Whamcloud - gitweb
b=23179 MDS node unresponsive
[fs/lustre-release.git] / lnet / lnet / api-ni.c
index 56afd45..3b97ea5 100644 (file)
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- * api/api-ni.c
- * Network Interface code
+ * GPL HEADER START
  *
- *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
- *  Copyright (c) 2001-2002 Sandia National Laboratories
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *   This file is part of Lustre, http://www.sf.net/projects/lustre/
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
  *
- *   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.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
  *
- *   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
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  *
- *   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.
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
  */
 
-#define DEBUG_SUBSYSTEM S_PORTALS
-#include <portals/api-support.h>
-
-int ptl_init;
+#define DEBUG_SUBSYSTEM S_LNET
+#include <lnet/lib-lnet.h>
 
-/* Put some magic in the NI handle so uninitialised/zeroed handles are easy
- * to spot */
-#define NI_HANDLE_MAGIC  0xebc0de00
-#define NI_HANDLE_MASK   0x000000ff
+#ifdef __KERNEL__
+#define D_LNI D_CONSOLE
+#else
+#define D_LNI D_CONFIG
+#endif
 
-static struct nal_t *ptl_nal_table[NAL_MAX_NR];
+lnet_t      the_lnet;                           /* THE state of the network */
 
 #ifdef __KERNEL__
-DECLARE_MUTEX(ptl_mutex);
 
-static void ptl_mutex_enter (void) 
+static char *ip2nets = "";
+CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
+                "LNET network <- IP table");
+
+static char *networks = "";
+CFS_MODULE_PARM(networks, "s", charp, 0444,
+                "local networks");
+
+static char *routes = "";
+CFS_MODULE_PARM(routes, "s", charp, 0444,
+                "routes to non-local networks");
+
+char *
+lnet_get_routes(void)
+{
+        return routes;
+}
+
+char *
+lnet_get_networks(void)
+{
+        char   *nets;
+        int     rc;
+
+        if (*networks != 0 && *ip2nets != 0) {
+                LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
+                                   "'ip2nets' but not both at once\n");
+                return NULL;
+        }
+
+        if (*ip2nets != 0) {
+                rc = lnet_parse_ip2nets(&nets, ip2nets);
+                return (rc == 0) ? nets : NULL;
+        }
+
+        if (*networks != 0)
+                return networks;
+
+        return "tcp";
+}
+
+void
+lnet_init_locks(void)
+{
+        cfs_spin_lock_init (&the_lnet.ln_lock);
+        cfs_waitq_init (&the_lnet.ln_waitq);
+        cfs_init_mutex(&the_lnet.ln_lnd_mutex);
+        cfs_init_mutex(&the_lnet.ln_api_mutex);
+}
+
+void
+lnet_fini_locks(void)
 {
-        down (&ptl_mutex);
 }
 
-static void ptl_mutex_exit (void)
+#else
+
+char *
+lnet_get_routes(void)
 {
-        up (&ptl_mutex);
+        char *str = getenv("LNET_ROUTES");
+
+        return (str == NULL) ? "" : str;
 }
+
+char *
+lnet_get_networks (void)
+{
+        static char       default_networks[256];
+        char             *networks = getenv ("LNET_NETWORKS");
+        char             *ip2nets  = getenv ("LNET_IP2NETS");
+        char             *str;
+        char             *sep;
+        int               len;
+        int               nob;
+        int               rc;
+        cfs_list_t       *tmp;
+
+#ifdef NOT_YET
+        if (networks != NULL && ip2nets != NULL) {
+                LCONSOLE_ERROR_MSG(0x103, "Please set EITHER 'LNET_NETWORKS' or"
+                                   " 'LNET_IP2NETS' but not both at once\n");
+                return NULL;
+        }
+
+        if (ip2nets != NULL) {
+                rc = lnet_parse_ip2nets(&networks, ip2nets);
+                return (rc == 0) ? networks : NULL;
+        }
 #else
-static void ptl_mutex_enter (void)
+        ip2nets = NULL;
+        rc = 0;
+#endif
+        if (networks != NULL)
+                return networks;
+
+        /* In userland, the default 'networks=' is the list of known net types */
+
+        len = sizeof(default_networks);
+        str = default_networks;
+        *str = 0;
+        sep = "";
+
+        cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
+                lnd_t *lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
+
+                nob = snprintf(str, len, "%s%s", sep,
+                               libcfs_lnd2str(lnd->lnd_type));
+                len -= nob;
+                if (len < 0) {
+                        /* overflowed the string; leave it where it was */
+                        *str = 0;
+                        break;
+                }
+
+                str += nob;
+                sep = ",";
+        }
+
+        return default_networks;
+}
+
+# ifndef HAVE_LIBPTHREAD
+
+void lnet_init_locks(void)
 {
+        the_lnet.ln_lock = 0;
+        the_lnet.ln_lnd_mutex = 0;
+        the_lnet.ln_api_mutex = 0;
 }
 
-static void ptl_mutex_exit (void) 
+void lnet_fini_locks(void)
 {
+        LASSERT (the_lnet.ln_api_mutex == 0);
+        LASSERT (the_lnet.ln_lnd_mutex == 0);
+        LASSERT (the_lnet.ln_lock == 0);
 }
+
+# else
+
+void lnet_init_locks(void)
+{
+        pthread_cond_init(&the_lnet.ln_cond, NULL);
+        pthread_mutex_init(&the_lnet.ln_lock, NULL);
+        pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
+        pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
+}
+
+void lnet_fini_locks(void)
+{
+        pthread_mutex_destroy(&the_lnet.ln_api_mutex);
+        pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
+        pthread_mutex_destroy(&the_lnet.ln_lock);
+        pthread_cond_destroy(&the_lnet.ln_cond);
+}
+
+# endif
 #endif
 
-nal_t *ptl_hndl2nal(ptl_handle_any_t *handle)
+void lnet_assert_wire_constants (void)
 {
-        unsigned int idx = handle->nal_idx;
+        /* Wire protocol assertions generated by 'wirecheck'
+         * running on Linux robert.bartonsoftware.com 2.6.8-1.521
+         * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
+         * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 
-        /* XXX we really rely on the caller NOT racing with interface
-         * setup/teardown.  That ensures her NI handle can't get
-         * invalidated out from under her (or worse, swapped for a
-         * completely different interface!) */
+        /* Constants... */
+        CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
+        CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
+        CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
+        CLASSERT (LNET_MSG_ACK == 0);
+        CLASSERT (LNET_MSG_PUT == 1);
+        CLASSERT (LNET_MSG_GET == 2);
+        CLASSERT (LNET_MSG_REPLY == 3);
+        CLASSERT (LNET_MSG_HELLO == 4);
 
-        LASSERT (ptl_init);
+        /* Checks for struct ptl_handle_wire_t */
+        CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
+        CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
+        CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
+        CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
+        CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 
-        if (((idx ^ NI_HANDLE_MAGIC) & ~NI_HANDLE_MASK) != 0)
-                return NULL;
+        /* Checks for struct lnet_magicversion_t */
+        CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
+        CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
+        CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
+        CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
+        CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
+        CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
+        CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 
-        idx &= NI_HANDLE_MASK;
-        
-        if (idx >= NAL_MAX_NR ||
-            ptl_nal_table[idx] == NULL ||
-            ptl_nal_table[idx]->nal_refct == 0)
-                return NULL;
+        /* Checks for struct lnet_hdr_t */
+        CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
+        CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
+        CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
+        CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
+        CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
+        CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
+        CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
+
+        /* Ack */
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
+
+        /* Put */
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 
-        return ptl_nal_table[idx];
+        /* Get */
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
+
+        /* Reply */
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
+
+        /* Hello */
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
+        CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
+        CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 }
 
-int ptl_register_nal (ptl_interface_t interface, nal_t *nal)
+lnd_t *
+lnet_find_lnd_by_type (int type)
 {
-        int    rc;
-        
-        ptl_mutex_enter();
-        
-        if (interface < 0 || interface >= NAL_MAX_NR)
-                rc = PTL_IFACE_INVALID;
-        else if (ptl_nal_table[interface] != NULL)
-                rc = PTL_IFACE_DUP;
-        else {
-                rc = PTL_OK;
-                ptl_nal_table[interface] = nal;
-                LASSERT(nal->nal_refct == 0);
+        lnd_t              *lnd;
+        cfs_list_t         *tmp;
+
+        /* holding lnd mutex */
+        cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
+                lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
+
+                if ((int)lnd->lnd_type == type)
+                        return lnd;
         }
 
-        ptl_mutex_exit();
-        return (rc);
+        return NULL;
+}
+
+void
+lnet_register_lnd (lnd_t *lnd)
+{
+        LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
+
+        LASSERT (the_lnet.ln_init);
+        LASSERT (libcfs_isknown_lnd(lnd->lnd_type));
+        LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
+
+        cfs_list_add_tail (&lnd->lnd_list, &the_lnet.ln_lnds);
+        lnd->lnd_refcount = 0;
+
+        CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
+
+        LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
+}
+
+void
+lnet_unregister_lnd (lnd_t *lnd)
+{
+        LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
+
+        LASSERT (the_lnet.ln_init);
+        LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
+        LASSERT (lnd->lnd_refcount == 0);
+
+        cfs_list_del (&lnd->lnd_list);
+        CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
+
+        LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
+}
+
+#ifndef LNET_USE_LIB_FREELIST
+
+int
+lnet_descriptor_setup (void)
+{
+        return 0;
+}
+
+void
+lnet_descriptor_cleanup (void)
+{
+}
+
+#else
+
+int
+lnet_freelist_init (lnet_freelist_t *fl, int n, int size)
+{
+        char *space;
+
+        LASSERT (n > 0);
+
+        size += offsetof (lnet_freeobj_t, fo_contents);
+
+        LIBCFS_ALLOC(space, n * size);
+        if (space == NULL)
+                return (-ENOMEM);
+
+        CFS_INIT_LIST_HEAD (&fl->fl_list);
+        fl->fl_objs = space;
+        fl->fl_nobjs = n;
+        fl->fl_objsize = size;
+
+        do
+        {
+                memset (space, 0, size);
+                cfs_list_add ((cfs_list_t *)space, &fl->fl_list);
+                space += size;
+        } while (--n != 0);
+
+        return (0);
 }
 
-void ptl_unregister_nal (ptl_interface_t interface)
+void
+lnet_freelist_fini (lnet_freelist_t *fl)
 {
-        LASSERT(interface >= 0 && interface < NAL_MAX_NR);
-        LASSERT(ptl_nal_table[interface] != NULL);
-        LASSERT(ptl_nal_table[interface]->nal_refct == 0);
-        
-        ptl_mutex_enter();
-        
-        ptl_nal_table[interface] = NULL;
+        cfs_list_t       *el;
+        int               count;
 
-        ptl_mutex_exit();
+        if (fl->fl_nobjs == 0)
+                return;
+
+        count = 0;
+        for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
+                count++;
+
+        LASSERT (count == fl->fl_nobjs);
+
+        LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
+        memset (fl, 0, sizeof (*fl));
 }
 
-int PtlInit(int *max_interfaces)
+int
+lnet_descriptor_setup (void)
 {
-        LASSERT(!strcmp(ptl_err_str[PTL_MAX_ERRNO], "PTL_MAX_ERRNO"));
+        /* NB on failure caller must still call lnet_descriptor_cleanup */
+        /*               ******                                         */
+        int        rc;
+
+        memset (&the_lnet.ln_free_mes,  0, sizeof (the_lnet.ln_free_mes));
+        memset (&the_lnet.ln_free_msgs, 0, sizeof (the_lnet.ln_free_msgs));
+        memset (&the_lnet.ln_free_mds,  0, sizeof (the_lnet.ln_free_mds));
+        memset (&the_lnet.ln_free_eqs,  0, sizeof (the_lnet.ln_free_eqs));
+
+        rc = lnet_freelist_init(&the_lnet.ln_free_mes,
+                                MAX_MES, sizeof (lnet_me_t));
+        if (rc != 0)
+                return (rc);
+
+        rc = lnet_freelist_init(&the_lnet.ln_free_msgs,
+                                MAX_MSGS, sizeof (lnet_msg_t));
+        if (rc != 0)
+                return (rc);
+
+        rc = lnet_freelist_init(&the_lnet.ln_free_mds,
+                                MAX_MDS, sizeof (lnet_libmd_t));
+        if (rc != 0)
+                return (rc);
+
+        rc = lnet_freelist_init(&the_lnet.ln_free_eqs,
+                                MAX_EQS, sizeof (lnet_eq_t));
+        return (rc);
+}
 
-        /* If this assertion fails, we need more bits in NI_HANDLE_MASK and
-         * to shift NI_HANDLE_MAGIC left appropriately */
-        LASSERT (NAL_MAX_NR <= (NI_HANDLE_MASK + 1));
-        
-        if (max_interfaces != NULL)
-                *max_interfaces = NAL_MAX_NR;
+void
+lnet_descriptor_cleanup (void)
+{
+        lnet_freelist_fini (&the_lnet.ln_free_mes);
+        lnet_freelist_fini (&the_lnet.ln_free_msgs);
+        lnet_freelist_fini (&the_lnet.ln_free_mds);
+        lnet_freelist_fini (&the_lnet.ln_free_eqs);
+}
 
-        ptl_mutex_enter();
+#endif
 
-        if (!ptl_init) {
-                /* NULL pointers, clear flags */
-                memset(ptl_nal_table, 0, sizeof(ptl_nal_table));
+__u64
+lnet_create_interface_cookie (void)
+{
+        /* NB the interface cookie in wire handles guards against delayed
+         * replies and ACKs appearing valid after reboot. Initialisation time,
+         * even if it's only implemented to millisecond resolution is probably
+         * easily good enough. */
+        struct timeval tv;
+        __u64          cookie;
 #ifndef __KERNEL__
-                /* Kernel NALs register themselves when their module loads,
-                 * and unregister themselves when their module is unloaded.
-                 * Userspace NALs, are plugged in explicitly here... */
-                {
-                        extern nal_t procapi_nal;
-
-                        /* XXX pretend it's socknal to keep liblustre happy... */
-                        ptl_nal_table[SOCKNAL] = &procapi_nal;
-                        LASSERT (procapi_nal.nal_refct == 0);
-                }
+        int            rc = gettimeofday (&tv, NULL);
+        LASSERT (rc == 0);
+#else
+        cfs_gettimeofday(&tv);
 #endif
-                ptl_init = 1;
+        cookie = tv.tv_sec;
+        cookie *= 1000000;
+        cookie += tv.tv_usec;
+        return cookie;
+}
+
+int
+lnet_setup_handle_hash (void)
+{
+        int       i;
+
+        /* Arbitrary choice of hash table size */
+#ifdef __KERNEL__
+        the_lnet.ln_lh_hash_size =
+                (2 * CFS_PAGE_SIZE) / sizeof (cfs_list_t);
+#else
+        the_lnet.ln_lh_hash_size = (MAX_MES + MAX_MDS + MAX_EQS)/4;
+#endif
+        LIBCFS_ALLOC(the_lnet.ln_lh_hash_table,
+                     the_lnet.ln_lh_hash_size * sizeof (cfs_list_t));
+        if (the_lnet.ln_lh_hash_table == NULL)
+                return (-ENOMEM);
+
+        for (i = 0; i < the_lnet.ln_lh_hash_size; i++)
+                CFS_INIT_LIST_HEAD (&the_lnet.ln_lh_hash_table[i]);
+
+        the_lnet.ln_next_object_cookie = LNET_COOKIE_TYPES;
+
+        return (0);
+}
+
+void
+lnet_cleanup_handle_hash (void)
+{
+        if (the_lnet.ln_lh_hash_table == NULL)
+                return;
+
+        LIBCFS_FREE(the_lnet.ln_lh_hash_table,
+                    the_lnet.ln_lh_hash_size * sizeof (cfs_list_t));
+}
+
+lnet_libhandle_t *
+lnet_lookup_cookie (__u64 cookie, int type)
+{
+        /* ALWAYS called with LNET_LOCK held */
+        cfs_list_t          *list;
+        cfs_list_t          *el;
+        unsigned int         hash;
+
+        if ((cookie & (LNET_COOKIE_TYPES - 1)) != type)
+                return (NULL);
+
+        hash = ((unsigned int)(cookie >> LNET_COOKIE_TYPE_BITS)) % the_lnet.ln_lh_hash_size;
+        list = &the_lnet.ln_lh_hash_table[hash];
+
+        cfs_list_for_each (el, list) {
+                lnet_libhandle_t *lh = cfs_list_entry (el, lnet_libhandle_t,
+                                                      lh_hash_chain);
+
+                if (lh->lh_cookie == cookie)
+                        return (lh);
         }
 
-        ptl_mutex_exit();
-        
-        return PTL_OK;
+        return (NULL);
 }
 
-void PtlFini(void)
+void
+lnet_initialise_handle (lnet_libhandle_t *lh, int type)
 {
-        nal_t  *nal;
-        int     i;
+        /* ALWAYS called with LNET_LOCK held */
+        unsigned int    hash;
 
-        ptl_mutex_enter();
+        LASSERT (type >= 0 && type < LNET_COOKIE_TYPES);
+        lh->lh_cookie = the_lnet.ln_next_object_cookie | type;
+        the_lnet.ln_next_object_cookie += LNET_COOKIE_TYPES;
 
-        if (ptl_init) {
-                for (i = 0; i < NAL_MAX_NR; i++) {
+        hash = ((unsigned int)(lh->lh_cookie >> LNET_COOKIE_TYPE_BITS)) % the_lnet.ln_lh_hash_size;
+        cfs_list_add (&lh->lh_hash_chain, &the_lnet.ln_lh_hash_table[hash]);
+}
 
-                        nal = ptl_nal_table[i];
-                        if (nal == NULL)
-                                continue;
-                        
-                        if (nal->nal_refct != 0) {
-                                CWARN("NAL %d has outstanding refcount %d\n",
-                                      i, nal->nal_refct);
-                                nal->nal_ni_fini(nal);
-                        }
-                        
-                        ptl_nal_table[i] = NULL;
+void
+lnet_invalidate_handle (lnet_libhandle_t *lh)
+{
+        /* ALWAYS called with LNET_LOCK held */
+        cfs_list_del (&lh->lh_hash_chain);
+}
+
+cfs_list_t *
+lnet_portal_mhash_alloc(void)
+{
+        cfs_list_t       *mhash;
+        int               i;
+
+        LIBCFS_ALLOC(mhash, sizeof(cfs_list_t) * LNET_PORTAL_HASH_SIZE);
+        if (mhash == NULL)
+                return NULL;
+
+        for (i = 0; i < LNET_PORTAL_HASH_SIZE; i++)
+                CFS_INIT_LIST_HEAD(&mhash[i]);
+
+        return mhash;
+}
+
+void
+lnet_portal_mhash_free(cfs_list_t *mhash)
+{
+        int     i;
+
+        for (i = 0; i < LNET_PORTAL_HASH_SIZE; i++) {
+                while (!cfs_list_empty(&mhash[i])) {
+                        lnet_me_t *me = cfs_list_entry(mhash[i].next,
+                                                       lnet_me_t, me_list);
+                        CERROR ("Active ME %p on exit portal mhash\n", me);
+                        cfs_list_del(&me->me_list);
+                        lnet_me_free(me);
                 }
+        }
+        LIBCFS_FREE(mhash, sizeof(cfs_list_t) * LNET_PORTAL_HASH_SIZE);
+}
+
+int
+lnet_init_finalizers(void)
+{
+#ifdef __KERNEL__
+        int    i;
+
+        the_lnet.ln_nfinalizers = (int) cfs_num_online_cpus();
 
-                ptl_init = 0;
+        LIBCFS_ALLOC(the_lnet.ln_finalizers,
+                     the_lnet.ln_nfinalizers *
+                     sizeof(*the_lnet.ln_finalizers));
+        if (the_lnet.ln_finalizers == NULL) {
+                CERROR("Can't allocate ln_finalizers\n");
+                return -ENOMEM;
         }
-        
-        ptl_mutex_exit();
+
+        for (i = 0; i < the_lnet.ln_nfinalizers; i++)
+                the_lnet.ln_finalizers[i] = NULL;
+#else
+        the_lnet.ln_finalizing = 0;
+#endif
+
+        CFS_INIT_LIST_HEAD(&the_lnet.ln_finalizeq);
+        return 0;
 }
 
-int PtlNIInit(ptl_interface_t interface, ptl_pid_t requested_pid,
-              ptl_ni_limits_t *desired_limits, ptl_ni_limits_t *actual_limits,
-              ptl_handle_ni_t *handle)
+void
+lnet_fini_finalizers(void)
 {
-        nal_t *nal;
+#ifdef __KERNEL__
         int    i;
-        int    rc;
 
-        if (!ptl_init)
-                return PTL_NO_INIT;
+        for (i = 0; i < the_lnet.ln_nfinalizers; i++)
+                LASSERT (the_lnet.ln_finalizers[i] == NULL);
+
+        LIBCFS_FREE(the_lnet.ln_finalizers,
+                    the_lnet.ln_nfinalizers *
+                    sizeof(*the_lnet.ln_finalizers));
+#else
+        LASSERT (!the_lnet.ln_finalizing);
+#endif
+        LASSERT (cfs_list_empty(&the_lnet.ln_finalizeq));
+}
 
-        ptl_mutex_enter ();
+#ifndef __KERNEL__
+/**
+ * Reserved API - do not use.
+ * Temporary workaround to allow uOSS and test programs force server
+ * mode in userspace. See comments near ln_server_mode_flag in
+ * lnet/lib-types.h */
 
-        if (interface == PTL_IFACE_DEFAULT) {
-                for (i = 0; i < NAL_MAX_NR; i++)
-                        if (ptl_nal_table[i] != NULL) {
-                                interface = i;
-                                break;
-                        }
-                /* NB if no interfaces are registered, 'interface' will
-                 * fail the valid test below */
-        }
-        
-        if (interface < 0 || 
-            interface >= NAL_MAX_NR ||
-            ptl_nal_table[interface] == NULL) {
-                GOTO(out, rc = PTL_IFACE_INVALID);
+void
+lnet_server_mode() {
+        the_lnet.ln_server_mode_flag = 1;
+}
+#endif
+
+int
+lnet_prepare(lnet_pid_t requested_pid)
+{
+        /* Prepare to bring up the network */
+        int               rc = 0;
+        int               i;
+
+        LASSERT (the_lnet.ln_refcount == 0);
+
+        the_lnet.ln_routing = 0;
+
+#ifdef __KERNEL__
+        LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
+        the_lnet.ln_pid = requested_pid;
+#else
+        if (the_lnet.ln_server_mode_flag) {/* server case (uOSS) */
+                LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
+
+                if (cfs_curproc_uid())/* Only root can run user-space server */
+                        return -EPERM;
+                the_lnet.ln_pid = requested_pid;
+
+        } else {/* client case (liblustre) */
+
+                /* My PID must be unique on this node and flag I'm userspace */
+                the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
         }
+#endif
+
+        rc = lnet_descriptor_setup();
+        if (rc != 0)
+                goto failed0;
+
+        memset(&the_lnet.ln_counters, 0,
+               sizeof(the_lnet.ln_counters));
+
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_active_msgs);
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_active_mds);
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_active_eqs);
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_test_peers);
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_nis);
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_zombie_nis);
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_remote_nets);
+        CFS_INIT_LIST_HEAD (&the_lnet.ln_routers);
 
-        nal = ptl_nal_table[interface];
-        nal->nal_handle.nal_idx = (NI_HANDLE_MAGIC & ~NI_HANDLE_MASK) | interface;
-        nal->nal_handle.cookie = 0;
-        
-        CDEBUG(D_OTHER, "Starting up NAL (%d) refs %d\n", interface, nal->nal_refct);
-        rc = nal->nal_ni_init(nal, requested_pid, desired_limits, actual_limits);
+        the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 
-        if (rc != PTL_OK) {
-                CERROR("Error %d starting up NAL %d, refs %d\n", rc,
-                       interface, nal->nal_refct);
-                GOTO(out, rc);
+        lnet_init_rtrpools();
+
+        rc = lnet_setup_handle_hash ();
+        if (rc != 0)
+                goto failed0;
+
+        rc = lnet_create_peer_table();
+        if (rc != 0)
+                goto failed1;
+
+        rc = lnet_init_finalizers();
+        if (rc != 0)
+                goto failed2;
+
+        the_lnet.ln_nportals = MAX_PORTALS;
+        LIBCFS_ALLOC(the_lnet.ln_portals,
+                     the_lnet.ln_nportals *
+                     sizeof(*the_lnet.ln_portals));
+        if (the_lnet.ln_portals == NULL) {
+                rc = -ENOMEM;
+                goto failed3;
         }
-        
-        if (nal->nal_refct != 0) {
-                /* Caller gets to know if this was the first ref or not */
-                rc = PTL_IFACE_DUP;
+
+        for (i = 0; i < the_lnet.ln_nportals; i++) {
+                CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_mlist));
+                CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_msgq));
+                the_lnet.ln_portals[i].ptl_options = 0;
         }
-        
-        nal->nal_refct++;
-        *handle = nal->nal_handle;
 
- out:
-        ptl_mutex_exit ();
+        return 0;
 
+ failed3:
+        lnet_fini_finalizers();
+ failed2:
+        lnet_destroy_peer_table();
+ failed1:
+        lnet_cleanup_handle_hash();
+ failed0:
+        lnet_descriptor_cleanup();
         return rc;
 }
 
-int PtlNIFini(ptl_handle_ni_t ni)
+int
+lnet_unprepare (void)
+{
+        int       idx;
+
+        /* NB no LNET_LOCK since this is the last reference.  All LND instances
+         * have shut down already, so it is safe to unlink and free all
+         * descriptors, even those that appear committed to a network op (eg MD
+         * with non-zero pending count) */
+
+        lnet_fail_nid(LNET_NID_ANY, 0);
+
+        LASSERT (cfs_list_empty(&the_lnet.ln_test_peers));
+        LASSERT (the_lnet.ln_refcount == 0);
+        LASSERT (cfs_list_empty(&the_lnet.ln_nis));
+        LASSERT (cfs_list_empty(&the_lnet.ln_zombie_nis));
+        LASSERT (the_lnet.ln_nzombie_nis == 0);
+
+        for (idx = 0; idx < the_lnet.ln_nportals; idx++) {
+                lnet_portal_t *ptl = &the_lnet.ln_portals[idx];
+
+                LASSERT (cfs_list_empty(&ptl->ptl_msgq));
+
+                while (!cfs_list_empty(&ptl->ptl_mlist)) {
+                        lnet_me_t *me = cfs_list_entry(ptl->ptl_mlist.next,
+                                                       lnet_me_t, me_list);
+                        CERROR ("Active ME %p on exit\n", me);
+                        cfs_list_del (&me->me_list);
+                        lnet_me_free (me);
+                }
+
+                if (ptl->ptl_mhash != NULL) {
+                        LASSERT (lnet_portal_is_unique(ptl));
+                        lnet_portal_mhash_free(ptl->ptl_mhash);
+                }
+        }
+
+        while (!cfs_list_empty (&the_lnet.ln_active_mds)) {
+                lnet_libmd_t *md = cfs_list_entry (the_lnet.ln_active_mds.next,
+                                                   lnet_libmd_t, md_list);
+
+                CERROR ("Active MD %p on exit\n", md);
+                cfs_list_del_init (&md->md_list);
+                lnet_md_free (md);
+        }
+
+        while (!cfs_list_empty (&the_lnet.ln_active_eqs)) {
+                lnet_eq_t *eq = cfs_list_entry (the_lnet.ln_active_eqs.next,
+                                                lnet_eq_t, eq_list);
+
+                CERROR ("Active EQ %p on exit\n", eq);
+                cfs_list_del (&eq->eq_list);
+                lnet_eq_free (eq);
+        }
+
+        while (!cfs_list_empty (&the_lnet.ln_active_msgs)) {
+                lnet_msg_t *msg = cfs_list_entry (the_lnet.ln_active_msgs.next,
+                                                  lnet_msg_t, msg_activelist);
+
+                CERROR ("Active msg %p on exit\n", msg);
+                LASSERT (msg->msg_onactivelist);
+                msg->msg_onactivelist = 0;
+                cfs_list_del (&msg->msg_activelist);
+                lnet_msg_free (msg);
+        }
+
+        LIBCFS_FREE(the_lnet.ln_portals,  
+                    the_lnet.ln_nportals * sizeof(*the_lnet.ln_portals));
+
+        lnet_free_rtrpools();
+        lnet_fini_finalizers();
+        lnet_destroy_peer_table();
+        lnet_cleanup_handle_hash();
+        lnet_descriptor_cleanup();
+
+        return (0);
+}
+
+lnet_ni_t  *
+lnet_net2ni_locked (__u32 net)
+{
+        cfs_list_t       *tmp;
+        lnet_ni_t        *ni;
+
+        cfs_list_for_each (tmp, &the_lnet.ln_nis) {
+                ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
+
+                if (LNET_NIDNET(ni->ni_nid) == net) {
+                        lnet_ni_addref_locked(ni);
+                        return ni;
+                }
+        }
+
+        return NULL;
+}
+
+int
+lnet_islocalnet (__u32 net)
+{
+        lnet_ni_t        *ni;
+
+        LNET_LOCK();
+        ni = lnet_net2ni_locked(net);
+        if (ni != NULL)
+                lnet_ni_decref_locked(ni);
+        LNET_UNLOCK();
+
+        return ni != NULL;
+}
+
+lnet_ni_t  *
+lnet_nid2ni_locked (lnet_nid_t nid)
+{
+        cfs_list_t       *tmp;
+        lnet_ni_t        *ni;
+
+        cfs_list_for_each (tmp, &the_lnet.ln_nis) {
+                ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
+
+                if (ni->ni_nid == nid) {
+                        lnet_ni_addref_locked(ni);
+                        return ni;
+                }
+        }
+
+        return NULL;
+}
+
+int
+lnet_islocalnid (lnet_nid_t nid)
+{
+        lnet_ni_t     *ni;
+
+        LNET_LOCK();
+        ni = lnet_nid2ni_locked(nid);
+        if (ni != NULL)
+                lnet_ni_decref_locked(ni);
+        LNET_UNLOCK();
+
+        return ni != NULL;
+}
+
+int
+lnet_count_acceptor_nis (void)
+{
+        /* Return the # of NIs that need the acceptor. */
+        int            count = 0;
+#if defined(__KERNEL__) || defined(HAVE_LIBPTHREAD)
+        cfs_list_t    *tmp;
+        lnet_ni_t     *ni;
+
+        LNET_LOCK();
+        cfs_list_for_each (tmp, &the_lnet.ln_nis) {
+                ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
+
+                if (ni->ni_lnd->lnd_accept != NULL)
+                        count++;
+        }
+
+        LNET_UNLOCK();
+
+#endif /* defined(__KERNEL__) || defined(HAVE_LIBPTHREAD) */
+        return count;
+}
+
+void
+lnet_shutdown_lndnis (void)
 {
-        nal_t *nal;
-        int    idx;
+        int                i;
+        int                islo;
+        lnet_ni_t         *ni;
+
+        /* NB called holding the global mutex */
+
+        /* All quiet on the API front */
+        LASSERT (!the_lnet.ln_shutdown);
+        LASSERT (the_lnet.ln_refcount == 0);
+        LASSERT (cfs_list_empty(&the_lnet.ln_zombie_nis));
+        LASSERT (the_lnet.ln_nzombie_nis == 0);
+        LASSERT (cfs_list_empty(&the_lnet.ln_remote_nets));
+
+        LNET_LOCK();
+        the_lnet.ln_shutdown = 1;               /* flag shutdown */
 
-        if (!ptl_init)
-                return PTL_NO_INIT;
+        /* Unlink NIs from the global table */
+        while (!cfs_list_empty(&the_lnet.ln_nis)) {
+                ni = cfs_list_entry(the_lnet.ln_nis.next,
+                                    lnet_ni_t, ni_list);
+                cfs_list_del (&ni->ni_list);
 
-        ptl_mutex_enter ();
+                the_lnet.ln_nzombie_nis++;
+                lnet_ni_decref_locked(ni); /* drop apini's ref */
+        }
+
+        /* Drop the cached eqwait NI. */
+        if (the_lnet.ln_eqwaitni != NULL) {
+                lnet_ni_decref_locked(the_lnet.ln_eqwaitni);
+                the_lnet.ln_eqwaitni = NULL;
+        }
 
-        nal = ptl_hndl2nal (&ni);
-        if (nal == NULL) {
-                ptl_mutex_exit ();
-                return PTL_HANDLE_INVALID;
+        /* Drop the cached loopback NI. */
+        if (the_lnet.ln_loni != NULL) {
+                lnet_ni_decref_locked(the_lnet.ln_loni);
+                the_lnet.ln_loni = NULL;
         }
 
-        idx = ni.nal_idx & NI_HANDLE_MASK;
+        LNET_UNLOCK();
 
-        LASSERT(nal->nal_refct > 0);
+        /* Clear lazy portals and drop delayed messages which hold refs
+         * on their lnet_msg_t::msg_rxpeer */
+        for (i = 0; i < the_lnet.ln_nportals; i++)
+                LNetClearLazyPortal(i);
 
-        nal->nal_refct--;
+        /* Clear the peer table and wait for all peers to go (they hold refs on
+         * their NIs) */
+        lnet_clear_peer_table();
 
-        /* nal_refct == 0 tells nal->shutdown to really shut down */
-        nal->nal_ni_fini(nal);
+        LNET_LOCK();
+        /* Now wait for the NI's I just nuked to show up on apini_zombie_nis
+         * and shut them down in guaranteed thread context */
+        i = 2;
+        while (the_lnet.ln_nzombie_nis != 0) {
 
-        ptl_mutex_exit ();
-        return PTL_OK;
+                while (cfs_list_empty(&the_lnet.ln_zombie_nis)) {
+                        LNET_UNLOCK();
+                        ++i;
+                        if ((i & (-i)) == i)
+                                CDEBUG(D_WARNING,"Waiting for %d zombie NIs\n",
+                                       the_lnet.ln_nzombie_nis);
+                        cfs_pause(cfs_time_seconds(1));
+                        LNET_LOCK();
+                }
+
+                ni = cfs_list_entry(the_lnet.ln_zombie_nis.next,
+                                    lnet_ni_t, ni_list);
+                cfs_list_del(&ni->ni_list);
+                ni->ni_lnd->lnd_refcount--;
+
+                LNET_UNLOCK();
+
+                islo = ni->ni_lnd->lnd_type == LOLND;
+
+                LASSERT (!cfs_in_interrupt ());
+                (ni->ni_lnd->lnd_shutdown)(ni);
+
+                /* can't deref lnd anymore now; it might have unregistered
+                 * itself...  */
+
+                if (!islo)
+                        CDEBUG(D_LNI, "Removed LNI %s\n",
+                               libcfs_nid2str(ni->ni_nid));
+
+                LIBCFS_FREE(ni, sizeof(*ni));
+
+                LNET_LOCK();
+                the_lnet.ln_nzombie_nis--;
+        }
+
+        the_lnet.ln_shutdown = 0;
+        LNET_UNLOCK();
+
+        if (the_lnet.ln_network_tokens != NULL) {
+                LIBCFS_FREE(the_lnet.ln_network_tokens,
+                            the_lnet.ln_network_tokens_nob);
+                the_lnet.ln_network_tokens = NULL;
+        }
+}
+
+int
+lnet_startup_lndnis (void)
+{
+        lnd_t             *lnd;
+        lnet_ni_t         *ni;
+        cfs_list_t         nilist;
+        int                rc = 0;
+        int                lnd_type;
+        int                nicount = 0;
+        char              *nets = lnet_get_networks();
+
+        CFS_INIT_LIST_HEAD(&nilist);
+
+        if (nets == NULL)
+                goto failed;
+
+        rc = lnet_parse_networks(&nilist, nets);
+        if (rc != 0)
+                goto failed;
+
+        while (!cfs_list_empty(&nilist)) {
+                ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
+                lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
+
+                LASSERT (libcfs_isknown_lnd(lnd_type));
+
+                LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
+                lnd = lnet_find_lnd_by_type(lnd_type);
+
+#ifdef __KERNEL__
+                if (lnd == NULL) {
+                        LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
+                        rc = cfs_request_module("%s",
+                                                libcfs_lnd2modname(lnd_type));
+                        LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
+
+                        lnd = lnet_find_lnd_by_type(lnd_type);
+                        if (lnd == NULL) {
+                                LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
+                                CERROR("Can't load LND %s, module %s, rc=%d\n",
+                                       libcfs_lnd2str(lnd_type),
+                                       libcfs_lnd2modname(lnd_type), rc);
+#ifndef HAVE_MODULE_LOADING_SUPPORT
+                                LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
+                                         "compiled with kernel module "
+                                         "loading support.");
+#endif
+                                goto failed;
+                        }
+                }
+#else
+                if (lnd == NULL) {
+                        LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
+                        CERROR("LND %s not supported\n",
+                               libcfs_lnd2str(lnd_type));
+                        goto failed;
+                }
+#endif
+
+                ni->ni_refcount = 1;
+
+                LNET_LOCK();
+                lnd->lnd_refcount++;
+                LNET_UNLOCK();
+
+                ni->ni_lnd = lnd;
+
+                rc = (lnd->lnd_startup)(ni);
+
+                LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
+
+                if (rc != 0) {
+                        LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
+                                           "\n",
+                                           rc, libcfs_lnd2str(lnd->lnd_type));
+                        LNET_LOCK();
+                        lnd->lnd_refcount--;
+                        LNET_UNLOCK();
+                        goto failed;
+                }
+
+                LASSERT (ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
+
+                cfs_list_del(&ni->ni_list);
+
+                LNET_LOCK();
+                cfs_list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
+                LNET_UNLOCK();
+
+                if (lnd->lnd_type == LOLND) {
+                        lnet_ni_addref(ni);
+                        LASSERT (the_lnet.ln_loni == NULL);
+                        the_lnet.ln_loni = ni;
+                        continue;
+                }
+
+#ifndef __KERNEL__
+                if (lnd->lnd_wait != NULL) {
+                        if (the_lnet.ln_eqwaitni == NULL) {
+                                lnet_ni_addref(ni);
+                                the_lnet.ln_eqwaitni = ni;
+                        }
+                } else {
+# ifndef HAVE_LIBPTHREAD
+                        LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
+                                           "single-threaded runtime\n",
+                                           libcfs_lnd2str(lnd_type));
+                        goto failed;
+# endif
+                }
+#endif
+                if (ni->ni_peertxcredits == 0 ||
+                    ni->ni_maxtxcredits == 0) {
+                        LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
+                                           libcfs_lnd2str(lnd->lnd_type),
+                                           ni->ni_peertxcredits == 0 ?
+                                           "" : "per-peer ");
+                        goto failed;
+                }
+
+                ni->ni_txcredits = ni->ni_mintxcredits = ni->ni_maxtxcredits;
+
+                CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
+                       libcfs_nid2str(ni->ni_nid),
+                       ni->ni_peertxcredits, ni->ni_txcredits,
+                       ni->ni_peerrtrcredits, ni->ni_peertimeout);
+
+                nicount++;
+        }
+
+        if (the_lnet.ln_eqwaitni != NULL && nicount > 1) {
+                lnd_type = the_lnet.ln_eqwaitni->ni_lnd->lnd_type;
+                LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
+                                   "\n",
+                                   libcfs_lnd2str(lnd_type));
+                goto failed;
+        }
+
+        return 0;
+
+ failed:
+        lnet_shutdown_lndnis();
+
+        while (!cfs_list_empty(&nilist)) {
+                ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
+                cfs_list_del(&ni->ni_list);
+                LIBCFS_FREE(ni, sizeof(*ni));
+        }
+
+        return -ENETDOWN;
+}
+
+/**
+ * Initialize LNet library.
+ *
+ * Only userspace program needs to call this function - it's automatically
+ * called in the kernel at module loading time. Caller has to call LNetFini()
+ * after a call to LNetInit(), if and only if the latter returned 0. It must
+ * be called exactly once.
+ *
+ * \return 0 on success, and -ve on failures.
+ */
+int
+LNetInit(void)
+{
+        lnet_assert_wire_constants ();
+        LASSERT (!the_lnet.ln_init);
+
+        memset(&the_lnet, 0, sizeof(the_lnet));
+
+        lnet_init_locks();
+        the_lnet.ln_refcount = 0;
+        the_lnet.ln_init = 1;
+        LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
+        CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
+        CFS_INIT_LIST_HEAD(&the_lnet.ln_zombie_rcd);
+
+#ifdef __KERNEL__
+        /* All LNDs apart from the LOLND are in separate modules.  They
+         * register themselves when their module loads, and unregister
+         * themselves when their module is unloaded. */
+#else
+        /* Register LNDs
+         * NB the order here determines default 'networks=' order */
+# ifdef CRAY_XT3
+        LNET_REGISTER_ULND(the_ptllnd);
+# endif
+# ifdef HAVE_LIBPTHREAD
+        LNET_REGISTER_ULND(the_tcplnd);
+# endif
+#endif
+        lnet_register_lnd(&the_lolnd);
+        return 0;
+}
+
+/**
+ * Finalize LNet library.
+ *
+ * Only userspace program needs to call this function. It can be called
+ * at most once.
+ *
+ * \pre LNetInit() called with success.
+ * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
+ */
+void
+LNetFini(void)
+{
+        LASSERT (the_lnet.ln_init);
+        LASSERT (the_lnet.ln_refcount == 0);
+
+        while (!cfs_list_empty(&the_lnet.ln_lnds))
+                lnet_unregister_lnd(cfs_list_entry(the_lnet.ln_lnds.next,
+                                                   lnd_t, lnd_list));
+        lnet_fini_locks();
+
+        the_lnet.ln_init = 0;
+}
+
+/**
+ * Set LNet PID and start LNet interfaces, routing, and forwarding.
+ *
+ * Userspace program should call this after a successful call to LNetInit().
+ * Users must call this function at least once before any other functions.
+ * For each successful call there must be a corresponding call to
+ * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
+ * ignored.
+ *
+ * The PID used by LNet may be different from the one requested.
+ * See LNetGetId().
+ *
+ * \param requested_pid PID requested by the caller.
+ *
+ * \return 0 on success, and non-zero error code on failures.
+ */
+int
+LNetNIInit(lnet_pid_t requested_pid)
+{
+        int         im_a_router = 0;
+        int         rc;
+
+        LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
+
+        LASSERT (the_lnet.ln_init);
+        CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
+
+        if (the_lnet.ln_refcount > 0) {
+                rc = the_lnet.ln_refcount++;
+                goto out;
+        }
+
+        lnet_get_tunables();
+
+        if (requested_pid == LNET_PID_ANY) {
+                /* Don't instantiate LNET just for me */
+                rc = -ENETDOWN;
+                goto failed0;
+        }
+
+        rc = lnet_prepare(requested_pid);
+        if (rc != 0)
+                goto failed0;
+
+        rc = lnet_startup_lndnis();
+        if (rc != 0)
+                goto failed1;
+
+        rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
+        if (rc != 0)
+                goto failed2;
+
+        rc = lnet_check_routes();
+        if (rc != 0)
+                goto failed2;
+
+        rc = lnet_alloc_rtrpools(im_a_router);
+        if (rc != 0)
+                goto failed2;
+
+        rc = lnet_acceptor_start();
+        if (rc != 0)
+                goto failed2;
+
+        the_lnet.ln_refcount = 1;
+        /* Now I may use my own API functions... */
+
+        /* NB router checker needs the_lnet.ln_ping_info in
+         * lnet_router_checker -> lnet_update_ni_status */
+        rc = lnet_ping_target_init();
+        if (rc != 0)
+                goto failed3;
+
+        rc = lnet_router_checker_start();
+        if (rc != 0)
+                goto failed4;
+
+        lnet_proc_init();
+        goto out;
+
+ failed4:
+        lnet_ping_target_fini();
+ failed3:
+        the_lnet.ln_refcount = 0;
+        lnet_acceptor_stop();
+ failed2:
+        lnet_destroy_routes();
+        lnet_shutdown_lndnis();
+ failed1:
+        lnet_unprepare();
+ failed0:
+        LASSERT (rc < 0);
+ out:
+        LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
+        return rc;
+}
+
+/**
+ * Stop LNet interfaces, routing, and forwarding.
+ *
+ * Users must call this function once for each successful call to LNetNIInit().
+ * Once the LNetNIFini() operation has been started, the results of pending
+ * API operations are undefined.
+ *
+ * \return always 0 for current implementation.
+ */
+int
+LNetNIFini()
+{
+        LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
+
+        LASSERT (the_lnet.ln_init);
+        LASSERT (the_lnet.ln_refcount > 0);
+
+        if (the_lnet.ln_refcount != 1) {
+                the_lnet.ln_refcount--;
+        } else {
+                LASSERT (!the_lnet.ln_niinit_self);
+
+                lnet_proc_fini();
+                lnet_router_checker_stop();
+                lnet_ping_target_fini();
+
+                /* Teardown fns that use my own API functions BEFORE here */
+                the_lnet.ln_refcount = 0;
+
+                lnet_acceptor_stop();
+                lnet_destroy_routes();
+                lnet_shutdown_lndnis();
+                lnet_unprepare();
+        }
+
+        LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
+        return 0;
+}
+
+/**
+ * This is an ugly hack to export IOC_LIBCFS_DEBUG_PEER and
+ * IOC_LIBCFS_PORTALS_COMPATIBILITY commands to users, by tweaking the LNet
+ * internal ioctl handler.
+ *
+ * IOC_LIBCFS_PORTALS_COMPATIBILITY is now deprecated, don't use it.
+ *
+ * \param cmd IOC_LIBCFS_DEBUG_PEER to print debugging data about a peer.
+ * The data will be printed to system console. Don't use it excessively.
+ * \param arg A pointer to lnet_process_id_t, process ID of the peer.
+ *
+ * \return Always return 0 when called by users directly (i.e., not via ioctl).
+ */
+int
+LNetCtl(unsigned int cmd, void *arg)
+{
+        struct libcfs_ioctl_data *data = arg;
+        lnet_process_id_t         id = {0};
+        lnet_ni_t                *ni;
+        int                       rc;
+
+        LASSERT (the_lnet.ln_init);
+        LASSERT (the_lnet.ln_refcount > 0);
+
+        switch (cmd) {
+        case IOC_LIBCFS_GET_NI:
+                rc = LNetGetId(data->ioc_count, &id);
+                data->ioc_nid = id.nid;
+                return rc;
+
+        case IOC_LIBCFS_FAIL_NID:
+                return lnet_fail_nid(data->ioc_nid, data->ioc_count);
+
+        case IOC_LIBCFS_ADD_ROUTE:
+                rc = lnet_add_route(data->ioc_net, data->ioc_count,
+                                    data->ioc_nid);
+                return (rc != 0) ? rc : lnet_check_routes();
+
+        case IOC_LIBCFS_DEL_ROUTE:
+                return lnet_del_route(data->ioc_net, data->ioc_nid);
+
+        case IOC_LIBCFS_GET_ROUTE:
+                return lnet_get_route(data->ioc_count,
+                                      &data->ioc_net, &data->ioc_count,
+                                      &data->ioc_nid, &data->ioc_flags);
+        case IOC_LIBCFS_NOTIFY_ROUTER:
+                return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
+                                   cfs_time_current() -
+                                   cfs_time_seconds(cfs_time_current_sec() -
+                                                    (time_t)data->ioc_u64[0]));
+
+        case IOC_LIBCFS_PORTALS_COMPATIBILITY:
+                /* This can be removed once lustre stops calling it */
+                return 0;
+
+        case IOC_LIBCFS_LNET_DIST:
+                rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
+                if (rc < 0 && rc != -EHOSTUNREACH)
+                        return rc;
+
+                data->ioc_u32[0] = rc;
+                return 0;
+
+        case IOC_LIBCFS_TESTPROTOCOMPAT:
+                LNET_LOCK();
+                the_lnet.ln_testprotocompat = data->ioc_flags;
+                LNET_UNLOCK();
+                return 0;
+
+        case IOC_LIBCFS_PING:
+                id.nid = data->ioc_nid;
+                id.pid = data->ioc_u32[0];
+                rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
+                               (lnet_process_id_t *)data->ioc_pbuf1,
+                               data->ioc_plen1/sizeof(lnet_process_id_t));
+                if (rc < 0)
+                        return rc;
+                data->ioc_count = rc;
+                return 0;
+
+        case IOC_LIBCFS_DEBUG_PEER: {
+                /* CAVEAT EMPTOR: this one designed for calling directly; not
+                 * via an ioctl */
+                id = *((lnet_process_id_t *) arg);
+
+                lnet_debug_peer(id.nid);
+
+                ni = lnet_net2ni(LNET_NIDNET(id.nid));
+                if (ni == NULL) {
+                        CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(id));
+                } else {
+                        if (ni->ni_lnd->lnd_ctl == NULL) {
+                                CDEBUG(D_WARNING, "No ctl for %s\n",
+                                       libcfs_id2str(id));
+                        } else {
+                                (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
+                        }
+
+                        lnet_ni_decref(ni);
+                }
+                return 0;
+        }
+
+        default:
+                ni = lnet_net2ni(data->ioc_net);
+                if (ni == NULL)
+                        return -EINVAL;
+
+                if (ni->ni_lnd->lnd_ctl == NULL)
+                        rc = -EINVAL;
+                else
+                        rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
+
+                lnet_ni_decref(ni);
+                return rc;
+        }
+        /* not reached */
+}
+
+/**
+ * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
+ * all interfaces share a same PID, as requested by LNetNIInit().
+ *
+ * \param index Index of the interface to look up.
+ * \param id On successful return, this location will hold the
+ * lnet_process_id_t ID of the interface.
+ *
+ * \retval 0 If an interface exists at \a index.
+ * \retval -ENOENT If no interface has been found.
+ */
+int
+LNetGetId(unsigned int index, lnet_process_id_t *id)
+{
+        lnet_ni_t        *ni;
+        cfs_list_t       *tmp;
+        int               rc = -ENOENT;
+
+        LASSERT (the_lnet.ln_init);
+        LASSERT (the_lnet.ln_refcount > 0);
+
+        LNET_LOCK();
+
+        cfs_list_for_each(tmp, &the_lnet.ln_nis) {
+                if (index-- != 0)
+                        continue;
+
+                ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
+
+                id->nid = ni->ni_nid;
+                id->pid = the_lnet.ln_pid;
+                rc = 0;
+                break;
+        }
+
+        LNET_UNLOCK();
+
+        return rc;
+}
+
+/**
+ * Print a string representation of handle \a h into buffer \a str of
+ * \a len bytes.
+ */
+void
+LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
+{
+        snprintf(str, len, LPX64, h.cookie);
+}
+
+static int
+lnet_create_ping_info(void)
+{
+        int               i;
+        int               n;
+        int               rc;
+        unsigned int      infosz;
+        lnet_ni_t        *ni;
+        lnet_process_id_t id;
+        lnet_ping_info_t *pinfo;
+
+        for (n = 0; ; n++) {
+                rc = LNetGetId(n, &id);
+                if (rc == -ENOENT)
+                        break;
+
+                LASSERT (rc == 0);
+        }
+
+        infosz = offsetof(lnet_ping_info_t, pi_ni[n]);
+        LIBCFS_ALLOC(pinfo, infosz);
+        if (pinfo == NULL) {
+                CERROR("Can't allocate ping info[%d]\n", n);
+                return -ENOMEM;
+        }
+
+        pinfo->pi_nnis    = n;
+        pinfo->pi_pid     = the_lnet.ln_pid;
+        pinfo->pi_magic   = LNET_PROTO_PING_MAGIC;
+        pinfo->pi_version = LNET_PROTO_PING_VERSION;
+
+        for (i = 0; i < n; i++) {
+                lnet_ni_status_t *ns = &pinfo->pi_ni[i];
+
+                rc = LNetGetId(i, &id);
+                LASSERT (rc == 0);
+
+                ns->ns_nid    = id.nid;
+                ns->ns_status = LNET_NI_STATUS_UP;
+
+                LNET_LOCK();
+
+                ni = lnet_nid2ni_locked(id.nid);
+                LASSERT (ni != NULL);
+                LASSERT (ni->ni_status == NULL);
+                ni->ni_status = ns;
+                lnet_ni_decref_locked(ni);
+
+                LNET_UNLOCK();
+        }
+
+        the_lnet.ln_ping_info = pinfo;
+        return 0;
+}
+
+static void
+lnet_destroy_ping_info(void)
+{
+        lnet_ni_t *ni;
+
+        LNET_LOCK();
+
+        cfs_list_for_each_entry (ni, &the_lnet.ln_nis, ni_list) {
+                ni->ni_status = NULL;
+        }
+
+        LNET_UNLOCK();
+
+        LIBCFS_FREE(the_lnet.ln_ping_info,
+                    offsetof(lnet_ping_info_t,
+                             pi_ni[the_lnet.ln_ping_info->pi_nnis]));
+        the_lnet.ln_ping_info = NULL;
+        return;
+}
+
+int
+lnet_ping_target_init(void)
+{
+        lnet_md_t         md = {0};
+        lnet_handle_me_t  meh;
+        lnet_process_id_t id;
+        int               rc;
+        int               rc2;
+        int               infosz;
+
+        rc = lnet_create_ping_info();
+        if (rc != 0)
+                return rc;
+
+        /* We can have a tiny EQ since we only need to see the unlink event on
+         * teardown, which by definition is the last one! */
+        rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
+        if (rc != 0) {
+                CERROR("Can't allocate ping EQ: %d\n", rc);
+                goto failed_0;
+        }
+
+        memset(&id, 0, sizeof(lnet_process_id_t));
+        id.nid = LNET_NID_ANY;
+        id.pid = LNET_PID_ANY;
+
+        rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
+                          LNET_PROTO_PING_MATCHBITS, 0,
+                          LNET_UNLINK, LNET_INS_AFTER,
+                          &meh);
+        if (rc != 0) {
+                CERROR("Can't create ping ME: %d\n", rc);
+                goto failed_1;
+        }
+
+        /* initialize md content */
+        infosz = offsetof(lnet_ping_info_t,
+                          pi_ni[the_lnet.ln_ping_info->pi_nnis]);
+        md.start     = the_lnet.ln_ping_info;
+        md.length    = infosz;
+        md.threshold = LNET_MD_THRESH_INF;
+        md.max_size  = 0;
+        md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
+                       LNET_MD_MANAGE_REMOTE;
+        md.user_ptr  = NULL;
+        md.eq_handle = the_lnet.ln_ping_target_eq;
+
+        rc = LNetMDAttach(meh, md,
+                          LNET_RETAIN,
+                          &the_lnet.ln_ping_target_md);
+        if (rc != 0) {
+                CERROR("Can't attach ping MD: %d\n", rc);
+                goto failed_2;
+        }
+
+        return 0;
+
+ failed_2:
+        rc2 = LNetMEUnlink(meh);
+        LASSERT (rc2 == 0);
+ failed_1:
+        rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
+        LASSERT (rc2 == 0);
+ failed_0:
+        lnet_destroy_ping_info();
+        return rc;
+}
+
+void
+lnet_ping_target_fini(void)
+{
+        lnet_event_t    event;
+        int             rc;
+        int             which;
+        int             timeout_ms = 1000;
+        cfs_sigset_t    blocked = cfs_block_allsigs();
+
+        LNetMDUnlink(the_lnet.ln_ping_target_md);
+        /* NB md could be busy; this just starts the unlink */
+
+        for (;;) {
+                rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
+                                timeout_ms, &event, &which);
+
+                /* I expect overflow... */
+                LASSERT (rc >= 0 || rc == -EOVERFLOW);
+
+                if (rc == 0) {
+                        /* timed out: provide a diagnostic */
+                        CWARN("Still waiting for ping MD to unlink\n");
+                        timeout_ms *= 2;
+                        continue;
+                }
+
+                /* Got a valid event */
+                if (event.unlinked)
+                        break;
+        }
+
+        rc = LNetEQFree(the_lnet.ln_ping_target_eq);
+        LASSERT (rc == 0);
+        lnet_destroy_ping_info();
+        cfs_restore_sigs(blocked);
+}
+
+int
+lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
+{
+        lnet_handle_eq_t     eqh;
+        lnet_handle_md_t     mdh;
+        lnet_event_t         event;
+        lnet_md_t            md = {0};
+        int                  which;
+        int                  unlinked = 0;
+        int                  replied = 0;
+        const int            a_long_time = 60000; /* mS */
+        int                  infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
+        lnet_ping_info_t    *info;
+        lnet_process_id_t    tmpid;
+        int                  i;
+        int                  nob;
+        int                  rc;
+        int                  rc2;
+        cfs_sigset_t         blocked;
+
+        if (n_ids <= 0 ||
+            id.nid == LNET_NID_ANY ||
+            timeout_ms > 500000 ||              /* arbitrary limit! */
+            n_ids > 20)                         /* arbitrary limit! */
+                return -EINVAL;
+
+        if (id.pid == LNET_PID_ANY)
+                id.pid = LUSTRE_SRV_LNET_PID;
+
+        LIBCFS_ALLOC(info, infosz);
+        if (info == NULL)
+                return -ENOMEM;
+
+        /* NB 2 events max (including any unlink event) */
+        rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
+        if (rc != 0) {
+                CERROR("Can't allocate EQ: %d\n", rc);
+                goto out_0;
+        }
+
+        /* initialize md content */
+        md.start     = info;
+        md.length    = infosz;
+        md.threshold = 2; /*GET/REPLY*/
+        md.max_size  = 0;
+        md.options   = LNET_MD_TRUNCATE;
+        md.user_ptr  = NULL;
+        md.eq_handle = eqh;
+
+        rc = LNetMDBind(md, LNET_UNLINK, &mdh);
+        if (rc != 0) {
+                CERROR("Can't bind MD: %d\n", rc);
+                goto out_1;
+        }
+
+        rc = LNetGet(LNET_NID_ANY, mdh, id,
+                     LNET_RESERVED_PORTAL,
+                     LNET_PROTO_PING_MATCHBITS, 0);
+
+        if (rc != 0) {
+                /* Don't CERROR; this could be deliberate! */
+
+                rc2 = LNetMDUnlink(mdh);
+                LASSERT (rc2 == 0);
+
+                /* NB must wait for the UNLINK event below... */
+                unlinked = 1;
+                timeout_ms = a_long_time;
+        }
+
+        do {
+                /* MUST block for unlink to complete */
+                if (unlinked)
+                        blocked = cfs_block_allsigs();
+
+                rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
+
+                if (unlinked)
+                        cfs_restore_sigs(blocked);
+
+                CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
+                       (rc2 <= 0) ? -1 : event.type,
+                       (rc2 <= 0) ? -1 : event.status,
+                       (rc2 > 0 && event.unlinked) ? " unlinked" : "");
+
+                LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
+
+                if (rc2 <= 0 || event.status != 0) {
+                        /* timeout or error */
+                        if (!replied && rc == 0)
+                                rc = (rc2 < 0) ? rc2 :
+                                     (rc2 == 0) ? -ETIMEDOUT :
+                                     event.status;
+
+                        if (!unlinked) {
+                                /* Ensure completion in finite time... */
+                                LNetMDUnlink(mdh);
+                                /* No assertion (racing with network) */
+                                unlinked = 1;
+                                timeout_ms = a_long_time;
+                        } else if (rc2 == 0) {
+                                /* timed out waiting for unlink */
+                                CWARN("ping %s: late network completion\n",
+                                      libcfs_id2str(id));
+                        }
+                } else if (event.type == LNET_EVENT_REPLY) {
+                        replied = 1;
+                        rc = event.mlength;
+                }
+
+        } while (rc2 <= 0 || !event.unlinked);
+
+        if (!replied) {
+                if (rc >= 0)
+                        CWARN("%s: Unexpected rc >= 0 but no reply!\n",
+                              libcfs_id2str(id));
+                rc = -EIO;
+                goto out_1;
+        }
+
+        nob = rc;
+        LASSERT (nob >= 0 && nob <= infosz);
+
+        rc = -EPROTO;                           /* if I can't parse... */
+
+        if (nob < 8) {
+                /* can't check magic/version */
+                CERROR("%s: ping info too short %d\n",
+                       libcfs_id2str(id), nob);
+                goto out_1;
+        }
+
+        if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
+                lnet_swap_pinginfo(info);
+        } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
+                CERROR("%s: Unexpected magic %08x\n", 
+                       libcfs_id2str(id), info->pi_magic);
+                goto out_1;
+        }
+
+        if (info->pi_version != LNET_PROTO_PING_VERSION) {
+                CERROR("%s: Unexpected version 0x%x\n",
+                       libcfs_id2str(id), info->pi_version);
+                goto out_1;
+        }
+
+        if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
+                CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
+                       nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
+                goto out_1;
+        }
+
+        if (info->pi_nnis < n_ids)
+                n_ids = info->pi_nnis;
+
+        if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
+                CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
+                       nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
+                goto out_1;
+        }
+
+        rc = -EFAULT;                           /* If I SEGV... */
+
+        for (i = 0; i < n_ids; i++) {
+                tmpid.pid = info->pi_pid;
+                tmpid.nid = info->pi_ni[i].ns_nid;
+#ifdef __KERNEL__
+                if (cfs_copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
+                        goto out_1;
+#else
+                ids[i] = tmpid;
+#endif
+        }
+        rc = info->pi_nnis;
+
+ out_1:
+        rc2 = LNetEQFree(eqh);
+        if (rc2 != 0)
+                CERROR("rc2 %d\n", rc2);
+        LASSERT (rc2 == 0);
+
+ out_0:
+        LIBCFS_FREE(info, infosz);
+        return rc;
 }