Whamcloud - gitweb
bad28431f5b6499d25da6e12ac395cff08013cc3
[fs/lustre-release.git] / lnet / lnet / peer.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  *
31  * lnet/lnet/peer.c
32  */
33
34 #define DEBUG_SUBSYSTEM S_LNET
35
36 #include <linux/sched.h>
37 #ifdef HAVE_SCHED_HEADERS
38 #include <linux/sched/signal.h>
39 #endif
40 #include <linux/uaccess.h>
41
42 #include <lnet/udsp.h>
43 #include <lnet/lib-lnet.h>
44 #include <uapi/linux/lnet/lnet-dlc.h>
45
46 /* Value indicating that recovery needs to re-check a peer immediately. */
47 #define LNET_REDISCOVER_PEER    (1)
48
49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
50
51 static void
52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
53 {
54         if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
55                 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
56                 lnet_peer_ni_decref_locked(lpni);
57         }
58 }
59
60 void
61 lnet_peer_net_added(struct lnet_net *net)
62 {
63         struct lnet_peer_ni *lpni, *tmp;
64
65         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
66                                  lpni_on_remote_peer_ni_list) {
67
68                 if (LNET_NID_NET(&lpni->lpni_nid) == net->net_id) {
69                         lpni->lpni_net = net;
70
71                         spin_lock(&lpni->lpni_lock);
72                         lpni->lpni_txcredits =
73                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
74                         lpni->lpni_mintxcredits = lpni->lpni_txcredits;
75                         lpni->lpni_rtrcredits =
76                                 lnet_peer_buffer_credits(lpni->lpni_net);
77                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
78                         spin_unlock(&lpni->lpni_lock);
79
80                         lnet_peer_remove_from_remote_list(lpni);
81                 }
82         }
83 }
84
85 static void
86 lnet_peer_tables_destroy(void)
87 {
88         struct lnet_peer_table  *ptable;
89         struct list_head        *hash;
90         int                     i;
91         int                     j;
92
93         if (!the_lnet.ln_peer_tables)
94                 return;
95
96         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
97                 hash = ptable->pt_hash;
98                 if (!hash) /* not intialized */
99                         break;
100
101                 LASSERT(list_empty(&ptable->pt_zombie_list));
102
103                 ptable->pt_hash = NULL;
104                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
105                         LASSERT(list_empty(&hash[j]));
106
107                 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
108         }
109
110         cfs_percpt_free(the_lnet.ln_peer_tables);
111         the_lnet.ln_peer_tables = NULL;
112 }
113
114 int
115 lnet_peer_tables_create(void)
116 {
117         struct lnet_peer_table  *ptable;
118         struct list_head        *hash;
119         int                     i;
120         int                     j;
121
122         the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
123                                                    sizeof(*ptable));
124         if (the_lnet.ln_peer_tables == NULL) {
125                 CERROR("Failed to allocate cpu-partition peer tables\n");
126                 return -ENOMEM;
127         }
128
129         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
130                 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
131                                  LNET_PEER_HASH_SIZE * sizeof(*hash));
132                 if (hash == NULL) {
133                         CERROR("Failed to create peer hash table\n");
134                         lnet_peer_tables_destroy();
135                         return -ENOMEM;
136                 }
137
138                 spin_lock_init(&ptable->pt_zombie_lock);
139                 INIT_LIST_HEAD(&ptable->pt_zombie_list);
140
141                 INIT_LIST_HEAD(&ptable->pt_peer_list);
142
143                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
144                         INIT_LIST_HEAD(&hash[j]);
145                 ptable->pt_hash = hash; /* sign of initialization */
146         }
147
148         return 0;
149 }
150
151 static struct lnet_peer_ni *
152 lnet_peer_ni_alloc(struct lnet_nid *nid)
153 {
154         struct lnet_peer_ni *lpni;
155         struct lnet_net *net;
156         int cpt;
157
158         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
159
160         LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
161         if (!lpni)
162                 return NULL;
163
164         INIT_LIST_HEAD(&lpni->lpni_txq);
165         INIT_LIST_HEAD(&lpni->lpni_hashlist);
166         INIT_LIST_HEAD(&lpni->lpni_peer_nis);
167         INIT_LIST_HEAD(&lpni->lpni_recovery);
168         INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
169         INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
170         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
171         kref_init(&lpni->lpni_kref);
172         lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
173
174         spin_lock_init(&lpni->lpni_lock);
175
176         if (lnet_peers_start_down())
177                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
178         else
179                 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
180         lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
181         lpni->lpni_nid = *nid;
182         lpni->lpni_cpt = cpt;
183         atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
184
185         net = lnet_get_net_locked(LNET_NID_NET(nid));
186         lpni->lpni_net = net;
187         if (net) {
188                 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
189                 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
190                 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
191                 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
192         } else {
193                 /*
194                  * This peer_ni is not on a local network, so we
195                  * cannot add the credits here. In case the net is
196                  * added later, add the peer_ni to the remote peer ni
197                  * list so it can be easily found and revisited.
198                  */
199                 /* FIXME: per-net implementation instead? */
200                 lnet_peer_ni_addref_locked(lpni);
201                 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
202                               &the_lnet.ln_remote_peer_ni_list);
203         }
204
205         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
206
207         return lpni;
208 }
209
210 static struct lnet_peer_net *
211 lnet_peer_net_alloc(__u32 net_id)
212 {
213         struct lnet_peer_net *lpn;
214
215         LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
216         if (!lpn)
217                 return NULL;
218
219         INIT_LIST_HEAD(&lpn->lpn_peer_nets);
220         INIT_LIST_HEAD(&lpn->lpn_peer_nis);
221         lpn->lpn_net_id = net_id;
222         lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
223
224         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
225
226         return lpn;
227 }
228
229 void
230 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
231 {
232         struct lnet_peer *lp;
233
234         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
235
236         LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
237         LASSERT(list_empty(&lpn->lpn_peer_nis));
238         LASSERT(list_empty(&lpn->lpn_peer_nets));
239         lp = lpn->lpn_peer;
240         lpn->lpn_peer = NULL;
241         LIBCFS_FREE(lpn, sizeof(*lpn));
242
243         lnet_peer_decref_locked(lp);
244 }
245
246 static struct lnet_peer *
247 lnet_peer_alloc(struct lnet_nid *nid)
248 {
249         struct lnet_peer *lp;
250
251         LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
252         if (!lp)
253                 return NULL;
254
255         INIT_LIST_HEAD(&lp->lp_rtrq);
256         INIT_LIST_HEAD(&lp->lp_routes);
257         INIT_LIST_HEAD(&lp->lp_peer_list);
258         INIT_LIST_HEAD(&lp->lp_peer_nets);
259         INIT_LIST_HEAD(&lp->lp_dc_list);
260         INIT_LIST_HEAD(&lp->lp_dc_pendq);
261         INIT_LIST_HEAD(&lp->lp_rtr_list);
262         init_waitqueue_head(&lp->lp_dc_waitq);
263         spin_lock_init(&lp->lp_lock);
264         lp->lp_primary_nid = *nid;
265         lp->lp_disc_src_nid = LNET_ANY_NID;
266         lp->lp_disc_dst_nid = LNET_ANY_NID;
267         if (lnet_peers_start_down())
268                 lp->lp_alive = false;
269         else
270                 lp->lp_alive = true;
271
272         /*
273          * all peers created on a router should have health on
274          * if it's not already on.
275          */
276         if (the_lnet.ln_routing && !lnet_health_sensitivity)
277                 lp->lp_health_sensitivity = 1;
278
279         /*
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
284          * myself.
285          */
286         if (nid_is_lo0(nid))
287                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
288         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
289
290         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
291
292         return lp;
293 }
294
295 void
296 lnet_destroy_peer_locked(struct lnet_peer *lp)
297 {
298         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
299
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));
305
306         if (lp->lp_data)
307                 lnet_ping_buffer_decref(lp->lp_data);
308
309         /*
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
313          * running.
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
320          */
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);
327
328         LIBCFS_FREE(lp, sizeof(*lp));
329 }
330
331 /*
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.
334  *
335  * Call with lnet_net_lock/EX held
336  */
337 static void
338 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
339 {
340         struct lnet_peer_table *ptable;
341         struct lnet_peer_net *lpn;
342         struct lnet_peer *lp;
343
344         /*
345          * Belts and suspenders: gracefully handle teardown of a
346          * partially connected peer_ni.
347          */
348         lpn = lpni->lpni_peer_net;
349
350         list_del_init(&lpni->lpni_peer_nis);
351         /*
352          * If there are no lpni's left, we detach lpn from
353          * lp_peer_nets, so it cannot be found anymore.
354          */
355         if (list_empty(&lpn->lpn_peer_nis))
356                 list_del_init(&lpn->lpn_peer_nets);
357
358         /* Update peer NID count. */
359         lp = lpn->lpn_peer;
360         lp->lp_nnis--;
361
362         /*
363          * If there are no more peer nets, make the peer unfindable
364          * via the peer_tables.
365          *
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.
369          */
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];
373                 ptable->pt_peers--;
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);
379         }
380         CDEBUG(D_NET, "peer %s NID %s\n",
381                 libcfs_nidstr(&lp->lp_primary_nid),
382                 libcfs_nidstr(&lpni->lpni_nid));
383 }
384
385 /* called with lnet_net_lock LNET_LOCK_EX held */
386 static int
387 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
388 {
389         struct lnet_peer_table *ptable = NULL;
390
391         /* don't remove a peer_ni if it's also a gateway */
392         if (lnet_isrouter(lpni) && !force) {
393                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
394                        libcfs_nidstr(&lpni->lpni_nid));
395                 return -EBUSY;
396         }
397
398         lnet_peer_remove_from_remote_list(lpni);
399
400         /* remove peer ni from the hash list. */
401         list_del_init(&lpni->lpni_hashlist);
402
403         /*
404          * indicate the peer is being deleted so the monitor thread can
405          * remove it from the recovery queue.
406          */
407         spin_lock(&lpni->lpni_lock);
408         lpni->lpni_state |= LNET_PEER_NI_DELETING;
409         spin_unlock(&lpni->lpni_lock);
410
411         /* decrement the ref count on the peer table */
412         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
413
414         /*
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.
418          *
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
422          * has its own lock.
423          */
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);
428
429         /* no need to keep this peer_ni on the hierarchy anymore */
430         lnet_peer_detach_peer_ni_locked(lpni);
431
432         /* remove hashlist reference on peer_ni */
433         lnet_peer_ni_decref_locked(lpni);
434
435         return 0;
436 }
437
438 void lnet_peer_uninit(void)
439 {
440         struct lnet_peer_ni *lpni, *tmp;
441
442         lnet_net_lock(LNET_LOCK_EX);
443
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);
448
449         lnet_peer_tables_destroy();
450
451         lnet_net_unlock(LNET_LOCK_EX);
452 }
453
454 static int
455 lnet_peer_del_locked(struct lnet_peer *peer)
456 {
457         struct lnet_peer_ni *lpni = NULL, *lpni2;
458         int rc = 0, rc2 = 0;
459
460         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&peer->lp_primary_nid));
461
462         spin_lock(&peer->lp_lock);
463         peer->lp_state |= LNET_PEER_MARK_DELETED;
464         spin_unlock(&peer->lp_lock);
465
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);
470                 if (rc != 0)
471                         rc2 = rc;
472                 lpni = lpni2;
473         }
474
475         return rc2;
476 }
477
478 /*
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
482  * the cleanup.
483  */
484 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
485 {
486         struct lnet_handle_md ping_mdh;
487         struct lnet_handle_md push_mdh;
488
489         LNetInvalidateMDHandle(&ping_mdh);
490         LNetInvalidateMDHandle(&push_mdh);
491
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);
496         }
497         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
498                 push_mdh = lp->lp_push_mdh;
499                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
500         }
501         spin_unlock(&lp->lp_lock);
502
503         if (!LNetMDHandleIsInvalid(ping_mdh))
504                 LNetMDUnlink(ping_mdh);
505         if (!LNetMDHandleIsInvalid(push_mdh))
506                 LNetMDUnlink(push_mdh);
507 }
508
509 static int
510 lnet_peer_del(struct lnet_peer *peer)
511 {
512         int rc;
513
514         lnet_peer_cancel_discovery(peer);
515         lnet_net_lock(LNET_LOCK_EX);
516         rc = lnet_peer_del_locked(peer);
517         lnet_net_unlock(LNET_LOCK_EX);
518
519         return rc;
520 }
521
522 /*
523  * Delete a NID from a peer. Call with ln_api_mutex held.
524  *
525  * Error codes:
526  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
527  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
528  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
529  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
530  */
531 static int
532 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
533 {
534         struct lnet_peer_ni *lpni;
535         struct lnet_nid primary_nid = lp->lp_primary_nid;
536         struct lnet_nid nid;
537         int rc = 0;
538         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
539
540         lnet_nid4_to_nid(nid4, &nid);
541         if (!(flags & LNET_PEER_CONFIGURED)) {
542                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
543                         rc = -EPERM;
544                         goto out;
545                 }
546         }
547
548         lpni = lnet_peer_ni_find_locked(&nid);
549         if (!lpni) {
550                 rc = -ENOENT;
551                 goto out;
552         }
553         lnet_peer_ni_decref_locked(lpni);
554         if (lp != lpni->lpni_peer_net->lpn_peer) {
555                 rc = -ECHILD;
556                 goto out;
557         }
558
559         /*
560          * This function only allows deletion of the primary NID if it
561          * is the only NID.
562          */
563         if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && !force) {
564                 rc = -EBUSY;
565                 goto out;
566         }
567
568         lnet_net_lock(LNET_LOCK_EX);
569
570         if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && force) {
571                 struct lnet_peer_ni *lpni2;
572                 /* assign the next peer_ni to be the primary */
573                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
574                 LASSERT(lpni2);
575                 lp->lp_primary_nid = lpni2->lpni_nid;
576         }
577         rc = lnet_peer_ni_del_locked(lpni, force);
578
579         lnet_net_unlock(LNET_LOCK_EX);
580
581 out:
582         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
583                libcfs_nidstr(&primary_nid), libcfs_nidstr(&nid),
584                flags, rc);
585
586         return rc;
587 }
588
589 static void
590 lnet_peer_table_cleanup_locked(struct lnet_net *net,
591                                struct lnet_peer_table *ptable)
592 {
593         int                      i;
594         struct lnet_peer_ni     *next;
595         struct lnet_peer_ni     *lpni;
596         struct lnet_peer        *peer;
597
598         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
599                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
600                                          lpni_hashlist) {
601                         if (net != NULL && net != lpni->lpni_net)
602                                 continue;
603
604                         peer = lpni->lpni_peer_net->lpn_peer;
605                         if (!nid_same(&peer->lp_primary_nid,
606                                        &lpni->lpni_nid)) {
607                                 lnet_peer_ni_del_locked(lpni, false);
608                                 continue;
609                         }
610                         /*
611                          * Removing the primary NID implies removing
612                          * the entire peer. Advance next beyond any
613                          * peer_ni that belongs to the same peer.
614                          */
615                         list_for_each_entry_from(next, &ptable->pt_hash[i],
616                                                  lpni_hashlist) {
617                                 if (next->lpni_peer_net->lpn_peer != peer)
618                                         break;
619                         }
620                         lnet_peer_del_locked(peer);
621                 }
622         }
623 }
624
625 static void
626 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
627 {
628         wait_var_event_warning(&ptable->pt_zombies,
629                                ptable->pt_zombies == 0,
630                                "Waiting for %d zombies on peer table\n",
631                                ptable->pt_zombies);
632 }
633
634 static void
635 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
636                                 struct lnet_peer_table *ptable)
637 {
638         struct lnet_peer_ni     *lp;
639         struct lnet_peer_ni     *tmp;
640         struct lnet_nid         gw_nid;
641         int                     i;
642
643         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
644                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
645                                          lpni_hashlist) {
646                         if (net != lp->lpni_net)
647                                 continue;
648
649                         if (!lnet_isrouter(lp))
650                                 continue;
651
652                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
653
654                         lnet_net_unlock(LNET_LOCK_EX);
655                         lnet_del_route(LNET_NET_ANY, &gw_nid);
656                         lnet_net_lock(LNET_LOCK_EX);
657                 }
658         }
659 }
660
661 void
662 lnet_peer_tables_cleanup(struct lnet_net *net)
663 {
664         int i;
665         struct lnet_peer_table *ptable;
666
667         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
668         /* If just deleting the peers for a NI, get rid of any routes these
669          * peers are gateways for. */
670         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
671                 lnet_net_lock(LNET_LOCK_EX);
672                 lnet_peer_table_del_rtrs_locked(net, ptable);
673                 lnet_net_unlock(LNET_LOCK_EX);
674         }
675
676         /* Start the cleanup process */
677         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
678                 lnet_net_lock(LNET_LOCK_EX);
679                 lnet_peer_table_cleanup_locked(net, ptable);
680                 lnet_net_unlock(LNET_LOCK_EX);
681         }
682
683         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
684                 lnet_peer_ni_finalize_wait(ptable);
685 }
686
687 static struct lnet_peer_ni *
688 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, struct lnet_nid *nid)
689 {
690         struct list_head        *peers;
691         struct lnet_peer_ni     *lp;
692
693         if (the_lnet.ln_state != LNET_STATE_RUNNING)
694                 return NULL;
695
696         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
697         list_for_each_entry(lp, peers, lpni_hashlist) {
698                 if (nid_same(&lp->lpni_nid, nid)) {
699                         lnet_peer_ni_addref_locked(lp);
700                         return lp;
701                 }
702         }
703
704         return NULL;
705 }
706
707 struct lnet_peer_ni *
708 lnet_peer_ni_find_locked(struct lnet_nid *nid)
709 {
710         struct lnet_peer_ni *lpni;
711         struct lnet_peer_table *ptable;
712         int cpt;
713
714         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
715
716         ptable = the_lnet.ln_peer_tables[cpt];
717         lpni = lnet_get_peer_ni_locked(ptable, nid);
718
719         return lpni;
720 }
721
722 struct lnet_peer_ni *
723 lnet_peer_ni_get_locked(struct lnet_peer *lp, struct lnet_nid *nid)
724 {
725         struct lnet_peer_net *lpn;
726         struct lnet_peer_ni *lpni;
727
728         lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
729         if (!lpn)
730                 return NULL;
731
732         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
733                 if (nid_same(&lpni->lpni_nid, nid))
734                         return lpni;
735         }
736
737         return NULL;
738 }
739
740 struct lnet_peer *
741 lnet_find_peer(struct lnet_nid *nid)
742 {
743         struct lnet_peer_ni *lpni;
744         struct lnet_peer *lp = NULL;
745         int cpt;
746
747         cpt = lnet_net_lock_current();
748         lpni = lnet_peer_ni_find_locked(nid);
749         if (lpni) {
750                 lp = lpni->lpni_peer_net->lpn_peer;
751                 lnet_peer_addref_locked(lp);
752                 lnet_peer_ni_decref_locked(lpni);
753         }
754         lnet_net_unlock(cpt);
755
756         return lp;
757 }
758
759 struct lnet_peer_net *
760 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
761 {
762         struct lnet_peer_net *net;
763
764         if (!prev_lpn_id) {
765                 /* no net id provided return the first net */
766                 net = list_first_entry_or_null(&lp->lp_peer_nets,
767                                                struct lnet_peer_net,
768                                                lpn_peer_nets);
769
770                 return net;
771         }
772
773         /* find the net after the one provided */
774         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
775                 if (net->lpn_net_id == prev_lpn_id) {
776                         /*
777                          * if we reached the end of the list loop to the
778                          * beginning.
779                          */
780                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
781                                 return list_first_entry_or_null(&lp->lp_peer_nets,
782                                                                 struct lnet_peer_net,
783                                                                 lpn_peer_nets);
784                         else
785                                 return list_next_entry(net, lpn_peer_nets);
786                 }
787         }
788
789         return NULL;
790 }
791
792 struct lnet_peer_ni *
793 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
794                              struct lnet_peer_net *peer_net,
795                              struct lnet_peer_ni *prev)
796 {
797         struct lnet_peer_ni *lpni;
798         struct lnet_peer_net *net = peer_net;
799
800         if (!prev) {
801                 if (!net) {
802                         if (list_empty(&peer->lp_peer_nets))
803                                 return NULL;
804
805                         net = list_entry(peer->lp_peer_nets.next,
806                                          struct lnet_peer_net,
807                                          lpn_peer_nets);
808                 }
809                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
810                                   lpni_peer_nis);
811
812                 return lpni;
813         }
814
815         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
816                 /*
817                  * if you reached the end of the peer ni list and the peer
818                  * net is specified then there are no more peer nis in that
819                  * net.
820                  */
821                 if (net)
822                         return NULL;
823
824                 /*
825                  * we reached the end of this net ni list. move to the
826                  * next net
827                  */
828                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
829                     &peer->lp_peer_nets)
830                         /* no more nets and no more NIs. */
831                         return NULL;
832
833                 /* get the next net */
834                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
835                                  struct lnet_peer_net,
836                                  lpn_peer_nets);
837                 /* get the ni on it */
838                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
839                                   lpni_peer_nis);
840
841                 return lpni;
842         }
843
844         /* there are more nis left */
845         lpni = list_entry(prev->lpni_peer_nis.next,
846                           struct lnet_peer_ni, lpni_peer_nis);
847
848         return lpni;
849 }
850
851 /* Call with the ln_api_mutex held */
852 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
853 {
854         struct lnet_process_id id;
855         struct lnet_peer_table *ptable;
856         struct lnet_peer *lp;
857         __u32 count = 0;
858         __u32 size = 0;
859         int lncpt;
860         int cpt;
861         __u32 i;
862         int rc;
863
864         rc = -ESHUTDOWN;
865         if (the_lnet.ln_state != LNET_STATE_RUNNING)
866                 goto done;
867
868         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
869
870         /*
871          * Count the number of peers, and return E2BIG if the buffer
872          * is too small. We'll also return the desired size.
873          */
874         rc = -E2BIG;
875         for (cpt = 0; cpt < lncpt; cpt++) {
876                 ptable = the_lnet.ln_peer_tables[cpt];
877                 count += ptable->pt_peers;
878         }
879         size = count * sizeof(*ids);
880         if (size > *sizep)
881                 goto done;
882
883         /*
884          * Walk the peer lists and copy out the primary nids.
885          * This is safe because the peer lists are only modified
886          * while the ln_api_mutex is held. So we don't need to
887          * hold the lnet_net_lock as well, and can therefore
888          * directly call copy_to_user().
889          */
890         rc = -EFAULT;
891         memset(&id, 0, sizeof(id));
892         id.pid = LNET_PID_LUSTRE;
893         i = 0;
894         for (cpt = 0; cpt < lncpt; cpt++) {
895                 ptable = the_lnet.ln_peer_tables[cpt];
896                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
897                         if (!nid_is_nid4(&lp->lp_primary_nid))
898                                 continue;
899                         if (i >= count)
900                                 goto done;
901                         id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
902                         if (copy_to_user(&ids[i], &id, sizeof(id)))
903                                 goto done;
904                         i++;
905                 }
906         }
907         rc = 0;
908 done:
909         *countp = count;
910         *sizep = size;
911         return rc;
912 }
913
914 /*
915  * Start pushes to peers that need to be updated for a configuration
916  * change on this node.
917  */
918 void
919 lnet_push_update_to_peers(int force)
920 {
921         struct lnet_peer_table *ptable;
922         struct lnet_peer *lp;
923         int lncpt;
924         int cpt;
925
926         lnet_net_lock(LNET_LOCK_EX);
927         if (lnet_peer_discovery_disabled)
928                 force = 0;
929         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
930         for (cpt = 0; cpt < lncpt; cpt++) {
931                 ptable = the_lnet.ln_peer_tables[cpt];
932                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
933                         if (force) {
934                                 spin_lock(&lp->lp_lock);
935                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
936                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
937                                 spin_unlock(&lp->lp_lock);
938                         }
939                         if (lnet_peer_needs_push(lp))
940                                 lnet_peer_queue_for_discovery(lp);
941                 }
942         }
943         lnet_net_unlock(LNET_LOCK_EX);
944         wake_up(&the_lnet.ln_dc_waitq);
945 }
946
947 /* find the NID in the preferred gateways for the remote peer
948  * return:
949  *      false: list is not empty and NID is not preferred
950  *      false: list is empty
951  *      true: nid is found in the list
952  */
953 bool
954 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
955                              struct lnet_nid *gw_nid)
956 {
957         struct lnet_nid_list *ne;
958
959         CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
960                libcfs_nidstr(&lpni->lpni_nid),
961                list_empty(&lpni->lpni_rtr_pref_nids));
962
963         if (list_empty(&lpni->lpni_rtr_pref_nids))
964                 return false;
965
966         /* iterate through all the preferred NIDs and see if any of them
967          * matches the provided gw_nid
968          */
969         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
970                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
971                        libcfs_nidstr(&ne->nl_nid),
972                        libcfs_nidstr(gw_nid));
973                 if (nid_same(&ne->nl_nid, gw_nid))
974                         return true;
975         }
976
977         return false;
978 }
979
980 void
981 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
982 {
983         struct list_head zombies;
984         struct lnet_nid_list *ne;
985         struct lnet_nid_list *tmp;
986         int cpt = lpni->lpni_cpt;
987
988         INIT_LIST_HEAD(&zombies);
989
990         lnet_net_lock(cpt);
991         list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
992         lnet_net_unlock(cpt);
993
994         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
995                 list_del(&ne->nl_list);
996                 LIBCFS_FREE(ne, sizeof(*ne));
997         }
998 }
999
1000 int
1001 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
1002                        struct lnet_nid *gw_nid)
1003 {
1004         int cpt = lpni->lpni_cpt;
1005         struct lnet_nid_list *ne = NULL;
1006
1007         /* This function is called with api_mutex held. When the api_mutex
1008          * is held the list can not be modified, as it is only modified as
1009          * a result of applying a UDSP and that happens under api_mutex
1010          * lock.
1011          */
1012         __must_hold(&the_lnet.ln_api_mutex);
1013
1014         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1015                 if (nid_same(&ne->nl_nid, gw_nid))
1016                         return -EEXIST;
1017         }
1018
1019         LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1020         if (!ne)
1021                 return -ENOMEM;
1022
1023         ne->nl_nid = *gw_nid;
1024
1025         /* Lock the cpt to protect against addition and checks in the
1026          * selection algorithm
1027          */
1028         lnet_net_lock(cpt);
1029         list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1030         lnet_net_unlock(cpt);
1031
1032         return 0;
1033 }
1034
1035 /*
1036  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1037  * this is a preferred point-to-point path. Call with lnet_net_lock in
1038  * shared mmode.
1039  */
1040 bool
1041 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1042 {
1043         struct lnet_nid_list *ne;
1044
1045         if (lpni->lpni_pref_nnids == 0)
1046                 return false;
1047         if (lpni->lpni_pref_nnids == 1)
1048                 return nid_same(&lpni->lpni_pref.nid, nid);
1049         list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1050                 if (nid_same(&ne->nl_nid, nid))
1051                         return true;
1052         }
1053         return false;
1054 }
1055
1056 /*
1057  * Set a single ni as preferred, provided no preferred ni is already
1058  * defined. Only to be used for non-multi-rail peer_ni.
1059  */
1060 int
1061 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni,
1062                                   struct lnet_nid *nid)
1063 {
1064         int rc = 0;
1065
1066         if (!nid)
1067                 return -EINVAL;
1068         spin_lock(&lpni->lpni_lock);
1069         if (LNET_NID_IS_ANY(nid)) {
1070                 rc = -EINVAL;
1071         } else if (lpni->lpni_pref_nnids > 0) {
1072                 rc = -EPERM;
1073         } else if (lpni->lpni_pref_nnids == 0) {
1074                 lpni->lpni_pref.nid = *nid;
1075                 lpni->lpni_pref_nnids = 1;
1076                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1077         }
1078         spin_unlock(&lpni->lpni_lock);
1079
1080         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1081                libcfs_nidstr(&lpni->lpni_nid), libcfs_nidstr(nid), rc);
1082         return rc;
1083 }
1084
1085 /*
1086  * Clear the preferred NID from a non-multi-rail peer_ni, provided
1087  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1088  */
1089 int
1090 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1091 {
1092         int rc = 0;
1093
1094         spin_lock(&lpni->lpni_lock);
1095         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1096                 lpni->lpni_pref_nnids = 0;
1097                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1098         } else if (lpni->lpni_pref_nnids == 0) {
1099                 rc = -ENOENT;
1100         } else {
1101                 rc = -EPERM;
1102         }
1103         spin_unlock(&lpni->lpni_lock);
1104
1105         CDEBUG(D_NET, "peer %s: %d\n",
1106                libcfs_nidstr(&lpni->lpni_nid), rc);
1107         return rc;
1108 }
1109
1110 void
1111 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1112 {
1113         lpni->lpni_sel_priority = priority;
1114 }
1115
1116 /*
1117  * Clear the preferred NIDs from a non-multi-rail peer.
1118  */
1119 void
1120 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1121 {
1122         struct lnet_peer_ni *lpni = NULL;
1123
1124         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1125                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1126 }
1127
1128 int
1129 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1130 {
1131         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1132         struct lnet_nid_list *ne1 = NULL;
1133         struct lnet_nid_list *ne2 = NULL;
1134         struct lnet_nid *tmp_nid = NULL;
1135         int rc = 0;
1136
1137         if (LNET_NID_IS_ANY(nid)) {
1138                 rc = -EINVAL;
1139                 goto out;
1140         }
1141
1142         if (lpni->lpni_pref_nnids == 1 &&
1143             nid_same(&lpni->lpni_pref.nid, nid)) {
1144                 rc = -EEXIST;
1145                 goto out;
1146         }
1147
1148         /* A non-MR node may have only one preferred NI per peer_ni */
1149         if (lpni->lpni_pref_nnids > 0 &&
1150             !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1151                 rc = -EPERM;
1152                 goto out;
1153         }
1154
1155         /* add the new preferred nid to the list of preferred nids */
1156         if (lpni->lpni_pref_nnids != 0) {
1157                 size_t alloc_size = sizeof(*ne1);
1158
1159                 if (lpni->lpni_pref_nnids == 1) {
1160                         tmp_nid = &lpni->lpni_pref.nid;
1161                         INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1162                 }
1163
1164                 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1165                         if (nid_same(&ne1->nl_nid, nid)) {
1166                                 rc = -EEXIST;
1167                                 goto out;
1168                         }
1169                 }
1170
1171                 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1172                                  alloc_size);
1173                 if (!ne1) {
1174                         rc = -ENOMEM;
1175                         goto out;
1176                 }
1177
1178                 /* move the originally stored nid to the list */
1179                 if (lpni->lpni_pref_nnids == 1) {
1180                         LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1181                                 lpni->lpni_cpt, alloc_size);
1182                         if (!ne2) {
1183                                 rc = -ENOMEM;
1184                                 goto out;
1185                         }
1186                         INIT_LIST_HEAD(&ne2->nl_list);
1187                         ne2->nl_nid = *tmp_nid;
1188                 }
1189                 ne1->nl_nid = *nid;
1190         }
1191
1192         lnet_net_lock(LNET_LOCK_EX);
1193         spin_lock(&lpni->lpni_lock);
1194         if (lpni->lpni_pref_nnids == 0) {
1195                 lpni->lpni_pref.nid = *nid;
1196         } else {
1197                 if (ne2)
1198                         list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1199                 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1200         }
1201         lpni->lpni_pref_nnids++;
1202         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1203         spin_unlock(&lpni->lpni_lock);
1204         lnet_net_unlock(LNET_LOCK_EX);
1205
1206 out:
1207         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1208                 spin_lock(&lpni->lpni_lock);
1209                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1210                 spin_unlock(&lpni->lpni_lock);
1211         }
1212         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1213                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1214         return rc;
1215 }
1216
1217 int
1218 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1219 {
1220         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1221         struct lnet_nid_list *ne = NULL;
1222         int rc = 0;
1223
1224         if (lpni->lpni_pref_nnids == 0) {
1225                 rc = -ENOENT;
1226                 goto out;
1227         }
1228
1229         if (lpni->lpni_pref_nnids == 1) {
1230                 if (!nid_same(&lpni->lpni_pref.nid, nid)) {
1231                         rc = -ENOENT;
1232                         goto out;
1233                 }
1234         } else {
1235                 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1236                         if (nid_same(&ne->nl_nid, nid))
1237                                 goto remove_nid_entry;
1238                 }
1239                 rc = -ENOENT;
1240                 ne = NULL;
1241                 goto out;
1242         }
1243
1244 remove_nid_entry:
1245         lnet_net_lock(LNET_LOCK_EX);
1246         spin_lock(&lpni->lpni_lock);
1247         if (lpni->lpni_pref_nnids == 1)
1248                 lpni->lpni_pref.nid = LNET_ANY_NID;
1249         else {
1250                 list_del_init(&ne->nl_list);
1251                 if (lpni->lpni_pref_nnids == 2) {
1252                         struct lnet_nid_list *ne, *tmp;
1253
1254                         list_for_each_entry_safe(ne, tmp,
1255                                                  &lpni->lpni_pref.nids,
1256                                                  nl_list) {
1257                                 lpni->lpni_pref.nid = ne->nl_nid;
1258                                 list_del_init(&ne->nl_list);
1259                                 LIBCFS_FREE(ne, sizeof(*ne));
1260                         }
1261                 }
1262         }
1263         lpni->lpni_pref_nnids--;
1264         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1265         spin_unlock(&lpni->lpni_lock);
1266         lnet_net_unlock(LNET_LOCK_EX);
1267
1268         if (ne)
1269                 LIBCFS_FREE(ne, sizeof(*ne));
1270 out:
1271         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1272                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1273         return rc;
1274 }
1275
1276 void
1277 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1278 {
1279         struct list_head zombies;
1280         struct lnet_nid_list *ne;
1281         struct lnet_nid_list *tmp;
1282
1283         INIT_LIST_HEAD(&zombies);
1284
1285         lnet_net_lock(LNET_LOCK_EX);
1286         if (lpni->lpni_pref_nnids == 1)
1287                 lpni->lpni_pref.nid = LNET_ANY_NID;
1288         else if (lpni->lpni_pref_nnids > 1)
1289                 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1290         lpni->lpni_pref_nnids = 0;
1291         lnet_net_unlock(LNET_LOCK_EX);
1292
1293         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1294                 list_del_init(&ne->nl_list);
1295                 LIBCFS_FREE(ne, sizeof(*ne));
1296         }
1297 }
1298
1299 void
1300 lnet_peer_primary_nid_locked(struct lnet_nid *nid, struct lnet_nid *result)
1301 {
1302         struct lnet_peer_ni *lpni;
1303
1304         *result = *nid;
1305         lpni = lnet_peer_ni_find_locked(nid);
1306         if (lpni) {
1307                 *result = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1308                 lnet_peer_ni_decref_locked(lpni);
1309         }
1310 }
1311
1312 bool
1313 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1314 __must_hold(&lp->lp_lock)
1315 {
1316         if (lnet_peer_discovery_disabled)
1317                 return true;
1318
1319         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1320             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1321                 return true;
1322         }
1323
1324         return false;
1325 }
1326
1327 /*
1328  * Peer Discovery
1329  */
1330 bool
1331 lnet_is_discovery_disabled(struct lnet_peer *lp)
1332 {
1333         bool rc = false;
1334
1335         spin_lock(&lp->lp_lock);
1336         rc = lnet_is_discovery_disabled_locked(lp);
1337         spin_unlock(&lp->lp_lock);
1338
1339         return rc;
1340 }
1341
1342 int
1343 LNetAddPeer(lnet_nid_t *nids, __u32 num_nids)
1344 {
1345         lnet_nid_t pnid = 0;
1346         bool mr;
1347         int i, rc;
1348
1349         if (!nids || num_nids < 1)
1350                 return -EINVAL;
1351
1352         rc = LNetNIInit(LNET_PID_ANY);
1353         if (rc < 0)
1354                 return rc;
1355
1356         mutex_lock(&the_lnet.ln_api_mutex);
1357
1358         mr = lnet_peer_discovery_disabled == 0;
1359
1360         rc = 0;
1361         for (i = 0; i < num_nids; i++) {
1362                 if (nids[i] == LNET_NID_LO_0)
1363                         continue;
1364
1365                 if (!pnid) {
1366                         pnid = nids[i];
1367                         rc = lnet_add_peer_ni(pnid, LNET_NID_ANY, mr, true);
1368                 } else if (lnet_peer_discovery_disabled) {
1369                         rc = lnet_add_peer_ni(nids[i], LNET_NID_ANY, mr, true);
1370                 } else {
1371                         rc = lnet_add_peer_ni(pnid, nids[i], mr, true);
1372                 }
1373
1374                 if (rc && rc != -EEXIST)
1375                         goto unlock;
1376         }
1377
1378 unlock:
1379         mutex_unlock(&the_lnet.ln_api_mutex);
1380
1381         LNetNIFini();
1382
1383         return rc == -EEXIST ? 0 : rc;
1384 }
1385 EXPORT_SYMBOL(LNetAddPeer);
1386
1387 void LNetPrimaryNID(struct lnet_nid *nid)
1388 {
1389         struct lnet_peer *lp;
1390         struct lnet_peer_ni *lpni;
1391         struct lnet_nid orig;
1392         int rc = 0;
1393         int cpt;
1394
1395         if (!nid || nid_is_lo0(nid))
1396                 return;
1397         orig = *nid;
1398
1399         cpt = lnet_net_lock_current();
1400         lpni = lnet_peerni_by_nid_locked(nid, NULL, cpt);
1401         if (IS_ERR(lpni)) {
1402                 rc = PTR_ERR(lpni);
1403                 goto out_unlock;
1404         }
1405         lp = lpni->lpni_peer_net->lpn_peer;
1406
1407         /* If discovery is disabled locally then we needn't bother running
1408          * discovery here because discovery will not modify whatever
1409          * primary NID is currently set for this peer. If the specified peer is
1410          * down then this discovery can introduce long delays into the mount
1411          * process, so skip it if it isn't necessary.
1412          */
1413         while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1414                 spin_lock(&lp->lp_lock);
1415                 /* force a full discovery cycle */
1416                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1417                 spin_unlock(&lp->lp_lock);
1418
1419                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1420                 if (rc)
1421                         goto out_decref;
1422                 /* The lpni (or lp) for this NID may have changed and our ref is
1423                  * the only thing keeping the old one around. Release the ref
1424                  * and lookup the lpni again
1425                  */
1426                 lnet_peer_ni_decref_locked(lpni);
1427                 lpni = lnet_peer_ni_find_locked(nid);
1428                 if (!lpni) {
1429                         rc = -ENOENT;
1430                         goto out_unlock;
1431                 }
1432                 lp = lpni->lpni_peer_net->lpn_peer;
1433
1434                 /* If we find that the peer has discovery disabled then we will
1435                  * not modify whatever primary NID is currently set for this
1436                  * peer. Thus, we can break out of this loop even if the peer
1437                  * is not fully up to date.
1438                  */
1439                 if (lnet_is_discovery_disabled(lp))
1440                         break;
1441         }
1442         *nid = lp->lp_primary_nid;
1443 out_decref:
1444         lnet_peer_ni_decref_locked(lpni);
1445 out_unlock:
1446         lnet_net_unlock(cpt);
1447
1448         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nidstr(&orig),
1449                libcfs_nidstr(nid), rc);
1450 }
1451 EXPORT_SYMBOL(LNetPrimaryNID);
1452
1453 struct lnet_peer_net *
1454 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1455 {
1456         struct lnet_peer_net *peer_net;
1457         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1458                 if (peer_net->lpn_net_id == net_id)
1459                         return peer_net;
1460         }
1461         return NULL;
1462 }
1463
1464 /*
1465  * Attach a peer_ni to a peer_net and peer. This function assumes
1466  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1467  * may be attached to a different peer, in which case it will be
1468  * properly detached first. The whole operation is done atomically.
1469  *
1470  * This function consumes the reference on lpni and Always returns 0.
1471  * This is the last function called from functions that do return an
1472  * int, so returning 0 here allows the compiler to do a tail call.
1473  */
1474 static int
1475 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1476                                 struct lnet_peer_net *lpn,
1477                                 struct lnet_peer_ni *lpni,
1478                                 unsigned flags)
1479 {
1480         struct lnet_peer_table *ptable;
1481         bool new_lpn = false;
1482         int rc;
1483
1484         /* Install the new peer_ni */
1485         lnet_net_lock(LNET_LOCK_EX);
1486         /* Add peer_ni to global peer table hash, if necessary. */
1487         if (list_empty(&lpni->lpni_hashlist)) {
1488                 int hash = lnet_nid2peerhash(&lpni->lpni_nid);
1489
1490                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1491                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1492                 ptable->pt_version++;
1493                 lnet_peer_ni_addref_locked(lpni);
1494         }
1495
1496         /* Detach the peer_ni from an existing peer, if necessary. */
1497         if (lpni->lpni_peer_net) {
1498                 LASSERT(lpni->lpni_peer_net != lpn);
1499                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1500                 lnet_peer_detach_peer_ni_locked(lpni);
1501                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1502                 lpni->lpni_peer_net = NULL;
1503         }
1504
1505         /* Add peer_ni to peer_net */
1506         lpni->lpni_peer_net = lpn;
1507         if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1508                 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1509         else
1510                 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1511         lnet_update_peer_net_healthv(lpni);
1512         lnet_peer_net_addref_locked(lpn);
1513
1514         /* Add peer_net to peer */
1515         if (!lpn->lpn_peer) {
1516                 new_lpn = true;
1517                 lpn->lpn_peer = lp;
1518                 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1519                         list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1520                 else
1521                         list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1522                 lnet_peer_addref_locked(lp);
1523         }
1524
1525         /* Add peer to global peer list, if necessary */
1526         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1527         if (list_empty(&lp->lp_peer_list)) {
1528                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1529                 ptable->pt_peers++;
1530         }
1531
1532
1533         /* Update peer state */
1534         spin_lock(&lp->lp_lock);
1535         if (flags & LNET_PEER_CONFIGURED) {
1536                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1537                         lp->lp_state |= LNET_PEER_CONFIGURED;
1538         }
1539         if (flags & LNET_PEER_MULTI_RAIL) {
1540                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1541                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1542                         lnet_peer_clr_non_mr_pref_nids(lp);
1543                 }
1544         }
1545         spin_unlock(&lp->lp_lock);
1546
1547         lp->lp_nnis++;
1548
1549         /* apply UDSPs */
1550         if (new_lpn) {
1551                 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1552                 if (rc)
1553                         CERROR("Failed to apply UDSPs on lpn %s\n",
1554                                libcfs_net2str(lpn->lpn_net_id));
1555         }
1556         rc = lnet_udsp_apply_policies_on_lpni(lpni);
1557         if (rc)
1558                 CERROR("Failed to apply UDSPs on lpni %s\n",
1559                        libcfs_nidstr(&lpni->lpni_nid));
1560
1561         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1562                libcfs_nidstr(&lp->lp_primary_nid),
1563                libcfs_nidstr(&lpni->lpni_nid), flags);
1564         lnet_peer_ni_decref_locked(lpni);
1565         lnet_net_unlock(LNET_LOCK_EX);
1566
1567         return 0;
1568 }
1569
1570 /*
1571  * Create a new peer, with nid as its primary nid.
1572  *
1573  * Call with the lnet_api_mutex held.
1574  */
1575 static int
1576 lnet_peer_add(struct lnet_nid *nid, unsigned int flags)
1577 {
1578         struct lnet_peer *lp;
1579         struct lnet_peer_net *lpn;
1580         struct lnet_peer_ni *lpni;
1581         int rc = 0;
1582
1583         LASSERT(nid);
1584
1585         /*
1586          * No need for the lnet_net_lock here, because the
1587          * lnet_api_mutex is held.
1588          */
1589         lpni = lnet_peer_ni_find_locked(nid);
1590         if (lpni) {
1591                 /* A peer with this NID already exists. */
1592                 lp = lpni->lpni_peer_net->lpn_peer;
1593                 lnet_peer_ni_decref_locked(lpni);
1594                 /*
1595                  * This is an error if the peer was configured and the
1596                  * primary NID differs or an attempt is made to change
1597                  * the Multi-Rail flag. Otherwise the assumption is
1598                  * that an existing peer is being modified.
1599                  */
1600                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1601                         if (!nid_same(&lp->lp_primary_nid, nid))
1602                                 rc = -EEXIST;
1603                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1604                                 rc = -EPERM;
1605                         goto out;
1606                 } else if (!(flags & LNET_PEER_CONFIGURED)) {
1607                         if (nid_same(&lp->lp_primary_nid, nid)) {
1608                                 rc = -EEXIST;
1609                                 goto out;
1610                         }
1611                 }
1612                 /* Delete and recreate as a configured peer. */
1613                 rc = lnet_peer_del(lp);
1614                 if (rc)
1615                         goto out;
1616         }
1617
1618         /* Create peer, peer_net, and peer_ni. */
1619         rc = -ENOMEM;
1620         lp = lnet_peer_alloc(nid);
1621         if (!lp)
1622                 goto out;
1623         lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1624         if (!lpn)
1625                 goto out_free_lp;
1626         lpni = lnet_peer_ni_alloc(nid);
1627         if (!lpni)
1628                 goto out_free_lpn;
1629
1630         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1631
1632 out_free_lpn:
1633         LIBCFS_FREE(lpn, sizeof(*lpn));
1634 out_free_lp:
1635         LIBCFS_FREE(lp, sizeof(*lp));
1636 out:
1637         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1638                libcfs_nidstr(nid), flags, rc);
1639         return rc;
1640 }
1641
1642 /*
1643  * Add a NID to a peer. Call with ln_api_mutex held.
1644  *
1645  * Error codes:
1646  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1647  *  -EEXIST:   The NID was configured by DLC for a different peer.
1648  *  -ENOMEM:   Out of memory.
1649  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1650  *             non-multi-rail peer.
1651  */
1652 static int
1653 lnet_peer_add_nid(struct lnet_peer *lp, struct lnet_nid *nid,
1654                   unsigned int flags)
1655 {
1656         struct lnet_peer_net *lpn;
1657         struct lnet_peer_ni *lpni;
1658         int rc = 0;
1659
1660         LASSERT(lp);
1661         LASSERT(nid);
1662
1663         /* A configured peer can only be updated through configuration. */
1664         if (!(flags & LNET_PEER_CONFIGURED)) {
1665                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1666                         rc = -EPERM;
1667                         goto out;
1668                 }
1669         }
1670
1671         /*
1672          * The MULTI_RAIL flag can be set but not cleared, because
1673          * that would leave the peer struct in an invalid state.
1674          */
1675         if (flags & LNET_PEER_MULTI_RAIL) {
1676                 spin_lock(&lp->lp_lock);
1677                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1678                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1679                         lnet_peer_clr_non_mr_pref_nids(lp);
1680                 }
1681                 spin_unlock(&lp->lp_lock);
1682         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1683                 rc = -EPERM;
1684                 goto out;
1685         }
1686
1687         lpni = lnet_peer_ni_find_locked(nid);
1688         if (lpni) {
1689                 /*
1690                  * A peer_ni already exists. This is only a problem if
1691                  * it is not connected to this peer and was configured
1692                  * by DLC.
1693                  */
1694                 if (lpni->lpni_peer_net->lpn_peer == lp)
1695                         goto out_free_lpni;
1696                 if (lnet_peer_ni_is_configured(lpni)) {
1697                         rc = -EEXIST;
1698                         goto out_free_lpni;
1699                 }
1700                 /* If this is the primary NID, destroy the peer. */
1701                 if (lnet_peer_ni_is_primary(lpni)) {
1702                         struct lnet_peer *rtr_lp =
1703                                 lpni->lpni_peer_net->lpn_peer;
1704                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1705                         /*
1706                          * if we're trying to delete a router it means
1707                          * we're moving this peer NI to a new peer so must
1708                          * transfer router properties to the new peer
1709                          */
1710                         if (rtr_refcount > 0) {
1711                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1712                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1713                         }
1714                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1715                         lnet_peer_ni_decref_locked(lpni);
1716                         lpni = lnet_peer_ni_alloc(nid);
1717                         if (!lpni) {
1718                                 rc = -ENOMEM;
1719                                 goto out_free_lpni;
1720                         }
1721                 }
1722         } else {
1723                 lpni = lnet_peer_ni_alloc(nid);
1724                 if (!lpni) {
1725                         rc = -ENOMEM;
1726                         goto out_free_lpni;
1727                 }
1728         }
1729
1730         /*
1731          * Get the peer_net. Check that we're not adding a second
1732          * peer_ni on a peer_net of a non-multi-rail peer.
1733          */
1734         lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
1735         if (!lpn) {
1736                 lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1737                 if (!lpn) {
1738                         rc = -ENOMEM;
1739                         goto out_free_lpni;
1740                 }
1741         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1742                 rc = -ENOTUNIQ;
1743                 goto out_free_lpni;
1744         }
1745
1746         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1747
1748 out_free_lpni:
1749         lnet_peer_ni_decref_locked(lpni);
1750 out:
1751         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1752                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid),
1753                flags, rc);
1754         return rc;
1755 }
1756
1757 /*
1758  * Update the primary NID of a peer, if possible.
1759  *
1760  * Call with the lnet_api_mutex held.
1761  */
1762 static int
1763 lnet_peer_set_primary_nid(struct lnet_peer *lp, struct lnet_nid *nid,
1764                           unsigned int flags)
1765 {
1766         struct lnet_nid old = lp->lp_primary_nid;
1767         int rc = 0;
1768
1769         if (nid_same(&lp->lp_primary_nid, nid))
1770                 goto out;
1771
1772         lp->lp_primary_nid = *nid;
1773
1774         rc = lnet_peer_add_nid(lp, nid, flags);
1775         if (rc) {
1776                 lp->lp_primary_nid = old;
1777                 goto out;
1778         }
1779 out:
1780         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1781                libcfs_nidstr(&old), libcfs_nidstr(nid), rc);
1782
1783         return rc;
1784 }
1785
1786 /*
1787  * lpni creation initiated due to traffic either sending or receiving.
1788  * Callers must hold ln_api_mutex
1789  * Ref taken on lnet_peer_ni returned by this function
1790  */
1791 static struct lnet_peer_ni *
1792 lnet_peer_ni_traffic_add(struct lnet_nid *nid, struct lnet_nid *pref)
1793 __must_hold(&the_lnet.ln_api_mutex)
1794 {
1795         struct lnet_peer *lp = NULL;
1796         struct lnet_peer_net *lpn = NULL;
1797         struct lnet_peer_ni *lpni;
1798         unsigned flags = 0;
1799         int rc = 0;
1800
1801         if (LNET_NID_IS_ANY(nid)) {
1802                 rc = -EINVAL;
1803                 goto out_err;
1804         }
1805
1806         /* lnet_net_lock is not needed here because ln_api_lock is held */
1807         lpni = lnet_peer_ni_find_locked(nid);
1808         if (lpni) {
1809                 /*
1810                  * We must have raced with another thread. Since we
1811                  * know next to nothing about a peer_ni created by
1812                  * traffic, we just assume everything is ok and
1813                  * return.
1814                  */
1815                 goto out;
1816         }
1817
1818         /* Create peer, peer_net, and peer_ni. */
1819         rc = -ENOMEM;
1820         lp = lnet_peer_alloc(nid);
1821         if (!lp)
1822                 goto out_err;
1823         lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1824         if (!lpn)
1825                 goto out_err;
1826         lpni = lnet_peer_ni_alloc(nid);
1827         if (!lpni)
1828                 goto out_err;
1829         lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1830
1831         /* lnet_peer_attach_peer_ni() always returns 0 */
1832         rc = lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1833
1834         lnet_peer_ni_addref_locked(lpni);
1835
1836 out_err:
1837         if (rc) {
1838                 if (lpn)
1839                         LIBCFS_FREE(lpn, sizeof(*lpn));
1840                 if (lp)
1841                         LIBCFS_FREE(lp, sizeof(*lp));
1842                 lpni = ERR_PTR(rc);
1843         }
1844 out:
1845         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(nid), rc);
1846         return lpni;
1847 }
1848
1849 /*
1850  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1851  *
1852  * This API handles the following combinations:
1853  *   Create a peer with its primary NI if only the prim_nid is provided
1854  *   Add a NID to a peer identified by the prim_nid. The peer identified
1855  *   by the prim_nid must already exist.
1856  *   The peer being created may be non-MR.
1857  *
1858  * The caller must hold ln_api_mutex. This prevents the peer from
1859  * being created/modified/deleted by a different thread.
1860  */
1861 int
1862 lnet_add_peer_ni(lnet_nid_t prim_nid4, lnet_nid_t nid4, bool mr, bool temp)
1863 {
1864         struct lnet_nid prim_nid, nid;
1865         struct lnet_peer *lp = NULL;
1866         struct lnet_peer_ni *lpni;
1867         unsigned int flags = 0;
1868
1869         /* The prim_nid must always be specified */
1870         if (prim_nid4 == LNET_NID_ANY)
1871                 return -EINVAL;
1872
1873         lnet_nid4_to_nid(prim_nid4, &prim_nid);
1874         lnet_nid4_to_nid(nid4, &nid);
1875
1876         if (!temp)
1877                 flags = LNET_PEER_CONFIGURED;
1878
1879         if (mr)
1880                 flags |= LNET_PEER_MULTI_RAIL;
1881
1882         /*
1883          * If nid isn't specified, we must create a new peer with
1884          * prim_nid as its primary nid.
1885          */
1886         if (nid4 == LNET_NID_ANY)
1887                 return lnet_peer_add(&prim_nid, flags);
1888
1889         /* Look up the prim_nid, which must exist. */
1890         lpni = lnet_peer_ni_find_locked(&prim_nid);
1891         if (!lpni)
1892                 return -ENOENT;
1893         lnet_peer_ni_decref_locked(lpni);
1894         lp = lpni->lpni_peer_net->lpn_peer;
1895
1896         /* Peer must have been configured. */
1897         if (!temp && !(lp->lp_state & LNET_PEER_CONFIGURED)) {
1898                 CDEBUG(D_NET, "peer %s was not configured\n",
1899                        libcfs_nidstr(&prim_nid));
1900                 return -ENOENT;
1901         }
1902
1903         /* Primary NID must match */
1904         if (!nid_same(&lp->lp_primary_nid, &prim_nid)) {
1905                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1906                        libcfs_nidstr(&prim_nid),
1907                        libcfs_nidstr(&lp->lp_primary_nid));
1908                 return -ENODEV;
1909         }
1910
1911         /* Multi-Rail flag must match. */
1912         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1913                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1914                        libcfs_nidstr(&prim_nid));
1915                 return -EPERM;
1916         }
1917
1918         return lnet_peer_add_nid(lp, &nid, flags);
1919 }
1920
1921 /*
1922  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1923  *
1924  * This API handles the following combinations:
1925  *   Delete a NI from a peer if both prim_nid and nid are provided.
1926  *   Delete a peer if only prim_nid is provided.
1927  *   Delete a peer if its primary nid is provided.
1928  *
1929  * The caller must hold ln_api_mutex. This prevents the peer from
1930  * being modified/deleted by a different thread.
1931  */
1932 int
1933 lnet_del_peer_ni(lnet_nid_t prim_nid4, lnet_nid_t nid)
1934 {
1935         struct lnet_peer *lp;
1936         struct lnet_peer_ni *lpni;
1937         unsigned int flags;
1938         struct lnet_nid prim_nid;
1939
1940         if (prim_nid4 == LNET_NID_ANY)
1941                 return -EINVAL;
1942         lnet_nid4_to_nid(prim_nid4, &prim_nid);
1943
1944         lpni = lnet_peer_ni_find_locked(&prim_nid);
1945         if (!lpni)
1946                 return -ENOENT;
1947         lnet_peer_ni_decref_locked(lpni);
1948         lp = lpni->lpni_peer_net->lpn_peer;
1949
1950         if (!nid_same(&prim_nid, &lp->lp_primary_nid)) {
1951                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1952                        libcfs_nidstr(&prim_nid),
1953                        libcfs_nidstr(&lp->lp_primary_nid));
1954                 return -ENODEV;
1955         }
1956
1957         lnet_net_lock(LNET_LOCK_EX);
1958         if (lp->lp_rtr_refcount > 0) {
1959                 lnet_net_unlock(LNET_LOCK_EX);
1960                 CERROR("%s is a router. Can not be deleted\n",
1961                        libcfs_nidstr(&prim_nid));
1962                 return -EBUSY;
1963         }
1964         lnet_net_unlock(LNET_LOCK_EX);
1965
1966         if (nid == LNET_NID_ANY || nid == lnet_nid_to_nid4(&lp->lp_primary_nid))
1967                 return lnet_peer_del(lp);
1968
1969         flags = LNET_PEER_CONFIGURED;
1970         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1971                 flags |= LNET_PEER_MULTI_RAIL;
1972
1973         return lnet_peer_del_nid(lp, nid, flags);
1974 }
1975
1976 void
1977 lnet_destroy_peer_ni_locked(struct kref *ref)
1978 {
1979         struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
1980                                                  lpni_kref);
1981         struct lnet_peer_table *ptable;
1982         struct lnet_peer_net *lpn;
1983
1984         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
1985
1986         LASSERT(kref_read(&lpni->lpni_kref) == 0);
1987         LASSERT(list_empty(&lpni->lpni_txq));
1988         LASSERT(lpni->lpni_txqnob == 0);
1989         LASSERT(list_empty(&lpni->lpni_peer_nis));
1990         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1991
1992         lpn = lpni->lpni_peer_net;
1993         lpni->lpni_peer_net = NULL;
1994         lpni->lpni_net = NULL;
1995
1996         if (!list_empty(&lpni->lpni_hashlist)) {
1997                 /* remove the peer ni from the zombie list */
1998                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1999                 spin_lock(&ptable->pt_zombie_lock);
2000                 list_del_init(&lpni->lpni_hashlist);
2001                 ptable->pt_zombies--;
2002                 spin_unlock(&ptable->pt_zombie_lock);
2003         }
2004
2005         if (lpni->lpni_pref_nnids > 1) {
2006                 struct lnet_nid_list *ne, *tmp;
2007
2008                 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
2009                                          nl_list) {
2010                         list_del_init(&ne->nl_list);
2011                         LIBCFS_FREE(ne, sizeof(*ne));
2012                 }
2013         }
2014         LIBCFS_FREE(lpni, sizeof(*lpni));
2015
2016         if (lpn)
2017                 lnet_peer_net_decref_locked(lpn);
2018 }
2019
2020 struct lnet_peer_ni *
2021 lnet_nid2peerni_ex(struct lnet_nid *nid)
2022 __must_hold(&the_lnet.ln_api_mutex)
2023 {
2024         struct lnet_peer_ni *lpni = NULL;
2025
2026         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2027                 return ERR_PTR(-ESHUTDOWN);
2028
2029         /*
2030          * find if a peer_ni already exists.
2031          * If so then just return that.
2032          */
2033         lpni = lnet_peer_ni_find_locked(nid);
2034         if (lpni)
2035                 return lpni;
2036
2037         lnet_net_unlock(LNET_LOCK_EX);
2038
2039         lpni = lnet_peer_ni_traffic_add(nid, NULL);
2040
2041         lnet_net_lock(LNET_LOCK_EX);
2042
2043         return lpni;
2044 }
2045
2046 /*
2047  * Get a peer_ni for the given nid, create it if necessary. Takes a
2048  * hold on the peer_ni.
2049  */
2050 struct lnet_peer_ni *
2051 lnet_peerni_by_nid_locked(struct lnet_nid *nid,
2052                         struct lnet_nid *pref, int cpt)
2053 {
2054         struct lnet_peer_ni *lpni = NULL;
2055
2056         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2057                 return ERR_PTR(-ESHUTDOWN);
2058
2059         /*
2060          * find if a peer_ni already exists.
2061          * If so then just return that.
2062          */
2063         lpni = lnet_peer_ni_find_locked(nid);
2064         if (lpni)
2065                 return lpni;
2066
2067         /*
2068          * Slow path:
2069          * use the lnet_api_mutex to serialize the creation of the peer_ni
2070          * and the creation/deletion of the local ni/net. When a local ni is
2071          * created, if there exists a set of peer_nis on that network,
2072          * they need to be traversed and updated. When a local NI is
2073          * deleted, which could result in a network being deleted, then
2074          * all peer nis on that network need to be removed as well.
2075          *
2076          * Creation through traffic should also be serialized with
2077          * creation through DLC.
2078          */
2079         lnet_net_unlock(cpt);
2080         mutex_lock(&the_lnet.ln_api_mutex);
2081         /*
2082          * the_lnet.ln_state is only modified under the ln_api_lock, so a single
2083          * check here is sufficent.
2084          */
2085         if (the_lnet.ln_state == LNET_STATE_RUNNING)
2086                 lpni = lnet_peer_ni_traffic_add(nid, pref);
2087
2088         mutex_unlock(&the_lnet.ln_api_mutex);
2089         lnet_net_lock(cpt);
2090
2091         /* Lock has been dropped, check again for shutdown. */
2092         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2093                 if (!IS_ERR_OR_NULL(lpni))
2094                         lnet_peer_ni_decref_locked(lpni);
2095                 lpni = ERR_PTR(-ESHUTDOWN);
2096         }
2097
2098         return lpni;
2099 }
2100
2101 bool
2102 lnet_peer_gw_discovery(struct lnet_peer *lp)
2103 {
2104         bool rc = false;
2105
2106         spin_lock(&lp->lp_lock);
2107         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2108                 rc = true;
2109         spin_unlock(&lp->lp_lock);
2110
2111         return rc;
2112 }
2113
2114 bool
2115 lnet_peer_is_uptodate(struct lnet_peer *lp)
2116 {
2117         bool rc;
2118
2119         spin_lock(&lp->lp_lock);
2120         rc = lnet_peer_is_uptodate_locked(lp);
2121         spin_unlock(&lp->lp_lock);
2122         return rc;
2123 }
2124
2125 /*
2126  * Is a peer uptodate from the point of view of discovery?
2127  *
2128  * If it is currently being processed, obviously not.
2129  * A forced Ping or Push is also handled by the discovery thread.
2130  *
2131  * Otherwise look at whether the peer needs rediscovering.
2132  */
2133 bool
2134 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2135 __must_hold(&lp->lp_lock)
2136 {
2137         bool rc;
2138
2139         if (lp->lp_state & (LNET_PEER_DISCOVERING |
2140                             LNET_PEER_FORCE_PING |
2141                             LNET_PEER_FORCE_PUSH)) {
2142                 rc = false;
2143         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2144                 rc = false;
2145         } else if (lnet_peer_needs_push(lp)) {
2146                 rc = false;
2147         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2148                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2149                         rc = true;
2150                 else
2151                         rc = false;
2152         } else {
2153                 rc = false;
2154         }
2155
2156         return rc;
2157 }
2158
2159 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2160 void
2161 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2162 {
2163         /* The discovery thread holds net_lock/EX and lp_lock when it splices
2164          * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2165          * when adding to the list and queuing the peer to ensure that we do not
2166          * strand any messages on the lp_dc_pendq. This scheme ensures the
2167          * message will be resent even if the peer is already being discovered.
2168          * Therefore we needn't check the return value of
2169          * lnet_peer_queue_for_discovery(lp).
2170          */
2171         lnet_net_lock(LNET_LOCK_EX);
2172         spin_lock(&lp->lp_lock);
2173         list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2174         spin_unlock(&lp->lp_lock);
2175         lnet_peer_queue_for_discovery(lp);
2176         lnet_net_unlock(LNET_LOCK_EX);
2177 }
2178
2179 /*
2180  * Queue a peer for the attention of the discovery thread.  Call with
2181  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2182  * -EALREADY if the peer was already queued.
2183  */
2184 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2185 {
2186         int rc;
2187
2188         spin_lock(&lp->lp_lock);
2189         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2190                 lp->lp_state |= LNET_PEER_DISCOVERING;
2191         spin_unlock(&lp->lp_lock);
2192         if (list_empty(&lp->lp_dc_list)) {
2193                 lnet_peer_addref_locked(lp);
2194                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2195                 wake_up(&the_lnet.ln_dc_waitq);
2196                 rc = 0;
2197         } else {
2198                 rc = -EALREADY;
2199         }
2200
2201         CDEBUG(D_NET, "Queue peer %s: %d\n",
2202                libcfs_nidstr(&lp->lp_primary_nid), rc);
2203
2204         return rc;
2205 }
2206
2207 /*
2208  * Discovery of a peer is complete. Wake all waiters on the peer.
2209  * Call with lnet_net_lock/EX held.
2210  */
2211 static void lnet_peer_discovery_complete(struct lnet_peer *lp, int dc_error)
2212 {
2213         struct lnet_msg *msg, *tmp;
2214         int rc = 0;
2215         LIST_HEAD(pending_msgs);
2216
2217         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2218                libcfs_nidstr(&lp->lp_primary_nid));
2219
2220         list_del_init(&lp->lp_dc_list);
2221         spin_lock(&lp->lp_lock);
2222         if (dc_error) {
2223                 lp->lp_dc_error = dc_error;
2224                 lp->lp_state &= ~LNET_PEER_DISCOVERING;
2225                 lp->lp_state |= LNET_PEER_REDISCOVER;
2226         }
2227         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2228         spin_unlock(&lp->lp_lock);
2229         wake_up(&lp->lp_dc_waitq);
2230
2231         if (lp->lp_rtr_refcount > 0)
2232                 lnet_router_discovery_complete(lp);
2233
2234         lnet_net_unlock(LNET_LOCK_EX);
2235
2236         /* iterate through all pending messages and send them again */
2237         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2238                 list_del_init(&msg->msg_list);
2239                 if (dc_error) {
2240                         lnet_finalize(msg, dc_error);
2241                         continue;
2242                 }
2243
2244                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2245                        lnet_msgtyp2str(msg->msg_type),
2246                        libcfs_idstr(&msg->msg_target));
2247                 rc = lnet_send(&msg->msg_src_nid_param, msg,
2248                                &msg->msg_rtr_nid_param);
2249                 if (rc < 0) {
2250                         CNETERR("Error sending %s to %s: %d\n",
2251                                lnet_msgtyp2str(msg->msg_type),
2252                                libcfs_idstr(&msg->msg_target), rc);
2253                         lnet_finalize(msg, rc);
2254                 }
2255         }
2256         lnet_net_lock(LNET_LOCK_EX);
2257         lnet_peer_decref_locked(lp);
2258 }
2259
2260 /*
2261  * Handle inbound push.
2262  * Like any event handler, called with lnet_res_lock/CPT held.
2263  */
2264 void lnet_peer_push_event(struct lnet_event *ev)
2265 {
2266         struct lnet_ping_buffer *pbuf;
2267         struct lnet_peer *lp;
2268
2269         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2270
2271         /* lnet_find_peer() adds a refcount */
2272         lp = lnet_find_peer(&ev->source.nid);
2273         if (!lp) {
2274                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2275                        libcfs_nidstr(&ev->initiator.nid),
2276                        libcfs_nidstr(&ev->source.nid));
2277                 pbuf->pb_needs_post = true;
2278                 return;
2279         }
2280
2281         /* Ensure peer state remains consistent while we modify it. */
2282         spin_lock(&lp->lp_lock);
2283
2284         /*
2285          * If some kind of error happened the contents of the message
2286          * cannot be used. Clear the NIDS_UPTODATE and set the
2287          * FORCE_PING flag to trigger a ping.
2288          */
2289         if (ev->status) {
2290                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2291                 lp->lp_state |= LNET_PEER_FORCE_PING;
2292                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2293                        ev->status,
2294                        libcfs_nidstr(&lp->lp_primary_nid),
2295                        libcfs_nidstr(&ev->source.nid));
2296                 goto out;
2297         }
2298
2299         /*
2300          * A push with invalid or corrupted info. Clear the UPTODATE
2301          * flag to trigger a ping.
2302          */
2303         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2304                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2305                 lp->lp_state |= LNET_PEER_FORCE_PING;
2306                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2307                        libcfs_nidstr(&lp->lp_primary_nid));
2308                 goto out;
2309         }
2310
2311         /*
2312          * Make sure we'll allocate the correct size ping buffer when
2313          * pinging the peer.
2314          */
2315         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2316                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2317
2318         /*
2319          * A non-Multi-Rail peer is not supposed to be capable of
2320          * sending a push.
2321          */
2322         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2323                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2324                        libcfs_nidstr(&lp->lp_primary_nid));
2325                 goto out;
2326         }
2327
2328         /*
2329          * The peer may have discovery disabled at its end. Set
2330          * NO_DISCOVERY as appropriate.
2331          */
2332         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2333                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2334                        libcfs_nidstr(&lp->lp_primary_nid));
2335                 /*
2336                  * Mark the peer for deletion if we already know about it
2337                  * and it's going from discovery set to no discovery set
2338                  */
2339                 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2340                                       LNET_PEER_DISCOVERING)) &&
2341                      lp->lp_state & LNET_PEER_DISCOVERED) {
2342                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2343                                libcfs_nidstr(&lp->lp_primary_nid),
2344                                lp->lp_state);
2345                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2346                 }
2347                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2348         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2349                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2350                        libcfs_nidstr(&lp->lp_primary_nid));
2351                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2352         }
2353
2354         /*
2355          * Update the MULTI_RAIL flag based on the push. If the peer
2356          * was configured with DLC then the setting should match what
2357          * DLC put in.
2358          * NB: We verified above that the MR feature bit is set in pi_features
2359          */
2360         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2361                 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2362                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2363         } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2364                 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2365                       libcfs_nidstr(&lp->lp_primary_nid));
2366         } else if (lnet_peer_discovery_disabled) {
2367                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2368                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2369         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2370                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2371                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2372         } else {
2373                 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2374                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2375                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2376                 lnet_peer_clr_non_mr_pref_nids(lp);
2377         }
2378
2379         /*
2380          * Check for truncation of the Put message. Clear the
2381          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2382          * and tell discovery to allocate a bigger buffer.
2383          */
2384         if (ev->mlength < ev->rlength) {
2385                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2386                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2387                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2388                 lp->lp_state |= LNET_PEER_FORCE_PING;
2389                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2390                        libcfs_nidstr(&lp->lp_primary_nid),
2391                        pbuf->pb_info.pi_nnis);
2392                 goto out;
2393         }
2394
2395         /* always assume new data */
2396         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2397         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2398
2399         /*
2400          * If there is data present that hasn't been processed yet,
2401          * we'll replace it if the Put contained newer data and it
2402          * fits. We're racing with a Ping or earlier Push in this
2403          * case.
2404          */
2405         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2406                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2407                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2408                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2409                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2410                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2411                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2412                               libcfs_nidstr(&lp->lp_primary_nid),
2413                               LNET_PING_BUFFER_SEQNO(pbuf),
2414                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2415                 }
2416                 goto out;
2417         }
2418
2419         /*
2420          * Allocate a buffer to copy the data. On a failure we drop
2421          * the Push and set FORCE_PING to force the discovery
2422          * thread to fix the problem by pinging the peer.
2423          */
2424         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2425         if (!lp->lp_data) {
2426                 lp->lp_state |= LNET_PEER_FORCE_PING;
2427                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2428                        libcfs_nidstr(&lp->lp_primary_nid),
2429                        LNET_PING_BUFFER_SEQNO(pbuf));
2430                 goto out;
2431         }
2432
2433         /* Success */
2434         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2435                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2436         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2437         CDEBUG(D_NET, "Received Push %s %u\n",
2438                libcfs_nidstr(&lp->lp_primary_nid),
2439                LNET_PING_BUFFER_SEQNO(pbuf));
2440
2441 out:
2442         /* We've processed this buffer. It can be reposted */
2443         pbuf->pb_needs_post = true;
2444
2445         /*
2446          * Queue the peer for discovery if not done, force it on the request
2447          * queue and wake the discovery thread if the peer was already queued,
2448          * because its status changed.
2449          */
2450         spin_unlock(&lp->lp_lock);
2451         lnet_net_lock(LNET_LOCK_EX);
2452         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2453                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2454                 wake_up(&the_lnet.ln_dc_waitq);
2455         }
2456         /* Drop refcount from lookup */
2457         lnet_peer_decref_locked(lp);
2458         lnet_net_unlock(LNET_LOCK_EX);
2459 }
2460
2461 /*
2462  * Clear the discovery error state, unless we're already discovering
2463  * this peer, in which case the error is current.
2464  */
2465 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2466 {
2467         spin_lock(&lp->lp_lock);
2468         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2469                 lp->lp_dc_error = 0;
2470         spin_unlock(&lp->lp_lock);
2471 }
2472
2473 /*
2474  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2475  * dropped/retaken within this function. An lnet_peer_ni is passed in
2476  * because discovery could tear down an lnet_peer.
2477  */
2478 int
2479 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2480 {
2481         DEFINE_WAIT(wait);
2482         struct lnet_peer *lp;
2483         int rc = 0;
2484         int count = 0;
2485
2486 again:
2487         lnet_net_unlock(cpt);
2488         lnet_net_lock(LNET_LOCK_EX);
2489         lp = lpni->lpni_peer_net->lpn_peer;
2490         lnet_peer_clear_discovery_error(lp);
2491
2492         /*
2493          * We're willing to be interrupted. The lpni can become a
2494          * zombie if we race with DLC, so we must check for that.
2495          */
2496         for (;;) {
2497                 /* Keep lp alive when the lnet_net_lock is unlocked */
2498                 lnet_peer_addref_locked(lp);
2499                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2500                 if (signal_pending(current))
2501                         break;
2502                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2503                         break;
2504                 /*
2505                  * Don't repeat discovery if discovery is disabled. This is
2506                  * done to ensure we can use discovery as a standard ping as
2507                  * well for backwards compatibility with routers which do not
2508                  * have discovery or have discovery disabled
2509                  */
2510                 if (lnet_is_discovery_disabled(lp) && count > 0)
2511                         break;
2512                 if (lp->lp_dc_error)
2513                         break;
2514                 if (lnet_peer_is_uptodate(lp))
2515                         break;
2516                 if (lp->lp_state & LNET_PEER_MARK_DELETED)
2517                         break;
2518                 lnet_peer_queue_for_discovery(lp);
2519                 count++;
2520                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2521
2522                 /*
2523                  * If caller requested a non-blocking operation then
2524                  * return immediately. Once discovery is complete any
2525                  * pending messages that were stopped due to discovery
2526                  * will be transmitted.
2527                  */
2528                 if (!block)
2529                         break;
2530
2531                 lnet_net_unlock(LNET_LOCK_EX);
2532                 schedule();
2533                 finish_wait(&lp->lp_dc_waitq, &wait);
2534                 lnet_net_lock(LNET_LOCK_EX);
2535                 lnet_peer_decref_locked(lp);
2536                 /* Peer may have changed */
2537                 lp = lpni->lpni_peer_net->lpn_peer;
2538         }
2539         finish_wait(&lp->lp_dc_waitq, &wait);
2540
2541         lnet_net_unlock(LNET_LOCK_EX);
2542         lnet_net_lock(cpt);
2543         lnet_peer_decref_locked(lp);
2544         /*
2545          * The peer may have changed, so re-check and rediscover if that turns
2546          * out to have been the case. The reference count on lp ensured that
2547          * even if it was unlinked from lpni the memory could not be recycled.
2548          * Thus the check below is sufficient to determine whether the peer
2549          * changed. If the peer changed, then lp must not be dereferenced.
2550          */
2551         if (lp != lpni->lpni_peer_net->lpn_peer)
2552                 goto again;
2553
2554         if (signal_pending(current))
2555                 rc = -EINTR;
2556         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2557                 rc = -ESHUTDOWN;
2558         else if (lp->lp_dc_error)
2559                 rc = lp->lp_dc_error;
2560         else if (!block)
2561                 CDEBUG(D_NET, "non-blocking discovery\n");
2562         else if (!lnet_peer_is_uptodate(lp) &&
2563                  !(lnet_is_discovery_disabled(lp) ||
2564                    (lp->lp_state & LNET_PEER_MARK_DELETED)))
2565                 goto again;
2566
2567         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2568                (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2569                libcfs_nidstr(&lpni->lpni_nid), rc,
2570                (!block) ? "pending discovery" : "discovery complete");
2571
2572         return rc;
2573 }
2574
2575 /* Handle an incoming ack for a push. */
2576 static void
2577 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2578 {
2579         struct lnet_ping_buffer *pbuf;
2580
2581         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2582         spin_lock(&lp->lp_lock);
2583         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2584         lp->lp_push_error = ev->status;
2585         if (ev->status)
2586                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2587         else
2588                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2589         spin_unlock(&lp->lp_lock);
2590
2591         CDEBUG(D_NET, "peer %s ev->status %d\n",
2592                libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2593 }
2594
2595 /* Handle a Reply message. This is the reply to a Ping message. */
2596 static void
2597 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2598 {
2599         struct lnet_ping_buffer *pbuf;
2600         int rc;
2601
2602         spin_lock(&lp->lp_lock);
2603
2604         lp->lp_disc_src_nid = ev->target.nid;
2605         lp->lp_disc_dst_nid = ev->source.nid;
2606
2607         /*
2608          * If some kind of error happened the contents of message
2609          * cannot be used. Set PING_FAILED to trigger a retry.
2610          */
2611         if (ev->status) {
2612                 lp->lp_state |= LNET_PEER_PING_FAILED;
2613                 lp->lp_ping_error = ev->status;
2614                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2615                        ev->status,
2616                        libcfs_nidstr(&lp->lp_primary_nid),
2617                        libcfs_nidstr(&ev->source.nid));
2618                 goto out;
2619         }
2620
2621         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2622         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2623                 lnet_swap_pinginfo(pbuf);
2624
2625         /*
2626          * A reply with invalid or corrupted info. Set PING_FAILED to
2627          * trigger a retry.
2628          */
2629         rc = lnet_ping_info_validate(&pbuf->pb_info);
2630         if (rc) {
2631                 lp->lp_state |= LNET_PEER_PING_FAILED;
2632                 lp->lp_ping_error = 0;
2633                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2634                        libcfs_nidstr(&lp->lp_primary_nid), rc);
2635                 goto out;
2636         }
2637
2638         /*
2639          * The peer may have discovery disabled at its end. Set
2640          * NO_DISCOVERY as appropriate.
2641          */
2642         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) ||
2643             lnet_peer_discovery_disabled) {
2644                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2645                        libcfs_nidstr(&lp->lp_primary_nid));
2646
2647                 /* Detect whether this peer has toggled discovery from on to
2648                  * off and whether we can delete and re-create the peer. Peers
2649                  * that were manually configured cannot be deleted by discovery.
2650                  * We need to delete this peer and re-create it if the peer was
2651                  * not configured manually, is currently considered DD capable,
2652                  * and either:
2653                  * 1. We've already discovered the peer (the peer has toggled
2654                  *    the discovery feature from on to off), or
2655                  * 2. The peer is considered MR, but it was not user configured
2656                  *    (this was a "temporary" peer created via the kernel APIs
2657                  *     that we're discovering for the first time)
2658                  */
2659                 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2660                                       LNET_PEER_NO_DISCOVERY)) &&
2661                     (lp->lp_state & (LNET_PEER_DISCOVERED |
2662                                      LNET_PEER_MULTI_RAIL))) {
2663                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2664                                libcfs_nidstr(&lp->lp_primary_nid),
2665                                lp->lp_state);
2666                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2667                 }
2668                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2669         } else {
2670                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2671                        libcfs_nidstr(&lp->lp_primary_nid));
2672                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2673         }
2674
2675         /*
2676          * Update the MULTI_RAIL flag based on the reply. If the peer
2677          * was configured with DLC then the setting should match what
2678          * DLC put in.
2679          */
2680         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2681                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2682                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2683                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2684                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2685                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2686                               libcfs_nidstr(&lp->lp_primary_nid));
2687                 } else if (lnet_peer_discovery_disabled) {
2688                         CDEBUG(D_NET,
2689                                "peer %s(%p) not MR: DD disabled locally\n",
2690                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2691                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2692                         CDEBUG(D_NET,
2693                                "peer %s(%p) not MR: DD disabled remotely\n",
2694                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2695                 } else {
2696                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2697                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2698                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2699                         lnet_peer_clr_non_mr_pref_nids(lp);
2700                 }
2701         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2702                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2703                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2704                               libcfs_nidstr(&lp->lp_primary_nid));
2705                 } else {
2706                         CERROR("Multi-Rail state vanished from %s\n",
2707                                libcfs_nidstr(&lp->lp_primary_nid));
2708                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2709                 }
2710         }
2711
2712         /*
2713          * Make sure we'll allocate the correct size ping buffer when
2714          * pinging the peer.
2715          */
2716         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2717                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2718
2719         /*
2720          * Check for truncation of the Reply. Clear PING_SENT and set
2721          * PING_FAILED to trigger a retry.
2722          */
2723         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2724                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2725                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2726                 lp->lp_state |= LNET_PEER_PING_FAILED;
2727                 lp->lp_ping_error = 0;
2728                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2729                        libcfs_nidstr(&lp->lp_primary_nid),
2730                        pbuf->pb_info.pi_nnis);
2731                 goto out;
2732         }
2733
2734         /*
2735          * Check the sequence numbers in the reply. These are only
2736          * available if the reply came from a Multi-Rail peer.
2737          */
2738         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2739             pbuf->pb_info.pi_nnis > 1 &&
2740             lnet_nid_to_nid4(&lp->lp_primary_nid) ==
2741             pbuf->pb_info.pi_ni[1].ns_nid) {
2742                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2743                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2744                                 libcfs_nidstr(&lp->lp_primary_nid),
2745                                 LNET_PING_BUFFER_SEQNO(pbuf),
2746                                 lp->lp_peer_seqno);
2747
2748                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2749         }
2750
2751         /* We're happy with the state of the data in the buffer. */
2752         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2753                libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2754                lp->lp_state);
2755         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2756                 lnet_ping_buffer_decref(lp->lp_data);
2757         else
2758                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2759         lnet_ping_buffer_addref(pbuf);
2760         lp->lp_data = pbuf;
2761 out:
2762         lp->lp_state &= ~LNET_PEER_PING_SENT;
2763         spin_unlock(&lp->lp_lock);
2764
2765         lnet_net_lock(LNET_LOCK_EX);
2766         /*
2767          * If this peer is a gateway, call the routing callback to
2768          * handle the ping reply
2769          */
2770         if (lp->lp_rtr_refcount > 0)
2771                 lnet_router_discovery_ping_reply(lp);
2772         lnet_net_unlock(LNET_LOCK_EX);
2773 }
2774
2775 /*
2776  * Send event handling. Only matters for error cases, where we clean
2777  * up state on the peer and peer_ni that would otherwise be updated in
2778  * the REPLY event handler for a successful Ping, and the ACK event
2779  * handler for a successful Push.
2780  */
2781 static int
2782 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2783 {
2784         int rc = 0;
2785
2786         if (!ev->status)
2787                 goto out;
2788
2789         spin_lock(&lp->lp_lock);
2790         if (ev->msg_type == LNET_MSG_GET) {
2791                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2792                 lp->lp_state |= LNET_PEER_PING_FAILED;
2793                 lp->lp_ping_error = ev->status;
2794         } else { /* ev->msg_type == LNET_MSG_PUT */
2795                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2796                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2797                 lp->lp_push_error = ev->status;
2798         }
2799         spin_unlock(&lp->lp_lock);
2800         rc = LNET_REDISCOVER_PEER;
2801 out:
2802         CDEBUG(D_NET, "%s Send to %s: %d\n",
2803                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2804                 libcfs_nidstr(&ev->target.nid), rc);
2805         return rc;
2806 }
2807
2808 /*
2809  * Unlink event handling. This event is only seen if a call to
2810  * LNetMDUnlink() caused the event to be unlinked. If this call was
2811  * made after the event was set up in LNetGet() or LNetPut() then we
2812  * assume the Ping or Push timed out.
2813  */
2814 static void
2815 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2816 {
2817         spin_lock(&lp->lp_lock);
2818         /* We've passed through LNetGet() */
2819         if (lp->lp_state & LNET_PEER_PING_SENT) {
2820                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2821                 lp->lp_state |= LNET_PEER_PING_FAILED;
2822                 lp->lp_ping_error = -ETIMEDOUT;
2823                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2824                         libcfs_nidstr(&lp->lp_primary_nid));
2825         }
2826         /* We've passed through LNetPut() */
2827         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2828                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2829                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2830                 lp->lp_push_error = -ETIMEDOUT;
2831                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2832                         libcfs_nidstr(&lp->lp_primary_nid));
2833         }
2834         spin_unlock(&lp->lp_lock);
2835 }
2836
2837 /*
2838  * Event handler for the discovery EQ.
2839  *
2840  * Called with lnet_res_lock(cpt) held. The cpt is the
2841  * lnet_cpt_of_cookie() of the md handle cookie.
2842  */
2843 static void lnet_discovery_event_handler(struct lnet_event *event)
2844 {
2845         struct lnet_peer *lp = event->md_user_ptr;
2846         struct lnet_ping_buffer *pbuf;
2847         int rc;
2848
2849         /* discovery needs to take another look */
2850         rc = LNET_REDISCOVER_PEER;
2851
2852         CDEBUG(D_NET, "Received event: %d\n", event->type);
2853
2854         switch (event->type) {
2855         case LNET_EVENT_ACK:
2856                 lnet_discovery_event_ack(lp, event);
2857                 break;
2858         case LNET_EVENT_REPLY:
2859                 lnet_discovery_event_reply(lp, event);
2860                 break;
2861         case LNET_EVENT_SEND:
2862                 /* Only send failure triggers a retry. */
2863                 rc = lnet_discovery_event_send(lp, event);
2864                 break;
2865         case LNET_EVENT_UNLINK:
2866                 /* LNetMDUnlink() was called */
2867                 lnet_discovery_event_unlink(lp, event);
2868                 break;
2869         default:
2870                 /* Invalid events. */
2871                 LBUG();
2872         }
2873         lnet_net_lock(LNET_LOCK_EX);
2874         if (event->unlinked) {
2875                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2876                 lnet_ping_buffer_decref(pbuf);
2877                 lnet_peer_decref_locked(lp);
2878         }
2879
2880         /* put peer back at end of request queue, if discovery not already
2881          * done */
2882         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2883             lnet_peer_queue_for_discovery(lp)) {
2884                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2885                 wake_up(&the_lnet.ln_dc_waitq);
2886         }
2887         lnet_net_unlock(LNET_LOCK_EX);
2888 }
2889
2890 /*
2891  * Build a peer from incoming data.
2892  *
2893  * The NIDs in the incoming data are supposed to be structured as follows:
2894  *  - loopback
2895  *  - primary NID
2896  *  - other NIDs in same net
2897  *  - NIDs in second net
2898  *  - NIDs in third net
2899  *  - ...
2900  * This due to the way the list of NIDs in the data is created.
2901  *
2902  * Note that this function will mark the peer uptodate unless an
2903  * ENOMEM is encontered. All other errors are due to a conflict
2904  * between the DLC configuration and what discovery sees. We treat DLC
2905  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2906  * peer from becoming stuck in discovery.
2907  */
2908 static int lnet_peer_merge_data(struct lnet_peer *lp,
2909                                 struct lnet_ping_buffer *pbuf)
2910 {
2911         struct lnet_peer_net *lpn;
2912         struct lnet_peer_ni *lpni;
2913         lnet_nid_t *curnis = NULL;
2914         struct lnet_ni_status *addnis = NULL;
2915         lnet_nid_t *delnis = NULL;
2916         struct lnet_nid nid;
2917         unsigned int flags;
2918         int ncurnis;
2919         int naddnis;
2920         int ndelnis;
2921         int nnis = 0;
2922         int i;
2923         int j;
2924         int rc;
2925
2926         flags = LNET_PEER_DISCOVERED;
2927         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2928                 flags |= LNET_PEER_MULTI_RAIL;
2929
2930         /*
2931          * Cache the routing feature for the peer; whether it is enabled
2932          * for disabled as reported by the remote peer.
2933          */
2934         spin_lock(&lp->lp_lock);
2935         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2936                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2937         else
2938                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2939         spin_unlock(&lp->lp_lock);
2940
2941         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2942         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2943         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2944         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2945         if (!curnis || !addnis || !delnis) {
2946                 rc = -ENOMEM;
2947                 goto out;
2948         }
2949         ncurnis = 0;
2950         naddnis = 0;
2951         ndelnis = 0;
2952
2953         /* Construct the list of NIDs present in peer. */
2954         lpni = NULL;
2955         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2956                 curnis[ncurnis++] = lnet_nid_to_nid4(&lpni->lpni_nid);
2957
2958         /*
2959          * Check for NIDs in pbuf not present in curnis[].
2960          * The loop starts at 1 to skip the loopback NID.
2961          */
2962         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2963                 for (j = 0; j < ncurnis; j++)
2964                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2965                                 break;
2966                 if (j == ncurnis)
2967                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2968         }
2969         /*
2970          * Check for NIDs in curnis[] not present in pbuf.
2971          * The nested loop starts at 1 to skip the loopback NID.
2972          *
2973          * But never add the loopback NID to delnis[]: if it is
2974          * present in curnis[] then this peer is for this node.
2975          */
2976         for (i = 0; i < ncurnis; i++) {
2977                 if (curnis[i] == LNET_NID_LO_0)
2978                         continue;
2979                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2980                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2981                                 /*
2982                                  * update the information we cache for the
2983                                  * peer with the latest information we
2984                                  * received
2985                                  */
2986                                 lnet_nid4_to_nid(curnis[i], &nid);
2987                                 lpni = lnet_peer_ni_find_locked(&nid);
2988                                 if (lpni) {
2989                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2990                                         lnet_peer_ni_decref_locked(lpni);
2991                                 }
2992                                 break;
2993                         }
2994                 }
2995                 if (j == pbuf->pb_info.pi_nnis)
2996                         delnis[ndelnis++] = curnis[i];
2997         }
2998
2999         /*
3000          * If we get here and the discovery is disabled then we don't want
3001          * to add or delete any NIs. We just updated the ones we have some
3002          * information on, and call it a day
3003          */
3004         rc = 0;
3005         if (lnet_is_discovery_disabled(lp))
3006                 goto out;
3007
3008         for (i = 0; i < naddnis; i++) {
3009                 lnet_nid4_to_nid(addnis[i].ns_nid, &nid);
3010                 rc = lnet_peer_add_nid(lp, &nid, flags);
3011                 if (rc) {
3012                         CERROR("Error adding NID %s to peer %s: %d\n",
3013                                libcfs_nid2str(addnis[i].ns_nid),
3014                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3015                         if (rc == -ENOMEM)
3016                                 goto out;
3017                 }
3018                 lpni = lnet_peer_ni_find_locked(&nid);
3019                 if (lpni) {
3020                         lpni->lpni_ns_status = addnis[i].ns_status;
3021                         lnet_peer_ni_decref_locked(lpni);
3022                 }
3023         }
3024
3025         for (i = 0; i < ndelnis; i++) {
3026                 /*
3027                  * for routers it's okay to delete the primary_nid because
3028                  * the upper layers don't really rely on it. So if we're
3029                  * being told that the router changed its primary_nid
3030                  * then it's okay to delete it.
3031                  */
3032                 if (lp->lp_rtr_refcount > 0)
3033                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3034                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
3035                 if (rc) {
3036                         CERROR("Error deleting NID %s from peer %s: %d\n",
3037                                libcfs_nid2str(delnis[i]),
3038                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3039                         if (rc == -ENOMEM)
3040                                 goto out;
3041                 }
3042         }
3043
3044         /* The peer net for the primary NID should be the first entry in the
3045          * peer's lp_peer_nets list, and the peer NI for the primary NID should
3046          * be the first entry in its peer net's lpn_peer_nis list.
3047          */
3048         lnet_nid4_to_nid(pbuf->pb_info.pi_ni[1].ns_nid, &nid);
3049         lpni = lnet_peer_ni_find_locked(&nid);
3050         if (!lpni) {
3051                 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3052                        libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
3053                 goto out;
3054         }
3055
3056         lnet_peer_ni_decref_locked(lpni);
3057
3058         lpn = lpni->lpni_peer_net;
3059         if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3060                 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3061
3062         if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3063                 list_move(&lpni->lpni_peer_nis,
3064                           &lpni->lpni_peer_net->lpn_peer_nis);
3065
3066         /*
3067          * Errors other than -ENOMEM are due to peers having been
3068          * configured with DLC. Ignore these because DLC overrides
3069          * Discovery.
3070          */
3071         rc = 0;
3072 out:
3073         CFS_FREE_PTR_ARRAY(curnis, nnis);
3074         CFS_FREE_PTR_ARRAY(addnis, nnis);
3075         CFS_FREE_PTR_ARRAY(delnis, nnis);
3076         lnet_ping_buffer_decref(pbuf);
3077         CDEBUG(D_NET, "peer %s (%p): %d\n",
3078                libcfs_nidstr(&lp->lp_primary_nid), lp, rc);
3079
3080         if (rc) {
3081                 spin_lock(&lp->lp_lock);
3082                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
3083                 lp->lp_state |= LNET_PEER_FORCE_PING;
3084                 spin_unlock(&lp->lp_lock);
3085         }
3086         return rc;
3087 }
3088
3089 /*
3090  * The data in pbuf says lp is its primary peer, but the data was
3091  * received by a different peer. Try to update lp with the data.
3092  */
3093 static int
3094 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
3095 {
3096         struct lnet_handle_md mdh;
3097
3098         /* Queue lp for discovery, and force it on the request queue. */
3099         lnet_net_lock(LNET_LOCK_EX);
3100         if (lnet_peer_queue_for_discovery(lp))
3101                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3102         lnet_net_unlock(LNET_LOCK_EX);
3103
3104         LNetInvalidateMDHandle(&mdh);
3105
3106         /*
3107          * Decide whether we can move the peer to the DATA_PRESENT state.
3108          *
3109          * We replace stale data for a multi-rail peer, repair PING_FAILED
3110          * status, and preempt FORCE_PING.
3111          *
3112          * If after that we have DATA_PRESENT, we merge it into this peer.
3113          */
3114         spin_lock(&lp->lp_lock);
3115         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3116                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3117                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3118                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3119                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3120                         lnet_ping_buffer_decref(pbuf);
3121                         pbuf = lp->lp_data;
3122                         lp->lp_data = NULL;
3123                 }
3124         }
3125         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3126                 lnet_ping_buffer_decref(lp->lp_data);
3127                 lp->lp_data = NULL;
3128                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3129         }
3130         if (lp->lp_state & LNET_PEER_PING_FAILED) {
3131                 mdh = lp->lp_ping_mdh;
3132                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3133                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3134                 lp->lp_ping_error = 0;
3135         }
3136         if (lp->lp_state & LNET_PEER_FORCE_PING)
3137                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3138         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3139         spin_unlock(&lp->lp_lock);
3140
3141         if (!LNetMDHandleIsInvalid(mdh))
3142                 LNetMDUnlink(mdh);
3143
3144         if (pbuf)
3145                 return lnet_peer_merge_data(lp, pbuf);
3146
3147         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3148         return 0;
3149 }
3150
3151 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
3152 {
3153         int i;
3154
3155         for (i = 0; i < pinfo->pi_nnis; i++) {
3156                 if (pinfo->pi_ni[i].ns_nid == nid)
3157                         return true;
3158         }
3159
3160         return false;
3161 }
3162
3163 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3164  * to the discovery queue a reference was taken that will prevent the peer from
3165  * actually being freed by this function. After this function exits the
3166  * discovery thread should call lnet_peer_discovery_complete() which will
3167  * drop that reference as well as wake any waiters that may also be holding a
3168  * ref on the peer
3169  */
3170 static int lnet_peer_deletion(struct lnet_peer *lp)
3171 __must_hold(&lp->lp_lock)
3172 {
3173         struct list_head rlist;
3174         struct lnet_route *route, *tmp;
3175         int sensitivity = lp->lp_health_sensitivity;
3176         int rc;
3177
3178         INIT_LIST_HEAD(&rlist);
3179
3180         lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3181                           LNET_PEER_FORCE_PUSH);
3182         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3183                libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3184
3185         /* no-op if lnet_peer_del() has already been called on this peer */
3186         if (lp->lp_state & LNET_PEER_MARK_DELETED)
3187                 return 0;
3188
3189         spin_unlock(&lp->lp_lock);
3190
3191         mutex_lock(&the_lnet.ln_api_mutex);
3192         if (the_lnet.ln_state != LNET_STATE_RUNNING ||
3193             the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
3194                 mutex_unlock(&the_lnet.ln_api_mutex);
3195                 spin_lock(&lp->lp_lock);
3196                 return -ESHUTDOWN;
3197         }
3198
3199         lnet_net_lock(LNET_LOCK_EX);
3200         /* remove the peer from the discovery work
3201          * queue if it's on there in preparation
3202          * of deleting it.
3203          */
3204         if (!list_empty(&lp->lp_dc_list))
3205                 list_del_init(&lp->lp_dc_list);
3206         list_for_each_entry_safe(route, tmp,
3207                                  &lp->lp_routes,
3208                                  lr_gwlist)
3209                 lnet_move_route(route, NULL, &rlist);
3210         lnet_net_unlock(LNET_LOCK_EX);
3211
3212         /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3213         rc = lnet_peer_del(lp);
3214         if (rc)
3215                 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3216                         libcfs_nidstr(&lp->lp_primary_nid), rc);
3217
3218         list_for_each_entry_safe(route, tmp,
3219                                  &rlist, lr_list) {
3220                 /* re-add these routes */
3221                 lnet_add_route(route->lr_net,
3222                                route->lr_hops,
3223                                &route->lr_nid,
3224                                route->lr_priority,
3225                                sensitivity);
3226                 LIBCFS_FREE(route, sizeof(*route));
3227         }
3228
3229         mutex_unlock(&the_lnet.ln_api_mutex);
3230
3231         spin_lock(&lp->lp_lock);
3232
3233         return 0;
3234 }
3235
3236 /*
3237  * Update a peer using the data received.
3238  */
3239 static int lnet_peer_data_present(struct lnet_peer *lp)
3240 __must_hold(&lp->lp_lock)
3241 {
3242         struct lnet_ping_buffer *pbuf;
3243         struct lnet_peer_ni *lpni;
3244         struct lnet_nid nid;
3245         unsigned int flags;
3246         int rc = 0;
3247
3248         pbuf = lp->lp_data;
3249         lp->lp_data = NULL;
3250         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3251         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3252         spin_unlock(&lp->lp_lock);
3253
3254         /*
3255          * Modifications of peer structures are done while holding the
3256          * ln_api_mutex. A global lock is required because we may be
3257          * modifying multiple peer structures, and a mutex greatly
3258          * simplifies memory management.
3259          *
3260          * The actual changes to the data structures must also protect
3261          * against concurrent lookups, for which the lnet_net_lock in
3262          * LNET_LOCK_EX mode is used.
3263          */
3264         mutex_lock(&the_lnet.ln_api_mutex);
3265         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3266                 rc = -ESHUTDOWN;
3267                 goto out;
3268         }
3269
3270         /*
3271          * If this peer is not on the peer list then it is being torn
3272          * down, and our reference count may be all that is keeping it
3273          * alive. Don't do any work on it.
3274          */
3275         if (list_empty(&lp->lp_peer_list)) {
3276                 lnet_ping_buffer_decref(pbuf);
3277                 goto out;
3278         }
3279
3280         flags = LNET_PEER_DISCOVERED;
3281         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3282                 flags |= LNET_PEER_MULTI_RAIL;
3283
3284         /*
3285          * Check whether the primary NID in the message matches the
3286          * primary NID of the peer. If it does, update the peer, if
3287          * it it does not, check whether there is already a peer with
3288          * that primary NID. If no such peer exists, try to update
3289          * the primary NID of the current peer (allowed if it was
3290          * created due to message traffic) and complete the update.
3291          * If the peer did exist, hand off the data to it.
3292          *
3293          * The peer for the loopback interface is a special case: this
3294          * is the peer for the local node, and we want to set its
3295          * primary NID to the correct value here. Moreover, this peer
3296          * can show up with only the loopback NID in the ping buffer.
3297          */
3298         if (pbuf->pb_info.pi_nnis <= 1) {
3299                 lnet_ping_buffer_decref(pbuf);
3300                 goto out;
3301         }
3302         lnet_nid4_to_nid(pbuf->pb_info.pi_ni[1].ns_nid, &nid);
3303         if (nid_is_lo0(&lp->lp_primary_nid)) {
3304                 rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3305                 if (rc)
3306                         lnet_ping_buffer_decref(pbuf);
3307                 else
3308                         rc = lnet_peer_merge_data(lp, pbuf);
3309         /*
3310          * if the primary nid of the peer is present in the ping info returned
3311          * from the peer, but it's not the local primary peer we have
3312          * cached and discovery is disabled, then we don't want to update
3313          * our local peer info, by adding or removing NIDs, we just want
3314          * to update the status of the nids that we currently have
3315          * recorded in that peer.
3316          */
3317         } else if (nid_same(&lp->lp_primary_nid, &nid) ||
3318                    (lnet_is_nid_in_ping_info(lnet_nid_to_nid4(&lp->lp_primary_nid),
3319                                              &pbuf->pb_info) &&
3320                     lnet_is_discovery_disabled(lp))) {
3321                 rc = lnet_peer_merge_data(lp, pbuf);
3322         } else {
3323                 lpni = lnet_peer_ni_find_locked(&nid);
3324                 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3325                         rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3326                         if (rc) {
3327                                 CERROR("Primary NID error %s versus %s: %d\n",
3328                                        libcfs_nidstr(&lp->lp_primary_nid),
3329                                        libcfs_nidstr(&nid), rc);
3330                                 lnet_ping_buffer_decref(pbuf);
3331                         } else {
3332                                 rc = lnet_peer_merge_data(lp, pbuf);
3333                         }
3334                         if (lpni)
3335                                 lnet_peer_ni_decref_locked(lpni);
3336                 } else {
3337                         struct lnet_peer *new_lp;
3338                         new_lp = lpni->lpni_peer_net->lpn_peer;
3339                         /*
3340                          * if lp has discovery/MR enabled that means new_lp
3341                          * should have discovery/MR enabled as well, since
3342                          * it's the same peer, which we're about to merge
3343                          */
3344                         spin_lock(&lp->lp_lock);
3345                         spin_lock(&new_lp->lp_lock);
3346                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3347                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3348                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3349                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3350                         /* If we're processing a ping reply then we may be
3351                          * about to send a push to the peer that we ping'd.
3352                          * Since the ping reply that we're processing was
3353                          * received by lp, we need to set the discovery source
3354                          * NID for new_lp to the NID stored in lp.
3355                          */
3356                         if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3357                                 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3358                                 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3359                         }
3360                         spin_unlock(&new_lp->lp_lock);
3361                         spin_unlock(&lp->lp_lock);
3362
3363                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
3364                         lnet_consolidate_routes_locked(lp, new_lp);
3365                         lnet_peer_ni_decref_locked(lpni);
3366                 }
3367         }
3368 out:
3369         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3370                libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3371                lp->lp_state);
3372         mutex_unlock(&the_lnet.ln_api_mutex);
3373
3374         spin_lock(&lp->lp_lock);
3375         /* Tell discovery to re-check the peer immediately. */
3376         if (!rc)
3377                 rc = LNET_REDISCOVER_PEER;
3378         return rc;
3379 }
3380
3381 /*
3382  * A ping failed. Clear the PING_FAILED state and set the
3383  * FORCE_PING state, to ensure a retry even if discovery is
3384  * disabled. This avoids being left with incorrect state.
3385  */
3386 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3387 __must_hold(&lp->lp_lock)
3388 {
3389         struct lnet_handle_md mdh;
3390         int rc;
3391
3392         mdh = lp->lp_ping_mdh;
3393         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3394         lp->lp_state &= ~LNET_PEER_PING_FAILED;
3395         lp->lp_state |= LNET_PEER_FORCE_PING;
3396         rc = lp->lp_ping_error;
3397         lp->lp_ping_error = 0;
3398         spin_unlock(&lp->lp_lock);
3399
3400         if (!LNetMDHandleIsInvalid(mdh))
3401                 LNetMDUnlink(mdh);
3402
3403         CDEBUG(D_NET, "peer %s:%d\n",
3404                libcfs_nidstr(&lp->lp_primary_nid), rc);
3405
3406         spin_lock(&lp->lp_lock);
3407         return rc ? rc : LNET_REDISCOVER_PEER;
3408 }
3409
3410 /* Active side of ping. */
3411 static int lnet_peer_send_ping(struct lnet_peer *lp)
3412 __must_hold(&lp->lp_lock)
3413 {
3414         int nnis;
3415         int rc;
3416         int cpt;
3417
3418         lp->lp_state |= LNET_PEER_PING_SENT;
3419         lp->lp_state &= ~LNET_PEER_FORCE_PING;
3420         spin_unlock(&lp->lp_lock);
3421
3422         cpt = lnet_net_lock_current();
3423         /* Refcount for MD. */
3424         lnet_peer_addref_locked(lp);
3425         lnet_net_unlock(cpt);
3426
3427         nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3428
3429         rc = lnet_send_ping(&lp->lp_primary_nid, &lp->lp_ping_mdh, nnis, lp,
3430                             the_lnet.ln_dc_handler, false);
3431
3432         /*
3433          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3434          * refcount on the peer, otherwise LNetMDUnlink will be called
3435          * which will eventually do that.
3436          */
3437         if (rc > 0) {
3438                 lnet_net_lock(cpt);
3439                 lnet_peer_decref_locked(lp);
3440                 lnet_net_unlock(cpt);
3441                 rc = -rc; /* change the rc to negative value */
3442                 goto fail_error;
3443         } else if (rc < 0) {
3444                 goto fail_error;
3445         }
3446
3447         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3448
3449         spin_lock(&lp->lp_lock);
3450         return 0;
3451
3452 fail_error:
3453         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3454         /*
3455          * The errors that get us here are considered hard errors and
3456          * cause Discovery to terminate. So we clear PING_SENT, but do
3457          * not set either PING_FAILED or FORCE_PING. In fact we need
3458          * to clear PING_FAILED, because the unlink event handler will
3459          * have set it if we called LNetMDUnlink() above.
3460          */
3461         spin_lock(&lp->lp_lock);
3462         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3463         return rc;
3464 }
3465
3466 /*
3467  * This function exists because you cannot call LNetMDUnlink() from an
3468  * event handler.
3469  */
3470 static int lnet_peer_push_failed(struct lnet_peer *lp)
3471 __must_hold(&lp->lp_lock)
3472 {
3473         struct lnet_handle_md mdh;
3474         int rc;
3475
3476         mdh = lp->lp_push_mdh;
3477         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3478         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3479         rc = lp->lp_push_error;
3480         lp->lp_push_error = 0;
3481         spin_unlock(&lp->lp_lock);
3482
3483         if (!LNetMDHandleIsInvalid(mdh))
3484                 LNetMDUnlink(mdh);
3485
3486         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3487         spin_lock(&lp->lp_lock);
3488         return rc ? rc : LNET_REDISCOVER_PEER;
3489 }
3490
3491 /*
3492  * Mark the peer as discovered.
3493  */
3494 static int lnet_peer_discovered(struct lnet_peer *lp)
3495 __must_hold(&lp->lp_lock)
3496 {
3497         lp->lp_state |= LNET_PEER_DISCOVERED;
3498         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3499                           LNET_PEER_REDISCOVER);
3500
3501         lp->lp_dc_error = 0;
3502
3503         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3504
3505         return 0;
3506 }
3507
3508 /* Active side of push. */
3509 static int lnet_peer_send_push(struct lnet_peer *lp)
3510 __must_hold(&lp->lp_lock)
3511 {
3512         struct lnet_ping_buffer *pbuf;
3513         struct lnet_processid id;
3514         struct lnet_md md;
3515         int cpt;
3516         int rc;
3517
3518         /* Don't push to a non-multi-rail peer. */
3519         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3520                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3521                 /* if peer's NIDs are uptodate then peer is discovered */
3522                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3523                         rc = lnet_peer_discovered(lp);
3524                         return rc;
3525                 }
3526
3527                 return 0;
3528         }
3529
3530         lp->lp_state |= LNET_PEER_PUSH_SENT;
3531         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3532         spin_unlock(&lp->lp_lock);
3533
3534         cpt = lnet_net_lock_current();
3535         pbuf = the_lnet.ln_ping_target;
3536         lnet_ping_buffer_addref(pbuf);
3537         lnet_net_unlock(cpt);
3538
3539         /* Push source MD */
3540         md.start     = &pbuf->pb_info;
3541         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3542         md.threshold = 2; /* Put/Ack */
3543         md.max_size  = 0;
3544         md.options   = LNET_MD_TRACK_RESPONSE;
3545         md.handler   = the_lnet.ln_dc_handler;
3546         md.user_ptr  = lp;
3547
3548         rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3549         if (rc) {
3550                 lnet_ping_buffer_decref(pbuf);
3551                 CERROR("Can't bind push source MD: %d\n", rc);
3552                 goto fail_error;
3553         }
3554
3555         cpt = lnet_net_lock_current();
3556         /* Refcount for MD. */
3557         lnet_peer_addref_locked(lp);
3558         id.pid = LNET_PID_LUSTRE;
3559         if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3560                 id.nid = lp->lp_disc_dst_nid;
3561         else
3562                 id.nid = lp->lp_primary_nid;
3563         lnet_net_unlock(cpt);
3564
3565         rc = LNetPut(&lp->lp_disc_src_nid, lp->lp_push_mdh,
3566                      LNET_ACK_REQ, &id, LNET_RESERVED_PORTAL,
3567                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3568
3569         /*
3570          * reset the discovery nid. There is no need to restrict sending
3571          * from that source, if we call lnet_push_update_to_peers(). It'll
3572          * get set to a specific NID, if we initiate discovery from the
3573          * scratch
3574          */
3575         lp->lp_disc_src_nid = LNET_ANY_NID;
3576         lp->lp_disc_dst_nid = LNET_ANY_NID;
3577
3578         if (rc)
3579                 goto fail_unlink;
3580
3581         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3582
3583         spin_lock(&lp->lp_lock);
3584         return 0;
3585
3586 fail_unlink:
3587         LNetMDUnlink(lp->lp_push_mdh);
3588         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3589 fail_error:
3590         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3591                lp, rc);
3592         /*
3593          * The errors that get us here are considered hard errors and
3594          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3595          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3596          * because the unlink event handler will have set it if we
3597          * called LNetMDUnlink() above.
3598          */
3599         spin_lock(&lp->lp_lock);
3600         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3601         return rc;
3602 }
3603
3604 /*
3605  * Wait for work to be queued or some other change that must be
3606  * attended to. Returns non-zero if the discovery thread should shut
3607  * down.
3608  */
3609 static int lnet_peer_discovery_wait_for_work(void)
3610 {
3611         int cpt;
3612         int rc = 0;
3613
3614         DEFINE_WAIT(wait);
3615
3616         cpt = lnet_net_lock_current();
3617         for (;;) {
3618                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3619                                 TASK_INTERRUPTIBLE);
3620                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3621                         break;
3622                 if (lnet_push_target_resize_needed() ||
3623                     the_lnet.ln_push_target->pb_needs_post)
3624                         break;
3625                 if (!list_empty(&the_lnet.ln_dc_request))
3626                         break;
3627                 if (!list_empty(&the_lnet.ln_msg_resend))
3628                         break;
3629                 lnet_net_unlock(cpt);
3630
3631                 /*
3632                  * wakeup max every second to check if there are peers that
3633                  * have been stuck on the working queue for greater than
3634                  * the peer timeout.
3635                  */
3636                 schedule_timeout(cfs_time_seconds(1));
3637                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3638                 cpt = lnet_net_lock_current();
3639         }
3640         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3641
3642         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3643                 rc = -ESHUTDOWN;
3644
3645         lnet_net_unlock(cpt);
3646
3647         CDEBUG(D_NET, "woken: %d\n", rc);
3648
3649         return rc;
3650 }
3651
3652 /*
3653  * Messages that were pending on a destroyed peer will be put on a global
3654  * resend list. The message resend list will be checked by
3655  * the discovery thread when it wakes up, and will resend messages. These
3656  * messages can still be sendable in the case the lpni which was the initial
3657  * cause of the message re-queue was transfered to another peer.
3658  *
3659  * It is possible that LNet could be shutdown while we're iterating
3660  * through the list. lnet_shudown_lndnets() will attempt to access the
3661  * resend list, but will have to wait until the spinlock is released, by
3662  * which time there shouldn't be any more messages on the resend list.
3663  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3664  * for the messages so they can be released. The other case is that
3665  * lnet_shudown_lndnets() can finalize all the messages before this
3666  * function can visit the resend list, in which case this function will be
3667  * a no-op.
3668  */
3669 static void lnet_resend_msgs(void)
3670 {
3671         struct lnet_msg *msg, *tmp;
3672         LIST_HEAD(resend);
3673         int rc;
3674
3675         spin_lock(&the_lnet.ln_msg_resend_lock);
3676         list_splice(&the_lnet.ln_msg_resend, &resend);
3677         spin_unlock(&the_lnet.ln_msg_resend_lock);
3678
3679         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3680                 list_del_init(&msg->msg_list);
3681                 rc = lnet_send(&msg->msg_src_nid_param, msg,
3682                                &msg->msg_rtr_nid_param);
3683                 if (rc < 0) {
3684                         CNETERR("Error sending %s to %s: %d\n",
3685                                lnet_msgtyp2str(msg->msg_type),
3686                                libcfs_idstr(&msg->msg_target), rc);
3687                         lnet_finalize(msg, rc);
3688                 }
3689         }
3690 }
3691
3692 /* The discovery thread. */
3693 static int lnet_peer_discovery(void *arg)
3694 {
3695         struct lnet_peer *lp;
3696         int rc;
3697
3698         wait_for_completion(&the_lnet.ln_started);
3699
3700         CDEBUG(D_NET, "started\n");
3701
3702         for (;;) {
3703                 if (lnet_peer_discovery_wait_for_work())
3704                         break;
3705
3706                 if (lnet_push_target_resize_needed())
3707                         lnet_push_target_resize();
3708                 else if (the_lnet.ln_push_target->pb_needs_post)
3709                         lnet_push_target_post(the_lnet.ln_push_target,
3710                                               &the_lnet.ln_push_target_md);
3711
3712                 lnet_resend_msgs();
3713
3714                 lnet_net_lock(LNET_LOCK_EX);
3715                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3716                         lnet_net_unlock(LNET_LOCK_EX);
3717                         break;
3718                 }
3719
3720                 /*
3721                  * Process all incoming discovery work requests.  When
3722                  * discovery must wait on a peer to change state, it
3723                  * is added to the tail of the ln_dc_working queue. A
3724                  * timestamp keeps track of when the peer was added,
3725                  * so we can time out discovery requests that take too
3726                  * long.
3727                  */
3728                 while (!list_empty(&the_lnet.ln_dc_request)) {
3729                         lp = list_first_entry(&the_lnet.ln_dc_request,
3730                                               struct lnet_peer, lp_dc_list);
3731                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3732                         /*
3733                          * set the time the peer was put on the dc_working
3734                          * queue. It shouldn't remain on the queue
3735                          * forever, in case the GET message (for ping)
3736                          * doesn't get a REPLY or the PUT message (for
3737                          * push) doesn't get an ACK.
3738                          */
3739                         lp->lp_last_queued = ktime_get_real_seconds();
3740                         lnet_net_unlock(LNET_LOCK_EX);
3741
3742                         if (lnet_push_target_resize_needed())
3743                                 lnet_push_target_resize();
3744                         else if (the_lnet.ln_push_target->pb_needs_post)
3745                                 lnet_push_target_post(the_lnet.ln_push_target,
3746                                                       &the_lnet.ln_push_target_md);
3747
3748                         /*
3749                          * Select an action depending on the state of
3750                          * the peer and whether discovery is disabled.
3751                          * The check whether discovery is disabled is
3752                          * done after the code that handles processing
3753                          * for arrived data, cleanup for failures, and
3754                          * forcing a Ping or Push.
3755                          */
3756                         spin_lock(&lp->lp_lock);
3757                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3758                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3759                                 lp->lp_state);
3760                         if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3761                                             LNET_PEER_MARK_DELETED))
3762                                 rc = lnet_peer_deletion(lp);
3763                         else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3764                                 rc = lnet_peer_data_present(lp);
3765                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3766                                 rc = lnet_peer_ping_failed(lp);
3767                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3768                                 rc = lnet_peer_push_failed(lp);
3769                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3770                                 rc = lnet_peer_send_ping(lp);
3771                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3772                                 rc = lnet_peer_send_push(lp);
3773                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3774                                 rc = lnet_peer_send_ping(lp);
3775                         else if (lnet_peer_needs_push(lp))
3776                                 rc = lnet_peer_send_push(lp);
3777                         else
3778                                 rc = lnet_peer_discovered(lp);
3779                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3780                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3781                                 lp->lp_state, rc);
3782
3783                         if (rc == LNET_REDISCOVER_PEER) {
3784                                 spin_unlock(&lp->lp_lock);
3785                                 lnet_net_lock(LNET_LOCK_EX);
3786                                 list_move(&lp->lp_dc_list,
3787                                           &the_lnet.ln_dc_request);
3788                         } else if (rc ||
3789                                    !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3790                                 spin_unlock(&lp->lp_lock);
3791                                 lnet_net_lock(LNET_LOCK_EX);
3792                                 lnet_peer_discovery_complete(lp, rc);
3793                         } else {
3794                                 spin_unlock(&lp->lp_lock);
3795                                 lnet_net_lock(LNET_LOCK_EX);
3796                         }
3797
3798                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3799                                 break;
3800
3801                 }
3802
3803                 lnet_net_unlock(LNET_LOCK_EX);
3804         }
3805
3806         CDEBUG(D_NET, "stopping\n");
3807         /*
3808          * Clean up before telling lnet_peer_discovery_stop() that
3809          * we're done. Use wake_up() below to somewhat reduce the
3810          * size of the thundering herd if there are multiple threads
3811          * waiting on discovery of a single peer.
3812          */
3813
3814         /* Queue cleanup 1: stop all pending pings and pushes. */
3815         lnet_net_lock(LNET_LOCK_EX);
3816         while (!list_empty(&the_lnet.ln_dc_working)) {
3817                 lp = list_first_entry(&the_lnet.ln_dc_working,
3818                                       struct lnet_peer, lp_dc_list);
3819                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3820                 lnet_net_unlock(LNET_LOCK_EX);
3821                 lnet_peer_cancel_discovery(lp);
3822                 lnet_net_lock(LNET_LOCK_EX);
3823         }
3824         lnet_net_unlock(LNET_LOCK_EX);
3825
3826         /* Queue cleanup 2: wait for the expired queue to clear. */
3827         while (!list_empty(&the_lnet.ln_dc_expired))
3828                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3829
3830         /* Queue cleanup 3: clear the request queue. */
3831         lnet_net_lock(LNET_LOCK_EX);
3832         while (!list_empty(&the_lnet.ln_dc_request)) {
3833                 lp = list_first_entry(&the_lnet.ln_dc_request,
3834                                       struct lnet_peer, lp_dc_list);
3835                 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3836         }
3837         lnet_net_unlock(LNET_LOCK_EX);
3838
3839         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3840         the_lnet.ln_dc_handler = NULL;
3841
3842         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3843         wake_up(&the_lnet.ln_dc_waitq);
3844
3845         CDEBUG(D_NET, "stopped\n");
3846
3847         return 0;
3848 }
3849
3850 /* ln_api_mutex is held on entry. */
3851 int lnet_peer_discovery_start(void)
3852 {
3853         struct task_struct *task;
3854         int rc = 0;
3855
3856         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3857                 return -EALREADY;
3858
3859         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3860         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3861         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3862         if (IS_ERR(task)) {
3863                 rc = PTR_ERR(task);
3864                 CERROR("Can't start peer discovery thread: %d\n", rc);
3865
3866                 the_lnet.ln_dc_handler = NULL;
3867
3868                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3869         }
3870
3871         CDEBUG(D_NET, "discovery start: %d\n", rc);
3872
3873         return rc;
3874 }
3875
3876 /* ln_api_mutex is held on entry. */
3877 void lnet_peer_discovery_stop(void)
3878 {
3879         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3880                 return;
3881
3882         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3883         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3884
3885         /* In the LNetNIInit() path we may be stopping discovery before it
3886          * entered its work loop
3887          */
3888         if (!completion_done(&the_lnet.ln_started))
3889                 complete(&the_lnet.ln_started);
3890         else
3891                 wake_up(&the_lnet.ln_dc_waitq);
3892
3893         mutex_unlock(&the_lnet.ln_api_mutex);
3894         wait_event(the_lnet.ln_dc_waitq,
3895                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3896         mutex_lock(&the_lnet.ln_api_mutex);
3897
3898         LASSERT(list_empty(&the_lnet.ln_dc_request));
3899         LASSERT(list_empty(&the_lnet.ln_dc_working));
3900         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3901
3902         CDEBUG(D_NET, "discovery stopped\n");
3903 }
3904
3905 /* Debugging */
3906
3907 void
3908 lnet_debug_peer(lnet_nid_t nid4)
3909 {
3910         char                    *aliveness = "NA";
3911         struct lnet_peer_ni     *lp;
3912         int                     cpt;
3913         struct lnet_nid nid;
3914
3915         lnet_nid4_to_nid(nid4, &nid);
3916         cpt = lnet_nid2cpt(&nid, NULL);
3917         lnet_net_lock(cpt);
3918
3919         lp = lnet_peerni_by_nid_locked(&nid, NULL, cpt);
3920         if (IS_ERR(lp)) {
3921                 lnet_net_unlock(cpt);
3922                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nidstr(&nid));
3923                 return;
3924         }
3925
3926         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3927                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3928
3929         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3930        &n