4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
35 #define DEBUG_SUBSYSTEM S_LNET
37 #include <linux/sched.h>
38 #ifdef HAVE_SCHED_HEADERS
39 #include <linux/sched/signal.h>
41 #include <linux/uaccess.h>
43 #include <lnet/udsp.h>
44 #include <lnet/lib-lnet.h>
45 #include <uapi/linux/lnet/lnet-dlc.h>
47 /* Value indicating that recovery needs to re-check a peer immediately. */
48 #define LNET_REDISCOVER_PEER (1)
50 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
53 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
55 if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
56 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
57 lnet_peer_ni_decref_locked(lpni);
62 lnet_peer_net_added(struct lnet_net *net)
64 struct lnet_peer_ni *lpni, *tmp;
66 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
67 lpni_on_remote_peer_ni_list) {
69 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
72 spin_lock(&lpni->lpni_lock);
73 lpni->lpni_txcredits =
74 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
75 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
76 lpni->lpni_rtrcredits =
77 lnet_peer_buffer_credits(lpni->lpni_net);
78 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
79 spin_unlock(&lpni->lpni_lock);
81 lnet_peer_remove_from_remote_list(lpni);
87 lnet_peer_tables_destroy(void)
89 struct lnet_peer_table *ptable;
90 struct list_head *hash;
94 if (!the_lnet.ln_peer_tables)
97 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
98 hash = ptable->pt_hash;
99 if (!hash) /* not intialized */
102 LASSERT(list_empty(&ptable->pt_zombie_list));
104 ptable->pt_hash = NULL;
105 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
106 LASSERT(list_empty(&hash[j]));
108 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
111 cfs_percpt_free(the_lnet.ln_peer_tables);
112 the_lnet.ln_peer_tables = NULL;
116 lnet_peer_tables_create(void)
118 struct lnet_peer_table *ptable;
119 struct list_head *hash;
123 the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
125 if (the_lnet.ln_peer_tables == NULL) {
126 CERROR("Failed to allocate cpu-partition peer tables\n");
130 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
131 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
132 LNET_PEER_HASH_SIZE * sizeof(*hash));
134 CERROR("Failed to create peer hash table\n");
135 lnet_peer_tables_destroy();
139 spin_lock_init(&ptable->pt_zombie_lock);
140 INIT_LIST_HEAD(&ptable->pt_zombie_list);
142 INIT_LIST_HEAD(&ptable->pt_peer_list);
144 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
145 INIT_LIST_HEAD(&hash[j]);
146 ptable->pt_hash = hash; /* sign of initialization */
152 static struct lnet_peer_ni *
153 lnet_peer_ni_alloc(lnet_nid_t nid)
155 struct lnet_peer_ni *lpni;
156 struct lnet_net *net;
159 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
161 LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
165 INIT_LIST_HEAD(&lpni->lpni_txq);
166 INIT_LIST_HEAD(&lpni->lpni_hashlist);
167 INIT_LIST_HEAD(&lpni->lpni_peer_nis);
168 INIT_LIST_HEAD(&lpni->lpni_recovery);
169 INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
170 INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
171 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
172 atomic_set(&lpni->lpni_refcount, 1);
173 lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
175 spin_lock_init(&lpni->lpni_lock);
177 if (lnet_peers_start_down())
178 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
180 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
181 lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
182 lpni->lpni_nid = nid;
183 lpni->lpni_cpt = cpt;
184 atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
186 net = lnet_get_net_locked(LNET_NIDNET(nid));
187 lpni->lpni_net = net;
189 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
190 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
191 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
192 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
195 * This peer_ni is not on a local network, so we
196 * cannot add the credits here. In case the net is
197 * added later, add the peer_ni to the remote peer ni
198 * list so it can be easily found and revisited.
200 /* FIXME: per-net implementation instead? */
201 lnet_peer_ni_addref_locked(lpni);
202 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
203 &the_lnet.ln_remote_peer_ni_list);
206 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
211 static struct lnet_peer_net *
212 lnet_peer_net_alloc(__u32 net_id)
214 struct lnet_peer_net *lpn;
216 LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
220 INIT_LIST_HEAD(&lpn->lpn_peer_nets);
221 INIT_LIST_HEAD(&lpn->lpn_peer_nis);
222 lpn->lpn_net_id = net_id;
223 lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
225 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
231 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
233 struct lnet_peer *lp;
235 CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
237 LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
238 LASSERT(list_empty(&lpn->lpn_peer_nis));
239 LASSERT(list_empty(&lpn->lpn_peer_nets));
241 lpn->lpn_peer = NULL;
242 LIBCFS_FREE(lpn, sizeof(*lpn));
244 lnet_peer_decref_locked(lp);
247 static struct lnet_peer *
248 lnet_peer_alloc(lnet_nid_t nid)
250 struct lnet_peer *lp;
252 LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
256 INIT_LIST_HEAD(&lp->lp_rtrq);
257 INIT_LIST_HEAD(&lp->lp_routes);
258 INIT_LIST_HEAD(&lp->lp_peer_list);
259 INIT_LIST_HEAD(&lp->lp_peer_nets);
260 INIT_LIST_HEAD(&lp->lp_dc_list);
261 INIT_LIST_HEAD(&lp->lp_dc_pendq);
262 INIT_LIST_HEAD(&lp->lp_rtr_list);
263 init_waitqueue_head(&lp->lp_dc_waitq);
264 spin_lock_init(&lp->lp_lock);
265 lp->lp_primary_nid = nid;
266 lp->lp_disc_src_nid = LNET_NID_ANY;
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
286 if (nid == LNET_NID_LO_0)
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_nid2str(lp->lp_primary_nid));
296 lnet_destroy_peer_locked(struct lnet_peer *lp)
298 CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(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_nid2str(lp->lp_primary_nid),
382 libcfs_nid2str(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_nid2str(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_nid2str(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)
512 lnet_peer_cancel_discovery(peer);
513 lnet_net_lock(LNET_LOCK_EX);
514 lnet_peer_del_locked(peer);
515 lnet_net_unlock(LNET_LOCK_EX);
521 * Delete a NID from a peer. Call with ln_api_mutex held.
524 * -EPERM: Non-DLC deletion from DLC-configured peer.
525 * -ENOENT: No lnet_peer_ni corresponding to the nid.
526 * -ECHILD: The lnet_peer_ni isn't connected to the peer.
527 * -EBUSY: The lnet_peer_ni is the primary, and not the only peer_ni.
530 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
532 struct lnet_peer_ni *lpni;
533 lnet_nid_t primary_nid = lp->lp_primary_nid;
535 bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
537 if (!(flags & LNET_PEER_CONFIGURED)) {
538 if (lp->lp_state & LNET_PEER_CONFIGURED) {
543 lpni = lnet_find_peer_ni_locked(nid);
548 lnet_peer_ni_decref_locked(lpni);
549 if (lp != lpni->lpni_peer_net->lpn_peer) {
555 * This function only allows deletion of the primary NID if it
558 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
563 lnet_net_lock(LNET_LOCK_EX);
565 if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
566 struct lnet_peer_ni *lpni2;
567 /* assign the next peer_ni to be the primary */
568 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
570 lp->lp_primary_nid = lpni2->lpni_nid;
572 rc = lnet_peer_ni_del_locked(lpni, force);
574 lnet_net_unlock(LNET_LOCK_EX);
577 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
578 libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
584 lnet_peer_table_cleanup_locked(struct lnet_net *net,
585 struct lnet_peer_table *ptable)
588 struct lnet_peer_ni *next;
589 struct lnet_peer_ni *lpni;
590 struct lnet_peer *peer;
592 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
593 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
595 if (net != NULL && net != lpni->lpni_net)
598 peer = lpni->lpni_peer_net->lpn_peer;
599 if (peer->lp_primary_nid != lpni->lpni_nid) {
600 lnet_peer_ni_del_locked(lpni, false);
604 * Removing the primary NID implies removing
605 * the entire peer. Advance next beyond any
606 * peer_ni that belongs to the same peer.
608 list_for_each_entry_from(next, &ptable->pt_hash[i],
610 if (next->lpni_peer_net->lpn_peer != peer)
613 lnet_peer_del_locked(peer);
619 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
621 wait_var_event_warning(&ptable->pt_zombies,
622 ptable->pt_zombies == 0,
623 "Waiting for %d zombies on peer table\n",
628 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
629 struct lnet_peer_table *ptable)
631 struct lnet_peer_ni *lp;
632 struct lnet_peer_ni *tmp;
636 for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
637 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
639 if (net != lp->lpni_net)
642 if (!lnet_isrouter(lp))
645 gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
647 lnet_net_unlock(LNET_LOCK_EX);
648 lnet_del_route(LNET_NET_ANY, gw_nid);
649 lnet_net_lock(LNET_LOCK_EX);
655 lnet_peer_tables_cleanup(struct lnet_net *net)
658 struct lnet_peer_table *ptable;
660 LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
661 /* If just deleting the peers for a NI, get rid of any routes these
662 * peers are gateways for. */
663 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
664 lnet_net_lock(LNET_LOCK_EX);
665 lnet_peer_table_del_rtrs_locked(net, ptable);
666 lnet_net_unlock(LNET_LOCK_EX);
669 /* Start the cleanup process */
670 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
671 lnet_net_lock(LNET_LOCK_EX);
672 lnet_peer_table_cleanup_locked(net, ptable);
673 lnet_net_unlock(LNET_LOCK_EX);
676 cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
677 lnet_peer_ni_finalize_wait(ptable);
680 static struct lnet_peer_ni *
681 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
683 struct list_head *peers;
684 struct lnet_peer_ni *lp;
686 if (the_lnet.ln_state != LNET_STATE_RUNNING)
689 peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
690 list_for_each_entry(lp, peers, lpni_hashlist) {
691 if (lp->lpni_nid == nid) {
692 lnet_peer_ni_addref_locked(lp);
700 struct lnet_peer_ni *
701 lnet_find_peer_ni_locked(lnet_nid_t nid)
703 struct lnet_peer_ni *lpni;
704 struct lnet_peer_table *ptable;
707 cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
709 ptable = the_lnet.ln_peer_tables[cpt];
710 lpni = lnet_get_peer_ni_locked(ptable, nid);
715 struct lnet_peer_ni *
716 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
718 struct lnet_peer_net *lpn;
719 struct lnet_peer_ni *lpni;
721 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
725 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
726 if (lpni->lpni_nid == nid)
734 lnet_find_peer(lnet_nid_t nid)
736 struct lnet_peer_ni *lpni;
737 struct lnet_peer *lp = NULL;
740 cpt = lnet_net_lock_current();
741 lpni = lnet_find_peer_ni_locked(nid);
743 lp = lpni->lpni_peer_net->lpn_peer;
744 lnet_peer_addref_locked(lp);
745 lnet_peer_ni_decref_locked(lpni);
747 lnet_net_unlock(cpt);
752 struct lnet_peer_net *
753 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
755 struct lnet_peer_net *net;
758 /* no net id provided return the first net */
759 net = list_first_entry_or_null(&lp->lp_peer_nets,
760 struct lnet_peer_net,
766 /* find the net after the one provided */
767 list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
768 if (net->lpn_net_id == prev_lpn_id) {
770 * if we reached the end of the list loop to the
773 if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
774 return list_first_entry_or_null(&lp->lp_peer_nets,
775 struct lnet_peer_net,
778 return list_next_entry(net, lpn_peer_nets);
785 struct lnet_peer_ni *
786 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
787 struct lnet_peer_net *peer_net,
788 struct lnet_peer_ni *prev)
790 struct lnet_peer_ni *lpni;
791 struct lnet_peer_net *net = peer_net;
795 if (list_empty(&peer->lp_peer_nets))
798 net = list_entry(peer->lp_peer_nets.next,
799 struct lnet_peer_net,
802 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
808 if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
810 * if you reached the end of the peer ni list and the peer
811 * net is specified then there are no more peer nis in that
818 * we reached the end of this net ni list. move to the
821 if (prev->lpni_peer_net->lpn_peer_nets.next ==
823 /* no more nets and no more NIs. */
826 /* get the next net */
827 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
828 struct lnet_peer_net,
830 /* get the ni on it */
831 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
837 /* there are more nis left */
838 lpni = list_entry(prev->lpni_peer_nis.next,
839 struct lnet_peer_ni, lpni_peer_nis);
844 /* Call with the ln_api_mutex held */
845 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
847 struct lnet_process_id id;
848 struct lnet_peer_table *ptable;
849 struct lnet_peer *lp;
858 if (the_lnet.ln_state != LNET_STATE_RUNNING)
861 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
864 * Count the number of peers, and return E2BIG if the buffer
865 * is too small. We'll also return the desired size.
868 for (cpt = 0; cpt < lncpt; cpt++) {
869 ptable = the_lnet.ln_peer_tables[cpt];
870 count += ptable->pt_peers;
872 size = count * sizeof(*ids);
877 * Walk the peer lists and copy out the primary nids.
878 * This is safe because the peer lists are only modified
879 * while the ln_api_mutex is held. So we don't need to
880 * hold the lnet_net_lock as well, and can therefore
881 * directly call copy_to_user().
884 memset(&id, 0, sizeof(id));
885 id.pid = LNET_PID_LUSTRE;
887 for (cpt = 0; cpt < lncpt; cpt++) {
888 ptable = the_lnet.ln_peer_tables[cpt];
889 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
892 id.nid = lp->lp_primary_nid;
893 if (copy_to_user(&ids[i], &id, sizeof(id)))
906 * Start pushes to peers that need to be updated for a configuration
907 * change on this node.
910 lnet_push_update_to_peers(int force)
912 struct lnet_peer_table *ptable;
913 struct lnet_peer *lp;
917 lnet_net_lock(LNET_LOCK_EX);
918 if (lnet_peer_discovery_disabled)
920 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
921 for (cpt = 0; cpt < lncpt; cpt++) {
922 ptable = the_lnet.ln_peer_tables[cpt];
923 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
925 spin_lock(&lp->lp_lock);
926 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
927 lp->lp_state |= LNET_PEER_FORCE_PUSH;
928 spin_unlock(&lp->lp_lock);
930 if (lnet_peer_needs_push(lp))
931 lnet_peer_queue_for_discovery(lp);
934 lnet_net_unlock(LNET_LOCK_EX);
935 wake_up(&the_lnet.ln_dc_waitq);
938 /* find the NID in the preferred gateways for the remote peer
940 * false: list is not empty and NID is not preferred
941 * false: list is empty
942 * true: nid is found in the list
945 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
948 struct lnet_nid_list *ne;
950 CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
951 libcfs_nid2str(lpni->lpni_nid),
952 list_empty(&lpni->lpni_rtr_pref_nids));
954 if (list_empty(&lpni->lpni_rtr_pref_nids))
957 /* iterate through all the preferred NIDs and see if any of them
958 * matches the provided gw_nid
960 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
961 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
962 libcfs_nid2str(ne->nl_nid),
963 libcfs_nid2str(gw_nid));
964 if (ne->nl_nid == gw_nid)
972 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
974 struct list_head zombies;
975 struct lnet_nid_list *ne;
976 struct lnet_nid_list *tmp;
977 int cpt = lpni->lpni_cpt;
979 INIT_LIST_HEAD(&zombies);
982 list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
983 lnet_net_unlock(cpt);
985 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
986 list_del(&ne->nl_list);
987 LIBCFS_FREE(ne, sizeof(*ne));
992 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
995 int cpt = lpni->lpni_cpt;
996 struct lnet_nid_list *ne = NULL;
998 /* This function is called with api_mutex held. When the api_mutex
999 * is held the list can not be modified, as it is only modified as
1000 * a result of applying a UDSP and that happens under api_mutex
1003 __must_hold(&the_lnet.ln_api_mutex);
1005 list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1006 if (ne->nl_nid == gw_nid)
1010 LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1014 ne->nl_nid = gw_nid;
1016 /* Lock the cpt to protect against addition and checks in the
1017 * selection algorithm
1020 list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1021 lnet_net_unlock(cpt);
1027 * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1028 * this is a preferred point-to-point path. Call with lnet_net_lock in
1032 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1034 struct lnet_nid_list *ne;
1036 if (lpni->lpni_pref_nnids == 0)
1038 if (lpni->lpni_pref_nnids == 1)
1039 return lpni->lpni_pref.nid == nid;
1040 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1041 if (ne->nl_nid == nid)
1048 * Set a single ni as preferred, provided no preferred ni is already
1049 * defined. Only to be used for non-multi-rail peer_ni.
1052 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1056 spin_lock(&lpni->lpni_lock);
1057 if (nid == LNET_NID_ANY) {
1059 } else if (lpni->lpni_pref_nnids > 0) {
1061 } else if (lpni->lpni_pref_nnids == 0) {
1062 lpni->lpni_pref.nid = nid;
1063 lpni->lpni_pref_nnids = 1;
1064 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1066 spin_unlock(&lpni->lpni_lock);
1068 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1069 libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
1074 * Clear the preferred NID from a non-multi-rail peer_ni, provided
1075 * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1078 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1082 spin_lock(&lpni->lpni_lock);
1083 if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1084 lpni->lpni_pref_nnids = 0;
1085 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1086 } else if (lpni->lpni_pref_nnids == 0) {
1091 spin_unlock(&lpni->lpni_lock);
1093 CDEBUG(D_NET, "peer %s: %d\n",
1094 libcfs_nid2str(lpni->lpni_nid), rc);
1099 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1101 lpni->lpni_sel_priority = priority;
1105 * Clear the preferred NIDs from a non-multi-rail peer.
1108 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1110 struct lnet_peer_ni *lpni = NULL;
1112 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1113 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1117 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1119 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1120 struct lnet_nid_list *ne1 = NULL;
1121 struct lnet_nid_list *ne2 = NULL;
1122 lnet_nid_t tmp_nid = LNET_NID_ANY;
1125 if (nid == LNET_NID_ANY) {
1130 if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1135 /* A non-MR node may have only one preferred NI per peer_ni */
1136 if (lpni->lpni_pref_nnids > 0 &&
1137 !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1142 /* add the new preferred nid to the list of preferred nids */
1143 if (lpni->lpni_pref_nnids != 0) {
1144 size_t alloc_size = sizeof(*ne1);
1146 if (lpni->lpni_pref_nnids == 1) {
1147 tmp_nid = lpni->lpni_pref.nid;
1148 INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1151 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1152 if (ne1->nl_nid == nid) {
1158 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1165 /* move the originally stored nid to the list */
1166 if (lpni->lpni_pref_nnids == 1) {
1167 LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1168 lpni->lpni_cpt, alloc_size);
1173 INIT_LIST_HEAD(&ne2->nl_list);
1174 ne2->nl_nid = tmp_nid;
1179 lnet_net_lock(LNET_LOCK_EX);
1180 spin_lock(&lpni->lpni_lock);
1181 if (lpni->lpni_pref_nnids == 0) {
1182 lpni->lpni_pref.nid = nid;
1185 list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1186 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1188 lpni->lpni_pref_nnids++;
1189 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1190 spin_unlock(&lpni->lpni_lock);
1191 lnet_net_unlock(LNET_LOCK_EX);
1194 if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1195 spin_lock(&lpni->lpni_lock);
1196 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1197 spin_unlock(&lpni->lpni_lock);
1199 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1200 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1205 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1207 struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1208 struct lnet_nid_list *ne = NULL;
1211 if (lpni->lpni_pref_nnids == 0) {
1216 if (lpni->lpni_pref_nnids == 1) {
1217 if (lpni->lpni_pref.nid != nid) {
1222 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1223 if (ne->nl_nid == nid)
1224 goto remove_nid_entry;
1232 lnet_net_lock(LNET_LOCK_EX);
1233 spin_lock(&lpni->lpni_lock);
1234 if (lpni->lpni_pref_nnids == 1)
1235 lpni->lpni_pref.nid = LNET_NID_ANY;
1237 list_del_init(&ne->nl_list);
1238 if (lpni->lpni_pref_nnids == 2) {
1239 struct lnet_nid_list *ne, *tmp;
1241 list_for_each_entry_safe(ne, tmp,
1242 &lpni->lpni_pref.nids,
1244 lpni->lpni_pref.nid = ne->nl_nid;
1245 list_del_init(&ne->nl_list);
1246 LIBCFS_FREE(ne, sizeof(*ne));
1250 lpni->lpni_pref_nnids--;
1251 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1252 spin_unlock(&lpni->lpni_lock);
1253 lnet_net_unlock(LNET_LOCK_EX);
1256 LIBCFS_FREE(ne, sizeof(*ne));
1258 CDEBUG(D_NET, "peer %s nid %s: %d\n",
1259 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1264 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1266 struct list_head zombies;
1267 struct lnet_nid_list *ne;
1268 struct lnet_nid_list *tmp;
1270 INIT_LIST_HEAD(&zombies);
1272 lnet_net_lock(LNET_LOCK_EX);
1273 if (lpni->lpni_pref_nnids == 1)
1274 lpni->lpni_pref.nid = LNET_NID_ANY;
1275 else if (lpni->lpni_pref_nnids > 1)
1276 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1277 lpni->lpni_pref_nnids = 0;
1278 lnet_net_unlock(LNET_LOCK_EX);
1280 list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1281 list_del_init(&ne->nl_list);
1282 LIBCFS_FREE(ne, sizeof(*ne));
1287 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1289 struct lnet_peer_ni *lpni;
1290 lnet_nid_t primary_nid = nid;
1292 lpni = lnet_find_peer_ni_locked(nid);
1294 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1295 lnet_peer_ni_decref_locked(lpni);
1302 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1303 __must_hold(&lp->lp_lock)
1305 if (lnet_peer_discovery_disabled)
1308 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1309 (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1320 lnet_is_discovery_disabled(struct lnet_peer *lp)
1324 spin_lock(&lp->lp_lock);
1325 rc = lnet_is_discovery_disabled_locked(lp);
1326 spin_unlock(&lp->lp_lock);
1332 LNetPrimaryNID(lnet_nid_t nid)
1334 struct lnet_peer *lp;
1335 struct lnet_peer_ni *lpni;
1336 lnet_nid_t primary_nid = nid;
1340 if (nid == LNET_NID_LO_0)
1341 return LNET_NID_LO_0;
1343 cpt = lnet_net_lock_current();
1344 lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1349 lp = lpni->lpni_peer_net->lpn_peer;
1351 while (!lnet_peer_is_uptodate(lp)) {
1352 spin_lock(&lp->lp_lock);
1353 /* force a full discovery cycle */
1354 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1355 spin_unlock(&lp->lp_lock);
1357 rc = lnet_discover_peer_locked(lpni, cpt, true);
1360 lp = lpni->lpni_peer_net->lpn_peer;
1362 /* Only try once if discovery is disabled */
1363 if (lnet_is_discovery_disabled(lp))
1366 primary_nid = lp->lp_primary_nid;
1368 lnet_peer_ni_decref_locked(lpni);
1370 lnet_net_unlock(cpt);
1372 CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1373 libcfs_nid2str(primary_nid), rc);
1376 EXPORT_SYMBOL(LNetPrimaryNID);
1378 struct lnet_peer_net *
1379 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1381 struct lnet_peer_net *peer_net;
1382 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1383 if (peer_net->lpn_net_id == net_id)
1390 * Attach a peer_ni to a peer_net and peer. This function assumes
1391 * peer_ni is not already attached to the peer_net/peer. The peer_ni
1392 * may be attached to a different peer, in which case it will be
1393 * properly detached first. The whole operation is done atomically.
1395 * This function consumes the reference on lpni and Always returns 0.
1396 * This is the last function called from functions that do return an
1397 * int, so returning 0 here allows the compiler to do a tail call.
1400 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1401 struct lnet_peer_net *lpn,
1402 struct lnet_peer_ni *lpni,
1405 struct lnet_peer_table *ptable;
1406 bool new_lpn = false;
1409 /* Install the new peer_ni */
1410 lnet_net_lock(LNET_LOCK_EX);
1411 /* Add peer_ni to global peer table hash, if necessary. */
1412 if (list_empty(&lpni->lpni_hashlist)) {
1413 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1415 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1416 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1417 ptable->pt_version++;
1418 lnet_peer_ni_addref_locked(lpni);
1421 /* Detach the peer_ni from an existing peer, if necessary. */
1422 if (lpni->lpni_peer_net) {
1423 LASSERT(lpni->lpni_peer_net != lpn);
1424 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1425 lnet_peer_detach_peer_ni_locked(lpni);
1426 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1427 lpni->lpni_peer_net = NULL;
1430 /* Add peer_ni to peer_net */
1431 lpni->lpni_peer_net = lpn;
1432 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1433 lnet_update_peer_net_healthv(lpni);
1434 lnet_peer_net_addref_locked(lpn);
1436 /* Add peer_net to peer */
1437 if (!lpn->lpn_peer) {
1440 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1441 lnet_peer_addref_locked(lp);
1444 /* Add peer to global peer list, if necessary */
1445 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1446 if (list_empty(&lp->lp_peer_list)) {
1447 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1452 /* Update peer state */
1453 spin_lock(&lp->lp_lock);
1454 if (flags & LNET_PEER_CONFIGURED) {
1455 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1456 lp->lp_state |= LNET_PEER_CONFIGURED;
1458 if (flags & LNET_PEER_MULTI_RAIL) {
1459 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1460 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1461 lnet_peer_clr_non_mr_pref_nids(lp);
1464 spin_unlock(&lp->lp_lock);
1470 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1472 CERROR("Failed to apply UDSPs on lpn %s\n",
1473 libcfs_net2str(lpn->lpn_net_id));
1475 rc = lnet_udsp_apply_policies_on_lpni(lpni);
1477 CERROR("Failed to apply UDSPs on lpni %s\n",
1478 libcfs_nid2str(lpni->lpni_nid));
1480 CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1481 libcfs_nid2str(lp->lp_primary_nid),
1482 libcfs_nid2str(lpni->lpni_nid), flags);
1483 lnet_peer_ni_decref_locked(lpni);
1484 lnet_net_unlock(LNET_LOCK_EX);
1490 * Create a new peer, with nid as its primary nid.
1492 * Call with the lnet_api_mutex held.
1495 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1497 struct lnet_peer *lp;
1498 struct lnet_peer_net *lpn;
1499 struct lnet_peer_ni *lpni;
1502 LASSERT(nid != LNET_NID_ANY);
1505 * No need for the lnet_net_lock here, because the
1506 * lnet_api_mutex is held.
1508 lpni = lnet_find_peer_ni_locked(nid);
1510 /* A peer with this NID already exists. */
1511 lp = lpni->lpni_peer_net->lpn_peer;
1512 lnet_peer_ni_decref_locked(lpni);
1514 * This is an error if the peer was configured and the
1515 * primary NID differs or an attempt is made to change
1516 * the Multi-Rail flag. Otherwise the assumption is
1517 * that an existing peer is being modified.
1519 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1520 if (lp->lp_primary_nid != nid)
1522 else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1526 /* Delete and recreate as a configured peer. */
1530 /* Create peer, peer_net, and peer_ni. */
1532 lp = lnet_peer_alloc(nid);
1535 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1538 lpni = lnet_peer_ni_alloc(nid);
1542 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1545 LIBCFS_FREE(lpn, sizeof(*lpn));
1547 LIBCFS_FREE(lp, sizeof(*lp));
1549 CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1550 libcfs_nid2str(nid), flags, rc);
1555 * Add a NID to a peer. Call with ln_api_mutex held.
1558 * -EPERM: Non-DLC addition to a DLC-configured peer.
1559 * -EEXIST: The NID was configured by DLC for a different peer.
1560 * -ENOMEM: Out of memory.
1561 * -ENOTUNIQ: Adding a second peer NID on a single network on a
1562 * non-multi-rail peer.
1565 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1567 struct lnet_peer_net *lpn;
1568 struct lnet_peer_ni *lpni;
1572 LASSERT(nid != LNET_NID_ANY);
1574 /* A configured peer can only be updated through configuration. */
1575 if (!(flags & LNET_PEER_CONFIGURED)) {
1576 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1583 * The MULTI_RAIL flag can be set but not cleared, because
1584 * that would leave the peer struct in an invalid state.
1586 if (flags & LNET_PEER_MULTI_RAIL) {
1587 spin_lock(&lp->lp_lock);
1588 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1589 lp->lp_state |= LNET_PEER_MULTI_RAIL;
1590 lnet_peer_clr_non_mr_pref_nids(lp);
1592 spin_unlock(&lp->lp_lock);
1593 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1598 lpni = lnet_find_peer_ni_locked(nid);
1601 * A peer_ni already exists. This is only a problem if
1602 * it is not connected to this peer and was configured
1605 if (lpni->lpni_peer_net->lpn_peer == lp)
1607 if (lnet_peer_ni_is_configured(lpni)) {
1611 /* If this is the primary NID, destroy the peer. */
1612 if (lnet_peer_ni_is_primary(lpni)) {
1613 struct lnet_peer *rtr_lp =
1614 lpni->lpni_peer_net->lpn_peer;
1615 int rtr_refcount = rtr_lp->lp_rtr_refcount;
1617 * if we're trying to delete a router it means
1618 * we're moving this peer NI to a new peer so must
1619 * transfer router properties to the new peer
1621 if (rtr_refcount > 0) {
1622 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1623 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1625 lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1626 lnet_peer_ni_decref_locked(lpni);
1627 lpni = lnet_peer_ni_alloc(nid);
1634 lpni = lnet_peer_ni_alloc(nid);
1642 * Get the peer_net. Check that we're not adding a second
1643 * peer_ni on a peer_net of a non-multi-rail peer.
1645 lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1647 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1652 } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1657 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1660 lnet_peer_ni_decref_locked(lpni);
1662 CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1663 libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1669 * Update the primary NID of a peer, if possible.
1671 * Call with the lnet_api_mutex held.
1674 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1676 lnet_nid_t old = lp->lp_primary_nid;
1679 if (lp->lp_primary_nid == nid)
1681 rc = lnet_peer_add_nid(lp, nid, flags);
1684 lp->lp_primary_nid = nid;
1686 CDEBUG(D_NET, "peer %s NID %s: %d\n",
1687 libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1692 * lpni creation initiated due to traffic either sending or receiving.
1695 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1697 struct lnet_peer *lp;
1698 struct lnet_peer_net *lpn;
1699 struct lnet_peer_ni *lpni;
1703 if (nid == LNET_NID_ANY) {
1708 /* lnet_net_lock is not needed here because ln_api_lock is held */
1709 lpni = lnet_find_peer_ni_locked(nid);
1712 * We must have raced with another thread. Since we
1713 * know next to nothing about a peer_ni created by
1714 * traffic, we just assume everything is ok and
1717 lnet_peer_ni_decref_locked(lpni);
1721 /* Create peer, peer_net, and peer_ni. */
1723 lp = lnet_peer_alloc(nid);
1726 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1729 lpni = lnet_peer_ni_alloc(nid);
1732 if (pref != LNET_NID_ANY)
1733 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1735 return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1738 LIBCFS_FREE(lpn, sizeof(*lpn));
1740 LIBCFS_FREE(lp, sizeof(*lp));
1742 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1747 * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1749 * This API handles the following combinations:
1750 * Create a peer with its primary NI if only the prim_nid is provided
1751 * Add a NID to a peer identified by the prim_nid. The peer identified
1752 * by the prim_nid must already exist.
1753 * The peer being created may be non-MR.
1755 * The caller must hold ln_api_mutex. This prevents the peer from
1756 * being created/modified/deleted by a different thread.
1759 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1761 struct lnet_peer *lp = NULL;
1762 struct lnet_peer_ni *lpni;
1765 /* The prim_nid must always be specified */
1766 if (prim_nid == LNET_NID_ANY)
1769 flags = LNET_PEER_CONFIGURED;
1771 flags |= LNET_PEER_MULTI_RAIL;
1774 * If nid isn't specified, we must create a new peer with
1775 * prim_nid as its primary nid.
1777 if (nid == LNET_NID_ANY)
1778 return lnet_peer_add(prim_nid, flags);
1780 /* Look up the prim_nid, which must exist. */
1781 lpni = lnet_find_peer_ni_locked(prim_nid);
1784 lnet_peer_ni_decref_locked(lpni);
1785 lp = lpni->lpni_peer_net->lpn_peer;
1787 /* Peer must have been configured. */
1788 if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1789 CDEBUG(D_NET, "peer %s was not configured\n",
1790 libcfs_nid2str(prim_nid));
1794 /* Primary NID must match */
1795 if (lp->lp_primary_nid != prim_nid) {
1796 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1797 libcfs_nid2str(prim_nid),
1798 libcfs_nid2str(lp->lp_primary_nid));
1802 /* Multi-Rail flag must match. */
1803 if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1804 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1805 libcfs_nid2str(prim_nid));
1809 return lnet_peer_add_nid(lp, nid, flags);
1813 * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1815 * This API handles the following combinations:
1816 * Delete a NI from a peer if both prim_nid and nid are provided.
1817 * Delete a peer if only prim_nid is provided.
1818 * Delete a peer if its primary nid is provided.
1820 * The caller must hold ln_api_mutex. This prevents the peer from
1821 * being modified/deleted by a different thread.
1824 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1826 struct lnet_peer *lp;
1827 struct lnet_peer_ni *lpni;
1830 if (prim_nid == LNET_NID_ANY)
1833 lpni = lnet_find_peer_ni_locked(prim_nid);
1836 lnet_peer_ni_decref_locked(lpni);
1837 lp = lpni->lpni_peer_net->lpn_peer;
1839 if (prim_nid != lp->lp_primary_nid) {
1840 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1841 libcfs_nid2str(prim_nid),
1842 libcfs_nid2str(lp->lp_primary_nid));
1846 lnet_net_lock(LNET_LOCK_EX);
1847 if (lp->lp_rtr_refcount > 0) {
1848 lnet_net_unlock(LNET_LOCK_EX);
1849 CERROR("%s is a router. Can not be deleted\n",
1850 libcfs_nid2str(prim_nid));
1853 lnet_net_unlock(LNET_LOCK_EX);
1855 if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1856 return lnet_peer_del(lp);
1858 flags = LNET_PEER_CONFIGURED;
1859 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1860 flags |= LNET_PEER_MULTI_RAIL;
1862 return lnet_peer_del_nid(lp, nid, flags);
1866 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1868 struct lnet_peer_table *ptable;
1869 struct lnet_peer_net *lpn;
1871 CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1873 LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1874 LASSERT(list_empty(&lpni->lpni_txq));
1875 LASSERT(lpni->lpni_txqnob == 0);
1876 LASSERT(list_empty(&lpni->lpni_peer_nis));
1877 LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1879 lpn = lpni->lpni_peer_net;
1880 lpni->lpni_peer_net = NULL;
1881 lpni->lpni_net = NULL;
1883 if (!list_empty(&lpni->lpni_hashlist)) {
1884 /* remove the peer ni from the zombie list */
1885 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1886 spin_lock(&ptable->pt_zombie_lock);
1887 list_del_init(&lpni->lpni_hashlist);
1888 ptable->pt_zombies--;
1889 spin_unlock(&ptable->pt_zombie_lock);
1892 if (lpni->lpni_pref_nnids > 1) {
1893 struct lnet_nid_list *ne, *tmp;
1895 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
1897 list_del_init(&ne->nl_list);
1898 LIBCFS_FREE(ne, sizeof(*ne));
1901 LIBCFS_FREE(lpni, sizeof(*lpni));
1904 lnet_peer_net_decref_locked(lpn);
1907 struct lnet_peer_ni *
1908 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1910 struct lnet_peer_ni *lpni = NULL;
1913 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1914 return ERR_PTR(-ESHUTDOWN);
1917 * find if a peer_ni already exists.
1918 * If so then just return that.
1920 lpni = lnet_find_peer_ni_locked(nid);
1924 lnet_net_unlock(cpt);
1926 rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1929 goto out_net_relock;
1932 lpni = lnet_find_peer_ni_locked(nid);
1942 * Get a peer_ni for the given nid, create it if necessary. Takes a
1943 * hold on the peer_ni.
1945 struct lnet_peer_ni *
1946 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1948 struct lnet_peer_ni *lpni = NULL;
1951 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1952 return ERR_PTR(-ESHUTDOWN);
1955 * find if a peer_ni already exists.
1956 * If so then just return that.
1958 lpni = lnet_find_peer_ni_locked(nid);
1964 * use the lnet_api_mutex to serialize the creation of the peer_ni
1965 * and the creation/deletion of the local ni/net. When a local ni is
1966 * created, if there exists a set of peer_nis on that network,
1967 * they need to be traversed and updated. When a local NI is
1968 * deleted, which could result in a network being deleted, then
1969 * all peer nis on that network need to be removed as well.
1971 * Creation through traffic should also be serialized with
1972 * creation through DLC.
1974 lnet_net_unlock(cpt);
1975 mutex_lock(&the_lnet.ln_api_mutex);
1977 * Shutdown is only set under the ln_api_lock, so a single
1978 * check here is sufficent.
1980 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1981 lpni = ERR_PTR(-ESHUTDOWN);
1982 goto out_mutex_unlock;
1985 rc = lnet_peer_ni_traffic_add(nid, pref);
1988 goto out_mutex_unlock;
1991 lpni = lnet_find_peer_ni_locked(nid);
1995 mutex_unlock(&the_lnet.ln_api_mutex);
1998 /* Lock has been dropped, check again for shutdown. */
1999 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2001 lnet_peer_ni_decref_locked(lpni);
2002 lpni = ERR_PTR(-ESHUTDOWN);
2009 lnet_peer_gw_discovery(struct lnet_peer *lp)
2013 spin_lock(&lp->lp_lock);
2014 if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2016 spin_unlock(&lp->lp_lock);
2022 lnet_peer_is_uptodate(struct lnet_peer *lp)
2026 spin_lock(&lp->lp_lock);
2027 rc = lnet_peer_is_uptodate_locked(lp);
2028 spin_unlock(&lp->lp_lock);
2033 * Is a peer uptodate from the point of view of discovery?
2035 * If it is currently being processed, obviously not.
2036 * A forced Ping or Push is also handled by the discovery thread.
2038 * Otherwise look at whether the peer needs rediscovering.
2041 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2042 __must_hold(&lp->lp_lock)
2046 if (lp->lp_state & (LNET_PEER_DISCOVERING |
2047 LNET_PEER_FORCE_PING |
2048 LNET_PEER_FORCE_PUSH)) {
2050 } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2052 } else if (lnet_peer_needs_push(lp)) {
2054 } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2055 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2067 * Queue a peer for the attention of the discovery thread. Call with
2068 * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2069 * -EALREADY if the peer was already queued.
2071 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2075 spin_lock(&lp->lp_lock);
2076 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2077 lp->lp_state |= LNET_PEER_DISCOVERING;
2078 spin_unlock(&lp->lp_lock);
2079 if (list_empty(&lp->lp_dc_list)) {
2080 lnet_peer_addref_locked(lp);
2081 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2082 wake_up(&the_lnet.ln_dc_waitq);
2088 CDEBUG(D_NET, "Queue peer %s: %d\n",
2089 libcfs_nid2str(lp->lp_primary_nid), rc);
2095 * Discovery of a peer is complete. Wake all waiters on the peer.
2096 * Call with lnet_net_lock/EX held.
2098 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
2100 struct lnet_msg *msg, *tmp;
2102 LIST_HEAD(pending_msgs);
2104 CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2105 libcfs_nid2str(lp->lp_primary_nid));
2107 list_del_init(&lp->lp_dc_list);
2108 spin_lock(&lp->lp_lock);
2109 list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2110 spin_unlock(&lp->lp_lock);
2111 wake_up_all(&lp->lp_dc_waitq);
2113 if (lp->lp_rtr_refcount > 0)
2114 lnet_router_discovery_complete(lp);
2116 lnet_net_unlock(LNET_LOCK_EX);
2118 /* iterate through all pending messages and send them again */
2119 list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2120 list_del_init(&msg->msg_list);
2121 if (lp->lp_dc_error) {
2122 lnet_finalize(msg, lp->lp_dc_error);
2126 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2127 lnet_msgtyp2str(msg->msg_type),
2128 libcfs_id2str(msg->msg_target));
2129 rc = lnet_send(msg->msg_src_nid_param, msg,
2130 msg->msg_rtr_nid_param);
2132 CNETERR("Error sending %s to %s: %d\n",
2133 lnet_msgtyp2str(msg->msg_type),
2134 libcfs_id2str(msg->msg_target), rc);
2135 lnet_finalize(msg, rc);
2138 lnet_net_lock(LNET_LOCK_EX);
2139 lnet_peer_decref_locked(lp);
2143 * Handle inbound push.
2144 * Like any event handler, called with lnet_res_lock/CPT held.
2146 void lnet_peer_push_event(struct lnet_event *ev)
2148 struct lnet_ping_buffer *pbuf;
2149 struct lnet_peer *lp;
2151 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2153 /* lnet_find_peer() adds a refcount */
2154 lp = lnet_find_peer(ev->source.nid);
2156 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2157 libcfs_nid2str(ev->initiator.nid),
2158 libcfs_nid2str(ev->source.nid));
2159 pbuf->pb_needs_post = true;
2163 /* Ensure peer state remains consistent while we modify it. */
2164 spin_lock(&lp->lp_lock);
2167 * If some kind of error happened the contents of the message
2168 * cannot be used. Clear the NIDS_UPTODATE and set the
2169 * FORCE_PING flag to trigger a ping.
2172 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2173 lp->lp_state |= LNET_PEER_FORCE_PING;
2174 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2176 libcfs_nid2str(lp->lp_primary_nid),
2177 libcfs_nid2str(ev->source.nid));
2182 * A push with invalid or corrupted info. Clear the UPTODATE
2183 * flag to trigger a ping.
2185 if (lnet_ping_info_validate(&pbuf->pb_info)) {
2186 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2187 lp->lp_state |= LNET_PEER_FORCE_PING;
2188 CDEBUG(D_NET, "Corrupted Push from %s\n",
2189 libcfs_nid2str(lp->lp_primary_nid));
2194 * Make sure we'll allocate the correct size ping buffer when
2197 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2198 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2201 * A non-Multi-Rail peer is not supposed to be capable of
2204 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2205 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2206 libcfs_nid2str(lp->lp_primary_nid));
2211 * The peer may have discovery disabled at its end. Set
2212 * NO_DISCOVERY as appropriate.
2214 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2215 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2216 libcfs_nid2str(lp->lp_primary_nid));
2218 * Mark the peer for deletion if we already know about it
2219 * and it's going from discovery set to no discovery set
2221 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2222 LNET_PEER_DISCOVERING)) &&
2223 lp->lp_state & LNET_PEER_DISCOVERED) {
2224 CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2225 libcfs_nid2str(lp->lp_primary_nid),
2227 lp->lp_state |= LNET_PEER_MARK_DELETION;
2229 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2230 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2231 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2232 libcfs_nid2str(lp->lp_primary_nid));
2233 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2237 * Update the MULTI_RAIL flag based on the push. If the peer
2238 * was configured with DLC then the setting should match what
2240 * NB: We verified above that the MR feature bit is set in pi_features
2242 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2243 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2244 libcfs_nid2str(lp->lp_primary_nid), lp);
2245 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2246 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2247 libcfs_nid2str(lp->lp_primary_nid));
2248 } else if (lnet_peer_discovery_disabled) {
2249 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2250 libcfs_nid2str(lp->lp_primary_nid), lp);
2251 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2252 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2253 libcfs_nid2str(lp->lp_primary_nid), lp);
2255 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2256 libcfs_nid2str(lp->lp_primary_nid), lp);
2257 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2258 lnet_peer_clr_non_mr_pref_nids(lp);
2262 * Check for truncation of the Put message. Clear the
2263 * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2264 * and tell discovery to allocate a bigger buffer.
2266 if (ev->mlength < ev->rlength) {
2267 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2268 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2269 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2270 lp->lp_state |= LNET_PEER_FORCE_PING;
2271 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2272 libcfs_nid2str(lp->lp_primary_nid),
2273 pbuf->pb_info.pi_nnis);
2277 /* always assume new data */
2278 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2279 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2282 * If there is data present that hasn't been processed yet,
2283 * we'll replace it if the Put contained newer data and it
2284 * fits. We're racing with a Ping or earlier Push in this
2287 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2288 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2289 LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2290 pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2291 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2292 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2293 CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2294 libcfs_nid2str(lp->lp_primary_nid),
2295 LNET_PING_BUFFER_SEQNO(pbuf),
2296 LNET_PING_BUFFER_SEQNO(lp->lp_data));
2302 * Allocate a buffer to copy the data. On a failure we drop
2303 * the Push and set FORCE_PING to force the discovery
2304 * thread to fix the problem by pinging the peer.
2306 lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2308 lp->lp_state |= LNET_PEER_FORCE_PING;
2309 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2310 libcfs_nid2str(lp->lp_primary_nid),
2311 LNET_PING_BUFFER_SEQNO(pbuf));
2316 memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2317 LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2318 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2319 CDEBUG(D_NET, "Received Push %s %u\n",
2320 libcfs_nid2str(lp->lp_primary_nid),
2321 LNET_PING_BUFFER_SEQNO(pbuf));
2324 /* We've processed this buffer. It can be reposted */
2325 pbuf->pb_needs_post = true;
2328 * Queue the peer for discovery if not done, force it on the request
2329 * queue and wake the discovery thread if the peer was already queued,
2330 * because its status changed.
2332 spin_unlock(&lp->lp_lock);
2333 lnet_net_lock(LNET_LOCK_EX);
2334 if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2335 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2336 wake_up(&the_lnet.ln_dc_waitq);
2338 /* Drop refcount from lookup */
2339 lnet_peer_decref_locked(lp);
2340 lnet_net_unlock(LNET_LOCK_EX);
2344 * Clear the discovery error state, unless we're already discovering
2345 * this peer, in which case the error is current.
2347 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2349 spin_lock(&lp->lp_lock);
2350 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2351 lp->lp_dc_error = 0;
2352 spin_unlock(&lp->lp_lock);
2356 * Peer discovery slow path. The ln_api_mutex is held on entry, and
2357 * dropped/retaken within this function. An lnet_peer_ni is passed in
2358 * because discovery could tear down an lnet_peer.
2361 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2364 struct lnet_peer *lp;
2369 lnet_net_unlock(cpt);
2370 lnet_net_lock(LNET_LOCK_EX);
2371 lp = lpni->lpni_peer_net->lpn_peer;
2372 lnet_peer_clear_discovery_error(lp);
2375 * We're willing to be interrupted. The lpni can become a
2376 * zombie if we race with DLC, so we must check for that.
2379 /* Keep lp alive when the lnet_net_lock is unlocked */
2380 lnet_peer_addref_locked(lp);
2381 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2382 if (signal_pending(current))
2384 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2387 * Don't repeat discovery if discovery is disabled. This is
2388 * done to ensure we can use discovery as a standard ping as
2389 * well for backwards compatibility with routers which do not
2390 * have discovery or have discovery disabled
2392 if (lnet_is_discovery_disabled(lp) && count > 0)
2394 if (lp->lp_dc_error)
2396 if (lnet_peer_is_uptodate(lp))
2398 lnet_peer_queue_for_discovery(lp);
2400 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2403 * If caller requested a non-blocking operation then
2404 * return immediately. Once discovery is complete any
2405 * pending messages that were stopped due to discovery
2406 * will be transmitted.
2411 lnet_net_unlock(LNET_LOCK_EX);
2413 finish_wait(&lp->lp_dc_waitq, &wait);
2414 lnet_net_lock(LNET_LOCK_EX);
2415 lnet_peer_decref_locked(lp);
2416 /* Peer may have changed */
2417 lp = lpni->lpni_peer_net->lpn_peer;
2419 finish_wait(&lp->lp_dc_waitq, &wait);
2421 lnet_net_unlock(LNET_LOCK_EX);
2423 lnet_peer_decref_locked(lp);
2425 * The peer may have changed, so re-check and rediscover if that turns
2426 * out to have been the case. The reference count on lp ensured that
2427 * even if it was unlinked from lpni the memory could not be recycled.
2428 * Thus the check below is sufficient to determine whether the peer
2429 * changed. If the peer changed, then lp must not be dereferenced.
2431 if (lp != lpni->lpni_peer_net->lpn_peer)
2434 if (signal_pending(current))
2436 else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2438 else if (lp->lp_dc_error)
2439 rc = lp->lp_dc_error;
2441 CDEBUG(D_NET, "non-blocking discovery\n");
2442 else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2445 CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2446 (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2447 libcfs_nid2str(lpni->lpni_nid), rc,
2448 (!block) ? "pending discovery" : "discovery complete");
2453 /* Handle an incoming ack for a push. */
2455 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2457 struct lnet_ping_buffer *pbuf;
2459 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2460 spin_lock(&lp->lp_lock);
2461 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2462 lp->lp_push_error = ev->status;
2464 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2466 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2467 spin_unlock(&lp->lp_lock);
2469 CDEBUG(D_NET, "peer %s ev->status %d\n",
2470 libcfs_nid2str(lp->lp_primary_nid), ev->status);
2473 /* Handle a Reply message. This is the reply to a Ping message. */
2475 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2477 struct lnet_ping_buffer *pbuf;
2480 spin_lock(&lp->lp_lock);
2482 lp->lp_disc_src_nid = ev->target.nid;
2485 * If some kind of error happened the contents of message
2486 * cannot be used. Set PING_FAILED to trigger a retry.
2489 lp->lp_state |= LNET_PEER_PING_FAILED;
2490 lp->lp_ping_error = ev->status;
2491 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2493 libcfs_nid2str(lp->lp_primary_nid),
2494 libcfs_nid2str(ev->source.nid));
2498 pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2499 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2500 lnet_swap_pinginfo(pbuf);
2503 * A reply with invalid or corrupted info. Set PING_FAILED to
2506 rc = lnet_ping_info_validate(&pbuf->pb_info);
2508 lp->lp_state |= LNET_PEER_PING_FAILED;
2509 lp->lp_ping_error = 0;
2510 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2511 libcfs_nid2str(lp->lp_primary_nid), rc);
2517 * The peer may have discovery disabled at its end. Set
2518 * NO_DISCOVERY as appropriate.
2520 if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2521 !lnet_peer_discovery_disabled) {
2522 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2523 libcfs_nid2str(lp->lp_primary_nid));
2524 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2526 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2527 libcfs_nid2str(lp->lp_primary_nid));
2528 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2532 * Update the MULTI_RAIL flag based on the reply. If the peer
2533 * was configured with DLC then the setting should match what
2536 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2537 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2538 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2539 libcfs_nid2str(lp->lp_primary_nid), lp);
2540 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2541 CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2542 libcfs_nid2str(lp->lp_primary_nid));
2543 } else if (lnet_peer_discovery_disabled) {
2545 "peer %s(%p) not MR: DD disabled locally\n",
2546 libcfs_nid2str(lp->lp_primary_nid), lp);
2547 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2549 "peer %s(%p) not MR: DD disabled remotely\n",
2550 libcfs_nid2str(lp->lp_primary_nid), lp);
2552 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2553 libcfs_nid2str(lp->lp_primary_nid), lp);
2554 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2555 lnet_peer_clr_non_mr_pref_nids(lp);
2557 } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2558 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2559 CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2560 libcfs_nid2str(lp->lp_primary_nid));
2562 CERROR("Multi-Rail state vanished from %s\n",
2563 libcfs_nid2str(lp->lp_primary_nid));
2564 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2569 * Make sure we'll allocate the correct size ping buffer when
2572 if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2573 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2576 * Check for truncation of the Reply. Clear PING_SENT and set
2577 * PING_FAILED to trigger a retry.
2579 if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2580 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2581 the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2582 lp->lp_state |= LNET_PEER_PING_FAILED;
2583 lp->lp_ping_error = 0;
2584 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2585 libcfs_nid2str(lp->lp_primary_nid),
2586 pbuf->pb_info.pi_nnis);
2591 * Check the sequence numbers in the reply. These are only
2592 * available if the reply came from a Multi-Rail peer.
2594 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2595 pbuf->pb_info.pi_nnis > 1 &&
2596 lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2597 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2598 CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2599 libcfs_nid2str(lp->lp_primary_nid),
2600 LNET_PING_BUFFER_SEQNO(pbuf),
2603 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2606 /* We're happy with the state of the data in the buffer. */
2607 CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2608 libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2609 if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2610 lnet_ping_buffer_decref(lp->lp_data);
2612 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2613 lnet_ping_buffer_addref(pbuf);
2616 lp->lp_state &= ~LNET_PEER_PING_SENT;
2617 spin_unlock(&lp->lp_lock);
2619 lnet_net_lock(LNET_LOCK_EX);
2621 * If this peer is a gateway, call the routing callback to
2622 * handle the ping reply
2624 if (lp->lp_rtr_refcount > 0)
2625 lnet_router_discovery_ping_reply(lp);
2626 lnet_net_unlock(LNET_LOCK_EX);
2630 * Send event handling. Only matters for error cases, where we clean
2631 * up state on the peer and peer_ni that would otherwise be updated in
2632 * the REPLY event handler for a successful Ping, and the ACK event
2633 * handler for a successful Push.
2636 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2643 spin_lock(&lp->lp_lock);
2644 if (ev->msg_type == LNET_MSG_GET) {
2645 lp->lp_state &= ~LNET_PEER_PING_SENT;
2646 lp->lp_state |= LNET_PEER_PING_FAILED;
2647 lp->lp_ping_error = ev->status;
2648 } else { /* ev->msg_type == LNET_MSG_PUT */
2649 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2650 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2651 lp->lp_push_error = ev->status;
2653 spin_unlock(&lp->lp_lock);
2654 rc = LNET_REDISCOVER_PEER;
2656 CDEBUG(D_NET, "%s Send to %s: %d\n",
2657 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2658 libcfs_nid2str(ev->target.nid), rc);
2663 * Unlink event handling. This event is only seen if a call to
2664 * LNetMDUnlink() caused the event to be unlinked. If this call was
2665 * made after the event was set up in LNetGet() or LNetPut() then we
2666 * assume the Ping or Push timed out.
2669 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2671 spin_lock(&lp->lp_lock);
2672 /* We've passed through LNetGet() */
2673 if (lp->lp_state & LNET_PEER_PING_SENT) {
2674 lp->lp_state &= ~LNET_PEER_PING_SENT;
2675 lp->lp_state |= LNET_PEER_PING_FAILED;
2676 lp->lp_ping_error = -ETIMEDOUT;
2677 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2678 libcfs_nid2str(lp->lp_primary_nid));
2680 /* We've passed through LNetPut() */
2681 if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2682 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2683 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2684 lp->lp_push_error = -ETIMEDOUT;
2685 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2686 libcfs_nid2str(lp->lp_primary_nid));
2688 spin_unlock(&lp->lp_lock);
2692 * Event handler for the discovery EQ.
2694 * Called with lnet_res_lock(cpt) held. The cpt is the
2695 * lnet_cpt_of_cookie() of the md handle cookie.
2697 static void lnet_discovery_event_handler(struct lnet_event *event)
2699 struct lnet_peer *lp = event->md_user_ptr;
2700 struct lnet_ping_buffer *pbuf;
2703 /* discovery needs to take another look */
2704 rc = LNET_REDISCOVER_PEER;
2706 CDEBUG(D_NET, "Received event: %d\n", event->type);
2708 switch (event->type) {
2709 case LNET_EVENT_ACK:
2710 lnet_discovery_event_ack(lp, event);
2712 case LNET_EVENT_REPLY:
2713 lnet_discovery_event_reply(lp, event);
2715 case LNET_EVENT_SEND:
2716 /* Only send failure triggers a retry. */
2717 rc = lnet_discovery_event_send(lp, event);
2719 case LNET_EVENT_UNLINK:
2720 /* LNetMDUnlink() was called */
2721 lnet_discovery_event_unlink(lp, event);
2724 /* Invalid events. */
2727 lnet_net_lock(LNET_LOCK_EX);
2728 if (event->unlinked) {
2729 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2730 lnet_ping_buffer_decref(pbuf);
2731 lnet_peer_decref_locked(lp);
2734 /* put peer back at end of request queue, if discovery not already
2736 if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2737 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2738 wake_up(&the_lnet.ln_dc_waitq);
2740 lnet_net_unlock(LNET_LOCK_EX);
2744 * Build a peer from incoming data.
2746 * The NIDs in the incoming data are supposed to be structured as follows:
2749 * - other NIDs in same net
2750 * - NIDs in second net
2751 * - NIDs in third net
2753 * This due to the way the list of NIDs in the data is created.
2755 * Note that this function will mark the peer uptodate unless an
2756 * ENOMEM is encontered. All other errors are due to a conflict
2757 * between the DLC configuration and what discovery sees. We treat DLC
2758 * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2759 * peer from becoming stuck in discovery.
2761 static int lnet_peer_merge_data(struct lnet_peer *lp,
2762 struct lnet_ping_buffer *pbuf)
2764 struct lnet_peer_ni *lpni;
2765 lnet_nid_t *curnis = NULL;
2766 struct lnet_ni_status *addnis = NULL;
2767 lnet_nid_t *delnis = NULL;
2777 flags = LNET_PEER_DISCOVERED;
2778 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2779 flags |= LNET_PEER_MULTI_RAIL;
2782 * Cache the routing feature for the peer; whether it is enabled
2783 * for disabled as reported by the remote peer.
2785 spin_lock(&lp->lp_lock);
2786 if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2787 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2789 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2790 spin_unlock(&lp->lp_lock);
2792 nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2793 CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2794 CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2795 CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2796 if (!curnis || !addnis || !delnis) {
2804 /* Construct the list of NIDs present in peer. */
2806 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2807 curnis[ncurnis++] = lpni->lpni_nid;
2810 * Check for NIDs in pbuf not present in curnis[].
2811 * The loop starts at 1 to skip the loopback NID.
2813 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2814 for (j = 0; j < ncurnis; j++)
2815 if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2818 addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2821 * Check for NIDs in curnis[] not present in pbuf.
2822 * The nested loop starts at 1 to skip the loopback NID.
2824 * But never add the loopback NID to delnis[]: if it is
2825 * present in curnis[] then this peer is for this node.
2827 for (i = 0; i < ncurnis; i++) {
2828 if (curnis[i] == LNET_NID_LO_0)
2830 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2831 if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2833 * update the information we cache for the
2834 * peer with the latest information we
2837 lpni = lnet_find_peer_ni_locked(curnis[i]);
2839 lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2840 lnet_peer_ni_decref_locked(lpni);
2845 if (j == pbuf->pb_info.pi_nnis)
2846 delnis[ndelnis++] = curnis[i];
2850 * If we get here and the discovery is disabled then we don't want
2851 * to add or delete any NIs. We just updated the ones we have some
2852 * information on, and call it a day
2855 if (lnet_is_discovery_disabled(lp))
2858 for (i = 0; i < naddnis; i++) {
2859 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2861 CERROR("Error adding NID %s to peer %s: %d\n",
2862 libcfs_nid2str(addnis[i].ns_nid),
2863 libcfs_nid2str(lp->lp_primary_nid), rc);
2867 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2869 lpni->lpni_ns_status = addnis[i].ns_status;
2870 lnet_peer_ni_decref_locked(lpni);
2874 for (i = 0; i < ndelnis; i++) {
2876 * for routers it's okay to delete the primary_nid because
2877 * the upper layers don't really rely on it. So if we're
2878 * being told that the router changed its primary_nid
2879 * then it's okay to delete it.
2881 if (lp->lp_rtr_refcount > 0)
2882 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2883 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2885 CERROR("Error deleting NID %s from peer %s: %d\n",
2886 libcfs_nid2str(delnis[i]),
2887 libcfs_nid2str(lp->lp_primary_nid), rc);
2893 * Errors other than -ENOMEM are due to peers having been
2894 * configured with DLC. Ignore these because DLC overrides
2899 CFS_FREE_PTR_ARRAY(curnis, nnis);
2900 CFS_FREE_PTR_ARRAY(addnis, nnis);
2901 CFS_FREE_PTR_ARRAY(delnis, nnis);
2902 lnet_ping_buffer_decref(pbuf);
2903 CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2906 spin_lock(&lp->lp_lock);
2907 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2908 lp->lp_state |= LNET_PEER_FORCE_PING;
2909 spin_unlock(&lp->lp_lock);
2915 * The data in pbuf says lp is its primary peer, but the data was
2916 * received by a different peer. Try to update lp with the data.
2919 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2921 struct lnet_handle_md mdh;
2923 /* Queue lp for discovery, and force it on the request queue. */
2924 lnet_net_lock(LNET_LOCK_EX);
2925 if (lnet_peer_queue_for_discovery(lp))
2926 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2927 lnet_net_unlock(LNET_LOCK_EX);
2929 LNetInvalidateMDHandle(&mdh);
2932 * Decide whether we can move the peer to the DATA_PRESENT state.
2934 * We replace stale data for a multi-rail peer, repair PING_FAILED
2935 * status, and preempt FORCE_PING.
2937 * If after that we have DATA_PRESENT, we merge it into this peer.
2939 spin_lock(&lp->lp_lock);
2940 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2941 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2942 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2943 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2944 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2945 lnet_ping_buffer_decref(pbuf);
2950 if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2951 lnet_ping_buffer_decref(lp->lp_data);
2953 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2955 if (lp->lp_state & LNET_PEER_PING_FAILED) {
2956 mdh = lp->lp_ping_mdh;
2957 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2958 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2959 lp->lp_ping_error = 0;
2961 if (lp->lp_state & LNET_PEER_FORCE_PING)
2962 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2963 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2964 spin_unlock(&lp->lp_lock);
2966 if (!LNetMDHandleIsInvalid(mdh))
2970 return lnet_peer_merge_data(lp, pbuf);
2972 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2976 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2980 for (i = 0; i < pinfo->pi_nnis; i++) {
2981 if (pinfo->pi_ni[i].ns_nid == nid)
2988 /* Delete a peer that has been marked for deletion. NB: when this peer was added
2989 * to the discovery queue a reference was taken that will prevent the peer from
2990 * actually being freed by this function. After this function exits the
2991 * discovery thread should call lnet_peer_discovery_complete() which will
2992 * drop that reference as well as wake any waiters that may also be holding a
2995 static int lnet_peer_deletion(struct lnet_peer *lp)
2996 __must_hold(&lp->lp_lock)
2998 struct list_head rlist;
2999 struct lnet_route *route, *tmp;
3000 int sensitivity = lp->lp_health_sensitivity;
3002 INIT_LIST_HEAD(&rlist);
3004 lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3005 LNET_PEER_FORCE_PUSH);
3006 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3007 libcfs_nid2str(lp->lp_primary_nid), lp, lp->lp_state);
3009 /* no-op if lnet_peer_del() has already been called on this peer */
3010 if (lp->lp_state & LNET_PEER_MARK_DELETED)
3013 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
3016 spin_unlock(&lp->lp_lock);
3018 mutex_lock(&the_lnet.ln_api_mutex);
3020 lnet_net_lock(LNET_LOCK_EX);
3021 /* remove the peer from the discovery work
3022 * queue if it's on there in preparation
3025 if (!list_empty(&lp->lp_dc_list))
3026 list_del(&lp->lp_dc_list);
3027 list_for_each_entry_safe(route, tmp,
3030 lnet_move_route(route, NULL, &rlist);
3031 lnet_net_unlock(LNET_LOCK_EX);
3033 /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3036 list_for_each_entry_safe(route, tmp,
3038 /* re-add these routes */
3039 lnet_add_route(route->lr_net,
3044 LIBCFS_FREE(route, sizeof(*route));
3047 mutex_unlock(&the_lnet.ln_api_mutex);
3049 spin_lock(&lp->lp_lock);
3055 * Update a peer using the data received.
3057 static int lnet_peer_data_present(struct lnet_peer *lp)
3058 __must_hold(&lp->lp_lock)
3060 struct lnet_ping_buffer *pbuf;
3061 struct lnet_peer_ni *lpni;
3062 lnet_nid_t nid = LNET_NID_ANY;
3068 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3069 lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3070 spin_unlock(&lp->lp_lock);
3073 * Modifications of peer structures are done while holding the
3074 * ln_api_mutex. A global lock is required because we may be
3075 * modifying multiple peer structures, and a mutex greatly
3076 * simplifies memory management.
3078 * The actual changes to the data structures must also protect
3079 * against concurrent lookups, for which the lnet_net_lock in
3080 * LNET_LOCK_EX mode is used.
3082 mutex_lock(&the_lnet.ln_api_mutex);
3083 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3089 * If this peer is not on the peer list then it is being torn
3090 * down, and our reference count may be all that is keeping it
3091 * alive. Don't do any work on it.
3093 if (list_empty(&lp->lp_peer_list))
3096 flags = LNET_PEER_DISCOVERED;
3097 if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3098 flags |= LNET_PEER_MULTI_RAIL;
3101 * Check whether the primary NID in the message matches the
3102 * primary NID of the peer. If it does, update the peer, if
3103 * it it does not, check whether there is already a peer with
3104 * that primary NID. If no such peer exists, try to update
3105 * the primary NID of the current peer (allowed if it was
3106 * created due to message traffic) and complete the update.
3107 * If the peer did exist, hand off the data to it.
3109 * The peer for the loopback interface is a special case: this
3110 * is the peer for the local node, and we want to set its
3111 * primary NID to the correct value here. Moreover, this peer
3112 * can show up with only the loopback NID in the ping buffer.
3114 if (pbuf->pb_info.pi_nnis <= 1)
3116 nid = pbuf->pb_info.pi_ni[1].ns_nid;
3117 if (lp->lp_primary_nid == LNET_NID_LO_0) {
3118 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3120 rc = lnet_peer_merge_data(lp, pbuf);
3122 * if the primary nid of the peer is present in the ping info returned
3123 * from the peer, but it's not the local primary peer we have
3124 * cached and discovery is disabled, then we don't want to update
3125 * our local peer info, by adding or removing NIDs, we just want
3126 * to update the status of the nids that we currently have
3127 * recorded in that peer.
3129 } else if (lp->lp_primary_nid == nid ||
3130 (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
3131 lnet_is_discovery_disabled(lp))) {
3132 rc = lnet_peer_merge_data(lp, pbuf);
3134 lpni = lnet_find_peer_ni_locked(nid);
3136 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3138 CERROR("Primary NID error %s versus %s: %d\n",
3139 libcfs_nid2str(lp->lp_primary_nid),
3140 libcfs_nid2str(nid), rc);
3142 rc = lnet_peer_merge_data(lp, pbuf);
3145 struct lnet_peer *new_lp;
3146 new_lp = lpni->lpni_peer_net->lpn_peer;
3148 * if lp has discovery/MR enabled that means new_lp
3149 * should have discovery/MR enabled as well, since
3150 * it's the same peer, which we're about to merge
3152 if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3153 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3154 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3155 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3157 rc = lnet_peer_set_primary_data(new_lp, pbuf);
3158 lnet_consolidate_routes_locked(lp, new_lp);
3159 lnet_peer_ni_decref_locked(lpni);
3163 CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
3165 mutex_unlock(&the_lnet.ln_api_mutex);
3167 spin_lock(&lp->lp_lock);
3168 /* Tell discovery to re-check the peer immediately. */
3170 rc = LNET_REDISCOVER_PEER;
3175 * A ping failed. Clear the PING_FAILED state and set the
3176 * FORCE_PING state, to ensure a retry even if discovery is
3177 * disabled. This avoids being left with incorrect state.
3179 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3180 __must_hold(&lp->lp_lock)
3182 struct lnet_handle_md mdh;
3185 mdh = lp->lp_ping_mdh;
3186 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3187 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3188 lp->lp_state |= LNET_PEER_FORCE_PING;
3189 rc = lp->lp_ping_error;
3190 lp->lp_ping_error = 0;
3191 spin_unlock(&lp->lp_lock);
3193 if (!LNetMDHandleIsInvalid(mdh))
3196 CDEBUG(D_NET, "peer %s:%d\n",
3197 libcfs_nid2str(lp->lp_primary_nid), rc);
3199 spin_lock(&lp->lp_lock);
3200 return rc ? rc : LNET_REDISCOVER_PEER;
3204 * Select NID to send a Ping or Push to.
3206 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
3208 struct lnet_peer_ni *lpni;
3210 /* Look for a direct-connected NID for this peer. */
3212 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3213 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
3218 return lpni->lpni_nid;
3220 /* Look for a routed-connected NID for this peer. */
3222 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3223 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
3228 return lpni->lpni_nid;
3230 return LNET_NID_ANY;
3233 /* Active side of ping. */
3234 static int lnet_peer_send_ping(struct lnet_peer *lp)
3235 __must_hold(&lp->lp_lock)
3242 lp->lp_state |= LNET_PEER_PING_SENT;
3243 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3244 spin_unlock(&lp->lp_lock);
3246 cpt = lnet_net_lock_current();
3247 /* Refcount for MD. */
3248 lnet_peer_addref_locked(lp);
3249 pnid = lnet_peer_select_nid(lp);
3250 lnet_net_unlock(cpt);
3252 nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3254 rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
3255 the_lnet.ln_dc_handler, false);
3258 * if LNetMDBind in lnet_send_ping fails we need to decrement the
3259 * refcount on the peer, otherwise LNetMDUnlink will be called
3260 * which will eventually do that.
3264 lnet_peer_decref_locked(lp);
3265 lnet_net_unlock(cpt);
3266 rc = -rc; /* change the rc to negative value */
3268 } else if (rc < 0) {
3272 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3274 spin_lock(&lp->lp_lock);
3278 CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3280 * The errors that get us here are considered hard errors and
3281 * cause Discovery to terminate. So we clear PING_SENT, but do
3282 * not set either PING_FAILED or FORCE_PING. In fact we need
3283 * to clear PING_FAILED, because the unlink event handler will
3284 * have set it if we called LNetMDUnlink() above.
3286 spin_lock(&lp->lp_lock);
3287 lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3292 * This function exists because you cannot call LNetMDUnlink() from an
3295 static int lnet_peer_push_failed(struct lnet_peer *lp)
3296 __must_hold(&lp->lp_lock)
3298 struct lnet_handle_md mdh;
3301 mdh = lp->lp_push_mdh;
3302 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3303 lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3304 rc = lp->lp_push_error;
3305 lp->lp_push_error = 0;
3306 spin_unlock(&lp->lp_lock);
3308 if (!LNetMDHandleIsInvalid(mdh))
3311 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3312 spin_lock(&lp->lp_lock);
3313 return rc ? rc : LNET_REDISCOVER_PEER;
3317 * Mark the peer as discovered.
3319 static int lnet_peer_discovered(struct lnet_peer *lp)
3320 __must_hold(&lp->lp_lock)
3322 lp->lp_state |= LNET_PEER_DISCOVERED;
3323 lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3324 LNET_PEER_REDISCOVER);
3326 lp->lp_dc_error = 0;
3328 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3333 /* Active side of push. */
3334 static int lnet_peer_send_push(struct lnet_peer *lp)
3335 __must_hold(&lp->lp_lock)
3337 struct lnet_ping_buffer *pbuf;
3338 struct lnet_process_id id;
3343 /* Don't push to a non-multi-rail peer. */
3344 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3345 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3346 /* if peer's NIDs are uptodate then peer is discovered */
3347 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3348 rc = lnet_peer_discovered(lp);
3355 lp->lp_state |= LNET_PEER_PUSH_SENT;
3356 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3357 spin_unlock(&lp->lp_lock);
3359 cpt = lnet_net_lock_current();
3360 pbuf = the_lnet.ln_ping_target;
3361 lnet_ping_buffer_addref(pbuf);
3362 lnet_net_unlock(cpt);
3364 /* Push source MD */
3365 md.start = &pbuf->pb_info;
3366 md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3367 md.threshold = 2; /* Put/Ack */
3369 md.options = LNET_MD_TRACK_RESPONSE;
3370 md.handler = the_lnet.ln_dc_handler;
3373 rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3375 lnet_ping_buffer_decref(pbuf);
3376 CERROR("Can't bind push source MD: %d\n", rc);
3379 cpt = lnet_net_lock_current();
3380 /* Refcount for MD. */
3381 lnet_peer_addref_locked(lp);
3382 id.pid = LNET_PID_LUSTRE;
3383 id.nid = lnet_peer_select_nid(lp);
3384 lnet_net_unlock(cpt);
3386 if (id.nid == LNET_NID_ANY) {
3391 rc = LNetPut(lp->lp_disc_src_nid, lp->lp_push_mdh,
3392 LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3393 LNET_PROTO_PING_MATCHBITS, 0, 0);
3396 * reset the discovery nid. There is no need to restrict sending
3397 * from that source, if we call lnet_push_update_to_peers(). It'll
3398 * get set to a specific NID, if we initiate discovery from the
3401 lp->lp_disc_src_nid = LNET_NID_ANY;
3406 CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3408 spin_lock(&lp->lp_lock);
3412 LNetMDUnlink(lp->lp_push_mdh);
3413 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3415 CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3417 * The errors that get us here are considered hard errors and
3418 * cause Discovery to terminate. So we clear PUSH_SENT, but do
3419 * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3420 * because the unlink event handler will have set it if we
3421 * called LNetMDUnlink() above.
3423 spin_lock(&lp->lp_lock);
3424 lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3429 * An unrecoverable error was encountered during discovery.
3430 * Set error status in peer and abort discovery.
3432 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3434 CDEBUG(D_NET, "Discovery error %s: %d\n",
3435 libcfs_nid2str(lp->lp_primary_nid), error);
3437 spin_lock(&lp->lp_lock);
3438 lp->lp_dc_error = error;
3439 lp->lp_state &= ~LNET_PEER_DISCOVERING;
3440 lp->lp_state |= LNET_PEER_REDISCOVER;
3441 spin_unlock(&lp->lp_lock);
3445 * Wait for work to be queued or some other change that must be
3446 * attended to. Returns non-zero if the discovery thread should shut
3449 static int lnet_peer_discovery_wait_for_work(void)
3456 cpt = lnet_net_lock_current();
3458 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3459 TASK_INTERRUPTIBLE);
3460 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3462 if (lnet_push_target_resize_needed() ||
3463 the_lnet.ln_push_target->pb_needs_post)
3465 if (!list_empty(&the_lnet.ln_dc_request))
3467 if (!list_empty(&the_lnet.ln_msg_resend))
3469 lnet_net_unlock(cpt);
3472 * wakeup max every second to check if there are peers that
3473 * have been stuck on the working queue for greater than
3476 schedule_timeout(cfs_time_seconds(1));
3477 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3478 cpt = lnet_net_lock_current();
3480 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3482 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3485 lnet_net_unlock(cpt);
3487 CDEBUG(D_NET, "woken: %d\n", rc);
3493 * Messages that were pending on a destroyed peer will be put on a global
3494 * resend list. The message resend list will be checked by
3495 * the discovery thread when it wakes up, and will resend messages. These
3496 * messages can still be sendable in the case the lpni which was the initial
3497 * cause of the message re-queue was transfered to another peer.
3499 * It is possible that LNet could be shutdown while we're iterating
3500 * through the list. lnet_shudown_lndnets() will attempt to access the
3501 * resend list, but will have to wait until the spinlock is released, by
3502 * which time there shouldn't be any more messages on the resend list.
3503 * During shutdown lnet_send() will fail and lnet_finalize() will be called
3504 * for the messages so they can be released. The other case is that
3505 * lnet_shudown_lndnets() can finalize all the messages before this
3506 * function can visit the resend list, in which case this function will be
3509 static void lnet_resend_msgs(void)
3511 struct lnet_msg *msg, *tmp;
3515 spin_lock(&the_lnet.ln_msg_resend_lock);
3516 list_splice(&the_lnet.ln_msg_resend, &resend);
3517 spin_unlock(&the_lnet.ln_msg_resend_lock);
3519 list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3520 list_del_init(&msg->msg_list);
3521 rc = lnet_send(msg->msg_src_nid_param, msg,
3522 msg->msg_rtr_nid_param);
3524 CNETERR("Error sending %s to %s: %d\n",
3525 lnet_msgtyp2str(msg->msg_type),
3526 libcfs_id2str(msg->msg_target), rc);
3527 lnet_finalize(msg, rc);
3532 /* The discovery thread. */
3533 static int lnet_peer_discovery(void *arg)
3535 struct lnet_peer *lp;
3538 wait_for_completion(&the_lnet.ln_started);
3540 CDEBUG(D_NET, "started\n");
3543 if (lnet_peer_discovery_wait_for_work())
3546 if (lnet_push_target_resize_needed())
3547 lnet_push_target_resize();
3548 else if (the_lnet.ln_push_target->pb_needs_post)
3549 lnet_push_target_post(the_lnet.ln_push_target,
3550 &the_lnet.ln_push_target_md);
3554 lnet_net_lock(LNET_LOCK_EX);
3555 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3556 lnet_net_unlock(LNET_LOCK_EX);
3561 * Process all incoming discovery work requests. When
3562 * discovery must wait on a peer to change state, it
3563 * is added to the tail of the ln_dc_working queue. A
3564 * timestamp keeps track of when the peer was added,
3565 * so we can time out discovery requests that take too
3568 while (!list_empty(&the_lnet.ln_dc_request)) {
3569 lp = list_first_entry(&the_lnet.ln_dc_request,
3570 struct lnet_peer, lp_dc_list);
3571 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3573 * set the time the peer was put on the dc_working
3574 * queue. It shouldn't remain on the queue
3575 * forever, in case the GET message (for ping)
3576 * doesn't get a REPLY or the PUT message (for
3577 * push) doesn't get an ACK.
3579 lp->lp_last_queued = ktime_get_real_seconds();
3580 lnet_net_unlock(LNET_LOCK_EX);
3582 if (lnet_push_target_resize_needed())
3583 lnet_push_target_resize();
3584 else if (the_lnet.ln_push_target->pb_needs_post)
3585 lnet_push_target_post(the_lnet.ln_push_target,
3586 &the_lnet.ln_push_target_md);
3589 * Select an action depending on the state of
3590 * the peer and whether discovery is disabled.
3591 * The check whether discovery is disabled is
3592 * done after the code that handles processing
3593 * for arrived data, cleanup for failures, and
3594 * forcing a Ping or Push.
3596 spin_lock(&lp->lp_lock);
3597 CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3598 libcfs_nid2str(lp->lp_primary_nid), lp,
3600 if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3601 LNET_PEER_MARK_DELETED))
3602 rc = lnet_peer_deletion(lp);
3603 else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3604 rc = lnet_peer_data_present(lp);
3605 else if (lp->lp_state & LNET_PEER_PING_FAILED)
3606 rc = lnet_peer_ping_failed(lp);
3607 else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3608 rc = lnet_peer_push_failed(lp);
3609 else if (lp->lp_state & LNET_PEER_FORCE_PING)
3610 rc = lnet_peer_send_ping(lp);
3611 else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3612 rc = lnet_peer_send_push(lp);
3613 else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3614 rc = lnet_peer_send_ping(lp);
3615 else if (lnet_peer_needs_push(lp))
3616 rc = lnet_peer_send_push(lp);
3618 rc = lnet_peer_discovered(lp);
3619 CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3620 libcfs_nid2str(lp->lp_primary_nid), lp,
3622 spin_unlock(&lp->lp_lock);
3624 lnet_net_lock(LNET_LOCK_EX);
3625 if (rc == LNET_REDISCOVER_PEER) {
3626 list_move(&lp->lp_dc_list,
3627 &the_lnet.ln_dc_request);
3629 lnet_peer_discovery_error(lp, rc);
3631 if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3632 lnet_peer_discovery_complete(lp);
3633 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3638 lnet_net_unlock(LNET_LOCK_EX);
3641 CDEBUG(D_NET, "stopping\n");
3643 * Clean up before telling lnet_peer_discovery_stop() that
3644 * we're done. Use wake_up() below to somewhat reduce the
3645 * size of the thundering herd if there are multiple threads
3646 * waiting on discovery of a single peer.
3649 /* Queue cleanup 1: stop all pending pings and pushes. */
3650 lnet_net_lock(LNET_LOCK_EX);
3651 while (!list_empty(&the_lnet.ln_dc_working)) {
3652 lp = list_first_entry(&the_lnet.ln_dc_working,
3653 struct lnet_peer, lp_dc_list);
3654 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3655 lnet_net_unlock(LNET_LOCK_EX);
3656 lnet_peer_cancel_discovery(lp);
3657 lnet_net_lock(LNET_LOCK_EX);
3659 lnet_net_unlock(LNET_LOCK_EX);
3661 /* Queue cleanup 2: wait for the expired queue to clear. */
3662 while (!list_empty(&the_lnet.ln_dc_expired))
3663 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3665 /* Queue cleanup 3: clear the request queue. */
3666 lnet_net_lock(LNET_LOCK_EX);
3667 while (!list_empty(&the_lnet.ln_dc_request)) {
3668 lp = list_first_entry(&the_lnet.ln_dc_request,
3669 struct lnet_peer, lp_dc_list);
3670 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3671 lnet_peer_discovery_complete(lp);
3673 lnet_net_unlock(LNET_LOCK_EX);
3675 lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3676 the_lnet.ln_dc_handler = NULL;
3678 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3679 wake_up(&the_lnet.ln_dc_waitq);
3681 CDEBUG(D_NET, "stopped\n");
3686 /* ln_api_mutex is held on entry. */
3687 int lnet_peer_discovery_start(void)
3689 struct task_struct *task;
3692 if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3695 the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3696 the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3697 task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3700 CERROR("Can't start peer discovery thread: %d\n", rc);
3702 the_lnet.ln_dc_handler = NULL;
3704 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3707 CDEBUG(D_NET, "discovery start: %d\n", rc);
3712 /* ln_api_mutex is held on entry. */
3713 void lnet_peer_discovery_stop(void)
3715 if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3718 LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3719 the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3721 /* In the LNetNIInit() path we may be stopping discovery before it
3722 * entered its work loop
3724 if (!completion_done(&the_lnet.ln_started))
3725 complete(&the_lnet.ln_started);
3727 wake_up(&the_lnet.ln_dc_waitq);
3729 wait_event(the_lnet.ln_dc_waitq,
3730 the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3732 LASSERT(list_empty(&the_lnet.ln_dc_request));
3733 LASSERT(list_empty(&the_lnet.ln_dc_working));
3734 LASSERT(list_empty(&the_lnet.ln_dc_expired));
3736 CDEBUG(D_NET, "discovery stopped\n");
3742 lnet_debug_peer(lnet_nid_t nid)
3744 char *aliveness = "NA";
3745 struct lnet_peer_ni *lp;
3748 cpt = lnet_cpt_of_nid(nid, NULL);
3751 lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3753 lnet_net_unlock(cpt);
3754 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3758 if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3759 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3761 CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3762 libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3763 aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3764 lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3765 lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3767 lnet_peer_ni_decref_locked(lp);
3769 lnet_net_unlock(cpt);
3772 /* Gathering information for userspace. */
3774 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3775 char aliveness[LNET_MAX_STR_LEN],
3776 __u32 *cpt_iter, __u32 *refcount,
3777 __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3778 __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3779 __u32 *peer_tx_qnob)
3781 struct lnet_peer_table *peer_table;
3782 struct lnet_peer_ni *lp;
3787 /* get the number of CPTs */
3788 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3790 /* if the cpt number to be examined is >= the number of cpts in
3791 * the system then indicate that there are no more cpts to examin
3793 if (*cpt_iter >= lncpt)
3796 /* get the current table */
3797 peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3798 /* if the ptable is NULL then there are no more cpts to examine */
3799 if (peer_table == NULL)
3802 lnet_net_lock(*cpt_iter);
3804 for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3805 struct list_head *peers = &peer_table->pt_hash[j];
3807 list_for_each_entry(lp, peers, lpni_hashlist) {
3808 if (peer_index-- > 0)
3811 snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3812 if (lnet_isrouter(lp) ||
3813 lnet_peer_aliveness_enabled(lp))
3814 snprintf(aliveness, LNET_MAX_STR_LEN,
3815 lnet_is_peer_ni_alive(lp) ? "up" : "down");
3817 *nid = lp->lpni_nid;
3818 *refcount = atomic_read(&lp->lpni_refcount);
3819 *ni_peer_tx_credits =
3820 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3821 *peer_tx_credits = lp->lpni_txcredits;
3822 *peer_rtr_credits = lp->lpni_rtrcredits;
3823 *peer_min_rtr_credits = lp->lpni_mintxcredits;
3824 *peer_tx_qnob = lp->lpni_txqnob;
3830 lnet_net_unlock(*cpt_iter);
3834 return found ? 0 : -ENOENT;
3837 /* ln_api_mutex is held, which keeps the peer list stable */
3838 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3840 struct lnet_ioctl_element_stats *lpni_stats;
3841 struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3842 struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3843 struct lnet_peer_ni_credit_info *lpni_info;
3844 struct lnet_peer_ni *lpni;
3845 struct lnet_peer *lp;
3850 lp = lnet_find_peer(cfg->prcfg_prim_nid);
3857 size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3858 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3859 size *= lp->lp_nnis;
3860 if (size > cfg->prcfg_size) {
3861 cfg->prcfg_size = size;
3866 cfg->prcfg_prim_nid = lp->lp_primary_nid;
3867 cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3868 cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3869 cfg->prcfg_count = lp->lp_nnis;
3870 cfg->prcfg_size = size;
3871 cfg->prcfg_state = lp->lp_state;
3873 /* Allocate helper buffers. */
3875 LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3878 LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3881 LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3882 if (!lpni_msg_stats)
3883 goto out_free_stats;
3884 LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3886 goto out_free_msg_stats;
3891 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3892 nid = lpni->lpni_nid;
3893 if (copy_to_user(bulk, &nid, sizeof(nid)))
3894 goto out_free_hstats;
3895 bulk += sizeof(nid);
3897 memset(lpni_info, 0, sizeof(*lpni_info));
3898 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3899 if (lnet_isrouter(lpni) ||
3900 lnet_peer_aliveness_enabled(lpni))
3901 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3902 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3904 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3905 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3906 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3907 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3908 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3909 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3910 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3911 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3912 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3913 goto out_free_hstats;
3914 bulk += sizeof(*lpni_info);
3916 memset(lpni_stats, 0, sizeof(*lpni_stats));
3917 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3918 LNET_STATS_TYPE_SEND);
3919 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3920 LNET_STATS_TYPE_RECV);
3921 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3922 LNET_STATS_TYPE_DROP);
3923 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3924 goto out_free_hstats;
3925 bulk += sizeof(*lpni_stats);
3926 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3927 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3928 goto out_free_hstats;
3929 bulk += sizeof(*lpni_msg_stats);
3930 lpni_hstats->hlpni_network_timeout =
3931 atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3932 lpni_hstats->hlpni_remote_dropped =
3933 atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3934 lpni_hstats->hlpni_remote_timeout =
3935 atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3936 lpni_hstats->hlpni_remote_error =
3937 atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3938 lpni_hstats->hlpni_health_value =
3939 atomic_read(&lpni->lpni_healthv);
3940 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3941 goto out_free_hstats;
3942 bulk += sizeof(*lpni_hstats);
3947 LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3949 LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3951 LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3953 LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3955 lnet_peer_decref_locked(lp);
3961 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3963 /* the mt could've shutdown and cleaned up the queues */
3964 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3967 if (list_empty(&lpni->lpni_recovery) &&
3968 atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3969 CDEBUG(D_NET, "lpni %s added to recovery queue. Health = %d\n",
3970 libcfs_nid2str(lpni->lpni_nid),
3971 atomic_read(&lpni->lpni_healthv));
3972 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3973 lnet_peer_ni_addref_locked(lpni);
3977 /* Call with the ln_api_mutex held */
3979 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3981 struct lnet_peer_table *ptable;
3982 struct lnet_peer *lp;
3983 struct lnet_peer_net *lpn;
3984 struct lnet_peer_ni *lpni;
3988 if (the_lnet.ln_state != LNET_STATE_RUNNING)
3992 lnet_net_lock(LNET_LOCK_EX);
3993 lpni = lnet_find_peer_ni_locked(nid);
3995 lnet_net_unlock(LNET_LOCK_EX);
3998 atomic_set(&lpni->lpni_healthv, value);
3999 lnet_peer_ni_add_to_recoveryq_locked(lpni);
4000 lnet_peer_ni_decref_locked(lpni);
4001 lnet_net_unlock(LNET_LOCK_EX);
4005 lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4008 * Walk all the peers and reset the healhv for each one to the
4011 lnet_net_lock(LNET_LOCK_EX);
4012 for (cpt = 0; cpt < lncpt; cpt++) {
4013 ptable = the_lnet.ln_peer_tables[cpt];
4014 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4015 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4016 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4018 atomic_set(&lpni->lpni_healthv, value);
4019 lnet_peer_ni_add_to_recoveryq_locked(lpni);
4024 lnet_net_unlock(LNET_LOCK_EX);