4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
35 #define DEBUG_SUBSYSTEM S_LNET
37 #include <linux/sched.h>
38 #ifdef HAVE_SCHED_HEADERS
39 #include <linux/sched/signal.h>
41 #include <linux/uaccess.h>
43 #include <lnet/lib-lnet.h>
44 #include <uapi/linux/lnet/lnet-dlc.h>
46 /* Value indicating that recovery needs to re-check a peer immediately. */
47 #define LNET_REDISCOVER_PEER (1)
49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
54 if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
55 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
56 lnet_peer_ni_decref_locked(lpni);
61 lnet_peer_net_added(struct lnet_net *net)
63 struct lnet_peer_ni *lpni, *tmp;
65 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
66 lpni_on_remote_peer_ni_list) {
68 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
71 spin_lock(&lpni->lpni_lock);
72 lpni->lpni_txcredits =
73 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
74 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
75 lpni->lpni_rtrcredits =
76 lnet_peer_buffer_credits(lpni->lpni_net);
77 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
78 spin_unlock(&lpni->lpni_lock);
80 lnet_peer_remove_from_remote_list(lpni);
86 lnet_peer_tables_destroy(void)
88 struct lnet_peer_table *ptable;
89 struct list_head *hash;
93 if (!the_lnet.ln_peer_tables)
96 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
97 hash = ptable->pt_hash;
98 if (!hash) /* not intialized */
101 LASSERT(list_empty(&ptable->pt_zombie_list));
103 ptable->pt_hash = NULL;
104 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
105 LASSERT(list_empty(&hash[j]));
107 LIBCFS_FREE(hash, LNET_PEER_HASH_SIZE * sizeof(*hash));
110 cfs_percpt_free(the_lnet.ln_peer_tables);
111 the_lnet.ln_peer_tables = NULL;
115 lnet_peer_tables_create(void)
117 struct lnet_peer_table *ptable;
118 struct list_head *hash;
122 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
124 if (the_lnet.ln_peer_tables == NULL) {
125 CERROR("Failed to allocate cpu-partition peer tables\n");
129 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131 LNET_PEER_HASH_SIZE * sizeof(*hash));
133 CERROR("Failed to create peer hash table\n");
134 lnet_peer_tables_destroy();
138 spin_lock_init(&ptable->pt_zombie_lock);
139 INIT_LIST_HEAD(&ptable->pt_zombie_list);
141 INIT_LIST_HEAD(&ptable->pt_peer_list);
143 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144 INIT_LIST_HEAD(&hash[j]);
145 ptable->pt_hash = hash; /* sign of initialization */
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(lnet_nid_t nid)
154 struct lnet_peer_ni *lpni;
155 struct lnet_net *net;
158 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
160 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
164 INIT_LIST_HEAD(&lpni->lpni_txq);
165 INIT_LIST_HEAD(&lpni->lpni_hashlist);
166 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167 INIT_LIST_HEAD(&lpni->lpni_recovery);
168 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171 spin_lock_init(&lpni->lpni_lock);
173 if (lnet_peers_start_down())
174 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
176 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
177 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
178 lpni->lpni_nid = nid;
179 lpni->lpni_cpt = cpt;
180 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
182 net = lnet_get_net_locked(LNET_NIDNET(nid));
183 lpni->lpni_net = net;
185 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
186 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
187 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
188 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
191 * This peer_ni is not on a local network, so we
192 * cannot add the credits here. In case the net is
193 * added later, add the peer_ni to the remote peer ni
194 * list so it can be easily found and revisited.
196 /* FIXME: per-net implementation instead? */
197 atomic_inc(&lpni->lpni_refcount);
198 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
199 &the_lnet.ln_remote_peer_ni_list);
202 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
207 static struct lnet_peer_net *
208 lnet_peer_net_alloc(__u32 net_id)
210 struct lnet_peer_net *lpn;
212 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
216 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
217 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
218 lpn->lpn_net_id = net_id;
220 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
226 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
228 struct lnet_peer *lp;
230 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
232 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
233 LASSERT(list_empty(&lpn->lpn_peer_nis));
234 LASSERT(list_empty(&lpn->lpn_peer_nets));
236 lpn->lpn_peer = NULL;
237 LIBCFS_FREE(lpn, sizeof(*lpn));
239 lnet_peer_decref_locked(lp);
242 static struct lnet_peer *
243 lnet_peer_alloc(lnet_nid_t nid)
245 struct lnet_peer *lp;
247 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
251 INIT_LIST_HEAD(&lp->lp_rtrq);
252 INIT_LIST_HEAD(&lp->lp_routes);
253 INIT_LIST_HEAD(&lp->lp_peer_list);
254 INIT_LIST_HEAD(&lp->lp_peer_nets);
255 INIT_LIST_HEAD(&lp->lp_dc_list);
256 INIT_LIST_HEAD(&lp->lp_dc_pendq);
257 INIT_LIST_HEAD(&lp->lp_rtr_list);
258 init_waitqueue_head(&lp->lp_dc_waitq);
259 spin_lock_init(&lp->lp_lock);
260 lp->lp_primary_nid = nid;
263 * all peers created on a router should have health on
264 * if it's not already on.
266 if (the_lnet.ln_routing && !lnet_health_sensitivity)
267 lp->lp_health_sensitivity = 1;
270 * Turn off discovery for loopback peer. If you're creating a peer
271 * for the loopback interface then that was initiated when we
272 * attempted to send a message over the loopback. There is no need
273 * to ever use a different interface when sending messages to
276 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
277 lp->lp_state = LNET_PEER_NO_DISCOVERY;
278 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
280 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
286 lnet_destroy_peer_locked(struct lnet_peer *lp)
288 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
290 LASSERT(atomic_read(&lp->lp_refcount) == 0);
291 LASSERT(lp->lp_rtr_refcount == 0);
292 LASSERT(list_empty(&lp->lp_peer_nets));
293 LASSERT(list_empty(&lp->lp_peer_list));
294 LASSERT(list_empty(&lp->lp_dc_list));
297 lnet_ping_buffer_decref(lp->lp_data);
300 * if there are messages still on the pending queue, then make
301 * sure to queue them on the ln_msg_resend list so they can be
302 * resent at a later point if the discovery thread is still
304 * If the discovery thread has stopped, then the wakeup will be a
305 * no-op, and it is expected the lnet_shutdown_lndnets() will
306 * eventually be called, which will traverse this list and
307 * finalize the messages on the list.
308 * We can not resend them now because we're holding the cpt lock.
309 * Releasing the lock can cause an inconsistent state
311 spin_lock(&the_lnet.ln_msg_resend_lock);
312 spin_lock(&lp->lp_lock);
313 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
314 spin_unlock(&lp->lp_lock);
315 spin_unlock(&the_lnet.ln_msg_resend_lock);
316 wake_up(&the_lnet.ln_dc_waitq);
318 LIBCFS_FREE(lp, sizeof(*lp));
322 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
323 * that peer_net, detach the peer_net from the peer.
325 * Call with lnet_net_lock/EX held
328 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
330 struct lnet_peer_table *ptable;
331 struct lnet_peer_net *lpn;
332 struct lnet_peer *lp;
335 * Belts and suspenders: gracefully handle teardown of a
336 * partially connected peer_ni.
338 lpn = lpni->lpni_peer_net;
340 list_del_init(&lpni->lpni_peer_nis);
342 * If there are no lpni's left, we detach lpn from
343 * lp_peer_nets, so it cannot be found anymore.
345 if (list_empty(&lpn->lpn_peer_nis))
346 list_del_init(&lpn->lpn_peer_nets);
348 /* Update peer NID count. */
353 * If there are no more peer nets, make the peer unfindable
354 * via the peer_tables.
356 * Otherwise, if the peer is DISCOVERED, tell discovery to
357 * take another look at it. This is a no-op if discovery for
358 * this peer did the detaching.
360 if (list_empty(&lp->lp_peer_nets)) {
361 list_del_init(&lp->lp_peer_list);
362 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
364 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
365 /* Discovery isn't running, nothing to do here. */
366 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
367 lnet_peer_queue_for_discovery(lp);
368 wake_up(&the_lnet.ln_dc_waitq);
370 CDEBUG(D_NET, "peer %s NID %s\n",
371 libcfs_nid2str(lp->lp_primary_nid),
372 libcfs_nid2str(lpni->lpni_nid));
375 /* called with lnet_net_lock LNET_LOCK_EX held */
377 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
379 struct lnet_peer_table *ptable = NULL;
381 /* don't remove a peer_ni if it's also a gateway */
382 if (lnet_isrouter(lpni) && !force) {
383 CERROR("Peer NI %s is a gateway. Can not delete it\n",
384 libcfs_nid2str(lpni->lpni_nid));
388 lnet_peer_remove_from_remote_list(lpni);
390 /* remove peer ni from the hash list. */
391 list_del_init(&lpni->lpni_hashlist);
394 * indicate the peer is being deleted so the monitor thread can
395 * remove it from the recovery queue.
397 spin_lock(&lpni->lpni_lock);
398 lpni->lpni_state |= LNET_PEER_NI_DELETING;
399 spin_unlock(&lpni->lpni_lock);
401 /* decrement the ref count on the peer table */
402 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
403 LASSERT(ptable->pt_number > 0);
407 * The peer_ni can no longer be found with a lookup. But there
408 * can be current users, so keep track of it on the zombie
409 * list until the reference count has gone to zero.
411 * The last reference may be lost in a place where the
412 * lnet_net_lock locks only a single cpt, and that cpt may not
413 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
416 spin_lock(&ptable->pt_zombie_lock);
417 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
418 ptable->pt_zombies++;
419 spin_unlock(&ptable->pt_zombie_lock);
421 /* no need to keep this peer_ni on the hierarchy anymore */
422 lnet_peer_detach_peer_ni_locked(lpni);
424 /* remove hashlist reference on peer_ni */
425 lnet_peer_ni_decref_locked(lpni);
430 void lnet_peer_uninit(void)
432 struct lnet_peer_ni *lpni, *tmp;
434 lnet_net_lock(LNET_LOCK_EX);
436 /* remove all peer_nis from the remote peer and the hash list */
437 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
438 lpni_on_remote_peer_ni_list)
439 lnet_peer_ni_del_locked(lpni, false);
441 lnet_peer_tables_destroy();
443 lnet_net_unlock(LNET_LOCK_EX);
447 lnet_peer_del_locked(struct lnet_peer *peer)
449 struct lnet_peer_ni *lpni = NULL, *lpni2;
452 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
454 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
455 while (lpni != NULL) {
456 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
457 rc = lnet_peer_ni_del_locked(lpni, false);
467 lnet_peer_del(struct lnet_peer *peer)
469 lnet_net_lock(LNET_LOCK_EX);
470 lnet_peer_del_locked(peer);
471 lnet_net_unlock(LNET_LOCK_EX);
477 * Delete a NID from a peer. Call with ln_api_mutex held.
480 * -EPERM: Non-DLC deletion from DLC-configured peer.
481 * -ENOENT: No lnet_peer_ni corresponding to the nid.
482 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
483 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
486 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
488 struct lnet_peer_ni *lpni;
489 lnet_nid_t primary_nid = lp->lp_primary_nid;
491 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
493 if (!(flags & LNET_PEER_CONFIGURED)) {
494 if (lp->lp_state & LNET_PEER_CONFIGURED) {
499 lpni = lnet_find_peer_ni_locked(nid);
504 lnet_peer_ni_decref_locked(lpni);
505 if (lp != lpni->lpni_peer_net->lpn_peer) {
511 * This function only allows deletion of the primary NID if it
514 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
519 lnet_net_lock(LNET_LOCK_EX);
521 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
522 struct lnet_peer_ni *lpni2;
523 /* assign the next peer_ni to be the primary */
524 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
526 lp->lp_primary_nid = lpni->lpni_nid;
528 rc = lnet_peer_ni_del_locked(lpni, force);
530 lnet_net_unlock(LNET_LOCK_EX);
533 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
534 libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
540 lnet_peer_table_cleanup_locked(struct lnet_net *net,
541 struct lnet_peer_table *ptable)
544 struct lnet_peer_ni *next;
545 struct lnet_peer_ni *lpni;
546 struct lnet_peer *peer;
548 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
549 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
551 if (net != NULL && net != lpni->lpni_net)
554 peer = lpni->lpni_peer_net->lpn_peer;
555 if (peer->lp_primary_nid != lpni->lpni_nid) {
556 lnet_peer_ni_del_locked(lpni, false);
560 * Removing the primary NID implies removing
561 * the entire peer. Advance next beyond any
562 * peer_ni that belongs to the same peer.
564 list_for_each_entry_from(next, &ptable->pt_hash[i],
566 if (next->lpni_peer_net->lpn_peer != peer)
569 lnet_peer_del_locked(peer);
575 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
579 spin_lock(&ptable->pt_zombie_lock);
580 while (ptable->pt_zombies) {
581 spin_unlock(&ptable->pt_zombie_lock);
583 if (is_power_of_2(i)) {
585 "Waiting for %d zombies on peer table\n",
588 set_current_state(TASK_UNINTERRUPTIBLE);
589 schedule_timeout(cfs_time_seconds(1) >> 1);
590 spin_lock(&ptable->pt_zombie_lock);
592 spin_unlock(&ptable->pt_zombie_lock);
596 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
597 struct lnet_peer_table *ptable)
599 struct lnet_peer_ni *lp;
600 struct lnet_peer_ni *tmp;
604 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
605 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
607 if (net != lp->lpni_net)
610 if (!lnet_isrouter(lp))
613 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
615 lnet_net_unlock(LNET_LOCK_EX);
616 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
617 lnet_net_lock(LNET_LOCK_EX);
623 lnet_peer_tables_cleanup(struct lnet_net *net)
626 struct lnet_peer_table *ptable;
628 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
629 /* If just deleting the peers for a NI, get rid of any routes these
630 * peers are gateways for. */
631 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
632 lnet_net_lock(LNET_LOCK_EX);
633 lnet_peer_table_del_rtrs_locked(net, ptable);
634 lnet_net_unlock(LNET_LOCK_EX);
637 /* Start the cleanup process */
638 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
639 lnet_net_lock(LNET_LOCK_EX);
640 lnet_peer_table_cleanup_locked(net, ptable);
641 lnet_net_unlock(LNET_LOCK_EX);
644 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
645 lnet_peer_ni_finalize_wait(ptable);
648 static struct lnet_peer_ni *
649 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
651 struct list_head *peers;
652 struct lnet_peer_ni *lp;
654 LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
656 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
657 list_for_each_entry(lp, peers, lpni_hashlist) {
658 if (lp->lpni_nid == nid) {
659 lnet_peer_ni_addref_locked(lp);
667 struct lnet_peer_ni *
668 lnet_find_peer_ni_locked(lnet_nid_t nid)
670 struct lnet_peer_ni *lpni;
671 struct lnet_peer_table *ptable;
674 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
676 ptable = the_lnet.ln_peer_tables[cpt];
677 lpni = lnet_get_peer_ni_locked(ptable, nid);
682 struct lnet_peer_ni *
683 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
685 struct lnet_peer_net *lpn;
686 struct lnet_peer_ni *lpni;
688 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
692 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
693 if (lpni->lpni_nid == nid)
701 lnet_find_peer(lnet_nid_t nid)
703 struct lnet_peer_ni *lpni;
704 struct lnet_peer *lp = NULL;
707 cpt = lnet_net_lock_current();
708 lpni = lnet_find_peer_ni_locked(nid);
710 lp = lpni->lpni_peer_net->lpn_peer;
711 lnet_peer_addref_locked(lp);
712 lnet_peer_ni_decref_locked(lpni);
714 lnet_net_unlock(cpt);
719 struct lnet_peer_net *
720 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
722 struct lnet_peer_net *net;
725 /* no net id provided return the first net */
726 net = list_first_entry_or_null(&lp->lp_peer_nets,
727 struct lnet_peer_net,
733 /* find the net after the one provided */
734 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
735 if (net->lpn_net_id == prev_lpn_id) {
737 * if we reached the end of the list loop to the
740 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
741 return list_first_entry_or_null(&lp->lp_peer_nets,
742 struct lnet_peer_net,
745 return list_next_entry(net, lpn_peer_nets);
752 struct lnet_peer_ni *
753 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
754 struct lnet_peer_net *peer_net,
755 struct lnet_peer_ni *prev)
757 struct lnet_peer_ni *lpni;
758 struct lnet_peer_net *net = peer_net;
762 if (list_empty(&peer->lp_peer_nets))
765 net = list_entry(peer->lp_peer_nets.next,
766 struct lnet_peer_net,
769 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
775 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
777 * if you reached the end of the peer ni list and the peer
778 * net is specified then there are no more peer nis in that
785 * we reached the end of this net ni list. move to the
788 if (prev->lpni_peer_net->lpn_peer_nets.next ==
790 /* no more nets and no more NIs. */
793 /* get the next net */
794 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
795 struct lnet_peer_net,
797 /* get the ni on it */
798 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
804 /* there are more nis left */
805 lpni = list_entry(prev->lpni_peer_nis.next,
806 struct lnet_peer_ni, lpni_peer_nis);
811 /* Call with the ln_api_mutex held */
812 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
814 struct lnet_process_id id;
815 struct lnet_peer_table *ptable;
816 struct lnet_peer *lp;
825 if (the_lnet.ln_state != LNET_STATE_RUNNING)
828 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
831 * Count the number of peers, and return E2BIG if the buffer
832 * is too small. We'll also return the desired size.
835 for (cpt = 0; cpt < lncpt; cpt++) {
836 ptable = the_lnet.ln_peer_tables[cpt];
837 count += ptable->pt_peers;
839 size = count * sizeof(*ids);
844 * Walk the peer lists and copy out the primary nids.
845 * This is safe because the peer lists are only modified
846 * while the ln_api_mutex is held. So we don't need to
847 * hold the lnet_net_lock as well, and can therefore
848 * directly call copy_to_user().
851 memset(&id, 0, sizeof(id));
852 id.pid = LNET_PID_LUSTRE;
854 for (cpt = 0; cpt < lncpt; cpt++) {
855 ptable = the_lnet.ln_peer_tables[cpt];
856 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
859 id.nid = lp->lp_primary_nid;
860 if (copy_to_user(&ids[i], &id, sizeof(id)))
873 * Start pushes to peers that need to be updated for a configuration
874 * change on this node.
877 lnet_push_update_to_peers(int force)
879 struct lnet_peer_table *ptable;
880 struct lnet_peer *lp;
884 lnet_net_lock(LNET_LOCK_EX);
885 if (lnet_peer_discovery_disabled)
887 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
888 for (cpt = 0; cpt < lncpt; cpt++) {
889 ptable = the_lnet.ln_peer_tables[cpt];
890 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
892 spin_lock(&lp->lp_lock);
893 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
894 lp->lp_state |= LNET_PEER_FORCE_PUSH;
895 spin_unlock(&lp->lp_lock);
897 if (lnet_peer_needs_push(lp))
898 lnet_peer_queue_for_discovery(lp);
901 lnet_net_unlock(LNET_LOCK_EX);
902 wake_up(&the_lnet.ln_dc_waitq);
906 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
907 * this is a preferred point-to-point path. Call with lnet_net_lock in
911 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
915 if (lpni->lpni_pref_nnids == 0)
917 if (lpni->lpni_pref_nnids == 1)
918 return lpni->lpni_pref.nid == nid;
919 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
920 if (lpni->lpni_pref.nids[i] == nid)
927 * Set a single ni as preferred, provided no preferred ni is already
928 * defined. Only to be used for non-multi-rail peer_ni.
931 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
935 spin_lock(&lpni->lpni_lock);
936 if (nid == LNET_NID_ANY) {
938 } else if (lpni->lpni_pref_nnids > 0) {
940 } else if (lpni->lpni_pref_nnids == 0) {
941 lpni->lpni_pref.nid = nid;
942 lpni->lpni_pref_nnids = 1;
943 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
945 spin_unlock(&lpni->lpni_lock);
947 CDEBUG(D_NET, "peer %s nid %s: %d\n",
948 libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
953 * Clear the preferred NID from a non-multi-rail peer_ni, provided
954 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
957 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
961 spin_lock(&lpni->lpni_lock);
962 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
963 lpni->lpni_pref_nnids = 0;
964 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
965 } else if (lpni->lpni_pref_nnids == 0) {
970 spin_unlock(&lpni->lpni_lock);
972 CDEBUG(D_NET, "peer %s: %d\n",
973 libcfs_nid2str(lpni->lpni_nid), rc);
978 * Clear the preferred NIDs from a non-multi-rail peer.
981 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
983 struct lnet_peer_ni *lpni = NULL;
985 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
986 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
990 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
992 lnet_nid_t *nids = NULL;
993 lnet_nid_t *oldnids = NULL;
994 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
999 if (nid == LNET_NID_ANY) {
1004 if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1009 /* A non-MR node may have only one preferred NI per peer_ni */
1010 if (lpni->lpni_pref_nnids > 0) {
1011 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1017 if (lpni->lpni_pref_nnids != 0) {
1018 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1019 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1024 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1025 if (lpni->lpni_pref.nids[i] == nid) {
1026 LIBCFS_FREE(nids, size);
1030 nids[i] = lpni->lpni_pref.nids[i];
1035 lnet_net_lock(LNET_LOCK_EX);
1036 spin_lock(&lpni->lpni_lock);
1037 if (lpni->lpni_pref_nnids == 0) {
1038 lpni->lpni_pref.nid = nid;
1040 oldnids = lpni->lpni_pref.nids;
1041 lpni->lpni_pref.nids = nids;
1043 lpni->lpni_pref_nnids++;
1044 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1045 spin_unlock(&lpni->lpni_lock);
1046 lnet_net_unlock(LNET_LOCK_EX);
1049 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1050 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1053 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1054 spin_lock(&lpni->lpni_lock);
1055 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1056 spin_unlock(&lpni->lpni_lock);
1058 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1059 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1064 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1066 lnet_nid_t *nids = NULL;
1067 lnet_nid_t *oldnids = NULL;
1068 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1073 if (lpni->lpni_pref_nnids == 0) {
1078 if (lpni->lpni_pref_nnids == 1) {
1079 if (lpni->lpni_pref.nid != nid) {
1083 } else if (lpni->lpni_pref_nnids == 2) {
1084 if (lpni->lpni_pref.nids[0] != nid &&
1085 lpni->lpni_pref.nids[1] != nid) {
1090 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1091 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1096 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1097 if (lpni->lpni_pref.nids[i] != nid)
1099 nids[j++] = lpni->lpni_pref.nids[i];
1101 /* Check if we actually removed a nid. */
1102 if (j == lpni->lpni_pref_nnids) {
1103 LIBCFS_FREE(nids, size);
1109 lnet_net_lock(LNET_LOCK_EX);
1110 spin_lock(&lpni->lpni_lock);
1111 if (lpni->lpni_pref_nnids == 1) {
1112 lpni->lpni_pref.nid = LNET_NID_ANY;
1113 } else if (lpni->lpni_pref_nnids == 2) {
1114 oldnids = lpni->lpni_pref.nids;
1115 if (oldnids[0] == nid)
1116 lpni->lpni_pref.nid = oldnids[1];
1118 lpni->lpni_pref.nid = oldnids[2];
1120 oldnids = lpni->lpni_pref.nids;
1121 lpni->lpni_pref.nids = nids;
1123 lpni->lpni_pref_nnids--;
1124 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1125 spin_unlock(&lpni->lpni_lock);
1126 lnet_net_unlock(LNET_LOCK_EX);
1129 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1130 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1133 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1134 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1139 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1141 struct lnet_peer_ni *lpni;
1142 lnet_nid_t primary_nid = nid;
1144 lpni = lnet_find_peer_ni_locked(nid);
1146 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1147 lnet_peer_ni_decref_locked(lpni);
1154 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1156 if (lnet_peer_discovery_disabled)
1159 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1160 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1171 lnet_is_discovery_disabled(struct lnet_peer *lp)
1175 spin_lock(&lp->lp_lock);
1176 rc = lnet_is_discovery_disabled_locked(lp);
1177 spin_unlock(&lp->lp_lock);
1183 LNetPrimaryNID(lnet_nid_t nid)
1185 struct lnet_peer *lp;
1186 struct lnet_peer_ni *lpni;
1187 lnet_nid_t primary_nid = nid;
1191 cpt = lnet_net_lock_current();
1192 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1197 lp = lpni->lpni_peer_net->lpn_peer;
1199 while (!lnet_peer_is_uptodate(lp)) {
1200 rc = lnet_discover_peer_locked(lpni, cpt, true);
1203 lp = lpni->lpni_peer_net->lpn_peer;
1205 /* Only try once if discovery is disabled */
1206 if (lnet_is_discovery_disabled(lp))
1209 primary_nid = lp->lp_primary_nid;
1211 lnet_peer_ni_decref_locked(lpni);
1213 lnet_net_unlock(cpt);
1215 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1216 libcfs_nid2str(primary_nid), rc);
1219 EXPORT_SYMBOL(LNetPrimaryNID);
1221 struct lnet_peer_net *
1222 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1224 struct lnet_peer_net *peer_net;
1225 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1226 if (peer_net->lpn_net_id == net_id)
1233 * Attach a peer_ni to a peer_net and peer. This function assumes
1234 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1235 * may be attached to a different peer, in which case it will be
1236 * properly detached first. The whole operation is done atomically.
1238 * Always returns 0. This is the last function called from functions
1239 * that do return an int, so returning 0 here allows the compiler to
1243 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1244 struct lnet_peer_net *lpn,
1245 struct lnet_peer_ni *lpni,
1248 struct lnet_peer_table *ptable;
1250 /* Install the new peer_ni */
1251 lnet_net_lock(LNET_LOCK_EX);
1252 /* Add peer_ni to global peer table hash, if necessary. */
1253 if (list_empty(&lpni->lpni_hashlist)) {
1254 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1256 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1257 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1258 ptable->pt_version++;
1259 ptable->pt_number++;
1260 /* This is the 1st refcount on lpni. */
1261 atomic_inc(&lpni->lpni_refcount);
1264 /* Detach the peer_ni from an existing peer, if necessary. */
1265 if (lpni->lpni_peer_net) {
1266 LASSERT(lpni->lpni_peer_net != lpn);
1267 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1268 lnet_peer_detach_peer_ni_locked(lpni);
1269 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1270 lpni->lpni_peer_net = NULL;
1273 /* Add peer_ni to peer_net */
1274 lpni->lpni_peer_net = lpn;
1275 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1276 lnet_peer_net_addref_locked(lpn);
1278 /* Add peer_net to peer */
1279 if (!lpn->lpn_peer) {
1281 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1282 lnet_peer_addref_locked(lp);
1285 /* Add peer to global peer list, if necessary */
1286 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1287 if (list_empty(&lp->lp_peer_list)) {
1288 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1293 /* Update peer state */
1294 spin_lock(&lp->lp_lock);
1295 if (flags & LNET_PEER_CONFIGURED) {
1296 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1297 lp->lp_state |= LNET_PEER_CONFIGURED;
1299 if (flags & LNET_PEER_MULTI_RAIL) {
1300 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1301 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1302 lnet_peer_clr_non_mr_pref_nids(lp);
1305 spin_unlock(&lp->lp_lock);
1308 lnet_net_unlock(LNET_LOCK_EX);
1310 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1311 libcfs_nid2str(lp->lp_primary_nid),
1312 libcfs_nid2str(lpni->lpni_nid), flags);
1318 * Create a new peer, with nid as its primary nid.
1320 * Call with the lnet_api_mutex held.
1323 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1325 struct lnet_peer *lp;
1326 struct lnet_peer_net *lpn;
1327 struct lnet_peer_ni *lpni;
1330 LASSERT(nid != LNET_NID_ANY);
1333 * No need for the lnet_net_lock here, because the
1334 * lnet_api_mutex is held.
1336 lpni = lnet_find_peer_ni_locked(nid);
1338 /* A peer with this NID already exists. */
1339 lp = lpni->lpni_peer_net->lpn_peer;
1340 lnet_peer_ni_decref_locked(lpni);
1342 * This is an error if the peer was configured and the
1343 * primary NID differs or an attempt is made to change
1344 * the Multi-Rail flag. Otherwise the assumption is
1345 * that an existing peer is being modified.
1347 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1348 if (lp->lp_primary_nid != nid)
1350 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1354 /* Delete and recreate as a configured peer. */
1358 /* Create peer, peer_net, and peer_ni. */
1360 lp = lnet_peer_alloc(nid);
1363 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1366 lpni = lnet_peer_ni_alloc(nid);
1370 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1373 LIBCFS_FREE(lpn, sizeof(*lpn));
1375 LIBCFS_FREE(lp, sizeof(*lp));
1377 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1378 libcfs_nid2str(nid), flags, rc);
1383 * Add a NID to a peer. Call with ln_api_mutex held.
1386 * -EPERM: Non-DLC addition to a DLC-configured peer.
1387 * -EEXIST: The NID was configured by DLC for a different peer.
1388 * -ENOMEM: Out of memory.
1389 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1390 * non-multi-rail peer.
1393 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1395 struct lnet_peer_net *lpn;
1396 struct lnet_peer_ni *lpni;
1400 LASSERT(nid != LNET_NID_ANY);
1402 /* A configured peer can only be updated through configuration. */
1403 if (!(flags & LNET_PEER_CONFIGURED)) {
1404 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1411 * The MULTI_RAIL flag can be set but not cleared, because
1412 * that would leave the peer struct in an invalid state.
1414 if (flags & LNET_PEER_MULTI_RAIL) {
1415 spin_lock(&lp->lp_lock);
1416 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1417 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1418 lnet_peer_clr_non_mr_pref_nids(lp);
1420 spin_unlock(&lp->lp_lock);
1421 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1426 lpni = lnet_find_peer_ni_locked(nid);
1429 * A peer_ni already exists. This is only a problem if
1430 * it is not connected to this peer and was configured
1433 lnet_peer_ni_decref_locked(lpni);
1434 if (lpni->lpni_peer_net->lpn_peer == lp)
1436 if (lnet_peer_ni_is_configured(lpni)) {
1440 /* If this is the primary NID, destroy the peer. */
1441 if (lnet_peer_ni_is_primary(lpni)) {
1442 struct lnet_peer *rtr_lp =
1443 lpni->lpni_peer_net->lpn_peer;
1444 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1446 * if we're trying to delete a router it means
1447 * we're moving this peer NI to a new peer so must
1448 * transfer router properties to the new peer
1450 if (rtr_refcount > 0) {
1451 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1452 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1454 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1455 lpni = lnet_peer_ni_alloc(nid);
1462 lpni = lnet_peer_ni_alloc(nid);
1470 * Get the peer_net. Check that we're not adding a second
1471 * peer_ni on a peer_net of a non-multi-rail peer.
1473 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1475 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1480 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1485 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1488 /* If the peer_ni was allocated above its peer_net pointer is NULL */
1489 if (!lpni->lpni_peer_net)
1490 LIBCFS_FREE(lpni, sizeof(*lpni));
1492 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1493 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1499 * Update the primary NID of a peer, if possible.
1501 * Call with the lnet_api_mutex held.
1504 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1506 lnet_nid_t old = lp->lp_primary_nid;
1509 if (lp->lp_primary_nid == nid)
1511 rc = lnet_peer_add_nid(lp, nid, flags);
1514 lp->lp_primary_nid = nid;
1516 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1517 libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1522 * lpni creation initiated due to traffic either sending or receiving.
1525 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1527 struct lnet_peer *lp;
1528 struct lnet_peer_net *lpn;
1529 struct lnet_peer_ni *lpni;
1531 * Assume peer is Multi-Rail capable and let discovery find out
1534 unsigned flags = LNET_PEER_MULTI_RAIL;
1537 if (nid == LNET_NID_ANY) {
1542 /* lnet_net_lock is not needed here because ln_api_lock is held */
1543 lpni = lnet_find_peer_ni_locked(nid);
1546 * We must have raced with another thread. Since we
1547 * know next to nothing about a peer_ni created by
1548 * traffic, we just assume everything is ok and
1551 lnet_peer_ni_decref_locked(lpni);
1555 /* Create peer, peer_net, and peer_ni. */
1557 lp = lnet_peer_alloc(nid);
1560 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1563 lpni = lnet_peer_ni_alloc(nid);
1566 if (pref != LNET_NID_ANY)
1567 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1569 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1572 LIBCFS_FREE(lpn, sizeof(*lpn));
1574 LIBCFS_FREE(lp, sizeof(*lp));
1576 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1581 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1583 * This API handles the following combinations:
1584 * Create a peer with its primary NI if only the prim_nid is provided
1585 * Add a NID to a peer identified by the prim_nid. The peer identified
1586 * by the prim_nid must already exist.
1587 * The peer being created may be non-MR.
1589 * The caller must hold ln_api_mutex. This prevents the peer from
1590 * being created/modified/deleted by a different thread.
1593 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1595 struct lnet_peer *lp = NULL;
1596 struct lnet_peer_ni *lpni;
1599 /* The prim_nid must always be specified */
1600 if (prim_nid == LNET_NID_ANY)
1603 flags = LNET_PEER_CONFIGURED;
1605 flags |= LNET_PEER_MULTI_RAIL;
1608 * If nid isn't specified, we must create a new peer with
1609 * prim_nid as its primary nid.
1611 if (nid == LNET_NID_ANY)
1612 return lnet_peer_add(prim_nid, flags);
1614 /* Look up the prim_nid, which must exist. */
1615 lpni = lnet_find_peer_ni_locked(prim_nid);
1618 lnet_peer_ni_decref_locked(lpni);
1619 lp = lpni->lpni_peer_net->lpn_peer;
1621 /* Peer must have been configured. */
1622 if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1623 CDEBUG(D_NET, "peer %s was not configured\n",
1624 libcfs_nid2str(prim_nid));
1628 /* Primary NID must match */
1629 if (lp->lp_primary_nid != prim_nid) {
1630 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1631 libcfs_nid2str(prim_nid),
1632 libcfs_nid2str(lp->lp_primary_nid));
1636 /* Multi-Rail flag must match. */
1637 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1638 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1639 libcfs_nid2str(prim_nid));
1643 return lnet_peer_add_nid(lp, nid, flags);
1647 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1649 * This API handles the following combinations:
1650 * Delete a NI from a peer if both prim_nid and nid are provided.
1651 * Delete a peer if only prim_nid is provided.
1652 * Delete a peer if its primary nid is provided.
1654 * The caller must hold ln_api_mutex. This prevents the peer from
1655 * being modified/deleted by a different thread.
1658 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1660 struct lnet_peer *lp;
1661 struct lnet_peer_ni *lpni;
1664 if (prim_nid == LNET_NID_ANY)
1667 lpni = lnet_find_peer_ni_locked(prim_nid);
1670 lnet_peer_ni_decref_locked(lpni);
1671 lp = lpni->lpni_peer_net->lpn_peer;
1673 if (prim_nid != lp->lp_primary_nid) {
1674 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1675 libcfs_nid2str(prim_nid),
1676 libcfs_nid2str(lp->lp_primary_nid));
1680 lnet_net_lock(LNET_LOCK_EX);
1681 if (lp->lp_rtr_refcount > 0) {
1682 lnet_net_unlock(LNET_LOCK_EX);
1683 CERROR("%s is a router. Can not be deleted\n",
1684 libcfs_nid2str(prim_nid));
1687 lnet_net_unlock(LNET_LOCK_EX);
1689 if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1690 return lnet_peer_del(lp);
1692 flags = LNET_PEER_CONFIGURED;
1693 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1694 flags |= LNET_PEER_MULTI_RAIL;
1696 return lnet_peer_del_nid(lp, nid, flags);
1700 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1702 struct lnet_peer_table *ptable;
1703 struct lnet_peer_net *lpn;
1705 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1707 LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1708 LASSERT(list_empty(&lpni->lpni_txq));
1709 LASSERT(lpni->lpni_txqnob == 0);
1710 LASSERT(list_empty(&lpni->lpni_peer_nis));
1711 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1713 lpn = lpni->lpni_peer_net;
1714 lpni->lpni_peer_net = NULL;
1715 lpni->lpni_net = NULL;
1717 /* remove the peer ni from the zombie list */
1718 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1719 spin_lock(&ptable->pt_zombie_lock);
1720 list_del_init(&lpni->lpni_hashlist);
1721 ptable->pt_zombies--;
1722 spin_unlock(&ptable->pt_zombie_lock);
1724 if (lpni->lpni_pref_nnids > 1) {
1725 LIBCFS_FREE(lpni->lpni_pref.nids,
1726 sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1728 LIBCFS_FREE(lpni, sizeof(*lpni));
1730 lnet_peer_net_decref_locked(lpn);
1733 struct lnet_peer_ni *
1734 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1736 struct lnet_peer_ni *lpni = NULL;
1739 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1740 return ERR_PTR(-ESHUTDOWN);
1743 * find if a peer_ni already exists.
1744 * If so then just return that.
1746 lpni = lnet_find_peer_ni_locked(nid);
1750 lnet_net_unlock(cpt);
1752 rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1755 goto out_net_relock;
1758 lpni = lnet_find_peer_ni_locked(nid);
1768 * Get a peer_ni for the given nid, create it if necessary. Takes a
1769 * hold on the peer_ni.
1771 struct lnet_peer_ni *
1772 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1774 struct lnet_peer_ni *lpni = NULL;
1777 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1778 return ERR_PTR(-ESHUTDOWN);
1781 * find if a peer_ni already exists.
1782 * If so then just return that.
1784 lpni = lnet_find_peer_ni_locked(nid);
1790 * use the lnet_api_mutex to serialize the creation of the peer_ni
1791 * and the creation/deletion of the local ni/net. When a local ni is
1792 * created, if there exists a set of peer_nis on that network,
1793 * they need to be traversed and updated. When a local NI is
1794 * deleted, which could result in a network being deleted, then
1795 * all peer nis on that network need to be removed as well.
1797 * Creation through traffic should also be serialized with
1798 * creation through DLC.
1800 lnet_net_unlock(cpt);
1801 mutex_lock(&the_lnet.ln_api_mutex);
1803 * Shutdown is only set under the ln_api_lock, so a single
1804 * check here is sufficent.
1806 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1807 lpni = ERR_PTR(-ESHUTDOWN);
1808 goto out_mutex_unlock;
1811 rc = lnet_peer_ni_traffic_add(nid, pref);
1814 goto out_mutex_unlock;
1817 lpni = lnet_find_peer_ni_locked(nid);
1821 mutex_unlock(&the_lnet.ln_api_mutex);
1824 /* Lock has been dropped, check again for shutdown. */
1825 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1827 lnet_peer_ni_decref_locked(lpni);
1828 lpni = ERR_PTR(-ESHUTDOWN);
1835 lnet_peer_gw_discovery(struct lnet_peer *lp)
1839 spin_lock(&lp->lp_lock);
1840 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1842 spin_unlock(&lp->lp_lock);
1848 * Is a peer uptodate from the point of view of discovery?
1850 * If it is currently being processed, obviously not.
1851 * A forced Ping or Push is also handled by the discovery thread.
1853 * Otherwise look at whether the peer needs rediscovering.
1856 lnet_peer_is_uptodate(struct lnet_peer *lp)
1860 spin_lock(&lp->lp_lock);
1861 if (lp->lp_state & (LNET_PEER_DISCOVERING |
1862 LNET_PEER_FORCE_PING |
1863 LNET_PEER_FORCE_PUSH)) {
1865 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1867 } else if (lnet_peer_needs_push(lp)) {
1869 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1870 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1877 spin_unlock(&lp->lp_lock);
1883 * Queue a peer for the attention of the discovery thread. Call with
1884 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1885 * -EALREADY if the peer was already queued.
1887 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1891 spin_lock(&lp->lp_lock);
1892 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1893 lp->lp_state |= LNET_PEER_DISCOVERING;
1894 spin_unlock(&lp->lp_lock);
1895 if (list_empty(&lp->lp_dc_list)) {
1896 lnet_peer_addref_locked(lp);
1897 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1898 wake_up(&the_lnet.ln_dc_waitq);
1904 CDEBUG(D_NET, "Queue peer %s: %d\n",
1905 libcfs_nid2str(lp->lp_primary_nid), rc);
1911 * Discovery of a peer is complete. Wake all waiters on the peer.
1912 * Call with lnet_net_lock/EX held.
1914 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1916 struct lnet_msg *msg, *tmp;
1918 struct list_head pending_msgs;
1920 INIT_LIST_HEAD(&pending_msgs);
1922 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1923 libcfs_nid2str(lp->lp_primary_nid));
1925 list_del_init(&lp->lp_dc_list);
1926 spin_lock(&lp->lp_lock);
1927 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1928 spin_unlock(&lp->lp_lock);
1929 wake_up_all(&lp->lp_dc_waitq);
1931 if (lp->lp_rtr_refcount > 0)
1932 lnet_router_discovery_complete(lp);
1934 lnet_net_unlock(LNET_LOCK_EX);
1936 /* iterate through all pending messages and send them again */
1937 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1938 list_del_init(&msg->msg_list);
1939 if (lp->lp_dc_error) {
1940 lnet_finalize(msg, lp->lp_dc_error);
1944 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1945 lnet_msgtyp2str(msg->msg_type),
1946 libcfs_id2str(msg->msg_target));
1947 rc = lnet_send(msg->msg_src_nid_param, msg,
1948 msg->msg_rtr_nid_param);
1950 CNETERR("Error sending %s to %s: %d\n",
1951 lnet_msgtyp2str(msg->msg_type),
1952 libcfs_id2str(msg->msg_target), rc);
1953 lnet_finalize(msg, rc);
1956 lnet_net_lock(LNET_LOCK_EX);
1957 lnet_peer_decref_locked(lp);
1961 * Handle inbound push.
1962 * Like any event handler, called with lnet_res_lock/CPT held.
1964 void lnet_peer_push_event(struct lnet_event *ev)
1966 struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1967 struct lnet_peer *lp;
1969 /* lnet_find_peer() adds a refcount */
1970 lp = lnet_find_peer(ev->source.nid);
1972 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1973 libcfs_nid2str(ev->initiator.nid),
1974 libcfs_nid2str(ev->source.nid));
1978 /* Ensure peer state remains consistent while we modify it. */
1979 spin_lock(&lp->lp_lock);
1982 * If some kind of error happened the contents of the message
1983 * cannot be used. Clear the NIDS_UPTODATE and set the
1984 * FORCE_PING flag to trigger a ping.
1987 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1988 lp->lp_state |= LNET_PEER_FORCE_PING;
1989 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1991 libcfs_nid2str(lp->lp_primary_nid),
1992 libcfs_nid2str(ev->source.nid));
1997 * A push with invalid or corrupted info. Clear the UPTODATE
1998 * flag to trigger a ping.
2000 if (lnet_ping_info_validate(&pbuf->pb_info)) {
2001 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2002 lp->lp_state |= LNET_PEER_FORCE_PING;
2003 CDEBUG(D_NET, "Corrupted Push from %s\n",
2004 libcfs_nid2str(lp->lp_primary_nid));
2009 * Make sure we'll allocate the correct size ping buffer when
2012 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2013 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2016 * A non-Multi-Rail peer is not supposed to be capable of
2019 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2020 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2021 libcfs_nid2str(lp->lp_primary_nid));
2026 * Check the MULTIRAIL flag. Complain if the peer was DLC
2027 * configured without it.
2029 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2030 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2031 CERROR("Push says %s is Multi-Rail, DLC says not\n",
2032 libcfs_nid2str(lp->lp_primary_nid));
2034 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2035 lnet_peer_clr_non_mr_pref_nids(lp);
2040 * The peer may have discovery disabled at its end. Set
2041 * NO_DISCOVERY as appropriate.
2043 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2044 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2045 libcfs_nid2str(lp->lp_primary_nid));
2046 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2047 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2048 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2049 libcfs_nid2str(lp->lp_primary_nid));
2050 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2054 * Check for truncation of the Put message. Clear the
2055 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2056 * and tell discovery to allocate a bigger buffer.
2058 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2059 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2060 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2061 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2062 lp->lp_state |= LNET_PEER_FORCE_PING;
2063 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2064 libcfs_nid2str(lp->lp_primary_nid),
2065 pbuf->pb_info.pi_nnis);
2069 /* always assume new data */
2070 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2071 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2074 * If there is data present that hasn't been processed yet,
2075 * we'll replace it if the Put contained newer data and it
2076 * fits. We're racing with a Ping or earlier Push in this
2079 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2080 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2081 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2082 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2083 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2084 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2085 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2086 libcfs_nid2str(lp->lp_primary_nid),
2087 LNET_PING_BUFFER_SEQNO(pbuf),
2088 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2094 * Allocate a buffer to copy the data. On a failure we drop
2095 * the Push and set FORCE_PING to force the discovery
2096 * thread to fix the problem by pinging the peer.
2098 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2100 lp->lp_state |= LNET_PEER_FORCE_PING;
2101 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2102 libcfs_nid2str(lp->lp_primary_nid),
2103 LNET_PING_BUFFER_SEQNO(pbuf));
2108 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2109 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2110 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2111 CDEBUG(D_NET, "Received Push %s %u\n",
2112 libcfs_nid2str(lp->lp_primary_nid),
2113 LNET_PING_BUFFER_SEQNO(pbuf));
2117 * Queue the peer for discovery if not done, force it on the request
2118 * queue and wake the discovery thread if the peer was already queued,
2119 * because its status changed.
2121 spin_unlock(&lp->lp_lock);
2122 lnet_net_lock(LNET_LOCK_EX);
2123 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2124 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2125 wake_up(&the_lnet.ln_dc_waitq);
2127 /* Drop refcount from lookup */
2128 lnet_peer_decref_locked(lp);
2129 lnet_net_unlock(LNET_LOCK_EX);
2133 * Clear the discovery error state, unless we're already discovering
2134 * this peer, in which case the error is current.
2136 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2138 spin_lock(&lp->lp_lock);
2139 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2140 lp->lp_dc_error = 0;
2141 spin_unlock(&lp->lp_lock);
2145 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2146 * dropped/retaken within this function. An lnet_peer_ni is passed in
2147 * because discovery could tear down an lnet_peer.
2150 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2153 struct lnet_peer *lp;
2158 lnet_net_unlock(cpt);
2159 lnet_net_lock(LNET_LOCK_EX);
2160 lp = lpni->lpni_peer_net->lpn_peer;
2161 lnet_peer_clear_discovery_error(lp);
2164 * We're willing to be interrupted. The lpni can become a
2165 * zombie if we race with DLC, so we must check for that.
2168 /* Keep lp alive when the lnet_net_lock is unlocked */
2169 lnet_peer_addref_locked(lp);
2170 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2171 if (signal_pending(current))
2173 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2176 * Don't repeat discovery if discovery is disabled. This is
2177 * done to ensure we can use discovery as a standard ping as
2178 * well for backwards compatibility with routers which do not
2179 * have discovery or have discovery disabled
2181 if (lnet_is_discovery_disabled(lp) && count > 0)
2183 if (lp->lp_dc_error)
2185 if (lnet_peer_is_uptodate(lp))
2187 lnet_peer_queue_for_discovery(lp);
2189 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2192 * If caller requested a non-blocking operation then
2193 * return immediately. Once discovery is complete any
2194 * pending messages that were stopped due to discovery
2195 * will be transmitted.
2200 lnet_net_unlock(LNET_LOCK_EX);
2202 finish_wait(&lp->lp_dc_waitq, &wait);
2203 lnet_net_lock(LNET_LOCK_EX);
2204 lnet_peer_decref_locked(lp);
2205 /* Peer may have changed */
2206 lp = lpni->lpni_peer_net->lpn_peer;
2208 finish_wait(&lp->lp_dc_waitq, &wait);
2210 lnet_net_unlock(LNET_LOCK_EX);
2212 lnet_peer_decref_locked(lp);
2214 * The peer may have changed, so re-check and rediscover if that turns
2215 * out to have been the case. The reference count on lp ensured that
2216 * even if it was unlinked from lpni the memory could not be recycled.
2217 * Thus the check below is sufficient to determine whether the peer
2218 * changed. If the peer changed, then lp must not be dereferenced.
2220 if (lp != lpni->lpni_peer_net->lpn_peer)
2223 if (signal_pending(current))
2225 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2227 else if (lp->lp_dc_error)
2228 rc = lp->lp_dc_error;
2230 CDEBUG(D_NET, "non-blocking discovery\n");
2231 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2234 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2235 (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2236 libcfs_nid2str(lpni->lpni_nid), rc,
2237 (!block) ? "pending discovery" : "discovery complete");
2242 /* Handle an incoming ack for a push. */
2244 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2246 struct lnet_ping_buffer *pbuf;
2248 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2249 spin_lock(&lp->lp_lock);
2250 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2251 lp->lp_push_error = ev->status;
2253 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2255 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2256 spin_unlock(&lp->lp_lock);
2258 CDEBUG(D_NET, "peer %s ev->status %d\n",
2259 libcfs_nid2str(lp->lp_primary_nid), ev->status);
2262 /* Handle a Reply message. This is the reply to a Ping message. */
2264 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2266 struct lnet_ping_buffer *pbuf;
2269 spin_lock(&lp->lp_lock);
2272 * If some kind of error happened the contents of message
2273 * cannot be used. Set PING_FAILED to trigger a retry.
2276 lp->lp_state |= LNET_PEER_PING_FAILED;
2277 lp->lp_ping_error = ev->status;
2278 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2280 libcfs_nid2str(lp->lp_primary_nid),
2281 libcfs_nid2str(ev->source.nid));
2285 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2286 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2287 lnet_swap_pinginfo(pbuf);
2290 * A reply with invalid or corrupted info. Set PING_FAILED to
2293 rc = lnet_ping_info_validate(&pbuf->pb_info);
2295 lp->lp_state |= LNET_PEER_PING_FAILED;
2296 lp->lp_ping_error = 0;
2297 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2298 libcfs_nid2str(lp->lp_primary_nid), rc);
2304 * Only enable the multi-rail feature on the peer if both sides of
2305 * the connection have discovery on
2307 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2308 CDEBUG(D_NET, "Peer %s has Multi-Rail feature enabled\n",
2309 libcfs_nid2str(lp->lp_primary_nid));
2310 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2312 CDEBUG(D_NET, "Peer %s has Multi-Rail feature disabled\n",
2313 libcfs_nid2str(lp->lp_primary_nid));
2314 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2318 * The peer may have discovery disabled at its end. Set
2319 * NO_DISCOVERY as appropriate.
2321 if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2322 !lnet_peer_discovery_disabled) {
2323 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2324 libcfs_nid2str(lp->lp_primary_nid));
2325 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2327 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2328 libcfs_nid2str(lp->lp_primary_nid));
2329 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2333 * Update the MULTI_RAIL flag based on the reply. If the peer
2334 * was configured with DLC then the setting should match what
2337 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2338 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2339 /* Everything's fine */
2340 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2341 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2342 libcfs_nid2str(lp->lp_primary_nid));
2345 * if discovery is disabled then we don't want to
2346 * update the state of the peer. All we'll do is
2347 * update the peer_nis which were reported back in
2351 if (!lnet_is_discovery_disabled_locked(lp)) {
2352 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2353 lnet_peer_clr_non_mr_pref_nids(lp);
2356 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2357 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2358 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2359 libcfs_nid2str(lp->lp_primary_nid));
2361 CERROR("Multi-Rail state vanished from %s\n",
2362 libcfs_nid2str(lp->lp_primary_nid));
2363 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2368 * Make sure we'll allocate the correct size ping buffer when
2371 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2372 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2375 * Check for truncation of the Reply. Clear PING_SENT and set
2376 * PING_FAILED to trigger a retry.
2378 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2379 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2380 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2381 lp->lp_state |= LNET_PEER_PING_FAILED;
2382 lp->lp_ping_error = 0;
2383 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2384 libcfs_nid2str(lp->lp_primary_nid),
2385 pbuf->pb_info.pi_nnis);
2390 * Check the sequence numbers in the reply. These are only
2391 * available if the reply came from a Multi-Rail peer.
2393 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2394 pbuf->pb_info.pi_nnis > 1 &&
2395 lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2396 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2397 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2398 libcfs_nid2str(lp->lp_primary_nid),
2399 LNET_PING_BUFFER_SEQNO(pbuf),
2402 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2405 /* We're happy with the state of the data in the buffer. */
2406 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2407 libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2408 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2409 lnet_ping_buffer_decref(lp->lp_data);
2411 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2412 lnet_ping_buffer_addref(pbuf);
2415 lp->lp_state &= ~LNET_PEER_PING_SENT;
2416 spin_unlock(&lp->lp_lock);
2420 * Send event handling. Only matters for error cases, where we clean
2421 * up state on the peer and peer_ni that would otherwise be updated in
2422 * the REPLY event handler for a successful Ping, and the ACK event
2423 * handler for a successful Push.
2426 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2433 spin_lock(&lp->lp_lock);
2434 if (ev->msg_type == LNET_MSG_GET) {
2435 lp->lp_state &= ~LNET_PEER_PING_SENT;
2436 lp->lp_state |= LNET_PEER_PING_FAILED;
2437 lp->lp_ping_error = ev->status;
2438 } else { /* ev->msg_type == LNET_MSG_PUT */
2439 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2440 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2441 lp->lp_push_error = ev->status;
2443 spin_unlock(&lp->lp_lock);
2444 rc = LNET_REDISCOVER_PEER;
2446 CDEBUG(D_NET, "%s Send to %s: %d\n",
2447 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2448 libcfs_nid2str(ev->target.nid), rc);
2453 * Unlink event handling. This event is only seen if a call to
2454 * LNetMDUnlink() caused the event to be unlinked. If this call was
2455 * made after the event was set up in LNetGet() or LNetPut() then we
2456 * assume the Ping or Push timed out.
2459 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2461 spin_lock(&lp->lp_lock);
2462 /* We've passed through LNetGet() */
2463 if (lp->lp_state & LNET_PEER_PING_SENT) {
2464 lp->lp_state &= ~LNET_PEER_PING_SENT;
2465 lp->lp_state |= LNET_PEER_PING_FAILED;
2466 lp->lp_ping_error = -ETIMEDOUT;
2467 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2468 libcfs_nid2str(lp->lp_primary_nid));
2470 /* We've passed through LNetPut() */
2471 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2472 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2473 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2474 lp->lp_push_error = -ETIMEDOUT;
2475 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2476 libcfs_nid2str(lp->lp_primary_nid));
2478 spin_unlock(&lp->lp_lock);
2482 * Event handler for the discovery EQ.
2484 * Called with lnet_res_lock(cpt) held. The cpt is the
2485 * lnet_cpt_of_cookie() of the md handle cookie.
2487 static void lnet_discovery_event_handler(struct lnet_event *event)
2489 struct lnet_peer *lp = event->md.user_ptr;
2490 struct lnet_ping_buffer *pbuf;
2493 /* discovery needs to take another look */
2494 rc = LNET_REDISCOVER_PEER;
2496 CDEBUG(D_NET, "Received event: %d\n", event->type);
2498 switch (event->type) {
2499 case LNET_EVENT_ACK:
2500 lnet_discovery_event_ack(lp, event);
2502 case LNET_EVENT_REPLY:
2503 lnet_discovery_event_reply(lp, event);
2505 case LNET_EVENT_SEND:
2506 /* Only send failure triggers a retry. */
2507 rc = lnet_discovery_event_send(lp, event);
2509 case LNET_EVENT_UNLINK:
2510 /* LNetMDUnlink() was called */
2511 lnet_discovery_event_unlink(lp, event);
2514 /* Invalid events. */
2517 lnet_net_lock(LNET_LOCK_EX);
2518 if (event->unlinked) {
2519 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2520 lnet_ping_buffer_decref(pbuf);
2521 lnet_peer_decref_locked(lp);
2524 /* put peer back at end of request queue, if discovery not already
2526 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2527 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2528 wake_up(&the_lnet.ln_dc_waitq);
2530 lnet_net_unlock(LNET_LOCK_EX);
2534 * Build a peer from incoming data.
2536 * The NIDs in the incoming data are supposed to be structured as follows:
2539 * - other NIDs in same net
2540 * - NIDs in second net
2541 * - NIDs in third net
2543 * This due to the way the list of NIDs in the data is created.
2545 * Note that this function will mark the peer uptodate unless an
2546 * ENOMEM is encontered. All other errors are due to a conflict
2547 * between the DLC configuration and what discovery sees. We treat DLC
2548 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2549 * peer from becoming stuck in discovery.
2551 static int lnet_peer_merge_data(struct lnet_peer *lp,
2552 struct lnet_ping_buffer *pbuf)
2554 struct lnet_peer_ni *lpni;
2555 lnet_nid_t *curnis = NULL;
2556 struct lnet_ni_status *addnis = NULL;
2557 lnet_nid_t *delnis = NULL;
2567 flags = LNET_PEER_DISCOVERED;
2568 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2569 flags |= LNET_PEER_MULTI_RAIL;
2572 * Cache the routing feature for the peer; whether it is enabled
2573 * for disabled as reported by the remote peer.
2575 spin_lock(&lp->lp_lock);
2576 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2577 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2579 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2580 spin_unlock(&lp->lp_lock);
2582 nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2583 LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2584 LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2585 LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2586 if (!curnis || !addnis || !delnis) {
2594 /* Construct the list of NIDs present in peer. */
2596 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2597 curnis[ncurnis++] = lpni->lpni_nid;
2600 * Check for NIDs in pbuf not present in curnis[].
2601 * The loop starts at 1 to skip the loopback NID.
2603 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2604 for (j = 0; j < ncurnis; j++)
2605 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2608 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2611 * Check for NIDs in curnis[] not present in pbuf.
2612 * The nested loop starts at 1 to skip the loopback NID.
2614 * But never add the loopback NID to delnis[]: if it is
2615 * present in curnis[] then this peer is for this node.
2617 for (i = 0; i < ncurnis; i++) {
2618 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2620 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2621 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2623 * update the information we cache for the
2624 * peer with the latest information we
2627 lpni = lnet_find_peer_ni_locked(curnis[i]);
2629 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2630 lnet_peer_ni_decref_locked(lpni);
2635 if (j == pbuf->pb_info.pi_nnis)
2636 delnis[ndelnis++] = curnis[i];
2640 * If we get here and the discovery is disabled then we don't want
2641 * to add or delete any NIs. We just updated the ones we have some
2642 * information on, and call it a day
2645 if (lnet_is_discovery_disabled(lp))
2648 for (i = 0; i < naddnis; i++) {
2649 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2651 CERROR("Error adding NID %s to peer %s: %d\n",
2652 libcfs_nid2str(addnis[i].ns_nid),
2653 libcfs_nid2str(lp->lp_primary_nid), rc);
2657 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2659 lpni->lpni_ns_status = addnis[i].ns_status;
2660 lnet_peer_ni_decref_locked(lpni);
2664 for (i = 0; i < ndelnis; i++) {
2666 * for routers it's okay to delete the primary_nid because
2667 * the upper layers don't really rely on it. So if we're
2668 * being told that the router changed its primary_nid
2669 * then it's okay to delete it.
2671 if (lp->lp_rtr_refcount > 0)
2672 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2673 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2675 CERROR("Error deleting NID %s from peer %s: %d\n",
2676 libcfs_nid2str(delnis[i]),
2677 libcfs_nid2str(lp->lp_primary_nid), rc);
2683 * Errors other than -ENOMEM are due to peers having been
2684 * configured with DLC. Ignore these because DLC overrides
2689 LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2690 LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2691 LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2692 lnet_ping_buffer_decref(pbuf);
2693 CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2696 spin_lock(&lp->lp_lock);
2697 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2698 lp->lp_state |= LNET_PEER_FORCE_PING;
2699 spin_unlock(&lp->lp_lock);
2705 * The data in pbuf says lp is its primary peer, but the data was
2706 * received by a different peer. Try to update lp with the data.
2709 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2711 struct lnet_handle_md mdh;
2713 /* Queue lp for discovery, and force it on the request queue. */
2714 lnet_net_lock(LNET_LOCK_EX);
2715 if (lnet_peer_queue_for_discovery(lp))
2716 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2717 lnet_net_unlock(LNET_LOCK_EX);
2719 LNetInvalidateMDHandle(&mdh);
2722 * Decide whether we can move the peer to the DATA_PRESENT state.
2724 * We replace stale data for a multi-rail peer, repair PING_FAILED
2725 * status, and preempt FORCE_PING.
2727 * If after that we have DATA_PRESENT, we merge it into this peer.
2729 spin_lock(&lp->lp_lock);
2730 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2731 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2732 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2733 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2734 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2735 lnet_ping_buffer_decref(pbuf);
2740 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2741 lnet_ping_buffer_decref(lp->lp_data);
2743 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2745 if (lp->lp_state & LNET_PEER_PING_FAILED) {
2746 mdh = lp->lp_ping_mdh;
2747 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2748 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2749 lp->lp_ping_error = 0;
2751 if (lp->lp_state & LNET_PEER_FORCE_PING)
2752 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2753 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2754 spin_unlock(&lp->lp_lock);
2756 if (!LNetMDHandleIsInvalid(mdh))
2760 return lnet_peer_merge_data(lp, pbuf);
2762 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2766 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2770 for (i = 0; i < pinfo->pi_nnis; i++) {
2771 if (pinfo->pi_ni[i].ns_nid == nid)
2779 * Update a peer using the data received.
2781 static int lnet_peer_data_present(struct lnet_peer *lp)
2782 __must_hold(&lp->lp_lock)
2784 struct lnet_ping_buffer *pbuf;
2785 struct lnet_peer_ni *lpni;
2786 lnet_nid_t nid = LNET_NID_ANY;
2792 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2793 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2794 spin_unlock(&lp->lp_lock);
2797 * Modifications of peer structures are done while holding the
2798 * ln_api_mutex. A global lock is required because we may be
2799 * modifying multiple peer structures, and a mutex greatly
2800 * simplifies memory management.
2802 * The actual changes to the data structures must also protect
2803 * against concurrent lookups, for which the lnet_net_lock in
2804 * LNET_LOCK_EX mode is used.
2806 mutex_lock(&the_lnet.ln_api_mutex);
2807 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2813 * If this peer is not on the peer list then it is being torn
2814 * down, and our reference count may be all that is keeping it
2815 * alive. Don't do any work on it.
2817 if (list_empty(&lp->lp_peer_list))
2820 flags = LNET_PEER_DISCOVERED;
2821 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2822 flags |= LNET_PEER_MULTI_RAIL;
2825 * Check whether the primary NID in the message matches the
2826 * primary NID of the peer. If it does, update the peer, if
2827 * it it does not, check whether there is already a peer with
2828 * that primary NID. If no such peer exists, try to update
2829 * the primary NID of the current peer (allowed if it was
2830 * created due to message traffic) and complete the update.
2831 * If the peer did exist, hand off the data to it.
2833 * The peer for the loopback interface is a special case: this
2834 * is the peer for the local node, and we want to set its
2835 * primary NID to the correct value here. Moreover, this peer
2836 * can show up with only the loopback NID in the ping buffer.
2838 if (pbuf->pb_info.pi_nnis <= 1)
2840 nid = pbuf->pb_info.pi_ni[1].ns_nid;
2841 if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2842 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2844 rc = lnet_peer_merge_data(lp, pbuf);
2846 * if the primary nid of the peer is present in the ping info returned
2847 * from the peer, but it's not the local primary peer we have
2848 * cached and discovery is disabled, then we don't want to update
2849 * our local peer info, by adding or removing NIDs, we just want
2850 * to update the status of the nids that we currently have
2851 * recorded in that peer.
2853 } else if (lp->lp_primary_nid == nid ||
2854 (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2855 lnet_is_discovery_disabled(lp))) {
2856 rc = lnet_peer_merge_data(lp, pbuf);
2858 lpni = lnet_find_peer_ni_locked(nid);
2860 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2862 CERROR("Primary NID error %s versus %s: %d\n",
2863 libcfs_nid2str(lp->lp_primary_nid),
2864 libcfs_nid2str(nid), rc);
2866 rc = lnet_peer_merge_data(lp, pbuf);
2869 struct lnet_peer *new_lp;
2870 new_lp = lpni->lpni_peer_net->lpn_peer;
2872 * if lp has discovery/MR enabled that means new_lp
2873 * should have discovery/MR enabled as well, since
2874 * it's the same peer, which we're about to merge
2876 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2877 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2878 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2879 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2881 rc = lnet_peer_set_primary_data(new_lp, pbuf);
2882 lnet_consolidate_routes_locked(lp, new_lp);
2883 lnet_peer_ni_decref_locked(lpni);
2887 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2889 mutex_unlock(&the_lnet.ln_api_mutex);
2891 spin_lock(&lp->lp_lock);
2892 /* Tell discovery to re-check the peer immediately. */
2894 rc = LNET_REDISCOVER_PEER;
2899 * A ping failed. Clear the PING_FAILED state and set the
2900 * FORCE_PING state, to ensure a retry even if discovery is
2901 * disabled. This avoids being left with incorrect state.
2903 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2904 __must_hold(&lp->lp_lock)
2906 struct lnet_handle_md mdh;
2909 mdh = lp->lp_ping_mdh;
2910 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2911 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2912 lp->lp_state |= LNET_PEER_FORCE_PING;
2913 rc = lp->lp_ping_error;
2914 lp->lp_ping_error = 0;
2915 spin_unlock(&lp->lp_lock);
2917 if (!LNetMDHandleIsInvalid(mdh))
2920 CDEBUG(D_NET, "peer %s:%d\n",
2921 libcfs_nid2str(lp->lp_primary_nid), rc);
2923 spin_lock(&lp->lp_lock);
2924 return rc ? rc : LNET_REDISCOVER_PEER;
2928 * Select NID to send a Ping or Push to.
2930 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2932 struct lnet_peer_ni *lpni;
2934 /* Look for a direct-connected NID for this peer. */
2936 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2937 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2942 return lpni->lpni_nid;
2944 /* Look for a routed-connected NID for this peer. */
2946 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2947 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2952 return lpni->lpni_nid;
2954 return LNET_NID_ANY;
2957 /* Active side of ping. */
2958 static int lnet_peer_send_ping(struct lnet_peer *lp)
2959 __must_hold(&lp->lp_lock)
2966 lp->lp_state |= LNET_PEER_PING_SENT;
2967 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2968 spin_unlock(&lp->lp_lock);
2970 cpt = lnet_net_lock_current();
2971 /* Refcount for MD. */
2972 lnet_peer_addref_locked(lp);
2973 pnid = lnet_peer_select_nid(lp);
2974 lnet_net_unlock(cpt);
2976 nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2978 rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2979 the_lnet.ln_dc_eqh, false);
2982 * if LNetMDBind in lnet_send_ping fails we need to decrement the
2983 * refcount on the peer, otherwise LNetMDUnlink will be called
2984 * which will eventually do that.
2988 lnet_peer_decref_locked(lp);
2989 lnet_net_unlock(cpt);
2990 rc = -rc; /* change the rc to negative value */
2992 } else if (rc < 0) {
2996 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2998 spin_lock(&lp->lp_lock);
3002 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3004 * The errors that get us here are considered hard errors and
3005 * cause Discovery to terminate. So we clear PING_SENT, but do
3006 * not set either PING_FAILED or FORCE_PING. In fact we need
3007 * to clear PING_FAILED, because the unlink event handler will
3008 * have set it if we called LNetMDUnlink() above.
3010 spin_lock(&lp->lp_lock);
3011 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3016 * This function exists because you cannot call LNetMDUnlink() from an
3019 static int lnet_peer_push_failed(struct lnet_peer *lp)
3020 __must_hold(&lp->lp_lock)
3022 struct lnet_handle_md mdh;
3025 mdh = lp->lp_push_mdh;
3026 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3027 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3028 rc = lp->lp_push_error;
3029 lp->lp_push_error = 0;
3030 spin_unlock(&lp->lp_lock);
3032 if (!LNetMDHandleIsInvalid(mdh))
3035 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3036 spin_lock(&lp->lp_lock);
3037 return rc ? rc : LNET_REDISCOVER_PEER;
3040 /* Active side of push. */
3041 static int lnet_peer_send_push(struct lnet_peer *lp)
3042 __must_hold(&lp->lp_lock)
3044 struct lnet_ping_buffer *pbuf;
3045 struct lnet_process_id id;
3050 /* Don't push to a non-multi-rail peer. */
3051 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3052 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3056 lp->lp_state |= LNET_PEER_PUSH_SENT;
3057 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3058 spin_unlock(&lp->lp_lock);
3060 cpt = lnet_net_lock_current();
3061 pbuf = the_lnet.ln_ping_target;
3062 lnet_ping_buffer_addref(pbuf);
3063 lnet_net_unlock(cpt);
3065 /* Push source MD */
3066 md.start = &pbuf->pb_info;
3067 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3068 md.threshold = 2; /* Put/Ack */
3071 md.eq_handle = the_lnet.ln_dc_eqh;
3074 rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3076 lnet_ping_buffer_decref(pbuf);
3077 CERROR("Can't bind push source MD: %d\n", rc);
3080 cpt = lnet_net_lock_current();
3081 /* Refcount for MD. */
3082 lnet_peer_addref_locked(lp);
3083 id.pid = LNET_PID_LUSTRE;
3084 id.nid = lnet_peer_select_nid(lp);
3085 lnet_net_unlock(cpt);
3087 if (id.nid == LNET_NID_ANY) {
3092 rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3093 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3094 LNET_PROTO_PING_MATCHBITS, 0, 0);
3099 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3101 spin_lock(&lp->lp_lock);
3105 LNetMDUnlink(lp->lp_push_mdh);
3106 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3108 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3110 * The errors that get us here are considered hard errors and
3111 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3112 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3113 * because the unlink event handler will have set it if we
3114 * called LNetMDUnlink() above.
3116 spin_lock(&lp->lp_lock);
3117 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3122 * An unrecoverable error was encountered during discovery.
3123 * Set error status in peer and abort discovery.
3125 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3127 CDEBUG(D_NET, "Discovery error %s: %d\n",
3128 libcfs_nid2str(lp->lp_primary_nid), error);
3130 spin_lock(&lp->lp_lock);
3131 lp->lp_dc_error = error;
3132 lp->lp_state &= ~LNET_PEER_DISCOVERING;
3133 lp->lp_state |= LNET_PEER_REDISCOVER;
3134 spin_unlock(&lp->lp_lock);
3138 * Mark the peer as discovered.
3140 static int lnet_peer_discovered(struct lnet_peer *lp)
3141 __must_hold(&lp->lp_lock)
3143 lp->lp_state |= LNET_PEER_DISCOVERED;
3144 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3145 LNET_PEER_REDISCOVER);
3147 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3154 * Discovering this peer is taking too long. Cancel any Ping or Push
3155 * that discovery is waiting on by unlinking the relevant MDs. The
3156 * lnet_discovery_event_handler() will proceed from here and complete
3159 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3161 struct lnet_handle_md ping_mdh;
3162 struct lnet_handle_md push_mdh;
3164 LNetInvalidateMDHandle(&ping_mdh);
3165 LNetInvalidateMDHandle(&push_mdh);
3167 spin_lock(&lp->lp_lock);
3168 if (lp->lp_state & LNET_PEER_PING_SENT) {
3169 ping_mdh = lp->lp_ping_mdh;
3170 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3172 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3173 push_mdh = lp->lp_push_mdh;
3174 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3176 spin_unlock(&lp->lp_lock);
3178 if (!LNetMDHandleIsInvalid(ping_mdh))
3179 LNetMDUnlink(ping_mdh);
3180 if (!LNetMDHandleIsInvalid(push_mdh))
3181 LNetMDUnlink(push_mdh);
3185 * Wait for work to be queued or some other change that must be
3186 * attended to. Returns non-zero if the discovery thread should shut
3189 static int lnet_peer_discovery_wait_for_work(void)
3196 cpt = lnet_net_lock_current();
3198 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3199 TASK_INTERRUPTIBLE);
3200 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3202 if (lnet_push_target_resize_needed())
3204 if (!list_empty(&the_lnet.ln_dc_request))
3206 if (!list_empty(&the_lnet.ln_msg_resend))
3208 lnet_net_unlock(cpt);
3211 * wakeup max every second to check if there are peers that
3212 * have been stuck on the working queue for greater than
3215 schedule_timeout(cfs_time_seconds(1));
3216 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3217 cpt = lnet_net_lock_current();
3219 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3221 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3224 lnet_net_unlock(cpt);
3226 CDEBUG(D_NET, "woken: %d\n", rc);
3232 * Messages that were pending on a destroyed peer will be put on a global
3233 * resend list. The message resend list will be checked by
3234 * the discovery thread when it wakes up, and will resend messages. These
3235 * messages can still be sendable in the case the lpni which was the initial
3236 * cause of the message re-queue was transfered to another peer.
3238 * It is possible that LNet could be shutdown while we're iterating
3239 * through the list. lnet_shudown_lndnets() will attempt to access the
3240 * resend list, but will have to wait until the spinlock is released, by
3241 * which time there shouldn't be any more messages on the resend list.
3242 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3243 * for the messages so they can be released. The other case is that
3244 * lnet_shudown_lndnets() can finalize all the messages before this
3245 * function can visit the resend list, in which case this function will be
3248 static void lnet_resend_msgs(void)
3250 struct lnet_msg *msg, *tmp;
3251 struct list_head resend;
3254 INIT_LIST_HEAD(&resend);
3256 spin_lock(&the_lnet.ln_msg_resend_lock);
3257 list_splice(&the_lnet.ln_msg_resend, &resend);
3258 spin_unlock(&the_lnet.ln_msg_resend_lock);
3260 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3261 list_del_init(&msg->msg_list);
3262 rc = lnet_send(msg->msg_src_nid_param, msg,
3263 msg->msg_rtr_nid_param);
3265 CNETERR("Error sending %s to %s: %d\n",
3266 lnet_msgtyp2str(msg->msg_type),
3267 libcfs_id2str(msg->msg_target), rc);
3268 lnet_finalize(msg, rc);
3273 /* The discovery thread. */
3274 static int lnet_peer_discovery(void *arg)
3276 struct lnet_peer *lp;
3279 wait_for_completion(&the_lnet.ln_started);
3281 CDEBUG(D_NET, "started\n");
3282 cfs_block_allsigs();
3285 if (lnet_peer_discovery_wait_for_work())
3290 if (lnet_push_target_resize_needed())
3291 lnet_push_target_resize();
3293 lnet_net_lock(LNET_LOCK_EX);
3294 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3298 * Process all incoming discovery work requests. When
3299 * discovery must wait on a peer to change state, it
3300 * is added to the tail of the ln_dc_working queue. A
3301 * timestamp keeps track of when the peer was added,
3302 * so we can time out discovery requests that take too
3305 while (!list_empty(&the_lnet.ln_dc_request)) {
3306 lp = list_first_entry(&the_lnet.ln_dc_request,
3307 struct lnet_peer, lp_dc_list);
3308 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3310 * set the time the peer was put on the dc_working
3311 * queue. It shouldn't remain on the queue
3312 * forever, in case the GET message (for ping)
3313 * doesn't get a REPLY or the PUT message (for
3314 * push) doesn't get an ACK.
3316 lp->lp_last_queued = ktime_get_real_seconds();
3317 lnet_net_unlock(LNET_LOCK_EX);
3320 * Select an action depending on the state of
3321 * the peer and whether discovery is disabled.
3322 * The check whether discovery is disabled is
3323 * done after the code that handles processing
3324 * for arrived data, cleanup for failures, and
3325 * forcing a Ping or Push.
3327 spin_lock(&lp->lp_lock);
3328 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3329 libcfs_nid2str(lp->lp_primary_nid), lp,
3331 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3332 rc = lnet_peer_data_present(lp);
3333 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3334 rc = lnet_peer_ping_failed(lp);
3335 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3336 rc = lnet_peer_push_failed(lp);
3337 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3338 rc = lnet_peer_send_ping(lp);
3339 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3340 rc = lnet_peer_send_push(lp);
3341 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3342 rc = lnet_peer_send_ping(lp);
3343 else if (lnet_peer_needs_push(lp))
3344 rc = lnet_peer_send_push(lp);
3346 rc = lnet_peer_discovered(lp);
3347 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3348 libcfs_nid2str(lp->lp_primary_nid), lp,
3350 spin_unlock(&lp->lp_lock);
3352 lnet_net_lock(LNET_LOCK_EX);
3353 if (rc == LNET_REDISCOVER_PEER) {
3354 list_move(&lp->lp_dc_list,
3355 &the_lnet.ln_dc_request);
3357 lnet_peer_discovery_error(lp, rc);
3359 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3360 lnet_peer_discovery_complete(lp);
3361 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3365 lnet_net_unlock(LNET_LOCK_EX);
3368 CDEBUG(D_NET, "stopping\n");
3370 * Clean up before telling lnet_peer_discovery_stop() that
3371 * we're done. Use wake_up() below to somewhat reduce the
3372 * size of the thundering herd if there are multiple threads
3373 * waiting on discovery of a single peer.
3376 /* Queue cleanup 1: stop all pending pings and pushes. */
3377 lnet_net_lock(LNET_LOCK_EX);
3378 while (!list_empty(&the_lnet.ln_dc_working)) {
3379 lp = list_first_entry(&the_lnet.ln_dc_working,
3380 struct lnet_peer, lp_dc_list);
3381 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3382 lnet_net_unlock(LNET_LOCK_EX);
3383 lnet_peer_cancel_discovery(lp);
3384 lnet_net_lock(LNET_LOCK_EX);
3386 lnet_net_unlock(LNET_LOCK_EX);
3388 /* Queue cleanup 2: wait for the expired queue to clear. */
3389 while (!list_empty(&the_lnet.ln_dc_expired))
3390 schedule_timeout(cfs_time_seconds(1));
3392 /* Queue cleanup 3: clear the request queue. */
3393 lnet_net_lock(LNET_LOCK_EX);
3394 while (!list_empty(&the_lnet.ln_dc_request)) {
3395 lp = list_first_entry(&the_lnet.ln_dc_request,
3396 struct lnet_peer, lp_dc_list);
3397 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3398 lnet_peer_discovery_complete(lp);
3400 lnet_net_unlock(LNET_LOCK_EX);
3402 LNetEQFree(the_lnet.ln_dc_eqh);
3403 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3405 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3406 wake_up(&the_lnet.ln_dc_waitq);
3408 CDEBUG(D_NET, "stopped\n");
3413 /* ln_api_mutex is held on entry. */
3414 int lnet_peer_discovery_start(void)
3416 struct task_struct *task;
3419 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3422 rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3424 CERROR("Can't allocate discovery EQ: %d\n", rc);
3428 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3429 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3432 CERROR("Can't start peer discovery thread: %d\n", rc);
3434 LNetEQFree(the_lnet.ln_dc_eqh);
3435 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3437 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3440 CDEBUG(D_NET, "discovery start: %d\n", rc);
3445 /* ln_api_mutex is held on entry. */
3446 void lnet_peer_discovery_stop(void)
3448 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3451 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3452 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3454 /* In the LNetNIInit() path we may be stopping discovery before it
3455 * entered its work loop
3457 if (!completion_done(&the_lnet.ln_started))
3458 complete(&the_lnet.ln_started);
3460 wake_up(&the_lnet.ln_dc_waitq);
3462 wait_event(the_lnet.ln_dc_waitq,
3463 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3465 LASSERT(list_empty(&the_lnet.ln_dc_request));
3466 LASSERT(list_empty(&the_lnet.ln_dc_working));
3467 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3469 CDEBUG(D_NET, "discovery stopped\n");
3475 lnet_debug_peer(lnet_nid_t nid)
3477 char *aliveness = "NA";
3478 struct lnet_peer_ni *lp;
3481 cpt = lnet_cpt_of_nid(nid, NULL);
3484 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3486 lnet_net_unlock(cpt);
3487 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3491 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3492 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3494 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3495 libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3496 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3497 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3498 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3500 lnet_peer_ni_decref_locked(lp);
3502 lnet_net_unlock(cpt);
3505 /* Gathering information for userspace. */
3507 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3508 char aliveness[LNET_MAX_STR_LEN],
3509 __u32 *cpt_iter, __u32 *refcount,
3510 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3511 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3512 __u32 *peer_tx_qnob)
3514 struct lnet_peer_table *peer_table;
3515 struct lnet_peer_ni *lp;
3520 /* get the number of CPTs */
3521 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3523 /* if the cpt number to be examined is >= the number of cpts in
3524 * the system then indicate that there are no more cpts to examin
3526 if (*cpt_iter >= lncpt)
3529 /* get the current table */
3530 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3531 /* if the ptable is NULL then there are no more cpts to examine */
3532 if (peer_table == NULL)
3535 lnet_net_lock(*cpt_iter);
3537 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3538 struct list_head *peers = &peer_table->pt_hash[j];
3540 list_for_each_entry(lp, peers, lpni_hashlist) {
3541 if (peer_index-- > 0)
3544 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3545 if (lnet_isrouter(lp) ||
3546 lnet_peer_aliveness_enabled(lp))
3547 snprintf(aliveness, LNET_MAX_STR_LEN,
3548 lnet_is_peer_ni_alive(lp) ? "up" : "down");
3550 *nid = lp->lpni_nid;
3551 *refcount = atomic_read(&lp->lpni_refcount);
3552 *ni_peer_tx_credits =
3553 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3554 *peer_tx_credits = lp->lpni_txcredits;
3555 *peer_rtr_credits = lp->lpni_rtrcredits;
3556 *peer_min_rtr_credits = lp->lpni_mintxcredits;
3557 *peer_tx_qnob = lp->lpni_txqnob;
3563 lnet_net_unlock(*cpt_iter);
3567 return found ? 0 : -ENOENT;
3570 /* ln_api_mutex is held, which keeps the peer list stable */
3571 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3573 struct lnet_ioctl_element_stats *lpni_stats;
3574 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3575 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3576 struct lnet_peer_ni_credit_info *lpni_info;
3577 struct lnet_peer_ni *lpni;
3578 struct lnet_peer *lp;
3583 lp = lnet_find_peer(cfg->prcfg_prim_nid);
3590 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3591 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3592 size *= lp->lp_nnis;
3593 if (size > cfg->prcfg_size) {
3594 cfg->prcfg_size = size;
3599 cfg->prcfg_prim_nid = lp->lp_primary_nid;
3600 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3601 cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3602 cfg->prcfg_count = lp->lp_nnis;
3603 cfg->prcfg_size = size;
3604 cfg->prcfg_state = lp->lp_state;
3606 /* Allocate helper buffers. */
3608 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3611 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3614 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3615 if (!lpni_msg_stats)
3616 goto out_free_stats;
3617 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3619 goto out_free_msg_stats;
3624 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3625 nid = lpni->lpni_nid;
3626 if (copy_to_user(bulk, &nid, sizeof(nid)))
3627 goto out_free_hstats;
3628 bulk += sizeof(nid);
3630 memset(lpni_info, 0, sizeof(*lpni_info));
3631 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3632 if (lnet_isrouter(lpni) ||
3633 lnet_peer_aliveness_enabled(lpni))
3634 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3635 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3637 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3638 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3639 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3640 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3641 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3642 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3643 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3644 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3645 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3646 goto out_free_hstats;
3647 bulk += sizeof(*lpni_info);
3649 memset(lpni_stats, 0, sizeof(*lpni_stats));
3650 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3651 LNET_STATS_TYPE_SEND);
3652 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3653 LNET_STATS_TYPE_RECV);
3654 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3655 LNET_STATS_TYPE_DROP);
3656 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3657 goto out_free_hstats;
3658 bulk += sizeof(*lpni_stats);
3659 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3660 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3661 goto out_free_hstats;
3662 bulk += sizeof(*lpni_msg_stats);
3663 lpni_hstats->hlpni_network_timeout =
3664 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3665 lpni_hstats->hlpni_remote_dropped =
3666 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3667 lpni_hstats->hlpni_remote_timeout =
3668 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3669 lpni_hstats->hlpni_remote_error =
3670 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3671 lpni_hstats->hlpni_health_value =
3672 atomic_read(&lpni->lpni_healthv);
3673 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3674 goto out_free_hstats;
3675 bulk += sizeof(*lpni_hstats);
3680 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3682 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3684 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3686 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3688 lnet_peer_decref_locked(lp);
3694 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3696 /* the mt could've shutdown and cleaned up the queues */
3697 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3700 if (list_empty(&lpni->lpni_recovery) &&
3701 atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3702 CERROR("lpni %s added to recovery queue. Health = %d\n",
3703 libcfs_nid2str(lpni->lpni_nid),
3704 atomic_read(&lpni->lpni_healthv));
3705 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3706 lnet_peer_ni_addref_locked(lpni);
3710 /* Call with the ln_api_mutex held */
3712 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3714 struct lnet_peer_table *ptable;
3715 struct lnet_peer *lp;
3716 struct lnet_peer_net *lpn;
3717 struct lnet_peer_ni *lpni;
3721 if (the_lnet.ln_state != LNET_STATE_RUNNING)
3725 lnet_net_lock(LNET_LOCK_EX);
3726 lpni = lnet_find_peer_ni_locked(nid);
3728 lnet_net_unlock(LNET_LOCK_EX);
3731 atomic_set(&lpni->lpni_healthv, value);
3732 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3733 lnet_peer_ni_decref_locked(lpni);
3734 lnet_net_unlock(LNET_LOCK_EX);
3738 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3741 * Walk all the peers and reset the healhv for each one to the
3744 lnet_net_lock(LNET_LOCK_EX);
3745 for (cpt = 0; cpt < lncpt; cpt++) {
3746 ptable = the_lnet.ln_peer_tables[cpt];
3747 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3748 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3749 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3751 atomic_set(&lpni->lpni_healthv, value);
3752 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3757 lnet_net_unlock(LNET_LOCK_EX);