4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
35 #define DEBUG_SUBSYSTEM S_LNET
37 #include <linux/sched.h>
38 #ifdef HAVE_SCHED_HEADERS
39 #include <linux/sched/signal.h>
41 #include <linux/uaccess.h>
43 #include <lnet/lib-lnet.h>
44 #include <uapi/linux/lnet/lnet-dlc.h>
46 /* Value indicating that recovery needs to re-check a peer immediately. */
47 #define LNET_REDISCOVER_PEER (1)
49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
54 if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
55 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
56 lnet_peer_ni_decref_locked(lpni);
61 lnet_peer_net_added(struct lnet_net *net)
63 struct lnet_peer_ni *lpni, *tmp;
65 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
66 lpni_on_remote_peer_ni_list) {
68 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
71 spin_lock(&lpni->lpni_lock);
72 lpni->lpni_txcredits =
73 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
74 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
75 lpni->lpni_rtrcredits =
76 lnet_peer_buffer_credits(lpni->lpni_net);
77 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
78 spin_unlock(&lpni->lpni_lock);
80 lnet_peer_remove_from_remote_list(lpni);
86 lnet_peer_tables_destroy(void)
88 struct lnet_peer_table *ptable;
89 struct list_head *hash;
93 if (!the_lnet.ln_peer_tables)
96 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
97 hash = ptable->pt_hash;
98 if (!hash) /* not intialized */
101 LASSERT(list_empty(&ptable->pt_zombie_list));
103 ptable->pt_hash = NULL;
104 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
105 LASSERT(list_empty(&hash[j]));
107 LIBCFS_FREE(hash, LNET_PEER_HASH_SIZE * sizeof(*hash));
110 cfs_percpt_free(the_lnet.ln_peer_tables);
111 the_lnet.ln_peer_tables = NULL;
115 lnet_peer_tables_create(void)
117 struct lnet_peer_table *ptable;
118 struct list_head *hash;
122 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
124 if (the_lnet.ln_peer_tables == NULL) {
125 CERROR("Failed to allocate cpu-partition peer tables\n");
129 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131 LNET_PEER_HASH_SIZE * sizeof(*hash));
133 CERROR("Failed to create peer hash table\n");
134 lnet_peer_tables_destroy();
138 spin_lock_init(&ptable->pt_zombie_lock);
139 INIT_LIST_HEAD(&ptable->pt_zombie_list);
141 INIT_LIST_HEAD(&ptable->pt_peer_list);
143 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144 INIT_LIST_HEAD(&hash[j]);
145 ptable->pt_hash = hash; /* sign of initialization */
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(lnet_nid_t nid)
154 struct lnet_peer_ni *lpni;
155 struct lnet_net *net;
158 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
160 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
164 INIT_LIST_HEAD(&lpni->lpni_txq);
165 INIT_LIST_HEAD(&lpni->lpni_hashlist);
166 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167 INIT_LIST_HEAD(&lpni->lpni_recovery);
168 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171 spin_lock_init(&lpni->lpni_lock);
173 if (lnet_peers_start_down())
174 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
176 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
177 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
178 lpni->lpni_nid = nid;
179 lpni->lpni_cpt = cpt;
180 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
182 net = lnet_get_net_locked(LNET_NIDNET(nid));
183 lpni->lpni_net = net;
185 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
186 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
187 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
188 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
191 * This peer_ni is not on a local network, so we
192 * cannot add the credits here. In case the net is
193 * added later, add the peer_ni to the remote peer ni
194 * list so it can be easily found and revisited.
196 /* FIXME: per-net implementation instead? */
197 atomic_inc(&lpni->lpni_refcount);
198 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
199 &the_lnet.ln_remote_peer_ni_list);
202 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
207 static struct lnet_peer_net *
208 lnet_peer_net_alloc(__u32 net_id)
210 struct lnet_peer_net *lpn;
212 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
216 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
217 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
218 lpn->lpn_net_id = net_id;
220 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
226 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
228 struct lnet_peer *lp;
230 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
232 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
233 LASSERT(list_empty(&lpn->lpn_peer_nis));
234 LASSERT(list_empty(&lpn->lpn_peer_nets));
236 lpn->lpn_peer = NULL;
237 LIBCFS_FREE(lpn, sizeof(*lpn));
239 lnet_peer_decref_locked(lp);
242 static struct lnet_peer *
243 lnet_peer_alloc(lnet_nid_t nid)
245 struct lnet_peer *lp;
247 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
251 INIT_LIST_HEAD(&lp->lp_rtrq);
252 INIT_LIST_HEAD(&lp->lp_routes);
253 INIT_LIST_HEAD(&lp->lp_peer_list);
254 INIT_LIST_HEAD(&lp->lp_peer_nets);
255 INIT_LIST_HEAD(&lp->lp_dc_list);
256 INIT_LIST_HEAD(&lp->lp_dc_pendq);
257 INIT_LIST_HEAD(&lp->lp_rtr_list);
258 init_waitqueue_head(&lp->lp_dc_waitq);
259 spin_lock_init(&lp->lp_lock);
260 lp->lp_primary_nid = nid;
263 * all peers created on a router should have health on
264 * if it's not already on.
266 if (the_lnet.ln_routing && !lnet_health_sensitivity)
267 lp->lp_health_sensitivity = 1;
270 * Turn off discovery for loopback peer. If you're creating a peer
271 * for the loopback interface then that was initiated when we
272 * attempted to send a message over the loopback. There is no need
273 * to ever use a different interface when sending messages to
276 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
277 lp->lp_state = LNET_PEER_NO_DISCOVERY;
278 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
280 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
286 lnet_destroy_peer_locked(struct lnet_peer *lp)
288 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
290 LASSERT(atomic_read(&lp->lp_refcount) == 0);
291 LASSERT(lp->lp_rtr_refcount == 0);
292 LASSERT(list_empty(&lp->lp_peer_nets));
293 LASSERT(list_empty(&lp->lp_peer_list));
294 LASSERT(list_empty(&lp->lp_dc_list));
297 lnet_ping_buffer_decref(lp->lp_data);
300 * if there are messages still on the pending queue, then make
301 * sure to queue them on the ln_msg_resend list so they can be
302 * resent at a later point if the discovery thread is still
304 * If the discovery thread has stopped, then the wakeup will be a
305 * no-op, and it is expected the lnet_shutdown_lndnets() will
306 * eventually be called, which will traverse this list and
307 * finalize the messages on the list.
308 * We can not resend them now because we're holding the cpt lock.
309 * Releasing the lock can cause an inconsistent state
311 spin_lock(&the_lnet.ln_msg_resend_lock);
312 spin_lock(&lp->lp_lock);
313 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
314 spin_unlock(&lp->lp_lock);
315 spin_unlock(&the_lnet.ln_msg_resend_lock);
316 wake_up(&the_lnet.ln_dc_waitq);
318 LIBCFS_FREE(lp, sizeof(*lp));
322 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
323 * that peer_net, detach the peer_net from the peer.
325 * Call with lnet_net_lock/EX held
328 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
330 struct lnet_peer_table *ptable;
331 struct lnet_peer_net *lpn;
332 struct lnet_peer *lp;
335 * Belts and suspenders: gracefully handle teardown of a
336 * partially connected peer_ni.
338 lpn = lpni->lpni_peer_net;
340 list_del_init(&lpni->lpni_peer_nis);
342 * If there are no lpni's left, we detach lpn from
343 * lp_peer_nets, so it cannot be found anymore.
345 if (list_empty(&lpn->lpn_peer_nis))
346 list_del_init(&lpn->lpn_peer_nets);
348 /* Update peer NID count. */
353 * If there are no more peer nets, make the peer unfindable
354 * via the peer_tables.
356 * Otherwise, if the peer is DISCOVERED, tell discovery to
357 * take another look at it. This is a no-op if discovery for
358 * this peer did the detaching.
360 if (list_empty(&lp->lp_peer_nets)) {
361 list_del_init(&lp->lp_peer_list);
362 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
364 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
365 /* Discovery isn't running, nothing to do here. */
366 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
367 lnet_peer_queue_for_discovery(lp);
368 wake_up(&the_lnet.ln_dc_waitq);
370 CDEBUG(D_NET, "peer %s NID %s\n",
371 libcfs_nid2str(lp->lp_primary_nid),
372 libcfs_nid2str(lpni->lpni_nid));
375 /* called with lnet_net_lock LNET_LOCK_EX held */
377 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
379 struct lnet_peer_table *ptable = NULL;
381 /* don't remove a peer_ni if it's also a gateway */
382 if (lnet_isrouter(lpni) && !force) {
383 CERROR("Peer NI %s is a gateway. Can not delete it\n",
384 libcfs_nid2str(lpni->lpni_nid));
388 lnet_peer_remove_from_remote_list(lpni);
390 /* remove peer ni from the hash list. */
391 list_del_init(&lpni->lpni_hashlist);
394 * indicate the peer is being deleted so the monitor thread can
395 * remove it from the recovery queue.
397 spin_lock(&lpni->lpni_lock);
398 lpni->lpni_state |= LNET_PEER_NI_DELETING;
399 spin_unlock(&lpni->lpni_lock);
401 /* decrement the ref count on the peer table */
402 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
403 LASSERT(ptable->pt_number > 0);
407 * The peer_ni can no longer be found with a lookup. But there
408 * can be current users, so keep track of it on the zombie
409 * list until the reference count has gone to zero.
411 * The last reference may be lost in a place where the
412 * lnet_net_lock locks only a single cpt, and that cpt may not
413 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
416 spin_lock(&ptable->pt_zombie_lock);
417 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
418 ptable->pt_zombies++;
419 spin_unlock(&ptable->pt_zombie_lock);
421 /* no need to keep this peer_ni on the hierarchy anymore */
422 lnet_peer_detach_peer_ni_locked(lpni);
424 /* remove hashlist reference on peer_ni */
425 lnet_peer_ni_decref_locked(lpni);
430 void lnet_peer_uninit(void)
432 struct lnet_peer_ni *lpni, *tmp;
434 lnet_net_lock(LNET_LOCK_EX);
436 /* remove all peer_nis from the remote peer and the hash list */
437 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
438 lpni_on_remote_peer_ni_list)
439 lnet_peer_ni_del_locked(lpni, false);
441 lnet_peer_tables_destroy();
443 lnet_net_unlock(LNET_LOCK_EX);
447 lnet_peer_del_locked(struct lnet_peer *peer)
449 struct lnet_peer_ni *lpni = NULL, *lpni2;
452 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
454 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
455 while (lpni != NULL) {
456 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
457 rc = lnet_peer_ni_del_locked(lpni, false);
467 lnet_peer_del(struct lnet_peer *peer)
469 lnet_net_lock(LNET_LOCK_EX);
470 lnet_peer_del_locked(peer);
471 lnet_net_unlock(LNET_LOCK_EX);
477 * Delete a NID from a peer. Call with ln_api_mutex held.
480 * -EPERM: Non-DLC deletion from DLC-configured peer.
481 * -ENOENT: No lnet_peer_ni corresponding to the nid.
482 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
483 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
486 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
488 struct lnet_peer_ni *lpni;
489 lnet_nid_t primary_nid = lp->lp_primary_nid;
491 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
493 if (!(flags & LNET_PEER_CONFIGURED)) {
494 if (lp->lp_state & LNET_PEER_CONFIGURED) {
499 lpni = lnet_find_peer_ni_locked(nid);
504 lnet_peer_ni_decref_locked(lpni);
505 if (lp != lpni->lpni_peer_net->lpn_peer) {
511 * This function only allows deletion of the primary NID if it
514 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
519 lnet_net_lock(LNET_LOCK_EX);
521 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
522 struct lnet_peer_ni *lpni2;
523 /* assign the next peer_ni to be the primary */
524 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
526 lp->lp_primary_nid = lpni->lpni_nid;
528 rc = lnet_peer_ni_del_locked(lpni, force);
530 lnet_net_unlock(LNET_LOCK_EX);
533 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
534 libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
540 lnet_peer_table_cleanup_locked(struct lnet_net *net,
541 struct lnet_peer_table *ptable)
544 struct lnet_peer_ni *next;
545 struct lnet_peer_ni *lpni;
546 struct lnet_peer *peer;
548 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
549 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
551 if (net != NULL && net != lpni->lpni_net)
554 peer = lpni->lpni_peer_net->lpn_peer;
555 if (peer->lp_primary_nid != lpni->lpni_nid) {
556 lnet_peer_ni_del_locked(lpni, false);
560 * Removing the primary NID implies removing
561 * the entire peer. Advance next beyond any
562 * peer_ni that belongs to the same peer.
564 list_for_each_entry_from(next, &ptable->pt_hash[i],
566 if (next->lpni_peer_net->lpn_peer != peer)
569 lnet_peer_del_locked(peer);
575 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
579 spin_lock(&ptable->pt_zombie_lock);
580 while (ptable->pt_zombies) {
581 spin_unlock(&ptable->pt_zombie_lock);
583 if (is_power_of_2(i)) {
585 "Waiting for %d zombies on peer table\n",
588 set_current_state(TASK_UNINTERRUPTIBLE);
589 schedule_timeout(cfs_time_seconds(1) >> 1);
590 spin_lock(&ptable->pt_zombie_lock);
592 spin_unlock(&ptable->pt_zombie_lock);
596 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
597 struct lnet_peer_table *ptable)
599 struct lnet_peer_ni *lp;
600 struct lnet_peer_ni *tmp;
604 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
605 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
607 if (net != lp->lpni_net)
610 if (!lnet_isrouter(lp))
613 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
615 lnet_net_unlock(LNET_LOCK_EX);
616 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
617 lnet_net_lock(LNET_LOCK_EX);
623 lnet_peer_tables_cleanup(struct lnet_net *net)
626 struct lnet_peer_table *ptable;
628 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
629 /* If just deleting the peers for a NI, get rid of any routes these
630 * peers are gateways for. */
631 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
632 lnet_net_lock(LNET_LOCK_EX);
633 lnet_peer_table_del_rtrs_locked(net, ptable);
634 lnet_net_unlock(LNET_LOCK_EX);
637 /* Start the cleanup process */
638 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
639 lnet_net_lock(LNET_LOCK_EX);
640 lnet_peer_table_cleanup_locked(net, ptable);
641 lnet_net_unlock(LNET_LOCK_EX);
644 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
645 lnet_peer_ni_finalize_wait(ptable);
648 static struct lnet_peer_ni *
649 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
651 struct list_head *peers;
652 struct lnet_peer_ni *lp;
654 LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
656 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
657 list_for_each_entry(lp, peers, lpni_hashlist) {
658 if (lp->lpni_nid == nid) {
659 lnet_peer_ni_addref_locked(lp);
667 struct lnet_peer_ni *
668 lnet_find_peer_ni_locked(lnet_nid_t nid)
670 struct lnet_peer_ni *lpni;
671 struct lnet_peer_table *ptable;
674 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
676 ptable = the_lnet.ln_peer_tables[cpt];
677 lpni = lnet_get_peer_ni_locked(ptable, nid);
682 struct lnet_peer_ni *
683 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
685 struct lnet_peer_net *lpn;
686 struct lnet_peer_ni *lpni;
688 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
692 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
693 if (lpni->lpni_nid == nid)
701 lnet_find_peer(lnet_nid_t nid)
703 struct lnet_peer_ni *lpni;
704 struct lnet_peer *lp = NULL;
707 cpt = lnet_net_lock_current();
708 lpni = lnet_find_peer_ni_locked(nid);
710 lp = lpni->lpni_peer_net->lpn_peer;
711 lnet_peer_addref_locked(lp);
712 lnet_peer_ni_decref_locked(lpni);
714 lnet_net_unlock(cpt);
719 struct lnet_peer_net *
720 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
722 struct lnet_peer_net *net;
725 /* no net id provided return the first net */
726 net = list_first_entry_or_null(&lp->lp_peer_nets,
727 struct lnet_peer_net,
733 /* find the net after the one provided */
734 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
735 if (net->lpn_net_id == prev_lpn_id) {
737 * if we reached the end of the list loop to the
740 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
741 return list_first_entry_or_null(&lp->lp_peer_nets,
742 struct lnet_peer_net,
745 return list_next_entry(net, lpn_peer_nets);
752 struct lnet_peer_ni *
753 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
754 struct lnet_peer_net *peer_net,
755 struct lnet_peer_ni *prev)
757 struct lnet_peer_ni *lpni;
758 struct lnet_peer_net *net = peer_net;
762 if (list_empty(&peer->lp_peer_nets))
765 net = list_entry(peer->lp_peer_nets.next,
766 struct lnet_peer_net,
769 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
775 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
777 * if you reached the end of the peer ni list and the peer
778 * net is specified then there are no more peer nis in that
785 * we reached the end of this net ni list. move to the
788 if (prev->lpni_peer_net->lpn_peer_nets.next ==
790 /* no more nets and no more NIs. */
793 /* get the next net */
794 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
795 struct lnet_peer_net,
797 /* get the ni on it */
798 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
804 /* there are more nis left */
805 lpni = list_entry(prev->lpni_peer_nis.next,
806 struct lnet_peer_ni, lpni_peer_nis);
811 /* Call with the ln_api_mutex held */
812 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
814 struct lnet_process_id id;
815 struct lnet_peer_table *ptable;
816 struct lnet_peer *lp;
825 if (the_lnet.ln_state != LNET_STATE_RUNNING)
828 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
831 * Count the number of peers, and return E2BIG if the buffer
832 * is too small. We'll also return the desired size.
835 for (cpt = 0; cpt < lncpt; cpt++) {
836 ptable = the_lnet.ln_peer_tables[cpt];
837 count += ptable->pt_peers;
839 size = count * sizeof(*ids);
844 * Walk the peer lists and copy out the primary nids.
845 * This is safe because the peer lists are only modified
846 * while the ln_api_mutex is held. So we don't need to
847 * hold the lnet_net_lock as well, and can therefore
848 * directly call copy_to_user().
851 memset(&id, 0, sizeof(id));
852 id.pid = LNET_PID_LUSTRE;
854 for (cpt = 0; cpt < lncpt; cpt++) {
855 ptable = the_lnet.ln_peer_tables[cpt];
856 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
859 id.nid = lp->lp_primary_nid;
860 if (copy_to_user(&ids[i], &id, sizeof(id)))
873 * Start pushes to peers that need to be updated for a configuration
874 * change on this node.
877 lnet_push_update_to_peers(int force)
879 struct lnet_peer_table *ptable;
880 struct lnet_peer *lp;
884 lnet_net_lock(LNET_LOCK_EX);
885 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
886 for (cpt = 0; cpt < lncpt; cpt++) {
887 ptable = the_lnet.ln_peer_tables[cpt];
888 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
890 spin_lock(&lp->lp_lock);
891 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
892 lp->lp_state |= LNET_PEER_FORCE_PUSH;
893 spin_unlock(&lp->lp_lock);
895 if (lnet_peer_needs_push(lp))
896 lnet_peer_queue_for_discovery(lp);
899 lnet_net_unlock(LNET_LOCK_EX);
900 wake_up(&the_lnet.ln_dc_waitq);
904 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
905 * this is a preferred point-to-point path. Call with lnet_net_lock in
909 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
913 if (lpni->lpni_pref_nnids == 0)
915 if (lpni->lpni_pref_nnids == 1)
916 return lpni->lpni_pref.nid == nid;
917 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
918 if (lpni->lpni_pref.nids[i] == nid)
925 * Set a single ni as preferred, provided no preferred ni is already
926 * defined. Only to be used for non-multi-rail peer_ni.
929 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
933 spin_lock(&lpni->lpni_lock);
934 if (nid == LNET_NID_ANY) {
936 } else if (lpni->lpni_pref_nnids > 0) {
938 } else if (lpni->lpni_pref_nnids == 0) {
939 lpni->lpni_pref.nid = nid;
940 lpni->lpni_pref_nnids = 1;
941 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
943 spin_unlock(&lpni->lpni_lock);
945 CDEBUG(D_NET, "peer %s nid %s: %d\n",
946 libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
951 * Clear the preferred NID from a non-multi-rail peer_ni, provided
952 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
955 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
959 spin_lock(&lpni->lpni_lock);
960 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
961 lpni->lpni_pref_nnids = 0;
962 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
963 } else if (lpni->lpni_pref_nnids == 0) {
968 spin_unlock(&lpni->lpni_lock);
970 CDEBUG(D_NET, "peer %s: %d\n",
971 libcfs_nid2str(lpni->lpni_nid), rc);
976 * Clear the preferred NIDs from a non-multi-rail peer.
979 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
981 struct lnet_peer_ni *lpni = NULL;
983 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
984 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
988 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
990 lnet_nid_t *nids = NULL;
991 lnet_nid_t *oldnids = NULL;
992 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
997 if (nid == LNET_NID_ANY) {
1002 if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1007 /* A non-MR node may have only one preferred NI per peer_ni */
1008 if (lpni->lpni_pref_nnids > 0) {
1009 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1015 if (lpni->lpni_pref_nnids != 0) {
1016 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1017 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1022 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1023 if (lpni->lpni_pref.nids[i] == nid) {
1024 LIBCFS_FREE(nids, size);
1028 nids[i] = lpni->lpni_pref.nids[i];
1033 lnet_net_lock(LNET_LOCK_EX);
1034 spin_lock(&lpni->lpni_lock);
1035 if (lpni->lpni_pref_nnids == 0) {
1036 lpni->lpni_pref.nid = nid;
1038 oldnids = lpni->lpni_pref.nids;
1039 lpni->lpni_pref.nids = nids;
1041 lpni->lpni_pref_nnids++;
1042 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1043 spin_unlock(&lpni->lpni_lock);
1044 lnet_net_unlock(LNET_LOCK_EX);
1047 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1048 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1051 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1052 spin_lock(&lpni->lpni_lock);
1053 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1054 spin_unlock(&lpni->lpni_lock);
1056 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1057 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1062 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1064 lnet_nid_t *nids = NULL;
1065 lnet_nid_t *oldnids = NULL;
1066 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1071 if (lpni->lpni_pref_nnids == 0) {
1076 if (lpni->lpni_pref_nnids == 1) {
1077 if (lpni->lpni_pref.nid != nid) {
1081 } else if (lpni->lpni_pref_nnids == 2) {
1082 if (lpni->lpni_pref.nids[0] != nid &&
1083 lpni->lpni_pref.nids[1] != nid) {
1088 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1089 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1094 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1095 if (lpni->lpni_pref.nids[i] != nid)
1097 nids[j++] = lpni->lpni_pref.nids[i];
1099 /* Check if we actually removed a nid. */
1100 if (j == lpni->lpni_pref_nnids) {
1101 LIBCFS_FREE(nids, size);
1107 lnet_net_lock(LNET_LOCK_EX);
1108 spin_lock(&lpni->lpni_lock);
1109 if (lpni->lpni_pref_nnids == 1) {
1110 lpni->lpni_pref.nid = LNET_NID_ANY;
1111 } else if (lpni->lpni_pref_nnids == 2) {
1112 oldnids = lpni->lpni_pref.nids;
1113 if (oldnids[0] == nid)
1114 lpni->lpni_pref.nid = oldnids[1];
1116 lpni->lpni_pref.nid = oldnids[2];
1118 oldnids = lpni->lpni_pref.nids;
1119 lpni->lpni_pref.nids = nids;
1121 lpni->lpni_pref_nnids--;
1122 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1123 spin_unlock(&lpni->lpni_lock);
1124 lnet_net_unlock(LNET_LOCK_EX);
1127 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1128 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1131 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1132 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1137 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1139 struct lnet_peer_ni *lpni;
1140 lnet_nid_t primary_nid = nid;
1142 lpni = lnet_find_peer_ni_locked(nid);
1144 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1145 lnet_peer_ni_decref_locked(lpni);
1152 LNetPrimaryNID(lnet_nid_t nid)
1154 struct lnet_peer *lp;
1155 struct lnet_peer_ni *lpni;
1156 lnet_nid_t primary_nid = nid;
1160 cpt = lnet_net_lock_current();
1161 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1166 lp = lpni->lpni_peer_net->lpn_peer;
1167 while (!lnet_peer_is_uptodate(lp)) {
1168 rc = lnet_discover_peer_locked(lpni, cpt, true);
1171 lp = lpni->lpni_peer_net->lpn_peer;
1173 primary_nid = lp->lp_primary_nid;
1175 lnet_peer_ni_decref_locked(lpni);
1177 lnet_net_unlock(cpt);
1179 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1180 libcfs_nid2str(primary_nid), rc);
1183 EXPORT_SYMBOL(LNetPrimaryNID);
1185 struct lnet_peer_net *
1186 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1188 struct lnet_peer_net *peer_net;
1189 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1190 if (peer_net->lpn_net_id == net_id)
1197 * Attach a peer_ni to a peer_net and peer. This function assumes
1198 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1199 * may be attached to a different peer, in which case it will be
1200 * properly detached first. The whole operation is done atomically.
1202 * Always returns 0. This is the last function called from functions
1203 * that do return an int, so returning 0 here allows the compiler to
1207 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1208 struct lnet_peer_net *lpn,
1209 struct lnet_peer_ni *lpni,
1212 struct lnet_peer_table *ptable;
1214 /* Install the new peer_ni */
1215 lnet_net_lock(LNET_LOCK_EX);
1216 /* Add peer_ni to global peer table hash, if necessary. */
1217 if (list_empty(&lpni->lpni_hashlist)) {
1218 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1220 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1221 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1222 ptable->pt_version++;
1223 ptable->pt_number++;
1224 /* This is the 1st refcount on lpni. */
1225 atomic_inc(&lpni->lpni_refcount);
1228 /* Detach the peer_ni from an existing peer, if necessary. */
1229 if (lpni->lpni_peer_net) {
1230 LASSERT(lpni->lpni_peer_net != lpn);
1231 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1232 lnet_peer_detach_peer_ni_locked(lpni);
1233 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1234 lpni->lpni_peer_net = NULL;
1237 /* Add peer_ni to peer_net */
1238 lpni->lpni_peer_net = lpn;
1239 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1240 lnet_peer_net_addref_locked(lpn);
1242 /* Add peer_net to peer */
1243 if (!lpn->lpn_peer) {
1245 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1246 lnet_peer_addref_locked(lp);
1249 /* Add peer to global peer list, if necessary */
1250 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1251 if (list_empty(&lp->lp_peer_list)) {
1252 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1257 /* Update peer state */
1258 spin_lock(&lp->lp_lock);
1259 if (flags & LNET_PEER_CONFIGURED) {
1260 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1261 lp->lp_state |= LNET_PEER_CONFIGURED;
1263 if (flags & LNET_PEER_MULTI_RAIL) {
1264 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1265 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1266 lnet_peer_clr_non_mr_pref_nids(lp);
1269 spin_unlock(&lp->lp_lock);
1272 lnet_net_unlock(LNET_LOCK_EX);
1274 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1275 libcfs_nid2str(lp->lp_primary_nid),
1276 libcfs_nid2str(lpni->lpni_nid), flags);
1282 * Create a new peer, with nid as its primary nid.
1284 * Call with the lnet_api_mutex held.
1287 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1289 struct lnet_peer *lp;
1290 struct lnet_peer_net *lpn;
1291 struct lnet_peer_ni *lpni;
1294 LASSERT(nid != LNET_NID_ANY);
1297 * No need for the lnet_net_lock here, because the
1298 * lnet_api_mutex is held.
1300 lpni = lnet_find_peer_ni_locked(nid);
1302 /* A peer with this NID already exists. */
1303 lp = lpni->lpni_peer_net->lpn_peer;
1304 lnet_peer_ni_decref_locked(lpni);
1306 * This is an error if the peer was configured and the
1307 * primary NID differs or an attempt is made to change
1308 * the Multi-Rail flag. Otherwise the assumption is
1309 * that an existing peer is being modified.
1311 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1312 if (lp->lp_primary_nid != nid)
1314 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1318 /* Delete and recreate as a configured peer. */
1322 /* Create peer, peer_net, and peer_ni. */
1324 lp = lnet_peer_alloc(nid);
1327 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1330 lpni = lnet_peer_ni_alloc(nid);
1334 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1337 LIBCFS_FREE(lpn, sizeof(*lpn));
1339 LIBCFS_FREE(lp, sizeof(*lp));
1341 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1342 libcfs_nid2str(nid), flags, rc);
1347 * Add a NID to a peer. Call with ln_api_mutex held.
1350 * -EPERM: Non-DLC addition to a DLC-configured peer.
1351 * -EEXIST: The NID was configured by DLC for a different peer.
1352 * -ENOMEM: Out of memory.
1353 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1354 * non-multi-rail peer.
1357 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1359 struct lnet_peer_net *lpn;
1360 struct lnet_peer_ni *lpni;
1364 LASSERT(nid != LNET_NID_ANY);
1366 /* A configured peer can only be updated through configuration. */
1367 if (!(flags & LNET_PEER_CONFIGURED)) {
1368 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1375 * The MULTI_RAIL flag can be set but not cleared, because
1376 * that would leave the peer struct in an invalid state.
1378 if (flags & LNET_PEER_MULTI_RAIL) {
1379 spin_lock(&lp->lp_lock);
1380 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1381 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1382 lnet_peer_clr_non_mr_pref_nids(lp);
1384 spin_unlock(&lp->lp_lock);
1385 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1390 lpni = lnet_find_peer_ni_locked(nid);
1393 * A peer_ni already exists. This is only a problem if
1394 * it is not connected to this peer and was configured
1397 lnet_peer_ni_decref_locked(lpni);
1398 if (lpni->lpni_peer_net->lpn_peer == lp)
1400 if (lnet_peer_ni_is_configured(lpni)) {
1404 /* If this is the primary NID, destroy the peer. */
1405 if (lnet_peer_ni_is_primary(lpni)) {
1406 struct lnet_peer *rtr_lp =
1407 lpni->lpni_peer_net->lpn_peer;
1408 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1410 * if we're trying to delete a router it means
1411 * we're moving this peer NI to a new peer so must
1412 * transfer router properties to the new peer
1414 if (rtr_refcount > 0) {
1415 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1416 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1418 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1419 lpni = lnet_peer_ni_alloc(nid);
1426 lpni = lnet_peer_ni_alloc(nid);
1434 * Get the peer_net. Check that we're not adding a second
1435 * peer_ni on a peer_net of a non-multi-rail peer.
1437 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1439 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1444 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1449 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1452 /* If the peer_ni was allocated above its peer_net pointer is NULL */
1453 if (!lpni->lpni_peer_net)
1454 LIBCFS_FREE(lpni, sizeof(*lpni));
1456 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1457 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1463 * Update the primary NID of a peer, if possible.
1465 * Call with the lnet_api_mutex held.
1468 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1470 lnet_nid_t old = lp->lp_primary_nid;
1473 if (lp->lp_primary_nid == nid)
1475 rc = lnet_peer_add_nid(lp, nid, flags);
1478 lp->lp_primary_nid = nid;
1480 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1481 libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1486 * lpni creation initiated due to traffic either sending or receiving.
1489 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1491 struct lnet_peer *lp;
1492 struct lnet_peer_net *lpn;
1493 struct lnet_peer_ni *lpni;
1495 * Assume peer is Multi-Rail capable and let discovery find out
1498 unsigned flags = LNET_PEER_MULTI_RAIL;
1501 if (nid == LNET_NID_ANY) {
1506 /* lnet_net_lock is not needed here because ln_api_lock is held */
1507 lpni = lnet_find_peer_ni_locked(nid);
1510 * We must have raced with another thread. Since we
1511 * know next to nothing about a peer_ni created by
1512 * traffic, we just assume everything is ok and
1515 lnet_peer_ni_decref_locked(lpni);
1519 /* Create peer, peer_net, and peer_ni. */
1521 lp = lnet_peer_alloc(nid);
1524 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1527 lpni = lnet_peer_ni_alloc(nid);
1530 if (pref != LNET_NID_ANY)
1531 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1533 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1536 LIBCFS_FREE(lpn, sizeof(*lpn));
1538 LIBCFS_FREE(lp, sizeof(*lp));
1540 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1545 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1547 * This API handles the following combinations:
1548 * Create a peer with its primary NI if only the prim_nid is provided
1549 * Add a NID to a peer identified by the prim_nid. The peer identified
1550 * by the prim_nid must already exist.
1551 * The peer being created may be non-MR.
1553 * The caller must hold ln_api_mutex. This prevents the peer from
1554 * being created/modified/deleted by a different thread.
1557 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1559 struct lnet_peer *lp = NULL;
1560 struct lnet_peer_ni *lpni;
1563 /* The prim_nid must always be specified */
1564 if (prim_nid == LNET_NID_ANY)
1567 flags = LNET_PEER_CONFIGURED;
1569 flags |= LNET_PEER_MULTI_RAIL;
1572 * If nid isn't specified, we must create a new peer with
1573 * prim_nid as its primary nid.
1575 if (nid == LNET_NID_ANY)
1576 return lnet_peer_add(prim_nid, flags);
1578 /* Look up the prim_nid, which must exist. */
1579 lpni = lnet_find_peer_ni_locked(prim_nid);
1582 lnet_peer_ni_decref_locked(lpni);
1583 lp = lpni->lpni_peer_net->lpn_peer;
1585 /* Peer must have been configured. */
1586 if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1587 CDEBUG(D_NET, "peer %s was not configured\n",
1588 libcfs_nid2str(prim_nid));
1592 /* Primary NID must match */
1593 if (lp->lp_primary_nid != prim_nid) {
1594 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1595 libcfs_nid2str(prim_nid),
1596 libcfs_nid2str(lp->lp_primary_nid));
1600 /* Multi-Rail flag must match. */
1601 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1602 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1603 libcfs_nid2str(prim_nid));
1607 return lnet_peer_add_nid(lp, nid, flags);
1611 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1613 * This API handles the following combinations:
1614 * Delete a NI from a peer if both prim_nid and nid are provided.
1615 * Delete a peer if only prim_nid is provided.
1616 * Delete a peer if its primary nid is provided.
1618 * The caller must hold ln_api_mutex. This prevents the peer from
1619 * being modified/deleted by a different thread.
1622 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1624 struct lnet_peer *lp;
1625 struct lnet_peer_ni *lpni;
1628 if (prim_nid == LNET_NID_ANY)
1631 lpni = lnet_find_peer_ni_locked(prim_nid);
1634 lnet_peer_ni_decref_locked(lpni);
1635 lp = lpni->lpni_peer_net->lpn_peer;
1637 if (prim_nid != lp->lp_primary_nid) {
1638 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1639 libcfs_nid2str(prim_nid),
1640 libcfs_nid2str(lp->lp_primary_nid));
1644 lnet_net_lock(LNET_LOCK_EX);
1645 if (lp->lp_rtr_refcount > 0) {
1646 lnet_net_unlock(LNET_LOCK_EX);
1647 CERROR("%s is a router. Can not be deleted\n",
1648 libcfs_nid2str(prim_nid));
1651 lnet_net_unlock(LNET_LOCK_EX);
1653 if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1654 return lnet_peer_del(lp);
1656 flags = LNET_PEER_CONFIGURED;
1657 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1658 flags |= LNET_PEER_MULTI_RAIL;
1660 return lnet_peer_del_nid(lp, nid, flags);
1664 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1666 struct lnet_peer_table *ptable;
1667 struct lnet_peer_net *lpn;
1669 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1671 LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1672 LASSERT(list_empty(&lpni->lpni_txq));
1673 LASSERT(lpni->lpni_txqnob == 0);
1674 LASSERT(list_empty(&lpni->lpni_peer_nis));
1675 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1677 lpn = lpni->lpni_peer_net;
1678 lpni->lpni_peer_net = NULL;
1679 lpni->lpni_net = NULL;
1681 /* remove the peer ni from the zombie list */
1682 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1683 spin_lock(&ptable->pt_zombie_lock);
1684 list_del_init(&lpni->lpni_hashlist);
1685 ptable->pt_zombies--;
1686 spin_unlock(&ptable->pt_zombie_lock);
1688 if (lpni->lpni_pref_nnids > 1) {
1689 LIBCFS_FREE(lpni->lpni_pref.nids,
1690 sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1692 LIBCFS_FREE(lpni, sizeof(*lpni));
1694 lnet_peer_net_decref_locked(lpn);
1697 struct lnet_peer_ni *
1698 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1700 struct lnet_peer_ni *lpni = NULL;
1703 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1704 return ERR_PTR(-ESHUTDOWN);
1707 * find if a peer_ni already exists.
1708 * If so then just return that.
1710 lpni = lnet_find_peer_ni_locked(nid);
1714 lnet_net_unlock(cpt);
1716 rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1719 goto out_net_relock;
1722 lpni = lnet_find_peer_ni_locked(nid);
1732 * Get a peer_ni for the given nid, create it if necessary. Takes a
1733 * hold on the peer_ni.
1735 struct lnet_peer_ni *
1736 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1738 struct lnet_peer_ni *lpni = NULL;
1741 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1742 return ERR_PTR(-ESHUTDOWN);
1745 * find if a peer_ni already exists.
1746 * If so then just return that.
1748 lpni = lnet_find_peer_ni_locked(nid);
1754 * use the lnet_api_mutex to serialize the creation of the peer_ni
1755 * and the creation/deletion of the local ni/net. When a local ni is
1756 * created, if there exists a set of peer_nis on that network,
1757 * they need to be traversed and updated. When a local NI is
1758 * deleted, which could result in a network being deleted, then
1759 * all peer nis on that network need to be removed as well.
1761 * Creation through traffic should also be serialized with
1762 * creation through DLC.
1764 lnet_net_unlock(cpt);
1765 mutex_lock(&the_lnet.ln_api_mutex);
1767 * Shutdown is only set under the ln_api_lock, so a single
1768 * check here is sufficent.
1770 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1771 lpni = ERR_PTR(-ESHUTDOWN);
1772 goto out_mutex_unlock;
1775 rc = lnet_peer_ni_traffic_add(nid, pref);
1778 goto out_mutex_unlock;
1781 lpni = lnet_find_peer_ni_locked(nid);
1785 mutex_unlock(&the_lnet.ln_api_mutex);
1788 /* Lock has been dropped, check again for shutdown. */
1789 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1791 lnet_peer_ni_decref_locked(lpni);
1792 lpni = ERR_PTR(-ESHUTDOWN);
1799 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1801 if (lnet_peer_discovery_disabled)
1804 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1805 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1816 lnet_is_discovery_disabled(struct lnet_peer *lp)
1820 spin_lock(&lp->lp_lock);
1821 rc = lnet_is_discovery_disabled_locked(lp);
1822 spin_unlock(&lp->lp_lock);
1828 lnet_peer_gw_discovery(struct lnet_peer *lp)
1832 spin_lock(&lp->lp_lock);
1833 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1835 spin_unlock(&lp->lp_lock);
1841 * Is a peer uptodate from the point of view of discovery?
1843 * If it is currently being processed, obviously not.
1844 * A forced Ping or Push is also handled by the discovery thread.
1846 * Otherwise look at whether the peer needs rediscovering.
1849 lnet_peer_is_uptodate(struct lnet_peer *lp)
1853 spin_lock(&lp->lp_lock);
1854 if (lp->lp_state & (LNET_PEER_DISCOVERING |
1855 LNET_PEER_FORCE_PING |
1856 LNET_PEER_FORCE_PUSH)) {
1858 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1860 } else if (lnet_peer_needs_push(lp)) {
1862 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1863 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1870 spin_unlock(&lp->lp_lock);
1876 * Queue a peer for the attention of the discovery thread. Call with
1877 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1878 * -EALREADY if the peer was already queued.
1880 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1884 spin_lock(&lp->lp_lock);
1885 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1886 lp->lp_state |= LNET_PEER_DISCOVERING;
1887 spin_unlock(&lp->lp_lock);
1888 if (list_empty(&lp->lp_dc_list)) {
1889 lnet_peer_addref_locked(lp);
1890 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1891 wake_up(&the_lnet.ln_dc_waitq);
1897 CDEBUG(D_NET, "Queue peer %s: %d\n",
1898 libcfs_nid2str(lp->lp_primary_nid), rc);
1904 * Discovery of a peer is complete. Wake all waiters on the peer.
1905 * Call with lnet_net_lock/EX held.
1907 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1909 struct lnet_msg *msg, *tmp;
1911 struct list_head pending_msgs;
1913 INIT_LIST_HEAD(&pending_msgs);
1915 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1916 libcfs_nid2str(lp->lp_primary_nid));
1918 list_del_init(&lp->lp_dc_list);
1919 spin_lock(&lp->lp_lock);
1920 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1921 spin_unlock(&lp->lp_lock);
1922 wake_up_all(&lp->lp_dc_waitq);
1924 if (lp->lp_rtr_refcount > 0)
1925 lnet_router_discovery_complete(lp);
1927 lnet_net_unlock(LNET_LOCK_EX);
1929 /* iterate through all pending messages and send them again */
1930 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1931 list_del_init(&msg->msg_list);
1932 if (lp->lp_dc_error) {
1933 lnet_finalize(msg, lp->lp_dc_error);
1937 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1938 lnet_msgtyp2str(msg->msg_type),
1939 libcfs_id2str(msg->msg_target));
1940 rc = lnet_send(msg->msg_src_nid_param, msg,
1941 msg->msg_rtr_nid_param);
1943 CNETERR("Error sending %s to %s: %d\n",
1944 lnet_msgtyp2str(msg->msg_type),
1945 libcfs_id2str(msg->msg_target), rc);
1946 lnet_finalize(msg, rc);
1949 lnet_net_lock(LNET_LOCK_EX);
1950 lnet_peer_decref_locked(lp);
1954 * Handle inbound push.
1955 * Like any event handler, called with lnet_res_lock/CPT held.
1957 void lnet_peer_push_event(struct lnet_event *ev)
1959 struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1960 struct lnet_peer *lp;
1962 /* lnet_find_peer() adds a refcount */
1963 lp = lnet_find_peer(ev->source.nid);
1965 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1966 libcfs_nid2str(ev->initiator.nid),
1967 libcfs_nid2str(ev->source.nid));
1971 /* Ensure peer state remains consistent while we modify it. */
1972 spin_lock(&lp->lp_lock);
1975 * If some kind of error happened the contents of the message
1976 * cannot be used. Clear the NIDS_UPTODATE and set the
1977 * FORCE_PING flag to trigger a ping.
1980 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1981 lp->lp_state |= LNET_PEER_FORCE_PING;
1982 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1984 libcfs_nid2str(lp->lp_primary_nid),
1985 libcfs_nid2str(ev->source.nid));
1990 * A push with invalid or corrupted info. Clear the UPTODATE
1991 * flag to trigger a ping.
1993 if (lnet_ping_info_validate(&pbuf->pb_info)) {
1994 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1995 lp->lp_state |= LNET_PEER_FORCE_PING;
1996 CDEBUG(D_NET, "Corrupted Push from %s\n",
1997 libcfs_nid2str(lp->lp_primary_nid));
2002 * Make sure we'll allocate the correct size ping buffer when
2005 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2006 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2009 * A non-Multi-Rail peer is not supposed to be capable of
2012 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2013 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2014 libcfs_nid2str(lp->lp_primary_nid));
2019 * Check the MULTIRAIL flag. Complain if the peer was DLC
2020 * configured without it.
2022 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2023 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2024 CERROR("Push says %s is Multi-Rail, DLC says not\n",
2025 libcfs_nid2str(lp->lp_primary_nid));
2027 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2028 lnet_peer_clr_non_mr_pref_nids(lp);
2033 * The peer may have discovery disabled at its end. Set
2034 * NO_DISCOVERY as appropriate.
2036 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2037 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2038 libcfs_nid2str(lp->lp_primary_nid));
2039 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2040 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2041 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2042 libcfs_nid2str(lp->lp_primary_nid));
2043 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2047 * Check for truncation of the Put message. Clear the
2048 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2049 * and tell discovery to allocate a bigger buffer.
2051 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2052 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2053 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2054 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2055 lp->lp_state |= LNET_PEER_FORCE_PING;
2056 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2057 libcfs_nid2str(lp->lp_primary_nid),
2058 pbuf->pb_info.pi_nnis);
2062 /* always assume new data */
2063 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2064 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2067 * If there is data present that hasn't been processed yet,
2068 * we'll replace it if the Put contained newer data and it
2069 * fits. We're racing with a Ping or earlier Push in this
2072 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2073 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2074 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2075 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2076 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2077 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2078 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2079 libcfs_nid2str(lp->lp_primary_nid),
2080 LNET_PING_BUFFER_SEQNO(pbuf),
2081 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2087 * Allocate a buffer to copy the data. On a failure we drop
2088 * the Push and set FORCE_PING to force the discovery
2089 * thread to fix the problem by pinging the peer.
2091 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2093 lp->lp_state |= LNET_PEER_FORCE_PING;
2094 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2095 libcfs_nid2str(lp->lp_primary_nid),
2096 LNET_PING_BUFFER_SEQNO(pbuf));
2101 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2102 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2103 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2104 CDEBUG(D_NET, "Received Push %s %u\n",
2105 libcfs_nid2str(lp->lp_primary_nid),
2106 LNET_PING_BUFFER_SEQNO(pbuf));
2110 * Queue the peer for discovery if not done, force it on the request
2111 * queue and wake the discovery thread if the peer was already queued,
2112 * because its status changed.
2114 spin_unlock(&lp->lp_lock);
2115 lnet_net_lock(LNET_LOCK_EX);
2116 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2117 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2118 wake_up(&the_lnet.ln_dc_waitq);
2120 /* Drop refcount from lookup */
2121 lnet_peer_decref_locked(lp);
2122 lnet_net_unlock(LNET_LOCK_EX);
2126 * Clear the discovery error state, unless we're already discovering
2127 * this peer, in which case the error is current.
2129 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2131 spin_lock(&lp->lp_lock);
2132 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2133 lp->lp_dc_error = 0;
2134 spin_unlock(&lp->lp_lock);
2138 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2139 * dropped/retaken within this function. An lnet_peer_ni is passed in
2140 * because discovery could tear down an lnet_peer.
2143 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2146 struct lnet_peer *lp;
2150 lnet_net_unlock(cpt);
2151 lnet_net_lock(LNET_LOCK_EX);
2152 lp = lpni->lpni_peer_net->lpn_peer;
2153 lnet_peer_clear_discovery_error(lp);
2156 * We're willing to be interrupted. The lpni can become a
2157 * zombie if we race with DLC, so we must check for that.
2160 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2161 if (signal_pending(current))
2163 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2165 if (lp->lp_dc_error)
2167 if (lnet_peer_is_uptodate(lp))
2169 lnet_peer_queue_for_discovery(lp);
2171 if (lnet_is_discovery_disabled(lp))
2174 * if caller requested a non-blocking operation then
2175 * return immediately. Once discovery is complete then the
2176 * peer ref will be decremented and any pending messages
2177 * that were stopped due to discovery will be transmitted.
2182 lnet_peer_addref_locked(lp);
2183 lnet_net_unlock(LNET_LOCK_EX);
2185 finish_wait(&lp->lp_dc_waitq, &wait);
2186 lnet_net_lock(LNET_LOCK_EX);
2187 lnet_peer_decref_locked(lp);
2188 /* Peer may have changed */
2189 lp = lpni->lpni_peer_net->lpn_peer;
2191 finish_wait(&lp->lp_dc_waitq, &wait);
2193 lnet_net_unlock(LNET_LOCK_EX);
2197 * If the peer has changed after we've discovered the older peer,
2198 * then we need to discovery the new peer to make sure the
2199 * interface information is up to date
2201 if (lp != lpni->lpni_peer_net->lpn_peer)
2204 if (signal_pending(current))
2206 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2208 else if (lp->lp_dc_error)
2209 rc = lp->lp_dc_error;
2211 CDEBUG(D_NET, "non-blocking discovery\n");
2212 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2215 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2216 (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2217 libcfs_nid2str(lpni->lpni_nid), rc,
2218 (!block) ? "pending discovery" : "discovery complete");
2223 /* Handle an incoming ack for a push. */
2225 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2227 struct lnet_ping_buffer *pbuf;
2229 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2230 spin_lock(&lp->lp_lock);
2231 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2232 lp->lp_push_error = ev->status;
2234 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2236 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2237 spin_unlock(&lp->lp_lock);
2239 CDEBUG(D_NET, "peer %s ev->status %d\n",
2240 libcfs_nid2str(lp->lp_primary_nid), ev->status);
2243 /* Handle a Reply message. This is the reply to a Ping message. */
2245 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2247 struct lnet_ping_buffer *pbuf;
2250 spin_lock(&lp->lp_lock);
2253 * If some kind of error happened the contents of message
2254 * cannot be used. Set PING_FAILED to trigger a retry.
2257 lp->lp_state |= LNET_PEER_PING_FAILED;
2258 lp->lp_ping_error = ev->status;
2259 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2261 libcfs_nid2str(lp->lp_primary_nid),
2262 libcfs_nid2str(ev->source.nid));
2266 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2267 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2268 lnet_swap_pinginfo(pbuf);
2271 * A reply with invalid or corrupted info. Set PING_FAILED to
2274 rc = lnet_ping_info_validate(&pbuf->pb_info);
2276 lp->lp_state |= LNET_PEER_PING_FAILED;
2277 lp->lp_ping_error = 0;
2278 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2279 libcfs_nid2str(lp->lp_primary_nid), rc);
2285 * Only enable the multi-rail feature on the peer if both sides of
2286 * the connection have discovery on
2288 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2289 CDEBUG(D_NET, "Peer %s has Multi-Rail feature enabled\n",
2290 libcfs_nid2str(lp->lp_primary_nid));
2291 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2293 CDEBUG(D_NET, "Peer %s has Multi-Rail feature disabled\n",
2294 libcfs_nid2str(lp->lp_primary_nid));
2295 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2299 * The peer may have discovery disabled at its end. Set
2300 * NO_DISCOVERY as appropriate.
2302 if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2303 !lnet_peer_discovery_disabled) {
2304 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2305 libcfs_nid2str(lp->lp_primary_nid));
2306 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2308 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2309 libcfs_nid2str(lp->lp_primary_nid));
2310 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2314 * Update the MULTI_RAIL flag based on the reply. If the peer
2315 * was configured with DLC then the setting should match what
2318 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2319 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2320 /* Everything's fine */
2321 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2322 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2323 libcfs_nid2str(lp->lp_primary_nid));
2326 * if discovery is disabled then we don't want to
2327 * update the state of the peer. All we'll do is
2328 * update the peer_nis which were reported back in
2332 if (!lnet_is_discovery_disabled_locked(lp)) {
2333 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2334 lnet_peer_clr_non_mr_pref_nids(lp);
2337 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2338 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2339 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2340 libcfs_nid2str(lp->lp_primary_nid));
2342 CERROR("Multi-Rail state vanished from %s\n",
2343 libcfs_nid2str(lp->lp_primary_nid));
2344 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2349 * Make sure we'll allocate the correct size ping buffer when
2352 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2353 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2356 * Check for truncation of the Reply. Clear PING_SENT and set
2357 * PING_FAILED to trigger a retry.
2359 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2360 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2361 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2362 lp->lp_state |= LNET_PEER_PING_FAILED;
2363 lp->lp_ping_error = 0;
2364 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2365 libcfs_nid2str(lp->lp_primary_nid),
2366 pbuf->pb_info.pi_nnis);
2371 * Check the sequence numbers in the reply. These are only
2372 * available if the reply came from a Multi-Rail peer.
2374 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2375 pbuf->pb_info.pi_nnis > 1 &&
2376 lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2377 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2378 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2379 libcfs_nid2str(lp->lp_primary_nid),
2380 LNET_PING_BUFFER_SEQNO(pbuf),
2383 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2386 /* We're happy with the state of the data in the buffer. */
2387 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2388 libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2389 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2390 lnet_ping_buffer_decref(lp->lp_data);
2392 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2393 lnet_ping_buffer_addref(pbuf);
2396 lp->lp_state &= ~LNET_PEER_PING_SENT;
2397 spin_unlock(&lp->lp_lock);
2401 * Send event handling. Only matters for error cases, where we clean
2402 * up state on the peer and peer_ni that would otherwise be updated in
2403 * the REPLY event handler for a successful Ping, and the ACK event
2404 * handler for a successful Push.
2407 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2414 spin_lock(&lp->lp_lock);
2415 if (ev->msg_type == LNET_MSG_GET) {
2416 lp->lp_state &= ~LNET_PEER_PING_SENT;
2417 lp->lp_state |= LNET_PEER_PING_FAILED;
2418 lp->lp_ping_error = ev->status;
2419 } else { /* ev->msg_type == LNET_MSG_PUT */
2420 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2421 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2422 lp->lp_push_error = ev->status;
2424 spin_unlock(&lp->lp_lock);
2425 rc = LNET_REDISCOVER_PEER;
2427 CDEBUG(D_NET, "%s Send to %s: %d\n",
2428 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2429 libcfs_nid2str(ev->target.nid), rc);
2434 * Unlink event handling. This event is only seen if a call to
2435 * LNetMDUnlink() caused the event to be unlinked. If this call was
2436 * made after the event was set up in LNetGet() or LNetPut() then we
2437 * assume the Ping or Push timed out.
2440 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2442 spin_lock(&lp->lp_lock);
2443 /* We've passed through LNetGet() */
2444 if (lp->lp_state & LNET_PEER_PING_SENT) {
2445 lp->lp_state &= ~LNET_PEER_PING_SENT;
2446 lp->lp_state |= LNET_PEER_PING_FAILED;
2447 lp->lp_ping_error = -ETIMEDOUT;
2448 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2449 libcfs_nid2str(lp->lp_primary_nid));
2451 /* We've passed through LNetPut() */
2452 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2453 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2454 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2455 lp->lp_push_error = -ETIMEDOUT;
2456 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2457 libcfs_nid2str(lp->lp_primary_nid));
2459 spin_unlock(&lp->lp_lock);
2463 * Event handler for the discovery EQ.
2465 * Called with lnet_res_lock(cpt) held. The cpt is the
2466 * lnet_cpt_of_cookie() of the md handle cookie.
2468 static void lnet_discovery_event_handler(struct lnet_event *event)
2470 struct lnet_peer *lp = event->md.user_ptr;
2471 struct lnet_ping_buffer *pbuf;
2474 /* discovery needs to take another look */
2475 rc = LNET_REDISCOVER_PEER;
2477 CDEBUG(D_NET, "Received event: %d\n", event->type);
2479 switch (event->type) {
2480 case LNET_EVENT_ACK:
2481 lnet_discovery_event_ack(lp, event);
2483 case LNET_EVENT_REPLY:
2484 lnet_discovery_event_reply(lp, event);
2486 case LNET_EVENT_SEND:
2487 /* Only send failure triggers a retry. */
2488 rc = lnet_discovery_event_send(lp, event);
2490 case LNET_EVENT_UNLINK:
2491 /* LNetMDUnlink() was called */
2492 lnet_discovery_event_unlink(lp, event);
2495 /* Invalid events. */
2498 lnet_net_lock(LNET_LOCK_EX);
2499 if (event->unlinked) {
2500 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2501 lnet_ping_buffer_decref(pbuf);
2502 lnet_peer_decref_locked(lp);
2505 /* put peer back at end of request queue, if discovery not already
2507 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2508 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2509 wake_up(&the_lnet.ln_dc_waitq);
2511 lnet_net_unlock(LNET_LOCK_EX);
2515 * Build a peer from incoming data.
2517 * The NIDs in the incoming data are supposed to be structured as follows:
2520 * - other NIDs in same net
2521 * - NIDs in second net
2522 * - NIDs in third net
2524 * This due to the way the list of NIDs in the data is created.
2526 * Note that this function will mark the peer uptodate unless an
2527 * ENOMEM is encontered. All other errors are due to a conflict
2528 * between the DLC configuration and what discovery sees. We treat DLC
2529 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2530 * peer from becoming stuck in discovery.
2532 static int lnet_peer_merge_data(struct lnet_peer *lp,
2533 struct lnet_ping_buffer *pbuf)
2535 struct lnet_peer_ni *lpni;
2536 lnet_nid_t *curnis = NULL;
2537 struct lnet_ni_status *addnis = NULL;
2538 lnet_nid_t *delnis = NULL;
2548 flags = LNET_PEER_DISCOVERED;
2549 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2550 flags |= LNET_PEER_MULTI_RAIL;
2553 * Cache the routing feature for the peer; whether it is enabled
2554 * for disabled as reported by the remote peer.
2556 spin_lock(&lp->lp_lock);
2557 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2558 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2560 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2561 spin_unlock(&lp->lp_lock);
2563 nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2564 LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2565 LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2566 LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2567 if (!curnis || !addnis || !delnis) {
2575 /* Construct the list of NIDs present in peer. */
2577 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2578 curnis[ncurnis++] = lpni->lpni_nid;
2581 * Check for NIDs in pbuf not present in curnis[].
2582 * The loop starts at 1 to skip the loopback NID.
2584 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2585 for (j = 0; j < ncurnis; j++)
2586 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2589 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2592 * Check for NIDs in curnis[] not present in pbuf.
2593 * The nested loop starts at 1 to skip the loopback NID.
2595 * But never add the loopback NID to delnis[]: if it is
2596 * present in curnis[] then this peer is for this node.
2598 for (i = 0; i < ncurnis; i++) {
2599 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2601 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2602 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2604 * update the information we cache for the
2605 * peer with the latest information we
2608 lpni = lnet_find_peer_ni_locked(curnis[i]);
2610 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2611 lnet_peer_ni_decref_locked(lpni);
2616 if (j == pbuf->pb_info.pi_nnis)
2617 delnis[ndelnis++] = curnis[i];
2621 * If we get here and the discovery is disabled then we don't want
2622 * to add or delete any NIs. We just updated the ones we have some
2623 * information on, and call it a day
2626 if (lnet_is_discovery_disabled(lp))
2629 for (i = 0; i < naddnis; i++) {
2630 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2632 CERROR("Error adding NID %s to peer %s: %d\n",
2633 libcfs_nid2str(addnis[i].ns_nid),
2634 libcfs_nid2str(lp->lp_primary_nid), rc);
2638 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2640 lpni->lpni_ns_status = addnis[i].ns_status;
2641 lnet_peer_ni_decref_locked(lpni);
2645 for (i = 0; i < ndelnis; i++) {
2647 * for routers it's okay to delete the primary_nid because
2648 * the upper layers don't really rely on it. So if we're
2649 * being told that the router changed its primary_nid
2650 * then it's okay to delete it.
2652 if (lp->lp_rtr_refcount > 0)
2653 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2654 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2656 CERROR("Error deleting NID %s from peer %s: %d\n",
2657 libcfs_nid2str(delnis[i]),
2658 libcfs_nid2str(lp->lp_primary_nid), rc);
2664 * Errors other than -ENOMEM are due to peers having been
2665 * configured with DLC. Ignore these because DLC overrides
2670 LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2671 LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2672 LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2673 lnet_ping_buffer_decref(pbuf);
2674 CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2677 spin_lock(&lp->lp_lock);
2678 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2679 lp->lp_state |= LNET_PEER_FORCE_PING;
2680 spin_unlock(&lp->lp_lock);
2686 * The data in pbuf says lp is its primary peer, but the data was
2687 * received by a different peer. Try to update lp with the data.
2690 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2692 struct lnet_handle_md mdh;
2694 /* Queue lp for discovery, and force it on the request queue. */
2695 lnet_net_lock(LNET_LOCK_EX);
2696 if (lnet_peer_queue_for_discovery(lp))
2697 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2698 lnet_net_unlock(LNET_LOCK_EX);
2700 LNetInvalidateMDHandle(&mdh);
2703 * Decide whether we can move the peer to the DATA_PRESENT state.
2705 * We replace stale data for a multi-rail peer, repair PING_FAILED
2706 * status, and preempt FORCE_PING.
2708 * If after that we have DATA_PRESENT, we merge it into this peer.
2710 spin_lock(&lp->lp_lock);
2711 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2712 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2713 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2714 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2715 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2716 lnet_ping_buffer_decref(pbuf);
2721 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2722 lnet_ping_buffer_decref(lp->lp_data);
2724 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2726 if (lp->lp_state & LNET_PEER_PING_FAILED) {
2727 mdh = lp->lp_ping_mdh;
2728 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2729 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2730 lp->lp_ping_error = 0;
2732 if (lp->lp_state & LNET_PEER_FORCE_PING)
2733 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2734 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2735 spin_unlock(&lp->lp_lock);
2737 if (!LNetMDHandleIsInvalid(mdh))
2741 return lnet_peer_merge_data(lp, pbuf);
2743 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2747 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2751 for (i = 0; i < pinfo->pi_nnis; i++) {
2752 if (pinfo->pi_ni[i].ns_nid == nid)
2760 * Update a peer using the data received.
2762 static int lnet_peer_data_present(struct lnet_peer *lp)
2763 __must_hold(&lp->lp_lock)
2765 struct lnet_ping_buffer *pbuf;
2766 struct lnet_peer_ni *lpni;
2767 lnet_nid_t nid = LNET_NID_ANY;
2773 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2774 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2775 spin_unlock(&lp->lp_lock);
2778 * Modifications of peer structures are done while holding the
2779 * ln_api_mutex. A global lock is required because we may be
2780 * modifying multiple peer structures, and a mutex greatly
2781 * simplifies memory management.
2783 * The actual changes to the data structures must also protect
2784 * against concurrent lookups, for which the lnet_net_lock in
2785 * LNET_LOCK_EX mode is used.
2787 mutex_lock(&the_lnet.ln_api_mutex);
2788 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2794 * If this peer is not on the peer list then it is being torn
2795 * down, and our reference count may be all that is keeping it
2796 * alive. Don't do any work on it.
2798 if (list_empty(&lp->lp_peer_list))
2801 flags = LNET_PEER_DISCOVERED;
2802 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2803 flags |= LNET_PEER_MULTI_RAIL;
2806 * Check whether the primary NID in the message matches the
2807 * primary NID of the peer. If it does, update the peer, if
2808 * it it does not, check whether there is already a peer with
2809 * that primary NID. If no such peer exists, try to update
2810 * the primary NID of the current peer (allowed if it was
2811 * created due to message traffic) and complete the update.
2812 * If the peer did exist, hand off the data to it.
2814 * The peer for the loopback interface is a special case: this
2815 * is the peer for the local node, and we want to set its
2816 * primary NID to the correct value here. Moreover, this peer
2817 * can show up with only the loopback NID in the ping buffer.
2819 if (pbuf->pb_info.pi_nnis <= 1)
2821 nid = pbuf->pb_info.pi_ni[1].ns_nid;
2822 if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2823 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2825 rc = lnet_peer_merge_data(lp, pbuf);
2827 * if the primary nid of the peer is present in the ping info returned
2828 * from the peer, but it's not the local primary peer we have
2829 * cached and discovery is disabled, then we don't want to update
2830 * our local peer info, by adding or removing NIDs, we just want
2831 * to update the status of the nids that we currently have
2832 * recorded in that peer.
2834 } else if (lp->lp_primary_nid == nid ||
2835 (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2836 lnet_is_discovery_disabled(lp))) {
2837 rc = lnet_peer_merge_data(lp, pbuf);
2839 lpni = lnet_find_peer_ni_locked(nid);
2841 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2843 CERROR("Primary NID error %s versus %s: %d\n",
2844 libcfs_nid2str(lp->lp_primary_nid),
2845 libcfs_nid2str(nid), rc);
2847 rc = lnet_peer_merge_data(lp, pbuf);
2850 struct lnet_peer *new_lp;
2851 new_lp = lpni->lpni_peer_net->lpn_peer;
2853 * if lp has discovery/MR enabled that means new_lp
2854 * should have discovery/MR enabled as well, since
2855 * it's the same peer, which we're about to merge
2857 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2858 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2859 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2860 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2862 rc = lnet_peer_set_primary_data(new_lp, pbuf);
2863 lnet_consolidate_routes_locked(lp, new_lp);
2864 lnet_peer_ni_decref_locked(lpni);
2868 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2870 mutex_unlock(&the_lnet.ln_api_mutex);
2872 spin_lock(&lp->lp_lock);
2873 /* Tell discovery to re-check the peer immediately. */
2875 rc = LNET_REDISCOVER_PEER;
2880 * A ping failed. Clear the PING_FAILED state and set the
2881 * FORCE_PING state, to ensure a retry even if discovery is
2882 * disabled. This avoids being left with incorrect state.
2884 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2885 __must_hold(&lp->lp_lock)
2887 struct lnet_handle_md mdh;
2890 mdh = lp->lp_ping_mdh;
2891 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2892 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2893 lp->lp_state |= LNET_PEER_FORCE_PING;
2894 rc = lp->lp_ping_error;
2895 lp->lp_ping_error = 0;
2896 spin_unlock(&lp->lp_lock);
2898 if (!LNetMDHandleIsInvalid(mdh))
2901 CDEBUG(D_NET, "peer %s:%d\n",
2902 libcfs_nid2str(lp->lp_primary_nid), rc);
2904 spin_lock(&lp->lp_lock);
2905 return rc ? rc : LNET_REDISCOVER_PEER;
2909 * Select NID to send a Ping or Push to.
2911 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2913 struct lnet_peer_ni *lpni;
2915 /* Look for a direct-connected NID for this peer. */
2917 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2918 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2923 return lpni->lpni_nid;
2925 /* Look for a routed-connected NID for this peer. */
2927 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2928 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2933 return lpni->lpni_nid;
2935 return LNET_NID_ANY;
2938 /* Active side of ping. */
2939 static int lnet_peer_send_ping(struct lnet_peer *lp)
2940 __must_hold(&lp->lp_lock)
2947 lp->lp_state |= LNET_PEER_PING_SENT;
2948 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2949 spin_unlock(&lp->lp_lock);
2951 cpt = lnet_net_lock_current();
2952 /* Refcount for MD. */
2953 lnet_peer_addref_locked(lp);
2954 pnid = lnet_peer_select_nid(lp);
2955 lnet_net_unlock(cpt);
2957 nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2959 rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2960 the_lnet.ln_dc_eqh, false);
2963 * if LNetMDBind in lnet_send_ping fails we need to decrement the
2964 * refcount on the peer, otherwise LNetMDUnlink will be called
2965 * which will eventually do that.
2969 lnet_peer_decref_locked(lp);
2970 lnet_net_unlock(cpt);
2971 rc = -rc; /* change the rc to negative value */
2973 } else if (rc < 0) {
2977 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2979 spin_lock(&lp->lp_lock);
2983 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2985 * The errors that get us here are considered hard errors and
2986 * cause Discovery to terminate. So we clear PING_SENT, but do
2987 * not set either PING_FAILED or FORCE_PING. In fact we need
2988 * to clear PING_FAILED, because the unlink event handler will
2989 * have set it if we called LNetMDUnlink() above.
2991 spin_lock(&lp->lp_lock);
2992 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
2997 * This function exists because you cannot call LNetMDUnlink() from an
3000 static int lnet_peer_push_failed(struct lnet_peer *lp)
3001 __must_hold(&lp->lp_lock)
3003 struct lnet_handle_md mdh;
3006 mdh = lp->lp_push_mdh;
3007 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3008 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3009 rc = lp->lp_push_error;
3010 lp->lp_push_error = 0;
3011 spin_unlock(&lp->lp_lock);
3013 if (!LNetMDHandleIsInvalid(mdh))
3016 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3017 spin_lock(&lp->lp_lock);
3018 return rc ? rc : LNET_REDISCOVER_PEER;
3021 /* Active side of push. */
3022 static int lnet_peer_send_push(struct lnet_peer *lp)
3023 __must_hold(&lp->lp_lock)
3025 struct lnet_ping_buffer *pbuf;
3026 struct lnet_process_id id;
3031 /* Don't push to a non-multi-rail peer. */
3032 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3033 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3037 lp->lp_state |= LNET_PEER_PUSH_SENT;
3038 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3039 spin_unlock(&lp->lp_lock);
3041 cpt = lnet_net_lock_current();
3042 pbuf = the_lnet.ln_ping_target;
3043 lnet_ping_buffer_addref(pbuf);
3044 lnet_net_unlock(cpt);
3046 /* Push source MD */
3047 md.start = &pbuf->pb_info;
3048 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3049 md.threshold = 2; /* Put/Ack */
3052 md.eq_handle = the_lnet.ln_dc_eqh;
3055 rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3057 lnet_ping_buffer_decref(pbuf);
3058 CERROR("Can't bind push source MD: %d\n", rc);
3061 cpt = lnet_net_lock_current();
3062 /* Refcount for MD. */
3063 lnet_peer_addref_locked(lp);
3064 id.pid = LNET_PID_LUSTRE;
3065 id.nid = lnet_peer_select_nid(lp);
3066 lnet_net_unlock(cpt);
3068 if (id.nid == LNET_NID_ANY) {
3073 rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3074 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3075 LNET_PROTO_PING_MATCHBITS, 0, 0);
3080 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3082 spin_lock(&lp->lp_lock);
3086 LNetMDUnlink(lp->lp_push_mdh);
3087 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3089 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3091 * The errors that get us here are considered hard errors and
3092 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3093 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3094 * because the unlink event handler will have set it if we
3095 * called LNetMDUnlink() above.
3097 spin_lock(&lp->lp_lock);
3098 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3103 * An unrecoverable error was encountered during discovery.
3104 * Set error status in peer and abort discovery.
3106 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3108 CDEBUG(D_NET, "Discovery error %s: %d\n",
3109 libcfs_nid2str(lp->lp_primary_nid), error);
3111 spin_lock(&lp->lp_lock);
3112 lp->lp_dc_error = error;
3113 lp->lp_state &= ~LNET_PEER_DISCOVERING;
3114 lp->lp_state |= LNET_PEER_REDISCOVER;
3115 spin_unlock(&lp->lp_lock);
3119 * Mark the peer as discovered.
3121 static int lnet_peer_discovered(struct lnet_peer *lp)
3122 __must_hold(&lp->lp_lock)
3124 lp->lp_state |= LNET_PEER_DISCOVERED;
3125 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3126 LNET_PEER_REDISCOVER);
3128 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3135 * Discovering this peer is taking too long. Cancel any Ping or Push
3136 * that discovery is waiting on by unlinking the relevant MDs. The
3137 * lnet_discovery_event_handler() will proceed from here and complete
3140 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3142 struct lnet_handle_md ping_mdh;
3143 struct lnet_handle_md push_mdh;
3145 LNetInvalidateMDHandle(&ping_mdh);
3146 LNetInvalidateMDHandle(&push_mdh);
3148 spin_lock(&lp->lp_lock);
3149 if (lp->lp_state & LNET_PEER_PING_SENT) {
3150 ping_mdh = lp->lp_ping_mdh;
3151 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3153 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3154 push_mdh = lp->lp_push_mdh;
3155 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3157 spin_unlock(&lp->lp_lock);
3159 if (!LNetMDHandleIsInvalid(ping_mdh))
3160 LNetMDUnlink(ping_mdh);
3161 if (!LNetMDHandleIsInvalid(push_mdh))
3162 LNetMDUnlink(push_mdh);
3166 * Wait for work to be queued or some other change that must be
3167 * attended to. Returns non-zero if the discovery thread should shut
3170 static int lnet_peer_discovery_wait_for_work(void)
3177 cpt = lnet_net_lock_current();
3179 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3180 TASK_INTERRUPTIBLE);
3181 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3183 if (lnet_push_target_resize_needed())
3185 if (!list_empty(&the_lnet.ln_dc_request))
3187 if (!list_empty(&the_lnet.ln_msg_resend))
3189 lnet_net_unlock(cpt);
3192 * wakeup max every second to check if there are peers that
3193 * have been stuck on the working queue for greater than
3196 schedule_timeout(cfs_time_seconds(1));
3197 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3198 cpt = lnet_net_lock_current();
3200 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3202 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3205 lnet_net_unlock(cpt);
3207 CDEBUG(D_NET, "woken: %d\n", rc);
3213 * Messages that were pending on a destroyed peer will be put on a global
3214 * resend list. The message resend list will be checked by
3215 * the discovery thread when it wakes up, and will resend messages. These
3216 * messages can still be sendable in the case the lpni which was the initial
3217 * cause of the message re-queue was transfered to another peer.
3219 * It is possible that LNet could be shutdown while we're iterating
3220 * through the list. lnet_shudown_lndnets() will attempt to access the
3221 * resend list, but will have to wait until the spinlock is released, by
3222 * which time there shouldn't be any more messages on the resend list.
3223 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3224 * for the messages so they can be released. The other case is that
3225 * lnet_shudown_lndnets() can finalize all the messages before this
3226 * function can visit the resend list, in which case this function will be
3229 static void lnet_resend_msgs(void)
3231 struct lnet_msg *msg, *tmp;
3232 struct list_head resend;
3235 INIT_LIST_HEAD(&resend);
3237 spin_lock(&the_lnet.ln_msg_resend_lock);
3238 list_splice(&the_lnet.ln_msg_resend, &resend);
3239 spin_unlock(&the_lnet.ln_msg_resend_lock);
3241 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3242 list_del_init(&msg->msg_list);
3243 rc = lnet_send(msg->msg_src_nid_param, msg,
3244 msg->msg_rtr_nid_param);
3246 CNETERR("Error sending %s to %s: %d\n",
3247 lnet_msgtyp2str(msg->msg_type),
3248 libcfs_id2str(msg->msg_target), rc);
3249 lnet_finalize(msg, rc);
3254 /* The discovery thread. */
3255 static int lnet_peer_discovery(void *arg)
3257 struct lnet_peer *lp;
3260 CDEBUG(D_NET, "started\n");
3261 cfs_block_allsigs();
3264 if (lnet_peer_discovery_wait_for_work())
3269 if (lnet_push_target_resize_needed())
3270 lnet_push_target_resize();
3272 lnet_net_lock(LNET_LOCK_EX);
3273 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3277 * Process all incoming discovery work requests. When
3278 * discovery must wait on a peer to change state, it
3279 * is added to the tail of the ln_dc_working queue. A
3280 * timestamp keeps track of when the peer was added,
3281 * so we can time out discovery requests that take too
3284 while (!list_empty(&the_lnet.ln_dc_request)) {
3285 lp = list_first_entry(&the_lnet.ln_dc_request,
3286 struct lnet_peer, lp_dc_list);
3287 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3289 * set the time the peer was put on the dc_working
3290 * queue. It shouldn't remain on the queue
3291 * forever, in case the GET message (for ping)
3292 * doesn't get a REPLY or the PUT message (for
3293 * push) doesn't get an ACK.
3295 lp->lp_last_queued = ktime_get_real_seconds();
3296 lnet_net_unlock(LNET_LOCK_EX);
3299 * Select an action depending on the state of
3300 * the peer and whether discovery is disabled.
3301 * The check whether discovery is disabled is
3302 * done after the code that handles processing
3303 * for arrived data, cleanup for failures, and
3304 * forcing a Ping or Push.
3306 spin_lock(&lp->lp_lock);
3307 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3308 libcfs_nid2str(lp->lp_primary_nid), lp,
3310 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3311 rc = lnet_peer_data_present(lp);
3312 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3313 rc = lnet_peer_ping_failed(lp);
3314 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3315 rc = lnet_peer_push_failed(lp);
3316 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3317 rc = lnet_peer_send_ping(lp);
3318 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3319 rc = lnet_peer_send_push(lp);
3320 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3321 rc = lnet_peer_send_ping(lp);
3322 else if (lnet_peer_needs_push(lp))
3323 rc = lnet_peer_send_push(lp);
3325 rc = lnet_peer_discovered(lp);
3326 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3327 libcfs_nid2str(lp->lp_primary_nid), lp,
3329 spin_unlock(&lp->lp_lock);
3331 lnet_net_lock(LNET_LOCK_EX);
3332 if (rc == LNET_REDISCOVER_PEER) {
3333 list_move(&lp->lp_dc_list,
3334 &the_lnet.ln_dc_request);
3336 lnet_peer_discovery_error(lp, rc);
3338 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3339 lnet_peer_discovery_complete(lp);
3340 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3344 lnet_net_unlock(LNET_LOCK_EX);
3347 CDEBUG(D_NET, "stopping\n");
3349 * Clean up before telling lnet_peer_discovery_stop() that
3350 * we're done. Use wake_up() below to somewhat reduce the
3351 * size of the thundering herd if there are multiple threads
3352 * waiting on discovery of a single peer.
3355 /* Queue cleanup 1: stop all pending pings and pushes. */
3356 lnet_net_lock(LNET_LOCK_EX);
3357 while (!list_empty(&the_lnet.ln_dc_working)) {
3358 lp = list_first_entry(&the_lnet.ln_dc_working,
3359 struct lnet_peer, lp_dc_list);
3360 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3361 lnet_net_unlock(LNET_LOCK_EX);
3362 lnet_peer_cancel_discovery(lp);
3363 lnet_net_lock(LNET_LOCK_EX);
3365 lnet_net_unlock(LNET_LOCK_EX);
3367 /* Queue cleanup 2: wait for the expired queue to clear. */
3368 while (!list_empty(&the_lnet.ln_dc_expired))
3369 schedule_timeout(cfs_time_seconds(1));
3371 /* Queue cleanup 3: clear the request queue. */
3372 lnet_net_lock(LNET_LOCK_EX);
3373 while (!list_empty(&the_lnet.ln_dc_request)) {
3374 lp = list_first_entry(&the_lnet.ln_dc_request,
3375 struct lnet_peer, lp_dc_list);
3376 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3377 lnet_peer_discovery_complete(lp);
3379 lnet_net_unlock(LNET_LOCK_EX);
3381 LNetEQFree(the_lnet.ln_dc_eqh);
3382 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3384 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3385 wake_up(&the_lnet.ln_dc_waitq);
3387 CDEBUG(D_NET, "stopped\n");
3392 /* ln_api_mutex is held on entry. */
3393 int lnet_peer_discovery_start(void)
3395 struct task_struct *task;
3398 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3401 rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3403 CERROR("Can't allocate discovery EQ: %d\n", rc);
3407 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3408 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3411 CERROR("Can't start peer discovery thread: %d\n", rc);
3413 LNetEQFree(the_lnet.ln_dc_eqh);
3414 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3416 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3419 CDEBUG(D_NET, "discovery start: %d\n", rc);
3424 /* ln_api_mutex is held on entry. */
3425 void lnet_peer_discovery_stop(void)
3427 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3430 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3431 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3432 wake_up(&the_lnet.ln_dc_waitq);
3434 wait_event(the_lnet.ln_dc_waitq,
3435 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3437 LASSERT(list_empty(&the_lnet.ln_dc_request));
3438 LASSERT(list_empty(&the_lnet.ln_dc_working));
3439 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3441 CDEBUG(D_NET, "discovery stopped\n");
3447 lnet_debug_peer(lnet_nid_t nid)
3449 char *aliveness = "NA";
3450 struct lnet_peer_ni *lp;
3453 cpt = lnet_cpt_of_nid(nid, NULL);
3456 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3458 lnet_net_unlock(cpt);
3459 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3463 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3464 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3466 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3467 libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3468 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3469 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3470 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3472 lnet_peer_ni_decref_locked(lp);
3474 lnet_net_unlock(cpt);
3477 /* Gathering information for userspace. */
3479 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3480 char aliveness[LNET_MAX_STR_LEN],
3481 __u32 *cpt_iter, __u32 *refcount,
3482 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3483 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3484 __u32 *peer_tx_qnob)
3486 struct lnet_peer_table *peer_table;
3487 struct lnet_peer_ni *lp;
3492 /* get the number of CPTs */
3493 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3495 /* if the cpt number to be examined is >= the number of cpts in
3496 * the system then indicate that there are no more cpts to examin
3498 if (*cpt_iter >= lncpt)
3501 /* get the current table */
3502 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3503 /* if the ptable is NULL then there are no more cpts to examine */
3504 if (peer_table == NULL)
3507 lnet_net_lock(*cpt_iter);
3509 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3510 struct list_head *peers = &peer_table->pt_hash[j];
3512 list_for_each_entry(lp, peers, lpni_hashlist) {
3513 if (peer_index-- > 0)
3516 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3517 if (lnet_isrouter(lp) ||
3518 lnet_peer_aliveness_enabled(lp))
3519 snprintf(aliveness, LNET_MAX_STR_LEN,
3520 lnet_is_peer_ni_alive(lp) ? "up" : "down");
3522 *nid = lp->lpni_nid;
3523 *refcount = atomic_read(&lp->lpni_refcount);
3524 *ni_peer_tx_credits =
3525 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3526 *peer_tx_credits = lp->lpni_txcredits;
3527 *peer_rtr_credits = lp->lpni_rtrcredits;
3528 *peer_min_rtr_credits = lp->lpni_mintxcredits;
3529 *peer_tx_qnob = lp->lpni_txqnob;
3535 lnet_net_unlock(*cpt_iter);
3539 return found ? 0 : -ENOENT;
3542 /* ln_api_mutex is held, which keeps the peer list stable */
3543 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3545 struct lnet_ioctl_element_stats *lpni_stats;
3546 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3547 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3548 struct lnet_peer_ni_credit_info *lpni_info;
3549 struct lnet_peer_ni *lpni;
3550 struct lnet_peer *lp;
3555 lp = lnet_find_peer(cfg->prcfg_prim_nid);
3562 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3563 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3564 size *= lp->lp_nnis;
3565 if (size > cfg->prcfg_size) {
3566 cfg->prcfg_size = size;
3571 cfg->prcfg_prim_nid = lp->lp_primary_nid;
3572 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3573 cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3574 cfg->prcfg_count = lp->lp_nnis;
3575 cfg->prcfg_size = size;
3576 cfg->prcfg_state = lp->lp_state;
3578 /* Allocate helper buffers. */
3580 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3583 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3586 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3587 if (!lpni_msg_stats)
3588 goto out_free_stats;
3589 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3591 goto out_free_msg_stats;
3596 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3597 nid = lpni->lpni_nid;
3598 if (copy_to_user(bulk, &nid, sizeof(nid)))
3599 goto out_free_hstats;
3600 bulk += sizeof(nid);
3602 memset(lpni_info, 0, sizeof(*lpni_info));
3603 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3604 if (lnet_isrouter(lpni) ||
3605 lnet_peer_aliveness_enabled(lpni))
3606 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3607 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3609 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3610 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3611 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3612 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3613 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3614 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3615 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3616 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3617 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3618 goto out_free_hstats;
3619 bulk += sizeof(*lpni_info);
3621 memset(lpni_stats, 0, sizeof(*lpni_stats));
3622 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3623 LNET_STATS_TYPE_SEND);
3624 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3625 LNET_STATS_TYPE_RECV);
3626 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3627 LNET_STATS_TYPE_DROP);
3628 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3629 goto out_free_hstats;
3630 bulk += sizeof(*lpni_stats);
3631 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3632 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3633 goto out_free_hstats;
3634 bulk += sizeof(*lpni_msg_stats);
3635 lpni_hstats->hlpni_network_timeout =
3636 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3637 lpni_hstats->hlpni_remote_dropped =
3638 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3639 lpni_hstats->hlpni_remote_timeout =
3640 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3641 lpni_hstats->hlpni_remote_error =
3642 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3643 lpni_hstats->hlpni_health_value =
3644 atomic_read(&lpni->lpni_healthv);
3645 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3646 goto out_free_hstats;
3647 bulk += sizeof(*lpni_hstats);
3652 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3654 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3656 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3658 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3660 lnet_peer_decref_locked(lp);
3666 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3668 /* the mt could've shutdown and cleaned up the queues */
3669 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3672 if (list_empty(&lpni->lpni_recovery) &&
3673 atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3674 CERROR("lpni %s added to recovery queue. Health = %d\n",
3675 libcfs_nid2str(lpni->lpni_nid),
3676 atomic_read(&lpni->lpni_healthv));
3677 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3678 lnet_peer_ni_addref_locked(lpni);
3682 /* Call with the ln_api_mutex held */
3684 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3686 struct lnet_peer_table *ptable;
3687 struct lnet_peer *lp;
3688 struct lnet_peer_net *lpn;
3689 struct lnet_peer_ni *lpni;
3693 if (the_lnet.ln_state != LNET_STATE_RUNNING)
3697 lnet_net_lock(LNET_LOCK_EX);
3698 lpni = lnet_find_peer_ni_locked(nid);
3700 lnet_net_unlock(LNET_LOCK_EX);
3703 atomic_set(&lpni->lpni_healthv, value);
3704 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3705 lnet_peer_ni_decref_locked(lpni);
3706 lnet_net_unlock(LNET_LOCK_EX);
3710 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3713 * Walk all the peers and reset the healhv for each one to the
3716 lnet_net_lock(LNET_LOCK_EX);
3717 for (cpt = 0; cpt < lncpt; cpt++) {
3718 ptable = the_lnet.ln_peer_tables[cpt];
3719 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3720 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3721 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3723 atomic_set(&lpni->lpni_healthv, value);
3724 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3729 lnet_net_unlock(LNET_LOCK_EX);