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 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
110 cfs_percpt_free(the_lnet.ln_peer_tables);
111 the_lnet.ln_peer_tables = NULL;
115 lnet_peer_tables_create(void)
117 struct lnet_peer_table *ptable;
118 struct list_head *hash;
122 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
124 if (the_lnet.ln_peer_tables == NULL) {
125 CERROR("Failed to allocate cpu-partition peer tables\n");
129 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131 LNET_PEER_HASH_SIZE * sizeof(*hash));
133 CERROR("Failed to create peer hash table\n");
134 lnet_peer_tables_destroy();
138 spin_lock_init(&ptable->pt_zombie_lock);
139 INIT_LIST_HEAD(&ptable->pt_zombie_list);
141 INIT_LIST_HEAD(&ptable->pt_peer_list);
143 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144 INIT_LIST_HEAD(&hash[j]);
145 ptable->pt_hash = hash; /* sign of initialization */
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(lnet_nid_t nid)
154 struct lnet_peer_ni *lpni;
155 struct lnet_net *net;
158 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
160 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
164 INIT_LIST_HEAD(&lpni->lpni_txq);
165 INIT_LIST_HEAD(&lpni->lpni_hashlist);
166 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167 INIT_LIST_HEAD(&lpni->lpni_recovery);
168 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171 spin_lock_init(&lpni->lpni_lock);
173 if (lnet_peers_start_down())
174 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
176 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
177 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
178 lpni->lpni_nid = nid;
179 lpni->lpni_cpt = cpt;
180 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
182 net = lnet_get_net_locked(LNET_NIDNET(nid));
183 lpni->lpni_net = net;
185 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
186 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
187 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
188 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
191 * This peer_ni is not on a local network, so we
192 * cannot add the credits here. In case the net is
193 * added later, add the peer_ni to the remote peer ni
194 * list so it can be easily found and revisited.
196 /* FIXME: per-net implementation instead? */
197 atomic_inc(&lpni->lpni_refcount);
198 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
199 &the_lnet.ln_remote_peer_ni_list);
202 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
207 static struct lnet_peer_net *
208 lnet_peer_net_alloc(__u32 net_id)
210 struct lnet_peer_net *lpn;
212 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
216 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
217 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
218 lpn->lpn_net_id = net_id;
220 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
226 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
228 struct lnet_peer *lp;
230 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
232 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
233 LASSERT(list_empty(&lpn->lpn_peer_nis));
234 LASSERT(list_empty(&lpn->lpn_peer_nets));
236 lpn->lpn_peer = NULL;
237 LIBCFS_FREE(lpn, sizeof(*lpn));
239 lnet_peer_decref_locked(lp);
242 static struct lnet_peer *
243 lnet_peer_alloc(lnet_nid_t nid)
245 struct lnet_peer *lp;
247 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
251 INIT_LIST_HEAD(&lp->lp_rtrq);
252 INIT_LIST_HEAD(&lp->lp_routes);
253 INIT_LIST_HEAD(&lp->lp_peer_list);
254 INIT_LIST_HEAD(&lp->lp_peer_nets);
255 INIT_LIST_HEAD(&lp->lp_dc_list);
256 INIT_LIST_HEAD(&lp->lp_dc_pendq);
257 INIT_LIST_HEAD(&lp->lp_rtr_list);
258 init_waitqueue_head(&lp->lp_dc_waitq);
259 spin_lock_init(&lp->lp_lock);
260 lp->lp_primary_nid = nid;
261 if (lnet_peers_start_down())
262 lp->lp_alive = false;
267 * all peers created on a router should have health on
268 * if it's not already on.
270 if (the_lnet.ln_routing && !lnet_health_sensitivity)
271 lp->lp_health_sensitivity = 1;
274 * Turn off discovery for loopback peer. If you're creating a peer
275 * for the loopback interface then that was initiated when we
276 * attempted to send a message over the loopback. There is no need
277 * to ever use a different interface when sending messages to
280 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
281 lp->lp_state = LNET_PEER_NO_DISCOVERY;
282 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
284 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
290 lnet_destroy_peer_locked(struct lnet_peer *lp)
292 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
294 LASSERT(atomic_read(&lp->lp_refcount) == 0);
295 LASSERT(lp->lp_rtr_refcount == 0);
296 LASSERT(list_empty(&lp->lp_peer_nets));
297 LASSERT(list_empty(&lp->lp_peer_list));
298 LASSERT(list_empty(&lp->lp_dc_list));
301 lnet_ping_buffer_decref(lp->lp_data);
304 * if there are messages still on the pending queue, then make
305 * sure to queue them on the ln_msg_resend list so they can be
306 * resent at a later point if the discovery thread is still
308 * If the discovery thread has stopped, then the wakeup will be a
309 * no-op, and it is expected the lnet_shutdown_lndnets() will
310 * eventually be called, which will traverse this list and
311 * finalize the messages on the list.
312 * We can not resend them now because we're holding the cpt lock.
313 * Releasing the lock can cause an inconsistent state
315 spin_lock(&the_lnet.ln_msg_resend_lock);
316 spin_lock(&lp->lp_lock);
317 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
318 spin_unlock(&lp->lp_lock);
319 spin_unlock(&the_lnet.ln_msg_resend_lock);
320 wake_up(&the_lnet.ln_dc_waitq);
322 LIBCFS_FREE(lp, sizeof(*lp));
326 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
327 * that peer_net, detach the peer_net from the peer.
329 * Call with lnet_net_lock/EX held
332 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
334 struct lnet_peer_table *ptable;
335 struct lnet_peer_net *lpn;
336 struct lnet_peer *lp;
339 * Belts and suspenders: gracefully handle teardown of a
340 * partially connected peer_ni.
342 lpn = lpni->lpni_peer_net;
344 list_del_init(&lpni->lpni_peer_nis);
346 * If there are no lpni's left, we detach lpn from
347 * lp_peer_nets, so it cannot be found anymore.
349 if (list_empty(&lpn->lpn_peer_nis))
350 list_del_init(&lpn->lpn_peer_nets);
352 /* Update peer NID count. */
357 * If there are no more peer nets, make the peer unfindable
358 * via the peer_tables.
360 * Otherwise, if the peer is DISCOVERED, tell discovery to
361 * take another look at it. This is a no-op if discovery for
362 * this peer did the detaching.
364 if (list_empty(&lp->lp_peer_nets)) {
365 list_del_init(&lp->lp_peer_list);
366 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
368 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
369 /* Discovery isn't running, nothing to do here. */
370 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
371 lnet_peer_queue_for_discovery(lp);
372 wake_up(&the_lnet.ln_dc_waitq);
374 CDEBUG(D_NET, "peer %s NID %s\n",
375 libcfs_nid2str(lp->lp_primary_nid),
376 libcfs_nid2str(lpni->lpni_nid));
379 /* called with lnet_net_lock LNET_LOCK_EX held */
381 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
383 struct lnet_peer_table *ptable = NULL;
385 /* don't remove a peer_ni if it's also a gateway */
386 if (lnet_isrouter(lpni) && !force) {
387 CERROR("Peer NI %s is a gateway. Can not delete it\n",
388 libcfs_nid2str(lpni->lpni_nid));
392 lnet_peer_remove_from_remote_list(lpni);
394 /* remove peer ni from the hash list. */
395 list_del_init(&lpni->lpni_hashlist);
398 * indicate the peer is being deleted so the monitor thread can
399 * remove it from the recovery queue.
401 spin_lock(&lpni->lpni_lock);
402 lpni->lpni_state |= LNET_PEER_NI_DELETING;
403 spin_unlock(&lpni->lpni_lock);
405 /* decrement the ref count on the peer table */
406 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
409 * The peer_ni can no longer be found with a lookup. But there
410 * can be current users, so keep track of it on the zombie
411 * list until the reference count has gone to zero.
413 * The last reference may be lost in a place where the
414 * lnet_net_lock locks only a single cpt, and that cpt may not
415 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
418 spin_lock(&ptable->pt_zombie_lock);
419 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
420 ptable->pt_zombies++;
421 spin_unlock(&ptable->pt_zombie_lock);
423 /* no need to keep this peer_ni on the hierarchy anymore */
424 lnet_peer_detach_peer_ni_locked(lpni);
426 /* remove hashlist reference on peer_ni */
427 lnet_peer_ni_decref_locked(lpni);
432 void lnet_peer_uninit(void)
434 struct lnet_peer_ni *lpni, *tmp;
436 lnet_net_lock(LNET_LOCK_EX);
438 /* remove all peer_nis from the remote peer and the hash list */
439 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
440 lpni_on_remote_peer_ni_list)
441 lnet_peer_ni_del_locked(lpni, false);
443 lnet_peer_tables_destroy();
445 lnet_net_unlock(LNET_LOCK_EX);
449 lnet_peer_del_locked(struct lnet_peer *peer)
451 struct lnet_peer_ni *lpni = NULL, *lpni2;
454 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
456 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
457 while (lpni != NULL) {
458 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
459 rc = lnet_peer_ni_del_locked(lpni, false);
469 lnet_peer_del(struct lnet_peer *peer)
471 lnet_net_lock(LNET_LOCK_EX);
472 lnet_peer_del_locked(peer);
473 lnet_net_unlock(LNET_LOCK_EX);
479 * Delete a NID from a peer. Call with ln_api_mutex held.
482 * -EPERM: Non-DLC deletion from DLC-configured peer.
483 * -ENOENT: No lnet_peer_ni corresponding to the nid.
484 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
485 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
488 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
490 struct lnet_peer_ni *lpni;
491 lnet_nid_t primary_nid = lp->lp_primary_nid;
493 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
495 if (!(flags & LNET_PEER_CONFIGURED)) {
496 if (lp->lp_state & LNET_PEER_CONFIGURED) {
501 lpni = lnet_find_peer_ni_locked(nid);
506 lnet_peer_ni_decref_locked(lpni);
507 if (lp != lpni->lpni_peer_net->lpn_peer) {
513 * This function only allows deletion of the primary NID if it
516 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
521 lnet_net_lock(LNET_LOCK_EX);
523 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
524 struct lnet_peer_ni *lpni2;
525 /* assign the next peer_ni to be the primary */
526 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
528 lp->lp_primary_nid = lpni2->lpni_nid;
530 rc = lnet_peer_ni_del_locked(lpni, force);
532 lnet_net_unlock(LNET_LOCK_EX);
535 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
536 libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
542 lnet_peer_table_cleanup_locked(struct lnet_net *net,
543 struct lnet_peer_table *ptable)
546 struct lnet_peer_ni *next;
547 struct lnet_peer_ni *lpni;
548 struct lnet_peer *peer;
550 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
551 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
553 if (net != NULL && net != lpni->lpni_net)
556 peer = lpni->lpni_peer_net->lpn_peer;
557 if (peer->lp_primary_nid != lpni->lpni_nid) {
558 lnet_peer_ni_del_locked(lpni, false);
562 * Removing the primary NID implies removing
563 * the entire peer. Advance next beyond any
564 * peer_ni that belongs to the same peer.
566 list_for_each_entry_from(next, &ptable->pt_hash[i],
568 if (next->lpni_peer_net->lpn_peer != peer)
571 lnet_peer_del_locked(peer);
577 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
579 wait_var_event_warning(&ptable->pt_zombies,
580 ptable->pt_zombies == 0,
581 "Waiting for %d zombies on peer table\n",
586 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
587 struct lnet_peer_table *ptable)
589 struct lnet_peer_ni *lp;
590 struct lnet_peer_ni *tmp;
594 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
595 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
597 if (net != lp->lpni_net)
600 if (!lnet_isrouter(lp))
603 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
605 lnet_net_unlock(LNET_LOCK_EX);
606 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
607 lnet_net_lock(LNET_LOCK_EX);
613 lnet_peer_tables_cleanup(struct lnet_net *net)
616 struct lnet_peer_table *ptable;
618 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
619 /* If just deleting the peers for a NI, get rid of any routes these
620 * peers are gateways for. */
621 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
622 lnet_net_lock(LNET_LOCK_EX);
623 lnet_peer_table_del_rtrs_locked(net, ptable);
624 lnet_net_unlock(LNET_LOCK_EX);
627 /* Start the cleanup process */
628 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
629 lnet_net_lock(LNET_LOCK_EX);
630 lnet_peer_table_cleanup_locked(net, ptable);
631 lnet_net_unlock(LNET_LOCK_EX);
634 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
635 lnet_peer_ni_finalize_wait(ptable);
638 static struct lnet_peer_ni *
639 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
641 struct list_head *peers;
642 struct lnet_peer_ni *lp;
644 if (the_lnet.ln_state != LNET_STATE_RUNNING)
647 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
648 list_for_each_entry(lp, peers, lpni_hashlist) {
649 if (lp->lpni_nid == nid) {
650 lnet_peer_ni_addref_locked(lp);
658 struct lnet_peer_ni *
659 lnet_find_peer_ni_locked(lnet_nid_t nid)
661 struct lnet_peer_ni *lpni;
662 struct lnet_peer_table *ptable;
665 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
667 ptable = the_lnet.ln_peer_tables[cpt];
668 lpni = lnet_get_peer_ni_locked(ptable, nid);
673 struct lnet_peer_ni *
674 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
676 struct lnet_peer_net *lpn;
677 struct lnet_peer_ni *lpni;
679 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
683 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
684 if (lpni->lpni_nid == nid)
692 lnet_find_peer(lnet_nid_t nid)
694 struct lnet_peer_ni *lpni;
695 struct lnet_peer *lp = NULL;
698 cpt = lnet_net_lock_current();
699 lpni = lnet_find_peer_ni_locked(nid);
701 lp = lpni->lpni_peer_net->lpn_peer;
702 lnet_peer_addref_locked(lp);
703 lnet_peer_ni_decref_locked(lpni);
705 lnet_net_unlock(cpt);
710 struct lnet_peer_net *
711 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
713 struct lnet_peer_net *net;
716 /* no net id provided return the first net */
717 net = list_first_entry_or_null(&lp->lp_peer_nets,
718 struct lnet_peer_net,
724 /* find the net after the one provided */
725 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
726 if (net->lpn_net_id == prev_lpn_id) {
728 * if we reached the end of the list loop to the
731 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
732 return list_first_entry_or_null(&lp->lp_peer_nets,
733 struct lnet_peer_net,
736 return list_next_entry(net, lpn_peer_nets);
743 struct lnet_peer_ni *
744 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
745 struct lnet_peer_net *peer_net,
746 struct lnet_peer_ni *prev)
748 struct lnet_peer_ni *lpni;
749 struct lnet_peer_net *net = peer_net;
753 if (list_empty(&peer->lp_peer_nets))
756 net = list_entry(peer->lp_peer_nets.next,
757 struct lnet_peer_net,
760 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
766 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
768 * if you reached the end of the peer ni list and the peer
769 * net is specified then there are no more peer nis in that
776 * we reached the end of this net ni list. move to the
779 if (prev->lpni_peer_net->lpn_peer_nets.next ==
781 /* no more nets and no more NIs. */
784 /* get the next net */
785 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
786 struct lnet_peer_net,
788 /* get the ni on it */
789 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
795 /* there are more nis left */
796 lpni = list_entry(prev->lpni_peer_nis.next,
797 struct lnet_peer_ni, lpni_peer_nis);
802 /* Call with the ln_api_mutex held */
803 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
805 struct lnet_process_id id;
806 struct lnet_peer_table *ptable;
807 struct lnet_peer *lp;
816 if (the_lnet.ln_state != LNET_STATE_RUNNING)
819 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
822 * Count the number of peers, and return E2BIG if the buffer
823 * is too small. We'll also return the desired size.
826 for (cpt = 0; cpt < lncpt; cpt++) {
827 ptable = the_lnet.ln_peer_tables[cpt];
828 count += ptable->pt_peers;
830 size = count * sizeof(*ids);
835 * Walk the peer lists and copy out the primary nids.
836 * This is safe because the peer lists are only modified
837 * while the ln_api_mutex is held. So we don't need to
838 * hold the lnet_net_lock as well, and can therefore
839 * directly call copy_to_user().
842 memset(&id, 0, sizeof(id));
843 id.pid = LNET_PID_LUSTRE;
845 for (cpt = 0; cpt < lncpt; cpt++) {
846 ptable = the_lnet.ln_peer_tables[cpt];
847 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
850 id.nid = lp->lp_primary_nid;
851 if (copy_to_user(&ids[i], &id, sizeof(id)))
864 * Start pushes to peers that need to be updated for a configuration
865 * change on this node.
868 lnet_push_update_to_peers(int force)
870 struct lnet_peer_table *ptable;
871 struct lnet_peer *lp;
875 lnet_net_lock(LNET_LOCK_EX);
876 if (lnet_peer_discovery_disabled)
878 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
879 for (cpt = 0; cpt < lncpt; cpt++) {
880 ptable = the_lnet.ln_peer_tables[cpt];
881 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
883 spin_lock(&lp->lp_lock);
884 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
885 lp->lp_state |= LNET_PEER_FORCE_PUSH;
886 spin_unlock(&lp->lp_lock);
888 if (lnet_peer_needs_push(lp))
889 lnet_peer_queue_for_discovery(lp);
892 lnet_net_unlock(LNET_LOCK_EX);
893 wake_up(&the_lnet.ln_dc_waitq);
897 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
898 * this is a preferred point-to-point path. Call with lnet_net_lock in
902 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
906 if (lpni->lpni_pref_nnids == 0)
908 if (lpni->lpni_pref_nnids == 1)
909 return lpni->lpni_pref.nid == nid;
910 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
911 if (lpni->lpni_pref.nids[i] == nid)
918 * Set a single ni as preferred, provided no preferred ni is already
919 * defined. Only to be used for non-multi-rail peer_ni.
922 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
926 spin_lock(&lpni->lpni_lock);
927 if (nid == LNET_NID_ANY) {
929 } else if (lpni->lpni_pref_nnids > 0) {
931 } else if (lpni->lpni_pref_nnids == 0) {
932 lpni->lpni_pref.nid = nid;
933 lpni->lpni_pref_nnids = 1;
934 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
936 spin_unlock(&lpni->lpni_lock);
938 CDEBUG(D_NET, "peer %s nid %s: %d\n",
939 libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
944 * Clear the preferred NID from a non-multi-rail peer_ni, provided
945 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
948 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
952 spin_lock(&lpni->lpni_lock);
953 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
954 lpni->lpni_pref_nnids = 0;
955 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
956 } else if (lpni->lpni_pref_nnids == 0) {
961 spin_unlock(&lpni->lpni_lock);
963 CDEBUG(D_NET, "peer %s: %d\n",
964 libcfs_nid2str(lpni->lpni_nid), rc);
969 * Clear the preferred NIDs from a non-multi-rail peer.
972 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
974 struct lnet_peer_ni *lpni = NULL;
976 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
977 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
981 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
983 lnet_nid_t *nids = NULL;
984 lnet_nid_t *oldnids = NULL;
985 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
990 if (nid == LNET_NID_ANY) {
995 if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1000 /* A non-MR node may have only one preferred NI per peer_ni */
1001 if (lpni->lpni_pref_nnids > 0) {
1002 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1008 if (lpni->lpni_pref_nnids != 0) {
1009 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1010 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1015 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1016 if (lpni->lpni_pref.nids[i] == nid) {
1017 LIBCFS_FREE(nids, size);
1021 nids[i] = lpni->lpni_pref.nids[i];
1026 lnet_net_lock(LNET_LOCK_EX);
1027 spin_lock(&lpni->lpni_lock);
1028 if (lpni->lpni_pref_nnids == 0) {
1029 lpni->lpni_pref.nid = nid;
1031 oldnids = lpni->lpni_pref.nids;
1032 lpni->lpni_pref.nids = nids;
1034 lpni->lpni_pref_nnids++;
1035 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1036 spin_unlock(&lpni->lpni_lock);
1037 lnet_net_unlock(LNET_LOCK_EX);
1040 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1041 CFS_FREE_PTR_ARRAY(oldnids, size);
1044 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1045 spin_lock(&lpni->lpni_lock);
1046 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1047 spin_unlock(&lpni->lpni_lock);
1049 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1050 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1055 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1057 lnet_nid_t *nids = NULL;
1058 lnet_nid_t *oldnids = NULL;
1059 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1064 if (lpni->lpni_pref_nnids == 0) {
1069 if (lpni->lpni_pref_nnids == 1) {
1070 if (lpni->lpni_pref.nid != nid) {
1074 } else if (lpni->lpni_pref_nnids == 2) {
1075 if (lpni->lpni_pref.nids[0] != nid &&
1076 lpni->lpni_pref.nids[1] != nid) {
1081 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1082 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1087 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1088 if (lpni->lpni_pref.nids[i] != nid)
1090 nids[j++] = lpni->lpni_pref.nids[i];
1092 /* Check if we actually removed a nid. */
1093 if (j == lpni->lpni_pref_nnids) {
1094 LIBCFS_FREE(nids, size);
1100 lnet_net_lock(LNET_LOCK_EX);
1101 spin_lock(&lpni->lpni_lock);
1102 if (lpni->lpni_pref_nnids == 1) {
1103 lpni->lpni_pref.nid = LNET_NID_ANY;
1104 } else if (lpni->lpni_pref_nnids == 2) {
1105 oldnids = lpni->lpni_pref.nids;
1106 if (oldnids[0] == nid)
1107 lpni->lpni_pref.nid = oldnids[1];
1109 lpni->lpni_pref.nid = oldnids[2];
1111 oldnids = lpni->lpni_pref.nids;
1112 lpni->lpni_pref.nids = nids;
1114 lpni->lpni_pref_nnids--;
1115 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1116 spin_unlock(&lpni->lpni_lock);
1117 lnet_net_unlock(LNET_LOCK_EX);
1120 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1121 CFS_FREE_PTR_ARRAY(oldnids, size);
1124 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1125 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1130 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1132 struct lnet_peer_ni *lpni;
1133 lnet_nid_t primary_nid = nid;
1135 lpni = lnet_find_peer_ni_locked(nid);
1137 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1138 lnet_peer_ni_decref_locked(lpni);
1145 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1146 __must_hold(&lp->lp_lock)
1148 if (lnet_peer_discovery_disabled)
1151 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1152 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1163 lnet_is_discovery_disabled(struct lnet_peer *lp)
1167 spin_lock(&lp->lp_lock);
1168 rc = lnet_is_discovery_disabled_locked(lp);
1169 spin_unlock(&lp->lp_lock);
1175 LNetPrimaryNID(lnet_nid_t nid)
1177 struct lnet_peer *lp;
1178 struct lnet_peer_ni *lpni;
1179 lnet_nid_t primary_nid = nid;
1183 cpt = lnet_net_lock_current();
1184 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1189 lp = lpni->lpni_peer_net->lpn_peer;
1191 while (!lnet_peer_is_uptodate(lp)) {
1192 rc = lnet_discover_peer_locked(lpni, cpt, true);
1195 lp = lpni->lpni_peer_net->lpn_peer;
1197 /* Only try once if discovery is disabled */
1198 if (lnet_is_discovery_disabled(lp))
1201 primary_nid = lp->lp_primary_nid;
1203 lnet_peer_ni_decref_locked(lpni);
1205 lnet_net_unlock(cpt);
1207 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1208 libcfs_nid2str(primary_nid), rc);
1211 EXPORT_SYMBOL(LNetPrimaryNID);
1213 struct lnet_peer_net *
1214 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1216 struct lnet_peer_net *peer_net;
1217 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1218 if (peer_net->lpn_net_id == net_id)
1225 * Attach a peer_ni to a peer_net and peer. This function assumes
1226 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1227 * may be attached to a different peer, in which case it will be
1228 * properly detached first. The whole operation is done atomically.
1230 * Always returns 0. This is the last function called from functions
1231 * that do return an int, so returning 0 here allows the compiler to
1235 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1236 struct lnet_peer_net *lpn,
1237 struct lnet_peer_ni *lpni,
1240 struct lnet_peer_table *ptable;
1242 /* Install the new peer_ni */
1243 lnet_net_lock(LNET_LOCK_EX);
1244 /* Add peer_ni to global peer table hash, if necessary. */
1245 if (list_empty(&lpni->lpni_hashlist)) {
1246 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1248 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1249 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1250 ptable->pt_version++;
1251 /* This is the 1st refcount on lpni. */
1252 atomic_inc(&lpni->lpni_refcount);
1255 /* Detach the peer_ni from an existing peer, if necessary. */
1256 if (lpni->lpni_peer_net) {
1257 LASSERT(lpni->lpni_peer_net != lpn);
1258 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1259 lnet_peer_detach_peer_ni_locked(lpni);
1260 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1261 lpni->lpni_peer_net = NULL;
1264 /* Add peer_ni to peer_net */
1265 lpni->lpni_peer_net = lpn;
1266 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1267 lnet_update_peer_net_healthv(lpni);
1268 lnet_peer_net_addref_locked(lpn);
1270 /* Add peer_net to peer */
1271 if (!lpn->lpn_peer) {
1273 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1274 lnet_peer_addref_locked(lp);
1277 /* Add peer to global peer list, if necessary */
1278 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1279 if (list_empty(&lp->lp_peer_list)) {
1280 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1285 /* Update peer state */
1286 spin_lock(&lp->lp_lock);
1287 if (flags & LNET_PEER_CONFIGURED) {
1288 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1289 lp->lp_state |= LNET_PEER_CONFIGURED;
1291 if (flags & LNET_PEER_MULTI_RAIL) {
1292 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1293 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1294 lnet_peer_clr_non_mr_pref_nids(lp);
1297 spin_unlock(&lp->lp_lock);
1300 lnet_net_unlock(LNET_LOCK_EX);
1302 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1303 libcfs_nid2str(lp->lp_primary_nid),
1304 libcfs_nid2str(lpni->lpni_nid), flags);
1310 * Create a new peer, with nid as its primary nid.
1312 * Call with the lnet_api_mutex held.
1315 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1317 struct lnet_peer *lp;
1318 struct lnet_peer_net *lpn;
1319 struct lnet_peer_ni *lpni;
1322 LASSERT(nid != LNET_NID_ANY);
1325 * No need for the lnet_net_lock here, because the
1326 * lnet_api_mutex is held.
1328 lpni = lnet_find_peer_ni_locked(nid);
1330 /* A peer with this NID already exists. */
1331 lp = lpni->lpni_peer_net->lpn_peer;
1332 lnet_peer_ni_decref_locked(lpni);
1334 * This is an error if the peer was configured and the
1335 * primary NID differs or an attempt is made to change
1336 * the Multi-Rail flag. Otherwise the assumption is
1337 * that an existing peer is being modified.
1339 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1340 if (lp->lp_primary_nid != nid)
1342 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1346 /* Delete and recreate as a configured peer. */
1350 /* Create peer, peer_net, and peer_ni. */
1352 lp = lnet_peer_alloc(nid);
1355 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1358 lpni = lnet_peer_ni_alloc(nid);
1362 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1365 LIBCFS_FREE(lpn, sizeof(*lpn));
1367 LIBCFS_FREE(lp, sizeof(*lp));
1369 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1370 libcfs_nid2str(nid), flags, rc);
1375 * Add a NID to a peer. Call with ln_api_mutex held.
1378 * -EPERM: Non-DLC addition to a DLC-configured peer.
1379 * -EEXIST: The NID was configured by DLC for a different peer.
1380 * -ENOMEM: Out of memory.
1381 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1382 * non-multi-rail peer.
1385 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1387 struct lnet_peer_net *lpn;
1388 struct lnet_peer_ni *lpni;
1392 LASSERT(nid != LNET_NID_ANY);
1394 /* A configured peer can only be updated through configuration. */
1395 if (!(flags & LNET_PEER_CONFIGURED)) {
1396 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1403 * The MULTI_RAIL flag can be set but not cleared, because
1404 * that would leave the peer struct in an invalid state.
1406 if (flags & LNET_PEER_MULTI_RAIL) {
1407 spin_lock(&lp->lp_lock);
1408 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1409 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1410 lnet_peer_clr_non_mr_pref_nids(lp);
1412 spin_unlock(&lp->lp_lock);
1413 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1418 lpni = lnet_find_peer_ni_locked(nid);
1421 * A peer_ni already exists. This is only a problem if
1422 * it is not connected to this peer and was configured
1425 lnet_peer_ni_decref_locked(lpni);
1426 if (lpni->lpni_peer_net->lpn_peer == lp)
1428 if (lnet_peer_ni_is_configured(lpni)) {
1432 /* If this is the primary NID, destroy the peer. */
1433 if (lnet_peer_ni_is_primary(lpni)) {
1434 struct lnet_peer *rtr_lp =
1435 lpni->lpni_peer_net->lpn_peer;
1436 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1438 * if we're trying to delete a router it means
1439 * we're moving this peer NI to a new peer so must
1440 * transfer router properties to the new peer
1442 if (rtr_refcount > 0) {
1443 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1444 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1446 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1447 lpni = lnet_peer_ni_alloc(nid);
1454 lpni = lnet_peer_ni_alloc(nid);
1462 * Get the peer_net. Check that we're not adding a second
1463 * peer_ni on a peer_net of a non-multi-rail peer.
1465 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1467 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1472 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1477 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1480 /* If the peer_ni was allocated above its peer_net pointer is NULL */
1481 if (!lpni->lpni_peer_net)
1482 LIBCFS_FREE(lpni, sizeof(*lpni));
1484 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1485 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1491 * Update the primary NID of a peer, if possible.
1493 * Call with the lnet_api_mutex held.
1496 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1498 lnet_nid_t old = lp->lp_primary_nid;
1501 if (lp->lp_primary_nid == nid)
1503 rc = lnet_peer_add_nid(lp, nid, flags);
1506 lp->lp_primary_nid = nid;
1508 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1509 libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1514 * lpni creation initiated due to traffic either sending or receiving.
1517 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1519 struct lnet_peer *lp;
1520 struct lnet_peer_net *lpn;
1521 struct lnet_peer_ni *lpni;
1525 if (nid == LNET_NID_ANY) {
1530 /* lnet_net_lock is not needed here because ln_api_lock is held */
1531 lpni = lnet_find_peer_ni_locked(nid);
1534 * We must have raced with another thread. Since we
1535 * know next to nothing about a peer_ni created by
1536 * traffic, we just assume everything is ok and
1539 lnet_peer_ni_decref_locked(lpni);
1543 /* Create peer, peer_net, and peer_ni. */
1545 lp = lnet_peer_alloc(nid);
1548 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1551 lpni = lnet_peer_ni_alloc(nid);
1554 if (pref != LNET_NID_ANY)
1555 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1557 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1560 LIBCFS_FREE(lpn, sizeof(*lpn));
1562 LIBCFS_FREE(lp, sizeof(*lp));
1564 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1569 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1571 * This API handles the following combinations:
1572 * Create a peer with its primary NI if only the prim_nid is provided
1573 * Add a NID to a peer identified by the prim_nid. The peer identified
1574 * by the prim_nid must already exist.
1575 * The peer being created may be non-MR.
1577 * The caller must hold ln_api_mutex. This prevents the peer from
1578 * being created/modified/deleted by a different thread.
1581 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1583 struct lnet_peer *lp = NULL;
1584 struct lnet_peer_ni *lpni;
1587 /* The prim_nid must always be specified */
1588 if (prim_nid == LNET_NID_ANY)
1591 flags = LNET_PEER_CONFIGURED;
1593 flags |= LNET_PEER_MULTI_RAIL;
1596 * If nid isn't specified, we must create a new peer with
1597 * prim_nid as its primary nid.
1599 if (nid == LNET_NID_ANY)
1600 return lnet_peer_add(prim_nid, flags);
1602 /* Look up the prim_nid, which must exist. */
1603 lpni = lnet_find_peer_ni_locked(prim_nid);
1606 lnet_peer_ni_decref_locked(lpni);
1607 lp = lpni->lpni_peer_net->lpn_peer;
1609 /* Peer must have been configured. */
1610 if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1611 CDEBUG(D_NET, "peer %s was not configured\n",
1612 libcfs_nid2str(prim_nid));
1616 /* Primary NID must match */
1617 if (lp->lp_primary_nid != prim_nid) {
1618 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1619 libcfs_nid2str(prim_nid),
1620 libcfs_nid2str(lp->lp_primary_nid));
1624 /* Multi-Rail flag must match. */
1625 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1626 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1627 libcfs_nid2str(prim_nid));
1631 return lnet_peer_add_nid(lp, nid, flags);
1635 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1637 * This API handles the following combinations:
1638 * Delete a NI from a peer if both prim_nid and nid are provided.
1639 * Delete a peer if only prim_nid is provided.
1640 * Delete a peer if its primary nid is provided.
1642 * The caller must hold ln_api_mutex. This prevents the peer from
1643 * being modified/deleted by a different thread.
1646 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1648 struct lnet_peer *lp;
1649 struct lnet_peer_ni *lpni;
1652 if (prim_nid == LNET_NID_ANY)
1655 lpni = lnet_find_peer_ni_locked(prim_nid);
1658 lnet_peer_ni_decref_locked(lpni);
1659 lp = lpni->lpni_peer_net->lpn_peer;
1661 if (prim_nid != lp->lp_primary_nid) {
1662 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1663 libcfs_nid2str(prim_nid),
1664 libcfs_nid2str(lp->lp_primary_nid));
1668 lnet_net_lock(LNET_LOCK_EX);
1669 if (lp->lp_rtr_refcount > 0) {
1670 lnet_net_unlock(LNET_LOCK_EX);
1671 CERROR("%s is a router. Can not be deleted\n",
1672 libcfs_nid2str(prim_nid));
1675 lnet_net_unlock(LNET_LOCK_EX);
1677 if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1678 return lnet_peer_del(lp);
1680 flags = LNET_PEER_CONFIGURED;
1681 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1682 flags |= LNET_PEER_MULTI_RAIL;
1684 return lnet_peer_del_nid(lp, nid, flags);
1688 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1690 struct lnet_peer_table *ptable;
1691 struct lnet_peer_net *lpn;
1693 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1695 LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1696 LASSERT(list_empty(&lpni->lpni_txq));
1697 LASSERT(lpni->lpni_txqnob == 0);
1698 LASSERT(list_empty(&lpni->lpni_peer_nis));
1699 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1701 lpn = lpni->lpni_peer_net;
1702 lpni->lpni_peer_net = NULL;
1703 lpni->lpni_net = NULL;
1705 /* remove the peer ni from the zombie list */
1706 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1707 spin_lock(&ptable->pt_zombie_lock);
1708 list_del_init(&lpni->lpni_hashlist);
1709 ptable->pt_zombies--;
1710 spin_unlock(&ptable->pt_zombie_lock);
1712 if (lpni->lpni_pref_nnids > 1)
1713 CFS_FREE_PTR_ARRAY(lpni->lpni_pref.nids, lpni->lpni_pref_nnids);
1715 LIBCFS_FREE(lpni, sizeof(*lpni));
1717 lnet_peer_net_decref_locked(lpn);
1720 struct lnet_peer_ni *
1721 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1723 struct lnet_peer_ni *lpni = NULL;
1726 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1727 return ERR_PTR(-ESHUTDOWN);
1730 * find if a peer_ni already exists.
1731 * If so then just return that.
1733 lpni = lnet_find_peer_ni_locked(nid);
1737 lnet_net_unlock(cpt);
1739 rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1742 goto out_net_relock;
1745 lpni = lnet_find_peer_ni_locked(nid);
1755 * Get a peer_ni for the given nid, create it if necessary. Takes a
1756 * hold on the peer_ni.
1758 struct lnet_peer_ni *
1759 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1761 struct lnet_peer_ni *lpni = NULL;
1764 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1765 return ERR_PTR(-ESHUTDOWN);
1768 * find if a peer_ni already exists.
1769 * If so then just return that.
1771 lpni = lnet_find_peer_ni_locked(nid);
1777 * use the lnet_api_mutex to serialize the creation of the peer_ni
1778 * and the creation/deletion of the local ni/net. When a local ni is
1779 * created, if there exists a set of peer_nis on that network,
1780 * they need to be traversed and updated. When a local NI is
1781 * deleted, which could result in a network being deleted, then
1782 * all peer nis on that network need to be removed as well.
1784 * Creation through traffic should also be serialized with
1785 * creation through DLC.
1787 lnet_net_unlock(cpt);
1788 mutex_lock(&the_lnet.ln_api_mutex);
1790 * Shutdown is only set under the ln_api_lock, so a single
1791 * check here is sufficent.
1793 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1794 lpni = ERR_PTR(-ESHUTDOWN);
1795 goto out_mutex_unlock;
1798 rc = lnet_peer_ni_traffic_add(nid, pref);
1801 goto out_mutex_unlock;
1804 lpni = lnet_find_peer_ni_locked(nid);
1808 mutex_unlock(&the_lnet.ln_api_mutex);
1811 /* Lock has been dropped, check again for shutdown. */
1812 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1814 lnet_peer_ni_decref_locked(lpni);
1815 lpni = ERR_PTR(-ESHUTDOWN);
1822 lnet_peer_gw_discovery(struct lnet_peer *lp)
1826 spin_lock(&lp->lp_lock);
1827 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1829 spin_unlock(&lp->lp_lock);
1835 lnet_peer_is_uptodate(struct lnet_peer *lp)
1839 spin_lock(&lp->lp_lock);
1840 rc = lnet_peer_is_uptodate_locked(lp);
1841 spin_unlock(&lp->lp_lock);
1846 * Is a peer uptodate from the point of view of discovery?
1848 * If it is currently being processed, obviously not.
1849 * A forced Ping or Push is also handled by the discovery thread.
1851 * Otherwise look at whether the peer needs rediscovering.
1854 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1855 __must_hold(&lp->lp_lock)
1859 if (lp->lp_state & (LNET_PEER_DISCOVERING |
1860 LNET_PEER_FORCE_PING |
1861 LNET_PEER_FORCE_PUSH)) {
1863 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1865 } else if (lnet_peer_needs_push(lp)) {
1867 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1868 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1880 * Queue a peer for the attention of the discovery thread. Call with
1881 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1882 * -EALREADY if the peer was already queued.
1884 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1888 spin_lock(&lp->lp_lock);
1889 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1890 lp->lp_state |= LNET_PEER_DISCOVERING;
1891 spin_unlock(&lp->lp_lock);
1892 if (list_empty(&lp->lp_dc_list)) {
1893 lnet_peer_addref_locked(lp);
1894 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1895 wake_up(&the_lnet.ln_dc_waitq);
1901 CDEBUG(D_NET, "Queue peer %s: %d\n",
1902 libcfs_nid2str(lp->lp_primary_nid), rc);
1908 * Discovery of a peer is complete. Wake all waiters on the peer.
1909 * Call with lnet_net_lock/EX held.
1911 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1913 struct lnet_msg *msg, *tmp;
1915 LIST_HEAD(pending_msgs);
1917 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1918 libcfs_nid2str(lp->lp_primary_nid));
1920 list_del_init(&lp->lp_dc_list);
1921 spin_lock(&lp->lp_lock);
1922 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1923 spin_unlock(&lp->lp_lock);
1924 wake_up_all(&lp->lp_dc_waitq);
1926 if (lp->lp_rtr_refcount > 0)
1927 lnet_router_discovery_complete(lp);
1929 lnet_net_unlock(LNET_LOCK_EX);
1931 /* iterate through all pending messages and send them again */
1932 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1933 list_del_init(&msg->msg_list);
1934 if (lp->lp_dc_error) {
1935 lnet_finalize(msg, lp->lp_dc_error);
1939 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1940 lnet_msgtyp2str(msg->msg_type),
1941 libcfs_id2str(msg->msg_target));
1942 rc = lnet_send(msg->msg_src_nid_param, msg,
1943 msg->msg_rtr_nid_param);
1945 CNETERR("Error sending %s to %s: %d\n",
1946 lnet_msgtyp2str(msg->msg_type),
1947 libcfs_id2str(msg->msg_target), rc);
1948 lnet_finalize(msg, rc);
1951 lnet_net_lock(LNET_LOCK_EX);
1952 lnet_peer_decref_locked(lp);
1956 * Handle inbound push.
1957 * Like any event handler, called with lnet_res_lock/CPT held.
1959 void lnet_peer_push_event(struct lnet_event *ev)
1961 struct lnet_ping_buffer *pbuf;
1962 struct lnet_peer *lp;
1964 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start + ev->offset);
1966 /* lnet_find_peer() adds a refcount */
1967 lp = lnet_find_peer(ev->source.nid);
1969 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1970 libcfs_nid2str(ev->initiator.nid),
1971 libcfs_nid2str(ev->source.nid));
1972 pbuf->pb_needs_post = true;
1976 /* Ensure peer state remains consistent while we modify it. */
1977 spin_lock(&lp->lp_lock);
1980 * If some kind of error happened the contents of the message
1981 * cannot be used. Clear the NIDS_UPTODATE and set the
1982 * FORCE_PING flag to trigger a ping.
1985 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1986 lp->lp_state |= LNET_PEER_FORCE_PING;
1987 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1989 libcfs_nid2str(lp->lp_primary_nid),
1990 libcfs_nid2str(ev->source.nid));
1995 * A push with invalid or corrupted info. Clear the UPTODATE
1996 * flag to trigger a ping.
1998 if (lnet_ping_info_validate(&pbuf->pb_info)) {
1999 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2000 lp->lp_state |= LNET_PEER_FORCE_PING;
2001 CDEBUG(D_NET, "Corrupted Push from %s\n",
2002 libcfs_nid2str(lp->lp_primary_nid));
2007 * Make sure we'll allocate the correct size ping buffer when
2010 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2011 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2014 * A non-Multi-Rail peer is not supposed to be capable of
2017 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2018 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2019 libcfs_nid2str(lp->lp_primary_nid));
2024 * The peer may have discovery disabled at its end. Set
2025 * NO_DISCOVERY as appropriate.
2027 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2028 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2029 libcfs_nid2str(lp->lp_primary_nid));
2030 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2031 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2032 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2033 libcfs_nid2str(lp->lp_primary_nid));
2034 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2038 * Update the MULTI_RAIL flag based on the push. If the peer
2039 * was configured with DLC then the setting should match what
2041 * NB: We verified above that the MR feature bit is set in pi_features
2043 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2044 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2045 libcfs_nid2str(lp->lp_primary_nid), lp);
2046 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2047 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2048 libcfs_nid2str(lp->lp_primary_nid));
2049 } else if (lnet_peer_discovery_disabled) {
2050 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2051 libcfs_nid2str(lp->lp_primary_nid), lp);
2052 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2053 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2054 libcfs_nid2str(lp->lp_primary_nid), lp);
2056 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2057 libcfs_nid2str(lp->lp_primary_nid), lp);
2058 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2059 lnet_peer_clr_non_mr_pref_nids(lp);
2063 * Check for truncation of the Put message. Clear the
2064 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2065 * and tell discovery to allocate a bigger buffer.
2067 if (ev->mlength < ev->rlength) {
2068 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2069 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2070 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2071 lp->lp_state |= LNET_PEER_FORCE_PING;
2072 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2073 libcfs_nid2str(lp->lp_primary_nid),
2074 pbuf->pb_info.pi_nnis);
2078 /* always assume new data */
2079 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2080 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2083 * If there is data present that hasn't been processed yet,
2084 * we'll replace it if the Put contained newer data and it
2085 * fits. We're racing with a Ping or earlier Push in this
2088 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2089 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2090 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2091 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2092 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2093 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2094 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2095 libcfs_nid2str(lp->lp_primary_nid),
2096 LNET_PING_BUFFER_SEQNO(pbuf),
2097 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2103 * Allocate a buffer to copy the data. On a failure we drop
2104 * the Push and set FORCE_PING to force the discovery
2105 * thread to fix the problem by pinging the peer.
2107 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2109 lp->lp_state |= LNET_PEER_FORCE_PING;
2110 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2111 libcfs_nid2str(lp->lp_primary_nid),
2112 LNET_PING_BUFFER_SEQNO(pbuf));
2117 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2118 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2119 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2120 CDEBUG(D_NET, "Received Push %s %u\n",
2121 libcfs_nid2str(lp->lp_primary_nid),
2122 LNET_PING_BUFFER_SEQNO(pbuf));
2125 /* We've processed this buffer. It can be reposted */
2126 pbuf->pb_needs_post = true;
2129 * Queue the peer for discovery if not done, force it on the request
2130 * queue and wake the discovery thread if the peer was already queued,
2131 * because its status changed.
2133 spin_unlock(&lp->lp_lock);
2134 lnet_net_lock(LNET_LOCK_EX);
2135 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2136 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2137 wake_up(&the_lnet.ln_dc_waitq);
2139 /* Drop refcount from lookup */
2140 lnet_peer_decref_locked(lp);
2141 lnet_net_unlock(LNET_LOCK_EX);
2145 * Clear the discovery error state, unless we're already discovering
2146 * this peer, in which case the error is current.
2148 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2150 spin_lock(&lp->lp_lock);
2151 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2152 lp->lp_dc_error = 0;
2153 spin_unlock(&lp->lp_lock);
2157 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2158 * dropped/retaken within this function. An lnet_peer_ni is passed in
2159 * because discovery could tear down an lnet_peer.
2162 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2165 struct lnet_peer *lp;
2170 lnet_net_unlock(cpt);
2171 lnet_net_lock(LNET_LOCK_EX);
2172 lp = lpni->lpni_peer_net->lpn_peer;
2173 lnet_peer_clear_discovery_error(lp);
2176 * We're willing to be interrupted. The lpni can become a
2177 * zombie if we race with DLC, so we must check for that.
2180 /* Keep lp alive when the lnet_net_lock is unlocked */
2181 lnet_peer_addref_locked(lp);
2182 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2183 if (signal_pending(current))
2185 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2188 * Don't repeat discovery if discovery is disabled. This is
2189 * done to ensure we can use discovery as a standard ping as
2190 * well for backwards compatibility with routers which do not
2191 * have discovery or have discovery disabled
2193 if (lnet_is_discovery_disabled(lp) && count > 0)
2195 if (lp->lp_dc_error)
2197 if (lnet_peer_is_uptodate(lp))
2199 lnet_peer_queue_for_discovery(lp);
2201 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2204 * If caller requested a non-blocking operation then
2205 * return immediately. Once discovery is complete any
2206 * pending messages that were stopped due to discovery
2207 * will be transmitted.
2212 lnet_net_unlock(LNET_LOCK_EX);
2214 finish_wait(&lp->lp_dc_waitq, &wait);
2215 lnet_net_lock(LNET_LOCK_EX);
2216 lnet_peer_decref_locked(lp);
2217 /* Peer may have changed */
2218 lp = lpni->lpni_peer_net->lpn_peer;
2220 finish_wait(&lp->lp_dc_waitq, &wait);
2222 lnet_net_unlock(LNET_LOCK_EX);
2224 lnet_peer_decref_locked(lp);
2226 * The peer may have changed, so re-check and rediscover if that turns
2227 * out to have been the case. The reference count on lp ensured that
2228 * even if it was unlinked from lpni the memory could not be recycled.
2229 * Thus the check below is sufficient to determine whether the peer
2230 * changed. If the peer changed, then lp must not be dereferenced.
2232 if (lp != lpni->lpni_peer_net->lpn_peer)
2235 if (signal_pending(current))
2237 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2239 else if (lp->lp_dc_error)
2240 rc = lp->lp_dc_error;
2242 CDEBUG(D_NET, "non-blocking discovery\n");
2243 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2246 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2247 (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2248 libcfs_nid2str(lpni->lpni_nid), rc,
2249 (!block) ? "pending discovery" : "discovery complete");
2254 /* Handle an incoming ack for a push. */
2256 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2258 struct lnet_ping_buffer *pbuf;
2260 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2261 spin_lock(&lp->lp_lock);
2262 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2263 lp->lp_push_error = ev->status;
2265 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2267 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2268 spin_unlock(&lp->lp_lock);
2270 CDEBUG(D_NET, "peer %s ev->status %d\n",
2271 libcfs_nid2str(lp->lp_primary_nid), ev->status);
2274 /* Handle a Reply message. This is the reply to a Ping message. */
2276 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2278 struct lnet_ping_buffer *pbuf;
2281 spin_lock(&lp->lp_lock);
2284 * If some kind of error happened the contents of message
2285 * cannot be used. Set PING_FAILED to trigger a retry.
2288 lp->lp_state |= LNET_PEER_PING_FAILED;
2289 lp->lp_ping_error = ev->status;
2290 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2292 libcfs_nid2str(lp->lp_primary_nid),
2293 libcfs_nid2str(ev->source.nid));
2297 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2298 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2299 lnet_swap_pinginfo(pbuf);
2302 * A reply with invalid or corrupted info. Set PING_FAILED to
2305 rc = lnet_ping_info_validate(&pbuf->pb_info);
2307 lp->lp_state |= LNET_PEER_PING_FAILED;
2308 lp->lp_ping_error = 0;
2309 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2310 libcfs_nid2str(lp->lp_primary_nid), rc);
2316 * The peer may have discovery disabled at its end. Set
2317 * NO_DISCOVERY as appropriate.
2319 if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2320 !lnet_peer_discovery_disabled) {
2321 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2322 libcfs_nid2str(lp->lp_primary_nid));
2323 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2325 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2326 libcfs_nid2str(lp->lp_primary_nid));
2327 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2331 * Update the MULTI_RAIL flag based on the reply. If the peer
2332 * was configured with DLC then the setting should match what
2335 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2336 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2337 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2338 libcfs_nid2str(lp->lp_primary_nid), lp);
2339 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2340 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2341 libcfs_nid2str(lp->lp_primary_nid));
2342 } else if (lnet_peer_discovery_disabled) {
2344 "peer %s(%p) not MR: DD disabled locally\n",
2345 libcfs_nid2str(lp->lp_primary_nid), lp);
2346 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2348 "peer %s(%p) not MR: DD disabled remotely\n",
2349 libcfs_nid2str(lp->lp_primary_nid), lp);
2351 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2352 libcfs_nid2str(lp->lp_primary_nid), lp);
2353 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2354 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);
2418 lnet_net_lock(LNET_LOCK_EX);
2420 * If this peer is a gateway, call the routing callback to
2421 * handle the ping reply
2423 if (lp->lp_rtr_refcount > 0)
2424 lnet_router_discovery_ping_reply(lp);
2425 lnet_net_unlock(LNET_LOCK_EX);
2429 * Send event handling. Only matters for error cases, where we clean
2430 * up state on the peer and peer_ni that would otherwise be updated in
2431 * the REPLY event handler for a successful Ping, and the ACK event
2432 * handler for a successful Push.
2435 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2442 spin_lock(&lp->lp_lock);
2443 if (ev->msg_type == LNET_MSG_GET) {
2444 lp->lp_state &= ~LNET_PEER_PING_SENT;
2445 lp->lp_state |= LNET_PEER_PING_FAILED;
2446 lp->lp_ping_error = ev->status;
2447 } else { /* ev->msg_type == LNET_MSG_PUT */
2448 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2449 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2450 lp->lp_push_error = ev->status;
2452 spin_unlock(&lp->lp_lock);
2453 rc = LNET_REDISCOVER_PEER;
2455 CDEBUG(D_NET, "%s Send to %s: %d\n",
2456 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2457 libcfs_nid2str(ev->target.nid), rc);
2462 * Unlink event handling. This event is only seen if a call to
2463 * LNetMDUnlink() caused the event to be unlinked. If this call was
2464 * made after the event was set up in LNetGet() or LNetPut() then we
2465 * assume the Ping or Push timed out.
2468 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2470 spin_lock(&lp->lp_lock);
2471 /* We've passed through LNetGet() */
2472 if (lp->lp_state & LNET_PEER_PING_SENT) {
2473 lp->lp_state &= ~LNET_PEER_PING_SENT;
2474 lp->lp_state |= LNET_PEER_PING_FAILED;
2475 lp->lp_ping_error = -ETIMEDOUT;
2476 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2477 libcfs_nid2str(lp->lp_primary_nid));
2479 /* We've passed through LNetPut() */
2480 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2481 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2482 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2483 lp->lp_push_error = -ETIMEDOUT;
2484 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2485 libcfs_nid2str(lp->lp_primary_nid));
2487 spin_unlock(&lp->lp_lock);
2491 * Event handler for the discovery EQ.
2493 * Called with lnet_res_lock(cpt) held. The cpt is the
2494 * lnet_cpt_of_cookie() of the md handle cookie.
2496 static void lnet_discovery_event_handler(struct lnet_event *event)
2498 struct lnet_peer *lp = event->md.user_ptr;
2499 struct lnet_ping_buffer *pbuf;
2502 /* discovery needs to take another look */
2503 rc = LNET_REDISCOVER_PEER;
2505 CDEBUG(D_NET, "Received event: %d\n", event->type);
2507 switch (event->type) {
2508 case LNET_EVENT_ACK:
2509 lnet_discovery_event_ack(lp, event);
2511 case LNET_EVENT_REPLY:
2512 lnet_discovery_event_reply(lp, event);
2514 case LNET_EVENT_SEND:
2515 /* Only send failure triggers a retry. */
2516 rc = lnet_discovery_event_send(lp, event);
2518 case LNET_EVENT_UNLINK:
2519 /* LNetMDUnlink() was called */
2520 lnet_discovery_event_unlink(lp, event);
2523 /* Invalid events. */
2526 lnet_net_lock(LNET_LOCK_EX);
2527 if (event->unlinked) {
2528 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2529 lnet_ping_buffer_decref(pbuf);
2530 lnet_peer_decref_locked(lp);
2533 /* put peer back at end of request queue, if discovery not already
2535 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2536 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2537 wake_up(&the_lnet.ln_dc_waitq);
2539 lnet_net_unlock(LNET_LOCK_EX);
2543 * Build a peer from incoming data.
2545 * The NIDs in the incoming data are supposed to be structured as follows:
2548 * - other NIDs in same net
2549 * - NIDs in second net
2550 * - NIDs in third net
2552 * This due to the way the list of NIDs in the data is created.
2554 * Note that this function will mark the peer uptodate unless an
2555 * ENOMEM is encontered. All other errors are due to a conflict
2556 * between the DLC configuration and what discovery sees. We treat DLC
2557 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2558 * peer from becoming stuck in discovery.
2560 static int lnet_peer_merge_data(struct lnet_peer *lp,
2561 struct lnet_ping_buffer *pbuf)
2563 struct lnet_peer_ni *lpni;
2564 lnet_nid_t *curnis = NULL;
2565 struct lnet_ni_status *addnis = NULL;
2566 lnet_nid_t *delnis = NULL;
2576 flags = LNET_PEER_DISCOVERED;
2577 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2578 flags |= LNET_PEER_MULTI_RAIL;
2581 * Cache the routing feature for the peer; whether it is enabled
2582 * for disabled as reported by the remote peer.
2584 spin_lock(&lp->lp_lock);
2585 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2586 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2588 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2589 spin_unlock(&lp->lp_lock);
2591 nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2592 CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2593 CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2594 CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2595 if (!curnis || !addnis || !delnis) {
2603 /* Construct the list of NIDs present in peer. */
2605 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2606 curnis[ncurnis++] = lpni->lpni_nid;
2609 * Check for NIDs in pbuf not present in curnis[].
2610 * The loop starts at 1 to skip the loopback NID.
2612 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2613 for (j = 0; j < ncurnis; j++)
2614 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2617 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2620 * Check for NIDs in curnis[] not present in pbuf.
2621 * The nested loop starts at 1 to skip the loopback NID.
2623 * But never add the loopback NID to delnis[]: if it is
2624 * present in curnis[] then this peer is for this node.
2626 for (i = 0; i < ncurnis; i++) {
2627 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2629 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2630 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2632 * update the information we cache for the
2633 * peer with the latest information we
2636 lpni = lnet_find_peer_ni_locked(curnis[i]);
2638 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2639 lnet_peer_ni_decref_locked(lpni);
2644 if (j == pbuf->pb_info.pi_nnis)
2645 delnis[ndelnis++] = curnis[i];
2649 * If we get here and the discovery is disabled then we don't want
2650 * to add or delete any NIs. We just updated the ones we have some
2651 * information on, and call it a day
2654 if (lnet_is_discovery_disabled(lp))
2657 for (i = 0; i < naddnis; i++) {
2658 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2660 CERROR("Error adding NID %s to peer %s: %d\n",
2661 libcfs_nid2str(addnis[i].ns_nid),
2662 libcfs_nid2str(lp->lp_primary_nid), rc);
2666 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2668 lpni->lpni_ns_status = addnis[i].ns_status;
2669 lnet_peer_ni_decref_locked(lpni);
2673 for (i = 0; i < ndelnis; i++) {
2675 * for routers it's okay to delete the primary_nid because
2676 * the upper layers don't really rely on it. So if we're
2677 * being told that the router changed its primary_nid
2678 * then it's okay to delete it.
2680 if (lp->lp_rtr_refcount > 0)
2681 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2682 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2684 CERROR("Error deleting NID %s from peer %s: %d\n",
2685 libcfs_nid2str(delnis[i]),
2686 libcfs_nid2str(lp->lp_primary_nid), rc);
2692 * Errors other than -ENOMEM are due to peers having been
2693 * configured with DLC. Ignore these because DLC overrides
2698 CFS_FREE_PTR_ARRAY(curnis, nnis);
2699 CFS_FREE_PTR_ARRAY(addnis, nnis);
2700 CFS_FREE_PTR_ARRAY(delnis, nnis);
2701 lnet_ping_buffer_decref(pbuf);
2702 CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2705 spin_lock(&lp->lp_lock);
2706 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2707 lp->lp_state |= LNET_PEER_FORCE_PING;
2708 spin_unlock(&lp->lp_lock);
2714 * The data in pbuf says lp is its primary peer, but the data was
2715 * received by a different peer. Try to update lp with the data.
2718 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2720 struct lnet_handle_md mdh;
2722 /* Queue lp for discovery, and force it on the request queue. */
2723 lnet_net_lock(LNET_LOCK_EX);
2724 if (lnet_peer_queue_for_discovery(lp))
2725 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2726 lnet_net_unlock(LNET_LOCK_EX);
2728 LNetInvalidateMDHandle(&mdh);
2731 * Decide whether we can move the peer to the DATA_PRESENT state.
2733 * We replace stale data for a multi-rail peer, repair PING_FAILED
2734 * status, and preempt FORCE_PING.
2736 * If after that we have DATA_PRESENT, we merge it into this peer.
2738 spin_lock(&lp->lp_lock);
2739 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2740 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2741 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2742 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2743 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2744 lnet_ping_buffer_decref(pbuf);
2749 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2750 lnet_ping_buffer_decref(lp->lp_data);
2752 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2754 if (lp->lp_state & LNET_PEER_PING_FAILED) {
2755 mdh = lp->lp_ping_mdh;
2756 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2757 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2758 lp->lp_ping_error = 0;
2760 if (lp->lp_state & LNET_PEER_FORCE_PING)
2761 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2762 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2763 spin_unlock(&lp->lp_lock);
2765 if (!LNetMDHandleIsInvalid(mdh))
2769 return lnet_peer_merge_data(lp, pbuf);
2771 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2775 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2779 for (i = 0; i < pinfo->pi_nnis; i++) {
2780 if (pinfo->pi_ni[i].ns_nid == nid)
2788 * Update a peer using the data received.
2790 static int lnet_peer_data_present(struct lnet_peer *lp)
2791 __must_hold(&lp->lp_lock)
2793 struct lnet_ping_buffer *pbuf;
2794 struct lnet_peer_ni *lpni;
2795 lnet_nid_t nid = LNET_NID_ANY;
2801 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2802 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2803 spin_unlock(&lp->lp_lock);
2806 * Modifications of peer structures are done while holding the
2807 * ln_api_mutex. A global lock is required because we may be
2808 * modifying multiple peer structures, and a mutex greatly
2809 * simplifies memory management.
2811 * The actual changes to the data structures must also protect
2812 * against concurrent lookups, for which the lnet_net_lock in
2813 * LNET_LOCK_EX mode is used.
2815 mutex_lock(&the_lnet.ln_api_mutex);
2816 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2822 * If this peer is not on the peer list then it is being torn
2823 * down, and our reference count may be all that is keeping it
2824 * alive. Don't do any work on it.
2826 if (list_empty(&lp->lp_peer_list))
2829 flags = LNET_PEER_DISCOVERED;
2830 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2831 flags |= LNET_PEER_MULTI_RAIL;
2834 * Check whether the primary NID in the message matches the
2835 * primary NID of the peer. If it does, update the peer, if
2836 * it it does not, check whether there is already a peer with
2837 * that primary NID. If no such peer exists, try to update
2838 * the primary NID of the current peer (allowed if it was
2839 * created due to message traffic) and complete the update.
2840 * If the peer did exist, hand off the data to it.
2842 * The peer for the loopback interface is a special case: this
2843 * is the peer for the local node, and we want to set its
2844 * primary NID to the correct value here. Moreover, this peer
2845 * can show up with only the loopback NID in the ping buffer.
2847 if (pbuf->pb_info.pi_nnis <= 1)
2849 nid = pbuf->pb_info.pi_ni[1].ns_nid;
2850 if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2851 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2853 rc = lnet_peer_merge_data(lp, pbuf);
2855 * if the primary nid of the peer is present in the ping info returned
2856 * from the peer, but it's not the local primary peer we have
2857 * cached and discovery is disabled, then we don't want to update
2858 * our local peer info, by adding or removing NIDs, we just want
2859 * to update the status of the nids that we currently have
2860 * recorded in that peer.
2862 } else if (lp->lp_primary_nid == nid ||
2863 (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2864 lnet_is_discovery_disabled(lp))) {
2865 rc = lnet_peer_merge_data(lp, pbuf);
2867 lpni = lnet_find_peer_ni_locked(nid);
2869 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2871 CERROR("Primary NID error %s versus %s: %d\n",
2872 libcfs_nid2str(lp->lp_primary_nid),
2873 libcfs_nid2str(nid), rc);
2875 rc = lnet_peer_merge_data(lp, pbuf);
2878 struct lnet_peer *new_lp;
2879 new_lp = lpni->lpni_peer_net->lpn_peer;
2881 * if lp has discovery/MR enabled that means new_lp
2882 * should have discovery/MR enabled as well, since
2883 * it's the same peer, which we're about to merge
2885 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2886 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2887 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2888 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2890 rc = lnet_peer_set_primary_data(new_lp, pbuf);
2891 lnet_consolidate_routes_locked(lp, new_lp);
2892 lnet_peer_ni_decref_locked(lpni);
2896 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2898 mutex_unlock(&the_lnet.ln_api_mutex);
2900 spin_lock(&lp->lp_lock);
2901 /* Tell discovery to re-check the peer immediately. */
2903 rc = LNET_REDISCOVER_PEER;
2908 * A ping failed. Clear the PING_FAILED state and set the
2909 * FORCE_PING state, to ensure a retry even if discovery is
2910 * disabled. This avoids being left with incorrect state.
2912 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2913 __must_hold(&lp->lp_lock)
2915 struct lnet_handle_md mdh;
2918 mdh = lp->lp_ping_mdh;
2919 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2920 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2921 lp->lp_state |= LNET_PEER_FORCE_PING;
2922 rc = lp->lp_ping_error;
2923 lp->lp_ping_error = 0;
2924 spin_unlock(&lp->lp_lock);
2926 if (!LNetMDHandleIsInvalid(mdh))
2929 CDEBUG(D_NET, "peer %s:%d\n",
2930 libcfs_nid2str(lp->lp_primary_nid), rc);
2932 spin_lock(&lp->lp_lock);
2933 return rc ? rc : LNET_REDISCOVER_PEER;
2937 * Select NID to send a Ping or Push to.
2939 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2941 struct lnet_peer_ni *lpni;
2943 /* Look for a direct-connected NID for this peer. */
2945 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2946 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2951 return lpni->lpni_nid;
2953 /* Look for a routed-connected NID for this peer. */
2955 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2956 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2961 return lpni->lpni_nid;
2963 return LNET_NID_ANY;
2966 /* Active side of ping. */
2967 static int lnet_peer_send_ping(struct lnet_peer *lp)
2968 __must_hold(&lp->lp_lock)
2975 lp->lp_state |= LNET_PEER_PING_SENT;
2976 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2977 spin_unlock(&lp->lp_lock);
2979 cpt = lnet_net_lock_current();
2980 /* Refcount for MD. */
2981 lnet_peer_addref_locked(lp);
2982 pnid = lnet_peer_select_nid(lp);
2983 lnet_net_unlock(cpt);
2985 nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2987 rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2988 the_lnet.ln_dc_eq, false);
2991 * if LNetMDBind in lnet_send_ping fails we need to decrement the
2992 * refcount on the peer, otherwise LNetMDUnlink will be called
2993 * which will eventually do that.
2997 lnet_peer_decref_locked(lp);
2998 lnet_net_unlock(cpt);
2999 rc = -rc; /* change the rc to negative value */
3001 } else if (rc < 0) {
3005 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3007 spin_lock(&lp->lp_lock);
3011 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3013 * The errors that get us here are considered hard errors and
3014 * cause Discovery to terminate. So we clear PING_SENT, but do
3015 * not set either PING_FAILED or FORCE_PING. In fact we need
3016 * to clear PING_FAILED, because the unlink event handler will
3017 * have set it if we called LNetMDUnlink() above.
3019 spin_lock(&lp->lp_lock);
3020 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3025 * This function exists because you cannot call LNetMDUnlink() from an
3028 static int lnet_peer_push_failed(struct lnet_peer *lp)
3029 __must_hold(&lp->lp_lock)
3031 struct lnet_handle_md mdh;
3034 mdh = lp->lp_push_mdh;
3035 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3036 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3037 rc = lp->lp_push_error;
3038 lp->lp_push_error = 0;
3039 spin_unlock(&lp->lp_lock);
3041 if (!LNetMDHandleIsInvalid(mdh))
3044 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3045 spin_lock(&lp->lp_lock);
3046 return rc ? rc : LNET_REDISCOVER_PEER;
3049 /* Active side of push. */
3050 static int lnet_peer_send_push(struct lnet_peer *lp)
3051 __must_hold(&lp->lp_lock)
3053 struct lnet_ping_buffer *pbuf;
3054 struct lnet_process_id id;
3059 /* Don't push to a non-multi-rail peer. */
3060 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3061 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3065 lp->lp_state |= LNET_PEER_PUSH_SENT;
3066 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3067 spin_unlock(&lp->lp_lock);
3069 cpt = lnet_net_lock_current();
3070 pbuf = the_lnet.ln_ping_target;
3071 lnet_ping_buffer_addref(pbuf);
3072 lnet_net_unlock(cpt);
3074 /* Push source MD */
3075 md.start = &pbuf->pb_info;
3076 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3077 md.threshold = 2; /* Put/Ack */
3080 md.eq_handle = the_lnet.ln_dc_eq;
3083 rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3085 lnet_ping_buffer_decref(pbuf);
3086 CERROR("Can't bind push source MD: %d\n", rc);
3089 cpt = lnet_net_lock_current();
3090 /* Refcount for MD. */
3091 lnet_peer_addref_locked(lp);
3092 id.pid = LNET_PID_LUSTRE;
3093 id.nid = lnet_peer_select_nid(lp);
3094 lnet_net_unlock(cpt);
3096 if (id.nid == LNET_NID_ANY) {
3101 rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3102 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3103 LNET_PROTO_PING_MATCHBITS, 0, 0);
3108 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3110 spin_lock(&lp->lp_lock);
3114 LNetMDUnlink(lp->lp_push_mdh);
3115 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3117 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3119 * The errors that get us here are considered hard errors and
3120 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3121 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3122 * because the unlink event handler will have set it if we
3123 * called LNetMDUnlink() above.
3125 spin_lock(&lp->lp_lock);
3126 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3131 * An unrecoverable error was encountered during discovery.
3132 * Set error status in peer and abort discovery.
3134 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3136 CDEBUG(D_NET, "Discovery error %s: %d\n",
3137 libcfs_nid2str(lp->lp_primary_nid), error);
3139 spin_lock(&lp->lp_lock);
3140 lp->lp_dc_error = error;
3141 lp->lp_state &= ~LNET_PEER_DISCOVERING;
3142 lp->lp_state |= LNET_PEER_REDISCOVER;
3143 spin_unlock(&lp->lp_lock);
3147 * Mark the peer as discovered.
3149 static int lnet_peer_discovered(struct lnet_peer *lp)
3150 __must_hold(&lp->lp_lock)
3152 lp->lp_state |= LNET_PEER_DISCOVERED;
3153 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3154 LNET_PEER_REDISCOVER);
3156 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3163 * Discovering this peer is taking too long. Cancel any Ping or Push
3164 * that discovery is waiting on by unlinking the relevant MDs. The
3165 * lnet_discovery_event_handler() will proceed from here and complete
3168 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3170 struct lnet_handle_md ping_mdh;
3171 struct lnet_handle_md push_mdh;
3173 LNetInvalidateMDHandle(&ping_mdh);
3174 LNetInvalidateMDHandle(&push_mdh);
3176 spin_lock(&lp->lp_lock);
3177 if (lp->lp_state & LNET_PEER_PING_SENT) {
3178 ping_mdh = lp->lp_ping_mdh;
3179 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3181 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3182 push_mdh = lp->lp_push_mdh;
3183 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3185 spin_unlock(&lp->lp_lock);
3187 if (!LNetMDHandleIsInvalid(ping_mdh))
3188 LNetMDUnlink(ping_mdh);
3189 if (!LNetMDHandleIsInvalid(push_mdh))
3190 LNetMDUnlink(push_mdh);
3194 * Wait for work to be queued or some other change that must be
3195 * attended to. Returns non-zero if the discovery thread should shut
3198 static int lnet_peer_discovery_wait_for_work(void)
3205 cpt = lnet_net_lock_current();
3207 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3208 TASK_INTERRUPTIBLE);
3209 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3211 if (lnet_push_target_resize_needed() ||
3212 the_lnet.ln_push_target->pb_needs_post)
3214 if (!list_empty(&the_lnet.ln_dc_request))
3216 if (!list_empty(&the_lnet.ln_msg_resend))
3218 lnet_net_unlock(cpt);
3221 * wakeup max every second to check if there are peers that
3222 * have been stuck on the working queue for greater than
3225 schedule_timeout(cfs_time_seconds(1));
3226 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3227 cpt = lnet_net_lock_current();
3229 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3231 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3234 lnet_net_unlock(cpt);
3236 CDEBUG(D_NET, "woken: %d\n", rc);
3242 * Messages that were pending on a destroyed peer will be put on a global
3243 * resend list. The message resend list will be checked by
3244 * the discovery thread when it wakes up, and will resend messages. These
3245 * messages can still be sendable in the case the lpni which was the initial
3246 * cause of the message re-queue was transfered to another peer.
3248 * It is possible that LNet could be shutdown while we're iterating
3249 * through the list. lnet_shudown_lndnets() will attempt to access the
3250 * resend list, but will have to wait until the spinlock is released, by
3251 * which time there shouldn't be any more messages on the resend list.
3252 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3253 * for the messages so they can be released. The other case is that
3254 * lnet_shudown_lndnets() can finalize all the messages before this
3255 * function can visit the resend list, in which case this function will be
3258 static void lnet_resend_msgs(void)
3260 struct lnet_msg *msg, *tmp;
3264 spin_lock(&the_lnet.ln_msg_resend_lock);
3265 list_splice(&the_lnet.ln_msg_resend, &resend);
3266 spin_unlock(&the_lnet.ln_msg_resend_lock);
3268 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3269 list_del_init(&msg->msg_list);
3270 rc = lnet_send(msg->msg_src_nid_param, msg,
3271 msg->msg_rtr_nid_param);
3273 CNETERR("Error sending %s to %s: %d\n",
3274 lnet_msgtyp2str(msg->msg_type),
3275 libcfs_id2str(msg->msg_target), rc);
3276 lnet_finalize(msg, rc);
3281 /* The discovery thread. */
3282 static int lnet_peer_discovery(void *arg)
3284 struct lnet_peer *lp;
3287 wait_for_completion(&the_lnet.ln_started);
3289 CDEBUG(D_NET, "started\n");
3292 if (lnet_peer_discovery_wait_for_work())
3295 if (lnet_push_target_resize_needed())
3296 lnet_push_target_resize();
3297 else if (the_lnet.ln_push_target->pb_needs_post)
3298 lnet_push_target_post(the_lnet.ln_push_target,
3299 &the_lnet.ln_push_target_md);
3303 lnet_net_lock(LNET_LOCK_EX);
3304 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3305 lnet_net_unlock(LNET_LOCK_EX);
3310 * Process all incoming discovery work requests. When
3311 * discovery must wait on a peer to change state, it
3312 * is added to the tail of the ln_dc_working queue. A
3313 * timestamp keeps track of when the peer was added,
3314 * so we can time out discovery requests that take too
3317 while (!list_empty(&the_lnet.ln_dc_request)) {
3318 lp = list_first_entry(&the_lnet.ln_dc_request,
3319 struct lnet_peer, lp_dc_list);
3320 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3322 * set the time the peer was put on the dc_working
3323 * queue. It shouldn't remain on the queue
3324 * forever, in case the GET message (for ping)
3325 * doesn't get a REPLY or the PUT message (for
3326 * push) doesn't get an ACK.
3328 lp->lp_last_queued = ktime_get_real_seconds();
3329 lnet_net_unlock(LNET_LOCK_EX);
3331 if (lnet_push_target_resize_needed())
3332 lnet_push_target_resize();
3333 else if (the_lnet.ln_push_target->pb_needs_post)
3334 lnet_push_target_post(the_lnet.ln_push_target,
3335 &the_lnet.ln_push_target_md);
3338 * Select an action depending on the state of
3339 * the peer and whether discovery is disabled.
3340 * The check whether discovery is disabled is
3341 * done after the code that handles processing
3342 * for arrived data, cleanup for failures, and
3343 * forcing a Ping or Push.
3345 spin_lock(&lp->lp_lock);
3346 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3347 libcfs_nid2str(lp->lp_primary_nid), lp,
3349 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3350 rc = lnet_peer_data_present(lp);
3351 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3352 rc = lnet_peer_ping_failed(lp);
3353 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3354 rc = lnet_peer_push_failed(lp);
3355 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3356 rc = lnet_peer_send_ping(lp);
3357 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3358 rc = lnet_peer_send_push(lp);
3359 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3360 rc = lnet_peer_send_ping(lp);
3361 else if (lnet_peer_needs_push(lp))
3362 rc = lnet_peer_send_push(lp);
3364 rc = lnet_peer_discovered(lp);
3365 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3366 libcfs_nid2str(lp->lp_primary_nid), lp,
3368 spin_unlock(&lp->lp_lock);
3370 lnet_net_lock(LNET_LOCK_EX);
3371 if (rc == LNET_REDISCOVER_PEER) {
3372 list_move(&lp->lp_dc_list,
3373 &the_lnet.ln_dc_request);
3375 lnet_peer_discovery_error(lp, rc);
3377 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3378 lnet_peer_discovery_complete(lp);
3379 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3383 lnet_net_unlock(LNET_LOCK_EX);
3386 CDEBUG(D_NET, "stopping\n");
3388 * Clean up before telling lnet_peer_discovery_stop() that
3389 * we're done. Use wake_up() below to somewhat reduce the
3390 * size of the thundering herd if there are multiple threads
3391 * waiting on discovery of a single peer.
3394 /* Queue cleanup 1: stop all pending pings and pushes. */
3395 lnet_net_lock(LNET_LOCK_EX);
3396 while (!list_empty(&the_lnet.ln_dc_working)) {
3397 lp = list_first_entry(&the_lnet.ln_dc_working,
3398 struct lnet_peer, lp_dc_list);
3399 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3400 lnet_net_unlock(LNET_LOCK_EX);
3401 lnet_peer_cancel_discovery(lp);
3402 lnet_net_lock(LNET_LOCK_EX);
3404 lnet_net_unlock(LNET_LOCK_EX);
3406 /* Queue cleanup 2: wait for the expired queue to clear. */
3407 while (!list_empty(&the_lnet.ln_dc_expired))
3408 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3410 /* Queue cleanup 3: clear the request queue. */
3411 lnet_net_lock(LNET_LOCK_EX);
3412 while (!list_empty(&the_lnet.ln_dc_request)) {
3413 lp = list_first_entry(&the_lnet.ln_dc_request,
3414 struct lnet_peer, lp_dc_list);
3415 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3416 lnet_peer_discovery_complete(lp);
3418 lnet_net_unlock(LNET_LOCK_EX);
3420 LNetEQFree(the_lnet.ln_dc_eq);
3421 the_lnet.ln_dc_eq = NULL;
3423 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3424 wake_up(&the_lnet.ln_dc_waitq);
3426 CDEBUG(D_NET, "stopped\n");
3431 /* ln_api_mutex is held on entry. */
3432 int lnet_peer_discovery_start(void)
3434 struct task_struct *task;
3437 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3440 the_lnet.ln_dc_eq = LNetEQAlloc(lnet_discovery_event_handler);
3441 if (IS_ERR(the_lnet.ln_dc_eq)) {
3442 rc = PTR_ERR(the_lnet.ln_dc_eq);
3443 CERROR("Can't allocate discovery EQ: %d\n", rc);
3447 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3448 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3451 CERROR("Can't start peer discovery thread: %d\n", rc);
3453 LNetEQFree(the_lnet.ln_dc_eq);
3454 the_lnet.ln_dc_eq = NULL;
3456 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3459 CDEBUG(D_NET, "discovery start: %d\n", rc);
3464 /* ln_api_mutex is held on entry. */
3465 void lnet_peer_discovery_stop(void)
3467 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3470 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3471 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3473 /* In the LNetNIInit() path we may be stopping discovery before it
3474 * entered its work loop
3476 if (!completion_done(&the_lnet.ln_started))
3477 complete(&the_lnet.ln_started);
3479 wake_up(&the_lnet.ln_dc_waitq);
3481 wait_event(the_lnet.ln_dc_waitq,
3482 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3484 LASSERT(list_empty(&the_lnet.ln_dc_request));
3485 LASSERT(list_empty(&the_lnet.ln_dc_working));
3486 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3488 CDEBUG(D_NET, "discovery stopped\n");
3494 lnet_debug_peer(lnet_nid_t nid)
3496 char *aliveness = "NA";
3497 struct lnet_peer_ni *lp;
3500 cpt = lnet_cpt_of_nid(nid, NULL);
3503 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3505 lnet_net_unlock(cpt);
3506 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3510 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3511 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3513 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3514 libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3515 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3516 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3517 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3519 lnet_peer_ni_decref_locked(lp);
3521 lnet_net_unlock(cpt);
3524 /* Gathering information for userspace. */
3526 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3527 char aliveness[LNET_MAX_STR_LEN],
3528 __u32 *cpt_iter, __u32 *refcount,
3529 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3530 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3531 __u32 *peer_tx_qnob)
3533 struct lnet_peer_table *peer_table;
3534 struct lnet_peer_ni *lp;
3539 /* get the number of CPTs */
3540 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3542 /* if the cpt number to be examined is >= the number of cpts in
3543 * the system then indicate that there are no more cpts to examin
3545 if (*cpt_iter >= lncpt)
3548 /* get the current table */
3549 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3550 /* if the ptable is NULL then there are no more cpts to examine */
3551 if (peer_table == NULL)
3554 lnet_net_lock(*cpt_iter);
3556 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3557 struct list_head *peers = &peer_table->pt_hash[j];
3559 list_for_each_entry(lp, peers, lpni_hashlist) {
3560 if (peer_index-- > 0)
3563 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3564 if (lnet_isrouter(lp) ||
3565 lnet_peer_aliveness_enabled(lp))
3566 snprintf(aliveness, LNET_MAX_STR_LEN,
3567 lnet_is_peer_ni_alive(lp) ? "up" : "down");
3569 *nid = lp->lpni_nid;
3570 *refcount = atomic_read(&lp->lpni_refcount);
3571 *ni_peer_tx_credits =
3572 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3573 *peer_tx_credits = lp->lpni_txcredits;
3574 *peer_rtr_credits = lp->lpni_rtrcredits;
3575 *peer_min_rtr_credits = lp->lpni_mintxcredits;
3576 *peer_tx_qnob = lp->lpni_txqnob;
3582 lnet_net_unlock(*cpt_iter);
3586 return found ? 0 : -ENOENT;
3589 /* ln_api_mutex is held, which keeps the peer list stable */
3590 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3592 struct lnet_ioctl_element_stats *lpni_stats;
3593 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3594 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3595 struct lnet_peer_ni_credit_info *lpni_info;
3596 struct lnet_peer_ni *lpni;
3597 struct lnet_peer *lp;
3602 lp = lnet_find_peer(cfg->prcfg_prim_nid);
3609 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3610 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3611 size *= lp->lp_nnis;
3612 if (size > cfg->prcfg_size) {
3613 cfg->prcfg_size = size;
3618 cfg->prcfg_prim_nid = lp->lp_primary_nid;
3619 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3620 cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3621 cfg->prcfg_count = lp->lp_nnis;
3622 cfg->prcfg_size = size;
3623 cfg->prcfg_state = lp->lp_state;
3625 /* Allocate helper buffers. */
3627 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3630 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3633 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3634 if (!lpni_msg_stats)
3635 goto out_free_stats;
3636 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3638 goto out_free_msg_stats;
3643 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3644 nid = lpni->lpni_nid;
3645 if (copy_to_user(bulk, &nid, sizeof(nid)))
3646 goto out_free_hstats;
3647 bulk += sizeof(nid);
3649 memset(lpni_info, 0, sizeof(*lpni_info));
3650 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3651 if (lnet_isrouter(lpni) ||
3652 lnet_peer_aliveness_enabled(lpni))
3653 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3654 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3656 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3657 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3658 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3659 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3660 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3661 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3662 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3663 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3664 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3665 goto out_free_hstats;
3666 bulk += sizeof(*lpni_info);
3668 memset(lpni_stats, 0, sizeof(*lpni_stats));
3669 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3670 LNET_STATS_TYPE_SEND);
3671 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3672 LNET_STATS_TYPE_RECV);
3673 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3674 LNET_STATS_TYPE_DROP);
3675 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3676 goto out_free_hstats;
3677 bulk += sizeof(*lpni_stats);
3678 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3679 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3680 goto out_free_hstats;
3681 bulk += sizeof(*lpni_msg_stats);
3682 lpni_hstats->hlpni_network_timeout =
3683 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3684 lpni_hstats->hlpni_remote_dropped =
3685 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3686 lpni_hstats->hlpni_remote_timeout =
3687 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3688 lpni_hstats->hlpni_remote_error =
3689 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3690 lpni_hstats->hlpni_health_value =
3691 atomic_read(&lpni->lpni_healthv);
3692 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3693 goto out_free_hstats;
3694 bulk += sizeof(*lpni_hstats);
3699 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3701 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3703 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3705 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3707 lnet_peer_decref_locked(lp);
3713 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3715 /* the mt could've shutdown and cleaned up the queues */
3716 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3719 if (list_empty(&lpni->lpni_recovery) &&
3720 atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3721 CDEBUG(D_NET, "lpni %s added to recovery queue. Health = %d\n",
3722 libcfs_nid2str(lpni->lpni_nid),
3723 atomic_read(&lpni->lpni_healthv));
3724 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3725 lnet_peer_ni_addref_locked(lpni);
3729 /* Call with the ln_api_mutex held */
3731 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3733 struct lnet_peer_table *ptable;
3734 struct lnet_peer *lp;
3735 struct lnet_peer_net *lpn;
3736 struct lnet_peer_ni *lpni;
3740 if (the_lnet.ln_state != LNET_STATE_RUNNING)
3744 lnet_net_lock(LNET_LOCK_EX);
3745 lpni = lnet_find_peer_ni_locked(nid);
3747 lnet_net_unlock(LNET_LOCK_EX);
3750 atomic_set(&lpni->lpni_healthv, value);
3751 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3752 lnet_peer_ni_decref_locked(lpni);
3753 lnet_net_unlock(LNET_LOCK_EX);
3757 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3760 * Walk all the peers and reset the healhv for each one to the
3763 lnet_net_lock(LNET_LOCK_EX);
3764 for (cpt = 0; cpt < lncpt; cpt++) {
3765 ptable = the_lnet.ln_peer_tables[cpt];
3766 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3767 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3768 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3770 atomic_set(&lpni->lpni_healthv, value);
3771 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3776 lnet_net_unlock(LNET_LOCK_EX);