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, 2014, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/peer.c
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_LNET
  36
  37 #include <lnet/lib-lnet.h>
  38 #include <uapi/linux/lnet/lnet-dlc.h>
  39
  40 /* Value indicating that recovery needs to re-check a peer immediately. */
  41 #define LNET_REDISCOVER_PEER    (1)
  42
  43 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
  44
  45 static void
  46 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
  47 {
  48         if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
  49                 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
  50                 lnet_peer_ni_decref_locked(lpni);
  51         }
  52 }
  53
  54 void
  55 lnet_peer_net_added(struct lnet_net *net)
  56 {
  57         struct lnet_peer_ni *lpni, *tmp;
  58
  59         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
  60                                  lpni_on_remote_peer_ni_list) {
  61
  62                 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
  63                         lpni->lpni_net = net;
  64
  65                         spin_lock(&lpni->lpni_lock);
  66                         lpni->lpni_txcredits =
  67                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
  68                         lpni->lpni_mintxcredits = lpni->lpni_txcredits;
  69                         lpni->lpni_rtrcredits =
  70                                 lnet_peer_buffer_credits(lpni->lpni_net);
  71                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
  72                         spin_unlock(&lpni->lpni_lock);
  73
  74                         lnet_peer_remove_from_remote_list(lpni);
  75                 }
  76         }
  77 }
  78
  79 static void
  80 lnet_peer_tables_destroy(void)
  81 {
  82         struct lnet_peer_table  *ptable;
  83         struct list_head        *hash;
  84         int                     i;
  85         int                     j;
  86
  87         if (!the_lnet.ln_peer_tables)
  88                 return;
  89
  90         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  91                 hash = ptable->pt_hash;
  92                 if (!hash) /* not intialized */
  93                         break;
  94
  95                 LASSERT(list_empty(&ptable->pt_zombie_list));
  96
  97                 ptable->pt_hash = NULL;
  98                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
  99                         LASSERT(list_empty(&hash[j]));
 100
 101                 LIBCFS_FREE(hash, LNET_PEER_HASH_SIZE * sizeof(*hash));
 102         }
 103
 104         cfs_percpt_free(the_lnet.ln_peer_tables);
 105         the_lnet.ln_peer_tables = NULL;
 106 }
 107
 108 int
 109 lnet_peer_tables_create(void)
 110 {
 111         struct lnet_peer_table  *ptable;
 112         struct list_head        *hash;
 113         int                     i;
 114         int                     j;
 115
 116         the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
 117                                                    sizeof(*ptable));
 118         if (the_lnet.ln_peer_tables == NULL) {
 119                 CERROR("Failed to allocate cpu-partition peer tables\n");
 120                 return -ENOMEM;
 121         }
 122
 123         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 124                 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
 125                                  LNET_PEER_HASH_SIZE * sizeof(*hash));
 126                 if (hash == NULL) {
 127                         CERROR("Failed to create peer hash table\n");
 128                         lnet_peer_tables_destroy();
 129                         return -ENOMEM;
 130                 }
 131
 132                 spin_lock_init(&ptable->pt_zombie_lock);
 133                 INIT_LIST_HEAD(&ptable->pt_zombie_list);
 134
 135                 INIT_LIST_HEAD(&ptable->pt_peer_list);
 136
 137                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 138                         INIT_LIST_HEAD(&hash[j]);
 139                 ptable->pt_hash = hash; /* sign of initialization */
 140         }
 141
 142         return 0;
 143 }
 144
 145 static struct lnet_peer_ni *
 146 lnet_peer_ni_alloc(lnet_nid_t nid)
 147 {
 148         struct lnet_peer_ni *lpni;
 149         struct lnet_net *net;
 150         int cpt;
 151
 152         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 153
 154         LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
 155         if (!lpni)
 156                 return NULL;
 157
 158         INIT_LIST_HEAD(&lpni->lpni_txq);
 159         INIT_LIST_HEAD(&lpni->lpni_rtrq);
 160         INIT_LIST_HEAD(&lpni->lpni_routes);
 161         INIT_LIST_HEAD(&lpni->lpni_hashlist);
 162         INIT_LIST_HEAD(&lpni->lpni_peer_nis);
 163         INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
 164
 165         spin_lock_init(&lpni->lpni_lock);
 166
 167         lpni->lpni_alive = !lnet_peers_start_down(); /* 1 bit!! */
 168         lpni->lpni_last_alive = cfs_time_current(); /* assumes alive */
 169         lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 170         lpni->lpni_nid = nid;
 171         lpni->lpni_cpt = cpt;
 172         lnet_set_peer_ni_health_locked(lpni, true);
 173
 174         net = lnet_get_net_locked(LNET_NIDNET(nid));
 175         lpni->lpni_net = net;
 176         if (net) {
 177                 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
 178                 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
 179                 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
 180                 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 181         } else {
 182                 /*
 183                  * This peer_ni is not on a local network, so we
 184                  * cannot add the credits here. In case the net is
 185                  * added later, add the peer_ni to the remote peer ni
 186                  * list so it can be easily found and revisited.
 187                  */
 188                 /* FIXME: per-net implementation instead? */
 189                 atomic_inc(&lpni->lpni_refcount);
 190                 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
 191                               &the_lnet.ln_remote_peer_ni_list);
 192         }
 193
 194         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
 195
 196         return lpni;
 197 }
 198
 199 static struct lnet_peer_net *
 200 lnet_peer_net_alloc(__u32 net_id)
 201 {
 202         struct lnet_peer_net *lpn;
 203
 204         LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
 205         if (!lpn)
 206                 return NULL;
 207
 208         INIT_LIST_HEAD(&lpn->lpn_peer_nets);
 209         INIT_LIST_HEAD(&lpn->lpn_peer_nis);
 210         lpn->lpn_net_id = net_id;
 211
 212         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 213
 214         return lpn;
 215 }
 216
 217 void
 218 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
 219 {
 220         struct lnet_peer *lp;
 221
 222         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 223
 224         LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
 225         LASSERT(list_empty(&lpn->lpn_peer_nis));
 226         LASSERT(list_empty(&lpn->lpn_peer_nets));
 227         lp = lpn->lpn_peer;
 228         lpn->lpn_peer = NULL;
 229         LIBCFS_FREE(lpn, sizeof(*lpn));
 230
 231         lnet_peer_decref_locked(lp);
 232 }
 233
 234 static struct lnet_peer *
 235 lnet_peer_alloc(lnet_nid_t nid)
 236 {
 237         struct lnet_peer *lp;
 238
 239         LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
 240         if (!lp)
 241                 return NULL;
 242
 243         INIT_LIST_HEAD(&lp->lp_peer_list);
 244         INIT_LIST_HEAD(&lp->lp_peer_nets);
 245         INIT_LIST_HEAD(&lp->lp_dc_list);
 246         INIT_LIST_HEAD(&lp->lp_dc_pendq);
 247         init_waitqueue_head(&lp->lp_dc_waitq);
 248         spin_lock_init(&lp->lp_lock);
 249         lp->lp_primary_nid = nid;
 250         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 251
 252         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 253
 254         return lp;
 255 }
 256
 257 void
 258 lnet_destroy_peer_locked(struct lnet_peer *lp)
 259 {
 260         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 261
 262         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 263         LASSERT(list_empty(&lp->lp_peer_nets));
 264         LASSERT(list_empty(&lp->lp_peer_list));
 265         LASSERT(list_empty(&lp->lp_dc_list));
 266
 267         if (lp->lp_data)
 268                 lnet_ping_buffer_decref(lp->lp_data);
 269
 270         LIBCFS_FREE(lp, sizeof(*lp));
 271 }
 272
 273 /*
 274  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 275  * that peer_net, detach the peer_net from the peer.
 276  *
 277  * Call with lnet_net_lock/EX held
 278  */
 279 static void
 280 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 281 {
 282         struct lnet_peer_table *ptable;
 283         struct lnet_peer_net *lpn;
 284         struct lnet_peer *lp;
 285
 286         /*
 287          * Belts and suspenders: gracefully handle teardown of a
 288          * partially connected peer_ni.
 289          */
 290         lpn = lpni->lpni_peer_net;
 291
 292         list_del_init(&lpni->lpni_peer_nis);
 293         /*
 294          * If there are no lpni's left, we detach lpn from
 295          * lp_peer_nets, so it cannot be found anymore.
 296          */
 297         if (list_empty(&lpn->lpn_peer_nis))
 298                 list_del_init(&lpn->lpn_peer_nets);
 299
 300         /* Update peer NID count. */
 301         lp = lpn->lpn_peer;
 302         lp->lp_nnis--;
 303
 304         /*
 305          * If there are no more peer nets, make the peer unfindable
 306          * via the peer_tables.
 307          *
 308          * Otherwise, if the peer is DISCOVERED, tell discovery to
 309          * take another look at it. This is a no-op if discovery for
 310          * this peer did the detaching.
 311          */
 312         if (list_empty(&lp->lp_peer_nets)) {
 313                 list_del_init(&lp->lp_peer_list);
 314                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 315                 ptable->pt_peers--;
 316         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 317                 /* Discovery isn't running, nothing to do here. */
 318         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 319                 lnet_peer_queue_for_discovery(lp);
 320                 wake_up(&the_lnet.ln_dc_waitq);
 321         }
 322         CDEBUG(D_NET, "peer %s NID %s\n",
 323                 libcfs_nid2str(lp->lp_primary_nid),
 324                 libcfs_nid2str(lpni->lpni_nid));
 325 }
 326
 327 /* called with lnet_net_lock LNET_LOCK_EX held */
 328 static int
 329 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni)
 330 {
 331         struct lnet_peer_table *ptable = NULL;
 332
 333         /* don't remove a peer_ni if it's also a gateway */
 334         if (lpni->lpni_rtr_refcount > 0) {
 335                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 336                        libcfs_nid2str(lpni->lpni_nid));
 337                 return -EBUSY;
 338         }
 339
 340         lnet_peer_remove_from_remote_list(lpni);
 341
 342         /* remove peer ni from the hash list. */
 343         list_del_init(&lpni->lpni_hashlist);
 344
 345         /* decrement the ref count on the peer table */
 346         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 347         LASSERT(ptable->pt_number > 0);
 348         ptable->pt_number--;
 349
 350         /*
 351          * The peer_ni can no longer be found with a lookup. But there
 352          * can be current users, so keep track of it on the zombie
 353          * list until the reference count has gone to zero.
 354          *
 355          * The last reference may be lost in a place where the
 356          * lnet_net_lock locks only a single cpt, and that cpt may not
 357          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 358          * has its own lock.
 359          */
 360         spin_lock(&ptable->pt_zombie_lock);
 361         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 362         ptable->pt_zombies++;
 363         spin_unlock(&ptable->pt_zombie_lock);
 364
 365         /* no need to keep this peer_ni on the hierarchy anymore */
 366         lnet_peer_detach_peer_ni_locked(lpni);
 367
 368         /* remove hashlist reference on peer_ni */
 369         lnet_peer_ni_decref_locked(lpni);
 370
 371         return 0;
 372 }
 373
 374 void lnet_peer_uninit(void)
 375 {
 376         struct lnet_peer_ni *lpni, *tmp;
 377
 378         lnet_net_lock(LNET_LOCK_EX);
 379
 380         /* remove all peer_nis from the remote peer and the hash list */
 381         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 382                                  lpni_on_remote_peer_ni_list)
 383                 lnet_peer_ni_del_locked(lpni);
 384
 385         lnet_peer_tables_destroy();
 386
 387         lnet_net_unlock(LNET_LOCK_EX);
 388 }
 389
 390 static int
 391 lnet_peer_del_locked(struct lnet_peer *peer)
 392 {
 393         struct lnet_peer_ni *lpni = NULL, *lpni2;
 394         int rc = 0, rc2 = 0;
 395
 396         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
 397
 398         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 399         while (lpni != NULL) {
 400                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 401                 rc = lnet_peer_ni_del_locked(lpni);
 402                 if (rc != 0)
 403                         rc2 = rc;
 404                 lpni = lpni2;
 405         }
 406
 407         return rc2;
 408 }
 409
 410 static int
 411 lnet_peer_del(struct lnet_peer *peer)
 412 {
 413         lnet_net_lock(LNET_LOCK_EX);
 414         lnet_peer_del_locked(peer);
 415         lnet_net_unlock(LNET_LOCK_EX);
 416
 417         return 0;
 418 }
 419
 420 /*
 421  * Delete a NID from a peer. Call with ln_api_mutex held.
 422  *
 423  * Error codes:
 424  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 425  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 426  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 427  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 428  */
 429 static int
 430 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
 431 {
 432         struct lnet_peer_ni *lpni;
 433         lnet_nid_t primary_nid = lp->lp_primary_nid;
 434         int rc = 0;
 435
 436         if (!(flags & LNET_PEER_CONFIGURED)) {
 437                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 438                         rc = -EPERM;
 439                         goto out;
 440                 }
 441         }
 442         lpni = lnet_find_peer_ni_locked(nid);
 443         if (!lpni) {
 444                 rc = -ENOENT;
 445                 goto out;
 446         }
 447         lnet_peer_ni_decref_locked(lpni);
 448         if (lp != lpni->lpni_peer_net->lpn_peer) {
 449                 rc = -ECHILD;
 450                 goto out;
 451         }
 452
 453         /*
 454          * This function only allows deletion of the primary NID if it
 455          * is the only NID.
 456          */
 457         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1) {
 458                 rc = -EBUSY;
 459                 goto out;
 460         }
 461
 462         lnet_net_lock(LNET_LOCK_EX);
 463         lnet_peer_ni_del_locked(lpni);
 464         lnet_net_unlock(LNET_LOCK_EX);
 465
 466 out:
 467         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 468                libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
 469
 470         return rc;
 471 }
 472
 473 static void
 474 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 475                                struct lnet_peer_table *ptable)
 476 {
 477         int                      i;
 478         struct lnet_peer_ni     *next;
 479         struct lnet_peer_ni     *lpni;
 480         struct lnet_peer        *peer;
 481
 482         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 483                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 484                                          lpni_hashlist) {
 485                         if (net != NULL && net != lpni->lpni_net)
 486                                 continue;
 487
 488                         peer = lpni->lpni_peer_net->lpn_peer;
 489                         if (peer->lp_primary_nid != lpni->lpni_nid) {
 490                                 lnet_peer_ni_del_locked(lpni);
 491                                 continue;
 492                         }
 493                         /*
 494                          * Removing the primary NID implies removing
 495                          * the entire peer. Advance next beyond any
 496                          * peer_ni that belongs to the same peer.
 497                          */
 498                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 499                                                  lpni_hashlist) {
 500                                 if (next->lpni_peer_net->lpn_peer != peer)
 501                                         break;
 502                         }
 503                         lnet_peer_del_locked(peer);
 504                 }
 505         }
 506 }
 507
 508 static void
 509 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 510 {
 511         int     i = 3;
 512
 513         spin_lock(&ptable->pt_zombie_lock);
 514         while (ptable->pt_zombies) {
 515                 spin_unlock(&ptable->pt_zombie_lock);
 516
 517                 if (is_power_of_2(i)) {
 518                         CDEBUG(D_WARNING,
 519                                "Waiting for %d zombies on peer table\n",
 520                                ptable->pt_zombies);
 521                 }
 522                 set_current_state(TASK_UNINTERRUPTIBLE);
 523                 schedule_timeout(cfs_time_seconds(1) >> 1);
 524                 spin_lock(&ptable->pt_zombie_lock);
 525         }
 526         spin_unlock(&ptable->pt_zombie_lock);
 527 }
 528
 529 static void
 530 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 531                                 struct lnet_peer_table *ptable)
 532 {
 533         struct lnet_peer_ni     *lp;
 534         struct lnet_peer_ni     *tmp;
 535         lnet_nid_t              lpni_nid;
 536         int                     i;
 537
 538         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 539                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 540                                          lpni_hashlist) {
 541                         if (net != lp->lpni_net)
 542                                 continue;
 543
 544                         if (lp->lpni_rtr_refcount == 0)
 545                                 continue;
 546
 547                         lpni_nid = lp->lpni_nid;
 548
 549                         lnet_net_unlock(LNET_LOCK_EX);
 550                         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), lpni_nid);
 551                         lnet_net_lock(LNET_LOCK_EX);
 552                 }
 553         }
 554 }
 555
 556 void
 557 lnet_peer_tables_cleanup(struct lnet_net *net)
 558 {
 559         int                             i;
 560         struct lnet_peer_table          *ptable;
 561
 562         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 563         /* If just deleting the peers for a NI, get rid of any routes these
 564          * peers are gateways for. */
 565         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 566                 lnet_net_lock(LNET_LOCK_EX);
 567                 lnet_peer_table_del_rtrs_locked(net, ptable);
 568                 lnet_net_unlock(LNET_LOCK_EX);
 569         }
 570
 571         /* Start the cleanup process */
 572         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 573                 lnet_net_lock(LNET_LOCK_EX);
 574                 lnet_peer_table_cleanup_locked(net, ptable);
 575                 lnet_net_unlock(LNET_LOCK_EX);
 576         }
 577
 578         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 579                 lnet_peer_ni_finalize_wait(ptable);
 580 }
 581
 582 static struct lnet_peer_ni *
 583 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 584 {
 585         struct list_head        *peers;
 586         struct lnet_peer_ni     *lp;
 587
 588         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 589
 590         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 591         list_for_each_entry(lp, peers, lpni_hashlist) {
 592                 if (lp->lpni_nid == nid) {
 593                         lnet_peer_ni_addref_locked(lp);
 594                         return lp;
 595                 }
 596         }
 597
 598         return NULL;
 599 }
 600
 601 struct lnet_peer_ni *
 602 lnet_find_peer_ni_locked(lnet_nid_t nid)
 603 {
 604         struct lnet_peer_ni *lpni;
 605         struct lnet_peer_table *ptable;
 606         int cpt;
 607
 608         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 609
 610         ptable = the_lnet.ln_peer_tables[cpt];
 611         lpni = lnet_get_peer_ni_locked(ptable, nid);
 612
 613         return lpni;
 614 }
 615
 616 struct lnet_peer *
 617 lnet_find_peer(lnet_nid_t nid)
 618 {
 619         struct lnet_peer_ni *lpni;
 620         struct lnet_peer *lp = NULL;
 621         int cpt;
 622
 623         cpt = lnet_net_lock_current();
 624         lpni = lnet_find_peer_ni_locked(nid);
 625         if (lpni) {
 626                 lp = lpni->lpni_peer_net->lpn_peer;
 627                 lnet_peer_addref_locked(lp);
 628                 lnet_peer_ni_decref_locked(lpni);
 629         }
 630         lnet_net_unlock(cpt);
 631
 632         return lp;
 633 }
 634
 635 struct lnet_peer_ni *
 636 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 637                              struct lnet_peer_net *peer_net,
 638                              struct lnet_peer_ni *prev)
 639 {
 640         struct lnet_peer_ni *lpni;
 641         struct lnet_peer_net *net = peer_net;
 642
 643         if (!prev) {
 644                 if (!net) {
 645                         if (list_empty(&peer->lp_peer_nets))
 646                                 return NULL;
 647
 648                         net = list_entry(peer->lp_peer_nets.next,
 649                                          struct lnet_peer_net,
 650                                          lpn_peer_nets);
 651                 }
 652                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 653                                   lpni_peer_nis);
 654
 655                 return lpni;
 656         }
 657
 658         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 659                 /*
 660                  * if you reached the end of the peer ni list and the peer
 661                  * net is specified then there are no more peer nis in that
 662                  * net.
 663                  */
 664                 if (net)
 665                         return NULL;
 666
 667                 /*
 668                  * we reached the end of this net ni list. move to the
 669                  * next net
 670                  */
 671                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 672                     &peer->lp_peer_nets)
 673                         /* no more nets and no more NIs. */
 674                         return NULL;
 675
 676                 /* get the next net */
 677                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 678                                  struct lnet_peer_net,
 679                                  lpn_peer_nets);
 680                 /* get the ni on it */
 681                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 682                                   lpni_peer_nis);
 683
 684                 return lpni;
 685         }
 686
 687         /* there are more nis left */
 688         lpni = list_entry(prev->lpni_peer_nis.next,
 689                           struct lnet_peer_ni, lpni_peer_nis);
 690
 691         return lpni;
 692 }
 693
 694 /* Call with the ln_api_mutex held */
 695 int
 696 lnet_get_peer_list(__u32 *countp, __u32 *sizep, lnet_process_id_t __user *ids)
 697 {
 698         lnet_process_id_t id;
 699         struct lnet_peer_table *ptable;
 700         struct lnet_peer *lp;
 701         __u32 count = 0;
 702         __u32 size = 0;
 703         int lncpt;
 704         int cpt;
 705         __u32 i;
 706         int rc;
 707
 708         rc = -ESHUTDOWN;
 709         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 710                 goto done;
 711
 712         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 713
 714         /*
 715          * Count the number of peers, and return E2BIG if the buffer
 716          * is too small. We'll also return the desired size.
 717          */
 718         rc = -E2BIG;
 719         for (cpt = 0; cpt < lncpt; cpt++) {
 720                 ptable = the_lnet.ln_peer_tables[cpt];
 721                 count += ptable->pt_peers;
 722         }
 723         size = count * sizeof(*ids);
 724         if (size > *sizep)
 725                 goto done;
 726
 727         /*
 728          * Walk the peer lists and copy out the primary nids.
 729          * This is safe because the peer lists are only modified
 730          * while the ln_api_mutex is held. So we don't need to
 731          * hold the lnet_net_lock as well, and can therefore
 732          * directly call copy_to_user().
 733          */
 734         rc = -EFAULT;
 735         memset(&id, 0, sizeof(id));
 736         id.pid = LNET_PID_LUSTRE;
 737         i = 0;
 738         for (cpt = 0; cpt < lncpt; cpt++) {
 739                 ptable = the_lnet.ln_peer_tables[cpt];
 740                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 741                         if (i >= count)
 742                                 goto done;
 743                         id.nid = lp->lp_primary_nid;
 744                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 745                                 goto done;
 746                         i++;
 747                 }
 748         }
 749         rc = 0;
 750 done:
 751         *countp = count;
 752         *sizep = size;
 753         return rc;
 754 }
 755
 756 /*
 757  * Start pushes to peers that need to be updated for a configuration
 758  * change on this node.
 759  */
 760 void
 761 lnet_push_update_to_peers(int force)
 762 {
 763         struct lnet_peer_table *ptable;
 764         struct lnet_peer *lp;
 765         int lncpt;
 766         int cpt;
 767
 768         lnet_net_lock(LNET_LOCK_EX);
 769         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 770         for (cpt = 0; cpt < lncpt; cpt++) {
 771                 ptable = the_lnet.ln_peer_tables[cpt];
 772                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 773                         if (force) {
 774                                 spin_lock(&lp->lp_lock);
 775                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 776                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 777                                 spin_unlock(&lp->lp_lock);
 778                         }
 779                         if (lnet_peer_needs_push(lp))
 780                                 lnet_peer_queue_for_discovery(lp);
 781                 }
 782         }
 783         lnet_net_unlock(LNET_LOCK_EX);
 784         wake_up(&the_lnet.ln_dc_waitq);
 785 }
 786
 787 /*
 788  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 789  * this is a preferred point-to-point path. Call with lnet_net_lock in
 790  * shared mmode.
 791  */
 792 bool
 793 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 794 {
 795         int i;
 796
 797         if (lpni->lpni_pref_nnids == 0)
 798                 return false;
 799         if (lpni->lpni_pref_nnids == 1)
 800                 return lpni->lpni_pref.nid == nid;
 801         for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 802                 if (lpni->lpni_pref.nids[i] == nid)
 803                         return true;
 804         }
 805         return false;
 806 }
 807
 808 /*
 809  * Set a single ni as preferred, provided no preferred ni is already
 810  * defined. Only to be used for non-multi-rail peer_ni.
 811  */
 812 int
 813 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 814 {
 815         int rc = 0;
 816
 817         spin_lock(&lpni->lpni_lock);
 818         if (nid == LNET_NID_ANY) {
 819                 rc = -EINVAL;
 820         } else if (lpni->lpni_pref_nnids > 0) {
 821                 rc = -EPERM;
 822         } else if (lpni->lpni_pref_nnids == 0) {
 823                 lpni->lpni_pref.nid = nid;
 824                 lpni->lpni_pref_nnids = 1;
 825                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
 826         }
 827         spin_unlock(&lpni->lpni_lock);
 828
 829         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 830                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
 831         return rc;
 832 }
 833
 834 /*
 835  * Clear the preferred NID from a non-multi-rail peer_ni, provided
 836  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
 837  */
 838 int
 839 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
 840 {
 841         int rc = 0;
 842
 843         spin_lock(&lpni->lpni_lock);
 844         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
 845                 lpni->lpni_pref_nnids = 0;
 846                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 847         } else if (lpni->lpni_pref_nnids == 0) {
 848                 rc = -ENOENT;
 849         } else {
 850                 rc = -EPERM;
 851         }
 852         spin_unlock(&lpni->lpni_lock);
 853
 854         CDEBUG(D_NET, "peer %s: %d\n",
 855                libcfs_nid2str(lpni->lpni_nid), rc);
 856         return rc;
 857 }
 858
 859 /*
 860  * Clear the preferred NIDs from a non-multi-rail peer.
 861  */
 862 void
 863 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
 864 {
 865         struct lnet_peer_ni *lpni = NULL;
 866
 867         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 868                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
 869 }
 870
 871 int
 872 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 873 {
 874         lnet_nid_t *nids = NULL;
 875         lnet_nid_t *oldnids = NULL;
 876         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 877         int size;
 878         int i;
 879         int rc = 0;
 880
 881         if (nid == LNET_NID_ANY) {
 882                 rc = -EINVAL;
 883                 goto out;
 884         }
 885
 886         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
 887                 rc = -EEXIST;
 888                 goto out;
 889         }
 890
 891         /* A non-MR node may have only one preferred NI per peer_ni */
 892         if (lpni->lpni_pref_nnids > 0) {
 893                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
 894                         rc = -EPERM;
 895                         goto out;
 896                 }
 897         }
 898
 899         if (lpni->lpni_pref_nnids != 0) {
 900                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
 901                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
 902                 if (!nids) {
 903                         rc = -ENOMEM;
 904                         goto out;
 905                 }
 906                 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 907                         if (lpni->lpni_pref.nids[i] == nid) {
 908                                 LIBCFS_FREE(nids, size);
 909                                 rc = -EEXIST;
 910                                 goto out;
 911                         }
 912                         nids[i] = lpni->lpni_pref.nids[i];
 913                 }
 914                 nids[i] = nid;
 915         }
 916
 917         lnet_net_lock(LNET_LOCK_EX);
 918         spin_lock(&lpni->lpni_lock);
 919         if (lpni->lpni_pref_nnids == 0) {
 920                 lpni->lpni_pref.nid = nid;
 921         } else {
 922                 oldnids = lpni->lpni_pref.nids;
 923                 lpni->lpni_pref.nids = nids;
 924         }
 925         lpni->lpni_pref_nnids++;
 926         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 927         spin_unlock(&lpni->lpni_lock);
 928         lnet_net_unlock(LNET_LOCK_EX);
 929
 930         if (oldnids) {
 931                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
 932                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
 933         }
 934 out:
 935         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
 936                 spin_lock(&lpni->lpni_lock);
 937                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 938                 spin_unlock(&lpni->lpni_lock);
 939         }
 940         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 941                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
 942         return rc;
 943 }
 944
 945 int
 946 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 947 {
 948         lnet_nid_t *nids = NULL;
 949         lnet_nid_t *oldnids = NULL;
 950         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 951         int size;
 952         int i, j;
 953         int rc = 0;
 954
 955         if (lpni->lpni_pref_nnids == 0) {
 956                 rc = -ENOENT;
 957                 goto out;
 958         }
 959
 960         if (lpni->lpni_pref_nnids == 1) {
 961                 if (lpni->lpni_pref.nid != nid) {
 962                         rc = -ENOENT;
 963                         goto out;
 964                 }
 965         } else if (lpni->lpni_pref_nnids == 2) {
 966                 if (lpni->lpni_pref.nids[0] != nid &&
 967                     lpni->lpni_pref.nids[1] != nid) {
 968                         rc = -ENOENT;
 969                         goto out;
 970                 }
 971         } else {
 972                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
 973                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
 974                 if (!nids) {
 975                         rc = -ENOMEM;
 976                         goto out;
 977                 }
 978                 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
 979                         if (lpni->lpni_pref.nids[i] != nid)
 980                                 continue;
 981                         nids[j++] = lpni->lpni_pref.nids[i];
 982                 }
 983                 /* Check if we actually removed a nid. */
 984                 if (j == lpni->lpni_pref_nnids) {
 985                         LIBCFS_FREE(nids, size);
 986                         rc = -ENOENT;
 987                         goto out;
 988                 }
 989         }
 990
 991         lnet_net_lock(LNET_LOCK_EX);
 992         spin_lock(&lpni->lpni_lock);
 993         if (lpni->lpni_pref_nnids == 1) {
 994                 lpni->lpni_pref.nid = LNET_NID_ANY;
 995         } else if (lpni->lpni_pref_nnids == 2) {
 996                 oldnids = lpni->lpni_pref.nids;
 997                 if (oldnids[0] == nid)
 998                         lpni->lpni_pref.nid = oldnids[1];
 999                 else
1000                         lpni->lpni_pref.nid = oldnids[2];
1001         } else {
1002                 oldnids = lpni->lpni_pref.nids;
1003                 lpni->lpni_pref.nids = nids;
1004         }
1005         lpni->lpni_pref_nnids--;
1006         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1007         spin_unlock(&lpni->lpni_lock);
1008         lnet_net_unlock(LNET_LOCK_EX);
1009
1010         if (oldnids) {
1011                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1012                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1013         }
1014 out:
1015         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1016                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1017         return rc;
1018 }
1019
1020 lnet_nid_t
1021 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1022 {
1023         struct lnet_peer_ni *lpni;
1024         lnet_nid_t primary_nid = nid;
1025
1026         lpni = lnet_find_peer_ni_locked(nid);
1027         if (lpni) {
1028                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1029                 lnet_peer_ni_decref_locked(lpni);
1030         }
1031
1032         return primary_nid;
1033 }
1034
1035 lnet_nid_t
1036 LNetPrimaryNID(lnet_nid_t nid)
1037 {
1038         struct lnet_peer *lp;
1039         struct lnet_peer_ni *lpni;
1040         lnet_nid_t primary_nid = nid;
1041         int rc = 0;
1042         int cpt;
1043
1044         cpt = lnet_net_lock_current();
1045         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1046         if (IS_ERR(lpni)) {
1047                 rc = PTR_ERR(lpni);
1048                 goto out_unlock;
1049         }
1050         lp = lpni->lpni_peer_net->lpn_peer;
1051         while (!lnet_peer_is_uptodate(lp)) {
1052                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1053                 if (rc)
1054                         goto out_decref;
1055                 lp = lpni->lpni_peer_net->lpn_peer;
1056         }
1057         primary_nid = lp->lp_primary_nid;
1058 out_decref:
1059         lnet_peer_ni_decref_locked(lpni);
1060 out_unlock:
1061         lnet_net_unlock(cpt);
1062
1063         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1064                libcfs_nid2str(primary_nid), rc);
1065         return primary_nid;
1066 }
1067 EXPORT_SYMBOL(LNetPrimaryNID);
1068
1069 struct lnet_peer_net *
1070 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1071 {
1072         struct lnet_peer_net *peer_net;
1073         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1074                 if (peer_net->lpn_net_id == net_id)
1075                         return peer_net;
1076         }
1077         return NULL;
1078 }
1079
1080 /*
1081  * Attach a peer_ni to a peer_net and peer. This function assumes
1082  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1083  * may be attached to a different peer, in which case it will be
1084  * properly detached first. The whole operation is done atomically.
1085  *
1086  * Always returns 0.  This is the last function called from functions
1087  * that do return an int, so returning 0 here allows the compiler to
1088  * do a tail call.
1089  */
1090 static int
1091 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1092                                 struct lnet_peer_net *lpn,
1093                                 struct lnet_peer_ni *lpni,
1094                                 unsigned flags)
1095 {
1096         struct lnet_peer_table *ptable;
1097
1098         /* Install the new peer_ni */
1099         lnet_net_lock(LNET_LOCK_EX);
1100         /* Add peer_ni to global peer table hash, if necessary. */
1101         if (list_empty(&lpni->lpni_hashlist)) {
1102                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1103
1104                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1105                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1106                 ptable->pt_version++;
1107                 ptable->pt_number++;
1108                 /* This is the 1st refcount on lpni. */
1109                 atomic_inc(&lpni->lpni_refcount);
1110         }
1111
1112         /* Detach the peer_ni from an existing peer, if necessary. */
1113         if (lpni->lpni_peer_net) {
1114                 LASSERT(lpni->lpni_peer_net != lpn);
1115                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1116                 lnet_peer_detach_peer_ni_locked(lpni);
1117                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1118                 lpni->lpni_peer_net = NULL;
1119         }
1120
1121         /* Add peer_ni to peer_net */
1122         lpni->lpni_peer_net = lpn;
1123         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1124         lnet_peer_net_addref_locked(lpn);
1125
1126         /* Add peer_net to peer */
1127         if (!lpn->lpn_peer) {
1128                 lpn->lpn_peer = lp;
1129                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1130                 lnet_peer_addref_locked(lp);
1131         }
1132
1133         /* Add peer to global peer list, if necessary */
1134         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1135         if (list_empty(&lp->lp_peer_list)) {
1136                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1137                 ptable->pt_peers++;
1138         }
1139
1140
1141         /* Update peer state */
1142         spin_lock(&lp->lp_lock);
1143         if (flags & LNET_PEER_CONFIGURED) {
1144                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1145                         lp->lp_state |= LNET_PEER_CONFIGURED;
1146         }
1147         if (flags & LNET_PEER_MULTI_RAIL) {
1148                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1149                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1150                         lnet_peer_clr_non_mr_pref_nids(lp);
1151                 }
1152         }
1153         spin_unlock(&lp->lp_lock);
1154
1155         lp->lp_nnis++;
1156         lnet_net_unlock(LNET_LOCK_EX);
1157
1158         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1159                libcfs_nid2str(lp->lp_primary_nid),
1160                libcfs_nid2str(lpni->lpni_nid), flags);
1161
1162         return 0;
1163 }
1164
1165 /*
1166  * Create a new peer, with nid as its primary nid.
1167  *
1168  * Call with the lnet_api_mutex held.
1169  */
1170 static int
1171 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1172 {
1173         struct lnet_peer *lp;
1174         struct lnet_peer_net *lpn;
1175         struct lnet_peer_ni *lpni;
1176         int rc = 0;
1177
1178         LASSERT(nid != LNET_NID_ANY);
1179
1180         /*
1181          * No need for the lnet_net_lock here, because the
1182          * lnet_api_mutex is held.
1183          */
1184         lpni = lnet_find_peer_ni_locked(nid);
1185         if (lpni) {
1186                 /* A peer with this NID already exists. */
1187                 lp = lpni->lpni_peer_net->lpn_peer;
1188                 lnet_peer_ni_decref_locked(lpni);
1189                 /*
1190                  * This is an error if the peer was configured and the
1191                  * primary NID differs or an attempt is made to change
1192                  * the Multi-Rail flag. Otherwise the assumption is
1193                  * that an existing peer is being modified.
1194                  */
1195                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1196                         if (lp->lp_primary_nid != nid)
1197                                 rc = -EEXIST;
1198                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1199                                 rc = -EPERM;
1200                         goto out;
1201                 }
1202                 /* Delete and recreate as a configured peer. */
1203                 lnet_peer_del(lp);
1204         }
1205
1206         /* Create peer, peer_net, and peer_ni. */
1207         rc = -ENOMEM;
1208         lp = lnet_peer_alloc(nid);
1209         if (!lp)
1210                 goto out;
1211         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1212         if (!lpn)
1213                 goto out_free_lp;
1214         lpni = lnet_peer_ni_alloc(nid);
1215         if (!lpni)
1216                 goto out_free_lpn;
1217
1218         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1219
1220 out_free_lpn:
1221         LIBCFS_FREE(lpn, sizeof(*lpn));
1222 out_free_lp:
1223         LIBCFS_FREE(lp, sizeof(*lp));
1224 out:
1225         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1226                libcfs_nid2str(nid), flags, rc);
1227         return rc;
1228 }
1229
1230 /*
1231  * Add a NID to a peer. Call with ln_api_mutex held.
1232  *
1233  * Error codes:
1234  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1235  *  -EEXIST:   The NID was configured by DLC for a different peer.
1236  *  -ENOMEM:   Out of memory.
1237  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1238  *             non-multi-rail peer.
1239  */
1240 static int
1241 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1242 {
1243         struct lnet_peer_net *lpn;
1244         struct lnet_peer_ni *lpni;
1245         int rc = 0;
1246
1247         LASSERT(lp);
1248         LASSERT(nid != LNET_NID_ANY);
1249
1250         /* A configured peer can only be updated through configuration. */
1251         if (!(flags & LNET_PEER_CONFIGURED)) {
1252                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1253                         rc = -EPERM;
1254                         goto out;
1255                 }
1256         }
1257
1258         /*
1259          * The MULTI_RAIL flag can be set but not cleared, because
1260          * that would leave the peer struct in an invalid state.
1261          */
1262         if (flags & LNET_PEER_MULTI_RAIL) {
1263                 spin_lock(&lp->lp_lock);
1264                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1265                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1266                         lnet_peer_clr_non_mr_pref_nids(lp);
1267                 }
1268                 spin_unlock(&lp->lp_lock);
1269         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1270                 rc = -EPERM;
1271                 goto out;
1272         }
1273
1274         lpni = lnet_find_peer_ni_locked(nid);
1275         if (lpni) {
1276                 /*
1277                  * A peer_ni already exists. This is only a problem if
1278                  * it is not connected to this peer and was configured
1279                  * by DLC.
1280                  */
1281                 lnet_peer_ni_decref_locked(lpni);
1282                 if (lpni->lpni_peer_net->lpn_peer == lp)
1283                         goto out;
1284                 if (lnet_peer_ni_is_configured(lpni)) {
1285                         rc = -EEXIST;
1286                         goto out;
1287                 }
1288                 /* If this is the primary NID, destroy the peer. */
1289                 if (lnet_peer_ni_is_primary(lpni)) {
1290                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1291                         lpni = lnet_peer_ni_alloc(nid);
1292                         if (!lpni) {
1293                                 rc = -ENOMEM;
1294                                 goto out;
1295                         }
1296                 }
1297         } else {
1298                 lpni = lnet_peer_ni_alloc(nid);
1299                 if (!lpni) {
1300                         rc = -ENOMEM;
1301                         goto out;
1302                 }
1303         }
1304
1305         /*
1306          * Get the peer_net. Check that we're not adding a second
1307          * peer_ni on a peer_net of a non-multi-rail peer.
1308          */
1309         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1310         if (!lpn) {
1311                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1312                 if (!lpn) {
1313                         rc = -ENOMEM;
1314                         goto out_free_lpni;
1315                 }
1316         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1317                 rc = -ENOTUNIQ;
1318                 goto out_free_lpni;
1319         }
1320
1321         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1322
1323 out_free_lpni:
1324         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1325         if (!lpni->lpni_peer_net)
1326                 LIBCFS_FREE(lpni, sizeof(*lpni));
1327 out:
1328         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1329                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1330                flags, rc);
1331         return rc;
1332 }
1333
1334 /*
1335  * Update the primary NID of a peer, if possible.
1336  *
1337  * Call with the lnet_api_mutex held.
1338  */
1339 static int
1340 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1341 {
1342         lnet_nid_t old = lp->lp_primary_nid;
1343         int rc = 0;
1344
1345         if (lp->lp_primary_nid == nid)
1346                 goto out;
1347         rc = lnet_peer_add_nid(lp, nid, flags);
1348         if (rc)
1349                 goto out;
1350         lp->lp_primary_nid = nid;
1351 out:
1352         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1353                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1354         return rc;
1355 }
1356
1357 /*
1358  * lpni creation initiated due to traffic either sending or receiving.
1359  */
1360 static int
1361 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1362 {
1363         struct lnet_peer *lp;
1364         struct lnet_peer_net *lpn;
1365         struct lnet_peer_ni *lpni;
1366         unsigned flags = 0;
1367         int rc = 0;
1368
1369         if (nid == LNET_NID_ANY) {
1370                 rc = -EINVAL;
1371                 goto out;
1372         }
1373
1374         /* lnet_net_lock is not needed here because ln_api_lock is held */
1375         lpni = lnet_find_peer_ni_locked(nid);
1376         if (lpni) {
1377                 /*
1378                  * We must have raced with another thread. Since we
1379                  * know next to nothing about a peer_ni created by
1380                  * traffic, we just assume everything is ok and
1381                  * return.
1382                  */
1383                 lnet_peer_ni_decref_locked(lpni);
1384                 goto out;
1385         }
1386
1387         /* Create peer, peer_net, and peer_ni. */
1388         rc = -ENOMEM;
1389         lp = lnet_peer_alloc(nid);
1390         if (!lp)
1391                 goto out;
1392         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1393         if (!lpn)
1394                 goto out_free_lp;
1395         lpni = lnet_peer_ni_alloc(nid);
1396         if (!lpni)
1397                 goto out_free_lpn;
1398         if (pref != LNET_NID_ANY)
1399                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1400
1401         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1402
1403 out_free_lpn:
1404         LIBCFS_FREE(lpn, sizeof(*lpn));
1405 out_free_lp:
1406         LIBCFS_FREE(lp, sizeof(*lp));
1407 out:
1408         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1409         return rc;
1410 }
1411
1412 /*
1413  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1414  *
1415  * This API handles the following combinations:
1416  *   Create a peer with its primary NI if only the prim_nid is provided
1417  *   Add a NID to a peer identified by the prim_nid. The peer identified
1418  *   by the prim_nid must already exist.
1419  *   The peer being created may be non-MR.
1420  *
1421  * The caller must hold ln_api_mutex. This prevents the peer from
1422  * being created/modified/deleted by a different thread.
1423  */
1424 int
1425 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1426 {
1427         struct lnet_peer *lp = NULL;
1428         struct lnet_peer_ni *lpni;
1429         unsigned flags;
1430
1431         /* The prim_nid must always be specified */
1432         if (prim_nid == LNET_NID_ANY)
1433                 return -EINVAL;
1434
1435         flags = LNET_PEER_CONFIGURED;
1436         if (mr)
1437                 flags |= LNET_PEER_MULTI_RAIL;
1438
1439         /*
1440          * If nid isn't specified, we must create a new peer with
1441          * prim_nid as its primary nid.
1442          */
1443         if (nid == LNET_NID_ANY)
1444                 return lnet_peer_add(prim_nid, flags);
1445
1446         /* Look up the prim_nid, which must exist. */
1447         lpni = lnet_find_peer_ni_locked(prim_nid);
1448         if (!lpni)
1449                 return -ENOENT;
1450         lnet_peer_ni_decref_locked(lpni);
1451         lp = lpni->lpni_peer_net->lpn_peer;
1452
1453         /* Peer must have been configured. */
1454         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1455                 CDEBUG(D_NET, "peer %s was not configured\n",
1456                        libcfs_nid2str(prim_nid));
1457                 return -ENOENT;
1458         }
1459
1460         /* Primary NID must match */
1461         if (lp->lp_primary_nid != prim_nid) {
1462                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1463                        libcfs_nid2str(prim_nid),
1464                        libcfs_nid2str(lp->lp_primary_nid));
1465                 return -ENODEV;
1466         }
1467
1468         /* Multi-Rail flag must match. */
1469         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1470                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1471                        libcfs_nid2str(prim_nid));
1472                 return -EPERM;
1473         }
1474
1475         return lnet_peer_add_nid(lp, nid, flags);
1476 }
1477
1478 /*
1479  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1480  *
1481  * This API handles the following combinations:
1482  *   Delete a NI from a peer if both prim_nid and nid are provided.
1483  *   Delete a peer if only prim_nid is provided.
1484  *   Delete a peer if its primary nid is provided.
1485  *
1486  * The caller must hold ln_api_mutex. This prevents the peer from
1487  * being modified/deleted by a different thread.
1488  */
1489 int
1490 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1491 {
1492         struct lnet_peer *lp;
1493         struct lnet_peer_ni *lpni;
1494         unsigned flags;
1495
1496         if (prim_nid == LNET_NID_ANY)
1497                 return -EINVAL;
1498
1499         lpni = lnet_find_peer_ni_locked(prim_nid);
1500         if (!lpni)
1501                 return -ENOENT;
1502         lnet_peer_ni_decref_locked(lpni);
1503         lp = lpni->lpni_peer_net->lpn_peer;
1504
1505         if (prim_nid != lp->lp_primary_nid) {
1506                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1507                        libcfs_nid2str(prim_nid),
1508                        libcfs_nid2str(lp->lp_primary_nid));
1509                 return -ENODEV;
1510         }
1511
1512         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1513                 return lnet_peer_del(lp);
1514
1515         flags = LNET_PEER_CONFIGURED;
1516         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1517                 flags |= LNET_PEER_MULTI_RAIL;
1518
1519         return lnet_peer_del_nid(lp, nid, flags);
1520 }
1521
1522 void
1523 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1524 {
1525         struct lnet_peer_table *ptable;
1526         struct lnet_peer_net *lpn;
1527
1528         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1529
1530         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1531         LASSERT(lpni->lpni_rtr_refcount == 0);
1532         LASSERT(list_empty(&lpni->lpni_txq));
1533         LASSERT(lpni->lpni_txqnob == 0);
1534
1535         lpn = lpni->lpni_peer_net;
1536         lpni->lpni_peer_net = NULL;
1537         lpni->lpni_net = NULL;
1538
1539         /* remove the peer ni from the zombie list */
1540         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1541         spin_lock(&ptable->pt_zombie_lock);
1542         list_del_init(&lpni->lpni_hashlist);
1543         ptable->pt_zombies--;
1544         spin_unlock(&ptable->pt_zombie_lock);
1545
1546         if (lpni->lpni_pref_nnids > 1) {
1547                 LIBCFS_FREE(lpni->lpni_pref.nids,
1548                         sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1549         }
1550         LIBCFS_FREE(lpni, sizeof(*lpni));
1551
1552         lnet_peer_net_decref_locked(lpn);
1553 }
1554
1555 struct lnet_peer_ni *
1556 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1557 {
1558         struct lnet_peer_ni *lpni = NULL;
1559         int rc;
1560
1561         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1562                 return ERR_PTR(-ESHUTDOWN);
1563
1564         /*
1565          * find if a peer_ni already exists.
1566          * If so then just return that.
1567          */
1568         lpni = lnet_find_peer_ni_locked(nid);
1569         if (lpni)
1570                 return lpni;
1571
1572         lnet_net_unlock(cpt);
1573
1574         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1575         if (rc) {
1576                 lpni = ERR_PTR(rc);
1577                 goto out_net_relock;
1578         }
1579
1580         lpni = lnet_find_peer_ni_locked(nid);
1581         LASSERT(lpni);
1582
1583 out_net_relock:
1584         lnet_net_lock(cpt);
1585
1586         return lpni;
1587 }
1588
1589 /*
1590  * Get a peer_ni for the given nid, create it if necessary. Takes a
1591  * hold on the peer_ni.
1592  */
1593 struct lnet_peer_ni *
1594 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1595 {
1596         struct lnet_peer_ni *lpni = NULL;
1597         int rc;
1598
1599         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1600                 return ERR_PTR(-ESHUTDOWN);
1601
1602         /*
1603          * find if a peer_ni already exists.
1604          * If so then just return that.
1605          */
1606         lpni = lnet_find_peer_ni_locked(nid);
1607         if (lpni)
1608                 return lpni;
1609
1610         /*
1611          * Slow path:
1612          * use the lnet_api_mutex to serialize the creation of the peer_ni
1613          * and the creation/deletion of the local ni/net. When a local ni is
1614          * created, if there exists a set of peer_nis on that network,
1615          * they need to be traversed and updated. When a local NI is
1616          * deleted, which could result in a network being deleted, then
1617          * all peer nis on that network need to be removed as well.
1618          *
1619          * Creation through traffic should also be serialized with
1620          * creation through DLC.
1621          */
1622         lnet_net_unlock(cpt);
1623         mutex_lock(&the_lnet.ln_api_mutex);
1624         /*
1625          * Shutdown is only set under the ln_api_lock, so a single
1626          * check here is sufficent.
1627          */
1628         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1629                 lpni = ERR_PTR(-ESHUTDOWN);
1630                 goto out_mutex_unlock;
1631         }
1632
1633         rc = lnet_peer_ni_traffic_add(nid, pref);
1634         if (rc) {
1635                 lpni = ERR_PTR(rc);
1636                 goto out_mutex_unlock;
1637         }
1638
1639         lpni = lnet_find_peer_ni_locked(nid);
1640         LASSERT(lpni);
1641
1642 out_mutex_unlock:
1643         mutex_unlock(&the_lnet.ln_api_mutex);
1644         lnet_net_lock(cpt);
1645
1646         /* Lock has been dropped, check again for shutdown. */
1647         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1648                 if (!IS_ERR(lpni))
1649                         lnet_peer_ni_decref_locked(lpni);
1650                 lpni = ERR_PTR(-ESHUTDOWN);
1651         }
1652
1653         return lpni;
1654 }
1655
1656 /*
1657  * Peer Discovery
1658  */
1659
1660 /*
1661  * Is a peer uptodate from the point of view of discovery?
1662  *
1663  * If it is currently being processed, obviously not.
1664  * A forced Ping or Push is also handled by the discovery thread.
1665  *
1666  * Otherwise look at whether the peer needs rediscovering.
1667  */
1668 bool
1669 lnet_peer_is_uptodate(struct lnet_peer *lp)
1670 {
1671         bool rc;
1672
1673         spin_lock(&lp->lp_lock);
1674         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1675                             LNET_PEER_FORCE_PING |
1676                             LNET_PEER_FORCE_PUSH)) {
1677                 rc = false;
1678         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
1679                 rc = true;
1680         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1681                 if (lnet_peer_discovery_disabled)
1682                         rc = true;
1683                 else
1684                         rc = false;
1685         } else if (lnet_peer_needs_push(lp)) {
1686                 rc = false;
1687         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1688                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1689                         rc = true;
1690                 else
1691                         rc = false;
1692         } else {
1693                 rc = false;
1694         }
1695         spin_unlock(&lp->lp_lock);
1696
1697         return rc;
1698 }
1699
1700 /*
1701  * Queue a peer for the attention of the discovery thread.  Call with
1702  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1703  * -EALREADY if the peer was already queued.
1704  */
1705 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1706 {
1707         int rc;
1708
1709         spin_lock(&lp->lp_lock);
1710         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1711                 lp->lp_state |= LNET_PEER_DISCOVERING;
1712         spin_unlock(&lp->lp_lock);
1713         if (list_empty(&lp->lp_dc_list)) {
1714                 lnet_peer_addref_locked(lp);
1715                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1716                 wake_up(&the_lnet.ln_dc_waitq);
1717                 rc = 0;
1718         } else {
1719                 rc = -EALREADY;
1720         }
1721
1722         CDEBUG(D_NET, "Queue peer %s: %d\n",
1723                libcfs_nid2str(lp->lp_primary_nid), rc);
1724
1725         return rc;
1726 }
1727
1728 /*
1729  * Discovery of a peer is complete. Wake all waiters on the peer.
1730  * Call with lnet_net_lock/EX held.
1731  */
1732 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1733 {
1734         struct lnet_msg *msg = NULL;
1735         int rc = 0;
1736         struct list_head pending_msgs;
1737
1738         INIT_LIST_HEAD(&pending_msgs);
1739
1740         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1741                libcfs_nid2str(lp->lp_primary_nid));
1742
1743         list_del_init(&lp->lp_dc_list);
1744         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1745         wake_up_all(&lp->lp_dc_waitq);
1746
1747         lnet_net_unlock(LNET_LOCK_EX);
1748
1749         /* iterate through all pending messages and send them again */
1750         list_for_each_entry(msg, &pending_msgs, msg_list) {
1751                 if (lp->lp_dc_error) {
1752                         lnet_finalize(msg, lp->lp_dc_error);
1753                         continue;
1754                 }
1755
1756                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1757                        lnet_msgtyp2str(msg->msg_type),
1758                        libcfs_id2str(msg->msg_target));
1759                 rc = lnet_send(msg->msg_src_nid_param, msg,
1760                                msg->msg_rtr_nid_param);
1761                 if (rc < 0) {
1762                         CNETERR("Error sending %s to %s: %d\n",
1763                                lnet_msgtyp2str(msg->msg_type),
1764                                libcfs_id2str(msg->msg_target), rc);
1765                         lnet_finalize(msg, rc);
1766                 }
1767         }
1768         lnet_net_lock(LNET_LOCK_EX);
1769         lnet_peer_decref_locked(lp);
1770 }
1771
1772 /*
1773  * Handle inbound push.
1774  * Like any event handler, called with lnet_res_lock/CPT held.
1775  */
1776 void lnet_peer_push_event(struct lnet_event *ev)
1777 {
1778         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1779         struct lnet_peer *lp;
1780
1781         /* lnet_find_peer() adds a refcount */
1782         lp = lnet_find_peer(ev->source.nid);
1783         if (!lp) {
1784                 CERROR("Push Put from unknown %s (source %s)\n",
1785                        libcfs_nid2str(ev->initiator.nid),
1786                        libcfs_nid2str(ev->source.nid));
1787                 return;
1788         }
1789
1790         /* Ensure peer state remains consistent while we modify it. */
1791         spin_lock(&lp->lp_lock);
1792
1793         /*
1794          * If some kind of error happened the contents of the message
1795          * cannot be used. Clear the NIDS_UPTODATE and set the
1796          * FORCE_PING flag to trigger a ping.
1797          */
1798         if (ev->status) {
1799                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1800                 lp->lp_state |= LNET_PEER_FORCE_PING;
1801                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1802                        ev->status,
1803                        libcfs_nid2str(lp->lp_primary_nid),
1804                        libcfs_nid2str(ev->source.nid));
1805                 goto out;
1806         }
1807
1808         /*
1809          * A push with invalid or corrupted info. Clear the UPTODATE
1810          * flag to trigger a ping.
1811          */
1812         if (lnet_ping_info_validate(&pbuf->pb_info)) {
1813                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1814                 lp->lp_state |= LNET_PEER_FORCE_PING;
1815                 CDEBUG(D_NET, "Corrupted Push from %s\n",
1816                        libcfs_nid2str(lp->lp_primary_nid));
1817                 goto out;
1818         }
1819
1820         /*
1821          * Make sure we'll allocate the correct size ping buffer when
1822          * pinging the peer.
1823          */
1824         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
1825                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
1826
1827         /*
1828          * A non-Multi-Rail peer is not supposed to be capable of
1829          * sending a push.
1830          */
1831         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
1832                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
1833                        libcfs_nid2str(lp->lp_primary_nid));
1834                 goto out;
1835         }
1836
1837         /*
1838          * Check the MULTIRAIL flag. Complain if the peer was DLC
1839          * configured without it.
1840          */
1841         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1842                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1843                         CERROR("Push says %s is Multi-Rail, DLC says not\n",
1844                                libcfs_nid2str(lp->lp_primary_nid));
1845                 } else {
1846                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1847                         lnet_peer_clr_non_mr_pref_nids(lp);
1848                 }
1849         }
1850
1851         /*
1852          * The peer may have discovery disabled at its end. Set
1853          * NO_DISCOVERY as appropriate.
1854          */
1855         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
1856                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
1857                        libcfs_nid2str(lp->lp_primary_nid));
1858                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
1859         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
1860                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
1861                        libcfs_nid2str(lp->lp_primary_nid));
1862                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
1863         }
1864
1865         /*
1866          * Check for truncation of the Put message. Clear the
1867          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
1868          * and tell discovery to allocate a bigger buffer.
1869          */
1870         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
1871                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
1872                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
1873                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1874                 lp->lp_state |= LNET_PEER_FORCE_PING;
1875                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
1876                        libcfs_nid2str(lp->lp_primary_nid),
1877                        pbuf->pb_info.pi_nnis);
1878                 goto out;
1879         }
1880
1881         /*
1882          * Check whether the Put data is stale. Stale data can just be
1883          * dropped.
1884          */
1885         if (pbuf->pb_info.pi_nnis > 1 &&
1886             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid &&
1887             LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno) {
1888                 CDEBUG(D_NET, "Stale Push from %s: got %u have %u\n",
1889                        libcfs_nid2str(lp->lp_primary_nid),
1890                        LNET_PING_BUFFER_SEQNO(pbuf),
1891                        lp->lp_peer_seqno);
1892                 goto out;
1893         }
1894
1895         /*
1896          * Check whether the Put data is new, in which case we clear
1897          * the UPTODATE flag and prepare to process it.
1898          *
1899          * If the Put data is current, and the peer is UPTODATE then
1900          * we assome everything is all right and drop the data as
1901          * stale.
1902          */
1903         if (LNET_PING_BUFFER_SEQNO(pbuf) > lp->lp_peer_seqno) {
1904                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
1905                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1906         } else if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
1907                 CDEBUG(D_NET, "Stale Push from %s: got %u have %u\n",
1908                        libcfs_nid2str(lp->lp_primary_nid),
1909                        LNET_PING_BUFFER_SEQNO(pbuf),
1910                        lp->lp_peer_seqno);
1911                 goto out;
1912         }
1913
1914         /*
1915          * If there is data present that hasn't been processed yet,
1916          * we'll replace it if the Put contained newer data and it
1917          * fits. We're racing with a Ping or earlier Push in this
1918          * case.
1919          */
1920         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
1921                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
1922                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
1923                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
1924                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
1925                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
1926                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
1927                               libcfs_nid2str(lp->lp_primary_nid),
1928                               LNET_PING_BUFFER_SEQNO(pbuf),
1929                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
1930                 }
1931                 goto out;
1932         }
1933
1934         /*
1935          * Allocate a buffer to copy the data. On a failure we drop
1936          * the Push and set FORCE_PING to force the discovery
1937          * thread to fix the problem by pinging the peer.
1938          */
1939         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
1940         if (!lp->lp_data) {
1941                 lp->lp_state |= LNET_PEER_FORCE_PING;
1942                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
1943                        libcfs_nid2str(lp->lp_primary_nid),
1944                        LNET_PING_BUFFER_SEQNO(pbuf));
1945                 goto out;
1946         }
1947
1948         /* Success */
1949         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
1950                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
1951         lp->lp_state |= LNET_PEER_DATA_PRESENT;
1952         CDEBUG(D_NET, "Received Push %s %u\n",
1953                libcfs_nid2str(lp->lp_primary_nid),
1954                LNET_PING_BUFFER_SEQNO(pbuf));
1955
1956 out:
1957         /*
1958          * Queue the peer for discovery, and wake the discovery thread
1959          * if the peer was already queued, because its status changed.
1960          */
1961         spin_unlock(&lp->lp_lock);
1962         lnet_net_lock(LNET_LOCK_EX);
1963         if (lnet_peer_queue_for_discovery(lp))
1964                 wake_up(&the_lnet.ln_dc_waitq);
1965         /* Drop refcount from lookup */
1966         lnet_peer_decref_locked(lp);
1967         lnet_net_unlock(LNET_LOCK_EX);
1968 }
1969
1970 /*
1971  * Clear the discovery error state, unless we're already discovering
1972  * this peer, in which case the error is current.
1973  */
1974 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
1975 {
1976         spin_lock(&lp->lp_lock);
1977         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1978                 lp->lp_dc_error = 0;
1979         spin_unlock(&lp->lp_lock);
1980 }
1981
1982 /*
1983  * Peer discovery slow path. The ln_api_mutex is held on entry, and
1984  * dropped/retaken within this function. An lnet_peer_ni is passed in
1985  * because discovery could tear down an lnet_peer.
1986  */
1987 int
1988 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
1989 {
1990         DEFINE_WAIT(wait);
1991         struct lnet_peer *lp;
1992         int rc = 0;
1993
1994 again:
1995         lnet_net_unlock(cpt);
1996         lnet_net_lock(LNET_LOCK_EX);
1997         lp = lpni->lpni_peer_net->lpn_peer;
1998         lnet_peer_clear_discovery_error(lp);
1999
2000         /*
2001          * We're willing to be interrupted. The lpni can become a
2002          * zombie if we race with DLC, so we must check for that.
2003          */
2004         for (;;) {
2005                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2006                 if (signal_pending(current))
2007                         break;
2008                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2009                         break;
2010                 if (lp->lp_dc_error)
2011                         break;
2012                 if (lnet_peer_is_uptodate(lp))
2013                         break;
2014                 lnet_peer_queue_for_discovery(lp);
2015                 /*
2016                  * if caller requested a non-blocking operation then
2017                  * return immediately. Once discovery is complete then the
2018                  * peer ref will be decremented and any pending messages
2019                  * that were stopped due to discovery will be transmitted.
2020                  */
2021                 if (!block)
2022                         break;
2023
2024                 lnet_peer_addref_locked(lp);
2025                 lnet_net_unlock(LNET_LOCK_EX);
2026                 schedule();
2027                 finish_wait(&lp->lp_dc_waitq, &wait);
2028                 lnet_net_lock(LNET_LOCK_EX);
2029                 lnet_peer_decref_locked(lp);
2030                 /* Peer may have changed */
2031                 lp = lpni->lpni_peer_net->lpn_peer;
2032         }
2033         finish_wait(&lp->lp_dc_waitq, &wait);
2034
2035         lnet_net_unlock(LNET_LOCK_EX);
2036         lnet_net_lock(cpt);
2037
2038         /*
2039          * If the peer has changed after we've discovered the older peer,
2040          * then we need to discovery the new peer to make sure the
2041          * interface information is up to date
2042          */
2043         if (lp != lpni->lpni_peer_net->lpn_peer)
2044                 goto again;
2045
2046         if (signal_pending(current))
2047                 rc = -EINTR;
2048         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2049                 rc = -ESHUTDOWN;
2050         else if (lp->lp_dc_error)
2051                 rc = lp->lp_dc_error;
2052         else if (!block)
2053                 CDEBUG(D_NET, "non-blocking discovery\n");
2054         else if (!lnet_peer_is_uptodate(lp))
2055                 goto again;
2056
2057         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2058                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2059                libcfs_nid2str(lpni->lpni_nid), rc,
2060                (!block) ? "pending discovery" : "discovery complete");
2061
2062         return rc;
2063 }
2064
2065 /* Handle an incoming ack for a push. */
2066 static void
2067 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2068 {
2069         struct lnet_ping_buffer *pbuf;
2070
2071         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2072         spin_lock(&lp->lp_lock);
2073         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2074         lp->lp_push_error = ev->status;
2075         if (ev->status)
2076                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2077         else
2078                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2079         spin_unlock(&lp->lp_lock);
2080
2081         CDEBUG(D_NET, "peer %s ev->status %d\n",
2082                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2083 }
2084
2085 /* Handle a Reply message. This is the reply to a Ping message. */
2086 static void
2087 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2088 {
2089         struct lnet_ping_buffer *pbuf;
2090         int rc;
2091
2092         spin_lock(&lp->lp_lock);
2093
2094         /*
2095          * If some kind of error happened the contents of message
2096          * cannot be used. Set PING_FAILED to trigger a retry.
2097          */
2098         if (ev->status) {
2099                 lp->lp_state |= LNET_PEER_PING_FAILED;
2100                 lp->lp_ping_error = ev->status;
2101                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2102                        ev->status,
2103                        libcfs_nid2str(lp->lp_primary_nid),
2104                        libcfs_nid2str(ev->source.nid));
2105                 goto out;
2106         }
2107
2108         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2109         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2110                 lnet_swap_pinginfo(pbuf);
2111
2112         /*
2113          * A reply with invalid or corrupted info. Set PING_FAILED to
2114          * trigger a retry.
2115          */
2116         rc = lnet_ping_info_validate(&pbuf->pb_info);
2117         if (rc) {
2118                 lp->lp_state |= LNET_PEER_PING_FAILED;
2119                 lp->lp_ping_error = 0;
2120                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2121                        libcfs_nid2str(lp->lp_primary_nid), rc);
2122                 goto out;
2123         }
2124
2125         /*
2126          * Update the MULTI_RAIL flag based on the reply. If the peer
2127          * was configured with DLC then the setting should match what
2128          * DLC put in.
2129          */
2130         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2131                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2132                         /* Everything's fine */
2133                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2134                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2135                               libcfs_nid2str(lp->lp_primary_nid));
2136                 } else {
2137                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2138                         lnet_peer_clr_non_mr_pref_nids(lp);
2139                 }
2140         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2141                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2142                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2143                               libcfs_nid2str(lp->lp_primary_nid));
2144                 } else {
2145                         CERROR("Multi-Rail state vanished from %s\n",
2146                                libcfs_nid2str(lp->lp_primary_nid));
2147                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2148                 }
2149         }
2150
2151         /*
2152          * Make sure we'll allocate the correct size ping buffer when
2153          * pinging the peer.
2154          */
2155         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2156                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2157
2158         /*
2159          * The peer may have discovery disabled at its end. Set
2160          * NO_DISCOVERY as appropriate.
2161          */
2162         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2163                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2164                        libcfs_nid2str(lp->lp_primary_nid));
2165                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2166         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2167                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2168                        libcfs_nid2str(lp->lp_primary_nid));
2169                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2170         }
2171
2172         /*
2173          * Check for truncation of the Reply. Clear PING_SENT and set
2174          * PING_FAILED to trigger a retry.
2175          */
2176         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2177                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2178                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2179                 lp->lp_state |= LNET_PEER_PING_FAILED;
2180                 lp->lp_ping_error = 0;
2181                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2182                        libcfs_nid2str(lp->lp_primary_nid),
2183                        pbuf->pb_info.pi_nnis);
2184                 goto out;
2185         }
2186
2187         /*
2188          * Check the sequence numbers in the reply. These are only
2189          * available if the reply came from a Multi-Rail peer.
2190          */
2191         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2192             pbuf->pb_info.pi_nnis > 1 &&
2193             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2194                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno) {
2195                         CDEBUG(D_NET, "Stale Reply from %s: got %u have %u\n",
2196                                 libcfs_nid2str(lp->lp_primary_nid),
2197                                 LNET_PING_BUFFER_SEQNO(pbuf),
2198                                 lp->lp_peer_seqno);
2199                         goto out;
2200                 }
2201
2202                 if (LNET_PING_BUFFER_SEQNO(pbuf) > lp->lp_peer_seqno)
2203                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2204         }
2205
2206         /* We're happy with the state of the data in the buffer. */
2207         CDEBUG(D_NET, "peer %s data present %u\n",
2208                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno);
2209         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2210                 lnet_ping_buffer_decref(lp->lp_data);
2211         else
2212                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2213         lnet_ping_buffer_addref(pbuf);
2214         lp->lp_data = pbuf;
2215 out:
2216         lp->lp_state &= ~LNET_PEER_PING_SENT;
2217         spin_unlock(&lp->lp_lock);
2218 }
2219
2220 /*
2221  * Send event handling. Only matters for error cases, where we clean
2222  * up state on the peer and peer_ni that would otherwise be updated in
2223  * the REPLY event handler for a successful Ping, and the ACK event
2224  * handler for a successful Push.
2225  */
2226 static int
2227 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2228 {
2229         int rc = 0;
2230
2231         if (!ev->status)
2232                 goto out;
2233
2234         spin_lock(&lp->lp_lock);
2235         if (ev->msg_type == LNET_MSG_GET) {
2236                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2237                 lp->lp_state |= LNET_PEER_PING_FAILED;
2238                 lp->lp_ping_error = ev->status;
2239         } else { /* ev->msg_type == LNET_MSG_PUT */
2240                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2241                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2242                 lp->lp_push_error = ev->status;
2243         }
2244         spin_unlock(&lp->lp_lock);
2245         rc = LNET_REDISCOVER_PEER;
2246 out:
2247         CDEBUG(D_NET, "%s Send to %s: %d\n",
2248                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2249                 libcfs_nid2str(ev->target.nid), rc);
2250         return rc;
2251 }
2252
2253 /*
2254  * Unlink event handling. This event is only seen if a call to
2255  * LNetMDUnlink() caused the event to be unlinked. If this call was
2256  * made after the event was set up in LNetGet() or LNetPut() then we
2257  * assume the Ping or Push timed out.
2258  */
2259 static void
2260 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2261 {
2262         spin_lock(&lp->lp_lock);
2263         /* We've passed through LNetGet() */
2264         if (lp->lp_state & LNET_PEER_PING_SENT) {
2265                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2266                 lp->lp_state |= LNET_PEER_PING_FAILED;
2267                 lp->lp_ping_error = -ETIMEDOUT;
2268                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2269                         libcfs_nid2str(lp->lp_primary_nid));
2270         }
2271         /* We've passed through LNetPut() */
2272         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2273                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2274                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2275                 lp->lp_push_error = -ETIMEDOUT;
2276                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2277                         libcfs_nid2str(lp->lp_primary_nid));
2278         }
2279         spin_unlock(&lp->lp_lock);
2280 }
2281
2282 /*
2283  * Event handler for the discovery EQ.
2284  *
2285  * Called with lnet_res_lock(cpt) held. The cpt is the
2286  * lnet_cpt_of_cookie() of the md handle cookie.
2287  */
2288 static void lnet_discovery_event_handler(lnet_event_t *event)
2289 {
2290         struct lnet_peer *lp = event->md.user_ptr;
2291         struct lnet_ping_buffer *pbuf;
2292         int rc;
2293
2294         /* discovery needs to take another look */
2295         rc = LNET_REDISCOVER_PEER;
2296
2297         CDEBUG(D_NET, "Received event: %d\n", event->type);
2298
2299         switch (event->type) {
2300         case LNET_EVENT_ACK:
2301                 lnet_discovery_event_ack(lp, event);
2302                 break;
2303         case LNET_EVENT_REPLY:
2304                 lnet_discovery_event_reply(lp, event);
2305                 break;
2306         case LNET_EVENT_SEND:
2307                 /* Only send failure triggers a retry. */
2308                 rc = lnet_discovery_event_send(lp, event);
2309                 break;
2310         case LNET_EVENT_UNLINK:
2311                 /* LNetMDUnlink() was called */
2312                 lnet_discovery_event_unlink(lp, event);
2313                 break;
2314         default:
2315                 /* Invalid events. */
2316                 LBUG();
2317         }
2318         lnet_net_lock(LNET_LOCK_EX);
2319         if (event->unlinked) {
2320                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2321                 lnet_ping_buffer_decref(pbuf);
2322                 lnet_peer_decref_locked(lp);
2323         }
2324         if (rc == LNET_REDISCOVER_PEER) {
2325                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2326                 wake_up(&the_lnet.ln_dc_waitq);
2327         }
2328         lnet_net_unlock(LNET_LOCK_EX);
2329 }
2330
2331 /*
2332  * Build a peer from incoming data.
2333  *
2334  * The NIDs in the incoming data are supposed to be structured as follows:
2335  *  - loopback
2336  *  - primary NID
2337  *  - other NIDs in same net
2338  *  - NIDs in second net
2339  *  - NIDs in third net
2340  *  - ...
2341  * This due to the way the list of NIDs in the data is created.
2342  *
2343  * Note that this function will mark the peer uptodate unless an
2344  * ENOMEM is encontered. All other errors are due to a conflict
2345  * between the DLC configuration and what discovery sees. We treat DLC
2346  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2347  * peer from becoming stuck in discovery.
2348  */
2349 static int lnet_peer_merge_data(struct lnet_peer *lp,
2350                                 struct lnet_ping_buffer *pbuf)
2351 {
2352         struct lnet_peer_ni *lpni;
2353         lnet_nid_t *curnis = NULL;
2354         lnet_nid_t *addnis = NULL;
2355         lnet_nid_t *delnis = NULL;
2356         unsigned flags;
2357         int ncurnis;
2358         int naddnis;
2359         int ndelnis;
2360         int nnis = 0;
2361         int i;
2362         int j;
2363         int rc;
2364
2365         flags = LNET_PEER_DISCOVERED;
2366         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2367                 flags |= LNET_PEER_MULTI_RAIL;
2368
2369         nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2370         LIBCFS_ALLOC(curnis, nnis * sizeof(lnet_nid_t));
2371         LIBCFS_ALLOC(addnis, nnis * sizeof(lnet_nid_t));
2372         LIBCFS_ALLOC(delnis, nnis * sizeof(lnet_nid_t));
2373         if (!curnis || !addnis || !delnis) {
2374                 rc = -ENOMEM;
2375                 goto out;
2376         }
2377         ncurnis = 0;
2378         naddnis = 0;
2379         ndelnis = 0;
2380
2381         /* Construct the list of NIDs present in peer. */
2382         lpni = NULL;
2383         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2384                 curnis[ncurnis++] = lpni->lpni_nid;
2385
2386         /*
2387          * Check for NIDs in pbuf not present in curnis[].
2388          * The loop starts at 1 to skip the loopback NID.
2389          */
2390         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2391                 for (j = 0; j < ncurnis; j++)
2392                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2393                                 break;
2394                 if (j == ncurnis)
2395                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i].ns_nid;
2396         }
2397         /*
2398          * Check for NIDs in curnis[] not present in pbuf.
2399          * The nested loop starts at 1 to skip the loopback NID.
2400          *
2401          * But never add the loopback NID to delnis[]: if it is
2402          * present in curnis[] then this peer is for this node.
2403          */
2404         for (i = 0; i < ncurnis; i++) {
2405                 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2406                         continue;
2407                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++)
2408                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid)
2409                                 break;
2410                 if (j == pbuf->pb_info.pi_nnis)
2411                         delnis[ndelnis++] = curnis[i];
2412         }
2413
2414         for (i = 0; i < naddnis; i++) {
2415                 rc = lnet_peer_add_nid(lp, addnis[i], flags);
2416                 if (rc) {
2417                         CERROR("Error adding NID %s to peer %s: %d\n",
2418                                libcfs_nid2str(addnis[i]),
2419                                libcfs_nid2str(lp->lp_primary_nid), rc);
2420                         if (rc == -ENOMEM)
2421                                 goto out;
2422                 }
2423         }
2424         for (i = 0; i < ndelnis; i++) {
2425                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2426                 if (rc) {
2427                         CERROR("Error deleting NID %s from peer %s: %d\n",
2428                                libcfs_nid2str(delnis[i]),
2429                                libcfs_nid2str(lp->lp_primary_nid), rc);
2430                         if (rc == -ENOMEM)
2431                                 goto out;
2432                 }
2433         }
2434         /*
2435          * Errors other than -ENOMEM are due to peers having been
2436          * configured with DLC. Ignore these because DLC overrides
2437          * Discovery.
2438          */
2439         rc = 0;
2440 out:
2441         LIBCFS_FREE(curnis, nnis * sizeof(lnet_nid_t));
2442         LIBCFS_FREE(addnis, nnis * sizeof(lnet_nid_t));
2443         LIBCFS_FREE(delnis, nnis * sizeof(lnet_nid_t));
2444         lnet_ping_buffer_decref(pbuf);
2445         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2446
2447         if (rc) {
2448                 spin_lock(&lp->lp_lock);
2449                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2450                 lp->lp_state |= LNET_PEER_FORCE_PING;
2451                 spin_unlock(&lp->lp_lock);
2452         }
2453         return rc;
2454 }
2455
2456 /*
2457  * The data in pbuf says lp is its primary peer, but the data was
2458  * received by a different peer. Try to update lp with the data.
2459  */
2460 static int
2461 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2462 {
2463         lnet_handle_md_t mdh;
2464
2465         /* Queue lp for discovery, and force it on the request queue. */
2466         lnet_net_lock(LNET_LOCK_EX);
2467         if (lnet_peer_queue_for_discovery(lp))
2468                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2469         lnet_net_unlock(LNET_LOCK_EX);
2470
2471         LNetInvalidateMDHandle(&mdh);
2472
2473         /*
2474          * Decide whether we can move the peer to the DATA_PRESENT state.
2475          *
2476          * We replace stale data for a multi-rail peer, repair PING_FAILED
2477          * status, and preempt FORCE_PING.
2478          *
2479          * If after that we have DATA_PRESENT, we merge it into this peer.
2480          */
2481         spin_lock(&lp->lp_lock);
2482         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2483                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2484                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2485                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2486                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2487                         lnet_ping_buffer_decref(pbuf);
2488                         pbuf = lp->lp_data;
2489                         lp->lp_data = NULL;
2490                 }
2491         }
2492         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2493                 lnet_ping_buffer_decref(lp->lp_data);
2494                 lp->lp_data = NULL;
2495                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2496         }
2497         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2498                 mdh = lp->lp_ping_mdh;
2499                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2500                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2501                 lp->lp_ping_error = 0;
2502         }
2503         if (lp->lp_state & LNET_PEER_FORCE_PING)
2504                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2505         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2506         spin_unlock(&lp->lp_lock);
2507
2508         if (!LNetMDHandleIsInvalid(mdh))
2509                 LNetMDUnlink(mdh);
2510
2511         if (pbuf)
2512                 return lnet_peer_merge_data(lp, pbuf);
2513
2514         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2515         return 0;
2516 }
2517
2518 /*
2519  * Update a peer using the data received.
2520  */
2521 static int lnet_peer_data_present(struct lnet_peer *lp)
2522 __must_hold(&lp->lp_lock)
2523 {
2524         struct lnet_ping_buffer *pbuf;
2525         struct lnet_peer_ni *lpni;
2526         lnet_nid_t nid = LNET_NID_ANY;
2527         unsigned flags;
2528         int rc = 0;
2529
2530         pbuf = lp->lp_data;
2531         lp->lp_data = NULL;
2532         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2533         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2534         spin_unlock(&lp->lp_lock);
2535
2536         /*
2537          * Modifications of peer structures are done while holding the
2538          * ln_api_mutex. A global lock is required because we may be
2539          * modifying multiple peer structures, and a mutex greatly
2540          * simplifies memory management.
2541          *
2542          * The actual changes to the data structures must also protect
2543          * against concurrent lookups, for which the lnet_net_lock in
2544          * LNET_LOCK_EX mode is used.
2545          */
2546         mutex_lock(&the_lnet.ln_api_mutex);
2547         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2548                 rc = -ESHUTDOWN;
2549                 goto out;
2550         }
2551
2552         /*
2553          * If this peer is not on the peer list then it is being torn
2554          * down, and our reference count may be all that is keeping it
2555          * alive. Don't do any work on it.
2556          */
2557         if (list_empty(&lp->lp_peer_list))
2558                 goto out;
2559
2560         flags = LNET_PEER_DISCOVERED;
2561         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2562                 flags |= LNET_PEER_MULTI_RAIL;
2563
2564         /*
2565          * Check whether the primary NID in the message matches the
2566          * primary NID of the peer. If it does, update the peer, if
2567          * it it does not, check whether there is already a peer with
2568          * that primary NID. If no such peer exists, try to update
2569          * the primary NID of the current peer (allowed if it was
2570          * created due to message traffic) and complete the update.
2571          * If the peer did exist, hand off the data to it.
2572          *
2573          * The peer for the loopback interface is a special case: this
2574          * is the peer for the local node, and we want to set its
2575          * primary NID to the correct value here.
2576          */
2577         if (pbuf->pb_info.pi_nnis > 1)
2578                 nid = pbuf->pb_info.pi_ni[1].ns_nid;
2579         if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2580                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2581                 if (!rc)
2582                         rc = lnet_peer_merge_data(lp, pbuf);
2583         } else if (lp->lp_primary_nid == nid) {
2584                 rc = lnet_peer_merge_data(lp, pbuf);
2585         } else {
2586                 lpni = lnet_find_peer_ni_locked(nid);
2587                 if (!lpni) {
2588                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2589                         if (rc) {
2590                                 CERROR("Primary NID error %s versus %s: %d\n",
2591                                        libcfs_nid2str(lp->lp_primary_nid),
2592                                        libcfs_nid2str(nid), rc);
2593                         } else {
2594                                 rc = lnet_peer_merge_data(lp, pbuf);
2595                         }
2596                 } else {
2597                         rc = lnet_peer_set_primary_data(
2598                                 lpni->lpni_peer_net->lpn_peer, pbuf);
2599                         lnet_peer_ni_decref_locked(lpni);
2600                 }
2601         }
2602 out:
2603         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2604         mutex_unlock(&the_lnet.ln_api_mutex);
2605
2606         spin_lock(&lp->lp_lock);
2607         /* Tell discovery to re-check the peer immediately. */
2608         if (!rc)
2609                 rc = LNET_REDISCOVER_PEER;
2610         return rc;
2611 }
2612
2613 /*
2614  * A ping failed. Clear the PING_FAILED state and set the
2615  * FORCE_PING state, to ensure a retry even if discovery is
2616  * disabled. This avoids being left with incorrect state.
2617  */
2618 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2619 __must_hold(&lp->lp_lock)
2620 {
2621         lnet_handle_md_t mdh;
2622         int rc;
2623
2624         mdh = lp->lp_ping_mdh;
2625         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2626         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2627         lp->lp_state |= LNET_PEER_FORCE_PING;
2628         rc = lp->lp_ping_error;
2629         lp->lp_ping_error = 0;
2630         spin_unlock(&lp->lp_lock);
2631
2632         if (!LNetMDHandleIsInvalid(mdh))
2633                 LNetMDUnlink(mdh);
2634
2635         CDEBUG(D_NET, "peer %s:%d\n",
2636                libcfs_nid2str(lp->lp_primary_nid), rc);
2637
2638         spin_lock(&lp->lp_lock);
2639         return rc ? rc : LNET_REDISCOVER_PEER;
2640 }
2641
2642 /*
2643  * Select NID to send a Ping or Push to.
2644  */
2645 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2646 {
2647         struct lnet_peer_ni *lpni;
2648
2649         /* Look for a direct-connected NID for this peer. */
2650         lpni = NULL;
2651         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2652                 if (!lnet_is_peer_ni_healthy_locked(lpni))
2653                         continue;
2654                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2655                         continue;
2656                 break;
2657         }
2658         if (lpni)
2659                 return lpni->lpni_nid;
2660
2661         /* Look for a routed-connected NID for this peer. */
2662         lpni = NULL;
2663         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2664                 if (!lnet_is_peer_ni_healthy_locked(lpni))
2665                         continue;
2666                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2667                         continue;
2668                 break;
2669         }
2670         if (lpni)
2671                 return lpni->lpni_nid;
2672
2673         return LNET_NID_ANY;
2674 }
2675
2676 /* Active side of ping. */
2677 static int lnet_peer_send_ping(struct lnet_peer *lp)
2678 __must_hold(&lp->lp_lock)
2679 {
2680         lnet_md_t md = { NULL };
2681         lnet_process_id_t id;
2682         struct lnet_ping_buffer *pbuf;
2683         int nnis;
2684         int rc;
2685         int cpt;
2686
2687         lp->lp_state |= LNET_PEER_PING_SENT;
2688         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2689         spin_unlock(&lp->lp_lock);
2690
2691         nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2692         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
2693         if (!pbuf) {
2694                 rc = -ENOMEM;
2695                 goto fail_error;
2696         }
2697
2698         /* initialize md content */
2699         md.start     = &pbuf->pb_info;
2700         md.length    = LNET_PING_INFO_SIZE(nnis);
2701         md.threshold = 2; /* GET/REPLY */
2702         md.max_size  = 0;
2703         md.options   = LNET_MD_TRUNCATE;
2704         md.user_ptr  = lp;
2705         md.eq_handle = the_lnet.ln_dc_eqh;
2706
2707         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_ping_mdh);
2708         if (rc != 0) {
2709                 lnet_ping_buffer_decref(pbuf);
2710                 CERROR("Can't bind MD: %d\n", rc);
2711                 goto fail_error;
2712         }
2713         cpt = lnet_net_lock_current();
2714         /* Refcount for MD. */
2715         lnet_peer_addref_locked(lp);
2716         id.pid = LNET_PID_LUSTRE;
2717         id.nid = lnet_peer_select_nid(lp);
2718         lnet_net_unlock(cpt);
2719
2720         if (id.nid == LNET_NID_ANY) {
2721                 rc = -EHOSTUNREACH;
2722                 goto fail_unlink_md;
2723         }
2724
2725         rc = LNetGet(LNET_NID_ANY, lp->lp_ping_mdh, id,
2726                      LNET_RESERVED_PORTAL,
2727                      LNET_PROTO_PING_MATCHBITS, 0);
2728
2729         if (rc)
2730                 goto fail_unlink_md;
2731
2732         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2733
2734         spin_lock(&lp->lp_lock);
2735         return 0;
2736
2737 fail_unlink_md:
2738         LNetMDUnlink(lp->lp_ping_mdh);
2739         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2740 fail_error:
2741         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2742         /*
2743          * The errors that get us here are considered hard errors and
2744          * cause Discovery to terminate. So we clear PING_SENT, but do
2745          * not set either PING_FAILED or FORCE_PING. In fact we need
2746          * to clear PING_FAILED, because the unlink event handler will
2747          * have set it if we called LNetMDUnlink() above.
2748          */
2749         spin_lock(&lp->lp_lock);
2750         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
2751         return rc;
2752 }
2753
2754 /*
2755  * This function exists because you cannot call LNetMDUnlink() from an
2756  * event handler.
2757  */
2758 static int lnet_peer_push_failed(struct lnet_peer *lp)
2759 __must_hold(&lp->lp_lock)
2760 {
2761         lnet_handle_md_t mdh;
2762         int rc;
2763
2764         mdh = lp->lp_push_mdh;
2765         LNetInvalidateMDHandle(&lp->lp_push_mdh);
2766         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
2767         rc = lp->lp_push_error;
2768         lp->lp_push_error = 0;
2769         spin_unlock(&lp->lp_lock);
2770
2771         if (!LNetMDHandleIsInvalid(mdh))
2772                 LNetMDUnlink(mdh);
2773
2774         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2775         spin_lock(&lp->lp_lock);
2776         return rc ? rc : LNET_REDISCOVER_PEER;
2777 }
2778
2779 /* Active side of push. */
2780 static int lnet_peer_send_push(struct lnet_peer *lp)
2781 __must_hold(&lp->lp_lock)
2782 {
2783         struct lnet_ping_buffer *pbuf;
2784         lnet_process_id_t id;
2785         lnet_md_t md;
2786         int cpt;
2787         int rc;
2788
2789         /* Don't push to a non-multi-rail peer. */
2790         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2791                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
2792                 return 0;
2793         }
2794
2795         lp->lp_state |= LNET_PEER_PUSH_SENT;
2796         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
2797         spin_unlock(&lp->lp_lock);
2798
2799         cpt = lnet_net_lock_current();
2800         pbuf = the_lnet.ln_ping_target;
2801         lnet_ping_buffer_addref(pbuf);
2802         lnet_net_unlock(cpt);
2803
2804         /* Push source MD */
2805         md.start     = &pbuf->pb_info;
2806         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
2807         md.threshold = 2; /* Put/Ack */
2808         md.max_size  = 0;
2809         md.options   = 0;
2810         md.eq_handle = the_lnet.ln_dc_eqh;
2811         md.user_ptr  = lp;
2812
2813         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
2814         if (rc) {
2815                 lnet_ping_buffer_decref(pbuf);
2816                 CERROR("Can't bind push source MD: %d\n", rc);
2817                 goto fail_error;
2818         }
2819         cpt = lnet_net_lock_current();
2820         /* Refcount for MD. */
2821         lnet_peer_addref_locked(lp);
2822         id.pid = LNET_PID_LUSTRE;
2823         id.nid = lnet_peer_select_nid(lp);
2824         lnet_net_unlock(cpt);
2825
2826         if (id.nid == LNET_NID_ANY) {
2827                 rc = -EHOSTUNREACH;
2828                 goto fail_unlink;
2829         }
2830
2831         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
2832                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
2833                      LNET_PROTO_PING_MATCHBITS, 0, 0);
2834
2835         if (rc)
2836                 goto fail_unlink;
2837
2838         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2839
2840         spin_lock(&lp->lp_lock);
2841         return 0;
2842
2843 fail_unlink:
2844         LNetMDUnlink(lp->lp_push_mdh);
2845         LNetInvalidateMDHandle(&lp->lp_push_mdh);
2846 fail_error:
2847         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2848         /*
2849          * The errors that get us here are considered hard errors and
2850          * cause Discovery to terminate. So we clear PUSH_SENT, but do
2851          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
2852          * because the unlink event handler will have set it if we
2853          * called LNetMDUnlink() above.
2854          */
2855         spin_lock(&lp->lp_lock);
2856         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
2857         return rc;
2858 }
2859
2860 /*
2861  * An unrecoverable error was encountered during discovery.
2862  * Set error status in peer and abort discovery.
2863  */
2864 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
2865 {
2866         CDEBUG(D_NET, "Discovery error %s: %d\n",
2867                libcfs_nid2str(lp->lp_primary_nid), error);
2868
2869         spin_lock(&lp->lp_lock);
2870         lp->lp_dc_error = error;
2871         lp->lp_state &= ~LNET_PEER_DISCOVERING;
2872         lp->lp_state |= LNET_PEER_REDISCOVER;
2873         spin_unlock(&lp->lp_lock);
2874 }
2875
2876 /*
2877  * Mark the peer as discovered.
2878  */
2879 static int lnet_peer_discovered(struct lnet_peer *lp)
2880 __must_hold(&lp->lp_lock)
2881 {
2882         lp->lp_state |= LNET_PEER_DISCOVERED;
2883         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
2884                           LNET_PEER_REDISCOVER);
2885
2886         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2887
2888         return 0;
2889 }
2890
2891 /*
2892  * Mark the peer as to be rediscovered.
2893  */
2894 static int lnet_peer_rediscover(struct lnet_peer *lp)
2895 __must_hold(&lp->lp_lock)
2896 {
2897         lp->lp_state |= LNET_PEER_REDISCOVER;
2898         lp->lp_state &= ~LNET_PEER_DISCOVERING;
2899
2900         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2901
2902         return 0;
2903 }
2904
2905 /*
2906  * Returns the first peer on the ln_dc_working queue if its timeout
2907  * has expired. Takes the current time as an argument so as to not
2908  * obsessively re-check the clock. The oldest discovery request will
2909  * be at the head of the queue.
2910  */
2911 static struct lnet_peer *lnet_peer_dc_timed_out(time64_t now)
2912 {
2913         struct lnet_peer *lp;
2914
2915         if (list_empty(&the_lnet.ln_dc_working))
2916                 return NULL;
2917         lp = list_first_entry(&the_lnet.ln_dc_working,
2918                               struct lnet_peer, lp_dc_list);
2919         if (now < lp->lp_last_queued + DEFAULT_PEER_TIMEOUT)
2920                 return NULL;
2921         return lp;
2922 }
2923
2924 /*
2925  * Discovering this peer is taking too long. Cancel any Ping or Push
2926  * that discovery is waiting on by unlinking the relevant MDs. The
2927  * lnet_discovery_event_handler() will proceed from here and complete
2928  * the cleanup.
2929  */
2930 static void lnet_peer_discovery_timeout(struct lnet_peer *lp)
2931 {
2932         struct lnet_handle_md ping_mdh;
2933         struct lnet_handle_md push_mdh;
2934
2935         LNetInvalidateMDHandle(&ping_mdh);
2936         LNetInvalidateMDHandle(&push_mdh);
2937
2938         spin_lock(&lp->lp_lock);
2939         if (lp->lp_state & LNET_PEER_PING_SENT) {
2940                 ping_mdh = lp->lp_ping_mdh;
2941                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2942         }
2943         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2944                 push_mdh = lp->lp_push_mdh;
2945                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
2946         }
2947         spin_unlock(&lp->lp_lock);
2948
2949         if (!LNetMDHandleIsInvalid(ping_mdh))
2950                 LNetMDUnlink(ping_mdh);
2951         if (!LNetMDHandleIsInvalid(push_mdh))
2952                 LNetMDUnlink(push_mdh);
2953 }
2954
2955 /*
2956  * Wait for work to be queued or some other change that must be
2957  * attended to. Returns non-zero if the discovery thread should shut
2958  * down.
2959  */
2960 static int lnet_peer_discovery_wait_for_work(void)
2961 {
2962         int cpt;
2963         int rc = 0;
2964
2965         DEFINE_WAIT(wait);
2966
2967         cpt = lnet_net_lock_current();
2968         for (;;) {
2969                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
2970                                 TASK_INTERRUPTIBLE);
2971                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
2972                         break;
2973                 if (lnet_push_target_resize_needed())
2974                         break;
2975                 if (!list_empty(&the_lnet.ln_dc_request))
2976                         break;
2977                 if (lnet_peer_dc_timed_out(ktime_get_real_seconds()))
2978                         break;
2979                 lnet_net_unlock(cpt);
2980
2981                 /*
2982                  * wakeup max every second to check if there are peers that
2983                  * have been stuck on the working queue for greater than
2984                  * the peer timeout.
2985                  */
2986                 schedule_timeout(cfs_time_seconds(1));
2987                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
2988                 cpt = lnet_net_lock_current();
2989         }
2990         finish_wait(&the_lnet.ln_dc_waitq, &wait);
2991
2992         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
2993                 rc = -ESHUTDOWN;
2994
2995         lnet_net_unlock(cpt);
2996
2997         CDEBUG(D_NET, "woken: %d\n", rc);
2998
2999         return rc;
3000 }
3001
3002 /* The discovery thread. */
3003 static int lnet_peer_discovery(void *arg)
3004 {
3005         struct lnet_peer *lp;
3006         time64_t now;
3007         int rc;
3008
3009         CDEBUG(D_NET, "started\n");
3010         cfs_block_allsigs();
3011
3012         for (;;) {
3013                 if (lnet_peer_discovery_wait_for_work())
3014                         break;
3015
3016                 if (lnet_push_target_resize_needed())
3017                         lnet_push_target_resize();
3018
3019                 lnet_net_lock(LNET_LOCK_EX);
3020                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3021                         break;
3022
3023                 /*
3024                  * Process all incoming discovery work requests.  When
3025                  * discovery must wait on a peer to change state, it
3026                  * is added to the tail of the ln_dc_working queue. A
3027                  * timestamp keeps track of when the peer was added,
3028                  * so we can time out discovery requests that take too
3029                  * long.
3030                  */
3031                 while (!list_empty(&the_lnet.ln_dc_request)) {
3032                         lp = list_first_entry(&the_lnet.ln_dc_request,
3033                                               struct lnet_peer, lp_dc_list);
3034                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3035                         /*
3036                          * set the time the peer was put on the dc_working
3037                          * queue. It shouldn't remain on the queue
3038                          * forever, in case the GET message (for ping)
3039                          * doesn't get a REPLY or the PUT message (for
3040                          * push) doesn't get an ACK.
3041                          *
3042                          * TODO: LNet Health will deal with this scenario
3043                          * in a generic way.
3044                          */
3045                         lp->lp_last_queued = ktime_get_real_seconds();
3046                         lnet_net_unlock(LNET_LOCK_EX);
3047
3048                         /*
3049                          * Select an action depending on the state of
3050                          * the peer and whether discovery is disabled.
3051                          * The check whether discovery is disabled is
3052                          * done after the code that handles processing
3053                          * for arrived data, cleanup for failures, and
3054                          * forcing a Ping or Push.
3055                          */
3056                         spin_lock(&lp->lp_lock);
3057                         CDEBUG(D_NET, "peer %s state %#x\n",
3058                                 libcfs_nid2str(lp->lp_primary_nid),
3059                                 lp->lp_state);
3060                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3061                                 rc = lnet_peer_data_present(lp);
3062                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3063                                 rc = lnet_peer_ping_failed(lp);
3064                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3065                                 rc = lnet_peer_push_failed(lp);
3066                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3067                                 rc = lnet_peer_send_ping(lp);
3068                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3069                                 rc = lnet_peer_send_push(lp);
3070                         else if (lnet_peer_discovery_disabled)
3071                                 rc = lnet_peer_rediscover(lp);
3072                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3073                                 rc = lnet_peer_send_ping(lp);
3074                         else if (lnet_peer_needs_push(lp))
3075                                 rc = lnet_peer_send_push(lp);
3076                         else
3077                                 rc = lnet_peer_discovered(lp);
3078                         CDEBUG(D_NET, "peer %s state %#x rc %d\n",
3079                                 libcfs_nid2str(lp->lp_primary_nid),
3080                                 lp->lp_state, rc);
3081                         spin_unlock(&lp->lp_lock);
3082
3083                         lnet_net_lock(LNET_LOCK_EX);
3084                         if (rc == LNET_REDISCOVER_PEER) {
3085                                 list_move(&lp->lp_dc_list,
3086                                           &the_lnet.ln_dc_request);
3087                         } else if (rc) {
3088                                 lnet_peer_discovery_error(lp, rc);
3089                         }
3090                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3091                                 lnet_peer_discovery_complete(lp);
3092                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3093                                 break;
3094                 }
3095
3096                 /*
3097                  * Now that the ln_dc_request queue has been emptied
3098                  * check the ln_dc_working queue for peers that are
3099                  * taking too long. Move all that are found to the
3100                  * ln_dc_expired queue and time out any pending
3101                  * Ping or Push. We have to drop the lnet_net_lock
3102                  * in the loop because lnet_peer_discovery_timeout()
3103                  * calls LNetMDUnlink().
3104                  */
3105                 now = ktime_get_real_seconds();
3106                 while ((lp = lnet_peer_dc_timed_out(now)) != NULL) {
3107                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3108                         lnet_net_unlock(LNET_LOCK_EX);
3109                         lnet_peer_discovery_timeout(lp);
3110                         lnet_net_lock(LNET_LOCK_EX);
3111                 }
3112
3113                 lnet_net_unlock(LNET_LOCK_EX);
3114         }
3115
3116         CDEBUG(D_NET, "stopping\n");
3117         /*
3118          * Clean up before telling lnet_peer_discovery_stop() that
3119          * we're done. Use wake_up() below to somewhat reduce the
3120          * size of the thundering herd if there are multiple threads
3121          * waiting on discovery of a single peer.
3122          */
3123         LNetEQFree(the_lnet.ln_dc_eqh);
3124         LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3125
3126         /* Queue cleanup 1: stop all pending pings and pushes. */
3127         lnet_net_lock(LNET_LOCK_EX);
3128         while (!list_empty(&the_lnet.ln_dc_working)) {
3129                 lp = list_first_entry(&the_lnet.ln_dc_working,
3130                                       struct lnet_peer, lp_dc_list);
3131                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3132                 lnet_net_unlock(LNET_LOCK_EX);
3133                 lnet_peer_discovery_timeout(lp);
3134                 lnet_net_lock(LNET_LOCK_EX);
3135         }
3136         lnet_net_unlock(LNET_LOCK_EX);
3137
3138         /* Queue cleanup 2: wait for the expired queue to clear. */
3139         while (!list_empty(&the_lnet.ln_dc_expired))
3140                 schedule_timeout(cfs_time_seconds(1));
3141
3142         /* Queue cleanup 3: clear the request queue. */
3143         lnet_net_lock(LNET_LOCK_EX);
3144         while (!list_empty(&the_lnet.ln_dc_request)) {
3145                 lp = list_first_entry(&the_lnet.ln_dc_request,
3146                                       struct lnet_peer, lp_dc_list);
3147                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3148                 lnet_peer_discovery_complete(lp);
3149         }
3150         lnet_net_unlock(LNET_LOCK_EX);
3151
3152         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3153         wake_up(&the_lnet.ln_dc_waitq);
3154
3155         CDEBUG(D_NET, "stopped\n");
3156
3157         return 0;
3158 }
3159
3160 /* ln_api_mutex is held on entry. */
3161 int lnet_peer_discovery_start(void)
3162 {
3163         struct task_struct *task;
3164         int rc;
3165
3166         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3167                 return -EALREADY;
3168
3169         rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3170         if (rc != 0) {
3171                 CERROR("Can't allocate discovery EQ: %d\n", rc);
3172                 return rc;
3173         }
3174
3175         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3176         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3177         if (IS_ERR(task)) {
3178                 rc = PTR_ERR(task);
3179                 CERROR("Can't start peer discovery thread: %d\n", rc);
3180
3181                 LNetEQFree(the_lnet.ln_dc_eqh);
3182                 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3183
3184                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3185         }
3186
3187         CDEBUG(D_NET, "discovery start: %d\n", rc);
3188
3189         return rc;
3190 }
3191
3192 /* ln_api_mutex is held on entry. */
3193 void lnet_peer_discovery_stop(void)
3194 {
3195         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3196                 return;
3197
3198         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3199         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3200         wake_up(&the_lnet.ln_dc_waitq);
3201
3202         wait_event(the_lnet.ln_dc_waitq,
3203                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3204
3205         LASSERT(list_empty(&the_lnet.ln_dc_request));
3206         LASSERT(list_empty(&the_lnet.ln_dc_working));
3207         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3208
3209         CDEBUG(D_NET, "discovery stopped\n");
3210 }
3211
3212 /* Debugging */
3213
3214 void
3215 lnet_debug_peer(lnet_nid_t nid)
3216 {
3217         char                    *aliveness = "NA";
3218         struct lnet_peer_ni     *lp;
3219         int                     cpt;
3220
3221         cpt = lnet_cpt_of_nid(nid, NULL);
3222         lnet_net_lock(cpt);
3223
3224         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3225         if (IS_ERR(lp)) {
3226                 lnet_net_unlock(cpt);
3227                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3228                 return;
3229         }
3230
3231         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3232                 aliveness = lp->lpni_alive ? "up" : "down";
3233
3234         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3235                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3236                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3237                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3238                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3239
3240         lnet_peer_ni_decref_locked(lp);
3241
3242         lnet_net_unlock(cpt);
3243 }
3244
3245 /* Gathering information for userspace. */
3246
3247 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3248                           char aliveness[LNET_MAX_STR_LEN],
3249                           __u32 *cpt_iter, __u32 *refcount,
3250                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3251                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3252                           __u32 *peer_tx_qnob)
3253 {
3254         struct lnet_peer_table          *peer_table;
3255         struct lnet_peer_ni             *lp;
3256         int                             j;
3257         int                             lncpt;
3258         bool                            found = false;
3259
3260         /* get the number of CPTs */
3261         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3262
3263         /* if the cpt number to be examined is >= the number of cpts in
3264          * the system then indicate that there are no more cpts to examin
3265          */
3266         if (*cpt_iter >= lncpt)
3267                 return -ENOENT;
3268
3269         /* get the current table */
3270         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3271         /* if the ptable is NULL then there are no more cpts to examine */
3272         if (peer_table == NULL)
3273                 return -ENOENT;
3274
3275         lnet_net_lock(*cpt_iter);
3276
3277         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3278                 struct list_head *peers = &peer_table->pt_hash[j];
3279
3280                 list_for_each_entry(lp, peers, lpni_hashlist) {
3281                         if (peer_index-- > 0)
3282                                 continue;
3283
3284                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3285                         if (lnet_isrouter(lp) ||
3286                                 lnet_peer_aliveness_enabled(lp))
3287                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3288                                          lp->lpni_alive ? "up" : "down");
3289
3290                         *nid = lp->lpni_nid;
3291                         *refcount = atomic_read(&lp->lpni_refcount);
3292                         *ni_peer_tx_credits =
3293                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3294                         *peer_tx_credits = lp->lpni_txcredits;
3295                         *peer_rtr_credits = lp->lpni_rtrcredits;
3296                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3297                         *peer_tx_qnob = lp->lpni_txqnob;
3298
3299                         found = true;
3300                 }
3301
3302         }
3303         lnet_net_unlock(*cpt_iter);
3304
3305         *cpt_iter = lncpt;
3306
3307         return found ? 0 : -ENOENT;
3308 }
3309
3310 /* ln_api_mutex is held, which keeps the peer list stable */
3311 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3312 {
3313         struct lnet_ioctl_element_stats *lpni_stats;
3314         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3315         struct lnet_peer_ni_credit_info *lpni_info;
3316         struct lnet_peer_ni *lpni;
3317         struct lnet_peer *lp;
3318         lnet_nid_t nid;
3319         __u32 size;
3320         int rc;
3321
3322         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3323
3324         if (!lp) {
3325                 rc = -ENOENT;
3326                 goto out;
3327         }
3328
3329         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3330                 + sizeof(*lpni_msg_stats);
3331         size *= lp->lp_nnis;
3332         if (size > cfg->prcfg_size) {
3333                 cfg->prcfg_size = size;
3334                 rc = -E2BIG;
3335                 goto out_lp_decref;
3336         }
3337
3338         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3339         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3340         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3341         cfg->prcfg_count = lp->lp_nnis;
3342         cfg->prcfg_size = size;
3343         cfg->prcfg_state = lp->lp_state;
3344
3345         /* Allocate helper buffers. */
3346         rc = -ENOMEM;
3347         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3348         if (!lpni_info)
3349                 goto out_lp_decref;
3350         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3351         if (!lpni_stats)
3352                 goto out_free_info;
3353         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3354         if (!lpni_msg_stats)
3355                 goto out_free_stats;
3356
3357
3358         lpni = NULL;
3359         rc = -EFAULT;
3360         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3361                 nid = lpni->lpni_nid;
3362                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3363                         goto out_free_msg_stats;
3364                 bulk += sizeof(nid);
3365
3366                 memset(lpni_info, 0, sizeof(*lpni_info));
3367                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3368                 if (lnet_isrouter(lpni) ||
3369                         lnet_peer_aliveness_enabled(lpni))
3370                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3371                                 lpni->lpni_alive ? "up" : "down");
3372
3373                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3374                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3375                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3376                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3377                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3378                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3379                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3380                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3381                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3382                         goto out_free_msg_stats;
3383                 bulk += sizeof(*lpni_info);
3384
3385                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3386                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3387                                                             LNET_STATS_TYPE_SEND);
3388                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3389                                                             LNET_STATS_TYPE_RECV);
3390                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3391                                                             LNET_STATS_TYPE_DROP);
3392                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3393                         goto out_free_msg_stats;
3394                 bulk += sizeof(*lpni_stats);
3395                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3396                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3397                         goto out_free_msg_stats;
3398                 bulk += sizeof(*lpni_msg_stats);
3399         }
3400         rc = 0;
3401
3402 out_free_msg_stats:
3403         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3404 out_free_stats:
3405         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3406 out_free_info:
3407         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3408 out_lp_decref:
3409         lnet_peer_decref_locked(lp);
3410 out:
3411         return rc;
3412 }