1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.sf.net/projects/lustre/
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22 #define DEBUG_SUBSYSTEM S_LNET
23 #include <lnet/lib-lnet.h>
26 #define D_LNI D_CONSOLE
28 #define D_LNI D_CONFIG
31 lnet_t the_lnet; /* THE state of the network */
35 static char *ip2nets = "";
36 CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
37 "LNET network <- IP table");
39 static char *networks = "";
40 CFS_MODULE_PARM(networks, "s", charp, 0444,
43 static char *routes = "";
44 CFS_MODULE_PARM(routes, "s", charp, 0444,
45 "routes to non-local networks");
47 static char *portals_compatibility = "none";
48 CFS_MODULE_PARM(portals_compatibility, "s", charp, 0444,
49 "wire protocol compatibility: 'strong'|'weak'|'none'");
58 lnet_get_networks(void)
63 if (*networks != 0 && *ip2nets != 0) {
64 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
65 "'ip2nets' but not both at once\n");
70 rc = lnet_parse_ip2nets(&nets, ip2nets);
71 return (rc == 0) ? nets : NULL;
81 lnet_get_portals_compatibility(void)
83 if (!strcmp(portals_compatibility, "none")) {
87 if (!strcmp(portals_compatibility, "weak")) {
89 LCONSOLE_WARN("Starting in weak portals-compatible mode\n");
92 if (!strcmp(portals_compatibility, "strong")) {
94 LCONSOLE_WARN("Starting in strong portals-compatible mode\n");
97 LCONSOLE_ERROR_MSG(0x102, "portals_compatibility=\"%s\" not supported\n",
98 portals_compatibility);
103 lnet_init_locks(void)
105 spin_lock_init (&the_lnet.ln_lock);
106 cfs_waitq_init (&the_lnet.ln_waitq);
107 init_mutex(&the_lnet.ln_lnd_mutex);
108 init_mutex(&the_lnet.ln_api_mutex);
112 lnet_fini_locks(void)
119 lnet_get_routes(void)
121 char *str = getenv("LNET_ROUTES");
123 return (str == NULL) ? "" : str;
127 lnet_get_networks (void)
129 static char default_networks[256];
130 char *networks = getenv ("LNET_NETWORKS");
131 char *ip2nets = getenv ("LNET_IP2NETS");
137 struct list_head *tmp;
140 if (networks != NULL && ip2nets != NULL) {
141 LCONSOLE_ERROR_MSG(0x103, "Please set EITHER 'LNET_NETWORKS' or"
142 " 'LNET_IP2NETS' but not both at once\n");
146 if (ip2nets != NULL) {
147 rc = lnet_parse_ip2nets(&networks, ip2nets);
148 return (rc == 0) ? networks : NULL;
154 if (networks != NULL)
157 /* In userland, the default 'networks=' is the list of known net types */
159 len = sizeof(default_networks);
160 str = default_networks;
164 list_for_each (tmp, &the_lnet.ln_lnds) {
165 lnd_t *lnd = list_entry(tmp, lnd_t, lnd_list);
167 nob = snprintf(str, len, "%s%s", sep,
168 libcfs_lnd2str(lnd->lnd_type));
171 /* overflowed the string; leave it where it was */
180 return default_networks;
184 lnet_get_portals_compatibility(void)
189 # ifndef HAVE_LIBPTHREAD
191 void lnet_init_locks(void)
193 the_lnet.ln_lock = 0;
194 the_lnet.ln_lnd_mutex = 0;
195 the_lnet.ln_api_mutex = 0;
198 void lnet_fini_locks(void)
200 LASSERT (the_lnet.ln_api_mutex == 0);
201 LASSERT (the_lnet.ln_lnd_mutex == 0);
202 LASSERT (the_lnet.ln_lock == 0);
207 void lnet_init_locks(void)
209 pthread_cond_init(&the_lnet.ln_cond, NULL);
210 pthread_mutex_init(&the_lnet.ln_lock, NULL);
211 pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
212 pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
215 void lnet_fini_locks(void)
217 pthread_mutex_destroy(&the_lnet.ln_api_mutex);
218 pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
219 pthread_mutex_destroy(&the_lnet.ln_lock);
220 pthread_cond_destroy(&the_lnet.ln_cond);
226 void lnet_assert_wire_constants (void)
228 /* Wire protocol assertions generated by 'wirecheck'
229 * running on Linux robert.bartonsoftware.com 2.6.8-1.521
230 * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
231 * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
234 CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
235 CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
236 CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
237 CLASSERT (LNET_MSG_ACK == 0);
238 CLASSERT (LNET_MSG_PUT == 1);
239 CLASSERT (LNET_MSG_GET == 2);
240 CLASSERT (LNET_MSG_REPLY == 3);
241 CLASSERT (LNET_MSG_HELLO == 4);
243 /* Checks for struct ptl_handle_wire_t */
244 CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
245 CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
246 CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
247 CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
248 CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
250 /* Checks for struct lnet_magicversion_t */
251 CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
252 CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
253 CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
254 CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
255 CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
256 CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
257 CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
259 /* Checks for struct lnet_hdr_t */
260 CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
261 CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
262 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
263 CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
264 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
265 CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
266 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
267 CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
268 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
269 CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
270 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
271 CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
272 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
273 CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
274 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
277 CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
278 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
279 CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
280 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
281 CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
282 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
285 CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
286 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
287 CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
288 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
289 CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
290 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
291 CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
292 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
293 CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
294 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
297 CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
298 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
299 CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
300 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
301 CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
302 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
303 CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
304 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
305 CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
306 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
309 CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
310 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
313 CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
314 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
315 CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
316 CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
320 lnet_find_lnd_by_type (int type)
323 struct list_head *tmp;
325 /* holding lnd mutex */
326 list_for_each (tmp, &the_lnet.ln_lnds) {
327 lnd = list_entry(tmp, lnd_t, lnd_list);
329 if (lnd->lnd_type == type)
337 lnet_register_lnd (lnd_t *lnd)
339 LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
341 LASSERT (the_lnet.ln_init);
342 LASSERT (libcfs_isknown_lnd(lnd->lnd_type));
343 LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
345 list_add_tail (&lnd->lnd_list, &the_lnet.ln_lnds);
346 lnd->lnd_refcount = 0;
348 CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
350 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
354 lnet_unregister_lnd (lnd_t *lnd)
356 LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
358 LASSERT (the_lnet.ln_init);
359 LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
360 LASSERT (lnd->lnd_refcount == 0);
362 list_del (&lnd->lnd_list);
363 CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
365 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
368 #ifndef LNET_USE_LIB_FREELIST
371 lnet_descriptor_setup (void)
377 lnet_descriptor_cleanup (void)
384 lnet_freelist_init (lnet_freelist_t *fl, int n, int size)
390 size += offsetof (lnet_freeobj_t, fo_contents);
392 LIBCFS_ALLOC(space, n * size);
396 CFS_INIT_LIST_HEAD (&fl->fl_list);
399 fl->fl_objsize = size;
403 memset (space, 0, size);
404 list_add ((struct list_head *)space, &fl->fl_list);
412 lnet_freelist_fini (lnet_freelist_t *fl)
414 struct list_head *el;
417 if (fl->fl_nobjs == 0)
421 for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
424 LASSERT (count == fl->fl_nobjs);
426 LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
427 memset (fl, 0, sizeof (fl));
431 lnet_descriptor_setup (void)
433 /* NB on failure caller must still call lnet_descriptor_cleanup */
437 memset (&the_lnet.ln_free_mes, 0, sizeof (the_lnet.ln_free_mes));
438 memset (&the_lnet.ln_free_msgs, 0, sizeof (the_lnet.ln_free_msgs));
439 memset (&the_lnet.ln_free_mds, 0, sizeof (the_lnet.ln_free_mds));
440 memset (&the_lnet.ln_free_eqs, 0, sizeof (the_lnet.ln_free_eqs));
442 rc = lnet_freelist_init(&the_lnet.ln_free_mes,
443 MAX_MES, sizeof (lnet_me_t));
447 rc = lnet_freelist_init(&the_lnet.ln_free_msgs,
448 MAX_MSGS, sizeof (lnet_msg_t));
452 rc = lnet_freelist_init(&the_lnet.ln_free_mds,
453 MAX_MDS, sizeof (lnet_libmd_t));
457 rc = lnet_freelist_init(&the_lnet.ln_free_eqs,
458 MAX_EQS, sizeof (lnet_eq_t));
463 lnet_descriptor_cleanup (void)
465 lnet_freelist_fini (&the_lnet.ln_free_mes);
466 lnet_freelist_fini (&the_lnet.ln_free_msgs);
467 lnet_freelist_fini (&the_lnet.ln_free_mds);
468 lnet_freelist_fini (&the_lnet.ln_free_eqs);
474 lnet_create_interface_cookie (void)
476 /* NB the interface cookie in wire handles guards against delayed
477 * replies and ACKs appearing valid after reboot. Initialisation time,
478 * even if it's only implemented to millisecond resolution is probably
479 * easily good enough. */
483 int rc = gettimeofday (&tv, NULL);
486 do_gettimeofday(&tv);
490 cookie += tv.tv_usec;
495 lnet_setup_handle_hash (void)
499 /* Arbitrary choice of hash table size */
501 the_lnet.ln_lh_hash_size = CFS_PAGE_SIZE / sizeof (struct list_head);
503 the_lnet.ln_lh_hash_size = (MAX_MES + MAX_MDS + MAX_EQS)/4;
505 LIBCFS_ALLOC(the_lnet.ln_lh_hash_table,
506 the_lnet.ln_lh_hash_size * sizeof (struct list_head));
507 if (the_lnet.ln_lh_hash_table == NULL)
510 for (i = 0; i < the_lnet.ln_lh_hash_size; i++)
511 CFS_INIT_LIST_HEAD (&the_lnet.ln_lh_hash_table[i]);
513 the_lnet.ln_next_object_cookie = LNET_COOKIE_TYPES;
519 lnet_cleanup_handle_hash (void)
521 if (the_lnet.ln_lh_hash_table == NULL)
524 LIBCFS_FREE(the_lnet.ln_lh_hash_table,
525 the_lnet.ln_lh_hash_size * sizeof (struct list_head));
529 lnet_lookup_cookie (__u64 cookie, int type)
531 /* ALWAYS called with LNET_LOCK held */
532 struct list_head *list;
533 struct list_head *el;
536 if ((cookie & (LNET_COOKIE_TYPES - 1)) != type)
539 hash = ((unsigned int)cookie) % the_lnet.ln_lh_hash_size;
540 list = &the_lnet.ln_lh_hash_table[hash];
542 list_for_each (el, list) {
543 lnet_libhandle_t *lh = list_entry (el, lnet_libhandle_t,
546 if (lh->lh_cookie == cookie)
554 lnet_initialise_handle (lnet_libhandle_t *lh, int type)
556 /* ALWAYS called with LNET_LOCK held */
559 LASSERT (type >= 0 && type < LNET_COOKIE_TYPES);
560 lh->lh_cookie = the_lnet.ln_next_object_cookie | type;
561 the_lnet.ln_next_object_cookie += LNET_COOKIE_TYPES;
563 hash = ((unsigned int)lh->lh_cookie) % the_lnet.ln_lh_hash_size;
564 list_add (&lh->lh_hash_chain, &the_lnet.ln_lh_hash_table[hash]);
568 lnet_invalidate_handle (lnet_libhandle_t *lh)
570 /* ALWAYS called with LNET_LOCK held */
571 list_del (&lh->lh_hash_chain);
575 lnet_init_finalizers(void)
580 the_lnet.ln_nfinalizers = num_online_cpus();
582 LIBCFS_ALLOC(the_lnet.ln_finalizers,
583 the_lnet.ln_nfinalizers *
584 sizeof(*the_lnet.ln_finalizers));
585 if (the_lnet.ln_finalizers == NULL) {
586 CERROR("Can't allocate ln_finalizers\n");
590 for (i = 0; i < the_lnet.ln_nfinalizers; i++)
591 the_lnet.ln_finalizers[i] = NULL;
593 the_lnet.ln_finalizing = 0;
596 CFS_INIT_LIST_HEAD(&the_lnet.ln_finalizeq);
601 lnet_fini_finalizers(void)
606 for (i = 0; i < the_lnet.ln_nfinalizers; i++)
607 LASSERT (the_lnet.ln_finalizers[i] == NULL);
609 LIBCFS_FREE(the_lnet.ln_finalizers,
610 the_lnet.ln_nfinalizers *
611 sizeof(*the_lnet.ln_finalizers));
613 LASSERT (!the_lnet.ln_finalizing);
615 LASSERT (list_empty(&the_lnet.ln_finalizeq));
619 lnet_prepare(lnet_pid_t requested_pid)
621 /* Prepare to bring up the network */
625 LASSERT (the_lnet.ln_refcount == 0);
627 the_lnet.ln_routing = 0;
630 LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
631 the_lnet.ln_pid = requested_pid;
633 /* My PID must be unique on this node and flag I'm userspace */
634 the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
637 rc = lnet_descriptor_setup();
641 memset(&the_lnet.ln_counters, 0,
642 sizeof(the_lnet.ln_counters));
644 CFS_INIT_LIST_HEAD (&the_lnet.ln_active_msgs);
645 CFS_INIT_LIST_HEAD (&the_lnet.ln_active_mds);
646 CFS_INIT_LIST_HEAD (&the_lnet.ln_active_eqs);
647 CFS_INIT_LIST_HEAD (&the_lnet.ln_test_peers);
648 CFS_INIT_LIST_HEAD (&the_lnet.ln_nis);
649 CFS_INIT_LIST_HEAD (&the_lnet.ln_zombie_nis);
650 CFS_INIT_LIST_HEAD (&the_lnet.ln_remote_nets);
651 CFS_INIT_LIST_HEAD (&the_lnet.ln_routers);
653 the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
655 lnet_init_rtrpools();
657 rc = lnet_setup_handle_hash ();
661 rc = lnet_create_peer_table();
665 rc = lnet_init_finalizers();
669 the_lnet.ln_nportals = MAX_PORTALS;
670 LIBCFS_ALLOC(the_lnet.ln_portals,
671 the_lnet.ln_nportals *
672 sizeof(*the_lnet.ln_portals));
673 if (the_lnet.ln_portals == NULL) {
678 for (i = 0; i < the_lnet.ln_nportals; i++) {
679 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_ml));
680 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_msgq));
681 the_lnet.ln_portals[i].ptl_options = 0;
687 lnet_fini_finalizers();
689 lnet_destroy_peer_table();
691 lnet_cleanup_handle_hash();
693 lnet_descriptor_cleanup();
698 lnet_unprepare (void)
702 /* NB no LNET_LOCK since this is the last reference. All LND instances
703 * have shut down already, so it is safe to unlink and free all
704 * descriptors, even those that appear committed to a network op (eg MD
705 * with non-zero pending count) */
707 lnet_fail_nid(LNET_NID_ANY, 0);
709 LASSERT (list_empty(&the_lnet.ln_test_peers));
710 LASSERT (the_lnet.ln_refcount == 0);
711 LASSERT (list_empty(&the_lnet.ln_nis));
712 LASSERT (list_empty(&the_lnet.ln_zombie_nis));
713 LASSERT (the_lnet.ln_nzombie_nis == 0);
715 for (idx = 0; idx < the_lnet.ln_nportals; idx++) {
717 LNetClearLazyPortal(idx);
718 LASSERT (list_empty(&the_lnet.ln_portals[idx].ptl_msgq));
720 while (!list_empty (&the_lnet.ln_portals[idx].ptl_ml)) {
721 lnet_me_t *me = list_entry (the_lnet.ln_portals[idx].ptl_ml.next,
724 CERROR ("Active me %p on exit\n", me);
725 list_del (&me->me_list);
730 while (!list_empty (&the_lnet.ln_active_mds)) {
731 lnet_libmd_t *md = list_entry (the_lnet.ln_active_mds.next,
732 lnet_libmd_t, md_list);
734 CERROR ("Active md %p on exit\n", md);
735 list_del (&md->md_list);
739 while (!list_empty (&the_lnet.ln_active_eqs)) {
740 lnet_eq_t *eq = list_entry (the_lnet.ln_active_eqs.next,
743 CERROR ("Active eq %p on exit\n", eq);
744 list_del (&eq->eq_list);
748 while (!list_empty (&the_lnet.ln_active_msgs)) {
749 lnet_msg_t *msg = list_entry (the_lnet.ln_active_msgs.next,
750 lnet_msg_t, msg_activelist);
752 CERROR ("Active msg %p on exit\n", msg);
753 LASSERT (msg->msg_onactivelist);
754 msg->msg_onactivelist = 0;
755 list_del (&msg->msg_activelist);
759 LIBCFS_FREE(the_lnet.ln_portals,
760 the_lnet.ln_nportals * sizeof(*the_lnet.ln_portals));
762 lnet_free_rtrpools();
763 lnet_fini_finalizers();
764 lnet_destroy_peer_table();
765 lnet_cleanup_handle_hash();
766 lnet_descriptor_cleanup();
772 lnet_net2ni_locked (__u32 net)
774 struct list_head *tmp;
777 list_for_each (tmp, &the_lnet.ln_nis) {
778 ni = list_entry(tmp, lnet_ni_t, ni_list);
780 if (lnet_ptlcompat_matchnet(LNET_NIDNET(ni->ni_nid), net)) {
781 lnet_ni_addref_locked(ni);
790 lnet_islocalnet (__u32 net)
795 ni = lnet_net2ni_locked(net);
797 lnet_ni_decref_locked(ni);
804 lnet_nid2ni_locked (lnet_nid_t nid)
806 struct list_head *tmp;
809 list_for_each (tmp, &the_lnet.ln_nis) {
810 ni = list_entry(tmp, lnet_ni_t, ni_list);
812 if (lnet_ptlcompat_matchnid(ni->ni_nid, nid)) {
813 lnet_ni_addref_locked(ni);
822 lnet_islocalnid (lnet_nid_t nid)
827 ni = lnet_nid2ni_locked(nid);
829 lnet_ni_decref_locked(ni);
836 lnet_count_acceptor_nis (lnet_ni_t **first_ni)
838 /* Return the # of NIs that need the acceptor. Return the first one in
839 * *first_ni so the acceptor can pass it connections "blind" to retain
840 * binary compatibility. */
843 struct list_head *tmp;
847 list_for_each (tmp, &the_lnet.ln_nis) {
848 ni = list_entry(tmp, lnet_ni_t, ni_list);
850 if (ni->ni_lnd->lnd_accept != NULL) {
851 /* This LND uses the acceptor */
852 if (count == 0 && first_ni != NULL) {
853 lnet_ni_addref_locked(ni);
866 lnet_shutdown_lndnis (void)
872 /* NB called holding the global mutex */
874 /* All quiet on the API front */
875 LASSERT (!the_lnet.ln_shutdown);
876 LASSERT (the_lnet.ln_refcount == 0);
877 LASSERT (list_empty(&the_lnet.ln_zombie_nis));
878 LASSERT (the_lnet.ln_nzombie_nis == 0);
879 LASSERT (list_empty(&the_lnet.ln_remote_nets));
882 the_lnet.ln_shutdown = 1; /* flag shutdown */
884 /* Unlink NIs from the global table */
885 while (!list_empty(&the_lnet.ln_nis)) {
886 ni = list_entry(the_lnet.ln_nis.next,
888 list_del (&ni->ni_list);
890 the_lnet.ln_nzombie_nis++;
891 lnet_ni_decref_locked(ni); /* drop apini's ref */
894 /* Drop the cached eqwait NI. */
895 if (the_lnet.ln_eqwaitni != NULL) {
896 lnet_ni_decref_locked(the_lnet.ln_eqwaitni);
897 the_lnet.ln_eqwaitni = NULL;
900 /* Drop the cached loopback NI. */
901 if (the_lnet.ln_loni != NULL) {
902 lnet_ni_decref_locked(the_lnet.ln_loni);
903 the_lnet.ln_loni = NULL;
907 /* Clear the peer table and wait for all peers to go (they hold refs on
910 lnet_clear_peer_table();
913 /* Now wait for the NI's I just nuked to show up on apini_zombie_nis
914 * and shut them down in guaranteed thread context */
916 while (the_lnet.ln_nzombie_nis != 0) {
918 while (list_empty(&the_lnet.ln_zombie_nis)) {
922 CDEBUG(D_WARNING,"Waiting for %d zombie NIs\n",
923 the_lnet.ln_nzombie_nis);
924 cfs_pause(cfs_time_seconds(1));
928 ni = list_entry(the_lnet.ln_zombie_nis.next,
930 list_del(&ni->ni_list);
931 ni->ni_lnd->lnd_refcount--;
935 islo = ni->ni_lnd->lnd_type == LOLND;
937 LASSERT (!in_interrupt ());
938 (ni->ni_lnd->lnd_shutdown)(ni);
940 /* can't deref lnd anymore now; it might have unregistered
944 CDEBUG(D_LNI, "Removed LNI %s\n",
945 libcfs_nid2str(ni->ni_nid));
947 LIBCFS_FREE(ni, sizeof(*ni));
950 the_lnet.ln_nzombie_nis--;
953 the_lnet.ln_shutdown = 0;
956 if (the_lnet.ln_network_tokens != NULL) {
957 LIBCFS_FREE(the_lnet.ln_network_tokens,
958 the_lnet.ln_network_tokens_nob);
959 the_lnet.ln_network_tokens = NULL;
964 lnet_startup_lndnis (void)
968 struct list_head nilist;
972 char *nets = lnet_get_networks();
974 INIT_LIST_HEAD(&nilist);
979 rc = lnet_parse_networks(&nilist, nets);
983 while (!list_empty(&nilist)) {
984 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
985 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
987 LASSERT (libcfs_isknown_lnd(lnd_type));
989 LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
990 lnd = lnet_find_lnd_by_type(lnd_type);
994 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
995 rc = request_module(libcfs_lnd2modname(lnd_type));
996 LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
998 lnd = lnet_find_lnd_by_type(lnd_type);
1000 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1001 CERROR("Can't load LND %s, module %s, rc=%d\n",
1002 libcfs_lnd2str(lnd_type),
1003 libcfs_lnd2modname(lnd_type), rc);
1005 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1006 "compiled with CONFIG_KMOD set for "
1007 "automatic module loading.");
1014 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1015 CERROR("LND %s not supported\n",
1016 libcfs_lnd2str(lnd_type));
1021 ni->ni_refcount = 1;
1024 lnd->lnd_refcount++;
1029 rc = (lnd->lnd_startup)(ni);
1031 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1034 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1036 rc, libcfs_lnd2str(lnd->lnd_type));
1038 lnd->lnd_refcount--;
1043 list_del(&ni->ni_list);
1046 list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1049 if (lnd->lnd_type == LOLND) {
1051 LASSERT (the_lnet.ln_loni == NULL);
1052 the_lnet.ln_loni = ni;
1057 if (lnd->lnd_wait != NULL) {
1058 if (the_lnet.ln_eqwaitni == NULL) {
1060 the_lnet.ln_eqwaitni = ni;
1063 # ifndef HAVE_LIBPTHREAD
1064 LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1065 "single-threaded runtime\n",
1066 libcfs_lnd2str(lnd_type));
1071 if (ni->ni_peertxcredits == 0 ||
1072 ni->ni_maxtxcredits == 0) {
1073 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1074 libcfs_lnd2str(lnd->lnd_type),
1075 ni->ni_peertxcredits == 0 ?
1080 ni->ni_txcredits = ni->ni_mintxcredits = ni->ni_maxtxcredits;
1082 CDEBUG(D_LNI, "Added LNI %s [%d/%d]\n",
1083 libcfs_nid2str(ni->ni_nid),
1084 ni->ni_peertxcredits, ni->ni_txcredits);
1086 /* Handle nidstrings for network 0 just like this one */
1087 if (the_lnet.ln_ptlcompat > 0) {
1089 LCONSOLE_ERROR_MSG(0x108, "Can't run > 1 "
1090 "network when portals_compatibility is "
1094 libcfs_setnet0alias(lnd->lnd_type);
1100 if (the_lnet.ln_eqwaitni != NULL && nicount > 1) {
1101 lnd_type = the_lnet.ln_eqwaitni->ni_lnd->lnd_type;
1102 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1104 libcfs_lnd2str(lnd_type));
1111 lnet_shutdown_lndnis();
1113 while (!list_empty(&nilist)) {
1114 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
1115 list_del(&ni->ni_list);
1116 LIBCFS_FREE(ni, sizeof(*ni));
1127 lnet_assert_wire_constants ();
1128 LASSERT (!the_lnet.ln_init);
1130 memset(&the_lnet, 0, sizeof(the_lnet));
1132 rc = lnet_get_portals_compatibility();
1137 CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1138 the_lnet.ln_ptlcompat = rc;
1139 the_lnet.ln_refcount = 0;
1140 the_lnet.ln_init = 1;
1143 /* All LNDs apart from the LOLND are in separate modules. They
1144 * register themselves when their module loads, and unregister
1145 * themselves when their module is unloaded. */
1148 * NB the order here determines default 'networks=' order */
1150 LNET_REGISTER_ULND(the_ptllnd);
1152 # ifdef HAVE_LIBPTHREAD
1153 LNET_REGISTER_ULND(the_tcplnd);
1156 lnet_register_lnd(&the_lolnd);
1163 LASSERT (the_lnet.ln_init);
1164 LASSERT (the_lnet.ln_refcount == 0);
1166 while (!list_empty(&the_lnet.ln_lnds))
1167 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1171 the_lnet.ln_init = 0;
1175 LNetNIInit(lnet_pid_t requested_pid)
1177 int im_a_router = 0;
1180 LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1182 LASSERT (the_lnet.ln_init);
1183 CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1185 if (the_lnet.ln_refcount > 0) {
1186 rc = the_lnet.ln_refcount++;
1190 if (requested_pid == LNET_PID_ANY) {
1191 /* Don't instantiate LNET just for me */
1196 rc = lnet_prepare(requested_pid);
1200 rc = lnet_startup_lndnis();
1204 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1208 rc = lnet_check_routes();
1212 rc = lnet_alloc_rtrpools(im_a_router);
1216 rc = lnet_acceptor_start();
1220 the_lnet.ln_refcount = 1;
1221 /* Now I may use my own API functions... */
1223 rc = lnet_router_checker_start();
1227 rc = lnet_ping_target_init();
1235 lnet_router_checker_stop();
1237 the_lnet.ln_refcount = 0;
1238 lnet_acceptor_stop();
1240 lnet_destroy_routes();
1241 lnet_shutdown_lndnis();
1247 LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1254 LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1256 LASSERT (the_lnet.ln_init);
1257 LASSERT (the_lnet.ln_refcount > 0);
1259 if (the_lnet.ln_refcount != 1) {
1260 the_lnet.ln_refcount--;
1262 LASSERT (!the_lnet.ln_niinit_self);
1265 lnet_ping_target_fini();
1266 lnet_router_checker_stop();
1268 /* Teardown fns that use my own API functions BEFORE here */
1269 the_lnet.ln_refcount = 0;
1271 lnet_acceptor_stop();
1272 lnet_destroy_routes();
1273 lnet_shutdown_lndnis();
1277 LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1282 LNetCtl(unsigned int cmd, void *arg)
1284 struct libcfs_ioctl_data *data = arg;
1285 lnet_process_id_t id;
1289 LASSERT (the_lnet.ln_init);
1290 LASSERT (the_lnet.ln_refcount > 0);
1293 case IOC_LIBCFS_GET_NI:
1294 rc = LNetGetId(data->ioc_count, &id);
1295 data->ioc_nid = id.nid;
1298 case IOC_LIBCFS_FAIL_NID:
1299 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1301 case IOC_LIBCFS_ADD_ROUTE:
1302 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1304 return (rc != 0) ? rc : lnet_check_routes();
1306 case IOC_LIBCFS_DEL_ROUTE:
1307 return lnet_del_route(data->ioc_net, data->ioc_nid);
1309 case IOC_LIBCFS_GET_ROUTE:
1310 return lnet_get_route(data->ioc_count,
1311 &data->ioc_net, &data->ioc_count,
1312 &data->ioc_nid, &data->ioc_flags);
1313 case IOC_LIBCFS_NOTIFY_ROUTER:
1314 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1315 (time_t)data->ioc_u64[0]);
1317 case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1318 return the_lnet.ln_ptlcompat;
1320 case IOC_LIBCFS_LNET_DIST:
1321 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1322 if (rc < 0 && rc != -EHOSTUNREACH)
1325 data->ioc_u32[0] = rc;
1328 case IOC_LIBCFS_TESTPROTOCOMPAT:
1330 the_lnet.ln_testprotocompat = data->ioc_flags;
1334 case IOC_LIBCFS_PING:
1335 rc = lnet_ping((lnet_process_id_t) {.nid = data->ioc_nid,
1336 .pid = data->ioc_u32[0]},
1337 data->ioc_u32[1], /* timeout */
1338 (lnet_process_id_t *)data->ioc_pbuf1,
1339 data->ioc_plen1/sizeof(lnet_process_id_t));
1342 data->ioc_count = rc;
1345 case IOC_LIBCFS_DEBUG_PEER: {
1346 /* CAVEAT EMPTOR: this one designed for calling directly; not
1348 lnet_process_id_t *id = arg;
1350 lnet_debug_peer(id->nid);
1352 ni = lnet_net2ni(LNET_NIDNET(id->nid));
1354 CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(*id));
1356 if (ni->ni_lnd->lnd_ctl == NULL) {
1357 CDEBUG(D_WARNING, "No ctl for %s\n",
1358 libcfs_id2str(*id));
1360 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1369 ni = lnet_net2ni(data->ioc_net);
1373 if (ni->ni_lnd->lnd_ctl == NULL)
1376 rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1385 LNetGetId(unsigned int index, lnet_process_id_t *id)
1388 struct list_head *tmp;
1391 LASSERT (the_lnet.ln_init);
1392 LASSERT (the_lnet.ln_refcount > 0);
1396 list_for_each(tmp, &the_lnet.ln_nis) {
1400 ni = list_entry(tmp, lnet_ni_t, ni_list);
1402 id->nid = ni->ni_nid;
1403 id->pid = the_lnet.ln_pid;
1414 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1416 snprintf(str, len, LPX64, h.cookie);
1421 lnet_ping_target_init(void)
1423 lnet_handle_me_t meh;
1424 lnet_process_id_t id;
1431 for (n = 0; ; n++) {
1432 rc = LNetGetId(n, &id);
1439 infosz = offsetof(lnet_ping_info_t, pi_nid[n]);
1440 LIBCFS_ALLOC(the_lnet.ln_ping_info, infosz);
1441 if (the_lnet.ln_ping_info == NULL) {
1442 CERROR("Can't allocate ping info[%d]\n", n);
1446 the_lnet.ln_ping_info->pi_magic = LNET_PROTO_PING_MAGIC;
1447 the_lnet.ln_ping_info->pi_version = LNET_PROTO_PING_VERSION;
1448 the_lnet.ln_ping_info->pi_pid = the_lnet.ln_pid;
1449 the_lnet.ln_ping_info->pi_nnids = n;
1451 for (i = 0; i < n; i++) {
1452 rc = LNetGetId(i, &id);
1454 the_lnet.ln_ping_info->pi_nid[i] = id.nid;
1457 /* We can have a tiny EQ since we only need to see the unlink event on
1458 * teardown, which by definition is the last one! */
1459 rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1461 CERROR("Can't allocate ping EQ: %d\n", rc);
1465 rc = LNetMEAttach(LNET_RESERVED_PORTAL,
1466 (lnet_process_id_t){.nid = LNET_NID_ANY,
1467 .pid = LNET_PID_ANY},
1468 LNET_PROTO_PING_MATCHBITS, 0LL,
1469 LNET_UNLINK, LNET_INS_AFTER,
1472 CERROR("Can't create ping ME: %d\n", rc);
1476 rc = LNetMDAttach(meh,
1477 (lnet_md_t){.start = the_lnet.ln_ping_info,
1479 .threshold = LNET_MD_THRESH_INF,
1480 .options = (LNET_MD_OP_GET |
1482 LNET_MD_MANAGE_REMOTE),
1483 .eq_handle = the_lnet.ln_ping_target_eq},
1485 &the_lnet.ln_ping_target_md);
1487 CERROR("Can't attach ping MD: %d\n", rc);
1494 rc2 = LNetMEUnlink(meh);
1497 rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1500 LIBCFS_FREE(the_lnet.ln_ping_info, infosz);
1506 lnet_ping_target_fini(void)
1511 int timeout_ms = 1000;
1512 cfs_sigset_t blocked = cfs_block_allsigs();
1514 LNetMDUnlink(the_lnet.ln_ping_target_md);
1515 /* NB md could be busy; this just starts the unlink */
1518 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1519 timeout_ms, &event, &which);
1521 /* I expect overflow... */
1522 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1525 /* timed out: provide a diagnostic */
1526 CWARN("Still waiting for ping MD to unlink\n");
1531 /* Got a valid event */
1536 rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1539 LIBCFS_FREE(the_lnet.ln_ping_info,
1540 offsetof(lnet_ping_info_t,
1541 pi_nid[the_lnet.ln_ping_info->pi_nnids]));
1543 cfs_restore_sigs(blocked);
1547 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1549 lnet_handle_eq_t eqh;
1550 lnet_handle_md_t mdh;
1555 const int a_long_time = 60000; /* mS */
1556 int infosz = offsetof(lnet_ping_info_t, pi_nid[n_ids]);
1557 lnet_ping_info_t *info;
1558 lnet_process_id_t tmpid;
1563 cfs_sigset_t blocked;
1566 id.nid == LNET_NID_ANY ||
1567 timeout_ms > 500000 || /* arbitrary limit! */
1568 n_ids > 20) /* arbitrary limit! */
1571 if (id.pid == LNET_PID_ANY)
1572 id.pid = LUSTRE_SRV_LNET_PID;
1574 LIBCFS_ALLOC(info, infosz);
1578 /* NB 2 events max (including any unlink event) */
1579 rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1581 CERROR("Can't allocate EQ: %d\n", rc);
1585 rc = LNetMDBind((lnet_md_t){.start = info,
1587 .threshold = 2, /* GET/REPLY */
1588 .options = LNET_MD_TRUNCATE,
1593 CERROR("Can't bind MD: %d\n", rc);
1597 rc = LNetGet(LNET_NID_ANY, mdh, id,
1598 LNET_RESERVED_PORTAL,
1599 LNET_PROTO_PING_MATCHBITS, 0);
1602 /* Don't CERROR; this could be deliberate! */
1604 rc2 = LNetMDUnlink(mdh);
1607 /* NB must wait for the UNLINK event below... */
1609 timeout_ms = a_long_time;
1613 /* MUST block for unlink to complete */
1615 blocked = cfs_block_allsigs();
1617 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1620 cfs_restore_sigs(blocked);
1622 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1623 (rc2 <= 0) ? -1 : event.type,
1624 (rc2 <= 0) ? -1 : event.status,
1625 (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1627 LASSERT (rc2 != -EOVERFLOW); /* can't miss anything */
1629 if (rc2 <= 0 || event.status != 0) {
1630 /* timeout or error */
1631 if (!replied && rc == 0)
1632 rc = (rc2 < 0) ? rc2 :
1633 (rc2 == 0) ? -ETIMEDOUT :
1637 /* Ensure completion in finite time... */
1639 /* No assertion (racing with network) */
1641 timeout_ms = a_long_time;
1642 } else if (rc2 == 0) {
1643 /* timed out waiting for unlink */
1644 CWARN("ping %s: late network completion\n",
1648 } else if (event.type == LNET_EVENT_REPLY) {
1653 } while (rc2 <= 0 || !event.unlinked);
1657 CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1664 LASSERT (nob >= 0 && nob <= infosz);
1666 rc = -EPROTO; /* if I can't parse... */
1669 /* can't check magic/version */
1670 CERROR("%s: ping info too short %d\n",
1671 libcfs_id2str(id), nob);
1675 if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
1676 /* NB I might be swabbing garbage until I check below, but it
1678 __swab32s(&info->pi_version);
1679 __swab32s(&info->pi_pid);
1680 __swab32s(&info->pi_nnids);
1681 for (i = 0; i < info->pi_nnids && i < n_ids; i++)
1682 __swab64s(&info->pi_nid[i]);
1684 } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
1685 CERROR("%s: Unexpected magic %08x\n",
1686 libcfs_id2str(id), info->pi_magic);
1690 if (info->pi_version != LNET_PROTO_PING_VERSION) {
1691 CERROR("%s: Unexpected version 0x%x\n",
1692 libcfs_id2str(id), info->pi_version);
1696 if (nob < offsetof(lnet_ping_info_t, pi_nid[0])) {
1697 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
1698 nob, (int)offsetof(lnet_ping_info_t, pi_nid[0]));
1702 if (info->pi_nnids < n_ids)
1703 n_ids = info->pi_nnids;
1705 if (nob < offsetof(lnet_ping_info_t, pi_nid[n_ids])) {
1706 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
1707 nob, (int)offsetof(lnet_ping_info_t, pi_nid[n_ids]));
1711 rc = -EFAULT; /* If I SEGV... */
1713 for (i = 0; i < n_ids; i++) {
1714 tmpid.pid = info->pi_pid;
1715 tmpid.nid = info->pi_nid[i];
1717 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
1723 rc = info->pi_nnids;
1726 rc2 = LNetEQFree(eqh);
1728 CERROR("rc2 %d\n", rc2);
1732 LIBCFS_FREE(info, infosz);