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/
34 #define DEBUG_SUBSYSTEM S_LNET
36 #include <linux/sched.h>
37 #ifdef HAVE_SCHED_HEADERS
38 #include <linux/sched/signal.h>
40 #include <linux/uaccess.h>
42 #include <lnet/udsp.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_NID_NET(&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 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
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(struct lnet_nid *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 INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
170 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171 kref_init(&lpni->lpni_kref);
172 lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
174 spin_lock_init(&lpni->lpni_lock);
176 if (lnet_peers_start_down())
177 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
179 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
180 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
181 lpni->lpni_nid = *nid;
182 lpni->lpni_cpt = cpt;
183 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
185 net = lnet_get_net_locked(LNET_NID_NET(nid));
186 lpni->lpni_net = net;
188 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
189 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
190 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
191 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
194 * This peer_ni is not on a local network, so we
195 * cannot add the credits here. In case the net is
196 * added later, add the peer_ni to the remote peer ni
197 * list so it can be easily found and revisited.
199 /* FIXME: per-net implementation instead? */
200 lnet_peer_ni_addref_locked(lpni);
201 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
202 &the_lnet.ln_remote_peer_ni_list);
205 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
210 static struct lnet_peer_net *
211 lnet_peer_net_alloc(__u32 net_id)
213 struct lnet_peer_net *lpn;
215 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
219 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
220 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
221 lpn->lpn_net_id = net_id;
222 lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
224 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
230 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
232 struct lnet_peer *lp;
234 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
236 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
237 LASSERT(list_empty(&lpn->lpn_peer_nis));
238 LASSERT(list_empty(&lpn->lpn_peer_nets));
240 lpn->lpn_peer = NULL;
241 LIBCFS_FREE(lpn, sizeof(*lpn));
243 lnet_peer_decref_locked(lp);
246 static struct lnet_peer *
247 lnet_peer_alloc(struct lnet_nid *nid)
249 struct lnet_peer *lp;
251 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
255 INIT_LIST_HEAD(&lp->lp_rtrq);
256 INIT_LIST_HEAD(&lp->lp_routes);
257 INIT_LIST_HEAD(&lp->lp_peer_list);
258 INIT_LIST_HEAD(&lp->lp_peer_nets);
259 INIT_LIST_HEAD(&lp->lp_dc_list);
260 INIT_LIST_HEAD(&lp->lp_dc_pendq);
261 INIT_LIST_HEAD(&lp->lp_rtr_list);
262 init_waitqueue_head(&lp->lp_dc_waitq);
263 spin_lock_init(&lp->lp_lock);
264 lp->lp_primary_nid = *nid;
265 lp->lp_disc_src_nid = LNET_ANY_NID;
266 lp->lp_disc_dst_nid = LNET_ANY_NID;
267 if (lnet_peers_start_down())
268 lp->lp_alive = false;
273 * all peers created on a router should have health on
274 * if it's not already on.
276 if (the_lnet.ln_routing && !lnet_health_sensitivity)
277 lp->lp_health_sensitivity = 1;
280 * Turn off discovery for loopback peer. If you're creating a peer
281 * for the loopback interface then that was initiated when we
282 * attempted to send a message over the loopback. There is no need
283 * to ever use a different interface when sending messages to
287 lp->lp_state = LNET_PEER_NO_DISCOVERY;
288 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
290 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
296 lnet_destroy_peer_locked(struct lnet_peer *lp)
298 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
300 LASSERT(atomic_read(&lp->lp_refcount) == 0);
301 LASSERT(lp->lp_rtr_refcount == 0);
302 LASSERT(list_empty(&lp->lp_peer_nets));
303 LASSERT(list_empty(&lp->lp_peer_list));
304 LASSERT(list_empty(&lp->lp_dc_list));
307 lnet_ping_buffer_decref(lp->lp_data);
310 * if there are messages still on the pending queue, then make
311 * sure to queue them on the ln_msg_resend list so they can be
312 * resent at a later point if the discovery thread is still
314 * If the discovery thread has stopped, then the wakeup will be a
315 * no-op, and it is expected the lnet_shutdown_lndnets() will
316 * eventually be called, which will traverse this list and
317 * finalize the messages on the list.
318 * We can not resend them now because we're holding the cpt lock.
319 * Releasing the lock can cause an inconsistent state
321 spin_lock(&the_lnet.ln_msg_resend_lock);
322 spin_lock(&lp->lp_lock);
323 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
324 spin_unlock(&lp->lp_lock);
325 spin_unlock(&the_lnet.ln_msg_resend_lock);
326 wake_up(&the_lnet.ln_dc_waitq);
328 LIBCFS_FREE(lp, sizeof(*lp));
332 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
333 * that peer_net, detach the peer_net from the peer.
335 * Call with lnet_net_lock/EX held
338 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
340 struct lnet_peer_table *ptable;
341 struct lnet_peer_net *lpn;
342 struct lnet_peer *lp;
345 * Belts and suspenders: gracefully handle teardown of a
346 * partially connected peer_ni.
348 lpn = lpni->lpni_peer_net;
350 list_del_init(&lpni->lpni_peer_nis);
352 * If there are no lpni's left, we detach lpn from
353 * lp_peer_nets, so it cannot be found anymore.
355 if (list_empty(&lpn->lpn_peer_nis))
356 list_del_init(&lpn->lpn_peer_nets);
358 /* Update peer NID count. */
363 * If there are no more peer nets, make the peer unfindable
364 * via the peer_tables.
366 * Otherwise, if the peer is DISCOVERED, tell discovery to
367 * take another look at it. This is a no-op if discovery for
368 * this peer did the detaching.
370 if (list_empty(&lp->lp_peer_nets)) {
371 list_del_init(&lp->lp_peer_list);
372 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
374 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
375 /* Discovery isn't running, nothing to do here. */
376 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
377 lnet_peer_queue_for_discovery(lp);
378 wake_up(&the_lnet.ln_dc_waitq);
380 CDEBUG(D_NET, "peer %s NID %s\n",
381 libcfs_nidstr(&lp->lp_primary_nid),
382 libcfs_nidstr(&lpni->lpni_nid));
385 /* called with lnet_net_lock LNET_LOCK_EX held */
387 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
389 struct lnet_peer_table *ptable = NULL;
391 /* don't remove a peer_ni if it's also a gateway */
392 if (lnet_isrouter(lpni) && !force) {
393 CERROR("Peer NI %s is a gateway. Can not delete it\n",
394 libcfs_nidstr(&lpni->lpni_nid));
398 lnet_peer_remove_from_remote_list(lpni);
400 /* remove peer ni from the hash list. */
401 list_del_init(&lpni->lpni_hashlist);
404 * indicate the peer is being deleted so the monitor thread can
405 * remove it from the recovery queue.
407 spin_lock(&lpni->lpni_lock);
408 lpni->lpni_state |= LNET_PEER_NI_DELETING;
409 spin_unlock(&lpni->lpni_lock);
411 /* decrement the ref count on the peer table */
412 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
415 * The peer_ni can no longer be found with a lookup. But there
416 * can be current users, so keep track of it on the zombie
417 * list until the reference count has gone to zero.
419 * The last reference may be lost in a place where the
420 * lnet_net_lock locks only a single cpt, and that cpt may not
421 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
424 spin_lock(&ptable->pt_zombie_lock);
425 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
426 ptable->pt_zombies++;
427 spin_unlock(&ptable->pt_zombie_lock);
429 /* no need to keep this peer_ni on the hierarchy anymore */
430 lnet_peer_detach_peer_ni_locked(lpni);
432 /* remove hashlist reference on peer_ni */
433 lnet_peer_ni_decref_locked(lpni);
438 void lnet_peer_uninit(void)
440 struct lnet_peer_ni *lpni, *tmp;
442 lnet_net_lock(LNET_LOCK_EX);
444 /* remove all peer_nis from the remote peer and the hash list */
445 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
446 lpni_on_remote_peer_ni_list)
447 lnet_peer_ni_del_locked(lpni, false);
449 lnet_peer_tables_destroy();
451 lnet_net_unlock(LNET_LOCK_EX);
455 lnet_peer_del_locked(struct lnet_peer *peer)
457 struct lnet_peer_ni *lpni = NULL, *lpni2;
460 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&peer->lp_primary_nid));
462 spin_lock(&peer->lp_lock);
463 peer->lp_state |= LNET_PEER_MARK_DELETED;
464 spin_unlock(&peer->lp_lock);
466 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
467 while (lpni != NULL) {
468 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
469 rc = lnet_peer_ni_del_locked(lpni, false);
479 * Discovering this peer is taking too long. Cancel any Ping or Push
480 * that discovery is waiting on by unlinking the relevant MDs. The
481 * lnet_discovery_event_handler() will proceed from here and complete
484 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
486 struct lnet_handle_md ping_mdh;
487 struct lnet_handle_md push_mdh;
489 LNetInvalidateMDHandle(&ping_mdh);
490 LNetInvalidateMDHandle(&push_mdh);
492 spin_lock(&lp->lp_lock);
493 if (lp->lp_state & LNET_PEER_PING_SENT) {
494 ping_mdh = lp->lp_ping_mdh;
495 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
497 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
498 push_mdh = lp->lp_push_mdh;
499 LNetInvalidateMDHandle(&lp->lp_push_mdh);
501 spin_unlock(&lp->lp_lock);
503 if (!LNetMDHandleIsInvalid(ping_mdh))
504 LNetMDUnlink(ping_mdh);
505 if (!LNetMDHandleIsInvalid(push_mdh))
506 LNetMDUnlink(push_mdh);
510 lnet_peer_del(struct lnet_peer *peer)
514 lnet_peer_cancel_discovery(peer);
515 lnet_net_lock(LNET_LOCK_EX);
516 rc = lnet_peer_del_locked(peer);
517 lnet_net_unlock(LNET_LOCK_EX);
523 * Delete a NID from a peer. Call with ln_api_mutex held.
526 * -EPERM: Non-DLC deletion from DLC-configured peer.
527 * -ENOENT: No lnet_peer_ni corresponding to the nid.
528 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
529 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
532 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
534 struct lnet_peer_ni *lpni;
535 struct lnet_nid primary_nid = lp->lp_primary_nid;
538 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
540 lnet_nid4_to_nid(nid4, &nid);
541 if (!(flags & LNET_PEER_CONFIGURED)) {
542 if (lp->lp_state & LNET_PEER_CONFIGURED) {
548 lpni = lnet_peer_ni_find_locked(&nid);
553 lnet_peer_ni_decref_locked(lpni);
554 if (lp != lpni->lpni_peer_net->lpn_peer) {
560 * This function only allows deletion of the primary NID if it
563 if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && !force) {
568 lnet_net_lock(LNET_LOCK_EX);
570 if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && force) {
571 struct lnet_peer_ni *lpni2;
572 /* assign the next peer_ni to be the primary */
573 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
575 lp->lp_primary_nid = lpni2->lpni_nid;
577 rc = lnet_peer_ni_del_locked(lpni, force);
579 lnet_net_unlock(LNET_LOCK_EX);
582 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
583 libcfs_nidstr(&primary_nid), libcfs_nidstr(&nid),
590 lnet_peer_table_cleanup_locked(struct lnet_net *net,
591 struct lnet_peer_table *ptable)
594 struct lnet_peer_ni *next;
595 struct lnet_peer_ni *lpni;
596 struct lnet_peer *peer;
598 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
599 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
601 if (net != NULL && net != lpni->lpni_net)
604 peer = lpni->lpni_peer_net->lpn_peer;
605 if (!nid_same(&peer->lp_primary_nid,
607 lnet_peer_ni_del_locked(lpni, false);
611 * Removing the primary NID implies removing
612 * the entire peer. Advance next beyond any
613 * peer_ni that belongs to the same peer.
615 list_for_each_entry_from(next, &ptable->pt_hash[i],
617 if (next->lpni_peer_net->lpn_peer != peer)
620 lnet_peer_del_locked(peer);
626 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
628 wait_var_event_warning(&ptable->pt_zombies,
629 ptable->pt_zombies == 0,
630 "Waiting for %d zombies on peer table\n",
635 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
636 struct lnet_peer_table *ptable)
638 struct lnet_peer_ni *lp;
639 struct lnet_peer_ni *tmp;
640 struct lnet_nid gw_nid;
643 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
644 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
646 if (net != lp->lpni_net)
649 if (!lnet_isrouter(lp))
652 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
654 lnet_net_unlock(LNET_LOCK_EX);
655 lnet_del_route(LNET_NET_ANY, &gw_nid);
656 lnet_net_lock(LNET_LOCK_EX);
662 lnet_peer_tables_cleanup(struct lnet_net *net)
665 struct lnet_peer_table *ptable;
667 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
668 /* If just deleting the peers for a NI, get rid of any routes these
669 * peers are gateways for. */
670 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
671 lnet_net_lock(LNET_LOCK_EX);
672 lnet_peer_table_del_rtrs_locked(net, ptable);
673 lnet_net_unlock(LNET_LOCK_EX);
676 /* Start the cleanup process */
677 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
678 lnet_net_lock(LNET_LOCK_EX);
679 lnet_peer_table_cleanup_locked(net, ptable);
680 lnet_net_unlock(LNET_LOCK_EX);
683 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
684 lnet_peer_ni_finalize_wait(ptable);
687 static struct lnet_peer_ni *
688 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, struct lnet_nid *nid)
690 struct list_head *peers;
691 struct lnet_peer_ni *lp;
693 if (the_lnet.ln_state != LNET_STATE_RUNNING)
696 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
697 list_for_each_entry(lp, peers, lpni_hashlist) {
698 if (nid_same(&lp->lpni_nid, nid)) {
699 lnet_peer_ni_addref_locked(lp);
707 struct lnet_peer_ni *
708 lnet_find_peer_ni_locked(lnet_nid_t nid4)
710 struct lnet_peer_ni *lpni;
711 struct lnet_peer_table *ptable;
715 lnet_nid4_to_nid(nid4, &nid);
717 cpt = lnet_nid_cpt_hash(&nid, LNET_CPT_NUMBER);
719 ptable = the_lnet.ln_peer_tables[cpt];
720 lpni = lnet_get_peer_ni_locked(ptable, &nid);
725 struct lnet_peer_ni *
726 lnet_peer_ni_find_locked(struct lnet_nid *nid)
728 struct lnet_peer_ni *lpni;
729 struct lnet_peer_table *ptable;
732 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
734 ptable = the_lnet.ln_peer_tables[cpt];
735 lpni = lnet_get_peer_ni_locked(ptable, nid);
740 struct lnet_peer_ni *
741 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
743 struct lnet_peer_net *lpn;
744 struct lnet_peer_ni *lpni;
746 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
750 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
751 if (lnet_nid_to_nid4(&lpni->lpni_nid) == nid)
758 struct lnet_peer_ni *
759 lnet_peer_ni_get_locked(struct lnet_peer *lp, struct lnet_nid *nid)
761 struct lnet_peer_net *lpn;
762 struct lnet_peer_ni *lpni;
764 lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
768 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
769 if (nid_same(&lpni->lpni_nid, nid))
777 lnet_find_peer4(lnet_nid_t nid)
779 struct lnet_peer_ni *lpni;
780 struct lnet_peer *lp = NULL;
783 cpt = lnet_net_lock_current();
784 lpni = lnet_find_peer_ni_locked(nid);
786 lp = lpni->lpni_peer_net->lpn_peer;
787 lnet_peer_addref_locked(lp);
788 lnet_peer_ni_decref_locked(lpni);
790 lnet_net_unlock(cpt);
796 lnet_find_peer(struct lnet_nid *nid)
798 struct lnet_peer_ni *lpni;
799 struct lnet_peer *lp = NULL;
802 cpt = lnet_net_lock_current();
803 lpni = lnet_peer_ni_find_locked(nid);
805 lp = lpni->lpni_peer_net->lpn_peer;
806 lnet_peer_addref_locked(lp);
807 lnet_peer_ni_decref_locked(lpni);
809 lnet_net_unlock(cpt);
814 struct lnet_peer_net *
815 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
817 struct lnet_peer_net *net;
820 /* no net id provided return the first net */
821 net = list_first_entry_or_null(&lp->lp_peer_nets,
822 struct lnet_peer_net,
828 /* find the net after the one provided */
829 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
830 if (net->lpn_net_id == prev_lpn_id) {
832 * if we reached the end of the list loop to the
835 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
836 return list_first_entry_or_null(&lp->lp_peer_nets,
837 struct lnet_peer_net,
840 return list_next_entry(net, lpn_peer_nets);
847 struct lnet_peer_ni *
848 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
849 struct lnet_peer_net *peer_net,
850 struct lnet_peer_ni *prev)
852 struct lnet_peer_ni *lpni;
853 struct lnet_peer_net *net = peer_net;
857 if (list_empty(&peer->lp_peer_nets))
860 net = list_entry(peer->lp_peer_nets.next,
861 struct lnet_peer_net,
864 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
870 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
872 * if you reached the end of the peer ni list and the peer
873 * net is specified then there are no more peer nis in that
880 * we reached the end of this net ni list. move to the
883 if (prev->lpni_peer_net->lpn_peer_nets.next ==
885 /* no more nets and no more NIs. */
888 /* get the next net */
889 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
890 struct lnet_peer_net,
892 /* get the ni on it */
893 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
899 /* there are more nis left */
900 lpni = list_entry(prev->lpni_peer_nis.next,
901 struct lnet_peer_ni, lpni_peer_nis);
906 /* Call with the ln_api_mutex held */
907 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
909 struct lnet_process_id id;
910 struct lnet_peer_table *ptable;
911 struct lnet_peer *lp;
920 if (the_lnet.ln_state != LNET_STATE_RUNNING)
923 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
926 * Count the number of peers, and return E2BIG if the buffer
927 * is too small. We'll also return the desired size.
930 for (cpt = 0; cpt < lncpt; cpt++) {
931 ptable = the_lnet.ln_peer_tables[cpt];
932 count += ptable->pt_peers;
934 size = count * sizeof(*ids);
939 * Walk the peer lists and copy out the primary nids.
940 * This is safe because the peer lists are only modified
941 * while the ln_api_mutex is held. So we don't need to
942 * hold the lnet_net_lock as well, and can therefore
943 * directly call copy_to_user().
946 memset(&id, 0, sizeof(id));
947 id.pid = LNET_PID_LUSTRE;
949 for (cpt = 0; cpt < lncpt; cpt++) {
950 ptable = the_lnet.ln_peer_tables[cpt];
951 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
952 if (!nid_is_nid4(&lp->lp_primary_nid))
956 id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
957 if (copy_to_user(&ids[i], &id, sizeof(id)))
970 * Start pushes to peers that need to be updated for a configuration
971 * change on this node.
974 lnet_push_update_to_peers(int force)
976 struct lnet_peer_table *ptable;
977 struct lnet_peer *lp;
981 lnet_net_lock(LNET_LOCK_EX);
982 if (lnet_peer_discovery_disabled)
984 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
985 for (cpt = 0; cpt < lncpt; cpt++) {
986 ptable = the_lnet.ln_peer_tables[cpt];
987 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
989 spin_lock(&lp->lp_lock);
990 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
991 lp->lp_state |= LNET_PEER_FORCE_PUSH;
992 spin_unlock(&lp->lp_lock);
994 if (lnet_peer_needs_push(lp))
995 lnet_peer_queue_for_discovery(lp);
998 lnet_net_unlock(LNET_LOCK_EX);
999 wake_up(&the_lnet.ln_dc_waitq);
1002 /* find the NID in the preferred gateways for the remote peer
1004 * false: list is not empty and NID is not preferred
1005 * false: list is empty
1006 * true: nid is found in the list
1009 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
1010 struct lnet_nid *gw_nid)
1012 struct lnet_nid_list *ne;
1014 CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
1015 libcfs_nidstr(&lpni->lpni_nid),
1016 list_empty(&lpni->lpni_rtr_pref_nids));
1018 if (list_empty(&lpni->lpni_rtr_pref_nids))
1021 /* iterate through all the preferred NIDs and see if any of them
1022 * matches the provided gw_nid
1024 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1025 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
1026 libcfs_nidstr(&ne->nl_nid),
1027 libcfs_nidstr(gw_nid));
1028 if (nid_same(&ne->nl_nid, gw_nid))
1036 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
1038 struct list_head zombies;
1039 struct lnet_nid_list *ne;
1040 struct lnet_nid_list *tmp;
1041 int cpt = lpni->lpni_cpt;
1043 INIT_LIST_HEAD(&zombies);
1046 list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
1047 lnet_net_unlock(cpt);
1049 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1050 list_del(&ne->nl_list);
1051 LIBCFS_FREE(ne, sizeof(*ne));
1056 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
1057 struct lnet_nid *gw_nid)
1059 int cpt = lpni->lpni_cpt;
1060 struct lnet_nid_list *ne = NULL;
1062 /* This function is called with api_mutex held. When the api_mutex
1063 * is held the list can not be modified, as it is only modified as
1064 * a result of applying a UDSP and that happens under api_mutex
1067 __must_hold(&the_lnet.ln_api_mutex);
1069 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1070 if (nid_same(&ne->nl_nid, gw_nid))
1074 LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1078 ne->nl_nid = *gw_nid;
1080 /* Lock the cpt to protect against addition and checks in the
1081 * selection algorithm
1084 list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1085 lnet_net_unlock(cpt);
1091 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1092 * this is a preferred point-to-point path. Call with lnet_net_lock in
1096 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1098 struct lnet_nid_list *ne;
1100 if (lpni->lpni_pref_nnids == 0)
1102 if (lpni->lpni_pref_nnids == 1)
1103 return nid_same(&lpni->lpni_pref.nid, nid);
1104 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1105 if (nid_same(&ne->nl_nid, nid))
1112 * Set a single ni as preferred, provided no preferred ni is already
1113 * defined. Only to be used for non-multi-rail peer_ni.
1116 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni,
1117 struct lnet_nid *nid)
1123 spin_lock(&lpni->lpni_lock);
1124 if (LNET_NID_IS_ANY(nid)) {
1126 } else if (lpni->lpni_pref_nnids > 0) {
1128 } else if (lpni->lpni_pref_nnids == 0) {
1129 lpni->lpni_pref.nid = *nid;
1130 lpni->lpni_pref_nnids = 1;
1131 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1133 spin_unlock(&lpni->lpni_lock);
1135 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1136 libcfs_nidstr(&lpni->lpni_nid), libcfs_nidstr(nid), rc);
1141 * Clear the preferred NID from a non-multi-rail peer_ni, provided
1142 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1145 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1149 spin_lock(&lpni->lpni_lock);
1150 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1151 lpni->lpni_pref_nnids = 0;
1152 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1153 } else if (lpni->lpni_pref_nnids == 0) {
1158 spin_unlock(&lpni->lpni_lock);
1160 CDEBUG(D_NET, "peer %s: %d\n",
1161 libcfs_nidstr(&lpni->lpni_nid), rc);
1166 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1168 lpni->lpni_sel_priority = priority;
1172 * Clear the preferred NIDs from a non-multi-rail peer.
1175 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1177 struct lnet_peer_ni *lpni = NULL;
1179 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1180 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1184 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1186 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1187 struct lnet_nid_list *ne1 = NULL;
1188 struct lnet_nid_list *ne2 = NULL;
1189 struct lnet_nid *tmp_nid = NULL;
1192 if (LNET_NID_IS_ANY(nid)) {
1197 if (lpni->lpni_pref_nnids == 1 &&
1198 nid_same(&lpni->lpni_pref.nid, nid)) {
1203 /* A non-MR node may have only one preferred NI per peer_ni */
1204 if (lpni->lpni_pref_nnids > 0 &&
1205 !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1210 /* add the new preferred nid to the list of preferred nids */
1211 if (lpni->lpni_pref_nnids != 0) {
1212 size_t alloc_size = sizeof(*ne1);
1214 if (lpni->lpni_pref_nnids == 1) {
1215 tmp_nid = &lpni->lpni_pref.nid;
1216 INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1219 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1220 if (nid_same(&ne1->nl_nid, nid)) {
1226 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1233 /* move the originally stored nid to the list */
1234 if (lpni->lpni_pref_nnids == 1) {
1235 LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1236 lpni->lpni_cpt, alloc_size);
1241 INIT_LIST_HEAD(&ne2->nl_list);
1242 ne2->nl_nid = *tmp_nid;
1247 lnet_net_lock(LNET_LOCK_EX);
1248 spin_lock(&lpni->lpni_lock);
1249 if (lpni->lpni_pref_nnids == 0) {
1250 lpni->lpni_pref.nid = *nid;
1253 list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1254 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1256 lpni->lpni_pref_nnids++;
1257 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1258 spin_unlock(&lpni->lpni_lock);
1259 lnet_net_unlock(LNET_LOCK_EX);
1262 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1263 spin_lock(&lpni->lpni_lock);
1264 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1265 spin_unlock(&lpni->lpni_lock);
1267 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1268 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1273 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1275 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1276 struct lnet_nid_list *ne = NULL;
1279 if (lpni->lpni_pref_nnids == 0) {
1284 if (lpni->lpni_pref_nnids == 1) {
1285 if (!nid_same(&lpni->lpni_pref.nid, nid)) {
1290 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1291 if (nid_same(&ne->nl_nid, nid))
1292 goto remove_nid_entry;
1300 lnet_net_lock(LNET_LOCK_EX);
1301 spin_lock(&lpni->lpni_lock);
1302 if (lpni->lpni_pref_nnids == 1)
1303 lpni->lpni_pref.nid = LNET_ANY_NID;
1305 list_del_init(&ne->nl_list);
1306 if (lpni->lpni_pref_nnids == 2) {
1307 struct lnet_nid_list *ne, *tmp;
1309 list_for_each_entry_safe(ne, tmp,
1310 &lpni->lpni_pref.nids,
1312 lpni->lpni_pref.nid = ne->nl_nid;
1313 list_del_init(&ne->nl_list);
1314 LIBCFS_FREE(ne, sizeof(*ne));
1318 lpni->lpni_pref_nnids--;
1319 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1320 spin_unlock(&lpni->lpni_lock);
1321 lnet_net_unlock(LNET_LOCK_EX);
1324 LIBCFS_FREE(ne, sizeof(*ne));
1326 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1327 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1332 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1334 struct list_head zombies;
1335 struct lnet_nid_list *ne;
1336 struct lnet_nid_list *tmp;
1338 INIT_LIST_HEAD(&zombies);
1340 lnet_net_lock(LNET_LOCK_EX);
1341 if (lpni->lpni_pref_nnids == 1)
1342 lpni->lpni_pref.nid = LNET_ANY_NID;
1343 else if (lpni->lpni_pref_nnids > 1)
1344 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1345 lpni->lpni_pref_nnids = 0;
1346 lnet_net_unlock(LNET_LOCK_EX);
1348 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1349 list_del_init(&ne->nl_list);
1350 LIBCFS_FREE(ne, sizeof(*ne));
1355 lnet_peer_primary_nid_locked(struct lnet_nid *nid, struct lnet_nid *result)
1357 struct lnet_peer_ni *lpni;
1360 lpni = lnet_peer_ni_find_locked(nid);
1362 *result = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1363 lnet_peer_ni_decref_locked(lpni);
1368 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1369 __must_hold(&lp->lp_lock)
1371 if (lnet_peer_discovery_disabled)
1374 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1375 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1386 lnet_is_discovery_disabled(struct lnet_peer *lp)
1390 spin_lock(&lp->lp_lock);
1391 rc = lnet_is_discovery_disabled_locked(lp);
1392 spin_unlock(&lp->lp_lock);
1398 LNetAddPeer(lnet_nid_t *nids, __u32 num_nids)
1400 lnet_nid_t pnid = 0;
1404 if (!nids || num_nids < 1)
1407 rc = LNetNIInit(LNET_PID_ANY);
1411 mutex_lock(&the_lnet.ln_api_mutex);
1413 mr = lnet_peer_discovery_disabled == 0;
1416 for (i = 0; i < num_nids; i++) {
1417 if (nids[i] == LNET_NID_LO_0)
1422 rc = lnet_add_peer_ni(pnid, LNET_NID_ANY, mr, true);
1423 } else if (lnet_peer_discovery_disabled) {
1424 rc = lnet_add_peer_ni(nids[i], LNET_NID_ANY, mr, true);
1426 rc = lnet_add_peer_ni(pnid, nids[i], mr, true);
1429 if (rc && rc != -EEXIST)
1434 mutex_unlock(&the_lnet.ln_api_mutex);
1438 return rc == -EEXIST ? 0 : rc;
1440 EXPORT_SYMBOL(LNetAddPeer);
1442 void LNetPrimaryNID(struct lnet_nid *nid)
1444 struct lnet_peer *lp;
1445 struct lnet_peer_ni *lpni;
1446 struct lnet_nid orig;
1450 if (!nid || nid_is_lo0(nid))
1454 cpt = lnet_net_lock_current();
1455 lpni = lnet_peerni_by_nid_locked(nid, NULL, cpt);
1460 lp = lpni->lpni_peer_net->lpn_peer;
1462 /* If discovery is disabled locally then we needn't bother running
1463 * discovery here because discovery will not modify whatever
1464 * primary NID is currently set for this peer. If the specified peer is
1465 * down then this discovery can introduce long delays into the mount
1466 * process, so skip it if it isn't necessary.
1468 while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1469 spin_lock(&lp->lp_lock);
1470 /* force a full discovery cycle */
1471 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1472 spin_unlock(&lp->lp_lock);
1474 rc = lnet_discover_peer_locked(lpni, cpt, true);
1477 /* The lpni (or lp) for this NID may have changed and our ref is
1478 * the only thing keeping the old one around. Release the ref
1479 * and lookup the lpni again
1481 lnet_peer_ni_decref_locked(lpni);
1482 lpni = lnet_peer_ni_find_locked(nid);
1487 lp = lpni->lpni_peer_net->lpn_peer;
1489 /* If we find that the peer has discovery disabled then we will
1490 * not modify whatever primary NID is currently set for this
1491 * peer. Thus, we can break out of this loop even if the peer
1492 * is not fully up to date.
1494 if (lnet_is_discovery_disabled(lp))
1497 *nid = lp->lp_primary_nid;
1499 lnet_peer_ni_decref_locked(lpni);
1501 lnet_net_unlock(cpt);
1503 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nidstr(&orig),
1504 libcfs_nidstr(nid), rc);
1506 EXPORT_SYMBOL(LNetPrimaryNID);
1508 struct lnet_peer_net *
1509 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1511 struct lnet_peer_net *peer_net;
1512 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1513 if (peer_net->lpn_net_id == net_id)
1520 * Attach a peer_ni to a peer_net and peer. This function assumes
1521 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1522 * may be attached to a different peer, in which case it will be
1523 * properly detached first. The whole operation is done atomically.
1525 * This function consumes the reference on lpni and Always returns 0.
1526 * This is the last function called from functions that do return an
1527 * int, so returning 0 here allows the compiler to do a tail call.
1530 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1531 struct lnet_peer_net *lpn,
1532 struct lnet_peer_ni *lpni,
1535 struct lnet_peer_table *ptable;
1536 bool new_lpn = false;
1539 /* Install the new peer_ni */
1540 lnet_net_lock(LNET_LOCK_EX);
1541 /* Add peer_ni to global peer table hash, if necessary. */
1542 if (list_empty(&lpni->lpni_hashlist)) {
1543 int hash = lnet_nid2peerhash(&lpni->lpni_nid);
1545 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1546 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1547 ptable->pt_version++;
1548 lnet_peer_ni_addref_locked(lpni);
1551 /* Detach the peer_ni from an existing peer, if necessary. */
1552 if (lpni->lpni_peer_net) {
1553 LASSERT(lpni->lpni_peer_net != lpn);
1554 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1555 lnet_peer_detach_peer_ni_locked(lpni);
1556 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1557 lpni->lpni_peer_net = NULL;
1560 /* Add peer_ni to peer_net */
1561 lpni->lpni_peer_net = lpn;
1562 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1563 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1565 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1566 lnet_update_peer_net_healthv(lpni);
1567 lnet_peer_net_addref_locked(lpn);
1569 /* Add peer_net to peer */
1570 if (!lpn->lpn_peer) {
1573 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1574 list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1576 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1577 lnet_peer_addref_locked(lp);
1580 /* Add peer to global peer list, if necessary */
1581 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1582 if (list_empty(&lp->lp_peer_list)) {
1583 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1588 /* Update peer state */
1589 spin_lock(&lp->lp_lock);
1590 if (flags & LNET_PEER_CONFIGURED) {
1591 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1592 lp->lp_state |= LNET_PEER_CONFIGURED;
1594 if (flags & LNET_PEER_MULTI_RAIL) {
1595 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1596 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1597 lnet_peer_clr_non_mr_pref_nids(lp);
1600 spin_unlock(&lp->lp_lock);
1606 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1608 CERROR("Failed to apply UDSPs on lpn %s\n",
1609 libcfs_net2str(lpn->lpn_net_id));
1611 rc = lnet_udsp_apply_policies_on_lpni(lpni);
1613 CERROR("Failed to apply UDSPs on lpni %s\n",
1614 libcfs_nidstr(&lpni->lpni_nid));
1616 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1617 libcfs_nidstr(&lp->lp_primary_nid),
1618 libcfs_nidstr(&lpni->lpni_nid), flags);
1619 lnet_peer_ni_decref_locked(lpni);
1620 lnet_net_unlock(LNET_LOCK_EX);
1626 * Create a new peer, with nid as its primary nid.
1628 * Call with the lnet_api_mutex held.
1631 lnet_peer_add(lnet_nid_t nid4, unsigned int flags)
1633 struct lnet_nid nid;
1634 struct lnet_peer *lp;
1635 struct lnet_peer_net *lpn;
1636 struct lnet_peer_ni *lpni;
1639 LASSERT(nid4 != LNET_NID_ANY);
1642 * No need for the lnet_net_lock here, because the
1643 * lnet_api_mutex is held.
1645 lpni = lnet_find_peer_ni_locked(nid4);
1647 /* A peer with this NID already exists. */
1648 lp = lpni->lpni_peer_net->lpn_peer;
1649 lnet_peer_ni_decref_locked(lpni);
1651 * This is an error if the peer was configured and the
1652 * primary NID differs or an attempt is made to change
1653 * the Multi-Rail flag. Otherwise the assumption is
1654 * that an existing peer is being modified.
1656 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1657 if (lnet_nid_to_nid4(&lp->lp_primary_nid) != nid4)
1659 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1662 } else if (!(flags & LNET_PEER_CONFIGURED)) {
1663 if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid4) {
1668 /* Delete and recreate as a configured peer. */
1669 rc = lnet_peer_del(lp);
1674 /* Create peer, peer_net, and peer_ni. */
1676 lnet_nid4_to_nid(nid4, &nid);
1677 lp = lnet_peer_alloc(&nid);
1680 lpn = lnet_peer_net_alloc(LNET_NID_NET(&nid));
1683 lpni = lnet_peer_ni_alloc(&nid);
1687 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1690 LIBCFS_FREE(lpn, sizeof(*lpn));
1692 LIBCFS_FREE(lp, sizeof(*lp));
1694 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1695 libcfs_nid2str(nid4), flags, rc);
1700 * Add a NID to a peer. Call with ln_api_mutex held.
1703 * -EPERM: Non-DLC addition to a DLC-configured peer.
1704 * -EEXIST: The NID was configured by DLC for a different peer.
1705 * -ENOMEM: Out of memory.
1706 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1707 * non-multi-rail peer.
1710 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
1712 struct lnet_peer_net *lpn;
1713 struct lnet_peer_ni *lpni;
1714 struct lnet_nid nid;
1718 LASSERT(nid4 != LNET_NID_ANY);
1720 lnet_nid4_to_nid(nid4, &nid);
1722 /* A configured peer can only be updated through configuration. */
1723 if (!(flags & LNET_PEER_CONFIGURED)) {
1724 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1731 * The MULTI_RAIL flag can be set but not cleared, because
1732 * that would leave the peer struct in an invalid state.
1734 if (flags & LNET_PEER_MULTI_RAIL) {
1735 spin_lock(&lp->lp_lock);
1736 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1737 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1738 lnet_peer_clr_non_mr_pref_nids(lp);
1740 spin_unlock(&lp->lp_lock);
1741 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1746 lpni = lnet_find_peer_ni_locked(nid4);
1749 * A peer_ni already exists. This is only a problem if
1750 * it is not connected to this peer and was configured
1753 if (lpni->lpni_peer_net->lpn_peer == lp)
1755 if (lnet_peer_ni_is_configured(lpni)) {
1759 /* If this is the primary NID, destroy the peer. */
1760 if (lnet_peer_ni_is_primary(lpni)) {
1761 struct lnet_peer *rtr_lp =
1762 lpni->lpni_peer_net->lpn_peer;
1763 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1765 * if we're trying to delete a router it means
1766 * we're moving this peer NI to a new peer so must
1767 * transfer router properties to the new peer
1769 if (rtr_refcount > 0) {
1770 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1771 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1773 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1774 lnet_peer_ni_decref_locked(lpni);
1775 lpni = lnet_peer_ni_alloc(&nid);
1782 lpni = lnet_peer_ni_alloc(&nid);
1790 * Get the peer_net. Check that we're not adding a second
1791 * peer_ni on a peer_net of a non-multi-rail peer.
1793 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid4));
1795 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid4));
1800 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1805 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1808 lnet_peer_ni_decref_locked(lpni);
1810 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1811 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nid2str(nid4),
1817 * Update the primary NID of a peer, if possible.
1819 * Call with the lnet_api_mutex held.
1822 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid,
1825 struct lnet_nid old = lp->lp_primary_nid;
1828 if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid)
1831 lnet_nid4_to_nid(nid, &lp->lp_primary_nid);
1833 rc = lnet_peer_add_nid(lp, nid, flags);
1835 lp->lp_primary_nid = old;
1839 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1840 libcfs_nidstr(&old), libcfs_nid2str(nid), rc);
1846 * lpni creation initiated due to traffic either sending or receiving.
1849 lnet_peer_ni_traffic_add(struct lnet_nid *nid, struct lnet_nid *pref)
1851 struct lnet_peer *lp;
1852 struct lnet_peer_net *lpn;
1853 struct lnet_peer_ni *lpni;
1857 if (LNET_NID_IS_ANY(nid)) {
1862 /* lnet_net_lock is not needed here because ln_api_lock is held */
1863 lpni = lnet_peer_ni_find_locked(nid);
1866 * We must have raced with another thread. Since we
1867 * know next to nothing about a peer_ni created by
1868 * traffic, we just assume everything is ok and
1871 lnet_peer_ni_decref_locked(lpni);
1875 /* Create peer, peer_net, and peer_ni. */
1877 lp = lnet_peer_alloc(nid);
1880 lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1883 lpni = lnet_peer_ni_alloc(nid);
1886 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1888 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1891 LIBCFS_FREE(lpn, sizeof(*lpn));
1893 LIBCFS_FREE(lp, sizeof(*lp));
1895 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(nid), rc);
1900 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1902 * This API handles the following combinations:
1903 * Create a peer with its primary NI if only the prim_nid is provided
1904 * Add a NID to a peer identified by the prim_nid. The peer identified
1905 * by the prim_nid must already exist.
1906 * The peer being created may be non-MR.
1908 * The caller must hold ln_api_mutex. This prevents the peer from
1909 * being created/modified/deleted by a different thread.
1912 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr, bool temp)
1914 struct lnet_peer *lp = NULL;
1915 struct lnet_peer_ni *lpni;
1916 unsigned int flags = 0;
1918 /* The prim_nid must always be specified */
1919 if (prim_nid == LNET_NID_ANY)
1923 flags = LNET_PEER_CONFIGURED;
1926 flags |= LNET_PEER_MULTI_RAIL;
1929 * If nid isn't specified, we must create a new peer with
1930 * prim_nid as its primary nid.
1932 if (nid == LNET_NID_ANY)
1933 return lnet_peer_add(prim_nid, flags);
1935 /* Look up the prim_nid, which must exist. */
1936 lpni = lnet_find_peer_ni_locked(prim_nid);
1939 lnet_peer_ni_decref_locked(lpni);
1940 lp = lpni->lpni_peer_net->lpn_peer;
1942 /* Peer must have been configured. */
1943 if (!temp && !(lp->lp_state & LNET_PEER_CONFIGURED)) {
1944 CDEBUG(D_NET, "peer %s was not configured\n",
1945 libcfs_nid2str(prim_nid));
1949 /* Primary NID must match */
1950 if (lnet_nid_to_nid4(&lp->lp_primary_nid) != prim_nid) {
1951 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1952 libcfs_nid2str(prim_nid),
1953 libcfs_nidstr(&lp->lp_primary_nid));
1957 /* Multi-Rail flag must match. */
1958 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1959 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1960 libcfs_nid2str(prim_nid));
1964 return lnet_peer_add_nid(lp, nid, flags);
1968 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1970 * This API handles the following combinations:
1971 * Delete a NI from a peer if both prim_nid and nid are provided.
1972 * Delete a peer if only prim_nid is provided.
1973 * Delete a peer if its primary nid is provided.
1975 * The caller must hold ln_api_mutex. This prevents the peer from
1976 * being modified/deleted by a different thread.
1979 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1981 struct lnet_peer *lp;
1982 struct lnet_peer_ni *lpni;
1985 if (prim_nid == LNET_NID_ANY)
1988 lpni = lnet_find_peer_ni_locked(prim_nid);
1991 lnet_peer_ni_decref_locked(lpni);
1992 lp = lpni->lpni_peer_net->lpn_peer;
1994 if (prim_nid != lnet_nid_to_nid4(&lp->lp_primary_nid)) {
1995 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1996 libcfs_nid2str(prim_nid),
1997 libcfs_nidstr(&lp->lp_primary_nid));
2001 lnet_net_lock(LNET_LOCK_EX);
2002 if (lp->lp_rtr_refcount > 0) {
2003 lnet_net_unlock(LNET_LOCK_EX);
2004 CERROR("%s is a router. Can not be deleted\n",
2005 libcfs_nid2str(prim_nid));
2008 lnet_net_unlock(LNET_LOCK_EX);
2010 if (nid == LNET_NID_ANY || nid == lnet_nid_to_nid4(&lp->lp_primary_nid))
2011 return lnet_peer_del(lp);
2013 flags = LNET_PEER_CONFIGURED;
2014 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2015 flags |= LNET_PEER_MULTI_RAIL;
2017 return lnet_peer_del_nid(lp, nid, flags);
2021 lnet_destroy_peer_ni_locked(struct kref *ref)
2023 struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
2025 struct lnet_peer_table *ptable;
2026 struct lnet_peer_net *lpn;
2028 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
2030 LASSERT(kref_read(&lpni->lpni_kref) == 0);
2031 LASSERT(list_empty(&lpni->lpni_txq));
2032 LASSERT(lpni->lpni_txqnob == 0);
2033 LASSERT(list_empty(&lpni->lpni_peer_nis));
2034 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
2036 lpn = lpni->lpni_peer_net;
2037 lpni->lpni_peer_net = NULL;
2038 lpni->lpni_net = NULL;
2040 if (!list_empty(&lpni->lpni_hashlist)) {
2041 /* remove the peer ni from the zombie list */
2042 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
2043 spin_lock(&ptable->pt_zombie_lock);
2044 list_del_init(&lpni->lpni_hashlist);
2045 ptable->pt_zombies--;
2046 spin_unlock(&ptable->pt_zombie_lock);
2049 if (lpni->lpni_pref_nnids > 1) {
2050 struct lnet_nid_list *ne, *tmp;
2052 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
2054 list_del_init(&ne->nl_list);
2055 LIBCFS_FREE(ne, sizeof(*ne));
2058 LIBCFS_FREE(lpni, sizeof(*lpni));
2061 lnet_peer_net_decref_locked(lpn);
2064 struct lnet_peer_ni *
2065 lnet_nid2peerni_ex(struct lnet_nid *nid, int cpt)
2067 struct lnet_peer_ni *lpni = NULL;
2070 if (the_lnet.ln_state != LNET_STATE_RUNNING)
2071 return ERR_PTR(-ESHUTDOWN);
2074 * find if a peer_ni already exists.
2075 * If so then just return that.
2077 lpni = lnet_peer_ni_find_locked(nid);
2081 lnet_net_unlock(cpt);
2083 rc = lnet_peer_ni_traffic_add(nid, NULL);
2086 goto out_net_relock;
2089 lpni = lnet_peer_ni_find_locked(nid);
2099 * Get a peer_ni for the given nid, create it if necessary. Takes a
2100 * hold on the peer_ni.
2102 struct lnet_peer_ni *
2103 lnet_peerni_by_nid_locked(struct lnet_nid *nid,
2104 struct lnet_nid *pref, int cpt)
2106 struct lnet_peer_ni *lpni = NULL;
2109 if (the_lnet.ln_state != LNET_STATE_RUNNING)
2110 return ERR_PTR(-ESHUTDOWN);
2113 * find if a peer_ni already exists.
2114 * If so then just return that.
2116 lpni = lnet_peer_ni_find_locked(nid);
2122 * use the lnet_api_mutex to serialize the creation of the peer_ni
2123 * and the creation/deletion of the local ni/net. When a local ni is
2124 * created, if there exists a set of peer_nis on that network,
2125 * they need to be traversed and updated. When a local NI is
2126 * deleted, which could result in a network being deleted, then
2127 * all peer nis on that network need to be removed as well.
2129 * Creation through traffic should also be serialized with
2130 * creation through DLC.
2132 lnet_net_unlock(cpt);
2133 mutex_lock(&the_lnet.ln_api_mutex);
2135 * Shutdown is only set under the ln_api_lock, so a single
2136 * check here is sufficent.
2138 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2139 lpni = ERR_PTR(-ESHUTDOWN);
2140 goto out_mutex_unlock;
2143 rc = lnet_peer_ni_traffic_add(nid, pref);
2146 goto out_mutex_unlock;
2149 lpni = lnet_peer_ni_find_locked(nid);
2153 mutex_unlock(&the_lnet.ln_api_mutex);
2156 /* Lock has been dropped, check again for shutdown. */
2157 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2159 lnet_peer_ni_decref_locked(lpni);
2160 lpni = ERR_PTR(-ESHUTDOWN);
2166 struct lnet_peer_ni *
2167 lnet_nid2peerni_locked(lnet_nid_t nid4, lnet_nid_t pref4, int cpt)
2169 struct lnet_nid nid, pref;
2171 lnet_nid4_to_nid(nid4, &nid);
2172 lnet_nid4_to_nid(pref4, &pref);
2173 if (pref4 == LNET_NID_ANY)
2174 return lnet_peerni_by_nid_locked(&nid, NULL, cpt);
2176 return lnet_peerni_by_nid_locked(&nid, &pref, cpt);
2180 lnet_peer_gw_discovery(struct lnet_peer *lp)
2184 spin_lock(&lp->lp_lock);
2185 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2187 spin_unlock(&lp->lp_lock);
2193 lnet_peer_is_uptodate(struct lnet_peer *lp)
2197 spin_lock(&lp->lp_lock);
2198 rc = lnet_peer_is_uptodate_locked(lp);
2199 spin_unlock(&lp->lp_lock);
2204 * Is a peer uptodate from the point of view of discovery?
2206 * If it is currently being processed, obviously not.
2207 * A forced Ping or Push is also handled by the discovery thread.
2209 * Otherwise look at whether the peer needs rediscovering.
2212 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2213 __must_hold(&lp->lp_lock)
2217 if (lp->lp_state & (LNET_PEER_DISCOVERING |
2218 LNET_PEER_FORCE_PING |
2219 LNET_PEER_FORCE_PUSH)) {
2221 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2223 } else if (lnet_peer_needs_push(lp)) {
2225 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2226 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2237 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2239 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2241 /* The discovery thread holds net_lock/EX and lp_lock when it splices
2242 * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2243 * when adding to the list and queuing the peer to ensure that we do not
2244 * strand any messages on the lp_dc_pendq. This scheme ensures the
2245 * message will be resent even if the peer is already being discovered.
2246 * Therefore we needn't check the return value of
2247 * lnet_peer_queue_for_discovery(lp).
2249 lnet_net_lock(LNET_LOCK_EX);
2250 spin_lock(&lp->lp_lock);
2251 list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2252 spin_unlock(&lp->lp_lock);
2253 lnet_peer_queue_for_discovery(lp);
2254 lnet_net_unlock(LNET_LOCK_EX);
2258 * Queue a peer for the attention of the discovery thread. Call with
2259 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2260 * -EALREADY if the peer was already queued.
2262 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2266 spin_lock(&lp->lp_lock);
2267 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2268 lp->lp_state |= LNET_PEER_DISCOVERING;
2269 spin_unlock(&lp->lp_lock);
2270 if (list_empty(&lp->lp_dc_list)) {
2271 lnet_peer_addref_locked(lp);
2272 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2273 wake_up(&the_lnet.ln_dc_waitq);
2279 CDEBUG(D_NET, "Queue peer %s: %d\n",
2280 libcfs_nidstr(&lp->lp_primary_nid), rc);
2286 * Discovery of a peer is complete. Wake all waiters on the peer.
2287 * Call with lnet_net_lock/EX held.
2289 static void lnet_peer_discovery_complete(struct lnet_peer *lp, int dc_error)
2291 struct lnet_msg *msg, *tmp;
2293 LIST_HEAD(pending_msgs);
2295 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2296 libcfs_nidstr(&lp->lp_primary_nid));
2298 list_del_init(&lp->lp_dc_list);
2299 spin_lock(&lp->lp_lock);
2301 lp->lp_dc_error = dc_error;
2302 lp->lp_state &= ~LNET_PEER_DISCOVERING;
2303 lp->lp_state |= LNET_PEER_REDISCOVER;
2305 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2306 spin_unlock(&lp->lp_lock);
2307 wake_up(&lp->lp_dc_waitq);
2309 if (lp->lp_rtr_refcount > 0)
2310 lnet_router_discovery_complete(lp);
2312 lnet_net_unlock(LNET_LOCK_EX);
2314 /* iterate through all pending messages and send them again */
2315 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2316 list_del_init(&msg->msg_list);
2318 lnet_finalize(msg, dc_error);
2322 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2323 lnet_msgtyp2str(msg->msg_type),
2324 libcfs_idstr(&msg->msg_target));
2325 rc = lnet_send(&msg->msg_src_nid_param, msg,
2326 &msg->msg_rtr_nid_param);
2328 CNETERR("Error sending %s to %s: %d\n",
2329 lnet_msgtyp2str(msg->msg_type),
2330 libcfs_idstr(&msg->msg_target), rc);
2331 lnet_finalize(msg, rc);
2334 lnet_net_lock(LNET_LOCK_EX);
2335 lnet_peer_decref_locked(lp);
2339 * Handle inbound push.
2340 * Like any event handler, called with lnet_res_lock/CPT held.
2342 void lnet_peer_push_event(struct lnet_event *ev)
2344 struct lnet_ping_buffer *pbuf;
2345 struct lnet_peer *lp;
2347 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2349 /* lnet_find_peer() adds a refcount */
2350 lp = lnet_find_peer(&ev->source.nid);
2352 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2353 libcfs_nidstr(&ev->initiator.nid),
2354 libcfs_nidstr(&ev->source.nid));
2355 pbuf->pb_needs_post = true;
2359 /* Ensure peer state remains consistent while we modify it. */
2360 spin_lock(&lp->lp_lock);
2363 * If some kind of error happened the contents of the message
2364 * cannot be used. Clear the NIDS_UPTODATE and set the
2365 * FORCE_PING flag to trigger a ping.
2368 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2369 lp->lp_state |= LNET_PEER_FORCE_PING;
2370 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2372 libcfs_nidstr(&lp->lp_primary_nid),
2373 libcfs_nidstr(&ev->source.nid));
2378 * A push with invalid or corrupted info. Clear the UPTODATE
2379 * flag to trigger a ping.
2381 if (lnet_ping_info_validate(&pbuf->pb_info)) {
2382 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2383 lp->lp_state |= LNET_PEER_FORCE_PING;
2384 CDEBUG(D_NET, "Corrupted Push from %s\n",
2385 libcfs_nidstr(&lp->lp_primary_nid));
2390 * Make sure we'll allocate the correct size ping buffer when
2393 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2394 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2397 * A non-Multi-Rail peer is not supposed to be capable of
2400 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2401 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2402 libcfs_nidstr(&lp->lp_primary_nid));
2407 * The peer may have discovery disabled at its end. Set
2408 * NO_DISCOVERY as appropriate.
2410 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2411 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2412 libcfs_nidstr(&lp->lp_primary_nid));
2414 * Mark the peer for deletion if we already know about it
2415 * and it's going from discovery set to no discovery set
2417 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2418 LNET_PEER_DISCOVERING)) &&
2419 lp->lp_state & LNET_PEER_DISCOVERED) {
2420 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2421 libcfs_nidstr(&lp->lp_primary_nid),
2423 lp->lp_state |= LNET_PEER_MARK_DELETION;
2425 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2426 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2427 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2428 libcfs_nidstr(&lp->lp_primary_nid));
2429 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2433 * Update the MULTI_RAIL flag based on the push. If the peer
2434 * was configured with DLC then the setting should match what
2436 * NB: We verified above that the MR feature bit is set in pi_features
2438 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2439 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2440 libcfs_nidstr(&lp->lp_primary_nid), lp);
2441 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2442 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2443 libcfs_nidstr(&lp->lp_primary_nid));
2444 } else if (lnet_peer_discovery_disabled) {
2445 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2446 libcfs_nidstr(&lp->lp_primary_nid), lp);
2447 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2448 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2449 libcfs_nidstr(&lp->lp_primary_nid), lp);
2451 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2452 libcfs_nidstr(&lp->lp_primary_nid), lp);
2453 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2454 lnet_peer_clr_non_mr_pref_nids(lp);
2458 * Check for truncation of the Put message. Clear the
2459 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2460 * and tell discovery to allocate a bigger buffer.
2462 if (ev->mlength < ev->rlength) {
2463 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2464 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2465 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2466 lp->lp_state |= LNET_PEER_FORCE_PING;
2467 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2468 libcfs_nidstr(&lp->lp_primary_nid),
2469 pbuf->pb_info.pi_nnis);
2473 /* always assume new data */
2474 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2475 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2478 * If there is data present that hasn't been processed yet,
2479 * we'll replace it if the Put contained newer data and it
2480 * fits. We're racing with a Ping or earlier Push in this
2483 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2484 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2485 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2486 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2487 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2488 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2489 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2490 libcfs_nidstr(&lp->lp_primary_nid),
2491 LNET_PING_BUFFER_SEQNO(pbuf),
2492 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2498 * Allocate a buffer to copy the data. On a failure we drop
2499 * the Push and set FORCE_PING to force the discovery
2500 * thread to fix the problem by pinging the peer.
2502 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2504 lp->lp_state |= LNET_PEER_FORCE_PING;
2505 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2506 libcfs_nidstr(&lp->lp_primary_nid),
2507 LNET_PING_BUFFER_SEQNO(pbuf));
2512 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2513 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2514 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2515 CDEBUG(D_NET, "Received Push %s %u\n",
2516 libcfs_nidstr(&lp->lp_primary_nid),
2517 LNET_PING_BUFFER_SEQNO(pbuf));
2520 /* We've processed this buffer. It can be reposted */
2521 pbuf->pb_needs_post = true;
2524 * Queue the peer for discovery if not done, force it on the request
2525 * queue and wake the discovery thread if the peer was already queued,
2526 * because its status changed.
2528 spin_unlock(&lp->lp_lock);
2529 lnet_net_lock(LNET_LOCK_EX);
2530 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2531 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2532 wake_up(&the_lnet.ln_dc_waitq);
2534 /* Drop refcount from lookup */
2535 lnet_peer_decref_locked(lp);
2536 lnet_net_unlock(LNET_LOCK_EX);
2540 * Clear the discovery error state, unless we're already discovering
2541 * this peer, in which case the error is current.
2543 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2545 spin_lock(&lp->lp_lock);
2546 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2547 lp->lp_dc_error = 0;
2548 spin_unlock(&lp->lp_lock);
2552 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2553 * dropped/retaken within this function. An lnet_peer_ni is passed in
2554 * because discovery could tear down an lnet_peer.
2557 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2560 struct lnet_peer *lp;
2565 lnet_net_unlock(cpt);
2566 lnet_net_lock(LNET_LOCK_EX);
2567 lp = lpni->lpni_peer_net->lpn_peer;
2568 lnet_peer_clear_discovery_error(lp);
2571 * We're willing to be interrupted. The lpni can become a
2572 * zombie if we race with DLC, so we must check for that.
2575 /* Keep lp alive when the lnet_net_lock is unlocked */
2576 lnet_peer_addref_locked(lp);
2577 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2578 if (signal_pending(current))
2580 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2583 * Don't repeat discovery if discovery is disabled. This is
2584 * done to ensure we can use discovery as a standard ping as
2585 * well for backwards compatibility with routers which do not
2586 * have discovery or have discovery disabled
2588 if (lnet_is_discovery_disabled(lp) && count > 0)
2590 if (lp->lp_dc_error)
2592 if (lnet_peer_is_uptodate(lp))
2594 if (lp->lp_state & LNET_PEER_MARK_DELETED)
2596 lnet_peer_queue_for_discovery(lp);
2598 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2601 * If caller requested a non-blocking operation then
2602 * return immediately. Once discovery is complete any
2603 * pending messages that were stopped due to discovery
2604 * will be transmitted.
2609 lnet_net_unlock(LNET_LOCK_EX);
2611 finish_wait(&lp->lp_dc_waitq, &wait);
2612 lnet_net_lock(LNET_LOCK_EX);
2613 lnet_peer_decref_locked(lp);
2614 /* Peer may have changed */
2615 lp = lpni->lpni_peer_net->lpn_peer;
2617 finish_wait(&lp->lp_dc_waitq, &wait);
2619 lnet_net_unlock(LNET_LOCK_EX);
2621 lnet_peer_decref_locked(lp);
2623 * The peer may have changed, so re-check and rediscover if that turns
2624 * out to have been the case. The reference count on lp ensured that
2625 * even if it was unlinked from lpni the memory could not be recycled.
2626 * Thus the check below is sufficient to determine whether the peer
2627 * changed. If the peer changed, then lp must not be dereferenced.
2629 if (lp != lpni->lpni_peer_net->lpn_peer)
2632 if (signal_pending(current))
2634 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2636 else if (lp->lp_dc_error)
2637 rc = lp->lp_dc_error;
2639 CDEBUG(D_NET, "non-blocking discovery\n");
2640 else if (!lnet_peer_is_uptodate(lp) &&
2641 !(lnet_is_discovery_disabled(lp) ||
2642 (lp->lp_state & LNET_PEER_MARK_DELETED)))
2645 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2646 (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2647 libcfs_nidstr(&lpni->lpni_nid), rc,
2648 (!block) ? "pending discovery" : "discovery complete");
2653 /* Handle an incoming ack for a push. */
2655 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2657 struct lnet_ping_buffer *pbuf;
2659 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2660 spin_lock(&lp->lp_lock);
2661 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2662 lp->lp_push_error = ev->status;
2664 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2666 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2667 spin_unlock(&lp->lp_lock);
2669 CDEBUG(D_NET, "peer %s ev->status %d\n",
2670 libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2673 /* Handle a Reply message. This is the reply to a Ping message. */
2675 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2677 struct lnet_ping_buffer *pbuf;
2680 spin_lock(&lp->lp_lock);
2682 lp->lp_disc_src_nid = ev->target.nid;
2683 lp->lp_disc_dst_nid = ev->source.nid;
2686 * If some kind of error happened the contents of message
2687 * cannot be used. Set PING_FAILED to trigger a retry.
2690 lp->lp_state |= LNET_PEER_PING_FAILED;
2691 lp->lp_ping_error = ev->status;
2692 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2694 libcfs_nidstr(&lp->lp_primary_nid),
2695 libcfs_nidstr(&ev->source.nid));
2699 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2700 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2701 lnet_swap_pinginfo(pbuf);
2704 * A reply with invalid or corrupted info. Set PING_FAILED to
2707 rc = lnet_ping_info_validate(&pbuf->pb_info);
2709 lp->lp_state |= LNET_PEER_PING_FAILED;
2710 lp->lp_ping_error = 0;
2711 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2712 libcfs_nidstr(&lp->lp_primary_nid), rc);
2717 * The peer may have discovery disabled at its end. Set
2718 * NO_DISCOVERY as appropriate.
2720 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) ||
2721 lnet_peer_discovery_disabled) {
2722 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2723 libcfs_nidstr(&lp->lp_primary_nid));
2725 /* Detect whether this peer has toggled discovery from on to
2726 * off and whether we can delete and re-create the peer. Peers
2727 * that were manually configured cannot be deleted by discovery.
2728 * We need to delete this peer and re-create it if the peer was
2729 * not configured manually, is currently considered DD capable,
2731 * 1. We've already discovered the peer (the peer has toggled
2732 * the discovery feature from on to off), or
2733 * 2. The peer is considered MR, but it was not user configured
2734 * (this was a "temporary" peer created via the kernel APIs
2735 * that we're discovering for the first time)
2737 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2738 LNET_PEER_NO_DISCOVERY)) &&
2739 (lp->lp_state & (LNET_PEER_DISCOVERED |
2740 LNET_PEER_MULTI_RAIL))) {
2741 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2742 libcfs_nidstr(&lp->lp_primary_nid),
2744 lp->lp_state |= LNET_PEER_MARK_DELETION;
2746 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2748 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2749 libcfs_nidstr(&lp->lp_primary_nid));
2750 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2754 * Update the MULTI_RAIL flag based on the reply. If the peer
2755 * was configured with DLC then the setting should match what
2758 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2759 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2760 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2761 libcfs_nidstr(&lp->lp_primary_nid), lp);
2762 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2763 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2764 libcfs_nidstr(&lp->lp_primary_nid));
2765 } else if (lnet_peer_discovery_disabled) {
2767 "peer %s(%p) not MR: DD disabled locally\n",
2768 libcfs_nidstr(&lp->lp_primary_nid), lp);
2769 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2771 "peer %s(%p) not MR: DD disabled remotely\n",
2772 libcfs_nidstr(&lp->lp_primary_nid), lp);
2774 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2775 libcfs_nidstr(&lp->lp_primary_nid), lp);
2776 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2777 lnet_peer_clr_non_mr_pref_nids(lp);
2779 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2780 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2781 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2782 libcfs_nidstr(&lp->lp_primary_nid));
2784 CERROR("Multi-Rail state vanished from %s\n",
2785 libcfs_nidstr(&lp->lp_primary_nid));
2786 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2791 * Make sure we'll allocate the correct size ping buffer when
2794 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2795 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2798 * Check for truncation of the Reply. Clear PING_SENT and set
2799 * PING_FAILED to trigger a retry.
2801 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2802 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2803 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2804 lp->lp_state |= LNET_PEER_PING_FAILED;
2805 lp->lp_ping_error = 0;
2806 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2807 libcfs_nidstr(&lp->lp_primary_nid),
2808 pbuf->pb_info.pi_nnis);
2813 * Check the sequence numbers in the reply. These are only
2814 * available if the reply came from a Multi-Rail peer.
2816 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2817 pbuf->pb_info.pi_nnis > 1 &&
2818 lnet_nid_to_nid4(&lp->lp_primary_nid) ==
2819 pbuf->pb_info.pi_ni[1].ns_nid) {
2820 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2821 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2822 libcfs_nidstr(&lp->lp_primary_nid),
2823 LNET_PING_BUFFER_SEQNO(pbuf),
2826 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2829 /* We're happy with the state of the data in the buffer. */
2830 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2831 libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2833 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2834 lnet_ping_buffer_decref(lp->lp_data);
2836 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2837 lnet_ping_buffer_addref(pbuf);
2840 lp->lp_state &= ~LNET_PEER_PING_SENT;
2841 spin_unlock(&lp->lp_lock);
2843 lnet_net_lock(LNET_LOCK_EX);
2845 * If this peer is a gateway, call the routing callback to
2846 * handle the ping reply
2848 if (lp->lp_rtr_refcount > 0)
2849 lnet_router_discovery_ping_reply(lp);
2850 lnet_net_unlock(LNET_LOCK_EX);
2854 * Send event handling. Only matters for error cases, where we clean
2855 * up state on the peer and peer_ni that would otherwise be updated in
2856 * the REPLY event handler for a successful Ping, and the ACK event
2857 * handler for a successful Push.
2860 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2867 spin_lock(&lp->lp_lock);
2868 if (ev->msg_type == LNET_MSG_GET) {
2869 lp->lp_state &= ~LNET_PEER_PING_SENT;
2870 lp->lp_state |= LNET_PEER_PING_FAILED;
2871 lp->lp_ping_error = ev->status;
2872 } else { /* ev->msg_type == LNET_MSG_PUT */
2873 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2874 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2875 lp->lp_push_error = ev->status;
2877 spin_unlock(&lp->lp_lock);
2878 rc = LNET_REDISCOVER_PEER;
2880 CDEBUG(D_NET, "%s Send to %s: %d\n",
2881 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2882 libcfs_nidstr(&ev->target.nid), rc);
2887 * Unlink event handling. This event is only seen if a call to
2888 * LNetMDUnlink() caused the event to be unlinked. If this call was
2889 * made after the event was set up in LNetGet() or LNetPut() then we
2890 * assume the Ping or Push timed out.
2893 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2895 spin_lock(&lp->lp_lock);
2896 /* We've passed through LNetGet() */
2897 if (lp->lp_state & LNET_PEER_PING_SENT) {
2898 lp->lp_state &= ~LNET_PEER_PING_SENT;
2899 lp->lp_state |= LNET_PEER_PING_FAILED;
2900 lp->lp_ping_error = -ETIMEDOUT;
2901 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2902 libcfs_nidstr(&lp->lp_primary_nid));
2904 /* We've passed through LNetPut() */
2905 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2906 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2907 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2908 lp->lp_push_error = -ETIMEDOUT;
2909 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2910 libcfs_nidstr(&lp->lp_primary_nid));
2912 spin_unlock(&lp->lp_lock);
2916 * Event handler for the discovery EQ.
2918 * Called with lnet_res_lock(cpt) held. The cpt is the
2919 * lnet_cpt_of_cookie() of the md handle cookie.
2921 static void lnet_discovery_event_handler(struct lnet_event *event)
2923 struct lnet_peer *lp = event->md_user_ptr;
2924 struct lnet_ping_buffer *pbuf;
2927 /* discovery needs to take another look */
2928 rc = LNET_REDISCOVER_PEER;
2930 CDEBUG(D_NET, "Received event: %d\n", event->type);
2932 switch (event->type) {
2933 case LNET_EVENT_ACK:
2934 lnet_discovery_event_ack(lp, event);
2936 case LNET_EVENT_REPLY:
2937 lnet_discovery_event_reply(lp, event);
2939 case LNET_EVENT_SEND:
2940 /* Only send failure triggers a retry. */
2941 rc = lnet_discovery_event_send(lp, event);
2943 case LNET_EVENT_UNLINK:
2944 /* LNetMDUnlink() was called */
2945 lnet_discovery_event_unlink(lp, event);
2948 /* Invalid events. */
2951 lnet_net_lock(LNET_LOCK_EX);
2952 if (event->unlinked) {
2953 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2954 lnet_ping_buffer_decref(pbuf);
2955 lnet_peer_decref_locked(lp);
2958 /* put peer back at end of request queue, if discovery not already
2960 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2961 lnet_peer_queue_for_discovery(lp)) {
2962 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2963 wake_up(&the_lnet.ln_dc_waitq);
2965 lnet_net_unlock(LNET_LOCK_EX);
2969 * Build a peer from incoming data.
2971 * The NIDs in the incoming data are supposed to be structured as follows:
2974 * - other NIDs in same net
2975 * - NIDs in second net
2976 * - NIDs in third net
2978 * This due to the way the list of NIDs in the data is created.
2980 * Note that this function will mark the peer uptodate unless an
2981 * ENOMEM is encontered. All other errors are due to a conflict
2982 * between the DLC configuration and what discovery sees. We treat DLC
2983 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2984 * peer from becoming stuck in discovery.
2986 static int lnet_peer_merge_data(struct lnet_peer *lp,
2987 struct lnet_ping_buffer *pbuf)
2989 struct lnet_peer_net *lpn;
2990 struct lnet_peer_ni *lpni;
2991 lnet_nid_t *curnis = NULL;
2992 struct lnet_ni_status *addnis = NULL;
2993 lnet_nid_t *delnis = NULL;
3003 flags = LNET_PEER_DISCOVERED;
3004 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3005 flags |= LNET_PEER_MULTI_RAIL;
3008 * Cache the routing feature for the peer; whether it is enabled
3009 * for disabled as reported by the remote peer.
3011 spin_lock(&lp->lp_lock);
3012 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
3013 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
3015 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
3016 spin_unlock(&lp->lp_lock);
3018 nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
3019 CFS_ALLOC_PTR_ARRAY(curnis, nnis);
3020 CFS_ALLOC_PTR_ARRAY(addnis, nnis);
3021 CFS_ALLOC_PTR_ARRAY(delnis, nnis);
3022 if (!curnis || !addnis || !delnis) {
3030 /* Construct the list of NIDs present in peer. */
3032 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
3033 curnis[ncurnis++] = lnet_nid_to_nid4(&lpni->lpni_nid);
3036 * Check for NIDs in pbuf not present in curnis[].
3037 * The loop starts at 1 to skip the loopback NID.
3039 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
3040 for (j = 0; j < ncurnis; j++)
3041 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
3044 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
3047 * Check for NIDs in curnis[] not present in pbuf.
3048 * The nested loop starts at 1 to skip the loopback NID.
3050 * But never add the loopback NID to delnis[]: if it is
3051 * present in curnis[] then this peer is for this node.
3053 for (i = 0; i < ncurnis; i++) {
3054 if (curnis[i] == LNET_NID_LO_0)
3056 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
3057 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
3059 * update the information we cache for the
3060 * peer with the latest information we
3063 lpni = lnet_find_peer_ni_locked(curnis[i]);
3065 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
3066 lnet_peer_ni_decref_locked(lpni);
3071 if (j == pbuf->pb_info.pi_nnis)
3072 delnis[ndelnis++] = curnis[i];
3076 * If we get here and the discovery is disabled then we don't want
3077 * to add or delete any NIs. We just updated the ones we have some
3078 * information on, and call it a day
3081 if (lnet_is_discovery_disabled(lp))
3084 for (i = 0; i < naddnis; i++) {
3085 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
3087 CERROR("Error adding NID %s to peer %s: %d\n",
3088 libcfs_nid2str(addnis[i].ns_nid),
3089 libcfs_nidstr(&lp->lp_primary_nid), rc);
3093 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
3095 lpni->lpni_ns_status = addnis[i].ns_status;
3096 lnet_peer_ni_decref_locked(lpni);
3100 for (i = 0; i < ndelnis; i++) {
3102 * for routers it's okay to delete the primary_nid because
3103 * the upper layers don't really rely on it. So if we're
3104 * being told that the router changed its primary_nid
3105 * then it's okay to delete it.
3107 if (lp->lp_rtr_refcount > 0)
3108 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3109 rc = lnet_peer_del_nid(lp, delnis[i], flags);
3111 CERROR("Error deleting NID %s from peer %s: %d\n",
3112 libcfs_nid2str(delnis[i]),
3113 libcfs_nidstr(&lp->lp_primary_nid), rc);
3119 /* The peer net for the primary NID should be the first entry in the
3120 * peer's lp_peer_nets list, and the peer NI for the primary NID should
3121 * be the first entry in its peer net's lpn_peer_nis list.
3123 lpni = lnet_find_peer_ni_locked(pbuf->pb_info.pi_ni[1].ns_nid);
3125 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3126 libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
3130 lnet_peer_ni_decref_locked(lpni);
3132 lpn = lpni->lpni_peer_net;
3133 if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3134 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3136 if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3137 list_move(&lpni->lpni_peer_nis,
3138 &lpni->lpni_peer_net->lpn_peer_nis);
3141 * Errors other than -ENOMEM are due to peers having been
3142 * configured with DLC. Ignore these because DLC overrides
3147 CFS_FREE_PTR_ARRAY(curnis, nnis);
3148 CFS_FREE_PTR_ARRAY(addnis, nnis);
3149 CFS_FREE_PTR_ARRAY(delnis, nnis);
3150 lnet_ping_buffer_decref(pbuf);
3151 CDEBUG(D_NET, "peer %s (%p): %d\n",
3152 libcfs_nidstr(&lp->lp_primary_nid), lp, rc);
3155 spin_lock(&lp->lp_lock);
3156 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
3157 lp->lp_state |= LNET_PEER_FORCE_PING;
3158 spin_unlock(&lp->lp_lock);
3164 * The data in pbuf says lp is its primary peer, but the data was
3165 * received by a different peer. Try to update lp with the data.
3168 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
3170 struct lnet_handle_md mdh;
3172 /* Queue lp for discovery, and force it on the request queue. */
3173 lnet_net_lock(LNET_LOCK_EX);
3174 if (lnet_peer_queue_for_discovery(lp))
3175 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3176 lnet_net_unlock(LNET_LOCK_EX);
3178 LNetInvalidateMDHandle(&mdh);
3181 * Decide whether we can move the peer to the DATA_PRESENT state.
3183 * We replace stale data for a multi-rail peer, repair PING_FAILED
3184 * status, and preempt FORCE_PING.
3186 * If after that we have DATA_PRESENT, we merge it into this peer.
3188 spin_lock(&lp->lp_lock);
3189 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3190 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3191 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3192 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3193 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3194 lnet_ping_buffer_decref(pbuf);
3199 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3200 lnet_ping_buffer_decref(lp->lp_data);
3202 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3204 if (lp->lp_state & LNET_PEER_PING_FAILED) {
3205 mdh = lp->lp_ping_mdh;
3206 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3207 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3208 lp->lp_ping_error = 0;
3210 if (lp->lp_state & LNET_PEER_FORCE_PING)
3211 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3212 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3213 spin_unlock(&lp->lp_lock);
3215 if (!LNetMDHandleIsInvalid(mdh))
3219 return lnet_peer_merge_data(lp, pbuf);
3221 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3225 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
3229 for (i = 0; i < pinfo->pi_nnis; i++) {
3230 if (pinfo->pi_ni[i].ns_nid == nid)
3237 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3238 * to the discovery queue a reference was taken that will prevent the peer from
3239 * actually being freed by this function. After this function exits the
3240 * discovery thread should call lnet_peer_discovery_complete() which will
3241 * drop that reference as well as wake any waiters that may also be holding a
3244 static int lnet_peer_deletion(struct lnet_peer *lp)
3245 __must_hold(&lp->lp_lock)
3247 struct list_head rlist;
3248 struct lnet_route *route, *tmp;
3249 int sensitivity = lp->lp_health_sensitivity;
3252 INIT_LIST_HEAD(&rlist);
3254 lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3255 LNET_PEER_FORCE_PUSH);
3256 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3257 libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3259 /* no-op if lnet_peer_del() has already been called on this peer */
3260 if (lp->lp_state & LNET_PEER_MARK_DELETED)
3263 spin_unlock(&lp->lp_lock);
3265 mutex_lock(&the_lnet.ln_api_mutex);
3266 if (the_lnet.ln_state != LNET_STATE_RUNNING ||
3267 the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
3268 mutex_unlock(&the_lnet.ln_api_mutex);
3269 spin_lock(&lp->lp_lock);
3273 lnet_net_lock(LNET_LOCK_EX);
3274 /* remove the peer from the discovery work
3275 * queue if it's on there in preparation
3278 if (!list_empty(&lp->lp_dc_list))
3279 list_del_init(&lp->lp_dc_list);
3280 list_for_each_entry_safe(route, tmp,
3283 lnet_move_route(route, NULL, &rlist);
3284 lnet_net_unlock(LNET_LOCK_EX);
3286 /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3287 rc = lnet_peer_del(lp);
3289 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3290 libcfs_nidstr(&lp->lp_primary_nid), rc);
3292 list_for_each_entry_safe(route, tmp,
3294 /* re-add these routes */
3295 lnet_add_route(route->lr_net,
3300 LIBCFS_FREE(route, sizeof(*route));
3303 mutex_unlock(&the_lnet.ln_api_mutex);
3305 spin_lock(&lp->lp_lock);
3311 * Update a peer using the data received.
3313 static int lnet_peer_data_present(struct lnet_peer *lp)
3314 __must_hold(&lp->lp_lock)
3316 struct lnet_ping_buffer *pbuf;
3317 struct lnet_peer_ni *lpni;
3318 lnet_nid_t nid = LNET_NID_ANY;
3324 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3325 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3326 spin_unlock(&lp->lp_lock);
3329 * Modifications of peer structures are done while holding the
3330 * ln_api_mutex. A global lock is required because we may be
3331 * modifying multiple peer structures, and a mutex greatly
3332 * simplifies memory management.
3334 * The actual changes to the data structures must also protect
3335 * against concurrent lookups, for which the lnet_net_lock in
3336 * LNET_LOCK_EX mode is used.
3338 mutex_lock(&the_lnet.ln_api_mutex);
3339 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3345 * If this peer is not on the peer list then it is being torn
3346 * down, and our reference count may be all that is keeping it
3347 * alive. Don't do any work on it.
3349 if (list_empty(&lp->lp_peer_list))
3352 flags = LNET_PEER_DISCOVERED;
3353 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3354 flags |= LNET_PEER_MULTI_RAIL;
3357 * Check whether the primary NID in the message matches the
3358 * primary NID of the peer. If it does, update the peer, if
3359 * it it does not, check whether there is already a peer with
3360 * that primary NID. If no such peer exists, try to update
3361 * the primary NID of the current peer (allowed if it was
3362 * created due to message traffic) and complete the update.
3363 * If the peer did exist, hand off the data to it.
3365 * The peer for the loopback interface is a special case: this
3366 * is the peer for the local node, and we want to set its
3367 * primary NID to the correct value here. Moreover, this peer
3368 * can show up with only the loopback NID in the ping buffer.
3370 if (pbuf->pb_info.pi_nnis <= 1) {
3371 lnet_ping_buffer_decref(pbuf);
3374 nid = pbuf->pb_info.pi_ni[1].ns_nid;
3375 if (nid_is_lo0(&lp->lp_primary_nid)) {
3376 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3378 rc = lnet_peer_merge_data(lp, pbuf);
3380 * if the primary nid of the peer is present in the ping info returned
3381 * from the peer, but it's not the local primary peer we have
3382 * cached and discovery is disabled, then we don't want to update
3383 * our local peer info, by adding or removing NIDs, we just want
3384 * to update the status of the nids that we currently have
3385 * recorded in that peer.
3387 } else if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid ||
3388 (lnet_is_nid_in_ping_info(lnet_nid_to_nid4(&lp->lp_primary_nid),
3390 lnet_is_discovery_disabled(lp))) {
3391 rc = lnet_peer_merge_data(lp, pbuf);
3393 lpni = lnet_find_peer_ni_locked(nid);
3394 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3395 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3397 CERROR("Primary NID error %s versus %s: %d\n",
3398 libcfs_nidstr(&lp->lp_primary_nid),
3399 libcfs_nid2str(nid), rc);
3401 rc = lnet_peer_merge_data(lp, pbuf);
3404 lnet_peer_ni_decref_locked(lpni);
3406 struct lnet_peer *new_lp;
3407 new_lp = lpni->lpni_peer_net->lpn_peer;
3409 * if lp has discovery/MR enabled that means new_lp
3410 * should have discovery/MR enabled as well, since
3411 * it's the same peer, which we're about to merge
3413 spin_lock(&lp->lp_lock);
3414 spin_lock(&new_lp->lp_lock);
3415 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3416 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3417 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3418 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3419 /* If we're processing a ping reply then we may be
3420 * about to send a push to the peer that we ping'd.
3421 * Since the ping reply that we're processing was
3422 * received by lp, we need to set the discovery source
3423 * NID for new_lp to the NID stored in lp.
3425 if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3426 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3427 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3429 spin_unlock(&new_lp->lp_lock);
3430 spin_unlock(&lp->lp_lock);
3432 rc = lnet_peer_set_primary_data(new_lp, pbuf);
3433 lnet_consolidate_routes_locked(lp, new_lp);
3434 lnet_peer_ni_decref_locked(lpni);
3438 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3439 libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3441 mutex_unlock(&the_lnet.ln_api_mutex);
3443 spin_lock(&lp->lp_lock);
3444 /* Tell discovery to re-check the peer immediately. */
3446 rc = LNET_REDISCOVER_PEER;
3451 * A ping failed. Clear the PING_FAILED state and set the
3452 * FORCE_PING state, to ensure a retry even if discovery is
3453 * disabled. This avoids being left with incorrect state.
3455 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3456 __must_hold(&lp->lp_lock)
3458 struct lnet_handle_md mdh;
3461 mdh = lp->lp_ping_mdh;
3462 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3463 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3464 lp->lp_state |= LNET_PEER_FORCE_PING;
3465 rc = lp->lp_ping_error;
3466 lp->lp_ping_error = 0;
3467 spin_unlock(&lp->lp_lock);
3469 if (!LNetMDHandleIsInvalid(mdh))
3472 CDEBUG(D_NET, "peer %s:%d\n",
3473 libcfs_nidstr(&lp->lp_primary_nid), rc);
3475 spin_lock(&lp->lp_lock);
3476 return rc ? rc : LNET_REDISCOVER_PEER;
3479 /* Active side of ping. */
3480 static int lnet_peer_send_ping(struct lnet_peer *lp)
3481 __must_hold(&lp->lp_lock)
3487 lp->lp_state |= LNET_PEER_PING_SENT;
3488 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3489 spin_unlock(&lp->lp_lock);
3491 cpt = lnet_net_lock_current();
3492 /* Refcount for MD. */
3493 lnet_peer_addref_locked(lp);
3494 lnet_net_unlock(cpt);
3496 nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3498 rc = lnet_send_ping(&lp->lp_primary_nid, &lp->lp_ping_mdh, nnis, lp,
3499 the_lnet.ln_dc_handler, false);
3502 * if LNetMDBind in lnet_send_ping fails we need to decrement the
3503 * refcount on the peer, otherwise LNetMDUnlink will be called
3504 * which will eventually do that.
3508 lnet_peer_decref_locked(lp);
3509 lnet_net_unlock(cpt);
3510 rc = -rc; /* change the rc to negative value */
3512 } else if (rc < 0) {
3516 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3518 spin_lock(&lp->lp_lock);
3522 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3524 * The errors that get us here are considered hard errors and
3525 * cause Discovery to terminate. So we clear PING_SENT, but do
3526 * not set either PING_FAILED or FORCE_PING. In fact we need
3527 * to clear PING_FAILED, because the unlink event handler will
3528 * have set it if we called LNetMDUnlink() above.
3530 spin_lock(&lp->lp_lock);
3531 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3536 * This function exists because you cannot call LNetMDUnlink() from an
3539 static int lnet_peer_push_failed(struct lnet_peer *lp)
3540 __must_hold(&lp->lp_lock)
3542 struct lnet_handle_md mdh;
3545 mdh = lp->lp_push_mdh;
3546 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3547 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3548 rc = lp->lp_push_error;
3549 lp->lp_push_error = 0;
3550 spin_unlock(&lp->lp_lock);
3552 if (!LNetMDHandleIsInvalid(mdh))
3555 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3556 spin_lock(&lp->lp_lock);
3557 return rc ? rc : LNET_REDISCOVER_PEER;
3561 * Mark the peer as discovered.
3563 static int lnet_peer_discovered(struct lnet_peer *lp)
3564 __must_hold(&lp->lp_lock)
3566 lp->lp_state |= LNET_PEER_DISCOVERED;
3567 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3568 LNET_PEER_REDISCOVER);
3570 lp->lp_dc_error = 0;
3572 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3577 /* Active side of push. */
3578 static int lnet_peer_send_push(struct lnet_peer *lp)
3579 __must_hold(&lp->lp_lock)
3581 struct lnet_ping_buffer *pbuf;
3582 struct lnet_processid id;
3587 /* Don't push to a non-multi-rail peer. */
3588 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3589 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3590 /* if peer's NIDs are uptodate then peer is discovered */
3591 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3592 rc = lnet_peer_discovered(lp);
3599 lp->lp_state |= LNET_PEER_PUSH_SENT;
3600 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3601 spin_unlock(&lp->lp_lock);
3603 cpt = lnet_net_lock_current();
3604 pbuf = the_lnet.ln_ping_target;
3605 lnet_ping_buffer_addref(pbuf);
3606 lnet_net_unlock(cpt);
3608 /* Push source MD */
3609 md.start = &pbuf->pb_info;
3610 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3611 md.threshold = 2; /* Put/Ack */
3613 md.options = LNET_MD_TRACK_RESPONSE;
3614 md.handler = the_lnet.ln_dc_handler;
3617 rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3619 lnet_ping_buffer_decref(pbuf);
3620 CERROR("Can't bind push source MD: %d\n", rc);
3624 cpt = lnet_net_lock_current();
3625 /* Refcount for MD. */
3626 lnet_peer_addref_locked(lp);
3627 id.pid = LNET_PID_LUSTRE;
3628 if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3629 id.nid = lp->lp_disc_dst_nid;
3631 id.nid = lp->lp_primary_nid;
3632 lnet_net_unlock(cpt);
3634 rc = LNetPut(&lp->lp_disc_src_nid, lp->lp_push_mdh,
3635 LNET_ACK_REQ, &id, LNET_RESERVED_PORTAL,
3636 LNET_PROTO_PING_MATCHBITS, 0, 0);
3639 * reset the discovery nid. There is no need to restrict sending
3640 * from that source, if we call lnet_push_update_to_peers(). It'll
3641 * get set to a specific NID, if we initiate discovery from the
3644 lp->lp_disc_src_nid = LNET_ANY_NID;
3645 lp->lp_disc_dst_nid = LNET_ANY_NID;
3650 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3652 spin_lock(&lp->lp_lock);
3656 LNetMDUnlink(lp->lp_push_mdh);
3657 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3659 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3662 * The errors that get us here are considered hard errors and
3663 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3664 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3665 * because the unlink event handler will have set it if we
3666 * called LNetMDUnlink() above.
3668 spin_lock(&lp->lp_lock);
3669 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3674 * Wait for work to be queued or some other change that must be
3675 * attended to. Returns non-zero if the discovery thread should shut
3678 static int lnet_peer_discovery_wait_for_work(void)
3685 cpt = lnet_net_lock_current();
3687 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3688 TASK_INTERRUPTIBLE);
3689 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3691 if (lnet_push_target_resize_needed() ||
3692 the_lnet.ln_push_target->pb_needs_post)
3694 if (!list_empty(&the_lnet.ln_dc_request))
3696 if (!list_empty(&the_lnet.ln_msg_resend))
3698 lnet_net_unlock(cpt);
3701 * wakeup max every second to check if there are peers that
3702 * have been stuck on the working queue for greater than
3705 schedule_timeout(cfs_time_seconds(1));
3706 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3707 cpt = lnet_net_lock_current();
3709 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3711 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3714 lnet_net_unlock(cpt);
3716 CDEBUG(D_NET, "woken: %d\n", rc);
3722 * Messages that were pending on a destroyed peer will be put on a global
3723 * resend list. The message resend list will be checked by
3724 * the discovery thread when it wakes up, and will resend messages. These
3725 * messages can still be sendable in the case the lpni which was the initial
3726 * cause of the message re-queue was transfered to another peer.
3728 * It is possible that LNet could be shutdown while we're iterating
3729 * through the list. lnet_shudown_lndnets() will attempt to access the
3730 * resend list, but will have to wait until the spinlock is released, by
3731 * which time there shouldn't be any more messages on the resend list.
3732 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3733 * for the messages so they can be released. The other case is that
3734 * lnet_shudown_lndnets() can finalize all the messages before this
3735 * function can visit the resend list, in which case this function will be
3738 static void lnet_resend_msgs(void)
3740 struct lnet_msg *msg, *tmp;
3744 spin_lock(&the_lnet.ln_msg_resend_lock);
3745 list_splice(&the_lnet.ln_msg_resend, &resend);
3746 spin_unlock(&the_lnet.ln_msg_resend_lock);
3748 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3749 list_del_init(&msg->msg_list);
3750 rc = lnet_send(&msg->msg_src_nid_param, msg,
3751 &msg->msg_rtr_nid_param);
3753 CNETERR("Error sending %s to %s: %d\n",
3754 lnet_msgtyp2str(msg->msg_type),
3755 libcfs_idstr(&msg->msg_target), rc);
3756 lnet_finalize(msg, rc);
3761 /* The discovery thread. */
3762 static int lnet_peer_discovery(void *arg)
3764 struct lnet_peer *lp;
3767 wait_for_completion(&the_lnet.ln_started);
3769 CDEBUG(D_NET, "started\n");
3772 if (lnet_peer_discovery_wait_for_work())
3775 if (lnet_push_target_resize_needed())
3776 lnet_push_target_resize();
3777 else if (the_lnet.ln_push_target->pb_needs_post)
3778 lnet_push_target_post(the_lnet.ln_push_target,
3779 &the_lnet.ln_push_target_md);
3783 lnet_net_lock(LNET_LOCK_EX);
3784 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3785 lnet_net_unlock(LNET_LOCK_EX);
3790 * Process all incoming discovery work requests. When
3791 * discovery must wait on a peer to change state, it
3792 * is added to the tail of the ln_dc_working queue. A
3793 * timestamp keeps track of when the peer was added,
3794 * so we can time out discovery requests that take too
3797 while (!list_empty(&the_lnet.ln_dc_request)) {
3798 lp = list_first_entry(&the_lnet.ln_dc_request,
3799 struct lnet_peer, lp_dc_list);
3800 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3802 * set the time the peer was put on the dc_working
3803 * queue. It shouldn't remain on the queue
3804 * forever, in case the GET message (for ping)
3805 * doesn't get a REPLY or the PUT message (for
3806 * push) doesn't get an ACK.
3808 lp->lp_last_queued = ktime_get_real_seconds();
3809 lnet_net_unlock(LNET_LOCK_EX);
3811 if (lnet_push_target_resize_needed())
3812 lnet_push_target_resize();
3813 else if (the_lnet.ln_push_target->pb_needs_post)
3814 lnet_push_target_post(the_lnet.ln_push_target,
3815 &the_lnet.ln_push_target_md);
3818 * Select an action depending on the state of
3819 * the peer and whether discovery is disabled.
3820 * The check whether discovery is disabled is
3821 * done after the code that handles processing
3822 * for arrived data, cleanup for failures, and
3823 * forcing a Ping or Push.
3825 spin_lock(&lp->lp_lock);
3826 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3827 libcfs_nidstr(&lp->lp_primary_nid), lp,
3829 if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3830 LNET_PEER_MARK_DELETED))
3831 rc = lnet_peer_deletion(lp);
3832 else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3833 rc = lnet_peer_data_present(lp);
3834 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3835 rc = lnet_peer_ping_failed(lp);
3836 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3837 rc = lnet_peer_push_failed(lp);
3838 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3839 rc = lnet_peer_send_ping(lp);
3840 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3841 rc = lnet_peer_send_push(lp);
3842 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3843 rc = lnet_peer_send_ping(lp);
3844 else if (lnet_peer_needs_push(lp))
3845 rc = lnet_peer_send_push(lp);
3847 rc = lnet_peer_discovered(lp);
3848 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3849 libcfs_nidstr(&lp->lp_primary_nid), lp,
3852 if (rc == LNET_REDISCOVER_PEER) {
3853 spin_unlock(&lp->lp_lock);
3854 lnet_net_lock(LNET_LOCK_EX);
3855 list_move(&lp->lp_dc_list,
3856 &the_lnet.ln_dc_request);
3858 !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3859 spin_unlock(&lp->lp_lock);
3860 lnet_net_lock(LNET_LOCK_EX);
3861 lnet_peer_discovery_complete(lp, rc);
3863 spin_unlock(&lp->lp_lock);
3864 lnet_net_lock(LNET_LOCK_EX);
3867 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3872 lnet_net_unlock(LNET_LOCK_EX);
3875 CDEBUG(D_NET, "stopping\n");
3877 * Clean up before telling lnet_peer_discovery_stop() that
3878 * we're done. Use wake_up() below to somewhat reduce the
3879 * size of the thundering herd if there are multiple threads
3880 * waiting on discovery of a single peer.
3883 /* Queue cleanup 1: stop all pending pings and pushes. */
3884 lnet_net_lock(LNET_LOCK_EX);
3885 while (!list_empty(&the_lnet.ln_dc_working)) {
3886 lp = list_first_entry(&the_lnet.ln_dc_working,
3887 struct lnet_peer, lp_dc_list);
3888 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3889 lnet_net_unlock(LNET_LOCK_EX);
3890 lnet_peer_cancel_discovery(lp);
3891 lnet_net_lock(LNET_LOCK_EX);
3893 lnet_net_unlock(LNET_LOCK_EX);
3895 /* Queue cleanup 2: wait for the expired queue to clear. */
3896 while (!list_empty(&the_lnet.ln_dc_expired))
3897 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3899 /* Queue cleanup 3: clear the request queue. */
3900 lnet_net_lock(LNET_LOCK_EX);
3901 while (!list_empty(&the_lnet.ln_dc_request)) {
3902 lp = list_first_entry(&the_lnet.ln_dc_request,
3903 struct lnet_peer, lp_dc_list);
3904 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3906 lnet_net_unlock(LNET_LOCK_EX);
3908 lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3909 the_lnet.ln_dc_handler = NULL;
3911 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3912 wake_up(&the_lnet.ln_dc_waitq);
3914 CDEBUG(D_NET, "stopped\n");
3919 /* ln_api_mutex is held on entry. */
3920 int lnet_peer_discovery_start(void)
3922 struct task_struct *task;
3925 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3928 the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3929 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3930 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3933 CERROR("Can't start peer discovery thread: %d\n", rc);
3935 the_lnet.ln_dc_handler = NULL;
3937 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3940 CDEBUG(D_NET, "discovery start: %d\n", rc);
3945 /* ln_api_mutex is held on entry. */
3946 void lnet_peer_discovery_stop(void)
3948 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3951 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3952 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3954 /* In the LNetNIInit() path we may be stopping discovery before it
3955 * entered its work loop
3957 if (!completion_done(&the_lnet.ln_started))
3958 complete(&the_lnet.ln_started);
3960 wake_up(&the_lnet.ln_dc_waitq);
3962 mutex_unlock(&the_lnet.ln_api_mutex);
3963 wait_event(the_lnet.ln_dc_waitq,
3964 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3965 mutex_lock(&the_lnet.ln_api_mutex);
3967 LASSERT(list_empty(&the_lnet.ln_dc_request));
3968 LASSERT(list_empty(&the_lnet.ln_dc_working));
3969 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3971 CDEBUG(D_NET, "discovery stopped\n");
3977 lnet_debug_peer(lnet_nid_t nid)
3979 char *aliveness = "NA";
3980 struct lnet_peer_ni *lp;
3983 cpt = lnet_cpt_of_nid(nid, NULL);
3986 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3988 lnet_net_unlock(cpt);
3989 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3993 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3994 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3996 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3997 libcfs_nidstr(&lp->lpni_nid), kref_read(&lp->lpni_kref),
3998 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3999 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
4000 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
4002 lnet_peer_ni_decref_locked(lp);
4004 lnet_net_unlock(cpt);
4007 /* Gathering information for userspace. */
4009 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
4010 char aliveness[LNET_MAX_STR_LEN],
4011 __u32 *cpt_iter, __u32 *refcount,
4012 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
4013 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
4014 __u32 *peer_tx_qnob)
4016 struct lnet_peer_table *peer_table;
4017 struct lnet_peer_ni *lp;
4022 /* get the number of CPTs */
4023 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4025 /* if the cpt number to be examined is >= the number of cpts in
4026 * the system then indicate that there are no more cpts to examin
4028 if (*cpt_iter >= lncpt)
4031 /* get the current table */
4032 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
4033 /* if the ptable is NULL then there are no more cpts to examine */
4034 if (peer_table == NULL)
4037 lnet_net_lock(*cpt_iter);
4039 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
4040 struct list_head *peers = &peer_table->pt_hash[j];
4042 list_for_each_entry(lp, peers, lpni_hashlist) {
4043 if (!nid_is_nid4(&lp->lpni_nid))
4045 if (peer_index-- > 0)
4048 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
4049 if (lnet_isrouter(lp) ||
4050 lnet_peer_aliveness_enabled(lp))
4051 snprintf(aliveness, LNET_MAX_STR_LEN,
4052 lnet_is_peer_ni_alive(lp) ? "up" : "down");
4054 *nid = lnet_nid_to_nid4(&lp->lpni_nid);
4055 *refcount = kref_read(&lp->lpni_kref);
4056 *ni_peer_tx_credits =
4057 lp->lpni_net->net_tunables.lct_peer_tx_credits;
4058 *peer_tx_credits = lp->lpni_txcredits;
4059 *peer_rtr_credits = lp->lpni_rtrcredits;
4060 *peer_min_rtr_credits = lp->lpni_mintxcredits;
4061 *peer_tx_qnob = lp->lpni_txqnob;
4067 lnet_net_unlock(*cpt_iter);
4071 return found ? 0 : -ENOENT;
4074 /* ln_api_mutex is held, which keeps the peer list stable */
4075 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
4077 struct lnet_ioctl_element_stats *lpni_stats;
4078 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
4079 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
4080 struct lnet_peer_ni_credit_info *lpni_info;
4081 struct lnet_peer_ni *lpni;
4082 struct lnet_peer *lp;
4087 lp = lnet_find_peer4(cfg->prcfg_prim_nid);
4094 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
4095 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
4096 size *= lp->lp_nnis;
4097 if (size > cfg->prcfg_size) {
4098 cfg->prcfg_size = size;
4103 cfg->prcfg_prim_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4104 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
4105 cfg->prcfg_cfg_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4106 cfg->prcfg_count = lp->lp_nnis;
4107 cfg->prcfg_size = size;
4108 cfg->prcfg_state = lp->lp_state;
4110 /* Allocate helper buffers. */
4112 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
4115 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
4118 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
4119 if (!lpni_msg_stats)
4120 goto out_free_stats;
4121 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
4123 goto out_free_msg_stats;
4128 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
4129 if (!nid_is_nid4(&lpni->lpni_nid))
4131 nid = lnet_nid_to_nid4(&lpni->lpni_nid);
4132 if (copy_to_user(bulk, &nid, sizeof(nid)))
4133 goto out_free_hstats;
4134 bulk += sizeof(nid);
4136 memset(lpni_info, 0, sizeof(*lpni_info));
4137 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
4138 if (lnet_isrouter(lpni) ||
4139 lnet_peer_aliveness_enabled(lpni))
4140 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
4141 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
4143 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
4144 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
4145 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
4146 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
4147 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
4148 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
4149 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
4150 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4151 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4152 goto out_free_hstats;
4153 bulk += sizeof(*lpni_info);
4155 memset(lpni_stats, 0, sizeof(*lpni_stats));
4156 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4157 LNET_STATS_TYPE_SEND);
4158 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4159 LNET_STATS_TYPE_RECV);
4160 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4161 LNET_STATS_TYPE_DROP);
4162 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4163 goto out_free_hstats;
4164 bulk += sizeof(*lpni_stats);
4165 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4166 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4167 goto out_free_hstats;
4168 bulk += sizeof(*lpni_msg_stats);
4169 lpni_hstats->hlpni_network_timeout =
4170 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4171 lpni_hstats->hlpni_remote_dropped =
4172 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4173 lpni_hstats->hlpni_remote_timeout =
4174 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4175 lpni_hstats->hlpni_remote_error =
4176 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4177 lpni_hstats->hlpni_health_value =
4178 atomic_read(&lpni->lpni_healthv);
4179 lpni_hstats->hlpni_ping_count = lpni->lpni_ping_count;
4180 lpni_hstats->hlpni_next_ping = lpni->lpni_next_ping;
4181 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4182 goto out_free_hstats;
4183 bulk += sizeof(*lpni_hstats);
4188 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4190 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4192 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4194 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4196 lnet_peer_decref_locked(lp);
4201 /* must hold net_lock/0 */
4203 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4204 struct list_head *recovery_queue,
4207 /* the mt could've shutdown and cleaned up the queues */
4208 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4211 if (!list_empty(&lpni->lpni_recovery))
4214 if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4217 if (!lpni->lpni_last_alive) {
4219 "lpni %s(%p) not eligible for recovery last alive %lld\n",
4220 libcfs_nidstr(&lpni->lpni_nid), lpni,
4221 lpni->lpni_last_alive);
4225 if (lnet_recovery_limit &&
4226 now > lpni->lpni_last_alive + lnet_recovery_limit) {
4227 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4228 libcfs_nidstr(&lpni->lpni_nid),
4229 lpni->lpni_last_alive);
4230 /* Reset the ping count so that if this peer NI is added back to
4231 * the recovery queue we will send the first ping right away.
4233 lpni->lpni_ping_count = 0;
4237 /* This peer NI is going on the recovery queue, so take a ref on it */
4238 lnet_peer_ni_addref_locked(lpni);
4240 lnet_peer_ni_set_next_ping(lpni, now);
4242 CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4243 libcfs_nidstr(&lpni->lpni_nid),
4244 lpni->lpni_ping_count,
4245 lpni->lpni_next_ping,
4246 lpni->lpni_last_alive,
4247 atomic_read(&lpni->lpni_healthv));
4249 list_add_tail(&lpni->lpni_recovery, recovery_queue);
4252 /* Call with the ln_api_mutex held */
4254 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
4256 struct lnet_peer_table *ptable;
4257 struct lnet_peer *lp;
4258 struct lnet_peer_net *lpn;
4259 struct lnet_peer_ni *lpni;
4264 if (the_lnet.ln_state != LNET_STATE_RUNNING)
4267 now = ktime_get_seconds();
4270 lnet_net_lock(LNET_LOCK_EX);
4271 lpni = lnet_find_peer_ni_locked(nid);
4273 lnet_net_unlock(LNET_LOCK_EX);
4276 lnet_set_lpni_healthv_locked(lpni, value);
4277 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4278 &the_lnet.ln_mt_peerNIRecovq, now);
4279 lnet_peer_ni_decref_locked(lpni);
4280 lnet_net_unlock(LNET_LOCK_EX);
4284 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4287 * Walk all the peers and reset the health value for each one to the
4290 lnet_net_lock(LNET_LOCK_EX);
4291 for (cpt = 0; cpt < lncpt; cpt++) {
4292 ptable = the_lnet.ln_peer_tables[cpt];
4293 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4294 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4295 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4297 lnet_set_lpni_healthv_locked(lpni,
4299 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4300 &the_lnet.ln_mt_peerNIRecovq, now);
4305 lnet_net_unlock(LNET_LOCK_EX);