4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
34 #define DEBUG_SUBSYSTEM S_LNET
36 #include <linux/sched.h>
37 #ifdef HAVE_SCHED_HEADERS
38 #include <linux/sched/signal.h>
40 #include <linux/uaccess.h>
42 #include <lnet/udsp.h>
43 #include <lnet/lib-lnet.h>
44 #include <uapi/linux/lnet/lnet-dlc.h>
46 /* Value indicating that recovery needs to re-check a peer immediately. */
47 #define LNET_REDISCOVER_PEER (1)
49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
54 if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
55 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
56 lnet_peer_ni_decref_locked(lpni);
61 lnet_peer_net_added(struct lnet_net *net)
63 struct lnet_peer_ni *lpni, *tmp;
65 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
66 lpni_on_remote_peer_ni_list) {
68 if (LNET_NID_NET(&lpni->lpni_nid) == net->net_id) {
71 spin_lock(&lpni->lpni_lock);
72 lpni->lpni_txcredits =
73 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
74 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
75 lpni->lpni_rtrcredits =
76 lnet_peer_buffer_credits(lpni->lpni_net);
77 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
78 spin_unlock(&lpni->lpni_lock);
80 lnet_peer_remove_from_remote_list(lpni);
86 lnet_peer_tables_destroy(void)
88 struct lnet_peer_table *ptable;
89 struct list_head *hash;
93 if (!the_lnet.ln_peer_tables)
96 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
97 hash = ptable->pt_hash;
98 if (!hash) /* not intialized */
101 LASSERT(list_empty(&ptable->pt_zombie_list));
103 ptable->pt_hash = NULL;
104 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
105 LASSERT(list_empty(&hash[j]));
107 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
110 cfs_percpt_free(the_lnet.ln_peer_tables);
111 the_lnet.ln_peer_tables = NULL;
115 lnet_peer_tables_create(void)
117 struct lnet_peer_table *ptable;
118 struct list_head *hash;
122 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
124 if (the_lnet.ln_peer_tables == NULL) {
125 CERROR("Failed to allocate cpu-partition peer tables\n");
129 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131 LNET_PEER_HASH_SIZE * sizeof(*hash));
133 CERROR("Failed to create peer hash table\n");
134 lnet_peer_tables_destroy();
138 spin_lock_init(&ptable->pt_zombie_lock);
139 INIT_LIST_HEAD(&ptable->pt_zombie_list);
141 INIT_LIST_HEAD(&ptable->pt_peer_list);
143 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144 INIT_LIST_HEAD(&hash[j]);
145 ptable->pt_hash = hash; /* sign of initialization */
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(struct lnet_nid *nid)
154 struct lnet_peer_ni *lpni;
155 struct lnet_net *net;
158 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
160 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
164 INIT_LIST_HEAD(&lpni->lpni_txq);
165 INIT_LIST_HEAD(&lpni->lpni_hashlist);
166 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167 INIT_LIST_HEAD(&lpni->lpni_recovery);
168 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169 INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
170 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171 kref_init(&lpni->lpni_kref);
172 lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
174 spin_lock_init(&lpni->lpni_lock);
176 if (lnet_peers_start_down())
177 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
179 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
180 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
181 lpni->lpni_nid = *nid;
182 lpni->lpni_cpt = cpt;
183 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
185 net = lnet_get_net_locked(LNET_NID_NET(nid));
186 lpni->lpni_net = net;
188 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
189 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
190 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
191 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
194 * This peer_ni is not on a local network, so we
195 * cannot add the credits here. In case the net is
196 * added later, add the peer_ni to the remote peer ni
197 * list so it can be easily found and revisited.
199 /* FIXME: per-net implementation instead? */
200 lnet_peer_ni_addref_locked(lpni);
201 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
202 &the_lnet.ln_remote_peer_ni_list);
205 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
210 static struct lnet_peer_net *
211 lnet_peer_net_alloc(__u32 net_id)
213 struct lnet_peer_net *lpn;
215 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
219 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
220 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
221 lpn->lpn_net_id = net_id;
222 lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
224 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
230 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
232 struct lnet_peer *lp;
234 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
236 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
237 LASSERT(list_empty(&lpn->lpn_peer_nis));
238 LASSERT(list_empty(&lpn->lpn_peer_nets));
240 lpn->lpn_peer = NULL;
241 LIBCFS_FREE(lpn, sizeof(*lpn));
243 lnet_peer_decref_locked(lp);
246 static struct lnet_peer *
247 lnet_peer_alloc(struct lnet_nid *nid)
249 struct lnet_peer *lp;
251 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
255 INIT_LIST_HEAD(&lp->lp_rtrq);
256 INIT_LIST_HEAD(&lp->lp_routes);
257 INIT_LIST_HEAD(&lp->lp_peer_list);
258 INIT_LIST_HEAD(&lp->lp_peer_nets);
259 INIT_LIST_HEAD(&lp->lp_dc_list);
260 INIT_LIST_HEAD(&lp->lp_dc_pendq);
261 INIT_LIST_HEAD(&lp->lp_rtr_list);
262 init_waitqueue_head(&lp->lp_dc_waitq);
263 spin_lock_init(&lp->lp_lock);
264 lp->lp_primary_nid = *nid;
265 lp->lp_disc_src_nid = LNET_ANY_NID;
266 lp->lp_disc_dst_nid = LNET_ANY_NID;
267 if (lnet_peers_start_down())
268 lp->lp_alive = false;
273 * all peers created on a router should have health on
274 * if it's not already on.
276 if (the_lnet.ln_routing && !lnet_health_sensitivity)
277 lp->lp_health_sensitivity = 1;
280 * Turn off discovery for loopback peer. If you're creating a peer
281 * for the loopback interface then that was initiated when we
282 * attempted to send a message over the loopback. There is no need
283 * to ever use a different interface when sending messages to
287 lp->lp_state = LNET_PEER_NO_DISCOVERY;
288 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
290 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
296 lnet_destroy_peer_locked(struct lnet_peer *lp)
298 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
300 LASSERT(atomic_read(&lp->lp_refcount) == 0);
301 LASSERT(lp->lp_rtr_refcount == 0);
302 LASSERT(list_empty(&lp->lp_peer_nets));
303 LASSERT(list_empty(&lp->lp_peer_list));
304 LASSERT(list_empty(&lp->lp_dc_list));
307 lnet_ping_buffer_decref(lp->lp_data);
310 * if there are messages still on the pending queue, then make
311 * sure to queue them on the ln_msg_resend list so they can be
312 * resent at a later point if the discovery thread is still
314 * If the discovery thread has stopped, then the wakeup will be a
315 * no-op, and it is expected the lnet_shutdown_lndnets() will
316 * eventually be called, which will traverse this list and
317 * finalize the messages on the list.
318 * We can not resend them now because we're holding the cpt lock.
319 * Releasing the lock can cause an inconsistent state
321 spin_lock(&the_lnet.ln_msg_resend_lock);
322 spin_lock(&lp->lp_lock);
323 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
324 spin_unlock(&lp->lp_lock);
325 spin_unlock(&the_lnet.ln_msg_resend_lock);
326 wake_up(&the_lnet.ln_dc_waitq);
328 LIBCFS_FREE(lp, sizeof(*lp));
332 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
333 * that peer_net, detach the peer_net from the peer.
335 * Call with lnet_net_lock/EX held
338 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
340 struct lnet_peer_table *ptable;
341 struct lnet_peer_net *lpn;
342 struct lnet_peer *lp;
345 * Belts and suspenders: gracefully handle teardown of a
346 * partially connected peer_ni.
348 lpn = lpni->lpni_peer_net;
350 list_del_init(&lpni->lpni_peer_nis);
352 * If there are no lpni's left, we detach lpn from
353 * lp_peer_nets, so it cannot be found anymore.
355 if (list_empty(&lpn->lpn_peer_nis))
356 list_del_init(&lpn->lpn_peer_nets);
358 /* Update peer NID count. */
363 * If there are no more peer nets, make the peer unfindable
364 * via the peer_tables.
366 * Otherwise, if the peer is DISCOVERED, tell discovery to
367 * take another look at it. This is a no-op if discovery for
368 * this peer did the detaching.
370 if (list_empty(&lp->lp_peer_nets)) {
371 list_del_init(&lp->lp_peer_list);
372 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
374 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
375 /* Discovery isn't running, nothing to do here. */
376 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
377 lnet_peer_queue_for_discovery(lp);
378 wake_up(&the_lnet.ln_dc_waitq);
380 CDEBUG(D_NET, "peer %s NID %s\n",
381 libcfs_nidstr(&lp->lp_primary_nid),
382 libcfs_nidstr(&lpni->lpni_nid));
385 /* called with lnet_net_lock LNET_LOCK_EX held */
387 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
389 struct lnet_peer_table *ptable = NULL;
391 /* don't remove a peer_ni if it's also a gateway */
392 if (lnet_isrouter(lpni) && !force) {
393 CERROR("Peer NI %s is a gateway. Can not delete it\n",
394 libcfs_nidstr(&lpni->lpni_nid));
398 lnet_peer_remove_from_remote_list(lpni);
400 /* remove peer ni from the hash list. */
401 list_del_init(&lpni->lpni_hashlist);
404 * indicate the peer is being deleted so the monitor thread can
405 * remove it from the recovery queue.
407 spin_lock(&lpni->lpni_lock);
408 lpni->lpni_state |= LNET_PEER_NI_DELETING;
409 spin_unlock(&lpni->lpni_lock);
411 /* decrement the ref count on the peer table */
412 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
415 * The peer_ni can no longer be found with a lookup. But there
416 * can be current users, so keep track of it on the zombie
417 * list until the reference count has gone to zero.
419 * The last reference may be lost in a place where the
420 * lnet_net_lock locks only a single cpt, and that cpt may not
421 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
424 spin_lock(&ptable->pt_zombie_lock);
425 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
426 ptable->pt_zombies++;
427 spin_unlock(&ptable->pt_zombie_lock);
429 /* no need to keep this peer_ni on the hierarchy anymore */
430 lnet_peer_detach_peer_ni_locked(lpni);
432 /* remove hashlist reference on peer_ni */
433 lnet_peer_ni_decref_locked(lpni);
438 void lnet_peer_uninit(void)
440 struct lnet_peer_ni *lpni, *tmp;
442 lnet_net_lock(LNET_LOCK_EX);
444 /* remove all peer_nis from the remote peer and the hash list */
445 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
446 lpni_on_remote_peer_ni_list)
447 lnet_peer_ni_del_locked(lpni, false);
449 lnet_peer_tables_destroy();
451 lnet_net_unlock(LNET_LOCK_EX);
455 lnet_peer_del_locked(struct lnet_peer *peer)
457 struct lnet_peer_ni *lpni = NULL, *lpni2;
460 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&peer->lp_primary_nid));
462 spin_lock(&peer->lp_lock);
463 peer->lp_state |= LNET_PEER_MARK_DELETED;
464 spin_unlock(&peer->lp_lock);
466 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
467 while (lpni != NULL) {
468 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
469 rc = lnet_peer_ni_del_locked(lpni, false);
479 * Discovering this peer is taking too long. Cancel any Ping or Push
480 * that discovery is waiting on by unlinking the relevant MDs. The
481 * lnet_discovery_event_handler() will proceed from here and complete
484 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
486 struct lnet_handle_md ping_mdh;
487 struct lnet_handle_md push_mdh;
489 LNetInvalidateMDHandle(&ping_mdh);
490 LNetInvalidateMDHandle(&push_mdh);
492 spin_lock(&lp->lp_lock);
493 if (lp->lp_state & LNET_PEER_PING_SENT) {
494 ping_mdh = lp->lp_ping_mdh;
495 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
497 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
498 push_mdh = lp->lp_push_mdh;
499 LNetInvalidateMDHandle(&lp->lp_push_mdh);
501 spin_unlock(&lp->lp_lock);
503 if (!LNetMDHandleIsInvalid(ping_mdh))
504 LNetMDUnlink(ping_mdh);
505 if (!LNetMDHandleIsInvalid(push_mdh))
506 LNetMDUnlink(push_mdh);
510 lnet_peer_del(struct lnet_peer *peer)
514 lnet_peer_cancel_discovery(peer);
515 lnet_net_lock(LNET_LOCK_EX);
516 rc = lnet_peer_del_locked(peer);
517 lnet_net_unlock(LNET_LOCK_EX);
523 * Delete a NID from a peer. Call with ln_api_mutex held.
526 * -EPERM: Non-DLC deletion from DLC-configured peer.
527 * -ENOENT: No lnet_peer_ni corresponding to the nid.
528 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
529 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
532 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
534 struct lnet_peer_ni *lpni;
535 struct lnet_nid primary_nid = lp->lp_primary_nid;
538 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
540 lnet_nid4_to_nid(nid4, &nid);
541 if (!(flags & LNET_PEER_CONFIGURED)) {
542 if (lp->lp_state & LNET_PEER_CONFIGURED) {
548 lpni = lnet_peer_ni_find_locked(&nid);
553 lnet_peer_ni_decref_locked(lpni);
554 if (lp != lpni->lpni_peer_net->lpn_peer) {
560 * This function only allows deletion of the primary NID if it
563 if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && !force) {
568 lnet_net_lock(LNET_LOCK_EX);
570 if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && force) {
571 struct lnet_peer_ni *lpni2;
572 /* assign the next peer_ni to be the primary */
573 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
575 lp->lp_primary_nid = lpni2->lpni_nid;
577 rc = lnet_peer_ni_del_locked(lpni, force);
579 lnet_net_unlock(LNET_LOCK_EX);
582 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
583 libcfs_nidstr(&primary_nid), libcfs_nidstr(&nid),
590 lnet_peer_table_cleanup_locked(struct lnet_net *net,
591 struct lnet_peer_table *ptable)
594 struct lnet_peer_ni *next;
595 struct lnet_peer_ni *lpni;
596 struct lnet_peer *peer;
598 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
599 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
601 if (net != NULL && net != lpni->lpni_net)
604 peer = lpni->lpni_peer_net->lpn_peer;
605 if (!nid_same(&peer->lp_primary_nid,
607 lnet_peer_ni_del_locked(lpni, false);
611 * Removing the primary NID implies removing
612 * the entire peer. Advance next beyond any
613 * peer_ni that belongs to the same peer.
615 list_for_each_entry_from(next, &ptable->pt_hash[i],
617 if (next->lpni_peer_net->lpn_peer != peer)
620 lnet_peer_del_locked(peer);
626 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
628 wait_var_event_warning(&ptable->pt_zombies,
629 ptable->pt_zombies == 0,
630 "Waiting for %d zombies on peer table\n",
635 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
636 struct lnet_peer_table *ptable)
638 struct lnet_peer_ni *lp;
639 struct lnet_peer_ni *tmp;
640 struct lnet_nid gw_nid;
643 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
644 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
646 if (net != lp->lpni_net)
649 if (!lnet_isrouter(lp))
652 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
654 lnet_net_unlock(LNET_LOCK_EX);
655 lnet_del_route(LNET_NET_ANY, &gw_nid);
656 lnet_net_lock(LNET_LOCK_EX);
662 lnet_peer_tables_cleanup(struct lnet_net *net)
665 struct lnet_peer_table *ptable;
667 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
668 /* If just deleting the peers for a NI, get rid of any routes these
669 * peers are gateways for. */
670 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
671 lnet_net_lock(LNET_LOCK_EX);
672 lnet_peer_table_del_rtrs_locked(net, ptable);
673 lnet_net_unlock(LNET_LOCK_EX);
676 /* Start the cleanup process */
677 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
678 lnet_net_lock(LNET_LOCK_EX);
679 lnet_peer_table_cleanup_locked(net, ptable);
680 lnet_net_unlock(LNET_LOCK_EX);
683 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
684 lnet_peer_ni_finalize_wait(ptable);
687 static struct lnet_peer_ni *
688 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, struct lnet_nid *nid)
690 struct list_head *peers;
691 struct lnet_peer_ni *lp;
693 if (the_lnet.ln_state != LNET_STATE_RUNNING)
696 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
697 list_for_each_entry(lp, peers, lpni_hashlist) {
698 if (nid_same(&lp->lpni_nid, nid)) {
699 lnet_peer_ni_addref_locked(lp);
707 struct lnet_peer_ni *
708 lnet_find_peer_ni_locked(lnet_nid_t nid4)
710 struct lnet_peer_ni *lpni;
711 struct lnet_peer_table *ptable;
715 lnet_nid4_to_nid(nid4, &nid);
717 cpt = lnet_nid_cpt_hash(&nid, LNET_CPT_NUMBER);
719 ptable = the_lnet.ln_peer_tables[cpt];
720 lpni = lnet_get_peer_ni_locked(ptable, &nid);
725 struct lnet_peer_ni *
726 lnet_peer_ni_find_locked(struct lnet_nid *nid)
728 struct lnet_peer_ni *lpni;
729 struct lnet_peer_table *ptable;
732 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
734 ptable = the_lnet.ln_peer_tables[cpt];
735 lpni = lnet_get_peer_ni_locked(ptable, nid);
740 struct lnet_peer_ni *
741 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
743 struct lnet_peer_net *lpn;
744 struct lnet_peer_ni *lpni;
746 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
750 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
751 if (lnet_nid_to_nid4(&lpni->lpni_nid) == nid)
758 struct lnet_peer_ni *
759 lnet_peer_ni_get_locked(struct lnet_peer *lp, struct lnet_nid *nid)
761 struct lnet_peer_net *lpn;
762 struct lnet_peer_ni *lpni;
764 lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
768 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
769 if (nid_same(&lpni->lpni_nid, nid))
777 lnet_find_peer4(lnet_nid_t nid)
779 struct lnet_peer_ni *lpni;
780 struct lnet_peer *lp = NULL;
783 cpt = lnet_net_lock_current();
784 lpni = lnet_find_peer_ni_locked(nid);
786 lp = lpni->lpni_peer_net->lpn_peer;
787 lnet_peer_addref_locked(lp);
788 lnet_peer_ni_decref_locked(lpni);
790 lnet_net_unlock(cpt);
796 lnet_find_peer(struct lnet_nid *nid)
798 struct lnet_peer_ni *lpni;
799 struct lnet_peer *lp = NULL;
802 cpt = lnet_net_lock_current();
803 lpni = lnet_peer_ni_find_locked(nid);
805 lp = lpni->lpni_peer_net->lpn_peer;
806 lnet_peer_addref_locked(lp);
807 lnet_peer_ni_decref_locked(lpni);
809 lnet_net_unlock(cpt);
814 struct lnet_peer_net *
815 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
817 struct lnet_peer_net *net;
820 /* no net id provided return the first net */
821 net = list_first_entry_or_null(&lp->lp_peer_nets,
822 struct lnet_peer_net,
828 /* find the net after the one provided */
829 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
830 if (net->lpn_net_id == prev_lpn_id) {
832 * if we reached the end of the list loop to the
835 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
836 return list_first_entry_or_null(&lp->lp_peer_nets,
837 struct lnet_peer_net,
840 return list_next_entry(net, lpn_peer_nets);
847 struct lnet_peer_ni *
848 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
849 struct lnet_peer_net *peer_net,
850 struct lnet_peer_ni *prev)
852 struct lnet_peer_ni *lpni;
853 struct lnet_peer_net *net = peer_net;
857 if (list_empty(&peer->lp_peer_nets))
860 net = list_entry(peer->lp_peer_nets.next,
861 struct lnet_peer_net,
864 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
870 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
872 * if you reached the end of the peer ni list and the peer
873 * net is specified then there are no more peer nis in that
880 * we reached the end of this net ni list. move to the
883 if (prev->lpni_peer_net->lpn_peer_nets.next ==
885 /* no more nets and no more NIs. */
888 /* get the next net */
889 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
890 struct lnet_peer_net,
892 /* get the ni on it */
893 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
899 /* there are more nis left */
900 lpni = list_entry(prev->lpni_peer_nis.next,
901 struct lnet_peer_ni, lpni_peer_nis);
906 /* Call with the ln_api_mutex held */
907 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
909 struct lnet_process_id id;
910 struct lnet_peer_table *ptable;
911 struct lnet_peer *lp;
920 if (the_lnet.ln_state != LNET_STATE_RUNNING)
923 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
926 * Count the number of peers, and return E2BIG if the buffer
927 * is too small. We'll also return the desired size.
930 for (cpt = 0; cpt < lncpt; cpt++) {
931 ptable = the_lnet.ln_peer_tables[cpt];
932 count += ptable->pt_peers;
934 size = count * sizeof(*ids);
939 * Walk the peer lists and copy out the primary nids.
940 * This is safe because the peer lists are only modified
941 * while the ln_api_mutex is held. So we don't need to
942 * hold the lnet_net_lock as well, and can therefore
943 * directly call copy_to_user().
946 memset(&id, 0, sizeof(id));
947 id.pid = LNET_PID_LUSTRE;
949 for (cpt = 0; cpt < lncpt; cpt++) {
950 ptable = the_lnet.ln_peer_tables[cpt];
951 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
952 if (!nid_is_nid4(&lp->lp_primary_nid))
956 id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
957 if (copy_to_user(&ids[i], &id, sizeof(id)))
970 * Start pushes to peers that need to be updated for a configuration
971 * change on this node.
974 lnet_push_update_to_peers(int force)
976 struct lnet_peer_table *ptable;
977 struct lnet_peer *lp;
981 lnet_net_lock(LNET_LOCK_EX);
982 if (lnet_peer_discovery_disabled)
984 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
985 for (cpt = 0; cpt < lncpt; cpt++) {
986 ptable = the_lnet.ln_peer_tables[cpt];
987 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
989 spin_lock(&lp->lp_lock);
990 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
991 lp->lp_state |= LNET_PEER_FORCE_PUSH;
992 spin_unlock(&lp->lp_lock);
994 if (lnet_peer_needs_push(lp))
995 lnet_peer_queue_for_discovery(lp);
998 lnet_net_unlock(LNET_LOCK_EX);
999 wake_up(&the_lnet.ln_dc_waitq);
1002 /* find the NID in the preferred gateways for the remote peer
1004 * false: list is not empty and NID is not preferred
1005 * false: list is empty
1006 * true: nid is found in the list
1009 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
1010 struct lnet_nid *gw_nid)
1012 struct lnet_nid_list *ne;
1014 CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
1015 libcfs_nidstr(&lpni->lpni_nid),
1016 list_empty(&lpni->lpni_rtr_pref_nids));
1018 if (list_empty(&lpni->lpni_rtr_pref_nids))
1021 /* iterate through all the preferred NIDs and see if any of them
1022 * matches the provided gw_nid
1024 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1025 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
1026 libcfs_nidstr(&ne->nl_nid),
1027 libcfs_nidstr(gw_nid));
1028 if (nid_same(&ne->nl_nid, gw_nid))
1036 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
1038 struct list_head zombies;
1039 struct lnet_nid_list *ne;
1040 struct lnet_nid_list *tmp;
1041 int cpt = lpni->lpni_cpt;
1043 INIT_LIST_HEAD(&zombies);
1046 list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
1047 lnet_net_unlock(cpt);
1049 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1050 list_del(&ne->nl_list);
1051 LIBCFS_FREE(ne, sizeof(*ne));
1056 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
1057 struct lnet_nid *gw_nid)
1059 int cpt = lpni->lpni_cpt;
1060 struct lnet_nid_list *ne = NULL;
1062 /* This function is called with api_mutex held. When the api_mutex
1063 * is held the list can not be modified, as it is only modified as
1064 * a result of applying a UDSP and that happens under api_mutex
1067 __must_hold(&the_lnet.ln_api_mutex);
1069 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1070 if (nid_same(&ne->nl_nid, gw_nid))
1074 LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1078 ne->nl_nid = *gw_nid;
1080 /* Lock the cpt to protect against addition and checks in the
1081 * selection algorithm
1084 list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1085 lnet_net_unlock(cpt);
1091 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1092 * this is a preferred point-to-point path. Call with lnet_net_lock in
1096 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1098 struct lnet_nid_list *ne;
1100 if (lpni->lpni_pref_nnids == 0)
1102 if (lpni->lpni_pref_nnids == 1)
1103 return nid_same(&lpni->lpni_pref.nid, nid);
1104 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1105 if (nid_same(&ne->nl_nid, nid))
1112 * Set a single ni as preferred, provided no preferred ni is already
1113 * defined. Only to be used for non-multi-rail peer_ni.
1116 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni,
1117 struct lnet_nid *nid)
1123 spin_lock(&lpni->lpni_lock);
1124 if (LNET_NID_IS_ANY(nid)) {
1126 } else if (lpni->lpni_pref_nnids > 0) {
1128 } else if (lpni->lpni_pref_nnids == 0) {
1129 lpni->lpni_pref.nid = *nid;
1130 lpni->lpni_pref_nnids = 1;
1131 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1133 spin_unlock(&lpni->lpni_lock);
1135 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1136 libcfs_nidstr(&lpni->lpni_nid), libcfs_nidstr(nid), rc);
1141 * Clear the preferred NID from a non-multi-rail peer_ni, provided
1142 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1145 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1149 spin_lock(&lpni->lpni_lock);
1150 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1151 lpni->lpni_pref_nnids = 0;
1152 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1153 } else if (lpni->lpni_pref_nnids == 0) {
1158 spin_unlock(&lpni->lpni_lock);
1160 CDEBUG(D_NET, "peer %s: %d\n",
1161 libcfs_nidstr(&lpni->lpni_nid), rc);
1166 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1168 lpni->lpni_sel_priority = priority;
1172 * Clear the preferred NIDs from a non-multi-rail peer.
1175 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1177 struct lnet_peer_ni *lpni = NULL;
1179 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1180 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1184 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1186 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1187 struct lnet_nid_list *ne1 = NULL;
1188 struct lnet_nid_list *ne2 = NULL;
1189 struct lnet_nid *tmp_nid = NULL;
1192 if (LNET_NID_IS_ANY(nid)) {
1197 if (lpni->lpni_pref_nnids == 1 &&
1198 nid_same(&lpni->lpni_pref.nid, nid)) {
1203 /* A non-MR node may have only one preferred NI per peer_ni */
1204 if (lpni->lpni_pref_nnids > 0 &&
1205 !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1210 /* add the new preferred nid to the list of preferred nids */
1211 if (lpni->lpni_pref_nnids != 0) {
1212 size_t alloc_size = sizeof(*ne1);
1214 if (lpni->lpni_pref_nnids == 1) {
1215 tmp_nid = &lpni->lpni_pref.nid;
1216 INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1219 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1220 if (nid_same(&ne1->nl_nid, nid)) {
1226 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1233 /* move the originally stored nid to the list */
1234 if (lpni->lpni_pref_nnids == 1) {
1235 LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1236 lpni->lpni_cpt, alloc_size);
1241 INIT_LIST_HEAD(&ne2->nl_list);
1242 ne2->nl_nid = *tmp_nid;
1247 lnet_net_lock(LNET_LOCK_EX);
1248 spin_lock(&lpni->lpni_lock);
1249 if (lpni->lpni_pref_nnids == 0) {
1250 lpni->lpni_pref.nid = *nid;
1253 list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1254 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1256 lpni->lpni_pref_nnids++;
1257 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1258 spin_unlock(&lpni->lpni_lock);
1259 lnet_net_unlock(LNET_LOCK_EX);
1262 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1263 spin_lock(&lpni->lpni_lock);
1264 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1265 spin_unlock(&lpni->lpni_lock);
1267 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1268 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1273 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1275 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1276 struct lnet_nid_list *ne = NULL;
1279 if (lpni->lpni_pref_nnids == 0) {
1284 if (lpni->lpni_pref_nnids == 1) {
1285 if (!nid_same(&lpni->lpni_pref.nid, nid)) {
1290 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1291 if (nid_same(&ne->nl_nid, nid))
1292 goto remove_nid_entry;
1300 lnet_net_lock(LNET_LOCK_EX);
1301 spin_lock(&lpni->lpni_lock);
1302 if (lpni->lpni_pref_nnids == 1)
1303 lpni->lpni_pref.nid = LNET_ANY_NID;
1305 list_del_init(&ne->nl_list);
1306 if (lpni->lpni_pref_nnids == 2) {
1307 struct lnet_nid_list *ne, *tmp;
1309 list_for_each_entry_safe(ne, tmp,
1310 &lpni->lpni_pref.nids,
1312 lpni->lpni_pref.nid = ne->nl_nid;
1313 list_del_init(&ne->nl_list);
1314 LIBCFS_FREE(ne, sizeof(*ne));
1318 lpni->lpni_pref_nnids--;
1319 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1320 spin_unlock(&lpni->lpni_lock);
1321 lnet_net_unlock(LNET_LOCK_EX);
1324 LIBCFS_FREE(ne, sizeof(*ne));
1326 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1327 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1332 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1334 struct list_head zombies;
1335 struct lnet_nid_list *ne;
1336 struct lnet_nid_list *tmp;
1338 INIT_LIST_HEAD(&zombies);
1340 lnet_net_lock(LNET_LOCK_EX);
1341 if (lpni->lpni_pref_nnids == 1)
1342 lpni->lpni_pref.nid = LNET_ANY_NID;
1343 else if (lpni->lpni_pref_nnids > 1)
1344 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1345 lpni->lpni_pref_nnids = 0;
1346 lnet_net_unlock(LNET_LOCK_EX);
1348 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1349 list_del_init(&ne->nl_list);
1350 LIBCFS_FREE(ne, sizeof(*ne));
1355 lnet_peer_primary_nid_locked(struct lnet_nid *nid, struct lnet_nid *result)
1357 struct lnet_peer_ni *lpni;
1360 lpni = lnet_peer_ni_find_locked(nid);
1362 *result = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1363 lnet_peer_ni_decref_locked(lpni);
1368 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1369 __must_hold(&lp->lp_lock)
1371 if (lnet_peer_discovery_disabled)
1374 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1375 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1386 lnet_is_discovery_disabled(struct lnet_peer *lp)
1390 spin_lock(&lp->lp_lock);
1391 rc = lnet_is_discovery_disabled_locked(lp);
1392 spin_unlock(&lp->lp_lock);
1398 LNetAddPeer(lnet_nid_t *nids, __u32 num_nids)
1400 lnet_nid_t pnid = 0;
1404 if (!nids || num_nids < 1)
1407 rc = LNetNIInit(LNET_PID_ANY);
1411 mutex_lock(&the_lnet.ln_api_mutex);
1413 mr = lnet_peer_discovery_disabled == 0;
1416 for (i = 0; i < num_nids; i++) {
1417 if (nids[i] == LNET_NID_LO_0)
1422 rc = lnet_add_peer_ni(pnid, LNET_NID_ANY, mr, true);
1423 } else if (lnet_peer_discovery_disabled) {
1424 rc = lnet_add_peer_ni(nids[i], LNET_NID_ANY, mr, true);
1426 rc = lnet_add_peer_ni(pnid, nids[i], mr, true);
1429 if (rc && rc != -EEXIST)
1434 mutex_unlock(&the_lnet.ln_api_mutex);
1438 return rc == -EEXIST ? 0 : rc;
1440 EXPORT_SYMBOL(LNetAddPeer);
1442 /* FIXME support large-addr nid */
1444 LNetPrimaryNID(lnet_nid_t nid)
1446 struct lnet_peer *lp;
1447 struct lnet_peer_ni *lpni;
1448 lnet_nid_t primary_nid = nid;
1452 if (nid == LNET_NID_LO_0)
1453 return LNET_NID_LO_0;
1455 cpt = lnet_net_lock_current();
1456 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1461 lp = lpni->lpni_peer_net->lpn_peer;
1463 /* If discovery is disabled locally then we needn't bother running
1464 * discovery here because discovery will not modify whatever
1465 * primary NID is currently set for this peer. If the specified peer is
1466 * down then this discovery can introduce long delays into the mount
1467 * process, so skip it if it isn't necessary.
1469 while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1470 spin_lock(&lp->lp_lock);
1471 /* force a full discovery cycle */
1472 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1473 spin_unlock(&lp->lp_lock);
1475 rc = lnet_discover_peer_locked(lpni, cpt, true);
1478 /* The lpni (or lp) for this NID may have changed and our ref is
1479 * the only thing keeping the old one around. Release the ref
1480 * and lookup the lpni again
1482 lnet_peer_ni_decref_locked(lpni);
1483 lpni = lnet_find_peer_ni_locked(nid);
1488 lp = lpni->lpni_peer_net->lpn_peer;
1490 /* If we find that the peer has discovery disabled then we will
1491 * not modify whatever primary NID is currently set for this
1492 * peer. Thus, we can break out of this loop even if the peer
1493 * is not fully up to date.
1495 if (lnet_is_discovery_disabled(lp))
1498 primary_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
1500 lnet_peer_ni_decref_locked(lpni);
1502 lnet_net_unlock(cpt);
1504 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1505 libcfs_nid2str(primary_nid), rc);
1508 EXPORT_SYMBOL(LNetPrimaryNID);
1510 struct lnet_peer_net *
1511 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1513 struct lnet_peer_net *peer_net;
1514 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1515 if (peer_net->lpn_net_id == net_id)
1522 * Attach a peer_ni to a peer_net and peer. This function assumes
1523 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1524 * may be attached to a different peer, in which case it will be
1525 * properly detached first. The whole operation is done atomically.
1527 * This function consumes the reference on lpni and Always returns 0.
1528 * This is the last function called from functions that do return an
1529 * int, so returning 0 here allows the compiler to do a tail call.
1532 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1533 struct lnet_peer_net *lpn,
1534 struct lnet_peer_ni *lpni,
1537 struct lnet_peer_table *ptable;
1538 bool new_lpn = false;
1541 /* Install the new peer_ni */
1542 lnet_net_lock(LNET_LOCK_EX);
1543 /* Add peer_ni to global peer table hash, if necessary. */
1544 if (list_empty(&lpni->lpni_hashlist)) {
1545 int hash = lnet_nid2peerhash(&lpni->lpni_nid);
1547 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1548 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1549 ptable->pt_version++;
1550 lnet_peer_ni_addref_locked(lpni);
1553 /* Detach the peer_ni from an existing peer, if necessary. */
1554 if (lpni->lpni_peer_net) {
1555 LASSERT(lpni->lpni_peer_net != lpn);
1556 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1557 lnet_peer_detach_peer_ni_locked(lpni);
1558 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1559 lpni->lpni_peer_net = NULL;
1562 /* Add peer_ni to peer_net */
1563 lpni->lpni_peer_net = lpn;
1564 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1565 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1567 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1568 lnet_update_peer_net_healthv(lpni);
1569 lnet_peer_net_addref_locked(lpn);
1571 /* Add peer_net to peer */
1572 if (!lpn->lpn_peer) {
1575 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1576 list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1578 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1579 lnet_peer_addref_locked(lp);
1582 /* Add peer to global peer list, if necessary */
1583 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1584 if (list_empty(&lp->lp_peer_list)) {
1585 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1590 /* Update peer state */
1591 spin_lock(&lp->lp_lock);
1592 if (flags & LNET_PEER_CONFIGURED) {
1593 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1594 lp->lp_state |= LNET_PEER_CONFIGURED;
1596 if (flags & LNET_PEER_MULTI_RAIL) {
1597 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1598 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1599 lnet_peer_clr_non_mr_pref_nids(lp);
1602 spin_unlock(&lp->lp_lock);
1608 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1610 CERROR("Failed to apply UDSPs on lpn %s\n",
1611 libcfs_net2str(lpn->lpn_net_id));
1613 rc = lnet_udsp_apply_policies_on_lpni(lpni);
1615 CERROR("Failed to apply UDSPs on lpni %s\n",
1616 libcfs_nidstr(&lpni->lpni_nid));
1618 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1619 libcfs_nidstr(&lp->lp_primary_nid),
1620 libcfs_nidstr(&lpni->lpni_nid), flags);
1621 lnet_peer_ni_decref_locked(lpni);
1622 lnet_net_unlock(LNET_LOCK_EX);
1628 * Create a new peer, with nid as its primary nid.
1630 * Call with the lnet_api_mutex held.
1633 lnet_peer_add(lnet_nid_t nid4, unsigned int flags)
1635 struct lnet_nid nid;
1636 struct lnet_peer *lp;
1637 struct lnet_peer_net *lpn;
1638 struct lnet_peer_ni *lpni;
1641 LASSERT(nid4 != LNET_NID_ANY);
1644 * No need for the lnet_net_lock here, because the
1645 * lnet_api_mutex is held.
1647 lpni = lnet_find_peer_ni_locked(nid4);
1649 /* A peer with this NID already exists. */
1650 lp = lpni->lpni_peer_net->lpn_peer;
1651 lnet_peer_ni_decref_locked(lpni);
1653 * This is an error if the peer was configured and the
1654 * primary NID differs or an attempt is made to change
1655 * the Multi-Rail flag. Otherwise the assumption is
1656 * that an existing peer is being modified.
1658 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1659 if (lnet_nid_to_nid4(&lp->lp_primary_nid) != nid4)
1661 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1664 } else if (!(flags & LNET_PEER_CONFIGURED)) {
1665 if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid4) {
1670 /* Delete and recreate as a configured peer. */
1671 rc = lnet_peer_del(lp);
1676 /* Create peer, peer_net, and peer_ni. */
1678 lnet_nid4_to_nid(nid4, &nid);
1679 lp = lnet_peer_alloc(&nid);
1682 lpn = lnet_peer_net_alloc(LNET_NID_NET(&nid));
1685 lpni = lnet_peer_ni_alloc(&nid);
1689 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1692 LIBCFS_FREE(lpn, sizeof(*lpn));
1694 LIBCFS_FREE(lp, sizeof(*lp));
1696 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1697 libcfs_nid2str(nid4), flags, rc);
1702 * Add a NID to a peer. Call with ln_api_mutex held.
1705 * -EPERM: Non-DLC addition to a DLC-configured peer.
1706 * -EEXIST: The NID was configured by DLC for a different peer.
1707 * -ENOMEM: Out of memory.
1708 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1709 * non-multi-rail peer.
1712 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
1714 struct lnet_peer_net *lpn;
1715 struct lnet_peer_ni *lpni;
1716 struct lnet_nid nid;
1720 LASSERT(nid4 != LNET_NID_ANY);
1722 lnet_nid4_to_nid(nid4, &nid);
1724 /* A configured peer can only be updated through configuration. */
1725 if (!(flags & LNET_PEER_CONFIGURED)) {
1726 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1733 * The MULTI_RAIL flag can be set but not cleared, because
1734 * that would leave the peer struct in an invalid state.
1736 if (flags & LNET_PEER_MULTI_RAIL) {
1737 spin_lock(&lp->lp_lock);
1738 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1739 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1740 lnet_peer_clr_non_mr_pref_nids(lp);
1742 spin_unlock(&lp->lp_lock);
1743 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1748 lpni = lnet_find_peer_ni_locked(nid4);
1751 * A peer_ni already exists. This is only a problem if
1752 * it is not connected to this peer and was configured
1755 if (lpni->lpni_peer_net->lpn_peer == lp)
1757 if (lnet_peer_ni_is_configured(lpni)) {
1761 /* If this is the primary NID, destroy the peer. */
1762 if (lnet_peer_ni_is_primary(lpni)) {
1763 struct lnet_peer *rtr_lp =
1764 lpni->lpni_peer_net->lpn_peer;
1765 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1767 * if we're trying to delete a router it means
1768 * we're moving this peer NI to a new peer so must
1769 * transfer router properties to the new peer
1771 if (rtr_refcount > 0) {
1772 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1773 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1775 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1776 lnet_peer_ni_decref_locked(lpni);
1777 lpni = lnet_peer_ni_alloc(&nid);
1784 lpni = lnet_peer_ni_alloc(&nid);
1792 * Get the peer_net. Check that we're not adding a second
1793 * peer_ni on a peer_net of a non-multi-rail peer.
1795 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid4));
1797 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid4));
1802 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1807 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1810 lnet_peer_ni_decref_locked(lpni);
1812 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1813 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nid2str(nid4),
1819 * Update the primary NID of a peer, if possible.
1821 * Call with the lnet_api_mutex held.
1824 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid,
1827 struct lnet_nid old = lp->lp_primary_nid;
1830 if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid)
1833 lnet_nid4_to_nid(nid, &lp->lp_primary_nid);
1835 rc = lnet_peer_add_nid(lp, nid, flags);
1837 lp->lp_primary_nid = old;
1841 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1842 libcfs_nidstr(&old), libcfs_nid2str(nid), rc);
1848 * lpni creation initiated due to traffic either sending or receiving.
1851 lnet_peer_ni_traffic_add(struct lnet_nid *nid, struct lnet_nid *pref)
1853 struct lnet_peer *lp;
1854 struct lnet_peer_net *lpn;
1855 struct lnet_peer_ni *lpni;
1859 if (LNET_NID_IS_ANY(nid)) {
1864 /* lnet_net_lock is not needed here because ln_api_lock is held */
1865 lpni = lnet_peer_ni_find_locked(nid);
1868 * We must have raced with another thread. Since we
1869 * know next to nothing about a peer_ni created by
1870 * traffic, we just assume everything is ok and
1873 lnet_peer_ni_decref_locked(lpni);
1877 /* Create peer, peer_net, and peer_ni. */
1879 lp = lnet_peer_alloc(nid);
1882 lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1885 lpni = lnet_peer_ni_alloc(nid);
1888 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1890 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1893 LIBCFS_FREE(lpn, sizeof(*lpn));
1895 LIBCFS_FREE(lp, sizeof(*lp));
1897 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(nid), rc);
1902 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1904 * This API handles the following combinations:
1905 * Create a peer with its primary NI if only the prim_nid is provided
1906 * Add a NID to a peer identified by the prim_nid. The peer identified
1907 * by the prim_nid must already exist.
1908 * The peer being created may be non-MR.
1910 * The caller must hold ln_api_mutex. This prevents the peer from
1911 * being created/modified/deleted by a different thread.
1914 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr, bool temp)
1916 struct lnet_peer *lp = NULL;
1917 struct lnet_peer_ni *lpni;
1918 unsigned int flags = 0;
1920 /* The prim_nid must always be specified */
1921 if (prim_nid == LNET_NID_ANY)
1925 flags = LNET_PEER_CONFIGURED;
1928 flags |= LNET_PEER_MULTI_RAIL;
1931 * If nid isn't specified, we must create a new peer with
1932 * prim_nid as its primary nid.
1934 if (nid == LNET_NID_ANY)
1935 return lnet_peer_add(prim_nid, flags);
1937 /* Look up the prim_nid, which must exist. */
1938 lpni = lnet_find_peer_ni_locked(prim_nid);
1941 lnet_peer_ni_decref_locked(lpni);
1942 lp = lpni->lpni_peer_net->lpn_peer;
1944 /* Peer must have been configured. */
1945 if (!temp && !(lp->lp_state & LNET_PEER_CONFIGURED)) {
1946 CDEBUG(D_NET, "peer %s was not configured\n",
1947 libcfs_nid2str(prim_nid));
1951 /* Primary NID must match */
1952 if (lnet_nid_to_nid4(&lp->lp_primary_nid) != prim_nid) {
1953 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1954 libcfs_nid2str(prim_nid),
1955 libcfs_nidstr(&lp->lp_primary_nid));
1959 /* Multi-Rail flag must match. */
1960 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1961 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1962 libcfs_nid2str(prim_nid));
1966 return lnet_peer_add_nid(lp, nid, flags);
1970 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1972 * This API handles the following combinations:
1973 * Delete a NI from a peer if both prim_nid and nid are provided.
1974 * Delete a peer if only prim_nid is provided.
1975 * Delete a peer if its primary nid is provided.
1977 * The caller must hold ln_api_mutex. This prevents the peer from
1978 * being modified/deleted by a different thread.
1981 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1983 struct lnet_peer *lp;
1984 struct lnet_peer_ni *lpni;
1987 if (prim_nid == LNET_NID_ANY)
1990 lpni = lnet_find_peer_ni_locked(prim_nid);
1993 lnet_peer_ni_decref_locked(lpni);
1994 lp = lpni->lpni_peer_net->lpn_peer;
1996 if (prim_nid != lnet_nid_to_nid4(&lp->lp_primary_nid)) {
1997 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1998 libcfs_nid2str(prim_nid),
1999 libcfs_nidstr(&lp->lp_primary_nid));
2003 lnet_net_lock(LNET_LOCK_EX);
2004 if (lp->lp_rtr_refcount > 0) {
2005 lnet_net_unlock(LNET_LOCK_EX);
2006 CERROR("%s is a router. Can not be deleted\n",
2007 libcfs_nid2str(prim_nid));
2010 lnet_net_unlock(LNET_LOCK_EX);
2012 if (nid == LNET_NID_ANY || nid == lnet_nid_to_nid4(&lp->lp_primary_nid))
2013 return lnet_peer_del(lp);
2015 flags = LNET_PEER_CONFIGURED;
2016 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2017 flags |= LNET_PEER_MULTI_RAIL;
2019 return lnet_peer_del_nid(lp, nid, flags);
2023 lnet_destroy_peer_ni_locked(struct kref *ref)
2025 struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
2027 struct lnet_peer_table *ptable;
2028 struct lnet_peer_net *lpn;
2030 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
2032 LASSERT(kref_read(&lpni->lpni_kref) == 0);
2033 LASSERT(list_empty(&lpni->lpni_txq));
2034 LASSERT(lpni->lpni_txqnob == 0);
2035 LASSERT(list_empty(&lpni->lpni_peer_nis));
2036 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
2038 lpn = lpni->lpni_peer_net;
2039 lpni->lpni_peer_net = NULL;
2040 lpni->lpni_net = NULL;
2042 if (!list_empty(&lpni->lpni_hashlist)) {
2043 /* remove the peer ni from the zombie list */
2044 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
2045 spin_lock(&ptable->pt_zombie_lock);
2046 list_del_init(&lpni->lpni_hashlist);
2047 ptable->pt_zombies--;
2048 spin_unlock(&ptable->pt_zombie_lock);
2051 if (lpni->lpni_pref_nnids > 1) {
2052 struct lnet_nid_list *ne, *tmp;
2054 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
2056 list_del_init(&ne->nl_list);
2057 LIBCFS_FREE(ne, sizeof(*ne));
2060 LIBCFS_FREE(lpni, sizeof(*lpni));
2063 lnet_peer_net_decref_locked(lpn);
2066 struct lnet_peer_ni *
2067 lnet_nid2peerni_ex(struct lnet_nid *nid, int cpt)
2069 struct lnet_peer_ni *lpni = NULL;
2072 if (the_lnet.ln_state != LNET_STATE_RUNNING)
2073 return ERR_PTR(-ESHUTDOWN);
2076 * find if a peer_ni already exists.
2077 * If so then just return that.
2079 lpni = lnet_peer_ni_find_locked(nid);
2083 lnet_net_unlock(cpt);
2085 rc = lnet_peer_ni_traffic_add(nid, NULL);
2088 goto out_net_relock;
2091 lpni = lnet_peer_ni_find_locked(nid);
2101 * Get a peer_ni for the given nid, create it if necessary. Takes a
2102 * hold on the peer_ni.
2104 struct lnet_peer_ni *
2105 lnet_peerni_by_nid_locked(struct lnet_nid *nid,
2106 struct lnet_nid *pref, int cpt)
2108 struct lnet_peer_ni *lpni = NULL;
2111 if (the_lnet.ln_state != LNET_STATE_RUNNING)
2112 return ERR_PTR(-ESHUTDOWN);
2115 * find if a peer_ni already exists.
2116 * If so then just return that.
2118 lpni = lnet_peer_ni_find_locked(nid);
2124 * use the lnet_api_mutex to serialize the creation of the peer_ni
2125 * and the creation/deletion of the local ni/net. When a local ni is
2126 * created, if there exists a set of peer_nis on that network,
2127 * they need to be traversed and updated. When a local NI is
2128 * deleted, which could result in a network being deleted, then
2129 * all peer nis on that network need to be removed as well.
2131 * Creation through traffic should also be serialized with
2132 * creation through DLC.
2134 lnet_net_unlock(cpt);
2135 mutex_lock(&the_lnet.ln_api_mutex);
2137 * Shutdown is only set under the ln_api_lock, so a single
2138 * check here is sufficent.
2140 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2141 lpni = ERR_PTR(-ESHUTDOWN);
2142 goto out_mutex_unlock;
2145 rc = lnet_peer_ni_traffic_add(nid, pref);
2148 goto out_mutex_unlock;
2151 lpni = lnet_peer_ni_find_locked(nid);
2155 mutex_unlock(&the_lnet.ln_api_mutex);
2158 /* Lock has been dropped, check again for shutdown. */
2159 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2161 lnet_peer_ni_decref_locked(lpni);
2162 lpni = ERR_PTR(-ESHUTDOWN);
2168 struct lnet_peer_ni *
2169 lnet_nid2peerni_locked(lnet_nid_t nid4, lnet_nid_t pref4, int cpt)
2171 struct lnet_nid nid, pref;
2173 lnet_nid4_to_nid(nid4, &nid);
2174 lnet_nid4_to_nid(pref4, &pref);
2175 if (pref4 == LNET_NID_ANY)
2176 return lnet_peerni_by_nid_locked(&nid, NULL, cpt);
2178 return lnet_peerni_by_nid_locked(&nid, &pref, cpt);
2182 lnet_peer_gw_discovery(struct lnet_peer *lp)
2186 spin_lock(&lp->lp_lock);
2187 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2189 spin_unlock(&lp->lp_lock);
2195 lnet_peer_is_uptodate(struct lnet_peer *lp)
2199 spin_lock(&lp->lp_lock);
2200 rc = lnet_peer_is_uptodate_locked(lp);
2201 spin_unlock(&lp->lp_lock);
2206 * Is a peer uptodate from the point of view of discovery?
2208 * If it is currently being processed, obviously not.
2209 * A forced Ping or Push is also handled by the discovery thread.
2211 * Otherwise look at whether the peer needs rediscovering.
2214 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2215 __must_hold(&lp->lp_lock)
2219 if (lp->lp_state & (LNET_PEER_DISCOVERING |
2220 LNET_PEER_FORCE_PING |
2221 LNET_PEER_FORCE_PUSH)) {
2223 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2225 } else if (lnet_peer_needs_push(lp)) {
2227 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2228 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2239 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2241 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2243 /* The discovery thread holds net_lock/EX and lp_lock when it splices
2244 * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2245 * when adding to the list and queuing the peer to ensure that we do not
2246 * strand any messages on the lp_dc_pendq. This scheme ensures the
2247 * message will be resent even if the peer is already being discovered.
2248 * Therefore we needn't check the return value of
2249 * lnet_peer_queue_for_discovery(lp).
2251 lnet_net_lock(LNET_LOCK_EX);
2252 spin_lock(&lp->lp_lock);
2253 list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2254 spin_unlock(&lp->lp_lock);
2255 lnet_peer_queue_for_discovery(lp);
2256 lnet_net_unlock(LNET_LOCK_EX);
2260 * Queue a peer for the attention of the discovery thread. Call with
2261 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2262 * -EALREADY if the peer was already queued.
2264 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2268 spin_lock(&lp->lp_lock);
2269 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2270 lp->lp_state |= LNET_PEER_DISCOVERING;
2271 spin_unlock(&lp->lp_lock);
2272 if (list_empty(&lp->lp_dc_list)) {
2273 lnet_peer_addref_locked(lp);
2274 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2275 wake_up(&the_lnet.ln_dc_waitq);
2281 CDEBUG(D_NET, "Queue peer %s: %d\n",
2282 libcfs_nidstr(&lp->lp_primary_nid), rc);
2288 * Discovery of a peer is complete. Wake all waiters on the peer.
2289 * Call with lnet_net_lock/EX held.
2291 static void lnet_peer_discovery_complete(struct lnet_peer *lp, int dc_error)
2293 struct lnet_msg *msg, *tmp;
2295 LIST_HEAD(pending_msgs);
2297 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2298 libcfs_nidstr(&lp->lp_primary_nid));
2300 list_del_init(&lp->lp_dc_list);
2301 spin_lock(&lp->lp_lock);
2303 lp->lp_dc_error = dc_error;
2304 lp->lp_state &= ~LNET_PEER_DISCOVERING;
2305 lp->lp_state |= LNET_PEER_REDISCOVER;
2307 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2308 spin_unlock(&lp->lp_lock);
2309 wake_up(&lp->lp_dc_waitq);
2311 if (lp->lp_rtr_refcount > 0)
2312 lnet_router_discovery_complete(lp);
2314 lnet_net_unlock(LNET_LOCK_EX);
2316 /* iterate through all pending messages and send them again */
2317 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2318 list_del_init(&msg->msg_list);
2320 lnet_finalize(msg, dc_error);
2324 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2325 lnet_msgtyp2str(msg->msg_type),
2326 libcfs_idstr(&msg->msg_target));
2327 rc = lnet_send(&msg->msg_src_nid_param, msg,
2328 &msg->msg_rtr_nid_param);
2330 CNETERR("Error sending %s to %s: %d\n",
2331 lnet_msgtyp2str(msg->msg_type),
2332 libcfs_idstr(&msg->msg_target), rc);
2333 lnet_finalize(msg, rc);
2336 lnet_net_lock(LNET_LOCK_EX);
2337 lnet_peer_decref_locked(lp);
2341 * Handle inbound push.
2342 * Like any event handler, called with lnet_res_lock/CPT held.
2344 void lnet_peer_push_event(struct lnet_event *ev)
2346 struct lnet_ping_buffer *pbuf;
2347 struct lnet_peer *lp;
2349 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2351 /* lnet_find_peer() adds a refcount */
2352 lp = lnet_find_peer(&ev->source.nid);
2354 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2355 libcfs_nidstr(&ev->initiator.nid),
2356 libcfs_nidstr(&ev->source.nid));
2357 pbuf->pb_needs_post = true;
2361 /* Ensure peer state remains consistent while we modify it. */
2362 spin_lock(&lp->lp_lock);
2365 * If some kind of error happened the contents of the message
2366 * cannot be used. Clear the NIDS_UPTODATE and set the
2367 * FORCE_PING flag to trigger a ping.
2370 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2371 lp->lp_state |= LNET_PEER_FORCE_PING;
2372 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2374 libcfs_nidstr(&lp->lp_primary_nid),
2375 libcfs_nidstr(&ev->source.nid));
2380 * A push with invalid or corrupted info. Clear the UPTODATE
2381 * flag to trigger a ping.
2383 if (lnet_ping_info_validate(&pbuf->pb_info)) {
2384 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2385 lp->lp_state |= LNET_PEER_FORCE_PING;
2386 CDEBUG(D_NET, "Corrupted Push from %s\n",
2387 libcfs_nidstr(&lp->lp_primary_nid));
2392 * Make sure we'll allocate the correct size ping buffer when
2395 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2396 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2399 * A non-Multi-Rail peer is not supposed to be capable of
2402 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2403 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2404 libcfs_nidstr(&lp->lp_primary_nid));
2409 * The peer may have discovery disabled at its end. Set
2410 * NO_DISCOVERY as appropriate.
2412 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2413 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2414 libcfs_nidstr(&lp->lp_primary_nid));
2416 * Mark the peer for deletion if we already know about it
2417 * and it's going from discovery set to no discovery set
2419 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2420 LNET_PEER_DISCOVERING)) &&
2421 lp->lp_state & LNET_PEER_DISCOVERED) {
2422 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2423 libcfs_nidstr(&lp->lp_primary_nid),
2425 lp->lp_state |= LNET_PEER_MARK_DELETION;
2427 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2428 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2429 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2430 libcfs_nidstr(&lp->lp_primary_nid));
2431 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2435 * Update the MULTI_RAIL flag based on the push. If the peer
2436 * was configured with DLC then the setting should match what
2438 * NB: We verified above that the MR feature bit is set in pi_features
2440 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2441 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2442 libcfs_nidstr(&lp->lp_primary_nid), lp);
2443 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2444 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2445 libcfs_nidstr(&lp->lp_primary_nid));
2446 } else if (lnet_peer_discovery_disabled) {
2447 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2448 libcfs_nidstr(&lp->lp_primary_nid), lp);
2449 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2450 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2451 libcfs_nidstr(&lp->lp_primary_nid), lp);
2453 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2454 libcfs_nidstr(&lp->lp_primary_nid), lp);
2455 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2456 lnet_peer_clr_non_mr_pref_nids(lp);
2460 * Check for truncation of the Put message. Clear the
2461 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2462 * and tell discovery to allocate a bigger buffer.
2464 if (ev->mlength < ev->rlength) {
2465 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2466 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2467 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2468 lp->lp_state |= LNET_PEER_FORCE_PING;
2469 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2470 libcfs_nidstr(&lp->lp_primary_nid),
2471 pbuf->pb_info.pi_nnis);
2475 /* always assume new data */
2476 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2477 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2480 * If there is data present that hasn't been processed yet,
2481 * we'll replace it if the Put contained newer data and it
2482 * fits. We're racing with a Ping or earlier Push in this
2485 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2486 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2487 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2488 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2489 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2490 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2491 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2492 libcfs_nidstr(&lp->lp_primary_nid),
2493 LNET_PING_BUFFER_SEQNO(pbuf),
2494 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2500 * Allocate a buffer to copy the data. On a failure we drop
2501 * the Push and set FORCE_PING to force the discovery
2502 * thread to fix the problem by pinging the peer.
2504 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2506 lp->lp_state |= LNET_PEER_FORCE_PING;
2507 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2508 libcfs_nidstr(&lp->lp_primary_nid),
2509 LNET_PING_BUFFER_SEQNO(pbuf));
2514 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2515 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2516 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2517 CDEBUG(D_NET, "Received Push %s %u\n",
2518 libcfs_nidstr(&lp->lp_primary_nid),
2519 LNET_PING_BUFFER_SEQNO(pbuf));
2522 /* We've processed this buffer. It can be reposted */
2523 pbuf->pb_needs_post = true;
2526 * Queue the peer for discovery if not done, force it on the request
2527 * queue and wake the discovery thread if the peer was already queued,
2528 * because its status changed.
2530 spin_unlock(&lp->lp_lock);
2531 lnet_net_lock(LNET_LOCK_EX);
2532 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2533 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2534 wake_up(&the_lnet.ln_dc_waitq);
2536 /* Drop refcount from lookup */
2537 lnet_peer_decref_locked(lp);
2538 lnet_net_unlock(LNET_LOCK_EX);
2542 * Clear the discovery error state, unless we're already discovering
2543 * this peer, in which case the error is current.
2545 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2547 spin_lock(&lp->lp_lock);
2548 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2549 lp->lp_dc_error = 0;
2550 spin_unlock(&lp->lp_lock);
2554 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2555 * dropped/retaken within this function. An lnet_peer_ni is passed in
2556 * because discovery could tear down an lnet_peer.
2559 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2562 struct lnet_peer *lp;
2567 lnet_net_unlock(cpt);
2568 lnet_net_lock(LNET_LOCK_EX);
2569 lp = lpni->lpni_peer_net->lpn_peer;
2570 lnet_peer_clear_discovery_error(lp);
2573 * We're willing to be interrupted. The lpni can become a
2574 * zombie if we race with DLC, so we must check for that.
2577 /* Keep lp alive when the lnet_net_lock is unlocked */
2578 lnet_peer_addref_locked(lp);
2579 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2580 if (signal_pending(current))
2582 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2585 * Don't repeat discovery if discovery is disabled. This is
2586 * done to ensure we can use discovery as a standard ping as
2587 * well for backwards compatibility with routers which do not
2588 * have discovery or have discovery disabled
2590 if (lnet_is_discovery_disabled(lp) && count > 0)
2592 if (lp->lp_dc_error)
2594 if (lnet_peer_is_uptodate(lp))
2596 lnet_peer_queue_for_discovery(lp);
2598 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2601 * If caller requested a non-blocking operation then
2602 * return immediately. Once discovery is complete any
2603 * pending messages that were stopped due to discovery
2604 * will be transmitted.
2609 lnet_net_unlock(LNET_LOCK_EX);
2611 finish_wait(&lp->lp_dc_waitq, &wait);
2612 lnet_net_lock(LNET_LOCK_EX);
2613 lnet_peer_decref_locked(lp);
2614 /* Peer may have changed */
2615 lp = lpni->lpni_peer_net->lpn_peer;
2617 finish_wait(&lp->lp_dc_waitq, &wait);
2619 lnet_net_unlock(LNET_LOCK_EX);
2621 lnet_peer_decref_locked(lp);
2623 * The peer may have changed, so re-check and rediscover if that turns
2624 * out to have been the case. The reference count on lp ensured that
2625 * even if it was unlinked from lpni the memory could not be recycled.
2626 * Thus the check below is sufficient to determine whether the peer
2627 * changed. If the peer changed, then lp must not be dereferenced.
2629 if (lp != lpni->lpni_peer_net->lpn_peer)
2632 if (signal_pending(current))
2634 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2636 else if (lp->lp_dc_error)
2637 rc = lp->lp_dc_error;
2639 CDEBUG(D_NET, "non-blocking discovery\n");
2640 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2643 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2644 (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2645 libcfs_nidstr(&lpni->lpni_nid), rc,
2646 (!block) ? "pending discovery" : "discovery complete");
2651 /* Handle an incoming ack for a push. */
2653 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2655 struct lnet_ping_buffer *pbuf;
2657 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2658 spin_lock(&lp->lp_lock);
2659 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2660 lp->lp_push_error = ev->status;
2662 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2664 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2665 spin_unlock(&lp->lp_lock);
2667 CDEBUG(D_NET, "peer %s ev->status %d\n",
2668 libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2671 /* Handle a Reply message. This is the reply to a Ping message. */
2673 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2675 struct lnet_ping_buffer *pbuf;
2678 spin_lock(&lp->lp_lock);
2680 lp->lp_disc_src_nid = ev->target.nid;
2681 lp->lp_disc_dst_nid = ev->source.nid;
2684 * If some kind of error happened the contents of message
2685 * cannot be used. Set PING_FAILED to trigger a retry.
2688 lp->lp_state |= LNET_PEER_PING_FAILED;
2689 lp->lp_ping_error = ev->status;
2690 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2692 libcfs_nidstr(&lp->lp_primary_nid),
2693 libcfs_nidstr(&ev->source.nid));
2697 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2698 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2699 lnet_swap_pinginfo(pbuf);
2702 * A reply with invalid or corrupted info. Set PING_FAILED to
2705 rc = lnet_ping_info_validate(&pbuf->pb_info);
2707 lp->lp_state |= LNET_PEER_PING_FAILED;
2708 lp->lp_ping_error = 0;
2709 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2710 libcfs_nidstr(&lp->lp_primary_nid), rc);
2715 * The peer may have discovery disabled at its end. Set
2716 * NO_DISCOVERY as appropriate.
2718 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) ||
2719 lnet_peer_discovery_disabled) {
2720 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2721 libcfs_nidstr(&lp->lp_primary_nid));
2723 /* Detect whether this peer has toggled discovery from on to
2724 * off and whether we can delete and re-create the peer. Peers
2725 * that were manually configured cannot be deleted by discovery.
2726 * We need to delete this peer and re-create it if the peer was
2727 * not configured manually, is currently considered DD capable,
2729 * 1. We've already discovered the peer (the peer has toggled
2730 * the discovery feature from on to off), or
2731 * 2. The peer is considered MR, but it was not user configured
2732 * (this was a "temporary" peer created via the kernel APIs
2733 * that we're discovering for the first time)
2735 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2736 LNET_PEER_NO_DISCOVERY)) &&
2737 (lp->lp_state & (LNET_PEER_DISCOVERED |
2738 LNET_PEER_MULTI_RAIL))) {
2739 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2740 libcfs_nidstr(&lp->lp_primary_nid),
2742 lp->lp_state |= LNET_PEER_MARK_DELETION;
2744 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2746 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2747 libcfs_nidstr(&lp->lp_primary_nid));
2748 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2752 * Update the MULTI_RAIL flag based on the reply. If the peer
2753 * was configured with DLC then the setting should match what
2756 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2757 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2758 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2759 libcfs_nidstr(&lp->lp_primary_nid), lp);
2760 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2761 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2762 libcfs_nidstr(&lp->lp_primary_nid));
2763 } else if (lnet_peer_discovery_disabled) {
2765 "peer %s(%p) not MR: DD disabled locally\n",
2766 libcfs_nidstr(&lp->lp_primary_nid), lp);
2767 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2769 "peer %s(%p) not MR: DD disabled remotely\n",
2770 libcfs_nidstr(&lp->lp_primary_nid), lp);
2772 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2773 libcfs_nidstr(&lp->lp_primary_nid), lp);
2774 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2775 lnet_peer_clr_non_mr_pref_nids(lp);
2777 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2778 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2779 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2780 libcfs_nidstr(&lp->lp_primary_nid));
2782 CERROR("Multi-Rail state vanished from %s\n",
2783 libcfs_nidstr(&lp->lp_primary_nid));
2784 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2789 * Make sure we'll allocate the correct size ping buffer when
2792 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2793 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2796 * Check for truncation of the Reply. Clear PING_SENT and set
2797 * PING_FAILED to trigger a retry.
2799 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2800 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2801 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2802 lp->lp_state |= LNET_PEER_PING_FAILED;
2803 lp->lp_ping_error = 0;
2804 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2805 libcfs_nidstr(&lp->lp_primary_nid),
2806 pbuf->pb_info.pi_nnis);
2811 * Check the sequence numbers in the reply. These are only
2812 * available if the reply came from a Multi-Rail peer.
2814 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2815 pbuf->pb_info.pi_nnis > 1 &&
2816 lnet_nid_to_nid4(&lp->lp_primary_nid) ==
2817 pbuf->pb_info.pi_ni[1].ns_nid) {
2818 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2819 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2820 libcfs_nidstr(&lp->lp_primary_nid),
2821 LNET_PING_BUFFER_SEQNO(pbuf),
2824 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2827 /* We're happy with the state of the data in the buffer. */
2828 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2829 libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2831 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2832 lnet_ping_buffer_decref(lp->lp_data);
2834 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2835 lnet_ping_buffer_addref(pbuf);
2838 lp->lp_state &= ~LNET_PEER_PING_SENT;
2839 spin_unlock(&lp->lp_lock);
2841 lnet_net_lock(LNET_LOCK_EX);
2843 * If this peer is a gateway, call the routing callback to
2844 * handle the ping reply
2846 if (lp->lp_rtr_refcount > 0)
2847 lnet_router_discovery_ping_reply(lp);
2848 lnet_net_unlock(LNET_LOCK_EX);
2852 * Send event handling. Only matters for error cases, where we clean
2853 * up state on the peer and peer_ni that would otherwise be updated in
2854 * the REPLY event handler for a successful Ping, and the ACK event
2855 * handler for a successful Push.
2858 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2865 spin_lock(&lp->lp_lock);
2866 if (ev->msg_type == LNET_MSG_GET) {
2867 lp->lp_state &= ~LNET_PEER_PING_SENT;
2868 lp->lp_state |= LNET_PEER_PING_FAILED;
2869 lp->lp_ping_error = ev->status;
2870 } else { /* ev->msg_type == LNET_MSG_PUT */
2871 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2872 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2873 lp->lp_push_error = ev->status;
2875 spin_unlock(&lp->lp_lock);
2876 rc = LNET_REDISCOVER_PEER;
2878 CDEBUG(D_NET, "%s Send to %s: %d\n",
2879 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2880 libcfs_nidstr(&ev->target.nid), rc);
2885 * Unlink event handling. This event is only seen if a call to
2886 * LNetMDUnlink() caused the event to be unlinked. If this call was
2887 * made after the event was set up in LNetGet() or LNetPut() then we
2888 * assume the Ping or Push timed out.
2891 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2893 spin_lock(&lp->lp_lock);
2894 /* We've passed through LNetGet() */
2895 if (lp->lp_state & LNET_PEER_PING_SENT) {
2896 lp->lp_state &= ~LNET_PEER_PING_SENT;
2897 lp->lp_state |= LNET_PEER_PING_FAILED;
2898 lp->lp_ping_error = -ETIMEDOUT;
2899 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2900 libcfs_nidstr(&lp->lp_primary_nid));
2902 /* We've passed through LNetPut() */
2903 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2904 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2905 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2906 lp->lp_push_error = -ETIMEDOUT;
2907 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2908 libcfs_nidstr(&lp->lp_primary_nid));
2910 spin_unlock(&lp->lp_lock);
2914 * Event handler for the discovery EQ.
2916 * Called with lnet_res_lock(cpt) held. The cpt is the
2917 * lnet_cpt_of_cookie() of the md handle cookie.
2919 static void lnet_discovery_event_handler(struct lnet_event *event)
2921 struct lnet_peer *lp = event->md_user_ptr;
2922 struct lnet_ping_buffer *pbuf;
2925 /* discovery needs to take another look */
2926 rc = LNET_REDISCOVER_PEER;
2928 CDEBUG(D_NET, "Received event: %d\n", event->type);
2930 switch (event->type) {
2931 case LNET_EVENT_ACK:
2932 lnet_discovery_event_ack(lp, event);
2934 case LNET_EVENT_REPLY:
2935 lnet_discovery_event_reply(lp, event);
2937 case LNET_EVENT_SEND:
2938 /* Only send failure triggers a retry. */
2939 rc = lnet_discovery_event_send(lp, event);
2941 case LNET_EVENT_UNLINK:
2942 /* LNetMDUnlink() was called */
2943 lnet_discovery_event_unlink(lp, event);
2946 /* Invalid events. */
2949 lnet_net_lock(LNET_LOCK_EX);
2950 if (event->unlinked) {
2951 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2952 lnet_ping_buffer_decref(pbuf);
2953 lnet_peer_decref_locked(lp);
2956 /* put peer back at end of request queue, if discovery not already
2958 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2959 lnet_peer_queue_for_discovery(lp)) {
2960 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2961 wake_up(&the_lnet.ln_dc_waitq);
2963 lnet_net_unlock(LNET_LOCK_EX);
2967 * Build a peer from incoming data.
2969 * The NIDs in the incoming data are supposed to be structured as follows:
2972 * - other NIDs in same net
2973 * - NIDs in second net
2974 * - NIDs in third net
2976 * This due to the way the list of NIDs in the data is created.
2978 * Note that this function will mark the peer uptodate unless an
2979 * ENOMEM is encontered. All other errors are due to a conflict
2980 * between the DLC configuration and what discovery sees. We treat DLC
2981 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2982 * peer from becoming stuck in discovery.
2984 static int lnet_peer_merge_data(struct lnet_peer *lp,
2985 struct lnet_ping_buffer *pbuf)
2987 struct lnet_peer_net *lpn;
2988 struct lnet_peer_ni *lpni;
2989 lnet_nid_t *curnis = NULL;
2990 struct lnet_ni_status *addnis = NULL;
2991 lnet_nid_t *delnis = NULL;
3001 flags = LNET_PEER_DISCOVERED;
3002 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3003 flags |= LNET_PEER_MULTI_RAIL;
3006 * Cache the routing feature for the peer; whether it is enabled
3007 * for disabled as reported by the remote peer.
3009 spin_lock(&lp->lp_lock);
3010 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
3011 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
3013 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
3014 spin_unlock(&lp->lp_lock);
3016 nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
3017 CFS_ALLOC_PTR_ARRAY(curnis, nnis);
3018 CFS_ALLOC_PTR_ARRAY(addnis, nnis);
3019 CFS_ALLOC_PTR_ARRAY(delnis, nnis);
3020 if (!curnis || !addnis || !delnis) {
3028 /* Construct the list of NIDs present in peer. */
3030 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
3031 curnis[ncurnis++] = lnet_nid_to_nid4(&lpni->lpni_nid);
3034 * Check for NIDs in pbuf not present in curnis[].
3035 * The loop starts at 1 to skip the loopback NID.
3037 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
3038 for (j = 0; j < ncurnis; j++)
3039 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
3042 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
3045 * Check for NIDs in curnis[] not present in pbuf.
3046 * The nested loop starts at 1 to skip the loopback NID.
3048 * But never add the loopback NID to delnis[]: if it is
3049 * present in curnis[] then this peer is for this node.
3051 for (i = 0; i < ncurnis; i++) {
3052 if (curnis[i] == LNET_NID_LO_0)
3054 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
3055 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
3057 * update the information we cache for the
3058 * peer with the latest information we
3061 lpni = lnet_find_peer_ni_locked(curnis[i]);
3063 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
3064 lnet_peer_ni_decref_locked(lpni);
3069 if (j == pbuf->pb_info.pi_nnis)
3070 delnis[ndelnis++] = curnis[i];
3074 * If we get here and the discovery is disabled then we don't want
3075 * to add or delete any NIs. We just updated the ones we have some
3076 * information on, and call it a day
3079 if (lnet_is_discovery_disabled(lp))
3082 for (i = 0; i < naddnis; i++) {
3083 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
3085 CERROR("Error adding NID %s to peer %s: %d\n",
3086 libcfs_nid2str(addnis[i].ns_nid),
3087 libcfs_nidstr(&lp->lp_primary_nid), rc);
3091 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
3093 lpni->lpni_ns_status = addnis[i].ns_status;
3094 lnet_peer_ni_decref_locked(lpni);
3098 for (i = 0; i < ndelnis; i++) {
3100 * for routers it's okay to delete the primary_nid because
3101 * the upper layers don't really rely on it. So if we're
3102 * being told that the router changed its primary_nid
3103 * then it's okay to delete it.
3105 if (lp->lp_rtr_refcount > 0)
3106 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3107 rc = lnet_peer_del_nid(lp, delnis[i], flags);
3109 CERROR("Error deleting NID %s from peer %s: %d\n",
3110 libcfs_nid2str(delnis[i]),
3111 libcfs_nidstr(&lp->lp_primary_nid), rc);
3117 /* The peer net for the primary NID should be the first entry in the
3118 * peer's lp_peer_nets list, and the peer NI for the primary NID should
3119 * be the first entry in its peer net's lpn_peer_nis list.
3121 lpni = lnet_find_peer_ni_locked(pbuf->pb_info.pi_ni[1].ns_nid);
3123 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3124 libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
3128 lnet_peer_ni_decref_locked(lpni);
3130 lpn = lpni->lpni_peer_net;
3131 if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3132 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3134 if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3135 list_move(&lpni->lpni_peer_nis,
3136 &lpni->lpni_peer_net->lpn_peer_nis);
3139 * Errors other than -ENOMEM are due to peers having been
3140 * configured with DLC. Ignore these because DLC overrides
3145 CFS_FREE_PTR_ARRAY(curnis, nnis);
3146 CFS_FREE_PTR_ARRAY(addnis, nnis);
3147 CFS_FREE_PTR_ARRAY(delnis, nnis);
3148 lnet_ping_buffer_decref(pbuf);
3149 CDEBUG(D_NET, "peer %s (%p): %d\n",
3150 libcfs_nidstr(&lp->lp_primary_nid), lp, rc);
3153 spin_lock(&lp->lp_lock);
3154 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
3155 lp->lp_state |= LNET_PEER_FORCE_PING;
3156 spin_unlock(&lp->lp_lock);
3162 * The data in pbuf says lp is its primary peer, but the data was
3163 * received by a different peer. Try to update lp with the data.
3166 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
3168 struct lnet_handle_md mdh;
3170 /* Queue lp for discovery, and force it on the request queue. */
3171 lnet_net_lock(LNET_LOCK_EX);
3172 if (lnet_peer_queue_for_discovery(lp))
3173 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3174 lnet_net_unlock(LNET_LOCK_EX);
3176 LNetInvalidateMDHandle(&mdh);
3179 * Decide whether we can move the peer to the DATA_PRESENT state.
3181 * We replace stale data for a multi-rail peer, repair PING_FAILED
3182 * status, and preempt FORCE_PING.
3184 * If after that we have DATA_PRESENT, we merge it into this peer.
3186 spin_lock(&lp->lp_lock);
3187 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3188 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3189 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3190 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3191 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3192 lnet_ping_buffer_decref(pbuf);
3197 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3198 lnet_ping_buffer_decref(lp->lp_data);
3200 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3202 if (lp->lp_state & LNET_PEER_PING_FAILED) {
3203 mdh = lp->lp_ping_mdh;
3204 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3205 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3206 lp->lp_ping_error = 0;
3208 if (lp->lp_state & LNET_PEER_FORCE_PING)
3209 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3210 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3211 spin_unlock(&lp->lp_lock);
3213 if (!LNetMDHandleIsInvalid(mdh))
3217 return lnet_peer_merge_data(lp, pbuf);
3219 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3223 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
3227 for (i = 0; i < pinfo->pi_nnis; i++) {
3228 if (pinfo->pi_ni[i].ns_nid == nid)
3235 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3236 * to the discovery queue a reference was taken that will prevent the peer from
3237 * actually being freed by this function. After this function exits the
3238 * discovery thread should call lnet_peer_discovery_complete() which will
3239 * drop that reference as well as wake any waiters that may also be holding a
3242 static int lnet_peer_deletion(struct lnet_peer *lp)
3243 __must_hold(&lp->lp_lock)
3245 struct list_head rlist;
3246 struct lnet_route *route, *tmp;
3247 int sensitivity = lp->lp_health_sensitivity;
3250 INIT_LIST_HEAD(&rlist);
3252 lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3253 LNET_PEER_FORCE_PUSH);
3254 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3255 libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3257 /* no-op if lnet_peer_del() has already been called on this peer */
3258 if (lp->lp_state & LNET_PEER_MARK_DELETED)
3261 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
3264 spin_unlock(&lp->lp_lock);
3266 mutex_lock(&the_lnet.ln_api_mutex);
3268 lnet_net_lock(LNET_LOCK_EX);
3269 /* remove the peer from the discovery work
3270 * queue if it's on there in preparation
3273 if (!list_empty(&lp->lp_dc_list))
3274 list_del_init(&lp->lp_dc_list);
3275 list_for_each_entry_safe(route, tmp,
3278 lnet_move_route(route, NULL, &rlist);
3279 lnet_net_unlock(LNET_LOCK_EX);
3281 /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3282 rc = lnet_peer_del(lp);
3284 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3285 libcfs_nidstr(&lp->lp_primary_nid), rc);
3287 list_for_each_entry_safe(route, tmp,
3289 /* re-add these routes */
3290 lnet_add_route(route->lr_net,
3295 LIBCFS_FREE(route, sizeof(*route));
3298 mutex_unlock(&the_lnet.ln_api_mutex);
3300 spin_lock(&lp->lp_lock);
3306 * Update a peer using the data received.
3308 static int lnet_peer_data_present(struct lnet_peer *lp)
3309 __must_hold(&lp->lp_lock)
3311 struct lnet_ping_buffer *pbuf;
3312 struct lnet_peer_ni *lpni;
3313 lnet_nid_t nid = LNET_NID_ANY;
3319 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3320 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3321 spin_unlock(&lp->lp_lock);
3324 * Modifications of peer structures are done while holding the
3325 * ln_api_mutex. A global lock is required because we may be
3326 * modifying multiple peer structures, and a mutex greatly
3327 * simplifies memory management.
3329 * The actual changes to the data structures must also protect
3330 * against concurrent lookups, for which the lnet_net_lock in
3331 * LNET_LOCK_EX mode is used.
3333 mutex_lock(&the_lnet.ln_api_mutex);
3334 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3340 * If this peer is not on the peer list then it is being torn
3341 * down, and our reference count may be all that is keeping it
3342 * alive. Don't do any work on it.
3344 if (list_empty(&lp->lp_peer_list))
3347 flags = LNET_PEER_DISCOVERED;
3348 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3349 flags |= LNET_PEER_MULTI_RAIL;
3352 * Check whether the primary NID in the message matches the
3353 * primary NID of the peer. If it does, update the peer, if
3354 * it it does not, check whether there is already a peer with
3355 * that primary NID. If no such peer exists, try to update
3356 * the primary NID of the current peer (allowed if it was
3357 * created due to message traffic) and complete the update.
3358 * If the peer did exist, hand off the data to it.
3360 * The peer for the loopback interface is a special case: this
3361 * is the peer for the local node, and we want to set its
3362 * primary NID to the correct value here. Moreover, this peer
3363 * can show up with only the loopback NID in the ping buffer.
3365 if (pbuf->pb_info.pi_nnis <= 1)
3367 nid = pbuf->pb_info.pi_ni[1].ns_nid;
3368 if (nid_is_lo0(&lp->lp_primary_nid)) {
3369 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3371 rc = lnet_peer_merge_data(lp, pbuf);
3373 * if the primary nid of the peer is present in the ping info returned
3374 * from the peer, but it's not the local primary peer we have
3375 * cached and discovery is disabled, then we don't want to update
3376 * our local peer info, by adding or removing NIDs, we just want
3377 * to update the status of the nids that we currently have
3378 * recorded in that peer.
3380 } else if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid ||
3381 (lnet_is_nid_in_ping_info(lnet_nid_to_nid4(&lp->lp_primary_nid),
3383 lnet_is_discovery_disabled(lp))) {
3384 rc = lnet_peer_merge_data(lp, pbuf);
3386 lpni = lnet_find_peer_ni_locked(nid);
3387 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3388 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3390 CERROR("Primary NID error %s versus %s: %d\n",
3391 libcfs_nidstr(&lp->lp_primary_nid),
3392 libcfs_nid2str(nid), rc);
3394 rc = lnet_peer_merge_data(lp, pbuf);
3397 lnet_peer_ni_decref_locked(lpni);
3399 struct lnet_peer *new_lp;
3400 new_lp = lpni->lpni_peer_net->lpn_peer;
3402 * if lp has discovery/MR enabled that means new_lp
3403 * should have discovery/MR enabled as well, since
3404 * it's the same peer, which we're about to merge
3406 spin_lock(&lp->lp_lock);
3407 spin_lock(&new_lp->lp_lock);
3408 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3409 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3410 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3411 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3412 /* If we're processing a ping reply then we may be
3413 * about to send a push to the peer that we ping'd.
3414 * Since the ping reply that we're processing was
3415 * received by lp, we need to set the discovery source
3416 * NID for new_lp to the NID stored in lp.
3418 if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3419 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3420 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3422 spin_unlock(&new_lp->lp_lock);
3423 spin_unlock(&lp->lp_lock);
3425 rc = lnet_peer_set_primary_data(new_lp, pbuf);
3426 lnet_consolidate_routes_locked(lp, new_lp);
3427 lnet_peer_ni_decref_locked(lpni);
3431 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3432 libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3434 mutex_unlock(&the_lnet.ln_api_mutex);
3436 spin_lock(&lp->lp_lock);
3437 /* Tell discovery to re-check the peer immediately. */
3439 rc = LNET_REDISCOVER_PEER;
3444 * A ping failed. Clear the PING_FAILED state and set the
3445 * FORCE_PING state, to ensure a retry even if discovery is
3446 * disabled. This avoids being left with incorrect state.
3448 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3449 __must_hold(&lp->lp_lock)
3451 struct lnet_handle_md mdh;
3454 mdh = lp->lp_ping_mdh;
3455 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3456 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3457 lp->lp_state |= LNET_PEER_FORCE_PING;
3458 rc = lp->lp_ping_error;
3459 lp->lp_ping_error = 0;
3460 spin_unlock(&lp->lp_lock);
3462 if (!LNetMDHandleIsInvalid(mdh))
3465 CDEBUG(D_NET, "peer %s:%d\n",
3466 libcfs_nidstr(&lp->lp_primary_nid), rc);
3468 spin_lock(&lp->lp_lock);
3469 return rc ? rc : LNET_REDISCOVER_PEER;
3472 /* Active side of ping. */
3473 static int lnet_peer_send_ping(struct lnet_peer *lp)
3474 __must_hold(&lp->lp_lock)
3480 lp->lp_state |= LNET_PEER_PING_SENT;
3481 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3482 spin_unlock(&lp->lp_lock);
3484 cpt = lnet_net_lock_current();
3485 /* Refcount for MD. */
3486 lnet_peer_addref_locked(lp);
3487 lnet_net_unlock(cpt);
3489 nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3491 rc = lnet_send_ping(&lp->lp_primary_nid, &lp->lp_ping_mdh, nnis, lp,
3492 the_lnet.ln_dc_handler, false);
3495 * if LNetMDBind in lnet_send_ping fails we need to decrement the
3496 * refcount on the peer, otherwise LNetMDUnlink will be called
3497 * which will eventually do that.
3501 lnet_peer_decref_locked(lp);
3502 lnet_net_unlock(cpt);
3503 rc = -rc; /* change the rc to negative value */
3505 } else if (rc < 0) {
3509 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3511 spin_lock(&lp->lp_lock);
3515 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3517 * The errors that get us here are considered hard errors and
3518 * cause Discovery to terminate. So we clear PING_SENT, but do
3519 * not set either PING_FAILED or FORCE_PING. In fact we need
3520 * to clear PING_FAILED, because the unlink event handler will
3521 * have set it if we called LNetMDUnlink() above.
3523 spin_lock(&lp->lp_lock);
3524 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3529 * This function exists because you cannot call LNetMDUnlink() from an
3532 static int lnet_peer_push_failed(struct lnet_peer *lp)
3533 __must_hold(&lp->lp_lock)
3535 struct lnet_handle_md mdh;
3538 mdh = lp->lp_push_mdh;
3539 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3540 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3541 rc = lp->lp_push_error;
3542 lp->lp_push_error = 0;
3543 spin_unlock(&lp->lp_lock);
3545 if (!LNetMDHandleIsInvalid(mdh))
3548 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3549 spin_lock(&lp->lp_lock);
3550 return rc ? rc : LNET_REDISCOVER_PEER;
3554 * Mark the peer as discovered.
3556 static int lnet_peer_discovered(struct lnet_peer *lp)
3557 __must_hold(&lp->lp_lock)
3559 lp->lp_state |= LNET_PEER_DISCOVERED;
3560 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3561 LNET_PEER_REDISCOVER);
3563 lp->lp_dc_error = 0;
3565 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3570 /* Active side of push. */
3571 static int lnet_peer_send_push(struct lnet_peer *lp)
3572 __must_hold(&lp->lp_lock)
3574 struct lnet_ping_buffer *pbuf;
3575 struct lnet_process_id id;
3580 /* Don't push to a non-multi-rail peer. */
3581 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3582 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3583 /* if peer's NIDs are uptodate then peer is discovered */
3584 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3585 rc = lnet_peer_discovered(lp);
3592 lp->lp_state |= LNET_PEER_PUSH_SENT;
3593 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3594 spin_unlock(&lp->lp_lock);
3596 cpt = lnet_net_lock_current();
3597 pbuf = the_lnet.ln_ping_target;
3598 lnet_ping_buffer_addref(pbuf);
3599 lnet_net_unlock(cpt);
3601 /* Push source MD */
3602 md.start = &pbuf->pb_info;
3603 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3604 md.threshold = 2; /* Put/Ack */
3606 md.options = LNET_MD_TRACK_RESPONSE;
3607 md.handler = the_lnet.ln_dc_handler;
3610 rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3612 lnet_ping_buffer_decref(pbuf);
3613 CERROR("Can't bind push source MD: %d\n", rc);
3617 cpt = lnet_net_lock_current();
3618 /* Refcount for MD. */
3619 lnet_peer_addref_locked(lp);
3620 id.pid = LNET_PID_LUSTRE;
3621 if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3622 id.nid = lnet_nid_to_nid4(&lp->lp_disc_dst_nid);
3624 id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
3625 lnet_net_unlock(cpt);
3627 rc = LNetPut(lnet_nid_to_nid4(&lp->lp_disc_src_nid), lp->lp_push_mdh,
3628 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3629 LNET_PROTO_PING_MATCHBITS, 0, 0);
3632 * reset the discovery nid. There is no need to restrict sending
3633 * from that source, if we call lnet_push_update_to_peers(). It'll
3634 * get set to a specific NID, if we initiate discovery from the
3637 lp->lp_disc_src_nid = LNET_ANY_NID;
3638 lp->lp_disc_dst_nid = LNET_ANY_NID;
3643 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3645 spin_lock(&lp->lp_lock);
3649 LNetMDUnlink(lp->lp_push_mdh);
3650 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3652 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3655 * The errors that get us here are considered hard errors and
3656 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3657 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3658 * because the unlink event handler will have set it if we
3659 * called LNetMDUnlink() above.
3661 spin_lock(&lp->lp_lock);
3662 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3667 * Wait for work to be queued or some other change that must be
3668 * attended to. Returns non-zero if the discovery thread should shut
3671 static int lnet_peer_discovery_wait_for_work(void)
3678 cpt = lnet_net_lock_current();
3680 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3681 TASK_INTERRUPTIBLE);
3682 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3684 if (lnet_push_target_resize_needed() ||
3685 the_lnet.ln_push_target->pb_needs_post)
3687 if (!list_empty(&the_lnet.ln_dc_request))
3689 if (!list_empty(&the_lnet.ln_msg_resend))
3691 lnet_net_unlock(cpt);
3694 * wakeup max every second to check if there are peers that
3695 * have been stuck on the working queue for greater than
3698 schedule_timeout(cfs_time_seconds(1));
3699 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3700 cpt = lnet_net_lock_current();
3702 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3704 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3707 lnet_net_unlock(cpt);
3709 CDEBUG(D_NET, "woken: %d\n", rc);
3715 * Messages that were pending on a destroyed peer will be put on a global
3716 * resend list. The message resend list will be checked by
3717 * the discovery thread when it wakes up, and will resend messages. These
3718 * messages can still be sendable in the case the lpni which was the initial
3719 * cause of the message re-queue was transfered to another peer.
3721 * It is possible that LNet could be shutdown while we're iterating
3722 * through the list. lnet_shudown_lndnets() will attempt to access the
3723 * resend list, but will have to wait until the spinlock is released, by
3724 * which time there shouldn't be any more messages on the resend list.
3725 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3726 * for the messages so they can be released. The other case is that
3727 * lnet_shudown_lndnets() can finalize all the messages before this
3728 * function can visit the resend list, in which case this function will be
3731 static void lnet_resend_msgs(void)
3733 struct lnet_msg *msg, *tmp;
3737 spin_lock(&the_lnet.ln_msg_resend_lock);
3738 list_splice(&the_lnet.ln_msg_resend, &resend);
3739 spin_unlock(&the_lnet.ln_msg_resend_lock);
3741 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3742 list_del_init(&msg->msg_list);
3743 rc = lnet_send(&msg->msg_src_nid_param, msg,
3744 &msg->msg_rtr_nid_param);
3746 CNETERR("Error sending %s to %s: %d\n",
3747 lnet_msgtyp2str(msg->msg_type),
3748 libcfs_idstr(&msg->msg_target), rc);
3749 lnet_finalize(msg, rc);
3754 /* The discovery thread. */
3755 static int lnet_peer_discovery(void *arg)
3757 struct lnet_peer *lp;
3760 wait_for_completion(&the_lnet.ln_started);
3762 CDEBUG(D_NET, "started\n");
3765 if (lnet_peer_discovery_wait_for_work())
3768 if (lnet_push_target_resize_needed())
3769 lnet_push_target_resize();
3770 else if (the_lnet.ln_push_target->pb_needs_post)
3771 lnet_push_target_post(the_lnet.ln_push_target,
3772 &the_lnet.ln_push_target_md);
3776 lnet_net_lock(LNET_LOCK_EX);
3777 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3778 lnet_net_unlock(LNET_LOCK_EX);
3783 * Process all incoming discovery work requests. When
3784 * discovery must wait on a peer to change state, it
3785 * is added to the tail of the ln_dc_working queue. A
3786 * timestamp keeps track of when the peer was added,
3787 * so we can time out discovery requests that take too
3790 while (!list_empty(&the_lnet.ln_dc_request)) {
3791 lp = list_first_entry(&the_lnet.ln_dc_request,
3792 struct lnet_peer, lp_dc_list);
3793 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3795 * set the time the peer was put on the dc_working
3796 * queue. It shouldn't remain on the queue
3797 * forever, in case the GET message (for ping)
3798 * doesn't get a REPLY or the PUT message (for
3799 * push) doesn't get an ACK.
3801 lp->lp_last_queued = ktime_get_real_seconds();
3802 lnet_net_unlock(LNET_LOCK_EX);
3804 if (lnet_push_target_resize_needed())
3805 lnet_push_target_resize();
3806 else if (the_lnet.ln_push_target->pb_needs_post)
3807 lnet_push_target_post(the_lnet.ln_push_target,
3808 &the_lnet.ln_push_target_md);
3811 * Select an action depending on the state of
3812 * the peer and whether discovery is disabled.
3813 * The check whether discovery is disabled is
3814 * done after the code that handles processing
3815 * for arrived data, cleanup for failures, and
3816 * forcing a Ping or Push.
3818 spin_lock(&lp->lp_lock);
3819 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3820 libcfs_nidstr(&lp->lp_primary_nid), lp,
3822 if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3823 LNET_PEER_MARK_DELETED))
3824 rc = lnet_peer_deletion(lp);
3825 else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3826 rc = lnet_peer_data_present(lp);
3827 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3828 rc = lnet_peer_ping_failed(lp);
3829 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3830 rc = lnet_peer_push_failed(lp);
3831 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3832 rc = lnet_peer_send_ping(lp);
3833 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3834 rc = lnet_peer_send_push(lp);
3835 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3836 rc = lnet_peer_send_ping(lp);
3837 else if (lnet_peer_needs_push(lp))
3838 rc = lnet_peer_send_push(lp);
3840 rc = lnet_peer_discovered(lp);
3841 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3842 libcfs_nidstr(&lp->lp_primary_nid), lp,
3845 if (rc == LNET_REDISCOVER_PEER) {
3846 spin_unlock(&lp->lp_lock);
3847 lnet_net_lock(LNET_LOCK_EX);
3848 list_move(&lp->lp_dc_list,
3849 &the_lnet.ln_dc_request);
3851 !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3852 spin_unlock(&lp->lp_lock);
3853 lnet_net_lock(LNET_LOCK_EX);
3854 lnet_peer_discovery_complete(lp, rc);
3856 spin_unlock(&lp->lp_lock);
3857 lnet_net_lock(LNET_LOCK_EX);
3860 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3865 lnet_net_unlock(LNET_LOCK_EX);
3868 CDEBUG(D_NET, "stopping\n");
3870 * Clean up before telling lnet_peer_discovery_stop() that
3871 * we're done. Use wake_up() below to somewhat reduce the
3872 * size of the thundering herd if there are multiple threads
3873 * waiting on discovery of a single peer.
3876 /* Queue cleanup 1: stop all pending pings and pushes. */
3877 lnet_net_lock(LNET_LOCK_EX);
3878 while (!list_empty(&the_lnet.ln_dc_working)) {
3879 lp = list_first_entry(&the_lnet.ln_dc_working,
3880 struct lnet_peer, lp_dc_list);
3881 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3882 lnet_net_unlock(LNET_LOCK_EX);
3883 lnet_peer_cancel_discovery(lp);
3884 lnet_net_lock(LNET_LOCK_EX);
3886 lnet_net_unlock(LNET_LOCK_EX);
3888 /* Queue cleanup 2: wait for the expired queue to clear. */
3889 while (!list_empty(&the_lnet.ln_dc_expired))
3890 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3892 /* Queue cleanup 3: clear the request queue. */
3893 lnet_net_lock(LNET_LOCK_EX);
3894 while (!list_empty(&the_lnet.ln_dc_request)) {
3895 lp = list_first_entry(&the_lnet.ln_dc_request,
3896 struct lnet_peer, lp_dc_list);
3897 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3899 lnet_net_unlock(LNET_LOCK_EX);
3901 lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3902 the_lnet.ln_dc_handler = NULL;
3904 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3905 wake_up(&the_lnet.ln_dc_waitq);
3907 CDEBUG(D_NET, "stopped\n");
3912 /* ln_api_mutex is held on entry. */
3913 int lnet_peer_discovery_start(void)
3915 struct task_struct *task;
3918 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3921 the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3922 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3923 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3926 CERROR("Can't start peer discovery thread: %d\n", rc);
3928 the_lnet.ln_dc_handler = NULL;
3930 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3933 CDEBUG(D_NET, "discovery start: %d\n", rc);
3938 /* ln_api_mutex is held on entry. */
3939 void lnet_peer_discovery_stop(void)
3941 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3944 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3945 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3947 /* In the LNetNIInit() path we may be stopping discovery before it
3948 * entered its work loop
3950 if (!completion_done(&the_lnet.ln_started))
3951 complete(&the_lnet.ln_started);
3953 wake_up(&the_lnet.ln_dc_waitq);
3955 wait_event(the_lnet.ln_dc_waitq,
3956 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3958 LASSERT(list_empty(&the_lnet.ln_dc_request));
3959 LASSERT(list_empty(&the_lnet.ln_dc_working));
3960 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3962 CDEBUG(D_NET, "discovery stopped\n");
3968 lnet_debug_peer(lnet_nid_t nid)
3970 char *aliveness = "NA";
3971 struct lnet_peer_ni *lp;
3974 cpt = lnet_cpt_of_nid(nid, NULL);
3977 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3979 lnet_net_unlock(cpt);
3980 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3984 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3985 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3987 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3988 libcfs_nidstr(&lp->lpni_nid), kref_read(&lp->lpni_kref),
3989 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3990 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3991 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3993 lnet_peer_ni_decref_locked(lp);
3995 lnet_net_unlock(cpt);
3998 /* Gathering information for userspace. */
4000 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
4001 char aliveness[LNET_MAX_STR_LEN],
4002 __u32 *cpt_iter, __u32 *refcount,
4003 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
4004 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
4005 __u32 *peer_tx_qnob)
4007 struct lnet_peer_table *peer_table;
4008 struct lnet_peer_ni *lp;
4013 /* get the number of CPTs */
4014 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4016 /* if the cpt number to be examined is >= the number of cpts in
4017 * the system then indicate that there are no more cpts to examin
4019 if (*cpt_iter >= lncpt)
4022 /* get the current table */
4023 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
4024 /* if the ptable is NULL then there are no more cpts to examine */
4025 if (peer_table == NULL)
4028 lnet_net_lock(*cpt_iter);
4030 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
4031 struct list_head *peers = &peer_table->pt_hash[j];
4033 list_for_each_entry(lp, peers, lpni_hashlist) {
4034 if (!nid_is_nid4(&lp->lpni_nid))
4036 if (peer_index-- > 0)
4039 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
4040 if (lnet_isrouter(lp) ||
4041 lnet_peer_aliveness_enabled(lp))
4042 snprintf(aliveness, LNET_MAX_STR_LEN,
4043 lnet_is_peer_ni_alive(lp) ? "up" : "down");
4045 *nid = lnet_nid_to_nid4(&lp->lpni_nid);
4046 *refcount = kref_read(&lp->lpni_kref);
4047 *ni_peer_tx_credits =
4048 lp->lpni_net->net_tunables.lct_peer_tx_credits;
4049 *peer_tx_credits = lp->lpni_txcredits;
4050 *peer_rtr_credits = lp->lpni_rtrcredits;
4051 *peer_min_rtr_credits = lp->lpni_mintxcredits;
4052 *peer_tx_qnob = lp->lpni_txqnob;
4058 lnet_net_unlock(*cpt_iter);
4062 return found ? 0 : -ENOENT;
4065 /* ln_api_mutex is held, which keeps the peer list stable */
4066 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
4068 struct lnet_ioctl_element_stats *lpni_stats;
4069 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
4070 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
4071 struct lnet_peer_ni_credit_info *lpni_info;
4072 struct lnet_peer_ni *lpni;
4073 struct lnet_peer *lp;
4078 lp = lnet_find_peer4(cfg->prcfg_prim_nid);
4085 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
4086 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
4087 size *= lp->lp_nnis;
4088 if (size > cfg->prcfg_size) {
4089 cfg->prcfg_size = size;
4094 cfg->prcfg_prim_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4095 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
4096 cfg->prcfg_cfg_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4097 cfg->prcfg_count = lp->lp_nnis;
4098 cfg->prcfg_size = size;
4099 cfg->prcfg_state = lp->lp_state;
4101 /* Allocate helper buffers. */
4103 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
4106 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
4109 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
4110 if (!lpni_msg_stats)
4111 goto out_free_stats;
4112 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
4114 goto out_free_msg_stats;
4119 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
4120 if (!nid_is_nid4(&lpni->lpni_nid))
4122 nid = lnet_nid_to_nid4(&lpni->lpni_nid);
4123 if (copy_to_user(bulk, &nid, sizeof(nid)))
4124 goto out_free_hstats;
4125 bulk += sizeof(nid);
4127 memset(lpni_info, 0, sizeof(*lpni_info));
4128 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
4129 if (lnet_isrouter(lpni) ||
4130 lnet_peer_aliveness_enabled(lpni))
4131 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
4132 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
4134 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
4135 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
4136 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
4137 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
4138 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
4139 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
4140 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
4141 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4142 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4143 goto out_free_hstats;
4144 bulk += sizeof(*lpni_info);
4146 memset(lpni_stats, 0, sizeof(*lpni_stats));
4147 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4148 LNET_STATS_TYPE_SEND);
4149 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4150 LNET_STATS_TYPE_RECV);
4151 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4152 LNET_STATS_TYPE_DROP);
4153 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4154 goto out_free_hstats;
4155 bulk += sizeof(*lpni_stats);
4156 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4157 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4158 goto out_free_hstats;
4159 bulk += sizeof(*lpni_msg_stats);
4160 lpni_hstats->hlpni_network_timeout =
4161 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4162 lpni_hstats->hlpni_remote_dropped =
4163 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4164 lpni_hstats->hlpni_remote_timeout =
4165 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4166 lpni_hstats->hlpni_remote_error =
4167 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4168 lpni_hstats->hlpni_health_value =
4169 atomic_read(&lpni->lpni_healthv);
4170 lpni_hstats->hlpni_ping_count = lpni->lpni_ping_count;
4171 lpni_hstats->hlpni_next_ping = lpni->lpni_next_ping;
4172 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4173 goto out_free_hstats;
4174 bulk += sizeof(*lpni_hstats);
4179 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4181 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4183 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4185 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4187 lnet_peer_decref_locked(lp);
4192 /* must hold net_lock/0 */
4194 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4195 struct list_head *recovery_queue,
4198 /* the mt could've shutdown and cleaned up the queues */
4199 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4202 if (!list_empty(&lpni->lpni_recovery))
4205 if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4208 if (!lpni->lpni_last_alive) {
4210 "lpni %s(%p) not eligible for recovery last alive %lld\n",
4211 libcfs_nidstr(&lpni->lpni_nid), lpni,
4212 lpni->lpni_last_alive);
4216 if (lnet_recovery_limit &&
4217 now > lpni->lpni_last_alive + lnet_recovery_limit) {
4218 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4219 libcfs_nidstr(&lpni->lpni_nid),
4220 lpni->lpni_last_alive);
4221 /* Reset the ping count so that if this peer NI is added back to
4222 * the recovery queue we will send the first ping right away.
4224 lpni->lpni_ping_count = 0;
4228 /* This peer NI is going on the recovery queue, so take a ref on it */
4229 lnet_peer_ni_addref_locked(lpni);
4231 lnet_peer_ni_set_next_ping(lpni, now);
4233 CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4234 libcfs_nidstr(&lpni->lpni_nid),
4235 lpni->lpni_ping_count,
4236 lpni->lpni_next_ping,
4237 lpni->lpni_last_alive,
4238 atomic_read(&lpni->lpni_healthv));
4240 list_add_tail(&lpni->lpni_recovery, recovery_queue);
4243 /* Call with the ln_api_mutex held */
4245 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
4247 struct lnet_peer_table *ptable;
4248 struct lnet_peer *lp;
4249 struct lnet_peer_net *lpn;
4250 struct lnet_peer_ni *lpni;
4255 if (the_lnet.ln_state != LNET_STATE_RUNNING)
4258 now = ktime_get_seconds();
4261 lnet_net_lock(LNET_LOCK_EX);
4262 lpni = lnet_find_peer_ni_locked(nid);
4264 lnet_net_unlock(LNET_LOCK_EX);
4267 lnet_set_lpni_healthv_locked(lpni, value);
4268 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4269 &the_lnet.ln_mt_peerNIRecovq, now);
4270 lnet_peer_ni_decref_locked(lpni);
4271 lnet_net_unlock(LNET_LOCK_EX);
4275 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4278 * Walk all the peers and reset the health value for each one to the
4281 lnet_net_lock(LNET_LOCK_EX);
4282 for (cpt = 0; cpt < lncpt; cpt++) {
4283 ptable = the_lnet.ln_peer_tables[cpt];
4284 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4285 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4286 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4288 lnet_set_lpni_healthv_locked(lpni,
4290 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4291 &the_lnet.ln_mt_peerNIRecovq, now);
4296 lnet_net_unlock(LNET_LOCK_EX);