From b5459c30cb6c6adae719a632a6459837f39d56f2 Mon Sep 17 00:00:00 2001 From: eeb Date: Thu, 28 Oct 2004 12:02:59 +0000 Subject: [PATCH] * Landed b1_2_singleportals --- lnet/klnds/iiblnd/.cvsignore | 10 + lnet/klnds/iiblnd/Makefile.in | 6 + lnet/klnds/iiblnd/Makefile.mk | 10 + lnet/klnds/iiblnd/autoMakefile.am | 15 + lnet/klnds/iiblnd/iiblnd.c | 1713 +++++++++++++++++++++ lnet/klnds/iiblnd/iiblnd.h | 892 +++++++++++ lnet/klnds/iiblnd/iiblnd_cb.c | 3018 +++++++++++++++++++++++++++++++++++++ 7 files changed, 5664 insertions(+) create mode 100644 lnet/klnds/iiblnd/.cvsignore create mode 100644 lnet/klnds/iiblnd/Makefile.in create mode 100644 lnet/klnds/iiblnd/Makefile.mk create mode 100644 lnet/klnds/iiblnd/autoMakefile.am create mode 100644 lnet/klnds/iiblnd/iiblnd.c create mode 100644 lnet/klnds/iiblnd/iiblnd.h create mode 100644 lnet/klnds/iiblnd/iiblnd_cb.c diff --git a/lnet/klnds/iiblnd/.cvsignore b/lnet/klnds/iiblnd/.cvsignore new file mode 100644 index 0000000..5ed596b --- /dev/null +++ b/lnet/klnds/iiblnd/.cvsignore @@ -0,0 +1,10 @@ +.deps +Makefile +.*.cmd +autoMakefile.in +autoMakefile +*.ko +*.mod.c +.*.flags +.tmp_versions +.depend diff --git a/lnet/klnds/iiblnd/Makefile.in b/lnet/klnds/iiblnd/Makefile.in new file mode 100644 index 0000000..e7934e2 --- /dev/null +++ b/lnet/klnds/iiblnd/Makefile.in @@ -0,0 +1,6 @@ +MODULES := kiibnal +kiibnal-objs := iibnal.o iibnal_cb.o + +EXTRA_POST_CFLAGS := @IIBCPPFLAGS@ + +@INCLUDE_RULES@ diff --git a/lnet/klnds/iiblnd/Makefile.mk b/lnet/klnds/iiblnd/Makefile.mk new file mode 100644 index 0000000..0459a20 --- /dev/null +++ b/lnet/klnds/iiblnd/Makefile.mk @@ -0,0 +1,10 @@ +# Copyright (C) 2001 Cluster File Systems, Inc. +# +# This code is issued under the GNU General Public License. +# See the file COPYING in this distribution + +include $(src)/../../Kernelenv + +obj-y += kiibnal.o +kiibnal-objs := iibnal.o iibnal_cb.o + diff --git a/lnet/klnds/iiblnd/autoMakefile.am b/lnet/klnds/iiblnd/autoMakefile.am new file mode 100644 index 0000000..251df66 --- /dev/null +++ b/lnet/klnds/iiblnd/autoMakefile.am @@ -0,0 +1,15 @@ +# Copyright (C) 2001 Cluster File Systems, Inc. +# +# This code is issued under the GNU General Public License. +# See the file COPYING in this distribution + +if MODULES +if !CRAY_PORTALS +if BUILD_IIBNAL +modulenet_DATA = kiibnal$(KMODEXT) +endif +endif +endif + +MOSTLYCLEANFILES = *.o *.ko *.mod.c +DIST_SOURCES = $(kiibnal-objs:%.o=%.c) iibnal.h diff --git a/lnet/klnds/iiblnd/iiblnd.c b/lnet/klnds/iiblnd/iiblnd.c new file mode 100644 index 0000000..09908c9 --- /dev/null +++ b/lnet/klnds/iiblnd/iiblnd.c @@ -0,0 +1,1713 @@ +/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- + * vim:expandtab:shiftwidth=8:tabstop=8: + * + * Copyright (C) 2004 Cluster File Systems, Inc. + * Author: Eric Barton + * + * This file is part of Lustre, http://www.lustre.org. + * + * Lustre is free software; you can redistribute it and/or + * modify it under the terms of version 2 of the GNU General Public + * License as published by the Free Software Foundation. + * + * Lustre 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 Lustre; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * + */ + +#include "iibnal.h" + +nal_t kibnal_api; +ptl_handle_ni_t kibnal_ni; +kib_tunables_t kibnal_tunables; + +kib_data_t kibnal_data = { + .kib_service_id = IBNAL_SERVICE_NUMBER, +}; + +#ifdef CONFIG_SYSCTL +#define IBNAL_SYSCTL 202 + +#define IBNAL_SYSCTL_TIMEOUT 1 + +static ctl_table kibnal_ctl_table[] = { + {IBNAL_SYSCTL_TIMEOUT, "timeout", + &kibnal_tunables.kib_io_timeout, sizeof (int), + 0644, NULL, &proc_dointvec}, + { 0 } +}; + +static ctl_table kibnal_top_ctl_table[] = { + {IBNAL_SYSCTL, "iibnal", NULL, 0, 0555, kibnal_ctl_table}, + { 0 } +}; +#endif + +#ifdef unused +void +print_service(IB_SERVICE_RECORD *service, char *tag, int rc) +{ + char name[32]; + + if (service == NULL) + { + CWARN("tag : %s\n" + "status : %d (NULL)\n", tag, rc); + return; + } + strncpy (name, service->ServiceName, sizeof(name)-1); + name[sizeof(name)-1] = 0; + + CWARN("tag : %s\n" + "status : %d\n" + "service id: "LPX64"\n" + "name : %s\n" + "NID : "LPX64"\n", tag, rc, + service->RID.ServiceID, name, + *kibnal_service_nid_field(service)); +} +#endif + +static void +kibnal_service_setunset_done (void *arg, FABRIC_OPERATION_DATA *fod, + FSTATUS frc, uint32 madrc) +{ + *(FSTATUS *)arg = frc; + up (&kibnal_data.kib_nid_signal); +} + +#if IBNAL_CHECK_ADVERT +static void +kibnal_service_query_done (void *arg, QUERY *qry, + QUERY_RESULT_VALUES *qry_result) +{ + FSTATUS frc = qry_result->Status; + + if (frc != FSUCCESS && + qry_result->ResultDataSize == 0) + frc = FERROR; + + *(FSTATUS *)arg = frc; + up (&kibnal_data.kib_nid_signal); +} + +static void +kibnal_check_advert (void) +{ + QUERY *qry; + IB_SERVICE_RECORD *svc; + FSTATUS frc; + FSTATUS frc2; + + PORTAL_ALLOC(qry, sizeof(*qry)); + if (qry == NULL) + return; + + memset (qry, 0, sizeof(*qry)); + qry->InputType = InputTypeServiceRecord; + qry->OutputType = OutputTypeServiceRecord; + qry->InputValue.ServiceRecordValue.ComponentMask = KIBNAL_SERVICE_KEY_MASK; + svc = &qry->InputValue.ServiceRecordValue.ServiceRecord; + kibnal_set_service_keys(svc, kibnal_data.kib_nid); + + frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd, + kibnal_data.kib_port_guid, + qry, + kibnal_service_query_done, + NULL, &frc2); + if (frc != FSUCCESS && frc != FPENDING) { + CERROR ("Immediate error %d checking SM service\n", frc); + } else { + down (&kibnal_data.kib_nid_signal); + frc = frc2; + + if (frc != 0) + CERROR ("Error %d checking SM service\n", rc); + } + + return (rc); +} +#endif + +static void fill_fod(FABRIC_OPERATION_DATA *fod, FABRIC_OPERATION_TYPE type) +{ + IB_SERVICE_RECORD *svc; + + memset (fod, 0, sizeof(*fod)); + fod->Type = type; + + svc = &fod->Value.ServiceRecordValue.ServiceRecord; + svc->RID.ServiceID = kibnal_data.kib_service_id; + svc->RID.ServiceGID.Type.Global.InterfaceID = kibnal_data.kib_port_guid; + svc->RID.ServiceGID.Type.Global.SubnetPrefix = DEFAULT_SUBNET_PREFIX; + svc->RID.ServiceP_Key = kibnal_data.kib_port_pkey; + svc->ServiceLease = 0xffffffff; + + kibnal_set_service_keys(svc, kibnal_data.kib_nid); +} + +static int +kibnal_advertise (void) +{ + FABRIC_OPERATION_DATA *fod; + IB_SERVICE_RECORD *svc; + FSTATUS frc; + FSTATUS frc2; + + LASSERT (kibnal_data.kib_nid != PTL_NID_ANY); + + PORTAL_ALLOC(fod, sizeof(*fod)); + if (fod == NULL) + return (-ENOMEM); + + fill_fod(fod, FabOpSetServiceRecord); + svc = &fod->Value.ServiceRecordValue.ServiceRecord; + + CDEBUG(D_NET, "Advertising service id "LPX64" %s:"LPX64"\n", + svc->RID.ServiceID, + svc->ServiceName, *kibnal_service_nid_field(svc)); + + frc = iibt_sd_port_fabric_operation(kibnal_data.kib_sd, + kibnal_data.kib_port_guid, + fod, kibnal_service_setunset_done, + NULL, &frc2); + + if (frc != FSUCCESS && frc != FPENDING) { + CERROR ("Immediate error %d advertising NID "LPX64"\n", + frc, kibnal_data.kib_nid); + goto out; + } + + down (&kibnal_data.kib_nid_signal); + + frc = frc2; + if (frc != FSUCCESS) + CERROR ("Error %d advertising BUD "LPX64"\n", + frc, kibnal_data.kib_nid); +out: + PORTAL_FREE(fod, sizeof(*fod)); + return (frc == FSUCCESS) ? 0 : -EINVAL; +} + +static void +kibnal_unadvertise (int expect_success) +{ + FABRIC_OPERATION_DATA *fod; + IB_SERVICE_RECORD *svc; + FSTATUS frc; + FSTATUS frc2; + + LASSERT (kibnal_data.kib_nid != PTL_NID_ANY); + + PORTAL_ALLOC(fod, sizeof(*fod)); + if (fod == NULL) + return; + + fill_fod(fod, FabOpDeleteServiceRecord); + svc = &fod->Value.ServiceRecordValue.ServiceRecord; + + CDEBUG(D_NET, "Unadvertising service %s:"LPX64"\n", + svc->ServiceName, *kibnal_service_nid_field(svc)); + + frc = iibt_sd_port_fabric_operation(kibnal_data.kib_sd, + kibnal_data.kib_port_guid, + fod, kibnal_service_setunset_done, + NULL, &frc2); + + if (frc != FSUCCESS && frc != FPENDING) { + CERROR ("Immediate error %d unadvertising NID "LPX64"\n", + frc, kibnal_data.kib_nid); + goto out; + } + + down (&kibnal_data.kib_nid_signal); + + if ((frc2 == FSUCCESS) == !!expect_success) + goto out; + + if (expect_success) + CERROR("Error %d unadvertising NID "LPX64"\n", + frc2, kibnal_data.kib_nid); + else + CWARN("Removed conflicting NID "LPX64"\n", + kibnal_data.kib_nid); + out: + PORTAL_FREE(fod, sizeof(*fod)); +} + +static int +kibnal_set_mynid(ptl_nid_t nid) +{ + struct timeval tv; + lib_ni_t *ni = &kibnal_lib.libnal_ni; + int rc; + FSTATUS frc; + + CDEBUG(D_IOCTL, "setting mynid to "LPX64" (old nid="LPX64")\n", + nid, ni->ni_pid.nid); + + do_gettimeofday(&tv); + + down (&kibnal_data.kib_nid_mutex); + + if (nid == kibnal_data.kib_nid) { + /* no change of NID */ + up (&kibnal_data.kib_nid_mutex); + return (0); + } + + CDEBUG(D_NET, "NID "LPX64"("LPX64")\n", + kibnal_data.kib_nid, nid); + + if (kibnal_data.kib_nid != PTL_NID_ANY) { + + kibnal_unadvertise (1); + + frc = iibt_cm_cancel(kibnal_data.kib_cep); + if (frc != FSUCCESS && frc != FPENDING) + CERROR ("Error %d stopping listener\n", frc); + + frc = iibt_cm_destroy_cep(kibnal_data.kib_cep); + if (frc != FSUCCESS) + CERROR ("Error %d destroying CEP\n", frc); + + kibnal_data.kib_cep = NULL; + } + + kibnal_data.kib_nid = ni->ni_pid.nid = nid; + kibnal_data.kib_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; + + /* Delete all existing peers and their connections after new + * NID/incarnation set to ensure no old connections in our brave + * new world. */ + kibnal_del_peer (PTL_NID_ANY, 0); + + if (kibnal_data.kib_nid == PTL_NID_ANY) { + /* No new NID to install */ + up (&kibnal_data.kib_nid_mutex); + return (0); + } + + /* remove any previous advert (crashed node etc) */ + kibnal_unadvertise(0); + + kibnal_data.kib_cep = iibt_cm_create_cep(CM_RC_TYPE); + if (kibnal_data.kib_cep == NULL) { + CERROR ("Can't create CEP\n"); + rc = -ENOMEM; + } else { + CM_LISTEN_INFO info; + memset (&info, 0, sizeof(info)); + info.ListenAddr.EndPt.SID = kibnal_data.kib_service_id; + + frc = iibt_cm_listen(kibnal_data.kib_cep, &info, + kibnal_listen_callback, NULL); + if (frc != FSUCCESS && frc != FPENDING) { + CERROR ("iibt_cm_listen error: %d\n", frc); + rc = -EINVAL; + } else { + rc = 0; + } + } + + if (rc == 0) { + rc = kibnal_advertise(); + if (rc == 0) { +#if IBNAL_CHECK_ADVERT + kibnal_check_advert(); +#endif + up (&kibnal_data.kib_nid_mutex); + return (0); + } + + iibt_cm_cancel (kibnal_data.kib_cep); + iibt_cm_destroy_cep (kibnal_data.kib_cep); + /* remove any peers that sprung up while I failed to + * advertise myself */ + kibnal_del_peer (PTL_NID_ANY, 0); + } + + kibnal_data.kib_nid = PTL_NID_ANY; + up (&kibnal_data.kib_nid_mutex); + return (rc); +} + +kib_peer_t * +kibnal_create_peer (ptl_nid_t nid) +{ + kib_peer_t *peer; + + LASSERT (nid != PTL_NID_ANY); + + PORTAL_ALLOC (peer, sizeof (*peer)); + if (peer == NULL) + return (NULL); + + memset(peer, 0, sizeof(*peer)); /* zero flags etc */ + + peer->ibp_nid = nid; + atomic_set (&peer->ibp_refcount, 1); /* 1 ref for caller */ + + INIT_LIST_HEAD (&peer->ibp_list); /* not in the peer table yet */ + INIT_LIST_HEAD (&peer->ibp_conns); + INIT_LIST_HEAD (&peer->ibp_tx_queue); + + peer->ibp_reconnect_time = jiffies; + peer->ibp_reconnect_interval = IBNAL_MIN_RECONNECT_INTERVAL; + + atomic_inc (&kibnal_data.kib_npeers); + return (peer); +} + +void +kibnal_destroy_peer (kib_peer_t *peer) +{ + + LASSERT (atomic_read (&peer->ibp_refcount) == 0); + LASSERT (peer->ibp_persistence == 0); + LASSERT (!kibnal_peer_active(peer)); + LASSERT (peer->ibp_connecting == 0); + LASSERT (list_empty (&peer->ibp_conns)); + LASSERT (list_empty (&peer->ibp_tx_queue)); + + PORTAL_FREE (peer, sizeof (*peer)); + + /* NB a peer's connections keep a reference on their peer until + * they are destroyed, so we can be assured that _all_ state to do + * with this peer has been cleaned up when its refcount drops to + * zero. */ + atomic_dec (&kibnal_data.kib_npeers); +} + +/* the caller is responsible for accounting for the additional reference + * that this creates */ +kib_peer_t * +kibnal_find_peer_locked (ptl_nid_t nid) +{ + struct list_head *peer_list = kibnal_nid2peerlist (nid); + struct list_head *tmp; + kib_peer_t *peer; + + list_for_each (tmp, peer_list) { + + peer = list_entry (tmp, kib_peer_t, ibp_list); + + LASSERT (peer->ibp_persistence != 0 || /* persistent peer */ + peer->ibp_connecting != 0 || /* creating conns */ + !list_empty (&peer->ibp_conns)); /* active conn */ + + if (peer->ibp_nid != nid) + continue; + + CDEBUG(D_NET, "got peer [%p] -> "LPX64" (%d)\n", + peer, nid, atomic_read (&peer->ibp_refcount)); + return (peer); + } + return (NULL); +} + +kib_peer_t * +kibnal_get_peer (ptl_nid_t nid) +{ + kib_peer_t *peer; + + read_lock (&kibnal_data.kib_global_lock); + peer = kibnal_find_peer_locked (nid); + if (peer != NULL) /* +1 ref for caller? */ + kib_peer_addref(peer); + read_unlock (&kibnal_data.kib_global_lock); + + return (peer); +} + +void +kibnal_unlink_peer_locked (kib_peer_t *peer) +{ + LASSERT (peer->ibp_persistence == 0); + LASSERT (list_empty(&peer->ibp_conns)); + + LASSERT (kibnal_peer_active(peer)); + list_del_init (&peer->ibp_list); + /* lose peerlist's ref */ + kib_peer_decref(peer); +} + +static int +kibnal_get_peer_info (int index, ptl_nid_t *nidp, int *persistencep) +{ + kib_peer_t *peer; + struct list_head *ptmp; + int i; + + read_lock (&kibnal_data.kib_global_lock); + + for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) { + + list_for_each (ptmp, &kibnal_data.kib_peers[i]) { + + peer = list_entry (ptmp, kib_peer_t, ibp_list); + LASSERT (peer->ibp_persistence != 0 || + peer->ibp_connecting != 0 || + !list_empty (&peer->ibp_conns)); + + if (index-- > 0) + continue; + + *nidp = peer->ibp_nid; + *persistencep = peer->ibp_persistence; + + read_unlock (&kibnal_data.kib_global_lock); + return (0); + } + } + + read_unlock (&kibnal_data.kib_global_lock); + return (-ENOENT); +} + +static int +kibnal_add_persistent_peer (ptl_nid_t nid) +{ + unsigned long flags; + kib_peer_t *peer; + kib_peer_t *peer2; + + if (nid == PTL_NID_ANY) + return (-EINVAL); + + peer = kibnal_create_peer (nid); + if (peer == NULL) + return (-ENOMEM); + + write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + + peer2 = kibnal_find_peer_locked (nid); + if (peer2 != NULL) { + kib_peer_decref (peer); + peer = peer2; + } else { + /* peer table takes existing ref on peer */ + list_add_tail (&peer->ibp_list, + kibnal_nid2peerlist (nid)); + } + + peer->ibp_persistence++; + + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + return (0); +} + +static void +kibnal_del_peer_locked (kib_peer_t *peer, int single_share) +{ + struct list_head *ctmp; + struct list_head *cnxt; + kib_conn_t *conn; + + if (!single_share) + peer->ibp_persistence = 0; + else if (peer->ibp_persistence > 0) + peer->ibp_persistence--; + + if (peer->ibp_persistence != 0) + return; + + list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry(ctmp, kib_conn_t, ibc_list); + + kibnal_close_conn_locked (conn, 0); + } + + /* NB peer unlinks itself when last conn is closed */ +} + +int +kibnal_del_peer (ptl_nid_t nid, int single_share) +{ + unsigned long flags; + struct list_head *ptmp; + struct list_head *pnxt; + kib_peer_t *peer; + int lo; + int hi; + int i; + int rc = -ENOENT; + + write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + + if (nid != PTL_NID_ANY) + lo = hi = kibnal_nid2peerlist(nid) - kibnal_data.kib_peers; + else { + lo = 0; + hi = kibnal_data.kib_peer_hash_size - 1; + } + + for (i = lo; i <= hi; i++) { + list_for_each_safe (ptmp, pnxt, &kibnal_data.kib_peers[i]) { + peer = list_entry (ptmp, kib_peer_t, ibp_list); + LASSERT (peer->ibp_persistence != 0 || + peer->ibp_connecting != 0 || + !list_empty (&peer->ibp_conns)); + + if (!(nid == PTL_NID_ANY || peer->ibp_nid == nid)) + continue; + + kibnal_del_peer_locked (peer, single_share); + rc = 0; /* matched something */ + + if (single_share) + goto out; + } + } + out: + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + + return (rc); +} + +static kib_conn_t * +kibnal_get_conn_by_idx (int index) +{ + kib_peer_t *peer; + struct list_head *ptmp; + kib_conn_t *conn; + struct list_head *ctmp; + int i; + + read_lock (&kibnal_data.kib_global_lock); + + for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) { + list_for_each (ptmp, &kibnal_data.kib_peers[i]) { + + peer = list_entry (ptmp, kib_peer_t, ibp_list); + LASSERT (peer->ibp_persistence > 0 || + peer->ibp_connecting != 0 || + !list_empty (&peer->ibp_conns)); + + list_for_each (ctmp, &peer->ibp_conns) { + if (index-- > 0) + continue; + + conn = list_entry (ctmp, kib_conn_t, ibc_list); + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); + read_unlock (&kibnal_data.kib_global_lock); + return (conn); + } + } + } + + read_unlock (&kibnal_data.kib_global_lock); + return (NULL); +} + +kib_conn_t * +kibnal_create_conn (void) +{ + kib_conn_t *conn; + int i; + __u64 vaddr = 0; + __u64 vaddr_base; + int page_offset; + int ipage; + int rc; + FSTATUS frc; + union { + IB_QP_ATTRIBUTES_CREATE qp_create; + IB_QP_ATTRIBUTES_MODIFY qp_attr; + } params; + + PORTAL_ALLOC (conn, sizeof (*conn)); + if (conn == NULL) { + CERROR ("Can't allocate connection\n"); + return (NULL); + } + + /* zero flags, NULL pointers etc... */ + memset (conn, 0, sizeof (*conn)); + + INIT_LIST_HEAD (&conn->ibc_tx_queue); + INIT_LIST_HEAD (&conn->ibc_active_txs); + spin_lock_init (&conn->ibc_lock); + + atomic_inc (&kibnal_data.kib_nconns); + /* well not really, but I call destroy() on failure, which decrements */ + + PORTAL_ALLOC (conn->ibc_rxs, IBNAL_RX_MSGS * sizeof (kib_rx_t)); + if (conn->ibc_rxs == NULL) + goto failed; + memset (conn->ibc_rxs, 0, IBNAL_RX_MSGS * sizeof(kib_rx_t)); + + rc = kibnal_alloc_pages(&conn->ibc_rx_pages, IBNAL_RX_MSG_PAGES, 1); + if (rc != 0) + goto failed; + + vaddr_base = vaddr = conn->ibc_rx_pages->ibp_vaddr; + + for (i = ipage = page_offset = 0; i < IBNAL_RX_MSGS; i++) { + struct page *page = conn->ibc_rx_pages->ibp_pages[ipage]; + kib_rx_t *rx = &conn->ibc_rxs[i]; + + rx->rx_conn = conn; + rx->rx_msg = (kib_msg_t *)(((char *)page_address(page)) + + page_offset); + + if (kibnal_whole_mem()) + rx->rx_vaddr = kibnal_page2phys(page) + + page_offset + + kibnal_data.kib_md.md_addr; + else + rx->rx_vaddr = vaddr; + + vaddr += IBNAL_MSG_SIZE; + LASSERT (vaddr <= vaddr_base + IBNAL_RX_MSG_BYTES); + + page_offset += IBNAL_MSG_SIZE; + LASSERT (page_offset <= PAGE_SIZE); + + if (page_offset == PAGE_SIZE) { + page_offset = 0; + ipage++; + LASSERT (ipage <= IBNAL_RX_MSG_PAGES); + } + } + + params.qp_create = (IB_QP_ATTRIBUTES_CREATE) { + .Type = QPTypeReliableConnected, + .SendQDepth = IBNAL_TX_MAX_SG * + IBNAL_MSG_QUEUE_SIZE, + .RecvQDepth = IBNAL_MSG_QUEUE_SIZE, + .SendDSListDepth = 1, + .RecvDSListDepth = 1, + .SendCQHandle = kibnal_data.kib_cq, + .RecvCQHandle = kibnal_data.kib_cq, + .PDHandle = kibnal_data.kib_pd, + .SendSignaledCompletions = TRUE, + }; + frc = iibt_qp_create(kibnal_data.kib_hca, ¶ms.qp_create, NULL, + &conn->ibc_qp, &conn->ibc_qp_attrs); + if (rc != 0) { + CERROR ("Failed to create queue pair: %d\n", rc); + goto failed; + } + + /* Mark QP created */ + conn->ibc_state = IBNAL_CONN_INIT_QP; + + params.qp_attr = (IB_QP_ATTRIBUTES_MODIFY) { + .RequestState = QPStateInit, + .Attrs = (IB_QP_ATTR_PORTGUID | + IB_QP_ATTR_PKEYINDEX | + IB_QP_ATTR_ACCESSCONTROL), + .PortGUID = kibnal_data.kib_port_guid, + .PkeyIndex = 0, + .AccessControl = { + .s = { + .RdmaWrite = 1, + .RdmaRead = 1, + }, + }, + }; + rc = iibt_qp_modify(conn->ibc_qp, ¶ms.qp_attr, NULL); + if (rc != 0) { + CERROR ("Failed to modify queue pair: %d\n", rc); + goto failed; + } + + /* 1 ref for caller */ + atomic_set (&conn->ibc_refcount, 1); + return (conn); + + failed: + kibnal_destroy_conn (conn); + return (NULL); +} + +void +kibnal_destroy_conn (kib_conn_t *conn) +{ + int rc; + FSTATUS frc; + + CDEBUG (D_NET, "connection %p\n", conn); + + LASSERT (atomic_read (&conn->ibc_refcount) == 0); + LASSERT (list_empty(&conn->ibc_tx_queue)); + LASSERT (list_empty(&conn->ibc_active_txs)); + LASSERT (conn->ibc_nsends_posted == 0); + LASSERT (conn->ibc_connreq == NULL); + + switch (conn->ibc_state) { + case IBNAL_CONN_DISCONNECTED: + /* called after connection sequence initiated */ + /* fall through */ + + case IBNAL_CONN_INIT_QP: + /* _destroy includes an implicit Reset of the QP which + * discards posted work */ + rc = iibt_qp_destroy(conn->ibc_qp); + if (rc != 0) + CERROR("Can't destroy QP: %d\n", rc); + /* fall through */ + + case IBNAL_CONN_INIT_NOTHING: + break; + + default: + LASSERT (0); + } + + if (conn->ibc_cep != NULL) { + frc = iibt_cm_destroy_cep(conn->ibc_cep); + if (frc != 0) + CERROR("Can't destroy CEP %p: %d\n", conn->ibc_cep, + frc); + } + + if (conn->ibc_rx_pages != NULL) + kibnal_free_pages(conn->ibc_rx_pages); + + if (conn->ibc_rxs != NULL) + PORTAL_FREE(conn->ibc_rxs, + IBNAL_RX_MSGS * sizeof(kib_rx_t)); + + if (conn->ibc_peer != NULL) + kib_peer_decref(conn->ibc_peer); + + PORTAL_FREE(conn, sizeof (*conn)); + + atomic_dec(&kibnal_data.kib_nconns); + + if (atomic_read (&kibnal_data.kib_nconns) == 0 && + kibnal_data.kib_shutdown) { + /* I just nuked the last connection on shutdown; wake up + * everyone so they can exit. */ + wake_up_all(&kibnal_data.kib_sched_waitq); + wake_up_all(&kibnal_data.kib_connd_waitq); + } +} + +void +kibnal_put_conn (kib_conn_t *conn) +{ + unsigned long flags; + + CDEBUG (D_NET, "putting conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + + LASSERT (atomic_read (&conn->ibc_refcount) > 0); + if (!atomic_dec_and_test (&conn->ibc_refcount)) + return; + + /* must disconnect before dropping the final ref */ + LASSERT (conn->ibc_state == IBNAL_CONN_DISCONNECTED); + + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + + list_add (&conn->ibc_list, &kibnal_data.kib_connd_conns); + wake_up (&kibnal_data.kib_connd_waitq); + + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); +} + +static int +kibnal_close_peer_conns_locked (kib_peer_t *peer, int why) +{ + kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + int count = 0; + + list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry (ctmp, kib_conn_t, ibc_list); + + count++; + kibnal_close_conn_locked (conn, why); + } + + return (count); +} + +int +kibnal_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation) +{ + kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + int count = 0; + + list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry (ctmp, kib_conn_t, ibc_list); + + if (conn->ibc_incarnation == incarnation) + continue; + + CDEBUG(D_NET, "Closing stale conn nid:"LPX64" incarnation:"LPX64"("LPX64")\n", + peer->ibp_nid, conn->ibc_incarnation, incarnation); + + count++; + kibnal_close_conn_locked (conn, -ESTALE); + } + + return (count); +} + +static int +kibnal_close_matching_conns (ptl_nid_t nid) +{ + unsigned long flags; + kib_peer_t *peer; + struct list_head *ptmp; + struct list_head *pnxt; + int lo; + int hi; + int i; + int count = 0; + + write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + + if (nid != PTL_NID_ANY) + lo = hi = kibnal_nid2peerlist(nid) - kibnal_data.kib_peers; + else { + lo = 0; + hi = kibnal_data.kib_peer_hash_size - 1; + } + + for (i = lo; i <= hi; i++) { + list_for_each_safe (ptmp, pnxt, &kibnal_data.kib_peers[i]) { + + peer = list_entry (ptmp, kib_peer_t, ibp_list); + LASSERT (peer->ibp_persistence != 0 || + peer->ibp_connecting != 0 || + !list_empty (&peer->ibp_conns)); + + if (!(nid == PTL_NID_ANY || nid == peer->ibp_nid)) + continue; + + count += kibnal_close_peer_conns_locked (peer, 0); + } + } + + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + + /* wildcards always succeed */ + if (nid == PTL_NID_ANY) + return (0); + + return (count == 0 ? -ENOENT : 0); +} + +static int +kibnal_cmd(struct portals_cfg *pcfg, void * private) +{ + int rc = -EINVAL; + ENTRY; + + LASSERT (pcfg != NULL); + + switch(pcfg->pcfg_command) { + case NAL_CMD_GET_PEER: { + ptl_nid_t nid = 0; + int share_count = 0; + + rc = kibnal_get_peer_info(pcfg->pcfg_count, + &nid, &share_count); + pcfg->pcfg_nid = nid; + pcfg->pcfg_size = 0; + pcfg->pcfg_id = 0; + pcfg->pcfg_misc = 0; + pcfg->pcfg_count = 0; + pcfg->pcfg_wait = share_count; + break; + } + case NAL_CMD_ADD_PEER: { + rc = kibnal_add_persistent_peer (pcfg->pcfg_nid); + break; + } + case NAL_CMD_DEL_PEER: { + rc = kibnal_del_peer (pcfg->pcfg_nid, + /* flags == single_share */ + pcfg->pcfg_flags != 0); + break; + } + case NAL_CMD_GET_CONN: { + kib_conn_t *conn = kibnal_get_conn_by_idx (pcfg->pcfg_count); + + if (conn == NULL) + rc = -ENOENT; + else { + rc = 0; + pcfg->pcfg_nid = conn->ibc_peer->ibp_nid; + pcfg->pcfg_id = 0; + pcfg->pcfg_misc = 0; + pcfg->pcfg_flags = 0; + kibnal_put_conn (conn); + } + break; + } + case NAL_CMD_CLOSE_CONNECTION: { + rc = kibnal_close_matching_conns (pcfg->pcfg_nid); + break; + } + case NAL_CMD_REGISTER_MYNID: { + if (pcfg->pcfg_nid == PTL_NID_ANY) + rc = -EINVAL; + else + rc = kibnal_set_mynid (pcfg->pcfg_nid); + break; + } + } + + RETURN(rc); +} + +void +kibnal_free_pages (kib_pages_t *p) +{ + int npages = p->ibp_npages; + int rc; + int i; + + if (p->ibp_mapped) { + rc = iibt_deregister_memory(p->ibp_handle); + if (rc != 0) + CERROR ("Deregister error: %d\n", rc); + } + + for (i = 0; i < npages; i++) + if (p->ibp_pages[i] != NULL) + __free_page(p->ibp_pages[i]); + + PORTAL_FREE (p, offsetof(kib_pages_t, ibp_pages[npages])); +} + +int +kibnal_alloc_pages (kib_pages_t **pp, int npages, int allow_write) +{ + kib_pages_t *p; + __u64 *phys_pages; + int i; + FSTATUS frc; + IB_ACCESS_CONTROL access; + + memset(&access, 0, sizeof(access)); + access.s.MWBindable = 1; + access.s.LocalWrite = 1; + access.s.RdmaRead = 1; + access.s.RdmaWrite = 1; + + PORTAL_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages])); + if (p == NULL) { + CERROR ("Can't allocate buffer %d\n", npages); + return (-ENOMEM); + } + + memset (p, 0, offsetof(kib_pages_t, ibp_pages[npages])); + p->ibp_npages = npages; + + for (i = 0; i < npages; i++) { + p->ibp_pages[i] = alloc_page (GFP_KERNEL); + if (p->ibp_pages[i] == NULL) { + CERROR ("Can't allocate page %d of %d\n", i, npages); + kibnal_free_pages(p); + return (-ENOMEM); + } + } + + if (kibnal_whole_mem()) + goto out; + + PORTAL_ALLOC(phys_pages, npages * sizeof(*phys_pages)); + if (phys_pages == NULL) { + CERROR ("Can't allocate physarray for %d pages\n", npages); + /* XXX free ibp_pages? */ + kibnal_free_pages(p); + return (-ENOMEM); + } + + /* if we were using the _contig_ registration variant we would have + * an array of PhysAddr/Length pairs, but the discontiguous variant + * just takes the PhysAddr */ + for (i = 0; i < npages; i++) + phys_pages[i] = kibnal_page2phys(p->ibp_pages[i]); + + frc = iibt_register_physical_memory(kibnal_data.kib_hca, + 0, /* requested vaddr */ + phys_pages, npages, + 0, /* offset */ + kibnal_data.kib_pd, + access, + &p->ibp_handle, &p->ibp_vaddr, + &p->ibp_lkey, &p->ibp_rkey); + + PORTAL_FREE(phys_pages, npages * sizeof(*phys_pages)); + + if (frc != FSUCCESS) { + CERROR ("Error %d mapping %d pages\n", frc, npages); + kibnal_free_pages(p); + return (-ENOMEM); + } + + CDEBUG(D_NET, "registered %d pages; handle: %p vaddr "LPX64" " + "lkey %x rkey %x\n", npages, p->ibp_handle, + p->ibp_vaddr, p->ibp_lkey, p->ibp_rkey); + + p->ibp_mapped = 1; +out: + *pp = p; + return (0); +} + +static int +kibnal_setup_tx_descs (void) +{ + int ipage = 0; + int page_offset = 0; + __u64 vaddr; + __u64 vaddr_base; + struct page *page; + kib_tx_t *tx; + int i; + int rc; + + /* pre-mapped messages are not bigger than 1 page */ + LASSERT (IBNAL_MSG_SIZE <= PAGE_SIZE); + + /* No fancy arithmetic when we do the buffer calculations */ + LASSERT (PAGE_SIZE % IBNAL_MSG_SIZE == 0); + + rc = kibnal_alloc_pages(&kibnal_data.kib_tx_pages, IBNAL_TX_MSG_PAGES, + 0); + if (rc != 0) + return (rc); + + /* ignored for the whole_mem case */ + vaddr = vaddr_base = kibnal_data.kib_tx_pages->ibp_vaddr; + + for (i = 0; i < IBNAL_TX_MSGS; i++) { + page = kibnal_data.kib_tx_pages->ibp_pages[ipage]; + tx = &kibnal_data.kib_tx_descs[i]; + + memset (tx, 0, sizeof(*tx)); /* zero flags etc */ + + tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) + + page_offset); + + if (kibnal_whole_mem()) + tx->tx_vaddr = kibnal_page2phys(page) + + page_offset + + kibnal_data.kib_md.md_addr; + else + tx->tx_vaddr = vaddr; + + tx->tx_isnblk = (i >= IBNAL_NTX); + tx->tx_mapped = KIB_TX_UNMAPPED; + + CDEBUG(D_NET, "Tx[%d] %p->%p - "LPX64"\n", + i, tx, tx->tx_msg, tx->tx_vaddr); + + if (tx->tx_isnblk) + list_add (&tx->tx_list, + &kibnal_data.kib_idle_nblk_txs); + else + list_add (&tx->tx_list, + &kibnal_data.kib_idle_txs); + + vaddr += IBNAL_MSG_SIZE; + LASSERT (vaddr <= vaddr_base + IBNAL_TX_MSG_BYTES); + + page_offset += IBNAL_MSG_SIZE; + LASSERT (page_offset <= PAGE_SIZE); + + if (page_offset == PAGE_SIZE) { + page_offset = 0; + ipage++; + LASSERT (ipage <= IBNAL_TX_MSG_PAGES); + } + } + + return (0); +} + +static void +kibnal_api_shutdown (nal_t *nal) +{ + int i; + int rc; + + if (nal->nal_refct != 0) { + /* This module got the first ref */ + PORTAL_MODULE_UNUSE; + return; + } + + CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n", + atomic_read (&portal_kmemory)); + + LASSERT(nal == &kibnal_api); + + switch (kibnal_data.kib_init) { + default: + CERROR ("Unexpected state %d\n", kibnal_data.kib_init); + LBUG(); + + case IBNAL_INIT_ALL: + /* stop calls to nal_cmd */ + libcfs_nal_cmd_unregister(IIBNAL); + /* No new peers */ + + /* resetting my NID to unadvertises me, removes my + * listener and nukes all current peers */ + kibnal_set_mynid (PTL_NID_ANY); + + /* Wait for all peer state to clean up (crazy) */ + i = 2; + while (atomic_read (&kibnal_data.kib_npeers) != 0) { + i++; + CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */ + "waiting for %d peers to disconnect (can take a few seconds)\n", + atomic_read (&kibnal_data.kib_npeers)); + set_current_state (TASK_UNINTERRUPTIBLE); + schedule_timeout (HZ); + } + /* fall through */ + + case IBNAL_INIT_CQ: + rc = iibt_cq_destroy(kibnal_data.kib_cq); + if (rc != 0) + CERROR ("Destroy CQ error: %d\n", rc); + /* fall through */ + + case IBNAL_INIT_TXD: + kibnal_free_pages (kibnal_data.kib_tx_pages); + /* fall through */ + + case IBNAL_INIT_MR: + if (kibnal_data.kib_md.md_handle != NULL) { + rc = iibt_deregister_memory(kibnal_data.kib_md.md_handle); + if (rc != FSUCCESS) + CERROR ("Deregister memory: %d\n", rc); + } + /* fall through */ + +#if IBNAL_FMR + case IBNAL_INIT_FMR: + rc = ib_fmr_pool_destroy (kibnal_data.kib_fmr_pool); + if (rc != 0) + CERROR ("Destroy FMR pool error: %d\n", rc); + /* fall through */ +#endif + case IBNAL_INIT_PD: + rc = iibt_pd_free(kibnal_data.kib_pd); + if (rc != 0) + CERROR ("Destroy PD error: %d\n", rc); + /* fall through */ + + case IBNAL_INIT_SD: + rc = iibt_sd_deregister(kibnal_data.kib_sd); + if (rc != 0) + CERROR ("Deregister SD error: %d\n", rc); + /* fall through */ + + case IBNAL_INIT_PORT: + /* XXX ??? */ + /* fall through */ + + case IBNAL_INIT_PORTATTRS: + PORTAL_FREE(kibnal_data.kib_hca_attrs.PortAttributesList, + kibnal_data.kib_hca_attrs.PortAttributesListSize); + /* fall through */ + + case IBNAL_INIT_HCA: + rc = iibt_close_hca(kibnal_data.kib_hca); + if (rc != 0) + CERROR ("Close HCA error: %d\n", rc); + /* fall through */ + + case IBNAL_INIT_LIB: + lib_fini(&kibnal_lib); + /* fall through */ + + case IBNAL_INIT_DATA: + /* Module refcount only gets to zero when all peers + * have been closed so all lists must be empty */ + LASSERT (atomic_read (&kibnal_data.kib_npeers) == 0); + LASSERT (kibnal_data.kib_peers != NULL); + for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) { + LASSERT (list_empty (&kibnal_data.kib_peers[i])); + } + LASSERT (atomic_read (&kibnal_data.kib_nconns) == 0); + LASSERT (list_empty (&kibnal_data.kib_sched_rxq)); + LASSERT (list_empty (&kibnal_data.kib_sched_txq)); + LASSERT (list_empty (&kibnal_data.kib_connd_conns)); + LASSERT (list_empty (&kibnal_data.kib_connd_peers)); + + /* flag threads to terminate; wake and wait for them to die */ + kibnal_data.kib_shutdown = 1; + wake_up_all (&kibnal_data.kib_sched_waitq); + wake_up_all (&kibnal_data.kib_connd_waitq); + + i = 2; + while (atomic_read (&kibnal_data.kib_nthreads) != 0) { + i++; + CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */ + "Waiting for %d threads to terminate\n", + atomic_read (&kibnal_data.kib_nthreads)); + set_current_state (TASK_INTERRUPTIBLE); + schedule_timeout (HZ); + } + /* fall through */ + + case IBNAL_INIT_NOTHING: + break; + } + + if (kibnal_data.kib_tx_descs != NULL) + PORTAL_FREE (kibnal_data.kib_tx_descs, + IBNAL_TX_MSGS * sizeof(kib_tx_t)); + + if (kibnal_data.kib_peers != NULL) + PORTAL_FREE (kibnal_data.kib_peers, + sizeof (struct list_head) * + kibnal_data.kib_peer_hash_size); + + CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n", + atomic_read (&portal_kmemory)); + printk(KERN_INFO "Lustre: Infinicon IB NAL unloaded (final mem %d)\n", + atomic_read(&portal_kmemory)); + + kibnal_data.kib_init = IBNAL_INIT_NOTHING; +} + +#define roundup_power(val, power) \ + ( (val + (__u64)(power - 1)) & ~((__u64)(power - 1)) ) + +/* this isn't very portable or sturdy in the face of funny mem/bus configs */ +static __u64 max_phys_mem(IB_CA_ATTRIBUTES *ca_attr) +{ + struct sysinfo si; + __u64 ret; + + /* XXX we don't bother with first-gen cards */ + if (ca_attr->VendorId == 0xd0b7 && ca_attr->DeviceId == 0x3101) + return 0ULL; + + si_meminfo(&si); + ret = (__u64)max(si.totalram, max_mapnr) * si.mem_unit; + return roundup_power(ret, 128 * 1024 * 1024); +} +#undef roundup_power + +static int +kibnal_api_startup (nal_t *nal, ptl_pid_t requested_pid, + ptl_ni_limits_t *requested_limits, + ptl_ni_limits_t *actual_limits) +{ + ptl_process_id_t process_id; + int pkmem = atomic_read(&portal_kmemory); + IB_PORT_ATTRIBUTES *pattr; + FSTATUS frc; + int rc; + int n; + int i; + + LASSERT (nal == &kibnal_api); + + if (nal->nal_refct != 0) { + if (actual_limits != NULL) + *actual_limits = kibnal_lib.libnal_ni.ni_actual_limits; + /* This module got the first ref */ + PORTAL_MODULE_USE; + return (PTL_OK); + } + + LASSERT (kibnal_data.kib_init == IBNAL_INIT_NOTHING); + + frc = IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2, + &kibnal_data.kib_interfaces); + if (frc != FSUCCESS) { + CERROR("IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2) = %d\n", + frc); + return -ENOSYS; + } + + init_MUTEX (&kibnal_data.kib_nid_mutex); + init_MUTEX_LOCKED (&kibnal_data.kib_nid_signal); + kibnal_data.kib_nid = PTL_NID_ANY; + + rwlock_init(&kibnal_data.kib_global_lock); + + kibnal_data.kib_peer_hash_size = IBNAL_PEER_HASH_SIZE; + PORTAL_ALLOC (kibnal_data.kib_peers, + sizeof (struct list_head) * kibnal_data.kib_peer_hash_size); + if (kibnal_data.kib_peers == NULL) { + goto failed; + } + for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) + INIT_LIST_HEAD(&kibnal_data.kib_peers[i]); + + spin_lock_init (&kibnal_data.kib_connd_lock); + INIT_LIST_HEAD (&kibnal_data.kib_connd_peers); + INIT_LIST_HEAD (&kibnal_data.kib_connd_conns); + init_waitqueue_head (&kibnal_data.kib_connd_waitq); + + spin_lock_init (&kibnal_data.kib_sched_lock); + INIT_LIST_HEAD (&kibnal_data.kib_sched_txq); + INIT_LIST_HEAD (&kibnal_data.kib_sched_rxq); + init_waitqueue_head (&kibnal_data.kib_sched_waitq); + + spin_lock_init (&kibnal_data.kib_tx_lock); + INIT_LIST_HEAD (&kibnal_data.kib_idle_txs); + INIT_LIST_HEAD (&kibnal_data.kib_idle_nblk_txs); + init_waitqueue_head(&kibnal_data.kib_idle_tx_waitq); + + PORTAL_ALLOC (kibnal_data.kib_tx_descs, + IBNAL_TX_MSGS * sizeof(kib_tx_t)); + if (kibnal_data.kib_tx_descs == NULL) { + CERROR ("Can't allocate tx descs\n"); + goto failed; + } + + /* lists/ptrs/locks initialised */ + kibnal_data.kib_init = IBNAL_INIT_DATA; + /*****************************************************/ + + process_id.pid = 0; + process_id.nid = kibnal_data.kib_nid; + + rc = lib_init(&kibnal_lib, nal, process_id, + requested_limits, actual_limits); + if (rc != PTL_OK) { + CERROR("lib_init failed: error %d\n", rc); + goto failed; + } + + /* lib interface initialised */ + kibnal_data.kib_init = IBNAL_INIT_LIB; + /*****************************************************/ + + for (i = 0; i < IBNAL_N_SCHED; i++) { + rc = kibnal_thread_start (kibnal_scheduler, (void *)i); + if (rc != 0) { + CERROR("Can't spawn iibnal scheduler[%d]: %d\n", + i, rc); + goto failed; + } + } + + rc = kibnal_thread_start (kibnal_connd, NULL); + if (rc != 0) { + CERROR ("Can't spawn iibnal connd: %d\n", rc); + goto failed; + } + + n = sizeof(kibnal_data.kib_hca_guids) / + sizeof(kibnal_data.kib_hca_guids[0]); + frc = iibt_get_hca_guids(&n, kibnal_data.kib_hca_guids); + if (frc != FSUCCESS) { + CERROR ("Can't get channel adapter guids: %d\n", frc); + goto failed; + } + if (n == 0) { + CERROR ("No channel adapters found\n"); + goto failed; + } + + /* Infinicon has per-HCA rather than per CQ completion handlers */ + frc = iibt_open_hca(kibnal_data.kib_hca_guids[0], + kibnal_ca_callback, + kibnal_ca_async_callback, + &kibnal_data.kib_hca, + &kibnal_data.kib_hca); + if (frc != FSUCCESS) { + CERROR ("Can't open CA[0]: %d\n", frc); + goto failed; + } + + /* Channel Adapter opened */ + kibnal_data.kib_init = IBNAL_INIT_HCA; + /*****************************************************/ + + kibnal_data.kib_hca_attrs.PortAttributesList = NULL; + kibnal_data.kib_hca_attrs.PortAttributesListSize = 0; + frc = iibt_query_hca(kibnal_data.kib_hca, + &kibnal_data.kib_hca_attrs, NULL); + if (frc != FSUCCESS) { + CERROR ("Can't size port attrs: %d\n", frc); + goto failed; + } + + PORTAL_ALLOC(kibnal_data.kib_hca_attrs.PortAttributesList, + kibnal_data.kib_hca_attrs.PortAttributesListSize); + if (kibnal_data.kib_hca_attrs.PortAttributesList == NULL) + goto failed; + + /* Port attrs allocated */ + kibnal_data.kib_init = IBNAL_INIT_PORTATTRS; + /*****************************************************/ + + frc = iibt_query_hca(kibnal_data.kib_hca, &kibnal_data.kib_hca_attrs, + NULL); + if (frc != FSUCCESS) { + CERROR ("Can't get port attrs for CA 0: %d\n", frc); + goto failed; + } + + for (i = 0, pattr = kibnal_data.kib_hca_attrs.PortAttributesList; + pattr != NULL; + i++, pattr = pattr->Next) { + switch (pattr->PortState) { + default: + CERROR("Unexpected port[%d] state %d\n", + i, pattr->PortState); + continue; + case PortStateDown: + CDEBUG(D_NET, "port[%d] Down\n", i); + continue; + case PortStateInit: + CDEBUG(D_NET, "port[%d] Init\n", i); + continue; + case PortStateArmed: + CDEBUG(D_NET, "port[%d] Armed\n", i); + continue; + + case PortStateActive: + CDEBUG(D_NET, "port[%d] Active\n", i); + kibnal_data.kib_port = i; + kibnal_data.kib_port_guid = pattr->GUID; + kibnal_data.kib_port_pkey = pattr->PkeyTable[0]; + break; + } + break; + } + + if (pattr == NULL) { + CERROR ("Can't find an active port\n"); + goto failed; + } + + CDEBUG(D_NET, "got guid "LPX64"\n", kibnal_data.kib_port_guid); + + /* Active port found */ + kibnal_data.kib_init = IBNAL_INIT_PORT; + /*****************************************************/ + + frc = iibt_sd_register(&kibnal_data.kib_sd, NULL); + if (frc != FSUCCESS) { + CERROR ("Can't register with SD: %d\n", frc); + goto failed; + } + + /* Registered with SD OK */ + kibnal_data.kib_init = IBNAL_INIT_SD; + /*****************************************************/ + + frc = iibt_pd_allocate(kibnal_data.kib_hca, 0, &kibnal_data.kib_pd); + if (frc != FSUCCESS) { + CERROR ("Can't create PD: %d\n", rc); + goto failed; + } + + /* flag PD initialised */ + kibnal_data.kib_init = IBNAL_INIT_PD; + /*****************************************************/ + +#if IBNAL_FMR + { + const int pool_size = IBNAL_NTX + IBNAL_NTX_NBLK; + struct ib_fmr_pool_param params = { + .max_pages_per_fmr = PTL_MTU/PAGE_SIZE, + .access = (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE | + IB_ACCESS_REMOTE_READ), + .pool_size = pool_size, + .dirty_watermark = (pool_size * 3)/4, + .flush_function = NULL, + .flush_arg = NULL, + .cache = 1, + }; + rc = ib_fmr_pool_create(kibnal_data.kib_pd, ¶ms, + &kibnal_data.kib_fmr_pool); + if (rc != 0) { + CERROR ("Can't create FMR pool size %d: %d\n", + pool_size, rc); + goto failed; + } + } + + /* flag FMR pool initialised */ + kibnal_data.kib_init = IBNAL_INIT_FMR; +#endif + /*****************************************************/ + if (IBNAL_WHOLE_MEM) { + IB_MR_PHYS_BUFFER phys; + IB_ACCESS_CONTROL access; + kib_md_t *md = &kibnal_data.kib_md; + + memset(&access, 0, sizeof(access)); + access.s.MWBindable = 1; + access.s.LocalWrite = 1; + access.s.RdmaRead = 1; + access.s.RdmaWrite = 1; + + phys.PhysAddr = 0; + phys.Length = max_phys_mem(&kibnal_data.kib_hca_attrs); + if (phys.Length == 0) { + CERROR ("couldn't determine the end of phys mem\n"); + goto failed; + } + + rc = iibt_register_contig_physical_memory(kibnal_data.kib_hca, + 0, + &phys, 1, + 0, + kibnal_data.kib_pd, + access, + &md->md_handle, + &md->md_addr, + &md->md_lkey, + &md->md_rkey); + if (rc != FSUCCESS) { + CERROR("registering physical memory failed: %d\n", + rc); + CERROR("falling back to registration per-rdma\n"); + md->md_handle = NULL; + } else { + CDEBUG(D_NET, "registered "LPU64" bytes of mem\n", + phys.Length); + kibnal_data.kib_init = IBNAL_INIT_MR; + } + } + + /*****************************************************/ + + rc = kibnal_setup_tx_descs(); + if (rc != 0) { + CERROR ("Can't register tx descs: %d\n", rc); + goto failed; + } + + /* flag TX descs initialised */ + kibnal_data.kib_init = IBNAL_INIT_TXD; + /*****************************************************/ + + { + uint32 nentries; + + frc = iibt_cq_create(kibnal_data.kib_hca, IBNAL_CQ_ENTRIES, + &kibnal_data.kib_cq, &kibnal_data.kib_cq, + &nentries); + if (frc != FSUCCESS) { + CERROR ("Can't create RX CQ: %d\n", frc); + goto failed; + } + + /* flag CQ initialised */ + kibnal_data.kib_init = IBNAL_INIT_CQ; + + if (nentries < IBNAL_CQ_ENTRIES) { + CERROR ("CQ only has %d entries, need %d\n", + nentries, IBNAL_CQ_ENTRIES); + goto failed; + } + + rc = iibt_cq_rearm(kibnal_data.kib_cq, CQEventSelNextWC); + if (rc != 0) { + CERROR ("Failed to re-arm completion queue: %d\n", rc); + goto failed; + } + } + + /*****************************************************/ + + rc = libcfs_nal_cmd_register(IIBNAL, &kibnal_cmd, NULL); + if (rc != 0) { + CERROR ("Can't initialise command interface (rc = %d)\n", rc); + goto failed; + } + + /* flag everything initialised */ + kibnal_data.kib_init = IBNAL_INIT_ALL; + /*****************************************************/ + + printk(KERN_INFO "Lustre: Infinicon IB NAL loaded " + "(initial mem %d)\n", pkmem); + + return (PTL_OK); + + failed: + kibnal_api_shutdown (&kibnal_api); + return (PTL_FAIL); +} + +void __exit +kibnal_module_fini (void) +{ +#ifdef CONFIG_SYSCTL + if (kibnal_tunables.kib_sysctl != NULL) + unregister_sysctl_table (kibnal_tunables.kib_sysctl); +#endif + PtlNIFini(kibnal_ni); + + ptl_unregister_nal(IIBNAL); +} + +int __init +kibnal_module_init (void) +{ + int rc; + + if (sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN) { + CERROR("sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN\n"); + return -EINVAL; + } + + /* the following must be sizeof(int) for proc_dointvec() */ + if (sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)) { + CERROR("sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)\n"); + return -EINVAL; + } + + kibnal_api.nal_ni_init = kibnal_api_startup; + kibnal_api.nal_ni_fini = kibnal_api_shutdown; + + /* Initialise dynamic tunables to defaults once only */ + kibnal_tunables.kib_io_timeout = IBNAL_IO_TIMEOUT; + + rc = ptl_register_nal(IIBNAL, &kibnal_api); + if (rc != PTL_OK) { + CERROR("Can't register IBNAL: %d\n", rc); + return (-ENOMEM); /* or something... */ + } + + /* Pure gateways want the NAL started up at module load time... */ + rc = PtlNIInit(IIBNAL, 0, NULL, NULL, &kibnal_ni); + if (rc != PTL_OK && rc != PTL_IFACE_DUP) { + ptl_unregister_nal(IIBNAL); + return (-ENODEV); + } + +#ifdef CONFIG_SYSCTL + /* Press on regardless even if registering sysctl doesn't work */ + kibnal_tunables.kib_sysctl = + register_sysctl_table (kibnal_top_ctl_table, 0); +#endif + return (0); +} + +MODULE_AUTHOR("Cluster File Systems, Inc. "); +MODULE_DESCRIPTION("Kernel Infinicon IB NAL v0.01"); +MODULE_LICENSE("GPL"); + +module_init(kibnal_module_init); +module_exit(kibnal_module_fini); + diff --git a/lnet/klnds/iiblnd/iiblnd.h b/lnet/klnds/iiblnd/iiblnd.h new file mode 100644 index 0000000..0a25a9a --- /dev/null +++ b/lnet/klnds/iiblnd/iiblnd.h @@ -0,0 +1,892 @@ +/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- + * vim:expandtab:shiftwidth=8:tabstop=8: + * + * Copyright (C) 2004 Cluster File Systems, Inc. + * Author: Eric Barton + * + * This file is part of Lustre, http://www.lustre.org. + * + * Lustre is free software; you can redistribute it and/or + * modify it under the terms of version 2 of the GNU General Public + * License as published by the Free Software Foundation. + * + * Lustre 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 Lustre; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * + */ + +#ifndef EXPORT_SYMTAB +# define EXPORT_SYMTAB +#endif + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include + +#define DEBUG_SUBSYSTEM S_IBNAL + +#include +#include +#include +#include + +#include + +#define GCC_VERSION (__GNUC__ * 10000 \ + + __GNUC_MINOR__ * 100 \ + + __GNUC_PATCHLEVEL__) + +/* Test for GCC > 3.2.2 */ +#if GCC_VERSION <= 30202 +/* GCC 3.2.2, and presumably several versions before it, will + * miscompile this driver. See + * http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9853. */ +#error Invalid GCC version. Must use GCC >= 3.2.3 +#endif + +#define IBNAL_SERVICE_NAME "iibnal" +#define IBNAL_SERVICE_NUMBER 0x11b9a1 + +#if CONFIG_SMP +# define IBNAL_N_SCHED num_online_cpus() /* # schedulers */ +#else +# define IBNAL_N_SCHED 1 /* # schedulers */ +#endif + +#define IBNAL_MIN_RECONNECT_INTERVAL HZ /* first failed connection retry... */ +#define IBNAL_MAX_RECONNECT_INTERVAL (60*HZ) /* ...exponentially increasing to this */ + +#define IBNAL_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */ + +#define IBNAL_MSG_QUEUE_SIZE 8 /* # messages/RDMAs in-flight */ +#define IBNAL_CREDIT_HIGHWATER 7 /* when to eagerly return credits */ +/* 7 indicates infinite retry attempts, Infinicon recommended 5 */ +#define IBNAL_RETRY 5 /* # times to retry */ +#define IBNAL_RNR_RETRY 5 /* */ +#define IBNAL_CM_RETRY 5 /* # times to retry connection */ +#define IBNAL_FLOW_CONTROL 1 +#define IBNAL_ACK_TIMEOUT 20 /* supposedly 4 secs */ + +#define IBNAL_NTX 64 /* # tx descs */ +/* this had to be dropped down so that we only register < 255 pages per + * region. this will change if we register all memory. */ +#define IBNAL_NTX_NBLK 128 /* # reserved tx descs */ + +#define IBNAL_PEER_HASH_SIZE 101 /* # peer lists */ + +#define IBNAL_RESCHED 100 /* # scheduler loops before reschedule */ + +#define IBNAL_CONCURRENT_PEERS 1000 /* # nodes all talking at once to me */ + +/* default vals for runtime tunables */ +#define IBNAL_IO_TIMEOUT 50 /* default comms timeout (seconds) */ + +/************************/ +/* derived constants... */ + +/* TX messages (shared by all connections) */ +#define IBNAL_TX_MSGS (IBNAL_NTX + IBNAL_NTX_NBLK) +#define IBNAL_TX_MSG_BYTES (IBNAL_TX_MSGS * IBNAL_MSG_SIZE) +#define IBNAL_TX_MSG_PAGES ((IBNAL_TX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) + +#define IBNAL_TX_MAX_SG (PTL_MD_MAX_IOV + 1) + +/* RX messages (per connection) */ +#define IBNAL_RX_MSGS IBNAL_MSG_QUEUE_SIZE +#define IBNAL_RX_MSG_BYTES (IBNAL_RX_MSGS * IBNAL_MSG_SIZE) +#define IBNAL_RX_MSG_PAGES ((IBNAL_RX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) + + +/* we may have up to 2 completions per transmit + + 1 completion per receive, per connection */ +#define IBNAL_CQ_ENTRIES ((2*IBNAL_TX_MSGS) + \ + (IBNAL_RX_MSGS * IBNAL_CONCURRENT_PEERS)) + +#define IBNAL_RDMA_BASE 0x0eeb0000 +#define IBNAL_FMR 0 +#define IBNAL_WHOLE_MEM 1 +#define IBNAL_CKSUM 0 +//#define IBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_PROCESS +#define IBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_INTERRUPT + +/* XXX I have no idea. */ +#define IBNAL_STARTING_PSN 1 + +typedef struct +{ + int kib_io_timeout; /* comms timeout (seconds) */ + struct ctl_table_header *kib_sysctl; /* sysctl interface */ +} kib_tunables_t; + +/* some of these have specific types in the stack that just map back + * to the uFOO types, like IB_{L,R}_KEY. */ +typedef struct +{ + int ibp_npages; /* # pages */ + int ibp_mapped; /* mapped? */ + __u64 ibp_vaddr; /* mapped region vaddr */ + __u32 ibp_lkey; /* mapped region lkey */ + __u32 ibp_rkey; /* mapped region rkey */ + IB_HANDLE ibp_handle; /* mapped region handle */ + struct page *ibp_pages[0]; +} kib_pages_t; + +typedef struct +{ + IB_HANDLE md_handle; + __u32 md_lkey; + __u32 md_rkey; + __u64 md_addr; +} kib_md_t __attribute__((packed)); + +typedef struct +{ + int kib_init; /* initialisation state */ + __u64 kib_incarnation; /* which one am I */ + int kib_shutdown; /* shut down? */ + atomic_t kib_nthreads; /* # live threads */ + + __u64 kib_service_id; /* service number I listen on */ + __u64 kib_port_guid; /* my GUID (lo 64 of GID)*/ + __u16 kib_port_pkey; /* my pkey, whatever that is */ + ptl_nid_t kib_nid; /* my NID */ + struct semaphore kib_nid_mutex; /* serialise NID ops */ + struct semaphore kib_nid_signal; /* signal completion */ + IB_HANDLE kib_cep; /* connection end point */ + + rwlock_t kib_global_lock; /* stabilize peer/conn ops */ + + struct list_head *kib_peers; /* hash table of all my known peers */ + int kib_peer_hash_size; /* size of kib_peers */ + atomic_t kib_npeers; /* # peers extant */ + atomic_t kib_nconns; /* # connections extant */ + + struct list_head kib_connd_conns; /* connections to progress */ + struct list_head kib_connd_peers; /* peers waiting for a connection */ + wait_queue_head_t kib_connd_waitq; /* connection daemons sleep here */ + unsigned long kib_connd_waketime; /* when connd will wake */ + spinlock_t kib_connd_lock; /* serialise */ + + wait_queue_head_t kib_sched_waitq; /* schedulers sleep here */ + struct list_head kib_sched_txq; /* tx requiring attention */ + struct list_head kib_sched_rxq; /* rx requiring attention */ + spinlock_t kib_sched_lock; /* serialise */ + + struct kib_tx *kib_tx_descs; /* all the tx descriptors */ + kib_pages_t *kib_tx_pages; /* premapped tx msg pages */ + + struct list_head kib_idle_txs; /* idle tx descriptors */ + struct list_head kib_idle_nblk_txs; /* idle reserved tx descriptors */ + wait_queue_head_t kib_idle_tx_waitq; /* block here for tx descriptor */ + __u64 kib_next_tx_cookie; /* RDMA completion cookie */ + spinlock_t kib_tx_lock; /* serialise */ + + IB_HANDLE kib_hca; /* The HCA */ + int kib_port; /* port on the device */ + IB_HANDLE kib_pd; /* protection domain */ + IB_HANDLE kib_sd; /* SD handle */ + IB_HANDLE kib_cq; /* completion queue */ + kib_md_t kib_md; /* full-mem registration */ + + void *kib_listen_handle; /* where I listen for connections */ + + IBT_INTERFACE_UNION kib_interfaces; /* The Infinicon IBT interface */ + + uint64 kib_hca_guids[8]; /* all the HCA guids */ + IB_CA_ATTRIBUTES kib_hca_attrs; /* where to get HCA attrs */ + FABRIC_OPERATION_DATA kib_fabopdata; /* (un)advertise service record */ +} kib_data_t; + +#define IBNAL_INIT_NOTHING 0 +#define IBNAL_INIT_DATA 1 +#define IBNAL_INIT_LIB 2 +#define IBNAL_INIT_HCA 3 +#define IBNAL_INIT_PORTATTRS 4 +#define IBNAL_INIT_PORT 5 +#define IBNAL_INIT_SD 6 +#define IBNAL_INIT_PD 7 +#define IBNAL_INIT_FMR 8 +#define IBNAL_INIT_MR 9 +#define IBNAL_INIT_TXD 10 +#define IBNAL_INIT_CQ 11 +#define IBNAL_INIT_ALL 12 + +/************************************************************************ + * Wire message structs. + * These are sent in sender's byte order (i.e. receiver flips). + * CAVEAT EMPTOR: other structs communicated between nodes (e.g. MAD + * private data and SM service info), is LE on the wire. + */ + +/* also kib_md_t above */ + +typedef struct +{ + __u32 rd_key; /* remote key */ + __u32 rd_nob; /* # of bytes */ + __u64 rd_addr; /* remote io vaddr */ +} kib_rdma_desc_t __attribute__((packed)); + +typedef struct +{ + ptl_hdr_t ibim_hdr; /* portals header */ + char ibim_payload[0]; /* piggy-backed payload */ +} kib_immediate_msg_t __attribute__((packed)); + +/* these arrays serve two purposes during rdma. they are built on the passive + * side and sent to the active side as remote arguments. On the active side + * the descs are used as a data structure on the way to local gather items. + * the different roles result in split local/remote meaning of desc->rd_key */ +typedef struct +{ + ptl_hdr_t ibrm_hdr; /* portals header */ + __u64 ibrm_cookie; /* opaque completion cookie */ + __u32 ibrm_num_descs; /* how many descs */ + kib_rdma_desc_t ibrm_desc[0]; /* where to suck/blow */ +} kib_rdma_msg_t __attribute__((packed)); + +#define kib_rdma_msg_len(num_descs) \ + offsetof(kib_msg_t, ibm_u.rdma.ibrm_desc[num_descs]) + +typedef struct +{ + __u64 ibcm_cookie; /* opaque completion cookie */ + __u32 ibcm_status; /* completion status */ +} kib_completion_msg_t __attribute__((packed)); + +typedef struct +{ + __u32 ibm_magic; /* I'm an openibnal message */ + __u16 ibm_version; /* this is my version number */ + __u8 ibm_type; /* msg type */ + __u8 ibm_credits; /* returned credits */ +#if IBNAL_CKSUM + __u32 ibm_nob; + __u32 ibm_cksum; +#endif + union { + kib_immediate_msg_t immediate; + kib_rdma_msg_t rdma; + kib_completion_msg_t completion; + } ibm_u __attribute__((packed)); +} kib_msg_t __attribute__((packed)); + +#define IBNAL_MSG_MAGIC 0x0be91b91 /* unique magic */ +#define IBNAL_MSG_VERSION 1 /* current protocol version */ + +#define IBNAL_MSG_NOOP 0xd0 /* nothing (just credits) */ +#define IBNAL_MSG_IMMEDIATE 0xd1 /* portals hdr + payload */ +#define IBNAL_MSG_PUT_RDMA 0xd2 /* portals PUT hdr + source rdma desc */ +#define IBNAL_MSG_PUT_DONE 0xd3 /* signal PUT rdma completion */ +#define IBNAL_MSG_GET_RDMA 0xd4 /* portals GET hdr + sink rdma desc */ +#define IBNAL_MSG_GET_DONE 0xd5 /* signal GET rdma completion */ + +/***********************************************************************/ + +typedef struct kib_rx /* receive message */ +{ + struct list_head rx_list; /* queue for attention */ + struct kib_conn *rx_conn; /* owning conn */ + int rx_rdma; /* RDMA completion posted? */ + int rx_posted; /* posted? */ + __u64 rx_vaddr; /* pre-mapped buffer (hca vaddr) */ + kib_msg_t *rx_msg; /* pre-mapped buffer (host vaddr) */ + IB_WORK_REQ rx_wrq; + IB_LOCAL_DATASEGMENT rx_gl; /* and it's memory */ +} kib_rx_t; + +typedef struct kib_tx /* transmit message */ +{ + struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */ + int tx_isnblk; /* I'm reserved for non-blocking sends */ + struct kib_conn *tx_conn; /* owning conn */ + int tx_mapped; /* mapped for RDMA? */ + int tx_sending; /* # tx callbacks outstanding */ + int tx_status; /* completion status */ + unsigned long tx_deadline; /* completion deadline */ + int tx_passive_rdma; /* peer sucks/blows */ + int tx_passive_rdma_wait; /* waiting for peer to complete */ + __u64 tx_passive_rdma_cookie; /* completion cookie */ + lib_msg_t *tx_libmsg[2]; /* lib msgs to finalize on completion */ + kib_md_t tx_md; /* RDMA mapping (active/passive) */ + __u64 tx_vaddr; /* pre-mapped buffer (hca vaddr) */ + kib_msg_t *tx_msg; /* pre-mapped buffer (host vaddr) */ + int tx_nsp; /* # send work items */ + IB_WORK_REQ tx_wrq[IBNAL_TX_MAX_SG]; /* send work items... */ + IB_LOCAL_DATASEGMENT tx_gl[IBNAL_TX_MAX_SG]; /* ...and their memory */ +} kib_tx_t; + +#define KIB_TX_UNMAPPED 0 +#define KIB_TX_MAPPED 1 +#define KIB_TX_MAPPED_FMR 2 + +typedef struct kib_wire_connreq +{ + __u32 wcr_magic; /* I'm an openibnal connreq */ + __u16 wcr_version; /* this is my version number */ + __u16 wcr_queue_depth; /* this is my receive queue size */ + __u64 wcr_nid; /* peer's NID */ + __u64 wcr_incarnation; /* peer's incarnation */ +} kib_wire_connreq_t; + +typedef struct kib_gid +{ + __u64 hi, lo; +} kib_gid_t; + +typedef struct kib_connreq +{ + /* connection-in-progress */ + struct kib_conn *cr_conn; + kib_wire_connreq_t cr_wcr; + __u64 cr_tid; + IB_SERVICE_RECORD cr_service; + kib_gid_t cr_gid; + IB_PATH_RECORD cr_path; + CM_REQUEST_INFO cr_cmreq; + CM_CONN_INFO cr_discarded; + CM_REJECT_INFO cr_rej_info; +} kib_connreq_t; + +typedef struct kib_conn +{ + struct kib_peer *ibc_peer; /* owning peer */ + struct list_head ibc_list; /* stash on peer's conn list */ + __u64 ibc_incarnation; /* which instance of the peer */ + atomic_t ibc_refcount; /* # users */ + int ibc_state; /* what's happening */ + atomic_t ibc_nob; /* # bytes buffered */ + int ibc_nsends_posted; /* # uncompleted sends */ + int ibc_credits; /* # credits I have */ + int ibc_outstanding_credits; /* # credits to return */ + int ibc_rcvd_disconnect;/* received discon request */ + int ibc_sent_disconnect;/* sent discon request */ + struct list_head ibc_tx_queue; /* send queue */ + struct list_head ibc_active_txs; /* active tx awaiting completion */ + spinlock_t ibc_lock; /* serialise */ + kib_rx_t *ibc_rxs; /* the rx descs */ + kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */ + IB_HANDLE ibc_qp; /* queue pair */ + IB_HANDLE ibc_cep; /* connection ID? */ + IB_QP_ATTRIBUTES_QUERY ibc_qp_attrs; /* QP attrs */ + kib_connreq_t *ibc_connreq; /* connection request state */ +} kib_conn_t; + +#define IBNAL_CONN_INIT_NOTHING 0 /* initial state */ +#define IBNAL_CONN_INIT_QP 1 /* ibc_qp set up */ +#define IBNAL_CONN_CONNECTING 2 /* started to connect */ +#define IBNAL_CONN_ESTABLISHED 3 /* connection established */ +#define IBNAL_CONN_SEND_DREQ 4 /* to send disconnect req */ +#define IBNAL_CONN_DREQ 5 /* sent disconnect req */ +#define IBNAL_CONN_DREP 6 /* sent disconnect rep */ +#define IBNAL_CONN_DISCONNECTED 7 /* no more QP or CM traffic */ + +#define KIB_ASSERT_CONN_STATE(conn, state) do { \ + LASSERTF((conn)->ibc_state == state, "%d\n", conn->ibc_state); \ +} while (0) + +#define KIB_ASSERT_CONN_STATE_RANGE(conn, low, high) do { \ + LASSERTF(low <= high, "%d %d\n", low, high); \ + LASSERTF((conn)->ibc_state >= low && (conn)->ibc_state <= high, \ + "%d\n", conn->ibc_state); \ +} while (0) + +typedef struct kib_peer +{ + struct list_head ibp_list; /* stash on global peer list */ + struct list_head ibp_connd_list; /* schedule on kib_connd_peers */ + ptl_nid_t ibp_nid; /* who's on the other end(s) */ + atomic_t ibp_refcount; /* # users */ + int ibp_persistence; /* "known" peer refs */ + struct list_head ibp_conns; /* all active connections */ + struct list_head ibp_tx_queue; /* msgs waiting for a conn */ + int ibp_connecting; /* connecting+accepting */ + unsigned long ibp_reconnect_time; /* when reconnect may be attempted */ + unsigned long ibp_reconnect_interval; /* exponential backoff */ +} kib_peer_t; + + +extern lib_nal_t kibnal_lib; +extern kib_data_t kibnal_data; +extern kib_tunables_t kibnal_tunables; + +/******************************************************************************/ +/* Infinicon IBT interface wrappers */ +#define IIBT_IF (kibnal_data.kib_interfaces.ver2) + +static inline FSTATUS +iibt_get_hca_guids(uint32 *hca_count, EUI64 *hca_guid_list) +{ + return IIBT_IF.GetCaGuids(hca_count, hca_guid_list); +} + +static inline FSTATUS +iibt_open_hca(EUI64 hca_guid, + IB_COMPLETION_CALLBACK completion_callback, + IB_ASYNC_EVENT_CALLBACK async_event_callback, + void *arg, + IB_HANDLE *handle) +{ + return IIBT_IF.Vpi.OpenCA(hca_guid, completion_callback, + async_event_callback, arg, handle); +} + +static inline FSTATUS +iibt_query_hca(IB_HANDLE hca_handle, IB_CA_ATTRIBUTES *attrs, void **argp) +{ + return IIBT_IF.Vpi.QueryCA(hca_handle, attrs, argp); +} + +static inline FSTATUS +iibt_close_hca(IB_HANDLE hca_handle) +{ + return IIBT_IF.Vpi.CloseCA(hca_handle); +} + +static inline FSTATUS +iibt_pd_allocate(IB_HANDLE hca_handle, __u32 max_avs, IB_HANDLE *pd_handle) +{ + return IIBT_IF.Vpi.AllocatePD(hca_handle, max_avs, pd_handle); +} + +static inline FSTATUS +iibt_pd_free(IB_HANDLE pd_handle) +{ + return IIBT_IF.Vpi.FreePD(pd_handle); +} + +static inline FSTATUS +iibt_register_physical_memory(IB_HANDLE hca_handle, + IB_VIRT_ADDR requested_io_va, + void *phys_buffers, uint64 nphys_buffers, + uint32 io_va_offset, IB_HANDLE pd_handle, + IB_ACCESS_CONTROL access, + IB_HANDLE *mem_handle, + IB_VIRT_ADDR *actual_io_va, + IB_L_KEY *lkey, IB_R_KEY *rkey) +{ + return IIBT_IF.Vpi.RegisterPhysMemRegion(hca_handle, requested_io_va, + phys_buffers, nphys_buffers, + io_va_offset, pd_handle, + access, + mem_handle, actual_io_va, + lkey, rkey); +} + +static inline FSTATUS +iibt_register_contig_physical_memory(IB_HANDLE hca_handle, + IB_VIRT_ADDR requested_io_va, + IB_MR_PHYS_BUFFER *phys_buffers, + uint64 nphys_buffers, + uint32 io_va_offset, IB_HANDLE pd_handle, + IB_ACCESS_CONTROL access, + IB_HANDLE *mem_handle, + IB_VIRT_ADDR *actual_io_va, + IB_L_KEY *lkey, IB_R_KEY *rkey) +{ + return IIBT_IF.Vpi.RegisterContigPhysMemRegion(hca_handle, + requested_io_va, + phys_buffers, + nphys_buffers, + io_va_offset, pd_handle, + access, + mem_handle, actual_io_va, + lkey, rkey); +} + +static inline FSTATUS +iibt_register_memory(IB_HANDLE hca_handle, + void *virt_addr, unsigned int length, + IB_HANDLE pd_handle, + IB_ACCESS_CONTROL access, + IB_HANDLE *mem_handle, + IB_L_KEY *lkey, IB_R_KEY *rkey) +{ + return IIBT_IF.Vpi.RegisterMemRegion(hca_handle, + virt_addr, length, + pd_handle, + access, + mem_handle, + lkey, rkey); +} + +static inline FSTATUS +iibt_deregister_memory(IB_HANDLE mem_handle) +{ + return IIBT_IF.Vpi.DeregisterMemRegion(mem_handle); +} + +static inline FSTATUS +iibt_cq_create(IB_HANDLE hca_handle, uint32 requested_size, + void *arg, IB_HANDLE *cq_handle, uint32 *actual_size) +{ + return IIBT_IF.Vpi.CreateCQ(hca_handle, requested_size, + arg, cq_handle, actual_size); +} + +static inline FSTATUS +iibt_cq_poll(IB_HANDLE cq_handle, IB_WORK_COMPLETION *wc) +{ + return IIBT_IF.Vpi.PollCQ(cq_handle, wc); +} + +static inline FSTATUS +iibt_cq_rearm(IB_HANDLE cq_handle, IB_CQ_EVENT_SELECT select) +{ + return IIBT_IF.Vpi.RearmCQ(cq_handle, select); +} + +static inline FSTATUS +iibt_cq_destroy(IB_HANDLE cq_handle) +{ + return IIBT_IF.Vpi.DestroyCQ(cq_handle); +} + +static inline FSTATUS +iibt_qp_create(IB_HANDLE hca_handle, IB_QP_ATTRIBUTES_CREATE *create_attr, + void *arg, IB_HANDLE *cq_handle, + IB_QP_ATTRIBUTES_QUERY *query_attr) +{ + return IIBT_IF.Vpi.CreateQP(hca_handle, create_attr, arg, cq_handle, + query_attr); +} + +static inline FSTATUS +iibt_qp_query(IB_HANDLE qp_handle, IB_QP_ATTRIBUTES_QUERY *query_attr, + void **arg_ptr) +{ + return IIBT_IF.Vpi.QueryQP(qp_handle, query_attr, arg_ptr); +} + +static inline FSTATUS +iibt_qp_modify(IB_HANDLE qp_handle, IB_QP_ATTRIBUTES_MODIFY *modify_attr, + IB_QP_ATTRIBUTES_QUERY *query_attr) +{ + return IIBT_IF.Vpi.ModifyQP(qp_handle, modify_attr, query_attr); +} + +static inline FSTATUS +iibt_qp_destroy(IB_HANDLE qp_handle) +{ + return IIBT_IF.Vpi.DestroyQP(qp_handle); +} + +static inline FSTATUS +iibt_postrecv(IB_HANDLE qp_handle, IB_WORK_REQ *work_req) +{ + return IIBT_IF.Vpi.PostRecv(qp_handle, work_req); +} + +static inline FSTATUS +iibt_postsend(IB_HANDLE qp_handle, IB_WORK_REQ *work_req) +{ + return IIBT_IF.Vpi.PostSend(qp_handle, work_req); +} + +static inline FSTATUS +iibt_sd_register(IB_HANDLE *sd_handle, CLIENT_CONTROL_PARAMETERS *p) +{ + return IIBT_IF.Sdi.Register(sd_handle, p); +} + +static inline FSTATUS +iibt_sd_deregister(IB_HANDLE sd_handle) +{ + return IIBT_IF.Sdi.Deregister(sd_handle); +} + +static inline FSTATUS +iibt_sd_port_fabric_operation(IB_HANDLE sd_handle, EUI64 port_guid, + FABRIC_OPERATION_DATA *fod, + PFABRIC_OPERATION_CALLBACK callback, + COMMAND_CONTROL_PARAMETERS *p, void *arg) +{ + return IIBT_IF.Sdi.PortFabricOperation(sd_handle, port_guid, + fod, callback, p, arg); +} + +static inline FSTATUS +iibt_sd_query_port_fabric_information(IB_HANDLE sd_handle, EUI64 port_guid, + QUERY *qry, + PQUERY_CALLBACK callback, + COMMAND_CONTROL_PARAMETERS *p, void *arg) +{ + return IIBT_IF.Sdi.QueryPortFabricInformation(sd_handle, port_guid, + qry, callback, p, arg); +} + +static inline IB_HANDLE +iibt_cm_create_cep(CM_CEP_TYPE type) +{ + return IIBT_IF.Cmi.CmCreateCEP(type); +} + +static inline FSTATUS +iibt_cm_modify_cep(IB_HANDLE cep, uint32 attr, char* value, uint32 len, + uint32 offset) +{ + return IIBT_IF.Cmi.CmModifyCEP(cep, attr, value, len, offset); +} + +static inline FSTATUS +iibt_cm_destroy_cep(IB_HANDLE cep_handle) +{ + return IIBT_IF.Cmi.CmDestroyCEP(cep_handle); +} + +static inline FSTATUS +iibt_cm_listen(IB_HANDLE cep, CM_LISTEN_INFO *info, + PFN_CM_CALLBACK callback, void *arg) +{ + return IIBT_IF.Cmi.CmListen(cep, info, callback, arg); +} + +static inline FSTATUS +iibt_cm_cancel(IB_HANDLE cep) +{ + return IIBT_IF.Cmi.CmCancel(cep); +} + +static inline FSTATUS +iibt_cm_accept(IB_HANDLE cep, + CM_CONN_INFO *send_info, CM_CONN_INFO *recv_info, + PFN_CM_CALLBACK callback, void *arg, + IB_HANDLE *new_cep) +{ + return IIBT_IF.Cmi.CmAccept(cep, + send_info, recv_info, + callback, arg, new_cep); +} + +static inline FSTATUS +iibt_cm_reject(IB_HANDLE cep, CM_REJECT_INFO *rej) +{ + return IIBT_IF.Cmi.CmReject(cep, rej); +} + +static inline FSTATUS +iibt_cm_disconnect(IB_HANDLE cep, CM_DREQUEST_INFO *req, + CM_DREPLY_INFO *reply) +{ + return IIBT_IF.Cmi.CmDisconnect(cep, req, reply); +} + +static inline FSTATUS +iibt_cm_connect (IB_HANDLE cep, CM_REQUEST_INFO *req, + PFN_CM_CALLBACK callback, void *arg) +{ + return IIBT_IF.Cmi.CmConnect (cep, req, callback, arg); +} + +static inline int wrq_signals_completion(IB_WORK_REQ *wrq) +{ + return wrq->Req.SendRC.Options.s.SignaledCompletion == 1; +} + + +/******************************************************************************/ + +/* these are purposely avoiding using local vars so they don't increase + * stack consumption. */ + +#define kib_peer_addref(peer) do { \ + LASSERTF(atomic_read(&peer->ibp_refcount) > 0, "%d\n", \ + atomic_read(&peer->ibp_refcount)); \ + CDEBUG(D_NET, "++peer[%p] -> "LPX64" (%d)\n", \ + peer, peer->ibp_nid, atomic_read (&peer->ibp_refcount)); \ + atomic_inc(&peer->ibp_refcount); \ +} while (0) + +#define kib_peer_decref(peer) do { \ + LASSERTF(atomic_read(&peer->ibp_refcount) > 0, "%d\n", \ + atomic_read(&peer->ibp_refcount)); \ + CDEBUG(D_NET, "--peer[%p] -> "LPX64" (%d)\n", \ + peer, peer->ibp_nid, atomic_read (&peer->ibp_refcount)); \ + if (atomic_dec_and_test (&peer->ibp_refcount)) { \ + CDEBUG (D_NET, "destroying peer "LPX64" %p\n", \ + peer->ibp_nid, peer); \ + kibnal_destroy_peer (peer); \ + } \ +} while (0) + +/******************************************************************************/ + +static inline struct list_head * +kibnal_nid2peerlist (ptl_nid_t nid) +{ + unsigned int hash = ((unsigned int)nid) % kibnal_data.kib_peer_hash_size; + + return (&kibnal_data.kib_peers [hash]); +} + +static inline int +kibnal_peer_active(kib_peer_t *peer) +{ + /* Am I in the peer hash table? */ + return (!list_empty(&peer->ibp_list)); +} + +static inline void +kibnal_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn) +{ + /* CAVEAT EMPTOR: tx takes caller's ref on conn */ + + LASSERT (tx->tx_nsp > 0); /* work items set up */ + LASSERT (tx->tx_conn == NULL); /* only set here */ + + tx->tx_conn = conn; + tx->tx_deadline = jiffies + kibnal_tunables.kib_io_timeout * HZ; + list_add_tail(&tx->tx_list, &conn->ibc_tx_queue); +} + +#define KIBNAL_SERVICE_KEY_MASK (IB_SERVICE_RECORD_COMP_SERVICENAME | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_1 | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_2 | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_3 | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_4 | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_5 | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_6 | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_7 | \ + IB_SERVICE_RECORD_COMP_SERVICEDATA8_8) + +static inline __u64* +kibnal_service_nid_field(IB_SERVICE_RECORD *srv) +{ + /* must be consistent with KIBNAL_SERVICE_KEY_MASK */ + return (__u64 *)srv->ServiceData8; +} + + +static inline void +kibnal_set_service_keys(IB_SERVICE_RECORD *srv, ptl_nid_t nid) +{ + LASSERT (strlen(IBNAL_SERVICE_NAME) < sizeof(srv->ServiceName)); + memset (srv->ServiceName, 0, sizeof(srv->ServiceName)); + strcpy (srv->ServiceName, IBNAL_SERVICE_NAME); + + *kibnal_service_nid_field(srv) = cpu_to_le64(nid); +} + +#if 0 +static inline void +kibnal_show_rdma_attr (kib_conn_t *conn) +{ + struct ib_qp_attribute qp_attr; + int rc; + + memset (&qp_attr, 0, sizeof(qp_attr)); + rc = ib_qp_query(conn->ibc_qp, &qp_attr); + if (rc != 0) { + CERROR ("Can't get qp attrs: %d\n", rc); + return; + } + + CWARN ("RDMA CAPABILITY: write %s read %s\n", + (qp_attr.valid_fields & TS_IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE) ? + (qp_attr.enable_rdma_write ? "enabled" : "disabled") : "invalid", + (qp_attr.valid_fields & TS_IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE) ? + (qp_attr.enable_rdma_read ? "enabled" : "disabled") : "invalid"); +} +#endif + +#if CONFIG_X86 +static inline __u64 +kibnal_page2phys (struct page *p) +{ + __u64 page_number = p - mem_map; + + return (page_number << PAGE_SHIFT); +} +#else +# error "no page->phys" +#endif + +/* CAVEAT EMPTOR: + * We rely on tx/rx descriptor alignment to allow us to use the lowest bit + * of the work request id as a flag to determine if the completion is for a + * transmit or a receive. It seems that that the CQ entry's 'op' field + * isn't always set correctly on completions that occur after QP teardown. */ + +static inline __u64 +kibnal_ptr2wreqid (void *ptr, int isrx) +{ + unsigned long lptr = (unsigned long)ptr; + + LASSERT ((lptr & 1) == 0); + return (__u64)(lptr | (isrx ? 1 : 0)); +} + +static inline void * +kibnal_wreqid2ptr (__u64 wreqid) +{ + return (void *)(((unsigned long)wreqid) & ~1UL); +} + +static inline int +kibnal_wreqid_is_rx (__u64 wreqid) +{ + return (wreqid & 1) != 0; +} + +static inline int +kibnal_whole_mem(void) +{ + return kibnal_data.kib_md.md_handle != NULL; +} + +extern kib_peer_t *kibnal_create_peer (ptl_nid_t nid); +extern void kibnal_destroy_peer (kib_peer_t *peer); +extern int kibnal_del_peer (ptl_nid_t nid, int single_share); +extern kib_peer_t *kibnal_find_peer_locked (ptl_nid_t nid); +extern void kibnal_unlink_peer_locked (kib_peer_t *peer); +extern int kibnal_close_stale_conns_locked (kib_peer_t *peer, + __u64 incarnation); +extern kib_conn_t *kibnal_create_conn (void); +extern void kibnal_put_conn (kib_conn_t *conn); +extern void kibnal_destroy_conn (kib_conn_t *conn); +void kibnal_listen_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg); + +extern int kibnal_alloc_pages (kib_pages_t **pp, int npages, int access); +extern void kibnal_free_pages (kib_pages_t *p); + +extern void kibnal_check_sends (kib_conn_t *conn); +extern void kibnal_close_conn_locked (kib_conn_t *conn, int error); +extern void kibnal_destroy_conn (kib_conn_t *conn); +extern int kibnal_thread_start (int (*fn)(void *arg), void *arg); +extern int kibnal_scheduler(void *arg); +extern int kibnal_connd (void *arg); +extern void kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob); +extern void kibnal_close_conn (kib_conn_t *conn, int why); +extern void kibnal_start_active_rdma (int type, int status, + kib_rx_t *rx, lib_msg_t *libmsg, + unsigned int niov, + struct iovec *iov, ptl_kiov_t *kiov, + size_t offset, size_t nob); + +void kibnal_ca_async_callback (void *ca_arg, IB_EVENT_RECORD *ev); +void kibnal_ca_callback (void *ca_arg, void *cq_arg); diff --git a/lnet/klnds/iiblnd/iiblnd_cb.c b/lnet/klnds/iiblnd/iiblnd_cb.c new file mode 100644 index 0000000..a827ba5 --- /dev/null +++ b/lnet/klnds/iiblnd/iiblnd_cb.c @@ -0,0 +1,3018 @@ +/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- + * vim:expandtab:shiftwidth=8:tabstop=8: + * + * Copyright (C) 2004 Cluster File Systems, Inc. + * Author: Eric Barton + * + * This file is part of Lustre, http://www.lustre.org. + * + * Lustre is free software; you can redistribute it and/or + * modify it under the terms of version 2 of the GNU General Public + * License as published by the Free Software Foundation. + * + * Lustre 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 Lustre; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * + */ + +#include "iibnal.h" + +/* + * LIB functions follow + * + */ +static void +kibnal_schedule_tx_done (kib_tx_t *tx) +{ + unsigned long flags; + + spin_lock_irqsave (&kibnal_data.kib_sched_lock, flags); + + list_add_tail(&tx->tx_list, &kibnal_data.kib_sched_txq); + wake_up (&kibnal_data.kib_sched_waitq); + + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); +} + +static void +kibnal_tx_done (kib_tx_t *tx) +{ + ptl_err_t ptlrc = (tx->tx_status == 0) ? PTL_OK : PTL_FAIL; + unsigned long flags; + int i; + FSTATUS frc; + + LASSERT (tx->tx_sending == 0); /* mustn't be awaiting callback */ + LASSERT (!tx->tx_passive_rdma_wait); /* mustn't be awaiting RDMA */ + + switch (tx->tx_mapped) { + default: + LBUG(); + + case KIB_TX_UNMAPPED: + break; + + case KIB_TX_MAPPED: + if (in_interrupt()) { + /* can't deregister memory in IRQ context... */ + kibnal_schedule_tx_done(tx); + return; + } + frc = iibt_deregister_memory(tx->tx_md.md_handle); + LASSERT (frc == FSUCCESS); + tx->tx_mapped = KIB_TX_UNMAPPED; + break; + +#if IBNAL_FMR + case KIB_TX_MAPPED_FMR: + if (in_interrupt() && tx->tx_status != 0) { + /* can't flush FMRs in IRQ context... */ + kibnal_schedule_tx_done(tx); + return; + } + + rc = ib_fmr_deregister(tx->tx_md.md_handle.fmr); + LASSERT (rc == 0); + + if (tx->tx_status != 0) + ib_fmr_pool_force_flush(kibnal_data.kib_fmr_pool); + tx->tx_mapped = KIB_TX_UNMAPPED; + break; +#endif + } + + for (i = 0; i < 2; i++) { + /* tx may have up to 2 libmsgs to finalise */ + if (tx->tx_libmsg[i] == NULL) + continue; + + lib_finalize (&kibnal_lib, NULL, tx->tx_libmsg[i], ptlrc); + tx->tx_libmsg[i] = NULL; + } + + if (tx->tx_conn != NULL) { + kibnal_put_conn (tx->tx_conn); + tx->tx_conn = NULL; + } + + tx->tx_nsp = 0; + tx->tx_passive_rdma = 0; + tx->tx_status = 0; + + spin_lock_irqsave (&kibnal_data.kib_tx_lock, flags); + + if (tx->tx_isnblk) { + list_add_tail (&tx->tx_list, &kibnal_data.kib_idle_nblk_txs); + } else { + list_add_tail (&tx->tx_list, &kibnal_data.kib_idle_txs); + wake_up (&kibnal_data.kib_idle_tx_waitq); + } + + spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags); +} + +static kib_tx_t * +kibnal_get_idle_tx (int may_block) +{ + unsigned long flags; + kib_tx_t *tx = NULL; + ENTRY; + + for (;;) { + spin_lock_irqsave (&kibnal_data.kib_tx_lock, flags); + + /* "normal" descriptor is free */ + if (!list_empty (&kibnal_data.kib_idle_txs)) { + tx = list_entry (kibnal_data.kib_idle_txs.next, + kib_tx_t, tx_list); + break; + } + + if (!may_block) { + /* may dip into reserve pool */ + if (list_empty (&kibnal_data.kib_idle_nblk_txs)) { + CERROR ("reserved tx desc pool exhausted\n"); + break; + } + + tx = list_entry (kibnal_data.kib_idle_nblk_txs.next, + kib_tx_t, tx_list); + break; + } + + /* block for idle tx */ + spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags); + + wait_event (kibnal_data.kib_idle_tx_waitq, + !list_empty (&kibnal_data.kib_idle_txs) || + kibnal_data.kib_shutdown); + } + + if (tx != NULL) { + list_del (&tx->tx_list); + + /* Allocate a new passive RDMA completion cookie. It might + * not be needed, but we've got a lock right now and we're + * unlikely to wrap... */ + tx->tx_passive_rdma_cookie = kibnal_data.kib_next_tx_cookie++; + + LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED); + LASSERT (tx->tx_nsp == 0); + LASSERT (tx->tx_sending == 0); + LASSERT (tx->tx_status == 0); + LASSERT (tx->tx_conn == NULL); + LASSERT (!tx->tx_passive_rdma); + LASSERT (!tx->tx_passive_rdma_wait); + LASSERT (tx->tx_libmsg[0] == NULL); + LASSERT (tx->tx_libmsg[1] == NULL); + } + + spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags); + + RETURN(tx); +} + +static int +kibnal_dist(lib_nal_t *nal, ptl_nid_t nid, unsigned long *dist) +{ + /* I would guess that if kibnal_get_peer (nid) == NULL, + and we're not routing, then 'nid' is very distant :) */ + if ( nal->libnal_ni.ni_pid.nid == nid ) { + *dist = 0; + } else { + *dist = 1; + } + + return 0; +} + +static void +kibnal_complete_passive_rdma(kib_conn_t *conn, __u64 cookie, int status) +{ + struct list_head *ttmp; + unsigned long flags; + int idle; + + spin_lock_irqsave (&conn->ibc_lock, flags); + + list_for_each (ttmp, &conn->ibc_active_txs) { + kib_tx_t *tx = list_entry(ttmp, kib_tx_t, tx_list); + + LASSERT (tx->tx_passive_rdma || + !tx->tx_passive_rdma_wait); + + LASSERT (tx->tx_passive_rdma_wait || + tx->tx_sending != 0); + + if (!tx->tx_passive_rdma_wait || + tx->tx_passive_rdma_cookie != cookie) + continue; + + CDEBUG(D_NET, "Complete %p "LPD64": %d\n", tx, cookie, status); + + tx->tx_status = status; + tx->tx_passive_rdma_wait = 0; + idle = (tx->tx_sending == 0); + + if (idle) + list_del (&tx->tx_list); + + spin_unlock_irqrestore (&conn->ibc_lock, flags); + + /* I could be racing with tx callbacks. It's whoever + * _makes_ tx idle that frees it */ + if (idle) + kibnal_tx_done (tx); + return; + } + + spin_unlock_irqrestore (&conn->ibc_lock, flags); + + CERROR ("Unmatched (late?) RDMA completion "LPX64" from "LPX64"\n", + cookie, conn->ibc_peer->ibp_nid); +} + +static __u32 +kibnal_lkey(kib_pages_t *ibp) +{ + if (kibnal_whole_mem()) + return kibnal_data.kib_md.md_lkey; + + return ibp->ibp_lkey; +} + +static void +kibnal_post_rx (kib_rx_t *rx, int do_credits) +{ + kib_conn_t *conn = rx->rx_conn; + int rc = 0; + unsigned long flags; + FSTATUS frc; + ENTRY; + + rx->rx_gl = (IB_LOCAL_DATASEGMENT) { + .Address = rx->rx_vaddr, + .Length = IBNAL_MSG_SIZE, + .Lkey = kibnal_lkey(conn->ibc_rx_pages), + }; + + rx->rx_wrq = (IB_WORK_REQ) { + .Operation = WROpRecv, + .DSListDepth = 1, + .MessageLen = IBNAL_MSG_SIZE, + .WorkReqId = kibnal_ptr2wreqid(rx, 1), + .DSList = &rx->rx_gl, + }; + + KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_ESTABLISHED, + IBNAL_CONN_DREP); + LASSERT (!rx->rx_posted); + rx->rx_posted = 1; + mb(); + + if (conn->ibc_state != IBNAL_CONN_ESTABLISHED) + rc = -ECONNABORTED; + else { + frc = iibt_postrecv(conn->ibc_qp, &rx->rx_wrq); + if (frc != FSUCCESS) { + CDEBUG(D_NET, "post failed %d\n", frc); + rc = -EINVAL; + } + CDEBUG(D_NET, "posted rx %p\n", &rx->rx_wrq); + } + + if (rc == 0) { + if (do_credits) { + spin_lock_irqsave(&conn->ibc_lock, flags); + conn->ibc_outstanding_credits++; + spin_unlock_irqrestore(&conn->ibc_lock, flags); + + kibnal_check_sends(conn); + } + EXIT; + return; + } + + if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) { + CERROR ("Error posting receive -> "LPX64": %d\n", + conn->ibc_peer->ibp_nid, rc); + kibnal_close_conn (rx->rx_conn, rc); + } else { + CDEBUG (D_NET, "Error posting receive -> "LPX64": %d\n", + conn->ibc_peer->ibp_nid, rc); + } + + /* Drop rx's ref */ + kibnal_put_conn (conn); + EXIT; +} + +#if IBNAL_CKSUM +static inline __u32 kibnal_cksum (void *ptr, int nob) +{ + char *c = ptr; + __u32 sum = 0; + + while (nob-- > 0) + sum = ((sum << 1) | (sum >> 31)) + *c++; + + return (sum); +} +#endif + +static void hexdump(char *string, void *ptr, int len) +{ + unsigned char *c = ptr; + int i; + + return; + + if (len < 0 || len > 2048) { + printk("XXX what the hell? %d\n",len); + return; + } + + printk("%d bytes of '%s' from 0x%p\n", len, string, ptr); + + for (i = 0; i < len;) { + printk("%02x",*(c++)); + i++; + if (!(i & 15)) { + printk("\n"); + } else if (!(i&1)) { + printk(" "); + } + } + + if(len & 15) { + printk("\n"); + } +} + +static void +kibnal_rx_callback (IB_WORK_COMPLETION *wc) +{ + kib_rx_t *rx = (kib_rx_t *)kibnal_wreqid2ptr(wc->WorkReqId); + kib_msg_t *msg = rx->rx_msg; + kib_conn_t *conn = rx->rx_conn; + int nob = wc->Length; + const int base_nob = offsetof(kib_msg_t, ibm_u); + int credits; + int flipped; + unsigned long flags; + __u32 i; +#if IBNAL_CKSUM + __u32 msg_cksum; + __u32 computed_cksum; +#endif + + /* we set the QP to erroring after we've finished disconnecting, + * maybe we should do so sooner. */ + KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_ESTABLISHED, + IBNAL_CONN_DISCONNECTED); + + CDEBUG(D_NET, "rx %p conn %p\n", rx, conn); + LASSERT (rx->rx_posted); + rx->rx_posted = 0; + mb(); + + /* receives complete with error in any case after we've started + * disconnecting */ + if (conn->ibc_state > IBNAL_CONN_ESTABLISHED) + goto failed; + + if (wc->Status != WRStatusSuccess) { + CERROR("Rx from "LPX64" failed: %d\n", + conn->ibc_peer->ibp_nid, wc->Status); + goto failed; + } + + if (nob < base_nob) { + CERROR ("Short rx from "LPX64": %d < expected %d\n", + conn->ibc_peer->ibp_nid, nob, base_nob); + goto failed; + } + + hexdump("rx", rx->rx_msg, sizeof(kib_msg_t)); + + /* Receiver does any byte flipping if necessary... */ + + if (msg->ibm_magic == IBNAL_MSG_MAGIC) { + flipped = 0; + } else { + if (msg->ibm_magic != __swab32(IBNAL_MSG_MAGIC)) { + CERROR ("Unrecognised magic: %08x from "LPX64"\n", + msg->ibm_magic, conn->ibc_peer->ibp_nid); + goto failed; + } + flipped = 1; + __swab16s (&msg->ibm_version); + LASSERT (sizeof(msg->ibm_type) == 1); + LASSERT (sizeof(msg->ibm_credits) == 1); + } + + if (msg->ibm_version != IBNAL_MSG_VERSION) { + CERROR ("Incompatible msg version %d (%d expected)\n", + msg->ibm_version, IBNAL_MSG_VERSION); + goto failed; + } + +#if IBNAL_CKSUM + if (nob != msg->ibm_nob) { + CERROR ("Unexpected # bytes %d (%d expected)\n", nob, msg->ibm_nob); + goto failed; + } + + msg_cksum = le32_to_cpu(msg->ibm_cksum); + msg->ibm_cksum = 0; + computed_cksum = kibnal_cksum (msg, nob); + + if (msg_cksum != computed_cksum) { + CERROR ("Checksum failure %d: (%d expected)\n", + computed_cksum, msg_cksum); +// goto failed; + } + CDEBUG(D_NET, "cksum %x, nob %d\n", computed_cksum, nob); +#endif + + /* Have I received credits that will let me send? */ + credits = msg->ibm_credits; + if (credits != 0) { + spin_lock_irqsave(&conn->ibc_lock, flags); + conn->ibc_credits += credits; + spin_unlock_irqrestore(&conn->ibc_lock, flags); + + kibnal_check_sends(conn); + } + + switch (msg->ibm_type) { + case IBNAL_MSG_NOOP: + kibnal_post_rx (rx, 1); + return; + + case IBNAL_MSG_IMMEDIATE: + if (nob < base_nob + sizeof (kib_immediate_msg_t)) { + CERROR ("Short IMMEDIATE from "LPX64": %d\n", + conn->ibc_peer->ibp_nid, nob); + goto failed; + } + break; + + case IBNAL_MSG_PUT_RDMA: + case IBNAL_MSG_GET_RDMA: + if (nob < base_nob + sizeof (kib_rdma_msg_t)) { + CERROR ("Short RDMA msg from "LPX64": %d\n", + conn->ibc_peer->ibp_nid, nob); + goto failed; + } + if (flipped) + __swab32(msg->ibm_u.rdma.ibrm_num_descs); + + CDEBUG(D_NET, "%d RDMA: cookie "LPX64":\n", + msg->ibm_type, msg->ibm_u.rdma.ibrm_cookie); + + if ((msg->ibm_u.rdma.ibrm_num_descs > PTL_MD_MAX_IOV) || + (kib_rdma_msg_len(msg->ibm_u.rdma.ibrm_num_descs) > + min(nob, IBNAL_MSG_SIZE))) { + CERROR ("num_descs %d too large\n", + msg->ibm_u.rdma.ibrm_num_descs); + goto failed; + } + + for(i = 0; i < msg->ibm_u.rdma.ibrm_num_descs; i++) { + kib_rdma_desc_t *desc = &msg->ibm_u.rdma.ibrm_desc[i]; + + if (flipped) { + __swab32(desc->rd_key); + __swab32(desc->rd_nob); + __swab64(desc->rd_addr); + } + + CDEBUG(D_NET, " key %x, " "addr "LPX64", nob %u\n", + desc->rd_key, desc->rd_addr, desc->rd_nob); + } + break; + + case IBNAL_MSG_PUT_DONE: + case IBNAL_MSG_GET_DONE: + if (nob < base_nob + sizeof (kib_completion_msg_t)) { + CERROR ("Short COMPLETION msg from "LPX64": %d\n", + conn->ibc_peer->ibp_nid, nob); + goto failed; + } + if (flipped) + __swab32s(&msg->ibm_u.completion.ibcm_status); + + CDEBUG(D_NET, "%d DONE: cookie "LPX64", status %d\n", + msg->ibm_type, msg->ibm_u.completion.ibcm_cookie, + msg->ibm_u.completion.ibcm_status); + + kibnal_complete_passive_rdma (conn, + msg->ibm_u.completion.ibcm_cookie, + msg->ibm_u.completion.ibcm_status); + kibnal_post_rx (rx, 1); + return; + + default: + CERROR ("Can't parse type from "LPX64": %d\n", + conn->ibc_peer->ibp_nid, msg->ibm_type); + goto failed; + } + + /* schedule for kibnal_rx() in thread context */ + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); + + list_add_tail (&rx->rx_list, &kibnal_data.kib_sched_rxq); + wake_up (&kibnal_data.kib_sched_waitq); + + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); + return; + + failed: + CDEBUG(D_NET, "rx %p conn %p\n", rx, conn); + kibnal_close_conn(conn, -ECONNABORTED); + + /* Don't re-post rx & drop its ref on conn */ + kibnal_put_conn(conn); +} + +void +kibnal_rx (kib_rx_t *rx) +{ + kib_msg_t *msg = rx->rx_msg; + + /* Clear flag so I can detect if I've sent an RDMA completion */ + rx->rx_rdma = 0; + + switch (msg->ibm_type) { + case IBNAL_MSG_GET_RDMA: + lib_parse(&kibnal_lib, &msg->ibm_u.rdma.ibrm_hdr, rx); + /* If the incoming get was matched, I'll have initiated the + * RDMA and the completion message... */ + if (rx->rx_rdma) + break; + + /* Otherwise, I'll send a failed completion now to prevent + * the peer's GET blocking for the full timeout. */ + CERROR ("Completing unmatched RDMA GET from "LPX64"\n", + rx->rx_conn->ibc_peer->ibp_nid); + kibnal_start_active_rdma (IBNAL_MSG_GET_DONE, -EIO, + rx, NULL, 0, NULL, NULL, 0, 0); + break; + + case IBNAL_MSG_PUT_RDMA: + lib_parse(&kibnal_lib, &msg->ibm_u.rdma.ibrm_hdr, rx); + if (rx->rx_rdma) + break; + /* This is most unusual, since even if lib_parse() didn't + * match anything, it should have asked us to read (and + * discard) the payload. The portals header must be + * inconsistent with this message type, so it's the + * sender's fault for sending garbage and she can time + * herself out... */ + CERROR ("Uncompleted RMDA PUT from "LPX64"\n", + rx->rx_conn->ibc_peer->ibp_nid); + break; + + case IBNAL_MSG_IMMEDIATE: + lib_parse(&kibnal_lib, &msg->ibm_u.immediate.ibim_hdr, rx); + LASSERT (!rx->rx_rdma); + break; + + default: + LBUG(); + break; + } + + kibnal_post_rx (rx, 1); +} + +static struct page * +kibnal_kvaddr_to_page (unsigned long vaddr) +{ + struct page *page; + + if (vaddr >= VMALLOC_START && + vaddr < VMALLOC_END) + page = vmalloc_to_page ((void *)vaddr); +#if CONFIG_HIGHMEM + else if (vaddr >= PKMAP_BASE && + vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) + page = vmalloc_to_page ((void *)vaddr); + /* in 2.4 ^ just walks the page tables */ +#endif + else + page = virt_to_page (vaddr); + + if (!VALID_PAGE (page)) + page = NULL; + + return page; +} + +static void +kibnal_fill_ibrm(kib_tx_t *tx, struct page *page, unsigned long page_offset, + unsigned long len, int active) +{ + kib_rdma_msg_t *ibrm = &tx->tx_msg->ibm_u.rdma; + kib_rdma_desc_t *desc; + + LASSERTF(ibrm->ibrm_num_descs < PTL_MD_MAX_IOV, "%u\n", + ibrm->ibrm_num_descs); + + desc = &ibrm->ibrm_desc[ibrm->ibrm_num_descs]; + if (active) + desc->rd_key = kibnal_data.kib_md.md_lkey; + else + desc->rd_key = kibnal_data.kib_md.md_rkey; + desc->rd_nob = len; /*PAGE_SIZE - kiov->kiov_offset; */ + desc->rd_addr = kibnal_page2phys(page) + page_offset + + kibnal_data.kib_md.md_addr; + + ibrm->ibrm_num_descs++; +} + +static int +kibnal_map_rdma_iov(kib_tx_t *tx, unsigned long vaddr, int nob, int active) +{ + struct page *page; + int page_offset, len; + + while (nob > 0) { + page = kibnal_kvaddr_to_page(vaddr); + if (page == NULL) + return -EFAULT; + + page_offset = vaddr & (PAGE_SIZE - 1); + len = min(nob, (int)PAGE_SIZE - page_offset); + + kibnal_fill_ibrm(tx, page, page_offset, len, active); + nob -= len; + vaddr += len; + } + return 0; +} + +static int +kibnal_map_iov (kib_tx_t *tx, IB_ACCESS_CONTROL access, + int niov, struct iovec *iov, int offset, int nob, int active) + +{ + void *vaddr; + FSTATUS frc; + + LASSERT (nob > 0); + LASSERT (niov > 0); + LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED); + + while (offset >= iov->iov_len) { + offset -= iov->iov_len; + niov--; + iov++; + LASSERT (niov > 0); + } + + if (nob > iov->iov_len - offset) { + CERROR ("Can't map multiple vaddr fragments\n"); + return (-EMSGSIZE); + } + + /* our large contiguous iov could be backed by multiple physical + * pages. */ + if (kibnal_whole_mem()) { + int rc; + tx->tx_msg->ibm_u.rdma.ibrm_num_descs = 0; + rc = kibnal_map_rdma_iov(tx, (unsigned long)iov->iov_base + + offset, nob, active); + if (rc != 0) { + CERROR ("Can't map iov: %d\n", rc); + return rc; + } + return 0; + } + + vaddr = (void *)(((unsigned long)iov->iov_base) + offset); + tx->tx_md.md_addr = (__u64)((unsigned long)vaddr); + + frc = iibt_register_memory(kibnal_data.kib_hca, vaddr, nob, + kibnal_data.kib_pd, access, + &tx->tx_md.md_handle, &tx->tx_md.md_lkey, + &tx->tx_md.md_rkey); + if (frc != 0) { + CERROR ("Can't map vaddr %p: %d\n", vaddr, frc); + return -EINVAL; + } + + tx->tx_mapped = KIB_TX_MAPPED; + return (0); +} + +static int +kibnal_map_kiov (kib_tx_t *tx, IB_ACCESS_CONTROL access, + int nkiov, ptl_kiov_t *kiov, + int offset, int nob, int active) +{ + __u64 *phys = NULL; + int page_offset; + int nphys; + int resid; + int phys_size = 0; + FSTATUS frc; + int i, rc = 0; + + CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob); + + LASSERT (nob > 0); + LASSERT (nkiov > 0); + LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED); + + while (offset >= kiov->kiov_len) { + offset -= kiov->kiov_len; + nkiov--; + kiov++; + LASSERT (nkiov > 0); + } + + page_offset = kiov->kiov_offset + offset; + nphys = 1; + + if (!kibnal_whole_mem()) { + phys_size = nkiov * sizeof (*phys); + PORTAL_ALLOC(phys, phys_size); + if (phys == NULL) { + CERROR ("Can't allocate tmp phys\n"); + return (-ENOMEM); + } + + phys[0] = kibnal_page2phys(kiov->kiov_page); + } else { + tx->tx_msg->ibm_u.rdma.ibrm_num_descs = 0; + kibnal_fill_ibrm(tx, kiov->kiov_page, kiov->kiov_offset, + kiov->kiov_len, active); + } + + resid = nob - (kiov->kiov_len - offset); + + while (resid > 0) { + kiov++; + nkiov--; + LASSERT (nkiov > 0); + + if (kiov->kiov_offset != 0 || + ((resid > PAGE_SIZE) && + kiov->kiov_len < PAGE_SIZE)) { + /* Can't have gaps */ + CERROR ("Can't make payload contiguous in I/O VM:" + "page %d, offset %d, len %d \n", nphys, + kiov->kiov_offset, kiov->kiov_len); + + for (i = -nphys; i < nkiov; i++) + { + CERROR("kiov[%d] %p +%d for %d\n", + i, kiov[i].kiov_page, kiov[i].kiov_offset, kiov[i].kiov_len); + } + + rc = -EINVAL; + goto out; + } + + if (nphys == PTL_MD_MAX_IOV) { + CERROR ("payload too big (%d)\n", nphys); + rc = -EMSGSIZE; + goto out; + } + + if (!kibnal_whole_mem()) { + LASSERT (nphys * sizeof (*phys) < phys_size); + phys[nphys] = kibnal_page2phys(kiov->kiov_page); + } else { + if (kib_rdma_msg_len(nphys) > IBNAL_MSG_SIZE) { + CERROR ("payload too big (%d)\n", nphys); + rc = -EMSGSIZE; + goto out; + } + kibnal_fill_ibrm(tx, kiov->kiov_page, + kiov->kiov_offset, kiov->kiov_len, + active); + } + + nphys ++; + resid -= PAGE_SIZE; + } + + if (kibnal_whole_mem()) + goto out; + +#if 0 + CWARN ("nphys %d, nob %d, page_offset %d\n", nphys, nob, page_offset); + for (i = 0; i < nphys; i++) + CWARN (" [%d] "LPX64"\n", i, phys[i]); +#endif + +#if IBNAL_FMR +#error "iibnal hasn't learned about FMR yet" + rc = ib_fmr_register_physical (kibnal_data.kib_fmr_pool, + phys, nphys, + &tx->tx_md.md_addr, + page_offset, + &tx->tx_md.md_handle.fmr, + &tx->tx_md.md_lkey, + &tx->tx_md.md_rkey); +#else + frc = iibt_register_physical_memory(kibnal_data.kib_hca, + IBNAL_RDMA_BASE, + phys, nphys, + 0, /* offset */ + kibnal_data.kib_pd, + access, + &tx->tx_md.md_handle, + &tx->tx_md.md_addr, + &tx->tx_md.md_lkey, + &tx->tx_md.md_rkey); +#endif + if (frc == FSUCCESS) { + CDEBUG(D_NET, "Mapped %d pages %d bytes @ offset %d: lkey %x, rkey %x\n", + nphys, nob, page_offset, tx->tx_md.md_lkey, tx->tx_md.md_rkey); +#if IBNAL_FMR + tx->tx_mapped = KIB_TX_MAPPED_FMR; +#else + tx->tx_mapped = KIB_TX_MAPPED; +#endif + } else { + CERROR ("Can't map phys: %d\n", rc); + rc = -EFAULT; + } + + out: + if (phys != NULL) + PORTAL_FREE(phys, phys_size); + return (rc); +} + +static kib_conn_t * +kibnal_find_conn_locked (kib_peer_t *peer) +{ + struct list_head *tmp; + + /* just return the first connection */ + list_for_each (tmp, &peer->ibp_conns) { + return (list_entry(tmp, kib_conn_t, ibc_list)); + } + + return (NULL); +} + +void +kibnal_check_sends (kib_conn_t *conn) +{ + unsigned long flags; + kib_tx_t *tx; + int rc; + int i; + int done; + int nwork; + ENTRY; + + spin_lock_irqsave (&conn->ibc_lock, flags); + + LASSERT (conn->ibc_nsends_posted <= IBNAL_MSG_QUEUE_SIZE); + + if (list_empty(&conn->ibc_tx_queue) && + conn->ibc_outstanding_credits >= IBNAL_CREDIT_HIGHWATER) { + spin_unlock_irqrestore(&conn->ibc_lock, flags); + + tx = kibnal_get_idle_tx(0); /* don't block */ + if (tx != NULL) + kibnal_init_tx_msg(tx, IBNAL_MSG_NOOP, 0); + + spin_lock_irqsave(&conn->ibc_lock, flags); + + if (tx != NULL) { + atomic_inc(&conn->ibc_refcount); + kibnal_queue_tx_locked(tx, conn); + } + } + + while (!list_empty (&conn->ibc_tx_queue)) { + tx = list_entry (conn->ibc_tx_queue.next, kib_tx_t, tx_list); + + /* We rely on this for QP sizing */ + LASSERT (tx->tx_nsp > 0 && tx->tx_nsp <= IBNAL_TX_MAX_SG); + + LASSERT (conn->ibc_outstanding_credits >= 0); + LASSERT (conn->ibc_outstanding_credits <= IBNAL_MSG_QUEUE_SIZE); + LASSERT (conn->ibc_credits >= 0); + LASSERT (conn->ibc_credits <= IBNAL_MSG_QUEUE_SIZE); + + /* Not on ibc_rdma_queue */ + LASSERT (!tx->tx_passive_rdma_wait); + + if (conn->ibc_nsends_posted == IBNAL_MSG_QUEUE_SIZE) + GOTO(out, 0); + + if (conn->ibc_credits == 0) /* no credits */ + GOTO(out, 1); + + if (conn->ibc_credits == 1 && /* last credit reserved for */ + conn->ibc_outstanding_credits == 0) /* giving back credits */ + GOTO(out, 2); + + list_del (&tx->tx_list); + + if (tx->tx_msg->ibm_type == IBNAL_MSG_NOOP && + (!list_empty(&conn->ibc_tx_queue) || + conn->ibc_outstanding_credits < IBNAL_CREDIT_HIGHWATER)) { + /* redundant NOOP */ + spin_unlock_irqrestore(&conn->ibc_lock, flags); + kibnal_tx_done(tx); + spin_lock_irqsave(&conn->ibc_lock, flags); + continue; + } + + tx->tx_msg->ibm_credits = conn->ibc_outstanding_credits; + conn->ibc_outstanding_credits = 0; + + conn->ibc_nsends_posted++; + conn->ibc_credits--; + + /* we only get a tx completion for the final rdma op */ + tx->tx_sending = min(tx->tx_nsp, 2); + tx->tx_passive_rdma_wait = tx->tx_passive_rdma; + list_add (&tx->tx_list, &conn->ibc_active_txs); +#if IBNAL_CKSUM + tx->tx_msg->ibm_cksum = 0; + tx->tx_msg->ibm_cksum = kibnal_cksum(tx->tx_msg, tx->tx_msg->ibm_nob); + CDEBUG(D_NET, "cksum %x, nob %d\n", tx->tx_msg->ibm_cksum, tx->tx_msg->ibm_nob); +#endif + spin_unlock_irqrestore (&conn->ibc_lock, flags); + + /* NB the gap between removing tx from the queue and sending it + * allows message re-ordering to occur */ + + LASSERT (tx->tx_nsp > 0); + + rc = -ECONNABORTED; + nwork = 0; + if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) { + tx->tx_status = 0; + /* Driver only accepts 1 item at a time */ + for (i = 0; i < tx->tx_nsp; i++) { + hexdump("tx", tx->tx_msg, sizeof(kib_msg_t)); + rc = iibt_postsend(conn->ibc_qp, + &tx->tx_wrq[i]); + if (rc != 0) + break; + if (wrq_signals_completion(&tx->tx_wrq[i])) + nwork++; + CDEBUG(D_NET, "posted tx wrq %p\n", + &tx->tx_wrq[i]); + } + } + + spin_lock_irqsave (&conn->ibc_lock, flags); + if (rc != 0) { + /* NB credits are transferred in the actual + * message, which can only be the last work item */ + conn->ibc_outstanding_credits += tx->tx_msg->ibm_credits; + conn->ibc_credits++; + conn->ibc_nsends_posted--; + + tx->tx_status = rc; + tx->tx_passive_rdma_wait = 0; + tx->tx_sending -= tx->tx_nsp - nwork; + + done = (tx->tx_sending == 0); + if (done) + list_del (&tx->tx_list); + + spin_unlock_irqrestore (&conn->ibc_lock, flags); + + if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) + CERROR ("Error %d posting transmit to "LPX64"\n", + rc, conn->ibc_peer->ibp_nid); + else + CDEBUG (D_NET, "Error %d posting transmit to " + LPX64"\n", rc, conn->ibc_peer->ibp_nid); + + kibnal_close_conn (conn, rc); + + if (done) + kibnal_tx_done (tx); + return; + } + + } + + EXIT; +out: + spin_unlock_irqrestore (&conn->ibc_lock, flags); +} + +static void +kibnal_tx_callback (IB_WORK_COMPLETION *wc) +{ + kib_tx_t *tx = (kib_tx_t *)kibnal_wreqid2ptr(wc->WorkReqId); + kib_conn_t *conn; + unsigned long flags; + int idle; + + conn = tx->tx_conn; + LASSERT (conn != NULL); + LASSERT (tx->tx_sending != 0); + + spin_lock_irqsave(&conn->ibc_lock, flags); + + CDEBUG(D_NET, "conn %p tx %p [%d/%d]: %d\n", conn, tx, + tx->tx_sending, tx->tx_nsp, wc->Status); + + /* I could be racing with rdma completion. Whoever makes 'tx' idle + * gets to free it, which also drops its ref on 'conn'. If it's + * not me, then I take an extra ref on conn so it can't disappear + * under me. */ + + tx->tx_sending--; + idle = (tx->tx_sending == 0) && /* This is the final callback */ + (!tx->tx_passive_rdma_wait); /* Not waiting for RDMA completion */ + if (idle) + list_del(&tx->tx_list); + + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); + + if (tx->tx_sending == 0) + conn->ibc_nsends_posted--; + + if (wc->Status != WRStatusSuccess && + tx->tx_status == 0) + tx->tx_status = -ECONNABORTED; + + spin_unlock_irqrestore(&conn->ibc_lock, flags); + + if (idle) + kibnal_tx_done (tx); + + if (wc->Status != WRStatusSuccess) { + CERROR ("Tx completion to "LPX64" failed: %d\n", + conn->ibc_peer->ibp_nid, wc->Status); + kibnal_close_conn (conn, -ENETDOWN); + } else { + /* can I shovel some more sends out the door? */ + kibnal_check_sends(conn); + } + + kibnal_put_conn (conn); +} + +void +kibnal_ca_async_callback (void *ca_arg, IB_EVENT_RECORD *ev) +{ + /* XXX flesh out. this seems largely for async errors */ + CERROR("type: %d code: %u\n", ev->EventType, ev->EventCode); +} + +void +kibnal_ca_callback (void *ca_arg, void *cq_arg) +{ + IB_HANDLE cq = *(IB_HANDLE *)cq_arg; + IB_HANDLE ca = *(IB_HANDLE *)ca_arg; + IB_WORK_COMPLETION wc; + int armed = 0; + + CDEBUG(D_NET, "ca %p cq %p\n", ca, cq); + + for(;;) { + while (iibt_cq_poll(cq, &wc) == FSUCCESS) { + if (kibnal_wreqid_is_rx(wc.WorkReqId)) + kibnal_rx_callback(&wc); + else + kibnal_tx_callback(&wc); + } + if (armed) + return; + if (iibt_cq_rearm(cq, CQEventSelNextWC) != FSUCCESS) { + CERROR("rearm failed?\n"); + return; + } + armed = 1; + } +} + +void +kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob) +{ + IB_LOCAL_DATASEGMENT *gl = &tx->tx_gl[tx->tx_nsp]; + IB_WORK_REQ *wrq = &tx->tx_wrq[tx->tx_nsp]; + int fence; + int nob = offsetof (kib_msg_t, ibm_u) + body_nob; + + LASSERT (tx->tx_nsp >= 0 && + tx->tx_nsp < sizeof(tx->tx_wrq)/sizeof(tx->tx_wrq[0])); + LASSERT (nob <= IBNAL_MSG_SIZE); + + tx->tx_msg->ibm_magic = IBNAL_MSG_MAGIC; + tx->tx_msg->ibm_version = IBNAL_MSG_VERSION; + tx->tx_msg->ibm_type = type; +#if IBNAL_CKSUM + tx->tx_msg->ibm_nob = nob; +#endif + /* Fence the message if it's bundled with an RDMA read */ + fence = (tx->tx_nsp > 0) && + (type == IBNAL_MSG_PUT_DONE); + + *gl = (IB_LOCAL_DATASEGMENT) { + .Address = tx->tx_vaddr, + .Length = IBNAL_MSG_SIZE, + .Lkey = kibnal_lkey(kibnal_data.kib_tx_pages), + }; + + wrq->WorkReqId = kibnal_ptr2wreqid(tx, 0); + wrq->Operation = WROpSend; + wrq->DSList = gl; + wrq->DSListDepth = 1; + wrq->MessageLen = nob; + wrq->Req.SendRC.ImmediateData = 0; + wrq->Req.SendRC.Options.s.SolicitedEvent = 1; + wrq->Req.SendRC.Options.s.SignaledCompletion = 1; + wrq->Req.SendRC.Options.s.ImmediateData = 0; + wrq->Req.SendRC.Options.s.Fence = fence; + + tx->tx_nsp++; +} + +static void +kibnal_queue_tx (kib_tx_t *tx, kib_conn_t *conn) +{ + unsigned long flags; + + spin_lock_irqsave(&conn->ibc_lock, flags); + + kibnal_queue_tx_locked (tx, conn); + + spin_unlock_irqrestore(&conn->ibc_lock, flags); + + kibnal_check_sends(conn); +} + +static void +kibnal_launch_tx (kib_tx_t *tx, ptl_nid_t nid) +{ + unsigned long flags; + kib_peer_t *peer; + kib_conn_t *conn; + rwlock_t *g_lock = &kibnal_data.kib_global_lock; + + /* If I get here, I've committed to send, so I complete the tx with + * failure on any problems */ + + LASSERT (tx->tx_conn == NULL); /* only set when assigned a conn */ + LASSERT (tx->tx_nsp > 0); /* work items have been set up */ + + read_lock (g_lock); + + peer = kibnal_find_peer_locked (nid); + if (peer == NULL) { + read_unlock (g_lock); + tx->tx_status = -EHOSTUNREACH; + kibnal_tx_done (tx); + return; + } + + conn = kibnal_find_conn_locked (peer); + if (conn != NULL) { + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); /* 1 ref for the tx */ + read_unlock (g_lock); + + kibnal_queue_tx (tx, conn); + return; + } + + /* Making one or more connections; I'll need a write lock... */ + read_unlock (g_lock); + write_lock_irqsave (g_lock, flags); + + peer = kibnal_find_peer_locked (nid); + if (peer == NULL) { + write_unlock_irqrestore (g_lock, flags); + tx->tx_status = -EHOSTUNREACH; + kibnal_tx_done (tx); + return; + } + + conn = kibnal_find_conn_locked (peer); + if (conn != NULL) { + /* Connection exists; queue message on it */ + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); /* 1 ref for the tx */ + write_unlock_irqrestore (g_lock, flags); + + kibnal_queue_tx (tx, conn); + return; + } + + if (peer->ibp_connecting == 0) { + if (!time_after_eq(jiffies, peer->ibp_reconnect_time)) { + write_unlock_irqrestore (g_lock, flags); + tx->tx_status = -EHOSTUNREACH; + kibnal_tx_done (tx); + return; + } + + peer->ibp_connecting = 1; + kib_peer_addref(peer); /* extra ref for connd */ + + spin_lock (&kibnal_data.kib_connd_lock); + + list_add_tail (&peer->ibp_connd_list, + &kibnal_data.kib_connd_peers); + wake_up (&kibnal_data.kib_connd_waitq); + + spin_unlock (&kibnal_data.kib_connd_lock); + } + + /* A connection is being established; queue the message... */ + list_add_tail (&tx->tx_list, &peer->ibp_tx_queue); + + write_unlock_irqrestore (g_lock, flags); +} + +static ptl_err_t +kibnal_start_passive_rdma (int type, ptl_nid_t nid, + lib_msg_t *libmsg, ptl_hdr_t *hdr) +{ + int nob = libmsg->md->length; + kib_tx_t *tx; + kib_msg_t *ibmsg; + int rc; + IB_ACCESS_CONTROL access = {0,}; + + LASSERT (type == IBNAL_MSG_PUT_RDMA || type == IBNAL_MSG_GET_RDMA); + LASSERT (nob > 0); + LASSERT (!in_interrupt()); /* Mapping could block */ + + access.s.MWBindable = 1; + access.s.LocalWrite = 1; + access.s.RdmaRead = 1; + access.s.RdmaWrite = 1; + + tx = kibnal_get_idle_tx (1); /* May block; caller is an app thread */ + LASSERT (tx != NULL); + + if ((libmsg->md->options & PTL_MD_KIOV) == 0) + rc = kibnal_map_iov (tx, access, + libmsg->md->md_niov, + libmsg->md->md_iov.iov, + 0, nob, 0); + else + rc = kibnal_map_kiov (tx, access, + libmsg->md->md_niov, + libmsg->md->md_iov.kiov, + 0, nob, 0); + + if (rc != 0) { + CERROR ("Can't map RDMA for "LPX64": %d\n", nid, rc); + goto failed; + } + + if (type == IBNAL_MSG_GET_RDMA) { + /* reply gets finalized when tx completes */ + tx->tx_libmsg[1] = lib_create_reply_msg(&kibnal_lib, + nid, libmsg); + if (tx->tx_libmsg[1] == NULL) { + CERROR ("Can't create reply for GET -> "LPX64"\n", + nid); + rc = -ENOMEM; + goto failed; + } + } + + tx->tx_passive_rdma = 1; + + ibmsg = tx->tx_msg; + + ibmsg->ibm_u.rdma.ibrm_hdr = *hdr; + ibmsg->ibm_u.rdma.ibrm_cookie = tx->tx_passive_rdma_cookie; + /* map_kiov alrady filled the rdma descs for the whole_mem case */ + if (!kibnal_whole_mem()) { + ibmsg->ibm_u.rdma.ibrm_desc[0].rd_key = tx->tx_md.md_rkey; + ibmsg->ibm_u.rdma.ibrm_desc[0].rd_addr = tx->tx_md.md_addr; + ibmsg->ibm_u.rdma.ibrm_desc[0].rd_nob = nob; + ibmsg->ibm_u.rdma.ibrm_num_descs = 1; + } + + kibnal_init_tx_msg (tx, type, + kib_rdma_msg_len(ibmsg->ibm_u.rdma.ibrm_num_descs)); + + CDEBUG(D_NET, "Passive: %p cookie "LPX64", key %x, addr " + LPX64", nob %d\n", + tx, tx->tx_passive_rdma_cookie, tx->tx_md.md_rkey, + tx->tx_md.md_addr, nob); + + /* libmsg gets finalized when tx completes. */ + tx->tx_libmsg[0] = libmsg; + + kibnal_launch_tx(tx, nid); + return (PTL_OK); + + failed: + tx->tx_status = rc; + kibnal_tx_done (tx); + return (PTL_FAIL); +} + +void +kibnal_start_active_rdma (int type, int status, + kib_rx_t *rx, lib_msg_t *libmsg, + unsigned int niov, + struct iovec *iov, ptl_kiov_t *kiov, + size_t offset, size_t nob) +{ + kib_msg_t *rxmsg = rx->rx_msg; + kib_msg_t *txmsg; + kib_tx_t *tx; + IB_ACCESS_CONTROL access = {0,}; + IB_WR_OP rdma_op; + int rc; + __u32 i; + + CDEBUG(D_NET, "type %d, status %d, niov %d, offset %d, nob %d\n", + type, status, niov, offset, nob); + + /* Called by scheduler */ + LASSERT (!in_interrupt ()); + + /* Either all pages or all vaddrs */ + LASSERT (!(kiov != NULL && iov != NULL)); + + /* No data if we're completing with failure */ + LASSERT (status == 0 || nob == 0); + + LASSERT (type == IBNAL_MSG_GET_DONE || + type == IBNAL_MSG_PUT_DONE); + + /* Flag I'm completing the RDMA. Even if I fail to send the + * completion message, I will have tried my best so further + * attempts shouldn't be tried. */ + LASSERT (!rx->rx_rdma); + rx->rx_rdma = 1; + + if (type == IBNAL_MSG_GET_DONE) { + rdma_op = WROpRdmaWrite; + LASSERT (rxmsg->ibm_type == IBNAL_MSG_GET_RDMA); + } else { + access.s.LocalWrite = 1; + rdma_op = WROpRdmaRead; + LASSERT (rxmsg->ibm_type == IBNAL_MSG_PUT_RDMA); + } + + tx = kibnal_get_idle_tx (0); /* Mustn't block */ + if (tx == NULL) { + CERROR ("tx descs exhausted on RDMA from "LPX64 + " completing locally with failure\n", + rx->rx_conn->ibc_peer->ibp_nid); + lib_finalize (&kibnal_lib, NULL, libmsg, PTL_NO_SPACE); + return; + } + LASSERT (tx->tx_nsp == 0); + + if (nob == 0) + GOTO(init_tx, 0); + + /* We actually need to transfer some data (the transfer + * size could get truncated to zero when the incoming + * message is matched) */ + if (kiov != NULL) + rc = kibnal_map_kiov (tx, access, niov, kiov, offset, nob, 1); + else + rc = kibnal_map_iov (tx, access, niov, iov, offset, nob, 1); + + if (rc != 0) { + CERROR ("Can't map RDMA -> "LPX64": %d\n", + rx->rx_conn->ibc_peer->ibp_nid, rc); + /* We'll skip the RDMA and complete with failure. */ + status = rc; + nob = 0; + GOTO(init_tx, rc); + } + + if (!kibnal_whole_mem()) { + tx->tx_msg->ibm_u.rdma.ibrm_desc[0].rd_key = tx->tx_md.md_lkey; + tx->tx_msg->ibm_u.rdma.ibrm_desc[0].rd_addr = tx->tx_md.md_addr; + tx->tx_msg->ibm_u.rdma.ibrm_desc[0].rd_nob = nob; + tx->tx_msg->ibm_u.rdma.ibrm_num_descs = 1; + } + + /* XXX ugh. different page-sized hosts. */ + if (tx->tx_msg->ibm_u.rdma.ibrm_num_descs != + rxmsg->ibm_u.rdma.ibrm_num_descs) { + CERROR("tx descs (%u) != rx descs (%u)\n", + tx->tx_msg->ibm_u.rdma.ibrm_num_descs, + rxmsg->ibm_u.rdma.ibrm_num_descs); + /* We'll skip the RDMA and complete with failure. */ + status = rc; + nob = 0; + GOTO(init_tx, rc); + } + + /* map_kiov filled in the rdma descs which describe our side of the + * rdma transfer. */ + /* ibrm_num_descs was verified in rx_callback */ + for(i = 0; i < rxmsg->ibm_u.rdma.ibrm_num_descs; i++) { + kib_rdma_desc_t *ldesc, *rdesc; /* local, remote */ + IB_LOCAL_DATASEGMENT *ds = &tx->tx_gl[i]; + IB_WORK_REQ *wrq = &tx->tx_wrq[i]; + + ldesc = &tx->tx_msg->ibm_u.rdma.ibrm_desc[i]; + rdesc = &rxmsg->ibm_u.rdma.ibrm_desc[i]; + + ds->Address = ldesc->rd_addr; + ds->Length = ldesc->rd_nob; + ds->Lkey = ldesc->rd_key; + + memset(wrq, 0, sizeof(*wrq)); + wrq->WorkReqId = kibnal_ptr2wreqid(tx, 0); + wrq->Operation = rdma_op; + wrq->DSList = ds; + wrq->DSListDepth = 1; + wrq->MessageLen = ds->Length; + wrq->Req.SendRC.ImmediateData = 0; + wrq->Req.SendRC.Options.s.SolicitedEvent = 0; + wrq->Req.SendRC.Options.s.SignaledCompletion = 0; + wrq->Req.SendRC.Options.s.ImmediateData = 0; + wrq->Req.SendRC.Options.s.Fence = 0; + wrq->Req.SendRC.RemoteDS.Address = rdesc->rd_addr; + wrq->Req.SendRC.RemoteDS.Rkey = rdesc->rd_key; + + /* only the last rdma post triggers tx completion */ + if (i == rxmsg->ibm_u.rdma.ibrm_num_descs - 1) + wrq->Req.SendRC.Options.s.SignaledCompletion = 1; + + tx->tx_nsp++; + } + +init_tx: + txmsg = tx->tx_msg; + + txmsg->ibm_u.completion.ibcm_cookie = rxmsg->ibm_u.rdma.ibrm_cookie; + txmsg->ibm_u.completion.ibcm_status = status; + + kibnal_init_tx_msg(tx, type, sizeof (kib_completion_msg_t)); + + if (status == 0 && nob != 0) { + LASSERT (tx->tx_nsp > 1); + /* RDMA: libmsg gets finalized when the tx completes. This + * is after the completion message has been sent, which in + * turn is after the RDMA has finished. */ + tx->tx_libmsg[0] = libmsg; + } else { + LASSERT (tx->tx_nsp == 1); + /* No RDMA: local completion happens now! */ + CDEBUG(D_WARNING,"No data: immediate completion\n"); + lib_finalize (&kibnal_lib, NULL, libmsg, + status == 0 ? PTL_OK : PTL_FAIL); + } + + /* +1 ref for this tx... */ + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + rx->rx_conn, rx->rx_conn->ibc_state, + rx->rx_conn->ibc_peer->ibp_nid, + atomic_read (&rx->rx_conn->ibc_refcount)); + atomic_inc (&rx->rx_conn->ibc_refcount); + /* ...and queue it up */ + kibnal_queue_tx(tx, rx->rx_conn); +} + +static ptl_err_t +kibnal_sendmsg(lib_nal_t *nal, + void *private, + lib_msg_t *libmsg, + ptl_hdr_t *hdr, + int type, + ptl_nid_t nid, + ptl_pid_t pid, + unsigned int payload_niov, + struct iovec *payload_iov, + ptl_kiov_t *payload_kiov, + size_t payload_offset, + size_t payload_nob) +{ + kib_msg_t *ibmsg; + kib_tx_t *tx; + int nob; + + /* NB 'private' is different depending on what we're sending.... */ + + CDEBUG(D_NET, "sending "LPSZ" bytes in %d frags to nid:"LPX64 + " pid %d\n", payload_nob, payload_niov, nid , pid); + + LASSERT (payload_nob == 0 || payload_niov > 0); + LASSERT (payload_niov <= PTL_MD_MAX_IOV); + + /* Thread context if we're sending payload */ + LASSERT (!in_interrupt() || payload_niov == 0); + /* payload is either all vaddrs or all pages */ + LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); + + switch (type) { + default: + LBUG(); + return (PTL_FAIL); + + case PTL_MSG_REPLY: { + /* reply's 'private' is the incoming receive */ + kib_rx_t *rx = private; + + /* RDMA reply expected? */ + if (rx->rx_msg->ibm_type == IBNAL_MSG_GET_RDMA) { + kibnal_start_active_rdma(IBNAL_MSG_GET_DONE, 0, + rx, libmsg, payload_niov, + payload_iov, payload_kiov, + payload_offset, payload_nob); + return (PTL_OK); + } + + /* Incoming message consistent with immediate reply? */ + if (rx->rx_msg->ibm_type != IBNAL_MSG_IMMEDIATE) { + CERROR ("REPLY to "LPX64" bad opbm type %d!!!\n", + nid, rx->rx_msg->ibm_type); + return (PTL_FAIL); + } + + /* Will it fit in a message? */ + nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]); + if (nob >= IBNAL_MSG_SIZE) { + CERROR("REPLY for "LPX64" too big (RDMA not requested): %d\n", + nid, payload_nob); + return (PTL_FAIL); + } + break; + } + + case PTL_MSG_GET: + /* might the REPLY message be big enough to need RDMA? */ + nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[libmsg->md->length]); + if (nob > IBNAL_MSG_SIZE) + return (kibnal_start_passive_rdma(IBNAL_MSG_GET_RDMA, + nid, libmsg, hdr)); + break; + + case PTL_MSG_ACK: + LASSERT (payload_nob == 0); + break; + + case PTL_MSG_PUT: + /* Is the payload big enough to need RDMA? */ + nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]); + if (nob > IBNAL_MSG_SIZE) + return (kibnal_start_passive_rdma(IBNAL_MSG_PUT_RDMA, + nid, libmsg, hdr)); + + break; + } + + tx = kibnal_get_idle_tx(!(type == PTL_MSG_ACK || + type == PTL_MSG_REPLY || + in_interrupt())); + if (tx == NULL) { + CERROR ("Can't send %d to "LPX64": tx descs exhausted%s\n", + type, nid, in_interrupt() ? " (intr)" : ""); + return (PTL_NO_SPACE); + } + + ibmsg = tx->tx_msg; + ibmsg->ibm_u.immediate.ibim_hdr = *hdr; + + if (payload_nob > 0) { + if (payload_kiov != NULL) + lib_copy_kiov2buf(ibmsg->ibm_u.immediate.ibim_payload, + payload_niov, payload_kiov, + payload_offset, payload_nob); + else + lib_copy_iov2buf(ibmsg->ibm_u.immediate.ibim_payload, + payload_niov, payload_iov, + payload_offset, payload_nob); + } + + kibnal_init_tx_msg (tx, IBNAL_MSG_IMMEDIATE, + offsetof(kib_immediate_msg_t, + ibim_payload[payload_nob])); + + /* libmsg gets finalized when tx completes */ + tx->tx_libmsg[0] = libmsg; + + kibnal_launch_tx(tx, nid); + return (PTL_OK); +} + +static ptl_err_t +kibnal_send (lib_nal_t *nal, void *private, lib_msg_t *cookie, + ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid, + unsigned int payload_niov, struct iovec *payload_iov, + size_t payload_offset, size_t payload_len) +{ + return (kibnal_sendmsg(nal, private, cookie, + hdr, type, nid, pid, + payload_niov, payload_iov, NULL, + payload_offset, payload_len)); +} + +static ptl_err_t +kibnal_send_pages (lib_nal_t *nal, void *private, lib_msg_t *cookie, + ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid, + unsigned int payload_niov, ptl_kiov_t *payload_kiov, + size_t payload_offset, size_t payload_len) +{ + return (kibnal_sendmsg(nal, private, cookie, + hdr, type, nid, pid, + payload_niov, NULL, payload_kiov, + payload_offset, payload_len)); +} + +static ptl_err_t +kibnal_recvmsg (lib_nal_t *nal, void *private, lib_msg_t *libmsg, + unsigned int niov, struct iovec *iov, ptl_kiov_t *kiov, + size_t offset, size_t mlen, size_t rlen) +{ + kib_rx_t *rx = private; + kib_msg_t *rxmsg = rx->rx_msg; + int msg_nob; + + LASSERT (mlen <= rlen); + LASSERT (!in_interrupt ()); + /* Either all pages or all vaddrs */ + LASSERT (!(kiov != NULL && iov != NULL)); + + switch (rxmsg->ibm_type) { + default: + LBUG(); + return (PTL_FAIL); + + case IBNAL_MSG_IMMEDIATE: + msg_nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]); + if (msg_nob > IBNAL_MSG_SIZE) { + CERROR ("Immediate message from "LPX64" too big: %d\n", + rxmsg->ibm_u.immediate.ibim_hdr.src_nid, rlen); + return (PTL_FAIL); + } + + if (kiov != NULL) + lib_copy_buf2kiov(niov, kiov, offset, + rxmsg->ibm_u.immediate.ibim_payload, + mlen); + else + lib_copy_buf2iov(niov, iov, offset, + rxmsg->ibm_u.immediate.ibim_payload, + mlen); + + lib_finalize (nal, NULL, libmsg, PTL_OK); + return (PTL_OK); + + case IBNAL_MSG_GET_RDMA: + /* We get called here just to discard any junk after the + * GET hdr. */ + LASSERT (libmsg == NULL); + lib_finalize (nal, NULL, libmsg, PTL_OK); + return (PTL_OK); + + case IBNAL_MSG_PUT_RDMA: + kibnal_start_active_rdma (IBNAL_MSG_PUT_DONE, 0, + rx, libmsg, + niov, iov, kiov, offset, mlen); + return (PTL_OK); + } +} + +static ptl_err_t +kibnal_recv (lib_nal_t *nal, void *private, lib_msg_t *msg, + unsigned int niov, struct iovec *iov, + size_t offset, size_t mlen, size_t rlen) +{ + return (kibnal_recvmsg (nal, private, msg, niov, iov, NULL, + offset, mlen, rlen)); +} + +static ptl_err_t +kibnal_recv_pages (lib_nal_t *nal, void *private, lib_msg_t *msg, + unsigned int niov, ptl_kiov_t *kiov, + size_t offset, size_t mlen, size_t rlen) +{ + return (kibnal_recvmsg (nal, private, msg, niov, NULL, kiov, + offset, mlen, rlen)); +} + +/***************************************************************************** + * the rest of this file concerns connection management. active connetions + * start with connect_peer, passive connections start with passive_callback. + * active disconnects start with conn_close, cm_callback starts passive + * disconnects and contains the guts of how the disconnect state machine + * progresses. + *****************************************************************************/ + +int +kibnal_thread_start (int (*fn)(void *arg), void *arg) +{ + long pid = kernel_thread (fn, arg, 0); + + if (pid < 0) + return ((int)pid); + + atomic_inc (&kibnal_data.kib_nthreads); + return (0); +} + +static void +kibnal_thread_fini (void) +{ + atomic_dec (&kibnal_data.kib_nthreads); +} + +/* this can be called by anyone at any time to close a connection. if + * the connection is still established it heads to the connd to start + * the disconnection in a safe context. It has no effect if called + * on a connection that is already disconnecting */ +void +kibnal_close_conn_locked (kib_conn_t *conn, int error) +{ + /* This just does the immmediate housekeeping, and schedules the + * connection for the connd to finish off. + * Caller holds kib_global_lock exclusively in irq context */ + kib_peer_t *peer = conn->ibc_peer; + + KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_CONNECTING, + IBNAL_CONN_DISCONNECTED); + + if (conn->ibc_state > IBNAL_CONN_ESTABLISHED) + return; /* already disconnecting */ + + CDEBUG (error == 0 ? D_NET : D_ERROR, + "closing conn to "LPX64": error %d\n", peer->ibp_nid, error); + + if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) { + /* kib_connd_conns takes ibc_list's ref */ + list_del (&conn->ibc_list); + } else { + /* new ref for kib_connd_conns */ + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); + } + + if (list_empty (&peer->ibp_conns) && + peer->ibp_persistence == 0) { + /* Non-persistent peer with no more conns... */ + kibnal_unlink_peer_locked (peer); + } + + conn->ibc_state = IBNAL_CONN_SEND_DREQ; + + spin_lock (&kibnal_data.kib_connd_lock); + + list_add_tail (&conn->ibc_list, &kibnal_data.kib_connd_conns); + wake_up (&kibnal_data.kib_connd_waitq); + + spin_unlock (&kibnal_data.kib_connd_lock); +} + +void +kibnal_close_conn (kib_conn_t *conn, int error) +{ + unsigned long flags; + + write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + + kibnal_close_conn_locked (conn, error); + + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); +} + +static void +kibnal_peer_connect_failed (kib_peer_t *peer, int active, int rc) +{ + LIST_HEAD (zombies); + kib_tx_t *tx; + unsigned long flags; + + LASSERT (rc != 0); + LASSERT (peer->ibp_reconnect_interval >= IBNAL_MIN_RECONNECT_INTERVAL); + + write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + + LASSERT (peer->ibp_connecting != 0); + peer->ibp_connecting--; + + if (peer->ibp_connecting != 0) { + /* another connection attempt under way (loopback?)... */ + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + return; + } + + if (list_empty(&peer->ibp_conns)) { + /* Say when active connection can be re-attempted */ + peer->ibp_reconnect_time = jiffies + peer->ibp_reconnect_interval; + /* Increase reconnection interval */ + peer->ibp_reconnect_interval = MIN (peer->ibp_reconnect_interval * 2, + IBNAL_MAX_RECONNECT_INTERVAL); + + /* Take peer's blocked blocked transmits; I'll complete + * them with error */ + while (!list_empty (&peer->ibp_tx_queue)) { + tx = list_entry (peer->ibp_tx_queue.next, + kib_tx_t, tx_list); + + list_del (&tx->tx_list); + list_add_tail (&tx->tx_list, &zombies); + } + + if (kibnal_peer_active(peer) && + (peer->ibp_persistence == 0)) { + /* failed connection attempt on non-persistent peer */ + kibnal_unlink_peer_locked (peer); + } + } else { + /* Can't have blocked transmits if there are connections */ + LASSERT (list_empty(&peer->ibp_tx_queue)); + } + + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + + if (!list_empty (&zombies)) + CERROR ("Deleting messages for "LPX64": connection failed\n", + peer->ibp_nid); + + while (!list_empty (&zombies)) { + tx = list_entry (zombies.next, kib_tx_t, tx_list); + + list_del (&tx->tx_list); + /* complete now */ + tx->tx_status = -EHOSTUNREACH; + kibnal_tx_done (tx); + } +} + +static void +kibnal_connreq_done (kib_conn_t *conn, int active, int status) +{ + int state = conn->ibc_state; + kib_peer_t *peer = conn->ibc_peer; + kib_tx_t *tx; + unsigned long flags; + int i; + + /* passive connection has no connreq & vice versa */ + LASSERTF(!active == !(conn->ibc_connreq != NULL), + "%d %p\n", active, conn->ibc_connreq); + if (active) { + PORTAL_FREE (conn->ibc_connreq, sizeof (*conn->ibc_connreq)); + conn->ibc_connreq = NULL; + } + + write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + + LASSERT (peer->ibp_connecting != 0); + + if (status == 0) { + /* connection established... */ + KIB_ASSERT_CONN_STATE(conn, IBNAL_CONN_CONNECTING); + conn->ibc_state = IBNAL_CONN_ESTABLISHED; + + if (!kibnal_peer_active(peer)) { + /* ...but peer deleted meantime */ + status = -ECONNABORTED; + } + } else { + KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_INIT_QP, + IBNAL_CONN_CONNECTING); + } + + if (status == 0) { + /* Everything worked! */ + + peer->ibp_connecting--; + + /* +1 ref for ibc_list; caller(== CM)'s ref remains until + * the IB_CM_IDLE callback */ + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); + list_add (&conn->ibc_list, &peer->ibp_conns); + + /* reset reconnect interval for next attempt */ + peer->ibp_reconnect_interval = IBNAL_MIN_RECONNECT_INTERVAL; + + /* post blocked sends to the new connection */ + spin_lock (&conn->ibc_lock); + + while (!list_empty (&peer->ibp_tx_queue)) { + tx = list_entry (peer->ibp_tx_queue.next, + kib_tx_t, tx_list); + + list_del (&tx->tx_list); + + /* +1 ref for each tx */ + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); + kibnal_queue_tx_locked (tx, conn); + } + + spin_unlock (&conn->ibc_lock); + + /* Nuke any dangling conns from a different peer instance... */ + kibnal_close_stale_conns_locked (conn->ibc_peer, + conn->ibc_incarnation); + + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + + /* queue up all the receives */ + for (i = 0; i < IBNAL_RX_MSGS; i++) { + /* +1 ref for rx desc */ + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_inc (&conn->ibc_refcount); + + CDEBUG(D_NET, "RX[%d] %p->%p - "LPX64"\n", + i, &conn->ibc_rxs[i], conn->ibc_rxs[i].rx_msg, + conn->ibc_rxs[i].rx_vaddr); + + kibnal_post_rx (&conn->ibc_rxs[i], 0); + } + + kibnal_check_sends (conn); + return; + } + + /* connection failed */ + if (state == IBNAL_CONN_CONNECTING) { + /* schedule for connd to close */ + kibnal_close_conn_locked (conn, status); + } else { + /* Don't have a CM comm_id; just wait for refs to drain */ + conn->ibc_state = IBNAL_CONN_DISCONNECTED; + } + + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + + kibnal_peer_connect_failed (conn->ibc_peer, active, status); + + /* If we didn't establish the connection we don't have to pass + * through the disconnect protocol before dropping the CM ref */ + if (state < IBNAL_CONN_CONNECTING) + kibnal_put_conn (conn); +} + +static int +kibnal_accept (kib_conn_t **connp, IB_HANDLE *cep, + ptl_nid_t nid, __u64 incarnation, int queue_depth) +{ + kib_conn_t *conn = kibnal_create_conn(); + kib_peer_t *peer; + kib_peer_t *peer2; + unsigned long flags; + + if (conn == NULL) + return (-ENOMEM); + + if (queue_depth != IBNAL_MSG_QUEUE_SIZE) { + CERROR("Can't accept "LPX64": bad queue depth %d (%d expected)\n", + nid, queue_depth, IBNAL_MSG_QUEUE_SIZE); + atomic_dec (&conn->ibc_refcount); + kibnal_destroy_conn(conn); + return (-EPROTO); + } + + /* assume 'nid' is a new peer */ + peer = kibnal_create_peer (nid); + if (peer == NULL) { + CDEBUG(D_NET, "--conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, conn->ibc_peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + atomic_dec (&conn->ibc_refcount); + kibnal_destroy_conn(conn); + return (-ENOMEM); + } + + write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + + peer2 = kibnal_find_peer_locked(nid); + if (peer2 == NULL) { + /* peer table takes my ref on peer */ + list_add_tail (&peer->ibp_list, kibnal_nid2peerlist(nid)); + } else { + kib_peer_decref (peer); + peer = peer2; + } + + kib_peer_addref(peer); /* +1 ref for conn */ + peer->ibp_connecting++; + + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + + conn->ibc_peer = peer; + conn->ibc_state = IBNAL_CONN_CONNECTING; + /* conn->ibc_cep is set when cm_accept is called */ + conn->ibc_incarnation = incarnation; + conn->ibc_credits = IBNAL_MSG_QUEUE_SIZE; + + *connp = conn; + return (0); +} + +static void kibnal_set_qp_state(IB_HANDLE *qp, IB_QP_STATE state) +{ + IB_QP_ATTRIBUTES_MODIFY modify_attr = {0,}; + FSTATUS frc; + + modify_attr.RequestState = state; + + frc = iibt_qp_modify(qp, &modify_attr, NULL); + if (frc != FSUCCESS) + CERROR("couldn't set qp state to %d, error %d\n", state, frc); +} + +static void kibnal_flush_pending(kib_conn_t *conn) +{ + LIST_HEAD (zombies); + struct list_head *tmp; + struct list_head *nxt; + kib_tx_t *tx; + unsigned long flags; + int done; + + /* NB we wait until the connection has closed before completing + * outstanding passive RDMAs so we can be sure the network can't + * touch the mapped memory any more. */ + KIB_ASSERT_CONN_STATE(conn, IBNAL_CONN_DISCONNECTED); + + /* set the QP to the error state so that we get flush callbacks + * on our posted receives which can then drop their conn refs */ + kibnal_set_qp_state(conn->ibc_qp, QPStateError); + + spin_lock_irqsave (&conn->ibc_lock, flags); + + /* grab passive RDMAs not waiting for the tx callback */ + list_for_each_safe (tmp, nxt, &conn->ibc_active_txs) { + tx = list_entry (tmp, kib_tx_t, tx_list); + + LASSERT (tx->tx_passive_rdma || + !tx->tx_passive_rdma_wait); + + LASSERT (tx->tx_passive_rdma_wait || + tx->tx_sending != 0); + + /* still waiting for tx callback? */ + if (!tx->tx_passive_rdma_wait) + continue; + + tx->tx_status = -ECONNABORTED; + tx->tx_passive_rdma_wait = 0; + done = (tx->tx_sending == 0); + + if (!done) + continue; + + list_del (&tx->tx_list); + list_add (&tx->tx_list, &zombies); + } + + /* grab all blocked transmits */ + list_for_each_safe (tmp, nxt, &conn->ibc_tx_queue) { + tx = list_entry (tmp, kib_tx_t, tx_list); + + list_del (&tx->tx_list); + list_add (&tx->tx_list, &zombies); + } + + spin_unlock_irqrestore (&conn->ibc_lock, flags); + + while (!list_empty(&zombies)) { + tx = list_entry (zombies.next, kib_tx_t, tx_list); + + list_del(&tx->tx_list); + kibnal_tx_done (tx); + } +} + +static void +kibnal_reject (IB_HANDLE cep, uint16_t reason) +{ + CM_REJECT_INFO *rej; + + PORTAL_ALLOC(rej, sizeof(*rej)); + if (rej == NULL) /* PORTAL_ALLOC() will CERROR on failure */ + return; + + rej->Reason = reason; + iibt_cm_reject(cep, rej); + PORTAL_FREE(rej, sizeof(*rej)); +} + +static FSTATUS +kibnal_qp_rts(IB_HANDLE qp_handle, __u32 qpn, __u8 resp_res, + IB_PATH_RECORD *path, __u8 init_depth, __u32 send_psn) +{ + IB_QP_ATTRIBUTES_MODIFY modify_attr; + FSTATUS frc; + ENTRY; + + modify_attr = (IB_QP_ATTRIBUTES_MODIFY) { + .RequestState = QPStateReadyToRecv, + .RecvPSN = IBNAL_STARTING_PSN, + .DestQPNumber = qpn, + .ResponderResources = resp_res, + .MinRnrTimer = UsecToRnrNakTimer(2000), /* 20 ms */ + .Attrs = (IB_QP_ATTR_RECVPSN | + IB_QP_ATTR_DESTQPNUMBER | + IB_QP_ATTR_RESPONDERRESOURCES | + IB_QP_ATTR_DESTAV | + IB_QP_ATTR_PATHMTU | + IB_QP_ATTR_MINRNRTIMER), + }; + GetAVFromPath(0, path, &modify_attr.PathMTU, NULL, + &modify_attr.DestAV); + + frc = iibt_qp_modify(qp_handle, &modify_attr, NULL); + if (frc != FSUCCESS) + RETURN(frc); + + modify_attr = (IB_QP_ATTRIBUTES_MODIFY) { + .RequestState = QPStateReadyToSend, + .FlowControl = TRUE, + .InitiatorDepth = init_depth, + .SendPSN = send_psn, + .LocalAckTimeout = path->PktLifeTime + 2, /* 2 or 1? */ + .RetryCount = IBNAL_RETRY, + .RnrRetryCount = IBNAL_RNR_RETRY, + .Attrs = (IB_QP_ATTR_FLOWCONTROL | + IB_QP_ATTR_INITIATORDEPTH | + IB_QP_ATTR_SENDPSN | + IB_QP_ATTR_LOCALACKTIMEOUT | + IB_QP_ATTR_RETRYCOUNT | + IB_QP_ATTR_RNRRETRYCOUNT), + }; + + frc = iibt_qp_modify(qp_handle, &modify_attr, NULL); + RETURN(frc); +} + +static void +kibnal_connect_reply (IB_HANDLE cep, CM_CONN_INFO *info, void *arg) +{ + IB_CA_ATTRIBUTES *ca_attr = &kibnal_data.kib_hca_attrs; + kib_conn_t *conn = arg; + kib_wire_connreq_t *wcr; + CM_REPLY_INFO *rep = &info->Info.Reply; + uint16_t reason; + FSTATUS frc; + + wcr = (kib_wire_connreq_t *)info->Info.Reply.PrivateData; + + if (wcr->wcr_magic != cpu_to_le32(IBNAL_MSG_MAGIC)) { + CERROR ("Can't connect "LPX64": bad magic %08x\n", + conn->ibc_peer->ibp_nid, le32_to_cpu(wcr->wcr_magic)); + GOTO(reject, reason = RC_USER_REJ); + } + + if (wcr->wcr_version != cpu_to_le16(IBNAL_MSG_VERSION)) { + CERROR ("Can't connect "LPX64": bad version %d\n", + conn->ibc_peer->ibp_nid, le16_to_cpu(wcr->wcr_magic)); + GOTO(reject, reason = RC_USER_REJ); + } + + if (wcr->wcr_queue_depth != cpu_to_le16(IBNAL_MSG_QUEUE_SIZE)) { + CERROR ("Can't connect "LPX64": bad queue depth %d\n", + conn->ibc_peer->ibp_nid, + le16_to_cpu(wcr->wcr_queue_depth)); + GOTO(reject, reason = RC_USER_REJ); + } + + if (le64_to_cpu(wcr->wcr_nid) != conn->ibc_peer->ibp_nid) { + CERROR ("Unexpected NID "LPX64" from "LPX64"\n", + le64_to_cpu(wcr->wcr_nid), conn->ibc_peer->ibp_nid); + GOTO(reject, reason = RC_USER_REJ); + } + + CDEBUG(D_NET, "Connection %p -> "LPX64" REP_RECEIVED.\n", + conn, conn->ibc_peer->ibp_nid); + + conn->ibc_incarnation = le64_to_cpu(wcr->wcr_incarnation); + conn->ibc_credits = IBNAL_MSG_QUEUE_SIZE; + + frc = kibnal_qp_rts(conn->ibc_qp, rep->QPN, + min_t(__u8, rep->ArbInitiatorDepth, + ca_attr->MaxQPResponderResources), + &conn->ibc_connreq->cr_path, + min_t(__u8, rep->ArbResponderResources, + ca_attr->MaxQPInitiatorDepth), + rep->StartingPSN); + if (frc != FSUCCESS) { + CERROR("Connection %p -> "LPX64" QP RTS/RTR failed: %d\n", + conn, conn->ibc_peer->ibp_nid, frc); + GOTO(reject, reason = RC_NO_QP); + } + + /* the callback arguments are ignored for an active accept */ + conn->ibc_connreq->cr_discarded.Status = FSUCCESS; + frc = iibt_cm_accept(cep, &conn->ibc_connreq->cr_discarded, + NULL, NULL, NULL, NULL); + if (frc != FCM_CONNECT_ESTABLISHED) { + CERROR("Connection %p -> "LPX64" CMAccept failed: %d\n", + conn, conn->ibc_peer->ibp_nid, frc); + kibnal_connreq_done (conn, 1, -ECONNABORTED); + /* XXX don't call reject after accept fails? */ + return; + } + + CDEBUG(D_NET, "Connection %p -> "LPX64" Established\n", + conn, conn->ibc_peer->ibp_nid); + + kibnal_connreq_done (conn, 1, 0); + return; + +reject: + kibnal_reject(cep, reason); + kibnal_connreq_done (conn, 1, -EPROTO); +} + +/* ib_cm.h has a wealth of information on the CM procedures */ +static void +kibnal_cm_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg) +{ + kib_conn_t *conn = arg; + + CDEBUG(D_NET, "status 0x%x\n", info->Status); + + /* Established Connection Notifier */ + switch (info->Status) { + default: + CERROR("unknown status %d on Connection %p -> "LPX64"\n", + info->Status, conn, conn->ibc_peer->ibp_nid); + LBUG(); + break; + + case FCM_CONNECT_REPLY: + kibnal_connect_reply(cep, info, arg); + break; + + case FCM_DISCONNECT_REQUEST: + /* XXX lock around these state management bits? */ + if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) + kibnal_close_conn (conn, 0); + conn->ibc_state = IBNAL_CONN_DREP; + iibt_cm_disconnect(conn->ibc_cep, NULL, NULL); + break; + + /* these both guarantee that no more cm callbacks will occur */ + case FCM_DISCONNECTED: /* aka FCM_DISCONNECT_TIMEOUT */ + case FCM_DISCONNECT_REPLY: + CDEBUG(D_NET, "Connection %p -> "LPX64" disconnect done.\n", + conn, conn->ibc_peer->ibp_nid); + + conn->ibc_state = IBNAL_CONN_DISCONNECTED; + kibnal_flush_pending(conn); + kibnal_put_conn(conn); /* Lose CM's ref */ + break; + } + + return; +} + +static int +kibnal_set_cm_flags(IB_HANDLE cep) +{ + FSTATUS frc; + uint32 value = 1; + + frc = iibt_cm_modify_cep(cep, CM_FLAG_TIMEWAIT_CALLBACK, + (char *)&value, sizeof(value), 0); + if (frc != FSUCCESS) { + CERROR("error setting timeout callback: %d\n", frc); + return -1; + } + +#if 0 + frc = iibt_cm_modify_cep(cep, CM_FLAG_ASYNC_ACCEPT, (char *)&value, + sizeof(value), 0); + if (frc != FSUCCESS) { + CERROR("error setting async accept: %d\n", frc); + return -1; + } +#endif + + return 0; +} + +void +kibnal_listen_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg) +{ + IB_CA_ATTRIBUTES *ca_attr = &kibnal_data.kib_hca_attrs; + IB_QP_ATTRIBUTES_QUERY *query; + CM_REQUEST_INFO *req; + CM_CONN_INFO *rep = NULL, *rcv = NULL; + kib_wire_connreq_t *wcr; + kib_conn_t *conn = NULL; + uint16_t reason = 0; + FSTATUS frc; + int rc = 0; + + LASSERT(cep); + LASSERT(info); + LASSERT(arg == NULL); /* no conn yet for passive */ + + CDEBUG(D_NET, "status 0x%x\n", info->Status); + + req = &info->Info.Request; + wcr = (kib_wire_connreq_t *)req->PrivateData; + + CDEBUG(D_NET, "%d from "LPX64"\n", info->Status, + le64_to_cpu(wcr->wcr_nid)); + + if (info->Status == FCM_CONNECT_CANCEL) + return; + + LASSERT (info->Status == FCM_CONNECT_REQUEST); + + if (wcr->wcr_magic != cpu_to_le32(IBNAL_MSG_MAGIC)) { + CERROR ("Can't accept: bad magic %08x\n", + le32_to_cpu(wcr->wcr_magic)); + GOTO(out, reason = RC_USER_REJ); + } + + if (wcr->wcr_version != cpu_to_le16(IBNAL_MSG_VERSION)) { + CERROR ("Can't accept: bad version %d\n", + le16_to_cpu(wcr->wcr_magic)); + GOTO(out, reason = RC_USER_REJ); + } + + rc = kibnal_accept(&conn, cep, + le64_to_cpu(wcr->wcr_nid), + le64_to_cpu(wcr->wcr_incarnation), + le16_to_cpu(wcr->wcr_queue_depth)); + if (rc != 0) { + CERROR ("Can't accept "LPX64": %d\n", + le64_to_cpu(wcr->wcr_nid), rc); + GOTO(out, reason = RC_NO_RESOURCES); + } + + frc = kibnal_qp_rts(conn->ibc_qp, req->CEPInfo.QPN, + min_t(__u8, req->CEPInfo.OfferedInitiatorDepth, + ca_attr->MaxQPResponderResources), + &req->PathInfo.Path, + min_t(__u8, req->CEPInfo.OfferedResponderResources, + ca_attr->MaxQPInitiatorDepth), + req->CEPInfo.StartingPSN); + + if (frc != FSUCCESS) { + CERROR ("Can't mark QP RTS/RTR "LPX64": %d\n", + le64_to_cpu(wcr->wcr_nid), frc); + GOTO(out, reason = RC_NO_QP); + } + + frc = iibt_qp_query(conn->ibc_qp, &conn->ibc_qp_attrs, NULL); + if (frc != FSUCCESS) { + CERROR ("Couldn't query qp attributes "LPX64": %d\n", + le64_to_cpu(wcr->wcr_nid), frc); + GOTO(out, reason = RC_NO_QP); + } + query = &conn->ibc_qp_attrs; + + PORTAL_ALLOC(rep, sizeof(*rep)); + PORTAL_ALLOC(rcv, sizeof(*rcv)); + if (rep == NULL || rcv == NULL) { + CERROR ("can't reply and receive buffers\n"); + GOTO(out, reason = RC_INSUFFICIENT_RESP_RES); + } + + /* don't try to deref this into the incoming wcr :) */ + wcr = (kib_wire_connreq_t *)rep->Info.Reply.PrivateData; + + rep->Info.Reply = (CM_REPLY_INFO) { + .QPN = query->QPNumber, + .QKey = query->Qkey, + .StartingPSN = query->RecvPSN, + .EndToEndFlowControl = query->FlowControl, + /* XXX Hmm. */ + .ArbInitiatorDepth = query->InitiatorDepth, + .ArbResponderResources = query->ResponderResources, + .TargetAckDelay = 0, + .FailoverAccepted = 0, + .RnRRetryCount = req->CEPInfo.RnrRetryCount, + }; + + *wcr = (kib_wire_connreq_t) { + .wcr_magic = cpu_to_le32(IBNAL_MSG_MAGIC), + .wcr_version = cpu_to_le16(IBNAL_MSG_VERSION), + .wcr_queue_depth = cpu_to_le32(IBNAL_MSG_QUEUE_SIZE), + .wcr_nid = cpu_to_le64(kibnal_data.kib_nid), + .wcr_incarnation = cpu_to_le64(kibnal_data.kib_incarnation), + }; + + frc = iibt_cm_accept(cep, rep, rcv, kibnal_cm_callback, conn, + &conn->ibc_cep); + + PORTAL_FREE(rep, sizeof(*rep)); + PORTAL_FREE(rcv, sizeof(*rcv)); + + if (frc != FCM_CONNECT_ESTABLISHED) { + /* XXX it seems we don't call reject after this point? */ + CERROR("iibt_cm_accept() failed: %d, aborting\n", frc); + rc = -ECONNABORTED; + goto out; + } + + if (kibnal_set_cm_flags(conn->ibc_cep)) { + rc = -ECONNABORTED; + goto out; + } + + CDEBUG(D_WARNING, "Connection %p -> "LPX64" ESTABLISHED.\n", + conn, conn->ibc_peer->ibp_nid); + +out: + if (reason) { + kibnal_reject(cep, reason); + rc = -ECONNABORTED; + } + if (conn != NULL) + kibnal_connreq_done(conn, 0, rc); + + return; +} + +static void +dump_path_records(PATH_RESULTS *results) +{ + IB_PATH_RECORD *path; + int i; + + for(i = 0; i < results->NumPathRecords; i++) { + path = &results->PathRecords[i]; + CDEBUG(D_NET, "%d: sgid "LPX64":"LPX64" dgid " + LPX64":"LPX64" pkey %x\n", + i, + path->SGID.Type.Global.SubnetPrefix, + path->SGID.Type.Global.InterfaceID, + path->DGID.Type.Global.SubnetPrefix, + path->DGID.Type.Global.InterfaceID, + path->P_Key); + } +} + +static void +kibnal_pathreq_callback (void *arg, QUERY *query, + QUERY_RESULT_VALUES *query_res) +{ + IB_CA_ATTRIBUTES *ca_attr = &kibnal_data.kib_hca_attrs; + kib_conn_t *conn = arg; + PATH_RESULTS *path; + FSTATUS frc; + + if (query_res->Status != FSUCCESS || query_res->ResultDataSize == 0) { + CERROR ("status %d data size %d\n", query_res->Status, + query_res->ResultDataSize); + kibnal_connreq_done (conn, 1, -EINVAL); + return; + } + + path = (PATH_RESULTS *)query_res->QueryResult; + + if (path->NumPathRecords < 1) { + CERROR ("expected path records: %d\n", path->NumPathRecords); + kibnal_connreq_done (conn, 1, -EINVAL); + return; + } + + dump_path_records(path); + + /* just using the first. this is probably a horrible idea. */ + conn->ibc_connreq->cr_path = path->PathRecords[0]; + + conn->ibc_cep = iibt_cm_create_cep(CM_RC_TYPE); + if (conn->ibc_cep == NULL) { + CERROR ("Can't create CEP\n"); + kibnal_connreq_done (conn, 1, -EINVAL); + return; + } + + if (kibnal_set_cm_flags(conn->ibc_cep)) { + kibnal_connreq_done (conn, 1, -EINVAL); + return; + } + + conn->ibc_connreq->cr_wcr = (kib_wire_connreq_t) { + .wcr_magic = cpu_to_le32(IBNAL_MSG_MAGIC), + .wcr_version = cpu_to_le16(IBNAL_MSG_VERSION), + .wcr_queue_depth = cpu_to_le16(IBNAL_MSG_QUEUE_SIZE), + .wcr_nid = cpu_to_le64(kibnal_data.kib_nid), + .wcr_incarnation = cpu_to_le64(kibnal_data.kib_incarnation), + }; + + conn->ibc_connreq->cr_cmreq = (CM_REQUEST_INFO) { + .SID = conn->ibc_connreq->cr_service.RID.ServiceID, + .CEPInfo = (CM_CEP_INFO) { + .CaGUID = kibnal_data.kib_hca_guids[0], + .EndToEndFlowControl = FALSE, + .PortGUID = conn->ibc_connreq->cr_path.SGID.Type.Global.InterfaceID, + .RetryCount = IBNAL_RETRY, + .RnrRetryCount = IBNAL_RNR_RETRY, + .AckTimeout = IBNAL_ACK_TIMEOUT, + .StartingPSN = IBNAL_STARTING_PSN, + .QPN = conn->ibc_qp_attrs.QPNumber, + .QKey = conn->ibc_qp_attrs.Qkey, + .OfferedResponderResources = ca_attr->MaxQPResponderResources, + .OfferedInitiatorDepth = ca_attr->MaxQPInitiatorDepth, + }, + .PathInfo = (CM_CEP_PATHINFO) { + .bSubnetLocal = TRUE, + .Path = conn->ibc_connreq->cr_path, + }, + }; + +#if 0 + /* XXX set timeout just like SDP!!!*/ + conn->ibc_connreq->cr_path.packet_life = 13; +#endif + /* Flag I'm getting involved with the CM... */ + conn->ibc_state = IBNAL_CONN_CONNECTING; + + CDEBUG(D_NET, "Connecting to, service id "LPX64", on "LPX64"\n", + conn->ibc_connreq->cr_service.RID.ServiceID, + *kibnal_service_nid_field(&conn->ibc_connreq->cr_service)); + + memset(conn->ibc_connreq->cr_cmreq.PrivateData, 0, + CM_REQUEST_INFO_USER_LEN); + memcpy(conn->ibc_connreq->cr_cmreq.PrivateData, + &conn->ibc_connreq->cr_wcr, sizeof(conn->ibc_connreq->cr_wcr)); + + /* kibnal_cm_callback gets my conn ref */ + frc = iibt_cm_connect(conn->ibc_cep, &conn->ibc_connreq->cr_cmreq, + kibnal_cm_callback, conn); + if (frc != FPENDING && frc != FSUCCESS) { + CERROR ("Connect: %d\n", frc); + /* Back out state change as connect failed */ + conn->ibc_state = IBNAL_CONN_INIT_QP; + kibnal_connreq_done (conn, 1, -EINVAL); + } +} + +static void +dump_service_records(SERVICE_RECORD_RESULTS *results) +{ + IB_SERVICE_RECORD *svc; + int i; + + for(i = 0; i < results->NumServiceRecords; i++) { + svc = &results->ServiceRecords[i]; + CDEBUG(D_NET, "%d: sid "LPX64" gid "LPX64":"LPX64" pkey %x\n", + i, + svc->RID.ServiceID, + svc->RID.ServiceGID.Type.Global.SubnetPrefix, + svc->RID.ServiceGID.Type.Global.InterfaceID, + svc->RID.ServiceP_Key); + } +} + + +static void +kibnal_service_get_callback (void *arg, QUERY *query, + QUERY_RESULT_VALUES *query_res) +{ + kib_conn_t *conn = arg; + SERVICE_RECORD_RESULTS *svc; + COMMAND_CONTROL_PARAMETERS sd_params; + QUERY path_query; + FSTATUS frc; + + if (query_res->Status != FSUCCESS || query_res->ResultDataSize == 0) { + CERROR ("status %d data size %d\n", query_res->Status, + query_res->ResultDataSize); + kibnal_connreq_done (conn, 1, -EINVAL); + return; + } + + svc = (SERVICE_RECORD_RESULTS *)query_res->QueryResult; + + if (svc->NumServiceRecords < 1) { + CERROR ("%d service records\n", svc->NumServiceRecords); + kibnal_connreq_done (conn, 1, -EINVAL); + return; + } + + dump_service_records(svc); + + conn->ibc_connreq->cr_service = svc->ServiceRecords[0]; + + CDEBUG(D_NET, "Got status %d, service id "LPX64", on "LPX64"\n", + query_res->Status , conn->ibc_connreq->cr_service.RID.ServiceID, + *kibnal_service_nid_field(&conn->ibc_connreq->cr_service)); + + memset(&path_query, 0, sizeof(path_query)); + path_query.InputType = InputTypePortGuidPair; + path_query.OutputType = OutputTypePathRecord; + path_query.InputValue.PortGuidPair.SourcePortGuid = kibnal_data.kib_port_guid; + path_query.InputValue.PortGuidPair.DestPortGuid = conn->ibc_connreq->cr_service.RID.ServiceGID.Type.Global.InterfaceID; + + memset(&sd_params, 0, sizeof(sd_params)); + sd_params.RetryCount = IBNAL_RETRY; + sd_params.Timeout = 10 * 1000; /* wait 10 seconds */ + + /* kibnal_service_get_callback gets my conn ref */ + + frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd, + kibnal_data.kib_port_guid, + &path_query, + kibnal_pathreq_callback, + &sd_params, conn); + if (frc == FPENDING) + return; + + CERROR ("Path record request failed: %d\n", frc); + kibnal_connreq_done (conn, 1, -EINVAL); +} + +static void +kibnal_connect_peer (kib_peer_t *peer) +{ + COMMAND_CONTROL_PARAMETERS sd_params; + QUERY query; + FSTATUS frc; + kib_conn_t *conn = kibnal_create_conn(); + + LASSERT (peer->ibp_connecting != 0); + + if (conn == NULL) { + CERROR ("Can't allocate conn\n"); + kibnal_peer_connect_failed (peer, 1, -ENOMEM); + return; + } + + conn->ibc_peer = peer; + kib_peer_addref(peer); + + PORTAL_ALLOC (conn->ibc_connreq, sizeof (*conn->ibc_connreq)); + if (conn->ibc_connreq == NULL) { + CERROR ("Can't allocate connreq\n"); + kibnal_connreq_done (conn, 1, -ENOMEM); + return; + } + + memset(conn->ibc_connreq, 0, sizeof (*conn->ibc_connreq)); + + kibnal_set_service_keys(&conn->ibc_connreq->cr_service, peer->ibp_nid); + + memset(&query, 0, sizeof(query)); + query.InputType = InputTypeServiceRecord; + query.OutputType = OutputTypeServiceRecord; + query.InputValue.ServiceRecordValue.ServiceRecord = conn->ibc_connreq->cr_service; + query.InputValue.ServiceRecordValue.ComponentMask = KIBNAL_SERVICE_KEY_MASK; + + memset(&sd_params, 0, sizeof(sd_params)); + sd_params.RetryCount = IBNAL_RETRY; + sd_params.Timeout = 10 * 1000; /* wait 10 seconds */ + + /* kibnal_service_get_callback gets my conn ref */ + frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd, + kibnal_data.kib_port_guid, + &query, + kibnal_service_get_callback, + &sd_params, conn); + if (frc == FPENDING) + return; + + CERROR ("iibt_sd_query_port_fabric_information(): %d\n", frc); + kibnal_connreq_done (conn, 1, frc); +} + +static int +kibnal_conn_timed_out (kib_conn_t *conn) +{ + kib_tx_t *tx; + struct list_head *ttmp; + unsigned long flags; + + spin_lock_irqsave (&conn->ibc_lock, flags); + + list_for_each (ttmp, &conn->ibc_tx_queue) { + tx = list_entry (ttmp, kib_tx_t, tx_list); + + LASSERT (!tx->tx_passive_rdma_wait); + LASSERT (tx->tx_sending == 0); + + if (time_after_eq (jiffies, tx->tx_deadline)) { + spin_unlock_irqrestore (&conn->ibc_lock, flags); + return 1; + } + } + + list_for_each (ttmp, &conn->ibc_active_txs) { + tx = list_entry (ttmp, kib_tx_t, tx_list); + + LASSERT (tx->tx_passive_rdma || + !tx->tx_passive_rdma_wait); + + LASSERT (tx->tx_passive_rdma_wait || + tx->tx_sending != 0); + + if (time_after_eq (jiffies, tx->tx_deadline)) { + spin_unlock_irqrestore (&conn->ibc_lock, flags); + return 1; + } + } + + spin_unlock_irqrestore (&conn->ibc_lock, flags); + + return 0; +} + +static void +kibnal_check_conns (int idx) +{ + struct list_head *peers = &kibnal_data.kib_peers[idx]; + struct list_head *ptmp; + kib_peer_t *peer; + kib_conn_t *conn; + struct list_head *ctmp; + + again: + /* NB. We expect to have a look at all the peers and not find any + * rdmas to time out, so we just use a shared lock while we + * take a look... */ + read_lock (&kibnal_data.kib_global_lock); + + list_for_each (ptmp, peers) { + peer = list_entry (ptmp, kib_peer_t, ibp_list); + + list_for_each (ctmp, &peer->ibp_conns) { + conn = list_entry (ctmp, kib_conn_t, ibc_list); + + KIB_ASSERT_CONN_STATE(conn, IBNAL_CONN_ESTABLISHED); + + /* In case we have enough credits to return via a + * NOOP, but there were no non-blocking tx descs + * free to do it last time... */ + kibnal_check_sends(conn); + + if (!kibnal_conn_timed_out(conn)) + continue; + + CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n", + conn, conn->ibc_state, peer->ibp_nid, + atomic_read (&conn->ibc_refcount)); + + atomic_inc (&conn->ibc_refcount); + read_unlock (&kibnal_data.kib_global_lock); + + CERROR("Timed out RDMA with "LPX64"\n", + peer->ibp_nid); + + kibnal_close_conn (conn, -ETIMEDOUT); + kibnal_put_conn (conn); + + /* start again now I've dropped the lock */ + goto again; + } + } + + read_unlock (&kibnal_data.kib_global_lock); +} + +static void +kib_connd_handle_state(kib_conn_t *conn) +{ + FSTATUS frc; + + switch (conn->ibc_state) { + /* all refs have gone, free and be done with it */ + case IBNAL_CONN_DISCONNECTED: + kibnal_destroy_conn (conn); + return; /* avoid put_conn */ + + case IBNAL_CONN_SEND_DREQ: + frc = iibt_cm_disconnect(conn->ibc_cep, NULL, NULL); + if (frc != FSUCCESS) /* XXX do real things */ + CERROR("disconnect failed: %d\n", frc); + conn->ibc_state = IBNAL_CONN_DREQ; + break; + + /* a callback got to the conn before we did */ + case IBNAL_CONN_DREP: + break; + + default: + CERROR ("Bad conn %p state: %d\n", conn, + conn->ibc_state); + LBUG(); + break; + } + + /* drop ref from close_conn */ + kibnal_put_conn(conn); +} + +int +kibnal_connd (void *arg) +{ + wait_queue_t wait; + unsigned long flags; + kib_conn_t *conn; + kib_peer_t *peer; + int timeout; + int i; + int peer_index = 0; + unsigned long deadline = jiffies; + + kportal_daemonize ("kibnal_connd"); + kportal_blockallsigs (); + + init_waitqueue_entry (&wait, current); + + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + + for (;;) { + if (!list_empty (&kibnal_data.kib_connd_conns)) { + conn = list_entry (kibnal_data.kib_connd_conns.next, + kib_conn_t, ibc_list); + list_del (&conn->ibc_list); + + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); + kib_connd_handle_state(conn); + + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + continue; + } + + if (!list_empty (&kibnal_data.kib_connd_peers)) { + peer = list_entry (kibnal_data.kib_connd_peers.next, + kib_peer_t, ibp_connd_list); + + list_del_init (&peer->ibp_connd_list); + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); + + kibnal_connect_peer (peer); + kib_peer_decref (peer); + + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + } + + /* shut down and nobody left to reap... */ + if (kibnal_data.kib_shutdown && + atomic_read(&kibnal_data.kib_nconns) == 0) + break; + + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); + + /* careful with the jiffy wrap... */ + while ((timeout = (int)(deadline - jiffies)) <= 0) { + const int n = 4; + const int p = 1; + int chunk = kibnal_data.kib_peer_hash_size; + + /* Time to check for RDMA timeouts on a few more + * peers: I do checks every 'p' seconds on a + * proportion of the peer table and I need to check + * every connection 'n' times within a timeout + * interval, to ensure I detect a timeout on any + * connection within (n+1)/n times the timeout + * interval. */ + + if (kibnal_tunables.kib_io_timeout > n * p) + chunk = (chunk * n * p) / + kibnal_tunables.kib_io_timeout; + if (chunk == 0) + chunk = 1; + + for (i = 0; i < chunk; i++) { + kibnal_check_conns (peer_index); + peer_index = (peer_index + 1) % + kibnal_data.kib_peer_hash_size; + } + + deadline += p * HZ; + } + + kibnal_data.kib_connd_waketime = jiffies + timeout; + + set_current_state (TASK_INTERRUPTIBLE); + add_wait_queue (&kibnal_data.kib_connd_waitq, &wait); + + if (!kibnal_data.kib_shutdown && + list_empty (&kibnal_data.kib_connd_conns) && + list_empty (&kibnal_data.kib_connd_peers)) + schedule_timeout (timeout); + + set_current_state (TASK_RUNNING); + remove_wait_queue (&kibnal_data.kib_connd_waitq, &wait); + + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + } + + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); + + kibnal_thread_fini (); + return (0); +} + +int +kibnal_scheduler(void *arg) +{ + long id = (long)arg; + char name[16]; + kib_rx_t *rx; + kib_tx_t *tx; + unsigned long flags; + int rc; + int counter = 0; + int did_something; + + snprintf(name, sizeof(name), "kibnal_sd_%02ld", id); + kportal_daemonize(name); + kportal_blockallsigs(); + + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); + + for (;;) { + did_something = 0; + + while (!list_empty(&kibnal_data.kib_sched_txq)) { + tx = list_entry(kibnal_data.kib_sched_txq.next, + kib_tx_t, tx_list); + list_del(&tx->tx_list); + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, + flags); + kibnal_tx_done(tx); + + spin_lock_irqsave(&kibnal_data.kib_sched_lock, + flags); + } + + if (!list_empty(&kibnal_data.kib_sched_rxq)) { + rx = list_entry(kibnal_data.kib_sched_rxq.next, + kib_rx_t, rx_list); + list_del(&rx->rx_list); + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, + flags); + + kibnal_rx(rx); + + did_something = 1; + spin_lock_irqsave(&kibnal_data.kib_sched_lock, + flags); + } + + /* shut down and no receives to complete... */ + if (kibnal_data.kib_shutdown && + atomic_read(&kibnal_data.kib_nconns) == 0) + break; + + /* nothing to do or hogging CPU */ + if (!did_something || counter++ == IBNAL_RESCHED) { + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, + flags); + counter = 0; + + if (!did_something) { + rc = wait_event_interruptible( + kibnal_data.kib_sched_waitq, + !list_empty(&kibnal_data.kib_sched_txq) || + !list_empty(&kibnal_data.kib_sched_rxq) || + (kibnal_data.kib_shutdown && + atomic_read (&kibnal_data.kib_nconns) == 0)); + } else { + our_cond_resched(); + } + + spin_lock_irqsave(&kibnal_data.kib_sched_lock, + flags); + } + } + + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); + + kibnal_thread_fini(); + return (0); +} + + +lib_nal_t kibnal_lib = { + libnal_data: &kibnal_data, /* NAL private data */ + libnal_send: kibnal_send, + libnal_send_pages: kibnal_send_pages, + libnal_recv: kibnal_recv, + libnal_recv_pages: kibnal_recv_pages, + libnal_dist: kibnal_dist +}; -- 1.8.3.1