4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
35 #define DEBUG_SUBSYSTEM S_LNET
37 #include <linux/sched.h>
38 #ifdef HAVE_SCHED_HEADERS
39 #include <linux/sched/signal.h>
41 #include <linux/uaccess.h>
43 #include <lnet/lib-lnet.h>
44 #include <uapi/linux/lnet/lnet-dlc.h>
46 /* Value indicating that recovery needs to re-check a peer immediately. */
47 #define LNET_REDISCOVER_PEER (1)
49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
54 if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
55 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
56 lnet_peer_ni_decref_locked(lpni);
61 lnet_peer_net_added(struct lnet_net *net)
63 struct lnet_peer_ni *lpni, *tmp;
65 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
66 lpni_on_remote_peer_ni_list) {
68 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
71 spin_lock(&lpni->lpni_lock);
72 lpni->lpni_txcredits =
73 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
74 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
75 lpni->lpni_rtrcredits =
76 lnet_peer_buffer_credits(lpni->lpni_net);
77 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
78 spin_unlock(&lpni->lpni_lock);
80 lnet_peer_remove_from_remote_list(lpni);
86 lnet_peer_tables_destroy(void)
88 struct lnet_peer_table *ptable;
89 struct list_head *hash;
93 if (!the_lnet.ln_peer_tables)
96 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
97 hash = ptable->pt_hash;
98 if (!hash) /* not intialized */
101 LASSERT(list_empty(&ptable->pt_zombie_list));
103 ptable->pt_hash = NULL;
104 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
105 LASSERT(list_empty(&hash[j]));
107 LIBCFS_FREE(hash, LNET_PEER_HASH_SIZE * sizeof(*hash));
110 cfs_percpt_free(the_lnet.ln_peer_tables);
111 the_lnet.ln_peer_tables = NULL;
115 lnet_peer_tables_create(void)
117 struct lnet_peer_table *ptable;
118 struct list_head *hash;
122 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
124 if (the_lnet.ln_peer_tables == NULL) {
125 CERROR("Failed to allocate cpu-partition peer tables\n");
129 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131 LNET_PEER_HASH_SIZE * sizeof(*hash));
133 CERROR("Failed to create peer hash table\n");
134 lnet_peer_tables_destroy();
138 spin_lock_init(&ptable->pt_zombie_lock);
139 INIT_LIST_HEAD(&ptable->pt_zombie_list);
141 INIT_LIST_HEAD(&ptable->pt_peer_list);
143 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144 INIT_LIST_HEAD(&hash[j]);
145 ptable->pt_hash = hash; /* sign of initialization */
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(lnet_nid_t nid)
154 struct lnet_peer_ni *lpni;
155 struct lnet_net *net;
158 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
160 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
164 INIT_LIST_HEAD(&lpni->lpni_txq);
165 INIT_LIST_HEAD(&lpni->lpni_hashlist);
166 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167 INIT_LIST_HEAD(&lpni->lpni_recovery);
168 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171 spin_lock_init(&lpni->lpni_lock);
173 if (lnet_peers_start_down())
174 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
176 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
177 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
178 lpni->lpni_nid = nid;
179 lpni->lpni_cpt = cpt;
180 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
182 net = lnet_get_net_locked(LNET_NIDNET(nid));
183 lpni->lpni_net = net;
185 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
186 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
187 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
188 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
191 * This peer_ni is not on a local network, so we
192 * cannot add the credits here. In case the net is
193 * added later, add the peer_ni to the remote peer ni
194 * list so it can be easily found and revisited.
196 /* FIXME: per-net implementation instead? */
197 atomic_inc(&lpni->lpni_refcount);
198 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
199 &the_lnet.ln_remote_peer_ni_list);
202 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
207 static struct lnet_peer_net *
208 lnet_peer_net_alloc(__u32 net_id)
210 struct lnet_peer_net *lpn;
212 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
216 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
217 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
218 lpn->lpn_net_id = net_id;
220 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
226 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
228 struct lnet_peer *lp;
230 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
232 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
233 LASSERT(list_empty(&lpn->lpn_peer_nis));
234 LASSERT(list_empty(&lpn->lpn_peer_nets));
236 lpn->lpn_peer = NULL;
237 LIBCFS_FREE(lpn, sizeof(*lpn));
239 lnet_peer_decref_locked(lp);
242 static struct lnet_peer *
243 lnet_peer_alloc(lnet_nid_t nid)
245 struct lnet_peer *lp;
247 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
251 INIT_LIST_HEAD(&lp->lp_rtrq);
252 INIT_LIST_HEAD(&lp->lp_routes);
253 INIT_LIST_HEAD(&lp->lp_peer_list);
254 INIT_LIST_HEAD(&lp->lp_peer_nets);
255 INIT_LIST_HEAD(&lp->lp_dc_list);
256 INIT_LIST_HEAD(&lp->lp_dc_pendq);
257 INIT_LIST_HEAD(&lp->lp_rtr_list);
258 init_waitqueue_head(&lp->lp_dc_waitq);
259 spin_lock_init(&lp->lp_lock);
260 lp->lp_primary_nid = nid;
261 if (lnet_peers_start_down())
262 lp->lp_alive = false;
267 * all peers created on a router should have health on
268 * if it's not already on.
270 if (the_lnet.ln_routing && !lnet_health_sensitivity)
271 lp->lp_health_sensitivity = 1;
274 * Turn off discovery for loopback peer. If you're creating a peer
275 * for the loopback interface then that was initiated when we
276 * attempted to send a message over the loopback. There is no need
277 * to ever use a different interface when sending messages to
280 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
281 lp->lp_state = LNET_PEER_NO_DISCOVERY;
282 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
284 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
290 lnet_destroy_peer_locked(struct lnet_peer *lp)
292 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
294 LASSERT(atomic_read(&lp->lp_refcount) == 0);
295 LASSERT(lp->lp_rtr_refcount == 0);
296 LASSERT(list_empty(&lp->lp_peer_nets));
297 LASSERT(list_empty(&lp->lp_peer_list));
298 LASSERT(list_empty(&lp->lp_dc_list));
301 lnet_ping_buffer_decref(lp->lp_data);
304 * if there are messages still on the pending queue, then make
305 * sure to queue them on the ln_msg_resend list so they can be
306 * resent at a later point if the discovery thread is still
308 * If the discovery thread has stopped, then the wakeup will be a
309 * no-op, and it is expected the lnet_shutdown_lndnets() will
310 * eventually be called, which will traverse this list and
311 * finalize the messages on the list.
312 * We can not resend them now because we're holding the cpt lock.
313 * Releasing the lock can cause an inconsistent state
315 spin_lock(&the_lnet.ln_msg_resend_lock);
316 spin_lock(&lp->lp_lock);
317 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
318 spin_unlock(&lp->lp_lock);
319 spin_unlock(&the_lnet.ln_msg_resend_lock);
320 wake_up(&the_lnet.ln_dc_waitq);
322 LIBCFS_FREE(lp, sizeof(*lp));
326 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
327 * that peer_net, detach the peer_net from the peer.
329 * Call with lnet_net_lock/EX held
332 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
334 struct lnet_peer_table *ptable;
335 struct lnet_peer_net *lpn;
336 struct lnet_peer *lp;
339 * Belts and suspenders: gracefully handle teardown of a
340 * partially connected peer_ni.
342 lpn = lpni->lpni_peer_net;
344 list_del_init(&lpni->lpni_peer_nis);
346 * If there are no lpni's left, we detach lpn from
347 * lp_peer_nets, so it cannot be found anymore.
349 if (list_empty(&lpn->lpn_peer_nis))
350 list_del_init(&lpn->lpn_peer_nets);
352 /* Update peer NID count. */
357 * If there are no more peer nets, make the peer unfindable
358 * via the peer_tables.
360 * Otherwise, if the peer is DISCOVERED, tell discovery to
361 * take another look at it. This is a no-op if discovery for
362 * this peer did the detaching.
364 if (list_empty(&lp->lp_peer_nets)) {
365 list_del_init(&lp->lp_peer_list);
366 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
368 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
369 /* Discovery isn't running, nothing to do here. */
370 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
371 lnet_peer_queue_for_discovery(lp);
372 wake_up(&the_lnet.ln_dc_waitq);
374 CDEBUG(D_NET, "peer %s NID %s\n",
375 libcfs_nid2str(lp->lp_primary_nid),
376 libcfs_nid2str(lpni->lpni_nid));
379 /* called with lnet_net_lock LNET_LOCK_EX held */
381 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
383 struct lnet_peer_table *ptable = NULL;
385 /* don't remove a peer_ni if it's also a gateway */
386 if (lnet_isrouter(lpni) && !force) {
387 CERROR("Peer NI %s is a gateway. Can not delete it\n",
388 libcfs_nid2str(lpni->lpni_nid));
392 lnet_peer_remove_from_remote_list(lpni);
394 /* remove peer ni from the hash list. */
395 list_del_init(&lpni->lpni_hashlist);
398 * indicate the peer is being deleted so the monitor thread can
399 * remove it from the recovery queue.
401 spin_lock(&lpni->lpni_lock);
402 lpni->lpni_state |= LNET_PEER_NI_DELETING;
403 spin_unlock(&lpni->lpni_lock);
405 /* decrement the ref count on the peer table */
406 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
409 * The peer_ni can no longer be found with a lookup. But there
410 * can be current users, so keep track of it on the zombie
411 * list until the reference count has gone to zero.
413 * The last reference may be lost in a place where the
414 * lnet_net_lock locks only a single cpt, and that cpt may not
415 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
418 spin_lock(&ptable->pt_zombie_lock);
419 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
420 ptable->pt_zombies++;
421 spin_unlock(&ptable->pt_zombie_lock);
423 /* no need to keep this peer_ni on the hierarchy anymore */
424 lnet_peer_detach_peer_ni_locked(lpni);
426 /* remove hashlist reference on peer_ni */
427 lnet_peer_ni_decref_locked(lpni);
432 void lnet_peer_uninit(void)
434 struct lnet_peer_ni *lpni, *tmp;
436 lnet_net_lock(LNET_LOCK_EX);
438 /* remove all peer_nis from the remote peer and the hash list */
439 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
440 lpni_on_remote_peer_ni_list)
441 lnet_peer_ni_del_locked(lpni, false);
443 lnet_peer_tables_destroy();
445 lnet_net_unlock(LNET_LOCK_EX);
449 lnet_peer_del_locked(struct lnet_peer *peer)
451 struct lnet_peer_ni *lpni = NULL, *lpni2;
454 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
456 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
457 while (lpni != NULL) {
458 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
459 rc = lnet_peer_ni_del_locked(lpni, false);
469 lnet_peer_del(struct lnet_peer *peer)
471 lnet_net_lock(LNET_LOCK_EX);
472 lnet_peer_del_locked(peer);
473 lnet_net_unlock(LNET_LOCK_EX);
479 * Delete a NID from a peer. Call with ln_api_mutex held.
482 * -EPERM: Non-DLC deletion from DLC-configured peer.
483 * -ENOENT: No lnet_peer_ni corresponding to the nid.
484 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
485 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
488 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
490 struct lnet_peer_ni *lpni;
491 lnet_nid_t primary_nid = lp->lp_primary_nid;
493 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
495 if (!(flags & LNET_PEER_CONFIGURED)) {
496 if (lp->lp_state & LNET_PEER_CONFIGURED) {
501 lpni = lnet_find_peer_ni_locked(nid);
506 lnet_peer_ni_decref_locked(lpni);
507 if (lp != lpni->lpni_peer_net->lpn_peer) {
513 * This function only allows deletion of the primary NID if it
516 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
521 lnet_net_lock(LNET_LOCK_EX);
523 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
524 struct lnet_peer_ni *lpni2;
525 /* assign the next peer_ni to be the primary */
526 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
528 lp->lp_primary_nid = lpni->lpni_nid;
530 rc = lnet_peer_ni_del_locked(lpni, force);
532 lnet_net_unlock(LNET_LOCK_EX);
535 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
536 libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
542 lnet_peer_table_cleanup_locked(struct lnet_net *net,
543 struct lnet_peer_table *ptable)
546 struct lnet_peer_ni *next;
547 struct lnet_peer_ni *lpni;
548 struct lnet_peer *peer;
550 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
551 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
553 if (net != NULL && net != lpni->lpni_net)
556 peer = lpni->lpni_peer_net->lpn_peer;
557 if (peer->lp_primary_nid != lpni->lpni_nid) {
558 lnet_peer_ni_del_locked(lpni, false);
562 * Removing the primary NID implies removing
563 * the entire peer. Advance next beyond any
564 * peer_ni that belongs to the same peer.
566 list_for_each_entry_from(next, &ptable->pt_hash[i],
568 if (next->lpni_peer_net->lpn_peer != peer)
571 lnet_peer_del_locked(peer);
577 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
581 spin_lock(&ptable->pt_zombie_lock);
582 while (ptable->pt_zombies) {
583 spin_unlock(&ptable->pt_zombie_lock);
585 if (is_power_of_2(i)) {
587 "Waiting for %d zombies on peer table\n",
590 set_current_state(TASK_UNINTERRUPTIBLE);
591 schedule_timeout(cfs_time_seconds(1) >> 1);
592 spin_lock(&ptable->pt_zombie_lock);
594 spin_unlock(&ptable->pt_zombie_lock);
598 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
599 struct lnet_peer_table *ptable)
601 struct lnet_peer_ni *lp;
602 struct lnet_peer_ni *tmp;
606 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
607 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
609 if (net != lp->lpni_net)
612 if (!lnet_isrouter(lp))
615 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
617 lnet_net_unlock(LNET_LOCK_EX);
618 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
619 lnet_net_lock(LNET_LOCK_EX);
625 lnet_peer_tables_cleanup(struct lnet_net *net)
628 struct lnet_peer_table *ptable;
630 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
631 /* If just deleting the peers for a NI, get rid of any routes these
632 * peers are gateways for. */
633 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
634 lnet_net_lock(LNET_LOCK_EX);
635 lnet_peer_table_del_rtrs_locked(net, ptable);
636 lnet_net_unlock(LNET_LOCK_EX);
639 /* Start the cleanup process */
640 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
641 lnet_net_lock(LNET_LOCK_EX);
642 lnet_peer_table_cleanup_locked(net, ptable);
643 lnet_net_unlock(LNET_LOCK_EX);
646 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
647 lnet_peer_ni_finalize_wait(ptable);
650 static struct lnet_peer_ni *
651 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
653 struct list_head *peers;
654 struct lnet_peer_ni *lp;
656 if (the_lnet.ln_state != LNET_STATE_RUNNING)
659 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
660 list_for_each_entry(lp, peers, lpni_hashlist) {
661 if (lp->lpni_nid == nid) {
662 lnet_peer_ni_addref_locked(lp);
670 struct lnet_peer_ni *
671 lnet_find_peer_ni_locked(lnet_nid_t nid)
673 struct lnet_peer_ni *lpni;
674 struct lnet_peer_table *ptable;
677 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
679 ptable = the_lnet.ln_peer_tables[cpt];
680 lpni = lnet_get_peer_ni_locked(ptable, nid);
685 struct lnet_peer_ni *
686 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
688 struct lnet_peer_net *lpn;
689 struct lnet_peer_ni *lpni;
691 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
695 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
696 if (lpni->lpni_nid == nid)
704 lnet_find_peer(lnet_nid_t nid)
706 struct lnet_peer_ni *lpni;
707 struct lnet_peer *lp = NULL;
710 cpt = lnet_net_lock_current();
711 lpni = lnet_find_peer_ni_locked(nid);
713 lp = lpni->lpni_peer_net->lpn_peer;
714 lnet_peer_addref_locked(lp);
715 lnet_peer_ni_decref_locked(lpni);
717 lnet_net_unlock(cpt);
722 struct lnet_peer_net *
723 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
725 struct lnet_peer_net *net;
728 /* no net id provided return the first net */
729 net = list_first_entry_or_null(&lp->lp_peer_nets,
730 struct lnet_peer_net,
736 /* find the net after the one provided */
737 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
738 if (net->lpn_net_id == prev_lpn_id) {
740 * if we reached the end of the list loop to the
743 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
744 return list_first_entry_or_null(&lp->lp_peer_nets,
745 struct lnet_peer_net,
748 return list_next_entry(net, lpn_peer_nets);
755 struct lnet_peer_ni *
756 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
757 struct lnet_peer_net *peer_net,
758 struct lnet_peer_ni *prev)
760 struct lnet_peer_ni *lpni;
761 struct lnet_peer_net *net = peer_net;
765 if (list_empty(&peer->lp_peer_nets))
768 net = list_entry(peer->lp_peer_nets.next,
769 struct lnet_peer_net,
772 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
778 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
780 * if you reached the end of the peer ni list and the peer
781 * net is specified then there are no more peer nis in that
788 * we reached the end of this net ni list. move to the
791 if (prev->lpni_peer_net->lpn_peer_nets.next ==
793 /* no more nets and no more NIs. */
796 /* get the next net */
797 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
798 struct lnet_peer_net,
800 /* get the ni on it */
801 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
807 /* there are more nis left */
808 lpni = list_entry(prev->lpni_peer_nis.next,
809 struct lnet_peer_ni, lpni_peer_nis);
814 /* Call with the ln_api_mutex held */
815 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
817 struct lnet_process_id id;
818 struct lnet_peer_table *ptable;
819 struct lnet_peer *lp;
828 if (the_lnet.ln_state != LNET_STATE_RUNNING)
831 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
834 * Count the number of peers, and return E2BIG if the buffer
835 * is too small. We'll also return the desired size.
838 for (cpt = 0; cpt < lncpt; cpt++) {
839 ptable = the_lnet.ln_peer_tables[cpt];
840 count += ptable->pt_peers;
842 size = count * sizeof(*ids);
847 * Walk the peer lists and copy out the primary nids.
848 * This is safe because the peer lists are only modified
849 * while the ln_api_mutex is held. So we don't need to
850 * hold the lnet_net_lock as well, and can therefore
851 * directly call copy_to_user().
854 memset(&id, 0, sizeof(id));
855 id.pid = LNET_PID_LUSTRE;
857 for (cpt = 0; cpt < lncpt; cpt++) {
858 ptable = the_lnet.ln_peer_tables[cpt];
859 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
862 id.nid = lp->lp_primary_nid;
863 if (copy_to_user(&ids[i], &id, sizeof(id)))
876 * Start pushes to peers that need to be updated for a configuration
877 * change on this node.
880 lnet_push_update_to_peers(int force)
882 struct lnet_peer_table *ptable;
883 struct lnet_peer *lp;
887 lnet_net_lock(LNET_LOCK_EX);
888 if (lnet_peer_discovery_disabled)
890 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
891 for (cpt = 0; cpt < lncpt; cpt++) {
892 ptable = the_lnet.ln_peer_tables[cpt];
893 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
895 spin_lock(&lp->lp_lock);
896 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
897 lp->lp_state |= LNET_PEER_FORCE_PUSH;
898 spin_unlock(&lp->lp_lock);
900 if (lnet_peer_needs_push(lp))
901 lnet_peer_queue_for_discovery(lp);
904 lnet_net_unlock(LNET_LOCK_EX);
905 wake_up(&the_lnet.ln_dc_waitq);
909 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
910 * this is a preferred point-to-point path. Call with lnet_net_lock in
914 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
918 if (lpni->lpni_pref_nnids == 0)
920 if (lpni->lpni_pref_nnids == 1)
921 return lpni->lpni_pref.nid == nid;
922 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
923 if (lpni->lpni_pref.nids[i] == nid)
930 * Set a single ni as preferred, provided no preferred ni is already
931 * defined. Only to be used for non-multi-rail peer_ni.
934 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
938 spin_lock(&lpni->lpni_lock);
939 if (nid == LNET_NID_ANY) {
941 } else if (lpni->lpni_pref_nnids > 0) {
943 } else if (lpni->lpni_pref_nnids == 0) {
944 lpni->lpni_pref.nid = nid;
945 lpni->lpni_pref_nnids = 1;
946 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
948 spin_unlock(&lpni->lpni_lock);
950 CDEBUG(D_NET, "peer %s nid %s: %d\n",
951 libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
956 * Clear the preferred NID from a non-multi-rail peer_ni, provided
957 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
960 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
964 spin_lock(&lpni->lpni_lock);
965 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
966 lpni->lpni_pref_nnids = 0;
967 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
968 } else if (lpni->lpni_pref_nnids == 0) {
973 spin_unlock(&lpni->lpni_lock);
975 CDEBUG(D_NET, "peer %s: %d\n",
976 libcfs_nid2str(lpni->lpni_nid), rc);
981 * Clear the preferred NIDs from a non-multi-rail peer.
984 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
986 struct lnet_peer_ni *lpni = NULL;
988 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
989 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
993 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
995 lnet_nid_t *nids = NULL;
996 lnet_nid_t *oldnids = NULL;
997 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1002 if (nid == LNET_NID_ANY) {
1007 if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1012 /* A non-MR node may have only one preferred NI per peer_ni */
1013 if (lpni->lpni_pref_nnids > 0) {
1014 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1020 if (lpni->lpni_pref_nnids != 0) {
1021 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1022 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1027 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1028 if (lpni->lpni_pref.nids[i] == nid) {
1029 LIBCFS_FREE(nids, size);
1033 nids[i] = lpni->lpni_pref.nids[i];
1038 lnet_net_lock(LNET_LOCK_EX);
1039 spin_lock(&lpni->lpni_lock);
1040 if (lpni->lpni_pref_nnids == 0) {
1041 lpni->lpni_pref.nid = nid;
1043 oldnids = lpni->lpni_pref.nids;
1044 lpni->lpni_pref.nids = nids;
1046 lpni->lpni_pref_nnids++;
1047 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1048 spin_unlock(&lpni->lpni_lock);
1049 lnet_net_unlock(LNET_LOCK_EX);
1052 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1053 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1056 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1057 spin_lock(&lpni->lpni_lock);
1058 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1059 spin_unlock(&lpni->lpni_lock);
1061 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1062 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1067 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1069 lnet_nid_t *nids = NULL;
1070 lnet_nid_t *oldnids = NULL;
1071 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1076 if (lpni->lpni_pref_nnids == 0) {
1081 if (lpni->lpni_pref_nnids == 1) {
1082 if (lpni->lpni_pref.nid != nid) {
1086 } else if (lpni->lpni_pref_nnids == 2) {
1087 if (lpni->lpni_pref.nids[0] != nid &&
1088 lpni->lpni_pref.nids[1] != nid) {
1093 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1094 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1099 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1100 if (lpni->lpni_pref.nids[i] != nid)
1102 nids[j++] = lpni->lpni_pref.nids[i];
1104 /* Check if we actually removed a nid. */
1105 if (j == lpni->lpni_pref_nnids) {
1106 LIBCFS_FREE(nids, size);
1112 lnet_net_lock(LNET_LOCK_EX);
1113 spin_lock(&lpni->lpni_lock);
1114 if (lpni->lpni_pref_nnids == 1) {
1115 lpni->lpni_pref.nid = LNET_NID_ANY;
1116 } else if (lpni->lpni_pref_nnids == 2) {
1117 oldnids = lpni->lpni_pref.nids;
1118 if (oldnids[0] == nid)
1119 lpni->lpni_pref.nid = oldnids[1];
1121 lpni->lpni_pref.nid = oldnids[2];
1123 oldnids = lpni->lpni_pref.nids;
1124 lpni->lpni_pref.nids = nids;
1126 lpni->lpni_pref_nnids--;
1127 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1128 spin_unlock(&lpni->lpni_lock);
1129 lnet_net_unlock(LNET_LOCK_EX);
1132 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1133 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1136 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1137 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1142 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1144 struct lnet_peer_ni *lpni;
1145 lnet_nid_t primary_nid = nid;
1147 lpni = lnet_find_peer_ni_locked(nid);
1149 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1150 lnet_peer_ni_decref_locked(lpni);
1157 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1158 __must_hold(&lp->lp_lock)
1160 if (lnet_peer_discovery_disabled)
1163 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1164 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1175 lnet_is_discovery_disabled(struct lnet_peer *lp)
1179 spin_lock(&lp->lp_lock);
1180 rc = lnet_is_discovery_disabled_locked(lp);
1181 spin_unlock(&lp->lp_lock);
1187 LNetPrimaryNID(lnet_nid_t nid)
1189 struct lnet_peer *lp;
1190 struct lnet_peer_ni *lpni;
1191 lnet_nid_t primary_nid = nid;
1195 cpt = lnet_net_lock_current();
1196 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1201 lp = lpni->lpni_peer_net->lpn_peer;
1203 while (!lnet_peer_is_uptodate(lp)) {
1204 rc = lnet_discover_peer_locked(lpni, cpt, true);
1207 lp = lpni->lpni_peer_net->lpn_peer;
1209 /* Only try once if discovery is disabled */
1210 if (lnet_is_discovery_disabled(lp))
1213 primary_nid = lp->lp_primary_nid;
1215 lnet_peer_ni_decref_locked(lpni);
1217 lnet_net_unlock(cpt);
1219 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1220 libcfs_nid2str(primary_nid), rc);
1223 EXPORT_SYMBOL(LNetPrimaryNID);
1225 struct lnet_peer_net *
1226 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1228 struct lnet_peer_net *peer_net;
1229 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1230 if (peer_net->lpn_net_id == net_id)
1237 * Attach a peer_ni to a peer_net and peer. This function assumes
1238 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1239 * may be attached to a different peer, in which case it will be
1240 * properly detached first. The whole operation is done atomically.
1242 * Always returns 0. This is the last function called from functions
1243 * that do return an int, so returning 0 here allows the compiler to
1247 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1248 struct lnet_peer_net *lpn,
1249 struct lnet_peer_ni *lpni,
1252 struct lnet_peer_table *ptable;
1254 /* Install the new peer_ni */
1255 lnet_net_lock(LNET_LOCK_EX);
1256 /* Add peer_ni to global peer table hash, if necessary. */
1257 if (list_empty(&lpni->lpni_hashlist)) {
1258 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1260 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1261 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1262 ptable->pt_version++;
1263 /* This is the 1st refcount on lpni. */
1264 atomic_inc(&lpni->lpni_refcount);
1267 /* Detach the peer_ni from an existing peer, if necessary. */
1268 if (lpni->lpni_peer_net) {
1269 LASSERT(lpni->lpni_peer_net != lpn);
1270 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1271 lnet_peer_detach_peer_ni_locked(lpni);
1272 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1273 lpni->lpni_peer_net = NULL;
1276 /* Add peer_ni to peer_net */
1277 lpni->lpni_peer_net = lpn;
1278 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1279 lnet_peer_net_addref_locked(lpn);
1281 /* Add peer_net to peer */
1282 if (!lpn->lpn_peer) {
1284 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1285 lnet_peer_addref_locked(lp);
1288 /* Add peer to global peer list, if necessary */
1289 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1290 if (list_empty(&lp->lp_peer_list)) {
1291 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1296 /* Update peer state */
1297 spin_lock(&lp->lp_lock);
1298 if (flags & LNET_PEER_CONFIGURED) {
1299 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1300 lp->lp_state |= LNET_PEER_CONFIGURED;
1302 if (flags & LNET_PEER_MULTI_RAIL) {
1303 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1304 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1305 lnet_peer_clr_non_mr_pref_nids(lp);
1308 spin_unlock(&lp->lp_lock);
1311 lnet_net_unlock(LNET_LOCK_EX);
1313 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1314 libcfs_nid2str(lp->lp_primary_nid),
1315 libcfs_nid2str(lpni->lpni_nid), flags);
1321 * Create a new peer, with nid as its primary nid.
1323 * Call with the lnet_api_mutex held.
1326 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1328 struct lnet_peer *lp;
1329 struct lnet_peer_net *lpn;
1330 struct lnet_peer_ni *lpni;
1333 LASSERT(nid != LNET_NID_ANY);
1336 * No need for the lnet_net_lock here, because the
1337 * lnet_api_mutex is held.
1339 lpni = lnet_find_peer_ni_locked(nid);
1341 /* A peer with this NID already exists. */
1342 lp = lpni->lpni_peer_net->lpn_peer;
1343 lnet_peer_ni_decref_locked(lpni);
1345 * This is an error if the peer was configured and the
1346 * primary NID differs or an attempt is made to change
1347 * the Multi-Rail flag. Otherwise the assumption is
1348 * that an existing peer is being modified.
1350 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1351 if (lp->lp_primary_nid != nid)
1353 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1357 /* Delete and recreate as a configured peer. */
1361 /* Create peer, peer_net, and peer_ni. */
1363 lp = lnet_peer_alloc(nid);
1366 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1369 lpni = lnet_peer_ni_alloc(nid);
1373 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1376 LIBCFS_FREE(lpn, sizeof(*lpn));
1378 LIBCFS_FREE(lp, sizeof(*lp));
1380 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1381 libcfs_nid2str(nid), flags, rc);
1386 * Add a NID to a peer. Call with ln_api_mutex held.
1389 * -EPERM: Non-DLC addition to a DLC-configured peer.
1390 * -EEXIST: The NID was configured by DLC for a different peer.
1391 * -ENOMEM: Out of memory.
1392 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1393 * non-multi-rail peer.
1396 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1398 struct lnet_peer_net *lpn;
1399 struct lnet_peer_ni *lpni;
1403 LASSERT(nid != LNET_NID_ANY);
1405 /* A configured peer can only be updated through configuration. */
1406 if (!(flags & LNET_PEER_CONFIGURED)) {
1407 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1414 * The MULTI_RAIL flag can be set but not cleared, because
1415 * that would leave the peer struct in an invalid state.
1417 if (flags & LNET_PEER_MULTI_RAIL) {
1418 spin_lock(&lp->lp_lock);
1419 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1420 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1421 lnet_peer_clr_non_mr_pref_nids(lp);
1423 spin_unlock(&lp->lp_lock);
1424 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1429 lpni = lnet_find_peer_ni_locked(nid);
1432 * A peer_ni already exists. This is only a problem if
1433 * it is not connected to this peer and was configured
1436 lnet_peer_ni_decref_locked(lpni);
1437 if (lpni->lpni_peer_net->lpn_peer == lp)
1439 if (lnet_peer_ni_is_configured(lpni)) {
1443 /* If this is the primary NID, destroy the peer. */
1444 if (lnet_peer_ni_is_primary(lpni)) {
1445 struct lnet_peer *rtr_lp =
1446 lpni->lpni_peer_net->lpn_peer;
1447 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1449 * if we're trying to delete a router it means
1450 * we're moving this peer NI to a new peer so must
1451 * transfer router properties to the new peer
1453 if (rtr_refcount > 0) {
1454 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1455 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1457 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1458 lpni = lnet_peer_ni_alloc(nid);
1465 lpni = lnet_peer_ni_alloc(nid);
1473 * Get the peer_net. Check that we're not adding a second
1474 * peer_ni on a peer_net of a non-multi-rail peer.
1476 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1478 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1483 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1488 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1491 /* If the peer_ni was allocated above its peer_net pointer is NULL */
1492 if (!lpni->lpni_peer_net)
1493 LIBCFS_FREE(lpni, sizeof(*lpni));
1495 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1496 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1502 * Update the primary NID of a peer, if possible.
1504 * Call with the lnet_api_mutex held.
1507 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1509 lnet_nid_t old = lp->lp_primary_nid;
1512 if (lp->lp_primary_nid == nid)
1514 rc = lnet_peer_add_nid(lp, nid, flags);
1517 lp->lp_primary_nid = nid;
1519 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1520 libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1525 * lpni creation initiated due to traffic either sending or receiving.
1528 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1530 struct lnet_peer *lp;
1531 struct lnet_peer_net *lpn;
1532 struct lnet_peer_ni *lpni;
1536 if (nid == LNET_NID_ANY) {
1541 /* lnet_net_lock is not needed here because ln_api_lock is held */
1542 lpni = lnet_find_peer_ni_locked(nid);
1545 * We must have raced with another thread. Since we
1546 * know next to nothing about a peer_ni created by
1547 * traffic, we just assume everything is ok and
1550 lnet_peer_ni_decref_locked(lpni);
1554 /* Create peer, peer_net, and peer_ni. */
1556 lp = lnet_peer_alloc(nid);
1559 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1562 lpni = lnet_peer_ni_alloc(nid);
1565 if (pref != LNET_NID_ANY)
1566 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1568 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1571 LIBCFS_FREE(lpn, sizeof(*lpn));
1573 LIBCFS_FREE(lp, sizeof(*lp));
1575 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1580 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1582 * This API handles the following combinations:
1583 * Create a peer with its primary NI if only the prim_nid is provided
1584 * Add a NID to a peer identified by the prim_nid. The peer identified
1585 * by the prim_nid must already exist.
1586 * The peer being created may be non-MR.
1588 * The caller must hold ln_api_mutex. This prevents the peer from
1589 * being created/modified/deleted by a different thread.
1592 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1594 struct lnet_peer *lp = NULL;
1595 struct lnet_peer_ni *lpni;
1598 /* The prim_nid must always be specified */
1599 if (prim_nid == LNET_NID_ANY)
1602 flags = LNET_PEER_CONFIGURED;
1604 flags |= LNET_PEER_MULTI_RAIL;
1607 * If nid isn't specified, we must create a new peer with
1608 * prim_nid as its primary nid.
1610 if (nid == LNET_NID_ANY)
1611 return lnet_peer_add(prim_nid, flags);
1613 /* Look up the prim_nid, which must exist. */
1614 lpni = lnet_find_peer_ni_locked(prim_nid);
1617 lnet_peer_ni_decref_locked(lpni);
1618 lp = lpni->lpni_peer_net->lpn_peer;
1620 /* Peer must have been configured. */
1621 if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1622 CDEBUG(D_NET, "peer %s was not configured\n",
1623 libcfs_nid2str(prim_nid));
1627 /* Primary NID must match */
1628 if (lp->lp_primary_nid != prim_nid) {
1629 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1630 libcfs_nid2str(prim_nid),
1631 libcfs_nid2str(lp->lp_primary_nid));
1635 /* Multi-Rail flag must match. */
1636 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1637 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1638 libcfs_nid2str(prim_nid));
1642 return lnet_peer_add_nid(lp, nid, flags);
1646 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1648 * This API handles the following combinations:
1649 * Delete a NI from a peer if both prim_nid and nid are provided.
1650 * Delete a peer if only prim_nid is provided.
1651 * Delete a peer if its primary nid is provided.
1653 * The caller must hold ln_api_mutex. This prevents the peer from
1654 * being modified/deleted by a different thread.
1657 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1659 struct lnet_peer *lp;
1660 struct lnet_peer_ni *lpni;
1663 if (prim_nid == LNET_NID_ANY)
1666 lpni = lnet_find_peer_ni_locked(prim_nid);
1669 lnet_peer_ni_decref_locked(lpni);
1670 lp = lpni->lpni_peer_net->lpn_peer;
1672 if (prim_nid != lp->lp_primary_nid) {
1673 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1674 libcfs_nid2str(prim_nid),
1675 libcfs_nid2str(lp->lp_primary_nid));
1679 lnet_net_lock(LNET_LOCK_EX);
1680 if (lp->lp_rtr_refcount > 0) {
1681 lnet_net_unlock(LNET_LOCK_EX);
1682 CERROR("%s is a router. Can not be deleted\n",
1683 libcfs_nid2str(prim_nid));
1686 lnet_net_unlock(LNET_LOCK_EX);
1688 if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1689 return lnet_peer_del(lp);
1691 flags = LNET_PEER_CONFIGURED;
1692 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1693 flags |= LNET_PEER_MULTI_RAIL;
1695 return lnet_peer_del_nid(lp, nid, flags);
1699 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1701 struct lnet_peer_table *ptable;
1702 struct lnet_peer_net *lpn;
1704 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1706 LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1707 LASSERT(list_empty(&lpni->lpni_txq));
1708 LASSERT(lpni->lpni_txqnob == 0);
1709 LASSERT(list_empty(&lpni->lpni_peer_nis));
1710 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1712 lpn = lpni->lpni_peer_net;
1713 lpni->lpni_peer_net = NULL;
1714 lpni->lpni_net = NULL;
1716 /* remove the peer ni from the zombie list */
1717 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1718 spin_lock(&ptable->pt_zombie_lock);
1719 list_del_init(&lpni->lpni_hashlist);
1720 ptable->pt_zombies--;
1721 spin_unlock(&ptable->pt_zombie_lock);
1723 if (lpni->lpni_pref_nnids > 1) {
1724 LIBCFS_FREE(lpni->lpni_pref.nids,
1725 sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1727 LIBCFS_FREE(lpni, sizeof(*lpni));
1729 lnet_peer_net_decref_locked(lpn);
1732 struct lnet_peer_ni *
1733 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1735 struct lnet_peer_ni *lpni = NULL;
1738 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1739 return ERR_PTR(-ESHUTDOWN);
1742 * find if a peer_ni already exists.
1743 * If so then just return that.
1745 lpni = lnet_find_peer_ni_locked(nid);
1749 lnet_net_unlock(cpt);
1751 rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1754 goto out_net_relock;
1757 lpni = lnet_find_peer_ni_locked(nid);
1767 * Get a peer_ni for the given nid, create it if necessary. Takes a
1768 * hold on the peer_ni.
1770 struct lnet_peer_ni *
1771 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1773 struct lnet_peer_ni *lpni = NULL;
1776 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1777 return ERR_PTR(-ESHUTDOWN);
1780 * find if a peer_ni already exists.
1781 * If so then just return that.
1783 lpni = lnet_find_peer_ni_locked(nid);
1789 * use the lnet_api_mutex to serialize the creation of the peer_ni
1790 * and the creation/deletion of the local ni/net. When a local ni is
1791 * created, if there exists a set of peer_nis on that network,
1792 * they need to be traversed and updated. When a local NI is
1793 * deleted, which could result in a network being deleted, then
1794 * all peer nis on that network need to be removed as well.
1796 * Creation through traffic should also be serialized with
1797 * creation through DLC.
1799 lnet_net_unlock(cpt);
1800 mutex_lock(&the_lnet.ln_api_mutex);
1802 * Shutdown is only set under the ln_api_lock, so a single
1803 * check here is sufficent.
1805 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1806 lpni = ERR_PTR(-ESHUTDOWN);
1807 goto out_mutex_unlock;
1810 rc = lnet_peer_ni_traffic_add(nid, pref);
1813 goto out_mutex_unlock;
1816 lpni = lnet_find_peer_ni_locked(nid);
1820 mutex_unlock(&the_lnet.ln_api_mutex);
1823 /* Lock has been dropped, check again for shutdown. */
1824 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1826 lnet_peer_ni_decref_locked(lpni);
1827 lpni = ERR_PTR(-ESHUTDOWN);
1834 lnet_peer_gw_discovery(struct lnet_peer *lp)
1838 spin_lock(&lp->lp_lock);
1839 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1841 spin_unlock(&lp->lp_lock);
1847 lnet_peer_is_uptodate(struct lnet_peer *lp)
1851 spin_lock(&lp->lp_lock);
1852 rc = lnet_peer_is_uptodate_locked(lp);
1853 spin_unlock(&lp->lp_lock);
1858 * Is a peer uptodate from the point of view of discovery?
1860 * If it is currently being processed, obviously not.
1861 * A forced Ping or Push is also handled by the discovery thread.
1863 * Otherwise look at whether the peer needs rediscovering.
1866 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1867 __must_hold(&lp->lp_lock)
1871 if (lp->lp_state & (LNET_PEER_DISCOVERING |
1872 LNET_PEER_FORCE_PING |
1873 LNET_PEER_FORCE_PUSH)) {
1875 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1877 } else if (lnet_peer_needs_push(lp)) {
1879 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1880 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1892 * Queue a peer for the attention of the discovery thread. Call with
1893 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1894 * -EALREADY if the peer was already queued.
1896 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1900 spin_lock(&lp->lp_lock);
1901 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1902 lp->lp_state |= LNET_PEER_DISCOVERING;
1903 spin_unlock(&lp->lp_lock);
1904 if (list_empty(&lp->lp_dc_list)) {
1905 lnet_peer_addref_locked(lp);
1906 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1907 wake_up(&the_lnet.ln_dc_waitq);
1913 CDEBUG(D_NET, "Queue peer %s: %d\n",
1914 libcfs_nid2str(lp->lp_primary_nid), rc);
1920 * Discovery of a peer is complete. Wake all waiters on the peer.
1921 * Call with lnet_net_lock/EX held.
1923 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1925 struct lnet_msg *msg, *tmp;
1927 LIST_HEAD(pending_msgs);
1929 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1930 libcfs_nid2str(lp->lp_primary_nid));
1932 list_del_init(&lp->lp_dc_list);
1933 spin_lock(&lp->lp_lock);
1934 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1935 spin_unlock(&lp->lp_lock);
1936 wake_up_all(&lp->lp_dc_waitq);
1938 if (lp->lp_rtr_refcount > 0)
1939 lnet_router_discovery_complete(lp);
1941 lnet_net_unlock(LNET_LOCK_EX);
1943 /* iterate through all pending messages and send them again */
1944 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1945 list_del_init(&msg->msg_list);
1946 if (lp->lp_dc_error) {
1947 lnet_finalize(msg, lp->lp_dc_error);
1951 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1952 lnet_msgtyp2str(msg->msg_type),
1953 libcfs_id2str(msg->msg_target));
1954 rc = lnet_send(msg->msg_src_nid_param, msg,
1955 msg->msg_rtr_nid_param);
1957 CNETERR("Error sending %s to %s: %d\n",
1958 lnet_msgtyp2str(msg->msg_type),
1959 libcfs_id2str(msg->msg_target), rc);
1960 lnet_finalize(msg, rc);
1963 lnet_net_lock(LNET_LOCK_EX);
1964 lnet_peer_decref_locked(lp);
1968 * Handle inbound push.
1969 * Like any event handler, called with lnet_res_lock/CPT held.
1971 void lnet_peer_push_event(struct lnet_event *ev)
1973 struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1974 struct lnet_peer *lp;
1976 /* lnet_find_peer() adds a refcount */
1977 lp = lnet_find_peer(ev->source.nid);
1979 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1980 libcfs_nid2str(ev->initiator.nid),
1981 libcfs_nid2str(ev->source.nid));
1985 /* Ensure peer state remains consistent while we modify it. */
1986 spin_lock(&lp->lp_lock);
1989 * If some kind of error happened the contents of the message
1990 * cannot be used. Clear the NIDS_UPTODATE and set the
1991 * FORCE_PING flag to trigger a ping.
1994 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1995 lp->lp_state |= LNET_PEER_FORCE_PING;
1996 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1998 libcfs_nid2str(lp->lp_primary_nid),
1999 libcfs_nid2str(ev->source.nid));
2004 * A push with invalid or corrupted info. Clear the UPTODATE
2005 * flag to trigger a ping.
2007 if (lnet_ping_info_validate(&pbuf->pb_info)) {
2008 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2009 lp->lp_state |= LNET_PEER_FORCE_PING;
2010 CDEBUG(D_NET, "Corrupted Push from %s\n",
2011 libcfs_nid2str(lp->lp_primary_nid));
2016 * Make sure we'll allocate the correct size ping buffer when
2019 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2020 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2023 * A non-Multi-Rail peer is not supposed to be capable of
2026 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2027 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2028 libcfs_nid2str(lp->lp_primary_nid));
2033 * Check the MULTIRAIL flag. Complain if the peer was DLC
2034 * configured without it.
2036 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2037 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2038 CERROR("Push says %s is Multi-Rail, DLC says not\n",
2039 libcfs_nid2str(lp->lp_primary_nid));
2041 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2042 lnet_peer_clr_non_mr_pref_nids(lp);
2047 * The peer may have discovery disabled at its end. Set
2048 * NO_DISCOVERY as appropriate.
2050 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2051 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2052 libcfs_nid2str(lp->lp_primary_nid));
2053 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2054 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2055 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2056 libcfs_nid2str(lp->lp_primary_nid));
2057 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2061 * Check for truncation of the Put message. Clear the
2062 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2063 * and tell discovery to allocate a bigger buffer.
2065 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2066 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2067 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2068 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2069 lp->lp_state |= LNET_PEER_FORCE_PING;
2070 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2071 libcfs_nid2str(lp->lp_primary_nid),
2072 pbuf->pb_info.pi_nnis);
2076 /* always assume new data */
2077 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2078 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2081 * If there is data present that hasn't been processed yet,
2082 * we'll replace it if the Put contained newer data and it
2083 * fits. We're racing with a Ping or earlier Push in this
2086 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2087 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2088 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2089 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2090 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2091 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2092 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2093 libcfs_nid2str(lp->lp_primary_nid),
2094 LNET_PING_BUFFER_SEQNO(pbuf),
2095 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2101 * Allocate a buffer to copy the data. On a failure we drop
2102 * the Push and set FORCE_PING to force the discovery
2103 * thread to fix the problem by pinging the peer.
2105 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2107 lp->lp_state |= LNET_PEER_FORCE_PING;
2108 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2109 libcfs_nid2str(lp->lp_primary_nid),
2110 LNET_PING_BUFFER_SEQNO(pbuf));
2115 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2116 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2117 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2118 CDEBUG(D_NET, "Received Push %s %u\n",
2119 libcfs_nid2str(lp->lp_primary_nid),
2120 LNET_PING_BUFFER_SEQNO(pbuf));
2124 * Queue the peer for discovery if not done, force it on the request
2125 * queue and wake the discovery thread if the peer was already queued,
2126 * because its status changed.
2128 spin_unlock(&lp->lp_lock);
2129 lnet_net_lock(LNET_LOCK_EX);
2130 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2131 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2132 wake_up(&the_lnet.ln_dc_waitq);
2134 /* Drop refcount from lookup */
2135 lnet_peer_decref_locked(lp);
2136 lnet_net_unlock(LNET_LOCK_EX);
2140 * Clear the discovery error state, unless we're already discovering
2141 * this peer, in which case the error is current.
2143 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2145 spin_lock(&lp->lp_lock);
2146 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2147 lp->lp_dc_error = 0;
2148 spin_unlock(&lp->lp_lock);
2152 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2153 * dropped/retaken within this function. An lnet_peer_ni is passed in
2154 * because discovery could tear down an lnet_peer.
2157 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2160 struct lnet_peer *lp;
2165 lnet_net_unlock(cpt);
2166 lnet_net_lock(LNET_LOCK_EX);
2167 lp = lpni->lpni_peer_net->lpn_peer;
2168 lnet_peer_clear_discovery_error(lp);
2171 * We're willing to be interrupted. The lpni can become a
2172 * zombie if we race with DLC, so we must check for that.
2175 /* Keep lp alive when the lnet_net_lock is unlocked */
2176 lnet_peer_addref_locked(lp);
2177 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2178 if (signal_pending(current))
2180 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2183 * Don't repeat discovery if discovery is disabled. This is
2184 * done to ensure we can use discovery as a standard ping as
2185 * well for backwards compatibility with routers which do not
2186 * have discovery or have discovery disabled
2188 if (lnet_is_discovery_disabled(lp) && count > 0)
2190 if (lp->lp_dc_error)
2192 if (lnet_peer_is_uptodate(lp))
2194 lnet_peer_queue_for_discovery(lp);
2196 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2199 * If caller requested a non-blocking operation then
2200 * return immediately. Once discovery is complete any
2201 * pending messages that were stopped due to discovery
2202 * will be transmitted.
2207 lnet_net_unlock(LNET_LOCK_EX);
2209 finish_wait(&lp->lp_dc_waitq, &wait);
2210 lnet_net_lock(LNET_LOCK_EX);
2211 lnet_peer_decref_locked(lp);
2212 /* Peer may have changed */
2213 lp = lpni->lpni_peer_net->lpn_peer;
2215 finish_wait(&lp->lp_dc_waitq, &wait);
2217 lnet_net_unlock(LNET_LOCK_EX);
2219 lnet_peer_decref_locked(lp);
2221 * The peer may have changed, so re-check and rediscover if that turns
2222 * out to have been the case. The reference count on lp ensured that
2223 * even if it was unlinked from lpni the memory could not be recycled.
2224 * Thus the check below is sufficient to determine whether the peer
2225 * changed. If the peer changed, then lp must not be dereferenced.
2227 if (lp != lpni->lpni_peer_net->lpn_peer)
2230 if (signal_pending(current))
2232 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2234 else if (lp->lp_dc_error)
2235 rc = lp->lp_dc_error;
2237 CDEBUG(D_NET, "non-blocking discovery\n");
2238 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2241 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2242 (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2243 libcfs_nid2str(lpni->lpni_nid), rc,
2244 (!block) ? "pending discovery" : "discovery complete");
2249 /* Handle an incoming ack for a push. */
2251 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2253 struct lnet_ping_buffer *pbuf;
2255 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2256 spin_lock(&lp->lp_lock);
2257 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2258 lp->lp_push_error = ev->status;
2260 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2262 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2263 spin_unlock(&lp->lp_lock);
2265 CDEBUG(D_NET, "peer %s ev->status %d\n",
2266 libcfs_nid2str(lp->lp_primary_nid), ev->status);
2269 /* Handle a Reply message. This is the reply to a Ping message. */
2271 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2273 struct lnet_ping_buffer *pbuf;
2276 spin_lock(&lp->lp_lock);
2279 * If some kind of error happened the contents of message
2280 * cannot be used. Set PING_FAILED to trigger a retry.
2283 lp->lp_state |= LNET_PEER_PING_FAILED;
2284 lp->lp_ping_error = ev->status;
2285 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2287 libcfs_nid2str(lp->lp_primary_nid),
2288 libcfs_nid2str(ev->source.nid));
2292 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2293 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2294 lnet_swap_pinginfo(pbuf);
2297 * A reply with invalid or corrupted info. Set PING_FAILED to
2300 rc = lnet_ping_info_validate(&pbuf->pb_info);
2302 lp->lp_state |= LNET_PEER_PING_FAILED;
2303 lp->lp_ping_error = 0;
2304 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2305 libcfs_nid2str(lp->lp_primary_nid), rc);
2311 * The peer may have discovery disabled at its end. Set
2312 * NO_DISCOVERY as appropriate.
2314 if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2315 !lnet_peer_discovery_disabled) {
2316 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2317 libcfs_nid2str(lp->lp_primary_nid));
2318 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2320 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2321 libcfs_nid2str(lp->lp_primary_nid));
2322 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2326 * Update the MULTI_RAIL flag based on the reply. If the peer
2327 * was configured with DLC then the setting should match what
2330 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2331 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2332 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2333 libcfs_nid2str(lp->lp_primary_nid), lp);
2334 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2335 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2336 libcfs_nid2str(lp->lp_primary_nid));
2337 } else if (lnet_peer_discovery_disabled) {
2339 "peer %s(%p) not MR: DD disabled locally\n",
2340 libcfs_nid2str(lp->lp_primary_nid), lp);
2341 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2343 "peer %s(%p) not MR: DD disabled remotely\n",
2344 libcfs_nid2str(lp->lp_primary_nid), lp);
2346 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2347 libcfs_nid2str(lp->lp_primary_nid), lp);
2348 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2349 lnet_peer_clr_non_mr_pref_nids(lp);
2351 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2352 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2353 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2354 libcfs_nid2str(lp->lp_primary_nid));
2356 CERROR("Multi-Rail state vanished from %s\n",
2357 libcfs_nid2str(lp->lp_primary_nid));
2358 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2363 * Make sure we'll allocate the correct size ping buffer when
2366 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2367 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2370 * Check for truncation of the Reply. Clear PING_SENT and set
2371 * PING_FAILED to trigger a retry.
2373 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2374 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2375 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2376 lp->lp_state |= LNET_PEER_PING_FAILED;
2377 lp->lp_ping_error = 0;
2378 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2379 libcfs_nid2str(lp->lp_primary_nid),
2380 pbuf->pb_info.pi_nnis);
2385 * Check the sequence numbers in the reply. These are only
2386 * available if the reply came from a Multi-Rail peer.
2388 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2389 pbuf->pb_info.pi_nnis > 1 &&
2390 lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2391 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2392 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2393 libcfs_nid2str(lp->lp_primary_nid),
2394 LNET_PING_BUFFER_SEQNO(pbuf),
2397 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2400 /* We're happy with the state of the data in the buffer. */
2401 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2402 libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2403 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2404 lnet_ping_buffer_decref(lp->lp_data);
2406 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2407 lnet_ping_buffer_addref(pbuf);
2410 lp->lp_state &= ~LNET_PEER_PING_SENT;
2411 spin_unlock(&lp->lp_lock);
2413 lnet_net_lock(LNET_LOCK_EX);
2415 * If this peer is a gateway, call the routing callback to
2416 * handle the ping reply
2418 if (lp->lp_rtr_refcount > 0)
2419 lnet_router_discovery_ping_reply(lp);
2420 lnet_net_unlock(LNET_LOCK_EX);
2424 * Send event handling. Only matters for error cases, where we clean
2425 * up state on the peer and peer_ni that would otherwise be updated in
2426 * the REPLY event handler for a successful Ping, and the ACK event
2427 * handler for a successful Push.
2430 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2437 spin_lock(&lp->lp_lock);
2438 if (ev->msg_type == LNET_MSG_GET) {
2439 lp->lp_state &= ~LNET_PEER_PING_SENT;
2440 lp->lp_state |= LNET_PEER_PING_FAILED;
2441 lp->lp_ping_error = ev->status;
2442 } else { /* ev->msg_type == LNET_MSG_PUT */
2443 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2444 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2445 lp->lp_push_error = ev->status;
2447 spin_unlock(&lp->lp_lock);
2448 rc = LNET_REDISCOVER_PEER;
2450 CDEBUG(D_NET, "%s Send to %s: %d\n",
2451 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2452 libcfs_nid2str(ev->target.nid), rc);
2457 * Unlink event handling. This event is only seen if a call to
2458 * LNetMDUnlink() caused the event to be unlinked. If this call was
2459 * made after the event was set up in LNetGet() or LNetPut() then we
2460 * assume the Ping or Push timed out.
2463 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2465 spin_lock(&lp->lp_lock);
2466 /* We've passed through LNetGet() */
2467 if (lp->lp_state & LNET_PEER_PING_SENT) {
2468 lp->lp_state &= ~LNET_PEER_PING_SENT;
2469 lp->lp_state |= LNET_PEER_PING_FAILED;
2470 lp->lp_ping_error = -ETIMEDOUT;
2471 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2472 libcfs_nid2str(lp->lp_primary_nid));
2474 /* We've passed through LNetPut() */
2475 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2476 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2477 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2478 lp->lp_push_error = -ETIMEDOUT;
2479 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2480 libcfs_nid2str(lp->lp_primary_nid));
2482 spin_unlock(&lp->lp_lock);
2486 * Event handler for the discovery EQ.
2488 * Called with lnet_res_lock(cpt) held. The cpt is the
2489 * lnet_cpt_of_cookie() of the md handle cookie.
2491 static void lnet_discovery_event_handler(struct lnet_event *event)
2493 struct lnet_peer *lp = event->md.user_ptr;
2494 struct lnet_ping_buffer *pbuf;
2497 /* discovery needs to take another look */
2498 rc = LNET_REDISCOVER_PEER;
2500 CDEBUG(D_NET, "Received event: %d\n", event->type);
2502 switch (event->type) {
2503 case LNET_EVENT_ACK:
2504 lnet_discovery_event_ack(lp, event);
2506 case LNET_EVENT_REPLY:
2507 lnet_discovery_event_reply(lp, event);
2509 case LNET_EVENT_SEND:
2510 /* Only send failure triggers a retry. */
2511 rc = lnet_discovery_event_send(lp, event);
2513 case LNET_EVENT_UNLINK:
2514 /* LNetMDUnlink() was called */
2515 lnet_discovery_event_unlink(lp, event);
2518 /* Invalid events. */
2521 lnet_net_lock(LNET_LOCK_EX);
2522 if (event->unlinked) {
2523 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2524 lnet_ping_buffer_decref(pbuf);
2525 lnet_peer_decref_locked(lp);
2528 /* put peer back at end of request queue, if discovery not already
2530 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2531 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2532 wake_up(&the_lnet.ln_dc_waitq);
2534 lnet_net_unlock(LNET_LOCK_EX);
2538 * Build a peer from incoming data.
2540 * The NIDs in the incoming data are supposed to be structured as follows:
2543 * - other NIDs in same net
2544 * - NIDs in second net
2545 * - NIDs in third net
2547 * This due to the way the list of NIDs in the data is created.
2549 * Note that this function will mark the peer uptodate unless an
2550 * ENOMEM is encontered. All other errors are due to a conflict
2551 * between the DLC configuration and what discovery sees. We treat DLC
2552 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2553 * peer from becoming stuck in discovery.
2555 static int lnet_peer_merge_data(struct lnet_peer *lp,
2556 struct lnet_ping_buffer *pbuf)
2558 struct lnet_peer_ni *lpni;
2559 lnet_nid_t *curnis = NULL;
2560 struct lnet_ni_status *addnis = NULL;
2561 lnet_nid_t *delnis = NULL;
2571 flags = LNET_PEER_DISCOVERED;
2572 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2573 flags |= LNET_PEER_MULTI_RAIL;
2576 * Cache the routing feature for the peer; whether it is enabled
2577 * for disabled as reported by the remote peer.
2579 spin_lock(&lp->lp_lock);
2580 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2581 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2583 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2584 spin_unlock(&lp->lp_lock);
2586 nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2587 LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2588 LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2589 LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2590 if (!curnis || !addnis || !delnis) {
2598 /* Construct the list of NIDs present in peer. */
2600 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2601 curnis[ncurnis++] = lpni->lpni_nid;
2604 * Check for NIDs in pbuf not present in curnis[].
2605 * The loop starts at 1 to skip the loopback NID.
2607 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2608 for (j = 0; j < ncurnis; j++)
2609 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2612 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2615 * Check for NIDs in curnis[] not present in pbuf.
2616 * The nested loop starts at 1 to skip the loopback NID.
2618 * But never add the loopback NID to delnis[]: if it is
2619 * present in curnis[] then this peer is for this node.
2621 for (i = 0; i < ncurnis; i++) {
2622 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2624 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2625 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2627 * update the information we cache for the
2628 * peer with the latest information we
2631 lpni = lnet_find_peer_ni_locked(curnis[i]);
2633 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2634 lnet_peer_ni_decref_locked(lpni);
2639 if (j == pbuf->pb_info.pi_nnis)
2640 delnis[ndelnis++] = curnis[i];
2644 * If we get here and the discovery is disabled then we don't want
2645 * to add or delete any NIs. We just updated the ones we have some
2646 * information on, and call it a day
2649 if (lnet_is_discovery_disabled(lp))
2652 for (i = 0; i < naddnis; i++) {
2653 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2655 CERROR("Error adding NID %s to peer %s: %d\n",
2656 libcfs_nid2str(addnis[i].ns_nid),
2657 libcfs_nid2str(lp->lp_primary_nid), rc);
2661 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2663 lpni->lpni_ns_status = addnis[i].ns_status;
2664 lnet_peer_ni_decref_locked(lpni);
2668 for (i = 0; i < ndelnis; i++) {
2670 * for routers it's okay to delete the primary_nid because
2671 * the upper layers don't really rely on it. So if we're
2672 * being told that the router changed its primary_nid
2673 * then it's okay to delete it.
2675 if (lp->lp_rtr_refcount > 0)
2676 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2677 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2679 CERROR("Error deleting NID %s from peer %s: %d\n",
2680 libcfs_nid2str(delnis[i]),
2681 libcfs_nid2str(lp->lp_primary_nid), rc);
2687 * Errors other than -ENOMEM are due to peers having been
2688 * configured with DLC. Ignore these because DLC overrides
2693 LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2694 LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2695 LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2696 lnet_ping_buffer_decref(pbuf);
2697 CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2700 spin_lock(&lp->lp_lock);
2701 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2702 lp->lp_state |= LNET_PEER_FORCE_PING;
2703 spin_unlock(&lp->lp_lock);
2709 * The data in pbuf says lp is its primary peer, but the data was
2710 * received by a different peer. Try to update lp with the data.
2713 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2715 struct lnet_handle_md mdh;
2717 /* Queue lp for discovery, and force it on the request queue. */
2718 lnet_net_lock(LNET_LOCK_EX);
2719 if (lnet_peer_queue_for_discovery(lp))
2720 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2721 lnet_net_unlock(LNET_LOCK_EX);
2723 LNetInvalidateMDHandle(&mdh);
2726 * Decide whether we can move the peer to the DATA_PRESENT state.
2728 * We replace stale data for a multi-rail peer, repair PING_FAILED
2729 * status, and preempt FORCE_PING.
2731 * If after that we have DATA_PRESENT, we merge it into this peer.
2733 spin_lock(&lp->lp_lock);
2734 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2735 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2736 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2737 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2738 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2739 lnet_ping_buffer_decref(pbuf);
2744 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2745 lnet_ping_buffer_decref(lp->lp_data);
2747 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2749 if (lp->lp_state & LNET_PEER_PING_FAILED) {
2750 mdh = lp->lp_ping_mdh;
2751 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2752 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2753 lp->lp_ping_error = 0;
2755 if (lp->lp_state & LNET_PEER_FORCE_PING)
2756 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2757 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2758 spin_unlock(&lp->lp_lock);
2760 if (!LNetMDHandleIsInvalid(mdh))
2764 return lnet_peer_merge_data(lp, pbuf);
2766 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2770 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2774 for (i = 0; i < pinfo->pi_nnis; i++) {
2775 if (pinfo->pi_ni[i].ns_nid == nid)
2783 * Update a peer using the data received.
2785 static int lnet_peer_data_present(struct lnet_peer *lp)
2786 __must_hold(&lp->lp_lock)
2788 struct lnet_ping_buffer *pbuf;
2789 struct lnet_peer_ni *lpni;
2790 lnet_nid_t nid = LNET_NID_ANY;
2796 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2797 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2798 spin_unlock(&lp->lp_lock);
2801 * Modifications of peer structures are done while holding the
2802 * ln_api_mutex. A global lock is required because we may be
2803 * modifying multiple peer structures, and a mutex greatly
2804 * simplifies memory management.
2806 * The actual changes to the data structures must also protect
2807 * against concurrent lookups, for which the lnet_net_lock in
2808 * LNET_LOCK_EX mode is used.
2810 mutex_lock(&the_lnet.ln_api_mutex);
2811 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2817 * If this peer is not on the peer list then it is being torn
2818 * down, and our reference count may be all that is keeping it
2819 * alive. Don't do any work on it.
2821 if (list_empty(&lp->lp_peer_list))
2824 flags = LNET_PEER_DISCOVERED;
2825 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2826 flags |= LNET_PEER_MULTI_RAIL;
2829 * Check whether the primary NID in the message matches the
2830 * primary NID of the peer. If it does, update the peer, if
2831 * it it does not, check whether there is already a peer with
2832 * that primary NID. If no such peer exists, try to update
2833 * the primary NID of the current peer (allowed if it was
2834 * created due to message traffic) and complete the update.
2835 * If the peer did exist, hand off the data to it.
2837 * The peer for the loopback interface is a special case: this
2838 * is the peer for the local node, and we want to set its
2839 * primary NID to the correct value here. Moreover, this peer
2840 * can show up with only the loopback NID in the ping buffer.
2842 if (pbuf->pb_info.pi_nnis <= 1)
2844 nid = pbuf->pb_info.pi_ni[1].ns_nid;
2845 if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2846 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2848 rc = lnet_peer_merge_data(lp, pbuf);
2850 * if the primary nid of the peer is present in the ping info returned
2851 * from the peer, but it's not the local primary peer we have
2852 * cached and discovery is disabled, then we don't want to update
2853 * our local peer info, by adding or removing NIDs, we just want
2854 * to update the status of the nids that we currently have
2855 * recorded in that peer.
2857 } else if (lp->lp_primary_nid == nid ||
2858 (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2859 lnet_is_discovery_disabled(lp))) {
2860 rc = lnet_peer_merge_data(lp, pbuf);
2862 lpni = lnet_find_peer_ni_locked(nid);
2864 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2866 CERROR("Primary NID error %s versus %s: %d\n",
2867 libcfs_nid2str(lp->lp_primary_nid),
2868 libcfs_nid2str(nid), rc);
2870 rc = lnet_peer_merge_data(lp, pbuf);
2873 struct lnet_peer *new_lp;
2874 new_lp = lpni->lpni_peer_net->lpn_peer;
2876 * if lp has discovery/MR enabled that means new_lp
2877 * should have discovery/MR enabled as well, since
2878 * it's the same peer, which we're about to merge
2880 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2881 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2882 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2883 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2885 rc = lnet_peer_set_primary_data(new_lp, pbuf);
2886 lnet_consolidate_routes_locked(lp, new_lp);
2887 lnet_peer_ni_decref_locked(lpni);
2891 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2893 mutex_unlock(&the_lnet.ln_api_mutex);
2895 spin_lock(&lp->lp_lock);
2896 /* Tell discovery to re-check the peer immediately. */
2898 rc = LNET_REDISCOVER_PEER;
2903 * A ping failed. Clear the PING_FAILED state and set the
2904 * FORCE_PING state, to ensure a retry even if discovery is
2905 * disabled. This avoids being left with incorrect state.
2907 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2908 __must_hold(&lp->lp_lock)
2910 struct lnet_handle_md mdh;
2913 mdh = lp->lp_ping_mdh;
2914 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2915 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2916 lp->lp_state |= LNET_PEER_FORCE_PING;
2917 rc = lp->lp_ping_error;
2918 lp->lp_ping_error = 0;
2919 spin_unlock(&lp->lp_lock);
2921 if (!LNetMDHandleIsInvalid(mdh))
2924 CDEBUG(D_NET, "peer %s:%d\n",
2925 libcfs_nid2str(lp->lp_primary_nid), rc);
2927 spin_lock(&lp->lp_lock);
2928 return rc ? rc : LNET_REDISCOVER_PEER;
2932 * Select NID to send a Ping or Push to.
2934 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2936 struct lnet_peer_ni *lpni;
2938 /* Look for a direct-connected NID for this peer. */
2940 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2941 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2946 return lpni->lpni_nid;
2948 /* Look for a routed-connected NID for this peer. */
2950 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2951 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2956 return lpni->lpni_nid;
2958 return LNET_NID_ANY;
2961 /* Active side of ping. */
2962 static int lnet_peer_send_ping(struct lnet_peer *lp)
2963 __must_hold(&lp->lp_lock)
2970 lp->lp_state |= LNET_PEER_PING_SENT;
2971 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2972 spin_unlock(&lp->lp_lock);
2974 cpt = lnet_net_lock_current();
2975 /* Refcount for MD. */
2976 lnet_peer_addref_locked(lp);
2977 pnid = lnet_peer_select_nid(lp);
2978 lnet_net_unlock(cpt);
2980 nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2982 rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2983 the_lnet.ln_dc_eqh, false);
2986 * if LNetMDBind in lnet_send_ping fails we need to decrement the
2987 * refcount on the peer, otherwise LNetMDUnlink will be called
2988 * which will eventually do that.
2992 lnet_peer_decref_locked(lp);
2993 lnet_net_unlock(cpt);
2994 rc = -rc; /* change the rc to negative value */
2996 } else if (rc < 0) {
3000 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3002 spin_lock(&lp->lp_lock);
3006 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3008 * The errors that get us here are considered hard errors and
3009 * cause Discovery to terminate. So we clear PING_SENT, but do
3010 * not set either PING_FAILED or FORCE_PING. In fact we need
3011 * to clear PING_FAILED, because the unlink event handler will
3012 * have set it if we called LNetMDUnlink() above.
3014 spin_lock(&lp->lp_lock);
3015 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3020 * This function exists because you cannot call LNetMDUnlink() from an
3023 static int lnet_peer_push_failed(struct lnet_peer *lp)
3024 __must_hold(&lp->lp_lock)
3026 struct lnet_handle_md mdh;
3029 mdh = lp->lp_push_mdh;
3030 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3031 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3032 rc = lp->lp_push_error;
3033 lp->lp_push_error = 0;
3034 spin_unlock(&lp->lp_lock);
3036 if (!LNetMDHandleIsInvalid(mdh))
3039 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3040 spin_lock(&lp->lp_lock);
3041 return rc ? rc : LNET_REDISCOVER_PEER;
3044 /* Active side of push. */
3045 static int lnet_peer_send_push(struct lnet_peer *lp)
3046 __must_hold(&lp->lp_lock)
3048 struct lnet_ping_buffer *pbuf;
3049 struct lnet_process_id id;
3054 /* Don't push to a non-multi-rail peer. */
3055 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3056 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3060 lp->lp_state |= LNET_PEER_PUSH_SENT;
3061 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3062 spin_unlock(&lp->lp_lock);
3064 cpt = lnet_net_lock_current();
3065 pbuf = the_lnet.ln_ping_target;
3066 lnet_ping_buffer_addref(pbuf);
3067 lnet_net_unlock(cpt);
3069 /* Push source MD */
3070 md.start = &pbuf->pb_info;
3071 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3072 md.threshold = 2; /* Put/Ack */
3075 md.eq_handle = the_lnet.ln_dc_eqh;
3078 rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3080 lnet_ping_buffer_decref(pbuf);
3081 CERROR("Can't bind push source MD: %d\n", rc);
3084 cpt = lnet_net_lock_current();
3085 /* Refcount for MD. */
3086 lnet_peer_addref_locked(lp);
3087 id.pid = LNET_PID_LUSTRE;
3088 id.nid = lnet_peer_select_nid(lp);
3089 lnet_net_unlock(cpt);
3091 if (id.nid == LNET_NID_ANY) {
3096 rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3097 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3098 LNET_PROTO_PING_MATCHBITS, 0, 0);
3103 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3105 spin_lock(&lp->lp_lock);
3109 LNetMDUnlink(lp->lp_push_mdh);
3110 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3112 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3114 * The errors that get us here are considered hard errors and
3115 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3116 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3117 * because the unlink event handler will have set it if we
3118 * called LNetMDUnlink() above.
3120 spin_lock(&lp->lp_lock);
3121 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3126 * An unrecoverable error was encountered during discovery.
3127 * Set error status in peer and abort discovery.
3129 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3131 CDEBUG(D_NET, "Discovery error %s: %d\n",
3132 libcfs_nid2str(lp->lp_primary_nid), error);
3134 spin_lock(&lp->lp_lock);
3135 lp->lp_dc_error = error;
3136 lp->lp_state &= ~LNET_PEER_DISCOVERING;
3137 lp->lp_state |= LNET_PEER_REDISCOVER;
3138 spin_unlock(&lp->lp_lock);
3142 * Mark the peer as discovered.
3144 static int lnet_peer_discovered(struct lnet_peer *lp)
3145 __must_hold(&lp->lp_lock)
3147 lp->lp_state |= LNET_PEER_DISCOVERED;
3148 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3149 LNET_PEER_REDISCOVER);
3151 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3158 * Discovering this peer is taking too long. Cancel any Ping or Push
3159 * that discovery is waiting on by unlinking the relevant MDs. The
3160 * lnet_discovery_event_handler() will proceed from here and complete
3163 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3165 struct lnet_handle_md ping_mdh;
3166 struct lnet_handle_md push_mdh;
3168 LNetInvalidateMDHandle(&ping_mdh);
3169 LNetInvalidateMDHandle(&push_mdh);
3171 spin_lock(&lp->lp_lock);
3172 if (lp->lp_state & LNET_PEER_PING_SENT) {
3173 ping_mdh = lp->lp_ping_mdh;
3174 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3176 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3177 push_mdh = lp->lp_push_mdh;
3178 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3180 spin_unlock(&lp->lp_lock);
3182 if (!LNetMDHandleIsInvalid(ping_mdh))
3183 LNetMDUnlink(ping_mdh);
3184 if (!LNetMDHandleIsInvalid(push_mdh))
3185 LNetMDUnlink(push_mdh);
3189 * Wait for work to be queued or some other change that must be
3190 * attended to. Returns non-zero if the discovery thread should shut
3193 static int lnet_peer_discovery_wait_for_work(void)
3200 cpt = lnet_net_lock_current();
3202 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3203 TASK_INTERRUPTIBLE);
3204 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3206 if (lnet_push_target_resize_needed())
3208 if (!list_empty(&the_lnet.ln_dc_request))
3210 if (!list_empty(&the_lnet.ln_msg_resend))
3212 lnet_net_unlock(cpt);
3215 * wakeup max every second to check if there are peers that
3216 * have been stuck on the working queue for greater than
3219 schedule_timeout(cfs_time_seconds(1));
3220 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3221 cpt = lnet_net_lock_current();
3223 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3225 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3228 lnet_net_unlock(cpt);
3230 CDEBUG(D_NET, "woken: %d\n", rc);
3236 * Messages that were pending on a destroyed peer will be put on a global
3237 * resend list. The message resend list will be checked by
3238 * the discovery thread when it wakes up, and will resend messages. These
3239 * messages can still be sendable in the case the lpni which was the initial
3240 * cause of the message re-queue was transfered to another peer.
3242 * It is possible that LNet could be shutdown while we're iterating
3243 * through the list. lnet_shudown_lndnets() will attempt to access the
3244 * resend list, but will have to wait until the spinlock is released, by
3245 * which time there shouldn't be any more messages on the resend list.
3246 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3247 * for the messages so they can be released. The other case is that
3248 * lnet_shudown_lndnets() can finalize all the messages before this
3249 * function can visit the resend list, in which case this function will be
3252 static void lnet_resend_msgs(void)
3254 struct lnet_msg *msg, *tmp;
3258 spin_lock(&the_lnet.ln_msg_resend_lock);
3259 list_splice(&the_lnet.ln_msg_resend, &resend);
3260 spin_unlock(&the_lnet.ln_msg_resend_lock);
3262 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3263 list_del_init(&msg->msg_list);
3264 rc = lnet_send(msg->msg_src_nid_param, msg,
3265 msg->msg_rtr_nid_param);
3267 CNETERR("Error sending %s to %s: %d\n",
3268 lnet_msgtyp2str(msg->msg_type),
3269 libcfs_id2str(msg->msg_target), rc);
3270 lnet_finalize(msg, rc);
3275 /* The discovery thread. */
3276 static int lnet_peer_discovery(void *arg)
3278 struct lnet_peer *lp;
3281 wait_for_completion(&the_lnet.ln_started);
3283 CDEBUG(D_NET, "started\n");
3284 cfs_block_allsigs();
3287 if (lnet_peer_discovery_wait_for_work())
3292 if (lnet_push_target_resize_needed())
3293 lnet_push_target_resize();
3295 lnet_net_lock(LNET_LOCK_EX);
3296 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3300 * Process all incoming discovery work requests. When
3301 * discovery must wait on a peer to change state, it
3302 * is added to the tail of the ln_dc_working queue. A
3303 * timestamp keeps track of when the peer was added,
3304 * so we can time out discovery requests that take too
3307 while (!list_empty(&the_lnet.ln_dc_request)) {
3308 lp = list_first_entry(&the_lnet.ln_dc_request,
3309 struct lnet_peer, lp_dc_list);
3310 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3312 * set the time the peer was put on the dc_working
3313 * queue. It shouldn't remain on the queue
3314 * forever, in case the GET message (for ping)
3315 * doesn't get a REPLY or the PUT message (for
3316 * push) doesn't get an ACK.
3318 lp->lp_last_queued = ktime_get_real_seconds();
3319 lnet_net_unlock(LNET_LOCK_EX);
3322 * Select an action depending on the state of
3323 * the peer and whether discovery is disabled.
3324 * The check whether discovery is disabled is
3325 * done after the code that handles processing
3326 * for arrived data, cleanup for failures, and
3327 * forcing a Ping or Push.
3329 spin_lock(&lp->lp_lock);
3330 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3331 libcfs_nid2str(lp->lp_primary_nid), lp,
3333 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3334 rc = lnet_peer_data_present(lp);
3335 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3336 rc = lnet_peer_ping_failed(lp);
3337 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3338 rc = lnet_peer_push_failed(lp);
3339 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3340 rc = lnet_peer_send_ping(lp);
3341 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3342 rc = lnet_peer_send_push(lp);
3343 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3344 rc = lnet_peer_send_ping(lp);
3345 else if (lnet_peer_needs_push(lp))
3346 rc = lnet_peer_send_push(lp);
3348 rc = lnet_peer_discovered(lp);
3349 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3350 libcfs_nid2str(lp->lp_primary_nid), lp,
3352 spin_unlock(&lp->lp_lock);
3354 lnet_net_lock(LNET_LOCK_EX);
3355 if (rc == LNET_REDISCOVER_PEER) {
3356 list_move(&lp->lp_dc_list,
3357 &the_lnet.ln_dc_request);
3359 lnet_peer_discovery_error(lp, rc);
3361 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3362 lnet_peer_discovery_complete(lp);
3363 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3367 lnet_net_unlock(LNET_LOCK_EX);
3370 CDEBUG(D_NET, "stopping\n");
3372 * Clean up before telling lnet_peer_discovery_stop() that
3373 * we're done. Use wake_up() below to somewhat reduce the
3374 * size of the thundering herd if there are multiple threads
3375 * waiting on discovery of a single peer.
3378 /* Queue cleanup 1: stop all pending pings and pushes. */
3379 lnet_net_lock(LNET_LOCK_EX);
3380 while (!list_empty(&the_lnet.ln_dc_working)) {
3381 lp = list_first_entry(&the_lnet.ln_dc_working,
3382 struct lnet_peer, lp_dc_list);
3383 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3384 lnet_net_unlock(LNET_LOCK_EX);
3385 lnet_peer_cancel_discovery(lp);
3386 lnet_net_lock(LNET_LOCK_EX);
3388 lnet_net_unlock(LNET_LOCK_EX);
3390 /* Queue cleanup 2: wait for the expired queue to clear. */
3391 while (!list_empty(&the_lnet.ln_dc_expired))
3392 schedule_timeout(cfs_time_seconds(1));
3394 /* Queue cleanup 3: clear the request queue. */
3395 lnet_net_lock(LNET_LOCK_EX);
3396 while (!list_empty(&the_lnet.ln_dc_request)) {
3397 lp = list_first_entry(&the_lnet.ln_dc_request,
3398 struct lnet_peer, lp_dc_list);
3399 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3400 lnet_peer_discovery_complete(lp);
3402 lnet_net_unlock(LNET_LOCK_EX);
3404 LNetEQFree(the_lnet.ln_dc_eqh);
3405 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3407 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3408 wake_up(&the_lnet.ln_dc_waitq);
3410 CDEBUG(D_NET, "stopped\n");
3415 /* ln_api_mutex is held on entry. */
3416 int lnet_peer_discovery_start(void)
3418 struct task_struct *task;
3421 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3424 rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3426 CERROR("Can't allocate discovery EQ: %d\n", rc);
3430 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3431 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3434 CERROR("Can't start peer discovery thread: %d\n", rc);
3436 LNetEQFree(the_lnet.ln_dc_eqh);
3437 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3439 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3442 CDEBUG(D_NET, "discovery start: %d\n", rc);
3447 /* ln_api_mutex is held on entry. */
3448 void lnet_peer_discovery_stop(void)
3450 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3453 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3454 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3456 /* In the LNetNIInit() path we may be stopping discovery before it
3457 * entered its work loop
3459 if (!completion_done(&the_lnet.ln_started))
3460 complete(&the_lnet.ln_started);
3462 wake_up(&the_lnet.ln_dc_waitq);
3464 wait_event(the_lnet.ln_dc_waitq,
3465 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3467 LASSERT(list_empty(&the_lnet.ln_dc_request));
3468 LASSERT(list_empty(&the_lnet.ln_dc_working));
3469 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3471 CDEBUG(D_NET, "discovery stopped\n");
3477 lnet_debug_peer(lnet_nid_t nid)
3479 char *aliveness = "NA";
3480 struct lnet_peer_ni *lp;
3483 cpt = lnet_cpt_of_nid(nid, NULL);
3486 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3488 lnet_net_unlock(cpt);
3489 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3493 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3494 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3496 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3497 libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3498 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3499 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3500 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3502 lnet_peer_ni_decref_locked(lp);
3504 lnet_net_unlock(cpt);
3507 /* Gathering information for userspace. */
3509 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3510 char aliveness[LNET_MAX_STR_LEN],
3511 __u32 *cpt_iter, __u32 *refcount,
3512 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3513 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3514 __u32 *peer_tx_qnob)
3516 struct lnet_peer_table *peer_table;
3517 struct lnet_peer_ni *lp;
3522 /* get the number of CPTs */
3523 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3525 /* if the cpt number to be examined is >= the number of cpts in
3526 * the system then indicate that there are no more cpts to examin
3528 if (*cpt_iter >= lncpt)
3531 /* get the current table */
3532 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3533 /* if the ptable is NULL then there are no more cpts to examine */
3534 if (peer_table == NULL)
3537 lnet_net_lock(*cpt_iter);
3539 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3540 struct list_head *peers = &peer_table->pt_hash[j];
3542 list_for_each_entry(lp, peers, lpni_hashlist) {
3543 if (peer_index-- > 0)
3546 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3547 if (lnet_isrouter(lp) ||
3548 lnet_peer_aliveness_enabled(lp))
3549 snprintf(aliveness, LNET_MAX_STR_LEN,
3550 lnet_is_peer_ni_alive(lp) ? "up" : "down");
3552 *nid = lp->lpni_nid;
3553 *refcount = atomic_read(&lp->lpni_refcount);
3554 *ni_peer_tx_credits =
3555 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3556 *peer_tx_credits = lp->lpni_txcredits;
3557 *peer_rtr_credits = lp->lpni_rtrcredits;
3558 *peer_min_rtr_credits = lp->lpni_mintxcredits;
3559 *peer_tx_qnob = lp->lpni_txqnob;
3565 lnet_net_unlock(*cpt_iter);
3569 return found ? 0 : -ENOENT;
3572 /* ln_api_mutex is held, which keeps the peer list stable */
3573 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3575 struct lnet_ioctl_element_stats *lpni_stats;
3576 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3577 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3578 struct lnet_peer_ni_credit_info *lpni_info;
3579 struct lnet_peer_ni *lpni;
3580 struct lnet_peer *lp;
3585 lp = lnet_find_peer(cfg->prcfg_prim_nid);
3592 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3593 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3594 size *= lp->lp_nnis;
3595 if (size > cfg->prcfg_size) {
3596 cfg->prcfg_size = size;
3601 cfg->prcfg_prim_nid = lp->lp_primary_nid;
3602 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3603 cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3604 cfg->prcfg_count = lp->lp_nnis;
3605 cfg->prcfg_size = size;
3606 cfg->prcfg_state = lp->lp_state;
3608 /* Allocate helper buffers. */
3610 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3613 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3616 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3617 if (!lpni_msg_stats)
3618 goto out_free_stats;
3619 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3621 goto out_free_msg_stats;
3626 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3627 nid = lpni->lpni_nid;
3628 if (copy_to_user(bulk, &nid, sizeof(nid)))
3629 goto out_free_hstats;
3630 bulk += sizeof(nid);
3632 memset(lpni_info, 0, sizeof(*lpni_info));
3633 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3634 if (lnet_isrouter(lpni) ||
3635 lnet_peer_aliveness_enabled(lpni))
3636 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3637 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3639 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3640 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3641 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3642 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3643 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3644 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3645 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3646 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3647 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3648 goto out_free_hstats;
3649 bulk += sizeof(*lpni_info);
3651 memset(lpni_stats, 0, sizeof(*lpni_stats));
3652 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3653 LNET_STATS_TYPE_SEND);
3654 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3655 LNET_STATS_TYPE_RECV);
3656 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3657 LNET_STATS_TYPE_DROP);
3658 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3659 goto out_free_hstats;
3660 bulk += sizeof(*lpni_stats);
3661 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3662 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3663 goto out_free_hstats;
3664 bulk += sizeof(*lpni_msg_stats);
3665 lpni_hstats->hlpni_network_timeout =
3666 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3667 lpni_hstats->hlpni_remote_dropped =
3668 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3669 lpni_hstats->hlpni_remote_timeout =
3670 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3671 lpni_hstats->hlpni_remote_error =
3672 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3673 lpni_hstats->hlpni_health_value =
3674 atomic_read(&lpni->lpni_healthv);
3675 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3676 goto out_free_hstats;
3677 bulk += sizeof(*lpni_hstats);
3682 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3684 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3686 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3688 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3690 lnet_peer_decref_locked(lp);
3696 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3698 /* the mt could've shutdown and cleaned up the queues */
3699 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3702 if (list_empty(&lpni->lpni_recovery) &&
3703 atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3704 CERROR("lpni %s added to recovery queue. Health = %d\n",
3705 libcfs_nid2str(lpni->lpni_nid),
3706 atomic_read(&lpni->lpni_healthv));
3707 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3708 lnet_peer_ni_addref_locked(lpni);
3712 /* Call with the ln_api_mutex held */
3714 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3716 struct lnet_peer_table *ptable;
3717 struct lnet_peer *lp;
3718 struct lnet_peer_net *lpn;
3719 struct lnet_peer_ni *lpni;
3723 if (the_lnet.ln_state != LNET_STATE_RUNNING)
3727 lnet_net_lock(LNET_LOCK_EX);
3728 lpni = lnet_find_peer_ni_locked(nid);
3730 lnet_net_unlock(LNET_LOCK_EX);
3733 atomic_set(&lpni->lpni_healthv, value);
3734 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3735 lnet_peer_ni_decref_locked(lpni);
3736 lnet_net_unlock(LNET_LOCK_EX);
3740 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3743 * Walk all the peers and reset the healhv for each one to the
3746 lnet_net_lock(LNET_LOCK_EX);
3747 for (cpt = 0; cpt < lncpt; cpt++) {
3748 ptable = the_lnet.ln_peer_tables[cpt];
3749 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3750 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3751 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3753 atomic_set(&lpni->lpni_healthv, value);
3754 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3759 lnet_net_unlock(LNET_LOCK_EX);