MI_CloseTheTrappedPacketIfCCis1_seq1"},
+{MI_PostDmaInterrupt,  "MI_PostDmaInterrupt"},
+{MI_InputUnLockQueue,  "MI_InputUnLockQueue"},
+{MI_WaitForUnLockDescRead,     "MI_WaitForUnLockDescRead"},
+{MI_SendEOPforRemoteDma,       "MI_SendEOPforRemoteDma"},
+{MI_LookAtRemoteAck,   "MI_LookAtRemoteAck"},
+{MI_InputWriteBlockQueue,      "MI_InputWriteBlockQueue"},
+{MI_WaitForSpStore,    "MI_WaitForSpStore"},
+{MI_TProcNext, "MI_TProcNext"},
+{MI_TProcStoppedRunning,       "MI_TProcStoppedRunning"},
+{MI_InputWriteBlock,   "MI_InputWriteBlock"},
+{MI_RunDmaOrDeqNonSysCntxDma,  "MI_RunDmaOrDeqNonSysCntxDma"},
+{MI_ExecuteDmaDescriptorForRun,        "MI_ExecuteDmaDescriptorForRun"},
+{MI_ConfirmQueueLock,  "MI_ConfirmQueueLock"},
+{MI_DmaInputIdentify,  "MI_DmaInputIdentify"},
+{MI_TProcStoppedRunning2,      "MI_TProcStoppedRunning2"},
+{MI_TProcStoppedRunning2_seq1, "MI_TProcStoppedRunning2_seq1"},
+{MI_TProcStoppedRunning2_seq2, "MI_TProcStoppedRunning2_seq2"},
+{MI_ThreadInputIdentify,       "MI_ThreadInputIdentify"},
+{MI_InputIdWriteAddrAndType3,  "MI_InputIdWriteAddrAndType3"},
+{MI_IProcTrappedWriteStatus,   "MI_IProcTrappedWriteStatus"},
+{MI_FinishTrappingEop, "MI_FinishTrappingEop"},
+{MI_InputTestTrans,    "MI_InputTestTrans"},
+{MI_TestAeqB3, "MI_TestAeqB3"},
+{MI_ThreadUpdateNonSysCntxBack,        "MI_ThreadUpdateNonSysCntxBack"},
+{MI_ThreadQueueOverflow,       "MI_ThreadQueueOverflow"},
+{MI_RunContext0Thread, "MI_RunContext0Thread"},
+{MI_RunContext0Thread_seq1,    "MI_RunContext0Thread_seq1"},
+{MI_RunContext0Thread_seq2,    "MI_RunContext0Thread_seq2"},
+{MI_RunDmaDesc,        "MI_RunDmaDesc"},
+{MI_RunDmaDesc_seq1,   "MI_RunDmaDesc_seq1"},
+{MI_RunDmaDesc_seq2,   "MI_RunDmaDesc_seq2"},
+{MI_TestAeqB,  "MI_TestAeqB"},
+{MI_WaitForNonCntxDmaDescRead, "MI_WaitForNonCntxDmaDescRead"},
+{MI_DmaQueueOverflow,  "MI_DmaQueueOverflow"},
+{MI_BlockCopyEvent,    "MI_BlockCopyEvent"},
+{MI_BlockCopyEventReadBlock,   "MI_BlockCopyEventReadBlock"},
+{MI_BlockCopyWaitForReadData,  "MI_BlockCopyWaitForReadData"},
+{MI_InputWriteWord,    "MI_InputWriteWord"},
+{MI_TraceSetEvents,    "MI_TraceSetEvents"},
+{MI_TraceSetEvents_seq1,       "MI_TraceSetEvents_seq1"},
+{MI_TraceSetEvents_seq2,       "MI_TraceSetEvents_seq2"},
+{MI_InputWriteDoubleWd,        "MI_InputWriteDoubleWd"},
+{MI_SendLockTransIfCCis1,      "MI_SendLockTransIfCCis1"},
+{MI_WaitForDmaRoutes1, "MI_WaitForDmaRoutes1"},
+{MI_LoadDmaContext,    "MI_LoadDmaContext"},
+{MI_InputTestAndSetWord,       "MI_InputTestAndSetWord"},
+{MI_InputTestAndSetWord_seq1,  "MI_InputTestAndSetWord_seq1"},
+{MI_GetDestEventValue, "MI_GetDestEventValue"},
+{MI_SendDmaIdentify,   "MI_SendDmaIdentify"},
+{MI_InputAtomicAddWord,        "MI_InputAtomicAddWord"},
+{MI_LoadBFromTransD0,  "MI_LoadBFromTransD0"},
+{MI_ConditionalWriteBackCCTrue,        "MI_ConditionalWriteBackCCTrue"},
+{MI_WaitOneTickForWakeup,      "MI_WaitOneTickForWakeup"},
+{MI_SendFinalUnlockTrans,      "MI_SendFinalUnlockTrans"},
+{MI_SendDmaEOP,        "MI_SendDmaEOP"},
+{MI_GenLastAddrForPsycho,      "MI_GenLastAddrForPsycho"},
+{MI_FailedAckIfCCis0,  "MI_FailedAckIfCCis0"},
+{MI_FailedAckIfCCis0_seq1,     "MI_FailedAckIfCCis0_seq1"},
+{MI_WriteDmaSysCntxDesc,       "MI_WriteDmaSysCntxDesc"},
+{MI_TimesliceDmaQueueOverflow, "MI_TimesliceDmaQueueOverflow"},
+{MI_DequeueNonSysCntxThread1,  "MI_DequeueNonSysCntxThread1"},
+{MI_DequeueNonSysCntxThread1_seq1,     "MI_DequeueNonSysCntxThread1_seq1"},
+{MI_TestThreadQueueEmpty,      "MI_TestThreadQueueEmpty"},
+{MI_ClearThreadQueueIfCC,      "MI_ClearThreadQueueIfCC"},
+{MI_DequeueSysCntxThread1,     "MI_DequeueSysCntxThread1"},
+{MI_DequeueSysCntxThread1_seq1,        "MI_DequeueSysCntxThread1_seq1"},
+{MI_TProcStartUpGeneric,       "MI_TProcStartUpGeneric"},
+{MI_WaitForPCload2,    "MI_WaitForPCload2"},
+{MI_WaitForNPCWrite,   "MI_WaitForNPCWrite"},
+{MI_WaitForEventWaitAddr,      "MI_WaitForEventWaitAddr"},
+{MI_WaitForWaitEventAccess,    "MI_WaitForWaitEventAccess"},
+{MI_WaitForWaitEventAccess_seq1,       "MI_WaitForWaitEventAccess_seq1"},
+{MI_WaitForWaitEventDesc,      "MI_WaitForWaitEventDesc"},
+{MI_WaitForEventReadTy0,       "MI_WaitForEventReadTy0"},
+{MI_SendCondTestFail,  "MI_SendCondTestFail"},
+{MI_InputMoveToNextTrans,      "MI_InputMoveToNextTrans"},
+{MI_ThreadUpdateSysCntxBack,   "MI_ThreadUpdateSysCntxBack"},
+{MI_FinishedSetEvent,  "MI_FinishedSetEvent"},
+{MI_EventIntUpdateBPtr,        "MI_EventIntUpdateBPtr"},
+{MI_EventQueueOverflow,        "MI_EventQueueOverflow"},
+{MI_MaskLowerSource,   "MI_MaskLowerSource"},
+{MI_DmaLoop,   "MI_DmaLoop"},
+{MI_SendNullSetEvent,  "MI_SendNullSetEvent"},
+{MI_SendFinalSetEvent, "MI_SendFinalSetEvent"},
+{MI_TestNackFailIsZero1,       "MI_TestNackFailIsZero1"},
+{MI_DmaPacketTimedOutOrPacketError,    "MI_DmaPacketTimedOutOrPacketError"},
+{MI_NextPacketIsLast,  "MI_NextPacketIsLast"},
+{MI_TestForZeroLengthDma,      "MI_TestForZeroLengthDma"},
+{MI_WaitForPCload,     "MI_WaitForPCload"},
+{MI_ReadInIns, "MI_ReadInIns"},
+{MI_WaitForInsRead,    "MI_WaitForInsRead"},
+{MI_WaitForLocals,     "MI_WaitForLocals"},
+{MI_WaitForOutsWrite,  "MI_WaitForOutsWrite"},
+{MI_WaitForWaitEvWrBack,       "MI_WaitForWaitEvWrBack"},
+{MI_WaitForLockRead,   "MI_WaitForLockRead"},
+{MI_TestQueueLock,     "MI_TestQueueLock"},
+{MI_InputIdWriteAddrAndType,   "MI_InputIdWriteAddrAndType"},
+{MI_InputIdWriteAddrAndType2,  "MI_InputIdWriteAddrAndType2"},
+{MI_ThreadInputIdentify2,      "MI_ThreadInputIdentify2"},
+{MI_WriteIntoTrapArea0,        "MI_WriteIntoTrapArea0"},
+{MI_GenQueueBlockWrAddr,       "MI_GenQueueBlockWrAddr"},
+{MI_InputDiscardFreeLock,      "MI_InputDiscardFreeLock"},
+{MI_WriteIntoTrapArea1,        "MI_WriteIntoTrapArea1"},
+{MI_WriteIntoTrapArea2,        "MI_WriteIntoTrapArea2"},
+{MI_ResetBPtrToBase,   "MI_ResetBPtrToBase"},
+{MI_InputDoTrap,       "MI_InputDoTrap"},
+{MI_RemoteDmaCntxt0Update,     "MI_RemoteDmaCntxt0Update"},
+{MI_ClearQueueLock,    "MI_ClearQueueLock"},
+{MI_IProcTrappedBlockWriteData,        "MI_IProcTrappedBlockWriteData"},
+{MI_FillContextFilter, "MI_FillContextFilter"},
+{MI_IProcTrapped4,     "MI_IProcTrapped4"},
+{MI_RunSysCntxDma,     "MI_RunSysCntxDma"},
+{MI_ChainedEventError, "MI_ChainedEventError"},
+{MI_InputTrappingEOP,  "MI_InputTrappingEOP"},
+{MI_CheckForRunIfZero, "MI_CheckForRunIfZero"},
+{MI_TestForBreakOrSuspend,     "MI_TestForBreakOrSuspend"},
+{MI_SwapForRunable,    "MI_SwapForRunable"},
Index: linux-2.4.21/include/elan3/neterr_rpc.h
===================================================================
--- linux-2.4.21.orig/include/elan3/neterr_rpc.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/neterr_rpc.h     2005-06-01 23:12:54.728419496 -0400
@@ -0,0 +1,68 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_NETERR_RPC_H
+#define __ELAN3_NETERR_RPC_H
+
+#ident "$Id: neterr_rpc.h,v 1.20 2003/06/26 16:05:22 fabien Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/neterr_rpc.h,v $*/
+
+#define NETERR_SERVICE "neterr-srv"
+#define NETERR_PROGRAM ((u_long) 170002)
+#define NETERR_VERSION ((u_long) 1)
+
+#define NETERR_NULL_RPC                0
+#define NETERR_FIXUP_RPC       1
+
+/* network error rpc timeout */
+#define NETERR_RPC_TIMEOUT     5
+
+/*
+ * XDR functions for Tru64 and Linux in userspace. 
+ *  NB Linux kernelspace xdr routines are in network_error.
+ *     and *must* be kept consistent.
+ */
+#if defined(DIGITAL_UNIX) || !defined(__KERNEL__)
+bool_t
+xdr_capability (XDR *xdrs, void *arg)
+{
+    ELAN_CAPABILITY *cap = (ELAN_CAPABILITY *) arg;
+
+    return (xdr_opaque (xdrs, (caddr_t) &cap->cap_userkey, sizeof (cap->cap_userkey)) &&
+           xdr_int (xdrs, &cap->cap_version) &&
+           xdr_u_short (xdrs, &cap->cap_type) &&
+           xdr_int (xdrs, &cap->cap_lowcontext) &&
+           xdr_int (xdrs, &cap->cap_highcontext) &&
+           xdr_int (xdrs, &cap->cap_mycontext) &&
+           xdr_int (xdrs, &cap->cap_lownode) &&
+           xdr_int (xdrs, &cap->cap_highnode) &&
+           xdr_u_int (xdrs, &cap->cap_railmask) &&
+           xdr_opaque (xdrs, (caddr_t) &cap->cap_bitmap[0], sizeof (cap->cap_bitmap)));
+}
+
+bool_t
+xdr_neterr_msg (XDR *xdrs, void *req)
+{
+    NETERR_MSG *msg = (NETERR_MSG *) req;
+
+    return (xdr_u_int (xdrs, &msg->Rail) &&
+           xdr_capability (xdrs, &msg->SrcCapability) &&
+           xdr_capability (xdrs, &msg->DstCapability) &&
+           xdr_u_int (xdrs, &msg->DstProcess) &&
+           xdr_u_int (xdrs, &msg->CookieAddr) &&
+           xdr_u_int (xdrs, &msg->CookieVProc) &&
+           xdr_u_int (xdrs, &msg->NextCookie) &&
+           xdr_u_int (xdrs, &msg->WaitForEop));
+}
+#endif /* INCLUDE_XDR_INLINE */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN3_NETERR_RPC_H */
Index: linux-2.4.21/include/elan3/perm.h
===================================================================
--- linux-2.4.21.orig/include/elan3/perm.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/perm.h   2005-06-01 23:12:54.728419496 -0400
@@ -0,0 +1,29 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_PERM_H
+#define __ELAN3_PERM_H
+
+#ident "$Id: perm.h,v 1.7 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/perm.h,v $*/
+
+#define ELAN3_PERM_NULL                0x00
+#define ELAN3_PERM_LOCAL_READ  0x04
+#define ELAN3_PERM_READ                0x08
+#define ELAN3_PERM_NOREMOTE    0x0c
+#define ELAN3_PERM_REMOTEREAD  0x10
+#define ELAN3_PERM_REMOTEWRITE 0x14
+#define ELAN3_PERM_REMOTEEVENT 0x18
+#define ELAN3_PERM_REMOTEALL   0x1c
+
+#endif /* __ELAN3_PERM_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/pte.h
===================================================================
--- linux-2.4.21.orig/include/elan3/pte.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/pte.h    2005-06-01 23:12:54.729419344 -0400
@@ -0,0 +1,139 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_PTE_H
+#define __ELAN3_PTE_H
+
+#ident "$Id: pte.h,v 1.26 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/pte.h,v $*/
+
+#ifdef __cplusplus
+extern          "C"
+{
+#endif
+
+#include <elan3/e3types.h>
+#include <elan3/perm.h>
+
+typedef E3_uint64 ELAN3_PTE;
+typedef E3_uint32 ELAN3_PTP;
+
+#define ELAN3_PTE_SIZE         (8)
+#define ELAN3_PTP_SIZE         (4)
+
+#define ELAN3_PTE_REF          ((E3_uint64) 1 << 63)           /* 63      - referenced bit */
+#define ELAN3_PTE_MOD          ((E3_uint64) 1 << 55)           /* 55      - modified bit */
+#define ELAN3_RM_MASK          (ELAN3_PTE_REF | ELAN3_PTE_MOD)
+
+#define ELAN3_PTE_PFN_MASK     0x0000fffffffff000ull           /* [12:48] - Physical address */
+
+#define ELAN3_PTE_BIG_ENDIAN   0x80                            /* 7       - big endian */
+#define ELAN3_PTE_64_BIT               0x40                            /* 6       - 64 bit pci address */
+#define ELAN3_PTE_LOCAL                0x20                            /* 5       - local sdram */
+
+#define ELAN3_PTE_PERM_MASK    0x1c                            /* [2:4]   - Permissions */
+#define ELAN3_PTE_PERM_SHIFT      2
+
+#define ELAN3_ET_MASK          0x3
+#define ELAN3_ET_INVALID               0x0                                     /* [0:1] */
+#define ELAN3_ET_PTP           0x1
+#define ELAN3_ET_PTE           0x2
+
+#define ELAN3_INVALID_PTP      ((ELAN3_PTP) 0)
+#define ELAN3_INVALID_PTE      ((ELAN3_PTE) 0)
+
+#define ELAN3_PTP_TYPE(ptp)    ((ptp) & ELAN3_ET_MASK)
+#define ELAN3_PTE_TYPE(pte)    ((pte) & ELAN3_ET_MASK)
+#define ELAN3_PTE_PERM(pte)    ((pte) & ELAN3_PTE_PERM_MASK)
+#define ELAN3_PTE_VALID(pte)   (((pte) & ELAN3_ET_MASK) == ELAN3_ET_PTE)
+#define ELAN3_PTE_ISREF(pte)   ((pte) & ELAN3_PTE_REF)
+#define ELAN3_PTE_ISMOD(pte)   ((pte) & ELAN3_PTE_MOD)
+#define ELAN3_PTE_WRITEABLE(pte)       (ELAN3_PERM_WRITEABLE(ELAN3_PTE_PERM(pte)))
+
+#define ELAN3_PERM_WRITEABLE(perm)     ((perm) == ELAN3_PERM_NOREMOTE || (perm) > ELAN3_PERM_REMOTEREAD)
+#define ELAN3_PERM_REMOTE(perm)                ((perm) > ELAN3_PERM_NOREMOTE)
+
+#define ELAN3_PERM_READONLY(perm)      ((perm) == ELAN3_PERM_NOREMOTE ? ELAN3_PERM_LOCAL_READ : \
+                                        (perm) > ELAN3_PERM_REMOTEREAD ? ELAN3_PERM_READ : (perm))
+#if PAGE_SHIFT == 12
+#  define ELAN3_PAGE_SHIFT     12
+#else
+#  define ELAN3_PAGE_SHIFT     13
+#endif
+
+#define ELAN3_PAGE_SIZE                (1 << ELAN3_PAGE_SHIFT)
+#define ELAN3_PAGE_OFFSET      (ELAN3_PAGE_SIZE-1)
+#define ELAN3_PAGE_MASK                (~ELAN3_PAGE_OFFSET)
+
+#if ELAN3_PAGE_SHIFT == 13
+#  define ELAN3_L3_SHIFT               5
+#else
+#  define ELAN3_L3_SHIFT               6
+#endif
+#define ELAN3_L2_SHIFT         6
+#define ELAN3_L1_SHIFT         8
+
+/* Number of entries in a given level ptbl */
+#define ELAN3_L3_ENTRIES               (1 << ELAN3_L3_SHIFT)
+#define ELAN3_L2_ENTRIES               (1 << ELAN3_L2_SHIFT)
+#define ELAN3_L1_ENTRIES               (1 << ELAN3_L1_SHIFT)
+
+/* Virtual address spanned by each entry */
+#define ELAN3_L3_SIZE          (1 << (ELAN3_PAGE_SHIFT))
+#define ELAN3_L2_SIZE          (1 << (ELAN3_L3_SHIFT+ELAN3_PAGE_SHIFT))
+#define ELAN3_L1_SIZE          (1 << (ELAN3_L3_SHIFT+ELAN3_L2_SHIFT+ELAN3_PAGE_SHIFT))
+
+/* Virtual address size of page table */
+#define ELAN3_L1_PTSIZE          (ELAN3_L1_ENTRIES * ELAN3_L1_SIZE)
+#define ELAN3_L3_PTSIZE                (ELAN3_L3_ENTRIES * ELAN3_L3_SIZE)
+#define ELAN3_L2_PTSIZE                (ELAN3_L2_ENTRIES * ELAN3_L2_SIZE)
+
+/* Mask for offset into page table */
+#define ELAN3_L1_PTOFFSET        ((ELAN3_L1_SIZE*ELAN3_L1_ENTRIES)-1)
+#define ELAN3_L3_PTOFFSET      ((ELAN3_L3_SIZE*ELAN3_L3_ENTRIES)-1)
+#define ELAN3_L2_PTOFFSET      ((ELAN3_L2_SIZE*ELAN3_L2_ENTRIES)-1)
+
+#define ELAN3_L1_INDEX(addr)   (((E3_Addr) (addr) & 0xFF000000) >> (ELAN3_L2_SHIFT+ELAN3_L3_SHIFT+ELAN3_PAGE_SHIFT))
+#define ELAN3_L2_INDEX(addr)   (((E3_Addr) (addr) & 0x00FD0000) >> (ELAN3_L3_SHIFT+ELAN3_PAGE_SHIFT))
+#define ELAN3_L3_INDEX(addr)   (((E3_Addr) (addr) & 0x0003F000) >> ELAN3_PAGE_SHIFT)
+
+#define        ELAN3_L1_BASE(addr)     (((E3_Addr)(addr)) & 0x00000000)
+#define        ELAN3_L2_BASE(addr)     (((E3_Addr)(addr)) & 0xFF000000)
+#define        ELAN3_L3_BASE(addr)     (((E3_Addr)(addr)) & 0xFFFC0000)
+
+/* Convert a page table pointer entry to the PT */
+#define PTP_TO_PT_PADDR(ptp)   ((E3_Addr)(ptp & 0xFFFFFFFC))
+
+#ifdef __KERNEL__
+/*
+ * incompatible access for permission macro.
+ */
+extern  u_char  elan3mmu_permissionTable[8];
+#define ELAN3_INCOMPAT_ACCESS(perm,access) (! (elan3mmu_permissionTable[(perm)>>ELAN3_PTE_PERM_SHIFT] & (1 << (access))))
+
+#define elan3_readptp(dev, ptp)                (elan3_sdram_readl (dev, ptp))
+#define elan3_writeptp(dev, ptp, value)        (elan3_sdram_writel (dev, ptp, value))
+#define elan3_readpte(dev, pte)                (elan3_sdram_readq (dev, pte))
+#define elan3_writepte(dev,pte, value) (elan3_sdram_writeq (dev, pte, value))
+
+#define elan3_invalidatepte(dev, pte)  (elan3_sdram_writel (dev, pte, 0))
+#define elan3_modifypte(dev,pte,new)   (elan3_sdram_writel (dev, pte, (int) (new)))
+#define elan3_clrref(dev,pte)          (elan3_sdram_writeb (dev, pte + 7)
+
+#endif /* __KERNEL__ */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __ELAN3_PTE_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/spinlock.h
===================================================================
--- linux-2.4.21.orig/include/elan3/spinlock.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/spinlock.h       2005-06-01 23:12:54.729419344 -0400
@@ -0,0 +1,195 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_SPINLOCK_
+#define _ELAN3_SPINLOCK_
+
+#ident "$Id: spinlock.h,v 1.31 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/spinlock.h,v $*/
+
+/*
+ * This spinlock is designed for main/elan processor interactions.
+ * The lock is split over Elan/Main memory in such a way that
+ * we don't end up busy-polling over the PCI.
+ * In the Elan memory we have two words; one is a sequence number
+ * and the other is a lock word for main.
+ * In main memory we have a copy of the sequence number which main polls when it is
+ * waiting for the Elan to drop the lock. Main polls this word until it becomes
+ * equal to the sequence number it sampled.
+ * The Elan drops the lock by writing the current sequence number to main memory.
+ * It is coded to always give priority to the Elan thread, and so when both go for the
+ * lock, main will back off first.
+ *
+ * 18/3/98
+ * This has been extended to avoid a starvation case where both the main and thread claim the
+ * lock and so both backoff (thread does a break). So now, main attempts to claim the
+ * lock by writing 'mainLock' then samples the 'sl_seq' and if it has the lock
+ * it sets 'mainGotLock'. The thread will now see the 'sl_mainLock' set, but will only
+ * backoff with a c_break_busywait() if 'mainGotLock' is set too.
+ */
+typedef struct elan3_spinlock_elan {
+    union {
+       volatile E3_uint64      mainLocks;              /* main writes this dble word */
+       struct {
+           volatile E3_uint32  mainLock;               /* main wants a lock */
+           volatile E3_uint32  mainGotLock;            /* main has the lock */
+       } s;
+    } sl_u;
+    volatile E3_uint32         sl_seq;                 /* thread owns this word */
+    volatile E3_uint32         sl_mainWait;            /* performance counter */
+    volatile E3_uint32         sl_elanWait;            /* performance counter */
+    volatile E3_uint32         sl_elanBusyWait;        /* performance counter */
+    /* NOTE: The lock/seq words must be within the same 32-byte Elan cache-line */
+    E3_uint64                   sl_pad[5];             /* pad to 64-bytes */
+} ELAN3_SPINLOCK_ELAN;
+
+#define sl_mainLocks sl_u.mainLocks
+#define sl_mainLock  sl_u.s.mainLock
+#define sl_mainGotLock sl_u.s.mainGotLock
+
+#define SL_MAIN_RECESSIVE      1
+#define SL_MAIN_DOMINANT       2
+
+/* Declare this as a main memory cache block for efficiency */
+typedef union elan3_spinlock_main {
+    volatile E3_uint32         sl_seq;                 /* copy of seq number updated by Elan */
+    volatile E3_uint32         sl_Int32[E3_BLK_SIZE/sizeof (E3_uint32)];
+} ELAN3_SPINLOCK_MAIN;
+
+/* Main/Main or Elan/Elan lock word */
+typedef volatile int   ELAN3_SPINLOCK;
+
+#ifdef __ELAN3__
+
+/* Main/Elan interlock */
+
+#define ELAN3_ME_SPINENTER(SLE,SL) do {\
+                       asm volatile ("! elan3_spinlock store barrier");\
+                       (SLE)->sl_seq++; \
+                       if ((SLE)->sl_mainLock) \
+                         elan3_me_spinblock(SLE, SL);\
+                       asm volatile ("! elan3_spinlock store barrier");\
+               } while (0)
+#define ELAN3_ME_SPINEXIT(SLE,SL) do {\
+                       asm volatile ("! elan3_spinlock store barrier");\
+                       (SL)->sl_seq = (SLE)->sl_seq;\
+                       asm volatile ("! elan3_spinlock store barrier");\
+               } while (0)
+
+
+/* Elan/Elan interlock */
+#define ELAN3_SPINENTER(L)     do {\
+                          asm volatile ("! store barrier");\
+                          if (c_swap ((L), 1)) elan3_spinenter(L);\
+                          asm volatile ("! store barrier");\
+                       } while (0)
+#define ELAN3_SPINEXIT(L)      do {\
+                          asm volatile ("! store barrier");\
+                          c_swap((L), 0);\
+                          asm volatile ("! store barrier");\
+                       } while (0)
+
+extern void elan3_me_spinblock (ELAN3_SPINLOCK_ELAN *sle, ELAN3_SPINLOCK_MAIN *sl);
+extern void elan3_spinenter (ELAN3_SPINLOCK *l);
+
+#else                     
+
+/* Main/Elan interlock */
+#ifdef DEBUG
+#define ELAN3_ME_SPINENTER(SDRAM,SLE,SL) do {\
+                       register E3_int32 maxLoops = 0x7fffffff;        \
+                       register E3_uint32 seq;\
+                       elan3_write32_sdram(SDRAM, (SLE) + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), SL_MAIN_RECESSIVE); \
+                       MEMBAR_STORELOAD(); \
+                       seq = elan3_read32_sdram(SDRAM, (SLE) + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)); \
+                       while (seq != (SL)->sl_seq) {\
+                           elan3_write32_sdram(SDRAM, (SLE) + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), 0); \
+                           while ((SL)->sl_seq == (seq-1) && maxLoops--) ; \
+                           if (maxLoops < 0) { \
+                               printf("Failed to get ME lock %lx/%lx seq %d sle_seq %d sl_seq %d\n", \
+                                      SL, SLE, seq, \
+                                      elan3_read32_sdram(SDRAM, (SLE) + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)), \
+                                      (SL)->sl_seq); \
+                           } \
+                           elan3_write32_sdram(SDRAM, (SLE) + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), SL_MAIN_RECESSIVE); \
+                           MEMBAR_STORELOAD(); \
+                           seq = elan3_read32_sdram(SDRAM, (SLE) + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)); \
+                       }\
+                       elan3_write32_sdram(SDRAM, (SLE) + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainGotLock), 1); \
+                       MEMBAR_LOADLOAD();\
+               } while (0)
+#else
+#define ELAN3_ME_SPINENTER(SDRAM,SLE,SL) do {\
+                       register E3_uint32 seq;\
+                       elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), SL_MAIN_RECESSIVE); \
+                       MEMBAR_STORELOAD(); \
+                       seq = elan3_read32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)); \
+                       while (seq != (SL)->sl_seq) {\
+                           elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), 0); \
+                           while ((SL)->sl_seq == (seq-1)) ; \
+                           elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), SL_MAIN_RECESSIVE); \
+                           MEMBAR_STORELOAD(); \
+                           seq = elan3_read32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)); \
+                       }\
+                       elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainGotLock), 1); \
+                       MEMBAR_LOADLOAD();\
+               } while (0)
+#endif
+#define ELAN3_ME_FORCEENTER(SDRAM,SLE,SL) do { \
+       register E3_uint32 seq; \
+       MEMBAR_STORELOAD(); \
+       elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), SL_MAIN_DOMINANT); \
+       MEMBAR_STORELOAD(); \
+       seq = elan3_read32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)); \
+       while (seq != (SL)->sl_seq) \
+       { \
+               /* NOTE: we MUST call elan3_usecspin here for kernel comms */\
+               while ((SL)->sl_seq == (seq)-1) \
+                       elan3_usecspin (1); \
+               seq = elan3_read32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)); \
+       } \
+       elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainGotLock), 1); \
+       MEMBAR_LOADLOAD(); \
+} while (0)
+
+#define ELAN3_ME_TRYENTER(SDRAM,SLE,SL,SEQ) do { \
+    elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLock), SL_MAIN_RECESSIVE); \
+    MEMBAR_STORELOAD(); \
+    SEQ = elan3_read32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_seq)); \
+} while (0)
+
+#define ELAN3_ME_CHECKENTER(SDRAM,SLE,SL,SEQ) do { \
+    if ((SEQ) == ((SL)->sl_seq)) { \
+        elan3_write32_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainGotLock), 1); \
+        MEMBAR_LOADLOAD();\
+    } \
+    else ELAN3_ME_SPINENTER(SLE,SL); \
+} while (0)
+       
+#define ELAN3_ME_SPINEXIT(SDRAM,SLE,SL) do {\
+                       MEMBAR_STORESTORE(); \
+                       elan3_write64_sdram(SDRAM, SLE + offsetof(ELAN3_SPINLOCK_ELAN, sl_mainLocks), 0); \
+                       MEMBAR_STORESTORE(); \
+               } while (0)
+
+
+/* Main/Main */
+#define ELAN3_SPINENTER(L)     do {\
+                          while (c_swap ((L), 1)) ; \
+                       } while (0)
+#define ELAN3_SPINEXIT(L)      do {\
+                          c_swap((L), 0);\
+                       } while (0)
+#endif /* _ELAN3_ */
+
+#endif /* _ELAN3_SPINLOCK_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/thread.h
===================================================================
--- linux-2.4.21.orig/include/elan3/thread.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/thread.h 2005-06-01 23:12:54.730419192 -0400
@@ -0,0 +1,137 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_THREAD_H
+#define _ELAN3_THREAD_H
+
+#ident "$Id: thread.h,v 1.17 2002/08/09 11:23:34 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/thread.h,v $*/
+
+/* Alignment for a stack frame */
+#define E3_STACK_ALIGN         (64)
+
+typedef struct _E3_Frame {
+    E3_uint32         fr_local[8];             /* saved locals (not used) */
+    E3_uint32         fr_arg[6];               /* saved arguements o0 -> o5 */
+    E3_Addr           fr_savefp;               /* saved frame pointer o6 */
+    E3_Addr           fr_savepc;               /* saved program counter o7 */
+    E3_Addr           fr_stret;                /* stuct return addr */
+    E3_uint32         fr_argd[6];              /* arg dump area */
+    E3_uint32         fr_argx[1];              /* array of args past the sixth */
+} E3_Frame;
+
+typedef struct _E3_Stack {
+    E3_uint32          Locals[8];
+    E3_uint32          Ins[8];
+    E3_uint32          Globals[8];
+    E3_uint32          Outs[8];
+} E3_Stack;
+
+typedef struct _E3_OutsRegs {
+   E3_uint32   o[8];                           /* o6 == pc, o7 == fptr */
+} E3_OutsRegs;
+
+/*
+ * "Magic" value for stack pointer to be ignored.
+ */
+#define VanishingStackPointer  0x42
+
+
+/*
+ * When the Elan traps the N & Z CC bits are held in the NPC
+ * and the V & C bits are in the PC
+ */
+#define PSR_C_BIT      (1)
+#define PSR_V_BIT      (2)
+#define PSR_Z_BIT      (1)
+#define PSR_N_BIT      (2)
+#define CC_MASK                (3)
+#define PC_MASK        (~3)
+#define SP_MASK                (~3)
+
+/*
+ * Threads processor Opcodes.
+ */
+#define OPCODE_MASK            (0xC1F80000)
+#define OPCODE_IMM             (1 << 13)
+
+#define OPCODE_CLASS(instr)    ((instr) & 0xC0000000)
+#define OPCODE_CLASS_0         0x00000000
+#define OPCODE_CLASS_1         0x40000000
+#define OPCODE_CLASS_2         0x80000000
+#define OPCODE_CLASS_3         0xC0000000
+
+#define OPCODE_CPOP            0x81B00000
+#define OPCODE_Ticc            0x81D00000
+
+#define OPCODE_FCODE_SHIFT     19
+#define OPCODE_FCODE_MASK      0x1f
+#define OPCODE_NOT_ALUOP       0x01000000
+
+#define OPCODE_SLL             0x81280000
+#define OPCODE_SRL             0x81300000
+#define OPCODE_SRA             0x81380000
+
+#define OPCODE_OPEN            0x81600000
+#define OPCODE_CLOSE           0x81680000
+#define OPCODE_BREAKTEST       0x81700000
+
+#define OPCODE_BREAK           0x81a00000
+#define OPCODE_SUSPEND         0x81a80000
+#define OPCODE_WAIT            0x81b00000
+
+#define OPCODE_JMPL            0x81c00000
+
+#define OPCODE_LD              0xC0000000
+#define OPCODE_LDD             0xC0180000
+
+#define OPCODE_LDBLOCK16       0xC0900000
+#define OPCODE_LDBLOCK32       0xC0800000
+#define OPCODE_LDBLOCK64       0xC0980000
+
+#define OPCODE_ST              0xC0200000
+#define OPCODE_STD             0xC0380000
+
+#define OPCODE_SWAP            0xC0780000
+
+#define OPCODE_STBLOCK16       0xC0b00000
+#define OPCODE_STBLOCK32       0xC0a00000
+#define OPCODE_STBLOCK64       0xC0b80000
+
+#define OPCODE_CLASS0_MASK     0xC1C00000
+#define OPCODE_SETHI           0x01000000
+#define OPCODE_BICC            0x00800000
+#define OPCODE_SENDREG         0x01800000
+#define OPCODE_SENDMEM         0x01c00000
+
+#define OPCODE_BICC_BN         0x00000000
+#define OPCODE_BICC_BE         0x02000000
+#define OPCODE_BICC_BLE                0x04000000
+#define OPCODE_BICC_BL         0x06000000
+#define OPCODE_BICC_BLEU       0x08000000
+#define OPCODE_BICC_BCS                0x0A000000
+#define OPCODE_BICC_BNEG       0x0C000000
+#define OPCODE_BICC_BVS                0x0E000000
+
+#define OPCODE_BICC_MASK       0x0E000000
+#define OPCODE_BICC_ANNUL      0x20000000
+
+#define INSTR_RS2(instr)       (((instr) >>  0) & 0x1F)
+#define INSTR_RS1(instr)       (((instr) >> 14) & 0x1F)
+#define INSTR_RD(instr)                (((instr) >> 25) & 0x1F)
+#define INSTR_IMM(instr)       (((instr) & 0x1000) ? ((instr) & 0xFFF) | 0xFFFFF000 : (instr) & 0xFFF)
+
+#define Ticc_COND(instr)       INSTR_RD(instr)
+#define Ticc_TA                        8
+
+#endif /* _ELAN3_THREAD_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/threadlinkage.h
===================================================================
--- linux-2.4.21.orig/include/elan3/threadlinkage.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/threadlinkage.h  2005-06-01 23:12:54.730419192 -0400
@@ -0,0 +1,103 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_THREADLINKAGE_H
+#define        __ELAN3_THREADLINKAGE_H
+
+#ident "$Id: threadlinkage.h,v 1.6 2002/08/09 11:23:34 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/threadlinkage.h,v $*/
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(_ASM) || defined(__LANGUAGE_ASSEMBLY__)
+
+/*
+ * Macro to define weak symbol aliases. These are similar to the ANSI-C
+ *     #pragma weak name = _name
+ * except a compiler can determine type. The assembler must be told. Hence,
+ * the second parameter must be the type of the symbol (i.e.: function,...)
+ */
+#define        ANSI_PRAGMA_WEAK(sym, stype)    \
+       .weak   sym; \
+       .type sym, #stype; \
+/* CSTYLED */ \
+sym    = _/**/sym
+
+/*
+ * ENTRY provides the standard procedure entry code
+ */
+#define        ENTRY(x) \
+       .section        ".text"; \
+       .align  4; \
+       .global x; \
+x:
+
+/*
+ * ENTRY2 is identical to ENTRY but provides two labels for the entry point.
+ */
+#define        ENTRY2(x, y) \
+       .section        ".text"; \
+       .align  4; \
+       .global x, y; \
+/* CSTYLED */ \
+x:     ; \
+y:
+
+
+/*
+ * ALTENTRY provides for additional entry points.
+ */
+#define        ALTENTRY(x) \
+       .global x; \
+x:
+
+/*
+ * DGDEF and DGDEF2 provide global data declarations.
+ *
+ * DGDEF provides a word aligned word of storage.
+ *
+ * DGDEF2 allocates "sz" bytes of storage with **NO** alignment.  This
+ * implies this macro is best used for byte arrays.
+ *
+ * DGDEF3 allocates "sz" bytes of storage with "algn" alignment.
+ */
+#define        DGDEF2(name, sz) \
+       .section        ".data"; \
+       .global name; \
+       .size   name, sz; \
+name:
+
+#define        DGDEF3(name, sz, algn) \
+       .section        ".data"; \
+       .align  algn; \
+       .global name; \
+       .size   name, sz; \
+name:
+
+#define        DGDEF(name)     DGDEF3(name, 4, 4)
+
+/*
+ * SET_SIZE trails a function and set the size for the ELF symbol table.
+ */
+#define        SET_SIZE(x) \
+       .size   x, (.-x)
+
+#endif /* _ASM || __LANGUAGE_ASSEMBLY__ */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __ELAN3_THREADLINKAGE_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/threadsyscall.h
===================================================================
--- linux-2.4.21.orig/include/elan3/threadsyscall.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/threadsyscall.h  2005-06-01 23:12:54.730419192 -0400
@@ -0,0 +1,64 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN3_SYSCALL_H
+#define __ELAN3_SYSCALL_H
+
+#ident "$Id: threadsyscall.h,v 1.12 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/threadsyscall.h,v $*/
+
+/* 
+ * This file contains the system calls supported from the Elan.
+ */
+#define ELAN3_DEBUG_TRAPNUM    5       /* thread debugging trap */
+#define ELAN3_ABORT_TRAPNUM    6       /* bad abort trap */
+#define ELAN3_ELANCALL_TRAPNUM 7       /* elansyscall trap */
+#define ELAN3_SYSCALL_TRAPNUM  8       /* new syscall trap */
+
+#define ELAN3_T_SYSCALL_CODE   0       /* offsets in struct elan3_t_syscall */
+#define ELAN3_T_SYSCALL_ERRNO  4
+
+#define ELAN3_SYS_open         1
+#define ELAN3_SYS_close                2
+#define ELAN3_SYS_write                3
+#define ELAN3_SYS_read         4
+#define ELAN3_SYS_poll         5
+#define ELAN3_SYS_ioctl                6
+#define ELAN3_SYS_lseek                7
+#define ELAN3_SYS_mmap         8
+#define ELAN3_SYS_munmap       9
+#define ELAN3_SYS_kill         10
+#define ELAN3_SYS_getpid       11
+
+#if !defined(SYS_getpid) && defined(__NR_getxpid) 
+#define SYS_getpid __NR_getxpid                /* for linux */
+#endif
+
+#if !defined(_ASM) && !defined(__LANGUAGE_ASSEMBLY__)
+
+extern int     elan3_t_open (const char *, int, ...);
+extern ssize_t elan3_t_write (int, const void *, unsigned);
+extern ssize_t elan3_t_read(int, void *, unsigned);
+extern int     elan3_t_ioctl(int, int, ...);
+extern int     elan3_t_close(int);
+extern off_t   elan3_t_lseek(int filedes, off_t offset, int whence);
+
+extern caddr_t elan3_t_mmap(caddr_t, size_t, int, int, int, off_t);
+extern int     elan3_t_munmap(caddr_t, size_t);
+
+extern int     elan3_t_getpid(void);
+extern void    elan3_t_abort(char *str);
+
+#endif /* !_ASM && ! __LANGUAGE_ASSEMBLY__ */
+
+#endif /* __ELAN3_SYSCALL_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/trtype.h
===================================================================
--- linux-2.4.21.orig/include/elan3/trtype.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/trtype.h 2005-06-01 23:12:54.731419040 -0400
@@ -0,0 +1,116 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN3_TRTYPE_H
+#define _ELAN3_TRTYPE_H
+
+#ident "$Id: trtype.h,v 1.13 2002/08/09 11:23:34 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/trtype.h,v $ */
+
+/*<15> ackNow  */
+#define TR_SENDACK     (1 << 15)
+
+#define TR_SIZE_SHIFT  12
+#define TR_SIZE_MASK   7
+
+/*<14:12> Size 0, 1, 2, 4, 8, 16, 32, 64  Double Words
+          Bit 14 is forced to zero currently so that only size 0, 1, 2, 4 are
+         allowed    */
+
+#define TR_SIZE0       (0 << TR_SIZE_SHIFT)
+#define TR_SIZE1       (1 << TR_SIZE_SHIFT)
+#define TR_SIZE2       (2 << TR_SIZE_SHIFT)
+#define TR_SIZE4       (3 << TR_SIZE_SHIFT)
+#define TR_SIZE8       (4 << TR_SIZE_SHIFT)
+
+#define TR_64_BIT_ADDR (1 << 11)
+#define TR_LAST_TRANS  (1 << 10)
+
+#define TR_WRITEBLOCK_BIT      (1 << 9)
+#define TR_WRITEBLOCK          (TR_WRITEBLOCK_BIT | TR_SIZE8)
+
+
+#define TR_WRITEBLOCK_SIZE     64
+
+/*
+ * write-block
+ */
+/*     WriteBlock      <8:7>   Data type
+                       <6:0>   Part write size */
+#define TR_TYPE_SHIFT  7
+#define TR_TYPE_MASK   ((1 << 2) - 1)
+
+#define TR_TYPE_BYTE   0
+#define TR_TYPE_SHORT  1
+#define TR_TYPE_WORD   2
+#define TR_TYPE_DWORD  3
+
+#define TR_PARTSIZE_MASK ((1 << 7) -1)
+
+#define TR_WAIT_FOR_EOP        (1 << 8)
+
+/*
+ * trace-route format 
+ */
+#define TR_TRACEROUTE0_CHANID(val)             ((val) & 1)                     /* 0     Chan Id */
+#define TR_TRACEROUTE0_LINKID(val)             (((val) >> 1) & 7)              /* 1:3   Link Id */
+#define TR_TRACEROUTE0_REVID(val)              (((val) >> 4) & 7)              /* 4:6   Revision ID */
+#define TR_TRACEROUTE0_BCAST_TOP_PIN(val)      (((val) >> 7) & 1)              /* 7     Broadcast Top Pin (REV B) */
+#define TR_TRACEROUTE0_LNR(val)                        ((val) >> 8)                    /* 8:15  Global Link Not Ready */
+
+#define TR_TRACEROUTE1_PRIO(val)               ((val & 0xF))                   /* 0:3   Arrival Priority (REV A) */
+#define TR_TRACEROUTE1_AGE(val)                        (((val) >> 4) & 0xF)            /* 4:7   Priority Held(Age) (REV A) */
+#define TR_TRACEROUTE1_ROUTE_SELECTED(val)     ((val) & 0xFF)                  /* 0:7   Arrival age (REV B) */
+#define TR_TRACEROUTE1_BCAST_TOP(val)          (((val) >> 8) & 7)              /* 8:10  Broadcast Top */
+#define TR_TRACEROUTE1_ADAPT(val)              (((val) >> 12) & 3)             /* 12:13 This Adaptive Value (REV A) */
+#define TR_TRACEROUTE1_BCAST_BOT(val)          (((val) >> 12) & 7)             /* 12:14 Broadcast Bottom (REV B) */
+
+#define TR_TRACEROUTE2_ARRIVAL_AGE(val)                ((val) & 0xF)                   /* 0:3   Arrival Age (REV B) */
+#define TR_TRACEROUTE2_CURR_AGE(val)           (((val) >> 4) & 0xF)            /* 4:7   Current Age (REV B) */
+#define TR_TRACEROUTE2_BUSY(val)               (((val) >> 8) & 0xFF)           /* 8:15  Busy (REV B) */
+
+#define TR_TRACEROUTE_SIZE     32
+#define TR_TRACEROUTE_ENTRIES  (TR_TRACEROUTE_SIZE/2)
+
+/*
+ * non-write block
+ */
+#define TR_OPCODE_MASK         (((1 << 8) - 1) |                       \
+                                (TR_SIZE_MASK << TR_SIZE_SHIFT) |      \
+                                TR_WRITEBLOCK_BIT)
+
+#define TR_NOP_TRANS           (0x0 | TR_SIZE0)
+#define TR_SETEVENT            (0x0 | TR_SIZE0 | TR_SENDACK | TR_LAST_TRANS)
+#define TR_REMOTEDMA           (0x1 | TR_SIZE4 | TR_SENDACK | TR_LAST_TRANS)
+#define TR_LOCKQUEUE           (0x2 | TR_SIZE0)
+#define TR_UNLOCKQUEUE         (0x3 | TR_SIZE0 | TR_SENDACK | TR_LAST_TRANS)
+
+#define TR_SENDDISCARD         (0x4 | TR_SIZE0)
+#define TR_TRACEROUTE          (0x5 | TR_SIZE4)
+
+#define TR_DMAIDENTIFY         (0x6 | TR_SIZE0)
+#define TR_THREADIDENTIFY      (0x7 | TR_SIZE1)
+
+#define TR_GTE                 (0x8 | TR_SIZE1)
+#define TR_LT                  (0x9 | TR_SIZE1)
+#define TR_EQ                  (0xA | TR_SIZE1)
+#define TR_NEQ                 (0xB | TR_SIZE1)
+
+#define TR_WRITEWORD           (0xC | TR_SIZE1)
+#define TR_WRITEDOUBLEWORD     (0xD | TR_SIZE1)
+#define TR_TESTANDWRITE        (0xE | TR_SIZE1)
+#define TR_ATOMICADDWORD       (0xF | TR_SIZE1 | TR_SENDACK | TR_LAST_TRANS)
+#define TR_OPCODE_TYPE_MASK    0xff
+
+
+#endif /* notdef _ELAN3_TRTYPE_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/urom_addrs.h
===================================================================
--- linux-2.4.21.orig/include/elan3/urom_addrs.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/urom_addrs.h     2005-06-01 23:12:54.731419040 -0400
@@ -0,0 +1,262 @@
+#define MI_WaitForRemoteDescRead 0x0
+#define MI_WaitForRemoteDescRead2 0x1
+#define MI_WaitForRemoteDescRead2_seq1 0x2
+#define MI_SendRemoteDmaRoutes 0x3
+#define MI_IProcTrapped 0x4
+#define MI_DProcTrapped 0x5
+#define MI_CProcTrapped 0x6
+#define MI_TProcTrapped 0x7
+#define MI_TestWhichDmaQueue 0x8
+#define MI_TestWhichDmaQueue_seq1 0x9
+#define MI_InputRemoteDmaUpdateBPtr 0xa
+#define MI_FixupQueueContextAndRemoteBit 0xb
+#define MI_FixupQueueContextAndRemoteBit_seq1 0xc
+#define MI_FixupQueueContextAndRemoteBit_seq2 0xd
+#define MI_FixupQueueContextAndRemoteBit_seq3 0xe
+#define MI_FixupQueueContextAndRemoteBit_seq4 0xf
+#define MI_RunDmaCommand 0x10
+#define MI_DoSendRemoteDmaDesc 0x11
+#define MI_DequeueNonSysCntxDma 0x12
+#define MI_WaitForRemoteDescRead1 0x13
+#define MI_RemoteDmaCommand 0x14
+#define MI_WaitForRemoteRoutes 0x15
+#define MI_DequeueSysCntxDma 0x16
+#define MI_ExecuteDmaDescriptorForQueue 0x17
+#define MI_ExecuteDmaDescriptor1 0x18
+#define MI_ExecuteDmaDescriptor1_seq1 0x19
+#define MI_ExecuteDmaDescriptor1_seq2 0x1a
+#define MI_ExecuteDmaDescriptor1_seq3 0x1b
+#define MI_GetNewSizeInProg 0x1c
+#define MI_GetNewSizeInProg_seq1 0x1d
+#define MI_FirstBlockRead 0x1e
+#define MI_ExtraFirstBlockRead 0x1f
+#define MI_UnimplementedError 0x20
+#define MI_UpdateDescriptor 0x21
+#define MI_UpdateDescriptor_seq1 0x22
+#define MI_UpdateDescriptor_seq2 0x23
+#define MI_UpdateDescriptor_seq3 0x24
+#define MI_UpdateDescriptor_seq4 0x25
+#define MI_UpdateDescriptor_seq5 0x26
+#define MI_GetNextSizeInProg 0x27
+#define MI_DoStopThisDma 0x28
+#define MI_DoStopThisDma_seq1 0x29
+#define MI_GenNewBytesToRead 0x2a
+#define MI_WaitForEventReadTy1 0x2b
+#define MI_WaitUpdateEvent 0x2c
+#define MI_WaitUpdateEvent_seq1 0x2d
+#define MI_DoSleepOneTickThenRunable 0x2e
+#define MI_RunEvent 0x2f
+#define MI_EnqueueThread 0x30
+#define MI_CheckContext0 0x31
+#define MI_EnqueueDma 0x32
+#define MI_CprocTrapping 0x33
+#define MI_CprocTrapping_seq1 0x34
+#define MI_WaitForRemoteRoutes1 0x35
+#define MI_SetEventCommand 0x36
+#define MI_DoSetEvent 0x37
+#define MI_DoRemoteSetEventNowOrTrapQueueingDma 0x38
+#define MI_DoRemoteSetEventNowOrTrapQueueingDma_seq1 0x39
+#define MI_SendRemoteDmaRoutes2 0x3a
+#define MI_WaitForRemoteRoutes2 0x3b
+#define MI_WaitEventCommandTy0 0x3c
+#define MI_DequeueNonSysCntxDma2 0x3d
+#define MI_WaitEventCommandTy1 0x3e
+#define MI_WaitEventCommandTy1_seq1 0x3f
+#define MI_DequeueNonSysCntxThread 0x40
+#define MI_DequeueSysCntxDma1 0x41
+#define MI_DequeueSysCntxThread 0x42
+#define MI_TestNonSysCntxDmaQueueEmpty 0x43
+#define MI_TestNonSysCntxDmaQueueEmpty_seq1 0x44
+#define MI_TestNonSysCntxDmaQueueEmpty_seq2 0x45
+#define MI_RunThreadCommand 0x46
+#define MI_SetEventWaitForLastAcess 0x47
+#define MI_SetEventReadWait 0x48
+#define MI_SetEventReadWait_seq1 0x49
+#define MI_TestEventType 0x4a
+#define MI_TestEventType_seq1 0x4b
+#define MI_TestEventBit2 0x4c
+#define MI_DmaDescOrBlockCopyOrChainedEvent 0x4d
+#define MI_RunThread 0x4e
+#define MI_RunThread1 0x4f
+#define MI_RunThread1_seq1 0x50
+#define MI_IncDmaSysCntxBPtr 0x51
+#define MI_IncDmaSysCntxBPtr_seq1 0x52
+#define MI_IncDmaSysCntxBPtr_seq2 0x53
+#define MI_WaitForCntxDmaDescRead 0x54
+#define MI_FillInContext 0x55
+#define MI_FillInContext_seq1 0x56
+#define MI_WriteNewDescToQueue 0x57
+#define MI_WriteNewDescToQueue_seq1 0x58
+#define MI_TestForQueueWrap 0x59
+#define MI_TestForQueueWrap_seq1 0x5a
+#define MI_TestQueueIsFull 0x5b
+#define MI_TestQueueIsFull_seq1 0x5c
+#define MI_TestQueueIsFull_seq2 0x5d
+#define MI_CheckPsychoShitFixup 0x5e
+#define MI_PsychoShitFixupForcedRead 0x5f
+#define MI_PrepareDMATimeSlice 0x60
+#define MI_PrepareDMATimeSlice_seq1 0x61
+#define MI_TProcRestartFromTrapOrTestEventBit2 0x62
+#define MI_TProcRestartFromTrapOrTestEventBit2_seq1 0x63
+#define MI_WaitForGlobalsRead 0x64
+#define MI_WaitForNPCRead 0x65
+#define MI_EventInterrupt 0x66
+#define MI_EventInterrupt_seq1 0x67
+#define MI_EventInterrupt_seq2 0x68
+#define MI_EventInterrupt_seq3 0x69
+#define MI_TestSysCntxDmaQueueEmpty 0x6a
+#define MI_TestSysCntxDmaQueueEmpty_seq1 0x6b
+#define MI_TestIfRemoteDesc 0x6c
+#define MI_DoDmaLocalSetEvent 0x6d
+#define MI_DoDmaLocalSetEvent_seq1 0x6e
+#define MI_DoDmaLocalSetEvent_seq2 0x6f
+#define MI_DmaLoop1 0x70
+#define MI_ExitDmaLoop 0x71
+#define MI_ExitDmaLoop_seq1 0x72
+#define MI_RemoteDmaTestPAckType 0x73
+#define MI_PacketDiscardOrTestFailRecIfCCis0 0x74
+#define MI_PacketDiscardOrTestFailRecIfCCis0_seq1 0x75
+#define MI_TestNackFailIsZero2 0x76
+#define MI_TestNackFailIsZero3 0x77
+#define MI_DmaFailCountError 0x78
+#define MI_TestDmaForSysCntx 0x79
+#define MI_TestDmaForSysCntx_seq1 0x7a
+#define MI_TestDmaForSysCntx_seq2 0x7b
+#define MI_TestAeqB2 0x7c
+#define MI_TestAeqB2_seq1 0x7d
+#define MI_GetNextDmaDescriptor 0x7e
+#define MI_DequeueSysCntxDma2 0x7f
+#define MI_InputSetEvent 0x80
+#define MI_PutBackSysCntxDma 0x81
+#define MI_PutBackSysCntxDma_seq1 0x82
+#define MI_PutBackSysCntxDma_seq2 0x83
+#define MI_InputRemoteDma 0x84
+#define MI_InputRemoteDma_seq1 0x85
+#define MI_WaitOneTickForWakeup1 0x86
+#define MI_SendRemoteDmaDesc 0x87
+#define MI_InputLockQueue 0x88
+#define MI_CloseTheTrappedPacketIfCCis1 0x89
+#define MI_CloseTheTrappedPacketIfCCis1_seq1 0x8a
+#define MI_PostDmaInterrupt 0x8b
+#define MI_InputUnLockQueue 0x8c
+#define MI_WaitForUnLockDescRead 0x8d
+#define MI_SendEOPforRemoteDma 0x8e
+#define MI_LookAtRemoteAck 0x8f
+#define MI_InputWriteBlockQueue 0x90
+#define MI_WaitForSpStore 0x91
+#define MI_TProcNext 0x92
+#define MI_TProcStoppedRunning 0x93
+#define MI_InputWriteBlock 0x94
+#define MI_RunDmaOrDeqNonSysCntxDma 0x95
+#define MI_ExecuteDmaDescriptorForRun 0x96
+#define MI_ConfirmQueueLock 0x97
+#define MI_DmaInputIdentify 0x98
+#define MI_TProcStoppedRunning2 0x99
+#define MI_TProcStoppedRunning2_seq1 0x9a
+#define MI_TProcStoppedRunning2_seq2 0x9b
+#define MI_ThreadInputIdentify 0x9c
+#define MI_InputIdWriteAddrAndType3 0x9d
+#define MI_IProcTrappedWriteStatus 0x9e
+#define MI_FinishTrappingEop 0x9f
+#define MI_InputTestTrans 0xa0
+#define MI_TestAeqB3 0xa1
+#define MI_ThreadUpdateNonSysCntxBack 0xa2
+#define MI_ThreadQueueOverflow 0xa3
+#define MI_RunContext0Thread 0xa4
+#define MI_RunContext0Thread_seq1 0xa5
+#define MI_RunContext0Thread_seq2 0xa6
+#define MI_RunDmaDesc 0xa7
+#define MI_RunDmaDesc_seq1 0xa8
+#define MI_RunDmaDesc_seq2 0xa9
+#define MI_TestAeqB 0xaa
+#define MI_WaitForNonCntxDmaDescRead 0xab
+#define MI_DmaQueueOverflow 0xac
+#define MI_BlockCopyEvent 0xad
+#define MI_BlockCopyEventReadBlock 0xae
+#define MI_BlockCopyWaitForReadData 0xaf
+#define MI_InputWriteWord 0xb0
+#define MI_TraceSetEvents 0xb1
+#define MI_TraceSetEvents_seq1 0xb2
+#define MI_TraceSetEvents_seq2 0xb3
+#define MI_InputWriteDoubleWd 0xb4
+#define MI_SendLockTransIfCCis1 0xb5
+#define MI_WaitForDmaRoutes1 0xb6
+#define MI_LoadDmaContext 0xb7
+#define MI_InputTestAndSetWord 0xb8
+#define MI_InputTestAndSetWord_seq1 0xb9
+#define MI_GetDestEventValue 0xba
+#define MI_SendDmaIdentify 0xbb
+#define MI_InputAtomicAddWord 0xbc
+#define MI_LoadBFromTransD0 0xbd
+#define MI_ConditionalWriteBackCCTrue 0xbe
+#define MI_WaitOneTickForWakeup 0xbf
+#define MI_SendFinalUnlockTrans 0xc0
+#define MI_SendDmaEOP 0xc1
+#define MI_GenLastAddrForPsycho 0xc2
+#define MI_FailedAckIfCCis0 0xc3
+#define MI_FailedAckIfCCis0_seq1 0xc4
+#define MI_WriteDmaSysCntxDesc 0xc5
+#define MI_TimesliceDmaQueueOverflow 0xc6
+#define MI_DequeueNonSysCntxThread1 0xc7
+#define MI_DequeueNonSysCntxThread1_seq1 0xc8
+#define MI_TestThreadQueueEmpty 0xc9
+#define MI_ClearThreadQueueIfCC 0xca
+#define MI_DequeueSysCntxThread1 0xcb
+#define MI_DequeueSysCntxThread1_seq1 0xcc
+#define MI_TProcStartUpGeneric 0xcd
+#define MI_WaitForPCload2 0xce
+#define MI_WaitForNPCWrite 0xcf
+#define MI_WaitForEventWaitAddr 0xd0
+#define MI_WaitForWaitEventAccess 0xd1
+#define MI_WaitForWaitEventAccess_seq1 0xd2
+#define MI_WaitForWaitEventDesc 0xd3
+#define MI_WaitForEventReadTy0 0xd4
+#define MI_SendCondTestFail 0xd5
+#define MI_InputMoveToNextTrans 0xd6
+#define MI_ThreadUpdateSysCntxBack 0xd7
+#define MI_FinishedSetEvent 0xd8
+#define MI_EventIntUpdateBPtr 0xd9
+#define MI_EventQueueOverflow 0xda
+#define MI_MaskLowerSource 0xdb
+#define MI_DmaLoop 0xdc
+#define MI_SendNullSetEvent 0xdd
+#define MI_SendFinalSetEvent 0xde
+#define MI_TestNackFailIsZero1 0xdf
+#define MI_DmaPacketTimedOutOrPacketError 0xe0
+#define MI_NextPacketIsLast 0xe1
+#define MI_TestForZeroLengthDma 0xe2
+#define MI_WaitForPCload 0xe3
+#define MI_ReadInIns 0xe4
+#define MI_WaitForInsRead 0xe5
+#define MI_WaitForLocals 0xe6
+#define MI_WaitForOutsWrite 0xe7
+#define MI_WaitForWaitEvWrBack 0xe8
+#define MI_WaitForLockRead 0xe9
+#define MI_TestQueueLock 0xea
+#define MI_InputIdWriteAddrAndType 0xeb
+#define MI_InputIdWriteAddrAndType2 0xec
+#define MI_ThreadInputIdentify2 0xed
+#define MI_WriteIntoTrapArea0 0xee
+#define MI_GenQueueBlockWrAddr 0xef
+#define MI_InputDiscardFreeLock 0xf0
+#define MI_WriteIntoTrapArea1 0xf1
+#define MI_WriteIntoTrapArea2 0xf2
+#define MI_ResetBPtrToBase 0xf3
+#define MI_InputDoTrap 0xf4
+#define MI_RemoteDmaCntxt0Update 0xf5
+#define MI_ClearQueueLock 0xf6
+#define MI_IProcTrappedBlockWriteData 0xf7
+#define MI_FillContextFilter 0xf8
+#define MI_IProcTrapped4 0xf9
+#define MI_RunSysCntxDma 0xfa
+#define MI_ChainedEventError 0xfb
+#define MI_InputTrappingEOP 0xfc
+#define MI_CheckForRunIfZero 0xfd
+#define MI_TestForBreakOrSuspend 0xfe
+#define MI_SwapForRunable 0xff
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/vmseg.h
===================================================================
--- linux-2.4.21.orig/include/elan3/vmseg.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/vmseg.h  2005-06-01 23:12:54.732418888 -0400
@@ -0,0 +1,75 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _VM_SEG_ELAN3_H
+#define _VM_SEG_ELAN3_H
+
+#ident "$Id: vmseg.h,v 1.20 2003/09/24 13:57:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/vmseg.h,v $*/
+
+#include <elan3/elanuregs.h>
+
+/*
+ * This segment maps Elan registers,  it is fixed size and has 8K 
+ * pages split up as follows
+ *
+ *    ----------------------------------------
+ *    |    Performance Counters (read-only)  |
+ *    ----------------------------------------
+ *    |    Flag Page (read-only)            |
+ *    ----------------------------------------
+ *    |    Command Port                             |
+ *    ----------------------------------------
+ */
+typedef volatile struct elan3_flagstats 
+{
+    u_int      CommandFlag;
+    u_int      PageFaults;
+    u_int      CProcTraps;
+    u_int      DProcTraps;
+    u_int      TProcTraps;
+    u_int      IProcTraps;
+    u_int      EopBadAcks;
+    u_int      EopResets;
+    u_int      DmaNetworkErrors;
+    u_int      DmaIdentifyNetworkErrors;
+    u_int      ThreadIdentifyNetworkErrors;
+    u_int      DmaRetries;
+    u_int      ThreadSystemCalls;
+    u_int      ThreadElanCalls;
+    u_int      LoadVirtualProcess;
+} ELAN3_FLAGSTATS;
+
+#ifdef DIGITAL_UNIX
+typedef volatile union elan3_flagpage
+{
+    u_char        Padding[8192];
+    ELAN3_FLAGSTATS Stats;
+} ELAN3_FLAGPAGE;
+
+typedef volatile struct elan3_vmseg
+{
+    E3_CommandPort CommandPort;
+    ELAN3_FLAGPAGE  FlagPage;
+    E3_User_Regs   UserRegs;
+} ELAN3_VMSEG;
+
+#define SEGELAN3_SIZE   (sizeof (ELAN3_VMSEG))
+
+#define SEGELAN3_COMMAND_PORT  0
+#define SEGELAN3_FLAG_PAGE     1
+#define SEGELAN3_PERF_COUNTERS 2
+
+#endif /* DIGITAL_UNIX */
+
+#endif /* _VM_SEG_ELAN3_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan3/vpd.h
===================================================================
--- linux-2.4.21.orig/include/elan3/vpd.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan3/vpd.h    2005-06-01 23:12:54.732418888 -0400
@@ -0,0 +1,47 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: vpd.h,v 1.5 2002/08/09 11:23:34 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan3mod/elan3/elan3/vpd.h,v $*/
+
+#ifndef __ELAN3_VPD_H
+#define __ELAN3_VPD_H
+
+#define LARGE_RESOURCE_BIT                     0x80
+
+#define SMALL_RESOURCE_COMPATIBLE_DEVICE_ID    0x3
+#define SMALL_RESOURCE_VENDOR_DEFINED          0xE
+#define SMALL_RESOURCE_END_TAG                 0xF
+
+#define LARGE_RESOURCE_STRING                  0x2
+#define LARGE_RESOURCE_VENDOR_DEFINED          0x4
+#define LARGE_RESOURCE_VITAL_PRODUCT_DATA      0x10
+
+#define VPD_PART_NUMBER                        "PN"
+#define VPD_FRU_PART_NUMBER            "FN"
+#define VPD_EC_LEVEL                   "EC"
+#define VPD_MANUFACTURE_ID             "MN"
+#define VPD_SERIAL_NUMBER              "SN"
+
+#define VPD_LOAD_ID                    "LI"
+#define VPD_ROM_LEVEL                  "RL"
+#define VPD_ALTERABLE_ROM_LEVEL                "RM"
+#define VPD_NETWORK_ADDRESS            "NA"
+#define VPD_DEVICE_DRIVER_LEVEL                "DD"
+#define VPD_DIAGNOSTIC_LEVEL           "DG"
+#define VPD_LOADABLE_MICROCODE_LEVEL   "LL"
+#define VPD_VENDOR_ID                  "VI"
+#define VPD_FUNCTION_NUMBER            "FU"
+#define VPD_SUBSYSTEM_VENDOR_ID                "SI"
+
+#endif /* __ELAN3_VPD_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/elan4/commands.h
===================================================================
--- linux-2.4.21.orig/include/elan4/commands.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/commands.h       2005-06-01 23:12:54.733418736 -0400
@@ -0,0 +1,247 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_COMMANDS_H
+#define __ELAN4_COMMANDS_H
+
+#ident "$Id: commands.h,v 1.29 2004/06/16 15:45:02 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/commands.h,v $*/
+
+/*
+ * This header file describes the command format for the Elan 4
+ *    See CommandFormat.doc
+ */
+
+/*
+ * Number of channels in traced elanlib_trace.c
+ */
+#define TRACE_MAX_CHANNELS     2
+
+/*
+ * Define encoding for the commands issued into the command queues
+ */
+#define RUN_THREAD_CMD       0x00
+#define OPEN_STEN_PKT_CMD    0x01
+#define WRITE_DWORD_CMD      0x02
+#define ADD_DWORD_CMD        0x03
+#define COPY64_CMD           0x05
+#define GUARD_CMD            0x06
+#define SET_EVENT_CMD        0x07
+#define SEND_TRANS_CMD       0x09
+#define INTERRUPT_CMD        0x0d
+#define RUN_DMA_CMD          0x0e
+#define SET_EVENTN_CMD       0x0f
+#define NOP_CMD                     0x17
+#define MAKE_EXT_CLEAN_CMD   0x37
+#define WAIT_EVENT_CMD       0x1f
+
+/*
+ * Define the portion of the data word the user is NOT
+ * allowed to use. This varies with Commmand type
+ */
+#define RUN_THREAD_CMD_MASK       0x03
+#define OPEN_STEN_PKT_CMD_MASK    0x0f
+#define WRITE_DWORD_CMD_MASK      0x07
+#define ADD_DWORD_CMD_MASK        0x07
+#define COPY64_CMD_MASK           0x0f
+#define GUARD_CMD_MASK            0x0f
+#define SET_EVENT_CMD_MASK        0x1f
+#define SEND_TRANS_CMD_MASK       0x1f
+#define INTERRUPT_CMD_MASK        0x0f
+#define RUN_DMA_CMD_MASK          0x0f
+#define SET_EVENTN_CMD_MASK       0x1f
+#define NOP_CMD_MASK             0x3f
+#define MAKE_EXT_CLEAN_MASK      0x3f
+#define WAIT_EVENT_CMD_MASK       0x1f
+
+#define COPY64_DATA_TYPE_SHIFT    0x4
+#define COPY64_DTYPE_BYTE        (0 << COPY64_DATA_TYPE_SHIFT)
+#define COPY64_DTYPE_SHORT       (1 << COPY64_DATA_TYPE_SHIFT)
+#define COPY64_DTYPE_WORD        (2 << COPY64_DATA_TYPE_SHIFT)
+#define COPY64_DTYPE_LONG        (3 << COPY64_DATA_TYPE_SHIFT)
+
+/*
+ * SET_EVENTN - word 1 has following form
+ * [63:5]      Event Address
+ * [4:0]       Part Set Value.
+ */
+#define SET_EVENT_PART_SET_MASK      0x1f
+
+/* OPEN_STEN_PKT_CMD 
+ *   [63:32]   Vproc
+ *   [31]      Use Test
+ *   [30:28]   unused
+ *   [27:21]   Test Acceptable PAck code
+ *   [20:16]   Test Ack Channel Number
+ *   [15:9]    Acceptable PAck code
+ *   [8:4]     Ack Channel Number (1 bit on Elan4)
+ *   [3:0]     Command type
+ */
+/* Acceptable PAck code */
+#define PACK_OK                        (1 << 0)
+#define PACK_TESTFAIL          (1 << 1)
+#define PACK_DISCARD           (1 << 2)
+#define RESTART_COUNT_ZERO     (1 << 3)
+#define PACK_ERROR             (1 << 7)
+#define PACK_TIMEOUT           (1 << 8)
+
+/*
+ *#ifndef USE_DIRTY_COMMANDS
+ *#define USE_DIRTY_COMMANDS
+ *#endif
+ */
+#ifdef USE_DIRTY_COMMANDS
+#define OPEN_PACKET_USED_MASK    0x00000000780f00e0ULL
+#define SEND_TRANS_USED_MASK     0xffffffff0000fff0ULL
+#define COPY64_WRITE_USED_MASK   0x000000000000000fULL
+#define MAIN_INT_USED_MASK       0x0000000000003ff0ULL
+#define GUARD_USED_MASK          0xfffffe007000fde0ULL
+#define DMA_TYPESIZE_USED_MASK   0x000000000000fff0ULL
+#define SETEVENTN_USED_MASK      0xffffffffffffffe0ULL
+#define NOP_USED_MASK            0xffffffffffffffc0ULL
+#define EXT_CLEAN_USED_MASK      0xffffffffffffffc0ULL
+#define WAIT_CNT_TYPE_USED_MASK  0x00000000fffff800ULL
+#else
+#define OPEN_PACKET_USED_MASK    0x0ULL
+#define SEND_TRANS_USED_MASK     0x0ULL
+#define COPY64_WRITE_USED_MASK   0x0ULL
+#define MAIN_INT_USED_MASK       0x0ULL
+#define GUARD_USED_MASK          0x0ULL
+#define DMA_TYPESIZE_USED_MASK   0x0ULL
+#define SETEVENTN_USED_MASK      0x0ULL
+#define NOP_USED_MASK            0x0ULL
+#define EXT_CLEAN_USED_MASK      0x0ULL
+#define WAIT_CNT_TYPE_USED_MASK  0x0ULL
+#endif
+
+#define OPEN_PACKET(chan, code, vproc) \
+       ((((chan) & 1) << 4) | (((code) & 0x7f) << 9) | ((E4_uint64)(vproc) << 32) | OPEN_STEN_PKT_CMD)
+
+#define OPEN_PACKET_TEST(chan, code, vproc, tchan, tcode) \
+       ((((chan) & 1) << 4) | (((code) & 0x7f) << 9) | ((E4_uint64)(vproc) << 32) | \
+        (((tchan) & 1) << 16) | (((tcode) & 0x7f) << 21) | (((E4_uint64) 1) << 31) | OPEN_STEN_PKT_CMD)
+
+/*
+ * GUARD_CMD
+ *   [63:41]   unused
+ *   [40]      Reset Restart Fail Count        // only performed if the Guard executes the next command.
+ *   [39:32]   New Restart Fail Count value
+ *   [31]      Use Test
+ *   [30:28]   unused
+ *   [27:21]   Test Acceptable PAck code
+ *   [20:16]   Test Ack Channel Number
+ *   [15:9]    unused
+ *   [8:4]     Ack Channel Number
+ *   [3:0]     Command type
+ */
+/* GUARD_CHANNEL(chan)
+ */
+#define GUARD_ALL_CHANNELS     ((1 << 9) | GUARD_CMD)
+#define GUARD_CHANNEL(chan)    ((((chan) & 1) << 4) | GUARD_CMD)
+#define GUARD_TEST(chan,code)  ((1ull << 31) | (((code) & 0x7f) << 21) | (((chan) & 1) << 16))
+#define GUARD_RESET(count)     ((1ull << 40) | ((((E4_uint64) count) & 0xff) << 32))
+
+#define GUARD_CHANNEL_TEST(chan,tchan,tcode) \
+       ((((chan) & 1) << 4) | (((tchan) & 1) << 16) | (((tcode) & 0x7f) << 21) | \
+        (((E4_uint64) 1) << 31) | GUARD_CMD)
+
+/*
+ * SEND_TRANS_CMD
+ * [63:32]     unused
+ * [31:16]     transaction type
+ * [15:4]      unused
+ * [3:0]       Command type
+ */
+#define SEND_TRANS(TransType)  (((TransType) << 16) | SEND_TRANS_CMD)
+
+/*
+ * Command port trace debug levels
+ */
+#define TRACE_CMD_BUFFER       0x01
+#define TRACE_CMD_TYPE         0x02
+#define TRACE_CHANNEL_OPENS    0x04
+#define TRACE_GUARDED_ATOMICS  0x08
+#define TRACE_CMD_TIMEOUT      0x10
+
+/*
+ * Commands that should be preceeded by a GUARD_CMD.
+ */
+#define IS_ATOMIC_CMD(cmd)                                                             \
+   ((cmd) == RUN_THREAD_CMD || (cmd) == ADD_DWORD_CMD || (cmd) == INTERRUPT_CMD ||     \
+    (cmd) == RUN_DMA_CMD    || (cmd) == SET_EVENT_CMD || (cmd) == SET_EVENTN_CMD ||    \
+    (cmd) == WAIT_EVENT_CMD)
+
+#ifndef _ASM
+
+/*
+ * These structures are used to build event copy command streams. They are intended to be included
+ * in a larger structure to form a self documenting command sequence that can be easily coped and manipulated.
+ */
+
+typedef struct e4_runthreadcmd
+{
+   E4_Addr     PC;
+   E4_uint64   r[6];
+} E4_RunThreadCmd;
+
+typedef E4_uint64 E4_OpenCmd;
+
+typedef struct e4_writecmd
+{
+   E4_Addr     WriteAddr;
+   E4_uint64   WriteValue;
+} E4_WriteCmd;
+
+typedef struct e4_addcmd
+{
+   E4_Addr     AddAddr;
+   E4_uint64   AddValue;
+} E4_AddCmd;
+
+typedef struct e4_copycmd
+{
+   E4_Addr     SrcAddr;
+   E4_Addr     DstAddr;
+} E4_CopyCmd;
+
+typedef E4_uint64 E4_GaurdCmd;
+typedef E4_uint64 E4_SetEventCmd;
+
+/*
+ * The data to this command must be declared as a vector after the use of this.
+ */
+typedef struct e4_sendtranscmd
+{
+   E4_Addr     Type;
+   E4_Addr     Addr;
+} E4_SendTransCmd;
+
+typedef E4_uint64 E4_IntCmd;
+
+/* The normal Dma struc can be used here. */
+
+typedef struct e4_seteventncmd
+{
+   E4_Addr     Event;
+   E4_Addr     SetCount;
+} E4_SetEventNCmd;
+
+typedef E4_uint64 E4_NopCmd;
+typedef E4_uint64 E4_MakeExtCleanCmd;
+
+typedef struct e4_waitcmd
+{
+   E4_Addr     ev_Event;
+   E4_Addr     ev_CountType;
+   E4_Addr     ev_Params[2];
+} E4_WaitCmd;
+
+#endif /* _ASM */
+
+#endif /* __ELAN4_COMMANDS_H  */
+
Index: linux-2.4.21/include/elan4/debug.h
===================================================================
--- linux-2.4.21.orig/include/elan4/debug.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/debug.h  2005-06-01 23:12:54.733418736 -0400
@@ -0,0 +1,113 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN4_ELANDEBUG_H
+#define _ELAN4_ELANDEBUG_H
+
+#ident "$Id: debug.h,v 1.19.6.1 2005/01/18 14:36:10 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/debug.h,v $ */
+
+/* values for "type" field - note a "ctxt" is permissible */
+/* and BUFFER/CONSOLE are for explict calls to elan4_debugf() */
+#define DBG_DEVICE     ((void *) 0)
+#define DBG_USER       ((void *) 1)
+
+#define DBG_BUFFER     ((void *) 62)
+#define DBG_CONSOLE    ((void *) 63)
+#define DBG_NTYPES     64
+
+/* values for "mode" field */
+#define DBG_CONFIG     0x00000001
+#define DBG_INTR       0x00000002
+#define DBG_MAININT    0x00000004
+#define DBG_SDRAM      0x00000008
+#define DBG_MMU                0x00000010
+#define DBG_REGISTER   0x00000020
+#define DBG_CQ         0x00000040
+#define DBG_NETWORK_CTX        0x00000080
+
+#define DBG_FLUSH      0x00000100
+#define DBG_FILE       0x00000200
+#define DBG_CONTROL    0x00000400
+#define DBG_MEM                0x00000800
+
+#define DBG_PERM       0x00001000
+#define DBG_FAULT      0x00002000
+#define DBG_SWAP       0x00004000
+#define DBG_TRAP       0x00008000
+#define DBG_DDCQ       0x00010000
+#define DBG_VP         0x00020000
+#define DBG_RESTART    0x00040000
+#define DBG_RESUME     0x00080000
+#define DBG_CPROC      0x00100000
+#define DBG_DPROC      0x00200000
+#define DBG_EPROC      0x00400000
+#define DBG_IPROC      0x00800000
+#define DBG_TPROC      0x01000000
+#define DBG_IOPROC     0x02000000
+#define DBG_ROUTE      0x04000000
+#define DBG_NETERR     0x08000000
+
+#define DBG_ALL                0x7FFFFFFF
+
+
+#ifdef DEBUG_PRINTF
+
+#  define PRINTF0(type,m,fmt)                  ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt)                  : (void)0)
+#  define PRINTF1(type,m,fmt,a)                        ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a)                : (void)0)
+#  define PRINTF2(type,m,fmt,a,b)              ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b)              : (void)0)
+#  define PRINTF3(type,m,fmt,a,b,c)            ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b,c)            : (void)0)
+#  define PRINTF4(type,m,fmt,a,b,c,d)          ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b,c,d)          : (void)0)
+#  define PRINTF5(type,m,fmt,a,b,c,d,e)                ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b,c,d,e)        : (void)0)
+#  define PRINTF6(type,m,fmt,a,b,c,d,e,f)      ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b,c,d,e,f)      : (void)0)
+#  define PRINTF7(type,m,fmt,a,b,c,d,e,f,g)    ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b,c,d,e,f,g)    : (void)0)
+#  define PRINTF8(type,m,fmt,a,b,c,d,e,f,g,h)  ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b,c,d,e,f,g,h)  : (void)0)
+#  define PRINTF9(type,m,fmt,a,b,c,d,e,f,g,h,i)        ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m,fmt,a,b,c,d,e,f,g,h,i): (void)0)
+#ifdef __GNUC__
+#  define PRINTF(type,m,args...)               ((elan4_debug&(m) || (type) == DBG_CONSOLE) ? elan4_debugf(type,m, ##args)              : (void)0)
+#endif
+#  define DBGCMD(type,m,cmd)                   ((elan4_debug&(m) || (type) == DBG_CONSOLE)  ? (void) (cmd) : (void) 0)
+
+#else
+
+#  define PRINTF0(type,m,fmt)                  (0)
+#  define PRINTF1(type,m,fmt,a)                        (0)
+#  define PRINTF2(type,m,fmt,a,b)              (0)
+#  define PRINTF3(type,m,fmt,a,b,c)            (0)
+#  define PRINTF4(type,m,fmt,a,b,c,d)          (0)
+#  define PRINTF5(type,m,fmt,a,b,c,d,e)                (0)
+#  define PRINTF6(type,m,fmt,a,b,c,d,e,f)      (0)
+#  define PRINTF7(type,m,fmt,a,b,c,d,e,f,g)    (0)
+#  define PRINTF8(type,m,fmt,a,b,c,d,e,f,g,h)  (0)
+#  define PRINTF9(type,m,fmt,a,b,c,d,e,f,g,h,i)        (0)
+#ifdef __GNUC__
+#  define PRINTF(type,m,args...)
+#endif
+#  define DBGCMD(type,m,cmd)                   ((void) 0)
+
+#endif /* DEBUG_PRINTF */
+
+extern unsigned   elan4_debug;
+extern unsigned   elan4_debug_toconsole;
+extern unsigned   elan4_debug_tobuffer;
+extern unsigned   elan4_debug_display_ctxt;
+extern unsigned   elan4_debug_ignore_ctxt;
+extern unsigned   elan4_debug_ignore_type;
+
+extern void      elan4_debug_init(void);
+extern void      elan4_debug_fini(void);
+extern void       elan4_debugf (void *type, int mode, char *fmt, ...);
+extern int        elan4_debug_snapshot (caddr_t ubuffer, int len);
+extern int       elan4_debug_display (void);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* _ELAN4_ELANDEBUG_H */
Index: linux-2.4.21/include/elan4/device.h
===================================================================
--- linux-2.4.21.orig/include/elan4/device.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/device.h 2005-06-01 23:12:54.735418432 -0400
@@ -0,0 +1,781 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_ELANDEV_H
+#define __ELAN4_ELANDEV_H
+
+#ident "$Id: device.h,v 1.68.2.1 2004/11/03 14:24:32 duncant Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/device.h,v $ */
+
+#include <elan/devinfo.h>
+#include <elan/capability.h>
+
+#include <elan4/pci.h>
+#include <elan4/sdram.h>
+#include <elan4/dma.h>
+#include <elan4/events.h>
+#include <elan4/registers.h>
+
+#include <elan4/mmu.h>
+#include <elan4/trap.h>
+#include <elan4/stats.h>
+#include <elan4/neterr.h>
+
+#ifdef CONFIG_MPSAS
+#include <elan4/mpsas.h>
+#endif
+
+#if defined(LINUX)
+#include <elan4/device_Linux.h>
+#elif defined(TRU64UNIX)
+#include <elan4/device_OSF1.h>
+#elif defined(SOLARIS)
+#include <elan4/device_SunOS.h>
+#endif
+
+/*
+ * Network context number allocation.
+ * [0]          neterr fixup system context
+ * [1]          kernel comms system context
+ * [2048-4095] kernel comms data contexts
+ */
+#define ELAN4_NETERR_CONTEXT_NUM       0x00                    /* network error fixup context number */
+#define ELAN4_KCOMM_CONTEXT_NUM                0x01                    /* kernel comms context number */
+#define ELAN4_KCOMM_BASE_CONTEXT_NUM   0x800                   /* kernel comms data transfer contexts */
+#define ELAN4_KCOMM_TOP_CONTEXT_NUM    0xfff
+
+#define ELAN4_SYSTEM_CONTEXT(ctx)  ((ctx) >= ELAN4_KCOMM_BASE_CONTEXT_NUM)
+
+typedef void (ELAN4_HALTFN)(struct elan4_dev *dev, void *arg);
+
+typedef struct elan4_haltop
+{
+    struct list_head    op_link;                               /* chain on a list */
+    E4_uint32          op_mask;                                /* Interrupt mask to see before calling function */
+    
+    ELAN4_HALTFN       *op_function;                           /* function to call */
+    void              *op_arg;                                 /* arguement to pass to function */
+} ELAN4_HALTOP;
+
+typedef void (ELAN4_DMA_FLUSHFN)(struct elan4_dev *dev, void *arg, int qfull);
+
+typedef struct elan4_dma_flushop
+{
+    struct list_head    op_link;                               /* chain on a list */
+    ELAN4_DMA_FLUSHFN  *op_function;                           /* function to call */
+    void              *op_arg;                                 /* arguement to pass to function */
+} ELAN4_DMA_FLUSHOP;
+
+typedef void (ELAN4_INTFN)(struct elan4_dev *dev, void *arg);
+
+typedef struct elan4_intop
+{
+    struct list_head    op_link;                               /* chain on a list */
+    ELAN4_INTFN        *op_function;                           /* function to call */
+    void              *op_arg;                                 /* arguement to pass to function */
+    E4_uint64          op_cookie;                              /* and main interrupt cookie */
+} ELAN4_INTOP;
+
+#define SDRAM_MIN_BLOCK_SHIFT  10
+#define SDRAM_NUM_FREE_LISTS   19                                      /* allows max 256 Mb block */
+#define SDRAM_MIN_BLOCK_SIZE   (1 << SDRAM_MIN_BLOCK_SHIFT)
+#define SDRAM_MAX_BLOCK_SIZE   (SDRAM_MIN_BLOCK_SIZE << (SDRAM_NUM_FREE_LISTS-1))
+
+#if PAGE_SHIFT < 13
+#define SDRAM_PAGE_SIZE                8192
+#define SDRAM_PGOFF_OFFSET     1
+#define SDRAM_PGOFF_MASK       (~SDRAM_PGOFF_OFFSET)
+#else
+#define SDRAM_PAGE_SIZE                PAGE_SIZE
+#define SDRAM_PGOFF_OFFSET     0
+#define SDRAM_PGOFF_MASK       (~SDRAM_PGOFF_OFFSET)
+#endif
+
+typedef struct elan4_sdram
+{
+    sdramaddr_t        b_base;                                 /* offset in sdram bar */
+    unsigned           b_size;                                 /* size of bank */
+    ioaddr_t           b_ioaddr;                               /* ioaddr where mapped into the kernel */
+    ELAN4_MAP_HANDLE   b_handle;                               /*    and mapping handle */
+    bitmap_t          *b_bitmaps[SDRAM_NUM_FREE_LISTS];        /* buddy allocator bitmaps */
+} ELAN4_SDRAM_BANK;
+
+/* command queue */
+typedef struct elan4_cq 
+{
+    struct elan4_cqa    *cq_cqa;                                       /* command queue allocator this belongs to */
+    unsigned            cq_idx;                                        /*  and which command queue this is */
+
+    sdramaddr_t                 cq_space;                                      /* sdram backing up command queue */
+    unsigned            cq_size;                                       /* size value */
+    unsigned            cq_perm;                                       /* permissions */
+    ioaddr_t            cq_mapping;                                    /* mapping of command queue page */
+    ELAN4_MAP_HANDLE    cq_handle;                                     /*    and mapping handle */
+} ELAN4_CQ;
+
+/* cqtype flags to elan4_alloccq() */
+#define CQ_Priority    (1 << 0)
+#define CQ_Reorder     (1 << 1)
+
+/* command queues are allocated in chunks,so that all the
+ * command ports are in a single system page */
+#define ELAN4_CQ_PER_CQA       MAX(1, (PAGESIZE/CQ_CommandMappingSize))
+
+/* maximum number of command queues per context */
+#define ELAN4_MAX_CQA          (256 / ELAN4_CQ_PER_CQA)
+
+typedef struct elan4_cqa
+{
+    struct list_head   cqa_link;                                       /* linked together */
+    bitmap_t           cqa_bitmap[BT_BITOUL(ELAN4_CQ_PER_CQA)];        /* bitmap of which are free */
+    unsigned int        cqa_type;                                      /* allocation type */
+    unsigned int       cqa_cqnum;                                      /* base cq number */
+    unsigned int       cqa_ref;                                        /* "mappings" to a queue */
+    unsigned int       cqa_idx;                                        /* index number */
+    ELAN4_CQ           cqa_cq[ELAN4_CQ_PER_CQA];                       /* command queue entries */
+} ELAN4_CQA;
+
+#define elan4_cq2num(cq)       ((cq)->cq_cqa->cqa_cqnum + (cq)->cq_idx)
+#define elan4_cq2idx(cq)       ((cq)->cq_cqa->cqa_idx * ELAN4_CQ_PER_CQA + (cq)->cq_idx)
+
+typedef struct elan4_ctxt
+{
+    struct elan4_dev      *ctxt_dev;                                   /* device we're associated with */
+    struct list_head       ctxt_link;                                  /* chained on device */
+    
+    struct elan4_trap_ops *ctxt_ops;                                   /* client specific operations */
+
+    signed                ctxt_num;                                    /* local context number */
+
+    struct list_head      ctxt_cqalist;                                /* link list of command queue allocators */
+    bitmap_t              ctxt_cqamap[BT_BITOUL(ELAN4_MAX_CQA)];       /*   bitmap for allocating cqa_idx */
+
+    ELAN4_HASH_ENTRY     **ctxt_mmuhash[2];                            /* software hash tables */
+    spinlock_t            ctxt_mmulock;                                /*   and spinlock. */
+} ELAN4_CTXT;
+
+typedef struct elan4_trap_ops
+{
+    void       (*op_eproc_trap) (ELAN4_CTXT *ctxt, E4_uint64 status);
+    void       (*op_cproc_trap) (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned cqnum);
+    void       (*op_dproc_trap) (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit);
+    void       (*op_tproc_trap) (ELAN4_CTXT *ctxt, E4_uint64 status);
+    void       (*op_iproc_trap) (ELAN4_CTXT *ctxt, E4_uint64 status, unsigned unit);
+    void       (*op_interrupt)  (ELAN4_CTXT *ctxt, E4_uint64 cookie);
+    void       (*op_neterrmsg)  (ELAN4_CTXT *ctxt, ELAN4_NETERR_MSG *msg);
+} ELAN4_TRAP_OPS;
+
+typedef struct elan4_route_table
+{
+    spinlock_t  tbl_lock;
+    unsigned   tbl_size;
+    sdramaddr_t tbl_entries;
+} ELAN4_ROUTE_TABLE;
+
+#ifdef ELAN4_LARGE_PAGE_SUPPORT
+#define NUM_HASH_TABLES                2
+#else
+#define NUM_HASH_TABLES                1
+#endif
+
+#define DEV_STASH_ROUTE_COUNT 20
+
+typedef struct elan4_route_ringbuf {
+    int start;
+    int end;
+    E4_VirtualProcessEntry routes[DEV_STASH_ROUTE_COUNT]; 
+} ELAN4_ROUTE_RINGBUF;
+
+#define elan4_ringbuf_init(ringbuf) memset(&ringbuf, 0, sizeof(ELAN4_ROUTE_RINGBUF));
+
+typedef struct elan4_dev
+{
+    ELAN4_CTXT          dev_ctxt;                                      /* context for device operations */
+
+    ELAN4_DEV_OSDEP     dev_osdep;                                     /* OS specific entries */
+
+    int                         dev_instance;                                  /* device number */
+    ELAN_DEVINFO        dev_devinfo;                                   /* device information (revision etc */
+    ELAN_POSITION       dev_position;                                  /* position connected to switch */
+    ELAN_DEV_IDX        dev_idx;                                       /* device idx registered with elanmod */
+
+    kmutex_t            dev_lock;                                      /* lock for device state/references */
+    unsigned            dev_state;                                     /* device state */
+    unsigned            dev_references;                                /*  # references */
+    unsigned            dev_features;                                  /* features supported */
+
+    ioaddr_t            dev_regs;                                      /* Mapping of device registers */
+    ELAN4_MAP_HANDLE    dev_regs_handle;
+    ioaddr_t            dev_rom;                                       /* Mapping of rom */
+    ELAN4_MAP_HANDLE    dev_rom_handle;
+    ioaddr_t            dev_i2c;                                       /* Mapping of I2C registers */
+    ELAN4_MAP_HANDLE    dev_i2c_handle;
+    
+    E4_uint64           dev_sdram_cfg;                                 /* SDRAM config value (from ROM) */
+    E4_uint64           dev_sdram_initial_ecc_val;                     /* power on ECC register value */
+    int                         dev_sdram_numbanks;                            /* # banks of sdram */
+    ELAN4_SDRAM_BANK    dev_sdram_banks[SDRAM_MAX_BANKS];              /* Mapping of sdram banks */
+    spinlock_t          dev_sdram_lock;                                /* spinlock for buddy allocator */
+    sdramaddr_t                 dev_sdram_freelists[SDRAM_NUM_FREE_LISTS];
+    unsigned            dev_sdram_freecounts[SDRAM_NUM_FREE_LISTS];
+
+    sdramaddr_t                 dev_cacheflush_space;                          /* sdram reserved for cache flush operation */
+
+    sdramaddr_t                 dev_faultarea;                                 /* fault areas for each unit */
+    sdramaddr_t                 dev_inputtraparea;                             /* trap area for trapped transactions */
+    sdramaddr_t                 dev_ctxtable;                                  /* context table (E4_ContextControlBlock) */
+    int                         dev_ctxtableshift;                             /* and size (in bits) */
+
+    E4_uint32           dev_syscontrol;                                /* copy of system control register */
+    spinlock_t          dev_syscontrol_lock;                           /*   spinlock to sequentialise modifications */
+    unsigned            dev_direct_map_pci_writes;                     /*   # counts for CONT_DIRECT_MAP_PCI_WRITES */
+
+    volatile E4_uint32  dev_intmask;                                   /* copy of interrupt mask register */
+    spinlock_t          dev_intmask_lock;                              /*   spinlock to sequentialise modifications */
+
+    /* i2c section */
+    spinlock_t          dev_i2c_lock;                                  /* spinlock for i2c operations */
+    unsigned int         dev_i2c_led_disabled;                         /* count of reasons led auto update disabled */
+
+    /* mmu section */
+    unsigned            dev_pagesizeval[NUM_HASH_TABLES];              /* page size value */
+    unsigned            dev_pageshift[NUM_HASH_TABLES];                        /* pageshift in bits. */
+    unsigned            dev_hashsize[NUM_HASH_TABLES];                 /* # entries in mmu hash table */
+    sdramaddr_t                 dev_hashtable[NUM_HASH_TABLES];                /* mmu hash table */
+    ELAN4_HASH_ENTRY   *dev_mmuhash[NUM_HASH_TABLES];                  /*   and software shadow */
+    ELAN4_HASH_ENTRY   **dev_mmufree[NUM_HASH_TABLES];                 /*   and partially free blocks */
+    ELAN4_HASH_ENTRY    *dev_mmufreelist;                              /*   and free blocks */
+    spinlock_t           dev_mmulock;
+    E4_uint16           dev_topaddr[4];                                /* top address values */
+    unsigned char       dev_topaddrvalid;
+    unsigned char       dev_topaddrmode;
+    unsigned char       dev_pteval;                                    /* allow setting of relaxed order/dont snoop attributes */
+
+    unsigned            dev_rsvd_hashmask[NUM_HASH_TABLES];
+    unsigned            dev_rsvd_hashval[NUM_HASH_TABLES];
+
+    /* run queues */
+    sdramaddr_t                 dev_comqlowpri;                                /* CProc low & high pri run queues */
+    sdramaddr_t                 dev_comqhighpri;
+
+    sdramaddr_t                 dev_dmaqlowpri;                                /* DProc,TProc,Interrupt queues */
+    sdramaddr_t                 dev_dmaqhighpri;
+    sdramaddr_t                 dev_threadqlowpri;
+    sdramaddr_t                 dev_threadqhighpri;
+    sdramaddr_t                 dev_interruptq;
+
+    E4_uint32           dev_interruptq_nfptr;                          /* cache next main interrupt fptr */
+    struct list_head     dev_interruptq_list;                          /*   list of operations to call when space in interruptq*/
+
+    /* command queue section */
+    sdramaddr_t                 dev_cqaddr;                                    /* SDRAM address of command queues */
+    unsigned            dev_cqoffset;                                  /* offset for command queue alignment constraints */
+    unsigned            dev_cqcount;                                   /* number of command queue descriptors */
+    bitmap_t           *dev_cqamap;                                    /* bitmap for allocation */
+    spinlock_t          dev_cqlock;                                    /* spinlock to protect bitmap */
+#ifdef CONFIG_MTRR
+    unsigned            dev_cqreorder;                                 /* offset for first re-ordering queue on revb */
+#endif
+
+    /* halt operation section */
+    struct list_head     dev_haltop_list;                              /* list of operations to call when units halted */
+    E4_uint32           dev_haltop_mask;                               /* mask of which ones to halt */
+    E4_uint32           dev_haltop_active;                             /* mask of which haltops are executing */
+    spinlock_t          dev_haltop_lock;                               /*    and their spinlock */
+
+    struct {
+       struct list_head list;                                          /* list of halt operations for DMAs */
+       ELAN4_CQ        *cq;                                            /*   and command queue's */
+       ELAN4_INTOP      intop;                                         /*   and main interrupt op */
+       E4_uint64        status;                                        /*   status register (when waiting for intop)*/
+    }                   dev_dma_flushop[2];
+
+    unsigned            dev_halt_all_count;                            /* count of reasons to halt all units */
+    unsigned            dev_halt_lowpri_count;                         /* count of reasons to halt lowpri queues */
+    unsigned            dev_halt_cproc_count;                          /* count of reasons to halt command processor */
+    unsigned            dev_halt_dproc_count;                          /* count of reasons to halt dma processor */
+    unsigned            dev_halt_tproc_count;                          /* count of reasons to halt thread processor */
+    unsigned            dev_discard_all_count;                         /* count of reasons to discard all packets */
+    unsigned            dev_discard_lowpri_count;                      /* count of reasons to discard non-system packets */
+    unsigned            dev_discard_highpri_count;                     /* count of reasons to discard system packets */
+
+    E4_uint32           dev_schedstatus;                               /* copy of schedule status register */
+
+    /* local context allocation section */
+    spinlock_t          dev_ctxlock;                                   /* spinlock to protect bitmap */
+    bitmap_t           *dev_ctxmap;                                    /* bitmap for local context allocation */
+
+    spinlock_t          dev_ctxt_lock;                                 /* spinlock to protect context list */
+    struct list_head     dev_ctxt_list;                                        /* linked list of contexts */
+
+    /* locks to sequentialise interrupt handling */
+    spinlock_t          dev_trap_lock;                                 /* spinlock while handling a trap */
+    spinlock_t          dev_requeue_lock;                              /* spinlock sequentialising cproc requeue */
+
+    /* error rate interrupt section */
+    long                dev_error_time;                                /* lbolt at start of sampling period */
+    unsigned            dev_errors_per_period;                         /* errors so far this sampling period */
+    timer_fn_t          dev_error_timeoutid;                           /* timeout to re-enable error interrupts */
+    timer_fn_t          dev_linkerr_timeoutid;                         /* timeout to clear link error led */
+
+    /* kernel threads */
+    unsigned            dev_stop_threads:1;                            /* kernel threads should exit */
+
+    /* main interrupt thread */
+    kcondvar_t          dev_mainint_wait;                              /* place for mainevent interrupt thread to sleep */
+    spinlock_t          dev_mainint_lock;                              /*   and it's spinlock */
+    unsigned            dev_mainint_started:1;
+    unsigned            dev_mainint_stopped:1;
+
+    /* device context - this is used to flush insert cache/instruction cache/dmas & threads */
+    ELAN4_CPROC_TRAP     dev_cproc_trap;                               /* space to extract cproc trap into */
+
+    struct list_head     dev_intop_list;                               /* list of main interrupt operations */
+    spinlock_t          dev_intop_lock;                                /*   and spinlock */
+    E4_uint64           dev_intop_cookie;                              /*   and next cookie to use */
+
+    spinlock_t          dev_flush_lock;                                /* spinlock for flushing */
+    kcondvar_t          dev_flush_wait;                                /*  and place to sleep */
+
+    ELAN4_CQ           *dev_flush_cq[COMMAND_INSERTER_CACHE_ENTRIES];  /* command queues to flush the insert cache */
+    ELAN4_INTOP          dev_flush_op[COMMAND_INSERTER_CACHE_ENTRIES]; /* and a main interrupt operation for each one */
+    unsigned            dev_flush_finished;                            /* flush command finished */
+
+    ELAN4_HALTOP        dev_iflush_haltop;                             /* halt operation for icache flush */
+    unsigned            dev_iflush_queued:1;                           /* icache haltop queued */
+
+    ELAN4_ROUTE_TABLE   *dev_routetable;                               /* virtual process table (for dma queue flush)*/
+    sdramaddr_t          dev_sdrampages[2];                            /* pages of sdram to hold suspend code sequence */
+    E4_Addr             dev_tproc_suspend;                             /*  st8suspend instruction */
+    E4_Addr             dev_tproc_space;                               /*     and target memory */
+
+    sdramaddr_t                 dev_neterr_inputq;                             /* network error input queue descriptor & event */
+    sdramaddr_t                 dev_neterr_slots;                              /* network error message slots */
+    ELAN4_CQ           *dev_neterr_msgcq;                              /* command queue for sending messages */
+    ELAN4_CQ           *dev_neterr_intcq;                              /* command queue for message received interrupt */
+    ELAN4_INTOP                 dev_neterr_intop;                              /*   and it's main interrupt operation */
+    E4_uint64           dev_neterr_queued;                             /* # message queued in msgcq */
+    spinlock_t           dev_neterr_lock;                              /*   and spinlock .... */
+
+    ELAN4_DEV_STATS     dev_stats;                                     /* device statistics */
+    E4_uint64           dev_sdramerrs[30];                             /* last few sdram errors for procfs */
+
+    spinlock_t          dev_error_routes_lock;
+    unsigned int       *dev_ack_errors;                                /* Map of source of dproc ack errors */
+    ELAN4_ROUTE_RINGBUF  dev_ack_error_routes;
+    unsigned int        *dev_dproc_timeout;                             /* Ditto dproc timeout errors */
+    ELAN4_ROUTE_RINGBUF  dev_dproc_timeout_routes;
+    unsigned int        *dev_cproc_timeout;                             /* Ditto cproc timeout errors */
+    ELAN4_ROUTE_RINGBUF  dev_cproc_timeout_routes;
+
+    struct list_head     dev_hc_list;                                   /* list of the allocated hash_chunks */
+
+    ELAN4_IPROC_TRAP     dev_iproc_trap;                               /* space for iproc trap */
+} ELAN4_DEV;
+
+/* values for dev_state */
+#define ELAN4_STATE_STOPPED            (1 << 0)                        /* device initialised but not started */
+#define ELAN4_STATE_STARTING           (1 << 1)                        /* device in process of starting */
+#define ELAN4_STATE_STARTED            (1 << 2)                        /* device started */
+#define ELAN4_STATE_STOPPING           (1 << 3)                        /* device in process of stopping */
+
+/* values for dev_features */
+#define ELAN4_FEATURE_NO_WRITE_COMBINE (1 << 0)                        /* don't allow write combinig at all */
+#define ELAN4_FEATURE_PCI_MAP          (1 << 1)                        /* must use pci mapping functions */
+#define ELAN4_FEATURE_NO_DWORD_READ    (1 << 2)                        /* must perform 64 bit PIO reads */
+
+extern __inline__ unsigned int
+__elan4_readb (ELAN4_DEV *dev, ioaddr_t addr)
+{
+    if (dev->dev_features & ELAN4_FEATURE_NO_DWORD_READ)
+    {
+       uint64_t val = readq ((void *) ((unsigned long) addr & ~7));
+       return ((val >> (((unsigned long) addr & 7) << 3)) & 0xff);
+    }
+    return readb (addr);
+}
+
+extern __inline__ unsigned int
+__elan4_readw (ELAN4_DEV *dev, ioaddr_t addr)
+{
+    if (dev->dev_features & ELAN4_FEATURE_NO_DWORD_READ)
+    {
+       uint64_t val = readq ((void *) ((unsigned long) addr & ~7));
+       return ((val >> (((unsigned long) addr & 7) << 3)) & 0xffff);
+    }
+    return readw (addr);
+}
+
+extern __inline__ unsigned int
+__elan4_readl (ELAN4_DEV *dev, ioaddr_t addr)
+{
+    if (dev->dev_features & ELAN4_FEATURE_NO_DWORD_READ)
+    {
+       uint64_t val = readq ((void *) ((unsigned long) addr & ~7));
+       return ((val >> (((unsigned long) addr & 7) << 3)) & 0xffffffff);
+    }
+    return readl (addr);
+}
+
+/* macros for accessing dev->dev_regs.Tags. */
+#define write_tag(dev,what,val)                writeq (val, dev->dev_regs + offsetof (E4_Registers, Tags.what))
+#define read_tag(dev,what)             readq (dev->dev_regs + offsetof (E4_Registers, Tags.what))
+
+/* macros for accessing dev->dev_regs.Regs. */
+#define write_reg64(dev,what,val)      writeq (val, dev->dev_regs + offsetof (E4_Registers, Regs.what))
+#define write_reg32(dev,what,val)      writel (val, dev->dev_regs + offsetof (E4_Registers, Regs.what))
+#define read_reg64(dev,what)           readq (dev->dev_regs + offsetof (E4_Registers, Regs.what))
+#define read_reg32(dev,what)           __elan4_readl (dev, dev->dev_regs + offsetof (E4_Registers, Regs.what))
+
+/* macros for accessing dev->dev_regs.uRegs. */
+#define write_ureg64(dev,what,val)     writeq (val, dev->dev_regs + offsetof (E4_Registers, uRegs.what))
+#define write_ureg32(dev,what,val)     writel (val, dev->dev_regs + offsetof (E4_Registers, uRegs.what))
+#define read_ureg64(dev,what)          readq (dev->dev_regs + offsetof (E4_Registers, uRegs.what))
+#define read_ureg32(dev,what)          __elan4_readl (dev, dev->dev_regs + offsetof (E4_Registers, uRegs.what))
+
+/* macros for accessing dev->dev_i2c */
+#define write_i2c(dev,what,val)                writeb (val, dev->dev_i2c + offsetof (E4_I2C, what))
+#define read_i2c(dev,what)             __elan4_readb (dev, dev->dev_i2c + offsetof (E4_I2C, what))
+
+/* macros for accessing dev->dev_rom */
+#define read_ebus_rom(dev,off)         __elan4_readb (dev, dev->dev_rom + off)
+
+/* PIO flush operations - ensure writes to registers/sdram are ordered */
+#ifdef CONFIG_IA64_SGI_SN2
+#define pioflush_reg(dev)              read_reg32(dev,InterruptReg)
+#define pioflush_sdram(dev)            elan4_sdram_readl(dev, 0)
+#else
+#define pioflush_reg(dev)              mb()
+#define pioflush_sdram(dev)            mb()
+#endif
+
+/* macros for manipulating the interrupt mask register */
+#define SET_INT_MASK(dev,value)        \
+do { \
+    write_reg32(dev, InterruptMask, (dev)->dev_intmask = (value)); \
+    pioflush_reg(dev);\
+} while (0)
+
+#define CHANGE_INT_MASK(dev, value) \
+do { \
+    if ((dev)->dev_intmask != (value)) \
+    {\
+       write_reg32 (dev, InterruptMask, (dev)->dev_intmask = (value));\
+       pioflush_reg(dev);\
+    }\
+} while (0)
+
+#define ENABLE_INT_MASK(dev,value) \
+do { \
+    unsigned long flags; \
+ \
+    spin_lock_irqsave (&(dev)->dev_intmask_lock, flags); \
+    write_reg32(dev, InterruptMask, (dev)->dev_intmask |= (value)); \
+    pioflush_reg(dev);\
+    spin_unlock_irqrestore (&(dev)->dev_intmask_lock, flags); \
+} while (0)
+
+#define DISABLE_INT_MASK(dev,value) \
+do { \
+    unsigned long flags; \
+ \
+    spin_lock_irqsave (&(dev)->dev_intmask_lock, flags); \
+    write_reg32(dev, InterruptMask, (dev)->dev_intmask &= ~(value)); \
+    pioflush_reg(dev);\
+    spin_unlock_irqrestore (&(dev)->dev_intmask_lock, flags); \
+} while (0)
+
+#define SET_SYSCONTROL(dev,what,value) \
+do { \
+    unsigned long flags; \
+\
+    spin_lock_irqsave (&(dev)->dev_syscontrol_lock, flags); \
+    if ((dev)->what++ == 0) \
+        write_reg64 (dev, SysControlReg, (dev)->dev_syscontrol |= (value)); \
+    pioflush_reg(dev);\
+    spin_unlock_irqrestore (&(dev)->dev_syscontrol_lock, flags); \
+} while (0)
+
+#define CLEAR_SYSCONTROL(dev,what,value) \
+do { \
+    unsigned long flags; \
+\
+    spin_lock_irqsave (&(dev)->dev_syscontrol_lock, flags); \
+    if (--(dev)->what == 0)\
+       write_reg64 (dev, SysControlReg, (dev)->dev_syscontrol &= ~(value)); \
+    pioflush_reg (dev); \
+    spin_unlock_irqrestore (&(dev)->dev_syscontrol_lock, flags); \
+} while (0)
+
+#define PULSE_SYSCONTROL(dev,value) \
+do { \
+    unsigned long flags; \
+\
+    spin_lock_irqsave (&(dev)->dev_syscontrol_lock, flags); \
+    write_reg64 (dev, SysControlReg, (dev)->dev_syscontrol | (value)); \
+    pioflush_reg (dev); \
+    spin_unlock_irqrestore (&(dev)->dev_syscontrol_lock, flags); \
+} while (0)
+
+#define CHANGE_SYSCONTROL(dev,add,sub) \
+do { \
+    unsigned long flags; \
+\
+    spin_lock_irqsave (&(dev)->dev_syscontrol_lock, flags); \
+    dev->dev_syscontrol |= (add);\
+    dev->dev_syscontrol &= ~(sub);\
+    write_reg64 (dev, SysControlReg, (dev)->dev_syscontrol);\
+    pioflush_reg (dev); \
+    spin_unlock_irqrestore (&(dev)->dev_syscontrol_lock, flags); \
+} while (0)
+
+#define SET_SCHED_STATUS(dev, value)\
+do {\
+    write_reg32 (dev, SchedStatus.Status, (dev)->dev_schedstatus = (value));\
+    pioflush_reg (dev);\
+} while (0)
+
+#define CHANGE_SCHED_STATUS(dev, value)\
+do {\
+    if ((dev)->dev_schedstatus != (value))\
+    {\
+       write_reg32 (dev, SchedStatus.Status, (dev)->dev_schedstatus = (value));\
+       pioflush_reg (dev);\
+    }\
+} while (0)
+
+#define PULSE_SCHED_RESTART(dev,value)\
+do {\
+    write_reg32 (dev, SchedStatus.Restart, value);\
+    pioflush_reg (dev);\
+} while (0)
+
+/* device context elan address space */
+#define DEVICE_TPROC_SUSPEND_ADDR              (0x1000000000000000ull)
+#define DEVICE_TPROC_SPACE_ADDR                        (0x1000000000000000ull + SDRAM_PAGE_SIZE)
+#if defined(__LITTLE_ENDIAN__)
+#  define DEVICE_TPROC_SUSPEND_INSTR           0xd3f040c0 /* st64suspend %r16, [%r1] */
+#else
+#  define DEVICE_TPROC_SUSPEND_INSTR           0xc040f0d3 /* st64suspend %r16, [%r1] */
+#endif
+
+#define DEVICE_NETERR_INPUTQ_ADDR              (0x2000000000000000ull)
+#define DEVICE_NETERR_INTCQ_ADDR               (0x2000000000000000ull + SDRAM_PAGE_SIZE)
+#define DEVICE_NETERR_SLOTS_ADDR               (0x2000000000000000ull + SDRAM_PAGE_SIZE*2)
+
+/*
+ * Interrupt operation cookie space
+ * [50:48]     type
+ * [47:0]      value
+ */
+#define INTOP_PERSISTENT                       (0x1000000000000ull)
+#define INTOP_ONESHOT                          (0x2000000000000ull)
+#define INTOP_TYPE_MASK                                (0x3000000000000ull)
+#define INTOP_VALUE_MASK                       (0x0ffffffffffffull)
+
+/* functions for accessing sdram - sdram.c */
+extern unsigned char      elan4_sdram_readb (ELAN4_DEV *dev, sdramaddr_t ptr);
+extern unsigned short     elan4_sdram_readw (ELAN4_DEV *dev, sdramaddr_t ptr);
+extern unsigned int       elan4_sdram_readl (ELAN4_DEV *dev, sdramaddr_t ptr);
+extern unsigned long long elan4_sdram_readq (ELAN4_DEV *dev, sdramaddr_t ptr);
+extern void               elan4_sdram_writeb (ELAN4_DEV *dev, sdramaddr_t ptr, unsigned char val);
+extern void               elan4_sdram_writew (ELAN4_DEV *dev, sdramaddr_t ptr, unsigned short val);
+extern void               elan4_sdram_writel (ELAN4_DEV *dev, sdramaddr_t ptr, unsigned int val);
+extern void               elan4_sdram_writeq (ELAN4_DEV *dev, sdramaddr_t ptr, unsigned long long val);
+
+extern void              elan4_sdram_zerob_sdram (ELAN4_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern void              elan4_sdram_zerow_sdram (ELAN4_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern void              elan4_sdram_zerol_sdram (ELAN4_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern void              elan4_sdram_zeroq_sdram (ELAN4_DEV *dev, sdramaddr_t ptr, int nbytes);
+
+extern void               elan4_sdram_copyb_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan4_sdram_copyw_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan4_sdram_copyl_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan4_sdram_copyq_from_sdram (ELAN4_DEV *dev, sdramaddr_t from, void *to, int nbytes);
+extern void               elan4_sdram_copyb_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+extern void               elan4_sdram_copyw_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+extern void               elan4_sdram_copyl_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+extern void               elan4_sdram_copyq_to_sdram (ELAN4_DEV *dev, void *from, sdramaddr_t to, int nbytes);
+
+/* device.c - configuration */
+extern unsigned int elan4_hash_0_size_val;
+extern unsigned int elan4_hash_1_size_val;
+extern unsigned int elan4_ctxt_table_shift;
+extern unsigned int elan4_ln2_max_cqs;
+extern unsigned int elan4_dmaq_highpri_size;
+extern unsigned int elan4_threadq_highpri_size;
+extern unsigned int elan4_dmaq_lowpri_size;
+extern unsigned int elan4_threadq_lowpri_size;
+extern unsigned int elan4_interruptq_size;
+
+extern unsigned int elan4_mainint_punt_loops;
+extern unsigned int elan4_mainint_resched_ticks;
+
+
+/* device.c */
+extern void               elan4_set_schedstatus (ELAN4_DEV *dev, E4_uint32 intreg);
+extern void               elan4_queue_haltop (ELAN4_DEV *dev, ELAN4_HALTOP *op);
+extern void              elan4_queue_intop (ELAN4_DEV *dev, ELAN4_CQ *cq, ELAN4_INTOP *op);
+extern void              elan4_register_intop (ELAN4_DEV *dev, ELAN4_INTOP *op);
+extern void              elan4_deregister_intop (ELAN4_DEV *dev, ELAN4_INTOP *op);
+extern void              elan4_queue_dma_flushop (ELAN4_DEV *dev, ELAN4_DMA_FLUSHOP *op, int hipri);
+extern void              elan4_queue_mainintop (ELAN4_DEV *dev, ELAN4_INTOP *op);
+
+extern int                elan4_1msi0 (ELAN4_DEV *dev);
+
+extern int                elan4_insertctxt (ELAN4_DEV *dev, ELAN4_CTXT *ctxt, ELAN4_TRAP_OPS *ops);
+extern void               elan4_removectxt (ELAN4_DEV *dev, ELAN4_CTXT *ctxt);
+extern ELAN4_CTXT        *elan4_localctxt (ELAN4_DEV *dev, unsigned num);
+extern ELAN4_CTXT        *elan4_networkctxt (ELAN4_DEV *dev, unsigned num);
+
+extern int                elan4_attach_filter (ELAN4_CTXT *ctxt, unsigned int ctxnum);
+extern void               elan4_detach_filter (ELAN4_CTXT *ctxt, unsigned int ctxnum);
+extern void              elan4_set_filter (ELAN4_CTXT *ctxt, unsigned int ctxnum, E4_uint32 state);
+extern void              elan4_set_routetable (ELAN4_CTXT *ctxt, ELAN4_ROUTE_TABLE *tbl);
+
+extern ELAN4_CQA *        elan4_getcqa (ELAN4_CTXT *ctxt, unsigned int idx);
+extern void               elan4_putcqa (ELAN4_CTXT *ctxt, unsigned int idx);
+extern ELAN4_CQ          *elan4_alloccq (ELAN4_CTXT *ctxt, unsigned cqsize, unsigned cqperm, unsigned cqtype);
+extern void               elan4_freecq (ELAN4_CTXT *ctxt, ELAN4_CQ *cq);
+extern void               elan4_restartcq (ELAN4_DEV *dev, ELAN4_CQ *cq);
+extern void               elan4_flushcq (ELAN4_DEV *dev, ELAN4_CQ *cq);
+extern void               elan4_updatecq (ELAN4_DEV *dev, ELAN4_CQ *cq, unsigned perm, unsigned restart);
+
+extern void              elan4_flush_icache (ELAN4_CTXT *ctxt);
+extern void              elan4_flush_icache_halted (ELAN4_CTXT *ctxt);
+
+extern int                elan4_initialise_device (ELAN4_DEV *dev);
+extern void               elan4_finalise_device (ELAN4_DEV *dev);
+extern int                elan4_start_device (ELAN4_DEV *dev);
+extern void               elan4_stop_device (ELAN4_DEV *dev);
+
+extern int               elan4_compute_position (ELAN_POSITION *pos, unsigned nodeid, unsigned numnodes, unsigned aritiyval);
+extern int               elan4_get_position (ELAN4_DEV *dev, ELAN_POSITION *pos);
+extern int               elan4_set_position (ELAN4_DEV *dev, ELAN_POSITION *pos);
+extern void              elan4_get_params   (ELAN4_DEV *dev, ELAN_PARAMS *params, unsigned short *mask);
+extern void              elan4_set_params   (ELAN4_DEV *dev, ELAN_PARAMS *params, unsigned short mask);
+
+
+extern int                elan4_read_vpd(ELAN4_DEV *dev, unsigned char *tag, unsigned char *result) ;
+
+
+/* device_osdep.c */
+extern unsigned int      elan4_pll_cfg;
+extern int               elan4_pll_div;
+extern int               elan4_mod45disable;
+
+extern int                elan4_pciinit (ELAN4_DEV *dev);
+extern void               elan4_pcifini (ELAN4_DEV *dev);
+extern void               elan4_pcierror (ELAN4_DEV *dev);
+
+extern ELAN4_DEV        *elan4_reference_device (int instance, int state);
+extern void              elan4_dereference_device (ELAN4_DEV *dev);
+
+extern ioaddr_t           elan4_map_device (ELAN4_DEV *dev, unsigned bar, unsigned off, unsigned size, ELAN4_MAP_HANDLE *handlep);
+extern void               elan4_unmap_device (ELAN4_DEV *dev, ioaddr_t ptr, unsigned size, ELAN4_MAP_HANDLE *handlep);
+extern unsigned long      elan4_resource_len (ELAN4_DEV *dev, unsigned bar);
+
+extern void               elan4_configure_mtrr (ELAN4_DEV *dev);
+extern void              elan4_unconfigure_mtrr (ELAN4_DEV *dev);
+
+/* i2c.c */
+extern int               i2c_disable_auto_led_update (ELAN4_DEV *dev);
+extern void              i2c_enable_auto_led_update (ELAN4_DEV *dev);
+extern int               i2c_write (ELAN4_DEV *dev, unsigned int addr, unsigned int count, unsigned char *data);
+extern int               i2c_read (ELAN4_DEV *dev, unsigned int addr, unsigned int count, unsigned char *data);
+extern int               i2c_writereg (ELAN4_DEV *dev, unsigned int addr, unsigned int reg, unsigned int count, unsigned char *data);
+extern int               i2c_readreg (ELAN4_DEV *dev, unsigned int addr, unsigned int reg, unsigned int count, unsigned char *data);
+extern int               i2c_read_rom (ELAN4_DEV *dev, unsigned int addr, unsigned int count, unsigned char *data);
+
+#if defined(__linux__)
+/* procfs_Linux.c */
+extern void              elan4_procfs_device_init (ELAN4_DEV *dev);
+extern void              elan4_procfs_device_fini (ELAN4_DEV *dev);
+extern void              elan4_procfs_init(void);
+extern void              elan4_procfs_fini(void);
+
+extern struct proc_dir_entry *elan4_procfs_root;
+extern struct proc_dir_entry *elan4_config_root;
+#endif
+
+/* sdram.c */
+extern void              elan4_sdram_init (ELAN4_DEV *dev);
+extern void               elan4_sdram_fini (ELAN4_DEV *dev);
+extern void               elan4_sdram_setup_delay_lines (ELAN4_DEV *dev);
+extern int                elan4_sdram_init_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank);
+extern void               elan4_sdram_fini_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank);
+extern void              elan4_sdram_add_bank (ELAN4_DEV *dev, ELAN4_SDRAM_BANK *bank);
+extern sdramaddr_t        elan4_sdram_alloc (ELAN4_DEV *dev, int nbytes);
+extern void               elan4_sdram_free (ELAN4_DEV *dev, sdramaddr_t ptr, int nbytes);
+extern void               elan4_sdram_flushcache (ELAN4_DEV *dev, sdramaddr_t base, int nbytes);
+extern char              *elan4_sdramerr2str (ELAN4_DEV *dev, E4_uint64 status, char *str);
+
+/* traps.c */
+extern void               elan4_display_eproc_trap (void *type, int mode, char *str, ELAN4_EPROC_TRAP *trap);
+extern void               elan4_display_cproc_trap (void *type, int mode, char *str, ELAN4_CPROC_TRAP *trap);
+extern void               elan4_display_dproc_trap (void *type, int mode, char *str, ELAN4_DPROC_TRAP *trap);
+extern void               elan4_display_tproc_trap (void *type, int mode, char *str, ELAN4_TPROC_TRAP *trap);
+extern void               elan4_display_iproc_trap (void *type, int mode, char *str, ELAN4_IPROC_TRAP *trap);
+
+
+extern void               elan4_extract_eproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_EPROC_TRAP *trap, int iswaitevent);
+extern void               elan4_extract_cproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_CPROC_TRAP *trap, unsigned cqnum);
+extern void               elan4_extract_dproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_DPROC_TRAP *trap, unsigned unit);
+extern void               elan4_extract_tproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_TPROC_TRAP *trap);
+extern void               elan4_extract_iproc_trap (ELAN4_DEV *dev, E4_uint64 status, ELAN4_IPROC_TRAP *trap, unsigned unit);
+extern void elan4_ringbuf_store(ELAN4_ROUTE_RINGBUF *ringbuf, E4_VirtualProcessEntry *route, ELAN4_DEV *dev);
+extern int                cproc_open_extract_vp (ELAN4_DEV *dev, ELAN4_CQ *cq);
+
+extern void               elan4_inspect_iproc_trap (ELAN4_IPROC_TRAP *trap);
+extern E4_uint64          elan4_trapped_open_command (ELAN4_DEV *dev, ELAN4_CQ *cq);
+
+/* mmu.c */
+extern void               elan4mmu_flush_tlb (ELAN4_DEV *dev);
+extern ELAN4_HASH_ENTRY  *elan4mmu_ptealloc (ELAN4_CTXT *ctxt, int tbl, E4_Addr vaddr, unsigned int *tagidxp);
+extern int                elan4mmu_pteload (ELAN4_CTXT *ctxt, int tbl, E4_Addr vaddr, E4_uint64 pte);
+extern void               elan4mmu_unload_range (ELAN4_CTXT *ctxt, int tbl, E4_Addr start, unsigned long len);
+extern void               elan4mmu_invalidate_ctxt (ELAN4_CTXT *ctxt);
+
+extern ELAN4_HASH_CACHE  *elan4mmu_reserve (ELAN4_CTXT *ctxt, int tbl, E4_Addr start, unsigned int npages, int cansleep);
+extern void               elan4mmu_release (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc);
+extern void               elan4mmu_set_pte (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc, unsigned int idx, E4_uint64 newpte);
+extern E4_uint64          elan4mmu_get_pte (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc, unsigned int idx);
+extern void               elan4mmu_clear_pte (ELAN4_CTXT *ctxt, ELAN4_HASH_CACHE *hc, unsigned int idx);
+
+/* mmu_osdep.c */
+extern int               elan4mmu_categorise_paddr (ELAN4_DEV *dev, physaddr_t *physp);
+extern int                elan4mmu_alloc_topaddr (ELAN4_DEV *dev, physaddr_t paddr, unsigned type);
+extern E4_uint64          elan4mmu_phys2pte (ELAN4_DEV *dev, physaddr_t paddr, unsigned perm);
+extern physaddr_t        elan4mmu_pte2phys (ELAN4_DEV *dev, E4_uint64 pte);
+
+/* neterr.c */
+extern int                elan4_neterr_init (ELAN4_DEV *dev);
+extern void               elan4_neterr_destroy (ELAN4_DEV *dev);
+extern int                elan4_neterr_sendmsg (ELAN4_DEV *dev, unsigned int nodeid, unsigned int retries, ELAN4_NETERR_MSG *msg);
+extern int                elan4_neterr_iproc_trap (ELAN4_DEV *dev, ELAN4_IPROC_TRAP *trap);
+
+/* routetable.c */
+extern ELAN4_ROUTE_TABLE *elan4_alloc_routetable (ELAN4_DEV *dev, unsigned size);
+extern void               elan4_free_routetable (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl);
+extern void               elan4_write_route (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl, unsigned vp, E4_VirtualProcessEntry *entry);
+extern void               elan4_read_route (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl, unsigned vp, E4_VirtualProcessEntry *entry);
+extern void               elan4_invalidate_route (ELAN4_DEV *dev, ELAN4_ROUTE_TABLE *tbl, unsigned vp);
+extern int                elan4_generate_route (ELAN_POSITION *pos, E4_VirtualProcessEntry *route, unsigned ctxnum,
+                                               unsigned lowid, unsigned highid, unsigned options);
+extern int               elan4_check_route (ELAN_POSITION *pos, ELAN_LOCATION location, E4_VirtualProcessEntry *route, unsigned flags);
+
+/* user.c */
+extern int        __categorise_command (E4_uint64 command, int *cmdSize);
+extern int        __whole_command (sdramaddr_t *commandPtr, sdramaddr_t insertPtr, unsigned int cqSize, unsigned int cmdSize);
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_ELANDEV_H */
Index: linux-2.4.21/include/elan4/device_Linux.h
===================================================================
--- linux-2.4.21.orig/include/elan4/device_Linux.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/device_Linux.h   2005-06-01 23:12:54.735418432 -0400
@@ -0,0 +1,97 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_ELANDEV_LINUX_H
+#define __ELAN4_ELANDEV_LINUX_H
+
+#ident "$Id: device_Linux.h,v 1.19 2004/08/09 14:02:37 daniel Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/device_Linux.h,v $*/
+
+#include <linux/coproc.h>
+
+#if defined(MPSAS)
+#include <elan4/mpsas.h>
+#endif
+
+#if defined(CONFIG_DEVFS_FS)
+#include <linux/devfs_fs_kernel.h>
+#endif
+
+#define ELAN4_MAJOR              61
+#define ELAN4_NAME               "elan4"
+#define ELAN4_MAX_CONTROLLER     16           /* limited to 4 bits */
+ 
+/* OS dependant component of ELAN4_DEV struct */
+typedef struct elan4_dev_osdep
+{
+    struct pci_dev       *pdev;                        /* PCI config data */
+
+    struct proc_dir_entry *procdir;
+    struct proc_dir_entry *configdir;
+    struct proc_dir_entry *statsdir;
+
+#if defined(CONFIG_DEVFS_FS)
+    devfs_handle_t devfs_control;
+    devfs_handle_t devfs_sdram;
+    devfs_handle_t devfs_user;
+#endif
+
+#if defined(CONFIG_MTRR)
+    int                           sdram_mtrr;
+    int                           regs_mtrr;
+#endif
+} ELAN4_DEV_OSDEP;
+
+/* /dev/elan/rmsX */
+
+/* /dev/elan4/controlX */
+typedef struct control_private
+{
+    struct elan4_dev   *pr_dev;
+    unsigned           pr_boundary_scan;
+} CONTROL_PRIVATE;
+
+/* /dev/elan4/sdramX */
+typedef struct mem_page
+{
+    struct mem_page *pg_next;
+    sdramaddr_t      pg_addr;
+    unsigned long    pg_pgoff;
+    unsigned        pg_ref;
+} MEM_PAGE;
+
+#define MEM_HASH_SIZE  32
+#define MEM_HASH(pgoff)        ((pgoff) & (MEM_HASH_SIZE-1))
+
+typedef struct mem_private
+{
+    struct elan4_dev *pr_dev;
+    MEM_PAGE         *pr_pages[MEM_HASH_SIZE];
+    spinlock_t        pr_lock;
+} MEM_PRIVATE;
+
+/* /dev/elan4/userX */
+typedef struct user_private
+{
+    atomic_t         pr_ref;
+    struct user_ctxt *pr_uctx;
+    struct mm_struct *pr_mm;
+    coproc_ops_t      pr_coproc;
+} USER_PRIVATE;
+
+/* No mapping handles on linux */
+typedef void *ELAN4_MAP_HANDLE;
+
+#define ELAN4_TASK_HANDLE()    ((unsigned long) current->mm)
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_ELANDEV_LINUX_H */
Index: linux-2.4.21/include/elan4/dma.h
===================================================================
--- linux-2.4.21.orig/include/elan4/dma.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/dma.h    2005-06-01 23:12:54.736418280 -0400
@@ -0,0 +1,82 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_DMA_H
+#define __ELAN4_DMA_H
+
+#ident "$Id: dma.h,v 1.16 2003/09/04 12:39:17 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/dma.h,v $*/
+
+#include <elan4/types.h>
+
+/* Alignment for a DMA descriptor */
+#define E4_DMA_ALIGN           (64)
+
+/* Maximum size of a single DMA ((1 << 31)-1) */
+#define E4_MAX_DMA_SIZE                (0x7fffffff)
+
+/* 
+ * dma_typeSize
+ *
+ * [63:32]     Size
+ * [31]                unused
+ * [30]                IsRemote
+ * [29]                QueueWrite
+ * [28]                ShmemWrite
+ * [27:26]     DataType
+ * [25]                Broadcast
+ * [24]                AlignPackets
+ * [23:16]     FailCount
+ * [15:14]     unused
+ * [13:0]      Context
+ */
+
+#define DMA_FailCount(val)     (((val) & 0xff) << 16)
+#define DMA_AlignPackets       (1 << 24)
+#define DMA_Broadcast          (1 << 25)
+#define DMA_ShMemWrite         (1 << 28)
+#define DMA_QueueWrite         (1 << 29)
+#define DMA_IsRemote           (1 << 30)
+#define DMA_Context(val)       ((unsigned) (val) & 0x3ff)
+#define DMA_ContextMask                0x3fffull
+#define Dma_TypeSizeMask       0xfffffffffff00000ull
+
+#define DMA_DataTypeByte       (E4_DATATYPE_BYTE  << 26)
+#define DMA_DataTypeShort      (E4_DATATYPE_SHORT << 26)
+#define DMA_DataTypeWord       (E4_DATATYPE_WORD  << 26)
+#define DMA_DataTypeLong       (E4_DATATYPE_DWORD << 26)
+
+#define E4_DMA_TYPE_SIZE(size, dataType, flags, failCount)     \
+    ((((E4_uint64)(size)) << 32) |  ((dataType) & DMA_DataTypeLong) | \
+     (flags) | DMA_FailCount(failCount))
+
+typedef volatile struct e4_dma
+{
+    E4_uint64          dma_typeSize;
+    E4_uint64          dma_cookie;
+    E4_uint64          dma_vproc;
+    E4_Addr            dma_srcAddr;
+    E4_Addr            dma_dstAddr;
+    E4_Addr            dma_srcEvent;
+    E4_Addr            dma_dstEvent;
+} E4_DMA;
+
+/* Same as above but padded to 64-bytes */
+typedef volatile struct e4_dma64
+{
+    E4_uint64          dma_typeSize;
+    E4_uint64          dma_cookie;
+    E4_uint64          dma_vproc;
+    E4_Addr            dma_srcAddr;
+    E4_Addr            dma_dstAddr;
+    E4_Addr            dma_srcEvent;
+    E4_Addr            dma_dstEvent;
+    E4_Addr            dma_pad;
+} E4_DMA64;
+
+#endif /* __ELAN4_DMA_H */
Index: linux-2.4.21/include/elan4/events.h
===================================================================
--- linux-2.4.21.orig/include/elan4/events.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/events.h 2005-06-01 23:12:54.736418280 -0400
@@ -0,0 +1,179 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_EVENTS_H
+#define __ELAN4_EVENTS_H
+
+#ident "$Id: events.h,v 1.22 2004/06/23 11:07:18 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/events.h,v $*/
+
+#define E4_EVENT_ALIGN         32
+#define E4_EVENTBLOCK_SIZE     64
+
+#ifndef _ASM
+/*
+ * Event locations must be aligned to a 32 byte boundary. It is very much more efficent to place
+ * them in elan local memory but is not essential.
+ */
+typedef struct _E4_Event
+{
+    volatile E4_uint64 ev_CountAndType;
+    E4_uint64          ev_Params[2];
+} E4_Event;
+
+/* Same as above but padded to correct Event alignment */
+typedef struct _E4_Event32
+{
+    volatile E4_uint64 ev_CountAndType;
+    E4_uint64          ev_Params[2];
+    E4_uint64          ev_pad;
+} E4_Event32;
+
+/*
+ * An E4_EVENTBLOCK_SIZE aligned block of Main or Elan memory
+ */
+typedef union _E4_Event_Blk
+{
+    /* Padded to 64-bytes in case a cache-line write is more efficient */
+    volatile E4_uint8  eb_unit8[E4_EVENTBLOCK_SIZE];
+    volatile E4_uint32 eb_uint32[E4_EVENTBLOCK_SIZE/sizeof(E4_uint32)];
+    volatile E4_uint64 eb_uint64[E4_EVENTBLOCK_SIZE/sizeof(E4_uint64)];
+} E4_Event_Blk;
+#define eb_done       eb_uint32[14]
+#define eb_done_dword eb_uint64[7]
+
+#endif /* ! _ASM */
+
+/*
+ * ev_CountAndType
+ *  [63:31]   Count
+ *  [10]      CopyType
+ *  [9:8]     DataType
+ *  [7:0]     CopySize
+ */
+#define E4_EVENT_TYPE_MASK     0x00000000ffffffffull
+#define E4_EVENT_COUNT_MASK    0xffffffff00000000ull
+#define E4_EVENT_COUNT_SHIFT   32
+#define E4_EVENT_COPY_TYPE_MASK        (1 << 10)
+#define E4_EVENT_DATA_TYPE_MASK        (3 << 8)
+#define E4_EVENT_COPY_SIZE_MASK        (0xff)
+
+/* CopyType */
+#define E4_EVENT_COPY          (0 << 10)
+#define E4_EVENT_WRITE         (1 << 10)
+
+/* DataType */
+#define E4_EVENT_DTYPE_BYTE    (0 << 8)
+#define E4_EVENT_DTYPE_SHORT   (1 << 8)
+#define E4_EVENT_DTYPE_WORD    (2 << 8)
+#define E4_EVENT_DTYPE_LONG    (3 << 8)
+
+#define EVENT_COUNT(EventPtr)  ((E4_int32)(elan4_load64 (&(EventPtr)->ev_CountAndType) >> E4_EVENT_COUNT_SHIFT))
+#define EVENT_TYPE(EventPtr)   ((E4_uint32)(elan4_load64 (&(EventPtr)->ev_CountAndType) & E4_EVENT_TYPE_MASK))
+
+#define E4_WAITEVENT_COUNT_TYPE_VALUE(Count, EventType, DataType, CopySize) \
+       (((E4_uint64)(Count) << E4_EVENT_COUNT_SHIFT) | (EventType) | (DataType) | (CopySize))
+
+#define E4_EVENT_TYPE_VALUE(EventType, DataType, CopySize)     \
+       ((EventType) | (DataType) | (CopySize))
+
+#define E4_EVENT_INIT_VALUE(InitialCount, EventType, DataType, CopySize)       \
+       (((E4_uint64)(InitialCount) << E4_EVENT_COUNT_SHIFT) | E4_EVENT_TYPE_VALUE(EventType, DataType, CopySize))
+
+#define ev_CopySource  ev_Params[0]
+#define ev_CopyDest    ev_Params[1]
+#define ev_WritePtr    ev_Params[0]
+#define ev_WriteValue  ev_Params[1]
+
+#define EVENT_BLK_READY(BLK) ((BLK)->eb_done != 0)
+#define EVENT_READY(EVENT)   ((E4_uint32)((((volatile E4_Event *) (EVENT))->ev_CountAndType) >> E4_EVENT_COUNT_SHIFT) >= 0)
+
+#define ELAN_WAIT_EVENT (0)
+#define ELAN_POLL_EVENT (-1)
+
+#define E4_BLK_PATTERN ((E4_uint32)0xfeedface)
+
+#define E4_INIT_COPY_EVENT(EVENT, BLK_ELAN, BLK, SIZE)                                                         \
+       do {                                                                                            \
+          elan4_store64 (E4_EVENT_INIT_VALUE(0, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, SIZE), &(EVENT)->ev_CountAndType); \
+           elan4_store64 ((BLK_ELAN), &(EVENT)->ev_CopySource); \
+          elan4_store64 ((BLK), &(EVENT)->ev_CopyDest); \
+       } while (0)
+
+#define E4_INIT_WRITE_EVENT(EVENT, DWORD)                                                              \
+       do {                                                                                            \
+           elan4_store64 (E4_EVENT_INIT_VALUE(0, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0), &(EVENT)->ev_CountAndType);  \
+           elan4_store64 ((DWORD), &(EVENT)->ev_WritePtr); \
+           elan4_store64 ((E4_Addr) (E4_BLK_PATTERN), &(EVENT)->ev_WriteValue); \
+       } while (0)
+
+#define E4_RESET_BLK_EVENT(BLK)                                        \
+       do {                                                            \
+               (BLK)->eb_done = (0);                                   \
+       } while (0)
+
+#define E4_PRIME_BLK_EVENT(EVENT, COUNT)                               \
+       do {                                                            \
+          elan4_store64 (E4_EVENT_INIT_VALUE(COUNT, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, 8), &(EVENT)->ev_CountAndType);\
+       } while (0)
+
+#define E4_PRIME_COPY_EVENT(EVENT, SIZE, COUNT)                                \
+       do {                                                            \
+          elan4_store64 (E4_EVENT_INIT_VALUE(COUNT, E4_EVENT_COPY, E4_EVENT_DTYPE_LONG, (SIZE >> 3)), &(EVENT)->ev_CountAndType);\
+       } while (0)
+
+#define E4_PRIME_WRITE_EVENT(EVENT, COUNT)                                     \
+       do {                                                                    \
+          elan4_store64 (E4_EVENT_INIT_VALUE(COUNT, E4_EVENT_WRITE, E4_EVENT_DTYPE_LONG, 0), &(EVENT)->ev_CountAndType);\
+       } while (0)
+
+#ifndef _ASM
+
+#define E4_INPUTQ_ALIGN                        32      /* Descriptor must be 32-byte aligned */
+
+typedef struct _E4_InputQueue
+{
+   volatile E4_Addr    q_bptr;         /* 64 bit aligned ptr to current back item */
+   E4_Addr             q_fptr;         /* 64 bit aligned ptr to current front item */
+   E4_uint64           q_control;      /* this defines the last item, item size, and offset back to the first item. */
+   E4_Addr             q_event;        /* queue event */
+} E4_InputQueue;
+
+#define E4_INPUTQ_LASTITEM_MASK        0x00000000ffffffffULL
+#define E4_INPUTQ_ITEMSIZE_MASK                0x000000ff00000000ULL
+#define E4_INPUTQ_LASTITEM_OFFSET_MASK 0xffffff0000000000ULL
+#define E4_INPUTQ_LASTITEM_SHIFT       0
+#define E4_INPUTQ_ITEMSIZE_SHIFT       32
+#define E4_INPUTQ_LASTITEM_OFFSET_SHIFT        40
+
+/*
+ * Macro to initialise the InputQueue control word given the FirstItem, LastItem & ItemSize
+ * FirstItem and LastItem are 64 bit double word aligned elan addresses.
+ */
+#define E4_InputQueueControl(FirstItem, LastItem, ItemSizeInBytes)\
+   (((((E4_uint64)(LastItem)))                                                      & E4_INPUTQ_LASTITEM_MASK) |\
+    ((((E4_uint64)(ItemSizeInBytes))        << (E4_INPUTQ_ITEMSIZE_SHIFT-3))        & E4_INPUTQ_ITEMSIZE_MASK)  |\
+    ((((E4_uint64)((FirstItem)-(LastItem))) << (E4_INPUTQ_LASTITEM_OFFSET_SHIFT-3)) & E4_INPUTQ_LASTITEM_OFFSET_MASK))    
+
+/* 
+ * LastItemOffset is a sign extended -ve quantity with LastItemOffset[26:3] == q_control[63:40]
+ * we sign extend this by setting LastItemOffset[63:27] to be #one.
+ */
+#define E4_InputQueueLastItemOffset(control)  ((((E4_int64) -1) << (64 - (E4_INPUTQ_LASTITEM_OFFSET_SHIFT-3))) | \
+                                              ((E4_int64) (((control) & E4_INPUTQ_LASTITEM_OFFSET_MASK) >> (E4_INPUTQ_LASTITEM_OFFSET_SHIFT-3))))
+#define E4_InputQueueItemSize(control)       (((control) & E4_INPUTQ_ITEMSIZE_MASK) >> (E4_INPUTQ_ITEMSIZE_SHIFT-3))
+
+/*
+ * Macro to increment the InputQ front pointer taking into account wrap 
+ */
+#define E4_InputQueueFptrIncrement(Q, FirstItem, LastItem, ItemSizeInBytes) \
+       ((Q)->q_fptr = ( ((Q)->q_fptr == (LastItem)) ? (FirstItem) : ((Q)->q_fptr + (ItemSizeInBytes))) )
+
+#endif /* _ASM */
+
+#endif /* __ELAN4_EVENTS_H */
Index: linux-2.4.21/include/elan4/i2c.h
===================================================================
--- linux-2.4.21.orig/include/elan4/i2c.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/i2c.h    2005-06-01 23:12:54.736418280 -0400
@@ -0,0 +1,47 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN4_I2C_H
+#define _ELAN4_I2C_H
+
+#ident "@(#)$Id: i2c.h,v 1.10 2003/12/02 16:11:22 lee Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/i2c.h,v $*/
+
+/* I2C address space - bits[7:1] */
+#define I2C_LED_I2C_ADDR                       0x20
+#define I2C_TEMP_ADDR                          0x48
+#define I2C_EEPROM_ADDR                                0x50
+
+#define I2C_WRITE_ADDR(addr)                   ((addr) << 1 | 0)
+#define I2C_READ_ADDR(addr)                    ((addr) << 1 | 1)
+
+/* I2C EEPROM appears as 8 I2C 256 byte devices */
+#define I2C_24LC16B_BLOCKSIZE                  (256)
+#define I2C_24LC16B_BLOCKADDR(addr)            ((addr) >> 8)
+#define I2C_24LC16B_BLOCKOFFSET(addr)          ((addr) & 0xff)
+
+#define I2C_ELAN_EEPROM_PCI_BASEADDR           0       /* PCI config starts at addr 0 in the EEPROM */
+#define I2C_ELAN_EEPROM_VPD_BASEADDR           256     /* VPD data start                            */
+#define I2C_ELAN_EEPROM_PCI_SIZE               256     /* PCI data max size                         */
+#define I2C_ELAN_EEPROM_VPD_SIZE               256     /* VPD data max size                         */
+
+#define I2C_ELAN_EEPROM_SIZE                   2048
+
+#define I2C_ELAN_EEPROM_DEVICE_ID              0xA0
+#define I2C_ELAN_EEPROM_FAIL_LIMIT              8
+
+#define I2C_ELAN_EEPROM_ADDR_BLOCKSIZE_SHIFT   0x8
+#define I2C_ELAN_EEPROM_ADDR_BLOCK_MASK                0x7
+#define I2C_ELAN_EEPROM_ADDR_BLOCK_SHIFT       0x1
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* _ELAN4_I2C_H */
Index: linux-2.4.21/include/elan4/intcookie.h
===================================================================
--- linux-2.4.21.orig/include/elan4/intcookie.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/intcookie.h      2005-06-01 23:12:54.737418128 -0400
@@ -0,0 +1,62 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: intcookie.h,v 1.10 2004/08/09 14:02:37 daniel Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/intcookie.h,v $*/
+
+#ifndef __ELAN4_INTCOOKIE_H
+#define __ELAN4_INTCOOKIE_H
+
+typedef E4_uint64 ELAN4_INTCOOKIE;
+
+#ifdef __KERNEL__
+
+typedef struct intcookie_entry
+{
+    struct intcookie_entry    *ent_next;
+    struct intcookie_entry    *ent_prev;
+
+    spinlock_t                ent_lock;
+    unsigned                  ent_ref;
+
+    ELAN4_INTCOOKIE           ent_cookie;
+    ELAN4_INTCOOKIE           ent_fired;
+    kcondvar_t                ent_wait;
+} INTCOOKIE_ENTRY;
+
+typedef struct intcookie_table
+{
+    struct intcookie_table    *tbl_next;
+    struct intcookie_table    *tbl_prev;
+
+    ELAN_CAPABILITY           *tbl_cap;
+
+    spinlock_t                tbl_lock;
+    unsigned                  tbl_ref;
+    INTCOOKIE_ENTRY           *tbl_entries;
+} INTCOOKIE_TABLE;
+
+extern void                intcookie_init(void);
+extern void                intcookie_fini(void);
+extern INTCOOKIE_TABLE    *intcookie_alloc_table (ELAN_CAPABILITY *cap);
+extern void                intcookie_free_table (INTCOOKIE_TABLE *tbl);
+extern int                 intcookie_alloc (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie);
+extern int                 intcookie_free (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie);
+extern int                 intcookie_fire (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie);
+extern int                 intcookie_fire_cap (ELAN_CAPABILITY *cap, ELAN4_INTCOOKIE cookie);
+extern int                 intcookie_wait (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie);
+extern int                 intcookie_arm (INTCOOKIE_TABLE *tbl, ELAN4_INTCOOKIE cookie);
+
+#endif /* __KERNEL */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_INTCOOKIE_H */
Index: linux-2.4.21/include/elan4/ioctl.h
===================================================================
--- linux-2.4.21.orig/include/elan4/ioctl.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/ioctl.h  2005-06-01 23:12:54.738417976 -0400
@@ -0,0 +1,320 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_IOCTL_H
+#define __ELAN4_IOCTL_H
+
+#ident "@(#)$Id: ioctl.h,v 1.27.6.2 2005/01/11 12:15:39 duncant Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/ioctl.h,v $*/
+
+#include <elan/devinfo.h>
+#include <elan/capability.h>
+
+#include <elan4/dma.h>
+#include <elan4/neterr.h>
+#include <elan4/registers.h>
+#include <elan4/intcookie.h>
+
+#define ELAN4IO_CONTROL_PATHNAME       "/dev/elan4/control%d"
+#define ELAN4IO_USER_PATHNAME          "/dev/elan4/user%d"
+#define ELAN4IO_SDRAM_PATHNAME         "/dev/elan4/sdram%d"
+#define ELAN4IO_MAX_PATHNAMELEN        32
+
+/*
+ * NOTE - ioctl values 0->0x1f are defined for 
+ *        generic/control usage.
+ */
+
+/* Macro to generate 'offset' to mmap "control" device */
+#define OFF_TO_BAR(off)                (((off) >> 28) & 0xF)
+#define OFF_TO_OFFSET(off)     ((off) & 0x0FFFFFFF)
+#define GEN_OFF(bar,off)       (((bar) << 28) | ((off) & 0x0FFFFFFF))
+
+/* Definiations for generic ioctls */
+#define ELAN4IO_GENERIC_BASE           0x00
+
+typedef struct elan4io_stats_struct
+{
+    int                       which;
+    unsigned long long ptr;                                    /* always pass pointer as 64 bit */
+} ELAN4IO_STATS_STRUCT;
+
+#define ELAN4IO_STATS                  _IOR ('e', ELAN4IO_GENERIC_BASE + 0, ELAN4IO_STATS_STRUCT)
+#define ELAN4IO_DEVINFO                        _IOR ('e', ELAN4IO_GENERIC_BASE + 1, ELAN_DEVINFO)
+#define ELAN4IO_POSITION               _IOR ('e', ELAN4IO_GENERIC_BASE + 2, ELAN_POSITION)
+
+
+/* 
+ * Definitions for /dev/elan4/controlX
+ */
+#define ELAN4IO_CONTROL_BASE           0x20
+
+#define ELAN4IO_GET_POSITION           _IOR ('e', ELAN4IO_CONTROL_BASE + 0, ELAN_POSITION)
+#define ELAN4IO_SET_POSITION           _IOW ('e', ELAN4IO_CONTROL_BASE + 1, ELAN_POSITION)
+#define ELAN4IO_DEBUG_SNAPSHOT         _IOW ('e', ELAN4IO_CONTROL_BASE + 2, )
+
+typedef struct elan4io_params_mask_struct
+{
+    unsigned short     p_mask;
+    ELAN_PARAMS                p_params;
+} ELAN4IO_PARAMS_STRUCT;
+#define ELAN4IO_GET_PARAMS             _IOR ('e', ELAN4IO_CONTROL_BASE + 3, ELAN4IO_PARAMS_STRUCT)
+#define ELAN4IO_SET_PARAMS             _IOW ('e', ELAN4IO_CONTROL_BASE + 4, ELAN4IO_PARAMS_STRUCT)
+
+/* old versions - implicit p_mask == 3 */
+#define ELAN4IO_OLD_GET_PARAMS         _IOR ('e', ELAN4IO_CONTROL_BASE + 3, ELAN_PARAMS)
+#define ELAN4IO_OLD_SET_PARAMS         _IOW ('e', ELAN4IO_CONTROL_BASE + 4, ELAN_PARAMS)
+
+/*
+ * Definitions for /dev/elan4/userX
+ */
+#define ELAN4IO_USER_BASE              0x40
+
+#define ELAN4IO_FREE                   _IO   ('e', ELAN4IO_USER_BASE + 0)
+#define ELAN4IO_ATTACH                 _IOWR ('e', ELAN4IO_USER_BASE + 1, ELAN_CAPABILITY)
+#define ELAN4IO_DETACH                 _IOWR ('e', ELAN4IO_USER_BASE + 2, ELAN_CAPABILITY)
+#define ELAN4IO_BLOCK_INPUTTER         _IO   ('e', ELAN4IO_USER_BASE + 3)
+
+typedef struct elan4io_add_p2pvp_struct 
+{
+    unsigned        vp_process;
+    ELAN_CAPABILITY vp_capability;
+} ELAN4IO_ADD_P2PVP_STRUCT;
+
+#define ELAN4IO_ADD_P2PVP              _IOW  ('e', ELAN4IO_USER_BASE + 4, ELAN4IO_ADD_P2PVP_STRUCT)
+
+typedef struct elan4io_add_bcastvp_struct
+{
+    unsigned int       vp_process;
+    unsigned int       vp_lowvp;
+    unsigned int       vp_highvp;
+} ELAN4IO_ADD_BCASTVP_STRUCT;
+
+#define ELAN4IO_ADD_BCASTVP            _IOW  ('e', ELAN4IO_USER_BASE + 5, ELAN4IO_ADD_BCASTVP_STRUCT)
+
+#define ELAN4IO_REMOVEVP               _IO   ('e', ELAN4IO_USER_BASE + 6)
+
+typedef struct elan4io_route_struct
+{
+    unsigned int          rt_process;
+    unsigned int          rt_error;
+    E4_VirtualProcessEntry rt_route;
+} ELAN4IO_ROUTE_STRUCT;
+
+#define ELAN4IO_SET_ROUTE              _IOW  ('e', ELAN4IO_USER_BASE + 7, ELAN4IO_ROUTE_STRUCT)
+#define ELAN4IO_RESET_ROUTE            _IOW  ('e', ELAN4IO_USER_BASE + 9, ELAN4IO_ROUTE_STRUCT)
+#define ELAN4IO_GET_ROUTE              _IOWR ('e', ELAN4IO_USER_BASE + 8, ELAN4IO_ROUTE_STRUCT)
+#define ELAN4IO_CHECK_ROUTE            _IOWR ('e', ELAN4IO_USER_BASE + 10, ELAN4IO_ROUTE_STRUCT)
+
+typedef struct elan4io_alloc_cq_struct
+{
+    unsigned int cq_size;                                      /* input: size of queue */
+    unsigned int cq_perm;                                      /* input: requested permissions */
+    unsigned int cq_type;                                      /* input: queue type */
+    unsigned int cq_indx;                                      /* output: queue number */
+} ELAN4IO_ALLOCCQ_STRUCT;
+
+#define ELAN4IO_ALLOCCQ                        _IOWR ('e', ELAN4IO_USER_BASE + 11, ELAN4IO_ALLOCCQ_STRUCT)
+#define ELAN4IO_FREECQ                 _IOWR ('e', ELAN4IO_USER_BASE + 12, unsigned)
+
+#define ELAN4IO_CQ_TYPE_REORDER                1                       /* revb reordering command queue */
+
+typedef struct elan4io_perm_struct
+{
+    E4_Addr            ps_eaddr;
+    E4_uint64          ps_len;
+    unsigned long      ps_maddr;
+    unsigned int       ps_perm;
+} ELAN4IO_PERM_STRUCT;
+
+typedef struct elan4io_perm_struct32
+{
+    E4_Addr            ps_eaddr;
+    E4_uint64          ps_len;
+    unsigned int       ps_maddr;
+    unsigned int       ps_perm;
+} ELAN4IO_PERM_STRUCT32;
+
+#define ELAN4IO_SETPERM                        _IOWR ('e', ELAN4IO_USER_BASE + 13, ELAN4IO_PERM_STRUCT)
+#define ELAN4IO_SETPERM32              _IOWR ('e', ELAN4IO_USER_BASE + 13, ELAN4IO_PERM_STRUCT32)
+#define ELAN4IO_CLRPERM                        _IOWR ('e', ELAN4IO_USER_BASE + 14, ELAN4IO_PERM_STRUCT)
+#define ELAN4IO_CLRPERM32              _IOWR ('e', ELAN4IO_USER_BASE + 14, ELAN4IO_PERM_STRUCT32)
+
+typedef struct elan4io_trapsig_struct
+{
+    int                ts_signo;
+} ELAN4IO_TRAPSIG_STRUCT;
+#define ELAN4IO_TRAPSIG                        _IOW  ('e', ELAN4IO_USER_BASE + 15, ELAN4IO_TRAPSIG_STRUCT)
+
+typedef struct elan4io_traphandler_struct
+{
+    unsigned int       th_nticks;                              /* number of ticks to sleep for next trap */
+    unsigned int       th_proc;                                        /* elan processor involved */
+    unsigned long      th_trapp;                               /* space to store trap */
+} ELAN4IO_TRAPHANDLER_STRUCT;
+
+typedef struct elan4io_traphandler_struct32
+{
+    unsigned int       th_nticks;                              /* number of ticks to sleep for next trap */
+    unsigned int       th_proc;                                        /* elan processor involved */
+    unsigned int       th_trapp;                               /* space to store trap */
+} ELAN4IO_TRAPHANDLER_STRUCT32;
+
+#define ELAN4IO_TRAPHANDLER            _IOW  ('e', ELAN4IO_USER_BASE + 16, ELAN4IO_TRAPHANDLER_STRUCT)
+#define ELAN4IO_TRAPHANDLER32          _IOW  ('e', ELAN4IO_USER_BASE + 16, ELAN4IO_TRAPHANDLER_STRUCT32)
+
+typedef struct elan4io_required_mappings_struct
+{
+    E4_Addr    rm_upage_addr;                                  /* elan address of user page */
+    E4_Addr    rm_trestart_addr;                               /* elan address of tproc restart trampoline */
+} ELAN4IO_REQUIRED_MAPPINGS_STRUCT;
+#define ELAN4IO_REQUIRED_MAPPINGS      _IOW  ('e', ELAN4IO_USER_BASE + 17, ELAN4IO_REQUIRED_MAPPINGS_STRUCT)
+
+typedef struct elan4io_resume_eproc_trap_struct
+{
+    E4_Addr             rs_addr;
+} ELAN4IO_RESUME_EPROC_TRAP_STRUCT;
+#define ELAN4IO_RESUME_EPROC_TRAP      _IOW  ('e', ELAN4IO_USER_BASE + 18, ELAN4IO_RESUME_EPROC_TRAP_STRUCT)
+
+typedef struct elan4io_resume_cproc_trap_struct
+{
+    unsigned int       rs_indx;
+} ELAN4IO_RESUME_CPROC_TRAP_STRUCT;
+#define ELAN4IO_RESUME_CPROC_TRAP      _IOW  ('e', ELAN4IO_USER_BASE + 19, ELAN4IO_RESUME_CPROC_TRAP_STRUCT)
+
+typedef struct elan4io_resume_dproc_trap_struct
+{
+    E4_DMA             rs_desc;
+} ELAN4IO_RESUME_DPROC_TRAP_STRUCT;
+#define ELAN4IO_RESUME_DPROC_TRAP      _IOW  ('e', ELAN4IO_USER_BASE + 20, ELAN4IO_RESUME_DPROC_TRAP_STRUCT)
+
+typedef struct elan4io_resume_tproc_trap_struct
+{
+    E4_ThreadRegs      rs_regs;
+} ELAN4IO_RESUME_TPROC_TRAP_STRUCT;
+#define ELAN4IO_RESUME_TPROC_TRAP      _IOW  ('e', ELAN4IO_USER_BASE + 21, ELAN4IO_RESUME_TPROC_TRAP_STRUCT)
+
+typedef struct elan4io_resume_iproc_trap_struct
+{
+    unsigned int       rs_channel;
+    unsigned int       rs_trans;
+    E4_IprocTrapHeader  rs_header;
+    E4_IprocTrapData    rs_data;
+} ELAN4IO_RESUME_IPROC_TRAP_STRUCT;
+#define ELAN4IO_RESUME_IPROC_TRAP      _IOW  ('e', ELAN4IO_USER_BASE + 22, ELAN4IO_RESUME_IPROC_TRAP_STRUCT)
+
+#define ELAN4IO_FLUSH_ICACHE           _IO   ('e', ELAN4IO_USER_BASE + 23)
+#define ELAN4IO_STOP_CTXT              _IO   ('e', ELAN4IO_USER_BASE + 24)
+
+#define ELAN4IO_ALLOC_INTCOOKIE                _IOW  ('e', ELAN4IO_USER_BASE + 25, ELAN4_INTCOOKIE)
+#define ELAN4IO_FREE_INTCOOKIE         _IOW  ('e', ELAN4IO_USER_BASE + 26, ELAN4_INTCOOKIE)
+#define ELAN4IO_ARM_INTCOOKIE          _IOW  ('e', ELAN4IO_USER_BASE + 27, ELAN4_INTCOOKIE)
+#define ELAN4IO_WAIT_INTCOOKIE         _IOW  ('e', ELAN4IO_USER_BASE + 28, ELAN4_INTCOOKIE)
+
+typedef struct elan4io_alloc_trap_queues_struct
+{
+    unsigned int       tq_ndproc_traps;
+    unsigned int       tq_neproc_traps;
+    unsigned int       tq_ntproc_traps;
+    unsigned int       tq_nthreads;
+    unsigned int       tq_ndmas;
+} ELAN4IO_ALLOC_TRAP_QUEUES_STRUCT;
+#define ELAN4IO_ALLOC_TRAP_QUEUES      _IOW  ('e', ELAN4IO_USER_BASE + 29, ELAN4IO_ALLOC_TRAP_QUEUES_STRUCT)
+
+typedef struct elan4io_neterr_msg_struct
+{
+    unsigned int       nm_vp;
+    unsigned int       nm_nctx;
+    unsigned int       nm_retries;
+    unsigned int        nm_pad;
+    ELAN4_NETERR_MSG    nm_msg;
+} ELAN4IO_NETERR_MSG_STRUCT;
+#define ELAN4IO_NETERR_MSG             _IOW ('e', ELAN4IO_USER_BASE + 30, ELAN4IO_NETERR_MSG_STRUCT)
+
+typedef struct elan4io_neterr_timer_struct 
+{
+    unsigned int       nt_usecs;
+} ELAN4IO_NETERR_TIMER_STUCT;
+
+#define ELAN4IO_NETERR_TIMER           _IO  ('e', ELAN4IO_USER_BASE + 31)
+
+typedef struct elan4io_neterr_fixup_struct
+{
+    E4_uint64          nf_cookie;
+    unsigned int       nf_waitforeop;
+    unsigned int       nf_sten;
+    unsigned int       nf_vp;
+    unsigned int       nf_pad;
+} ELAN4IO_NETERR_FIXUP_STRUCT;
+
+#define ELAN4IO_NETERR_FIXUP           _IOW ('e', ELAN4IO_USER_BASE + 32, ELAN4IO_NETERR_FIXUP_STRUCT)
+
+typedef struct elan4io_firecap_struct 
+{
+    ELAN_CAPABILITY     fc_capability;
+    ELAN4_INTCOOKIE     fc_cookie;
+} ELAN4IO_FIRECAP_STRUCT;
+
+#define ELAN4IO_FIRE_INTCOOKIE         _IOW  ('e', ELAN4IO_USER_BASE + 33, ELAN4IO_FIRECAP_STRUCT)
+
+#define ELAN4IO_ALLOC_INTCOOKIE_TABLE  _IOW  ('e', ELAN4IO_USER_BASE + 34, ELAN_CAPABILITY)
+#define ELAN4IO_FREE_INTCOOKIE_TABLE   _IO   ('e', ELAN4IO_USER_BASE + 35)
+
+typedef struct elan4io_translation
+{
+    E4_Addr            tr_addr;
+    unsigned long      tr_len;
+    unsigned int       tr_access;
+} ELAN4IO_TRANSLATION_STRUCT;
+
+#define ELAN4IO_LOAD_TRANSLATION       _IOW  ('e', ELAN4IO_USER_BASE + 36, ELAN4IO_TRANSLATION_STRUCT)
+#define ELAN4IO_UNLOAD_TRANSLATION     _IOW  ('e', ELAN4IO_USER_BASE + 37, ELAN4IO_TRANSLATION_STRUCT)
+
+typedef struct elan4io_dumpcq_struct32
+{
+    E4_uint64 cq_space;      /* output: sdram addr of q, used to decode ptrs */
+    E4_uint32 cq_size;       /* output: The real size of the command queue */
+    E4_uint32 bufsize;       /* input: The size of the buffer to dump to */
+    E4_uint32 cq_indx;       /* input: index of cq to dump */
+    unsigned int buffer;     /* input: user address of rgs->buffer to dump to */
+} ELAN4IO_DUMPCQ_STRUCT32;
+
+typedef struct elan4io_dumpcq_struct
+{
+    E4_uint64 cq_space;      /* output: sdram addr of q, used to decode ptrs */
+    E4_uint32 cq_size;       /* output: The real size of the command queue */
+    E4_uint32 bufsize;       /* input: The size of the buffer to dump to */
+    E4_uint32 cq_indx;       /* input: index of cq to dump */
+    unsigned long buffer;    /* input: user address of rgs->buffer to dump to */
+} ELAN4IO_DUMPCQ_STRUCT;
+
+#define ELAN4IO_DUMPCQ                 _IOWR ('e', ELAN4IO_USER_BASE + 38, ELAN4IO_DUMPCQ_STRUCT)
+#define ELAN4IO_DUMPCQ32                       _IOWR ('e', ELAN4IO_USER_BASE + 38, ELAN4IO_DUMPCQ_STRUCT32)
+
+/* mmap offsets - - we define the file offset space as follows:
+ *
+ * page 0 - 4095 - command queues
+ * page 4096    - device user registers
+ * page 4097    - flag page/user stats
+ * page 4098    - device stats
+ * page 4099     - tproc trampoline
+ */
+
+#define ELAN4_OFF_COMMAND_QUEUES       0
+#define ELAN4_OFF_USER_REGS            4096
+#define ELAN4_OFF_USER_PAGE            4097
+#define ELAN4_OFF_DEVICE_STATS         4098
+#define ELAN4_OFF_TPROC_TRAMPOLINE     4099
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_IOCTL_H */
Index: linux-2.4.21/include/elan4/mmu.h
===================================================================
--- linux-2.4.21.orig/include/elan4/mmu.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/mmu.h    2005-06-01 23:12:54.738417976 -0400
@@ -0,0 +1,94 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: mmu.h,v 1.11 2004/04/21 12:04:24 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/mmu.h,v $*/
+
+
+#ifndef __ELAN4_MMU_H
+#define __ELAN4_MMU_H
+
+typedef struct elan4_hash_entry
+{
+    struct elan4_hash_entry    *he_next;
+    struct elan4_hash_entry    *he_prev;
+
+    sdramaddr_t                         he_entry;
+    
+    struct elan4_hash_entry    *he_chain[2];
+    E4_uint64                   he_tag[2];
+    E4_uint32                   he_pte[2];
+} ELAN4_HASH_ENTRY;
+
+#define ELAN4_HENT_CHUNKS      16              /* SDRAM_MIN_BLOCK_SIZE/sizeof (E4_HashTableEntry) */
+
+typedef struct elan4_hash_chunk
+{
+    struct list_head            hc_link;
+    ELAN4_HASH_ENTRY           hc_hents[ELAN4_HENT_CHUNKS];
+} ELAN4_HASH_CHUNK;
+
+typedef struct elan4_hash_cache
+{
+    E4_Addr           hc_start;
+    E4_Addr           hc_end;
+    int                      hc_tbl;
+
+    ELAN4_HASH_ENTRY *hc_hes[1];
+} ELAN4_HASH_CACHE;
+
+/* 
+ * he_pte is really 4 bytes of pte "type" one for each pte
+ * entry - however we declare it as an "int" so we can
+ * easily determine that all 4 entries are invalid 
+ */
+#define HE_SET_PTE(he,tagidx,pteidx,val)       (((E4_uint8 *) &(he->he_pte[tagidx]))[pteidx] = (val))
+#define HE_GET_PTE(he,tagidx,pteidx)           (((E4_uint8 *) &(he->he_pte[tagidx]))[pteidx])
+
+/*
+ * he_tag has the following form :
+ *     [63:27] tag
+ *     [20:17]  pte valid
+ *     [16]     locked
+ *     [15]     copy
+ *     [14]     valid
+ *     [13:0]  context
+ */
+
+#define HE_TAG_VALID           (1 << 14)
+#define HE_TAG_COPY            (1 << 15)
+#define HE_TAG_LOCKED          (1 << 16)
+
+#define INVALID_CONTEXT                0
+
+extern u_char elan4_permtable[];
+#define ELAN4_INCOMPAT_ACCESS(perm,access) ((elan4_permtable[(perm)] & (1 << (access))) == 0)
+extern u_char elan4_permreadonly[];
+#define ELAN4_PERM_READONLY(perm)        (elan4_permreadonly[(perm)])
+
+/* return code from elan4mmu_categorise_paddr */
+#define ELAN4MMU_PADDR_SDRAM           0
+#define ELAN4MMU_PADDR_COMMAND         1
+#define ELAN4MMU_PADDR_LOCALPCI                2
+#define ELAN4MMU_PADDR_PAGE            3
+#define ELAN4MMU_PADDR_OTHER           4
+
+extern int elan4_debug_mmu;
+
+#ifdef DEBUG_PRINTF
+#  define MPRINTF(ctxt,lvl,args...)    (elan4_debug_mmu > (lvl) ? elan4_debugf(ctxt,DBG_MMU, ##args) : (void)0)
+#else
+#  define MPRINTF(ctxt,lvl,args...)    ((void) 0)
+#endif
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_MMU_H */
Index: linux-2.4.21/include/elan4/neterr.h
===================================================================
--- linux-2.4.21.orig/include/elan4/neterr.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/neterr.h 2005-06-01 23:12:54.738417976 -0400
@@ -0,0 +1,40 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2004 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_NETERR_H
+#define __ELAN4_NETERR_H
+
+#ident "@(#)$Id: neterr.h,v 1.1 2004/01/19 14:38:34 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/elan4mod/neterr.h,v $*/
+
+typedef struct elan4_neterr_msg
+{
+    E4_uint8           msg_type;
+    E4_uint8           msg_waitforeop;
+    E4_uint16          msg_context;                            /* network context # message sent to */
+    E4_int16           msg_found;                              /* # cookie found (response) */
+
+    ELAN_LOCATION      msg_sender;                             /* nodeid/context # message sent from */
+    E4_uint32          msg_pad;
+
+    E4_uint64          msg_cookies[6];                         /* 64 bit cookies from identify packets */
+} ELAN4_NETERR_MSG;
+
+#define ELAN4_NETERR_MSG_SIZE          sizeof (ELAN4_NETERR_MSG)
+#define ELAN4_NETERR_MSG_REQUEST       1
+#define ELAN4_NETERR_MSG_RESPONSE      2
+
+#define ELAN4_NETERR_MAX_COOKIES       (sizeof (((ELAN4_NETERR_MSG *) 0)->msg_cookies) / \
+                                        sizeof (((ELAN4_NETERR_MSG *) 0)->msg_cookies[0]))
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_NETERR_H */
Index: linux-2.4.21/include/elan4/pci.h
===================================================================
--- linux-2.4.21.orig/include/elan4/pci.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/pci.h    2005-06-01 23:12:54.739417824 -0400
@@ -0,0 +1,227 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_PCI_H 
+#define __ELAN4_PCI_H
+
+#ident "$Id: pci.h,v 1.32 2003/09/04 12:39:17 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/pci.h,v $*/
+
+/* Elan has 2 64 bit bars */
+#define ELAN4_BAR_SDRAM                        0
+#define ELAN4_BAR_REGISTERS            2
+
+#define PCI_VENDOR_ID_QUADRICS         0x14fc
+#define PCI_DEVICE_ID_ELAN3            0x0000
+#define   PCI_REVISION_ID_ELAN3_REVA   0x0000
+#define   PCI_REVISION_ID_ELAN3_REVB   0x0001
+#define PCI_DEVICE_ID_ELAN4            0x0001
+#define   PCI_REVISION_ID_ELAN4_REVA   0x0000
+#define   PCI_REVISION_ID_ELAN4_REVB   0x0001
+
+/* support standard pseudo bars */
+#define ELAN4_PSEUDO_BAR_ROM           8
+
+/* Elan PCI control
+ configuration space register. ElanControlRegister */
+#define PCI_ELAN_PARITY_ADDR_LO                0x40
+#define PCI_ELAN_PARITY_ADDR_HI                0x44
+#define PCI_ELAN_PARITY_TYPE           0x48
+#define PCI_ELAN_CONTROL               0x4c
+#define PCI_ELAN_PLL_CONTROL           0x50
+#define PCI_ELAN_SPLIT_MESSAGE_ATTR    0x54
+#define PCI_ELAN_SPLIT_MESSAGE_VALUE   0x54
+#define PCI_ELAN_RAMBIST_FAILED                0x54
+#define PCI_ELAN_TOPPHYSADDR(i)                (0x58 + ((i)<<1))
+
+/*
+ * [31]           PciM66EN             This is set it the bus is running in PCI2.3 - 66MHz mode.
+ * [30:28] InitPattern         This gives the PCI-X startup mode. See "Pci intialisation patterns" below.
+ * [27]           notBusIs64Bits       If set the bus is running 32 bits wide. If Clear it is a 64 bit bus.
+ * [26:24] RamBistCntl         Used to control the Elan4 RAM BIST. Not acitive it zero.
+ * [23]           RamBistFinished      Only used when performing the RAM BIST test.
+ * [22]           SelectSplitMessAttr  See ECTRL_SELECT_SPLIT_MESS_ATTR below.
+ * [21]           ReceivedSplitCompError See ECTRL_REC_SPLIT_COMP_MESSAGE below
+ * [20:16] WriteHighPriTime    Used with ReadHighPriTime to control the ratio of PCI master write to PCI master
+ *                             read bandwidth under heavy load. The high the value of WriteHighPriTime the longer
+ *                             the PCI write bursts will be allowed without interruption from a read transfer.
+ * [15]    DisableCouplingTest This is only used as part of the RAM BIST test. It effects the testing of the main
+ *                             cache tag RAMS.
+ * [14:13] Not used            Will read as zero.
+ * [12:8]  ReadHighPriTime     Used with WriteHighPriTime to control the ratio of PCI master write to PCI master
+ *                             read bandwidth under heavy load. The high the value of ReadHighPriTime the longer
+ *                             the PCI read bursts will be allowed without interruption from a write transfer.
+ * [7] EnableLatencyCountReset  This bit effect the behaviour of disconnects due to the removal of GNT# after the latency
+ *                             counter has expired. If set it will allow the latency counter to be reset each time the
+ *                             GNT# is reasserted. If asserted it should provided improved bandwidth on the PCI bus
+ *                             without increasing the maximum latency another device would have for access to the bus.
+ *                             It will increase the average latency of other devices.
+ * [6] ExtraMasterAddrBits     This bit used to control the physical PCI addresses generated by the MMU.
+ * [5] ReducedPciDecode                If set the PCI local memory BAR will decode 256Mbytes of PCI address space. If clear it
+ *                             will decode 2Gbyte of PCI address space.
+ * [4] ConfigInEBusRom         If set the constant values of the Elan4 PCI configuration space will be taken from the
+ *                             EEPROM. If clear the internal values will be used.
+ * [3] EnableRd2_2Bursts       This bit only effects the behaviour of burst reads when the PCI bus is operating in
+ *                             PCI-2.2 mode. It allows adjacent reads to be merged into longer bursts for higher
+ *                             performance.
+ * [2] SoftIntReset            If set this bit will cause the Elan4 to reset itself with the exception of the PCI
+ *                             configuation space. All internal state machines will be put into the reset state.
+ * [1] EnableWrBursts          This bit allows much longer PCI-X write bursts. If set it will stop the Elan4 from
+ *                             being completely PCI-X compliant as the Elan4 may request a long PCI-X write burst that
+ *                             it does not complete. However it should significantly increase the maximum PCI-X write
+ *                             bandwidth and is unlikely to cause problems with many PCI-X bridge chips.
+ * [0] InvertMSIPriority       This bit effect the way MSI interrupts are generated. It provides flexiblity to generate
+ *                             the MSI interrupts in a different way to allow for different implimentations of MSI
+ *                             logic and still give the correct priority of Elan4 interrupts.
+ *
+ *     {PciM66EN, InitPattern, notBusIs64Bits, RamBistCntl, RamBistFinished,
+ *      SelectSplitMessAttr, ReceivedSplitCompError, WriteHighPriTime,
+ *      DisableCouplingTest, 2'h0, ReadHighPriTime,
+ *      EnableLatencyCountReset, ExtraMasterAddrBits, ReducedPciDecode, ConfigInEBusRom,
+ *      EnableRd2_2Bursts, SoftIntReset, EnableWrBursts, InvertMSIPriority}
+ */
+
+#define ECTRL_INVERT_MSI_PRIO          (1 << 0)
+#define ECTRL_ENABLE_WRITEBURSTS       (1 << 1)
+#define ECTRL_SOFTWARE_INTERNAL_RESET  (1 << 2)
+#define ECTRL_ENABLE_2_2READBURSTS     (1 << 3)
+#define ECTRL_CONFIG_IN_EBUS_ROM       (1 << 4)
+#define ECTRL_28_NOT_30_BIT_LOCAL_BAR  (1 << 5)
+#define ECTRL_ExtraMasterAddrBits      (1 << 6)
+#define ECTRL_ENABLE_LATENCY_RESET      (1 << 7)
+#define ECTRL_DISABLE_COUPLING_TEST    (1 << 15)
+
+/*
+ * Ratio of the following two registers set the relative bandwidth given to intputer data
+ * versus other PCI pci traffic when scheduling new PCI master accesses.
+ */
+#define ECTRL_OTHER_HIGH_PRI_TIME_SHIFT        (8)     /* Sets top 4 bits of 8 bit counter */
+#define ECTRL_OTHER_HIGH_PRI_TIME_MASK (0x1f)
+
+
+#define ECTRL_IPROC_HIGH_PRI_TIME_SHIFT        (16)    /* Sets top 4 bits of 8 bit counter */
+#define ECTRL_IPROC_HIGH_PRI_TIME_MASK (0x1f)
+
+/*
+ * This is set if a split completion message is received.
+ * This will cause a PCI error interrupt.
+ * This error is cleared by writting a 1 to this bit.
+ */
+#define ECTRL_REC_SPLIT_COMP_MESSAGE   (1 << 21)
+/*
+ * This bit is used to select reading of either the Split message attribute value when
+ * set or the split completion message data value from 0x54 in the config space
+ * if the ECTRL_REC_SPLIT_COMP_MESSAGE bit is set. 0x54 returns the the BistFailed flags
+ * if any of the BIST control bits are set (bits 26 to 24)
+ */
+#define ECTRL_SELECT_SPLIT_MESS_ATTR   (1 << 22)
+
+// Internal RAM bist control bits.
+// Three bits of state control the RAM BIST (Built in self test).
+//
+// These bits must not be set unless the ECTRL_SOFTWARE_INTERNAL_RESET bit has also been set!
+//
+// For a normal fast ram test assert ECTRL_BIST_FAST_TEST. 
+// For a data retention test first write ECTRL_START_RETENTION_TEST then wait the retention period of
+// at least 1ms and preferably much longer then write ECTRL_CONTINUE_RETENTION_TEST then wait
+// again and finallly write ECTRL_FINISH_RETENTION_TEST.
+// 
+// The read only bit ECTRL_BIST_FINISHED_TEST can be polled to check that the test has compleated.
+#define ECTRL_BIST_CTRL_SHIFT          (24)
+#define ECTRL_BIST_CTRL_MASK           (7 << 24)
+
+#define ECTRL_BIST_FAST_TEST           ((7 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)     // old scheme
+#define ECTRL_START_RETENTION_TEST     ((1 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)
+#define ECTRL_CONTINUE_RETENTION_TEST  ((3 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)
+#define ECTRL_FINISH_RETENTION_TEST    ((7 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)
+
+#define ECTRL_BIST_KICK_OFF            ((1 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)     // new scheme
+#define ECTRL_BIST_MOVE_ON_ODD         ((3 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)
+#define ECTRL_BIST_MOVE_ON_EVEN                ((5 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)
+#define ECTRL_BIST_SCREAM_THROUGH      ((7 << 24) | ECTRL_SOFTWARE_INTERNAL_RESET)
+
+#define ECTRL_CLEAR_BIST_TEST          (0 << 24)
+#define ECTRL_BIST_FINISHED_TEST       (1 << 23)
+
+// Read only current PCI bus type.
+#define ECTRL_RUNNING_32BIT_MODE       (1 << 27)
+#define ECTRL_INITIALISATION_MODE      (7 << 28)
+#define ECTRL_RUNNING_M66EN_MODE       (1 << 31)
+
+#define ECTRL_INIT_PATTERN_SHIFT       (28)
+#define ECTRL_INIT_PATTERN_MASK                (0x7)
+
+// Pci intialisation patterns
+#define Pci2_2                         (0 << 28)
+#define PciX50To66MHz                  (1 << 28)
+#define PciX66to100MHz                 (2 << 28)
+#define PciX100to133MHz                        (3 << 28)
+#define PciXReserved1                  (4 << 28)
+#define PciXReserved2                  (5 << 28)
+#define PciXReserved3                  (6 << 28)
+#define PciXReserved4                  (7 << 28)
+
+/* Elan PCI pll and pad control configuration space register. ElanPllControlReg */
+// This overrides the default PCI pll control settings.
+#define PciPll_FeedForwardISel0                (1 << 0)        // Lsi name Z0
+#define PciPll_FeedForwardISel1                (1 << 1)        // Lsi name Z1
+#define PciPll_ChargePumpISel0         (1 << 2)        // Lsi name P0
+#define PciPll_ChargePumpISel1         (1 << 3)        // Lsi name P1
+#define PciPll_EnableAutoReset         (1 << 4)        // Lsi name ENARST
+#define PciPll_RSEL200500              (1 << 5)        // Lsi name Range Select, 0: 100 - 250MHz, 1: 200 - 500MHz
+#define PciPll_DivideFeedback          (1 << 6)        // Just used for test - This divides the shortcut feedback to the PCI PLL so that it can lock to the tester clock.
+#define PciPll_CutFeedback             (1 << 7)        // Just used for test - This disables the shortcut feedback.
+
+// This overrides the default PCI BZ controler settings.
+#define PciBZ_UPDI                     (0xf << 8)
+#define PciBZ_WAIT_INT                 (0xf << 12)
+
+// This overrides the default Sys and SDRam pll control settings.
+#define SysPll_FeedForwardISel0                (1 << 16)       // Lsi name P0     
+#define SysPll_FeedForwardISel1                (1 << 17)       // Lsi name P1     
+#define SysPll_ChargePumpISel0         (1 << 18)       // Lsi name Z0    
+#define SysPll_ChargePumpISel1         (1 << 19)       // Lsi name Z1    
+#define SysPll_EnableAutoReset         (1 << 20)       // Lsi name ENARST
+#define SysPll_DivPhaseCompInBy2       (1 << 21)       // Lsi name NODIV (Should be DIV)
+#define SysPll_PllTestClkSel           (1 << 22)       // If asserted the master clock source is not taken from the pll.
+
+#define Pll_ForceEBusADTristate                (1 << 23)       // Required to enable the testing of EnableAutoReset. Enables use of EBusAD[7] (rev A)
+#define Pll_LinkErrDirectToSDA         (1 << 23)       // Access to link error flag for triggering (rev B)
+
+
+#define ECTRL_SYS_CLOCK_RATIO_SHIFT    (24)
+// Config: with 800MHz         Speeds are 266 200 160 133.
+//         0 = 133/133 (1:1)   6:6     1
+//        1 = 160/133 (6:5)    5:6     1.2
+//         2 = 200/133 (3:2)   4:6     1.5
+//        3 = 266/133 (2:1)    3:6     2
+//        4 = 200/200 (1:1)    4:4     1
+//        5 = 266/200 (4:3)    3:4     1.33
+
+// Config: with 600MHz         Speeds are 200 150 120 100
+//         0 = 100/100 (1:1)   6:6     1
+//        1 = 120/100 (6:5)    5:6     1.2
+//         2 = 150/100 (3:2)   4:6     1.5
+//        3 = 200/100 (2:1)    3:6     2
+//        4 = 150/150 (1:1)    4:4     1
+//        5 = 200/150 (4:3)    3:4     1.33
+
+#define ECTRL_SYS_CLOCK_RATIO_SHIFT    (24)
+#define ECTRL_SYS_CLOCK_RATIO_1_1Slow  (0 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_6_5      (1 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_3_2      (2 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_2_1      (3 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_1_1Fast  (4 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_4_3      (5 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_MAX_NORMAL     (6)                                     /* used to generate a valid random value */
+#define GET_RANDOM_CLOCK_RATIO         (Random(ECTRL_SYS_CLOCK_MAX_NORMAL) << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_PLL_TEST (6 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_TEST     (7 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+#define ECTRL_SYS_CLOCK_RATIO_MASK     (7 << ECTRL_SYS_CLOCK_RATIO_SHIFT)
+
+#endif /* __ELAN4_PCI_H */
Index: linux-2.4.21/include/elan4/registers.h
===================================================================
--- linux-2.4.21.orig/include/elan4/registers.h 2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/registers.h      2005-06-01 23:12:54.742417368 -0400
@@ -0,0 +1,1588 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN4_REGISTERS_H
+#define _ELAN4_REGISTERS_H
+
+#ident "$Id: registers.h,v 1.117.2.1 2004/10/04 14:26:18 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/registers.h,v $*/
+
+/*
+ * Header file for internal slave mapping of the ELAN4 registers
+ */
+
+#define E4_CACHELINE_SIZE      (64)
+#define E4_STACK_ALIGN         (64)
+
+#ifndef _ASM
+
+#include <elan4/types.h>
+#include <elan4/dma.h>
+#include <elan4/userregs.h>
+
+typedef volatile struct _E4_CacheSets
+{
+   E4_uint64   Set0[1024];     /* 8k bytes per set */
+   E4_uint64   Set1[1024];     /* 8k bytes per set */
+   E4_uint64   Set2[1024];     /* 8k bytes per set */
+   E4_uint64   Set3[1024];     /* 8k bytes per set */
+} E4_CacheSets;
+
+typedef union e4_cache_tag
+{
+   struct {
+       E4_uint32 pad0;                 /* Undefined value when read */
+#if (BYTE_ORDER == LITTLE_ENDIAN) || defined(__LITTLE_ENDIAN__)
+       E4_uint32 :10;                                          /* 0-9   - reserved */
+       E4_uint32 LineError:1;                                  /* 10    - line error */
+       E4_uint32 Modified:1;                                   /* 11    - modified */
+       E4_uint32 FillPending:1;                                        /* 12    - fill pending */
+       E4_uint32 AddrTag30to13:18;                             /* 30-13 - tag */
+       E4_uint32 :1;                                           /* 31    -  */
+#else
+       E4_uint32 :1;                                           /* 31    -  */
+       E4_uint32 AddrTag30to13:18;                             /* 30-13 - tag */
+       E4_uint32 FillPending:1;                                        /* 12    - fill pending */
+       E4_uint32 Modified:1;                                   /* 11    - modified */
+       E4_uint32 LineError:1;                                  /* 10    - line error */
+       E4_uint32 :10;                                          /* 0-9   - reserved */
+#endif
+   } s;
+   E4_uint64   Value;
+} E4_CacheTag;
+
+typedef volatile struct _E4_CacheTags
+{
+   E4_CacheTag Tags[4][128];   /* 8k bytes per set, 64 byte cache line */
+} E4_CacheTags;
+
+#define E4_NumCacheSets                4
+#define E4_NumCacheLines       128
+#define E4_CacheLineSize       64
+#define E4_CacheSize           (E4_NumCacheSets * E4_NumCacheLines * E4_CacheLineSize)
+#define E4_CacheSetSize        (E4_NumCacheLines * E4_CacheLineSize)
+
+/*
+ * Run Queue pointers 
+ *
+ * [62:35]     FrontPointer[30:3]
+ * [33:32]     Size Value
+ * [30:3]      BackPointer[30:3]
+ */
+#define E4_QueuePtrMask                (0x7ffffff8ULL)
+#define E4_QueueSizeMask       3
+#define E4_QueueEntrySize       sizeof (E4_uint64)
+
+#define E4_Queue8KBytes                0
+#define E4_Queue64KBytes       1
+#define E4_Queue512KBytes      2
+#define E4_Queue4MBytes                3
+
+#define E4_QueueFrontValue(val,size)   ((val) | (size))
+#define E4_QueueValue(queue,size)      (((E4_uint64) E4_QueueFrontValue(queue,size)) << 32 | ((E4_uint64) (queue)))
+
+#define E4_QueueFrontPointer(val)      /* extract queue front pointer from register */\
+       (((val) >> 32) & E4_QueuePtrMask)
+#define E4_QueueBackPointer(val)       /* extract queue back pointer from register */ \
+       ((val) & E4_QueuePtrMask)
+#define E4_QueueSizeValue(val)         /* extract queue size value from register */ \
+       (((val) >> 32) & E4_QueueSizeMask)
+#define E4_QueueSize(value)            /* queue size in bytes from size value */ \
+       (1 << (((value)*3) + 13))
+#define E4_QueueOffsetMask(fptr)\
+        ((8192 << (((fptr) & E4_QueueSizeMask) << 3)) - 1)
+#define E4_QueueOffset(fptr)\
+        ((fptr) & E4_QueueOffsetMask(fptr))
+#define E4_QueueFrontPointerInc(fptr)   \
+        ( ((fptr) & ~E4_QueueOffsetMask(fptr)) | ((E4_QueueOffset(fptr) + 8) & E4_QueueOffsetMask(fptr)) )
+
+typedef union _E4_QueuePtr
+{
+   E4_uint64   Value;
+   struct {
+       E4_uint32 Back;
+       E4_uint32 Front;
+   } s;
+} E4_QueuePtr;
+
+/*
+ * DMA processor status register.
+ *
+ * [48]                FirstSendTrans          Set for the first packet of a dma.
+ * [47:46]     TimeSliceCount          Time left to timeslice.
+ * [45]                DmaLastPacket           Set for the last packet of a dma.
+ * [44]                CurrPrefetchDma         Dma descriptor the prefetcher is valid for.
+ * [43:39]     PrefetcherState         Dma prefetcher's state machines value.
+ * [38:33]     PacketAssemblyState     Packet assembler's state machines value.
+ * [32:31]     PrefetcherWakeupFnt     Dma prefetcher's wakeup function.
+ * [30:28]     PacketAssWakeupFnt      Packet assembler's wakeup function.
+ * [27]                AckBufferValid          Packet ack is valid.
+ * [26]                PrefetchedDataProblem   Had either a data read fault or data error. Valid if AckBufferValid.
+ * [25]                PrefetcherHalting       Prefetch data about to stop for halt. Valid if AckBufferValid.
+ * [24]                PacketTimeout           Packet timeout. Sent an EopError. Valid if AckBufferValid set.
+ * [23:22]     PacketAckValue          Packet ack type. Valid if AckBufferValid set.
+ * [21:20]     FaultUnitNo             Set if the dma prefetcher has faulted.
+ * [19:17]     TrapType                Packet assembler's trap type.
+ * [16]                PrefetcherFault         Set if the dma prefetcher has faulted for this DMA unit.
+ * [15]                Remote                  The Dma had been issued remotly
+ * [14]                Priority                Running at high priority.
+ * [13:0]      Context                 procs current context.
+ */
+
+#define DPROC_FirstSendTrans(s)                ((unsigned)((s) >> 48) & 1)
+#define DPROC_TimeSliceCount(s)                ((unsigned)(((s) >> 46) & 3)
+#define DPROC_DmaLastPacket(s)         ((unsigned)((s) >> 45) & 1)
+#define DPROC_CurrPrefetchDma(s)       ((unsigned)((s) >> 44) & 1)
+#define DPROC_PrefetcerState(s)                ((unsigned)((s) >> 39) & 0x1f)
+#define DPROC_PacketAssemblerState(s)  ((unsigned)((s) >> 33) & 0x1f)
+#define DPROC_PrefetcherWakeupFn(s)    ((unsigned)((s) >> 31) & 3)
+#define DPROC_PacketAssemblerWakeupFn(s)((unsigned)((s) >> 28) & 3)
+#define DPROC_AckBufferValid(s)                ((unsigned)((s) >> 27) & 1)
+#define DPROC_PrefetcherDataProblem(s) ((unsigned)((s) >> 26) & 1)
+#define DPROC_PrefetcherHalting(s)     ((unsigned)((s) >> 25) & 1)
+#define DPROC_PacketTimeout(s)         ((unsigned)((s) >> 24) & 1)
+#define DPROC_PacketAckValue(s)                ((unsigned)((s) >> 22) & 3)
+#define DPROC_FaultUnitNo(s)           ((unsigned)((s) >> 20) & 3)
+#define DPROC_TrapType(s)              ((unsigned)((s) >> 17) & 7)
+#define DPROC_PrefetcherFault(s)       ((unsigned)((s) >> 16) & 1)
+#define DPROC_Remote(s)                        ((unsigned)((s) >> 15) & 1)
+#define DPROC_Priority(s)              ((unsigned)((s) >> 14) & 1)
+#define DPROC_Context(s)               ((unsigned)(s) & 0x3fff)
+
+/*
+ * Command processor status register.
+ *
+ * [26:21]     CPState         procs current state.
+ * [20]                WakeupFnt       procs wakeup function.
+ * [19:16]     TrapValue       procs trap value.
+ * [15]                Remote          Issued remotely.
+ * [14]                Priority        Running at high priority.
+ * [13:0]      Context         procs current context.
+ */
+
+#define CPROC_TrapType(s)              ((unsigned)((s) >> 16) & 0xf)
+#define CPROC_Remote(s)                        ((unsigned)((s) >> 15) & 0x1)
+#define CPROC_Priority(s)              ((unsigned)((s) >> 14) & 0x1)
+#define CPROC_Context(s)               ((unsigned)(s) & 0x3fff)
+
+/*
+ * Event processor status register.
+ *
+ * [34:30]     CPState         event procs current state.
+ * [29:28]     WakeupFnt       event procs wakeup function.
+ * [27:20]     EventCopySize   This is the number of DWords to still be copied on a copy dword event.
+ * [19]                EProcPort1Fault CUN_EventProc1 has taken a translation fault.
+ * [18]                EProcPort0Fault CUN_EventProc0 has taken a translation fault.
+ * [17:16]     TrapValue       event proc's trap value.
+ * [15]                Remote          Issued remotely.
+ * [14]                Priority        Running at high priority.
+ * [13:0]      Context         procs current context.
+ */
+
+#define EPROC_CPState(s)               ((unsigned)((s) >> 30) & 0x1f)
+#define EPROC_WakeupFunction(s)                ((unsigned)((s) >> 28) & 3)
+#define EPROC_CopySize(s)              ((unsigned)((s) >> 20) & 0xFF)
+#define EPROC_Port1Fault(s)            ((unsigned)((s) >> 19) & 1)
+#define EPROC_Port0Fault(s)            ((unsigned)((s) >> 18) & 1)
+#define EPROC_TrapType(s)              ((unsigned)((s) >> 16) & 3)
+#define EPROC_Remote(s)                        ((unsigned)((s) >> 15) & 1)
+#define EPROC_Priority(s)              ((unsigned)((s) >> 14) & 1)
+#define EPROC_Context(s)               ((unsigned)(s) & 0x3fff)
+
+/*
+ * Thread processor status register.
+ *
+ * [39:24]     MemPortBusy             16 bits of port busy flags for all FFU memory ports.
+ * [23:21]     Reads as zero
+ * [20:18]     TQState                 State vector for thread queuing proc.
+ * [17]                HighRunQueueFull        High priority run queue is full
+ * [16]                LowRunQueueFull         Low priority run queue is full
+ * [15]                ReadyHigh               More runable threads at high priority
+ * [14]                ReadyLow                More runable threads at low priority
+ * [13:0]      Context                 procs current context.
+ */
+#define TPROC_HighRunQueueFull(s)      ((unsigned)((s) >> 17) & 1)
+#define TPROC_LowRunQueueFull(s)       ((unsigned)((s) >> 16) & 1)
+#define TPROC_ReadyHigh(s)             ((unsigned)((s) >> 15) & 1)
+#define TPROC_ReadyLow(s)              ((unsigned)((s) >> 14) & 1)
+#define TPROC_Context(s)               ((unsigned)((s) & 0x3fff))
+
+/*
+ * Input processor status register
+ *
+ * [55]                Last Trans (~EOP)
+ * [54]                First Trans (~EOP)
+ * [53]                Channel (~EOP) 
+ * [52]                Bad Length (~EOP)
+ * [51:50]     Trans CRC Status (~EOP)
+ * [49:48]     EOP type
+ * [47]                EOP trap
+ * [46]                Trapping priority
+ * [45]                Trapping Channel
+ * [44:43]     Bad ack sent
+ * [42:41]     Good ack sent
+ * [40]                Queueing Packet (~EOP)
+ * [39:36]     Channel trapped bits
+ * [35:32]     IProc Trap Value
+ * [31:16]     Network Context (~EOP)
+ * [15:0]      Transaction Type (~EOP)
+ */
+#define IPROC_LastTrans(s)             ((unsigned)((s) >> 55) & 0x1)
+#define IPROC_FirstTrans(s)            ((unsigned)((s) >> 54) & 0x1)
+#define IPROC_Channel(s)               ((unsigned)((s) >> 53) & 0x1)
+#define IPROC_BadLength(s)             ((unsigned)((s) >> 52) & 0x1)
+#define IPROC_TransCRCStatus(s)                ((unsigned)((s) >> 50) & 0x3)
+#define IPROC_EOPType(s)               ((unsigned)((s) >> 48) & 0x3)
+#define IPROC_EOPTrap(s)               ((unsigned)((s) >> 47) & 0x1)
+#define IPROC_InputterPri(s)           ((unsigned)((s) >> 46) & 0x1)
+#define IPROC_InputterChan(s)          ((unsigned)((s) >> 45) & 0x1)
+#define IPROC_BadAckSent(s)            ((unsigned)((s) >> 43) & 0x3)
+#define IPROC_GoodAckSent(s)           ((unsigned)((s) >> 41) & 0x3)
+#define IPROC_QueueingPacket(s)                ((unsigned)((s) >> 40) & 0x1)
+#define IPROC_ChannelTrapped(s)                ((unsigned)((s) >> 36) & 0xF)
+#define IPROC_TrapValue(s)             ((unsigned)((s) >> 32) & 0xF)
+#define IPROC_NetworkContext(s)                ((unsigned)((s) >> 16) & 0xFFFF)
+#define IPROC_TransactionType(s)       ((unsigned)(s) & 0xFFFF)
+
+/* values for IPROC_TransCRCStatus */
+#define CRC_STATUS_GOOD    (0)
+#define CRC_STATUS_DISCARD (1)
+#define CRC_STATUS_ERROR   (2)
+#define CRC_STATUS_BAD     (3)
+
+/* values for IPROC_EOPType */
+#define EOP_GOOD          (1)
+#define EOP_BADACK        (2)
+#define EOP_ERROR_RESET           (3)
+
+/*
+ * Interrupt register bits
+ *
+ * There are up to four sources of interrupt for the MSI port.
+ * The Elan will request 4 ports but may only get either 2 or 1 port. The Interrupts are assigned
+ * as shown below:
+ * No Of MSI ints      Low Prioity                                                     High Prioity
+ *     4               Event Ints      OtherInts               Inputer Ints            Hard Error ints.
+ *                i.e.                 Dproc, Tproc, Sten.     HighPri and LowPri      Link errs, ECC errs,
+ *
+ *     2               Event Ints      All other interrupts.
+ *     1               All together.
+ * 
+ * It is not safe to change the number of sources of interrupt while there may be outstanding,
+ * unserviced interrupts pending.
+ * There two forms of encoding. This has been provided in case an MSI implimentation assumes either
+ * a high value to have a high priority or a low value to have a high priority. This is controled
+ * by a bit in the Elan Pci Control register.
+ */
+#define INT_LinkPortKeyFail            (1<<18)
+#define INT_PciMemErr                  (1<<17)
+#define INT_SDRamInt                   (1<<16)
+#define INT_LinkError                  (1<<15)
+#define INT_IProcCh1HighPri            (1<<14)
+#define INT_IProcCh0HighPri            (1<<13)
+#define INT_IProcCh1LowPri             (1<<12)
+#define INT_IProcCh0LowPri             (1<<11)
+#define INT_DiscardingHighPri          (1<<10)
+#define INT_DiscardingLowPri           (1<<9)
+#define INT_CProcHalted                        (1<<8)
+#define INT_TProcHalted                        (1<<7)
+#define INT_DProcHalted                        (1<<6)
+#define INT_EProc                      (1<<5)
+#define INT_TProc                      (1<<4)
+#define INT_CProc                      (1<<3)
+#define INT_Dma1Proc                   (1<<2)
+#define INT_Dma0Proc                   (1<<1)
+#define INT_MainInterrupt              (1<<0)
+
+#define INT_Units              (INT_EProc | INT_TProc | INT_CProc | INT_Dma1Proc | INT_Dma0Proc)
+#define INT_Inputters          (INT_IProcCh1HighPri | INT_IProcCh0HighPri | INT_IProcCh1LowPri | INT_IProcCh0LowPri)
+#define INT_Discarding         (INT_DiscardingHighPri | INT_DiscardingLowPri)
+#define INT_Halted             (INT_CProcHalted | INT_TProcHalted | INT_DProcHalted)
+#define INT_ErrorInterrupts    (INT_LinkPortKeyFail | INT_PciMemErr | INT_SDRamInt | INT_LinkError)
+
+#define INT_MSI0               INT_MainInterrupt
+#define INT_MSI1               (INT_Units | INT_Discarding | INT_Halted)
+#define INT_MSI2               (INT_Inputters)
+#define INT_MSI3               (INT_ErrorInterrupts)
+
+#define E4_INTERRUPT_REG_SHIFT 32
+#define E4_INTERRUPT_MASK_MASK (0xffffffffULL)
+
+/*
+ * Trap type values - see trapvalues.v
+ */
+
+#define CommandProcInserterError               0x1
+#define CommandProcPermissionTrap              0x2
+#define CommandProcSendTransInvalid            0x3
+#define CommandProcSendTransExpected           0x4
+#define CommandProcDmaQueueOverflow            0x5
+#define CommandProcInterruptQueueOverflow      0x6
+#define CommandProcMemoryFault                 0x7
+#define CommandProcRouteFetchFault             0x8
+#define CommandProcFailCountZero               0x9
+#define CommandProcAddressAlignment            0xa
+#define CommandProcWaitTrap                    0xb
+#define CommandProcMultipleGuards              0xc
+#define CommandProcOpenOnGuardedChan           0xd
+#define CommandProcThreadQueueOverflow         0xe
+#define CommandProcBadData                      0xf
+
+#define DmaProcNoFault                         0x0
+#define DmaProcRouteFetchFault                 0x1
+#define DmaProcFailCountError                  0x2
+#define DmaProcPacketAckError                  0x3
+#define DmaProcRunQueueReadFault               0x4
+#define DmaProcQueueOverflow                   0x5
+
+#define EventProcNoFault                       0x0
+#define EventProcAddressAlignment              0x1
+#define EventProcMemoryFault                   0x2
+#define EventProcCountWrapError                        0x3
+
+#define InputNoFault                           0x0
+#define InputAddressAlignment                  0x1
+#define InputMemoryFault                       0x2
+#define InputInvalidTransType                  0x3
+#define InputDmaQueueOverflow                  0x4
+#define InputEventEngineTrapped                        0x5
+#define InputCrcErrorAfterPAckOk               0x6
+#define InputEopErrorOnWaitForEop              0x7
+#define InputEopErrorTrap                      0x8
+#define InputDiscardAfterAckOk                 0x9
+ 
+typedef struct _E4_Sched_Status
+{
+    E4_uint32 Status;
+    E4_uint32 Restart;
+} E4_Sched_Status;
+ 
+typedef struct _E4_Input_Ptrs
+{
+    E4_uint32 ContextFilterTable;
+    E4_uint32 TrapBasePtr;
+} E4_Input_Ptrs;
+
+#define SCH_StopLowPriQueues           (1 << 0)
+#define SCH_DProcHalt                  (1 << 1)
+#define SCH_TProcHalt                  (1 << 2)
+#define SCH_CProcHalt                  (1 << 3)
+
+#define SCH_CProcTimeout600ns          (1 << 4)
+#define SCH_CProcTimeout1p4us          (2 << 4)
+#define SCH_CProcTimeout3p0us          (3 << 4)
+#define SCH_CProcTimeout6p2us          (4 << 4)
+#define SCH_CProcTimeout12p6us         (5 << 4)
+#define SCH_CProcTimeout25p4us         (6 << 4)
+#define SCH_CProcTimeout51p0us         (7 << 4)
+#define SCH_DiscardLowPriInput         (1 << 7)
+#define SCH_DiscardHighPriInput                (1 << 8)
+
+#define SCH_DProcTimeslice64us         (0 << 9)
+#define SCH_DProcTimeslice128us                (1 << 9)
+#define SCH_DProcTimeslice256us                (2 << 9)
+#define SCH_DProcTimeslice512us                (3 << 9)
+
+#define SCH_Halt                       (SCH_StopLowPriQueues | SCH_DProcHalt | SCH_TProcHalt | SCH_CProcHalt)
+#define SCH_Discard                    (SCH_DiscardLowPriInput | SCH_DiscardHighPriInput)
+
+#define SCH_RestartCProc               (1 << 0)
+#define SCH_RestartTProc               (1 << 1)
+#define SCH_RestartEProc               (1 << 2)
+#define SCH_RestartDma0Proc            (1 << 3)
+#define SCH_RestartDma1Proc            (1 << 4)
+#define SCH_RestartDmaPrefetchProc     (1 << 5)
+#define SCH_RestartCh0LowPriInput      (1 << 6)
+#define SCH_RestartCh1LowPriInput      (1 << 7)
+#define SCH_RestartCh0HighPriInput     (1 << 8)
+#define SCH_RestartCh1HighPriInput     (1 << 9)
+#define SCH_ClearLinkErrorInt          (1 << 10)
+#define SCH_ContextFilterFlush         (1 << 11)
+
+/*
+ * Link state bits.
+ */
+#define LS_LinkNotReady                (1 << 0) /* Link is in reset or recovering from an error */
+#define LS_Locked              (1 << 1) /* Linkinput PLL is locked */
+#define LS_LockError           (1 << 2) /* Linkinput PLL was unable to lock onto the input clock. */
+#define LS_DeskewError         (1 << 3) /* Linkinput was unable to Deskew all the inputs. (Broken wire?) */
+#define LS_PhaseError          (1 << 4) /* Linkinput Phase alignment error. */
+#define LS_DataError           (1 << 5) /* Received value was neither good data or a token. */
+#define LS_FifoOvFlow0         (1 << 6) /* Channel 0 input fifo overflowed. */
+#define LS_FifoOvFlow1         (1 << 7) /* Channel 1 input fifo overflowed. */
+#define LS_Mod45Changed                (1 << 8) /* Mod45 bit has changed. Error setr to force reset. */
+#define LS_PAckNotSeenError    (1 << 9) /* PAck value not returned for this packet. */
+
+/*
+ * Link State Constant defines, used for writing to LinkSetValue
+ */
+
+#define LRS_DataDel0           0x0
+#define LRS_DataDel1           0x1
+#define LRS_DataDel2           0x2
+#define LRS_DataDel3           0x3
+#define LRS_DataDel4           0x4
+#define LRS_DataDel5           0x5
+#define LRS_DataDel6           0x6
+#define LRS_DataDel7           0x7
+#define LRS_DataDel8           0x8
+#define LRS_LinkInValue                0x9
+#define LRS_PllDelValue                0xA
+#define LRS_ClockEven          0xB
+#define LRS_ErrorVal8to0       0xC
+#define LRS_ErrorVal17to9      0xD
+#define LRS_ErrorVal26to18     0xE
+#define LRS_ErrorVal35to27     0xF
+#define LRS_NumLinkDels         0x10
+
+#define LRS_Pllfast             0x40
+
+typedef struct _E4_CommandControl
+{
+    volatile E4_uint32 CommandQueueDescsBase;
+    volatile E4_uint32 CommandRequeuePtr;
+} E4_CommandControl;
+
+#define E4_CommandRequeueBusy          0x80000000      /* Test against read value of CommandRequeuePtr */
+#define E4_CommandRequeueHighPri       0x1             /* Will requeue onto the high pri queue */
+#define E4_QueueDescPtrMask            0x7fffffe0
+
+typedef struct _E4_CommandQueueDesc
+{
+    E4_uint64 CQ_QueuePtrs;
+    E4_uint64 CQ_HoldingValue;         /* 32 bit value for 32 bit accesses or OutOfOrderMask*/
+    E4_uint64 CQ_AckBuffers;           /* Space for 32 4 bit ack buffer values. */
+    E4_uint64 CQ_Control;
+} E4_CommandQueueDesc;
+
+/*
+ * Rev A - CQ_QueuePtrs
+ * [63]                Unused          Should be set to zero.
+ * [62:51]     Unused          (reads as top of InsertPtr)
+ * [50:35]     CompletedPtr    Completed pointer. This is alligned to a byte address.
+ * [34]                Trapped         Will be set if the command has trapped.
+ * [33:32]     Size            Size of queue.
+ * [31]                Used            Will be set if the descriptor has been changed and written back by the elan.
+ * [30:3]      InsertPtr       Insert pointer. This is alligned to a byte address.
+ * [2]         TimedOut        Will be set if the queue timedout executing a command.
+ * [1]         Priority        When set the queue runs at high priority.
+ * [0]         Error           If this becomes set all new data written to the queue is * discarded.
+ *
+ * Rev B - CQ_QueuePtrs
+ * [63]                TimedOut        Will be set if the queue timedout executing a command.
+ * [62]                Priority        When set the queue runs at high priority.
+ * [61]                QueueType       1=will accept unordered 64 bit PCI writes. 0=will accept ordered 32 or 64 bit PCI writes.
+ * [60:51]     Unused          (reads as top of InsertPtr)
+ * [50:35]     CompletedPtr    Completed pointer. This is alligned to a byte address.
+ * [34]                Trapped         Will be set if the command has trapped.
+ * [33:32]     Size            Size of queue.
+ * [31]                Used            Will be set if the descriptor has been changed and written back by the elan.
+ * [30:3]      InsertPtr       Insert pointer. This is alligned to a byte address.
+ * [2]         OrderControl    Holds bit 8 of last PCI accesses. Used by a reordering queue.
+ * [1:0]       ErrorType       This field has the current error status of the queue.
+ */
+
+/* Common between revA and RevB */
+#define CQ_PtrMask             (0x7ffffff8)                    /* 31 bit sdram address */
+#define CQ_PtrOffsetMask       (0x7fff8)
+#define CQ_PtrBaseMask         (0x7ff80000)
+
+#define CQ_InsertPtrShift      (3 - 3)                         /* InsertPtr is 64 bit aligned */
+#define CQ_SizeShift           (32)
+#  define CQ_Size1K            0
+#  define CQ_Size8K            1
+#  define CQ_Size64K           2
+#  define CQ_Size512K          3
+#  define CQ_SizeMask          3
+
+#define CQ_CompletedPtrShift   (35 - 3)                        /* CompletedPtr is 64 but aligned */
+
+#define CQ_Used                        (1ull << 31)
+#define CQ_Trapped             (1ull << 34)
+
+#define CQ_QueuePtrsValue(Size,Inserter,Completer) \
+       (((E4_uint64) (Size) << CQ_SizeShift) | \
+        ((E4_uint64) (Inserter) << CQ_InsertPtrShift) | \
+        ((E4_uint64) (Completer) << CQ_CompletedPtrShift))
+
+#define CQ_InsertPtr(QueuePtrs) \
+       (((E4_uint64) QueuePtrs) & CQ_PtrMask)
+
+#define CQ_CompletedPtr(QueuePtrs) \
+       (((E4_uint32)((QueuePtrs) >> CQ_CompletedPtrShift) & CQ_PtrOffsetMask) | \
+        (CQ_InsertPtr(QueuePtrs) & CQ_PtrBaseMask))
+
+#define CQ_Size(SizeVal)               (1024 * (1 << ((SizeVal)*3)))
+
+/* Rev A specific */
+#define CQ_RevA_Error                  (1 << 0)
+#define CQ_RevA_Priority               (1 << 1)
+#define CQ_RevA_TimedOut               (1 << 2)
+
+/* Rev B specific */
+#define CQ_RevB_ErrorType(QueuePtr)    ((QueuePtr) & (3 << 0))
+#  define CQ_RevB_NoError              (0ull << 0)
+#  define CQ_RevB_Overflowed           (1ull << 0)
+#  define CQ_RevB_InvalidWriteSize     (2ull << 0)
+#  define CQ_RevB_InvalidWriteOrder    (3ull << 0)
+#define CQ_RevB_OrderControl           (1ull << 2)
+
+#define CQ_RevB_QueueType(QueuePtr)    ((QueuePtr) & (1ull << 61))
+#  define CQ_RevB_ReorderingQueue      (1ull << 61)
+#  define CQ_RevB_32bitWriteQueue      (0ull << 61)
+
+#define CQ_RevB_Priority               (1ull << 62)
+#define CQ_RevB_TimedOut               (1ull << 62)
+
+/* 
+ * CQ_AckBuffers - Packet Ack Values
+ */
+#define PackOk                 (0x0)
+#define PackTestFail           (0x1)
+#define PackDiscard            (0x2)
+#define PackError              (0x7)
+#define PackTimeout            (0x8)
+#define PackWaiting            (0xF)
+#define PackValue(val,chan)    (((val) >> ((chan) * 4)) & 0xf)
+
+/*
+ * CQ_Control
+ * [63:35]     ExtractPtr
+ * [34]                Unused
+ * [33:32]     ChannelNotCompleted
+ * [31:24]     Permissions
+ * [23:16]     RestartCount            Decremented after each restart. Will trap when zero
+ * [15:14]     Unused                  Should be set to zero
+ * [13:0]      Context
+ */
+#define CQ_Context(Control)            ((E4_uint32) ((Control) >>  0) & 0x3fff)
+#define CQ_RestartCount(Control)       ((E4_uint32) ((Control) >> 16) & 0x7f)
+#define CQ_ChannelNotCompleted(Control)        ((E4_uint32) ((Control) >> 32) & 3)
+#define CQ_ExtractPtr(Control)         ((E4_uint32) ((Control) >> 32) & 0xFFFFFFF8)
+
+#define CQ_RestartCountShift           16
+
+#define CQ_SetEventEnableBit   (1 << 24)
+#define CQ_WaitEventEnableBit  (1 << 25)
+#define CQ_ModifyEnableBit     (1 << 26)
+#define CQ_WriteEnableBit      (1 << 27)
+#define CQ_ThreadStartEnableBit        (1 << 28)
+#define CQ_DmaStartEnableBit   (1 << 29)
+#define CQ_STENEnableBit       (1 << 30)
+#define CQ_InterruptEnableBit  (1 << 31)
+#define CQ_EnableAllBits        (0xFF000000)
+#define CQ_PermissionMask      (0xFF000000)
+
+#define CQ_ControlValue(Cntx, RestartCount, Permissions) \
+       (((Cntx) & 0x3fff) | (((RestartCount) & 0xff) << 16) | ((Permissions) & CQ_PermissionMask))
+
+/*
+ * This file describes the slave address map of Elan4.
+ *
+ * Elan4 has two PCI 64 bit base address registers. One is setup for elan
+ * local memory and the other is for the command port, elan registers and ebus.
+ *
+ * This file describes the command port, elan registers and ebus BAR. This is a
+ * 26 bit base address register and is split up as follows:
+ * 1 The ebus requires 21 bits of address. 26'h3e00000 to 26'h3ffffff
+ * 2 The control regsiters requires 16 bits of address. 26'h3df0000 to 26'h3dfffff
+ * 3 The command port has the rest. This give just under 8k command ports or about 123 per
+ *   processor of a 64 node SMP.
+ */
+
+/* BAR1 contains the command queues followed by the registers and the Ebus - and is 26 bits */
+/* each command queue has an 8K page associated with it */
+#define CQ_CommandMappingSize          (1 << 13)
+#define CQ_NumCommandDescs             ((1 << (26 - 13)))
+#define CQ_CommandDescsAlignment       ((1 << (26 - 13)) * sizeof (E4_CommandQueueDesc))
+
+/* control reg bits i.e. E4_DataBusMap.SysControlReg */
+#define CONT_EN_ALL_SETS               (1ULL << 0) /* enable cache */
+#define CONT_MMU_ENABLE                        (1ULL << 1) /* bit 0 enables mmu */
+#define CONT_CACHE_HASH_TABLE          (1ULL << 2) /* cache up hash table entries */
+#define CONT_CACHE_CHAINS              (1ULL << 3) /* cache up chain entries */
+#define CONT_CACHE_ROOT_CNTX           (1ULL << 4) /* cache root context table for routes and filters. */
+#define CONT_CACHE_STEN_ROUTES         (1ULL << 5) /* cache up sten packet routes */
+#define CONT_CACHE_DMA_ROUTES          (1ULL << 6) /* cache up dma packet routes */
+
+#define CONT_CACHE_NONE                0ULL
+#define CONT_CACHE_ALL         (CONT_CACHE_HASH_TABLE | CONT_CACHE_CHAINS | CONT_CACHE_ROOT_CNTX | \
+                                CONT_CACHE_STEN_ROUTES | CONT_CACHE_DMA_ROUTES)
+
+/* This controls the format size and position of the MMU hash tables. */
+#define CONT_INHIBIT_MAX_CHAIN_ITEMS   (1ULL << 7)     /* Prevents the MaxChainItems value of 1024 from forcing a translation miss */
+#define CONT_TABLE0_MASK_SIZE_SHIFT    8               /* Defines the size of hash table 0 */
+#define CONT_TABLE0_PAGE_SIZE_SHIFT    13              /* Set the page size for hash table 0 */
+#define CONT_TABLE1_MASK_SIZE_SHIFT    16              /* Defines the size of hash table 1 */
+#define CONT_TABLE1_PAGE_SIZE_SHIFT    21              /* Set the page size for hash table 1 */
+#define CONT_TWO_HASH_TABLES           (1ULL << 24)    /* Sets the MMU to use two hash tables. If not set only 0 used. */
+#define CONT_2K_NOT_1K_DMA_PACKETS     (1ULL << 25)    /* Used to select the default DMA packet size. */
+#define CONT_ALIGN_ALL_DMA_PACKETS     (1ULL << 26)    /* Will force all dma packets to be aligned to a page.*/
+#define CONT_DIRECT_MAP_PCI_WRITES     (1ULL << 27)    /* Will force pci writes to write and flush the dcache.*/
+#define CONT_TLB_FLUSH                 (1ULL << 28)    /* Invalidates the TLB and indicates when flushed */
+#define CONT_CLEAR_WALK_WROTE_TABLES   (1ULL << 29)    /* Used to guarantee that the elan is using new PTE values. */
+#define CONT_ROUTE_FLUSH               (1ULL << 30)    /* Invalidates all route cache entries. */
+#define CONT_CLEAR_LINKPORT_INT                (1ULL << 31)    /* Clears the Linkport key fail interrupt. Reads as 0. */
+#define CONT_CLEAR_SDRAM_ERROR         (1ULL << 32)    /* Clears an EEC error interrupt. Reads as 0. */
+
+/*
+ * These are extra control bits used for testing the DLLs of the SDRAM interface. Most of the Sdram
+ * control bits are defined in xsdram.h
+ */
+#define SDRAM_FIXED_DLL_DELAY_SHIFT    47
+#define SDRAM_FIXED_DLL_DELAY_BITS     5
+#define SDRAM_FIXED_DLL_DELAY_MASK     ((1ULL << SDRAM_FIXED_DLL_DELAY_BITS) - 1ULL)
+#define SDRAM_FIXED_DLL_DELAY(Value)   ((SDRAM_FIXED_DLL_DELAY_MASK & (Value)) << SDRAM_FIXED_DLL_DELAY_SHIFT)
+#define SDRAM_FIXED_DELAY_ENABLE       (1ULL << 52)
+#define SDRAM_GET_DLL_DELAY(Value)     (((Value) >> SDRAM_FIXED_DLL_DELAY_SHIFT) & SDRAM_FIXED_DLL_DELAY_MASK)
+
+#define SDRAM_DLL_CORRECTION_FACTOR    3       /* This is to allow for SSO and ringing on the DQ lines */
+
+#define PAGE_SIZE_4K   0x0
+#define PAGE_SIZE_8K   0x1
+#define PAGE_SIZE_64K  0x2
+#define PAGE_SIZE_512K 0x3
+#define PAGE_SIZE_2M   0x4
+#define PAGE_SIZE_4M   0x5
+#define PAGE_SIZE_64M  0x6
+#define PAGE_SIZE_512M 0x7
+
+#define PAGE_SIZE_MASK 0x7
+#define PAGE_MASK_MASK 0x1f
+
+/* control reg bits i.e. E4_DataBusMap.LinkControlReg */
+#define LCONT_REVA_GREEN_LED           (1 << 0)
+#define LCONT_REVA_YELLOW_LED          (1 << 1)
+#define LCONT_REVA_RED_LED             (1 << 2)
+#define LCONT_REVA_ENABLE_LED_DRIVE    (1 << 3) /* Enable manual setting of the Leds to the bits set above. */
+
+#define LCONT_REVB_DISABLE_TLB_PREFETCH        (1 << 0)
+#define LCONT_REVB_DISABLE_CRC_ERROR_CHECKING  (1 << 1)
+
+
+#define LCONT_EN_SYS_WRITES            (1 << 4) /* Enable linkport writes to sys registers. i.e. all of E4_DataBusMap. */
+#define LCONT_EN_SYS_READS             (1 << 5) /* Enable linkport reads from sys registers. i.e. all of E4_DataBusMap. */
+#define LCONT_EN_USER_WRITES           (1 << 6) /* Enable linkport writes to user registers. i.e. all of E4_User_Regs. */
+#define LCONT_EN_USER_READS            (1 << 7) /* Enable linkport reads from user registers. i.e. all of E4_User_Regs. */
+
+#define LCONT_TEST_VALUE_MASK          0x3ff    /* Value used for test writes and link boundary scan. */
+#define LCONT_TEST_VALUE_SHIFT         8
+#define LCONT_TEST_VALUE(Value)                ((LCONT_LINK_STATE_MASK & (Value)) << LCONT_TEST_VALUE_SHIFT)
+
+/*
+ * State read from LINK_STATE when TEST_VALUE is set to the following values.
+ * TEST_VALUE    LINK_STATE read       TEST_VALUE        LINK_STATE read
+ *    000     -   Data delay count 0      008       -  Data delay count 8
+ *    001     -   Data delay count 1      009       -  Link in value
+ *    002     -   Data delay count 2      00a       -  PLL delay
+ *    003     -   Data delay count 3      00b       -  Clock Delay
+ *    004     -   Data delay count 4      00c       ?  ErrorVal8to0
+ *    005     -   Data delay count 5      00d       ?  ErrorVal17to9
+ *    006     -   Data delay count 6      00e       ?  ErrorVal26to18
+ *    007     -   Data delay count 7      00f       ?  ErrorVal35to27
+ */
+
+#define LCONT_TEST_CONTROL_MASK                0x3     /* Selects and controls the action of the LINK_STATE value. */
+#define LCONT_TEST_CONTROL_SHIFT       18
+
+#define LCONT_READ_ERRORS              0       /* {Mod45RequestChanged, FifoOverflowError, DataError, PhaseError,
+                                                *      DeskewError, LockError, Locked, LinkNotReady} */
+#define LCONT_READ_STATE               1       /* Read valus addressed by TEST_CONTROL value */
+#define LCONT_FIX_LINK_DELAYS          2       /* Sets delays to TEST_CONTROL value */
+#define LCONT_BOUNDARY_SCAN            3       /* Puts link into boundary scan. Outputs TEST_CONTROL value to link,
+                                                * reads LINK_STATE from link. */ 
+
+#define LCONT_LINK_STATE_MASK          0x3ff   /* Read only */
+#define LCONT_LINK_STATE_SHIFT         20      /* Read only */
+#define LCONT_LINK_STATE(ControlRegValue)      (LCONT_LINK_STATE_MASK & ((ControlRegValue) >> LCONT_LINK_STATE_SHIFT))
+
+/* control reg bits i.e. E4_DataBusMap.LinkContSettings */
+#define LCONT_MOD45_DISABLE            (1 << 0) /* is set the link will try to run in TNB mode. */
+#define LCONT_CONFIG_PHASE_MASK                0x7     /* This set the delay through the phase alignment buffer. */
+#define LCONT_CONFIG_PHASE_SHIFT       1
+
+#define LCONT_PLL_REF_VAL_BITS_MASK    0x7f    /* This is the divide value on the LinkIn clock to form the comms PLL */
+#define LCONT_PLL_REF_VAL_BITS_SHIFT   4       /* reference clock. Div value is (n - 2). e.g. to Divide by 7 set to 5. */
+
+#define LCONT_FORCE_COMMSCLK_LOCAL     (1 << 11) /* This must be set at one end of a back to back Elan configuration. */
+#define LCONT_LVDS_VOLTAGE_BITS_MASK   0x3     /* This is used to set the voltage swing on the LVDS link output pads. */
+#define LCONT_LVDS_VOLTAGE_BITS_SHIFT  12      /* reference clock. Div value is (n - 2). e.g. to Divide by 7 set to 5. */
+
+#define LCONT_VOD_170                  0       /* Approximate differential voltage swing in mV of link outputs into */
+#define LCONT_VOD_360                  1       /* a 100 ohm diferential load. */
+#define LCONT_VOD_460                  2
+#define LCONT_VOD_550                  3
+
+#define LCONT_LVDS_TERMINATION_MASK    0x3     /* This set the resistor values of the internal single ended termation */
+#define LCONT_LVDS_TERMINATION_SHIFT   14      /* resistors of the link input and comms input clcok. */
+
+#define LCONT_TERM_55_OHM              0       /* Resistor values for internal termination of LVDS pads. */
+#define LCONT_TERM_50_OHM              1
+#define LCONT_TERM_AUTO_OHM            2       /* Should normally be set to auto. */
+#define LCONT_TERM_45_OHM              3
+
+#define LCONT_LVDS_EN_TERM_UPDATE      (1 << 47) /* This should be asserted and deasserted if LCONT_LVDS_TERMINATION is changed. */
+
+/* Macros used to access and construct MMU hash table and chain entries. */
+/*
+ * Each hash entry is made up of a 64 byte block. Each entry hash two tags where each
+ * tag has 4 PTE's. PTE's 0 to 2 use the bottom 48 bits of a 64 bit word and PTE 3
+ * uses the top 16 bits of 3 64 bit words.
+ *
+ * These macros can be used to build a single PTE. PTE3 needs to be built into a 48 bit
+ * object before they can be used.
+ */
+#define PTE_ENTRY_MASK         0x0000ffffffffffffULL
+#define PTE_TYPE_MASK          0x000000000000000fULL   
+#define PTE_PERM_MASK          0x00000000000000f0ULL
+#define PTE_PERM_TYPE_MASK     0x00000000000000ffULL
+#define PTE_REF_MASK           0x0000000000000100ULL
+#define PTE_PPN_MASK           0x00007ffffffffe00ULL
+#define PTE_MOD_MASK           0x0000800000000000ULL
+#define PTE_TOPADDR_MASK       0x0000600000000000ULL
+
+#define PTE_MOD_SHIFT          47
+#define PTE_PPN_SHIFT          9
+#define PTE_REF_SHIFT          8
+#define PTE_PERM_SHIFT         4
+#define PTE_TYPE_SHIFT         0
+
+#define PTE_PADDR_SHIFT                (12 - 9)                /* Physical addresses are shifted down 3 this to go into the PTE */
+
+
+/* Values required for tag 3 */
+#define PTE_REF_3                      0x0100000000000000ULL
+#define PTE_MOD_3                      0x8000000000000000ULL
+#define PTE_ENTRY_MASK_3               0xffff000000000000ULL
+#define PTE_PERM_TYPE_MASK_3           0x00ff000000000000ULL
+#define PTE_ENTRY_3_FOR_0(NewPte)      ((NewPte << (48)) & PTE_ENTRY_MASK_3)
+#define PTE_ENTRY_3_FOR_1(NewPte)      ((NewPte << (32)) & PTE_ENTRY_MASK_3)
+#define PTE_ENTRY_3_FOR_2(NewPte)      ((NewPte << (16)) & PTE_ENTRY_MASK_3)
+
+/* Values required for the tags */
+#define TAG_CONTEXT_MASK               0x0000000000003fffULL
+#define TAG_ADDRESS_MASK               0xfffffffff8000000ULL
+#define TAG_CHAINPTR_18TO6_MASK                0x0000000007ffc000ULL
+#define TAG_CHAINPTR_LOW_SHIFT         (14 - 6)
+#define TAG_CHAINPTR_30TO19_MASK       0x0000000003ffc000ULL
+#define TAG_CHAINPTR_HIGH_SHIFT                (19 - 14)
+#define TAG_COPY_BIT                   0x0000000004000000ULL
+
+/*
+ * This takes number loaded into the control register and returns the page size as a power of two.
+ */
+
+#define E4_PAGE_SIZE_TABLE             E4_uint32 const PageSizeTable[] = {12, 13, 16, 19, 21, 22, 26, 29}
+#define E4_PAGE_SIZE_TABLE_SIZE                (sizeof(PageSizeTable)/sizeof(PageSizeTable[0]))
+
+/*
+ * This macro generates a hash block index.
+ *
+ * Cntx                 This is the 14 bit context. It should not be larger than 14 bits.
+ * VAddr        This is the 64 bit virtual address. It does not require any masking and can be a byte address.
+ * PageSize     This is the value loaded into the control register for this hash table.
+ * HashTableMask This should be set mask out upper bits past the end of the hash table.
+ */
+#define E4MMU_SHIFT_ADDR(VAddr, Shift) \
+    ((((E4_uint32)(VAddr)) >> (Shift)) | (((E4_uint32)((VAddr) >> 32)) << (32 - (Shift))))
+
+#define E4MMU_CONTEXT_SCRAMBLE(Cntx) \
+             ((((Cntx) << 8) | ((Cntx) >> 6)) ^ (((Cntx) << 15) | ((Cntx) << 1)))
+
+#define E4MMU_HASH_INDEX(Cntx, VAddr, PageShift, HashTableMask)                \
+           ((E4MMU_SHIFT_ADDR(VAddr, (PageShift) + 2) ^ E4MMU_CONTEXT_SCRAMBLE(Cntx)) & (HashTableMask))
+
+#define E4MMU_TAG(vaddr,ctx)   (((vaddr) & TAG_ADDRESS_MASK) | ((ctx) & TAG_CONTEXT_MASK))
+
+#define E4MMU_TAG2VADDR(tag,hashidx,PageShift,HashTableMask)   \
+               (((tag) & TAG_ADDRESS_MASK) | ((((hashidx) ^ E4MMU_CONTEXT_SCRAMBLE((tag) & TAG_CONTEXT_MASK)) & (HashTableMask)) << ((PageShift + 2))))
+
+/*
+ * Detailed bit descriptions for the tags and PTE's are better done with the macros
+ * defined above.
+ */
+typedef struct _E4_HashTableEntry
+{
+   E4_uint64   Tag[2];
+   E4_uint64   TagPTE[2][3];
+} E4_HashTableEntry;
+
+#define E4MMU_TAG_OFFSET(tag)          ((tag) << 3)
+#define E4MMU_PTE_LOW_OFFSET(tag,pte)  ((((tag)*3 + (pte) + 2) << 3))
+#define E4MMU_PTE_HIGH_OFFSET(tag,pte) ((((tag)*3 + (pte) + 2) << 3) + 4)
+#define E4MMU_PTE3_WORD0_OFFSET(tag)   ((((tag)*3 + 2) << 3) + 6)
+#define E4MMU_PTE3_WORD1_OFFSET(tag)   ((((tag)*3 + 3) << 3) + 6)
+#define E4MMU_PTE3_WORD2_OFFSET(tag)   ((((tag)*3 + 4) << 3) + 6)
+
+
+/*
+ * Hash0AddrBits is the size of the hash table in bytes as a power of 2.
+ * e.g. 11 would give 32 hash entries where each entry is 64 bytes.
+ */
+#define SETUP_HASH_TABLES(Hash0PageSize, Hash0AddrBits, Hash1PageSize, Hash1AddrBits)  \
+                         (((Hash0PageSize) << CONT_TABLE0_PAGE_SIZE_SHIFT) |   \
+                          ((Hash0AddrBits) << CONT_TABLE0_MASK_SIZE_SHIFT) |   \
+                          ((Hash1PageSize) << CONT_TABLE1_PAGE_SIZE_SHIFT) |   \
+                          ((Hash1AddrBits) << CONT_TABLE1_MASK_SIZE_SHIFT))
+
+/* ECC status register */
+#define ECC_Addr(s)                    ((s) & 0x7ffffff8ULL)
+#define ECC_Syndrome(s)                        (((s) >> 32) & 0xffffULL)
+#define ECC_RisingDQSSyndrome(s)       (((s) >> 32) & 0xffULL)
+#define ECC_FallingDQSSyndrome(s)      (((s) >> 40) & 0xffULL)
+#define ECC_UncorrectableErr(s)        (((s) >> 48) & 1ULL)
+#define ECC_MultUncorrectErrs(s)       (((s) >> 49) & 1ULL)
+#define ECC_CorrectableErr(s)          (((s) >> 50) & 1ULL)
+#define ECC_MultCorrectErrs(s)         (((s) >> 51) & 1ULL)
+
+/* Permission type saved in a PTE. This is a four bit field */
+#define PERM_Disabled          0x0
+#define PERM_Unused            0x1
+#define PERM_LocDataRead       0x2
+#define PERM_LocDataWrite      0x3
+#define PERM_LocRead           0x4
+#define PERM_LocExecute                0x5
+#define PERM_ReadOnly          0x6
+#define PERM_LocWrite          0x7
+#define PERM_LocEventOnly      0x8
+#define PERM_LocEventWrite     0x9
+#define PERM_RemoteEvent       0xa
+#define PERM_RemoteAll         0xb
+#define PERM_RemoteReadOnly    0xc
+#define PERM_RemoteWriteLocRead        0xd
+#define PERM_DataReadWrite     0xe
+#define PERM_NoFault           0xf
+
+#define PERM_Mask              0xf
+
+/* Permission type hints to device driver */
+#define PERM_Preload           0x10
+
+#define PTE_SetPerm(Perm)      (((Perm) & PERM_Mask) << 4)
+
+/* Control info saved in the lookup field of the TLB */
+#define PTE_PciNotLocal                (1ULL << 0)             /* Directs the access to the PCI interface */
+#define PTE_BigEndian          (1ULL << 1)             /* Valid for PCI entries only */
+#define PTE_RelaxedOrder       (1ULL << 2)             /* Valid for PCI entries only */
+#define PTE_DontSnoop          (1ULL << 3)             /* Valid for PCI entries only */
+
+#define PTE_UseFixedSet                (1ULL << 1)             /* Value for non PCI entries only */
+#define PTE_CommandQueue       (1ULL << 2)             /* Value for non PCI entries only */
+#define PTE_SetFixedSetNo(Set) ((((Set) & 3) << 2) | PTE_UseFixedSet)
+
+#define PTE_TypeBitsMask       (0xfULL)
+#define PTE_PermissionTypeMask (0xfULL << 4)
+#define PTE_Referenced         (1ULL << 8)
+#define PTE_PhysicalPageNoMask (0x7ffffffffe00ULL)
+#define PTE_Modified           (1ULL << 47)
+
+#define PTE_PhysicalAddrShiftIntoPTE   (12 - 9)
+
+/* define page table entry bit fields */
+#define TLB_PageSizeBits       (3 << 0)
+#define TLB_ACCBits            (7 << 2)
+#define TLB_LocalBit           (1 << 5)
+#define TLB_PCI64BitTargetBit  (1 << 6)
+#define TLB_PCIBigEndianBit    (1 << 7)
+
+#define TLB_ModifiedBit                (1 << 55)
+#define TLB_ReferencedBit      (1 << 63)
+
+/* Used to read values from the tlb. */
+#define TLB_TlbReadCntBitsSh   56
+#define TLB_UseSelAddrSh       (1ULL << 60)
+#define TLB_WriteTlbLine       (1ULL << 61)
+
+#define TLB_SEL_LINE(LineNo) (TLB_UseSelAddrSh | \
+                             ((E4_uint64)((LineNo) & 0xf) << TLB_TlbReadCntBitsSh))
+
+#define TLB_NUM_ENTRIES                16
+/*
+ * The following macros are used with the test access port (TlbLineValue) for the TLBs.
+ */
+#define TLV_DoPciAccess                        (1ULL << 0)
+#define TLV_CommandAccess              (1ULL << 1)
+#define TLV_DoCacheAccess              (1ULL << 2)
+#define TLV_notStartTLBWalk            (1ULL << 3)
+#define TLV_UseFixedSet                        (1ULL << 4)
+#define TLV_BigEndian                  (1ULL << 4)
+#define TLV_RelaxedOrder               (1ULL << 5)
+#define TLV_DontSnoop                  (1ULL << 6)
+#define TLV_FixedSetNo_MASK            (3ULL << 5)
+#define TLV_PciTypeBits_MASK           (7ULL << 4)
+#define TLV_LookupBits_MASK            (0x7fULL)
+#define TLV_MissErr                    (1ULL << 7)
+#define TLV_TypeBits                   (0xffULL)
+
+#define TLV_PhysicalAddr_MASK          (0x3fffffffff000ULL)
+
+#define TLV_TlbTesting                 (1ULL << 51)
+#define TLV_SelectUnitsTlbRead         (1ULL << 52)
+#define TLV_SelectTProcTlbRead         (1ULL << 53)
+
+#define TLV_TlbLineSelect_MASK         (0xf)
+#define TLV_UnitsTlbLineSelect_SHIFT   (54)
+#define TLV_TProcTlbLineSelect_SHIFT   (59)
+#define TLV_EnableUnitsTlbRead         (1ULL << 58)
+#define TLV_EnableTProcTlbRead         (1ULL << 63)
+
+/*
+ * Use this macro to enable direct testing of the Units TLB.
+ * When Line is in the range 0 to 15 a TLB line is selected for reading or writing.
+ * When Line is set to -1 the tlb will be activated to perform a match.
+ */
+#define TLV_UnitsTlbLineSel(Line) (((Line) == -1) ? 0ULL : \
+    (TLV_EnableUnitsTlbRead | ((E4_uint64)((Line) & TLV_TlbLineSelect_MASK) << TLV_UnitsTlbLineSelect_SHIFT)))
+#define TLV_TProcTlbLineSel(Line) (((Line) == -1) ? 0ULL : \
+    (TLV_EnableTProcTlbRead | ((E4_uint64)((Line) & TLV_TlbLineSelect_MASK) << TLV_TProcTlbLineSelect_SHIFT)))
+ 
+/* 
+ * Thread_Trap_State
+ *  see f_RegFileControl.v TProcStatus
+ */
+#define TS_HaltThread                (1 << 0)
+#define TS_TrapForTooManyInstructions (1 << 1)
+#define TS_InstAccessException       (1 << 2)
+#define TS_Unimplemented             (1 << 3)
+#define TS_DataAccessException       (1 << 4)
+#define TS_DataAlignmentError        (1 << 5)
+#define TS_TrapForUsingBadData       (1 << 6)
+#define TS_TrapTypeMask                      (0x7f)
+#define TS_DataPortNo(ts)            (((ts) >> 7) & 7)
+#define TS_TrappedFlag               (1 << 10)
+#define TS_MemLock                   (1 << 11)
+#define TS_XCCshift                  12
+#define TS_XCCmask                   0xff
+#define TS_ICC(ts)                   (((ts) >> 12) & 15)
+#define TS_XCC(ts)                   (((ts) >> 16) & 15)
+#define TS_InstValid_F               (1 << 20)
+#define TS_InstValid_R               (1 << 21)
+#define TS_InstValid_E               (1 << 22)
+#define TS_InstValid_W               (1 << 23)
+#define TS_HighPriority                      (1 << 24)
+#define TS_RemoteThread                      (1 << 25)
+#define TS_TProcTranslationInProgress (1 << 26)
+#define TS_MemLock_E                 (1 << 27)
+
+/* Thread run queue entries */
+typedef struct E4_ThreadRegs
+{
+    E4_uint64 Registers[7];
+} E4_ThreadRegs;
+
+typedef struct E4_TProcQueueEntry
+{
+    E4_ThreadRegs      Regs;                   /* XXXX: jon check this */
+    E4_uint64          Context;                /* XXXX: jon check this */
+} E4_TProcQueueEntry;
+
+typedef struct E4_DProcQueueEntry
+{
+    E4_DMA             Desc;
+    E4_uint64          Pad;
+} E4_DProcQueueEntry;
+
+/*
+ * Packet acknowledge values.
+ */
+#define E4_PAckOk      0
+#define E4_PAckTestFail        1
+#define E4_PAckDiscard 2
+#define E4_PAckError   3
+
+/*
+ * return values from breaktest instruction.
+ */
+#define ICC_CARRY_BIT           (0x1ULL << 0)  /* Breaktest: Load pending         */
+#define ICC_ZERO_BIT            (0x1ULL << 1)  /* Breaktest: Time to break        */
+#define ICC_SIGNED_BIT          (0x1ULL << 2)  /* Breaktest: Another thread ready */
+#define ICC_TPROC_RDY_LOW_PRI   (0x1ULL << 3)
+#define ICC_TPROC_RDY_HIGH_PRI  (0x1ULL << 4)
+#define ICC_RUNNING_HIGH_PRI    (0x1ULL << 5)
+#define ICC_RUNNING_AS_REMOTE   (0x1ULL << 6)
+#define ICC_TIME_TO_BREAK       (0x1ULL << 7)
+#define ICC_RS1LOAD_PENDING     (0x1ULL << 8)
+#define ICC_TPROC_HALT          (0x1ULL << 9)
+
+/*
+ * Main Interrupt cookies
+ * [63:14]     user cookie
+ * [13:0]      context
+ */
+#define E4_MAIN_INT_SHIFT              14
+#define E4_MAIN_INT_COOKIE(cookie)     ((cookie) >> E4_MAIN_INT_SHIFT)
+#define E4_MAIN_INT_CTX(cookie)                ((cookie) & 0x3FFF)
+
+typedef E4_uint64 E4_MainIntEntry;
+
+#define E4_MainIntEntrySize    sizeof (E4_MainIntEntry)
+
+/*
+ * The internal databus is 64 bits wide.
+ * All writes to the internal registers MUST be made with 64 bit write operations.
+ * These can be made up of pairs 32 bit writes on the PCI bus. The writes will be
+ * treated as nops if they are performed with two separate 32 bit writes.
+ */
+typedef volatile struct _E4_DataBusMap
+{
+   E4_uint64           InputTrans[4][16];                                                                      /* 0x000 */
+
+   E4_uint64           Dma0TransAddr;                                                                          /* 0x200 */
+   E4_DMA              Dma0Desc;       /* Current Dma0 registers */                                            /* 0x208 */
+
+   E4_uint64           Dma1TransAddr;                                                                          /* 0x240 */
+   E4_DMA              Dma1Desc;       /* Current Dma1 registers */                                            /* 0x248 */
+  
+   E4_uint64           Dma0LastPacketSize;                                                                     /* 0x280 */
+   E4_uint64           Dma0ThisPacketSize;                                                                     /* 0x288 */
+   E4_uint64           Dma0DescSizeInProg;                                                                     /* 0x290 */
+   E4_uint64           Dma0BytesToPrefetch;                                                                    /* 0x298 */
+   E4_uint64           Dma0PrefetchAddr;                                                                       /* 0x2a0 */
+   E4_uint64           EventCountAndType;                                                                      /* 0x2a8 */
+   E4_uint64           EventParameters[2];                                                                     /* 0x2b0 */
+  
+   E4_uint64           Dma1LastPacketSize;                                                                     /* 0x2c0 */
+   E4_uint64           Dma1ThisPacketSize;                                                                     /* 0x2c8 */
+   E4_uint64           Dma1DescSizeInProg;                                                                     /* 0x2d0 */
+   E4_uint64           Dma1BytesToPrefetch;                                                                    /* 0x2d8 */
+   E4_uint64           Dma1PrefetchAddr;                                                                       /* 0x2e0 */
+   E4_Input_Ptrs       InputTrapAndFilter;                                                                     /* 0x2e8 */
+   E4_uint64           EventAddress;                                                                           /* 0x2f0 */
+   E4_QueuePtr         MainIntQueuePtrs;                                                                       /* 0x2f8 */
+   
+   E4_uint64           Event_Copy[16];                                                                         /* 0x300 */
+
+   E4_uint64           CommandCopy[7];                                                                         /* 0x380 */
+   E4_uint64           CommandHold;                                                                            /* 0x3b8 */
+
+   E4_uint64           InputQueueDesc[4];                                                                      /* 0x3c0 */
+
+   /* Run queue Pointers */
+   E4_uint64           DProcLowPriPtrs;                                                                        /* 0x3e0 */
+   E4_uint64           DProcHighPriPtrs;                                                                       /* 0x3e8 */
+   E4_uint64           TProcLowPriPtrs;                                                                        /* 0x3f0 */
+   E4_uint64           TProcHighPriPtrs;                                                                       /* 0x3f8 */
+
+   E4_uint64           CProcStatus;                                                                            /* 0x400 */
+   E4_uint64           TProcStatus;                                                                            /* 0x408 */
+   E4_uint64           IProcStatus;                                                                            /* 0x410 */
+   E4_uint64           EProcStatus;                                                                            /* 0x418 */
+   E4_uint64           DProc0Status;                                                                           /* 0x420 */
+   E4_uint64           DProc1Status;                                                                           /* 0x428 */
+   E4_Sched_Status     SchedStatus;                                                                            /* 0x430 */
+
+   E4_uint64           LoadIProcCntxFilter;    /* Will load one of 4 cntx filter regs. Write only */           /* 0x438 */
+
+   E4_CommandControl   CommandControl;                                                                         /* 0x440 */
+   E4_uint64           CommandCacheTestPort;                                                                   /* 0x448 */
+   E4_uint64           CommandLowPriRunPtrs;                                                                   /* 0x450 */
+   E4_uint64           CommandHighPriRunPtrs;                                                                  /* 0x458 */
+   E4_uint64           CommandSchedDataPort[4];                                                                /* 0x460 */
+
+   E4_uint64           DmaRouteBuffer[2][2];   /* Write only. Should not be written to. */                     /* 0x480 */
+   E4_uint64           StenRouteBuffer[2];     /* Write only. Should not be written to. */                     /* 0x4a0 */
+   E4_uint64           pad4[0x098 - 0x096];                                                                    /* 0x4b0 */
+
+   E4_uint64           DmaAlignmentPort[8];    /* Write only. Should only be written to clear the prev reg. */ /* 0x4c0 */
+
+   E4_uint64           MmuBlockEntry[8];       /* Used for hash table and chain fetches */                     /* 0x500 */
+   E4_uint64           WriteUnitsTlbLine[3];                                                                   /* 0x550 */
+   E4_uint64           pad5;                                                                                   /* 0x540 */
+   E4_uint64           WriteTProcTlbLine[3];                                                                   /* 0x568 */
+   E4_uint64           pad6;                                                                                   /* 0x540 */
+
+   E4_uint64           MmuTableBasePtrs;       /* Both tables packed into a single 64 bit value */             /* 0x580 */
+   E4_uint64           MmuFaultAndRootCntxPtr; /* Both packed into a single 64 bit value */                    /* 0x588 */
+   E4_uint64           UnitsVAddr;                                                                             /* 0x590 */
+   E4_uint64           TProcVAddr;                                                                             /* 0x598 */
+   E4_uint64           UnitsCntx;                                                                              /* 0x5a0 */
+   E4_uint64           TProcCntx;              /* Read only. Writes access VProcCacheWritePort */              /* 0x5a8 */
+   E4_uint64           FaultAddrReg;                                                                           /* 0x5b0 */
+   E4_uint64           FaultTypeAndContextReg;                                                                 /* 0x5b8 */
+
+   E4_uint32           SysControlReg;                                                                          /* 0x5c0 */
+   E4_uint32           CacheTagValue;                                                                          /* 0x5c4 */
+   E4_uint64           TlbLineValue;                                                                           /* 0x5c8 */
+   E4_uint64           SDRamConfigReg;                                                                         /* 0x5d0 */
+   E4_uint32           InterruptMask;                                                                          /* 0x5d8 */
+   E4_uint32           InterruptReg;                                                                           /* 0x5dc */
+   E4_uint64           SDRamECCStatus;                                                                         /* 0x5e0 */
+   E4_uint32           LinkControlReg;                                                                         /* 0x5e8 */
+   E4_uint32           LinkContSettings;                                                                       /* 0x5ec */
+   E4_uint64           LinkPortKey;                                                                            /* 0x5f0 */
+   E4_uint64           LinkPortLock;                                                                           /* 0x5f8 */
+
+   E4_uint64           SDRamWriteBuffer[4][8];                                                                 /* 0x600 */
+   E4_uint64           SDRamReadBuffer[4][8];                                                                  /* 0x700 */
+
+   E4_uint64           TProcRegs[64];                                                                          /* 0x800 */
+   E4_uint64           TProcStartUp[8];        /* Not to be used except by the elan itself */                  /* 0xa00 */
+
+   E4_uint64           LoadPending;                                                                            /* 0xa40 */
+   E4_uint64           StortPending;                                                                           /* 0xa48 */
+   E4_uint64           DirtyBits;                                                                              /* 0xa50 */
+   E4_uint64           BadBits;                                                                                /* 0xa58 */
+
+   E4_uint64           ICachePort_Cntl_Addr;                                                                   /* 0xa60 */
+   E4_uint64           Thread_Trap_State;                                                                      /* 0xa68 */
+
+/* Instruction buffer (4 * 32 bit words) */
+   E4_uint64           nPC_W;                                                                                  /* 0xa70 */
+   E4_uint64           PC_W;                                                                                   /* 0xa78 */
+
+   E4_uint64           ICacheFillData[8];                                                                      /* 0xa80 */
+   E4_uint64           ICachePort[8];                                                                          /* 0xac0 */
+
+   E4_uint64           PciDataBufs[4][8];                                                                      /* 0xb00 */
+
+   E4_uint64           CommandQueueBuffer[128];                                                                /* 0xc00 */
+} E4_DataBusMap;
+
+#define LINK_PORT_LOCK_VALUE   0x123456789abcdef0ULL
+
+/*
+ * These macros are used to setup the thread pcoessors ICache.
+ */
+#define E4_ICacheTagAddrShift          6
+#define E4_AccessICacheRams            1
+#define E4_InvalidTagValue             0xffffffffffffffffULL
+#define E4_ICacheSizeInBytes           (1024*16)
+#define E4_ICacheLineSizeInBytes       (64)
+#define E4_ICacheLines                 (E4_ICacheSizeInBytes/E4_ICacheLineSizeInBytes)
+#define E4_ICachePortSize              ( (sizeof((E4_DataBusMap *) 0)->ICachePort) /   \
+                                         (sizeof((E4_DataBusMap *) 0)->ICachePort[0]))
+
+#define E4_ICacheFixupInsn             0xc0b02f95ull           /* st1 [%r0 +  0xf95] */
+#define E4_ICacheFixupAddr             0xf95ull
+#define E4_ICacheFixupOffset           0xfc0
+
+/*
+ * Event interrupt
+ */
+typedef volatile union _E4_EventInt
+{
+   E4_uint64    ForceAlign;
+   struct {
+       E4_uint32 IntCookie;
+       E4_uint32 EventContext; /* Bits 16 to 28 */
+    } s;
+} E4_EventInt;
+
+/*
+ * The following are used to interpret a fault status register.
+ */
+
+/*
+ * FSR[14:0] - AccessType
+ *
+ * T = Type bit
+ * S = size bit. Size is in units of 64 bits or 8 bytes.
+ * E = Byte end pointer. Used to define the last written byte of the last 64 bits written.
+ * D = Data type bit. Used for endian conversion in the PCI interface.
+ * C = Used by the cache to decide if this access should allocate a cache line.
+ * d = Set if dma read or write data data. This is used to guarantee order at the PCI interface.
+ * A = Access type used to check permissions by the MMU in a virtual access.
+ * P = Part Write. If set some byte enables may be used. Effects the action of a cache miss.
+ */
+
+/* FSR[7:0] */
+/* bit 7 => virtual write */
+#define AT_VirtualWriteAccBit          (1 << 7)                /* AAADDdC1EEESSSS = Virtual Write */
+#define AT_VirtualWriteSizeMask                0xf                     /* size of write access (0 => 128 bytes) */
+#define AT_VirtualWriteEndPtrShift     4                       /* end byte pointer for part write block */
+#define AT_VirtualWriteEndPtrMask      0x7
+
+/* else bit 6 => virtual read */
+#define AT_VirtualReadAccBit           (1 << 6)                /* AAADDdC01SSSSSS = Virtual Read */
+#define AT_VirtualReadSizeMask         0x3f                    /* size of read access (0 => 512 bytes) */
+
+/* else => special access */
+#define AT_SelBitsMask                 0xf                     /* Bits to select the type of acces from */
+#define AT_SelBitsShift                        0x4
+#define AT_SpecialRd                   (0x0 << 4)              /* AAADDdC0000TTTT = Special read Access */
+#define AT_SpecialWr                   (0x1 << 4)              /* AAADDdC0001TTTT = Special write Access */
+#define AT_PhysicalRd                  (0x2 << 4)              /* AAADDdC00100SSS = Physical Read */
+#define AT_PhysicalWr                  (0x3 << 4)              /* AAADDdC0011PSSS = Physical write */
+
+#define AT_OtherSizeMask               0xf                     /* Size bits used by all other accesses. 0=128 bytes */
+#define AT_SpecialBitsMask             0xf                     /* Bits used to define the special access types */
+#define AT_CacheSizeBitsMask           0x7                     /* Size bits used for local accesses. 0=64 */
+#define AT_CachePhysPartWriteBit       0x8                     /* This bit is set if the access is a part write to the cache */
+
+/* Special memory access operations */
+#define AT_RegAccess                   0x0
+#define AT_GetCntxFilter               0xe                     /* Only used by special reads */
+#define AT_RouteFetch                  0xf                     /* Only used by special reads */
+
+/* FSR[9:8] */
+#define AT_NonAlloc                    (1 << 8)                /* 1=Do not fill cache with this data */
+#define AT_DmaData                     (1 << 9)                /* This is a DMA read access. Required to guarantee dma read order. */
+
+/* FSR[11:10] - Data Type - defines data type for endian conversion in PCI interface*/
+#define AT_BlkDataTyMask               0x3
+#define AT_BlkDataTyShift              10
+
+#define AT_BlkDataType(FSR)            (((FSR) >> AT_BlkDataTyShift) & AT_BlkDataTyMask)
+#define AT_TypeByte                    0x0
+#define AT_TypeHWord                   0x1
+#define AT_TypeWord                    0x2
+#define AT_TypeDWord                   0x3
+
+/* FSR[14:12] - Access Permissions */
+#define AT_PermBitsMask                        0x7
+#define AT_PermBitsShift               12
+
+#define AT_Perm(FSR)                   (((FSR) >> AT_PermBitsShift) & AT_PermBitsMask)
+#define AT_PermLocalDataRead           0x0
+#define AT_PermLocalDataWrite          0x1
+#define AT_PermRemoteRead              0x2
+#define AT_PermRemoteWrite             0x3
+#define AT_PermExecute                 0x4
+#define AT_PermLocalEvent              0x5
+#define AT_PermRemoteEvent             0x7
+
+/* FSR[22:15] - reason for fault */
+
+#define FSR_WalkForThread              (1 << 15) /* The thread processor caused the fault */
+#define FSR_Walking                    (1 << 16) /* The fault was caused during a hash table access */
+#define FSR_NoTranslationsFound                (1 << 17) /* The hash table did not contain a matching tag */
+#define FSR_WalkingProtectionFault     (1 << 18) /* A protection fault was detected while walking */
+#define FSR_HashTable1                 (1 << 19) /* Was accessing hash table 1 not 0 */
+#define FSR_RouteVProcErr              (1 << 20) /* This is an invalid vproc for a route fetch */
+#define FSR_FaultForBadData            (1 << 21) /* Bad data (double bit ECC error) while performing a walk access */
+#define FSR_FaultForMaxChainCount      (1 << 22) /* The Elan4 has walked a chain of 1024 items. */
+
+typedef volatile struct _E4_FaultSave
+{
+    E4_uint64 FSRAndFaultContext;                 /* Bits 0-31 : FaultContext. Bits 32-63 : FaultStatus Register */
+    E4_uint64 FaultAddress;
+} E4_FaultSave;
+
+#define FaultSaveContext(FSRAndFaultContext)   ((E4_uint32) ((FSRAndFaultContext) & 0xFFFFFFFF))
+#define FaultSaveFSR(FSRAndFaultContext)       ((E4_uint32) ((FSRAndFaultContext) >> 32))
+
+typedef union E4_TrTypeCntx
+{
+   E4_uint32 TypeContext;
+   struct
+   {
+#if (BYTE_ORDER == LITTLE_ENDIAN) || defined(__LITTLE_ENDIAN__)
+      E4_uint32 Type:16;               /* Transaction type field */
+      E4_uint32 Context:13;            /* Transaction context */
+      E4_uint32 TypeCntxInvalid:1;     /* Bit  29 */
+      E4_uint32 StatusRegValid:1;      /* Bit  30 */
+      E4_uint32 LastTrappedTrans:1;    /* Bit  31 */
+#else
+      E4_uint32 LastTrappedTrans:1;    /* Bit  31 */
+      E4_uint32 StatusRegValid:1;      /* Bit  30 */
+      E4_uint32 TypeCntxInvalid:1;     /* Bit  29 */
+      E4_uint32 Context:13;            /* Transaction context */
+      E4_uint32 Type:16;               /* Transaction type field */
+#endif
+   } s;
+} E4_TrTypeCntx;
+
+#define MAX_TRAPPED_TRANS      28
+#define TRANS_DATA_DWORDS      16
+#define TRANS_DATA_BYTES       128
+#define NO_OF_INPUT_CHANNELS   4
+
+#define CH0_LOW_PRI_CHAN       0
+#define CH1_LOW_PRI_CHAN       1
+#define CH0_HIGH_PRI_CHAN      2
+#define CH1_HIGH_PRI_CHAN      3
+
+/* Words have been swapped for big endian access when fetched with dword access from elan.*/
+typedef struct _E4_IprocTrapHeader
+{
+   E4_uint64   TrAddr;
+   E4_uint64   IProcStatusCntxAndTrType;
+} E4_IprocTrapHeader;
+
+typedef struct _E4_IprocTrapData
+{
+   E4_uint64 Data[TRANS_DATA_DWORDS];
+} E4_IprocTrapData;
+
+/*
+ * This struct defines the trap state for the inputers. It requires a contiguous 16K byte block of local memory.
+ * The channel bits have been grouped to the low end of the address to force all Identify cookies to use the
+ * same cache line.
+ */
+typedef struct _E4_IprocTrapState
+{
+   E4_IprocTrapData   TrData[MAX_TRAPPED_TRANS][NO_OF_INPUT_CHANNELS];
+   E4_IprocTrapHeader TrHeader[MAX_TRAPPED_TRANS][NO_OF_INPUT_CHANNELS];
+   E4_uint64         pad[8*NO_OF_INPUT_CHANNELS];
+} E4_IprocTrapState;
+
+/*
+ * 64 kbytes of elan local memory. Must be aligned on a 64k boundary
+ */
+#define E4_LowPriQueueSize     0x400
+#define E4_HighPriQueueSize    0x100
+
+typedef struct _E4_FaultSaveArea
+{
+   E4_FaultSave                TProcData[8];
+   E4_FaultSave                TProcInst;
+   E4_FaultSave                Dummy[7];
+   E4_FaultSave                SchedProc;
+   E4_FaultSave                DProc;
+   E4_FaultSave                EventProc;
+   E4_FaultSave                IProc;
+   E4_FaultSave                DProcData[4];
+   E4_FaultSave                QReadData[8];
+} E4_FaultSaveArea;
+
+/* Macros to manipulate event queue pointers */
+/*     generate index in EventIntQueue */
+#define E4_EVENT_INTQ_INDEX(fptr)      (((fptr) & 0x1fff) >> 3)
+/*     generate next fptr */
+#define E4_EVENT_INTQ_NEXT(fptr)       ((((fptr) + 8) & ~0x4000) | 0x2000)
+
+typedef struct _E4_CommandPort
+{
+   volatile E4_uint64 Command[1024];   /* a whole 8k page */
+} E4_CommandPort;
+
+/*
+ * This is the allocation of unit numbers within the ELAN. It is used to extract the fault address
+ * and fault type after a unit has trapped on a memory fetch. Only units that can generate traps
+ * have been included.
+ */
+#define CUN_TProcData0         0x00
+#define CUN_TProcData1         0x01
+#define CUN_TProcData2         0x02
+#define CUN_TProcData3         0x03
+#define CUN_TProcData4         0x04
+#define CUN_TProcData5         0x05
+#define CUN_TProcData6         0x06
+#define CUN_TProcData7         0x07
+#define CUN_TProcInst          0x08
+
+/* memory current unit numbers
+ * TProc data bus */
+#define CUN_DProcPA0           0x10
+#define CUN_DProcPA1           0x11
+#define CUN_DProcPrefetch      0x12
+#define CUN_CommandProc                0x13
+#define CUN_DProcData0         0x14    /* Dma prefetch reads. */
+#define CUN_DProcData1         0x15    /* Dma prefetch reads. */
+#define CUN_DProcData2         0x16    /* Dma prefetch reads. */
+#define CUN_DProcData3         0x17    /* Dma prefetch reads. */
+
+#define CUN_IProcLowPri                0x18
+#define CUN_IProcHighPri       0x19
+#define CUN_Spare0             0x1A
+#define CUN_Spare1             0x1B
+#define CUN_Spare2             0x1C
+#define CUN_ThreadQueue                0x1D
+#define CUN_EventProc0         0x1e
+#define CUN_EventProc1         0x1f
+
+#define CUN_Entries            0x20
+
+typedef struct E4_Registers
+{
+   E4_CacheTags                Tags;                           /* 4k bytes  c000 -> cfff */
+   E4_DataBusMap       Regs;                           /* 4k bytes  d000 -> dfff */
+   E4_User_Regs                uRegs;                          /* 8k bytes  e000 -> ffff */
+} E4_Registers;
+
+#define I2cCntl_I2cPortWrite           (0 << 0)
+#define I2cCntl_I2cPortRead            (1 << 0)
+#define I2cCntl_I2cPortGenStopBit      (1 << 1)
+#define I2cCntl_I2cPortGenRestartBit   (1 << 2)
+#define I2cCntl_I2cPortAccFailed       (1 << 3)
+#define I2cCntl_I2cStopped             (1 << 4)
+#define I2cCntl_I2cWakeupFailed                (1 << 5)
+#define I2cCntl_I2cFastMode            (1 << 6)
+#define I2cCntl_I2cPortBusy            (1 << 7)
+
+#define I2cCntl_LedI2cRegBase_Mask     0x7f
+#define I2cCntl_I2cUpdatingLedReg      (1 << 7)
+
+#define I2cCntl_InvertLedValues                (1 << 0)                /* read/write */
+#define I2cCntl_LedRegWriteFailed      (1 << 1)                /* read only */
+#define I2cCntl_EEPromLoadFailed       (1 << 2)                /* read only */
+#define I2cCntl_InhibitI2CRom          (1 << 3)                /* read only */
+#define I2cCntl_BadRomCrc              (1 << 4)                /* read only */
+#define I2cCntl_MapInI2cConfigData     (1 << 5)                /* read/write */
+#define I2cCntl_SampleNewLedValues     (1 << 6)                /* read/write */
+#define I2cCntl_ClearLinkError         (1 << 7)                /* write only */
+
+typedef struct E4_I2C
+{
+   volatile E4_uint8    I2cWrData;
+   volatile E4_uint8    I2cRdData;
+   volatile E4_uint8    I2cPortControl;
+   volatile E4_uint8   I2cLedBase;
+   volatile E4_uint8    I2cStatus;
+   volatile E4_uint8    I2cLedsValue;
+   volatile E4_uint16  I2cPad;
+ 
+   E4_uint8            pad[256 - sizeof(E4_uint64)];
+
+   E4_uint8            UnchangedElan4ConfigRegs[256];
+   E4_uint8            I2cRomConfigShadowValues[256];
+   E4_uint8            ChangedElan4ConfigRegs[256];
+} E4_I2C;
+
+typedef struct _E4_ContextControlBlock 
+{
+    E4_uint32 Filter;                  /* Use a Network context to index for this value */
+    E4_uint32 VirtualProcessTable;     /* Use a local context to index for this value */
+} E4_ContextControlBlock;
+
+/*
+ * Filter
+ *   [13:0]    Context
+ *   [14]      DiscardAll
+ *   [15]      AckAll
+ *   [16]      HighPri
+ *   [17]      CountStats
+ *   [31:18]   Unused
+ */
+#define E4_FILTER_STATS                (1 << 17)
+#define E4_FILTER_HIGH_PRI     (1 << 16)
+#define E4_FILTER_ACKOK_ALL    (1 << 15)
+#define E4_FILTER_DISCARD_ALL  (1 << 14)
+#define E4_FILTER_CONTEXT_MASK (0x3FFF)
+
+/*
+ * VirtualProcessTable
+ *   [8:0]     Unused  
+ *   [12:9]    Size       num vp entries = 512 << Size
+ *   [30:13]   Pointer
+ *   [31]      Valid
+ */
+#define E4_VPT_MIN_ENTRIES      512
+#define E4_VPT_VALID           ((unsigned)1 << 31)
+#define E4_VPT_PTR_SHIFT       0
+#define E4_VPT_SIZE_SHIFT      9
+#define E4_VPT_SIZE_MASK        0xf
+#define E4_VPT_NUM_VP(vpt_val)  (E4_VPT_MIN_ENTRIES << (((vpt_val) >> E4_VPT_SIZE_SHIFT) & E4_VPT_SIZE_MASK))
+#define E4_VPT_VALUE(ptr,size) (((ptr) << E4_VPT_PTR_SHIFT) | ((size) << E4_VPT_SIZE_SHIFT))
+
+
+/* Virtual Process Table */
+typedef struct _E4_VirtualProcessEntry
+{
+    E4_uint64  Values[2];
+} E4_VirtualProcessEntry;
+
+/*
+ * Entries have the following format - rtX is a packed route 
+ *
+ * |rt11|rt10|rt9 |rt8 |rt7 |rt6 |rt5 |rt4 |rt3 |rt2 |rt2 |rt0 |PAAADD       RRRRRR|
+ * |output context     |rt23|rt22|rt21|rt20|rt19|rt18|rt17|rt16|rt15|rt14|rt13|rt12|
+ */
+
+#define ROUTE_CTXT_SHIFT       48
+#define ROUTE_CTXT_MASK                (~((1ull << ROUTE_CTXT_SHIFT)-1))
+#define ROUTE_CTXT_VALUE(ctx)  (((E4_uint64) ctx) << ROUTE_CTXT_SHIFT)
+
+#define ROUTE_PACKED_OFFSET    16
+#define ROUTE_NUM_PACKED       24
+
+/* defines for first flit of a route */
+#define FIRST_TIMEOUT(Val)     ((Val) << 14)                   /* [15:14]  */
+#define FIRST_SYSTEM_PACKET     (1 << 13)                       /* [13]     */
+#define FIRST_FLOOD_PACKET      (1 << 12)                       /* [12]     */
+#define FIRST_HIGH_PRI         (1 << 11)                       /* [11]    */
+#define FIRST_AGE(Val)         ((Val) << 7)                    /* [10:7] */
+#define FIRST_OPTIONS_MASK     (0xFF80)
+
+/* [6:0] unpacked 1st route value */
+#define FIRST_INVALID          (0)
+#define FIRST_ROUTE(Val)       (0x08 | (Val))
+#define FIRST_ADAPTIVE         (0x30)
+#define FIRST_BCAST_TREE       (0x20)
+#define FIRST_MYLINK           (0x10)
+#define FIRST_BCAST(Top, Bot)  (0x40 | ((Top) << 3) | (Bot))
+
+/* defines for 3 bit packed entries for subsequent flits */
+#define PACKED_INVALID         (0)
+#define PACKED_ROUTE(Val)      (8 | (Val))
+#define PACKED_ADAPTIVE                (3)
+#define PACKED_BCAST_TREE      (2)
+#define PACKED_MYLINK          (1)
+#define PACKED_BCAST0(Top,Bot) (4 | (Bot & 3))
+#define PACKED_BCAST1(Top,Bot) ((Top << 1) | (Bot >> 2))
+
+#endif /* _ASM */
+/* The MMU root context pointer has a mask to bounds check 
+ * it - this is computed as follows.
+ */
+#define E4_CONTEXT_MASK(num)   (((num) >= 0x2000) ? 0x00 :     \
+                                ((num) >= 0x1000) ? 0x80 :     \
+                                ((num) >= 0x0800) ? 0xc0 :     \
+                                ((num) >= 0x0400) ? 0xe0 :     \
+                                ((num) >= 0x0200) ? 0xf0 :     \
+                                ((num) >= 0x0100) ? 0xf8 :     \
+                                ((num) >= 0x0080) ? 0xfc :     \
+                                ((num) >= 0x0040) ? 0xfe : 0xff)
+/*
+ * This generates the size field for a virtual process table.
+ * Size defined as 2^n no of 8K pages.
+ * Single cycle route fetches are possible if the minimum vproc table size is 8k.
+ */
+#define E4_GEN_VPT_SIZE(Size)  (((Size) & E4_VPT_SIZE_MASK) << E4_VPT_SIZE_SHIFT)
+
+#define COMMAND_RUN_QUEUE_BITS         (13 + 2) /* 8K entries of 4 bytes. This is fixed in hardware. */
+#define COMMAND_DESCS_SPACE_BITS       (13 + 5) /* 8K entries of 32 bytes. This is fixed in hardware. */
+#define COMMAND_INSERTER_CACHE_ENTRIES 16
+
+#define COM_TEST_PORT_ADDR_MASK                0xfULL
+#define COM_TEST_PORT_ADDR_SH          0
+
+/*
+ * The flush register is accessed through the CommandControl register.
+ * The address is naturally alligned. It also positions the command descriptors in memory.
+ * When no command queues need flushing it should be or with COM_FLUSH_INVALID. This sets
+ * it to the top command queue descriptor. This cannot be accessed from the PCI.
+ */
+#define COM_ENABLE_DEQUEUE             (1 << 4)
+#define COM_FLUSH_DESCRIPTOR_MASK      0x7fffffe0ULL
+#define COM_FLUSH_INVALID              0x0003ffe0ULL
+
+
+/*
+ * Elan4 BAR1 is split up as follows :
+ *
+ * RevA
+ *     0x3f00000 EBUS other
+ *     0x3e00000 EBUS ROM
+ *     0x3dfc000 registers
+ *     0x0000000 command ports
+ *
+ * RevB
+ *     0x3ffc000 registers
+ *     0x3ff8000 padding
+ *     0x3ff6000 i2c registers
+ *     0x0000000 command ports
+ */
+#define ELAN4_BAR1_SIZE                        (1 << 26)       /* 64M */
+#define ELAN4_REG_SIZE                 (1 << 14)       /* 16K */
+
+#define ELAN4_REVA_EBUS_SIZE           (1 << 21)       /* 2M */
+#define ELAN4_REVA_EBUS_OFFSET         (ELAN4_BAR1_SIZE - ELAN4_REVA_EBUS_SIZE)
+#define ELAN4_REVA_REG_OFFSET          (ELAN4_REVA_EBUS_OFFSET - ELAN4_REG_SIZE)
+#define ELAN4_REVA_NUM_COMMAND_QUEUES  (ELAN4_REVA_REG_OFFSET >> 13)
+
+#define ELAN4_REVA_EBUS_ROM_SIZE       (1 << 20)       /* 1M */
+#define ELAN4_REVA_EBUS_ROM_OFFSET     0
+
+#define ELAN4_REVB_I2C_PADDING         (1 << 14)       /* 16K */
+#define ELAN4_REVB_I2C_SIZE            (1 << 13)       /* 8k */
+#define ELAN4_REVB_REG_OFFSET          (ELAN4_BAR1_SIZE - ELAN4_REG_SIZE)
+#define ELAN4_REVB_I2C_OFFSET          (ELAN4_REVB_REG_OFFSET - ELAN4_REVB_I2C_PADDING - ELAN4_REVB_I2C_SIZE)
+#define ELAN4_REVB_NUM_COMMAND_QUEUES  (ELAN4_REVB_I2C_OFFSET >> 13)
+
+#endif /* notdef _ELAN4_REGISTERS_H */
Index: linux-2.4.21/include/elan4/sdram.h
===================================================================
--- linux-2.4.21.orig/include/elan4/sdram.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/sdram.h  2005-06-01 23:12:54.743417216 -0400
@@ -0,0 +1,41 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_SDRAM_H
+#define __ELAN4_SDRAM_H
+
+#ident "$Id: sdram.h,v 1.8 2003/09/24 13:55:55 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/sdram.h,v $*/
+
+/* Include header file generated by sdram configuration program */
+#include <elan4/xsdram.h> 
+
+/* SDRAM bank shift definitions */
+#define SDRAM_0_CS_SHIFT       25
+#define SDRAM_1_CS_SHIFT       27
+#define SDRAM_2_CS_SHIFT       28
+#define SDRAM_3_CS_SHIFT       29
+
+#define SDRAM_BANK_SHIFT(cfg) \
+       (((cfg >> SDRAM_RamSize_SH) & 3) == 0 ? SDRAM_0_CS_SHIFT : \
+        ((cfg >> SDRAM_RamSize_SH) & 3) == 1 ? SDRAM_1_CS_SHIFT : \
+        ((cfg >> SDRAM_RamSize_SH) & 3) == 2 ? SDRAM_2_CS_SHIFT : SDRAM_3_CS_SHIFT)
+
+#define SDRAM_BANK_SIZE(cfg)           (1ULL << SDRAM_BANK_SHIFT(cfg))
+#define SDRAM_BANK_OFFSET(cfg,bank)    ((unsigned long long)(bank) << SDRAM_BANK_SHIFT(cfg))
+#define SDRAM_NUM_BANKS(cfg)           (4)
+#define SDRAM_MAX_BANKS                        4
+
+/* When the elan access sdram it passes eaddr[12] as sdramaddr[12] when
+ * running with a 4k page size, however PCI accesses pass paddr[12], so
+ * we must ensure that sdram pages are allocated such that eaddr[12] is the
+ * same as paddr[12] - the easiest way is to allocate sdram in 8k chunks and
+ * ensure that maddr[12] == eaddr[12] == pgoff[0] */
+#define SDRAM_MIN_PAGE_SIZE            (8192)
+
+#endif /* __ELAN4_SDRAM_H */
Index: linux-2.4.21/include/elan4/stats.h
===================================================================
--- linux-2.4.21.orig/include/elan4/stats.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/stats.h  2005-06-01 23:12:54.743417216 -0400
@@ -0,0 +1,83 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: stats.h,v 1.10.12.1 2004/10/06 11:09:12 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/stats.h,v $*/
+
+#ifndef __ELAN4_STATS_H
+#define __ELAN4_STATS_H
+
+#define ELAN4_DEV_STATS_BUCKETS                8
+
+
+typedef struct elan4_dev_stats
+{
+    unsigned long      s_interrupts;
+    
+    unsigned long       s_mainints[ELAN4_DEV_STATS_BUCKETS];
+    unsigned long      s_mainint_punts;
+    unsigned long      s_mainint_rescheds;
+
+    unsigned long       s_haltints;
+
+    unsigned long      s_cproc_traps;
+    unsigned long      s_dproc_traps;
+    unsigned long      s_eproc_traps;
+    unsigned long      s_iproc_traps;
+    unsigned long      s_tproc_traps;
+
+    unsigned long       s_cproc_trap_types[0x10];
+    unsigned long       s_dproc_trap_types[6];
+    unsigned long       s_eproc_trap_types[4];
+    unsigned long       s_iproc_trap_types[0xa];
+    unsigned long       s_tproc_trap_types[7];
+
+    unsigned long       s_correctable_errors;
+    unsigned long       s_multiple_errors;
+    
+    unsigned long       s_link_errors;
+    unsigned long       s_lock_errors;
+    unsigned long       s_deskew_errors;
+    unsigned long       s_phase_errors;
+    unsigned long      s_data_errors;
+    unsigned long      s_fifo_overflow0;
+    unsigned long      s_fifo_overflow1;
+    unsigned long       s_mod45changed;
+    unsigned long       s_pack_not_seen;
+    unsigned long       s_linkport_keyfail;
+
+    unsigned long      s_eop_reset;
+    unsigned long       s_bad_length;
+    unsigned long       s_crc_bad;
+    unsigned long       s_crc_error;
+
+    unsigned long      s_cproc_timeout;
+    unsigned long      s_dproc_timeout;
+
+    unsigned long      s_sdram_bytes_free;
+} ELAN4_DEV_STATS;
+
+#define MainIntBuckets         ((int[ELAN4_DEV_STATS_BUCKETS-1]) {1, 2, 3, 4, 8, 16, 32})
+
+#define BumpDevStat(dev,stat)  ((dev)->dev_stats.stat++)
+#define BucketDevStat(dev,stat,n,bucket)       ((n) <= (bucket)[0] ? (dev)->dev_stats.stat[0]++ : \
+                                                (n) <= (bucket)[1] ? (dev)->dev_stats.stat[1]++ : \
+                                                (n) <= (bucket)[2] ? (dev)->dev_stats.stat[2]++ : \
+                                                (n) <= (bucket)[3] ? (dev)->dev_stats.stat[3]++ : \
+                                                (n) <= (bucket)[4] ? (dev)->dev_stats.stat[4]++ : \
+                                                (n) <= (bucket)[5] ? (dev)->dev_stats.stat[5]++ : \
+                                                (n) <= (bucket)[6] ? (dev)->dev_stats.stat[6]++ : \
+                                                                     (dev)->dev_stats.stat[7]++)
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /*__ELAN4_STATS_H */
Index: linux-2.4.21/include/elan4/tprintf.h
===================================================================
--- linux-2.4.21.orig/include/elan4/tprintf.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/tprintf.h        2005-06-01 23:12:54.743417216 -0400
@@ -0,0 +1,24 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_TPRINTF_H
+#define __ELAN4_TPRINTF_H
+
+#ident "$Id: tprintf.h,v 1.6 2003/09/04 12:39:17 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/tprintf.h,v $*/
+
+
+#ifdef _ASM
+#define TPRINTF0(string)           add %r0, __LINE__, %r0
+#define TPRINTF1(string,reg)       add reg, __LINE__, %r0
+#else
+#define TPRINTF0(string)           asm volatile ("add %%r0, %0, %%r0" : : "i" (__LINE__))
+#define TPRINTF1(string, value)            asm volatile ("add %0,   %1, %%r0" : : "r" (value), "i" (__LINE__))
+#endif /* _ASM */
+
+#endif /* __ELAN4_TPRINTF_H */
Index: linux-2.4.21/include/elan4/trap.h
===================================================================
--- linux-2.4.21.orig/include/elan4/trap.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/trap.h   2005-06-01 23:12:54.743417216 -0400
@@ -0,0 +1,95 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: trap.h,v 1.10 2003/10/07 12:11:10 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/trap.h,v $*/
+
+#ifndef __ELAN4_TRAP_H
+#define __ELAN4_TRAP_H
+
+/*
+ * If the EProc Faults whilst performing an action (e.g. Read/Write on the data src or dest Addr)
+ *  the Eproc increments the Addr(s) by a block size (64 bytes):
+ *  1: Fault on Read: 
+ *                     Src EventAddr = Read Addr + block
+ *  2: Fault on Write:
+ *                     Src EventAddr = Read Addr + block
+ *                     Dst EventAddr = Read Addr + block
+ *                     Size          = Size - block ndwords
+ *  We must rewind the addr correctly to completely the transfer successfully
+ */
+#define EVENT_COPY_NDWORDS     0x8
+#define EVENT_COPY_BLOCK_SIZE  0x40
+
+typedef struct elan4_eproc_trap
+{
+    E4_uint64          tr_status;
+    E4_FaultSave       tr_faultarea;
+    E4_Event           tr_event;
+    E4_Addr            tr_eventaddr;
+} ELAN4_EPROC_TRAP;
+
+typedef struct elan4_cproc_trap
+{
+    E4_uint64          tr_status;                                      /* cproc status register */
+    E4_uint64          tr_command;                                     /* cproc command */
+    E4_CommandQueueDesc tr_qdesc;                                      /* copy of command queue descriptor */
+    E4_FaultSave       tr_faultarea;                                   /* fault area for mmu traps */
+    ELAN4_EPROC_TRAP   tr_eventtrap;                                   /* associated event trap (waitevent) */
+} ELAN4_CPROC_TRAP;
+
+typedef struct elan4_dproc_trap
+{
+    E4_DMA             tr_desc;
+    E4_FaultSave       tr_packAssemFault;
+    E4_FaultSave       tr_prefetchFault;
+    E4_uint64          tr_status;
+} ELAN4_DPROC_TRAP;
+
+typedef struct elan4_tproc_trap
+{
+    E4_uint64          tr_regs[64];
+    E4_FaultSave       tr_dataFault;
+    E4_FaultSave       tr_instFault;
+    E4_uint64          tr_status;
+    E4_uint64          tr_state;
+    E4_Addr            tr_pc;
+    E4_Addr            tr_npc;
+    E4_uint64          tr_dirty;
+    E4_uint64          tr_bad;
+} ELAN4_TPROC_TRAP;
+
+typedef struct elan4_iproc_trap
+{
+    E4_uint32            tr_numTransactions;
+    E4_uint32            tr_flags;
+    E4_uint32            tr_trappedTrans;
+    E4_uint32            tr_waitForEopTrans;
+    E4_uint32            tr_identifyTrans;
+    E4_uint32            tr_pad;
+
+    E4_FaultSave          tr_faultarea;
+    E4_IprocTrapHeader    tr_transactions[MAX_TRAPPED_TRANS];
+    E4_IprocTrapData      tr_dataBuffers[MAX_TRAPPED_TRANS];
+} ELAN4_IPROC_TRAP;
+
+#define TR_FLAG_ACK_SENT       (1 << 0)
+#define TR_FLAG_EOP_ERROR      (1 << 1)
+#define TR_FLAG_BAD_TRANS      (1 << 2)
+#define TR_FLAG_DMA_PACKET     (1 << 3)
+#define TR_FLAG_EOP_BAD                (1 << 4)
+#define TR_FLAG_TOOMANY_TRANS  (1 << 5)
+
+#define TR_TRANS_INVALID       (0xffffffff)
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_TRAP_H */
Index: linux-2.4.21/include/elan4/trtype.h
===================================================================
--- linux-2.4.21.orig/include/elan4/trtype.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/trtype.h 2005-06-01 23:12:54.744417064 -0400
@@ -0,0 +1,112 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _ELAN4_TRTYPE_H
+#define _ELAN4_TRTYPE_H
+
+#ident "$Id: trtype.h,v 1.20 2004/02/06 10:38:21 mike Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/trtype.h,v $*/
+
+/*<15:11> Size field is used to give the number of additional 64 bit data values.
+         A value from 0 to 16 inclusive is valid. */
+
+#include <elan4/types.h>
+
+#define TR_SIZE_SHIFT          (11)
+#define TR_SIZE_MASK           (0x1f << TR_SIZE_SHIFT)
+#define SET_TR_SIZE(Size)      (((Size) << TR_SIZE_SHIFT) & TR_SIZE_MASK)
+
+/* <10:9> Last Transaction and AckNow bits, marks the last transaction and
+          enables a PACK_OK to be sent. */
+#define TR_LAST_AND_SEND_ACK   (3 << 9)
+
+
+/* <8>  Only valid on the last transaction. Delays execution until an EOP_GOOD is received.
+ *      Any other EOP type will abort execution of this transaction. */
+#define TR_WAIT_FOR_EOP                (1 << 8)
+
+/*
+ * Data type. This is used by transactions of variable data type. It controls any endian
+ * converion required if the destiantion host processor has a big endian memory format.
+ */
+/*     WriteBlock      <8:7>   Data type
+                       <6:0>   Part write size */
+#define TR_DATATYPE_SHIFT      (6)
+#define TR_DATATYPE_MASK       ((1 << 2) - 1)
+
+#define TR_DATATYPE_BYTE       E4_DATATYPE_BYTE        
+#define TR_DATATYPE_SHORT      E4_DATATYPE_SHORT
+#define TR_DATATYPE_WORD       E4_DATATYPE_WORD        
+#define TR_DATATYPE_DWORD      E4_DATATYPE_DWORD
+
+/* <5:0> Transaction Type
+ *       For Writeblock <5:3> 000 => Write, 0001 => Read
+ *                      <2:0> End Byte Addr */
+#define TR_OPCODE_MASK         0x3F
+#define TR_BLOCK_OPCODE_MASK   0x38
+
+#define TR_WRITEBLOCK          0x0
+#define TR_ENDBYTE_MASK                0x7
+#define TR_WRITE(Size, EndByte, DataType)                                              \
+                       (0x0 | SET_TR_SIZE(Size) | ((EndByte) & TR_ENDBYTE_MASK) |      \
+                        (((DataType) & TR_DATATYPE_MASK) << TR_DATATYPE_SHIFT))
+
+#define TR_NOP_TRANS           (0x10 | SET_TR_SIZE(0))
+#define TR_SETEVENT            0x10
+#define TR_SETEVENT_NOIDENT    (TR_SETEVENT | SET_TR_SIZE(0) | TR_LAST_AND_SEND_ACK)
+#define TR_SETEVENT_IDENTIFY   (TR_SETEVENT | SET_TR_SIZE(1) | TR_LAST_AND_SEND_ACK)
+#define TR_REMOTEDMA           (0x11 | SET_TR_SIZE(7) | TR_LAST_AND_SEND_ACK)
+#define TR_SENDDISCARD         (0x12 | SET_TR_SIZE(0))
+
+/*
+ * Conditional transactions that might return PAckTestFail.
+ * All will allow further exection of the packet if ([Address] operator DataValue) is true.
+ * e.g. for TR_GTE further execution if ([Address] >= DataValue) is true.
+ * These should be used where a definite TRUE/FALSE answer is required.
+ */
+#define TR_GTE                 (0x14 | SET_TR_SIZE(1))
+#define TR_LT                  (0x15 | SET_TR_SIZE(1))
+#define TR_EQ                  (0x16 | SET_TR_SIZE(1))
+#define TR_NEQ                 (0x17 | SET_TR_SIZE(1))
+
+/*
+ * Conditional transactions that might return PAckDiscard.
+ * All will allow further exection of the packet if ([Address] operator DataValue) is true.
+ * e.g. for TR_GTE further execution if ([Address] >= DataValue) is true.
+ * These should be used where eventually a TRUE answer is expected but the node might not be ready yet.
+ * These can be mixed with the normal conditionals to allow a single packet to test for readyness and
+ * a TRUE/FALSE answer.
+ */
+#define TR_GTE_DISCARD         (0x34 | SET_TR_SIZE(1))
+#define TR_LT_DISCARD          (0x35 | SET_TR_SIZE(1))
+#define TR_EQ_DISCARD          (0x36 | SET_TR_SIZE(1))
+#define TR_NEQ_DISCARD         (0x37 | SET_TR_SIZE(1))
+
+#define TR_TRACEROUTE_TRANS    0x18
+#define TR_TRACEROUTE(Size)    (TR_TRACEROUTE_TRANS | (TR_DATATYPE_WORD << TR_DATATYPE_SHIFT) |SET_TR_SIZE(Size))
+#define TR_IDENTIFY            (0x19 | SET_TR_SIZE(0))
+
+#define TR_ADDWORD             (0x1c | SET_TR_SIZE(2) | TR_LAST_AND_SEND_ACK)
+#define TR_INPUT_Q_COMMIT      (0x1d | SET_TR_SIZE(1) | TR_LAST_AND_SEND_ACK)
+#define TR_TESTANDWRITE        (0x1e | SET_TR_SIZE(3) | TR_LAST_AND_SEND_ACK)
+#define TR_INPUT_Q_GETINDEX    (0x1f | SET_TR_SIZE(0))
+
+
+
+/* TraceRoute formate */
+#define TR_TRACEROUTE0_CHANID(val)             ((val) & 1)                     /* 0     Chan Id */
+#define TR_TRACEROUTE0_LINKID(val)             (((val) >> 1) & 7)              /* 1:3   Link Id */
+#define TR_TRACEROUTE0_REVID(val)              (((val) >> 4) & 7)              /* 4:6   Revision Id */
+#define TR_TRACEROUTE0_BCAST_PIN(val)          (((val) >> 7) & 1)              /* 7     Bcast Top Pin */
+#define TR_TRACEROUTE0_LNR(val)                        (((val) >> 8) & 0xFF)           /* 8:15  Global Link Not Ready */
+
+#define TR_TRACEROUTE1_ROUTES_SELECTED(val)    ((val & 0xFF))                  /* 0:7   Routes Selected */
+#define TR_TRACEROUTE1_BCAST_TOP(val)          (((val) >> 8) & 7)              /* 8:10  Broadcast Top */
+#define TR_TRACEROUTE1_BCAST_BOTTOM(val)       (((val) >> 12) & 7)             /* 12:14 Broadcast Bottom */
+
+#endif /* _ELAN4_TRANSACTIONTYPE_H */
Index: linux-2.4.21/include/elan4/types.h
===================================================================
--- linux-2.4.21.orig/include/elan4/types.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/types.h  2005-06-01 23:12:54.744417064 -0400
@@ -0,0 +1,69 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_TYPES_H
+#define __ELAN4_TYPES_H
+
+#ident "@(#)$Id: types.h,v 1.9 2003/09/04 12:39:17 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/types.h,v $*/
+
+#include <qsnet/config.h>
+/*
+ * "flip" values for correctly indexing into
+ * block data which was copied from the Elan
+ * using 64 bit accesses.
+ */
+#if defined(__LITTLE_ENDIAN__)
+#  define ByteEndianFlip  0
+#  define ShortEndianFlip 0
+#  define WordEndianFlip  0
+#else
+#  define ByteEndianFlip  7
+#  define ShortEndianFlip 3
+#  define WordEndianFlip  1
+#endif
+
+
+#ifndef _ASM
+
+typedef signed int        E4_int;
+typedef unsigned int              E4_uint;
+
+typedef signed char       E4_int8;
+typedef unsigned char     E4_uint8;
+
+typedef signed short      E4_int16;
+typedef unsigned short            E4_uint16;
+
+typedef signed int        E4_int32;
+typedef unsigned int              E4_uint32;
+
+#ifdef _LP64
+typedef signed long        E4_int64;
+typedef unsigned long      E4_uint64;
+#else
+typedef signed long long   E4_int64;
+typedef unsigned long long E4_uint64;
+#endif
+
+/* 64-bit Elan4 */
+typedef E4_uint64         E4_Addr;
+typedef E4_uint32         E4_LocPhysAddr;      /* Really 31 bits */
+
+#define OneK   (1024)
+#define EightK (8*OneK)
+
+#define E4_DATATYPE_BYTE       0
+#define E4_DATATYPE_SHORT      1
+#define E4_DATATYPE_WORD       2
+#define E4_DATATYPE_DWORD      3
+
+#endif /* _ASM */
+
+#endif /* __ELAN4_TYPES_H */
+
Index: linux-2.4.21/include/elan4/user.h
===================================================================
--- linux-2.4.21.orig/include/elan4/user.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/user.h   2005-06-01 23:12:54.745416912 -0400
@@ -0,0 +1,344 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: user.h,v 1.37.2.2 2004/11/18 17:54:17 duncant Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/user.h,v $*/
+
+#ifndef __ELAN4_USER_H
+#define __ELAN4_USER_H
+
+#include <elan/capability.h>
+#include <elan4/usertrap.h>
+#include <elan4/intcookie.h>
+
+typedef struct trap_queue
+{
+    unsigned   q_back;                 /* Next free space */
+    unsigned   q_front;                /* First object to remove */
+    unsigned   q_size;                 /* Size of queue */
+    unsigned   q_count;                /* Current number of entries */
+    unsigned   q_slop;                 /* FULL <=> (count+slop) == size */
+} RING_QUEUE;
+
+#define RING_QUEUE_INIT(q,num,slop)    ((q).q_size = (num), (q).q_slop = (slop), (q).q_front = (q).q_back = 0, (q).q_count = 0)
+#define RING_QUEUE_FULL(q)             ((q).q_count >= ((q).q_size - (q).q_slop))
+#define RING_QUEUE_REALLY_FULL(q)      ((q).q_count == (q).q_size)
+#define RING_QUEUE_EMPTY(q)            ((q).q_count == 0)
+#define RING_QUEUE_NEXT(q,indx)                ((indx) = (((indx)+1) % (q).q_size))
+#define RING_QUEUE_PREV(q,indx)                ((indx) = (((indx)+(q).q_size-1) % (q).q_size))
+#define RING_QUEUE_ADD(q)              (RING_QUEUE_NEXT(q ,(q).q_back),  (++(q).q_count) >= ((q).q_size - (q).q_slop))
+#define RING_QUEUE_REMOVE(q)           (RING_QUEUE_NEXT(q, (q).q_front), (--(q).q_count) == 0)
+#define RING_QUEUE_ADD_FRONT(q)                (RING_QUEUE_PREV(q, (q).q_front), (++(q).q_count) >= ((q).q_size - (q).q_slop))
+#define RING_QUEUE_ENTRY(qArea,indx)   (&(qArea)[(indx)])
+#define RING_QUEUE_FRONT(q,qArea)      RING_QUEUE_ENTRY(qArea, (q).q_front)
+#define RING_QUEUE_BACK(q,qArea)       RING_QUEUE_ENTRY(qArea, (q).q_back)
+#define RING_QUEUE_ITERATE(q,idx)      for (idx = (q).q_front; idx != (q).q_back; idx = (((idx) + 1) % (q).q_size))
+
+typedef struct user_rgn
+{
+    struct user_rgn *rgn_mnext;                                        /* Doubly linked list of regions */
+    struct user_rgn *rgn_mprev;                                        /*   sorted on main address */ 
+    virtaddr_t       rgn_mbase;                                        /* main address of base of region */
+
+    struct user_rgn *rgn_enext;                                        /* Doubly linked list of regions */
+    struct user_rgn *rgn_eprev;                                        /*   sorted on elan address */
+    E4_Addr         rgn_ebase;                                 /* elan address of base of region */
+
+    unsigned long    rgn_len;                                  /* length of region */
+    unsigned        rgn_perm;                                  /* elan access permission */
+} USER_RGN;
+
+typedef struct user_vpseg
+{ 
+    struct list_head   vps_link;
+
+    unsigned short     vps_process;                            /* virtual process number */
+    unsigned short     vps_entries;                            /*   and # virtual processes */
+
+    unsigned           vps_type;
+    union
+    {
+       struct {
+           ELAN_CAPABILITY        *cap;
+           E4_VirtualProcessEntry *routes;
+       } p2p;
+#define vps_p2p_cap    vps_u.p2p.cap
+#define vps_p2p_routes  vps_u.p2p.routes
+
+       struct {
+           unsigned short lowvp;
+           unsigned short highvp;
+       } bcast;
+#define vps_bcast_lowvp                vps_u.bcast.lowvp
+#define vps_bcast_highvp       vps_u.bcast.highvp
+    } vps_u;
+} USER_VPSEG;
+
+/* values for vps_type */
+#define USER_VPSEG_P2P         0
+#define USER_VPSEG_BCAST       1
+
+typedef struct user_cq
+{
+    struct list_head ucq_link;
+
+    ELAN4_CQ       *ucq_cq;                                    /* the real command queue */
+
+    unsigned char    ucq_state;                                        /* command queue state */
+    unsigned char    ucq_errored;                              /* command queue has errored */
+    unsigned char    ucq_flags;                                        /* flags */
+    ELAN4_CPROC_TRAP ucq_trap;                                 /* trap state */
+
+    atomic_t        ucq_ref;                                   /* # references to this cq (mmaps) */
+} USER_CQ;
+
+/* values for ucq_state */
+#define UCQ_RUNNING                     0                      /* command queue is running */
+#define UCQ_TRAPPED                     1                      /* command queue has trapped */
+#define UCQ_NEEDS_RESTART                2                     /* command queue has trapped, and needs restarting */
+#define UCQ_STOPPED                     3                      /* command queue has trapped, and delivered to user */
+
+/* values for ucq_flags */
+#define UCQ_SYSTEM             (1 << 0)
+#define UCQ_REORDER            (1 << 1)
+
+extern int num_fault_save;
+extern int min_fault_pages;
+extern int max_fault_pages;
+
+typedef struct fault_save
+{
+    struct fault_save           *next;
+    E4_Addr                      addr;
+    E4_uint32                    count;
+} FAULT_SAVE;
+
+typedef struct user_iproc_trap
+{
+    unsigned char     ut_state;
+    ELAN4_IPROC_TRAP  ut_trap;
+} USER_IPROC_TRAP;
+
+/* values for ut_state */
+#define UTS_IPROC_RUNNING                      0
+#define UTS_IPROC_TRAPPED                      1
+#define UTS_IPROC_RESOLVING                    2
+#define UTS_IPROC_EXECUTE_PACKET               3
+#define UTS_IPROC_EXECUTING                    4
+#define UTS_IPROC_NETWORK_ERROR                        5
+#define UTS_IPROC_STOPPED                      6
+
+typedef struct user_ctxt_entry
+{
+    struct list_head    cent_link;                                     /* entry chained on context */
+    ELAN_CAPABILITY    *cent_cap;                                      /* capability we attached with */
+} USER_CTXT_ENTRY;
+
+typedef struct user_ctxt
+{
+    ELAN4_CTXT         uctx_ctxt;                              /* is also an elan context */
+
+    spinlock_t        uctx_spinlock;                           /* spinlock for items used with interrupt handler */
+    kcondvar_t        uctx_wait;                               /* place to sleep (traphandler/swapout/swapin/neterr fixup) */
+
+    unsigned          uctx_status;                             /* status                               (uctx_spinlock) */
+
+    pid_t             uctx_trap_pid;                           /* pid to deliver signals to on trap */
+    int                       uctx_trap_signo;                         /* signal number to deliver */
+    unsigned          uctx_trap_state;                         /* state of trap handling code */
+    unsigned          uctx_trap_count;                         /* count of "thread" in user_trap_handler() */
+
+    unsigned          uctx_int_count;                          /* # interrupts since last zeroed */
+    unsigned long      uctx_int_start;                         /* tick when int_count last zeroed */
+    unsigned long      uctx_int_delay;                         /* # ticks to delay next wakeup */
+    struct timer_list  uctx_int_timer;                         /* and timer to use to delay signal */
+
+    struct timer_list  uctx_neterr_timer;                      /* network error timer */
+
+    struct list_head   uctx_vpseg_list;                                /* list of vp segments we've got */
+    kmutex_t           uctx_vpseg_lock;                                /*   and lock to protect it. */
+    ELAN4_ROUTE_TABLE *uctx_routetable;                                /* our virtual process table */
+    ELAN_POSITION      uctx_position;                          /* position in network */
+
+    struct list_head   uctx_cent_list;                                 /* list of attached network contexts */
+
+    USER_CQ          *uctx_ddcq;                               /* command queue for re-issueing traps */
+    E4_uint64         uctx_ddcq_insertcnt;                     /* # dwords inserted into command queue */
+    E4_uint64          uctx_ddcq_completed;                    /* last "completed" write was here */
+    int                       uctx_ddcq_intr;                          /* count of outstanding ddcq interrupts */
+
+    ELAN4_HALTOP       uctx_haltop;                            /* halt operation for flushing */
+    ELAN4_DMA_FLUSHOP  uctx_dma_flushop;                       /* flush operation for flushing dma runqueue */
+
+    INTCOOKIE_TABLE   *uctx_intcookie_table;                   /* table of interrupt cookies (shared with other uctxs for this task) */
+
+    kmutex_t          uctx_cqlock;                             /* lock for create/destory cqs */
+    struct list_head   uctx_cqlist;                            /* list of command queues               (uctx_cqlock,uctx_spinlock) */
+
+    ELAN4_DPROC_TRAP  *uctx_dprocTraps;                                /* queue of dproc traps to resolve/reissue */
+    RING_QUEUE        uctx_dprocTrapQ;
+
+    ELAN4_TPROC_TRAP  *uctx_tprocTraps;                                /* queue of tproc traps to resolve/reissue */
+    RING_QUEUE         uctx_tprocTrapQ;
+
+    ELAN4_EPROC_TRAP  *uctx_eprocTraps;                                /* queue of eproc traps to resolve */
+    RING_QUEUE        uctx_eprocTrapQ;
+
+    USER_IPROC_TRAP    uctx_iprocTrap[2];                      /* input trap state, 1 per virtual channel */
+
+    E4_DMA           *uctx_dmas;                               /* queue of dmas to restart */
+    RING_QUEUE         uctx_dmaQ;
+    
+    E4_ThreadRegs     *uctx_threads;                           /* queue of threads to restart */
+    RING_QUEUE         uctx_threadQ;
+
+    ELAN4_NETERR_MSG  *uctx_msgs;                              /* queue of neterr messages */
+    RING_QUEUE        uctx_msgQ;
+    kmutex_t          uctx_rgnmutex;                           /* lock for create/destroy regions */
+    spinlock_t        uctx_rgnlock;                            /* spinlock to protect linked lists */
+    USER_RGN         *uctx_mrgns;                              /* Doubly linked list of memory regions (uctx_rgnlock) */
+    USER_RGN         *uctx_mtail;                              /* Last memory region on list           (uctx_rgnlock) */
+    USER_RGN         *uctx_mrgnlast;                           /* Last region 'hit'                    (uctx_rgnlock) */
+
+    USER_RGN         *uctx_ergns;                              /* Doubly linked list of memory regions (uctx_rgnlock) */
+    USER_RGN         *uctx_etail;                              /* Last memory region on list           (uctx_rgnlock) */
+    USER_RGN         *uctx_ergnlast;                           /* Last region 'hit'                    (uctx_rgnlock) */
+
+    ELAN4_USER_PAGE   *uctx_upage;                             /* kernel page shared with user */
+    sdramaddr_t               uctx_trampoline;                         /* sdram page for tproc trampoline */
+
+    E4_Addr           uctx_upage_addr;                         /*   elan addr page mapped into */
+    E4_Addr           uctx_trestart_addr;                      /* address of thread restart code */
+    FAULT_SAVE         *uctx_faults;
+    FAULT_SAVE         *uctx_fault_list;
+    int                 uctx_num_fault_save;
+    spinlock_t          uctx_fault_lock;
+} USER_CTXT;
+
+/* bit values for uctx_status */
+#define UCTX_EXITING                           (1 << 0)                /* context is exiting. */
+#define UCTX_USER_FILTERING                    (1 << 1)                /* user requested context filter */
+#define UCTX_USER_STOPPED                      (1 << 2)                /* user requested stop */
+
+#define UCTX_SWAPPING                          (1 << 3)                /* context is swapping out */
+#define UCTX_SWAPPED                           (1 << 4)                /* context is swapped out */
+
+#define UCTX_STOPPING                          (1 << 5)                /* stopping elan from running this context */
+#define UCTX_STOPPED                           (1 << 6)                /* elan no longer running this context */
+
+#define UCTX_EPROC_QUEUE_FULL                  (1 << 7)                /* reasons for stopping running */
+#define UCTX_DPROC_QUEUE_FULL                  (1 << 8)
+#define UCTX_TPROC_QUEUE_FULL                  (1 << 9)
+#define UCTX_IPROC_CH0_TRAPPED                 (1 << 10)
+#define UCTX_IPROC_CH1_TRAPPED                 (1 << 11)
+
+#define UCTX_NETERR_TIMER                      (1 << 12)
+#define UCTX_NETERR_FIXUP                      (1 << 13)
+
+#define UCTX_EPROC_QUEUE_OVERFLOW              (1 << 14)
+#define UCTX_DPROC_QUEUE_OVERFLOW              (1 << 15)
+#define UCTX_TPROC_QUEUE_OVERFLOW              (1 << 16)
+
+#define UCTX_EPROC_QUEUE_ERROR                 (1 << 17)
+#define UCTX_DPROC_QUEUE_ERROR                 (1 << 18)
+#define UCTX_TPROC_QUEUE_ERROR                 (1 << 19)
+
+#define UCTX_STOPPED_REASONS                   (UCTX_EPROC_QUEUE_FULL | UCTX_DPROC_QUEUE_FULL | UCTX_TPROC_QUEUE_FULL)
+#define UCTX_SWAPPED_REASONS                   (UCTX_EXITING | UCTX_USER_STOPPED | UCTX_NETERR_FIXUP)
+#define UCTX_NACKING_REASONS                   (UCTX_USER_FILTERING | UCTX_IPROC_CH0_TRAPPED | UCTX_IPROC_CH1_TRAPPED)
+
+#define UCTX_OVERFLOW_REASONS                  (UCTX_EPROC_QUEUE_OVERFLOW | UCTX_DPROC_QUEUE_OVERFLOW | UCTX_TPROC_QUEUE_OVERFLOW)
+#define UCTX_ERROR_REASONS                     (UCTX_EPROC_QUEUE_ERROR | UCTX_DPROC_QUEUE_ERROR | UCTX_TPROC_QUEUE_ERROR)
+
+#define UCTX_RUNNABLE(uctx)                    (((uctx)->uctx_status & (UCTX_SWAPPED_REASONS | UCTX_STOPPED_REASONS)) == 0)
+#define UCTX_NACKING(uctx)                     (((uctx)->uctx_status & (UCTX_SWAPPED_REASONS | UCTX_STOPPED_REASONS | UCTX_NACKING_REASONS)) != 0)
+
+/* values for uctx_trap_signalled */
+#define UCTX_TRAP_IDLE                         0
+#define UCTX_TRAP_SLEEPING                     1
+#define UCTX_TRAP_SIGNALLED                    2
+#define UCTX_TRAP_ACTIVE                       3
+
+extern int        user_p2p_route_options;
+extern int        user_bcast_route_options;
+extern int       user_dproc_retry_count;
+extern int       user_cproc_retry_count;
+
+extern USER_CTXT *user_alloc (ELAN4_DEV *dev);
+extern void       user_free (USER_CTXT *uctx);
+extern void       user_swapout (USER_CTXT *uctx, unsigned reason);
+extern void       user_swapin (USER_CTXT *uctx, unsigned reason);
+extern int        user_attach (USER_CTXT *uctx, ELAN_CAPABILITY *cap);
+extern void       user_detach (USER_CTXT *uctx, ELAN_CAPABILITY *cap);
+extern void       user_block_inputter (USER_CTXT *uctx, unsigned blocked);
+extern int        user_alloc_trap_queues (USER_CTXT *uctx, unsigned ndproc_traps, unsigned neproc_traps, 
+                                         unsigned ntproc_traps, unsigned nthreads, unsigned ndmas);
+
+extern int        user_add_p2pvp (USER_CTXT *uctx, unsigned process, ELAN_CAPABILITY *cap);
+extern int        user_add_bcastvp (USER_CTXT *uctx, unsigned process, unsigned lowvp, unsigned highvp);
+extern int        user_removevp (USER_CTXT *uctx, unsigned process);
+
+extern int        user_set_route (USER_CTXT *uctx, unsigned process, E4_VirtualProcessEntry *route);
+extern int        user_reset_route (USER_CTXT *uctx, unsigned process);
+extern int        user_get_route (USER_CTXT *uctx, unsigned process, E4_VirtualProcessEntry *route);
+extern int        user_check_route (USER_CTXT *uctx, unsigned process, E4_VirtualProcessEntry *route, unsigned *error);
+extern int       user_send_neterr_msg (USER_CTXT *uctx, unsigned int vp, unsigned int nctx, unsigned int retries, ELAN4_NETERR_MSG *msg);
+extern int        user_neterr_sten (USER_CTXT *uctx, unsigned int vp, E4_uint64 cookie, int waitforeop);
+extern int        user_neterr_dma (USER_CTXT *uctx, unsigned int vp, E4_uint64 cookie, int waitforeop);
+
+extern int        user_resume_eproc_trap (USER_CTXT *uctx, E4_Addr addr);
+extern int        user_resume_cproc_trap (USER_CTXT *uctx, unsigned indx);
+extern int        user_resume_dproc_trap (USER_CTXT *uctx, E4_DMA *dma);
+extern int        user_resume_tproc_trap (USER_CTXT *uctx, E4_ThreadRegs *regs);
+extern int        user_resume_iproc_trap (USER_CTXT *uctx, unsigned channel, unsigned trans,
+                                         E4_IprocTrapHeader *hdrp, E4_IprocTrapData *datap);
+
+extern int        user_trap_handler (USER_CTXT *uctx, ELAN4_USER_TRAP *utrapp, int nticks);
+extern USER_CQ   *user_findcq (USER_CTXT *uctx, unsigned num);
+extern USER_CQ   *user_alloccq (USER_CTXT *uctx, unsigned size, unsigned perm, unsigned flags);
+extern void       user_freecq (USER_CTXT *uctx, USER_CQ *cq);
+extern void       user_dropcq (USER_CTXT *uctx, USER_CQ *cq);
+
+/* user_osdep.c */
+extern int        user_load_range (USER_CTXT *uctx, E4_Addr addr, unsigned long nbytes, E4_uint32 fsr);
+extern void       user_update_main (USER_CTXT *uctx, struct mm_struct *mm, unsigned long start, unsigned long len);
+extern void       user_unload_main (USER_CTXT *uctx, unsigned long start, unsigned long len);
+
+
+/* regions.c */
+extern USER_RGN  *user_findrgn_elan (USER_CTXT *uctx, E4_Addr addr, int tail);
+extern USER_RGN  *user_findrgn_main (USER_CTXT *uctx, virtaddr_t addr, int tail);
+extern USER_RGN  *user_rgnat_elan (USER_CTXT *uctx, E4_Addr addr);
+extern USER_RGN  *user_rgnat_main (USER_CTXT *uctx, virtaddr_t addr);
+extern int        user_setperm (USER_CTXT *uctx, virtaddr_t maddr, E4_Addr eaddr, unsigned long len, unsigned perm);
+extern void       user_clrperm (USER_CTXT *uctx, E4_Addr addr, unsigned long len);
+extern int        user_checkperm (USER_CTXT *uctx, E4_Addr raddr, unsigned long rsize, unsigned access);
+extern virtaddr_t user_elan2main (USER_CTXT *uctx, E4_Addr addr);
+extern E4_Addr    user_main2elan (USER_CTXT *uctx, virtaddr_t addr);
+extern void       user_preload_main (USER_CTXT *uctx, virtaddr_t addr, unsigned long len);
+extern void       user_freergns (USER_CTXT *uctx);
+
+/* user_ddcq.c */
+extern int        user_ddcq_check (USER_CTXT *uctx, unsigned num);
+extern int        user_ddcq_flush (USER_CTXT *uctx);
+extern void       user_ddcq_intr (USER_CTXT *uctx);
+extern void       user_ddcq_write_dword (USER_CTXT *uctx, E4_Addr addr, E4_uint64 value);
+extern void       user_ddcq_interrupt (USER_CTXT *uctx, E4_uint64 cookie);
+extern void       user_ddcq_run_dma (USER_CTXT *uctx, E4_DMA *dma);
+extern void       user_ddcq_run_thread (USER_CTXT *uctx, E4_ThreadRegs *regs);
+extern void       user_ddcq_setevent (USER_CTXT *uctx, E4_Addr addr);
+extern void       user_ddcq_seteventn (USER_CTXT *uctx, E4_Addr addr, E4_uint32 count);
+extern void       user_ddcq_waitevent (USER_CTXT *uctx, E4_Addr addr, E4_uint64 CountAndType, E4_uint64 Param0, E4_uint64 Param1);
+
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_USER_H */
Index: linux-2.4.21/include/elan4/userregs.h
===================================================================
--- linux-2.4.21.orig/include/elan4/userregs.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/userregs.h       2005-06-01 23:12:54.746416760 -0400
@@ -0,0 +1,383 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_USERREGS_H
+#define __ELAN4_USERREGS_H
+
+#ident "$Id: userregs.h,v 1.14.2.1 2004/10/07 10:57:40 addy Exp $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/userregs.h,v $*/
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Statistic control reg values
+ * Each 4-bit nibble of the control word specifies what statistic
+ * is to be recorded in each of the 8 statistic counters
+ */
+#define COUNT_REG0_SHIFT   32ull
+#define COUNT_REG1_SHIFT   36ull
+#define COUNT_REG2_SHIFT   40ull
+#define COUNT_REG3_SHIFT   44ull
+#define COUNT_REG4_SHIFT   48ull
+#define COUNT_REG5_SHIFT   52ull
+#define COUNT_REG6_SHIFT   56ull
+#define COUNT_REG7_SHIFT   60ull
+
+
+/* Count reg 0 */
+#define STC_INPUT_NON_WRITE_BLOCKS     (0x0ull << COUNT_REG0_SHIFT)
+#define STP_DMA_EOP_WAIT_ACK           (0x1ull << COUNT_REG0_SHIFT)
+#define STP_TPROC_RUNNING              (0x2ull << COUNT_REG0_SHIFT)
+#define STC_STEN_PKTS_OPEN              (0x3ull << COUNT_REG0_SHIFT)
+#define STP_CPROC_HOLDS_FFU_DP         (0x4ull << COUNT_REG0_SHIFT)
+#define STC_TLB_TABLE_WALKS             (0x5ull << COUNT_REG0_SHIFT)
+#define STC_CACHE_HITS                  (0x6ull << COUNT_REG0_SHIFT)
+#define STC_PCI_SLAVE_READS             (0x7ull << COUNT_REG0_SHIFT)
+#define STP_PCI_WAITING_FOR_GNT         (0x8ull << COUNT_REG0_SHIFT)
+#define STP_SYS_CLOCK_RATE0            (0xfull << COUNT_REG0_SHIFT)
+
+#define STATS_REG0_NAMES {                     \
+          "STC_INPUT_NON_WRITE_BLOCKS",        \
+          "STP_DMA_EOP_WAIT_ACK",              \
+          "STP_TPROC_RUNNING",                 \
+          "STC_STEN_PKTS_OPEN",                \
+          "STP_CPROC_HOLDS_FFU_DP",            \
+          "STC_TLB_TABLE_WALKS",               \
+          "STC_CACHE_HITS",                    \
+          "STC_PCI_SLAVE_READS",               \
+          "STP_PCI_WAITING_FOR_GNT",           \
+          "STP_SYS_CLOCK_RATE0"                \
+}
+
+/* Count reg 1 */
+#define STC_INPUT_WRITE_BLOCKS         (0x0ull << COUNT_REG1_SHIFT)
+#define STP_DMA_DATA_TRANSMITTING      (0x1ull << COUNT_REG1_SHIFT)
+#define STC_CPROC_VALUES_EXE           (0x2ull << COUNT_REG1_SHIFT)
+#define STC_STEN_TRANS_SENT            (0x3ull << COUNT_REG1_SHIFT)
+#define STP_TPROC_DQ_HOLDS_FFU_DP      (0x4ull << COUNT_REG1_SHIFT)
+#define STC_TPROC_TLB_HITS             (0x5ull << COUNT_REG1_SHIFT)
+#define STC_CACHE_ALLOC_MISSES         (0x6ull << COUNT_REG1_SHIFT)
+#define STP_PCI_MASTER_READ_WAITING    (0x7ull << COUNT_REG1_SHIFT)
+#define STP_PCI_WAITING_FOR_DEVSEL      (0x8ull << COUNT_REG1_SHIFT)
+#define STP_SYS_CLOCK_RATE1            (0xfull << COUNT_REG1_SHIFT)
+
+#define STATS_REG1_NAMES {                    \
+          "STC_INPUT_WRITE_BLOCKS",            \
+          "STP_DMA_DATA_TRANSMITTING",         \
+          "STC_CPROC_VALUES_EXE",              \
+          "STC_STEN_TRANS_SENT",               \
+          "STP_TPROC_DQ_HOLDS_FFU_DP",         \
+          "STC_TPROC_TLB_HITS",                \
+          "STC_CACHE_ALLOC_MISSES",            \
+          "STP_PCI_MASTER_READ_WAITING",       \
+          "STP_PCI_WAITING_FOR_DEVSEL",        \
+          "STP_SYS_CLOCK_RATE1"                \
+}
+
+/* Count reg 2 */
+#define STC_INPUT_PKTS                 (0x0ull << COUNT_REG2_SHIFT)
+#define STP_DMA_WAITING_MEM            (0x1ull << COUNT_REG2_SHIFT)
+#define STC_CPROC_TRANSFERS             (0x2ull << COUNT_REG2_SHIFT)
+#define STP_STEN_WAIT_NETWORK_BUSY     (0x3ull << COUNT_REG2_SHIFT)
+#define STP_IPROC_HOLDS_FFU_DP         (0x4ull << COUNT_REG2_SHIFT)
+#define STC_UNITS_TLB_HITS             (0x5ull << COUNT_REG2_SHIFT)
+#define STC_CACHE_NON_ALLOC_MISSES      (0x6ull << COUNT_REG2_SHIFT)
+#define STP_PCI_MASTER_WRITE_WAITING   (0x7ull << COUNT_REG2_SHIFT)
+#define STC_PCI_OUT_OF_ORDER_SPLIT_COMP (0x8ull << COUNT_REG2_SHIFT)
+#define STP_SYS_CLOCK_RATE2            (0xfull << COUNT_REG2_SHIFT)
+
+#define STATS_REG2_NAMES {                    \
+          "STC_INPUT_PKTS",                    \
+          "STP_DMA_WAITING_MEM",               \
+          "STC_CPROC_TRANSFERS",               \
+          "STP_STEN_WAIT_NETWORK_BUSY",        \
+          "STP_IPROC_HOLDS_FFU_DP",            \
+          "STC_UNITS_TLB_HITS",                \
+          "STC_CACHE_NON_ALLOC_MISSES",        \
+          "STP_PCI_MASTER_WRITE_WAITING",      \
+          "STC_PCI_OUT_OF_ORDER_SPLIT_COMP",   \
+          "STP_SYS_CLOCK_RATE2"                \
+}
+
+/* Count reg 3 */
+#define STC_INPUT_PKTS_REJECTED         (0x0ull << COUNT_REG3_SHIFT)
+#define STP_DMA_WAIT_NETWORK_BUSY       (0x1ull << COUNT_REG3_SHIFT)
+#define STC_CPROC_PREFETCH_SDRAM        (0x2ull << COUNT_REG3_SHIFT)
+#define STP_STEN_BLOCKED_ACKS_OR_VC     (0x3ull << COUNT_REG3_SHIFT)
+#define STP_EPROC_HOLDS_FFU_DP          (0x4ull << COUNT_REG3_SHIFT)
+#define STP_TPROC_BLOCKED_MEMSYS        (0x5ull << COUNT_REG3_SHIFT)
+#define STC_CACHE_WRITE_BACKS           (0x6ull << COUNT_REG3_SHIFT)
+#define STP_PCI_SLAVE_READ_WAITING      (0x7ull << COUNT_REG3_SHIFT)
+#define STP_PCI_IDLE_CYCLES            (0x8ull << COUNT_REG3_SHIFT)
+#define STP_SYS_CLOCK_RATE3            (0xfull << COUNT_REG3_SHIFT)
+
+#define STATS_REG3_NAMES {                    \
+          "STC_INPUT_PKTS_REJECTED",           \
+          "STP_DMA_WAIT_NETWORK_BUSY",         \
+          "STC_CPROC_PREFETCH_SDRAM",          \
+          "STP_STEN_BLOCKED_ACKS_OR_VC",       \
+          "STP_EPROC_HOLDS_FFU_DP",            \
+          "STP_TPROC_BLOCKED_MEMSYS",          \
+          "STC_CACHE_WRITE_BACKS",             \
+          "STP_PCI_SLAVE_READ_WAITING",        \
+          "STP_PCI_IDLE_CYCLES",               \
+          "STP_SYS_CLOCK_RATE3"                \
+}
+
+/* Count reg 4 */
+#define STP_INPUT_DATA_TRANSMITTING    (0x0ull << COUNT_REG4_SHIFT)
+#define STC_DMA_PKTS_ACCEPTED          (0x1ull << COUNT_REG4_SHIFT)
+#define STC_CPROC_FLUSH_REQ_SDRAM      (0x2ull << COUNT_REG4_SHIFT)
+#define STP_STEN_EOP_WAIT_ACK          (0x3ull << COUNT_REG4_SHIFT)
+#define STP_DMA_HOLDS_FFU_DP           (0x4ull << COUNT_REG4_SHIFT)
+#define STP_UNIT_BLOCKED_MEMSYS        (0x5ull << COUNT_REG4_SHIFT)
+#define STC_PCI_MASTER_READS           (0x6ull << COUNT_REG4_SHIFT)
+#define STP_PCI_SLAVE_WRITE_WAITING    (0x7ull << COUNT_REG4_SHIFT)
+#define STC_INPUT_PACKETS_DISCARDED    (0x8ull << COUNT_REG4_SHIFT)
+#define STP_SYS_CLOCK_RATE4            (0xfull << COUNT_REG4_SHIFT)
+
+#define STATS_REG4_NAMES {                    \
+          "STP_INPUT_DATA_TRANSMITTING",       \
+          "STC_DMA_PKTS_ACCEPTED",             \
+          "STC_CPROC_FLUSH_REQ_SDRAM",         \
+          "STP_STEN_EOP_WAIT_ACK",             \
+          "STP_DMA_HOLDS_FFU_DP",              \
+          "STP_UNIT_BLOCKED_MEMSYS",           \
+          "STC_PCI_MASTER_READS",              \
+          "STP_PCI_SLAVE_WRITE_WAITING",       \
+          "STC_INPUT_PACKETS_DISCARDED",       \
+          "STP_SYS_CLOCK_RATE4"                \
+}
+
+/* Count reg 5 */
+#define STP_INPUT_WAITING_NETWORK_DATA  (0x0ull << COUNT_REG5_SHIFT)
+#define STC_DMA_PKTS_REJECTED           (0x1ull << COUNT_REG5_SHIFT)
+#define STC_CPROC_INSERT_CACHE_MISSES   (0x2ull << COUNT_REG5_SHIFT)
+#define STP_STEN_TRANSMITTING_DATA      (0x3ull << COUNT_REG5_SHIFT)
+#define FFU_BLOCKED_DIFF_FFU_PROC       (0x4ull << COUNT_REG5_SHIFT)
+#define STP_TABLE_WALKS_BLOCKED_MEMSYS  (0x5ull << COUNT_REG5_SHIFT)
+#define STC_PCI_MASTER_WRITES           (0x6ull << COUNT_REG5_SHIFT)
+#define STP_PCI_MASTER_HOLDS_BUS        (0x7ull << COUNT_REG5_SHIFT)
+#define STC_PCI_NO_SPLIT_COMPS         (0x8ull << COUNT_REG5_SHIFT)
+#define STP_SYS_CLOCK_RATE5            (0xfull << COUNT_REG5_SHIFT)
+
+#define STATS_REG5_NAMES {                    \
+          "STP_INPUT_WAITING_NETWORK_DATA",    \
+          "STC_DMA_PKTS_REJECTED",             \
+          "STC_CPROC_INSERT_CACHE_MISSES",     \
+          "STP_STEN_TRANSMITTING_DATA",        \
+          "FFU_BLOCKED_DIFF_FFU_PROC",         \
+          "STP_TABLE_WALKS_BLOCKED_MEMSYS",    \
+          "STC_PCI_MASTER_WRITES",             \
+          "STP_PCI_MASTER_HOLDS_BUS",          \
+          "STC_PCI_NO_SPLIT_COMPS",            \
+          "STP_SYS_CLOCK_RATE5"                \
+}
+
+/* Count reg 6 */
+#define STP_INPUT_BLOCKED_WAITING_TRANS (0x0ull << COUNT_REG6_SHIFT)
+#define STP_TPROC_INST_STALL           (0x1ull << COUNT_REG6_SHIFT)
+#define STP_CPROC_WAITING_DESCHED      (0x2ull << COUNT_REG6_SHIFT)
+#define STP_STEN_PKT_OPEN_WAITING_DATA (0x3ull << COUNT_REG6_SHIFT)
+#define STP_TLB_HASH_TABLE_ACCESSES    (0x4ull << COUNT_REG6_SHIFT)
+#define STP_PCI_SLAVE_BLOCKED_MEMSYS   (0x5ull << COUNT_REG6_SHIFT)
+#define STP_PCI_TRANSFERRING_DATA       (0x6ull << COUNT_REG6_SHIFT)
+#define STP_PCI_MASTER_WAITING_BUS      (0x7ull << COUNT_REG6_SHIFT)
+#define STP_PCI_READ_LATENCY           (0x8ull << COUNT_REG6_SHIFT)
+#define STP_SYS_CLOCK_RATE6            (0xfull << COUNT_REG6_SHIFT)
+
+#define STATS_REG6_NAMES {                    \
+          "STP_INPUT_BLOCKED_WAITING_TRANS",   \
+          "STP_TPROC_INST_STALL",              \
+          "STP_CPROC_WAITING_DESCHED",         \
+          "STP_STEN_PKT_OPEN_WAITING_DATA",    \
+          "STP_TLB_HASH_TABLE_ACCESSES",       \
+          "STP_PCI_SLAVE_BLOCKED_MEMSYS",      \
+          "STP_PCI_TRANSFERRING_DATA",         \
+          "STP_PCI_MASTER_WAITING_BUS",        \
+          "STP_PCI_READ_LATENCY",              \
+          "STP_SYS_CLOCK_RATE6"                \
+}
+
+/* Count reg 7 */
+#define STC_INPUT_CTX_FILTER_FILL       (0x0ull << COUNT_REG7_SHIFT)   
+#define STP_TPROC_LOAD_STORE_STALL      (0x1ull << COUNT_REG7_SHIFT)
+#define STC_CPROC_TIMEOUTS              (0x2ull << COUNT_REG7_SHIFT)
+#define STP_STEN_BLOCKED_NETWORK        (0x3ull << COUNT_REG7_SHIFT)
+#define STP_TLB_CHAIN_ACCESSES          (0x4ull << COUNT_REG7_SHIFT)
+#define STP_CPROC_SCHED_BLOCKED_MEMSYS  (0x5ull << COUNT_REG7_SHIFT)
+#define STC_PCI_SLAVE_WRITES            (0x6ull << COUNT_REG7_SHIFT)
+#define STC_PCI_DISCONNECTS_RETRIES     (0x7ull << COUNT_REG7_SHIFT)
+#define STC_RING_OSCILLATOR            (0x8ull << COUNT_REG7_SHIFT)
+#define STP_SYS_CLOCK_RATE7            (0xfull << COUNT_REG7_SHIFT)
+
+#define STATS_REG7_NAMES {                    \
+          "STC_INPUT_CTX_FILTER_FILL",         \
+          "STP_TPROC_LOAD_STORE_STALL",        \
+          "STC_CPROC_TIMEOUTS",                \
+          "STP_STEN_BLOCKED_NETWORK",          \
+          "STP_TLB_CHAIN_ACCESSES",            \
+          "STP_CPROC_SCHED_BLOCKED_MEMSYS",    \
+          "STC_PCI_SLAVE_WRITES",              \
+          "STC_PCI_DISCONNECTS_RETRIES",       \
+          "STC_RING_OSCILLATOR",               \
+          "STP_SYS_CLOCK_RATE7"                \
+}
+
+#define STATS_REG_NAMES { \
+    STATS_REG0_NAMES, \
+    STATS_REG1_NAMES, \
+    STATS_REG2_NAMES, \
+    STATS_REG3_NAMES, \
+    STATS_REG4_NAMES, \
+    STATS_REG5_NAMES, \
+    STATS_REG6_NAMES, \
+    STATS_REG7_NAMES, \
+}
+
+
+#define INPUT_PERF_STATS        (STC_INPUT_NON_WRITE_BLOCKS | STC_INPUT_WRITE_BLOCKS |              \
+                                STC_INPUT_PKTS | STC_INPUT_PKTS_REJECTED |                         \
+                                 STC_INPUT_CTX_FILTER_FILL | STP_INPUT_DATA_TRANSMITTING |           \
+                                STP_INPUT_WAITING_NETWORK_DATA | STP_INPUT_BLOCKED_WAITING_TRANS | STC_INPUT_PACKETS_DISCARDED) 
+
+#define DMA_PERF_STATS          (STC_DMA_PKTS_ACCEPTED | STC_DMA_PKTS_REJECTED |                    \
+                                 STP_DMA_EOP_WAIT_ACK | STP_DMA_DATA_TRANSMITTING |                 \
+                                STP_DMA_WAITING_MEM | STP_DMA_WAIT_NETWORK_BUSY)                 
+
+
+#define TPROC_PERF_STATS        (STP_TPROC_RUNNING | STP_TPROC_INST_STALL |                         \
+                                 STP_TPROC_LOAD_STORE_STALL)
+
+#define CPROC_PERF_STATS        (STC_CPROC_VALUES_EXE | STC_CPROC_TRANSFERS |                       \
+                                STC_CPROC_PREFETCH_SDRAM | STC_CPROC_FLUSH_REQ_SDRAM |             \
+                                STC_CPROC_INSERT_CACHE_MISSES | STP_CPROC_WAITING_DESCHED |        \
+                                STC_CPROC_TIMEOUTS)
+
+#define STEN_PERF_STATS         (STC_STEN_PKTS_OPEN | STC_STEN_TRANS_SENT |                         \
+                                STP_STEN_WAIT_NETWORK_BUSY | STP_STEN_BLOCKED_ACKS_OR_VC |         \
+                                STP_STEN_EOP_WAIT_ACK | STP_STEN_TRANSMITTING_DATA |               \
+                                STP_STEN_PKT_OPEN_WAITING_DATA | STP_STEN_BLOCKED_NETWORK)
+
+#define FFU_PREF_STATS          (STP_CPROC_HOLDS_FFU_DP | STP_TPROC_DQ_HOLDS_FFU_DP |               \
+                                STP_IPROC_HOLDS_FFU_DP | STP_EPROC_HOLDS_FFU_DP |                  \
+                                STP_DMA_HOLDS_FFU_DP | FFU_BLOCKED_DIFF_FFU_PROC)
+
+#define TABLE_WALK_PERF_STATS   (STC_TPROC_TLB_HITS | STC_UNITS_TLB_HITS |                          \
+                                STP_TLB_HASH_TABLE_ACCESSES | STP_TLB_CHAIN_ACCESSES |             \
+                                STC_TLB_TABLE_WALKS)
+
+#define ADDRESS_ARB_PERF_STATS  (STP_UNIT_BLOCKED_MEMSYS | STP_TPROC_BLOCKED_MEMSYS |               \
+                                STP_TABLE_WALKS_BLOCKED_MEMSYS | STP_CPROC_SCHED_BLOCKED_MEMSYS |  \
+                                STP_PCI_SLAVE_BLOCKED_MEMSYS)
+
+#define CACHE_PERF_STATS        (STC_CACHE_HITS | STC_CACHE_ALLOC_MISSES |                          \
+                                STC_CACHE_NON_ALLOC_MISSES | STC_CACHE_WRITE_BACKS)
+
+
+#define PCI_PERF_STATS          (STC_PCI_SLAVE_READS | STP_PCI_MASTER_READ_WAITING |                \
+                                 STP_PCI_MASTER_WRITE_WAITING | STP_PCI_SLAVE_READ_WAITING |        \
+                                 STP_PCI_SLAVE_WRITE_WAITING | STC_PCI_MASTER_WRITES |              \
+                                 STP_PCI_TRANSFERRING_DATA | STC_PCI_SLAVE_WRITES)
+
+#define PCIBUS_PERF_STATS       (STP_PCI_WAITING_FOR_GNT | STP_PCI_WAITING_FOR_DEVSEL |                    \
+                                STC_PCI_OUT_OF_ORDER_SPLIT_COMP | STP_PCI_IDLE_CYCLES |            \
+                                STC_PCI_MASTER_READS | STP_PCI_MASTER_HOLDS_BUS |                  \
+                                STP_PCI_MASTER_WAITING_BUS | STC_PCI_DISCONNECTS_RETRIES)
+
+                                
+    extern const char *elan_stats_names[8][10];
+
+#define ELAN_STATS_NAME(COUNT, CONTROL) (elan_stats_names[(COUNT)][(CONTROL) & 7])
+
+    typedef volatile union e4_StatsControl
+    {
+       E4_uint64 StatsControl;
+       struct
+       {
+#if (BYTE_ORDER == LITTLE_ENDIAN) || defined(__LITTLE_ENDIAN__)
+           E4_uint32 StatCont0:4;
+           E4_uint32 StatCont1:4;
+           E4_uint32 StatCont2:4;
+           E4_uint32 StatCont3:4;
+           E4_uint32 StatCont4:4;
+           E4_uint32 StatCont5:4;
+           E4_uint32 StatCont6:4;
+           E4_uint32 StatCont7:4;
+#else
+           E4_uint32 StatCont7:4;
+           E4_uint32 StatCont6:4;
+           E4_uint32 StatCont5:4;
+
+           E4_uint32 StatCont4:4;
+           E4_uint32 StatCont3:4;
+           E4_uint32 StatCont2:4;
+           E4_uint32 StatCont1:4;
+           E4_uint32 StatCont0:4;
+#endif
+           E4_uint32 pad;
+       } s;
+    } E4_StatsControl;
+
+typedef volatile union e4_StatsCount
+{
+   E4_uint64    ClockStat; 
+   struct
+   {
+       E4_uint32 ClockLSW;     /* read only */
+       E4_uint32 StatsCount;
+   } s;
+} E4_StatsCount;
+
+typedef volatile union e4_clock
+{
+   E4_uint64 NanoSecClock;
+   struct
+   {
+      E4_uint32 ClockLSW;
+      E4_uint32 ClockMSW;
+   } s;
+} E4_Clock;
+#define E4_TIME( X ) ((X).NanoSecClock)
+
+#define ELAN4_COMMS_CLOCK_FREQUENCY    660             /* In Mhz. This is half the bit rate. */
+#define ELAN4_CLOCK_ADD_VALUE          200             /* For 200ns increment rate */
+#define ELAN4_CLOCK_COMMS_DIV_VALUE    (((ELAN4_COMMS_CLOCK_FREQUENCY * ELAN4_CLOCK_ADD_VALUE) / (1000 * 4)) - 1)
+#define ELAN4_CLOCK_TICK_RATE          ((ELAN4_CLOCK_ADD_VALUE << 8) + ELAN4_CLOCK_COMMS_DIV_VALUE)
+
+typedef volatile union e4_clocktickrate
+{
+   E4_uint64 NanoSecClock;
+   struct
+   {
+      E4_uint32 pad1;
+      E4_uint32 TickRates;
+   } s;
+} E4_ClockTickRate;
+
+/*
+ * This is made into an 8k byte object.
+ */
+typedef volatile struct _E4_User_Regs
+{
+   E4_StatsCount       StatCounts[8];
+   E4_StatsCount       InstCount;
+   E4_Clock            Clock;
+   E4_StatsControl     StatCont;
+   E4_ClockTickRate    ClockTickRate;
+   E4_uint8            pad1[EightK - ((sizeof(E4_StatsCount)*9)+sizeof(E4_StatsControl)+
+                                       sizeof(E4_Clock)+sizeof(E4_ClockTickRate))];
+} E4_User_Regs;
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __ELAN4_USERREGS_H */
Index: linux-2.4.21/include/elan4/usertrap.h
===================================================================
--- linux-2.4.21.orig/include/elan4/usertrap.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/usertrap.h       2005-06-01 23:12:54.746416760 -0400
@@ -0,0 +1,114 @@
+/*
+ *    Copyright (c) 2001-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: usertrap.h,v 1.17 2004/05/05 09:08:35 david Exp $"
+/*      $Source: /cvs/master/quadrics/elan4mod/usertrap.h,v $*/
+
+#ifndef __ELAN4_USERTRAP_H
+#define __ELAN4_USERTRAP_H
+
+#ifndef _ASM
+typedef struct elan4_user_page
+{
+    E4_uint64          upage_ddcq_completed;
+} ELAN4_USER_PAGE;
+
+typedef struct elan4_user_trap
+{
+    int                                ut_type;
+    unsigned                   ut_proc;
+    unsigned                   ut_args[4];
+
+    union {
+       ELAN4_EPROC_TRAP        eproc;
+       ELAN4_CPROC_TRAP        cproc;
+       ELAN4_DPROC_TRAP        dproc;
+       ELAN4_IPROC_TRAP        iproc;
+       ELAN4_TPROC_TRAP        tproc;
+       ELAN4_NETERR_MSG        msg;
+    }                  ut_trap;
+} ELAN4_USER_TRAP;
+
+#endif /* _ASM */
+
+
+/* value for ut_type */
+#define UTS_FINISHED           0                               /* all pending traps have been handled */
+#define UTS_RESCHEDULE         1                               /* must return to user mode and re-enter */
+#define UTS_UNIMP_INSTR                2                               /* unimplemented thread instruction */
+#define UTS_EXECUTE_PACKET     3                               /* iproc trap needs packet executing */
+#define UTS_NETWORK_ERROR_TRAP 4                               /* network error on this trap */
+#define UTS_NETWORK_ERROR_MSG  5                               /* network error message  */
+#define UTS_NETWORK_ERROR_TIMER        6                               /* network error timer expired */
+
+#define UTS_EFAULT             -1                              /* failed to copyout trap */
+#define UTS_INVALID_ADDR       -2                              /* all -ve codes mean trap could not be resolved. */
+#define UTS_INVALID_VPROC      -3
+#define UTS_INVALID_COMMAND    -4
+#define UTS_BAD_TRAP           -5
+#define UTS_ALIGNMENT_ERROR    -6
+#define UTS_QUEUE_OVERFLOW     -7
+#define UTS_QUEUE_ERROR                -8
+#define UTS_INVALID_TRANS      -9
+#define UTS_PERMISSION_DENIED  -10
+#define UTS_CPROC_ERROR                -11
+#define UTS_INVALID_COOKIE     -12
+#define UTS_NETERR_ERROR       -13
+
+/* "special" values for registering handlers */
+#define UTS_ALL_TRAPS          -9999
+
+/* value for ut_proc */
+#define UTS_NOPROC             0
+#define UTS_EPROC              1
+#define UTS_CPROC              2
+#define UTS_DPROC              3
+#define UTS_TPROC              4
+#define UTS_IPROC              5
+#define UTS_NETERR_MSG         6
+
+/* unimplemented trap numbers for thread processor */
+#define ELAN4_T_TRAP_INSTR(t)  (0x80202000 | ((t) & 0xFF))
+
+#define ELAN4_T_SYSCALL_TRAP   1
+#  define ELAN4_T_OPEN         0
+#  define ELAN4_T_WRITE                1
+#  define ELAN4_T_READ         2
+#  define ELAN4_T_IOCTL                3
+#  define ELAN4_T_LSEEK                4
+#  define ELAN4_T_POLL         5
+#  define ELAN4_T_CLOSE                6
+#  define ELAN4_T_KILL         7
+#  define ELAN4_T_MMAP         8
+#  define ELAN4_T_MUNMAP       9
+#  define ELAN4_T_ABORT                100
+#  define ELAN4_T_DEBUG                101
+#  define ELAN4_T_REGDUMP      102
+
+#define ELAN4_T_REGDUMP_TRAP   2
+
+#define ELAN4_T_LIBELAN_TRAP   3
+#  define ELAN4_T_TPORT_NEWBUF 0
+#  define ELAN4_T_TPORT_GC     1
+#  define ELAN4_T_TPORT_DEBUG  2
+
+#define ELAN4_T_ALLOC_TRAP     4
+#  define ELAN4_T_ALLOC_ELAN   0
+#  define ELAN4_T_ALLOC_MAIN   1
+#  define ELAN4_T_FREE_ELAN    2
+#  define ELAN4_T_FREE_MAIN    3
+
+/* reserved main interrupt cookies */
+#define ELAN4_INT_COOKIE_DDCQ  0
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
+#endif /* __ELAN4_USERTRAP_H */
Index: linux-2.4.21/include/elan4/xsdram.h
===================================================================
--- linux-2.4.21.orig/include/elan4/xsdram.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/elan4/xsdram.h 2005-06-01 23:12:54.747416608 -0400
@@ -0,0 +1,59 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2003 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __ELAN4_XSDRAM_H
+#define __ELAN4_XSDRAM_H
+
+#ident "@(#)$Id: xsdram.h,v 1.13 2004/03/05 12:32:04 jon Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/elan4hdr/xsdram.h,v $*/
+
+/* SAMSUNG K4H281638D-TCB3 */
+
+#define SDRAM_tRCF_1_SH         0
+#define SDRAM_tRP_1_SH          4
+#define SDRAM_tRCD_SH           8
+#define SDRAM_tRRD_SH           12
+#define SDRAM_tEndWr_SH         16
+#define SDRAM_tEndRd_SH         20
+#define SDRAM_Burst_SH          24
+#define SDRAM_CL_SH             28
+#define SDRAM_DsblBypass       (1ULL << 31)
+#define SDRAM_RefreshRate_SH    32
+#define SDRAM_RamSize_SH        34
+#define SDRAM_ReadLtncy_1_SH    36
+#define SDRAM_RdOffset_SH       40
+#define SDRAM_FlightDelay_SH    42
+
+#define SDRAM_ENABLE_ECC       (1ULL << 44) // Enables error detecting on the ECC.
+#define SDRAM_SDRAM_TESTING    (1ULL << 45) // Switches to test mode for checking EEC data bits
+#define SDRAM_SETUP            (1ULL << 46) // Writes SDram control reg when set. Also starts
+
+#define SDRAM_CS_MODE0          0ULL         // 64Mbit, 128Mbit, 256Mbit, 512Mbit or 1Gbit (16-bit output)
+#define SDRAM_CS_MODE1          1ULL         // 64Mbit, 128Mbit, 256Mbit or 512Mbit (8-bit output)
+#define SDRAM_CS_MODE2          2ULL         // 2Gbit (16-bit output) or 1Gbit (8-bit output)
+#define SDRAM_CS_MODE3          3ULL         // 4Gbit (16-bit output) or 2Gbit (8-bit output)
+
+#if defined(LINUX) && !defined(CONFIG_MPSAS)
+#define SDRAM_STARTUP_VALUE   ((0xbULL << SDRAM_tRCF_1_SH)      | (0x2ULL << SDRAM_tRP_1_SH)       | \
+                               (0x3ULL << SDRAM_tRCD_SH)        | (0x2ULL << SDRAM_tRRD_SH)        | \
+                               (0xaULL << SDRAM_tEndWr_SH)      | (0x6ULL << SDRAM_tEndRd_SH)      | \
+                               (0x8ULL << SDRAM_Burst_SH)       | (0x6ULL << SDRAM_CL_SH)          | \
+                               (0x2ULL << SDRAM_RefreshRate_SH) | (0x3ULL << SDRAM_RamSize_SH)     | \
+                               (0x1ULL << SDRAM_RdOffset_SH)    | (0x1ULL << SDRAM_FlightDelay_SH) | \
+                               (0x4ULL << SDRAM_ReadLtncy_1_SH))
+#else
+#define SDRAM_STARTUP_VALUE   ((0xbULL << SDRAM_tRCF_1_SH)      | (0x2ULL << SDRAM_tRP_1_SH)       | \
+                               (0x3ULL << SDRAM_tRCD_SH)        | (0x2ULL << SDRAM_tRRD_SH)        | \
+                               (0xaULL << SDRAM_tEndWr_SH)      | (0x6ULL << SDRAM_tEndRd_SH)      | \
+                               (0x8ULL << SDRAM_Burst_SH)       | (0x6ULL << SDRAM_CL_SH)          | \
+                               (0x0ULL << SDRAM_RefreshRate_SH) | (0x0ULL << SDRAM_RamSize_SH)     | \
+                               (0x1ULL << SDRAM_RdOffset_SH)    | (0x1ULL << SDRAM_FlightDelay_SH) | \
+                               (0x4ULL << SDRAM_ReadLtncy_1_SH) | SDRAM_ENABLE_ECC | SDRAM_SETUP)
+#endif
+
+#endif /* __ELAN4_XSDRAM_H */
Index: linux-2.4.21/include/jtag/jtagio.h
===================================================================
--- linux-2.4.21.orig/include/jtag/jtagio.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/jtag/jtagio.h  2005-06-01 23:12:54.747416608 -0400
@@ -0,0 +1,106 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "$Id: jtagio.h,v 1.7.8.1 2005/01/27 15:21:47 lee Exp $"
+/*             $Source: /cvs/master/quadrics/jtagmod/jtagio.h,v $*/
+
+
+#ifndef __SYS_JTAGMOD_H
+#define __SYS_JTAGMOD_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define JTAG_MAX_CHIPS         8
+#define JTAG_MAX_INSTR_LEN     8
+#define JTAG_MAX_BITS           (JTAG_MAX_CHIPS * JTAG_MAX_INSTR_LEN)
+#define JTAG_MAX_DATA_LEN      1024
+
+#define JTAG_BYPASS            0xFF
+
+#define I2C_ADDR_LEN           7                               /* 7 bits of address */
+#define I2C_DATA_LEN           8                               /* 8 bits of data */
+#define I2C_MAX_DATA_LEN       9                               /* and upto 9 bytes worth */
+
+#define BITS_PER_BYTE          8
+#define JTAG_NBYTES(nbits)     (((nbits)+BITS_PER_BYTE-1)/BITS_PER_BYTE)
+#define JTAG_BIT(v, num)       (((v)[(num) / BITS_PER_BYTE] >> ((num) % BITS_PER_BYTE)) & 1)
+#define JTAG_SET_BIT(v, num)   ((v)[(num) / BITS_PER_BYTE] |= (1 << ((num) % BITS_PER_BYTE)))
+#define JTAG_CLR_BIT(v, num)   ((v)[(num) / BITS_PER_BYTE] &= ~(1 << ((num) % BITS_PER_BYTE)))
+
+#define RING_CLOCK_CARD                (0x3D)
+#define RING_CLOCK_SHIFT       (0x3E)
+#define RING_JTAG_LOOPBACK     (0x3F)
+#define RING_MAX               (0x40)
+
+#define RING_QUAD_BIT          (0x40)
+#define RING_I2C_BIT           (0x80)
+
+#define VALID_JTAG_RING(ring)  ((ring) < 0x20 || (ring) == RING_JTAG_LOOPBACK)
+#define VALID_I2C_RING(ring)   ((ring) < 0x20 || (ring) == RING_CLOCK_CARD)
+
+
+typedef struct jtag_value
+{
+    u_char     bytes[JTAG_NBYTES(JTAG_MAX_DATA_LEN)];
+} JTAG_VALUE;
+
+/* arguements to JTAG_SHIFT_IR/JTAG_SHIFT_DR */
+typedef struct jtag_reset_args
+{
+    u_int      ring;
+} JTAG_RESET_ARGS;
+
+typedef struct jtag_shift_args
+{
+    u_int      ring;
+    u_int      nbits;
+    u_char     *value;
+} JTAG_SHIFT_ARGS;
+
+typedef struct i2c_args
+{
+    u_int      ring;
+    u_int      device;
+    u_int      reg;
+    u_int      count;
+    u_int      ok;
+    u_char     data[I2C_MAX_DATA_LEN];
+} I2C_ARGS;
+
+/* values for 'ok' - the return value from i2c_xx functions */
+#define I2C_OP_SUCCESS         0
+#define I2C_OP_ERROR           1
+#define I2C_OP_NOT_IDLE                2
+#define I2C_OP_NO_DEVICE       3
+#define I2C_OP_WRITE_TO_BIG    4
+#define I2C_OP_BAD_RESOURCE    5
+
+typedef struct i2c_clock_shift_args
+{
+    u_int      t;
+    u_int      n;
+    u_int      m;
+} I2C_CLOCK_SHIFT_ARGS;
+
+#define JTAG_RESET             _IOWR('j', '0', JTAG_RESET_ARGS)
+#define JTAG_SHIFT_IR          _IOWR('j', '1', JTAG_SHIFT_ARGS)
+#define JTAG_SHIFT_DR          _IOWR('j', '2', JTAG_SHIFT_ARGS)
+
+#define I2C_CLOCK_SHIFT                _IOWR('j', '4', I2C_CLOCK_SHIFT_ARGS)
+#define I2C_WRITE              _IOWR('j', '5', I2C_ARGS)
+#define I2C_READ               _IOWR('j', '6', I2C_ARGS)
+#define I2C_WRITEREG           _IOWR('j', '7', I2C_ARGS)
+#define I2C_READREG            _IOWR('j', '8', I2C_ARGS)
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __SYS_JTAGMOD_H */
Index: linux-2.4.21/include/linux/coproc.h
===================================================================
--- linux-2.4.21.orig/include/linux/coproc.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/linux/coproc.h 2005-06-01 23:12:54.748416456 -0400
@@ -0,0 +1,206 @@
+/*
+ *    Copyright (C) 2002, 2003 Quadrics Ltd.
+ *
+ *    This program is free software; you can redistribute it and/or modify
+ *    it under the terms of the GNU General Public License as published by
+ *    the Free Software Foundation; either version 2 of the License, or
+ *    (at your option) any later version.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    GNU General Public License for more details.
+ *
+ *    You should have received a copy of the GNU General Public License
+ *    along with this program; if not, write to the Free Software
+ *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ *
+ */
+
+/*
+ * Callbacks for coprocessor page table updates.
+ */
+
+#ifndef __LINUX_COPROC_H__
+#define __LINUX_COPROC_H__
+
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/list.h>
+#include <linux/slab.h> /* kmalloc */
+
+typedef struct coproc_ops_struct {
+       struct list_head list;
+       void *arg;
+
+       void (*release)(void *arg, struct mm_struct *mm);
+       void (*sync_range)(void *arg, struct mm_struct *mm, unsigned long start, unsigned long end);
+       void (*invalidate_range)(void *arg, struct mm_struct *mm, unsigned long start, unsigned long end);
+       void (*update_range)(void *arg, struct mm_struct *mm, unsigned long start, unsigned long end);
+
+       void (*change_protection)(void *arg, struct mm_struct *mm, unsigned long start, unsigned long end, pgprot_t newprot);
+
+       void (*sync_page)(void *arg, struct vm_area_struct *vma, unsigned long address);
+       void (*invalidate_page)(void *arg, struct vm_area_struct *vma, unsigned long address);
+       void (*update_page)(void *arg, struct vm_area_struct *vma, unsigned long address);
+
+} coproc_ops_t;
+
+extern __inline__ void 
+register_coproc_ops(struct mm_struct *mm, coproc_ops_t *cp)
+{
+       if (mm->coproc_ops == NULL) {
+               mm->coproc_ops = (struct list_head *)
+                   kmalloc(sizeof(struct list_head), GFP_KERNEL);
+               INIT_LIST_HEAD(mm->coproc_ops);
+       }
+       list_add(&cp->list, mm->coproc_ops);
+}
+
+extern __inline__ void 
+unregister_coproc_ops(struct mm_struct *mm, coproc_ops_t *cp)
+{
+       list_del(&cp->list);
+       if (list_empty(mm->coproc_ops)) {
+               kfree(mm->coproc_ops);
+               mm->coproc_ops = NULL;
+       }
+}
+
+extern __inline__ void 
+coproc_release(struct mm_struct *mm)
+{
+       struct list_head *head = mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               while (! list_empty(head)) {
+                   lp = head->next;
+                   cp = list_entry(lp, coproc_ops_t, list);
+                   
+                   list_del (&cp->list);
+                   
+                   if (cp->release)
+                       cp->release(cp->arg, mm);
+               }
+               kfree(head);
+               mm->coproc_ops = NULL;
+       }
+}
+
+extern __inline__ void 
+coproc_sync_range(struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+       struct list_head *head = mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               for (lp = head->next; lp != head; lp = lp->next) {
+                       cp = list_entry(lp, coproc_ops_t, list);
+                       if (cp->sync_range)
+                               cp->sync_range(cp->arg, mm, start, end);
+               }               
+       }
+}
+
+extern __inline__ void 
+coproc_invalidate_range(struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+       struct list_head *head = mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               for (lp = head->next; lp != head; lp = lp->next) {
+                       cp = list_entry(lp, coproc_ops_t, list);
+                       if (cp->invalidate_range)
+                               cp->invalidate_range(cp->arg, mm, start, end);
+               }               
+       }
+}
+
+extern __inline__ void 
+coproc_update_range(struct mm_struct *mm, unsigned long start, unsigned long end)
+{
+       struct list_head *head = mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               for (lp = head->next; lp != head; lp = lp->next) {
+                       cp = list_entry(lp, coproc_ops_t, list);
+                       if (cp->update_range)
+                               cp->update_range(cp->arg, mm, start, end);
+               }               
+       }
+}
+
+extern __inline__ void 
+coproc_change_protection (struct mm_struct *mm, unsigned long start, unsigned long end, pgprot_t newprot)
+{
+       struct list_head *head = mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               for (lp = head->next; lp != head; lp = lp->next) {
+                       cp = list_entry(lp, coproc_ops_t, list);
+                       if (cp->change_protection)
+                               cp->change_protection(cp->arg, mm, start, end, newprot);
+               }               
+       }
+}
+
+extern __inline__ void 
+coproc_sync_page(struct vm_area_struct *vma, unsigned long addr)
+{
+       struct list_head *head = vma->vm_mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               for (lp = head->next; lp != head; lp = lp->next) {
+                       cp = list_entry(lp, coproc_ops_t, list);
+                       if (cp->sync_page)
+                               cp->sync_page(cp->arg, vma, addr);
+               }               
+       }
+}
+
+extern __inline__ void 
+coproc_invalidate_page(struct vm_area_struct *vma, unsigned long addr)
+{
+       struct list_head *head = vma->vm_mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               for (lp = head->next; lp != head; lp = lp->next) {
+                       cp = list_entry(lp, coproc_ops_t, list);
+                       if (cp->invalidate_page)
+                               cp->invalidate_page(cp->arg, vma, addr);
+               }               
+       }
+}
+
+extern __inline__ void 
+coproc_update_page(struct vm_area_struct *vma, unsigned long addr)
+{
+       struct list_head *head = vma->vm_mm->coproc_ops;
+        struct list_head *lp;
+        coproc_ops_t *cp;
+
+       if (head) {
+               for (lp = head->next; lp != head; lp = lp->next) {
+                       cp = list_entry(lp, coproc_ops_t, list);
+                       if (cp->update_page)
+                               cp->update_page(cp->arg, vma, addr);
+               }               
+       }
+}
+
+
+#endif /* __LINUX_COPROC_H__ */
Index: linux-2.4.21/include/linux/ptrack.h
===================================================================
--- linux-2.4.21.orig/include/linux/ptrack.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/linux/ptrack.h 2005-06-01 23:12:54.748416456 -0400
@@ -0,0 +1,53 @@
+/*
+ *    Copyright (C) 2000  Regents of the University of California
+ *
+ *    This program is free software; you can redistribute it and/or modify
+ *    it under the terms of the GNU General Public License as published by
+ *    the Free Software Foundation; either version 2 of the License, or
+ *    (at your option) any later version.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    GNU General Public License for more details.
+ *
+ *    You should have received a copy of the GNU General Public License
+ *    along with this program; if not, write to the Free Software
+ *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ * Derived from exit_actn.c by
+ *    Copyright (C) 2003 Quadrics Ltd.
+ *
+ */
+#ifndef __LINUX_PTRACK_H
+#define __LINUX_PTRACK_H
+
+/* 
+ * Process tracking - this allows a module to keep track of processes
+ * in order that it can manage all tasks derived from a single process.
+ */
+
+#define PTRACK_PHASE_CLONE             1
+#define PTRACK_PHASE_CLONE_FAIL        2
+#define PTRACK_PHASE_EXEC              3
+#define PTRACK_PHASE_EXIT              4
+
+typedef int (*ptrack_callback_t)(void *arg, int phase, struct task_struct *child);
+
+#define PTRACK_FINISHED                0
+#define PTRACK_INNHERIT                1
+#define PTRACK_DENIED          2
+
+struct ptrack_desc {
+       struct list_head        link;
+       ptrack_callback_t       callback;
+       void                   *arg;
+};
+
+extern int     ptrack_register (ptrack_callback_t callback, void *arg);
+extern void    ptrack_deregister (ptrack_callback_t callback, void *arg);
+extern int     ptrack_registered (ptrack_callback_t callback, void *arg);
+
+extern int     ptrack_call_callbacks (int phase, struct task_struct *child);
+
+#endif /* __LINUX_PTRACK_H */
Index: linux-2.4.21/include/linux/sched.h
===================================================================
--- linux-2.4.21.orig/include/linux/sched.h     2005-06-01 22:52:05.000000000 -0400
+++ linux-2.4.21/include/linux/sched.h  2005-06-01 23:12:54.749416304 -0400
@@ -30,6 +30,8 @@
 #include <linux/pid.h>
 #include <linux/kernel_stat.h>
 
+#include <linux/list.h>
+
 struct exec_domain;
 extern int exec_shield;
 extern int exec_shield_randomize;
@@ -322,6 +324,9 @@
 #endif
        /* Architecture-specific MM context */
        mm_context_t context;
+       
+       /* Support page table updates on adapter cards with on-board MMU */
+       struct list_head *coproc_ops;
 
        /* coredumping support */
        int core_waiters;
@@ -342,6 +347,7 @@
        mmap_sem:       __RWSEM_INITIALIZER(name.mmap_sem), \
        page_table_lock: SPIN_LOCK_UNLOCKED,            \
        mmlist:         LIST_HEAD_INIT(name.mmlist),    \
+       coproc_ops:     NULL,                           \
        rlimit_rss:     RLIM_INFINITY,                  \
 }
 
@@ -572,6 +578,9 @@
 /* context-switch lock */
        spinlock_t switch_lock;
 
+/* process tracking callbacks */
+       struct list_head ptrack_list;
+
 /* journalling filesystem info */
        void *journal_info;
 
@@ -740,6 +749,7 @@
     blocked:           {{0}},                                          \
     alloc_lock:                SPIN_LOCK_UNLOCKED,                             \
     switch_lock:       SPIN_LOCK_UNLOCKED,                             \
+    ptrack_list:       LIST_HEAD_INIT(tsk.ptrack_list),                \
     journal_info:      NULL,                                           \
     real_stack:                &tsk,                                           \
 }
Index: linux-2.4.21/include/qsnet/autoconf.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/autoconf.h  2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/autoconf.h       2005-06-01 23:12:54.750416152 -0400
@@ -0,0 +1,38 @@
+/*
+ *    Copyright (c) 2004 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ * NOTE: This file has been automatically generated:
+ *       node   : milano
+ *       kernel : /src/linux/qsnet/linux-2.4.21
+ *       date   : Wed May  4 18:24:23 EDT 2005
+ *
+ */
+
+#include <linux/version.h>
+#undef NO_RMAP
+#define        AC
+#undef NO_O1_SCHED
+#undef NO_NPTL
+#define        NO_ABI
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 5, 0)
+#define        PROCESS_ACCT
+#endif
+#undef RSS_ATOMIC
+#undef NO_COPROC
+#define        NO_IOPROC
+#undef NO_PTRACK
+#define        NO_PANIC_NOTIFIER
+#undef NO_SHM_CLEANUP
+#undef NO_PDE
+
+
+#define        CONFIG_EIP
+#define        CONFIG_ELAN
+#define        CONFIG_ELAN3
+#define        CONFIG_ELAN4
+#define        CONFIG_EP
+#define        CONFIG_JTAG
+#define        CONFIG_QSNET
+#define        CONFIG_RMS
Index: linux-2.4.21/include/qsnet/condvar.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/condvar.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/condvar.h        2005-06-01 23:12:54.750416152 -0400
@@ -0,0 +1,140 @@
+/*
+ *    Copyright (C) 2000  Regents of the University of California
+ *
+ *    This program is free software; you can redistribute it and/or modify
+ *    it under the terms of the GNU General Public License as published by
+ *    the Free Software Foundation; either version 2 of the License, or
+ *    (at your option) any later version.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    GNU General Public License for more details.
+ *
+ *    You should have received a copy of the GNU General Public License
+ *    along with this program; if not, write to the Free Software
+ *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+
+#if    !defined(_LINUX_CONDVAR_H)
+#define        _LINUX_CONDVAR_H
+
+#if    defined(__KERNEL__)
+
+#include <linux/list.h>
+#include <qsnet/debug.h>
+
+#define CV_RET_SIGPENDING      0
+#define CV_RET_TIMEOUT         (-1)
+#define CV_RET_NORMAL          1
+
+struct kcondvar_task {
+       struct task_struct      *task;          /* need to wrap task in this */
+       struct list_head        list;           /*   to thread as a list */
+       int                     blocked;
+};
+
+typedef struct {
+       struct list_head        task_list;      /* list of kcondvar_task's */
+} kcondvar_t;
+
+#define kcondvar_wait(c,l,fl)                  debug_kcondvar_wait(c, l, fl, 0,  TASK_UNINTERRUPTIBLE)
+#define kcondvar_waitsig(c,l,fl)               debug_kcondvar_wait(c, l, fl, 0,  TASK_INTERRUPTIBLE)
+#define kcondvar_timedwait(c,l,fl,to)          debug_kcondvar_wait(c, l, fl, to, TASK_UNINTERRUPTIBLE)
+#define kcondvar_timedwaitsig(c,l,fl,to)       debug_kcondvar_wait(c, l, fl, to, TASK_INTERRUPTIBLE)
+#define kcondvar_wakeupone(c,l)                        kcondvar_wakeup(c, l, 0)
+#define kcondvar_wakeupall(c,l)                        kcondvar_wakeup(c, l, 1)
+ 
+extern __inline__ void
+kcondvar_init(kcondvar_t *c)
+{
+       INIT_LIST_HEAD(&c->task_list);
+}
+
+extern __inline__ void
+kcondvar_destroy(kcondvar_t *c)
+{
+       ASSERT(list_empty(&c->task_list));
+}
+
+/*
+ * We thread a struct kcondvar_task, allocated on the stack, onto the kcondvar_t's
+ * task_list, and take it off again when we wake up.
+ */
+extern __inline__ int
+debug_kcondvar_wait(kcondvar_t *c, spinlock_t *l, unsigned long *fl, long tmo, int state)
+{
+       struct kcondvar_task cvt;
+       int ret = CV_RET_NORMAL;
+
+       ASSERT(!in_interrupt());                /* we can block */
+       ASSERT(SPINLOCK_HELD(l));               /* enter holding lock */
+
+       cvt.task = current;
+       cvt.blocked = 1;
+       list_add(&cvt.list, &c->task_list);
+       do {
+              /* Note: we avoid using TASK_UNINTERRUPTIBLE here because avenrun()
+               * (linux/kernel/timer.c:calc_load())
+               * computation treats it like TASK_RUNNABLE hence creates false high
+               * load averages when we create kernel threads.
+               * The cvt.blocked flag distinguishes a signal wakeup from a kcondvar_wakeup.
+               *
+               * However, if we do take a signal we could end up busily spinning here, if
+               * we ignore it (state == TASK_UNINTERRUPTIBLE) so once we see a signal
+               * pending we do sleep TASK_UNINTERRUPTIBLE to stop a busy spin.
+               * I have now blocked all signals for kernel threads to prevent this
+               * happening but other users of kcondvar_wait may still hit this spin.
+               */
+               set_current_state (signal_pending(current) ? state : TASK_INTERRUPTIBLE);
+
+               if (fl)
+                   spin_unlock_irqrestore(l, *fl);
+               else
+                   spin_unlock(l);
+               if (tmo) {
+                       if (tmo <= jiffies || !schedule_timeout(tmo - jiffies))
+                               ret = CV_RET_TIMEOUT;
+               } else
+                       schedule();
+               if (fl)
+                   spin_lock_irqsave (l, *fl);
+               else
+                   spin_lock(l);
+               
+               /* signal_pending - Only exit the loop if the user was waiting TASK_INTERRUPTIBLE */
+               if ((state == TASK_INTERRUPTIBLE) && signal_pending(current))
+                       ret = CV_RET_SIGPENDING;
+
+       } while (cvt.blocked && ret == CV_RET_NORMAL);
+       list_del(&cvt.list);
+
+       /* Reset task state in case we didn't sleep above */
+       set_current_state (TASK_RUNNING);
+
+       return ret;                             /* return holding lock */
+}
+
+extern __inline__ void
+kcondvar_wakeup(kcondvar_t *c, spinlock_t *l, int wakeall)
+{
+       struct list_head *lp;
+       struct kcondvar_task *cvtp;
+
+       ASSERT(SPINLOCK_HELD(l));                       /* already holding lock */
+       for (lp = c->task_list.next; lp != &c->task_list; lp = lp->next) {
+               cvtp = list_entry(lp, struct kcondvar_task, list);
+               if (cvtp->blocked) {
+                       cvtp->blocked = 0;
+                       /* wake_up_process added to kernel/ksyms.c */
+                       wake_up_process(cvtp->task); 
+                       if (!wakeall)
+                               break;
+               }
+       }
+}                                              /* return still holding lock */
+
+
+#endif /* __KERNEL__ */
+#endif /* _LINUX_CONDVAR_H */
Index: linux-2.4.21/include/qsnet/config.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/config.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/config.h 2005-06-01 23:12:54.751416000 -0400
@@ -0,0 +1,195 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _QSNET_CONFIG_H
+#define _QSNET_CONFIG_H
+
+#ident "$Id: config.h,v 1.23 2003/07/24 21:31:19 robin Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/config.h,v $*/
+
+
+/*
+ * QSNET standard defines :
+ *
+ *   Target operating system defines
+ *             SOLARIS
+ *             TRU64UNIX/DIGITAL_UNIX
+ *             LINUX
+ *
+ *   Target processor defines
+ *             SPARC
+ *             ALPHA
+ *             I386
+ *             IA64
+ *             X86_64
+ *
+ *   Byte order defines
+ *             __LITTLE_ENDIAN__
+ *             __BIG_ENDIAN__
+ *
+ *   Data size defines
+ *             _LP64                   - LP64 - long/pointer is 64 bits
+ *             _ILP32                  - LP32 - long/pointer is 32 bits
+ *
+ *   Elan defines for main processor
+ *             __MAIN_LITTLE_ENDIAN__  - main byte order (for thread code)
+ *             __MAIN_BIG_ENDIAN__
+ *             _MAIN_LP64              - main long size (for thread code)
+ *             _MAIN_ILP32
+ *
+ *   Compiling for kernel (defined in makefile)
+ *             _KERNEL
+ *
+ */
+
+#if defined(__LP64__) && !defined(_LP64)
+#  define _LP64
+#endif
+
+#if defined(__arch64__) && !defined(_LP64) && !defined(_ILP32)
+#  define _LP64
+#endif
+
+#if defined(__alpha__) && !defined(_LP64) && !defined(_ILP32)
+#  define _LP64
+#endif
+
+#if !defined(__arch64__) && !defined(_ILP32) && !defined(_LP64)
+#  define _ILP32
+#endif
+
+#if defined(__ELAN__) || defined(__ELAN3__)
+
+#define __LITTLE_ENDIAN__
+
+#if defined(__host_solaris) && defined(__host_sparc)
+#define SOLARIS
+#define SPARC
+#define SOLARIS_SPARC
+#define _MAIN_ILP32
+#define __MAIN_BIG_ENDIAN__
+
+#elif defined(__host_osf)
+#define TRU64UNIX
+#define DIGITAL_UNIX
+#define ALPHA
+#define _MAIN_LP64
+#define __MAIN_LITTLE_ENDIAN__
+
+#elif defined(__host_linux) && defined(__host_alpha)
+#define LINUX
+#define ALPHA
+#define LINUX_ALPHA
+#define _MAIN_LP64
+#define __MAIN_LITTLE_ENDIAN__
+
+#elif defined(__host_linux) && defined(__host_sparc)
+#define LINUX
+#define SPARC
+#define LINUX_SPARC
+#define __MAIN_BIG_ENDIAN__
+#ifdef __KERNEL__
+#  define _MAIN_LP64
+#else
+#  define _MAIN_ILP32
+#endif
+
+#elif defined(__host_linux) && defined(__host_i386)
+#define LINUX
+#define I386
+#define LINUX_I386
+#define _MAIN_ILP32
+#define __MAIN_LITTLE_ENDIAN__
+
+#elif defined(__host_linux) && defined(__host_ia64)
+#define LINUX
+#define IA64
+#define LINUX_IA64
+#define _MAIN_LP64
+#define __MAIN_LITTLE_ENDIAN__
+
+#elif defined(__host_linux) && defined(__host_x86_64)
+#define LINUX
+#define X86_64
+#define LINUX_X86_64
+#define _MAIN_LP64
+#define __MAIN_LITTLE_ENDIAN__
+
+#else
+#error Cannot determine operating system/processor architecture.
+#endif
+
+#else /* !defined(__ELAN3__) */
+
+#if (defined(sun) || defined(__sun)) && defined(sparc) && !defined(__sparcv9)  /* Sun Solaris 5.6 */
+#define SOLARIS
+#define SPARC
+#define SOLARIS_SPARC
+#ifndef __BIG_ENDIAN__
+#define __BIG_ENDIAN__
+#endif 
+
+#elif (defined(sun) || defined(__sun)) && defined(sparc) && defined(__sparcv9) /* Sun Solaris 5.7 */
+#define SOLARIS
+#define SPARC
+#define SOLARIS_SPARC
+#define __BIG_ENDIAN__
+
+#elif defined(__osf__) && defined(__alpha)                                     /* Digital Unix */
+#define TRU64UNIX
+#define DIGITAL_UNIX
+#define ALPHA
+#define __LITTLE_ENDIAN__
+
+#elif (defined(linux) || defined(__linux__)) && defined(__alpha)               /* Linux Alpha */
+
+#define LINUX
+#define ALPHA
+#define LINUX_ALPHA
+#define __LITTLE_ENDIAN__
+
+#elif (defined(linux) || defined(__linux__)) && defined(__sparc)               /* Linux Sparc */
+
+#define LINUX
+#define SPARC
+#define LINUX_SPARC
+#define __BIG_ENDIAN__
+
+#elif (defined(linux) || defined(__linux__)) && defined(__i386)                        /* Linux i386 */
+
+#define LINUX
+#define I386
+#define LINUX_I386
+#define __LITTLE_ENDIAN__
+
+#elif (defined(linux) || defined(__linux__)) && defined(__ia64)                        /* Linux ia64 */
+
+#define LINUX
+#define IA64
+#define LINUX_IA64
+#define __LITTLE_ENDIAN__
+
+#elif (defined(linux) || defined(__linux__)) && defined(__x86_64)                      /* Linux x86_64 */
+
+#define LINUX
+#define X86_64
+#define LINUX_X86_64
+#define __LITTLE_ENDIAN__
+
+#elif defined(__QNXNTO__)
+#define QNX
+#define I386
+#define __LITTLE_ENDIAN__
+#else
+#error Cannot determine operating system/processor architecture.
+#endif
+
+#endif
+
+#include <qsnet/workarounds.h>
+
+#endif /* _QSNET_CONFIG_H */
Index: linux-2.4.21/include/qsnet/crwlock.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/crwlock.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/crwlock.h        2005-06-01 23:12:54.751416000 -0400
@@ -0,0 +1,207 @@
+/* 
+ *    Copyright (C) 2000  Regents of the University of California
+ *
+ *    This program is free software; you can redistribute it and/or modify
+ *    it under the terms of the GNU General Public License as published by
+ *    the Free Software Foundation; either version 2 of the License, or
+ *    (at your option) any later version.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    GNU General Public License for more details.
+ *
+ *    You should have received a copy of the GNU General Public License
+ *    along with this program; if not, write to the Free Software
+ *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+
+/*
+ *    Complex - Reader/Writer locks
+ *    Ref: "UNIX Systems for Modern Architectures", by Curt Schimmel, 
+ *    sec 11.6.3. 
+ *
+ *    This implementation is based on semaphores and may not be called from 
+ *    interrupt handlers.
+ *
+ */
+
+#if    !defined(_LINUX_RWLOCK_H)
+#define        _LINUX_RWLOCK_H
+
+#if    defined(__KERNEL__)
+
+typedef enum { RD, WRT, ANY } crwlock_type_t;
+
+#define crwlock_write_held(l) debug_crwlock_held(l, WRT, __BASE_FILE__,__LINE__)
+#define crwlock_read_held(l) debug_crwlock_held(l, RD, __BASE_FILE__, __LINE__)
+#define crwlock_held(l)      debug_crwlock_held(l, ANY, __BASE_FILE__, __LINE__)
+
+#define crwlock_read(l)             debug_crwlock_read(l, __BASE_FILE__, __LINE__)
+#define crwlock_write(l)     debug_crwlock_write(l, __BASE_FILE__, __LINE__)
+#define crwlock_done(l)      debug_crwlock_done(l, __BASE_FILE__, __LINE__)
+
+#if     defined(DEBUG_RWLOCK) && defined(__alpha__) && !defined(DEBUG_SPINLOCK)
+#define DEBUG_SPINLOCK
+#endif
+
+#include <linux/spinlock.h>
+#include <asm/semaphore.h>
+#include <qsnet/debug.h>
+#include <qsnet/mutex.h>
+#include <linux/version.h>
+
+#if    !defined(DEBUG_SPINLOCK)
+#define debug_spin_lock(lock, file, line)       spin_lock(lock)
+#endif
+
+typedef struct {
+        spinlock_t             m_lock;         /* protects cnt fields below */
+        int                     m_rdcnt;        /* # of rdrs in crit section */
+        int                     m_wrcnt;        /* # of wrtrs in crit section */
+        int                     m_rdwcnt;       /* # of waiting readers */
+        int                     m_wrwcnt;       /* # of waiting writers */
+        struct semaphore        m_rdwait;       /* sema where readers wait */
+        struct semaphore        m_wrwait;       /* sema where writers wait */
+        pid_t                  m_wrholder;     /* task holding write lock */
+} crwlock_t;
+ 
+extern __inline__ void 
+crwlock_init(crwlock_t *l)
+{
+       l->m_lock = SPIN_LOCK_UNLOCKED;
+#if    LINUX_VERSION_CODE < KERNEL_VERSION(2,3,0)
+       l->m_rdwait = MUTEX_LOCKED;
+       l->m_wrwait = MUTEX_LOCKED;
+#else
+       sema_init(&l->m_rdwait,0);
+       sema_init(&l->m_wrwait,0);
+#endif
+       l->m_rdcnt = l->m_wrcnt = l->m_rdwcnt = l->m_wrwcnt = 0;
+       l->m_wrholder = PID_NONE;
+}
+
+extern __inline__ void 
+crwlock_destroy(crwlock_t *l)
+{
+       ASSERT(l->m_rdcnt == 0 && l->m_wrcnt == 0);
+}
+
+/*
+ * If a writer has the lock presently or there are writers waiting,
+ * then we have to wait.
+ */
+extern __inline__ void 
+debug_crwlock_read(crwlock_t *l, char *file, int line)
+{
+       ASSERT(!in_interrupt());
+       spin_lock(&l->m_lock);
+       if (l->m_wrcnt || l->m_wrwcnt) {
+               l->m_rdwcnt++;
+               spin_unlock(&l->m_lock);
+               down(&l->m_rdwait); /* P */
+       } else {
+               l->m_rdcnt++;
+               spin_unlock(&l->m_lock);
+       }
+}
+
+/*
+ * If we're the last reader, and a writer is waiting,
+ * then let the writer go now.
+ */
+/* private */
+extern __inline__ void 
+debug_crwlock_read_done(crwlock_t *l, char *file, int line)
+{
+       spin_lock(&l->m_lock);
+       l->m_rdcnt--;
+       if (l->m_wrwcnt && l->m_rdcnt == 0) {
+               l->m_wrcnt = 1;
+               l->m_wrwcnt--;
+               spin_unlock(&l->m_lock);
+               up(&l->m_wrwait); /* V */       
+               return;
+       }
+       spin_unlock(&l->m_lock);
+}
+
+extern __inline__ void 
+debug_crwlock_write(crwlock_t *l, char *file, int line)
+{
+       ASSERT(!in_interrupt());
+       spin_lock(&l->m_lock);
+       if (l->m_wrcnt || l->m_rdcnt) {         /* block if lock is in use */
+               l->m_wrwcnt++;
+               spin_unlock(&l->m_lock);
+               down(&l->m_wrwait); /* P */
+       } else {                                /* lock is not in use */
+               l->m_wrcnt = 1;
+               spin_unlock(&l->m_lock);
+       }
+       l->m_wrholder = current->pid;
+}
+
+/* private */
+extern __inline__ void
+debug_crwlock_write_done(crwlock_t *l, char *file, int line)
+{
+       int rdrs;
+
+       spin_lock(&l->m_lock);
+       l->m_wrholder = PID_NONE;
+       if (l->m_rdwcnt) {                      /* let any readers go first */
+               l->m_wrcnt = 0;
+               rdrs = l->m_rdwcnt;
+               l->m_rdcnt = rdrs;
+               l->m_rdwcnt = 0;
+               spin_unlock(&l->m_lock);
+               while (rdrs--)
+                       up(&l->m_rdwait); /* V */
+       } else if (l->m_wrwcnt) {               /* or let any writer go */
+               l->m_wrwcnt--;
+               spin_unlock(&l->m_lock);
+               up(&l->m_wrwait); /* V */
+       } else {                                /* nobody waiting, unlock */
+               l->m_wrcnt = 0;
+               spin_unlock(&l->m_lock);
+       }
+}
+
+extern __inline__ void
+debug_crwlock_done(crwlock_t *l, char *file, int line)
+{
+       if (l->m_wrholder == current->pid)
+               debug_crwlock_write_done(l, file, line);
+       else
+               debug_crwlock_read_done(l, file, line);
+}
+
+/*
+ * Return nonzero if lock is held
+ */
+extern __inline__ int  
+debug_crwlock_held(crwlock_t *l, crwlock_type_t t, char *file, int line)
+{
+       int res;
+
+       spin_lock(&l->m_lock);
+       switch(t) {
+               case RD:
+                       res = l->m_rdcnt;
+                       break;
+               case WRT:
+                       res = l->m_wrcnt;
+                       break;
+               case ANY:
+                       res = l->m_wrcnt + l->m_rdcnt;
+                       break;
+       }
+       spin_unlock(&l->m_lock);
+
+       return res;
+}
+
+#endif /* __KERNEL__ */
+#endif /* _LINUX_RWLOCK_H */
Index: linux-2.4.21/include/qsnet/ctrl_linux.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/ctrl_linux.h        2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/ctrl_linux.h     2005-06-01 23:12:54.751416000 -0400
@@ -0,0 +1,37 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __QSNET_CTRL_LINUX_H
+#define __QSNET_CTRL_LINUX_H
+
+#ident "$Id: ctrl_linux.h,v 1.3 2003/03/26 09:32:03 mike Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/ctrl_linux.h,v $*/
+
+#define QSNETIO_USER_BASE              0x40
+
+#define QSNETIO_DEBUG_DUMP             _IO   ('e', QSNETIO_USER_BASE + 0)
+
+typedef struct qsnetio_debug_buffer_struct
+{
+       caddr_t addr; 
+       size_t  len;
+} QSNETIO_DEBUG_BUFFER_STRUCT;
+#define QSNETIO_DEBUG_BUFFER           _IOWR ('e', QSNETIO_USER_BASE + 1, QSNETIO_DEBUG_BUFFER_STRUCT)
+
+typedef struct qsnetio_debug_kmem_struct
+{
+       void *handle;
+} QSNETIO_DEBUG_KMEM_STRUCT;
+#define QSNETIO_DEBUG_KMEM             _IOWR   ('e', QSNETIO_USER_BASE + 2, QSNETIO_DEBUG_KMEM_STRUCT)
+
+#endif /* __QSNET_CTRL_LINUX_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/debug.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/debug.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/debug.h  2005-06-01 23:12:54.752415848 -0400
@@ -0,0 +1,68 @@
+/*
+ *    Copyright (C) 2000  Regents of the University of California
+ *
+ *    This program is free software; you can redistribute it and/or modify
+ *    it under the terms of the GNU General Public License as published by
+ *    the Free Software Foundation; either version 2 of the License, or
+ *    (at your option) any later version.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    GNU General Public License for more details.
+ *
+ *    You should have received a copy of the GNU General Public License
+ *    along with this program; if not, write to the Free Software
+ *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+#ifndef _QSNET_DEBUG_H
+#define _QSNET_DEBUG_H
+
+#if defined(DIGITAL_UNIX) 
+#include <kern/assert.h>
+#elif defined(LINUX)
+extern int qsnet_assfail (char *ex, const char *func, char *file, int line);
+
+#define ASSERT(EX)     do { \
+       if (!(EX) && qsnet_assfail (#EX, __FUNCTION__, __BASE_FILE__, __LINE__)) { \
+               BUG(); \
+       } \
+} while (0)
+#endif /* DIGITAL_UNIX */
+
+/* debug.c */
+extern void qsnet_debug_init(void);
+extern void qsnet_debug_fini(void);
+extern void qsnet_debug_disable(int);
+extern void qsnet_debug_alloc(void);
+
+#define QSNET_DEBUG_BUFFER  ((unsigned int)(0x01))
+#define QSNET_DEBUG_CONSOLE ((unsigned int)(0x02))
+#define QSNET_DEBUG_BUF_CON ( QSNET_DEBUG_BUFFER | QSNET_DEBUG_CONSOLE )
+
+#ifdef __GNUC__
+extern void qsnet_debugf      (unsigned int mode, char *fmt, ...)
+       __attribute__ ((format (printf,2,3)));
+extern void kqsnet_debugf      (char *fmt, ...)
+       __attribute__ ((format (printf,1,2)));
+#else
+extern void qsnet_debugf      (unsigned int mode, char *fmt, ...);
+extern void kqsnet_debugf     (char *fmt, ...);
+#endif
+extern void qsnet_vdebugf     (unsigned int mode, char * prefix, char *fmt,  va_list ap);
+extern int  qsnet_debug_buffer(caddr_t ubuffer, int len);
+extern int  qsnet_debug_dump  (void);
+extern int  qsnet_debug_kmem  (void *handle);
+
+extern void qsnet_debug_buffer_on(void);
+extern void qsnet_debug_buffer_clear(void);
+extern void qsnet_debug_buffer_mark(char *str);
+
+#endif /* _QSNET_DEBUG_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/fence.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/fence.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/fence.h  2005-06-01 23:12:54.752415848 -0400
@@ -0,0 +1,178 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+/* $Id: fence.h,v 1.21.6.4 2004/11/23 14:34:45 addy Exp $ */
+/*             $Source: /cvs/master/quadrics/qsnet/fence.h,v $*/
+
+#ifndef _CONFIG_FENCE_H
+#define _CONFIG_FENCE_H
+
+#ident "$Id: fence.h,v 1.21.6.4 2004/11/23 14:34:45 addy Exp $"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(__ELAN__) || defined(__ELAN3__)
+
+/* no memory barriers required on elan3/elan4 */
+
+#elif defined QSNET_MEMBARS_ASSERT
+
+#include <assert.h>
+#define MEMBAR_MEMISSUE()       assert(0);
+#define MEMBAR_SYNC()           assert(0);
+#define MEMBAR_STORELOAD()      assert(0);
+#define MEMBAR_LOADSTORE()      assert(0);
+#define MEMBAR_STORESTORE()     assert(0);
+#define MEMBAR_LOADLOAD()       assert(0);
+#define MEMBAR_VISIBLE()        assert(0);
+#define MEMBAR_DRAIN()          assert(0);
+    
+#elif defined(__alpha)
+
+/* Memory barrier instructions */
+#if defined(__DECC) || defined(__DECXX)
+long   asm( const char *,...);
+#pragma intrinsic( asm )
+#define MEMBAR_MEMISSUE()      asm("mb")
+#define MEMBAR_SYNC()          asm("mb")
+#define MEMBAR_STORELOAD()     asm("wmb")
+#define MEMBAR_LOADSTORE()     asm("mb")
+#define MEMBAR_STORESTORE()    asm("wmb")
+#define MEMBAR_LOADLOAD()      asm("mb")
+#define MEMBAR_VISIBLE()       asm("")
+#define MEMBAR_DRAIN()                 asm("wmb")
+
+#else
+/* Assume gcc */
+#define MEMBAR_MEMISSUE()      asm volatile ("mb"::)
+#define MEMBAR_SYNC()          asm volatile ("mb"::)
+#define MEMBAR_STORELOAD()     asm volatile ("wmb"::)
+#define MEMBAR_LOADSTORE()     asm volatile ("mb"::)
+#define MEMBAR_STORESTORE()    asm volatile ("wmb"::)
+#define MEMBAR_LOADLOAD()      asm volatile ("mb"::)
+#define MEMBAR_VISIBLE()       asm volatile (""   ::: "memory")
+#define MEMBAR_DRAIN()         asm volatile ("wmb"::: "memory")
+
+#endif /* __DECC */
+
+#elif defined(__sparc)
+
+/* UltraSPARC with WRITE MERGING enabled */
+#define MEMBAR_MEMISSUE()      asm volatile ("membar #MemIssue");
+#define MEMBAR_SYNC()          asm volatile ("membar #Sync");
+#define MEMBAR_STORELOAD()     asm volatile ("membar #StoreLoad");
+#define MEMBAR_LOADSTORE()     asm volatile ("membar #LoadStore");
+#define MEMBAR_STORESTORE()    asm volatile ("membar #StoreStore");
+#define MEMBAR_LOADLOAD()      asm volatile ("membar #LoadLoad");
+#define MEMBAR_VISIBLE()       asm volatile (""::: "memory")
+#define MEMBAR_DRAIN()         asm volatile (""::: "memory")
+
+#elif defined(__linux__)
+
+#if defined(__INTEL_COMPILER)
+
+/* NB: Intel compiler version 8.0 now also defines __GNUC__ unless you set the -no-gcc cmdline option
+ * I've moved the check for __INTEL_COMPILER to be first to get around this
+ */
+#ifdef __ECC
+
+#include <ia64intrin.h>
+
+#define MEMBAR_MEMISSUE()       __mf()
+#define MEMBAR_SYNC()           __mf()
+#define MEMBAR_STORELOAD()      __mf()
+#define MEMBAR_LOADSTORE()      __mf()
+#define MEMBAR_STORESTORE()     __mf()
+#define MEMBAR_LOADLOAD()       __mf()
+#define MEMBAR_VISIBLE()       __mf()
+#define MEMBAR_DRAIN()         __mf()
+
+#else
+
+#warning Membars not implemented with this compiler.
+#define MEMBAR_MEMISSUE()       ;
+#define MEMBAR_SYNC()           ;
+#define MEMBAR_STORELOAD()      ;
+#define MEMBAR_LOADSTORE()      ;
+#define MEMBAR_STORESTORE()     ;
+#define MEMBAR_LOADLOAD()       ;
+#define MEMBAR_VISIBLE()        ;
+#define MEMBAR_DRAIN()          ;
+
+#endif /* __ECC */
+
+#elif defined(__GNUC__)
+
+#ifndef __ia64
+
+/* These are needed by <asm/system.h> on AMD64 */
+#include <asm/types.h>
+#include <asm/bitops.h>
+
+#ifndef __cplusplus
+/* this header file has a parameter called "new" - great huh */
+#include <asm/system.h>
+#endif
+
+#else
+#  define mb()        __asm__ __volatile__ ("mf" ::: "memory")
+#  define rmb()       mb()
+#  define wmb()       mb()
+#endif /* !__ia64 */
+
+#if defined(__x86_64) || defined(__i386)
+/* For some reason the AMD64 definition (glibc-devel 2.3.X) of this 
+ * is not useful (compiler only directive) so we overload it here
+ */
+/* I don't trust the IA32 header files either as with mtrr enabled
+ * we really need a membar and not a compiler directive
+ * NB: sfence is only available with X86_FEATURE_XMM CPUs
+ */
+#undef wmb
+#define wmb()    asm volatile("sfence":::"memory");
+#endif /* __x86_64 */
+
+#define MEMBAR_MEMISSUE()      mb()
+#define MEMBAR_SYNC()          mb()
+#define MEMBAR_STORELOAD()     wmb()
+#define MEMBAR_LOADSTORE()     mb()
+#define MEMBAR_STORESTORE()    wmb()
+#define MEMBAR_LOADLOAD()      mb()
+
+#ifdef __ia64
+#define MEMBAR_VISIBLE()       asm volatile ("mf.a;;mf;;"::: "memory")
+#define MEMBAR_DRAIN()         asm volatile ("mf;"::: "memory")
+#else
+#define MEMBAR_VISIBLE()       asm volatile (""::: "memory")
+#define MEMBAR_DRAIN()         wmb()
+#endif
+
+#else /* elif __GNUC__ */
+
+#error Membars not implemented for this architecture/compiler.
+
+#endif /* __INTEL_COMPILER */
+
+#else /* elif __linux__ */
+
+#error Membars not implemented for this architecture/compiler.
+
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _CONFIG_FENCE_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/kernel.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/kernel.h    2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/kernel.h 2005-06-01 23:12:54.752415848 -0400
@@ -0,0 +1,38 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __QSNET_KERNEL_H
+#define __QSNET_KERNEL_H
+
+#ident "$Id: kernel.h,v 1.8 2003/03/14 10:18:22 mike Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/kernel.h,v $*/
+
+#include <qsnet/config.h>
+#include <qsnet/types.h>
+
+#if defined(SOLARIS)
+#include <qsnet/kernel_solaris.h>
+#endif
+
+#if defined(DIGITAL_UNIX)
+#include <qsnet/kernel_dunix.h>
+#endif
+
+#if defined(LINUX)
+#include <qsnet/kernel_linux.h>
+#endif
+
+#include <qsnet/debug.h>
+
+#endif /* __QSNET_KERNEL_H */
+
+
+
+
+
+
+
Index: linux-2.4.21/include/qsnet/kernel_linux.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/kernel_linux.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/kernel_linux.h   2005-06-01 23:12:54.753415696 -0400
@@ -0,0 +1,354 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __QSNET_KERNEL_LINUX_H
+#define __QSNET_KERNEL_LINUX_H
+
+#ident "$Id: kernel_linux.h,v 1.62.6.5 2005/01/18 14:37:22 david Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/kernel_linux.h,v $*/
+
+#if defined(MODVERSIONS)
+#include <linux/modversions.h>
+#endif
+
+#include <linux/autoconf.h>
+#include <linux/module.h>
+
+
+/* ASSERT(spin_is_locked(l)) would always fail on UP kernels */
+#if defined(CONFIG_SMP)
+#define SPINLOCK_HELD(l)       spin_is_locked(l)
+#else
+#define SPINLOCK_HELD(l)       (1) 
+#endif
+
+#include <asm/io.h>
+#include <asm/uaccess.h>
+
+#include <linux/types.h>
+#include <linux/time.h>
+
+#include <linux/delay.h>
+#include <linux/smp_lock.h>
+#include <linux/spinlock.h>
+#include <linux/module.h>
+
+#include <linux/coproc.h>      /* Quadrics added */
+
+#include <linux/highmem.h>
+
+#include <qsnet/mutex.h>
+#include <qsnet/condvar.h>
+#include <qsnet/crwlock.h>
+
+#if defined(LINUX_ALPHA)
+#  include <asm/core_tsunami.h>        /* for TSUNAMI_MEM */
+#endif
+
+#if LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)
+#      undef   MOD_INC_USE_COUNT
+#      undef   MOD_DEC_USE_COUNT
+#      define  MOD_INC_USE_COUNT
+#      define  MOD_DEC_USE_COUNT
+#endif
+
+#define MIN(a,b)       ((a) > (b) ? (b) : (a))
+#define MAX(a,b)       ((a) > (b) ? (a) : (b))
+
+/* stray types */
+typedef u64               u_longlong_t;
+typedef unsigned long     uintptr_t;
+typedef int               bool_t;
+
+typedef unsigned long     virtaddr_t;                          /* virtual address */
+typedef unsigned long      ioaddr_t;                           /* io address */
+typedef unsigned long      sdramaddr_t;                                /* elan sdram offset */
+
+/* 386 kernel can be compiled with PAE enabled to use a 44 bit physical address */
+#if defined(CONFIG_X86_PAE)
+typedef unsigned long long physaddr_t;
+#else
+typedef unsigned long     physaddr_t;
+#endif
+
+/* ticks since reboot, and tick freq */
+#define lbolt          jiffies 
+#define hz             HZ
+
+/* System page size and friends */
+#define PAGESIZE       PAGE_SIZE
+#define PAGESHIFT      PAGE_SHIFT
+#define PAGEOFFSET     (PAGE_SIZE - 1)
+#define PAGEMASK       PAGE_MASK
+
+#define PAGE_ALIGNED(a)        (((a) & PAGE_MASK) == a)
+
+/* convert between bytes and pages */
+#define btop(b)         ((unsigned long)(b) >> PAGE_SHIFT)     /* rnd down */ 
+#define btopr(b)        btop(PAGE_ALIGN((unsigned long) b))    /* rnd up */
+#define ptob(p)                ((unsigned long)(p) << PAGE_SHIFT)
+
+/* round up sz to the nearest multiple of blk */
+#define roundup(sz,blk) ((blk) * ((sz) / (blk) + ((sz) % (blk) ? 1 : 0)))      
+
+/* send a signal to a process */
+#define psignal(pr,sig)        send_sig(sig,pr,0)
+
+/* microsecond delay */
+#define DELAY(us)      udelay(us)
+
+/* macro macros */
+#define MACRO_BEGIN     do {
+#define MACRO_END       } while (0)
+
+/* D-Unix compatable errno values */
+#define ESUCCESS        0
+#define EFAIL           255
+
+/* ASSERT(NO_LOCKS_HELD) will be a no-op */
+#define NO_LOCKS_HELD  1
+
+/* misc */
+typedef int            label_t;
+#define on_fault(ljp)  ((ljp) == NULL)
+#define _NOTE(X)
+#define no_fault()     ((void) 0)
+#define panicstr       0
+
+/* return from system call is -EXXX on linux */
+#define set_errno(e)   (-(e))
+
+/* 
+ * BSD-style byte ops 
+ */
+
+#define bcmp(src1,src2,len)            memcmp(src1,src2,len)
+#define bzero(dst,len)                 memset(dst,0,len)
+#define bcopy(src,dst,len)             memcpy(dst,src,len)
+
+#define preemptable_start              do { long must_yield_at = lbolt + (hz/10);
+#define preemptable_end                        } while (0)
+#define preemptable_check()            do {\
+                                            if ((lbolt - must_yield_at) > 0)\
+                                           {\
+                                               preemptable_yield() ; \
+                                               must_yield_at = lbolt + (hz/10);\
+                                           }\
+                                       } while (0)
+
+#define preemptable_yield()            schedule()
+
+#define CURPROC()                       current
+#define CURTHREAD()                     current
+#define SUSER()                                suser()
+
+/* 64 bit IO operations on 32 bit intel cpus using MMX */
+#if defined(LINUX_I386)
+extern u64         qsnet_readq (volatile u64 *ptr);
+extern void        qsnet_writeq (u64 value, volatile u64 *ptr);
+
+#define readq(ptr)             qsnet_readq((void *) ptr)
+#define writeq(val,ptr)                qsnet_writeq(val, (void *)ptr)
+#endif
+
+/*
+ * Memory barriers
+ */
+#ifndef mmiob
+#  define mmiob()                      mb()
+#endif
+
+/* 
+ * Exit handlers
+ */
+#define HANDLER_REGISTER(func,arg,flags)   xa_handler_register(func,arg,flags)
+#define HANDLER_UNREGISTER(func,arg,flags) xa_handler_unregister(func,arg,flags)
+
+/* 
+ * KMEM_GETPAGES and KMEM_ALLOC both call kmem_alloc, which 
+ * translates the call to kmalloc if < PAGE_SIZE, or vmalloc 
+ * if >= PAGE_SIZE.  vmalloc will always return a page-aligned 
+ * region rounded up to the nearest page, while kmalloc will 
+ * return bits and pieces of a page.
+ */
+
+#ifdef KMEM_DEBUG
+extern void          *qsnet_kmem_alloc_debug(int len, int sleep, int zerofill, char *file, int line);
+extern void           qsnet_kmem_free_debug(void *ptr, int len, char *file, int line);
+#define KMEM_ALLOC(ptr,type,len,sleep) \
+       { KMEM_ASSERT(sleep); (ptr)=(type)qsnet_kmem_alloc_debug(len,sleep,0,__FILE__,__LINE__); }
+#define KMEM_ZALLOC(ptr,type,len,sleep) \
+       { KMEM_ASSERT(sleep); (ptr)=(type)qsnet_kmem_alloc_debug(len,sleep,1,__FILE__,__LINE__); }
+
+#define KMEM_FREE(ptr,len)               qsnet_kmem_free_debug((void *)ptr,len,__FILE__,__LINE__)
+
+#else
+
+extern void          *qsnet_kmem_alloc(int len, int sleep, int zerofill);
+extern void           qsnet_kmem_free(void *ptr, int len);
+
+#define KMEM_ALLOC(ptr,type,len,sleep) \
+       { KMEM_ASSERT(sleep); (ptr)=(type)qsnet_kmem_alloc(len,sleep,0); }
+#define KMEM_ZALLOC(ptr,type,len,sleep) \
+       { KMEM_ASSERT(sleep); (ptr)=(type)qsnet_kmem_alloc(len,sleep,1); }
+
+#define KMEM_FREE(ptr,len)               qsnet_kmem_free((void *)ptr,len)
+
+#endif
+extern void       qsnet_kmem_display(void *handle);
+extern physaddr_t kmem_to_phys(void *ptr);
+
+#define KMEM_ASSERT(sleep)              ASSERT(!(in_interrupt() && sleep))
+
+
+#define KMEM_GETPAGES(ptr,type,pgs,sleep) KMEM_ZALLOC(ptr,type,ptob(pgs),sleep)
+#define KMEM_FREEPAGES(ptr,pgs)          KMEM_FREE(ptr,ptob(pgs));
+
+/*
+ * Copying from user space -> kernel space (perms checked)
+ */
+#define copyin(up,kp,size)             copy_from_user(kp,up,size)
+#define copyin_noerr(up,kp,size)       copy_from_user(kp,up,size)
+
+/* get_user() gets xfer width right */
+#define fulinux(ret, up)               (get_user(ret, (up)) == 0 ? ret : -1)
+#define fulinuxp(ret, up)              (get_user(ret, (up)) == 0 ? ret : NULL)
+
+extern __inline__ int fubyte    (u8  *up) { u8  ret;   return fulinux(ret, up);}
+extern __inline__ int fusword   (u16 *up) { u16 ret;   return fulinux(ret, up);}
+extern __inline__ int fuword    (u32 *up) { u32 ret;   return fulinux(ret, up);}
+#if BITS_PER_LONG > 32
+extern __inline__ u64 fulonglong(u64 *up) { u64 ret;   return fulinux(ret, up);}
+#else
+extern __inline__ u64 fulonglong(u64 *up) { return ((u64) fuword((u32 *)up) | (((u64) fuword(((u32 *)up)+1))<<32)); }
+#endif
+extern __inline__ void *fuptr (void **up) { void *ret; return fulinuxp(ret,up);}
+
+#define fubyte_noerr(up)               fubyte(up)
+#define fusword_noerr(up)              fusword(up)
+#define fuword_noerr(up)               fuword(up)
+#define fulonglong_noerr(up)           fulonglong(up)
+#define fuptr_noerr(up)                        fuptr(up)
+
+extern __inline__ int copyinstr(char *up, char *kp, int max, int *size)
+{ 
+       for (*size = 1; *size <= max; (*size)++) {
+               if (get_user(*kp, up++) != 0)
+                       return EFAULT;  /* bad user space addr */
+               if (*kp++ == '\0')
+                       return 0;       /* success */
+       }
+       *size = max;
+       return ENAMETOOLONG;            /* runaway string */
+}
+ 
+/*
+ * Copying from kernel space -> user space (perms checked)
+ */
+
+#define copyout(kp,up,size)            copy_to_user(up,kp,size)
+#define copyout_noerr(kp,up,size)      copy_to_user(up,kp,size)
+
+/* put_user() gets xfer width right */
+#define sulinux(val, up)               (put_user(val, (up)) == 0 ? 0 : -1)
+
+extern __inline__ int subyte    (u8  *up, u8  val) { return sulinux(val, up); }
+extern __inline__ int susword   (u16 *up, u16 val) { return sulinux(val, up); }
+extern __inline__ int suword    (u32 *up, u32 val) { return sulinux(val, up); }
+#if BITS_PER_LONG > 32
+extern __inline__ int sulonglong(u64 *up, u64 val) { return sulinux(val, up); }
+#else
+extern __inline__ int sulonglong(u64 *up, u64 val) { return (suword((u32 *) up, (u32) val) == 0 ? 
+                                                            suword(((u32 *) up)+1, (u32) (val >> 32)) : -1); }
+#endif
+extern __inline__ int suptr   (void **up,void *val){ return sulinux(val, up); }
+
+#define subyte_noerr(up,val)           subyte(up,val)  
+#define susword_noerr(up,val)          susword(up,val) 
+#define suword_noerr(up,val)           suword(up,val)  
+#define sulonglong_noerr(up,val)       sulonglong(up,val)      
+#define suptr_noerr(up,val)            suptr(up,val)   
+
+/*
+ * /proc/qsnet interface
+ */
+extern inline int
+str_append(char *buf, char *add, int size)
+{
+#define TRUNC_MSG       "[Output truncated]\n"
+       int full = 0;
+       int max = size - strlen(TRUNC_MSG) - strlen(add) - 1;
+
+       if (strlen(buf) > max) {
+               strcat(buf, TRUNC_MSG);
+               full = 1;
+       } else
+               strcat(buf, add);
+       return full;
+}
+
+/* Spinlocks */
+#define spin_lock_destroy(l)           ((void) 0)
+
+/* Complex - Reader/Writer locks - we added <linux/crwlock.h> */
+typedef crwlock_t                      krwlock_t;
+#define krwlock_init(l)                        crwlock_init(l)
+#define krwlock_destroy(l)             crwlock_destroy(l)
+#define krwlock_write(l)               crwlock_write(l)
+#define krwlock_read(l)                        crwlock_read(l)
+#define krwlock_done(l)                        crwlock_done(l)
+#define krwlock_is_locked(l)           crwlock_held(l)
+#define krwlock_is_write_locked(l)     crwlock_write_held(l)
+#define krwlock_is_read_locked(l)      crwlock_read_held(l)
+
+/*
+ * Timeouts - Solaris style.
+ */
+typedef struct timer_list timer_fn_t;
+
+extern inline void
+schedule_timer_fn(timer_fn_t *timer, void (*fun)(void *), void *arg, long hz_delay)
+{
+       init_timer(timer);
+
+       timer->function = (void (*)(unsigned long)) fun;
+       timer->data     = (unsigned long) arg;
+       timer->expires  = jiffies + hz_delay;
+
+       add_timer(timer);
+}
+
+/* returns 1 if timer_fn was cancelled */
+extern inline int
+cancel_timer_fn(timer_fn_t *timer)
+{
+    return (del_timer_sync(timer));
+}
+
+extern inline int
+timer_fn_queued(timer_fn_t *timer)
+{
+    return (timer_pending (timer));
+}
+/*
+ * Hold/release CPU's.
+ */
+
+extern void    cpu_hold_all(void);
+extern void    cpu_release_all(void);
+#define CAPTURE_CPUS()         cpu_hold_all()
+#define RELEASE_CPUS()         cpu_release_all()
+
+#define IASSERT ASSERT
+
+#endif /* __QSNET_KERNEL_LINUX_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/kpte.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/kpte.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/kpte.h   2005-06-01 23:12:54.753415696 -0400
@@ -0,0 +1,107 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2004 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __QSNET_KPTE_H
+#define __QSNET_KPTE_H
+
+#ident "@(#)$Id: kpte.h,v 1.1.2.1 2004/11/02 10:45:29 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/qsnet/kpte.h,v $*/
+
+#include <qsnet/autoconf.h>
+
+#ifdef NO_RMAP
+#      define pte_offset_kernel pte_offset
+#      define pte_offset_map    pte_offset
+#       define pte_unmap(A)      do { ; } while (0)
+#endif
+
+/* 
+ * Pte stuff
+ */
+static __inline__ struct mm_struct *
+get_kern_mm(void)
+{
+        return &init_mm;
+}
+
+static __inline__ pte_t *
+find_pte_map(struct mm_struct *mm, unsigned long vaddr)
+{
+        pgd_t *pgd;
+        pmd_t *pmd;
+       pte_t *ptep;
+
+/* XXXX - handle hugh tlb code */
+       pgd = pgd_offset(mm, vaddr);
+       if (pgd_none(*pgd) || pgd_bad(*pgd))
+               goto out;
+       
+       pmd = pmd_offset(pgd, vaddr);
+       if (pmd_none(*pmd) || pmd_bad (*pmd))
+               goto out;
+
+       ptep = pte_offset_map (pmd, vaddr);
+       if (! ptep)
+               goto out;
+       
+       if (pte_present (*ptep))
+               return ptep;
+
+       pte_unmap (ptep);
+out:
+       return NULL;
+}
+
+static __inline__ pte_t *
+find_pte_kernel(unsigned long vaddr)
+{
+        pgd_t *pgd;
+        pmd_t *pmd;
+       pte_t *pte;
+
+       pgd = pgd_offset_k(vaddr);
+       if (pgd && !pgd_none(*pgd)) {
+           pmd = pmd_offset(pgd, vaddr);
+           if (pmd && pmd_present(*pmd)) {
+               pte = pte_offset_kernel(pmd, vaddr);
+               if (pte && pte_present(*pte))
+                   return (pte);
+           }
+       }
+       return (NULL);
+}
+
+static __inline__ physaddr_t
+pte_phys(pte_t pte)
+{
+#if defined(LINUX_ALPHA)
+       /* RedHat 7.1 2.4.3-12 
+        * They have now enabled Monster windows on Tsunami
+        * and so can use the Main's phys pte value 
+        */
+       return (pte_val(pte) >> (32-PAGE_SHIFT));
+#elif defined(LINUX_I386)
+       return (pte_val(pte) & ~((1 << PAGE_SHIFT)-1));
+#elif defined(LINUX_SPARC)
+       return (pte_val(pte) & _PAGE_PADDR);
+#elif defined(LINUX_IA64)
+       return (pte_val(pte) & _PFN_MASK);
+#elif defined(LINUX_X86_64)
+       return (pte_val(pte) & ~((1 << PAGE_SHIFT)-1) & ~_PAGE_NX);
+#else
+#error Unknown architecture
+#endif
+}
+
+#endif /* __QSNET_KPTE_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/kthread.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/kthread.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/kthread.h        2005-06-01 23:12:54.754415544 -0400
@@ -0,0 +1,71 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *    Copyright (c) 2002-2004 by Quadrics Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __QSNET_KTHREAD_H
+#define __QSNET_KTHREAD_H
+
+#ident "@(#)$Id: kthread.h,v 1.1 2004/10/28 11:50:29 david Exp $ $Name: QSNETMODULES-4-30_20050128 $"
+/*      $Source: /cvs/master/quadrics/qsnet/kthread.h,v $*/
+
+#include <qsnet/autoconf.h>
+
+/* 
+ * kernel threads 
+ */
+extern __inline__ void
+kernel_thread_init(char *comm)
+{
+#if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
+#ifndef NO_NPTL
+#      define sigmask_lock                     sighand->siglock
+#endif
+       lock_kernel();
+       daemonize();
+        reparent_to_init();
+
+        /* avoid getting signals */
+        spin_lock_irq(&current->sigmask_lock);
+        flush_signals(current);
+        sigfillset(&current->blocked);
+       
+#ifdef NO_NPTL
+        recalc_sigpending(current);
+#else
+        recalc_sigpending();
+#endif
+
+        spin_unlock_irq(&current->sigmask_lock);
+
+       /* set our name for identification purposes */
+       strncpy(current->comm, comm, sizeof(current->comm));
+
+       unlock_kernel();
+#else
+       daemonize(comm);
+#endif
+}
+
+extern __inline__ void *
+kernel_thread_wrap(caddr_t stk, int stksize, void (*proc)(void *), void *arg)
+{
+        ASSERT(stk == NULL && stksize == 0);
+        kernel_thread((int (*)(void *))proc, arg, CLONE_FS | CLONE_FILES | CLONE_SIGHAND);
+        return (void *)1; /* non-null value */
+}
+
+#define kernel_thread_create(proc,arg)  kernel_thread_wrap(NULL,0,(void (*)(void *))proc,arg)
+#define kernel_thread_exit()           ((void) 0)
+#define kernel_thread_become_highpri() ((void) 0)
+
+#endif /* __QSNET_KTHREAD_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/list.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/list.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/list.h   2005-06-01 23:12:54.754415544 -0400
@@ -0,0 +1,80 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Limited.
+ * 
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: list.h,v 1.5 2003/10/27 13:55:33 david Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/list.h,v $*/
+
+#ifndef __QSNET_LIST_H
+#define __QSNET_LIST_H
+
+/* Implementation of doubly linked lists - compatible with linux */
+struct list_head 
+{
+    struct list_head *next;
+    struct list_head *prev;
+};
+
+#if !defined(LINUX)
+#if ! defined( offsetof ) 
+#define offsetof(T,F) ((int )&(((T *)0)->F))
+#endif
+
+#define LIST_HEAD_INIT(name) { &(name), &(name) }
+
+#define LIST_HEAD(name) \
+       struct list_head name = LIST_HEAD_INIT(name)
+#endif
+
+#define list_entry(ptr, type, off) \
+       ((type *) ((unsigned long)(ptr) - offsetof (type,off)))
+
+#define INIT_LIST_HEAD(list) \
+MACRO_BEGIN \
+       (list)->next = (list)->prev = (list); \
+MACRO_END
+
+#define list_add(new, list) \
+MACRO_BEGIN \
+       (list)->next->prev = (new); \
+       (new)->next = (list)->next; \
+       (new)->prev = (list); \
+       (list)->next = (new); \
+MACRO_END
+
+#define list_add_tail(new, list) \
+MACRO_BEGIN \
+       (list)->prev->next = new; \
+       (new)->prev = (list)->prev; \
+       (new)->next = (list); \
+       (list)->prev = (new); \
+MACRO_END
+
+#define list_del(entry) \
+MACRO_BEGIN \
+       (entry)->prev->next = (entry)->next; \
+       (entry)->next->prev = (entry)->prev; \
+MACRO_END
+
+#define list_del_init(entry) \
+MACRO_BEGIN \
+       (entry)->prev->next = (entry)->next; \
+       (entry)->next->prev = (entry)->prev; \
+       (entry)->next = (entry)->prev = (entry); \
+MACRO_END
+
+#define list_empty(list) \
+       ((list)->next == (list))
+
+#define list_for_each(pos,list) \
+       for (pos = (list)->next; pos != (list); \
+            pos = (pos)->next)
+
+#define list_for_each_safe(pos,n,list) \
+       for (pos = (list)->next, n = (pos)->next; pos != (list); \
+            pos = n, n = (pos)->next)
+
+#endif /* __QSNET_LIST_H */
Index: linux-2.4.21/include/qsnet/mutex.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/mutex.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/mutex.h  2005-06-01 23:12:54.754415544 -0400
@@ -0,0 +1,91 @@
+/*
+ *    Copyright (C) 2000  Regents of the University of California
+ *
+ *    This program is free software; you can redistribute it and/or modify
+ *    it under the terms of the GNU General Public License as published by
+ *    the Free Software Foundation; either version 2 of the License, or
+ *    (at your option) any later version.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    GNU General Public License for more details.
+ *
+ *    You should have received a copy of the GNU General Public License
+ *    along with this program; if not, write to the Free Software
+ *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+
+#if    !defined(_LINUX_MUTEX_H)
+#define        _LINUX_MUTEX_H
+#if    defined(__KERNEL__)
+
+#include <asm/smp.h>
+#include <linux/spinlock.h>
+#include <asm/semaphore.h>
+#include <qsnet/debug.h>
+#include <linux/interrupt.h>
+#include <linux/version.h>
+
+#define PID_NONE       0
+
+typedef struct
+{
+    struct semaphore sem;
+    pid_t           holder;
+} kmutex_t;
+
+extern __inline__ void
+kmutex_init (kmutex_t *l)
+{
+#if    LINUX_VERSION_CODE < KERNEL_VERSION(2,3,0)
+    l->sem = MUTEX;
+#else
+    init_MUTEX(&l->sem);
+#endif
+    l->holder = PID_NONE;
+}
+
+extern __inline__ void
+kmutex_destroy (kmutex_t *l) 
+{
+    ASSERT (l->holder == PID_NONE); 
+}
+
+extern __inline__ void
+kmutex_lock (kmutex_t *l) 
+{ 
+    ASSERT(l->holder != current->pid);
+    down (&l->sem);
+    l->holder = current->pid; 
+}
+
+extern __inline__ void
+kmutex_unlock (kmutex_t *l) 
+{
+    ASSERT(l->holder == current->pid);
+
+    l->holder = PID_NONE;
+    up (&l->sem);
+}
+
+extern __inline__ int
+kmutex_trylock (kmutex_t *l) 
+{
+    if (down_trylock (&l->sem) == 0) 
+    {
+       l->holder = current->pid;
+       return (1);
+    }
+    return (0);
+}
+
+extern __inline__ int
+kmutex_is_locked (kmutex_t *l) 
+{
+    return (l->holder == current->pid);
+}
+
+#endif /* __KERNEL__ */
+#endif /* _LINUX_MUTEX_H */
Index: linux-2.4.21/include/qsnet/procfs_linux.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/procfs_linux.h      2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/procfs_linux.h   2005-06-01 23:12:54.755415392 -0400
@@ -0,0 +1,234 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __PROCFS_LINUX_H
+#define __PROCFS_LINUX_H
+
+#ident "$Id: procfs_linux.h,v 1.6.2.6 2004/12/06 17:36:24 robin Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/procfs_linux.h,v $ */
+
+#if defined(__KERNEL__)
+
+#include <qsnet/kernel_linux.h>
+#include <qsnet/autoconf.h>
+#include <linux/proc_fs.h>
+
+extern gid_t qsnet_procfs_gid;
+
+/* borrowed from fs/proc/proc_misc - helper for proc_read_int */
+static inline int 
+qsnet_proc_calc_metrics(char *page, char **start, off_t off, int count, int *eof, int len)
+{
+       if (len <= off+count) *eof = 1;
+       *start = page + off;
+       len -= off;
+       if (len>count) len = count;
+       if (len<0) len = 0;
+       return len;
+}
+
+static inline int
+qsnet_proc_write_int(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       char tmpbuf[16];
+       int  res = count;
+       
+       if (count > sizeof(tmpbuf) - 1)
+               return (-EINVAL);
+       
+       MOD_INC_USE_COUNT;
+       if (copy_from_user(tmpbuf, buf, count))
+               res = -EFAULT;
+       else
+       {
+               tmpbuf[count] = '\0'; 
+               *(int *)data = simple_strtoul(tmpbuf, NULL, 0);
+       }
+       MOD_DEC_USE_COUNT;
+       
+       return (res);
+}
+
+static inline int
+qsnet_proc_read_int(char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       int len, res;
+       
+       MOD_INC_USE_COUNT;
+       
+       len = sprintf(page, "%d\n", *(int *)data);
+       res = qsnet_proc_calc_metrics(page, start, off, count, eof, len);
+       
+       MOD_DEC_USE_COUNT;
+       return (res);
+}
+
+static inline struct proc_dir_entry *
+qsnet_proc_register_int(struct proc_dir_entry *dir, char *path, int *var, int read_only)
+{
+       struct proc_dir_entry *p;
+       
+       p = create_proc_entry(path, read_only ? S_IRUGO : S_IRUGO|S_IWUSR|S_IWGRP, dir);
+       if (p) {
+               if (! read_only) 
+                       p->write_proc = qsnet_proc_write_int;
+               p->read_proc  = qsnet_proc_read_int;
+               p->data       = var;
+               p->owner      = THIS_MODULE;
+               p->gid        = qsnet_procfs_gid;
+       }
+       return p;
+}
+
+static inline int
+qsnet_proc_write_hex(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       char tmpbuf[16];
+       int  res = count;
+       
+       if (count > sizeof(tmpbuf) - 1)
+               return (-EINVAL);
+       
+       MOD_INC_USE_COUNT;
+       if (copy_from_user(tmpbuf, buf, count))
+               res = -EFAULT;
+       else
+       {
+               tmpbuf[count] = '\0'; 
+               *(int *)data = simple_strtoul(tmpbuf, NULL, 0);
+       }
+       MOD_DEC_USE_COUNT;
+       
+       return (res);
+}
+
+static inline int
+qsnet_proc_read_hex(char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       int len, res;
+       
+       MOD_INC_USE_COUNT;
+       
+       len = sprintf(page, "0x%x\n", *(int *)data);
+       res = qsnet_proc_calc_metrics(page, start, off, count, eof, len);
+       
+       MOD_DEC_USE_COUNT;
+       return (res);
+}
+
+static inline struct proc_dir_entry *
+qsnet_proc_register_hex(struct proc_dir_entry *dir, char *path, int *var, int read_only)
+{
+       struct proc_dir_entry *p;
+       
+       p = create_proc_entry(path, read_only ? S_IRUGO : S_IRUGO|S_IWUSR|S_IWGRP, dir);
+       if (p) {
+               if (! read_only) 
+                       p->write_proc = qsnet_proc_write_hex;
+               p->read_proc  = qsnet_proc_read_hex;
+               p->data       = var;
+               p->owner      = THIS_MODULE;
+               p->gid        = qsnet_procfs_gid;
+       }
+       return p;
+}
+
+#define QSNET_PROC_STR_LEN_MAX ((int)256)
+
+static inline int
+qsnet_proc_write_str(struct file *file, const char *buf, unsigned long count, void *data)
+{
+       int  res = count;
+       
+       if (count > (QSNET_PROC_STR_LEN_MAX - 1))
+               return (-EINVAL);
+       
+       MOD_INC_USE_COUNT;
+       if (copy_from_user((char *)data, buf, count))
+               res = -EFAULT;
+       else
+       {
+               ((char *)data)[count] = '\0'; 
+               /* remove linefeed */
+               if ( (count) && (((char *)data)[count -1] == '\n'))
+                       ((char *)data)[count -1] = '\0';
+       }
+       MOD_DEC_USE_COUNT;
+       
+       return (res);
+}
+
+static inline int
+qsnet_proc_read_str(char *page, char **start, off_t off, int count, int *eof, void *data)
+{
+       int len, res;
+       
+       if ( strlen(data) > (count + 1))
+               return (-EINVAL);       
+
+       MOD_INC_USE_COUNT;
+       
+       /* cant output too much */
+       if ( strlen(data) > (count + 1))
+       {
+               MOD_DEC_USE_COUNT;
+               return (-EINVAL);       
+       }
+
+
+       len = sprintf(page, "%s\n", (char *)data);
+       if (len > count)
+       {
+               MOD_DEC_USE_COUNT;
+               return (-EINVAL);       
+       }
+
+       res = qsnet_proc_calc_metrics(page, start, off, count, eof, len);
+       
+       MOD_DEC_USE_COUNT;
+       return (res);
+}
+
+static inline struct proc_dir_entry *
+qsnet_proc_register_str(struct proc_dir_entry *dir, char *path, char *var, int read_only)
+{
+       struct proc_dir_entry *p;
+       
+       p = create_proc_entry(path, read_only ? S_IRUGO : S_IRUGO|S_IWUSR|S_IWGRP, dir);
+       if (p) {
+               if (! read_only) 
+                       p->write_proc = qsnet_proc_write_str;
+               p->read_proc  = qsnet_proc_read_str;
+               p->data       = var;
+               p->owner      = THIS_MODULE;
+               p->gid        = qsnet_procfs_gid;
+       }
+       return p;
+}
+
+extern struct proc_dir_entry *qsnet_procfs_root; 
+extern struct proc_dir_entry *qsnet_procfs_config;
+
+#ifdef NO_PDE
+static inline struct proc_dir_entry *PDE(const struct inode *inode)
+{
+    return inode->u.generic_ip;
+}
+#endif
+#endif /* __KERNEL__ */
+
+#define QSNET_PROCFS_IOCTL      "/proc/qsnet/ioctl"
+#define QSNET_PROCFS_KMEM_DEBUG "/proc/qsnet/kmem_debug"
+#define QSNET_PROCFS_VERSION    "/proc/qsnet/version"
+
+#endif /* __PROCFS_LINUX_H */
+
+/*
+ * Local variables:
+ * c-file-style: "linux"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/pthread.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/pthread.h   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/pthread.h        2005-06-01 23:12:54.755415392 -0400
@@ -0,0 +1,59 @@
+/*
+ *    Copyright (c) 2003 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+/* $Id: pthread.h,v 1.5 2004/06/07 10:47:06 addy Exp $ */
+/*             $Source: /cvs/master/quadrics/qsnet/pthread.h,v $*/
+
+#ifndef _CONFIG_PTHREAD_H
+#define _CONFIG_PTHREAD_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(__ELAN__)
+
+/* No pthread support on Elan co-processor */
+
+#define MUTEX                   unsigned long long
+#define MUTEX_INIT(X)          ;
+#define        MUTEX_LOCK(X)           ;
+#define        MUTEX_UNLOCK(X)         ;
+
+#else
+#if defined(DIGITAL_UNIX)
+#include <tis.h>
+#define MUTEX                  pthread_mutex_t
+#define MUTEX_INIT(X)          tis_mutex_init(X)
+#define        MUTEX_LOCK(X)           tis_mutex_lock(X)
+#define        MUTEX_UNLOCK(X)         tis_mutex_unlock(X)
+#define        MUTEX_TRYLOCK(X)        (tis_mutex_trylock(X) == 0)
+
+#else /* Linux... */
+
+/* Use standard pthread calls */
+#include <pthread.h>
+#define MUTEX                  pthread_mutex_t
+#define MUTEX_INIT(X)          pthread_mutex_init(X, NULL)
+#define        MUTEX_LOCK(X)           pthread_mutex_lock(X)
+#define        MUTEX_UNLOCK(X)         pthread_mutex_unlock(X)
+#define        MUTEX_TRYLOCK(X)        (pthread_mutex_trylock(X) == 0)
+
+#endif /* DIGITAL_UNIX */
+#endif /* __ELAN__ */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _CONFIG_PTHREAD_H */
+
+/*
+ * Local variables:
+ * c-file-style: "stroustrup"
+ * End:
+ */
Index: linux-2.4.21/include/qsnet/statsformat.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/statsformat.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/statsformat.h    2005-06-01 23:12:54.756415240 -0400
@@ -0,0 +1,25 @@
+#ifndef _QSNET_STATSFORMAT_H
+#define _QSNET_STATSFORMAT_H
+
+#ident "$Id: statsformat.h,v 1.2 2003/05/22 19:37:14 addy Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/statsformat.h,v $*/
+
+#include <qsnet/config.h>
+
+/*
+ * format of an Elan stats record
+ *
+ * type    char(8), type of statistic, e.g. FPAGE, ELAN3, TPORT
+ * time    uint64, 10 digits, time in millisecs since counters initialised
+ * device  uint, 2 digits, Elan device id
+ * name    char(32), name of the statistic
+ * value   uint64, current value of statistic
+ */
+    
+#ifdef _ILP32
+#define ELAN_STATSFORMAT "%-8s %10llu %2d %-32s %llu\n"
+#else
+#define ELAN_STATSFORMAT "%-8s %10lu %2d %-32s %lu\n"
+#endif
+
+#endif
Index: linux-2.4.21/include/qsnet/types.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/types.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/types.h  2005-06-01 23:12:54.756415240 -0400
@@ -0,0 +1,90 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef __QSNET_TYPES_H
+#define __QSNET_TYPES_H
+
+#ident "$Id: types.h,v 1.16 2003/08/01 16:21:38 addy Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/types.h,v $*/
+
+/*
+ * Include typedefs for ISO/IEC 9899:1990 standard types
+ *
+ *
+ *    The following integer typedefs are used:
+ *
+ *     int8_t, int16_t, int32_t, int64_t, intptr_t
+ *     uint8_t, uint16_t, uint32_t, uint64_t, uintptr_t
+ *     uchar_t, ushort_t, uint_t, ulong_t
+ *
+ *    <sys/types.h> also defines the following:
+ *     u_char, u_short, u_int, u_long, caddr_t
+ */
+
+#include <qsnet/config.h>
+
+#if defined(SOLARIS) && defined(__KERNEL__)
+#  include <sys/inttypes.h>
+#endif
+
+#if defined(SOLARIS) && !defined(__KERNEL__)
+#  include <inttypes.h>
+#  include <sys/types.h>
+#endif
+
+#if defined(DIGITAL_UNIX) && defined(__KERNEL__)
+#  include <sys/bitypes.h>
+#endif
+
+#if defined(DIGITAL_UNIX) && !defined(__KERNEL__)
+#  include <inttypes.h>
+#  include <sys/types.h>
+#endif
+
+#if defined(LINUX) && defined(__KERNEL__)
+#  include <linux/types.h>
+#endif
+
+#if defined(LINUX) && !defined(__KERNEL__)
+#  include <stdint.h>
+#  include <inttypes.h>
+#  include <sys/types.h>
+
+typedef unsigned char  uchar_t;
+typedef unsigned short ushort_t;
+typedef unsigned int   uint_t;
+typedef unsigned long  ulong_t;
+#endif
+
+#if defined(QNX)
+#  include <inttypes.h>
+#  include <sys/types.h>
+#endif
+
+/* Define a type that will represent a Main CPU pointer
+ * on both the Main and the Elan
+ */
+#ifdef __ELAN__
+
+#if defined(_MAIN_LP64)
+#define QSNET_MAIN_PTR uint64_t
+#else
+#define QSNET_MAIN_PTR uint32_t
+#endif
+
+#else
+
+#ifdef _LP64
+#define QSNET_MAIN_PTR uint64_t
+#else
+#define QSNET_MAIN_PTR uint32_t
+#endif
+
+#endif
+
+
+#endif /* __QSNET_TYPES_H */
Index: linux-2.4.21/include/qsnet/workarounds.h
===================================================================
--- linux-2.4.21.orig/include/qsnet/workarounds.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/qsnet/workarounds.h    2005-06-01 23:12:54.756415240 -0400
@@ -0,0 +1,24 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ifndef _QSNET_WORKAROUNDS_H
+#define _QSNET_WORKAROUNDS_H
+
+#ident "$Id: workarounds.h,v 1.11 2002/08/09 11:15:55 addy Exp $"
+/*      $Source: /cvs/master/quadrics/qsnet/workarounds.h,v $ */
+
+/* Elan workarounds */
+#undef  ELAN_REVA_SUPPORTED    /* rev a elans no longer supported. */
+#undef  ELITE_REVA_SUPPORTED   /* removed since RMS disables broadcast on rev A elites. */
+#define ELAN_REVB_BUG_1
+/* WORKAROUND for GNAT hw-elan3/3263 */
+#define ELAN_REVB_BUG_2
+
+/* WORKAROUND for GNATs ic-elan3/3637 & ic-elan3/3550 */
+#define ELAN_REVB_BUG_3
+
+#endif /* _QSNET_WORKAROUNDS_H */
Index: linux-2.4.21/include/rms/rmscall.h
===================================================================
--- linux-2.4.21.orig/include/rms/rmscall.h     2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/rms/rmscall.h  2005-06-01 23:12:54.757415088 -0400
@@ -0,0 +1,144 @@
+/*
+ * Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ * For licensing information please see the supplied COPYING file
+ *
+ * rmscall.h:  user interface to rms kernel module
+ *
+ * $Id: rmscall.h,v 1.25 2004/05/14 08:55:57 duncan Exp $
+ * $Source: /cvs/master/quadrics/rmsmod/rmscall.h,v $
+ *
+ */
+
+#ifndef RMSCALL_H_INCLUDED
+#define RMSCALL_H_INCLUDED 1
+
+#ident "$Id: rmscall.h,v 1.25 2004/05/14 08:55:57 duncan Exp $"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * flags for rms_fork_register
+ *
+ * RMS_IOF is not in a public header file 
+ */
+#define RMS_IOF 1              /* inherit on fork */
+
+#ifndef __KERNEL__
+#include <sys/types.h>
+#endif
+
+#include <qsnet/types.h>
+#include <elan/capability.h>
+
+#define MAXCOREPATHLEN 32
+
+#if defined(SOLARIS)
+typedef long long rmstime_t;
+#else  /* DIGITAL_UNIX */
+typedef long rmstime_t;
+#endif
+
+typedef enum {
+    
+    PRG_RUNNING  = 0x01,       /* program is running                  */
+    PRG_ZOMBIE   = 0x02,       /* last process on a node has exited   */
+    PRG_NODE     = 0x04,       /* stats are complete for this node    */
+    PRG_KILLED   = 0x08,       /* program was killed                  */
+    PRG_SUSPEND  = 0x10                /* program is suspended                */
+
+} PRGSTATUS_FLAGS;
+
+/*
+ * program time statistics extended in version 5 of the kernel module
+ */
+typedef struct {
+    rmstime_t etime;           /* elapsed cpu time (milli-secs)       */
+    rmstime_t atime;           /* allocated cpu time (cpu milli-secs) */
+    rmstime_t utime;           /* user cpu time (cpu milli-secs)      */
+    rmstime_t stime;           /* system cpu time (cpu milli-secs)    */
+    int ncpus;                 /* number of cpus allocated            */
+    int flags;                 /* program status flags                */
+    int mem;                   /* max memory size in MBytes           */
+    int pageflts;              /* number of page faults               */
+    rmstime_t memint;          /* memory integral                     */
+} prgstats_old_t;
+
+typedef struct {
+    uint64_t etime;            /* elapsed cpu time (milli-secs)       */
+    uint64_t atime;            /* allocated cpu time (cpu milli-secs) */
+    uint64_t utime;            /* user cpu time (cpu milli-secs)      */
+    uint64_t stime;            /* system cpu time (cpu milli-secs)    */
+    uint64_t pageflts;         /* number of page faults               */
+    uint64_t memint;           /* memory integral                     */
+    uint64_t ebytes;           /* data transferred by the Elan(s)     */
+    uint64_t exfers;           /* number of Elan data transfers       */
+    uint64_t spare64[4];       /* expansion space                     */
+    int ncpus;                 /* number of cpus allocated            */
+    int flags;                 /* program status flags                */
+    int mem;                   /* max memory size in MBytes           */
+    int spare32[5];             /* expansion space                     */
+} prgstats_t;
+
+int  rmsmod_init(void);
+void rmsmod_fini(void);
+
+int rms_setcorepath(caddr_t path);
+int rms_getcorepath(pid_t pid, caddr_t path, int maxlen);
+int rms_prgcreate(int id, uid_t uid, int cpus);
+int rms_prgdestroy(int id);
+int rms_prgids(int maxids, int *prgids, int *nprgs);
+int rms_prginfo(int id, int maxpids, pid_t *pids, int *nprocs);
+int rms_prgaddcap(int id, int index, ELAN_CAPABILITY *cap);
+
+int rms_prgsuspend(int id);
+int rms_prgresume(int id);
+int rms_prgsignal(int id, int signo);
+
+int rms_getprgid(pid_t pid, int *id);
+int rms_ncaps(int *ncaps);
+int rms_getcap(int index, ELAN_CAPABILITY *cap);
+int rms_mycap(int *index);
+int rms_setcap(int index, int ctx);
+int rms_prefcap(int nprocess, int *index);
+
+int   rms_prggetstats(int id, prgstats_t *stats);
+void  rms_accumulatestats(prgstats_t *total, prgstats_t *stats);
+char *rms_statsreport(prgstats_t *stats, char *buf);
+
+int rms_elaninitdone(int vp);
+int rms_prgelanpids(int id, int maxpids, int *vps, pid_t *pids, int *npids);
+int rms_setelanstats(int id, uint64_t ebytes, uint64_t exfers);
+
+int rms_setpset(int psid);
+int rms_getpset(int id, int *psid);
+int rms_modversion();
+
+#ifdef __cplusplus
+}
+#endif
+
+
+#if defined(__KERNEL__)
+
+int rms_init(void);
+int rms_fini(void);
+int rms_reconfigure(void);
+
+extern int rms_debug;
+
+#if 1
+#define DBG(x) do if (rms_debug) x ; while (0)
+#else
+#define DBG(x)
+#endif
+
+#endif
+
+#endif /* RMSCALL_H_INCLUDED */
+
+
+
+
Index: linux-2.4.21/include/rms/rmsio.h
===================================================================
--- linux-2.4.21.orig/include/rms/rmsio.h       2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/include/rms/rmsio.h    2005-06-01 23:12:54.757415088 -0400
@@ -0,0 +1,185 @@
+/*
+ *    Copyright (c) 1996-2002 by Quadrics Supercomputers World Ltd.
+ *
+ *    For licensing information please see the supplied COPYING file
+ *
+ */
+
+#ident "@(#)$Id: rmsio.h,v 1.6 2004/05/14 08:55:57 duncan Exp $"
+/*      $Source: /cvs/master/quadrics/rmsmod/rmsio.h,v $*/
+
+
+#ifndef __RMSMOD_RMSIO_H
+#define __RMSMOD_RMSIO_H
+
+/* arg is corepath string */
+#define RMSIO_SETCOREPATH      _IOW ('r', 1, char)
+
+typedef struct rmsio_getcorepath_struct
+{
+    pid_t              pid;
+    char               *corepath;
+    int                        maxlen;
+} RMSIO_GETCOREPATH_STRUCT;
+#define RMSIO_GETCOREPATH      _IOW ('r', 2, RMSIO_GETCOREPATH_STRUCT)
+
+typedef struct rmsio_prgcreate_struct
+{
+    int                        id;
+    uid_t              uid;
+    int                        cpus;
+} RMSIO_PRGCREATE_STRUCT;
+#define RMSIO_PRGCREATE                _IOW ('r', 3, RMSIO_PRGCREATE_STRUCT)
+
+typedef struct rmsio_prginfo_struct
+{
+    int                        id;
+    int                        maxpids;
+    pid_t              *pids;
+    int                        *nprocs;
+} RMSIO_PRGINFO_STRUCT;
+#define RMSIO_PRGINFO          _IOW ('r', 4, RMSIO_PRGINFO_STRUCT)
+
+typedef struct rmsio_prgsignal_struct
+{
+    int                        id;
+    int                        signo;
+} RMSIO_PRGSIGNAL_STRUCT;
+#define RMSIO_PRGSIGNAL                _IOW ('r', 5, RMSIO_PRGSIGNAL_STRUCT)
+
+typedef struct rmsio_prgaddcap_struct
+{
+    int                        id;
+    int                        index;
+    ELAN_CAPABILITY    *cap;
+} RMSIO_PRGADDCAP_STRUCT;
+#define RMSIO_PRGADDCAP                _IOW ('r', 6, RMSIO_PRGADDCAP_STRUCT)
+typedef struct rmsio_setcap_struct
+{
+    int                        index;
+    int                        ctx;
+} RMSIO_SETCAP_STRUCT;
+#define RMSIO_SETCAP           _IOW ('r', 7, RMSIO_SETCAP_STRUCT)
+
+typedef struct rmsio_getcap_struct
+{
+    int                        index;
+    ELAN_CAPABILITY     *cap;
+} RMSIO_GETCAP_STRUCT;
+#define RMSIO_GETCAP           _IOW ('r', 8, RMSIO_GETCAP_STRUCT)
+
+typedef struct rmsio_getcap_struct32
+{
+    int                        index;
+    unsigned int        capptr;
+} RMSIO_GETCAP_STRUCT32;
+#define RMSIO_GETCAP32         _IOW ('r', 8, RMSIO_GETCAP_STRUCT32)
+
+/* arg is pointer to ncaps */
+#define RMSIO_NCAPS            _IOW ('r', 9, int)
+
+typedef struct rmsio_prggetstats_struct
+{
+    int                        id;
+    prgstats_old_t     *stats;
+} RMSIO_PRGGETSTATS_STRUCT;
+#define RMSIO_PRGGETSTATS      _IOW ('r', 10, RMSIO_PRGGETSTATS_STRUCT)
+
+/* arg is program id */
+#define RMSIO_PRGSUSPEND       _IOW ('r', 11, int)
+#define RMSIO_PRGRESUME                _IOW ('r', 12, int)
+#define RMSIO_PRGDESTROY       _IOW ('r', 13, int)
+
+typedef struct rmsio_getprgid_struct
+{
+    pid_t              pid;
+    int                        *id;
+} RMSIO_GETPRGID_STRUCT;
+#define RMSIO_GETPRGID         _IOW ('r', 14, RMSIO_GETPRGID_STRUCT)
+
+typedef struct rmsio_getprgid_struct32
+{
+    pid_t              pid;
+    unsigned int       idptr;
+} RMSIO_GETPRGID_STRUCT32;
+#define RMSIO_GETPRGID32       _IOW ('r', 14, RMSIO_GETPRGID_STRUCT32)
+
+/* arg is pointer to index */
+#define RMSIO_GETMYCAP         _IOW ('r', 15, int)
+
+typedef struct rmsio_prgids_struct
+{
+    int                        maxids;
+    int                        *prgids;
+    int                        *nprgs;
+} RMSIO_PRGIDS_STRUCT;
+#define RMSIO_PRGIDS           _IOW ('r', 16, RMSIO_PRGIDS_STRUCT)
+
+/* arg is pointer to vp */
+#define RMSIO_ELANINITDONE     _IOW ('r', 17, int)
+
+typedef struct rmsio_prgelanpids_struct
+{
+    int    id;
+    int    maxpids;
+    int   *vps;
+    int   *pids;
+    int   *npids;
+} RMSIO_PRGELANPIDS_STRUCT;
+#define RMSIO_PRGELANPIDS      _IOW ('r', 18, RMSIO_PRGELANPIDS_STRUCT)
+
+typedef struct rmsio_setpset_struct
+{
+    int    id;
+    int    psid;
+} RMSIO_SETPSET_STRUCT;
+#define RMSIO_SETPSET          _IOW ('r', 19, RMSIO_SETPSET_STRUCT)
+
+typedef struct rmsio_getpset_struct
+{
+    int    id;
+    int   *psid;
+} RMSIO_GETPSET_STRUCT;
+#define RMSIO_GETPSET          _IOW ('r', 20, RMSIO_GETPSET_STRUCT)
+
+/*
+ * have to pass a pointer to the stats, the switch
+ * statement goes wrong in the module of the size
+ * is too large
+ */
+typedef struct {
+    uint64_t ebytes;
+    uint64_t exfers;
+} elanstats_t;
+
+typedef struct rmsio_setelanstats_struct
+{
+    int    id;
+    elanstats_t *estats;
+} RMSIO_SETELANSTATS_STRUCT;
+#define RMSIO_SETELANSTATS      _IOW ('r', 21, RMSIO_SETELANSTATS_STRUCT)
+
+typedef struct rmsio_prggetstats2_struct
+{
+    int                        id;
+    prgstats_t         *stats;
+} RMSIO_PRGGETSTATS2_STRUCT;
+#define RMSIO_PRGGETSTATS2     _IOW ('r', 22, RMSIO_PRGGETSTATS2_STRUCT)
+
+typedef struct rmsio_modversion_struct
+{
+    int *version;
+} RMSIO_MODVERSION_STRUCT;
+#define RMSIO_MODVERSION       _IOW ('r', 23, RMSIO_MODVERSION_STRUCT)
+
+
+#endif /* __RMSMOD_RMSIO_H */
+
+
+
+
+
+
+
+
+
Index: linux-2.4.21/ipc/shm.c
===================================================================
--- linux-2.4.21.orig/ipc/shm.c 2005-06-01 22:51:50.000000000 -0400
+++ linux-2.4.21/ipc/shm.c      2005-06-01 23:12:54.758414936 -0400
@@ -723,6 +723,44 @@
        return retval;
 }
 
+/*
+ * Mark all segments created by this process for destruction
+ */
+asmlinkage int shm_cleanup ()
+{
+       int i;
+
+       down(&shm_ids.sem);
+
+       for(i = 0; i <= shm_ids.max_id; i++) {
+               struct shmid_kernel* shp;
+
+               shp = shm_lock(i);
+               if(shp!=NULL) {
+
+                   /* Mark this segment for destruction if we created it */
+                   if (current->pid == shp->shm_cprid)
+                   {
+                       /* Copy of IPC_RMID code */
+                       if (shp->shm_nattch){
+                               shp->shm_flags |= SHM_DEST;
+                               /* Do not find it any more */
+                               shp->shm_perm.key = IPC_PRIVATE;
+                       } else {
+                               shm_destroy(shp);
+                               continue;
+                       }
+                   }
+
+                   shm_unlock(i);
+               }
+       }
+
+       up(&shm_ids.sem);
+
+       return 0;
+}
+
 #ifdef CONFIG_PROC_FS
 static int sysvipc_shm_read_proc(char *buffer, char **start, off_t offset, int length, int *eof, void *data)
 {
Index: linux-2.4.21/kernel/exit.c
===================================================================
--- linux-2.4.21.orig/kernel/exit.c     2005-06-01 22:58:09.055062312 -0400
+++ linux-2.4.21/kernel/exit.c  2005-06-01 23:12:54.759414784 -0400
@@ -19,6 +19,7 @@
 #include <linux/file.h>
 #include <linux/binfmts.h>
 #include <linux/ptrace.h>
+#include <linux/ptrack.h>
 #include <linux/mount.h>
 #include <linux/process_timing.h>
 #include <asm/uaccess.h>
@@ -705,6 +706,10 @@
        if (current->tux_info)
                current->tux_exit();
        acct_process(code);
+
+       /* Notify any ptrack callbacks of the process exit */
+       ptrack_call_callbacks(PTRACK_PHASE_EXIT, NULL);
+
        if (isaudit(tsk))
                audit_exit(tsk, code);
        __exit_mm(tsk);
Index: linux-2.4.21/kernel/fork.c
===================================================================
--- linux-2.4.21.orig/kernel/fork.c     2005-06-01 22:58:09.055062312 -0400
+++ linux-2.4.21/kernel/fork.c  2005-06-01 23:12:54.760414632 -0400
@@ -14,6 +14,7 @@
 #include <linux/config.h>
 #include <linux/slab.h>
 #include <linux/init.h>
+#include <linux/ptrack.h>
 #include <linux/unistd.h>
 #include <linux/smp_lock.h>
 #include <linux/module.h>
@@ -308,6 +309,7 @@
        /* unlimited stack is larger than TASK_SIZE */
        mm->non_executable_cache = NON_EXECUTABLE_CACHE(current);
        mm->pgd = pgd_alloc(mm);
+       mm->coproc_ops = NULL;
        mm->def_flags = 0;
        if (mm->pgd)
                return mm;
@@ -1110,6 +1112,12 @@
                        p->vfork_done = &vfork;
                        init_completion(&vfork);
                }
+               
+               if (ptrack_call_callbacks (PTRACK_PHASE_CLONE, p)) {
+                       /* start up with an immediate SIGKILL. */
+                       sigaddset (&p->pending.signal, SIGKILL);
+                       p->sigpending = 1;
+               }
 
                if ((p->ptrace & PT_PTRACED) || (clone_flags & CLONE_STOPPED)) {
                        /*
Index: linux-2.4.21/kernel/ksyms.c
===================================================================
--- linux-2.4.21.orig/kernel/ksyms.c    2005-06-01 23:12:40.911519984 -0400
+++ linux-2.4.21/kernel/ksyms.c 2005-06-01 23:12:54.760414632 -0400
@@ -43,6 +43,7 @@
 #include <linux/mmzone.h>
 #include <linux/mm.h>
 #include <linux/capability.h>
+#include <linux/ptrack.h>
 #include <linux/highuid.h>
 #include <linux/brlock.h>
 #include <linux/fs.h>
@@ -104,6 +105,10 @@
 
 #endif
 
+EXPORT_SYMBOL_GPL(ptrack_register);
+EXPORT_SYMBOL_GPL(ptrack_deregister);
+EXPORT_SYMBOL_GPL(ptrack_registered);
+
 /* process memory management */
 EXPORT_SYMBOL(do_mmap_pgoff);
 EXPORT_SYMBOL(do_munmap);
@@ -113,6 +118,7 @@
 EXPORT_SYMBOL(exit_files);
 EXPORT_SYMBOL(exit_fs);
 EXPORT_SYMBOL(exit_sighand);
+EXPORT_SYMBOL(make_pages_present);
 EXPORT_SYMBOL(unshare_files);
 EXPORT_SYMBOL(mmput);
 
@@ -589,6 +595,10 @@
 EXPORT_SYMBOL(kernel_read);
 EXPORT_SYMBOL(open_exec);
 
+/* QSW Shared-memory cleanup hook for rmsmod */
+extern int shm_cleanup();
+EXPORT_SYMBOL_GPL(shm_cleanup);
+
 /* Miscellaneous access points */
 EXPORT_SYMBOL(si_meminfo);
 
Index: linux-2.4.21/kernel/Makefile
===================================================================
--- linux-2.4.21.orig/kernel/Makefile   2005-06-01 22:51:53.000000000 -0400
+++ linux-2.4.21/kernel/Makefile        2005-06-01 23:12:54.760414632 -0400
@@ -18,6 +18,10 @@
            signal.o sys.o kmod.o context.o \
            futex.o pid.o kksymoops.o
 
+# Quadrics additions
+export-objs += ptrack.o
+obj-y += ptrack.o
+
 obj-$(CONFIG_UID16) += uid16.o
 obj-$(CONFIG_MODULES) += ksyms.o
 obj-$(CONFIG_COMPAT) += compat.o
Index: linux-2.4.21/kernel/ptrack.c
===================================================================
--- linux-2.4.21.orig/kernel/ptrack.c   2004-02-23 16:02:56.000000000 -0500
+++ linux-2.4.21/kernel/ptrack.c        2005-06-01 23:12:54.761414480 -0400
@@ -0,0 +1,143 @@
+/*
+ *    Copyright (C) 2000  Regents of the University of California
+ *
+ *    This program is free software; you can redistribute it and/or modify
+ *    it under the terms of the GNU General Public License as published by
+ *    the Free Software Foundation; either version 2 of the License, or
+ *    (at your option) any later version.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    GNU General Public License for more details.
+ *
+ *    You should have received a copy of the GNU General Public License
+ *    along with this program; if not, write to the Free Software
+ *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ * Derived from exit_actn.c by
+ *    Copyright (C) 2003 Quadrics Ltd.
+ */
+
+
+#include <linux/spinlock.h>
+#include <linux/sched.h>
+#include <linux/ptrack.h>
+#include <linux/slab.h>
+#include <linux/list.h>
+
+int
+ptrack_register (ptrack_callback_t callback, void *arg)
+{
+       struct ptrack_desc *desc = kmalloc (sizeof (struct ptrack_desc), GFP_KERNEL);
+       
+       if (desc == NULL)
+               return -ENOMEM;
+
+       desc->callback = callback;
+       desc->arg      = arg;
+       
+       list_add_tail (&desc->link, &current->ptrack_list);
+       
+       return 0;
+}
+
+void
+ptrack_deregister (ptrack_callback_t callback, void *arg)
+{      
+       struct list_head *el, *nel;
+       
+       list_for_each_safe (el, nel, &current->ptrack_list) {
+               struct ptrack_desc *desc = list_entry (el, struct ptrack_desc, link);
+               
+               if (desc->callback == callback && desc->arg == arg) {
+                       list_del (&desc->link);
+                       kfree (desc);
+               }
+       }
+}
+
+int
+ptrack_registered (ptrack_callback_t callback, void *arg)
+{
+       struct list_head *el;
+       
+       list_for_each (el, &current->ptrack_list) {
+               struct ptrack_desc *desc = list_entry (el, struct ptrack_desc, link);
+               
+               if (desc->callback == callback && desc->arg == arg)
+                       return 1;
+       }
+       return 0;
+}      
+        
+int
+ptrack_call_callbacks (int phase, struct task_struct *child)
+{
+       struct list_head *el, *nel;
+       struct ptrack_desc *new;
+       int res;
+
+       if (phase == PTRACK_PHASE_CLONE)
+               INIT_LIST_HEAD (&child->ptrack_list);
+
+       /* if init process, ignore */
+       if (current->pid == 0)
+           return 0;
+       
+       list_for_each_safe (el, nel, &current->ptrack_list) {
+               struct ptrack_desc *desc = list_entry (el, struct ptrack_desc, link);
+               
+              res = desc->callback (desc->arg, phase, child);
+               
+               switch (phase)
+               {
+               case PTRACK_PHASE_EXIT:
+                       list_del (&desc->link);
+                       kfree (desc);
+                       break;
+                       
+               case PTRACK_PHASE_CLONE:
+                      switch (res)
+                      {
+                      case PTRACK_FINISHED:
+                              break;
+
+                      case PTRACK_INNHERIT:
+                              if ((new = kmalloc (sizeof (struct ptrack_desc), GFP_ATOMIC)) == NULL)
+                              {
+                                      /* allocation failed - notify that this process is not going
+                                       * to be started by signalling clone failure.
+                                       */
+                                      desc->callback (desc->arg, PTRACK_PHASE_CLONE_FAIL, child);
+                                      
+                                      goto failed;
+                              }
+
+                               new->callback = desc->callback;
+                               new->arg      = desc->arg;
+                               
+                               list_add_tail (&new->link, &child->ptrack_list);
+                              break;
+
+                      case PTRACK_DENIED:
+                              goto failed;
+                       }
+                      break;
+               }
+       }
+
+       return 0;
+
+ failed:
+       while (! list_empty (&child->ptrack_list))
+       {
+              struct ptrack_desc *desc = list_entry (child->ptrack_list.next, struct ptrack_desc, link);
+              
+              desc->callback (desc->arg, PTRACK_PHASE_CLONE_FAIL, child);
+
+              list_del (&desc->link);
+              kfree (desc);
+       }
+       return 1;
+}
Index: linux-2.4.21/mm/filemap.c
===================================================================
--- linux-2.4.21.orig/mm/filemap.c      2005-06-01 23:12:41.100491256 -0400
+++ linux-2.4.21/mm/filemap.c   2005-06-01 23:12:54.763414176 -0400
@@ -22,6 +22,7 @@
 #include <linux/swapctl.h>
 #include <linux/init.h>
 #include <linux/mm.h>
+#include <linux/coproc.h>
 #include <linux/mm_inline.h>
 #include <linux/iobuf.h>
 #include <linux/bootmem.h>
@@ -2468,6 +2469,7 @@
        flush_cache_range(vma, end - size, end);
        if (address >= end)
                BUG();
+       coproc_sync_range (vma->vm_mm, address, end);
        do {
                error |= filemap_sync_pmd_range(dir, address, end - address, vma, flags);
                address = (address + PGDIR_SIZE) & PGDIR_MASK;
Index: linux-2.4.21/mm/memory.c
===================================================================
--- linux-2.4.21.orig/mm/memory.c       2005-06-01 22:52:04.000000000 -0400
+++ linux-2.4.21/mm/memory.c    2005-06-01 23:13:59.371592240 -0400
@@ -42,6 +42,7 @@
 #include <linux/smp_lock.h>
 #include <linux/swapctl.h>
 #include <linux/iobuf.h>
+#include <linux/coproc.h>
 #include <linux/highmem.h>
 #include <linux/pagemap.h>
 #include <linux/module.h>
@@ -632,6 +633,7 @@
                BUG_ON(address >= end);
 
                spin_lock(&mm->page_table_lock);
+               coproc_invalidate_range (mm, address, end);
                flush_cache_range(vma, start, end);
                tlb = tlb_gather_mmu(vma);
 
@@ -1302,6 +1304,7 @@
                BUG();
 
        spin_lock(&mm->page_table_lock);
+       coproc_invalidate_range (mm, beg, end);
        do {
                pmd_t *pmd = pmd_alloc(mm, dir, address);
                error = -ENOMEM;
@@ -1313,6 +1316,7 @@
                address = (address + PGDIR_SIZE) & PGDIR_MASK;
                dir++;
        } while (address && (address < end));
+       coproc_update_range(mm, beg, end);
        spin_unlock(&mm->page_table_lock);
        flush_tlb_range(vma, beg, end);
        return error;
@@ -1391,6 +1395,7 @@
                BUG();
 
        spin_lock(&mm->page_table_lock);
+       coproc_invalidate_range(mm, beg, end);
        do {
                pmd_t *pmd = pmd_alloc(mm, dir, from);
                error = -ENOMEM;
@@ -1402,6 +1407,7 @@
                from = (from + PGDIR_SIZE) & PGDIR_MASK;
                dir++;
        } while (from && (from < end));
+       coproc_update_range(mm, beg, end);
        spin_unlock(&mm->page_table_lock);
        flush_tlb_range(vma, beg, end);
        return error;
@@ -1497,8 +1503,10 @@
                        unlock_page(old_page);
                        flush_cache_page(vma, address);
                        entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)), vma);
+                       coproc_invalidate_page(vma, address);
                        establish_pte(vma, address, page_table, entry);
                        pte_unmap(page_table);
+                       coproc_update_page(vma, address);
                        spin_unlock(&mm->page_table_lock);
                        return 1;       /* Minor fault */
                }
@@ -1528,6 +1536,7 @@
                if (PageReserved(old_page))
                        ++mm->rss;
                page_remove_rmap(old_page, page_table);
+               coproc_invalidate_page(vma, address);
                break_cow(vma, new_page, address, page_table);
                pte_chain = page_add_rmap(new_page, page_table, pte_chain);
                lru_cache_add(new_page);
@@ -1536,6 +1545,7 @@
                new_page = old_page;
        }
        pte_unmap(page_table);
+       coproc_update_page(vma, address);
        spin_unlock(&mm->page_table_lock);
        if (old_page_locked)
                unlock_page(old_page);
@@ -1748,6 +1758,7 @@
        /* No need to invalidate - it was non-present before */
        update_mmu_cache(vma, address, pte);
        pte_unmap(page_table);
+       coproc_update_page(vma, address);
        spin_unlock(&mm->page_table_lock);
        pte_chain_free(pte_chain);
        return ret;
@@ -1804,6 +1815,7 @@
        /* No need to invalidate - it was non-present before */
        update_mmu_cache(vma, addr, entry);
        pte_unmap(page_table);
+       coproc_update_page(vma, addr);
        spin_unlock(&mm->page_table_lock);
        ret = 1;        /* Minor fault */
        goto out;
@@ -1902,6 +1914,7 @@
 
        /* no need to invalidate: a not-present page shouldn't be cached */
        update_mmu_cache(vma, address, entry);
+       coproc_update_page(vma, address);
        spin_unlock(&mm->page_table_lock);
        pte_chain_free(pte_chain);
        return 2;       /* Major fault */
@@ -1958,8 +1971,10 @@
                entry = pte_mkdirty(entry);
        }
        entry = pte_mkyoung(entry);
+       coproc_invalidate_page(vma, address);
        establish_pte(vma, address, pte, entry);
        pte_unmap(pte);
+       coproc_update_page(vma, address);
        spin_unlock(&mm->page_table_lock);
        return 1;
 }
Index: linux-2.4.21/mm/mmap.c
===================================================================
--- linux-2.4.21.orig/mm/mmap.c 2005-06-01 22:51:50.000000000 -0400
+++ linux-2.4.21/mm/mmap.c      2005-06-01 23:12:54.767413568 -0400
@@ -30,6 +30,7 @@
 #include <linux/init.h>
 #include <linux/file.h>
 #include <linux/fs.h>
+#include <linux/coproc.h>
 #include <linux/personality.h>
 #include <linux/compiler.h>
 #include <linux/profile.h>
@@ -1450,6 +1451,7 @@
        release_segments(mm);
  
        spin_lock(&mm->page_table_lock);
+       coproc_release(mm);
        mpnt = mm->mmap;
        mm->mmap = mm->mmap_cache = NULL;
        mm->mm_rb = RB_ROOT;
Index: linux-2.4.21/mm/mprotect.c
===================================================================
--- linux-2.4.21.orig/mm/mprotect.c     2005-06-01 22:51:50.000000000 -0400
+++ linux-2.4.21/mm/mprotect.c  2005-06-01 23:12:54.767413568 -0400
@@ -24,6 +24,7 @@
 #include <linux/smp_lock.h>
 #include <linux/shm.h>
 #include <linux/mman.h>
+#include <linux/coproc.h>
 #include <linux/highmem.h>
 #include <linux/hugetlb.h>
 
@@ -106,6 +107,7 @@
        if (start >= end)
                BUG();
        spin_lock(&current->mm->page_table_lock);
+       coproc_change_protection (current->mm, start, end, newprot);
        do {
                change_pmd_range(vma, dir, start, end - start, newprot);
                start = (start + PGDIR_SIZE) & PGDIR_MASK;
Index: linux-2.4.21/mm/mremap.c
===================================================================
--- linux-2.4.21.orig/mm/mremap.c       2005-06-01 22:51:50.000000000 -0400
+++ linux-2.4.21/mm/mremap.c    2005-06-01 23:12:54.768413416 -0400
@@ -26,6 +26,7 @@
 #include <linux/shm.h>
 #include <linux/mman.h>
 #include <linux/swap.h>
+#include <linux/coproc.h>
 #include <linux/highmem.h>
 #include <linux/hugetlb.h>
 
@@ -160,7 +161,10 @@
        unsigned long new_addr, unsigned long old_addr, unsigned long len)
 {
        unsigned long offset = len;
+       struct mm_struct *mm = vma->vm_mm;
 
+       coproc_invalidate_range(mm, old_addr, old_addr+len);
+       coproc_invalidate_range(mm, new_addr, new_addr+len);
        flush_cache_range(vma, old_addr, old_addr + len);
 
        /*
Index: linux-2.4.21/mm/rmap.c
===================================================================
--- linux-2.4.21.orig/mm/rmap.c 2005-06-01 22:51:50.000000000 -0400
+++ linux-2.4.21/mm/rmap.c      2005-06-01 23:12:54.768413416 -0400
@@ -26,6 +26,7 @@
 #include <linux/slab.h>
 #include <linux/init.h>
 #include <linux/cache.h>
+#include <linux/coproc.h>
 
 #include <asm/pgalloc.h>
 #include <asm/rmap.h>
@@ -449,6 +450,7 @@
        }
 
        /* Nuke the page table entry. */
+       coproc_invalidate_page(vma, address);
        pte = vm_ptep_get_and_clear(vma, address, ptep);
        flush_tlb_page(vma, address);
        flush_cache_page(vma, address);