4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
34 #define DEBUG_SUBSYSTEM S_LNET
36 #include <linux/sched.h>
37 #ifdef HAVE_SCHED_HEADERS
38 #include <linux/sched/signal.h>
40 #include <linux/uaccess.h>
42 #include <lnet/udsp.h>
43 #include <lnet/lib-lnet.h>
44 #include <uapi/linux/lnet/lnet-dlc.h>
46 /* Value indicating that recovery needs to re-check a peer immediately. */
47 #define LNET_REDISCOVER_PEER (1)
49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
54 if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
55 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
56 lnet_peer_ni_decref_locked(lpni);
61 lnet_peer_net_added(struct lnet_net *net)
63 struct lnet_peer_ni *lpni, *tmp;
65 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
66 lpni_on_remote_peer_ni_list) {
68 if (LNET_NID_NET(&lpni->lpni_nid) == net->net_id) {
71 spin_lock(&lpni->lpni_lock);
72 lpni->lpni_txcredits =
73 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
74 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
75 lpni->lpni_rtrcredits =
76 lnet_peer_buffer_credits(lpni->lpni_net);
77 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
78 spin_unlock(&lpni->lpni_lock);
80 lnet_peer_remove_from_remote_list(lpni);
86 lnet_peer_tables_destroy(void)
88 struct lnet_peer_table *ptable;
89 struct list_head *hash;
93 if (!the_lnet.ln_peer_tables)
96 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
97 hash = ptable->pt_hash;
98 if (!hash) /* not intialized */
101 LASSERT(list_empty(&ptable->pt_zombie_list));
103 ptable->pt_hash = NULL;
104 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
105 LASSERT(list_empty(&hash[j]));
107 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
110 cfs_percpt_free(the_lnet.ln_peer_tables);
111 the_lnet.ln_peer_tables = NULL;
115 lnet_peer_tables_create(void)
117 struct lnet_peer_table *ptable;
118 struct list_head *hash;
122 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
124 if (the_lnet.ln_peer_tables == NULL) {
125 CERROR("Failed to allocate cpu-partition peer tables\n");
129 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131 LNET_PEER_HASH_SIZE * sizeof(*hash));
133 CERROR("Failed to create peer hash table\n");
134 lnet_peer_tables_destroy();
138 spin_lock_init(&ptable->pt_zombie_lock);
139 INIT_LIST_HEAD(&ptable->pt_zombie_list);
141 INIT_LIST_HEAD(&ptable->pt_peer_list);
143 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144 INIT_LIST_HEAD(&hash[j]);
145 ptable->pt_hash = hash; /* sign of initialization */
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(struct lnet_nid *nid)
154 struct lnet_peer_ni *lpni;
155 struct lnet_net *net;
158 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
160 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
164 INIT_LIST_HEAD(&lpni->lpni_txq);
165 INIT_LIST_HEAD(&lpni->lpni_hashlist);
166 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167 INIT_LIST_HEAD(&lpni->lpni_recovery);
168 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169 INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
170 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171 kref_init(&lpni->lpni_kref);
172 lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
174 spin_lock_init(&lpni->lpni_lock);
176 if (lnet_peers_start_down())
177 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
179 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
180 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
181 lpni->lpni_nid = *nid;
182 lpni->lpni_cpt = cpt;
183 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
185 net = lnet_get_net_locked(LNET_NID_NET(nid));
186 lpni->lpni_net = net;
188 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
189 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
190 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
191 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
194 * This peer_ni is not on a local network, so we
195 * cannot add the credits here. In case the net is
196 * added later, add the peer_ni to the remote peer ni
197 * list so it can be easily found and revisited.
199 /* FIXME: per-net implementation instead? */
200 lnet_peer_ni_addref_locked(lpni);
201 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
202 &the_lnet.ln_remote_peer_ni_list);
205 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
210 static struct lnet_peer_net *
211 lnet_peer_net_alloc(__u32 net_id)
213 struct lnet_peer_net *lpn;
215 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
219 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
220 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
221 lpn->lpn_net_id = net_id;
222 lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
224 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
230 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
232 struct lnet_peer *lp;
234 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
236 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
237 LASSERT(list_empty(&lpn->lpn_peer_nis));
238 LASSERT(list_empty(&lpn->lpn_peer_nets));
240 lpn->lpn_peer = NULL;
241 LIBCFS_FREE(lpn, sizeof(*lpn));
243 lnet_peer_decref_locked(lp);
246 static struct lnet_peer *
247 lnet_peer_alloc(struct lnet_nid *nid)
249 struct lnet_peer *lp;
251 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
255 INIT_LIST_HEAD(&lp->lp_rtrq);
256 INIT_LIST_HEAD(&lp->lp_routes);
257 INIT_LIST_HEAD(&lp->lp_peer_list);
258 INIT_LIST_HEAD(&lp->lp_peer_nets);
259 INIT_LIST_HEAD(&lp->lp_dc_list);
260 INIT_LIST_HEAD(&lp->lp_dc_pendq);
261 INIT_LIST_HEAD(&lp->lp_rtr_list);
262 init_waitqueue_head(&lp->lp_dc_waitq);
263 spin_lock_init(&lp->lp_lock);
264 lp->lp_primary_nid = *nid;
265 lp->lp_disc_src_nid = LNET_ANY_NID;
266 lp->lp_disc_dst_nid = LNET_ANY_NID;
267 if (lnet_peers_start_down())
268 lp->lp_alive = false;
273 * all peers created on a router should have health on
274 * if it's not already on.
276 if (the_lnet.ln_routing && !lnet_health_sensitivity)
277 lp->lp_health_sensitivity = 1;
280 * Turn off discovery for loopback peer. If you're creating a peer
281 * for the loopback interface then that was initiated when we
282 * attempted to send a message over the loopback. There is no need
283 * to ever use a different interface when sending messages to
287 lp->lp_state = LNET_PEER_NO_DISCOVERY;
288 lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
290 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
296 lnet_destroy_peer_locked(struct lnet_peer *lp)
298 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
300 LASSERT(atomic_read(&lp->lp_refcount) == 0);
301 LASSERT(lp->lp_rtr_refcount == 0);
302 LASSERT(list_empty(&lp->lp_peer_nets));
303 LASSERT(list_empty(&lp->lp_peer_list));
304 LASSERT(list_empty(&lp->lp_dc_list));
307 lnet_ping_buffer_decref(lp->lp_data);
310 * if there are messages still on the pending queue, then make
311 * sure to queue them on the ln_msg_resend list so they can be
312 * resent at a later point if the discovery thread is still
314 * If the discovery thread has stopped, then the wakeup will be a
315 * no-op, and it is expected the lnet_shutdown_lndnets() will
316 * eventually be called, which will traverse this list and
317 * finalize the messages on the list.
318 * We can not resend them now because we're holding the cpt lock.
319 * Releasing the lock can cause an inconsistent state
321 spin_lock(&the_lnet.ln_msg_resend_lock);
322 spin_lock(&lp->lp_lock);
323 list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
324 spin_unlock(&lp->lp_lock);
325 spin_unlock(&the_lnet.ln_msg_resend_lock);
326 wake_up(&the_lnet.ln_dc_waitq);
328 LIBCFS_FREE(lp, sizeof(*lp));
332 * Detach a peer_ni from its peer_net. If this was the last peer_ni on
333 * that peer_net, detach the peer_net from the peer.
335 * Call with lnet_net_lock/EX held
338 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
340 struct lnet_peer_table *ptable;
341 struct lnet_peer_net *lpn;
342 struct lnet_peer *lp;
345 * Belts and suspenders: gracefully handle teardown of a
346 * partially connected peer_ni.
348 lpn = lpni->lpni_peer_net;
350 list_del_init(&lpni->lpni_peer_nis);
352 * If there are no lpni's left, we detach lpn from
353 * lp_peer_nets, so it cannot be found anymore.
355 if (list_empty(&lpn->lpn_peer_nis))
356 list_del_init(&lpn->lpn_peer_nets);
358 /* Update peer NID count. */
363 * If there are no more peer nets, make the peer unfindable
364 * via the peer_tables.
366 * Otherwise, if the peer is DISCOVERED, tell discovery to
367 * take another look at it. This is a no-op if discovery for
368 * this peer did the detaching.
370 if (list_empty(&lp->lp_peer_nets)) {
371 list_del_init(&lp->lp_peer_list);
372 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
374 } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
375 /* Discovery isn't running, nothing to do here. */
376 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
377 lnet_peer_queue_for_discovery(lp);
378 wake_up(&the_lnet.ln_dc_waitq);
380 CDEBUG(D_NET, "peer %s NID %s\n",
381 libcfs_nidstr(&lp->lp_primary_nid),
382 libcfs_nidstr(&lpni->lpni_nid));
385 /* called with lnet_net_lock LNET_LOCK_EX held */
387 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
389 struct lnet_peer_table *ptable = NULL;
391 /* don't remove a peer_ni if it's also a gateway */
392 if (lnet_isrouter(lpni) && !force) {
393 CERROR("Peer NI %s is a gateway. Can not delete it\n",
394 libcfs_nidstr(&lpni->lpni_nid));
398 lnet_peer_remove_from_remote_list(lpni);
400 /* remove peer ni from the hash list. */
401 list_del_init(&lpni->lpni_hashlist);
404 * indicate the peer is being deleted so the monitor thread can
405 * remove it from the recovery queue.
407 spin_lock(&lpni->lpni_lock);
408 lpni->lpni_state |= LNET_PEER_NI_DELETING;
409 spin_unlock(&lpni->lpni_lock);
411 /* decrement the ref count on the peer table */
412 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
415 * The peer_ni can no longer be found with a lookup. But there
416 * can be current users, so keep track of it on the zombie
417 * list until the reference count has gone to zero.
419 * The last reference may be lost in a place where the
420 * lnet_net_lock locks only a single cpt, and that cpt may not
421 * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
424 spin_lock(&ptable->pt_zombie_lock);
425 list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
426 ptable->pt_zombies++;
427 spin_unlock(&ptable->pt_zombie_lock);
429 /* no need to keep this peer_ni on the hierarchy anymore */
430 lnet_peer_detach_peer_ni_locked(lpni);
432 /* remove hashlist reference on peer_ni */
433 lnet_peer_ni_decref_locked(lpni);
438 void lnet_peer_uninit(void)
440 struct lnet_peer_ni *lpni, *tmp;
442 lnet_net_lock(LNET_LOCK_EX);
444 /* remove all peer_nis from the remote peer and the hash list */
445 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
446 lpni_on_remote_peer_ni_list)
447 lnet_peer_ni_del_locked(lpni, false);
449 lnet_peer_tables_destroy();
451 lnet_net_unlock(LNET_LOCK_EX);
455 lnet_peer_del_locked(struct lnet_peer *peer)
457 struct lnet_peer_ni *lpni = NULL, *lpni2;
460 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&peer->lp_primary_nid));
462 spin_lock(&peer->lp_lock);
463 peer->lp_state |= LNET_PEER_MARK_DELETED;
464 spin_unlock(&peer->lp_lock);
466 lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
467 while (lpni != NULL) {
468 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
469 rc = lnet_peer_ni_del_locked(lpni, false);
479 * Discovering this peer is taking too long. Cancel any Ping or Push
480 * that discovery is waiting on by unlinking the relevant MDs. The
481 * lnet_discovery_event_handler() will proceed from here and complete
484 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
486 struct lnet_handle_md ping_mdh;
487 struct lnet_handle_md push_mdh;
489 LNetInvalidateMDHandle(&ping_mdh);
490 LNetInvalidateMDHandle(&push_mdh);
492 spin_lock(&lp->lp_lock);
493 if (lp->lp_state & LNET_PEER_PING_SENT) {
494 ping_mdh = lp->lp_ping_mdh;
495 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
497 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
498 push_mdh = lp->lp_push_mdh;
499 LNetInvalidateMDHandle(&lp->lp_push_mdh);
501 spin_unlock(&lp->lp_lock);
503 if (!LNetMDHandleIsInvalid(ping_mdh))
504 LNetMDUnlink(ping_mdh);
505 if (!LNetMDHandleIsInvalid(push_mdh))
506 LNetMDUnlink(push_mdh);
510 lnet_peer_del(struct lnet_peer *peer)
514 lnet_peer_cancel_discovery(peer);
515 lnet_net_lock(LNET_LOCK_EX);
516 rc = lnet_peer_del_locked(peer);
517 lnet_net_unlock(LNET_LOCK_EX);
523 * Delete a NID from a peer. Call with ln_api_mutex held.
526 * -EPERM: Non-DLC deletion from DLC-configured peer.
527 * -ENOENT: No lnet_peer_ni corresponding to the nid.
528 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
529 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
532 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
534 struct lnet_peer_ni *lpni;
535 struct lnet_nid primary_nid = lp->lp_primary_nid;
538 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
540 lnet_nid4_to_nid(nid4, &nid);
541 if (!(flags & LNET_PEER_CONFIGURED)) {
542 if (lp->lp_state & LNET_PEER_CONFIGURED) {
548 lpni = lnet_peer_ni_find_locked(&nid);
553 lnet_peer_ni_decref_locked(lpni);
554 if (lp != lpni->lpni_peer_net->lpn_peer) {
560 * This function only allows deletion of the primary NID if it
563 if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && !force) {
568 lnet_net_lock(LNET_LOCK_EX);
570 if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && force) {
571 struct lnet_peer_ni *lpni2;
572 /* assign the next peer_ni to be the primary */
573 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
575 lp->lp_primary_nid = lpni2->lpni_nid;
577 rc = lnet_peer_ni_del_locked(lpni, force);
579 lnet_net_unlock(LNET_LOCK_EX);
582 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
583 libcfs_nidstr(&primary_nid), libcfs_nidstr(&nid),
590 lnet_peer_table_cleanup_locked(struct lnet_net *net,
591 struct lnet_peer_table *ptable)
594 struct lnet_peer_ni *next;
595 struct lnet_peer_ni *lpni;
596 struct lnet_peer *peer;
598 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
599 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
601 if (net != NULL && net != lpni->lpni_net)
604 peer = lpni->lpni_peer_net->lpn_peer;
605 if (!nid_same(&peer->lp_primary_nid,
607 lnet_peer_ni_del_locked(lpni, false);
611 * Removing the primary NID implies removing
612 * the entire peer. Advance next beyond any
613 * peer_ni that belongs to the same peer.
615 list_for_each_entry_from(next, &ptable->pt_hash[i],
617 if (next->lpni_peer_net->lpn_peer != peer)
620 lnet_peer_del_locked(peer);
626 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
628 wait_var_event_warning(&ptable->pt_zombies,
629 ptable->pt_zombies == 0,
630 "Waiting for %d zombies on peer table\n",
635 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
636 struct lnet_peer_table *ptable)
638 struct lnet_peer_ni *lp;
639 struct lnet_peer_ni *tmp;
643 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
644 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
646 if (net != lp->lpni_net)
649 if (!lnet_isrouter(lp))
652 /* FIXME handle large-addr nid */
653 gw_nid = lnet_nid_to_nid4(
654 &lp->lpni_peer_net->lpn_peer->lp_primary_nid);
656 lnet_net_unlock(LNET_LOCK_EX);
657 lnet_del_route(LNET_NET_ANY, gw_nid);
658 lnet_net_lock(LNET_LOCK_EX);
664 lnet_peer_tables_cleanup(struct lnet_net *net)
667 struct lnet_peer_table *ptable;
669 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
670 /* If just deleting the peers for a NI, get rid of any routes these
671 * peers are gateways for. */
672 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
673 lnet_net_lock(LNET_LOCK_EX);
674 lnet_peer_table_del_rtrs_locked(net, ptable);
675 lnet_net_unlock(LNET_LOCK_EX);
678 /* Start the cleanup process */
679 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
680 lnet_net_lock(LNET_LOCK_EX);
681 lnet_peer_table_cleanup_locked(net, ptable);
682 lnet_net_unlock(LNET_LOCK_EX);
685 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
686 lnet_peer_ni_finalize_wait(ptable);
689 static struct lnet_peer_ni *
690 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, struct lnet_nid *nid)
692 struct list_head *peers;
693 struct lnet_peer_ni *lp;
695 if (the_lnet.ln_state != LNET_STATE_RUNNING)
698 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
699 list_for_each_entry(lp, peers, lpni_hashlist) {
700 if (nid_same(&lp->lpni_nid, nid)) {
701 lnet_peer_ni_addref_locked(lp);
709 struct lnet_peer_ni *
710 lnet_find_peer_ni_locked(lnet_nid_t nid4)
712 struct lnet_peer_ni *lpni;
713 struct lnet_peer_table *ptable;
717 lnet_nid4_to_nid(nid4, &nid);
719 cpt = lnet_nid_cpt_hash(&nid, LNET_CPT_NUMBER);
721 ptable = the_lnet.ln_peer_tables[cpt];
722 lpni = lnet_get_peer_ni_locked(ptable, &nid);
727 struct lnet_peer_ni *
728 lnet_peer_ni_find_locked(struct lnet_nid *nid)
730 struct lnet_peer_ni *lpni;
731 struct lnet_peer_table *ptable;
734 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
736 ptable = the_lnet.ln_peer_tables[cpt];
737 lpni = lnet_get_peer_ni_locked(ptable, nid);
742 struct lnet_peer_ni *
743 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
745 struct lnet_peer_net *lpn;
746 struct lnet_peer_ni *lpni;
748 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
752 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
753 if (lnet_nid_to_nid4(&lpni->lpni_nid) == nid)
760 struct lnet_peer_ni *
761 lnet_peer_ni_get_locked(struct lnet_peer *lp, struct lnet_nid *nid)
763 struct lnet_peer_net *lpn;
764 struct lnet_peer_ni *lpni;
766 lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
770 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
771 if (nid_same(&lpni->lpni_nid, nid))
779 lnet_find_peer(lnet_nid_t nid)
781 struct lnet_peer_ni *lpni;
782 struct lnet_peer *lp = NULL;
785 cpt = lnet_net_lock_current();
786 lpni = lnet_find_peer_ni_locked(nid);
788 lp = lpni->lpni_peer_net->lpn_peer;
789 lnet_peer_addref_locked(lp);
790 lnet_peer_ni_decref_locked(lpni);
792 lnet_net_unlock(cpt);
797 struct lnet_peer_net *
798 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
800 struct lnet_peer_net *net;
803 /* no net id provided return the first net */
804 net = list_first_entry_or_null(&lp->lp_peer_nets,
805 struct lnet_peer_net,
811 /* find the net after the one provided */
812 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
813 if (net->lpn_net_id == prev_lpn_id) {
815 * if we reached the end of the list loop to the
818 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
819 return list_first_entry_or_null(&lp->lp_peer_nets,
820 struct lnet_peer_net,
823 return list_next_entry(net, lpn_peer_nets);
830 struct lnet_peer_ni *
831 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
832 struct lnet_peer_net *peer_net,
833 struct lnet_peer_ni *prev)
835 struct lnet_peer_ni *lpni;
836 struct lnet_peer_net *net = peer_net;
840 if (list_empty(&peer->lp_peer_nets))
843 net = list_entry(peer->lp_peer_nets.next,
844 struct lnet_peer_net,
847 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
853 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
855 * if you reached the end of the peer ni list and the peer
856 * net is specified then there are no more peer nis in that
863 * we reached the end of this net ni list. move to the
866 if (prev->lpni_peer_net->lpn_peer_nets.next ==
868 /* no more nets and no more NIs. */
871 /* get the next net */
872 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
873 struct lnet_peer_net,
875 /* get the ni on it */
876 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
882 /* there are more nis left */
883 lpni = list_entry(prev->lpni_peer_nis.next,
884 struct lnet_peer_ni, lpni_peer_nis);
889 /* Call with the ln_api_mutex held */
890 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
892 struct lnet_process_id id;
893 struct lnet_peer_table *ptable;
894 struct lnet_peer *lp;
903 if (the_lnet.ln_state != LNET_STATE_RUNNING)
906 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
909 * Count the number of peers, and return E2BIG if the buffer
910 * is too small. We'll also return the desired size.
913 for (cpt = 0; cpt < lncpt; cpt++) {
914 ptable = the_lnet.ln_peer_tables[cpt];
915 count += ptable->pt_peers;
917 size = count * sizeof(*ids);
922 * Walk the peer lists and copy out the primary nids.
923 * This is safe because the peer lists are only modified
924 * while the ln_api_mutex is held. So we don't need to
925 * hold the lnet_net_lock as well, and can therefore
926 * directly call copy_to_user().
929 memset(&id, 0, sizeof(id));
930 id.pid = LNET_PID_LUSTRE;
932 for (cpt = 0; cpt < lncpt; cpt++) {
933 ptable = the_lnet.ln_peer_tables[cpt];
934 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
935 if (!nid_is_nid4(&lp->lp_primary_nid))
939 id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
940 if (copy_to_user(&ids[i], &id, sizeof(id)))
953 * Start pushes to peers that need to be updated for a configuration
954 * change on this node.
957 lnet_push_update_to_peers(int force)
959 struct lnet_peer_table *ptable;
960 struct lnet_peer *lp;
964 lnet_net_lock(LNET_LOCK_EX);
965 if (lnet_peer_discovery_disabled)
967 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
968 for (cpt = 0; cpt < lncpt; cpt++) {
969 ptable = the_lnet.ln_peer_tables[cpt];
970 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
972 spin_lock(&lp->lp_lock);
973 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
974 lp->lp_state |= LNET_PEER_FORCE_PUSH;
975 spin_unlock(&lp->lp_lock);
977 if (lnet_peer_needs_push(lp))
978 lnet_peer_queue_for_discovery(lp);
981 lnet_net_unlock(LNET_LOCK_EX);
982 wake_up(&the_lnet.ln_dc_waitq);
985 /* find the NID in the preferred gateways for the remote peer
987 * false: list is not empty and NID is not preferred
988 * false: list is empty
989 * true: nid is found in the list
992 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
993 struct lnet_nid *gw_nid)
995 struct lnet_nid_list *ne;
997 CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
998 libcfs_nidstr(&lpni->lpni_nid),
999 list_empty(&lpni->lpni_rtr_pref_nids));
1001 if (list_empty(&lpni->lpni_rtr_pref_nids))
1004 /* iterate through all the preferred NIDs and see if any of them
1005 * matches the provided gw_nid
1007 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1008 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
1009 libcfs_nidstr(&ne->nl_nid),
1010 libcfs_nidstr(gw_nid));
1011 if (nid_same(&ne->nl_nid, gw_nid))
1019 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
1021 struct list_head zombies;
1022 struct lnet_nid_list *ne;
1023 struct lnet_nid_list *tmp;
1024 int cpt = lpni->lpni_cpt;
1026 INIT_LIST_HEAD(&zombies);
1029 list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
1030 lnet_net_unlock(cpt);
1032 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1033 list_del(&ne->nl_list);
1034 LIBCFS_FREE(ne, sizeof(*ne));
1039 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
1040 struct lnet_nid *gw_nid)
1042 int cpt = lpni->lpni_cpt;
1043 struct lnet_nid_list *ne = NULL;
1045 /* This function is called with api_mutex held. When the api_mutex
1046 * is held the list can not be modified, as it is only modified as
1047 * a result of applying a UDSP and that happens under api_mutex
1050 __must_hold(&the_lnet.ln_api_mutex);
1052 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1053 if (nid_same(&ne->nl_nid, gw_nid))
1057 LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1061 ne->nl_nid = *gw_nid;
1063 /* Lock the cpt to protect against addition and checks in the
1064 * selection algorithm
1067 list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1068 lnet_net_unlock(cpt);
1074 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1075 * this is a preferred point-to-point path. Call with lnet_net_lock in
1079 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1081 struct lnet_nid_list *ne;
1083 if (lpni->lpni_pref_nnids == 0)
1085 if (lpni->lpni_pref_nnids == 1)
1086 return nid_same(&lpni->lpni_pref.nid, nid);
1087 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1088 if (nid_same(&ne->nl_nid, nid))
1095 * Set a single ni as preferred, provided no preferred ni is already
1096 * defined. Only to be used for non-multi-rail peer_ni.
1099 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni,
1100 struct lnet_nid *nid)
1106 spin_lock(&lpni->lpni_lock);
1107 if (LNET_NID_IS_ANY(nid)) {
1109 } else if (lpni->lpni_pref_nnids > 0) {
1111 } else if (lpni->lpni_pref_nnids == 0) {
1112 lpni->lpni_pref.nid = *nid;
1113 lpni->lpni_pref_nnids = 1;
1114 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1116 spin_unlock(&lpni->lpni_lock);
1118 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1119 libcfs_nidstr(&lpni->lpni_nid), libcfs_nidstr(nid), rc);
1124 * Clear the preferred NID from a non-multi-rail peer_ni, provided
1125 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1128 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1132 spin_lock(&lpni->lpni_lock);
1133 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1134 lpni->lpni_pref_nnids = 0;
1135 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1136 } else if (lpni->lpni_pref_nnids == 0) {
1141 spin_unlock(&lpni->lpni_lock);
1143 CDEBUG(D_NET, "peer %s: %d\n",
1144 libcfs_nidstr(&lpni->lpni_nid), rc);
1149 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1151 lpni->lpni_sel_priority = priority;
1155 * Clear the preferred NIDs from a non-multi-rail peer.
1158 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1160 struct lnet_peer_ni *lpni = NULL;
1162 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1163 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1167 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1169 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1170 struct lnet_nid_list *ne1 = NULL;
1171 struct lnet_nid_list *ne2 = NULL;
1172 struct lnet_nid *tmp_nid = NULL;
1175 if (LNET_NID_IS_ANY(nid)) {
1180 if (lpni->lpni_pref_nnids == 1 &&
1181 nid_same(&lpni->lpni_pref.nid, nid)) {
1186 /* A non-MR node may have only one preferred NI per peer_ni */
1187 if (lpni->lpni_pref_nnids > 0 &&
1188 !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1193 /* add the new preferred nid to the list of preferred nids */
1194 if (lpni->lpni_pref_nnids != 0) {
1195 size_t alloc_size = sizeof(*ne1);
1197 if (lpni->lpni_pref_nnids == 1) {
1198 tmp_nid = &lpni->lpni_pref.nid;
1199 INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1202 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1203 if (nid_same(&ne1->nl_nid, nid)) {
1209 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1216 /* move the originally stored nid to the list */
1217 if (lpni->lpni_pref_nnids == 1) {
1218 LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1219 lpni->lpni_cpt, alloc_size);
1224 INIT_LIST_HEAD(&ne2->nl_list);
1225 ne2->nl_nid = *tmp_nid;
1230 lnet_net_lock(LNET_LOCK_EX);
1231 spin_lock(&lpni->lpni_lock);
1232 if (lpni->lpni_pref_nnids == 0) {
1233 lpni->lpni_pref.nid = *nid;
1236 list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1237 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1239 lpni->lpni_pref_nnids++;
1240 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1241 spin_unlock(&lpni->lpni_lock);
1242 lnet_net_unlock(LNET_LOCK_EX);
1245 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1246 spin_lock(&lpni->lpni_lock);
1247 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1248 spin_unlock(&lpni->lpni_lock);
1250 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1251 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1256 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1258 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1259 struct lnet_nid_list *ne = NULL;
1262 if (lpni->lpni_pref_nnids == 0) {
1267 if (lpni->lpni_pref_nnids == 1) {
1268 if (!nid_same(&lpni->lpni_pref.nid, nid)) {
1273 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1274 if (nid_same(&ne->nl_nid, nid))
1275 goto remove_nid_entry;
1283 lnet_net_lock(LNET_LOCK_EX);
1284 spin_lock(&lpni->lpni_lock);
1285 if (lpni->lpni_pref_nnids == 1)
1286 lpni->lpni_pref.nid = LNET_ANY_NID;
1288 list_del_init(&ne->nl_list);
1289 if (lpni->lpni_pref_nnids == 2) {
1290 struct lnet_nid_list *ne, *tmp;
1292 list_for_each_entry_safe(ne, tmp,
1293 &lpni->lpni_pref.nids,
1295 lpni->lpni_pref.nid = ne->nl_nid;
1296 list_del_init(&ne->nl_list);
1297 LIBCFS_FREE(ne, sizeof(*ne));
1301 lpni->lpni_pref_nnids--;
1302 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1303 spin_unlock(&lpni->lpni_lock);
1304 lnet_net_unlock(LNET_LOCK_EX);
1307 LIBCFS_FREE(ne, sizeof(*ne));
1309 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1310 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1315 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1317 struct list_head zombies;
1318 struct lnet_nid_list *ne;
1319 struct lnet_nid_list *tmp;
1321 INIT_LIST_HEAD(&zombies);
1323 lnet_net_lock(LNET_LOCK_EX);
1324 if (lpni->lpni_pref_nnids == 1)
1325 lpni->lpni_pref.nid = LNET_ANY_NID;
1326 else if (lpni->lpni_pref_nnids > 1)
1327 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1328 lpni->lpni_pref_nnids = 0;
1329 lnet_net_unlock(LNET_LOCK_EX);
1331 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1332 list_del_init(&ne->nl_list);
1333 LIBCFS_FREE(ne, sizeof(*ne));
1338 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1340 /* FIXME handle large-addr nid */
1341 struct lnet_peer_ni *lpni;
1342 lnet_nid_t primary_nid = nid;
1344 lpni = lnet_find_peer_ni_locked(nid);
1346 primary_nid = lnet_nid_to_nid4(
1347 &lpni->lpni_peer_net->lpn_peer->lp_primary_nid);
1348 lnet_peer_ni_decref_locked(lpni);
1355 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1356 __must_hold(&lp->lp_lock)
1358 if (lnet_peer_discovery_disabled)
1361 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1362 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1373 lnet_is_discovery_disabled(struct lnet_peer *lp)
1377 spin_lock(&lp->lp_lock);
1378 rc = lnet_is_discovery_disabled_locked(lp);
1379 spin_unlock(&lp->lp_lock);
1385 LNetAddPeer(lnet_nid_t *nids, __u32 num_nids)
1387 lnet_nid_t pnid = 0;
1391 if (!nids || num_nids < 1)
1394 rc = LNetNIInit(LNET_PID_ANY);
1398 mutex_lock(&the_lnet.ln_api_mutex);
1400 mr = lnet_peer_discovery_disabled == 0;
1403 for (i = 0; i < num_nids; i++) {
1404 if (nids[i] == LNET_NID_LO_0)
1409 rc = lnet_add_peer_ni(pnid, LNET_NID_ANY, mr, true);
1410 } else if (lnet_peer_discovery_disabled) {
1411 rc = lnet_add_peer_ni(nids[i], LNET_NID_ANY, mr, true);
1413 rc = lnet_add_peer_ni(pnid, nids[i], mr, true);
1416 if (rc && rc != -EEXIST)
1421 mutex_unlock(&the_lnet.ln_api_mutex);
1425 return rc == -EEXIST ? 0 : rc;
1427 EXPORT_SYMBOL(LNetAddPeer);
1429 /* FIXME support large-addr nid */
1431 LNetPrimaryNID(lnet_nid_t nid)
1433 struct lnet_peer *lp;
1434 struct lnet_peer_ni *lpni;
1435 lnet_nid_t primary_nid = nid;
1439 if (nid == LNET_NID_LO_0)
1440 return LNET_NID_LO_0;
1442 cpt = lnet_net_lock_current();
1443 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1448 lp = lpni->lpni_peer_net->lpn_peer;
1450 /* If discovery is disabled locally then we needn't bother running
1451 * discovery here because discovery will not modify whatever
1452 * primary NID is currently set for this peer. If the specified peer is
1453 * down then this discovery can introduce long delays into the mount
1454 * process, so skip it if it isn't necessary.
1456 while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1457 spin_lock(&lp->lp_lock);
1458 /* force a full discovery cycle */
1459 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1460 spin_unlock(&lp->lp_lock);
1462 rc = lnet_discover_peer_locked(lpni, cpt, true);
1465 /* The lpni (or lp) for this NID may have changed and our ref is
1466 * the only thing keeping the old one around. Release the ref
1467 * and lookup the lpni again
1469 lnet_peer_ni_decref_locked(lpni);
1470 lpni = lnet_find_peer_ni_locked(nid);
1475 lp = lpni->lpni_peer_net->lpn_peer;
1477 /* If we find that the peer has discovery disabled then we will
1478 * not modify whatever primary NID is currently set for this
1479 * peer. Thus, we can break out of this loop even if the peer
1480 * is not fully up to date.
1482 if (lnet_is_discovery_disabled(lp))
1485 primary_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
1487 lnet_peer_ni_decref_locked(lpni);
1489 lnet_net_unlock(cpt);
1491 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1492 libcfs_nid2str(primary_nid), rc);
1495 EXPORT_SYMBOL(LNetPrimaryNID);
1497 struct lnet_peer_net *
1498 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1500 struct lnet_peer_net *peer_net;
1501 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1502 if (peer_net->lpn_net_id == net_id)
1509 * Attach a peer_ni to a peer_net and peer. This function assumes
1510 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1511 * may be attached to a different peer, in which case it will be
1512 * properly detached first. The whole operation is done atomically.
1514 * This function consumes the reference on lpni and Always returns 0.
1515 * This is the last function called from functions that do return an
1516 * int, so returning 0 here allows the compiler to do a tail call.
1519 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1520 struct lnet_peer_net *lpn,
1521 struct lnet_peer_ni *lpni,
1524 struct lnet_peer_table *ptable;
1525 bool new_lpn = false;
1528 /* Install the new peer_ni */
1529 lnet_net_lock(LNET_LOCK_EX);
1530 /* Add peer_ni to global peer table hash, if necessary. */
1531 if (list_empty(&lpni->lpni_hashlist)) {
1532 int hash = lnet_nid2peerhash(&lpni->lpni_nid);
1534 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1535 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1536 ptable->pt_version++;
1537 lnet_peer_ni_addref_locked(lpni);
1540 /* Detach the peer_ni from an existing peer, if necessary. */
1541 if (lpni->lpni_peer_net) {
1542 LASSERT(lpni->lpni_peer_net != lpn);
1543 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1544 lnet_peer_detach_peer_ni_locked(lpni);
1545 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1546 lpni->lpni_peer_net = NULL;
1549 /* Add peer_ni to peer_net */
1550 lpni->lpni_peer_net = lpn;
1551 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1552 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1554 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1555 lnet_update_peer_net_healthv(lpni);
1556 lnet_peer_net_addref_locked(lpn);
1558 /* Add peer_net to peer */
1559 if (!lpn->lpn_peer) {
1562 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1563 list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1565 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1566 lnet_peer_addref_locked(lp);
1569 /* Add peer to global peer list, if necessary */
1570 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1571 if (list_empty(&lp->lp_peer_list)) {
1572 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1577 /* Update peer state */
1578 spin_lock(&lp->lp_lock);
1579 if (flags & LNET_PEER_CONFIGURED) {
1580 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1581 lp->lp_state |= LNET_PEER_CONFIGURED;
1583 if (flags & LNET_PEER_MULTI_RAIL) {
1584 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1585 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1586 lnet_peer_clr_non_mr_pref_nids(lp);
1589 spin_unlock(&lp->lp_lock);
1595 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1597 CERROR("Failed to apply UDSPs on lpn %s\n",
1598 libcfs_net2str(lpn->lpn_net_id));
1600 rc = lnet_udsp_apply_policies_on_lpni(lpni);
1602 CERROR("Failed to apply UDSPs on lpni %s\n",
1603 libcfs_nidstr(&lpni->lpni_nid));
1605 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1606 libcfs_nidstr(&lp->lp_primary_nid),
1607 libcfs_nidstr(&lpni->lpni_nid), flags);
1608 lnet_peer_ni_decref_locked(lpni);
1609 lnet_net_unlock(LNET_LOCK_EX);
1615 * Create a new peer, with nid as its primary nid.
1617 * Call with the lnet_api_mutex held.
1620 lnet_peer_add(lnet_nid_t nid4, unsigned int flags)
1622 struct lnet_nid nid;
1623 struct lnet_peer *lp;
1624 struct lnet_peer_net *lpn;
1625 struct lnet_peer_ni *lpni;
1628 LASSERT(nid4 != LNET_NID_ANY);
1631 * No need for the lnet_net_lock here, because the
1632 * lnet_api_mutex is held.
1634 lpni = lnet_find_peer_ni_locked(nid4);
1636 /* A peer with this NID already exists. */
1637 lp = lpni->lpni_peer_net->lpn_peer;
1638 lnet_peer_ni_decref_locked(lpni);
1640 * This is an error if the peer was configured and the
1641 * primary NID differs or an attempt is made to change
1642 * the Multi-Rail flag. Otherwise the assumption is
1643 * that an existing peer is being modified.
1645 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1646 if (lnet_nid_to_nid4(&lp->lp_primary_nid) != nid4)
1648 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1651 } else if (!(flags & LNET_PEER_CONFIGURED)) {
1652 if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid4) {
1657 /* Delete and recreate as a configured peer. */
1658 rc = lnet_peer_del(lp);
1663 /* Create peer, peer_net, and peer_ni. */
1665 lnet_nid4_to_nid(nid4, &nid);
1666 lp = lnet_peer_alloc(&nid);
1669 lpn = lnet_peer_net_alloc(LNET_NID_NET(&nid));
1672 lpni = lnet_peer_ni_alloc(&nid);
1676 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1679 LIBCFS_FREE(lpn, sizeof(*lpn));
1681 LIBCFS_FREE(lp, sizeof(*lp));
1683 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1684 libcfs_nid2str(nid4), flags, rc);
1689 * Add a NID to a peer. Call with ln_api_mutex held.
1692 * -EPERM: Non-DLC addition to a DLC-configured peer.
1693 * -EEXIST: The NID was configured by DLC for a different peer.
1694 * -ENOMEM: Out of memory.
1695 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1696 * non-multi-rail peer.
1699 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
1701 struct lnet_peer_net *lpn;
1702 struct lnet_peer_ni *lpni;
1703 struct lnet_nid nid;
1707 LASSERT(nid4 != LNET_NID_ANY);
1709 lnet_nid4_to_nid(nid4, &nid);
1711 /* A configured peer can only be updated through configuration. */
1712 if (!(flags & LNET_PEER_CONFIGURED)) {
1713 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1720 * The MULTI_RAIL flag can be set but not cleared, because
1721 * that would leave the peer struct in an invalid state.
1723 if (flags & LNET_PEER_MULTI_RAIL) {
1724 spin_lock(&lp->lp_lock);
1725 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1726 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1727 lnet_peer_clr_non_mr_pref_nids(lp);
1729 spin_unlock(&lp->lp_lock);
1730 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1735 lpni = lnet_find_peer_ni_locked(nid4);
1738 * A peer_ni already exists. This is only a problem if
1739 * it is not connected to this peer and was configured
1742 if (lpni->lpni_peer_net->lpn_peer == lp)
1744 if (lnet_peer_ni_is_configured(lpni)) {
1748 /* If this is the primary NID, destroy the peer. */
1749 if (lnet_peer_ni_is_primary(lpni)) {
1750 struct lnet_peer *rtr_lp =
1751 lpni->lpni_peer_net->lpn_peer;
1752 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1754 * if we're trying to delete a router it means
1755 * we're moving this peer NI to a new peer so must
1756 * transfer router properties to the new peer
1758 if (rtr_refcount > 0) {
1759 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1760 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1762 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1763 lnet_peer_ni_decref_locked(lpni);
1764 lpni = lnet_peer_ni_alloc(&nid);
1771 lpni = lnet_peer_ni_alloc(&nid);
1779 * Get the peer_net. Check that we're not adding a second
1780 * peer_ni on a peer_net of a non-multi-rail peer.
1782 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid4));
1784 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid4));
1789 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1794 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1797 lnet_peer_ni_decref_locked(lpni);
1799 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1800 libcfs_nidstr(&lp->lp_primary_nid), libcfs_nid2str(nid4),
1806 * Update the primary NID of a peer, if possible.
1808 * Call with the lnet_api_mutex held.
1811 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid,
1814 struct lnet_nid old = lp->lp_primary_nid;
1817 if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid)
1820 lnet_nid4_to_nid(nid, &lp->lp_primary_nid);
1822 rc = lnet_peer_add_nid(lp, nid, flags);
1824 lp->lp_primary_nid = old;
1828 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1829 libcfs_nidstr(&old), libcfs_nid2str(nid), rc);
1835 * lpni creation initiated due to traffic either sending or receiving.
1838 lnet_peer_ni_traffic_add(struct lnet_nid *nid, struct lnet_nid *pref)
1840 struct lnet_peer *lp;
1841 struct lnet_peer_net *lpn;
1842 struct lnet_peer_ni *lpni;
1846 if (LNET_NID_IS_ANY(nid)) {
1851 /* lnet_net_lock is not needed here because ln_api_lock is held */
1852 lpni = lnet_peer_ni_find_locked(nid);
1855 * We must have raced with another thread. Since we
1856 * know next to nothing about a peer_ni created by
1857 * traffic, we just assume everything is ok and
1860 lnet_peer_ni_decref_locked(lpni);
1864 /* Create peer, peer_net, and peer_ni. */
1866 lp = lnet_peer_alloc(nid);
1869 lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1872 lpni = lnet_peer_ni_alloc(nid);
1875 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1877 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1880 LIBCFS_FREE(lpn, sizeof(*lpn));
1882 LIBCFS_FREE(lp, sizeof(*lp));
1884 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(nid), rc);
1889 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1891 * This API handles the following combinations:
1892 * Create a peer with its primary NI if only the prim_nid is provided
1893 * Add a NID to a peer identified by the prim_nid. The peer identified
1894 * by the prim_nid must already exist.
1895 * The peer being created may be non-MR.
1897 * The caller must hold ln_api_mutex. This prevents the peer from
1898 * being created/modified/deleted by a different thread.
1901 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr, bool temp)
1903 struct lnet_peer *lp = NULL;
1904 struct lnet_peer_ni *lpni;
1905 unsigned int flags = 0;
1907 /* The prim_nid must always be specified */
1908 if (prim_nid == LNET_NID_ANY)
1912 flags = LNET_PEER_CONFIGURED;
1915 flags |= LNET_PEER_MULTI_RAIL;
1918 * If nid isn't specified, we must create a new peer with
1919 * prim_nid as its primary nid.
1921 if (nid == LNET_NID_ANY)
1922 return lnet_peer_add(prim_nid, flags);
1924 /* Look up the prim_nid, which must exist. */
1925 lpni = lnet_find_peer_ni_locked(prim_nid);
1928 lnet_peer_ni_decref_locked(lpni);
1929 lp = lpni->lpni_peer_net->lpn_peer;
1931 /* Peer must have been configured. */
1932 if (!temp && !(lp->lp_state & LNET_PEER_CONFIGURED)) {
1933 CDEBUG(D_NET, "peer %s was not configured\n",
1934 libcfs_nid2str(prim_nid));
1938 /* Primary NID must match */
1939 if (lnet_nid_to_nid4(&lp->lp_primary_nid) != prim_nid) {
1940 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1941 libcfs_nid2str(prim_nid),
1942 libcfs_nidstr(&lp->lp_primary_nid));
1946 /* Multi-Rail flag must match. */
1947 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1948 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1949 libcfs_nid2str(prim_nid));
1953 return lnet_peer_add_nid(lp, nid, flags);
1957 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1959 * This API handles the following combinations:
1960 * Delete a NI from a peer if both prim_nid and nid are provided.
1961 * Delete a peer if only prim_nid is provided.
1962 * Delete a peer if its primary nid is provided.
1964 * The caller must hold ln_api_mutex. This prevents the peer from
1965 * being modified/deleted by a different thread.
1968 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1970 struct lnet_peer *lp;
1971 struct lnet_peer_ni *lpni;
1974 if (prim_nid == LNET_NID_ANY)
1977 lpni = lnet_find_peer_ni_locked(prim_nid);
1980 lnet_peer_ni_decref_locked(lpni);
1981 lp = lpni->lpni_peer_net->lpn_peer;
1983 if (prim_nid != lnet_nid_to_nid4(&lp->lp_primary_nid)) {
1984 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1985 libcfs_nid2str(prim_nid),
1986 libcfs_nidstr(&lp->lp_primary_nid));
1990 lnet_net_lock(LNET_LOCK_EX);
1991 if (lp->lp_rtr_refcount > 0) {
1992 lnet_net_unlock(LNET_LOCK_EX);
1993 CERROR("%s is a router. Can not be deleted\n",
1994 libcfs_nid2str(prim_nid));
1997 lnet_net_unlock(LNET_LOCK_EX);
1999 if (nid == LNET_NID_ANY || nid == lnet_nid_to_nid4(&lp->lp_primary_nid))
2000 return lnet_peer_del(lp);
2002 flags = LNET_PEER_CONFIGURED;
2003 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2004 flags |= LNET_PEER_MULTI_RAIL;
2006 return lnet_peer_del_nid(lp, nid, flags);
2010 lnet_destroy_peer_ni_locked(struct kref *ref)
2012 struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
2014 struct lnet_peer_table *ptable;
2015 struct lnet_peer_net *lpn;
2017 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
2019 LASSERT(kref_read(&lpni->lpni_kref) == 0);
2020 LASSERT(list_empty(&lpni->lpni_txq));
2021 LASSERT(lpni->lpni_txqnob == 0);
2022 LASSERT(list_empty(&lpni->lpni_peer_nis));
2023 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
2025 lpn = lpni->lpni_peer_net;
2026 lpni->lpni_peer_net = NULL;
2027 lpni->lpni_net = NULL;
2029 if (!list_empty(&lpni->lpni_hashlist)) {
2030 /* remove the peer ni from the zombie list */
2031 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
2032 spin_lock(&ptable->pt_zombie_lock);
2033 list_del_init(&lpni->lpni_hashlist);
2034 ptable->pt_zombies--;
2035 spin_unlock(&ptable->pt_zombie_lock);
2038 if (lpni->lpni_pref_nnids > 1) {
2039 struct lnet_nid_list *ne, *tmp;
2041 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
2043 list_del_init(&ne->nl_list);
2044 LIBCFS_FREE(ne, sizeof(*ne));
2047 LIBCFS_FREE(lpni, sizeof(*lpni));
2050 lnet_peer_net_decref_locked(lpn);
2053 struct lnet_peer_ni *
2054 lnet_nid2peerni_ex(struct lnet_nid *nid, int cpt)
2056 struct lnet_peer_ni *lpni = NULL;
2059 if (the_lnet.ln_state != LNET_STATE_RUNNING)
2060 return ERR_PTR(-ESHUTDOWN);
2063 * find if a peer_ni already exists.
2064 * If so then just return that.
2066 lpni = lnet_peer_ni_find_locked(nid);
2070 lnet_net_unlock(cpt);
2072 rc = lnet_peer_ni_traffic_add(nid, NULL);
2075 goto out_net_relock;
2078 lpni = lnet_peer_ni_find_locked(nid);
2088 * Get a peer_ni for the given nid, create it if necessary. Takes a
2089 * hold on the peer_ni.
2091 struct lnet_peer_ni *
2092 lnet_peerni_by_nid_locked(struct lnet_nid *nid,
2093 struct lnet_nid *pref, int cpt)
2095 struct lnet_peer_ni *lpni = NULL;
2098 if (the_lnet.ln_state != LNET_STATE_RUNNING)
2099 return ERR_PTR(-ESHUTDOWN);
2102 * find if a peer_ni already exists.
2103 * If so then just return that.
2105 lpni = lnet_peer_ni_find_locked(nid);
2111 * use the lnet_api_mutex to serialize the creation of the peer_ni
2112 * and the creation/deletion of the local ni/net. When a local ni is
2113 * created, if there exists a set of peer_nis on that network,
2114 * they need to be traversed and updated. When a local NI is
2115 * deleted, which could result in a network being deleted, then
2116 * all peer nis on that network need to be removed as well.
2118 * Creation through traffic should also be serialized with
2119 * creation through DLC.
2121 lnet_net_unlock(cpt);
2122 mutex_lock(&the_lnet.ln_api_mutex);
2124 * Shutdown is only set under the ln_api_lock, so a single
2125 * check here is sufficent.
2127 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2128 lpni = ERR_PTR(-ESHUTDOWN);
2129 goto out_mutex_unlock;
2132 rc = lnet_peer_ni_traffic_add(nid, pref);
2135 goto out_mutex_unlock;
2138 lpni = lnet_peer_ni_find_locked(nid);
2142 mutex_unlock(&the_lnet.ln_api_mutex);
2145 /* Lock has been dropped, check again for shutdown. */
2146 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2148 lnet_peer_ni_decref_locked(lpni);
2149 lpni = ERR_PTR(-ESHUTDOWN);
2155 struct lnet_peer_ni *
2156 lnet_nid2peerni_locked(lnet_nid_t nid4, lnet_nid_t pref4, int cpt)
2158 struct lnet_nid nid, pref;
2160 lnet_nid4_to_nid(nid4, &nid);
2161 lnet_nid4_to_nid(pref4, &pref);
2162 if (pref4 == LNET_NID_ANY)
2163 return lnet_peerni_by_nid_locked(&nid, NULL, cpt);
2165 return lnet_peerni_by_nid_locked(&nid, &pref, cpt);
2169 lnet_peer_gw_discovery(struct lnet_peer *lp)
2173 spin_lock(&lp->lp_lock);
2174 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2176 spin_unlock(&lp->lp_lock);
2182 lnet_peer_is_uptodate(struct lnet_peer *lp)
2186 spin_lock(&lp->lp_lock);
2187 rc = lnet_peer_is_uptodate_locked(lp);
2188 spin_unlock(&lp->lp_lock);
2193 * Is a peer uptodate from the point of view of discovery?
2195 * If it is currently being processed, obviously not.
2196 * A forced Ping or Push is also handled by the discovery thread.
2198 * Otherwise look at whether the peer needs rediscovering.
2201 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2202 __must_hold(&lp->lp_lock)
2206 if (lp->lp_state & (LNET_PEER_DISCOVERING |
2207 LNET_PEER_FORCE_PING |
2208 LNET_PEER_FORCE_PUSH)) {
2210 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2212 } else if (lnet_peer_needs_push(lp)) {
2214 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2215 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2226 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2228 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2230 /* The discovery thread holds net_lock/EX and lp_lock when it splices
2231 * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2232 * when adding to the list and queuing the peer to ensure that we do not
2233 * strand any messages on the lp_dc_pendq. This scheme ensures the
2234 * message will be resent even if the peer is already being discovered.
2235 * Therefore we needn't check the return value of
2236 * lnet_peer_queue_for_discovery(lp).
2238 lnet_net_lock(LNET_LOCK_EX);
2239 spin_lock(&lp->lp_lock);
2240 list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2241 spin_unlock(&lp->lp_lock);
2242 lnet_peer_queue_for_discovery(lp);
2243 lnet_net_unlock(LNET_LOCK_EX);
2247 * Queue a peer for the attention of the discovery thread. Call with
2248 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2249 * -EALREADY if the peer was already queued.
2251 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2255 spin_lock(&lp->lp_lock);
2256 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2257 lp->lp_state |= LNET_PEER_DISCOVERING;
2258 spin_unlock(&lp->lp_lock);
2259 if (list_empty(&lp->lp_dc_list)) {
2260 lnet_peer_addref_locked(lp);
2261 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2262 wake_up(&the_lnet.ln_dc_waitq);
2268 CDEBUG(D_NET, "Queue peer %s: %d\n",
2269 libcfs_nidstr(&lp->lp_primary_nid), rc);
2275 * Discovery of a peer is complete. Wake all waiters on the peer.
2276 * Call with lnet_net_lock/EX held.
2278 static void lnet_peer_discovery_complete(struct lnet_peer *lp, int dc_error)
2280 struct lnet_msg *msg, *tmp;
2282 LIST_HEAD(pending_msgs);
2284 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2285 libcfs_nidstr(&lp->lp_primary_nid));
2287 list_del_init(&lp->lp_dc_list);
2288 spin_lock(&lp->lp_lock);
2290 lp->lp_dc_error = dc_error;
2291 lp->lp_state &= ~LNET_PEER_DISCOVERING;
2292 lp->lp_state |= LNET_PEER_REDISCOVER;
2294 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2295 spin_unlock(&lp->lp_lock);
2296 wake_up(&lp->lp_dc_waitq);
2298 if (lp->lp_rtr_refcount > 0)
2299 lnet_router_discovery_complete(lp);
2301 lnet_net_unlock(LNET_LOCK_EX);
2303 /* iterate through all pending messages and send them again */
2304 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2305 list_del_init(&msg->msg_list);
2307 lnet_finalize(msg, dc_error);
2311 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2312 lnet_msgtyp2str(msg->msg_type),
2313 libcfs_id2str(msg->msg_target));
2314 rc = lnet_send(msg->msg_src_nid_param, msg,
2315 msg->msg_rtr_nid_param);
2317 CNETERR("Error sending %s to %s: %d\n",
2318 lnet_msgtyp2str(msg->msg_type),
2319 libcfs_id2str(msg->msg_target), rc);
2320 lnet_finalize(msg, rc);
2323 lnet_net_lock(LNET_LOCK_EX);
2324 lnet_peer_decref_locked(lp);
2328 * Handle inbound push.
2329 * Like any event handler, called with lnet_res_lock/CPT held.
2331 void lnet_peer_push_event(struct lnet_event *ev)
2333 struct lnet_ping_buffer *pbuf;
2334 struct lnet_peer *lp;
2336 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2338 /* lnet_find_peer() adds a refcount */
2339 lp = lnet_find_peer(ev->source.nid);
2341 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2342 libcfs_nid2str(ev->initiator.nid),
2343 libcfs_nid2str(ev->source.nid));
2344 pbuf->pb_needs_post = true;
2348 /* Ensure peer state remains consistent while we modify it. */
2349 spin_lock(&lp->lp_lock);
2352 * If some kind of error happened the contents of the message
2353 * cannot be used. Clear the NIDS_UPTODATE and set the
2354 * FORCE_PING flag to trigger a ping.
2357 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2358 lp->lp_state |= LNET_PEER_FORCE_PING;
2359 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2361 libcfs_nidstr(&lp->lp_primary_nid),
2362 libcfs_nid2str(ev->source.nid));
2367 * A push with invalid or corrupted info. Clear the UPTODATE
2368 * flag to trigger a ping.
2370 if (lnet_ping_info_validate(&pbuf->pb_info)) {
2371 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2372 lp->lp_state |= LNET_PEER_FORCE_PING;
2373 CDEBUG(D_NET, "Corrupted Push from %s\n",
2374 libcfs_nidstr(&lp->lp_primary_nid));
2379 * Make sure we'll allocate the correct size ping buffer when
2382 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2383 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2386 * A non-Multi-Rail peer is not supposed to be capable of
2389 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2390 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2391 libcfs_nidstr(&lp->lp_primary_nid));
2396 * The peer may have discovery disabled at its end. Set
2397 * NO_DISCOVERY as appropriate.
2399 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2400 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2401 libcfs_nidstr(&lp->lp_primary_nid));
2403 * Mark the peer for deletion if we already know about it
2404 * and it's going from discovery set to no discovery set
2406 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2407 LNET_PEER_DISCOVERING)) &&
2408 lp->lp_state & LNET_PEER_DISCOVERED) {
2409 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2410 libcfs_nidstr(&lp->lp_primary_nid),
2412 lp->lp_state |= LNET_PEER_MARK_DELETION;
2414 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2415 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2416 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2417 libcfs_nidstr(&lp->lp_primary_nid));
2418 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2422 * Update the MULTI_RAIL flag based on the push. If the peer
2423 * was configured with DLC then the setting should match what
2425 * NB: We verified above that the MR feature bit is set in pi_features
2427 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2428 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2429 libcfs_nidstr(&lp->lp_primary_nid), lp);
2430 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2431 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2432 libcfs_nidstr(&lp->lp_primary_nid));
2433 } else if (lnet_peer_discovery_disabled) {
2434 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2435 libcfs_nidstr(&lp->lp_primary_nid), lp);
2436 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2437 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2438 libcfs_nidstr(&lp->lp_primary_nid), lp);
2440 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2441 libcfs_nidstr(&lp->lp_primary_nid), lp);
2442 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2443 lnet_peer_clr_non_mr_pref_nids(lp);
2447 * Check for truncation of the Put message. Clear the
2448 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2449 * and tell discovery to allocate a bigger buffer.
2451 if (ev->mlength < ev->rlength) {
2452 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2453 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2454 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2455 lp->lp_state |= LNET_PEER_FORCE_PING;
2456 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2457 libcfs_nidstr(&lp->lp_primary_nid),
2458 pbuf->pb_info.pi_nnis);
2462 /* always assume new data */
2463 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2464 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2467 * If there is data present that hasn't been processed yet,
2468 * we'll replace it if the Put contained newer data and it
2469 * fits. We're racing with a Ping or earlier Push in this
2472 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2473 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2474 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2475 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2476 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2477 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2478 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2479 libcfs_nidstr(&lp->lp_primary_nid),
2480 LNET_PING_BUFFER_SEQNO(pbuf),
2481 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2487 * Allocate a buffer to copy the data. On a failure we drop
2488 * the Push and set FORCE_PING to force the discovery
2489 * thread to fix the problem by pinging the peer.
2491 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2493 lp->lp_state |= LNET_PEER_FORCE_PING;
2494 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2495 libcfs_nidstr(&lp->lp_primary_nid),
2496 LNET_PING_BUFFER_SEQNO(pbuf));
2501 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2502 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2503 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2504 CDEBUG(D_NET, "Received Push %s %u\n",
2505 libcfs_nidstr(&lp->lp_primary_nid),
2506 LNET_PING_BUFFER_SEQNO(pbuf));
2509 /* We've processed this buffer. It can be reposted */
2510 pbuf->pb_needs_post = true;
2513 * Queue the peer for discovery if not done, force it on the request
2514 * queue and wake the discovery thread if the peer was already queued,
2515 * because its status changed.
2517 spin_unlock(&lp->lp_lock);
2518 lnet_net_lock(LNET_LOCK_EX);
2519 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2520 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2521 wake_up(&the_lnet.ln_dc_waitq);
2523 /* Drop refcount from lookup */
2524 lnet_peer_decref_locked(lp);
2525 lnet_net_unlock(LNET_LOCK_EX);
2529 * Clear the discovery error state, unless we're already discovering
2530 * this peer, in which case the error is current.
2532 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2534 spin_lock(&lp->lp_lock);
2535 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2536 lp->lp_dc_error = 0;
2537 spin_unlock(&lp->lp_lock);
2541 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2542 * dropped/retaken within this function. An lnet_peer_ni is passed in
2543 * because discovery could tear down an lnet_peer.
2546 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2549 struct lnet_peer *lp;
2554 lnet_net_unlock(cpt);
2555 lnet_net_lock(LNET_LOCK_EX);
2556 lp = lpni->lpni_peer_net->lpn_peer;
2557 lnet_peer_clear_discovery_error(lp);
2560 * We're willing to be interrupted. The lpni can become a
2561 * zombie if we race with DLC, so we must check for that.
2564 /* Keep lp alive when the lnet_net_lock is unlocked */
2565 lnet_peer_addref_locked(lp);
2566 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2567 if (signal_pending(current))
2569 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2572 * Don't repeat discovery if discovery is disabled. This is
2573 * done to ensure we can use discovery as a standard ping as
2574 * well for backwards compatibility with routers which do not
2575 * have discovery or have discovery disabled
2577 if (lnet_is_discovery_disabled(lp) && count > 0)
2579 if (lp->lp_dc_error)
2581 if (lnet_peer_is_uptodate(lp))
2583 lnet_peer_queue_for_discovery(lp);
2585 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2588 * If caller requested a non-blocking operation then
2589 * return immediately. Once discovery is complete any
2590 * pending messages that were stopped due to discovery
2591 * will be transmitted.
2596 lnet_net_unlock(LNET_LOCK_EX);
2598 finish_wait(&lp->lp_dc_waitq, &wait);
2599 lnet_net_lock(LNET_LOCK_EX);
2600 lnet_peer_decref_locked(lp);
2601 /* Peer may have changed */
2602 lp = lpni->lpni_peer_net->lpn_peer;
2604 finish_wait(&lp->lp_dc_waitq, &wait);
2606 lnet_net_unlock(LNET_LOCK_EX);
2608 lnet_peer_decref_locked(lp);
2610 * The peer may have changed, so re-check and rediscover if that turns
2611 * out to have been the case. The reference count on lp ensured that
2612 * even if it was unlinked from lpni the memory could not be recycled.
2613 * Thus the check below is sufficient to determine whether the peer
2614 * changed. If the peer changed, then lp must not be dereferenced.
2616 if (lp != lpni->lpni_peer_net->lpn_peer)
2619 if (signal_pending(current))
2621 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2623 else if (lp->lp_dc_error)
2624 rc = lp->lp_dc_error;
2626 CDEBUG(D_NET, "non-blocking discovery\n");
2627 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2630 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2631 (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2632 libcfs_nidstr(&lpni->lpni_nid), rc,
2633 (!block) ? "pending discovery" : "discovery complete");
2638 /* Handle an incoming ack for a push. */
2640 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2642 struct lnet_ping_buffer *pbuf;
2644 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2645 spin_lock(&lp->lp_lock);
2646 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2647 lp->lp_push_error = ev->status;
2649 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2651 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2652 spin_unlock(&lp->lp_lock);
2654 CDEBUG(D_NET, "peer %s ev->status %d\n",
2655 libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2658 /* Handle a Reply message. This is the reply to a Ping message. */
2660 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2662 struct lnet_ping_buffer *pbuf;
2665 spin_lock(&lp->lp_lock);
2667 lnet_nid4_to_nid(ev->target.nid, &lp->lp_disc_src_nid);
2668 lnet_nid4_to_nid(ev->source.nid, &lp->lp_disc_dst_nid);
2671 * If some kind of error happened the contents of message
2672 * cannot be used. Set PING_FAILED to trigger a retry.
2675 lp->lp_state |= LNET_PEER_PING_FAILED;
2676 lp->lp_ping_error = ev->status;
2677 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2679 libcfs_nidstr(&lp->lp_primary_nid),
2680 libcfs_nid2str(ev->source.nid));
2684 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2685 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2686 lnet_swap_pinginfo(pbuf);
2689 * A reply with invalid or corrupted info. Set PING_FAILED to
2692 rc = lnet_ping_info_validate(&pbuf->pb_info);
2694 lp->lp_state |= LNET_PEER_PING_FAILED;
2695 lp->lp_ping_error = 0;
2696 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2697 libcfs_nidstr(&lp->lp_primary_nid), rc);
2702 * The peer may have discovery disabled at its end. Set
2703 * NO_DISCOVERY as appropriate.
2705 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) ||
2706 lnet_peer_discovery_disabled) {
2707 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2708 libcfs_nidstr(&lp->lp_primary_nid));
2710 /* Detect whether this peer has toggled discovery from on to
2711 * off and whether we can delete and re-create the peer. Peers
2712 * that were manually configured cannot be deleted by discovery.
2713 * We need to delete this peer and re-create it if the peer was
2714 * not configured manually, is currently considered DD capable,
2716 * 1. We've already discovered the peer (the peer has toggled
2717 * the discovery feature from on to off), or
2718 * 2. The peer is considered MR, but it was not user configured
2719 * (this was a "temporary" peer created via the kernel APIs
2720 * that we're discovering for the first time)
2722 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2723 LNET_PEER_NO_DISCOVERY)) &&
2724 (lp->lp_state & (LNET_PEER_DISCOVERED |
2725 LNET_PEER_MULTI_RAIL))) {
2726 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2727 libcfs_nidstr(&lp->lp_primary_nid),
2729 lp->lp_state |= LNET_PEER_MARK_DELETION;
2731 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2733 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2734 libcfs_nidstr(&lp->lp_primary_nid));
2735 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2739 * Update the MULTI_RAIL flag based on the reply. If the peer
2740 * was configured with DLC then the setting should match what
2743 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2744 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2745 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2746 libcfs_nidstr(&lp->lp_primary_nid), lp);
2747 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2748 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2749 libcfs_nidstr(&lp->lp_primary_nid));
2750 } else if (lnet_peer_discovery_disabled) {
2752 "peer %s(%p) not MR: DD disabled locally\n",
2753 libcfs_nidstr(&lp->lp_primary_nid), lp);
2754 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2756 "peer %s(%p) not MR: DD disabled remotely\n",
2757 libcfs_nidstr(&lp->lp_primary_nid), lp);
2759 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2760 libcfs_nidstr(&lp->lp_primary_nid), lp);
2761 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2762 lnet_peer_clr_non_mr_pref_nids(lp);
2764 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2765 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2766 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2767 libcfs_nidstr(&lp->lp_primary_nid));
2769 CERROR("Multi-Rail state vanished from %s\n",
2770 libcfs_nidstr(&lp->lp_primary_nid));
2771 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2776 * Make sure we'll allocate the correct size ping buffer when
2779 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2780 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2783 * Check for truncation of the Reply. Clear PING_SENT and set
2784 * PING_FAILED to trigger a retry.
2786 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2787 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2788 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2789 lp->lp_state |= LNET_PEER_PING_FAILED;
2790 lp->lp_ping_error = 0;
2791 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2792 libcfs_nidstr(&lp->lp_primary_nid),
2793 pbuf->pb_info.pi_nnis);
2798 * Check the sequence numbers in the reply. These are only
2799 * available if the reply came from a Multi-Rail peer.
2801 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2802 pbuf->pb_info.pi_nnis > 1 &&
2803 lnet_nid_to_nid4(&lp->lp_primary_nid) ==
2804 pbuf->pb_info.pi_ni[1].ns_nid) {
2805 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2806 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2807 libcfs_nidstr(&lp->lp_primary_nid),
2808 LNET_PING_BUFFER_SEQNO(pbuf),
2811 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2814 /* We're happy with the state of the data in the buffer. */
2815 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2816 libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2818 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2819 lnet_ping_buffer_decref(lp->lp_data);
2821 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2822 lnet_ping_buffer_addref(pbuf);
2825 lp->lp_state &= ~LNET_PEER_PING_SENT;
2826 spin_unlock(&lp->lp_lock);
2828 lnet_net_lock(LNET_LOCK_EX);
2830 * If this peer is a gateway, call the routing callback to
2831 * handle the ping reply
2833 if (lp->lp_rtr_refcount > 0)
2834 lnet_router_discovery_ping_reply(lp);
2835 lnet_net_unlock(LNET_LOCK_EX);
2839 * Send event handling. Only matters for error cases, where we clean
2840 * up state on the peer and peer_ni that would otherwise be updated in
2841 * the REPLY event handler for a successful Ping, and the ACK event
2842 * handler for a successful Push.
2845 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2852 spin_lock(&lp->lp_lock);
2853 if (ev->msg_type == LNET_MSG_GET) {
2854 lp->lp_state &= ~LNET_PEER_PING_SENT;
2855 lp->lp_state |= LNET_PEER_PING_FAILED;
2856 lp->lp_ping_error = ev->status;
2857 } else { /* ev->msg_type == LNET_MSG_PUT */
2858 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2859 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2860 lp->lp_push_error = ev->status;
2862 spin_unlock(&lp->lp_lock);
2863 rc = LNET_REDISCOVER_PEER;
2865 CDEBUG(D_NET, "%s Send to %s: %d\n",
2866 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2867 libcfs_nid2str(ev->target.nid), rc);
2872 * Unlink event handling. This event is only seen if a call to
2873 * LNetMDUnlink() caused the event to be unlinked. If this call was
2874 * made after the event was set up in LNetGet() or LNetPut() then we
2875 * assume the Ping or Push timed out.
2878 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2880 spin_lock(&lp->lp_lock);
2881 /* We've passed through LNetGet() */
2882 if (lp->lp_state & LNET_PEER_PING_SENT) {
2883 lp->lp_state &= ~LNET_PEER_PING_SENT;
2884 lp->lp_state |= LNET_PEER_PING_FAILED;
2885 lp->lp_ping_error = -ETIMEDOUT;
2886 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2887 libcfs_nidstr(&lp->lp_primary_nid));
2889 /* We've passed through LNetPut() */
2890 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2891 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2892 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2893 lp->lp_push_error = -ETIMEDOUT;
2894 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2895 libcfs_nidstr(&lp->lp_primary_nid));
2897 spin_unlock(&lp->lp_lock);
2901 * Event handler for the discovery EQ.
2903 * Called with lnet_res_lock(cpt) held. The cpt is the
2904 * lnet_cpt_of_cookie() of the md handle cookie.
2906 static void lnet_discovery_event_handler(struct lnet_event *event)
2908 struct lnet_peer *lp = event->md_user_ptr;
2909 struct lnet_ping_buffer *pbuf;
2912 /* discovery needs to take another look */
2913 rc = LNET_REDISCOVER_PEER;
2915 CDEBUG(D_NET, "Received event: %d\n", event->type);
2917 switch (event->type) {
2918 case LNET_EVENT_ACK:
2919 lnet_discovery_event_ack(lp, event);
2921 case LNET_EVENT_REPLY:
2922 lnet_discovery_event_reply(lp, event);
2924 case LNET_EVENT_SEND:
2925 /* Only send failure triggers a retry. */
2926 rc = lnet_discovery_event_send(lp, event);
2928 case LNET_EVENT_UNLINK:
2929 /* LNetMDUnlink() was called */
2930 lnet_discovery_event_unlink(lp, event);
2933 /* Invalid events. */
2936 lnet_net_lock(LNET_LOCK_EX);
2937 if (event->unlinked) {
2938 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2939 lnet_ping_buffer_decref(pbuf);
2940 lnet_peer_decref_locked(lp);
2943 /* put peer back at end of request queue, if discovery not already
2945 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2946 lnet_peer_queue_for_discovery(lp)) {
2947 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2948 wake_up(&the_lnet.ln_dc_waitq);
2950 lnet_net_unlock(LNET_LOCK_EX);
2954 * Build a peer from incoming data.
2956 * The NIDs in the incoming data are supposed to be structured as follows:
2959 * - other NIDs in same net
2960 * - NIDs in second net
2961 * - NIDs in third net
2963 * This due to the way the list of NIDs in the data is created.
2965 * Note that this function will mark the peer uptodate unless an
2966 * ENOMEM is encontered. All other errors are due to a conflict
2967 * between the DLC configuration and what discovery sees. We treat DLC
2968 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2969 * peer from becoming stuck in discovery.
2971 static int lnet_peer_merge_data(struct lnet_peer *lp,
2972 struct lnet_ping_buffer *pbuf)
2974 struct lnet_peer_net *lpn;
2975 struct lnet_peer_ni *lpni;
2976 lnet_nid_t *curnis = NULL;
2977 struct lnet_ni_status *addnis = NULL;
2978 lnet_nid_t *delnis = NULL;
2988 flags = LNET_PEER_DISCOVERED;
2989 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2990 flags |= LNET_PEER_MULTI_RAIL;
2993 * Cache the routing feature for the peer; whether it is enabled
2994 * for disabled as reported by the remote peer.
2996 spin_lock(&lp->lp_lock);
2997 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2998 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
3000 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
3001 spin_unlock(&lp->lp_lock);
3003 nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
3004 CFS_ALLOC_PTR_ARRAY(curnis, nnis);
3005 CFS_ALLOC_PTR_ARRAY(addnis, nnis);
3006 CFS_ALLOC_PTR_ARRAY(delnis, nnis);
3007 if (!curnis || !addnis || !delnis) {
3015 /* Construct the list of NIDs present in peer. */
3017 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
3018 curnis[ncurnis++] = lnet_nid_to_nid4(&lpni->lpni_nid);
3021 * Check for NIDs in pbuf not present in curnis[].
3022 * The loop starts at 1 to skip the loopback NID.
3024 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
3025 for (j = 0; j < ncurnis; j++)
3026 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
3029 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
3032 * Check for NIDs in curnis[] not present in pbuf.
3033 * The nested loop starts at 1 to skip the loopback NID.
3035 * But never add the loopback NID to delnis[]: if it is
3036 * present in curnis[] then this peer is for this node.
3038 for (i = 0; i < ncurnis; i++) {
3039 if (curnis[i] == LNET_NID_LO_0)
3041 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
3042 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
3044 * update the information we cache for the
3045 * peer with the latest information we
3048 lpni = lnet_find_peer_ni_locked(curnis[i]);
3050 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
3051 lnet_peer_ni_decref_locked(lpni);
3056 if (j == pbuf->pb_info.pi_nnis)
3057 delnis[ndelnis++] = curnis[i];
3061 * If we get here and the discovery is disabled then we don't want
3062 * to add or delete any NIs. We just updated the ones we have some
3063 * information on, and call it a day
3066 if (lnet_is_discovery_disabled(lp))
3069 for (i = 0; i < naddnis; i++) {
3070 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
3072 CERROR("Error adding NID %s to peer %s: %d\n",
3073 libcfs_nid2str(addnis[i].ns_nid),
3074 libcfs_nidstr(&lp->lp_primary_nid), rc);
3078 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
3080 lpni->lpni_ns_status = addnis[i].ns_status;
3081 lnet_peer_ni_decref_locked(lpni);
3085 for (i = 0; i < ndelnis; i++) {
3087 * for routers it's okay to delete the primary_nid because
3088 * the upper layers don't really rely on it. So if we're
3089 * being told that the router changed its primary_nid
3090 * then it's okay to delete it.
3092 if (lp->lp_rtr_refcount > 0)
3093 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3094 rc = lnet_peer_del_nid(lp, delnis[i], flags);
3096 CERROR("Error deleting NID %s from peer %s: %d\n",
3097 libcfs_nid2str(delnis[i]),
3098 libcfs_nidstr(&lp->lp_primary_nid), rc);
3104 /* The peer net for the primary NID should be the first entry in the
3105 * peer's lp_peer_nets list, and the peer NI for the primary NID should
3106 * be the first entry in its peer net's lpn_peer_nis list.
3108 lpni = lnet_find_peer_ni_locked(pbuf->pb_info.pi_ni[1].ns_nid);
3110 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3111 libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
3115 lnet_peer_ni_decref_locked(lpni);
3117 lpn = lpni->lpni_peer_net;
3118 if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3119 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3121 if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3122 list_move(&lpni->lpni_peer_nis,
3123 &lpni->lpni_peer_net->lpn_peer_nis);
3126 * Errors other than -ENOMEM are due to peers having been
3127 * configured with DLC. Ignore these because DLC overrides
3132 CFS_FREE_PTR_ARRAY(curnis, nnis);
3133 CFS_FREE_PTR_ARRAY(addnis, nnis);
3134 CFS_FREE_PTR_ARRAY(delnis, nnis);
3135 lnet_ping_buffer_decref(pbuf);
3136 CDEBUG(D_NET, "peer %s (%p): %d\n",
3137 libcfs_nidstr(&lp->lp_primary_nid), lp, rc);
3140 spin_lock(&lp->lp_lock);
3141 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
3142 lp->lp_state |= LNET_PEER_FORCE_PING;
3143 spin_unlock(&lp->lp_lock);
3149 * The data in pbuf says lp is its primary peer, but the data was
3150 * received by a different peer. Try to update lp with the data.
3153 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
3155 struct lnet_handle_md mdh;
3157 /* Queue lp for discovery, and force it on the request queue. */
3158 lnet_net_lock(LNET_LOCK_EX);
3159 if (lnet_peer_queue_for_discovery(lp))
3160 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3161 lnet_net_unlock(LNET_LOCK_EX);
3163 LNetInvalidateMDHandle(&mdh);
3166 * Decide whether we can move the peer to the DATA_PRESENT state.
3168 * We replace stale data for a multi-rail peer, repair PING_FAILED
3169 * status, and preempt FORCE_PING.
3171 * If after that we have DATA_PRESENT, we merge it into this peer.
3173 spin_lock(&lp->lp_lock);
3174 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3175 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3176 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3177 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3178 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3179 lnet_ping_buffer_decref(pbuf);
3184 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3185 lnet_ping_buffer_decref(lp->lp_data);
3187 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3189 if (lp->lp_state & LNET_PEER_PING_FAILED) {
3190 mdh = lp->lp_ping_mdh;
3191 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3192 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3193 lp->lp_ping_error = 0;
3195 if (lp->lp_state & LNET_PEER_FORCE_PING)
3196 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3197 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3198 spin_unlock(&lp->lp_lock);
3200 if (!LNetMDHandleIsInvalid(mdh))
3204 return lnet_peer_merge_data(lp, pbuf);
3206 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3210 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
3214 for (i = 0; i < pinfo->pi_nnis; i++) {
3215 if (pinfo->pi_ni[i].ns_nid == nid)
3222 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3223 * to the discovery queue a reference was taken that will prevent the peer from
3224 * actually being freed by this function. After this function exits the
3225 * discovery thread should call lnet_peer_discovery_complete() which will
3226 * drop that reference as well as wake any waiters that may also be holding a
3229 static int lnet_peer_deletion(struct lnet_peer *lp)
3230 __must_hold(&lp->lp_lock)
3232 struct list_head rlist;
3233 struct lnet_route *route, *tmp;
3234 int sensitivity = lp->lp_health_sensitivity;
3237 INIT_LIST_HEAD(&rlist);
3239 lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3240 LNET_PEER_FORCE_PUSH);
3241 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3242 libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3244 /* no-op if lnet_peer_del() has already been called on this peer */
3245 if (lp->lp_state & LNET_PEER_MARK_DELETED)
3248 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
3251 spin_unlock(&lp->lp_lock);
3253 mutex_lock(&the_lnet.ln_api_mutex);
3255 lnet_net_lock(LNET_LOCK_EX);
3256 /* remove the peer from the discovery work
3257 * queue if it's on there in preparation
3260 if (!list_empty(&lp->lp_dc_list))
3261 list_del_init(&lp->lp_dc_list);
3262 list_for_each_entry_safe(route, tmp,
3265 lnet_move_route(route, NULL, &rlist);
3266 lnet_net_unlock(LNET_LOCK_EX);
3268 /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3269 rc = lnet_peer_del(lp);
3271 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3272 libcfs_nidstr(&lp->lp_primary_nid), rc);
3274 list_for_each_entry_safe(route, tmp,
3276 /* re-add these routes */
3277 lnet_add_route(route->lr_net,
3282 LIBCFS_FREE(route, sizeof(*route));
3285 mutex_unlock(&the_lnet.ln_api_mutex);
3287 spin_lock(&lp->lp_lock);
3293 * Update a peer using the data received.
3295 static int lnet_peer_data_present(struct lnet_peer *lp)
3296 __must_hold(&lp->lp_lock)
3298 struct lnet_ping_buffer *pbuf;
3299 struct lnet_peer_ni *lpni;
3300 lnet_nid_t nid = LNET_NID_ANY;
3306 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3307 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3308 spin_unlock(&lp->lp_lock);
3311 * Modifications of peer structures are done while holding the
3312 * ln_api_mutex. A global lock is required because we may be
3313 * modifying multiple peer structures, and a mutex greatly
3314 * simplifies memory management.
3316 * The actual changes to the data structures must also protect
3317 * against concurrent lookups, for which the lnet_net_lock in
3318 * LNET_LOCK_EX mode is used.
3320 mutex_lock(&the_lnet.ln_api_mutex);
3321 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3327 * If this peer is not on the peer list then it is being torn
3328 * down, and our reference count may be all that is keeping it
3329 * alive. Don't do any work on it.
3331 if (list_empty(&lp->lp_peer_list))
3334 flags = LNET_PEER_DISCOVERED;
3335 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3336 flags |= LNET_PEER_MULTI_RAIL;
3339 * Check whether the primary NID in the message matches the
3340 * primary NID of the peer. If it does, update the peer, if
3341 * it it does not, check whether there is already a peer with
3342 * that primary NID. If no such peer exists, try to update
3343 * the primary NID of the current peer (allowed if it was
3344 * created due to message traffic) and complete the update.
3345 * If the peer did exist, hand off the data to it.
3347 * The peer for the loopback interface is a special case: this
3348 * is the peer for the local node, and we want to set its
3349 * primary NID to the correct value here. Moreover, this peer
3350 * can show up with only the loopback NID in the ping buffer.
3352 if (pbuf->pb_info.pi_nnis <= 1)
3354 nid = pbuf->pb_info.pi_ni[1].ns_nid;
3355 if (nid_is_lo0(&lp->lp_primary_nid)) {
3356 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3358 rc = lnet_peer_merge_data(lp, pbuf);
3360 * if the primary nid of the peer is present in the ping info returned
3361 * from the peer, but it's not the local primary peer we have
3362 * cached and discovery is disabled, then we don't want to update
3363 * our local peer info, by adding or removing NIDs, we just want
3364 * to update the status of the nids that we currently have
3365 * recorded in that peer.
3367 } else if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid ||
3368 (lnet_is_nid_in_ping_info(lnet_nid_to_nid4(&lp->lp_primary_nid),
3370 lnet_is_discovery_disabled(lp))) {
3371 rc = lnet_peer_merge_data(lp, pbuf);
3373 lpni = lnet_find_peer_ni_locked(nid);
3374 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3375 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3377 CERROR("Primary NID error %s versus %s: %d\n",
3378 libcfs_nidstr(&lp->lp_primary_nid),
3379 libcfs_nid2str(nid), rc);
3381 rc = lnet_peer_merge_data(lp, pbuf);
3384 lnet_peer_ni_decref_locked(lpni);
3386 struct lnet_peer *new_lp;
3387 new_lp = lpni->lpni_peer_net->lpn_peer;
3389 * if lp has discovery/MR enabled that means new_lp
3390 * should have discovery/MR enabled as well, since
3391 * it's the same peer, which we're about to merge
3393 spin_lock(&lp->lp_lock);
3394 spin_lock(&new_lp->lp_lock);
3395 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3396 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3397 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3398 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3399 /* If we're processing a ping reply then we may be
3400 * about to send a push to the peer that we ping'd.
3401 * Since the ping reply that we're processing was
3402 * received by lp, we need to set the discovery source
3403 * NID for new_lp to the NID stored in lp.
3405 if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3406 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3407 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3409 spin_unlock(&new_lp->lp_lock);
3410 spin_unlock(&lp->lp_lock);
3412 rc = lnet_peer_set_primary_data(new_lp, pbuf);
3413 lnet_consolidate_routes_locked(lp, new_lp);
3414 lnet_peer_ni_decref_locked(lpni);
3418 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3419 libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3421 mutex_unlock(&the_lnet.ln_api_mutex);
3423 spin_lock(&lp->lp_lock);
3424 /* Tell discovery to re-check the peer immediately. */
3426 rc = LNET_REDISCOVER_PEER;
3431 * A ping failed. Clear the PING_FAILED state and set the
3432 * FORCE_PING state, to ensure a retry even if discovery is
3433 * disabled. This avoids being left with incorrect state.
3435 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3436 __must_hold(&lp->lp_lock)
3438 struct lnet_handle_md mdh;
3441 mdh = lp->lp_ping_mdh;
3442 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3443 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3444 lp->lp_state |= LNET_PEER_FORCE_PING;
3445 rc = lp->lp_ping_error;
3446 lp->lp_ping_error = 0;
3447 spin_unlock(&lp->lp_lock);
3449 if (!LNetMDHandleIsInvalid(mdh))
3452 CDEBUG(D_NET, "peer %s:%d\n",
3453 libcfs_nidstr(&lp->lp_primary_nid), rc);
3455 spin_lock(&lp->lp_lock);
3456 return rc ? rc : LNET_REDISCOVER_PEER;
3459 /* Active side of ping. */
3460 static int lnet_peer_send_ping(struct lnet_peer *lp)
3461 __must_hold(&lp->lp_lock)
3467 lp->lp_state |= LNET_PEER_PING_SENT;
3468 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3469 spin_unlock(&lp->lp_lock);
3471 cpt = lnet_net_lock_current();
3472 /* Refcount for MD. */
3473 lnet_peer_addref_locked(lp);
3474 lnet_net_unlock(cpt);
3476 nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3478 rc = lnet_send_ping(lnet_nid_to_nid4(&lp->lp_primary_nid),
3479 &lp->lp_ping_mdh, nnis, lp,
3480 the_lnet.ln_dc_handler, false);
3483 * if LNetMDBind in lnet_send_ping fails we need to decrement the
3484 * refcount on the peer, otherwise LNetMDUnlink will be called
3485 * which will eventually do that.
3489 lnet_peer_decref_locked(lp);
3490 lnet_net_unlock(cpt);
3491 rc = -rc; /* change the rc to negative value */
3493 } else if (rc < 0) {
3497 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3499 spin_lock(&lp->lp_lock);
3503 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3505 * The errors that get us here are considered hard errors and
3506 * cause Discovery to terminate. So we clear PING_SENT, but do
3507 * not set either PING_FAILED or FORCE_PING. In fact we need
3508 * to clear PING_FAILED, because the unlink event handler will
3509 * have set it if we called LNetMDUnlink() above.
3511 spin_lock(&lp->lp_lock);
3512 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3517 * This function exists because you cannot call LNetMDUnlink() from an
3520 static int lnet_peer_push_failed(struct lnet_peer *lp)
3521 __must_hold(&lp->lp_lock)
3523 struct lnet_handle_md mdh;
3526 mdh = lp->lp_push_mdh;
3527 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3528 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3529 rc = lp->lp_push_error;
3530 lp->lp_push_error = 0;
3531 spin_unlock(&lp->lp_lock);
3533 if (!LNetMDHandleIsInvalid(mdh))
3536 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3537 spin_lock(&lp->lp_lock);
3538 return rc ? rc : LNET_REDISCOVER_PEER;
3542 * Mark the peer as discovered.
3544 static int lnet_peer_discovered(struct lnet_peer *lp)
3545 __must_hold(&lp->lp_lock)
3547 lp->lp_state |= LNET_PEER_DISCOVERED;
3548 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3549 LNET_PEER_REDISCOVER);
3551 lp->lp_dc_error = 0;
3553 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3558 /* Active side of push. */
3559 static int lnet_peer_send_push(struct lnet_peer *lp)
3560 __must_hold(&lp->lp_lock)
3562 struct lnet_ping_buffer *pbuf;
3563 struct lnet_process_id id;
3568 /* Don't push to a non-multi-rail peer. */
3569 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3570 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3571 /* if peer's NIDs are uptodate then peer is discovered */
3572 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3573 rc = lnet_peer_discovered(lp);
3580 lp->lp_state |= LNET_PEER_PUSH_SENT;
3581 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3582 spin_unlock(&lp->lp_lock);
3584 cpt = lnet_net_lock_current();
3585 pbuf = the_lnet.ln_ping_target;
3586 lnet_ping_buffer_addref(pbuf);
3587 lnet_net_unlock(cpt);
3589 /* Push source MD */
3590 md.start = &pbuf->pb_info;
3591 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3592 md.threshold = 2; /* Put/Ack */
3594 md.options = LNET_MD_TRACK_RESPONSE;
3595 md.handler = the_lnet.ln_dc_handler;
3598 rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3600 lnet_ping_buffer_decref(pbuf);
3601 CERROR("Can't bind push source MD: %d\n", rc);
3605 cpt = lnet_net_lock_current();
3606 /* Refcount for MD. */
3607 lnet_peer_addref_locked(lp);
3608 id.pid = LNET_PID_LUSTRE;
3609 if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3610 id.nid = lnet_nid_to_nid4(&lp->lp_disc_dst_nid);
3612 id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
3613 lnet_net_unlock(cpt);
3615 rc = LNetPut(lnet_nid_to_nid4(&lp->lp_disc_src_nid), lp->lp_push_mdh,
3616 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3617 LNET_PROTO_PING_MATCHBITS, 0, 0);
3620 * reset the discovery nid. There is no need to restrict sending
3621 * from that source, if we call lnet_push_update_to_peers(). It'll
3622 * get set to a specific NID, if we initiate discovery from the
3625 lp->lp_disc_src_nid = LNET_ANY_NID;
3626 lp->lp_disc_dst_nid = LNET_ANY_NID;
3631 CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3633 spin_lock(&lp->lp_lock);
3637 LNetMDUnlink(lp->lp_push_mdh);
3638 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3640 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3643 * The errors that get us here are considered hard errors and
3644 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3645 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3646 * because the unlink event handler will have set it if we
3647 * called LNetMDUnlink() above.
3649 spin_lock(&lp->lp_lock);
3650 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3655 * Wait for work to be queued or some other change that must be
3656 * attended to. Returns non-zero if the discovery thread should shut
3659 static int lnet_peer_discovery_wait_for_work(void)
3666 cpt = lnet_net_lock_current();
3668 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3669 TASK_INTERRUPTIBLE);
3670 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3672 if (lnet_push_target_resize_needed() ||
3673 the_lnet.ln_push_target->pb_needs_post)
3675 if (!list_empty(&the_lnet.ln_dc_request))
3677 if (!list_empty(&the_lnet.ln_msg_resend))
3679 lnet_net_unlock(cpt);
3682 * wakeup max every second to check if there are peers that
3683 * have been stuck on the working queue for greater than
3686 schedule_timeout(cfs_time_seconds(1));
3687 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3688 cpt = lnet_net_lock_current();
3690 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3692 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3695 lnet_net_unlock(cpt);
3697 CDEBUG(D_NET, "woken: %d\n", rc);
3703 * Messages that were pending on a destroyed peer will be put on a global
3704 * resend list. The message resend list will be checked by
3705 * the discovery thread when it wakes up, and will resend messages. These
3706 * messages can still be sendable in the case the lpni which was the initial
3707 * cause of the message re-queue was transfered to another peer.
3709 * It is possible that LNet could be shutdown while we're iterating
3710 * through the list. lnet_shudown_lndnets() will attempt to access the
3711 * resend list, but will have to wait until the spinlock is released, by
3712 * which time there shouldn't be any more messages on the resend list.
3713 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3714 * for the messages so they can be released. The other case is that
3715 * lnet_shudown_lndnets() can finalize all the messages before this
3716 * function can visit the resend list, in which case this function will be
3719 static void lnet_resend_msgs(void)
3721 struct lnet_msg *msg, *tmp;
3725 spin_lock(&the_lnet.ln_msg_resend_lock);
3726 list_splice(&the_lnet.ln_msg_resend, &resend);
3727 spin_unlock(&the_lnet.ln_msg_resend_lock);
3729 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3730 list_del_init(&msg->msg_list);
3731 rc = lnet_send(msg->msg_src_nid_param, msg,
3732 msg->msg_rtr_nid_param);
3734 CNETERR("Error sending %s to %s: %d\n",
3735 lnet_msgtyp2str(msg->msg_type),
3736 libcfs_id2str(msg->msg_target), rc);
3737 lnet_finalize(msg, rc);
3742 /* The discovery thread. */
3743 static int lnet_peer_discovery(void *arg)
3745 struct lnet_peer *lp;
3748 wait_for_completion(&the_lnet.ln_started);
3750 CDEBUG(D_NET, "started\n");
3753 if (lnet_peer_discovery_wait_for_work())
3756 if (lnet_push_target_resize_needed())
3757 lnet_push_target_resize();
3758 else if (the_lnet.ln_push_target->pb_needs_post)
3759 lnet_push_target_post(the_lnet.ln_push_target,
3760 &the_lnet.ln_push_target_md);
3764 lnet_net_lock(LNET_LOCK_EX);
3765 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3766 lnet_net_unlock(LNET_LOCK_EX);
3771 * Process all incoming discovery work requests. When
3772 * discovery must wait on a peer to change state, it
3773 * is added to the tail of the ln_dc_working queue. A
3774 * timestamp keeps track of when the peer was added,
3775 * so we can time out discovery requests that take too
3778 while (!list_empty(&the_lnet.ln_dc_request)) {
3779 lp = list_first_entry(&the_lnet.ln_dc_request,
3780 struct lnet_peer, lp_dc_list);
3781 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3783 * set the time the peer was put on the dc_working
3784 * queue. It shouldn't remain on the queue
3785 * forever, in case the GET message (for ping)
3786 * doesn't get a REPLY or the PUT message (for
3787 * push) doesn't get an ACK.
3789 lp->lp_last_queued = ktime_get_real_seconds();
3790 lnet_net_unlock(LNET_LOCK_EX);
3792 if (lnet_push_target_resize_needed())
3793 lnet_push_target_resize();
3794 else if (the_lnet.ln_push_target->pb_needs_post)
3795 lnet_push_target_post(the_lnet.ln_push_target,
3796 &the_lnet.ln_push_target_md);
3799 * Select an action depending on the state of
3800 * the peer and whether discovery is disabled.
3801 * The check whether discovery is disabled is
3802 * done after the code that handles processing
3803 * for arrived data, cleanup for failures, and
3804 * forcing a Ping or Push.
3806 spin_lock(&lp->lp_lock);
3807 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3808 libcfs_nidstr(&lp->lp_primary_nid), lp,
3810 if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3811 LNET_PEER_MARK_DELETED))
3812 rc = lnet_peer_deletion(lp);
3813 else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3814 rc = lnet_peer_data_present(lp);
3815 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3816 rc = lnet_peer_ping_failed(lp);
3817 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3818 rc = lnet_peer_push_failed(lp);
3819 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3820 rc = lnet_peer_send_ping(lp);
3821 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3822 rc = lnet_peer_send_push(lp);
3823 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3824 rc = lnet_peer_send_ping(lp);
3825 else if (lnet_peer_needs_push(lp))
3826 rc = lnet_peer_send_push(lp);
3828 rc = lnet_peer_discovered(lp);
3829 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3830 libcfs_nidstr(&lp->lp_primary_nid), lp,
3833 if (rc == LNET_REDISCOVER_PEER) {
3834 spin_unlock(&lp->lp_lock);
3835 lnet_net_lock(LNET_LOCK_EX);
3836 list_move(&lp->lp_dc_list,
3837 &the_lnet.ln_dc_request);
3839 !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3840 spin_unlock(&lp->lp_lock);
3841 lnet_net_lock(LNET_LOCK_EX);
3842 lnet_peer_discovery_complete(lp, rc);
3844 spin_unlock(&lp->lp_lock);
3845 lnet_net_lock(LNET_LOCK_EX);
3848 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3853 lnet_net_unlock(LNET_LOCK_EX);
3856 CDEBUG(D_NET, "stopping\n");
3858 * Clean up before telling lnet_peer_discovery_stop() that
3859 * we're done. Use wake_up() below to somewhat reduce the
3860 * size of the thundering herd if there are multiple threads
3861 * waiting on discovery of a single peer.
3864 /* Queue cleanup 1: stop all pending pings and pushes. */
3865 lnet_net_lock(LNET_LOCK_EX);
3866 while (!list_empty(&the_lnet.ln_dc_working)) {
3867 lp = list_first_entry(&the_lnet.ln_dc_working,
3868 struct lnet_peer, lp_dc_list);
3869 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3870 lnet_net_unlock(LNET_LOCK_EX);
3871 lnet_peer_cancel_discovery(lp);
3872 lnet_net_lock(LNET_LOCK_EX);
3874 lnet_net_unlock(LNET_LOCK_EX);
3876 /* Queue cleanup 2: wait for the expired queue to clear. */
3877 while (!list_empty(&the_lnet.ln_dc_expired))
3878 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3880 /* Queue cleanup 3: clear the request queue. */
3881 lnet_net_lock(LNET_LOCK_EX);
3882 while (!list_empty(&the_lnet.ln_dc_request)) {
3883 lp = list_first_entry(&the_lnet.ln_dc_request,
3884 struct lnet_peer, lp_dc_list);
3885 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3887 lnet_net_unlock(LNET_LOCK_EX);
3889 lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3890 the_lnet.ln_dc_handler = NULL;
3892 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3893 wake_up(&the_lnet.ln_dc_waitq);
3895 CDEBUG(D_NET, "stopped\n");
3900 /* ln_api_mutex is held on entry. */
3901 int lnet_peer_discovery_start(void)
3903 struct task_struct *task;
3906 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3909 the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3910 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3911 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3914 CERROR("Can't start peer discovery thread: %d\n", rc);
3916 the_lnet.ln_dc_handler = NULL;
3918 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3921 CDEBUG(D_NET, "discovery start: %d\n", rc);
3926 /* ln_api_mutex is held on entry. */
3927 void lnet_peer_discovery_stop(void)
3929 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3932 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3933 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3935 /* In the LNetNIInit() path we may be stopping discovery before it
3936 * entered its work loop
3938 if (!completion_done(&the_lnet.ln_started))
3939 complete(&the_lnet.ln_started);
3941 wake_up(&the_lnet.ln_dc_waitq);
3943 wait_event(the_lnet.ln_dc_waitq,
3944 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3946 LASSERT(list_empty(&the_lnet.ln_dc_request));
3947 LASSERT(list_empty(&the_lnet.ln_dc_working));
3948 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3950 CDEBUG(D_NET, "discovery stopped\n");
3956 lnet_debug_peer(lnet_nid_t nid)
3958 char *aliveness = "NA";
3959 struct lnet_peer_ni *lp;
3962 cpt = lnet_cpt_of_nid(nid, NULL);
3965 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3967 lnet_net_unlock(cpt);
3968 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3972 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3973 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3975 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3976 libcfs_nidstr(&lp->lpni_nid), kref_read(&lp->lpni_kref),
3977 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3978 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3979 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3981 lnet_peer_ni_decref_locked(lp);
3983 lnet_net_unlock(cpt);
3986 /* Gathering information for userspace. */
3988 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3989 char aliveness[LNET_MAX_STR_LEN],
3990 __u32 *cpt_iter, __u32 *refcount,
3991 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3992 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3993 __u32 *peer_tx_qnob)
3995 struct lnet_peer_table *peer_table;
3996 struct lnet_peer_ni *lp;
4001 /* get the number of CPTs */
4002 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4004 /* if the cpt number to be examined is >= the number of cpts in
4005 * the system then indicate that there are no more cpts to examin
4007 if (*cpt_iter >= lncpt)
4010 /* get the current table */
4011 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
4012 /* if the ptable is NULL then there are no more cpts to examine */
4013 if (peer_table == NULL)
4016 lnet_net_lock(*cpt_iter);
4018 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
4019 struct list_head *peers = &peer_table->pt_hash[j];
4021 list_for_each_entry(lp, peers, lpni_hashlist) {
4022 if (!nid_is_nid4(&lp->lpni_nid))
4024 if (peer_index-- > 0)
4027 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
4028 if (lnet_isrouter(lp) ||
4029 lnet_peer_aliveness_enabled(lp))
4030 snprintf(aliveness, LNET_MAX_STR_LEN,
4031 lnet_is_peer_ni_alive(lp) ? "up" : "down");
4033 *nid = lnet_nid_to_nid4(&lp->lpni_nid);
4034 *refcount = kref_read(&lp->lpni_kref);
4035 *ni_peer_tx_credits =
4036 lp->lpni_net->net_tunables.lct_peer_tx_credits;
4037 *peer_tx_credits = lp->lpni_txcredits;
4038 *peer_rtr_credits = lp->lpni_rtrcredits;
4039 *peer_min_rtr_credits = lp->lpni_mintxcredits;
4040 *peer_tx_qnob = lp->lpni_txqnob;
4046 lnet_net_unlock(*cpt_iter);
4050 return found ? 0 : -ENOENT;
4053 /* ln_api_mutex is held, which keeps the peer list stable */
4054 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
4056 struct lnet_ioctl_element_stats *lpni_stats;
4057 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
4058 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
4059 struct lnet_peer_ni_credit_info *lpni_info;
4060 struct lnet_peer_ni *lpni;
4061 struct lnet_peer *lp;
4066 lp = lnet_find_peer(cfg->prcfg_prim_nid);
4073 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
4074 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
4075 size *= lp->lp_nnis;
4076 if (size > cfg->prcfg_size) {
4077 cfg->prcfg_size = size;
4082 cfg->prcfg_prim_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4083 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
4084 cfg->prcfg_cfg_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4085 cfg->prcfg_count = lp->lp_nnis;
4086 cfg->prcfg_size = size;
4087 cfg->prcfg_state = lp->lp_state;
4089 /* Allocate helper buffers. */
4091 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
4094 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
4097 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
4098 if (!lpni_msg_stats)
4099 goto out_free_stats;
4100 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
4102 goto out_free_msg_stats;
4107 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
4108 if (!nid_is_nid4(&lpni->lpni_nid))
4110 nid = lnet_nid_to_nid4(&lpni->lpni_nid);
4111 if (copy_to_user(bulk, &nid, sizeof(nid)))
4112 goto out_free_hstats;
4113 bulk += sizeof(nid);
4115 memset(lpni_info, 0, sizeof(*lpni_info));
4116 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
4117 if (lnet_isrouter(lpni) ||
4118 lnet_peer_aliveness_enabled(lpni))
4119 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
4120 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
4122 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
4123 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
4124 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
4125 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
4126 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
4127 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
4128 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
4129 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4130 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4131 goto out_free_hstats;
4132 bulk += sizeof(*lpni_info);
4134 memset(lpni_stats, 0, sizeof(*lpni_stats));
4135 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4136 LNET_STATS_TYPE_SEND);
4137 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4138 LNET_STATS_TYPE_RECV);
4139 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4140 LNET_STATS_TYPE_DROP);
4141 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4142 goto out_free_hstats;
4143 bulk += sizeof(*lpni_stats);
4144 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4145 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4146 goto out_free_hstats;
4147 bulk += sizeof(*lpni_msg_stats);
4148 lpni_hstats->hlpni_network_timeout =
4149 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4150 lpni_hstats->hlpni_remote_dropped =
4151 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4152 lpni_hstats->hlpni_remote_timeout =
4153 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4154 lpni_hstats->hlpni_remote_error =
4155 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4156 lpni_hstats->hlpni_health_value =
4157 atomic_read(&lpni->lpni_healthv);
4158 lpni_hstats->hlpni_ping_count = lpni->lpni_ping_count;
4159 lpni_hstats->hlpni_next_ping = lpni->lpni_next_ping;
4160 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4161 goto out_free_hstats;
4162 bulk += sizeof(*lpni_hstats);
4167 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4169 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4171 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4173 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4175 lnet_peer_decref_locked(lp);
4180 /* must hold net_lock/0 */
4182 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4183 struct list_head *recovery_queue,
4186 /* the mt could've shutdown and cleaned up the queues */
4187 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4190 if (!list_empty(&lpni->lpni_recovery))
4193 if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4196 if (!lpni->lpni_last_alive) {
4198 "lpni %s(%p) not eligible for recovery last alive %lld\n",
4199 libcfs_nidstr(&lpni->lpni_nid), lpni,
4200 lpni->lpni_last_alive);
4204 if (lnet_recovery_limit &&
4205 now > lpni->lpni_last_alive + lnet_recovery_limit) {
4206 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4207 libcfs_nidstr(&lpni->lpni_nid),
4208 lpni->lpni_last_alive);
4209 /* Reset the ping count so that if this peer NI is added back to
4210 * the recovery queue we will send the first ping right away.
4212 lpni->lpni_ping_count = 0;
4216 /* This peer NI is going on the recovery queue, so take a ref on it */
4217 lnet_peer_ni_addref_locked(lpni);
4219 lnet_peer_ni_set_next_ping(lpni, now);
4221 CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4222 libcfs_nidstr(&lpni->lpni_nid),
4223 lpni->lpni_ping_count,
4224 lpni->lpni_next_ping,
4225 lpni->lpni_last_alive,
4226 atomic_read(&lpni->lpni_healthv));
4228 list_add_tail(&lpni->lpni_recovery, recovery_queue);
4231 /* Call with the ln_api_mutex held */
4233 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
4235 struct lnet_peer_table *ptable;
4236 struct lnet_peer *lp;
4237 struct lnet_peer_net *lpn;
4238 struct lnet_peer_ni *lpni;
4243 if (the_lnet.ln_state != LNET_STATE_RUNNING)
4246 now = ktime_get_seconds();
4249 lnet_net_lock(LNET_LOCK_EX);
4250 lpni = lnet_find_peer_ni_locked(nid);
4252 lnet_net_unlock(LNET_LOCK_EX);
4255 lnet_set_lpni_healthv_locked(lpni, value);
4256 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4257 &the_lnet.ln_mt_peerNIRecovq, now);
4258 lnet_peer_ni_decref_locked(lpni);
4259 lnet_net_unlock(LNET_LOCK_EX);
4263 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4266 * Walk all the peers and reset the health value for each one to the
4269 lnet_net_lock(LNET_LOCK_EX);
4270 for (cpt = 0; cpt < lncpt; cpt++) {
4271 ptable = the_lnet.ln_peer_tables[cpt];
4272 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4273 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4274 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4276 lnet_set_lpni_healthv_locked(lpni,
4278 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4279 &the_lnet.ln_mt_peerNIRecovq, now);
4284 lnet_net_unlock(LNET_LOCK_EX);