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_NIDNET(lpni->lpni_nid) == net->net_id) {
71 spin_lock(&lpni->lpni_lock);
72 lpni->lpni_txcredits =
73 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
74 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
75 lpni->lpni_rtrcredits =
76 lnet_peer_buffer_credits(lpni->lpni_net);
77 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
78 spin_unlock(&lpni->lpni_lock);
80 lnet_peer_remove_from_remote_list(lpni);
86 lnet_peer_tables_destroy(void)
88 struct lnet_peer_table *ptable;
89 struct list_head *hash;
93 if (!the_lnet.ln_peer_tables)
96 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
97 hash = ptable->pt_hash;
98 if (!hash) /* not intialized */
101 LASSERT(list_empty(&ptable->pt_zombie_list));
103 ptable->pt_hash = NULL;
104 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
105 LASSERT(list_empty(&hash[j]));
107 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
110 cfs_percpt_free(the_lnet.ln_peer_tables);
111 the_lnet.ln_peer_tables = NULL;
115 lnet_peer_tables_create(void)
117 struct lnet_peer_table *ptable;
118 struct list_head *hash;
122 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
124 if (the_lnet.ln_peer_tables == NULL) {
125 CERROR("Failed to allocate cpu-partition peer tables\n");
129 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131 LNET_PEER_HASH_SIZE * sizeof(*hash));
133 CERROR("Failed to create peer hash table\n");
134 lnet_peer_tables_destroy();
138 spin_lock_init(&ptable->pt_zombie_lock);
139 INIT_LIST_HEAD(&ptable->pt_zombie_list);
141 INIT_LIST_HEAD(&ptable->pt_peer_list);
143 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144 INIT_LIST_HEAD(&hash[j]);
145 ptable->pt_hash = hash; /* sign of initialization */
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(lnet_nid_t nid)
154 struct lnet_peer_ni *lpni;
155 struct lnet_net *net;
158 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
160 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
164 INIT_LIST_HEAD(&lpni->lpni_txq);
165 INIT_LIST_HEAD(&lpni->lpni_hashlist);
166 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167 INIT_LIST_HEAD(&lpni->lpni_recovery);
168 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169 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_NIDNET(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_nid2str(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(lnet_nid_t 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_NID_ANY;
266 if (lnet_peers_start_down())
267 lp->lp_alive = false;
272 * all peers created on a router should have health on
273 * if it's not already on.
275 if (the_lnet.ln_routing && !lnet_health_sensitivity)
276 lp->lp_health_sensitivity = 1;
279 * Turn off discovery for loopback peer. If you're creating a peer
280 * for the loopback interface then that was initiated when we
281 * attempted to send a message over the loopback. There is no need
282 * to ever use a different interface when sending messages to
285 if (nid == LNET_NID_LO_0)
286 lp->lp_state = LNET_PEER_NO_DISCOVERY;
287 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
289 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
295 lnet_destroy_peer_locked(struct lnet_peer *lp)
297 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
299 LASSERT(atomic_read(&lp->lp_refcount) == 0);
300 LASSERT(lp->lp_rtr_refcount == 0);
301 LASSERT(list_empty(&lp->lp_peer_nets));
302 LASSERT(list_empty(&lp->lp_peer_list));
303 LASSERT(list_empty(&lp->lp_dc_list));
306 lnet_ping_buffer_decref(lp->lp_data);
309 * if there are messages still on the pending queue, then make
310 * sure to queue them on the ln_msg_resend list so they can be
311 * resent at a later point if the discovery thread is still
313 * If the discovery thread has stopped, then the wakeup will be a
314 * no-op, and it is expected the lnet_shutdown_lndnets() will
315 * eventually be called, which will traverse this list and
316 * finalize the messages on the list.
317 * We can not resend them now because we're holding the cpt lock.
318 * Releasing the lock can cause an inconsistent state
320 spin_lock(&the_lnet.ln_msg_resend_lock);
321 spin_lock(&lp->lp_lock);
322 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
323 spin_unlock(&lp->lp_lock);
324 spin_unlock(&the_lnet.ln_msg_resend_lock);
325 wake_up(&the_lnet.ln_dc_waitq);
327 LIBCFS_FREE(lp, sizeof(*lp));
331 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
332 * that peer_net, detach the peer_net from the peer.
334 * Call with lnet_net_lock/EX held
337 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
339 struct lnet_peer_table *ptable;
340 struct lnet_peer_net *lpn;
341 struct lnet_peer *lp;
344 * Belts and suspenders: gracefully handle teardown of a
345 * partially connected peer_ni.
347 lpn = lpni->lpni_peer_net;
349 list_del_init(&lpni->lpni_peer_nis);
351 * If there are no lpni's left, we detach lpn from
352 * lp_peer_nets, so it cannot be found anymore.
354 if (list_empty(&lpn->lpn_peer_nis))
355 list_del_init(&lpn->lpn_peer_nets);
357 /* Update peer NID count. */
362 * If there are no more peer nets, make the peer unfindable
363 * via the peer_tables.
365 * Otherwise, if the peer is DISCOVERED, tell discovery to
366 * take another look at it. This is a no-op if discovery for
367 * this peer did the detaching.
369 if (list_empty(&lp->lp_peer_nets)) {
370 list_del_init(&lp->lp_peer_list);
371 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
373 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
374 /* Discovery isn't running, nothing to do here. */
375 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
376 lnet_peer_queue_for_discovery(lp);
377 wake_up(&the_lnet.ln_dc_waitq);
379 CDEBUG(D_NET, "peer %s NID %s\n",
380 libcfs_nid2str(lp->lp_primary_nid),
381 libcfs_nid2str(lpni->lpni_nid));
384 /* called with lnet_net_lock LNET_LOCK_EX held */
386 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
388 struct lnet_peer_table *ptable = NULL;
390 /* don't remove a peer_ni if it's also a gateway */
391 if (lnet_isrouter(lpni) && !force) {
392 CERROR("Peer NI %s is a gateway. Can not delete it\n",
393 libcfs_nid2str(lpni->lpni_nid));
397 lnet_peer_remove_from_remote_list(lpni);
399 /* remove peer ni from the hash list. */
400 list_del_init(&lpni->lpni_hashlist);
403 * indicate the peer is being deleted so the monitor thread can
404 * remove it from the recovery queue.
406 spin_lock(&lpni->lpni_lock);
407 lpni->lpni_state |= LNET_PEER_NI_DELETING;
408 spin_unlock(&lpni->lpni_lock);
410 /* decrement the ref count on the peer table */
411 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
414 * The peer_ni can no longer be found with a lookup. But there
415 * can be current users, so keep track of it on the zombie
416 * list until the reference count has gone to zero.
418 * The last reference may be lost in a place where the
419 * lnet_net_lock locks only a single cpt, and that cpt may not
420 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
423 spin_lock(&ptable->pt_zombie_lock);
424 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
425 ptable->pt_zombies++;
426 spin_unlock(&ptable->pt_zombie_lock);
428 /* no need to keep this peer_ni on the hierarchy anymore */
429 lnet_peer_detach_peer_ni_locked(lpni);
431 /* remove hashlist reference on peer_ni */
432 lnet_peer_ni_decref_locked(lpni);
437 void lnet_peer_uninit(void)
439 struct lnet_peer_ni *lpni, *tmp;
441 lnet_net_lock(LNET_LOCK_EX);
443 /* remove all peer_nis from the remote peer and the hash list */
444 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
445 lpni_on_remote_peer_ni_list)
446 lnet_peer_ni_del_locked(lpni, false);
448 lnet_peer_tables_destroy();
450 lnet_net_unlock(LNET_LOCK_EX);
454 lnet_peer_del_locked(struct lnet_peer *peer)
456 struct lnet_peer_ni *lpni = NULL, *lpni2;
459 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
461 spin_lock(&peer->lp_lock);
462 peer->lp_state |= LNET_PEER_MARK_DELETED;
463 spin_unlock(&peer->lp_lock);
465 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
466 while (lpni != NULL) {
467 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
468 rc = lnet_peer_ni_del_locked(lpni, false);
478 * Discovering this peer is taking too long. Cancel any Ping or Push
479 * that discovery is waiting on by unlinking the relevant MDs. The
480 * lnet_discovery_event_handler() will proceed from here and complete
483 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
485 struct lnet_handle_md ping_mdh;
486 struct lnet_handle_md push_mdh;
488 LNetInvalidateMDHandle(&ping_mdh);
489 LNetInvalidateMDHandle(&push_mdh);
491 spin_lock(&lp->lp_lock);
492 if (lp->lp_state & LNET_PEER_PING_SENT) {
493 ping_mdh = lp->lp_ping_mdh;
494 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
496 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
497 push_mdh = lp->lp_push_mdh;
498 LNetInvalidateMDHandle(&lp->lp_push_mdh);
500 spin_unlock(&lp->lp_lock);
502 if (!LNetMDHandleIsInvalid(ping_mdh))
503 LNetMDUnlink(ping_mdh);
504 if (!LNetMDHandleIsInvalid(push_mdh))
505 LNetMDUnlink(push_mdh);
509 lnet_peer_del(struct lnet_peer *peer)
511 lnet_peer_cancel_discovery(peer);
512 lnet_net_lock(LNET_LOCK_EX);
513 lnet_peer_del_locked(peer);
514 lnet_net_unlock(LNET_LOCK_EX);
520 * Delete a NID from a peer. Call with ln_api_mutex held.
523 * -EPERM: Non-DLC deletion from DLC-configured peer.
524 * -ENOENT: No lnet_peer_ni corresponding to the nid.
525 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
526 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
529 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
531 struct lnet_peer_ni *lpni;
532 lnet_nid_t primary_nid = lp->lp_primary_nid;
534 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
536 if (!(flags & LNET_PEER_CONFIGURED)) {
537 if (lp->lp_state & LNET_PEER_CONFIGURED) {
542 lpni = lnet_find_peer_ni_locked(nid);
547 lnet_peer_ni_decref_locked(lpni);
548 if (lp != lpni->lpni_peer_net->lpn_peer) {
554 * This function only allows deletion of the primary NID if it
557 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
562 lnet_net_lock(LNET_LOCK_EX);
564 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
565 struct lnet_peer_ni *lpni2;
566 /* assign the next peer_ni to be the primary */
567 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
569 lp->lp_primary_nid = lpni2->lpni_nid;
571 rc = lnet_peer_ni_del_locked(lpni, force);
573 lnet_net_unlock(LNET_LOCK_EX);
576 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
577 libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
583 lnet_peer_table_cleanup_locked(struct lnet_net *net,
584 struct lnet_peer_table *ptable)
587 struct lnet_peer_ni *next;
588 struct lnet_peer_ni *lpni;
589 struct lnet_peer *peer;
591 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
592 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
594 if (net != NULL && net != lpni->lpni_net)
597 peer = lpni->lpni_peer_net->lpn_peer;
598 if (peer->lp_primary_nid != lpni->lpni_nid) {
599 lnet_peer_ni_del_locked(lpni, false);
603 * Removing the primary NID implies removing
604 * the entire peer. Advance next beyond any
605 * peer_ni that belongs to the same peer.
607 list_for_each_entry_from(next, &ptable->pt_hash[i],
609 if (next->lpni_peer_net->lpn_peer != peer)
612 lnet_peer_del_locked(peer);
618 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
620 wait_var_event_warning(&ptable->pt_zombies,
621 ptable->pt_zombies == 0,
622 "Waiting for %d zombies on peer table\n",
627 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
628 struct lnet_peer_table *ptable)
630 struct lnet_peer_ni *lp;
631 struct lnet_peer_ni *tmp;
635 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
636 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
638 if (net != lp->lpni_net)
641 if (!lnet_isrouter(lp))
644 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
646 lnet_net_unlock(LNET_LOCK_EX);
647 lnet_del_route(LNET_NET_ANY, gw_nid);
648 lnet_net_lock(LNET_LOCK_EX);
654 lnet_peer_tables_cleanup(struct lnet_net *net)
657 struct lnet_peer_table *ptable;
659 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
660 /* If just deleting the peers for a NI, get rid of any routes these
661 * peers are gateways for. */
662 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
663 lnet_net_lock(LNET_LOCK_EX);
664 lnet_peer_table_del_rtrs_locked(net, ptable);
665 lnet_net_unlock(LNET_LOCK_EX);
668 /* Start the cleanup process */
669 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
670 lnet_net_lock(LNET_LOCK_EX);
671 lnet_peer_table_cleanup_locked(net, ptable);
672 lnet_net_unlock(LNET_LOCK_EX);
675 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
676 lnet_peer_ni_finalize_wait(ptable);
679 static struct lnet_peer_ni *
680 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
682 struct list_head *peers;
683 struct lnet_peer_ni *lp;
685 if (the_lnet.ln_state != LNET_STATE_RUNNING)
688 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
689 list_for_each_entry(lp, peers, lpni_hashlist) {
690 if (lp->lpni_nid == nid) {
691 lnet_peer_ni_addref_locked(lp);
699 struct lnet_peer_ni *
700 lnet_find_peer_ni_locked(lnet_nid_t nid)
702 struct lnet_peer_ni *lpni;
703 struct lnet_peer_table *ptable;
706 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
708 ptable = the_lnet.ln_peer_tables[cpt];
709 lpni = lnet_get_peer_ni_locked(ptable, nid);
714 struct lnet_peer_ni *
715 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
717 struct lnet_peer_net *lpn;
718 struct lnet_peer_ni *lpni;
720 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
724 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
725 if (lpni->lpni_nid == nid)
733 lnet_find_peer(lnet_nid_t nid)
735 struct lnet_peer_ni *lpni;
736 struct lnet_peer *lp = NULL;
739 cpt = lnet_net_lock_current();
740 lpni = lnet_find_peer_ni_locked(nid);
742 lp = lpni->lpni_peer_net->lpn_peer;
743 lnet_peer_addref_locked(lp);
744 lnet_peer_ni_decref_locked(lpni);
746 lnet_net_unlock(cpt);
751 struct lnet_peer_net *
752 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
754 struct lnet_peer_net *net;
757 /* no net id provided return the first net */
758 net = list_first_entry_or_null(&lp->lp_peer_nets,
759 struct lnet_peer_net,
765 /* find the net after the one provided */
766 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
767 if (net->lpn_net_id == prev_lpn_id) {
769 * if we reached the end of the list loop to the
772 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
773 return list_first_entry_or_null(&lp->lp_peer_nets,
774 struct lnet_peer_net,
777 return list_next_entry(net, lpn_peer_nets);
784 struct lnet_peer_ni *
785 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
786 struct lnet_peer_net *peer_net,
787 struct lnet_peer_ni *prev)
789 struct lnet_peer_ni *lpni;
790 struct lnet_peer_net *net = peer_net;
794 if (list_empty(&peer->lp_peer_nets))
797 net = list_entry(peer->lp_peer_nets.next,
798 struct lnet_peer_net,
801 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
807 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
809 * if you reached the end of the peer ni list and the peer
810 * net is specified then there are no more peer nis in that
817 * we reached the end of this net ni list. move to the
820 if (prev->lpni_peer_net->lpn_peer_nets.next ==
822 /* no more nets and no more NIs. */
825 /* get the next net */
826 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
827 struct lnet_peer_net,
829 /* get the ni on it */
830 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
836 /* there are more nis left */
837 lpni = list_entry(prev->lpni_peer_nis.next,
838 struct lnet_peer_ni, lpni_peer_nis);
843 /* Call with the ln_api_mutex held */
844 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
846 struct lnet_process_id id;
847 struct lnet_peer_table *ptable;
848 struct lnet_peer *lp;
857 if (the_lnet.ln_state != LNET_STATE_RUNNING)
860 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
863 * Count the number of peers, and return E2BIG if the buffer
864 * is too small. We'll also return the desired size.
867 for (cpt = 0; cpt < lncpt; cpt++) {
868 ptable = the_lnet.ln_peer_tables[cpt];
869 count += ptable->pt_peers;
871 size = count * sizeof(*ids);
876 * Walk the peer lists and copy out the primary nids.
877 * This is safe because the peer lists are only modified
878 * while the ln_api_mutex is held. So we don't need to
879 * hold the lnet_net_lock as well, and can therefore
880 * directly call copy_to_user().
883 memset(&id, 0, sizeof(id));
884 id.pid = LNET_PID_LUSTRE;
886 for (cpt = 0; cpt < lncpt; cpt++) {
887 ptable = the_lnet.ln_peer_tables[cpt];
888 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
891 id.nid = lp->lp_primary_nid;
892 if (copy_to_user(&ids[i], &id, sizeof(id)))
905 * Start pushes to peers that need to be updated for a configuration
906 * change on this node.
909 lnet_push_update_to_peers(int force)
911 struct lnet_peer_table *ptable;
912 struct lnet_peer *lp;
916 lnet_net_lock(LNET_LOCK_EX);
917 if (lnet_peer_discovery_disabled)
919 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
920 for (cpt = 0; cpt < lncpt; cpt++) {
921 ptable = the_lnet.ln_peer_tables[cpt];
922 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
924 spin_lock(&lp->lp_lock);
925 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
926 lp->lp_state |= LNET_PEER_FORCE_PUSH;
927 spin_unlock(&lp->lp_lock);
929 if (lnet_peer_needs_push(lp))
930 lnet_peer_queue_for_discovery(lp);
933 lnet_net_unlock(LNET_LOCK_EX);
934 wake_up(&the_lnet.ln_dc_waitq);
937 /* find the NID in the preferred gateways for the remote peer
939 * false: list is not empty and NID is not preferred
940 * false: list is empty
941 * true: nid is found in the list
944 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
947 struct lnet_nid_list *ne;
949 CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
950 libcfs_nid2str(lpni->lpni_nid),
951 list_empty(&lpni->lpni_rtr_pref_nids));
953 if (list_empty(&lpni->lpni_rtr_pref_nids))
956 /* iterate through all the preferred NIDs and see if any of them
957 * matches the provided gw_nid
959 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
960 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
961 libcfs_nid2str(ne->nl_nid),
962 libcfs_nid2str(gw_nid));
963 if (ne->nl_nid == gw_nid)
971 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
973 struct list_head zombies;
974 struct lnet_nid_list *ne;
975 struct lnet_nid_list *tmp;
976 int cpt = lpni->lpni_cpt;
978 INIT_LIST_HEAD(&zombies);
981 list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
982 lnet_net_unlock(cpt);
984 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
985 list_del(&ne->nl_list);
986 LIBCFS_FREE(ne, sizeof(*ne));
991 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
994 int cpt = lpni->lpni_cpt;
995 struct lnet_nid_list *ne = NULL;
997 /* This function is called with api_mutex held. When the api_mutex
998 * is held the list can not be modified, as it is only modified as
999 * a result of applying a UDSP and that happens under api_mutex
1002 __must_hold(&the_lnet.ln_api_mutex);
1004 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1005 if (ne->nl_nid == gw_nid)
1009 LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1013 ne->nl_nid = gw_nid;
1015 /* Lock the cpt to protect against addition and checks in the
1016 * selection algorithm
1019 list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1020 lnet_net_unlock(cpt);
1026 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1027 * this is a preferred point-to-point path. Call with lnet_net_lock in
1031 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1033 struct lnet_nid_list *ne;
1035 if (lpni->lpni_pref_nnids == 0)
1037 if (lpni->lpni_pref_nnids == 1)
1038 return lpni->lpni_pref.nid == nid;
1039 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1040 if (ne->nl_nid == nid)
1047 * Set a single ni as preferred, provided no preferred ni is already
1048 * defined. Only to be used for non-multi-rail peer_ni.
1051 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1055 spin_lock(&lpni->lpni_lock);
1056 if (nid == LNET_NID_ANY) {
1058 } else if (lpni->lpni_pref_nnids > 0) {
1060 } else if (lpni->lpni_pref_nnids == 0) {
1061 lpni->lpni_pref.nid = nid;
1062 lpni->lpni_pref_nnids = 1;
1063 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1065 spin_unlock(&lpni->lpni_lock);
1067 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1068 libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
1073 * Clear the preferred NID from a non-multi-rail peer_ni, provided
1074 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1077 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1081 spin_lock(&lpni->lpni_lock);
1082 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1083 lpni->lpni_pref_nnids = 0;
1084 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1085 } else if (lpni->lpni_pref_nnids == 0) {
1090 spin_unlock(&lpni->lpni_lock);
1092 CDEBUG(D_NET, "peer %s: %d\n",
1093 libcfs_nid2str(lpni->lpni_nid), rc);
1098 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1100 lpni->lpni_sel_priority = priority;
1104 * Clear the preferred NIDs from a non-multi-rail peer.
1107 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1109 struct lnet_peer_ni *lpni = NULL;
1111 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1112 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1116 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1118 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1119 struct lnet_nid_list *ne1 = NULL;
1120 struct lnet_nid_list *ne2 = NULL;
1121 lnet_nid_t tmp_nid = LNET_NID_ANY;
1124 if (nid == LNET_NID_ANY) {
1129 if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1134 /* A non-MR node may have only one preferred NI per peer_ni */
1135 if (lpni->lpni_pref_nnids > 0 &&
1136 !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1141 /* add the new preferred nid to the list of preferred nids */
1142 if (lpni->lpni_pref_nnids != 0) {
1143 size_t alloc_size = sizeof(*ne1);
1145 if (lpni->lpni_pref_nnids == 1) {
1146 tmp_nid = lpni->lpni_pref.nid;
1147 INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1150 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1151 if (ne1->nl_nid == nid) {
1157 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1164 /* move the originally stored nid to the list */
1165 if (lpni->lpni_pref_nnids == 1) {
1166 LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1167 lpni->lpni_cpt, alloc_size);
1172 INIT_LIST_HEAD(&ne2->nl_list);
1173 ne2->nl_nid = tmp_nid;
1178 lnet_net_lock(LNET_LOCK_EX);
1179 spin_lock(&lpni->lpni_lock);
1180 if (lpni->lpni_pref_nnids == 0) {
1181 lpni->lpni_pref.nid = nid;
1184 list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1185 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1187 lpni->lpni_pref_nnids++;
1188 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1189 spin_unlock(&lpni->lpni_lock);
1190 lnet_net_unlock(LNET_LOCK_EX);
1193 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1194 spin_lock(&lpni->lpni_lock);
1195 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1196 spin_unlock(&lpni->lpni_lock);
1198 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1199 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1204 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1206 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1207 struct lnet_nid_list *ne = NULL;
1210 if (lpni->lpni_pref_nnids == 0) {
1215 if (lpni->lpni_pref_nnids == 1) {
1216 if (lpni->lpni_pref.nid != nid) {
1221 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1222 if (ne->nl_nid == nid)
1223 goto remove_nid_entry;
1231 lnet_net_lock(LNET_LOCK_EX);
1232 spin_lock(&lpni->lpni_lock);
1233 if (lpni->lpni_pref_nnids == 1)
1234 lpni->lpni_pref.nid = LNET_NID_ANY;
1236 list_del_init(&ne->nl_list);
1237 if (lpni->lpni_pref_nnids == 2) {
1238 struct lnet_nid_list *ne, *tmp;
1240 list_for_each_entry_safe(ne, tmp,
1241 &lpni->lpni_pref.nids,
1243 lpni->lpni_pref.nid = ne->nl_nid;
1244 list_del_init(&ne->nl_list);
1245 LIBCFS_FREE(ne, sizeof(*ne));
1249 lpni->lpni_pref_nnids--;
1250 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1251 spin_unlock(&lpni->lpni_lock);
1252 lnet_net_unlock(LNET_LOCK_EX);
1255 LIBCFS_FREE(ne, sizeof(*ne));
1257 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1258 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1263 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1265 struct list_head zombies;
1266 struct lnet_nid_list *ne;
1267 struct lnet_nid_list *tmp;
1269 INIT_LIST_HEAD(&zombies);
1271 lnet_net_lock(LNET_LOCK_EX);
1272 if (lpni->lpni_pref_nnids == 1)
1273 lpni->lpni_pref.nid = LNET_NID_ANY;
1274 else if (lpni->lpni_pref_nnids > 1)
1275 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1276 lpni->lpni_pref_nnids = 0;
1277 lnet_net_unlock(LNET_LOCK_EX);
1279 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1280 list_del_init(&ne->nl_list);
1281 LIBCFS_FREE(ne, sizeof(*ne));
1286 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1288 struct lnet_peer_ni *lpni;
1289 lnet_nid_t primary_nid = nid;
1291 lpni = lnet_find_peer_ni_locked(nid);
1293 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1294 lnet_peer_ni_decref_locked(lpni);
1301 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1302 __must_hold(&lp->lp_lock)
1304 if (lnet_peer_discovery_disabled)
1307 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1308 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1319 lnet_is_discovery_disabled(struct lnet_peer *lp)
1323 spin_lock(&lp->lp_lock);
1324 rc = lnet_is_discovery_disabled_locked(lp);
1325 spin_unlock(&lp->lp_lock);
1331 LNetPrimaryNID(lnet_nid_t nid)
1333 struct lnet_peer *lp;
1334 struct lnet_peer_ni *lpni;
1335 lnet_nid_t primary_nid = nid;
1339 if (nid == LNET_NID_LO_0)
1340 return LNET_NID_LO_0;
1342 cpt = lnet_net_lock_current();
1343 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1348 lp = lpni->lpni_peer_net->lpn_peer;
1350 /* If discovery is disabled locally then we needn't bother running
1351 * discovery here because discovery will not modify whatever
1352 * primary NID is currently set for this peer. If the specified peer is
1353 * down then this discovery can introduce long delays into the mount
1354 * process, so skip it if it isn't necessary.
1356 while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1357 spin_lock(&lp->lp_lock);
1358 /* force a full discovery cycle */
1359 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1360 spin_unlock(&lp->lp_lock);
1362 rc = lnet_discover_peer_locked(lpni, cpt, true);
1365 /* The lpni (or lp) for this NID may have changed and our ref is
1366 * the only thing keeping the old one around. Release the ref
1367 * and lookup the lpni again
1369 lnet_peer_ni_decref_locked(lpni);
1370 lpni = lnet_find_peer_ni_locked(nid);
1375 lp = lpni->lpni_peer_net->lpn_peer;
1377 /* If we find that the peer has discovery disabled then we will
1378 * not modify whatever primary NID is currently set for this
1379 * peer. Thus, we can break out of this loop even if the peer
1380 * is not fully up to date.
1382 if (lnet_is_discovery_disabled(lp))
1385 primary_nid = lp->lp_primary_nid;
1387 lnet_peer_ni_decref_locked(lpni);
1389 lnet_net_unlock(cpt);
1391 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1392 libcfs_nid2str(primary_nid), rc);
1395 EXPORT_SYMBOL(LNetPrimaryNID);
1397 struct lnet_peer_net *
1398 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1400 struct lnet_peer_net *peer_net;
1401 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1402 if (peer_net->lpn_net_id == net_id)
1409 * Attach a peer_ni to a peer_net and peer. This function assumes
1410 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1411 * may be attached to a different peer, in which case it will be
1412 * properly detached first. The whole operation is done atomically.
1414 * This function consumes the reference on lpni and Always returns 0.
1415 * This is the last function called from functions that do return an
1416 * int, so returning 0 here allows the compiler to do a tail call.
1419 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1420 struct lnet_peer_net *lpn,
1421 struct lnet_peer_ni *lpni,
1424 struct lnet_peer_table *ptable;
1425 bool new_lpn = false;
1428 /* Install the new peer_ni */
1429 lnet_net_lock(LNET_LOCK_EX);
1430 /* Add peer_ni to global peer table hash, if necessary. */
1431 if (list_empty(&lpni->lpni_hashlist)) {
1432 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1434 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1435 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1436 ptable->pt_version++;
1437 lnet_peer_ni_addref_locked(lpni);
1440 /* Detach the peer_ni from an existing peer, if necessary. */
1441 if (lpni->lpni_peer_net) {
1442 LASSERT(lpni->lpni_peer_net != lpn);
1443 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1444 lnet_peer_detach_peer_ni_locked(lpni);
1445 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1446 lpni->lpni_peer_net = NULL;
1449 /* Add peer_ni to peer_net */
1450 lpni->lpni_peer_net = lpn;
1451 if (lp->lp_primary_nid == lpni->lpni_nid)
1452 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1454 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1455 lnet_update_peer_net_healthv(lpni);
1456 lnet_peer_net_addref_locked(lpn);
1458 /* Add peer_net to peer */
1459 if (!lpn->lpn_peer) {
1462 if (lp->lp_primary_nid == lpni->lpni_nid)
1463 list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1465 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1466 lnet_peer_addref_locked(lp);
1469 /* Add peer to global peer list, if necessary */
1470 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1471 if (list_empty(&lp->lp_peer_list)) {
1472 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1477 /* Update peer state */
1478 spin_lock(&lp->lp_lock);
1479 if (flags & LNET_PEER_CONFIGURED) {
1480 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1481 lp->lp_state |= LNET_PEER_CONFIGURED;
1483 if (flags & LNET_PEER_MULTI_RAIL) {
1484 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1485 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1486 lnet_peer_clr_non_mr_pref_nids(lp);
1489 spin_unlock(&lp->lp_lock);
1495 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1497 CERROR("Failed to apply UDSPs on lpn %s\n",
1498 libcfs_net2str(lpn->lpn_net_id));
1500 rc = lnet_udsp_apply_policies_on_lpni(lpni);
1502 CERROR("Failed to apply UDSPs on lpni %s\n",
1503 libcfs_nid2str(lpni->lpni_nid));
1505 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1506 libcfs_nid2str(lp->lp_primary_nid),
1507 libcfs_nid2str(lpni->lpni_nid), flags);
1508 lnet_peer_ni_decref_locked(lpni);
1509 lnet_net_unlock(LNET_LOCK_EX);
1515 * Create a new peer, with nid as its primary nid.
1517 * Call with the lnet_api_mutex held.
1520 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1522 struct lnet_peer *lp;
1523 struct lnet_peer_net *lpn;
1524 struct lnet_peer_ni *lpni;
1527 LASSERT(nid != LNET_NID_ANY);
1530 * No need for the lnet_net_lock here, because the
1531 * lnet_api_mutex is held.
1533 lpni = lnet_find_peer_ni_locked(nid);
1535 /* A peer with this NID already exists. */
1536 lp = lpni->lpni_peer_net->lpn_peer;
1537 lnet_peer_ni_decref_locked(lpni);
1539 * This is an error if the peer was configured and the
1540 * primary NID differs or an attempt is made to change
1541 * the Multi-Rail flag. Otherwise the assumption is
1542 * that an existing peer is being modified.
1544 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1545 if (lp->lp_primary_nid != nid)
1547 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1551 /* Delete and recreate as a configured peer. */
1555 /* Create peer, peer_net, and peer_ni. */
1557 lp = lnet_peer_alloc(nid);
1560 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1563 lpni = lnet_peer_ni_alloc(nid);
1567 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1570 LIBCFS_FREE(lpn, sizeof(*lpn));
1572 LIBCFS_FREE(lp, sizeof(*lp));
1574 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1575 libcfs_nid2str(nid), flags, rc);
1580 * Add a NID to a peer. Call with ln_api_mutex held.
1583 * -EPERM: Non-DLC addition to a DLC-configured peer.
1584 * -EEXIST: The NID was configured by DLC for a different peer.
1585 * -ENOMEM: Out of memory.
1586 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1587 * non-multi-rail peer.
1590 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1592 struct lnet_peer_net *lpn;
1593 struct lnet_peer_ni *lpni;
1597 LASSERT(nid != LNET_NID_ANY);
1599 /* A configured peer can only be updated through configuration. */
1600 if (!(flags & LNET_PEER_CONFIGURED)) {
1601 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1608 * The MULTI_RAIL flag can be set but not cleared, because
1609 * that would leave the peer struct in an invalid state.
1611 if (flags & LNET_PEER_MULTI_RAIL) {
1612 spin_lock(&lp->lp_lock);
1613 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1614 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1615 lnet_peer_clr_non_mr_pref_nids(lp);
1617 spin_unlock(&lp->lp_lock);
1618 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1623 lpni = lnet_find_peer_ni_locked(nid);
1626 * A peer_ni already exists. This is only a problem if
1627 * it is not connected to this peer and was configured
1630 if (lpni->lpni_peer_net->lpn_peer == lp)
1632 if (lnet_peer_ni_is_configured(lpni)) {
1636 /* If this is the primary NID, destroy the peer. */
1637 if (lnet_peer_ni_is_primary(lpni)) {
1638 struct lnet_peer *rtr_lp =
1639 lpni->lpni_peer_net->lpn_peer;
1640 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1642 * if we're trying to delete a router it means
1643 * we're moving this peer NI to a new peer so must
1644 * transfer router properties to the new peer
1646 if (rtr_refcount > 0) {
1647 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1648 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1650 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1651 lnet_peer_ni_decref_locked(lpni);
1652 lpni = lnet_peer_ni_alloc(nid);
1659 lpni = lnet_peer_ni_alloc(nid);
1667 * Get the peer_net. Check that we're not adding a second
1668 * peer_ni on a peer_net of a non-multi-rail peer.
1670 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1672 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1677 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1682 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1685 lnet_peer_ni_decref_locked(lpni);
1687 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1688 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1694 * Update the primary NID of a peer, if possible.
1696 * Call with the lnet_api_mutex held.
1699 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1701 lnet_nid_t old = lp->lp_primary_nid;
1704 if (lp->lp_primary_nid == nid)
1707 lp->lp_primary_nid = nid;
1709 rc = lnet_peer_add_nid(lp, nid, flags);
1711 lp->lp_primary_nid = old;
1715 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1716 libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1721 * lpni creation initiated due to traffic either sending or receiving.
1724 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1726 struct lnet_peer *lp;
1727 struct lnet_peer_net *lpn;
1728 struct lnet_peer_ni *lpni;
1732 if (nid == LNET_NID_ANY) {
1737 /* lnet_net_lock is not needed here because ln_api_lock is held */
1738 lpni = lnet_find_peer_ni_locked(nid);
1741 * We must have raced with another thread. Since we
1742 * know next to nothing about a peer_ni created by
1743 * traffic, we just assume everything is ok and
1746 lnet_peer_ni_decref_locked(lpni);
1750 /* Create peer, peer_net, and peer_ni. */
1752 lp = lnet_peer_alloc(nid);
1755 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1758 lpni = lnet_peer_ni_alloc(nid);
1761 if (pref != LNET_NID_ANY)
1762 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1764 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1767 LIBCFS_FREE(lpn, sizeof(*lpn));
1769 LIBCFS_FREE(lp, sizeof(*lp));
1771 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1776 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1778 * This API handles the following combinations:
1779 * Create a peer with its primary NI if only the prim_nid is provided
1780 * Add a NID to a peer identified by the prim_nid. The peer identified
1781 * by the prim_nid must already exist.
1782 * The peer being created may be non-MR.
1784 * The caller must hold ln_api_mutex. This prevents the peer from
1785 * being created/modified/deleted by a different thread.
1788 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1790 struct lnet_peer *lp = NULL;
1791 struct lnet_peer_ni *lpni;
1794 /* The prim_nid must always be specified */
1795 if (prim_nid == LNET_NID_ANY)
1798 flags = LNET_PEER_CONFIGURED;
1800 flags |= LNET_PEER_MULTI_RAIL;
1803 * If nid isn't specified, we must create a new peer with
1804 * prim_nid as its primary nid.
1806 if (nid == LNET_NID_ANY)
1807 return lnet_peer_add(prim_nid, flags);
1809 /* Look up the prim_nid, which must exist. */
1810 lpni = lnet_find_peer_ni_locked(prim_nid);
1813 lnet_peer_ni_decref_locked(lpni);
1814 lp = lpni->lpni_peer_net->lpn_peer;
1816 /* Peer must have been configured. */
1817 if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1818 CDEBUG(D_NET, "peer %s was not configured\n",
1819 libcfs_nid2str(prim_nid));
1823 /* Primary NID must match */
1824 if (lp->lp_primary_nid != prim_nid) {
1825 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1826 libcfs_nid2str(prim_nid),
1827 libcfs_nid2str(lp->lp_primary_nid));
1831 /* Multi-Rail flag must match. */
1832 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1833 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1834 libcfs_nid2str(prim_nid));
1838 return lnet_peer_add_nid(lp, nid, flags);
1842 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1844 * This API handles the following combinations:
1845 * Delete a NI from a peer if both prim_nid and nid are provided.
1846 * Delete a peer if only prim_nid is provided.
1847 * Delete a peer if its primary nid is provided.
1849 * The caller must hold ln_api_mutex. This prevents the peer from
1850 * being modified/deleted by a different thread.
1853 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1855 struct lnet_peer *lp;
1856 struct lnet_peer_ni *lpni;
1859 if (prim_nid == LNET_NID_ANY)
1862 lpni = lnet_find_peer_ni_locked(prim_nid);
1865 lnet_peer_ni_decref_locked(lpni);
1866 lp = lpni->lpni_peer_net->lpn_peer;
1868 if (prim_nid != lp->lp_primary_nid) {
1869 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1870 libcfs_nid2str(prim_nid),
1871 libcfs_nid2str(lp->lp_primary_nid));
1875 lnet_net_lock(LNET_LOCK_EX);
1876 if (lp->lp_rtr_refcount > 0) {
1877 lnet_net_unlock(LNET_LOCK_EX);
1878 CERROR("%s is a router. Can not be deleted\n",
1879 libcfs_nid2str(prim_nid));
1882 lnet_net_unlock(LNET_LOCK_EX);
1884 if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1885 return lnet_peer_del(lp);
1887 flags = LNET_PEER_CONFIGURED;
1888 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1889 flags |= LNET_PEER_MULTI_RAIL;
1891 return lnet_peer_del_nid(lp, nid, flags);
1895 lnet_destroy_peer_ni_locked(struct kref *ref)
1897 struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
1899 struct lnet_peer_table *ptable;
1900 struct lnet_peer_net *lpn;
1902 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1904 LASSERT(kref_read(&lpni->lpni_kref) == 0);
1905 LASSERT(list_empty(&lpni->lpni_txq));
1906 LASSERT(lpni->lpni_txqnob == 0);
1907 LASSERT(list_empty(&lpni->lpni_peer_nis));
1908 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1910 lpn = lpni->lpni_peer_net;
1911 lpni->lpni_peer_net = NULL;
1912 lpni->lpni_net = NULL;
1914 if (!list_empty(&lpni->lpni_hashlist)) {
1915 /* remove the peer ni from the zombie list */
1916 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1917 spin_lock(&ptable->pt_zombie_lock);
1918 list_del_init(&lpni->lpni_hashlist);
1919 ptable->pt_zombies--;
1920 spin_unlock(&ptable->pt_zombie_lock);
1923 if (lpni->lpni_pref_nnids > 1) {
1924 struct lnet_nid_list *ne, *tmp;
1926 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
1928 list_del_init(&ne->nl_list);
1929 LIBCFS_FREE(ne, sizeof(*ne));
1932 LIBCFS_FREE(lpni, sizeof(*lpni));
1935 lnet_peer_net_decref_locked(lpn);
1938 struct lnet_peer_ni *
1939 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1941 struct lnet_peer_ni *lpni = NULL;
1944 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1945 return ERR_PTR(-ESHUTDOWN);
1948 * find if a peer_ni already exists.
1949 * If so then just return that.
1951 lpni = lnet_find_peer_ni_locked(nid);
1955 lnet_net_unlock(cpt);
1957 rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1960 goto out_net_relock;
1963 lpni = lnet_find_peer_ni_locked(nid);
1973 * Get a peer_ni for the given nid, create it if necessary. Takes a
1974 * hold on the peer_ni.
1976 struct lnet_peer_ni *
1977 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1979 struct lnet_peer_ni *lpni = NULL;
1982 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1983 return ERR_PTR(-ESHUTDOWN);
1986 * find if a peer_ni already exists.
1987 * If so then just return that.
1989 lpni = lnet_find_peer_ni_locked(nid);
1995 * use the lnet_api_mutex to serialize the creation of the peer_ni
1996 * and the creation/deletion of the local ni/net. When a local ni is
1997 * created, if there exists a set of peer_nis on that network,
1998 * they need to be traversed and updated. When a local NI is
1999 * deleted, which could result in a network being deleted, then
2000 * all peer nis on that network need to be removed as well.
2002 * Creation through traffic should also be serialized with
2003 * creation through DLC.
2005 lnet_net_unlock(cpt);
2006 mutex_lock(&the_lnet.ln_api_mutex);
2008 * Shutdown is only set under the ln_api_lock, so a single
2009 * check here is sufficent.
2011 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2012 lpni = ERR_PTR(-ESHUTDOWN);
2013 goto out_mutex_unlock;
2016 rc = lnet_peer_ni_traffic_add(nid, pref);
2019 goto out_mutex_unlock;
2022 lpni = lnet_find_peer_ni_locked(nid);
2026 mutex_unlock(&the_lnet.ln_api_mutex);
2029 /* Lock has been dropped, check again for shutdown. */
2030 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2032 lnet_peer_ni_decref_locked(lpni);
2033 lpni = ERR_PTR(-ESHUTDOWN);
2040 lnet_peer_gw_discovery(struct lnet_peer *lp)
2044 spin_lock(&lp->lp_lock);
2045 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2047 spin_unlock(&lp->lp_lock);
2053 lnet_peer_is_uptodate(struct lnet_peer *lp)
2057 spin_lock(&lp->lp_lock);
2058 rc = lnet_peer_is_uptodate_locked(lp);
2059 spin_unlock(&lp->lp_lock);
2064 * Is a peer uptodate from the point of view of discovery?
2066 * If it is currently being processed, obviously not.
2067 * A forced Ping or Push is also handled by the discovery thread.
2069 * Otherwise look at whether the peer needs rediscovering.
2072 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2073 __must_hold(&lp->lp_lock)
2077 if (lp->lp_state & (LNET_PEER_DISCOVERING |
2078 LNET_PEER_FORCE_PING |
2079 LNET_PEER_FORCE_PUSH)) {
2081 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2083 } else if (lnet_peer_needs_push(lp)) {
2085 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2086 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2097 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2099 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2101 /* The discovery thread holds net_lock/EX and lp_lock when it splices
2102 * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2103 * when adding to the list and queuing the peer to ensure that we do not
2104 * strand any messages on the lp_dc_pendq. This scheme ensures the
2105 * message will be resent even if the peer is already being discovered.
2106 * Therefore we needn't check the return value of
2107 * lnet_peer_queue_for_discovery(lp).
2109 lnet_net_lock(LNET_LOCK_EX);
2110 spin_lock(&lp->lp_lock);
2111 list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2112 spin_unlock(&lp->lp_lock);
2113 lnet_peer_queue_for_discovery(lp);
2114 lnet_net_unlock(LNET_LOCK_EX);
2118 * Queue a peer for the attention of the discovery thread. Call with
2119 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2120 * -EALREADY if the peer was already queued.
2122 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2126 spin_lock(&lp->lp_lock);
2127 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2128 lp->lp_state |= LNET_PEER_DISCOVERING;
2129 spin_unlock(&lp->lp_lock);
2130 if (list_empty(&lp->lp_dc_list)) {
2131 lnet_peer_addref_locked(lp);
2132 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2133 wake_up(&the_lnet.ln_dc_waitq);
2139 CDEBUG(D_NET, "Queue peer %s: %d\n",
2140 libcfs_nid2str(lp->lp_primary_nid), rc);
2146 * Discovery of a peer is complete. Wake all waiters on the peer.
2147 * Call with lnet_net_lock/EX held.
2149 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
2151 struct lnet_msg *msg, *tmp;
2153 LIST_HEAD(pending_msgs);
2155 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2156 libcfs_nid2str(lp->lp_primary_nid));
2158 list_del_init(&lp->lp_dc_list);
2159 spin_lock(&lp->lp_lock);
2160 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2161 spin_unlock(&lp->lp_lock);
2162 wake_up(&lp->lp_dc_waitq);
2164 if (lp->lp_rtr_refcount > 0)
2165 lnet_router_discovery_complete(lp);
2167 lnet_net_unlock(LNET_LOCK_EX);
2169 /* iterate through all pending messages and send them again */
2170 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2171 list_del_init(&msg->msg_list);
2172 if (lp->lp_dc_error) {
2173 lnet_finalize(msg, lp->lp_dc_error);
2177 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2178 lnet_msgtyp2str(msg->msg_type),
2179 libcfs_id2str(msg->msg_target));
2180 rc = lnet_send(msg->msg_src_nid_param, msg,
2181 msg->msg_rtr_nid_param);
2183 CNETERR("Error sending %s to %s: %d\n",
2184 lnet_msgtyp2str(msg->msg_type),
2185 libcfs_id2str(msg->msg_target), rc);
2186 lnet_finalize(msg, rc);
2189 lnet_net_lock(LNET_LOCK_EX);
2190 lnet_peer_decref_locked(lp);
2194 * Handle inbound push.
2195 * Like any event handler, called with lnet_res_lock/CPT held.
2197 void lnet_peer_push_event(struct lnet_event *ev)
2199 struct lnet_ping_buffer *pbuf;
2200 struct lnet_peer *lp;
2202 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2204 /* lnet_find_peer() adds a refcount */
2205 lp = lnet_find_peer(ev->source.nid);
2207 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2208 libcfs_nid2str(ev->initiator.nid),
2209 libcfs_nid2str(ev->source.nid));
2210 pbuf->pb_needs_post = true;
2214 /* Ensure peer state remains consistent while we modify it. */
2215 spin_lock(&lp->lp_lock);
2218 * If some kind of error happened the contents of the message
2219 * cannot be used. Clear the NIDS_UPTODATE and set the
2220 * FORCE_PING flag to trigger a ping.
2223 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2224 lp->lp_state |= LNET_PEER_FORCE_PING;
2225 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2227 libcfs_nid2str(lp->lp_primary_nid),
2228 libcfs_nid2str(ev->source.nid));
2233 * A push with invalid or corrupted info. Clear the UPTODATE
2234 * flag to trigger a ping.
2236 if (lnet_ping_info_validate(&pbuf->pb_info)) {
2237 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2238 lp->lp_state |= LNET_PEER_FORCE_PING;
2239 CDEBUG(D_NET, "Corrupted Push from %s\n",
2240 libcfs_nid2str(lp->lp_primary_nid));
2245 * Make sure we'll allocate the correct size ping buffer when
2248 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2249 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2252 * A non-Multi-Rail peer is not supposed to be capable of
2255 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2256 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2257 libcfs_nid2str(lp->lp_primary_nid));
2262 * The peer may have discovery disabled at its end. Set
2263 * NO_DISCOVERY as appropriate.
2265 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2266 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2267 libcfs_nid2str(lp->lp_primary_nid));
2269 * Mark the peer for deletion if we already know about it
2270 * and it's going from discovery set to no discovery set
2272 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2273 LNET_PEER_DISCOVERING)) &&
2274 lp->lp_state & LNET_PEER_DISCOVERED) {
2275 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2276 libcfs_nid2str(lp->lp_primary_nid),
2278 lp->lp_state |= LNET_PEER_MARK_DELETION;
2280 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2281 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2282 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2283 libcfs_nid2str(lp->lp_primary_nid));
2284 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2288 * Update the MULTI_RAIL flag based on the push. If the peer
2289 * was configured with DLC then the setting should match what
2291 * NB: We verified above that the MR feature bit is set in pi_features
2293 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2294 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2295 libcfs_nid2str(lp->lp_primary_nid), lp);
2296 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2297 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2298 libcfs_nid2str(lp->lp_primary_nid));
2299 } else if (lnet_peer_discovery_disabled) {
2300 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2301 libcfs_nid2str(lp->lp_primary_nid), lp);
2302 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2303 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2304 libcfs_nid2str(lp->lp_primary_nid), lp);
2306 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2307 libcfs_nid2str(lp->lp_primary_nid), lp);
2308 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2309 lnet_peer_clr_non_mr_pref_nids(lp);
2313 * Check for truncation of the Put message. Clear the
2314 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2315 * and tell discovery to allocate a bigger buffer.
2317 if (ev->mlength < ev->rlength) {
2318 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2319 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2320 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2321 lp->lp_state |= LNET_PEER_FORCE_PING;
2322 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2323 libcfs_nid2str(lp->lp_primary_nid),
2324 pbuf->pb_info.pi_nnis);
2328 /* always assume new data */
2329 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2330 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2333 * If there is data present that hasn't been processed yet,
2334 * we'll replace it if the Put contained newer data and it
2335 * fits. We're racing with a Ping or earlier Push in this
2338 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2339 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2340 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2341 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2342 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2343 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2344 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2345 libcfs_nid2str(lp->lp_primary_nid),
2346 LNET_PING_BUFFER_SEQNO(pbuf),
2347 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2353 * Allocate a buffer to copy the data. On a failure we drop
2354 * the Push and set FORCE_PING to force the discovery
2355 * thread to fix the problem by pinging the peer.
2357 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2359 lp->lp_state |= LNET_PEER_FORCE_PING;
2360 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2361 libcfs_nid2str(lp->lp_primary_nid),
2362 LNET_PING_BUFFER_SEQNO(pbuf));
2367 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2368 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2369 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2370 CDEBUG(D_NET, "Received Push %s %u\n",
2371 libcfs_nid2str(lp->lp_primary_nid),
2372 LNET_PING_BUFFER_SEQNO(pbuf));
2375 /* We've processed this buffer. It can be reposted */
2376 pbuf->pb_needs_post = true;
2379 * Queue the peer for discovery if not done, force it on the request
2380 * queue and wake the discovery thread if the peer was already queued,
2381 * because its status changed.
2383 spin_unlock(&lp->lp_lock);
2384 lnet_net_lock(LNET_LOCK_EX);
2385 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2386 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2387 wake_up(&the_lnet.ln_dc_waitq);
2389 /* Drop refcount from lookup */
2390 lnet_peer_decref_locked(lp);
2391 lnet_net_unlock(LNET_LOCK_EX);
2395 * Clear the discovery error state, unless we're already discovering
2396 * this peer, in which case the error is current.
2398 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2400 spin_lock(&lp->lp_lock);
2401 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2402 lp->lp_dc_error = 0;
2403 spin_unlock(&lp->lp_lock);
2407 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2408 * dropped/retaken within this function. An lnet_peer_ni is passed in
2409 * because discovery could tear down an lnet_peer.
2412 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2415 struct lnet_peer *lp;
2420 lnet_net_unlock(cpt);
2421 lnet_net_lock(LNET_LOCK_EX);
2422 lp = lpni->lpni_peer_net->lpn_peer;
2423 lnet_peer_clear_discovery_error(lp);
2426 * We're willing to be interrupted. The lpni can become a
2427 * zombie if we race with DLC, so we must check for that.
2430 /* Keep lp alive when the lnet_net_lock is unlocked */
2431 lnet_peer_addref_locked(lp);
2432 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2433 if (signal_pending(current))
2435 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2438 * Don't repeat discovery if discovery is disabled. This is
2439 * done to ensure we can use discovery as a standard ping as
2440 * well for backwards compatibility with routers which do not
2441 * have discovery or have discovery disabled
2443 if (lnet_is_discovery_disabled(lp) && count > 0)
2445 if (lp->lp_dc_error)
2447 if (lnet_peer_is_uptodate(lp))
2449 lnet_peer_queue_for_discovery(lp);
2451 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2454 * If caller requested a non-blocking operation then
2455 * return immediately. Once discovery is complete any
2456 * pending messages that were stopped due to discovery
2457 * will be transmitted.
2462 lnet_net_unlock(LNET_LOCK_EX);
2464 finish_wait(&lp->lp_dc_waitq, &wait);
2465 lnet_net_lock(LNET_LOCK_EX);
2466 lnet_peer_decref_locked(lp);
2467 /* Peer may have changed */
2468 lp = lpni->lpni_peer_net->lpn_peer;
2470 finish_wait(&lp->lp_dc_waitq, &wait);
2472 lnet_net_unlock(LNET_LOCK_EX);
2474 lnet_peer_decref_locked(lp);
2476 * The peer may have changed, so re-check and rediscover if that turns
2477 * out to have been the case. The reference count on lp ensured that
2478 * even if it was unlinked from lpni the memory could not be recycled.
2479 * Thus the check below is sufficient to determine whether the peer
2480 * changed. If the peer changed, then lp must not be dereferenced.
2482 if (lp != lpni->lpni_peer_net->lpn_peer)
2485 if (signal_pending(current))
2487 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2489 else if (lp->lp_dc_error)
2490 rc = lp->lp_dc_error;
2492 CDEBUG(D_NET, "non-blocking discovery\n");
2493 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2496 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2497 (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2498 libcfs_nid2str(lpni->lpni_nid), rc,
2499 (!block) ? "pending discovery" : "discovery complete");
2504 /* Handle an incoming ack for a push. */
2506 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2508 struct lnet_ping_buffer *pbuf;
2510 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2511 spin_lock(&lp->lp_lock);
2512 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2513 lp->lp_push_error = ev->status;
2515 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2517 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2518 spin_unlock(&lp->lp_lock);
2520 CDEBUG(D_NET, "peer %s ev->status %d\n",
2521 libcfs_nid2str(lp->lp_primary_nid), ev->status);
2524 /* Handle a Reply message. This is the reply to a Ping message. */
2526 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2528 struct lnet_ping_buffer *pbuf;
2531 spin_lock(&lp->lp_lock);
2533 lp->lp_disc_src_nid = ev->target.nid;
2536 * If some kind of error happened the contents of message
2537 * cannot be used. Set PING_FAILED to trigger a retry.
2540 lp->lp_state |= LNET_PEER_PING_FAILED;
2541 lp->lp_ping_error = ev->status;
2542 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2544 libcfs_nid2str(lp->lp_primary_nid),
2545 libcfs_nid2str(ev->source.nid));
2549 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2550 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2551 lnet_swap_pinginfo(pbuf);
2554 * A reply with invalid or corrupted info. Set PING_FAILED to
2557 rc = lnet_ping_info_validate(&pbuf->pb_info);
2559 lp->lp_state |= LNET_PEER_PING_FAILED;
2560 lp->lp_ping_error = 0;
2561 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2562 libcfs_nid2str(lp->lp_primary_nid), rc);
2568 * The peer may have discovery disabled at its end. Set
2569 * NO_DISCOVERY as appropriate.
2571 if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2572 !lnet_peer_discovery_disabled) {
2573 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2574 libcfs_nid2str(lp->lp_primary_nid));
2575 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2577 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2578 libcfs_nid2str(lp->lp_primary_nid));
2579 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2583 * Update the MULTI_RAIL flag based on the reply. If the peer
2584 * was configured with DLC then the setting should match what
2587 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2588 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2589 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2590 libcfs_nid2str(lp->lp_primary_nid), lp);
2591 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2592 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2593 libcfs_nid2str(lp->lp_primary_nid));
2594 } else if (lnet_peer_discovery_disabled) {
2596 "peer %s(%p) not MR: DD disabled locally\n",
2597 libcfs_nid2str(lp->lp_primary_nid), lp);
2598 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2600 "peer %s(%p) not MR: DD disabled remotely\n",
2601 libcfs_nid2str(lp->lp_primary_nid), lp);
2603 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2604 libcfs_nid2str(lp->lp_primary_nid), lp);
2605 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2606 lnet_peer_clr_non_mr_pref_nids(lp);
2608 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2609 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2610 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2611 libcfs_nid2str(lp->lp_primary_nid));
2613 CERROR("Multi-Rail state vanished from %s\n",
2614 libcfs_nid2str(lp->lp_primary_nid));
2615 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2620 * Make sure we'll allocate the correct size ping buffer when
2623 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2624 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2627 * Check for truncation of the Reply. Clear PING_SENT and set
2628 * PING_FAILED to trigger a retry.
2630 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2631 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2632 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2633 lp->lp_state |= LNET_PEER_PING_FAILED;
2634 lp->lp_ping_error = 0;
2635 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2636 libcfs_nid2str(lp->lp_primary_nid),
2637 pbuf->pb_info.pi_nnis);
2642 * Check the sequence numbers in the reply. These are only
2643 * available if the reply came from a Multi-Rail peer.
2645 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2646 pbuf->pb_info.pi_nnis > 1 &&
2647 lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2648 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2649 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2650 libcfs_nid2str(lp->lp_primary_nid),
2651 LNET_PING_BUFFER_SEQNO(pbuf),
2654 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2657 /* We're happy with the state of the data in the buffer. */
2658 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2659 libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2660 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2661 lnet_ping_buffer_decref(lp->lp_data);
2663 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2664 lnet_ping_buffer_addref(pbuf);
2667 lp->lp_state &= ~LNET_PEER_PING_SENT;
2668 spin_unlock(&lp->lp_lock);
2670 lnet_net_lock(LNET_LOCK_EX);
2672 * If this peer is a gateway, call the routing callback to
2673 * handle the ping reply
2675 if (lp->lp_rtr_refcount > 0)
2676 lnet_router_discovery_ping_reply(lp);
2677 lnet_net_unlock(LNET_LOCK_EX);
2681 * Send event handling. Only matters for error cases, where we clean
2682 * up state on the peer and peer_ni that would otherwise be updated in
2683 * the REPLY event handler for a successful Ping, and the ACK event
2684 * handler for a successful Push.
2687 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2694 spin_lock(&lp->lp_lock);
2695 if (ev->msg_type == LNET_MSG_GET) {
2696 lp->lp_state &= ~LNET_PEER_PING_SENT;
2697 lp->lp_state |= LNET_PEER_PING_FAILED;
2698 lp->lp_ping_error = ev->status;
2699 } else { /* ev->msg_type == LNET_MSG_PUT */
2700 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2701 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2702 lp->lp_push_error = ev->status;
2704 spin_unlock(&lp->lp_lock);
2705 rc = LNET_REDISCOVER_PEER;
2707 CDEBUG(D_NET, "%s Send to %s: %d\n",
2708 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2709 libcfs_nid2str(ev->target.nid), rc);
2714 * Unlink event handling. This event is only seen if a call to
2715 * LNetMDUnlink() caused the event to be unlinked. If this call was
2716 * made after the event was set up in LNetGet() or LNetPut() then we
2717 * assume the Ping or Push timed out.
2720 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2722 spin_lock(&lp->lp_lock);
2723 /* We've passed through LNetGet() */
2724 if (lp->lp_state & LNET_PEER_PING_SENT) {
2725 lp->lp_state &= ~LNET_PEER_PING_SENT;
2726 lp->lp_state |= LNET_PEER_PING_FAILED;
2727 lp->lp_ping_error = -ETIMEDOUT;
2728 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2729 libcfs_nid2str(lp->lp_primary_nid));
2731 /* We've passed through LNetPut() */
2732 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2733 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2734 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2735 lp->lp_push_error = -ETIMEDOUT;
2736 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2737 libcfs_nid2str(lp->lp_primary_nid));
2739 spin_unlock(&lp->lp_lock);
2743 * Event handler for the discovery EQ.
2745 * Called with lnet_res_lock(cpt) held. The cpt is the
2746 * lnet_cpt_of_cookie() of the md handle cookie.
2748 static void lnet_discovery_event_handler(struct lnet_event *event)
2750 struct lnet_peer *lp = event->md_user_ptr;
2751 struct lnet_ping_buffer *pbuf;
2754 /* discovery needs to take another look */
2755 rc = LNET_REDISCOVER_PEER;
2757 CDEBUG(D_NET, "Received event: %d\n", event->type);
2759 switch (event->type) {
2760 case LNET_EVENT_ACK:
2761 lnet_discovery_event_ack(lp, event);
2763 case LNET_EVENT_REPLY:
2764 lnet_discovery_event_reply(lp, event);
2766 case LNET_EVENT_SEND:
2767 /* Only send failure triggers a retry. */
2768 rc = lnet_discovery_event_send(lp, event);
2770 case LNET_EVENT_UNLINK:
2771 /* LNetMDUnlink() was called */
2772 lnet_discovery_event_unlink(lp, event);
2775 /* Invalid events. */
2778 lnet_net_lock(LNET_LOCK_EX);
2779 if (event->unlinked) {
2780 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2781 lnet_ping_buffer_decref(pbuf);
2782 lnet_peer_decref_locked(lp);
2785 /* put peer back at end of request queue, if discovery not already
2787 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2788 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2789 wake_up(&the_lnet.ln_dc_waitq);
2791 lnet_net_unlock(LNET_LOCK_EX);
2795 * Build a peer from incoming data.
2797 * The NIDs in the incoming data are supposed to be structured as follows:
2800 * - other NIDs in same net
2801 * - NIDs in second net
2802 * - NIDs in third net
2804 * This due to the way the list of NIDs in the data is created.
2806 * Note that this function will mark the peer uptodate unless an
2807 * ENOMEM is encontered. All other errors are due to a conflict
2808 * between the DLC configuration and what discovery sees. We treat DLC
2809 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2810 * peer from becoming stuck in discovery.
2812 static int lnet_peer_merge_data(struct lnet_peer *lp,
2813 struct lnet_ping_buffer *pbuf)
2815 struct lnet_peer_net *lpn;
2816 struct lnet_peer_ni *lpni;
2817 lnet_nid_t *curnis = NULL;
2818 struct lnet_ni_status *addnis = NULL;
2819 lnet_nid_t *delnis = NULL;
2829 flags = LNET_PEER_DISCOVERED;
2830 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2831 flags |= LNET_PEER_MULTI_RAIL;
2834 * Cache the routing feature for the peer; whether it is enabled
2835 * for disabled as reported by the remote peer.
2837 spin_lock(&lp->lp_lock);
2838 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2839 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2841 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2842 spin_unlock(&lp->lp_lock);
2844 nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2845 CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2846 CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2847 CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2848 if (!curnis || !addnis || !delnis) {
2856 /* Construct the list of NIDs present in peer. */
2858 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2859 curnis[ncurnis++] = lpni->lpni_nid;
2862 * Check for NIDs in pbuf not present in curnis[].
2863 * The loop starts at 1 to skip the loopback NID.
2865 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2866 for (j = 0; j < ncurnis; j++)
2867 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2870 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2873 * Check for NIDs in curnis[] not present in pbuf.
2874 * The nested loop starts at 1 to skip the loopback NID.
2876 * But never add the loopback NID to delnis[]: if it is
2877 * present in curnis[] then this peer is for this node.
2879 for (i = 0; i < ncurnis; i++) {
2880 if (curnis[i] == LNET_NID_LO_0)
2882 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2883 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2885 * update the information we cache for the
2886 * peer with the latest information we
2889 lpni = lnet_find_peer_ni_locked(curnis[i]);
2891 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2892 lnet_peer_ni_decref_locked(lpni);
2897 if (j == pbuf->pb_info.pi_nnis)
2898 delnis[ndelnis++] = curnis[i];
2902 * If we get here and the discovery is disabled then we don't want
2903 * to add or delete any NIs. We just updated the ones we have some
2904 * information on, and call it a day
2907 if (lnet_is_discovery_disabled(lp))
2910 for (i = 0; i < naddnis; i++) {
2911 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2913 CERROR("Error adding NID %s to peer %s: %d\n",
2914 libcfs_nid2str(addnis[i].ns_nid),
2915 libcfs_nid2str(lp->lp_primary_nid), rc);
2919 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2921 lpni->lpni_ns_status = addnis[i].ns_status;
2922 lnet_peer_ni_decref_locked(lpni);
2926 for (i = 0; i < ndelnis; i++) {
2928 * for routers it's okay to delete the primary_nid because
2929 * the upper layers don't really rely on it. So if we're
2930 * being told that the router changed its primary_nid
2931 * then it's okay to delete it.
2933 if (lp->lp_rtr_refcount > 0)
2934 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2935 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2937 CERROR("Error deleting NID %s from peer %s: %d\n",
2938 libcfs_nid2str(delnis[i]),
2939 libcfs_nid2str(lp->lp_primary_nid), rc);
2945 /* The peer net for the primary NID should be the first entry in the
2946 * peer's lp_peer_nets list, and the peer NI for the primary NID should
2947 * be the first entry in its peer net's lpn_peer_nis list.
2949 lpni = lnet_find_peer_ni_locked(pbuf->pb_info.pi_ni[1].ns_nid);
2951 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
2952 libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
2956 lnet_peer_ni_decref_locked(lpni);
2958 lpn = lpni->lpni_peer_net;
2959 if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
2960 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
2962 if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
2963 list_move(&lpni->lpni_peer_nis,
2964 &lpni->lpni_peer_net->lpn_peer_nis);
2967 * Errors other than -ENOMEM are due to peers having been
2968 * configured with DLC. Ignore these because DLC overrides
2973 CFS_FREE_PTR_ARRAY(curnis, nnis);
2974 CFS_FREE_PTR_ARRAY(addnis, nnis);
2975 CFS_FREE_PTR_ARRAY(delnis, nnis);
2976 lnet_ping_buffer_decref(pbuf);
2977 CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2980 spin_lock(&lp->lp_lock);
2981 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2982 lp->lp_state |= LNET_PEER_FORCE_PING;
2983 spin_unlock(&lp->lp_lock);
2989 * The data in pbuf says lp is its primary peer, but the data was
2990 * received by a different peer. Try to update lp with the data.
2993 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2995 struct lnet_handle_md mdh;
2997 /* Queue lp for discovery, and force it on the request queue. */
2998 lnet_net_lock(LNET_LOCK_EX);
2999 if (lnet_peer_queue_for_discovery(lp))
3000 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3001 lnet_net_unlock(LNET_LOCK_EX);
3003 LNetInvalidateMDHandle(&mdh);
3006 * Decide whether we can move the peer to the DATA_PRESENT state.
3008 * We replace stale data for a multi-rail peer, repair PING_FAILED
3009 * status, and preempt FORCE_PING.
3011 * If after that we have DATA_PRESENT, we merge it into this peer.
3013 spin_lock(&lp->lp_lock);
3014 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3015 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3016 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3017 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3018 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3019 lnet_ping_buffer_decref(pbuf);
3024 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3025 lnet_ping_buffer_decref(lp->lp_data);
3027 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3029 if (lp->lp_state & LNET_PEER_PING_FAILED) {
3030 mdh = lp->lp_ping_mdh;
3031 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3032 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3033 lp->lp_ping_error = 0;
3035 if (lp->lp_state & LNET_PEER_FORCE_PING)
3036 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3037 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3038 spin_unlock(&lp->lp_lock);
3040 if (!LNetMDHandleIsInvalid(mdh))
3044 return lnet_peer_merge_data(lp, pbuf);
3046 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3050 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
3054 for (i = 0; i < pinfo->pi_nnis; i++) {
3055 if (pinfo->pi_ni[i].ns_nid == nid)
3062 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3063 * to the discovery queue a reference was taken that will prevent the peer from
3064 * actually being freed by this function. After this function exits the
3065 * discovery thread should call lnet_peer_discovery_complete() which will
3066 * drop that reference as well as wake any waiters that may also be holding a
3069 static int lnet_peer_deletion(struct lnet_peer *lp)
3070 __must_hold(&lp->lp_lock)
3072 struct list_head rlist;
3073 struct lnet_route *route, *tmp;
3074 int sensitivity = lp->lp_health_sensitivity;
3076 INIT_LIST_HEAD(&rlist);
3078 lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3079 LNET_PEER_FORCE_PUSH);
3080 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3081 libcfs_nid2str(lp->lp_primary_nid), lp, lp->lp_state);
3083 /* no-op if lnet_peer_del() has already been called on this peer */
3084 if (lp->lp_state & LNET_PEER_MARK_DELETED)
3087 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
3090 spin_unlock(&lp->lp_lock);
3092 mutex_lock(&the_lnet.ln_api_mutex);
3094 lnet_net_lock(LNET_LOCK_EX);
3095 /* remove the peer from the discovery work
3096 * queue if it's on there in preparation
3099 if (!list_empty(&lp->lp_dc_list))
3100 list_del(&lp->lp_dc_list);
3101 list_for_each_entry_safe(route, tmp,
3104 lnet_move_route(route, NULL, &rlist);
3105 lnet_net_unlock(LNET_LOCK_EX);
3107 /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3110 list_for_each_entry_safe(route, tmp,
3112 /* re-add these routes */
3113 lnet_add_route(route->lr_net,
3118 LIBCFS_FREE(route, sizeof(*route));
3121 mutex_unlock(&the_lnet.ln_api_mutex);
3123 spin_lock(&lp->lp_lock);
3129 * Update a peer using the data received.
3131 static int lnet_peer_data_present(struct lnet_peer *lp)
3132 __must_hold(&lp->lp_lock)
3134 struct lnet_ping_buffer *pbuf;
3135 struct lnet_peer_ni *lpni;
3136 lnet_nid_t nid = LNET_NID_ANY;
3142 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3143 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3144 spin_unlock(&lp->lp_lock);
3147 * Modifications of peer structures are done while holding the
3148 * ln_api_mutex. A global lock is required because we may be
3149 * modifying multiple peer structures, and a mutex greatly
3150 * simplifies memory management.
3152 * The actual changes to the data structures must also protect
3153 * against concurrent lookups, for which the lnet_net_lock in
3154 * LNET_LOCK_EX mode is used.
3156 mutex_lock(&the_lnet.ln_api_mutex);
3157 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3163 * If this peer is not on the peer list then it is being torn
3164 * down, and our reference count may be all that is keeping it
3165 * alive. Don't do any work on it.
3167 if (list_empty(&lp->lp_peer_list))
3170 flags = LNET_PEER_DISCOVERED;
3171 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3172 flags |= LNET_PEER_MULTI_RAIL;
3175 * Check whether the primary NID in the message matches the
3176 * primary NID of the peer. If it does, update the peer, if
3177 * it it does not, check whether there is already a peer with
3178 * that primary NID. If no such peer exists, try to update
3179 * the primary NID of the current peer (allowed if it was
3180 * created due to message traffic) and complete the update.
3181 * If the peer did exist, hand off the data to it.
3183 * The peer for the loopback interface is a special case: this
3184 * is the peer for the local node, and we want to set its
3185 * primary NID to the correct value here. Moreover, this peer
3186 * can show up with only the loopback NID in the ping buffer.
3188 if (pbuf->pb_info.pi_nnis <= 1)
3190 nid = pbuf->pb_info.pi_ni[1].ns_nid;
3191 if (lp->lp_primary_nid == LNET_NID_LO_0) {
3192 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3194 rc = lnet_peer_merge_data(lp, pbuf);
3196 * if the primary nid of the peer is present in the ping info returned
3197 * from the peer, but it's not the local primary peer we have
3198 * cached and discovery is disabled, then we don't want to update
3199 * our local peer info, by adding or removing NIDs, we just want
3200 * to update the status of the nids that we currently have
3201 * recorded in that peer.
3203 } else if (lp->lp_primary_nid == nid ||
3204 (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
3205 lnet_is_discovery_disabled(lp))) {
3206 rc = lnet_peer_merge_data(lp, pbuf);
3208 lpni = lnet_find_peer_ni_locked(nid);
3209 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3210 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3212 CERROR("Primary NID error %s versus %s: %d\n",
3213 libcfs_nid2str(lp->lp_primary_nid),
3214 libcfs_nid2str(nid), rc);
3216 rc = lnet_peer_merge_data(lp, pbuf);
3219 lnet_peer_ni_decref_locked(lpni);
3221 struct lnet_peer *new_lp;
3222 new_lp = lpni->lpni_peer_net->lpn_peer;
3224 * if lp has discovery/MR enabled that means new_lp
3225 * should have discovery/MR enabled as well, since
3226 * it's the same peer, which we're about to merge
3228 spin_lock(&lp->lp_lock);
3229 spin_lock(&new_lp->lp_lock);
3230 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3231 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3232 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3233 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3234 /* If we're processing a ping reply then we may be
3235 * about to send a push to the peer that we ping'd.
3236 * Since the ping reply that we're processing was
3237 * received by lp, we need to set the discovery source
3238 * NID for new_lp to the NID stored in lp.
3240 if (lp->lp_disc_src_nid != LNET_NID_ANY)
3241 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3242 spin_unlock(&new_lp->lp_lock);
3243 spin_unlock(&lp->lp_lock);
3245 rc = lnet_peer_set_primary_data(new_lp, pbuf);
3246 lnet_consolidate_routes_locked(lp, new_lp);
3247 lnet_peer_ni_decref_locked(lpni);
3251 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
3253 mutex_unlock(&the_lnet.ln_api_mutex);
3255 spin_lock(&lp->lp_lock);
3256 /* Tell discovery to re-check the peer immediately. */
3258 rc = LNET_REDISCOVER_PEER;
3263 * A ping failed. Clear the PING_FAILED state and set the
3264 * FORCE_PING state, to ensure a retry even if discovery is
3265 * disabled. This avoids being left with incorrect state.
3267 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3268 __must_hold(&lp->lp_lock)
3270 struct lnet_handle_md mdh;
3273 mdh = lp->lp_ping_mdh;
3274 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3275 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3276 lp->lp_state |= LNET_PEER_FORCE_PING;
3277 rc = lp->lp_ping_error;
3278 lp->lp_ping_error = 0;
3279 spin_unlock(&lp->lp_lock);
3281 if (!LNetMDHandleIsInvalid(mdh))
3284 CDEBUG(D_NET, "peer %s:%d\n",
3285 libcfs_nid2str(lp->lp_primary_nid), rc);
3287 spin_lock(&lp->lp_lock);
3288 return rc ? rc : LNET_REDISCOVER_PEER;
3292 * Select NID to send a Ping or Push to.
3294 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
3296 struct lnet_peer_ni *lpni;
3298 /* Look for a direct-connected NID for this peer. */
3300 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3301 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
3306 return lpni->lpni_nid;
3308 /* Look for a routed-connected NID for this peer. */
3310 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3311 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
3316 return lpni->lpni_nid;
3318 return LNET_NID_ANY;
3321 /* Active side of ping. */
3322 static int lnet_peer_send_ping(struct lnet_peer *lp)
3323 __must_hold(&lp->lp_lock)
3330 lp->lp_state |= LNET_PEER_PING_SENT;
3331 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3332 spin_unlock(&lp->lp_lock);
3334 cpt = lnet_net_lock_current();
3335 /* Refcount for MD. */
3336 lnet_peer_addref_locked(lp);
3337 pnid = lnet_peer_select_nid(lp);
3338 lnet_net_unlock(cpt);
3340 nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3342 rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
3343 the_lnet.ln_dc_handler, false);
3346 * if LNetMDBind in lnet_send_ping fails we need to decrement the
3347 * refcount on the peer, otherwise LNetMDUnlink will be called
3348 * which will eventually do that.
3352 lnet_peer_decref_locked(lp);
3353 lnet_net_unlock(cpt);
3354 rc = -rc; /* change the rc to negative value */
3356 } else if (rc < 0) {
3360 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3362 spin_lock(&lp->lp_lock);
3366 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3368 * The errors that get us here are considered hard errors and
3369 * cause Discovery to terminate. So we clear PING_SENT, but do
3370 * not set either PING_FAILED or FORCE_PING. In fact we need
3371 * to clear PING_FAILED, because the unlink event handler will
3372 * have set it if we called LNetMDUnlink() above.
3374 spin_lock(&lp->lp_lock);
3375 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3380 * This function exists because you cannot call LNetMDUnlink() from an
3383 static int lnet_peer_push_failed(struct lnet_peer *lp)
3384 __must_hold(&lp->lp_lock)
3386 struct lnet_handle_md mdh;
3389 mdh = lp->lp_push_mdh;
3390 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3391 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3392 rc = lp->lp_push_error;
3393 lp->lp_push_error = 0;
3394 spin_unlock(&lp->lp_lock);
3396 if (!LNetMDHandleIsInvalid(mdh))
3399 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3400 spin_lock(&lp->lp_lock);
3401 return rc ? rc : LNET_REDISCOVER_PEER;
3405 * Mark the peer as discovered.
3407 static int lnet_peer_discovered(struct lnet_peer *lp)
3408 __must_hold(&lp->lp_lock)
3410 lp->lp_state |= LNET_PEER_DISCOVERED;
3411 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3412 LNET_PEER_REDISCOVER);
3414 lp->lp_dc_error = 0;
3416 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3421 /* Active side of push. */
3422 static int lnet_peer_send_push(struct lnet_peer *lp)
3423 __must_hold(&lp->lp_lock)
3425 struct lnet_ping_buffer *pbuf;
3426 struct lnet_process_id id;
3431 /* Don't push to a non-multi-rail peer. */
3432 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3433 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3434 /* if peer's NIDs are uptodate then peer is discovered */
3435 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3436 rc = lnet_peer_discovered(lp);
3443 lp->lp_state |= LNET_PEER_PUSH_SENT;
3444 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3445 spin_unlock(&lp->lp_lock);
3447 cpt = lnet_net_lock_current();
3448 pbuf = the_lnet.ln_ping_target;
3449 lnet_ping_buffer_addref(pbuf);
3450 lnet_net_unlock(cpt);
3452 /* Push source MD */
3453 md.start = &pbuf->pb_info;
3454 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3455 md.threshold = 2; /* Put/Ack */
3457 md.options = LNET_MD_TRACK_RESPONSE;
3458 md.handler = the_lnet.ln_dc_handler;
3461 rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3463 lnet_ping_buffer_decref(pbuf);
3464 CERROR("Can't bind push source MD: %d\n", rc);
3467 cpt = lnet_net_lock_current();
3468 /* Refcount for MD. */
3469 lnet_peer_addref_locked(lp);
3470 id.pid = LNET_PID_LUSTRE;
3471 id.nid = lnet_peer_select_nid(lp);
3472 lnet_net_unlock(cpt);
3474 if (id.nid == LNET_NID_ANY) {
3479 rc = LNetPut(lp->lp_disc_src_nid, lp->lp_push_mdh,
3480 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3481 LNET_PROTO_PING_MATCHBITS, 0, 0);
3484 * reset the discovery nid. There is no need to restrict sending
3485 * from that source, if we call lnet_push_update_to_peers(). It'll
3486 * get set to a specific NID, if we initiate discovery from the
3489 lp->lp_disc_src_nid = LNET_NID_ANY;
3494 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3496 spin_lock(&lp->lp_lock);
3500 LNetMDUnlink(lp->lp_push_mdh);
3501 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3503 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3505 * The errors that get us here are considered hard errors and
3506 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3507 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3508 * because the unlink event handler will have set it if we
3509 * called LNetMDUnlink() above.
3511 spin_lock(&lp->lp_lock);
3512 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3517 * An unrecoverable error was encountered during discovery.
3518 * Set error status in peer and abort discovery.
3520 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3522 CDEBUG(D_NET, "Discovery error %s: %d\n",
3523 libcfs_nid2str(lp->lp_primary_nid), error);
3525 spin_lock(&lp->lp_lock);
3526 lp->lp_dc_error = error;
3527 lp->lp_state &= ~LNET_PEER_DISCOVERING;
3528 lp->lp_state |= LNET_PEER_REDISCOVER;
3529 spin_unlock(&lp->lp_lock);
3533 * Wait for work to be queued or some other change that must be
3534 * attended to. Returns non-zero if the discovery thread should shut
3537 static int lnet_peer_discovery_wait_for_work(void)
3544 cpt = lnet_net_lock_current();
3546 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3547 TASK_INTERRUPTIBLE);
3548 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3550 if (lnet_push_target_resize_needed() ||
3551 the_lnet.ln_push_target->pb_needs_post)
3553 if (!list_empty(&the_lnet.ln_dc_request))
3555 if (!list_empty(&the_lnet.ln_msg_resend))
3557 lnet_net_unlock(cpt);
3560 * wakeup max every second to check if there are peers that
3561 * have been stuck on the working queue for greater than
3564 schedule_timeout(cfs_time_seconds(1));
3565 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3566 cpt = lnet_net_lock_current();
3568 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3570 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3573 lnet_net_unlock(cpt);
3575 CDEBUG(D_NET, "woken: %d\n", rc);
3581 * Messages that were pending on a destroyed peer will be put on a global
3582 * resend list. The message resend list will be checked by
3583 * the discovery thread when it wakes up, and will resend messages. These
3584 * messages can still be sendable in the case the lpni which was the initial
3585 * cause of the message re-queue was transfered to another peer.
3587 * It is possible that LNet could be shutdown while we're iterating
3588 * through the list. lnet_shudown_lndnets() will attempt to access the
3589 * resend list, but will have to wait until the spinlock is released, by
3590 * which time there shouldn't be any more messages on the resend list.
3591 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3592 * for the messages so they can be released. The other case is that
3593 * lnet_shudown_lndnets() can finalize all the messages before this
3594 * function can visit the resend list, in which case this function will be
3597 static void lnet_resend_msgs(void)
3599 struct lnet_msg *msg, *tmp;
3603 spin_lock(&the_lnet.ln_msg_resend_lock);
3604 list_splice(&the_lnet.ln_msg_resend, &resend);
3605 spin_unlock(&the_lnet.ln_msg_resend_lock);
3607 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3608 list_del_init(&msg->msg_list);
3609 rc = lnet_send(msg->msg_src_nid_param, msg,
3610 msg->msg_rtr_nid_param);
3612 CNETERR("Error sending %s to %s: %d\n",
3613 lnet_msgtyp2str(msg->msg_type),
3614 libcfs_id2str(msg->msg_target), rc);
3615 lnet_finalize(msg, rc);
3620 /* The discovery thread. */
3621 static int lnet_peer_discovery(void *arg)
3623 struct lnet_peer *lp;
3626 wait_for_completion(&the_lnet.ln_started);
3628 CDEBUG(D_NET, "started\n");
3631 if (lnet_peer_discovery_wait_for_work())
3634 if (lnet_push_target_resize_needed())
3635 lnet_push_target_resize();
3636 else if (the_lnet.ln_push_target->pb_needs_post)
3637 lnet_push_target_post(the_lnet.ln_push_target,
3638 &the_lnet.ln_push_target_md);
3642 lnet_net_lock(LNET_LOCK_EX);
3643 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3644 lnet_net_unlock(LNET_LOCK_EX);
3649 * Process all incoming discovery work requests. When
3650 * discovery must wait on a peer to change state, it
3651 * is added to the tail of the ln_dc_working queue. A
3652 * timestamp keeps track of when the peer was added,
3653 * so we can time out discovery requests that take too
3656 while (!list_empty(&the_lnet.ln_dc_request)) {
3657 lp = list_first_entry(&the_lnet.ln_dc_request,
3658 struct lnet_peer, lp_dc_list);
3659 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3661 * set the time the peer was put on the dc_working
3662 * queue. It shouldn't remain on the queue
3663 * forever, in case the GET message (for ping)
3664 * doesn't get a REPLY or the PUT message (for
3665 * push) doesn't get an ACK.
3667 lp->lp_last_queued = ktime_get_real_seconds();
3668 lnet_net_unlock(LNET_LOCK_EX);
3670 if (lnet_push_target_resize_needed())
3671 lnet_push_target_resize();
3672 else if (the_lnet.ln_push_target->pb_needs_post)
3673 lnet_push_target_post(the_lnet.ln_push_target,
3674 &the_lnet.ln_push_target_md);
3677 * Select an action depending on the state of
3678 * the peer and whether discovery is disabled.
3679 * The check whether discovery is disabled is
3680 * done after the code that handles processing
3681 * for arrived data, cleanup for failures, and
3682 * forcing a Ping or Push.
3684 spin_lock(&lp->lp_lock);
3685 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3686 libcfs_nid2str(lp->lp_primary_nid), lp,
3688 if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3689 LNET_PEER_MARK_DELETED))
3690 rc = lnet_peer_deletion(lp);
3691 else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3692 rc = lnet_peer_data_present(lp);
3693 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3694 rc = lnet_peer_ping_failed(lp);
3695 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3696 rc = lnet_peer_push_failed(lp);
3697 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3698 rc = lnet_peer_send_ping(lp);
3699 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3700 rc = lnet_peer_send_push(lp);
3701 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3702 rc = lnet_peer_send_ping(lp);
3703 else if (lnet_peer_needs_push(lp))
3704 rc = lnet_peer_send_push(lp);
3706 rc = lnet_peer_discovered(lp);
3707 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3708 libcfs_nid2str(lp->lp_primary_nid), lp,
3710 spin_unlock(&lp->lp_lock);
3712 lnet_net_lock(LNET_LOCK_EX);
3713 if (rc == LNET_REDISCOVER_PEER) {
3714 list_move(&lp->lp_dc_list,
3715 &the_lnet.ln_dc_request);
3717 lnet_peer_discovery_error(lp, rc);
3719 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3720 lnet_peer_discovery_complete(lp);
3721 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3726 lnet_net_unlock(LNET_LOCK_EX);
3729 CDEBUG(D_NET, "stopping\n");
3731 * Clean up before telling lnet_peer_discovery_stop() that
3732 * we're done. Use wake_up() below to somewhat reduce the
3733 * size of the thundering herd if there are multiple threads
3734 * waiting on discovery of a single peer.
3737 /* Queue cleanup 1: stop all pending pings and pushes. */
3738 lnet_net_lock(LNET_LOCK_EX);
3739 while (!list_empty(&the_lnet.ln_dc_working)) {
3740 lp = list_first_entry(&the_lnet.ln_dc_working,
3741 struct lnet_peer, lp_dc_list);
3742 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3743 lnet_net_unlock(LNET_LOCK_EX);
3744 lnet_peer_cancel_discovery(lp);
3745 lnet_net_lock(LNET_LOCK_EX);
3747 lnet_net_unlock(LNET_LOCK_EX);
3749 /* Queue cleanup 2: wait for the expired queue to clear. */
3750 while (!list_empty(&the_lnet.ln_dc_expired))
3751 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3753 /* Queue cleanup 3: clear the request queue. */
3754 lnet_net_lock(LNET_LOCK_EX);
3755 while (!list_empty(&the_lnet.ln_dc_request)) {
3756 lp = list_first_entry(&the_lnet.ln_dc_request,
3757 struct lnet_peer, lp_dc_list);
3758 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3759 lnet_peer_discovery_complete(lp);
3761 lnet_net_unlock(LNET_LOCK_EX);
3763 lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3764 the_lnet.ln_dc_handler = NULL;
3766 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3767 wake_up(&the_lnet.ln_dc_waitq);
3769 CDEBUG(D_NET, "stopped\n");
3774 /* ln_api_mutex is held on entry. */
3775 int lnet_peer_discovery_start(void)
3777 struct task_struct *task;
3780 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3783 the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3784 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3785 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3788 CERROR("Can't start peer discovery thread: %d\n", rc);
3790 the_lnet.ln_dc_handler = NULL;
3792 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3795 CDEBUG(D_NET, "discovery start: %d\n", rc);
3800 /* ln_api_mutex is held on entry. */
3801 void lnet_peer_discovery_stop(void)
3803 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3806 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3807 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3809 /* In the LNetNIInit() path we may be stopping discovery before it
3810 * entered its work loop
3812 if (!completion_done(&the_lnet.ln_started))
3813 complete(&the_lnet.ln_started);
3815 wake_up(&the_lnet.ln_dc_waitq);
3817 wait_event(the_lnet.ln_dc_waitq,
3818 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3820 LASSERT(list_empty(&the_lnet.ln_dc_request));
3821 LASSERT(list_empty(&the_lnet.ln_dc_working));
3822 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3824 CDEBUG(D_NET, "discovery stopped\n");
3830 lnet_debug_peer(lnet_nid_t nid)
3832 char *aliveness = "NA";
3833 struct lnet_peer_ni *lp;
3836 cpt = lnet_cpt_of_nid(nid, NULL);
3839 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3841 lnet_net_unlock(cpt);
3842 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3846 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3847 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3849 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3850 libcfs_nid2str(lp->lpni_nid), kref_read(&lp->lpni_kref),
3851 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3852 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3853 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3855 lnet_peer_ni_decref_locked(lp);
3857 lnet_net_unlock(cpt);
3860 /* Gathering information for userspace. */
3862 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3863 char aliveness[LNET_MAX_STR_LEN],
3864 __u32 *cpt_iter, __u32 *refcount,
3865 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3866 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3867 __u32 *peer_tx_qnob)
3869 struct lnet_peer_table *peer_table;
3870 struct lnet_peer_ni *lp;
3875 /* get the number of CPTs */
3876 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3878 /* if the cpt number to be examined is >= the number of cpts in
3879 * the system then indicate that there are no more cpts to examin
3881 if (*cpt_iter >= lncpt)
3884 /* get the current table */
3885 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3886 /* if the ptable is NULL then there are no more cpts to examine */
3887 if (peer_table == NULL)
3890 lnet_net_lock(*cpt_iter);
3892 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3893 struct list_head *peers = &peer_table->pt_hash[j];
3895 list_for_each_entry(lp, peers, lpni_hashlist) {
3896 if (peer_index-- > 0)
3899 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3900 if (lnet_isrouter(lp) ||
3901 lnet_peer_aliveness_enabled(lp))
3902 snprintf(aliveness, LNET_MAX_STR_LEN,
3903 lnet_is_peer_ni_alive(lp) ? "up" : "down");
3905 *nid = lp->lpni_nid;
3906 *refcount = kref_read(&lp->lpni_kref);
3907 *ni_peer_tx_credits =
3908 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3909 *peer_tx_credits = lp->lpni_txcredits;
3910 *peer_rtr_credits = lp->lpni_rtrcredits;
3911 *peer_min_rtr_credits = lp->lpni_mintxcredits;
3912 *peer_tx_qnob = lp->lpni_txqnob;
3918 lnet_net_unlock(*cpt_iter);
3922 return found ? 0 : -ENOENT;
3925 /* ln_api_mutex is held, which keeps the peer list stable */
3926 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3928 struct lnet_ioctl_element_stats *lpni_stats;
3929 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3930 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3931 struct lnet_peer_ni_credit_info *lpni_info;
3932 struct lnet_peer_ni *lpni;
3933 struct lnet_peer *lp;
3938 lp = lnet_find_peer(cfg->prcfg_prim_nid);
3945 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3946 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3947 size *= lp->lp_nnis;
3948 if (size > cfg->prcfg_size) {
3949 cfg->prcfg_size = size;
3954 cfg->prcfg_prim_nid = lp->lp_primary_nid;
3955 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3956 cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3957 cfg->prcfg_count = lp->lp_nnis;
3958 cfg->prcfg_size = size;
3959 cfg->prcfg_state = lp->lp_state;
3961 /* Allocate helper buffers. */
3963 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3966 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3969 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3970 if (!lpni_msg_stats)
3971 goto out_free_stats;
3972 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3974 goto out_free_msg_stats;
3979 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3980 nid = lpni->lpni_nid;
3981 if (copy_to_user(bulk, &nid, sizeof(nid)))
3982 goto out_free_hstats;
3983 bulk += sizeof(nid);
3985 memset(lpni_info, 0, sizeof(*lpni_info));
3986 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3987 if (lnet_isrouter(lpni) ||
3988 lnet_peer_aliveness_enabled(lpni))
3989 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3990 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3992 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
3993 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3994 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3995 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3996 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3997 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3998 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3999 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4000 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4001 goto out_free_hstats;
4002 bulk += sizeof(*lpni_info);
4004 memset(lpni_stats, 0, sizeof(*lpni_stats));
4005 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4006 LNET_STATS_TYPE_SEND);
4007 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4008 LNET_STATS_TYPE_RECV);
4009 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4010 LNET_STATS_TYPE_DROP);
4011 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4012 goto out_free_hstats;
4013 bulk += sizeof(*lpni_stats);
4014 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4015 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4016 goto out_free_hstats;
4017 bulk += sizeof(*lpni_msg_stats);
4018 lpni_hstats->hlpni_network_timeout =
4019 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4020 lpni_hstats->hlpni_remote_dropped =
4021 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4022 lpni_hstats->hlpni_remote_timeout =
4023 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4024 lpni_hstats->hlpni_remote_error =
4025 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4026 lpni_hstats->hlpni_health_value =
4027 atomic_read(&lpni->lpni_healthv);
4028 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4029 goto out_free_hstats;
4030 bulk += sizeof(*lpni_hstats);
4035 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4037 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4039 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4041 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4043 lnet_peer_decref_locked(lp);
4048 /* must hold net_lock/0 */
4050 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4051 struct list_head *recovery_queue,
4054 /* the mt could've shutdown and cleaned up the queues */
4055 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4058 if (!list_empty(&lpni->lpni_recovery))
4061 if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4064 if (!lpni->lpni_last_alive) {
4066 "lpni %s(%p) not eligible for recovery last alive %lld\n",
4067 libcfs_nid2str(lpni->lpni_nid), lpni,
4068 lpni->lpni_last_alive);
4072 if (now > lpni->lpni_last_alive + lnet_recovery_limit) {
4073 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4074 libcfs_nid2str(lpni->lpni_nid),
4075 lpni->lpni_last_alive);
4076 /* Reset the ping count so that if this peer NI is added back to
4077 * the recovery queue we will send the first ping right away.
4079 lpni->lpni_ping_count = 0;
4083 /* This peer NI is going on the recovery queue, so take a ref on it */
4084 lnet_peer_ni_addref_locked(lpni);
4086 lnet_peer_ni_set_next_ping(lpni, now);
4088 CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4089 libcfs_nid2str(lpni->lpni_nid),
4090 lpni->lpni_ping_count,
4091 lpni->lpni_next_ping,
4092 lpni->lpni_last_alive,
4093 atomic_read(&lpni->lpni_healthv));
4095 list_add_tail(&lpni->lpni_recovery, recovery_queue);
4098 /* Call with the ln_api_mutex held */
4100 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
4102 struct lnet_peer_table *ptable;
4103 struct lnet_peer *lp;
4104 struct lnet_peer_net *lpn;
4105 struct lnet_peer_ni *lpni;
4110 if (the_lnet.ln_state != LNET_STATE_RUNNING)
4113 now = ktime_get_seconds();
4116 lnet_net_lock(LNET_LOCK_EX);
4117 lpni = lnet_find_peer_ni_locked(nid);
4119 lnet_net_unlock(LNET_LOCK_EX);
4122 atomic_set(&lpni->lpni_healthv, value);
4123 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4124 &the_lnet.ln_mt_peerNIRecovq, now);
4125 lnet_peer_ni_decref_locked(lpni);
4126 lnet_net_unlock(LNET_LOCK_EX);
4130 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4133 * Walk all the peers and reset the health value for each one to the
4136 lnet_net_lock(LNET_LOCK_EX);
4137 for (cpt = 0; cpt < lncpt; cpt++) {
4138 ptable = the_lnet.ln_peer_tables[cpt];
4139 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4140 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4141 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4143 atomic_set(&lpni->lpni_healthv, value);
4144 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4145 &the_lnet.ln_mt_peerNIRecovq, now);
4150 lnet_net_unlock(LNET_LOCK_EX);