lnet/lnet/peer.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * 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 <linux/sched.h>
  38 #ifdef HAVE_SCHED_HEADERS
  39 #include <linux/sched/signal.h>
  40 #endif
  41 #include <linux/uaccess.h>
  42
  43 #include <lnet/lib-lnet.h>
  44 #include <uapi/linux/lnet/lnet-dlc.h>
  45
  46 /* Value indicating that recovery needs to re-check a peer immediately. */
  47 #define LNET_REDISCOVER_PEER    (1)
  48
  49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
  50
  51 static void
  52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
  53 {
  54         if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
  55                 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
  56                 lnet_peer_ni_decref_locked(lpni);
  57         }
  58 }
  59
  60 void
  61 lnet_peer_net_added(struct lnet_net *net)
  62 {
  63         struct lnet_peer_ni *lpni, *tmp;
  64
  65         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
  66                                  lpni_on_remote_peer_ni_list) {
  67
  68                 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
  69                         lpni->lpni_net = net;
  70
  71                         spin_lock(&lpni->lpni_lock);
  72                         lpni->lpni_txcredits =
  73                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
  74                         lpni->lpni_mintxcredits = lpni->lpni_txcredits;
  75                         lpni->lpni_rtrcredits =
  76                                 lnet_peer_buffer_credits(lpni->lpni_net);
  77                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
  78                         spin_unlock(&lpni->lpni_lock);
  79
  80                         lnet_peer_remove_from_remote_list(lpni);
  81                 }
  82         }
  83 }
  84
  85 static void
  86 lnet_peer_tables_destroy(void)
  87 {
  88         struct lnet_peer_table  *ptable;
  89         struct list_head        *hash;
  90         int                     i;
  91         int                     j;
  92
  93         if (!the_lnet.ln_peer_tables)
  94                 return;
  95
  96         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  97                 hash = ptable->pt_hash;
  98                 if (!hash) /* not intialized */
  99                         break;
 100
 101                 LASSERT(list_empty(&ptable->pt_zombie_list));
 102
 103                 ptable->pt_hash = NULL;
 104                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 105                         LASSERT(list_empty(&hash[j]));
 106
 107                 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
 108         }
 109
 110         cfs_percpt_free(the_lnet.ln_peer_tables);
 111         the_lnet.ln_peer_tables = NULL;
 112 }
 113
 114 int
 115 lnet_peer_tables_create(void)
 116 {
 117         struct lnet_peer_table  *ptable;
 118         struct list_head        *hash;
 119         int                     i;
 120         int                     j;
 121
 122         the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
 123                                                    sizeof(*ptable));
 124         if (the_lnet.ln_peer_tables == NULL) {
 125                 CERROR("Failed to allocate cpu-partition peer tables\n");
 126                 return -ENOMEM;
 127         }
 128
 129         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 130                 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
 131                                  LNET_PEER_HASH_SIZE * sizeof(*hash));
 132                 if (hash == NULL) {
 133                         CERROR("Failed to create peer hash table\n");
 134                         lnet_peer_tables_destroy();
 135                         return -ENOMEM;
 136                 }
 137
 138                 spin_lock_init(&ptable->pt_zombie_lock);
 139                 INIT_LIST_HEAD(&ptable->pt_zombie_list);
 140
 141                 INIT_LIST_HEAD(&ptable->pt_peer_list);
 142
 143                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 144                         INIT_LIST_HEAD(&hash[j]);
 145                 ptable->pt_hash = hash; /* sign of initialization */
 146         }
 147
 148         return 0;
 149 }
 150
 151 static struct lnet_peer_ni *
 152 lnet_peer_ni_alloc(lnet_nid_t nid)
 153 {
 154         struct lnet_peer_ni *lpni;
 155         struct lnet_net *net;
 156         int cpt;
 157
 158         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 159
 160         LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
 161         if (!lpni)
 162                 return NULL;
 163
 164         INIT_LIST_HEAD(&lpni->lpni_txq);
 165         INIT_LIST_HEAD(&lpni->lpni_hashlist);
 166         INIT_LIST_HEAD(&lpni->lpni_peer_nis);
 167         INIT_LIST_HEAD(&lpni->lpni_recovery);
 168         INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
 169         INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
 170         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
 171         atomic_set(&lpni->lpni_refcount, 1);
 172         lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
 173
 174         spin_lock_init(&lpni->lpni_lock);
 175
 176         if (lnet_peers_start_down())
 177                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 178         else
 179                 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
 180         lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 181         lpni->lpni_nid = nid;
 182         lpni->lpni_cpt = cpt;
 183         atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
 184
 185         net = lnet_get_net_locked(LNET_NIDNET(nid));
 186         lpni->lpni_net = net;
 187         if (net) {
 188                 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
 189                 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
 190                 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
 191                 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 192         } else {
 193                 /*
 194                  * This peer_ni is not on a local network, so we
 195                  * cannot add the credits here. In case the net is
 196                  * added later, add the peer_ni to the remote peer ni
 197                  * list so it can be easily found and revisited.
 198                  */
 199                 /* FIXME: per-net implementation instead? */
 200                 lnet_peer_ni_addref_locked(lpni);
 201                 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
 202                               &the_lnet.ln_remote_peer_ni_list);
 203         }
 204
 205         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
 206
 207         return lpni;
 208 }
 209
 210 static struct lnet_peer_net *
 211 lnet_peer_net_alloc(__u32 net_id)
 212 {
 213         struct lnet_peer_net *lpn;
 214
 215         LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
 216         if (!lpn)
 217                 return NULL;
 218
 219         INIT_LIST_HEAD(&lpn->lpn_peer_nets);
 220         INIT_LIST_HEAD(&lpn->lpn_peer_nis);
 221         lpn->lpn_net_id = net_id;
 222         lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
 223
 224         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 225
 226         return lpn;
 227 }
 228
 229 void
 230 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
 231 {
 232         struct lnet_peer *lp;
 233
 234         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 235
 236         LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
 237         LASSERT(list_empty(&lpn->lpn_peer_nis));
 238         LASSERT(list_empty(&lpn->lpn_peer_nets));
 239         lp = lpn->lpn_peer;
 240         lpn->lpn_peer = NULL;
 241         LIBCFS_FREE(lpn, sizeof(*lpn));
 242
 243         lnet_peer_decref_locked(lp);
 244 }
 245
 246 static struct lnet_peer *
 247 lnet_peer_alloc(lnet_nid_t nid)
 248 {
 249         struct lnet_peer *lp;
 250
 251         LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
 252         if (!lp)
 253                 return NULL;
 254
 255         INIT_LIST_HEAD(&lp->lp_rtrq);
 256         INIT_LIST_HEAD(&lp->lp_routes);
 257         INIT_LIST_HEAD(&lp->lp_peer_list);
 258         INIT_LIST_HEAD(&lp->lp_peer_nets);
 259         INIT_LIST_HEAD(&lp->lp_dc_list);
 260         INIT_LIST_HEAD(&lp->lp_dc_pendq);
 261         INIT_LIST_HEAD(&lp->lp_rtr_list);
 262         init_waitqueue_head(&lp->lp_dc_waitq);
 263         spin_lock_init(&lp->lp_lock);
 264         lp->lp_primary_nid = nid;
 265         lp->lp_disc_src_nid = LNET_NID_ANY;
 266         if (lnet_peers_start_down())
 267                 lp->lp_alive = false;
 268         else
 269                 lp->lp_alive = true;
 270
 271         /*
 272          * all peers created on a router should have health on
 273          * if it's not already on.
 274          */
 275         if (the_lnet.ln_routing && !lnet_health_sensitivity)
 276                 lp->lp_health_sensitivity = 1;
 277
 278         /*
 279          * Turn off discovery for loopback peer. If you're creating a peer
 280          * for the loopback interface then that was initiated when we
 281          * attempted to send a message over the loopback. There is no need
 282          * to ever use a different interface when sending messages to
 283          * myself.
 284          */
 285         if (nid == LNET_NID_LO_0)
 286                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 287         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 288
 289         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 290
 291         return lp;
 292 }
 293
 294 void
 295 lnet_destroy_peer_locked(struct lnet_peer *lp)
 296 {
 297         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 298
 299         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 300         LASSERT(lp->lp_rtr_refcount == 0);
 301         LASSERT(list_empty(&lp->lp_peer_nets));
 302         LASSERT(list_empty(&lp->lp_peer_list));
 303         LASSERT(list_empty(&lp->lp_dc_list));
 304
 305         if (lp->lp_data)
 306                 lnet_ping_buffer_decref(lp->lp_data);
 307
 308         /*
 309          * if there are messages still on the pending queue, then make
 310          * sure to queue them on the ln_msg_resend list so they can be
 311          * resent at a later point if the discovery thread is still
 312          * running.
 313          * If the discovery thread has stopped, then the wakeup will be a
 314          * no-op, and it is expected the lnet_shutdown_lndnets() will
 315          * eventually be called, which will traverse this list and
 316          * finalize the messages on the list.
 317          * We can not resend them now because we're holding the cpt lock.
 318          * Releasing the lock can cause an inconsistent state
 319          */
 320         spin_lock(&the_lnet.ln_msg_resend_lock);
 321         spin_lock(&lp->lp_lock);
 322         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 323         spin_unlock(&lp->lp_lock);
 324         spin_unlock(&the_lnet.ln_msg_resend_lock);
 325         wake_up(&the_lnet.ln_dc_waitq);
 326
 327         LIBCFS_FREE(lp, sizeof(*lp));
 328 }
 329
 330 /*
 331  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 332  * that peer_net, detach the peer_net from the peer.
 333  *
 334  * Call with lnet_net_lock/EX held
 335  */
 336 static void
 337 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 338 {
 339         struct lnet_peer_table *ptable;
 340         struct lnet_peer_net *lpn;
 341         struct lnet_peer *lp;
 342
 343         /*
 344          * Belts and suspenders: gracefully handle teardown of a
 345          * partially connected peer_ni.
 346          */
 347         lpn = lpni->lpni_peer_net;
 348
 349         list_del_init(&lpni->lpni_peer_nis);
 350         /*
 351          * If there are no lpni's left, we detach lpn from
 352          * lp_peer_nets, so it cannot be found anymore.
 353          */
 354         if (list_empty(&lpn->lpn_peer_nis))
 355                 list_del_init(&lpn->lpn_peer_nets);
 356
 357         /* Update peer NID count. */
 358         lp = lpn->lpn_peer;
 359         lp->lp_nnis--;
 360
 361         /*
 362          * If there are no more peer nets, make the peer unfindable
 363          * via the peer_tables.
 364          *
 365          * Otherwise, if the peer is DISCOVERED, tell discovery to
 366          * take another look at it. This is a no-op if discovery for
 367          * this peer did the detaching.
 368          */
 369         if (list_empty(&lp->lp_peer_nets)) {
 370                 list_del_init(&lp->lp_peer_list);
 371                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 372                 ptable->pt_peers--;
 373         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 374                 /* Discovery isn't running, nothing to do here. */
 375         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 376                 lnet_peer_queue_for_discovery(lp);
 377                 wake_up(&the_lnet.ln_dc_waitq);
 378         }
 379         CDEBUG(D_NET, "peer %s NID %s\n",
 380                 libcfs_nid2str(lp->lp_primary_nid),
 381                 libcfs_nid2str(lpni->lpni_nid));
 382 }
 383
 384 /* called with lnet_net_lock LNET_LOCK_EX held */
 385 static int
 386 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 387 {
 388         struct lnet_peer_table *ptable = NULL;
 389
 390         /* don't remove a peer_ni if it's also a gateway */
 391         if (lnet_isrouter(lpni) && !force) {
 392                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 393                        libcfs_nid2str(lpni->lpni_nid));
 394                 return -EBUSY;
 395         }
 396
 397         lnet_peer_remove_from_remote_list(lpni);
 398
 399         /* remove peer ni from the hash list. */
 400         list_del_init(&lpni->lpni_hashlist);
 401
 402         /*
 403          * indicate the peer is being deleted so the monitor thread can
 404          * remove it from the recovery queue.
 405          */
 406         spin_lock(&lpni->lpni_lock);
 407         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 408         spin_unlock(&lpni->lpni_lock);
 409
 410         /* decrement the ref count on the peer table */
 411         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 412
 413         /*
 414          * The peer_ni can no longer be found with a lookup. But there
 415          * can be current users, so keep track of it on the zombie
 416          * list until the reference count has gone to zero.
 417          *
 418          * The last reference may be lost in a place where the
 419          * lnet_net_lock locks only a single cpt, and that cpt may not
 420          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 421          * has its own lock.
 422          */
 423         spin_lock(&ptable->pt_zombie_lock);
 424         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 425         ptable->pt_zombies++;
 426         spin_unlock(&ptable->pt_zombie_lock);
 427
 428         /* no need to keep this peer_ni on the hierarchy anymore */
 429         lnet_peer_detach_peer_ni_locked(lpni);
 430
 431         /* remove hashlist reference on peer_ni */
 432         lnet_peer_ni_decref_locked(lpni);
 433
 434         return 0;
 435 }
 436
 437 void lnet_peer_uninit(void)
 438 {
 439         struct lnet_peer_ni *lpni, *tmp;
 440
 441         lnet_net_lock(LNET_LOCK_EX);
 442
 443         /* remove all peer_nis from the remote peer and the hash list */
 444         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 445                                  lpni_on_remote_peer_ni_list)
 446                 lnet_peer_ni_del_locked(lpni, false);
 447
 448         lnet_peer_tables_destroy();
 449
 450         lnet_net_unlock(LNET_LOCK_EX);
 451 }
 452
 453 static int
 454 lnet_peer_del_locked(struct lnet_peer *peer)
 455 {
 456         struct lnet_peer_ni *lpni = NULL, *lpni2;
 457         int rc = 0, rc2 = 0;
 458
 459         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
 460
 461         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 462         while (lpni != NULL) {
 463                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 464                 rc = lnet_peer_ni_del_locked(lpni, false);
 465                 if (rc != 0)
 466                         rc2 = rc;
 467                 lpni = lpni2;
 468         }
 469
 470         return rc2;
 471 }
 472
 473 static int
 474 lnet_peer_del(struct lnet_peer *peer)
 475 {
 476         lnet_net_lock(LNET_LOCK_EX);
 477         lnet_peer_del_locked(peer);
 478         lnet_net_unlock(LNET_LOCK_EX);
 479
 480         return 0;
 481 }
 482
 483 /*
 484  * Delete a NID from a peer. Call with ln_api_mutex held.
 485  *
 486  * Error codes:
 487  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 488  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 489  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 490  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 491  */
 492 static int
 493 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
 494 {
 495         struct lnet_peer_ni *lpni;
 496         lnet_nid_t primary_nid = lp->lp_primary_nid;
 497         int rc = 0;
 498         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 499
 500         if (!(flags & LNET_PEER_CONFIGURED)) {
 501                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 502                         rc = -EPERM;
 503                         goto out;
 504                 }
 505         }
 506         lpni = lnet_find_peer_ni_locked(nid);
 507         if (!lpni) {
 508                 rc = -ENOENT;
 509                 goto out;
 510         }
 511         lnet_peer_ni_decref_locked(lpni);
 512         if (lp != lpni->lpni_peer_net->lpn_peer) {
 513                 rc = -ECHILD;
 514                 goto out;
 515         }
 516
 517         /*
 518          * This function only allows deletion of the primary NID if it
 519          * is the only NID.
 520          */
 521         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
 522                 rc = -EBUSY;
 523                 goto out;
 524         }
 525
 526         lnet_net_lock(LNET_LOCK_EX);
 527
 528         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
 529                 struct lnet_peer_ni *lpni2;
 530                 /* assign the next peer_ni to be the primary */
 531                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 532                 LASSERT(lpni2);
 533                 lp->lp_primary_nid = lpni2->lpni_nid;
 534         }
 535         rc = lnet_peer_ni_del_locked(lpni, force);
 536
 537         lnet_net_unlock(LNET_LOCK_EX);
 538
 539 out:
 540         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 541                libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
 542
 543         return rc;
 544 }
 545
 546 static void
 547 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 548                                struct lnet_peer_table *ptable)
 549 {
 550         int                      i;
 551         struct lnet_peer_ni     *next;
 552         struct lnet_peer_ni     *lpni;
 553         struct lnet_peer        *peer;
 554
 555         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 556                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 557                                          lpni_hashlist) {
 558                         if (net != NULL && net != lpni->lpni_net)
 559                                 continue;
 560
 561                         peer = lpni->lpni_peer_net->lpn_peer;
 562                         if (peer->lp_primary_nid != lpni->lpni_nid) {
 563                                 lnet_peer_ni_del_locked(lpni, false);
 564                                 continue;
 565                         }
 566                         /*
 567                          * Removing the primary NID implies removing
 568                          * the entire peer. Advance next beyond any
 569                          * peer_ni that belongs to the same peer.
 570                          */
 571                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 572                                                  lpni_hashlist) {
 573                                 if (next->lpni_peer_net->lpn_peer != peer)
 574                                         break;
 575                         }
 576                         lnet_peer_del_locked(peer);
 577                 }
 578         }
 579 }
 580
 581 static void
 582 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 583 {
 584         wait_var_event_warning(&ptable->pt_zombies,
 585                                ptable->pt_zombies == 0,
 586                                "Waiting for %d zombies on peer table\n",
 587                                ptable->pt_zombies);
 588 }
 589
 590 static void
 591 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 592                                 struct lnet_peer_table *ptable)
 593 {
 594         struct lnet_peer_ni     *lp;
 595         struct lnet_peer_ni     *tmp;
 596         lnet_nid_t              gw_nid;
 597         int                     i;
 598
 599         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 600                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 601                                          lpni_hashlist) {
 602                         if (net != lp->lpni_net)
 603                                 continue;
 604
 605                         if (!lnet_isrouter(lp))
 606                                 continue;
 607
 608                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 609
 610                         lnet_net_unlock(LNET_LOCK_EX);
 611                         lnet_del_route(LNET_NET_ANY, gw_nid);
 612                         lnet_net_lock(LNET_LOCK_EX);
 613                 }
 614         }
 615 }
 616
 617 void
 618 lnet_peer_tables_cleanup(struct lnet_net *net)
 619 {
 620         int i;
 621         struct lnet_peer_table *ptable;
 622
 623         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 624         /* If just deleting the peers for a NI, get rid of any routes these
 625          * peers are gateways for. */
 626         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 627                 lnet_net_lock(LNET_LOCK_EX);
 628                 lnet_peer_table_del_rtrs_locked(net, ptable);
 629                 lnet_net_unlock(LNET_LOCK_EX);
 630         }
 631
 632         /* Start the cleanup process */
 633         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 634                 lnet_net_lock(LNET_LOCK_EX);
 635                 lnet_peer_table_cleanup_locked(net, ptable);
 636                 lnet_net_unlock(LNET_LOCK_EX);
 637         }
 638
 639         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 640                 lnet_peer_ni_finalize_wait(ptable);
 641 }
 642
 643 static struct lnet_peer_ni *
 644 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 645 {
 646         struct list_head        *peers;
 647         struct lnet_peer_ni     *lp;
 648
 649         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 650                 return NULL;
 651
 652         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 653         list_for_each_entry(lp, peers, lpni_hashlist) {
 654                 if (lp->lpni_nid == nid) {
 655                         lnet_peer_ni_addref_locked(lp);
 656                         return lp;
 657                 }
 658         }
 659
 660         return NULL;
 661 }
 662
 663 struct lnet_peer_ni *
 664 lnet_find_peer_ni_locked(lnet_nid_t nid)
 665 {
 666         struct lnet_peer_ni *lpni;
 667         struct lnet_peer_table *ptable;
 668         int cpt;
 669
 670         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 671
 672         ptable = the_lnet.ln_peer_tables[cpt];
 673         lpni = lnet_get_peer_ni_locked(ptable, nid);
 674
 675         return lpni;
 676 }
 677
 678 struct lnet_peer_ni *
 679 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 680 {
 681         struct lnet_peer_net *lpn;
 682         struct lnet_peer_ni *lpni;
 683
 684         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 685         if (!lpn)
 686                 return NULL;
 687
 688         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 689                 if (lpni->lpni_nid == nid)
 690                         return lpni;
 691         }
 692
 693         return NULL;
 694 }
 695
 696 struct lnet_peer *
 697 lnet_find_peer(lnet_nid_t nid)
 698 {
 699         struct lnet_peer_ni *lpni;
 700         struct lnet_peer *lp = NULL;
 701         int cpt;
 702
 703         cpt = lnet_net_lock_current();
 704         lpni = lnet_find_peer_ni_locked(nid);
 705         if (lpni) {
 706                 lp = lpni->lpni_peer_net->lpn_peer;
 707                 lnet_peer_addref_locked(lp);
 708                 lnet_peer_ni_decref_locked(lpni);
 709         }
 710         lnet_net_unlock(cpt);
 711
 712         return lp;
 713 }
 714
 715 struct lnet_peer_net *
 716 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 717 {
 718         struct lnet_peer_net *net;
 719
 720         if (!prev_lpn_id) {
 721                 /* no net id provided return the first net */
 722                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 723                                                struct lnet_peer_net,
 724                                                lpn_peer_nets);
 725
 726                 return net;
 727         }
 728
 729         /* find the net after the one provided */
 730         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 731                 if (net->lpn_net_id == prev_lpn_id) {
 732                         /*
 733                          * if we reached the end of the list loop to the
 734                          * beginning.
 735                          */
 736                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 737                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 738                                                                 struct lnet_peer_net,
 739                                                                 lpn_peer_nets);
 740                         else
 741                                 return list_next_entry(net, lpn_peer_nets);
 742                 }
 743         }
 744
 745         return NULL;
 746 }
 747
 748 struct lnet_peer_ni *
 749 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 750                              struct lnet_peer_net *peer_net,
 751                              struct lnet_peer_ni *prev)
 752 {
 753         struct lnet_peer_ni *lpni;
 754         struct lnet_peer_net *net = peer_net;
 755
 756         if (!prev) {
 757                 if (!net) {
 758                         if (list_empty(&peer->lp_peer_nets))
 759                                 return NULL;
 760
 761                         net = list_entry(peer->lp_peer_nets.next,
 762                                          struct lnet_peer_net,
 763                                          lpn_peer_nets);
 764                 }
 765                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 766                                   lpni_peer_nis);
 767
 768                 return lpni;
 769         }
 770
 771         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 772                 /*
 773                  * if you reached the end of the peer ni list and the peer
 774                  * net is specified then there are no more peer nis in that
 775                  * net.
 776                  */
 777                 if (net)
 778                         return NULL;
 779
 780                 /*
 781                  * we reached the end of this net ni list. move to the
 782                  * next net
 783                  */
 784                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 785                     &peer->lp_peer_nets)
 786                         /* no more nets and no more NIs. */
 787                         return NULL;
 788
 789                 /* get the next net */
 790                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 791                                  struct lnet_peer_net,
 792                                  lpn_peer_nets);
 793                 /* get the ni on it */
 794                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 795                                   lpni_peer_nis);
 796
 797                 return lpni;
 798         }
 799
 800         /* there are more nis left */
 801         lpni = list_entry(prev->lpni_peer_nis.next,
 802                           struct lnet_peer_ni, lpni_peer_nis);
 803
 804         return lpni;
 805 }
 806
 807 /* Call with the ln_api_mutex held */
 808 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 809 {
 810         struct lnet_process_id id;
 811         struct lnet_peer_table *ptable;
 812         struct lnet_peer *lp;
 813         __u32 count = 0;
 814         __u32 size = 0;
 815         int lncpt;
 816         int cpt;
 817         __u32 i;
 818         int rc;
 819
 820         rc = -ESHUTDOWN;
 821         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 822                 goto done;
 823
 824         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 825
 826         /*
 827          * Count the number of peers, and return E2BIG if the buffer
 828          * is too small. We'll also return the desired size.
 829          */
 830         rc = -E2BIG;
 831         for (cpt = 0; cpt < lncpt; cpt++) {
 832                 ptable = the_lnet.ln_peer_tables[cpt];
 833                 count += ptable->pt_peers;
 834         }
 835         size = count * sizeof(*ids);
 836         if (size > *sizep)
 837                 goto done;
 838
 839         /*
 840          * Walk the peer lists and copy out the primary nids.
 841          * This is safe because the peer lists are only modified
 842          * while the ln_api_mutex is held. So we don't need to
 843          * hold the lnet_net_lock as well, and can therefore
 844          * directly call copy_to_user().
 845          */
 846         rc = -EFAULT;
 847         memset(&id, 0, sizeof(id));
 848         id.pid = LNET_PID_LUSTRE;
 849         i = 0;
 850         for (cpt = 0; cpt < lncpt; cpt++) {
 851                 ptable = the_lnet.ln_peer_tables[cpt];
 852                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 853                         if (i >= count)
 854                                 goto done;
 855                         id.nid = lp->lp_primary_nid;
 856                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 857                                 goto done;
 858                         i++;
 859                 }
 860         }
 861         rc = 0;
 862 done:
 863         *countp = count;
 864         *sizep = size;
 865         return rc;
 866 }
 867
 868 /*
 869  * Start pushes to peers that need to be updated for a configuration
 870  * change on this node.
 871  */
 872 void
 873 lnet_push_update_to_peers(int force)
 874 {
 875         struct lnet_peer_table *ptable;
 876         struct lnet_peer *lp;
 877         int lncpt;
 878         int cpt;
 879
 880         lnet_net_lock(LNET_LOCK_EX);
 881         if (lnet_peer_discovery_disabled)
 882                 force = 0;
 883         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 884         for (cpt = 0; cpt < lncpt; cpt++) {
 885                 ptable = the_lnet.ln_peer_tables[cpt];
 886                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 887                         if (force) {
 888                                 spin_lock(&lp->lp_lock);
 889                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 890                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 891                                 spin_unlock(&lp->lp_lock);
 892                         }
 893                         if (lnet_peer_needs_push(lp))
 894                                 lnet_peer_queue_for_discovery(lp);
 895                 }
 896         }
 897         lnet_net_unlock(LNET_LOCK_EX);
 898         wake_up(&the_lnet.ln_dc_waitq);
 899 }
 900
 901 /* find the NID in the preferred gateways for the remote peer
 902  * return:
 903  *      false: list is not empty and NID is not preferred
 904  *      false: list is empty
 905  *      true: nid is found in the list
 906  */
 907 bool
 908 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
 909                              lnet_nid_t gw_nid)
 910 {
 911         struct lnet_nid_list *ne;
 912
 913         CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
 914                libcfs_nid2str(lpni->lpni_nid),
 915                list_empty(&lpni->lpni_rtr_pref_nids));
 916
 917         if (list_empty(&lpni->lpni_rtr_pref_nids))
 918                 return false;
 919
 920         /* iterate through all the preferred NIDs and see if any of them
 921          * matches the provided gw_nid
 922          */
 923         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
 924                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
 925                        libcfs_nid2str(ne->nl_nid),
 926                        libcfs_nid2str(gw_nid));
 927                 if (ne->nl_nid == gw_nid)
 928                         return true;
 929         }
 930
 931         return false;
 932 }
 933
 934 void
 935 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
 936 {
 937         struct list_head zombies;
 938         struct lnet_nid_list *ne;
 939         struct lnet_nid_list *tmp;
 940         int cpt = lpni->lpni_cpt;
 941
 942         INIT_LIST_HEAD(&zombies);
 943
 944         lnet_net_lock(cpt);
 945         list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
 946         lnet_net_unlock(cpt);
 947
 948         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
 949                 list_del(&ne->nl_list);
 950                 LIBCFS_FREE(ne, sizeof(*ne));
 951         }
 952 }
 953
 954 int
 955 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
 956                        lnet_nid_t gw_nid)
 957 {
 958         int cpt = lpni->lpni_cpt;
 959         struct lnet_nid_list *ne = NULL;
 960
 961         /* This function is called with api_mutex held. When the api_mutex
 962          * is held the list can not be modified, as it is only modified as
 963          * a result of applying a UDSP and that happens under api_mutex
 964          * lock.
 965          */
 966         __must_hold(&the_lnet.ln_api_mutex);
 967
 968         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
 969                 if (ne->nl_nid == gw_nid)
 970                         return -EEXIST;
 971         }
 972
 973         LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
 974         if (!ne)
 975                 return -ENOMEM;
 976
 977         ne->nl_nid = gw_nid;
 978
 979         /* Lock the cpt to protect against addition and checks in the
 980          * selection algorithm
 981          */
 982         lnet_net_lock(cpt);
 983         list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
 984         lnet_net_unlock(cpt);
 985
 986         return 0;
 987 }
 988
 989 /*
 990  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 991  * this is a preferred point-to-point path. Call with lnet_net_lock in
 992  * shared mmode.
 993  */
 994 bool
 995 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 996 {
 997         struct lnet_nid_list *ne;
 998
 999         if (lpni->lpni_pref_nnids == 0)
1000                 return false;
1001         if (lpni->lpni_pref_nnids == 1)
1002                 return lpni->lpni_pref.nid == nid;
1003         list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1004                 if (ne->nl_nid == nid)
1005                         return true;
1006         }
1007         return false;
1008 }
1009
1010 /*
1011  * Set a single ni as preferred, provided no preferred ni is already
1012  * defined. Only to be used for non-multi-rail peer_ni.
1013  */
1014 int
1015 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1016 {
1017         int rc = 0;
1018
1019         spin_lock(&lpni->lpni_lock);
1020         if (nid == LNET_NID_ANY) {
1021                 rc = -EINVAL;
1022         } else if (lpni->lpni_pref_nnids > 0) {
1023                 rc = -EPERM;
1024         } else if (lpni->lpni_pref_nnids == 0) {
1025                 lpni->lpni_pref.nid = nid;
1026                 lpni->lpni_pref_nnids = 1;
1027                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1028         }
1029         spin_unlock(&lpni->lpni_lock);
1030
1031         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1032                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
1033         return rc;
1034 }
1035
1036 /*
1037  * Clear the preferred NID from a non-multi-rail peer_ni, provided
1038  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1039  */
1040 int
1041 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1042 {
1043         int rc = 0;
1044
1045         spin_lock(&lpni->lpni_lock);
1046         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1047                 lpni->lpni_pref_nnids = 0;
1048                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1049         } else if (lpni->lpni_pref_nnids == 0) {
1050                 rc = -ENOENT;
1051         } else {
1052                 rc = -EPERM;
1053         }
1054         spin_unlock(&lpni->lpni_lock);
1055
1056         CDEBUG(D_NET, "peer %s: %d\n",
1057                libcfs_nid2str(lpni->lpni_nid), rc);
1058         return rc;
1059 }
1060
1061 void
1062 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1063 {
1064         lpni->lpni_sel_priority = priority;
1065 }
1066
1067 /*
1068  * Clear the preferred NIDs from a non-multi-rail peer.
1069  */
1070 void
1071 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1072 {
1073         struct lnet_peer_ni *lpni = NULL;
1074
1075         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1076                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1077 }
1078
1079 int
1080 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1081 {
1082         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1083         struct lnet_nid_list *ne1 = NULL;
1084         struct lnet_nid_list *ne2 = NULL;
1085         lnet_nid_t tmp_nid = LNET_NID_ANY;
1086         int rc = 0;
1087
1088         if (nid == LNET_NID_ANY) {
1089                 rc = -EINVAL;
1090                 goto out;
1091         }
1092
1093         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1094                 rc = -EEXIST;
1095                 goto out;
1096         }
1097
1098         /* A non-MR node may have only one preferred NI per peer_ni */
1099         if (lpni->lpni_pref_nnids > 0 &&
1100             !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1101                 rc = -EPERM;
1102                 goto out;
1103         }
1104
1105         /* add the new preferred nid to the list of preferred nids */
1106         if (lpni->lpni_pref_nnids != 0) {
1107                 size_t alloc_size = sizeof(*ne1);
1108
1109                 if (lpni->lpni_pref_nnids == 1) {
1110                         tmp_nid = lpni->lpni_pref.nid;
1111                         INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1112                 }
1113
1114                 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1115                         if (ne1->nl_nid == nid) {
1116                                 rc = -EEXIST;
1117                                 goto out;
1118                         }
1119                 }
1120
1121                 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1122                                  alloc_size);
1123                 if (!ne1) {
1124                         rc = -ENOMEM;
1125                         goto out;
1126                 }
1127
1128                 /* move the originally stored nid to the list */
1129                 if (lpni->lpni_pref_nnids == 1) {
1130                         LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1131                                 lpni->lpni_cpt, alloc_size);
1132                         if (!ne2) {
1133                                 rc = -ENOMEM;
1134                                 goto out;
1135                         }
1136                         INIT_LIST_HEAD(&ne2->nl_list);
1137                         ne2->nl_nid = tmp_nid;
1138                 }
1139                 ne1->nl_nid = nid;
1140         }
1141
1142         lnet_net_lock(LNET_LOCK_EX);
1143         spin_lock(&lpni->lpni_lock);
1144         if (lpni->lpni_pref_nnids == 0) {
1145                 lpni->lpni_pref.nid = nid;
1146         } else {
1147                 if (ne2)
1148                         list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1149                 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1150         }
1151         lpni->lpni_pref_nnids++;
1152         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1153         spin_unlock(&lpni->lpni_lock);
1154         lnet_net_unlock(LNET_LOCK_EX);
1155
1156 out:
1157         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1158                 spin_lock(&lpni->lpni_lock);
1159                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1160                 spin_unlock(&lpni->lpni_lock);
1161         }
1162         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1163                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1164         return rc;
1165 }
1166
1167 int
1168 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1169 {
1170         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1171         struct lnet_nid_list *ne = NULL;
1172         int rc = 0;
1173
1174         if (lpni->lpni_pref_nnids == 0) {
1175                 rc = -ENOENT;
1176                 goto out;
1177         }
1178
1179         if (lpni->lpni_pref_nnids == 1) {
1180                 if (lpni->lpni_pref.nid != nid) {
1181                         rc = -ENOENT;
1182                         goto out;
1183                 }
1184         } else {
1185                 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1186                         if (ne->nl_nid == nid)
1187                                 goto remove_nid_entry;
1188                 }
1189                 rc = -ENOENT;
1190                 ne = NULL;
1191                 goto out;
1192         }
1193
1194 remove_nid_entry:
1195         lnet_net_lock(LNET_LOCK_EX);
1196         spin_lock(&lpni->lpni_lock);
1197         if (lpni->lpni_pref_nnids == 1)
1198                 lpni->lpni_pref.nid = LNET_NID_ANY;
1199         else {
1200                 list_del_init(&ne->nl_list);
1201                 if (lpni->lpni_pref_nnids == 2) {
1202                         struct lnet_nid_list *ne, *tmp;
1203
1204                         list_for_each_entry_safe(ne, tmp,
1205                                                  &lpni->lpni_pref.nids,
1206                                                  nl_list) {
1207                                 lpni->lpni_pref.nid = ne->nl_nid;
1208                                 list_del_init(&ne->nl_list);
1209                                 LIBCFS_FREE(ne, sizeof(*ne));
1210                         }
1211                 }
1212         }
1213         lpni->lpni_pref_nnids--;
1214         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1215         spin_unlock(&lpni->lpni_lock);
1216         lnet_net_unlock(LNET_LOCK_EX);
1217
1218         if (ne)
1219                 LIBCFS_FREE(ne, sizeof(*ne));
1220 out:
1221         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1222                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1223         return rc;
1224 }
1225
1226 void
1227 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1228 {
1229         struct list_head zombies;
1230         struct lnet_nid_list *ne;
1231         struct lnet_nid_list *tmp;
1232
1233         INIT_LIST_HEAD(&zombies);
1234
1235         lnet_net_lock(LNET_LOCK_EX);
1236         if (lpni->lpni_pref_nnids == 1)
1237                 lpni->lpni_pref.nid = LNET_NID_ANY;
1238         else if (lpni->lpni_pref_nnids > 1)
1239                 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1240         lpni->lpni_pref_nnids = 0;
1241         lnet_net_unlock(LNET_LOCK_EX);
1242
1243         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1244                 list_del_init(&ne->nl_list);
1245                 LIBCFS_FREE(ne, sizeof(*ne));
1246         }
1247 }
1248
1249 lnet_nid_t
1250 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1251 {
1252         struct lnet_peer_ni *lpni;
1253         lnet_nid_t primary_nid = nid;
1254
1255         lpni = lnet_find_peer_ni_locked(nid);
1256         if (lpni) {
1257                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1258                 lnet_peer_ni_decref_locked(lpni);
1259         }
1260
1261         return primary_nid;
1262 }
1263
1264 bool
1265 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1266 __must_hold(&lp->lp_lock)
1267 {
1268         if (lnet_peer_discovery_disabled)
1269                 return true;
1270
1271         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1272             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1273                 return true;
1274         }
1275
1276         return false;
1277 }
1278
1279 /*
1280  * Peer Discovery
1281  */
1282 bool
1283 lnet_is_discovery_disabled(struct lnet_peer *lp)
1284 {
1285         bool rc = false;
1286
1287         spin_lock(&lp->lp_lock);
1288         rc = lnet_is_discovery_disabled_locked(lp);
1289         spin_unlock(&lp->lp_lock);
1290
1291         return rc;
1292 }
1293
1294 lnet_nid_t
1295 LNetPrimaryNID(lnet_nid_t nid)
1296 {
1297         struct lnet_peer *lp;
1298         struct lnet_peer_ni *lpni;
1299         lnet_nid_t primary_nid = nid;
1300         int rc = 0;
1301         int cpt;
1302
1303         if (nid == LNET_NID_LO_0)
1304                 return LNET_NID_LO_0;
1305
1306         cpt = lnet_net_lock_current();
1307         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1308         if (IS_ERR(lpni)) {
1309                 rc = PTR_ERR(lpni);
1310                 goto out_unlock;
1311         }
1312         lp = lpni->lpni_peer_net->lpn_peer;
1313
1314         while (!lnet_peer_is_uptodate(lp)) {
1315                 spin_lock(&lp->lp_lock);
1316                 /* force a full discovery cycle */
1317                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1318                 spin_unlock(&lp->lp_lock);
1319
1320                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1321                 if (rc)
1322                         goto out_decref;
1323                 lp = lpni->lpni_peer_net->lpn_peer;
1324
1325                 /* Only try once if discovery is disabled */
1326                 if (lnet_is_discovery_disabled(lp))
1327                         break;
1328         }
1329         primary_nid = lp->lp_primary_nid;
1330 out_decref:
1331         lnet_peer_ni_decref_locked(lpni);
1332 out_unlock:
1333         lnet_net_unlock(cpt);
1334
1335         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1336                libcfs_nid2str(primary_nid), rc);
1337         return primary_nid;
1338 }
1339 EXPORT_SYMBOL(LNetPrimaryNID);
1340
1341 struct lnet_peer_net *
1342 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1343 {
1344         struct lnet_peer_net *peer_net;
1345         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1346                 if (peer_net->lpn_net_id == net_id)
1347                         return peer_net;
1348         }
1349         return NULL;
1350 }
1351
1352 /*
1353  * Attach a peer_ni to a peer_net and peer. This function assumes
1354  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1355  * may be attached to a different peer, in which case it will be
1356  * properly detached first. The whole operation is done atomically.
1357  *
1358  * This function consumes the reference on lpni and Always returns 0.
1359  * This is the last function called from functions that do return an
1360  * int, so returning 0 here allows the compiler to do a tail call.
1361  */
1362 static int
1363 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1364                                 struct lnet_peer_net *lpn,
1365                                 struct lnet_peer_ni *lpni,
1366                                 unsigned flags)
1367 {
1368         struct lnet_peer_table *ptable;
1369
1370         /* Install the new peer_ni */
1371         lnet_net_lock(LNET_LOCK_EX);
1372         /* Add peer_ni to global peer table hash, if necessary. */
1373         if (list_empty(&lpni->lpni_hashlist)) {
1374                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1375
1376                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1377                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1378                 ptable->pt_version++;
1379                 lnet_peer_ni_addref_locked(lpni);
1380         }
1381
1382         /* Detach the peer_ni from an existing peer, if necessary. */
1383         if (lpni->lpni_peer_net) {
1384                 LASSERT(lpni->lpni_peer_net != lpn);
1385                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1386                 lnet_peer_detach_peer_ni_locked(lpni);
1387                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1388                 lpni->lpni_peer_net = NULL;
1389         }
1390
1391         /* Add peer_ni to peer_net */
1392         lpni->lpni_peer_net = lpn;
1393         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1394         lnet_update_peer_net_healthv(lpni);
1395         lnet_peer_net_addref_locked(lpn);
1396
1397         /* Add peer_net to peer */
1398         if (!lpn->lpn_peer) {
1399                 lpn->lpn_peer = lp;
1400                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1401                 lnet_peer_addref_locked(lp);
1402         }
1403
1404         /* Add peer to global peer list, if necessary */
1405         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1406         if (list_empty(&lp->lp_peer_list)) {
1407                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1408                 ptable->pt_peers++;
1409         }
1410
1411
1412         /* Update peer state */
1413         spin_lock(&lp->lp_lock);
1414         if (flags & LNET_PEER_CONFIGURED) {
1415                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1416                         lp->lp_state |= LNET_PEER_CONFIGURED;
1417         }
1418         if (flags & LNET_PEER_MULTI_RAIL) {
1419                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1420                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1421                         lnet_peer_clr_non_mr_pref_nids(lp);
1422                 }
1423         }
1424         spin_unlock(&lp->lp_lock);
1425
1426         lp->lp_nnis++;
1427
1428         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1429                libcfs_nid2str(lp->lp_primary_nid),
1430                libcfs_nid2str(lpni->lpni_nid), flags);
1431         lnet_peer_ni_decref_locked(lpni);
1432         lnet_net_unlock(LNET_LOCK_EX);
1433
1434         return 0;
1435 }
1436
1437 /*
1438  * Create a new peer, with nid as its primary nid.
1439  *
1440  * Call with the lnet_api_mutex held.
1441  */
1442 static int
1443 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1444 {
1445         struct lnet_peer *lp;
1446         struct lnet_peer_net *lpn;
1447         struct lnet_peer_ni *lpni;
1448         int rc = 0;
1449
1450         LASSERT(nid != LNET_NID_ANY);
1451
1452         /*
1453          * No need for the lnet_net_lock here, because the
1454          * lnet_api_mutex is held.
1455          */
1456         lpni = lnet_find_peer_ni_locked(nid);
1457         if (lpni) {
1458                 /* A peer with this NID already exists. */
1459                 lp = lpni->lpni_peer_net->lpn_peer;
1460                 lnet_peer_ni_decref_locked(lpni);
1461                 /*
1462                  * This is an error if the peer was configured and the
1463                  * primary NID differs or an attempt is made to change
1464                  * the Multi-Rail flag. Otherwise the assumption is
1465                  * that an existing peer is being modified.
1466                  */
1467                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1468                         if (lp->lp_primary_nid != nid)
1469                                 rc = -EEXIST;
1470                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1471                                 rc = -EPERM;
1472                         goto out;
1473                 }
1474                 /* Delete and recreate as a configured peer. */
1475                 lnet_peer_del(lp);
1476         }
1477
1478         /* Create peer, peer_net, and peer_ni. */
1479         rc = -ENOMEM;
1480         lp = lnet_peer_alloc(nid);
1481         if (!lp)
1482                 goto out;
1483         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1484         if (!lpn)
1485                 goto out_free_lp;
1486         lpni = lnet_peer_ni_alloc(nid);
1487         if (!lpni)
1488                 goto out_free_lpn;
1489
1490         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1491
1492 out_free_lpn:
1493         LIBCFS_FREE(lpn, sizeof(*lpn));
1494 out_free_lp:
1495         LIBCFS_FREE(lp, sizeof(*lp));
1496 out:
1497         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1498                libcfs_nid2str(nid), flags, rc);
1499         return rc;
1500 }
1501
1502 /*
1503  * Add a NID to a peer. Call with ln_api_mutex held.
1504  *
1505  * Error codes:
1506  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1507  *  -EEXIST:   The NID was configured by DLC for a different peer.
1508  *  -ENOMEM:   Out of memory.
1509  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1510  *             non-multi-rail peer.
1511  */
1512 static int
1513 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1514 {
1515         struct lnet_peer_net *lpn;
1516         struct lnet_peer_ni *lpni;
1517         int rc = 0;
1518
1519         LASSERT(lp);
1520         LASSERT(nid != LNET_NID_ANY);
1521
1522         /* A configured peer can only be updated through configuration. */
1523         if (!(flags & LNET_PEER_CONFIGURED)) {
1524                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1525                         rc = -EPERM;
1526                         goto out;
1527                 }
1528         }
1529
1530         /*
1531          * The MULTI_RAIL flag can be set but not cleared, because
1532          * that would leave the peer struct in an invalid state.
1533          */
1534         if (flags & LNET_PEER_MULTI_RAIL) {
1535                 spin_lock(&lp->lp_lock);
1536                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1537                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1538                         lnet_peer_clr_non_mr_pref_nids(lp);
1539                 }
1540                 spin_unlock(&lp->lp_lock);
1541         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1542                 rc = -EPERM;
1543                 goto out;
1544         }
1545
1546         lpni = lnet_find_peer_ni_locked(nid);
1547         if (lpni) {
1548                 /*
1549                  * A peer_ni already exists. This is only a problem if
1550                  * it is not connected to this peer and was configured
1551                  * by DLC.
1552                  */
1553                 if (lpni->lpni_peer_net->lpn_peer == lp)
1554                         goto out_free_lpni;
1555                 if (lnet_peer_ni_is_configured(lpni)) {
1556                         rc = -EEXIST;
1557                         goto out_free_lpni;
1558                 }
1559                 /* If this is the primary NID, destroy the peer. */
1560                 if (lnet_peer_ni_is_primary(lpni)) {
1561                         struct lnet_peer *rtr_lp =
1562                                 lpni->lpni_peer_net->lpn_peer;
1563                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1564                         /*
1565                          * if we're trying to delete a router it means
1566                          * we're moving this peer NI to a new peer so must
1567                          * transfer router properties to the new peer
1568                          */
1569                         if (rtr_refcount > 0) {
1570                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1571                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1572                         }
1573                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1574                         lnet_peer_ni_decref_locked(lpni);
1575                         lpni = lnet_peer_ni_alloc(nid);
1576                         if (!lpni) {
1577                                 rc = -ENOMEM;
1578                                 goto out_free_lpni;
1579                         }
1580                 }
1581         } else {
1582                 lpni = lnet_peer_ni_alloc(nid);
1583                 if (!lpni) {
1584                         rc = -ENOMEM;
1585                         goto out_free_lpni;
1586                 }
1587         }
1588
1589         /*
1590          * Get the peer_net. Check that we're not adding a second
1591          * peer_ni on a peer_net of a non-multi-rail peer.
1592          */
1593         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1594         if (!lpn) {
1595                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1596                 if (!lpn) {
1597                         rc = -ENOMEM;
1598                         goto out_free_lpni;
1599                 }
1600         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1601                 rc = -ENOTUNIQ;
1602                 goto out_free_lpni;
1603         }
1604
1605         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1606
1607 out_free_lpni:
1608         lnet_peer_ni_decref_locked(lpni);
1609 out:
1610         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1611                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1612                flags, rc);
1613         return rc;
1614 }
1615
1616 /*
1617  * Update the primary NID of a peer, if possible.
1618  *
1619  * Call with the lnet_api_mutex held.
1620  */
1621 static int
1622 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1623 {
1624         lnet_nid_t old = lp->lp_primary_nid;
1625         int rc = 0;
1626
1627         if (lp->lp_primary_nid == nid)
1628                 goto out;
1629         rc = lnet_peer_add_nid(lp, nid, flags);
1630         if (rc)
1631                 goto out;
1632         lp->lp_primary_nid = nid;
1633 out:
1634         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1635                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1636         return rc;
1637 }
1638
1639 /*
1640  * lpni creation initiated due to traffic either sending or receiving.
1641  */
1642 static int
1643 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1644 {
1645         struct lnet_peer *lp;
1646         struct lnet_peer_net *lpn;
1647         struct lnet_peer_ni *lpni;
1648         unsigned flags = 0;
1649         int rc = 0;
1650
1651         if (nid == LNET_NID_ANY) {
1652                 rc = -EINVAL;
1653                 goto out;
1654         }
1655
1656         /* lnet_net_lock is not needed here because ln_api_lock is held */
1657         lpni = lnet_find_peer_ni_locked(nid);
1658         if (lpni) {
1659                 /*
1660                  * We must have raced with another thread. Since we
1661                  * know next to nothing about a peer_ni created by
1662                  * traffic, we just assume everything is ok and
1663                  * return.
1664                  */
1665                 lnet_peer_ni_decref_locked(lpni);
1666                 goto out;
1667         }
1668
1669         /* Create peer, peer_net, and peer_ni. */
1670         rc = -ENOMEM;
1671         lp = lnet_peer_alloc(nid);
1672         if (!lp)
1673                 goto out;
1674         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1675         if (!lpn)
1676                 goto out_free_lp;
1677         lpni = lnet_peer_ni_alloc(nid);
1678         if (!lpni)
1679                 goto out_free_lpn;
1680         if (pref != LNET_NID_ANY)
1681                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1682
1683         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1684
1685 out_free_lpn:
1686         LIBCFS_FREE(lpn, sizeof(*lpn));
1687 out_free_lp:
1688         LIBCFS_FREE(lp, sizeof(*lp));
1689 out:
1690         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1691         return rc;
1692 }
1693
1694 /*
1695  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1696  *
1697  * This API handles the following combinations:
1698  *   Create a peer with its primary NI if only the prim_nid is provided
1699  *   Add a NID to a peer identified by the prim_nid. The peer identified
1700  *   by the prim_nid must already exist.
1701  *   The peer being created may be non-MR.
1702  *
1703  * The caller must hold ln_api_mutex. This prevents the peer from
1704  * being created/modified/deleted by a different thread.
1705  */
1706 int
1707 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1708 {
1709         struct lnet_peer *lp = NULL;
1710         struct lnet_peer_ni *lpni;
1711         unsigned flags;
1712
1713         /* The prim_nid must always be specified */
1714         if (prim_nid == LNET_NID_ANY)
1715                 return -EINVAL;
1716
1717         flags = LNET_PEER_CONFIGURED;
1718         if (mr)
1719                 flags |= LNET_PEER_MULTI_RAIL;
1720
1721         /*
1722          * If nid isn't specified, we must create a new peer with
1723          * prim_nid as its primary nid.
1724          */
1725         if (nid == LNET_NID_ANY)
1726                 return lnet_peer_add(prim_nid, flags);
1727
1728         /* Look up the prim_nid, which must exist. */
1729         lpni = lnet_find_peer_ni_locked(prim_nid);
1730         if (!lpni)
1731                 return -ENOENT;
1732         lnet_peer_ni_decref_locked(lpni);
1733         lp = lpni->lpni_peer_net->lpn_peer;
1734
1735         /* Peer must have been configured. */
1736         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1737                 CDEBUG(D_NET, "peer %s was not configured\n",
1738                        libcfs_nid2str(prim_nid));
1739                 return -ENOENT;
1740         }
1741
1742         /* Primary NID must match */
1743         if (lp->lp_primary_nid != prim_nid) {
1744                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1745                        libcfs_nid2str(prim_nid),
1746                        libcfs_nid2str(lp->lp_primary_nid));
1747                 return -ENODEV;
1748         }
1749
1750         /* Multi-Rail flag must match. */
1751         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1752                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1753                        libcfs_nid2str(prim_nid));
1754                 return -EPERM;
1755         }
1756
1757         return lnet_peer_add_nid(lp, nid, flags);
1758 }
1759
1760 /*
1761  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1762  *
1763  * This API handles the following combinations:
1764  *   Delete a NI from a peer if both prim_nid and nid are provided.
1765  *   Delete a peer if only prim_nid is provided.
1766  *   Delete a peer if its primary nid is provided.
1767  *
1768  * The caller must hold ln_api_mutex. This prevents the peer from
1769  * being modified/deleted by a different thread.
1770  */
1771 int
1772 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1773 {
1774         struct lnet_peer *lp;
1775         struct lnet_peer_ni *lpni;
1776         unsigned flags;
1777
1778         if (prim_nid == LNET_NID_ANY)
1779                 return -EINVAL;
1780
1781         lpni = lnet_find_peer_ni_locked(prim_nid);
1782         if (!lpni)
1783                 return -ENOENT;
1784         lnet_peer_ni_decref_locked(lpni);
1785         lp = lpni->lpni_peer_net->lpn_peer;
1786
1787         if (prim_nid != lp->lp_primary_nid) {
1788                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1789                        libcfs_nid2str(prim_nid),
1790                        libcfs_nid2str(lp->lp_primary_nid));
1791                 return -ENODEV;
1792         }
1793
1794         lnet_net_lock(LNET_LOCK_EX);
1795         if (lp->lp_rtr_refcount > 0) {
1796                 lnet_net_unlock(LNET_LOCK_EX);
1797                 CERROR("%s is a router. Can not be deleted\n",
1798                        libcfs_nid2str(prim_nid));
1799                 return -EBUSY;
1800         }
1801         lnet_net_unlock(LNET_LOCK_EX);
1802
1803         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1804                 return lnet_peer_del(lp);
1805
1806         flags = LNET_PEER_CONFIGURED;
1807         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1808                 flags |= LNET_PEER_MULTI_RAIL;
1809
1810         return lnet_peer_del_nid(lp, nid, flags);
1811 }
1812
1813 void
1814 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1815 {
1816         struct lnet_peer_table *ptable;
1817         struct lnet_peer_net *lpn;
1818
1819         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1820
1821         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1822         LASSERT(list_empty(&lpni->lpni_txq));
1823         LASSERT(lpni->lpni_txqnob == 0);
1824         LASSERT(list_empty(&lpni->lpni_peer_nis));
1825         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1826
1827         lpn = lpni->lpni_peer_net;
1828         lpni->lpni_peer_net = NULL;
1829         lpni->lpni_net = NULL;
1830
1831         if (!list_empty(&lpni->lpni_hashlist)) {
1832                 /* remove the peer ni from the zombie list */
1833                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1834                 spin_lock(&ptable->pt_zombie_lock);
1835                 list_del_init(&lpni->lpni_hashlist);
1836                 ptable->pt_zombies--;
1837                 spin_unlock(&ptable->pt_zombie_lock);
1838         }
1839
1840         if (lpni->lpni_pref_nnids > 1) {
1841                 struct lnet_nid_list *ne, *tmp;
1842
1843                 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
1844                                          nl_list) {
1845                         list_del_init(&ne->nl_list);
1846                         LIBCFS_FREE(ne, sizeof(*ne));
1847                 }
1848         }
1849         LIBCFS_FREE(lpni, sizeof(*lpni));
1850
1851         if (lpn)
1852                 lnet_peer_net_decref_locked(lpn);
1853 }
1854
1855 struct lnet_peer_ni *
1856 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1857 {
1858         struct lnet_peer_ni *lpni = NULL;
1859         int rc;
1860
1861         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1862                 return ERR_PTR(-ESHUTDOWN);
1863
1864         /*
1865          * find if a peer_ni already exists.
1866          * If so then just return that.
1867          */
1868         lpni = lnet_find_peer_ni_locked(nid);
1869         if (lpni)
1870                 return lpni;
1871
1872         lnet_net_unlock(cpt);
1873
1874         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1875         if (rc) {
1876                 lpni = ERR_PTR(rc);
1877                 goto out_net_relock;
1878         }
1879
1880         lpni = lnet_find_peer_ni_locked(nid);
1881         LASSERT(lpni);
1882
1883 out_net_relock:
1884         lnet_net_lock(cpt);
1885
1886         return lpni;
1887 }
1888
1889 /*
1890  * Get a peer_ni for the given nid, create it if necessary. Takes a
1891  * hold on the peer_ni.
1892  */
1893 struct lnet_peer_ni *
1894 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1895 {
1896         struct lnet_peer_ni *lpni = NULL;
1897         int rc;
1898
1899         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1900                 return ERR_PTR(-ESHUTDOWN);
1901
1902         /*
1903          * find if a peer_ni already exists.
1904          * If so then just return that.
1905          */
1906         lpni = lnet_find_peer_ni_locked(nid);
1907         if (lpni)
1908                 return lpni;
1909
1910         /*
1911          * Slow path:
1912          * use the lnet_api_mutex to serialize the creation of the peer_ni
1913          * and the creation/deletion of the local ni/net. When a local ni is
1914          * created, if there exists a set of peer_nis on that network,
1915          * they need to be traversed and updated. When a local NI is
1916          * deleted, which could result in a network being deleted, then
1917          * all peer nis on that network need to be removed as well.
1918          *
1919          * Creation through traffic should also be serialized with
1920          * creation through DLC.
1921          */
1922         lnet_net_unlock(cpt);
1923         mutex_lock(&the_lnet.ln_api_mutex);
1924         /*
1925          * Shutdown is only set under the ln_api_lock, so a single
1926          * check here is sufficent.
1927          */
1928         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1929                 lpni = ERR_PTR(-ESHUTDOWN);
1930                 goto out_mutex_unlock;
1931         }
1932
1933         rc = lnet_peer_ni_traffic_add(nid, pref);
1934         if (rc) {
1935                 lpni = ERR_PTR(rc);
1936                 goto out_mutex_unlock;
1937         }
1938
1939         lpni = lnet_find_peer_ni_locked(nid);
1940         LASSERT(lpni);
1941
1942 out_mutex_unlock:
1943         mutex_unlock(&the_lnet.ln_api_mutex);
1944         lnet_net_lock(cpt);
1945
1946         /* Lock has been dropped, check again for shutdown. */
1947         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1948                 if (!IS_ERR(lpni))
1949                         lnet_peer_ni_decref_locked(lpni);
1950                 lpni = ERR_PTR(-ESHUTDOWN);
1951         }
1952
1953         return lpni;
1954 }
1955
1956 bool
1957 lnet_peer_gw_discovery(struct lnet_peer *lp)
1958 {
1959         bool rc = false;
1960
1961         spin_lock(&lp->lp_lock);
1962         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1963                 rc = true;
1964         spin_unlock(&lp->lp_lock);
1965
1966         return rc;
1967 }
1968
1969 bool
1970 lnet_peer_is_uptodate(struct lnet_peer *lp)
1971 {
1972         bool rc;
1973
1974         spin_lock(&lp->lp_lock);
1975         rc = lnet_peer_is_uptodate_locked(lp);
1976         spin_unlock(&lp->lp_lock);
1977         return rc;
1978 }
1979
1980 /*
1981  * Is a peer uptodate from the point of view of discovery?
1982  *
1983  * If it is currently being processed, obviously not.
1984  * A forced Ping or Push is also handled by the discovery thread.
1985  *
1986  * Otherwise look at whether the peer needs rediscovering.
1987  */
1988 bool
1989 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1990 __must_hold(&lp->lp_lock)
1991 {
1992         bool rc;
1993
1994         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1995                             LNET_PEER_FORCE_PING |
1996                             LNET_PEER_FORCE_PUSH)) {
1997                 rc = false;
1998         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1999                 rc = false;
2000         } else if (lnet_peer_needs_push(lp)) {
2001                 rc = false;
2002         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2003                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2004                         rc = true;
2005                 else
2006                         rc = false;
2007         } else {
2008                 rc = false;
2009         }
2010
2011         return rc;
2012 }
2013
2014 /*
2015  * Queue a peer for the attention of the discovery thread.  Call with
2016  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2017  * -EALREADY if the peer was already queued.
2018  */
2019 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2020 {
2021         int rc;
2022
2023         spin_lock(&lp->lp_lock);
2024         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2025                 lp->lp_state |= LNET_PEER_DISCOVERING;
2026         spin_unlock(&lp->lp_lock);
2027         if (list_empty(&lp->lp_dc_list)) {
2028                 lnet_peer_addref_locked(lp);
2029                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2030                 wake_up(&the_lnet.ln_dc_waitq);
2031                 rc = 0;
2032         } else {
2033                 rc = -EALREADY;
2034         }
2035
2036         CDEBUG(D_NET, "Queue peer %s: %d\n",
2037                libcfs_nid2str(lp->lp_primary_nid), rc);
2038
2039         return rc;
2040 }
2041
2042 /*
2043  * Discovery of a peer is complete. Wake all waiters on the peer.
2044  * Call with lnet_net_lock/EX held.
2045  */
2046 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
2047 {
2048         struct lnet_msg *msg, *tmp;
2049         int rc = 0;
2050         LIST_HEAD(pending_msgs);
2051
2052         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2053                libcfs_nid2str(lp->lp_primary_nid));
2054
2055         list_del_init(&lp->lp_dc_list);
2056         spin_lock(&lp->lp_lock);
2057         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2058         spin_unlock(&lp->lp_lock);
2059         wake_up_all(&lp->lp_dc_waitq);
2060
2061         if (lp->lp_rtr_refcount > 0)
2062                 lnet_router_discovery_complete(lp);
2063
2064         lnet_net_unlock(LNET_LOCK_EX);
2065
2066         /* iterate through all pending messages and send them again */
2067         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2068                 list_del_init(&msg->msg_list);
2069                 if (lp->lp_dc_error) {
2070                         lnet_finalize(msg, lp->lp_dc_error);
2071                         continue;
2072                 }
2073
2074                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2075                        lnet_msgtyp2str(msg->msg_type),
2076                        libcfs_id2str(msg->msg_target));
2077                 rc = lnet_send(msg->msg_src_nid_param, msg,
2078                                msg->msg_rtr_nid_param);
2079                 if (rc < 0) {
2080                         CNETERR("Error sending %s to %s: %d\n",
2081                                lnet_msgtyp2str(msg->msg_type),
2082                                libcfs_id2str(msg->msg_target), rc);
2083                         lnet_finalize(msg, rc);
2084                 }
2085         }
2086         lnet_net_lock(LNET_LOCK_EX);
2087         lnet_peer_decref_locked(lp);
2088 }
2089
2090 /*
2091  * Handle inbound push.
2092  * Like any event handler, called with lnet_res_lock/CPT held.
2093  */
2094 void lnet_peer_push_event(struct lnet_event *ev)
2095 {
2096         struct lnet_ping_buffer *pbuf;
2097         struct lnet_peer *lp;
2098
2099         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2100
2101         /* lnet_find_peer() adds a refcount */
2102         lp = lnet_find_peer(ev->source.nid);
2103         if (!lp) {
2104                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2105                        libcfs_nid2str(ev->initiator.nid),
2106                        libcfs_nid2str(ev->source.nid));
2107                 pbuf->pb_needs_post = true;
2108                 return;
2109         }
2110
2111         /* Ensure peer state remains consistent while we modify it. */
2112         spin_lock(&lp->lp_lock);
2113
2114         /*
2115          * If some kind of error happened the contents of the message
2116          * cannot be used. Clear the NIDS_UPTODATE and set the
2117          * FORCE_PING flag to trigger a ping.
2118          */
2119         if (ev->status) {
2120                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2121                 lp->lp_state |= LNET_PEER_FORCE_PING;
2122                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2123                        ev->status,
2124                        libcfs_nid2str(lp->lp_primary_nid),
2125                        libcfs_nid2str(ev->source.nid));
2126                 goto out;
2127         }
2128
2129         /*
2130          * A push with invalid or corrupted info. Clear the UPTODATE
2131          * flag to trigger a ping.
2132          */
2133         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2134                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2135                 lp->lp_state |= LNET_PEER_FORCE_PING;
2136                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2137                        libcfs_nid2str(lp->lp_primary_nid));
2138                 goto out;
2139         }
2140
2141         /*
2142          * Make sure we'll allocate the correct size ping buffer when
2143          * pinging the peer.
2144          */
2145         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2146                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2147
2148         /*
2149          * A non-Multi-Rail peer is not supposed to be capable of
2150          * sending a push.
2151          */
2152         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2153                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2154                        libcfs_nid2str(lp->lp_primary_nid));
2155                 goto out;
2156         }
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                 /*
2166                  * Mark the peer for deletion if we already know about it
2167                  * and it's going from discovery set to no discovery set
2168                  */
2169                 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2170                                       LNET_PEER_DISCOVERING)) &&
2171                      lp->lp_state & LNET_PEER_DISCOVERED) {
2172                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2173                                libcfs_nid2str(lp->lp_primary_nid),
2174                                lp->lp_state);
2175                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2176                 }
2177                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2178         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2179                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2180                        libcfs_nid2str(lp->lp_primary_nid));
2181                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2182         }
2183
2184         /*
2185          * Update the MULTI_RAIL flag based on the push. If the peer
2186          * was configured with DLC then the setting should match what
2187          * DLC put in.
2188          * NB: We verified above that the MR feature bit is set in pi_features
2189          */
2190         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2191                 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2192                        libcfs_nid2str(lp->lp_primary_nid), lp);
2193         } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2194                 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2195                       libcfs_nid2str(lp->lp_primary_nid));
2196         } else if (lnet_peer_discovery_disabled) {
2197                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2198                        libcfs_nid2str(lp->lp_primary_nid), lp);
2199         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2200                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2201                        libcfs_nid2str(lp->lp_primary_nid), lp);
2202         } else {
2203                 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2204                        libcfs_nid2str(lp->lp_primary_nid), lp);
2205                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2206                 lnet_peer_clr_non_mr_pref_nids(lp);
2207         }
2208
2209         /*
2210          * Check for truncation of the Put message. Clear the
2211          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2212          * and tell discovery to allocate a bigger buffer.
2213          */
2214         if (ev->mlength < ev->rlength) {
2215                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2216                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2217                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2218                 lp->lp_state |= LNET_PEER_FORCE_PING;
2219                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2220                        libcfs_nid2str(lp->lp_primary_nid),
2221                        pbuf->pb_info.pi_nnis);
2222                 goto out;
2223         }
2224
2225         /* always assume new data */
2226         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2227         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2228
2229         /*
2230          * If there is data present that hasn't been processed yet,
2231          * we'll replace it if the Put contained newer data and it
2232          * fits. We're racing with a Ping or earlier Push in this
2233          * case.
2234          */
2235         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2236                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2237                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2238                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2239                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2240                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2241                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2242                               libcfs_nid2str(lp->lp_primary_nid),
2243                               LNET_PING_BUFFER_SEQNO(pbuf),
2244                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2245                 }
2246                 goto out;
2247         }
2248
2249         /*
2250          * Allocate a buffer to copy the data. On a failure we drop
2251          * the Push and set FORCE_PING to force the discovery
2252          * thread to fix the problem by pinging the peer.
2253          */
2254         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2255         if (!lp->lp_data) {
2256                 lp->lp_state |= LNET_PEER_FORCE_PING;
2257                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2258                        libcfs_nid2str(lp->lp_primary_nid),
2259                        LNET_PING_BUFFER_SEQNO(pbuf));
2260                 goto out;
2261         }
2262
2263         /* Success */
2264         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2265                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2266         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2267         CDEBUG(D_NET, "Received Push %s %u\n",
2268                libcfs_nid2str(lp->lp_primary_nid),
2269                LNET_PING_BUFFER_SEQNO(pbuf));
2270
2271 out:
2272         /* We've processed this buffer. It can be reposted */
2273         pbuf->pb_needs_post = true;
2274
2275         /*
2276          * Queue the peer for discovery if not done, force it on the request
2277          * queue and wake the discovery thread if the peer was already queued,
2278          * because its status changed.
2279          */
2280         spin_unlock(&lp->lp_lock);
2281         lnet_net_lock(LNET_LOCK_EX);
2282         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2283                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2284                 wake_up(&the_lnet.ln_dc_waitq);
2285         }
2286         /* Drop refcount from lookup */
2287         lnet_peer_decref_locked(lp);
2288         lnet_net_unlock(LNET_LOCK_EX);
2289 }
2290
2291 /*
2292  * Clear the discovery error state, unless we're already discovering
2293  * this peer, in which case the error is current.
2294  */
2295 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2296 {
2297         spin_lock(&lp->lp_lock);
2298         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2299                 lp->lp_dc_error = 0;
2300         spin_unlock(&lp->lp_lock);
2301 }
2302
2303 /*
2304  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2305  * dropped/retaken within this function. An lnet_peer_ni is passed in
2306  * because discovery could tear down an lnet_peer.
2307  */
2308 int
2309 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2310 {
2311         DEFINE_WAIT(wait);
2312         struct lnet_peer *lp;
2313         int rc = 0;
2314         int count = 0;
2315
2316 again:
2317         lnet_net_unlock(cpt);
2318         lnet_net_lock(LNET_LOCK_EX);
2319         lp = lpni->lpni_peer_net->lpn_peer;
2320         lnet_peer_clear_discovery_error(lp);
2321
2322         /*
2323          * We're willing to be interrupted. The lpni can become a
2324          * zombie if we race with DLC, so we must check for that.
2325          */
2326         for (;;) {
2327                 /* Keep lp alive when the lnet_net_lock is unlocked */
2328                 lnet_peer_addref_locked(lp);
2329                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2330                 if (signal_pending(current))
2331                         break;
2332                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2333                         break;
2334                 /*
2335                  * Don't repeat discovery if discovery is disabled. This is
2336                  * done to ensure we can use discovery as a standard ping as
2337                  * well for backwards compatibility with routers which do not
2338                  * have discovery or have discovery disabled
2339                  */
2340                 if (lnet_is_discovery_disabled(lp) && count > 0)
2341                         break;
2342                 if (lp->lp_dc_error)
2343                         break;
2344                 if (lnet_peer_is_uptodate(lp))
2345                         break;
2346                 lnet_peer_queue_for_discovery(lp);
2347                 count++;
2348                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2349
2350                 /*
2351                  * If caller requested a non-blocking operation then
2352                  * return immediately. Once discovery is complete any
2353                  * pending messages that were stopped due to discovery
2354                  * will be transmitted.
2355                  */
2356                 if (!block)
2357                         break;
2358
2359                 lnet_net_unlock(LNET_LOCK_EX);
2360                 schedule();
2361                 finish_wait(&lp->lp_dc_waitq, &wait);
2362                 lnet_net_lock(LNET_LOCK_EX);
2363                 lnet_peer_decref_locked(lp);
2364                 /* Peer may have changed */
2365                 lp = lpni->lpni_peer_net->lpn_peer;
2366         }
2367         finish_wait(&lp->lp_dc_waitq, &wait);
2368
2369         lnet_net_unlock(LNET_LOCK_EX);
2370         lnet_net_lock(cpt);
2371         lnet_peer_decref_locked(lp);
2372         /*
2373          * The peer may have changed, so re-check and rediscover if that turns
2374          * out to have been the case. The reference count on lp ensured that
2375          * even if it was unlinked from lpni the memory could not be recycled.
2376          * Thus the check below is sufficient to determine whether the peer
2377          * changed. If the peer changed, then lp must not be dereferenced.
2378          */
2379         if (lp != lpni->lpni_peer_net->lpn_peer)
2380                 goto again;
2381
2382         if (signal_pending(current))
2383                 rc = -EINTR;
2384         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2385                 rc = -ESHUTDOWN;
2386         else if (lp->lp_dc_error)
2387                 rc = lp->lp_dc_error;
2388         else if (!block)
2389                 CDEBUG(D_NET, "non-blocking discovery\n");
2390         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2391                 goto again;
2392
2393         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2394                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2395                libcfs_nid2str(lpni->lpni_nid), rc,
2396                (!block) ? "pending discovery" : "discovery complete");
2397
2398         return rc;
2399 }
2400
2401 /* Handle an incoming ack for a push. */
2402 static void
2403 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2404 {
2405         struct lnet_ping_buffer *pbuf;
2406
2407         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2408         spin_lock(&lp->lp_lock);
2409         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2410         lp->lp_push_error = ev->status;
2411         if (ev->status)
2412                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2413         else
2414                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2415         spin_unlock(&lp->lp_lock);
2416
2417         CDEBUG(D_NET, "peer %s ev->status %d\n",
2418                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2419 }
2420
2421 /* Handle a Reply message. This is the reply to a Ping message. */
2422 static void
2423 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2424 {
2425         struct lnet_ping_buffer *pbuf;
2426         int rc;
2427
2428         spin_lock(&lp->lp_lock);
2429
2430         lp->lp_disc_src_nid = ev->target.nid;
2431
2432         /*
2433          * If some kind of error happened the contents of message
2434          * cannot be used. Set PING_FAILED to trigger a retry.
2435          */
2436         if (ev->status) {
2437                 lp->lp_state |= LNET_PEER_PING_FAILED;
2438                 lp->lp_ping_error = ev->status;
2439                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2440                        ev->status,
2441                        libcfs_nid2str(lp->lp_primary_nid),
2442                        libcfs_nid2str(ev->source.nid));
2443                 goto out;
2444         }
2445
2446         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2447         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2448                 lnet_swap_pinginfo(pbuf);
2449
2450         /*
2451          * A reply with invalid or corrupted info. Set PING_FAILED to
2452          * trigger a retry.
2453          */
2454         rc = lnet_ping_info_validate(&pbuf->pb_info);
2455         if (rc) {
2456                 lp->lp_state |= LNET_PEER_PING_FAILED;
2457                 lp->lp_ping_error = 0;
2458                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2459                        libcfs_nid2str(lp->lp_primary_nid), rc);
2460                 goto out;
2461         }
2462
2463
2464         /*
2465          * The peer may have discovery disabled at its end. Set
2466          * NO_DISCOVERY as appropriate.
2467          */
2468         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2469             !lnet_peer_discovery_disabled) {
2470                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2471                        libcfs_nid2str(lp->lp_primary_nid));
2472                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2473         } else {
2474                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2475                        libcfs_nid2str(lp->lp_primary_nid));
2476                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2477         }
2478
2479         /*
2480          * Update the MULTI_RAIL flag based on the reply. If the peer
2481          * was configured with DLC then the setting should match what
2482          * DLC put in.
2483          */
2484         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2485                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2486                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2487                                libcfs_nid2str(lp->lp_primary_nid), lp);
2488                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2489                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2490                               libcfs_nid2str(lp->lp_primary_nid));
2491                 } else if (lnet_peer_discovery_disabled) {
2492                         CDEBUG(D_NET,
2493                                "peer %s(%p) not MR: DD disabled locally\n",
2494                                libcfs_nid2str(lp->lp_primary_nid), lp);
2495                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2496                         CDEBUG(D_NET,
2497                                "peer %s(%p) not MR: DD disabled remotely\n",
2498                                libcfs_nid2str(lp->lp_primary_nid), lp);
2499                 } else {
2500                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2501                                libcfs_nid2str(lp->lp_primary_nid), lp);
2502                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2503                         lnet_peer_clr_non_mr_pref_nids(lp);
2504                 }
2505         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2506                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2507                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2508                               libcfs_nid2str(lp->lp_primary_nid));
2509                 } else {
2510                         CERROR("Multi-Rail state vanished from %s\n",
2511                                libcfs_nid2str(lp->lp_primary_nid));
2512                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2513                 }
2514         }
2515
2516         /*
2517          * Make sure we'll allocate the correct size ping buffer when
2518          * pinging the peer.
2519          */
2520         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2521                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2522
2523         /*
2524          * Check for truncation of the Reply. Clear PING_SENT and set
2525          * PING_FAILED to trigger a retry.
2526          */
2527         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2528                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2529                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2530                 lp->lp_state |= LNET_PEER_PING_FAILED;
2531                 lp->lp_ping_error = 0;
2532                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2533                        libcfs_nid2str(lp->lp_primary_nid),
2534                        pbuf->pb_info.pi_nnis);
2535                 goto out;
2536         }
2537
2538         /*
2539          * Check the sequence numbers in the reply. These are only
2540          * available if the reply came from a Multi-Rail peer.
2541          */
2542         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2543             pbuf->pb_info.pi_nnis > 1 &&
2544             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2545                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2546                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2547                                 libcfs_nid2str(lp->lp_primary_nid),
2548                                 LNET_PING_BUFFER_SEQNO(pbuf),
2549                                 lp->lp_peer_seqno);
2550
2551                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2552         }
2553
2554         /* We're happy with the state of the data in the buffer. */
2555         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2556                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2557         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2558                 lnet_ping_buffer_decref(lp->lp_data);
2559         else
2560                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2561         lnet_ping_buffer_addref(pbuf);
2562         lp->lp_data = pbuf;
2563 out:
2564         lp->lp_state &= ~LNET_PEER_PING_SENT;
2565         spin_unlock(&lp->lp_lock);
2566
2567         lnet_net_lock(LNET_LOCK_EX);
2568         /*
2569          * If this peer is a gateway, call the routing callback to
2570          * handle the ping reply
2571          */
2572         if (lp->lp_rtr_refcount > 0)
2573                 lnet_router_discovery_ping_reply(lp);
2574         lnet_net_unlock(LNET_LOCK_EX);
2575 }
2576
2577 /*
2578  * Send event handling. Only matters for error cases, where we clean
2579  * up state on the peer and peer_ni that would otherwise be updated in
2580  * the REPLY event handler for a successful Ping, and the ACK event
2581  * handler for a successful Push.
2582  */
2583 static int
2584 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2585 {
2586         int rc = 0;
2587
2588         if (!ev->status)
2589                 goto out;
2590
2591         spin_lock(&lp->lp_lock);
2592         if (ev->msg_type == LNET_MSG_GET) {
2593                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2594                 lp->lp_state |= LNET_PEER_PING_FAILED;
2595                 lp->lp_ping_error = ev->status;
2596         } else { /* ev->msg_type == LNET_MSG_PUT */
2597                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2598                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2599                 lp->lp_push_error = ev->status;
2600         }
2601         spin_unlock(&lp->lp_lock);
2602         rc = LNET_REDISCOVER_PEER;
2603 out:
2604         CDEBUG(D_NET, "%s Send to %s: %d\n",
2605                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2606                 libcfs_nid2str(ev->target.nid), rc);
2607         return rc;
2608 }
2609
2610 /*
2611  * Unlink event handling. This event is only seen if a call to
2612  * LNetMDUnlink() caused the event to be unlinked. If this call was
2613  * made after the event was set up in LNetGet() or LNetPut() then we
2614  * assume the Ping or Push timed out.
2615  */
2616 static void
2617 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2618 {
2619         spin_lock(&lp->lp_lock);
2620         /* We've passed through LNetGet() */
2621         if (lp->lp_state & LNET_PEER_PING_SENT) {
2622                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2623                 lp->lp_state |= LNET_PEER_PING_FAILED;
2624                 lp->lp_ping_error = -ETIMEDOUT;
2625                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2626                         libcfs_nid2str(lp->lp_primary_nid));
2627         }
2628         /* We've passed through LNetPut() */
2629         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2630                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2631                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2632                 lp->lp_push_error = -ETIMEDOUT;
2633                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2634                         libcfs_nid2str(lp->lp_primary_nid));
2635         }
2636         spin_unlock(&lp->lp_lock);
2637 }
2638
2639 /*
2640  * Event handler for the discovery EQ.
2641  *
2642  * Called with lnet_res_lock(cpt) held. The cpt is the
2643  * lnet_cpt_of_cookie() of the md handle cookie.
2644  */
2645 static void lnet_discovery_event_handler(struct lnet_event *event)
2646 {
2647         struct lnet_peer *lp = event->md_user_ptr;
2648         struct lnet_ping_buffer *pbuf;
2649         int rc;
2650
2651         /* discovery needs to take another look */
2652         rc = LNET_REDISCOVER_PEER;
2653
2654         CDEBUG(D_NET, "Received event: %d\n", event->type);
2655
2656         switch (event->type) {
2657         case LNET_EVENT_ACK:
2658                 lnet_discovery_event_ack(lp, event);
2659                 break;
2660         case LNET_EVENT_REPLY:
2661                 lnet_discovery_event_reply(lp, event);
2662                 break;
2663         case LNET_EVENT_SEND:
2664                 /* Only send failure triggers a retry. */
2665                 rc = lnet_discovery_event_send(lp, event);
2666                 break;
2667         case LNET_EVENT_UNLINK:
2668                 /* LNetMDUnlink() was called */
2669                 lnet_discovery_event_unlink(lp, event);
2670                 break;
2671         default:
2672                 /* Invalid events. */
2673                 LBUG();
2674         }
2675         lnet_net_lock(LNET_LOCK_EX);
2676         if (event->unlinked) {
2677                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2678                 lnet_ping_buffer_decref(pbuf);
2679                 lnet_peer_decref_locked(lp);
2680         }
2681
2682         /* put peer back at end of request queue, if discovery not already
2683          * done */
2684         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2685                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2686                 wake_up(&the_lnet.ln_dc_waitq);
2687         }
2688         lnet_net_unlock(LNET_LOCK_EX);
2689 }
2690
2691 /*
2692  * Build a peer from incoming data.
2693  *
2694  * The NIDs in the incoming data are supposed to be structured as follows:
2695  *  - loopback
2696  *  - primary NID
2697  *  - other NIDs in same net
2698  *  - NIDs in second net
2699  *  - NIDs in third net
2700  *  - ...
2701  * This due to the way the list of NIDs in the data is created.
2702  *
2703  * Note that this function will mark the peer uptodate unless an
2704  * ENOMEM is encontered. All other errors are due to a conflict
2705  * between the DLC configuration and what discovery sees. We treat DLC
2706  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2707  * peer from becoming stuck in discovery.
2708  */
2709 static int lnet_peer_merge_data(struct lnet_peer *lp,
2710                                 struct lnet_ping_buffer *pbuf)
2711 {
2712         struct lnet_peer_ni *lpni;
2713         lnet_nid_t *curnis = NULL;
2714         struct lnet_ni_status *addnis = NULL;
2715         lnet_nid_t *delnis = NULL;
2716         unsigned flags;
2717         int ncurnis;
2718         int naddnis;
2719         int ndelnis;
2720         int nnis = 0;
2721         int i;
2722         int j;
2723         int rc;
2724
2725         flags = LNET_PEER_DISCOVERED;
2726         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2727                 flags |= LNET_PEER_MULTI_RAIL;
2728
2729         /*
2730          * Cache the routing feature for the peer; whether it is enabled
2731          * for disabled as reported by the remote peer.
2732          */
2733         spin_lock(&lp->lp_lock);
2734         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2735                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2736         else
2737                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2738         spin_unlock(&lp->lp_lock);
2739
2740         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2741         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2742         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2743         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2744         if (!curnis || !addnis || !delnis) {
2745                 rc = -ENOMEM;
2746                 goto out;
2747         }
2748         ncurnis = 0;
2749         naddnis = 0;
2750         ndelnis = 0;
2751
2752         /* Construct the list of NIDs present in peer. */
2753         lpni = NULL;
2754         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2755                 curnis[ncurnis++] = lpni->lpni_nid;
2756
2757         /*
2758          * Check for NIDs in pbuf not present in curnis[].
2759          * The loop starts at 1 to skip the loopback NID.
2760          */
2761         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2762                 for (j = 0; j < ncurnis; j++)
2763                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2764                                 break;
2765                 if (j == ncurnis)
2766                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2767         }
2768         /*
2769          * Check for NIDs in curnis[] not present in pbuf.
2770          * The nested loop starts at 1 to skip the loopback NID.
2771          *
2772          * But never add the loopback NID to delnis[]: if it is
2773          * present in curnis[] then this peer is for this node.
2774          */
2775         for (i = 0; i < ncurnis; i++) {
2776                 if (curnis[i] == LNET_NID_LO_0)
2777                         continue;
2778                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2779                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2780                                 /*
2781                                  * update the information we cache for the
2782                                  * peer with the latest information we
2783                                  * received
2784                                  */
2785                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2786                                 if (lpni) {
2787                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2788                                         lnet_peer_ni_decref_locked(lpni);
2789                                 }
2790                                 break;
2791                         }
2792                 }
2793                 if (j == pbuf->pb_info.pi_nnis)
2794                         delnis[ndelnis++] = curnis[i];
2795         }
2796
2797         /*
2798          * If we get here and the discovery is disabled then we don't want
2799          * to add or delete any NIs. We just updated the ones we have some
2800          * information on, and call it a day
2801          */
2802         rc = 0;
2803         if (lnet_is_discovery_disabled(lp))
2804                 goto out;
2805
2806         for (i = 0; i < naddnis; i++) {
2807                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2808                 if (rc) {
2809                         CERROR("Error adding NID %s to peer %s: %d\n",
2810                                libcfs_nid2str(addnis[i].ns_nid),
2811                                libcfs_nid2str(lp->lp_primary_nid), rc);
2812                         if (rc == -ENOMEM)
2813                                 goto out;
2814                 }
2815                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2816                 if (lpni) {
2817                         lpni->lpni_ns_status = addnis[i].ns_status;
2818                         lnet_peer_ni_decref_locked(lpni);
2819                 }
2820         }
2821
2822         for (i = 0; i < ndelnis; i++) {
2823                 /*
2824                  * for routers it's okay to delete the primary_nid because
2825                  * the upper layers don't really rely on it. So if we're
2826                  * being told that the router changed its primary_nid
2827                  * then it's okay to delete it.
2828                  */
2829                 if (lp->lp_rtr_refcount > 0)
2830                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2831                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2832                 if (rc) {
2833                         CERROR("Error deleting NID %s from peer %s: %d\n",
2834                                libcfs_nid2str(delnis[i]),
2835                                libcfs_nid2str(lp->lp_primary_nid), rc);
2836                         if (rc == -ENOMEM)
2837                                 goto out;
2838                 }
2839         }
2840         /*
2841          * Errors other than -ENOMEM are due to peers having been
2842          * configured with DLC. Ignore these because DLC overrides
2843          * Discovery.
2844          */
2845         rc = 0;
2846 out:
2847         CFS_FREE_PTR_ARRAY(curnis, nnis);
2848         CFS_FREE_PTR_ARRAY(addnis, nnis);
2849         CFS_FREE_PTR_ARRAY(delnis, nnis);
2850         lnet_ping_buffer_decref(pbuf);
2851         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2852
2853         if (rc) {
2854                 spin_lock(&lp->lp_lock);
2855                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2856                 lp->lp_state |= LNET_PEER_FORCE_PING;
2857                 spin_unlock(&lp->lp_lock);
2858         }
2859         return rc;
2860 }
2861
2862 /*
2863  * The data in pbuf says lp is its primary peer, but the data was
2864  * received by a different peer. Try to update lp with the data.
2865  */
2866 static int
2867 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2868 {
2869         struct lnet_handle_md mdh;
2870
2871         /* Queue lp for discovery, and force it on the request queue. */
2872         lnet_net_lock(LNET_LOCK_EX);
2873         if (lnet_peer_queue_for_discovery(lp))
2874                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2875         lnet_net_unlock(LNET_LOCK_EX);
2876
2877         LNetInvalidateMDHandle(&mdh);
2878
2879         /*
2880          * Decide whether we can move the peer to the DATA_PRESENT state.
2881          *
2882          * We replace stale data for a multi-rail peer, repair PING_FAILED
2883          * status, and preempt FORCE_PING.
2884          *
2885          * If after that we have DATA_PRESENT, we merge it into this peer.
2886          */
2887         spin_lock(&lp->lp_lock);
2888         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2889                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2890                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2891                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2892                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2893                         lnet_ping_buffer_decref(pbuf);
2894                         pbuf = lp->lp_data;
2895                         lp->lp_data = NULL;
2896                 }
2897         }
2898         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2899                 lnet_ping_buffer_decref(lp->lp_data);
2900                 lp->lp_data = NULL;
2901                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2902         }
2903         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2904                 mdh = lp->lp_ping_mdh;
2905                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2906                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2907                 lp->lp_ping_error = 0;
2908         }
2909         if (lp->lp_state & LNET_PEER_FORCE_PING)
2910                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2911         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2912         spin_unlock(&lp->lp_lock);
2913
2914         if (!LNetMDHandleIsInvalid(mdh))
2915                 LNetMDUnlink(mdh);
2916
2917         if (pbuf)
2918                 return lnet_peer_merge_data(lp, pbuf);
2919
2920         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2921         return 0;
2922 }
2923
2924 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2925 {
2926         int i;
2927
2928         for (i = 0; i < pinfo->pi_nnis; i++) {
2929                 if (pinfo->pi_ni[i].ns_nid == nid)
2930                         return true;
2931         }
2932
2933         return false;
2934 }
2935
2936 /*
2937  * Update a peer using the data received.
2938  */
2939 static int lnet_peer_data_present(struct lnet_peer *lp)
2940 __must_hold(&lp->lp_lock)
2941 {
2942         struct lnet_ping_buffer *pbuf;
2943         struct lnet_peer_ni *lpni;
2944         lnet_nid_t nid = LNET_NID_ANY;
2945         unsigned flags;
2946         int rc = 0;
2947
2948         pbuf = lp->lp_data;
2949         lp->lp_data = NULL;
2950         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2951         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2952         spin_unlock(&lp->lp_lock);
2953
2954         /*
2955          * Modifications of peer structures are done while holding the
2956          * ln_api_mutex. A global lock is required because we may be
2957          * modifying multiple peer structures, and a mutex greatly
2958          * simplifies memory management.
2959          *
2960          * The actual changes to the data structures must also protect
2961          * against concurrent lookups, for which the lnet_net_lock in
2962          * LNET_LOCK_EX mode is used.
2963          */
2964         mutex_lock(&the_lnet.ln_api_mutex);
2965         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2966                 rc = -ESHUTDOWN;
2967                 goto out;
2968         }
2969
2970         /*
2971          * If this peer is not on the peer list then it is being torn
2972          * down, and our reference count may be all that is keeping it
2973          * alive. Don't do any work on it.
2974          */
2975         if (list_empty(&lp->lp_peer_list))
2976                 goto out;
2977
2978         flags = LNET_PEER_DISCOVERED;
2979         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2980                 flags |= LNET_PEER_MULTI_RAIL;
2981
2982         /*
2983          * Check whether the primary NID in the message matches the
2984          * primary NID of the peer. If it does, update the peer, if
2985          * it it does not, check whether there is already a peer with
2986          * that primary NID. If no such peer exists, try to update
2987          * the primary NID of the current peer (allowed if it was
2988          * created due to message traffic) and complete the update.
2989          * If the peer did exist, hand off the data to it.
2990          *
2991          * The peer for the loopback interface is a special case: this
2992          * is the peer for the local node, and we want to set its
2993          * primary NID to the correct value here. Moreover, this peer
2994          * can show up with only the loopback NID in the ping buffer.
2995          */
2996         if (pbuf->pb_info.pi_nnis <= 1)
2997                 goto out;
2998         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2999         if (lp->lp_primary_nid == LNET_NID_LO_0) {
3000                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3001                 if (!rc)
3002                         rc = lnet_peer_merge_data(lp, pbuf);
3003         /*
3004          * if the primary nid of the peer is present in the ping info returned
3005          * from the peer, but it's not the local primary peer we have
3006          * cached and discovery is disabled, then we don't want to update
3007          * our local peer info, by adding or removing NIDs, we just want
3008          * to update the status of the nids that we currently have
3009          * recorded in that peer.
3010          */
3011         } else if (lp->lp_primary_nid == nid ||
3012                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
3013                     lnet_is_discovery_disabled(lp))) {
3014                 rc = lnet_peer_merge_data(lp, pbuf);
3015         } else {
3016                 lpni = lnet_find_peer_ni_locked(nid);
3017                 if (!lpni) {
3018                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
3019                         if (rc) {
3020                                 CERROR("Primary NID error %s versus %s: %d\n",
3021                                        libcfs_nid2str(lp->lp_primary_nid),
3022                                        libcfs_nid2str(nid), rc);
3023                         } else {
3024                                 rc = lnet_peer_merge_data(lp, pbuf);
3025                         }
3026                 } else {
3027                         struct lnet_peer *new_lp;
3028                         new_lp = lpni->lpni_peer_net->lpn_peer;
3029                         /*
3030                          * if lp has discovery/MR enabled that means new_lp
3031                          * should have discovery/MR enabled as well, since
3032                          * it's the same peer, which we're about to merge
3033                          */
3034                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3035                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3036                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3037                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3038
3039                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
3040                         lnet_consolidate_routes_locked(lp, new_lp);
3041                         lnet_peer_ni_decref_locked(lpni);
3042                 }
3043         }
3044 out:
3045         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
3046                lp->lp_state);
3047         mutex_unlock(&the_lnet.ln_api_mutex);
3048
3049         spin_lock(&lp->lp_lock);
3050         /* Tell discovery to re-check the peer immediately. */
3051         if (!rc)
3052                 rc = LNET_REDISCOVER_PEER;
3053         return rc;
3054 }
3055
3056 /*
3057  * A ping failed. Clear the PING_FAILED state and set the
3058  * FORCE_PING state, to ensure a retry even if discovery is
3059  * disabled. This avoids being left with incorrect state.
3060  */
3061 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3062 __must_hold(&lp->lp_lock)
3063 {
3064         struct lnet_handle_md mdh;
3065         int rc;
3066
3067         mdh = lp->lp_ping_mdh;
3068         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3069         lp->lp_state &= ~LNET_PEER_PING_FAILED;
3070         lp->lp_state |= LNET_PEER_FORCE_PING;
3071         rc = lp->lp_ping_error;
3072         lp->lp_ping_error = 0;
3073         spin_unlock(&lp->lp_lock);
3074
3075         if (!LNetMDHandleIsInvalid(mdh))
3076                 LNetMDUnlink(mdh);
3077
3078         CDEBUG(D_NET, "peer %s:%d\n",
3079                libcfs_nid2str(lp->lp_primary_nid), rc);
3080
3081         spin_lock(&lp->lp_lock);
3082         return rc ? rc : LNET_REDISCOVER_PEER;
3083 }
3084
3085 /*
3086  * Select NID to send a Ping or Push to.
3087  */
3088 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
3089 {
3090         struct lnet_peer_ni *lpni;
3091
3092         /* Look for a direct-connected NID for this peer. */
3093         lpni = NULL;
3094         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3095                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
3096                         continue;
3097                 break;
3098         }
3099         if (lpni)
3100                 return lpni->lpni_nid;
3101
3102         /* Look for a routed-connected NID for this peer. */
3103         lpni = NULL;
3104         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3105                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
3106                         continue;
3107                 break;
3108         }
3109         if (lpni)
3110                 return lpni->lpni_nid;
3111
3112         return LNET_NID_ANY;
3113 }
3114
3115 /* Active side of ping. */
3116 static int lnet_peer_send_ping(struct lnet_peer *lp)
3117 __must_hold(&lp->lp_lock)
3118 {
3119         lnet_nid_t pnid;
3120         int nnis;
3121         int rc;
3122         int cpt;
3123
3124         lp->lp_state |= LNET_PEER_PING_SENT;
3125         lp->lp_state &= ~LNET_PEER_FORCE_PING;
3126         spin_unlock(&lp->lp_lock);
3127
3128         cpt = lnet_net_lock_current();
3129         /* Refcount for MD. */
3130         lnet_peer_addref_locked(lp);
3131         pnid = lnet_peer_select_nid(lp);
3132         lnet_net_unlock(cpt);
3133
3134         nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3135
3136         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
3137                             the_lnet.ln_dc_handler, false);
3138
3139         /*
3140          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3141          * refcount on the peer, otherwise LNetMDUnlink will be called
3142          * which will eventually do that.
3143          */
3144         if (rc > 0) {
3145                 lnet_net_lock(cpt);
3146                 lnet_peer_decref_locked(lp);
3147                 lnet_net_unlock(cpt);
3148                 rc = -rc; /* change the rc to negative value */
3149                 goto fail_error;
3150         } else if (rc < 0) {
3151                 goto fail_error;
3152         }
3153
3154         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3155
3156         spin_lock(&lp->lp_lock);
3157         return 0;
3158
3159 fail_error:
3160         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3161         /*
3162          * The errors that get us here are considered hard errors and
3163          * cause Discovery to terminate. So we clear PING_SENT, but do
3164          * not set either PING_FAILED or FORCE_PING. In fact we need
3165          * to clear PING_FAILED, because the unlink event handler will
3166          * have set it if we called LNetMDUnlink() above.
3167          */
3168         spin_lock(&lp->lp_lock);
3169         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3170         return rc;
3171 }
3172
3173 /*
3174  * This function exists because you cannot call LNetMDUnlink() from an
3175  * event handler.
3176  */
3177 static int lnet_peer_push_failed(struct lnet_peer *lp)
3178 __must_hold(&lp->lp_lock)
3179 {
3180         struct lnet_handle_md mdh;
3181         int rc;
3182
3183         mdh = lp->lp_push_mdh;
3184         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3185         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3186         rc = lp->lp_push_error;
3187         lp->lp_push_error = 0;
3188         spin_unlock(&lp->lp_lock);
3189
3190         if (!LNetMDHandleIsInvalid(mdh))
3191                 LNetMDUnlink(mdh);
3192
3193         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3194         spin_lock(&lp->lp_lock);
3195         return rc ? rc : LNET_REDISCOVER_PEER;
3196 }
3197
3198 /*
3199  * Mark the peer as discovered.
3200  */
3201 static int lnet_peer_discovered(struct lnet_peer *lp)
3202 __must_hold(&lp->lp_lock)
3203 {
3204         lp->lp_state |= LNET_PEER_DISCOVERED;
3205         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3206                           LNET_PEER_REDISCOVER);
3207
3208         lp->lp_dc_error = 0;
3209
3210         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3211
3212         return 0;
3213 }
3214
3215 /* Active side of push. */
3216 static int lnet_peer_send_push(struct lnet_peer *lp)
3217 __must_hold(&lp->lp_lock)
3218 {
3219         struct lnet_ping_buffer *pbuf;
3220         struct lnet_process_id id;
3221         struct lnet_md md;
3222         int cpt;
3223         int rc;
3224
3225         /* Don't push to a non-multi-rail peer. */
3226         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3227                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3228                 /* if peer's NIDs are uptodate then peer is discovered */
3229                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3230                         rc = lnet_peer_discovered(lp);
3231                         return rc;
3232                 }
3233
3234                 return 0;
3235         }
3236
3237         lp->lp_state |= LNET_PEER_PUSH_SENT;
3238         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3239         spin_unlock(&lp->lp_lock);
3240
3241         cpt = lnet_net_lock_current();
3242         pbuf = the_lnet.ln_ping_target;
3243         lnet_ping_buffer_addref(pbuf);
3244         lnet_net_unlock(cpt);
3245
3246         /* Push source MD */
3247         md.start     = &pbuf->pb_info;
3248         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3249         md.threshold = 2; /* Put/Ack */
3250         md.max_size  = 0;
3251         md.options   = LNET_MD_TRACK_RESPONSE;
3252         md.handler   = the_lnet.ln_dc_handler;
3253         md.user_ptr  = lp;
3254
3255         rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3256         if (rc) {
3257                 lnet_ping_buffer_decref(pbuf);
3258                 CERROR("Can't bind push source MD: %d\n", rc);
3259                 goto fail_error;
3260         }
3261         cpt = lnet_net_lock_current();
3262         /* Refcount for MD. */
3263         lnet_peer_addref_locked(lp);
3264         id.pid = LNET_PID_LUSTRE;
3265         id.nid = lnet_peer_select_nid(lp);
3266         lnet_net_unlock(cpt);
3267
3268         if (id.nid == LNET_NID_ANY) {
3269                 rc = -EHOSTUNREACH;
3270                 goto fail_unlink;
3271         }
3272
3273         rc = LNetPut(lp->lp_disc_src_nid, lp->lp_push_mdh,
3274                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3275                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3276
3277         /*
3278          * reset the discovery nid. There is no need to restrict sending
3279          * from that source, if we call lnet_push_update_to_peers(). It'll
3280          * get set to a specific NID, if we initiate discovery from the
3281          * scratch
3282          */
3283         lp->lp_disc_src_nid = LNET_NID_ANY;
3284
3285         if (rc)
3286                 goto fail_unlink;
3287
3288         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3289
3290         spin_lock(&lp->lp_lock);
3291         return 0;
3292
3293 fail_unlink:
3294         LNetMDUnlink(lp->lp_push_mdh);
3295         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3296 fail_error:
3297         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3298         /*
3299          * The errors that get us here are considered hard errors and
3300          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3301          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3302          * because the unlink event handler will have set it if we
3303          * called LNetMDUnlink() above.
3304          */
3305         spin_lock(&lp->lp_lock);
3306         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3307         return rc;
3308 }
3309
3310 /*
3311  * An unrecoverable error was encountered during discovery.
3312  * Set error status in peer and abort discovery.
3313  */
3314 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3315 {
3316         CDEBUG(D_NET, "Discovery error %s: %d\n",
3317                libcfs_nid2str(lp->lp_primary_nid), error);
3318
3319         spin_lock(&lp->lp_lock);
3320         lp->lp_dc_error = error;
3321         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3322         lp->lp_state |= LNET_PEER_REDISCOVER;
3323         spin_unlock(&lp->lp_lock);
3324 }
3325
3326 /*
3327  * Discovering this peer is taking too long. Cancel any Ping or Push
3328  * that discovery is waiting on by unlinking the relevant MDs. The
3329  * lnet_discovery_event_handler() will proceed from here and complete
3330  * the cleanup.
3331  */
3332 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3333 {
3334         struct lnet_handle_md ping_mdh;
3335         struct lnet_handle_md push_mdh;
3336
3337         LNetInvalidateMDHandle(&ping_mdh);
3338         LNetInvalidateMDHandle(&push_mdh);
3339
3340         spin_lock(&lp->lp_lock);
3341         if (lp->lp_state & LNET_PEER_PING_SENT) {
3342                 ping_mdh = lp->lp_ping_mdh;
3343                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3344         }
3345         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3346                 push_mdh = lp->lp_push_mdh;
3347                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3348         }
3349         spin_unlock(&lp->lp_lock);
3350
3351         if (!LNetMDHandleIsInvalid(ping_mdh))
3352                 LNetMDUnlink(ping_mdh);
3353         if (!LNetMDHandleIsInvalid(push_mdh))
3354                 LNetMDUnlink(push_mdh);
3355 }
3356
3357 /*
3358  * Wait for work to be queued or some other change that must be
3359  * attended to. Returns non-zero if the discovery thread should shut
3360  * down.
3361  */
3362 static int lnet_peer_discovery_wait_for_work(void)
3363 {
3364         int cpt;
3365         int rc = 0;
3366
3367         DEFINE_WAIT(wait);
3368
3369         cpt = lnet_net_lock_current();
3370         for (;;) {
3371                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3372                                 TASK_INTERRUPTIBLE);
3373                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3374                         break;
3375                 if (lnet_push_target_resize_needed() ||
3376                     the_lnet.ln_push_target->pb_needs_post)
3377                         break;
3378                 if (!list_empty(&the_lnet.ln_dc_request))
3379                         break;
3380                 if (!list_empty(&the_lnet.ln_msg_resend))
3381                         break;
3382                 lnet_net_unlock(cpt);
3383
3384                 /*
3385                  * wakeup max every second to check if there are peers that
3386                  * have been stuck on the working queue for greater than
3387                  * the peer timeout.
3388                  */
3389                 schedule_timeout(cfs_time_seconds(1));
3390                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3391                 cpt = lnet_net_lock_current();
3392         }
3393         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3394
3395         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3396                 rc = -ESHUTDOWN;
3397
3398         lnet_net_unlock(cpt);
3399
3400         CDEBUG(D_NET, "woken: %d\n", rc);
3401
3402         return rc;
3403 }
3404
3405 /*
3406  * Messages that were pending on a destroyed peer will be put on a global
3407  * resend list. The message resend list will be checked by
3408  * the discovery thread when it wakes up, and will resend messages. These
3409  * messages can still be sendable in the case the lpni which was the initial
3410  * cause of the message re-queue was transfered to another peer.
3411  *
3412  * It is possible that LNet could be shutdown while we're iterating
3413  * through the list. lnet_shudown_lndnets() will attempt to access the
3414  * resend list, but will have to wait until the spinlock is released, by
3415  * which time there shouldn't be any more messages on the resend list.
3416  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3417  * for the messages so they can be released. The other case is that
3418  * lnet_shudown_lndnets() can finalize all the messages before this
3419  * function can visit the resend list, in which case this function will be
3420  * a no-op.
3421  */
3422 static void lnet_resend_msgs(void)
3423 {
3424         struct lnet_msg *msg, *tmp;
3425         LIST_HEAD(resend);
3426         int rc;
3427
3428         spin_lock(&the_lnet.ln_msg_resend_lock);
3429         list_splice(&the_lnet.ln_msg_resend, &resend);
3430         spin_unlock(&the_lnet.ln_msg_resend_lock);
3431
3432         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3433                 list_del_init(&msg->msg_list);
3434                 rc = lnet_send(msg->msg_src_nid_param, msg,
3435                                msg->msg_rtr_nid_param);
3436                 if (rc < 0) {
3437                         CNETERR("Error sending %s to %s: %d\n",
3438                                lnet_msgtyp2str(msg->msg_type),
3439                                libcfs_id2str(msg->msg_target), rc);
3440                         lnet_finalize(msg, rc);
3441                 }
3442         }
3443 }
3444
3445 /* The discovery thread. */
3446 static int lnet_peer_discovery(void *arg)
3447 {
3448         struct lnet_peer *lp;
3449         int rc;
3450
3451         wait_for_completion(&the_lnet.ln_started);
3452
3453         CDEBUG(D_NET, "started\n");
3454
3455         for (;;) {
3456                 if (lnet_peer_discovery_wait_for_work())
3457                         break;
3458
3459                 if (lnet_push_target_resize_needed())
3460                         lnet_push_target_resize();
3461                 else if (the_lnet.ln_push_target->pb_needs_post)
3462                         lnet_push_target_post(the_lnet.ln_push_target,
3463                                               &the_lnet.ln_push_target_md);
3464
3465                 lnet_resend_msgs();
3466
3467                 lnet_net_lock(LNET_LOCK_EX);
3468                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3469                         lnet_net_unlock(LNET_LOCK_EX);
3470                         break;
3471                 }
3472
3473                 /*
3474                  * Process all incoming discovery work requests.  When
3475                  * discovery must wait on a peer to change state, it
3476                  * is added to the tail of the ln_dc_working queue. A
3477                  * timestamp keeps track of when the peer was added,
3478                  * so we can time out discovery requests that take too
3479                  * long.
3480                  */
3481                 while (!list_empty(&the_lnet.ln_dc_request)) {
3482                         lp = list_first_entry(&the_lnet.ln_dc_request,
3483                                               struct lnet_peer, lp_dc_list);
3484                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3485                         /*
3486                          * set the time the peer was put on the dc_working
3487                          * queue. It shouldn't remain on the queue
3488                          * forever, in case the GET message (for ping)
3489                          * doesn't get a REPLY or the PUT message (for
3490                          * push) doesn't get an ACK.
3491                          */
3492                         lp->lp_last_queued = ktime_get_real_seconds();
3493                         lnet_net_unlock(LNET_LOCK_EX);
3494
3495                         if (lnet_push_target_resize_needed())
3496                                 lnet_push_target_resize();
3497                         else if (the_lnet.ln_push_target->pb_needs_post)
3498                                 lnet_push_target_post(the_lnet.ln_push_target,
3499                                                       &the_lnet.ln_push_target_md);
3500
3501                         /*
3502                          * Select an action depending on the state of
3503                          * the peer and whether discovery is disabled.
3504                          * The check whether discovery is disabled is
3505                          * done after the code that handles processing
3506                          * for arrived data, cleanup for failures, and
3507                          * forcing a Ping or Push.
3508                          */
3509                         spin_lock(&lp->lp_lock);
3510                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3511                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3512                                 lp->lp_state);
3513                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3514                                 rc = lnet_peer_data_present(lp);
3515                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3516                                 rc = lnet_peer_ping_failed(lp);
3517                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3518                                 rc = lnet_peer_push_failed(lp);
3519                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3520                                 rc = lnet_peer_send_ping(lp);
3521                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3522                                 rc = lnet_peer_send_push(lp);
3523                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3524                                 rc = lnet_peer_send_ping(lp);
3525                         else if (lnet_peer_needs_push(lp))
3526                                 rc = lnet_peer_send_push(lp);
3527                         else
3528                                 rc = lnet_peer_discovered(lp);
3529                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3530                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3531                                 lp->lp_state, rc);
3532                         spin_unlock(&lp->lp_lock);
3533
3534                         lnet_net_lock(LNET_LOCK_EX);
3535                         if (rc == LNET_REDISCOVER_PEER) {
3536                                 list_move(&lp->lp_dc_list,
3537                                           &the_lnet.ln_dc_request);
3538                         } else if (rc) {
3539                                 lnet_peer_discovery_error(lp, rc);
3540                         }
3541                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3542                                 lnet_peer_discovery_complete(lp);
3543                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3544                                 break;
3545
3546                         if (lp->lp_state & LNET_PEER_MARK_DELETION) {
3547                                 struct list_head rlist;
3548                                 struct lnet_route *route, *tmp;
3549                                 int sensitivity = lp->lp_health_sensitivity;
3550
3551                                 INIT_LIST_HEAD(&rlist);
3552
3553                                 /*
3554                                  * remove the peer from the discovery work
3555                                  * queue if it's on there in preparation
3556                                  * of deleting it.
3557                                  */
3558                                 if (!list_empty(&lp->lp_dc_list))
3559                                         list_del(&lp->lp_dc_list);
3560
3561                                 lnet_net_unlock(LNET_LOCK_EX);
3562
3563                                 mutex_lock(&the_lnet.ln_api_mutex);
3564
3565                                 lnet_net_lock(LNET_LOCK_EX);
3566                                 list_for_each_entry_safe(route, tmp,
3567                                                          &lp->lp_routes,
3568                                                          lr_gwlist)
3569                                         lnet_move_route(route, NULL, &rlist);
3570                                 lnet_net_unlock(LNET_LOCK_EX);
3571
3572                                 /* delete the peer */
3573                                 lnet_peer_del(lp);
3574
3575                                 list_for_each_entry_safe(route, tmp,
3576                                                          &rlist, lr_list) {
3577                                         /* re-add these routes */
3578                                         lnet_add_route(route->lr_net,
3579                                                        route->lr_hops,
3580                                                        route->lr_nid,
3581                                                        route->lr_priority,
3582                                                        sensitivity);
3583                                         LIBCFS_FREE(route, sizeof(*route));
3584                                 }
3585                                 mutex_unlock(&the_lnet.ln_api_mutex);
3586
3587                                 lnet_net_lock(LNET_LOCK_EX);
3588                         }
3589                 }
3590
3591                 lnet_net_unlock(LNET_LOCK_EX);
3592         }
3593
3594         CDEBUG(D_NET, "stopping\n");
3595         /*
3596          * Clean up before telling lnet_peer_discovery_stop() that
3597          * we're done. Use wake_up() below to somewhat reduce the
3598          * size of the thundering herd if there are multiple threads
3599          * waiting on discovery of a single peer.
3600          */
3601
3602         /* Queue cleanup 1: stop all pending pings and pushes. */
3603         lnet_net_lock(LNET_LOCK_EX);
3604         while (!list_empty(&the_lnet.ln_dc_working)) {
3605                 lp = list_first_entry(&the_lnet.ln_dc_working,
3606                                       struct lnet_peer, lp_dc_list);
3607                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3608                 lnet_net_unlock(LNET_LOCK_EX);
3609                 lnet_peer_cancel_discovery(lp);
3610                 lnet_net_lock(LNET_LOCK_EX);
3611         }
3612         lnet_net_unlock(LNET_LOCK_EX);
3613
3614         /* Queue cleanup 2: wait for the expired queue to clear. */
3615         while (!list_empty(&the_lnet.ln_dc_expired))
3616                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3617
3618         /* Queue cleanup 3: clear the request queue. */
3619         lnet_net_lock(LNET_LOCK_EX);
3620         while (!list_empty(&the_lnet.ln_dc_request)) {
3621                 lp = list_first_entry(&the_lnet.ln_dc_request,
3622                                       struct lnet_peer, lp_dc_list);
3623                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3624                 lnet_peer_discovery_complete(lp);
3625         }
3626         lnet_net_unlock(LNET_LOCK_EX);
3627
3628         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3629         the_lnet.ln_dc_handler = NULL;
3630
3631         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3632         wake_up(&the_lnet.ln_dc_waitq);
3633
3634         CDEBUG(D_NET, "stopped\n");
3635
3636         return 0;
3637 }
3638
3639 /* ln_api_mutex is held on entry. */
3640 int lnet_peer_discovery_start(void)
3641 {
3642         struct task_struct *task;
3643         int rc = 0;
3644
3645         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3646                 return -EALREADY;
3647
3648         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3649         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3650         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3651         if (IS_ERR(task)) {
3652                 rc = PTR_ERR(task);
3653                 CERROR("Can't start peer discovery thread: %d\n", rc);
3654
3655                 the_lnet.ln_dc_handler = NULL;
3656
3657                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3658         }
3659
3660         CDEBUG(D_NET, "discovery start: %d\n", rc);
3661
3662         return rc;
3663 }
3664
3665 /* ln_api_mutex is held on entry. */
3666 void lnet_peer_discovery_stop(void)
3667 {
3668         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3669                 return;
3670
3671         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3672         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3673
3674         /* In the LNetNIInit() path we may be stopping discovery before it
3675          * entered its work loop
3676          */
3677         if (!completion_done(&the_lnet.ln_started))
3678                 complete(&the_lnet.ln_started);
3679         else
3680                 wake_up(&the_lnet.ln_dc_waitq);
3681
3682         wait_event(the_lnet.ln_dc_waitq,
3683                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3684
3685         LASSERT(list_empty(&the_lnet.ln_dc_request));
3686         LASSERT(list_empty(&the_lnet.ln_dc_working));
3687         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3688
3689         CDEBUG(D_NET, "discovery stopped\n");
3690 }
3691
3692 /* Debugging */
3693
3694 void
3695 lnet_debug_peer(lnet_nid_t nid)
3696 {
3697         char                    *aliveness = "NA";
3698         struct lnet_peer_ni     *lp;
3699         int                     cpt;
3700
3701         cpt = lnet_cpt_of_nid(nid, NULL);
3702         lnet_net_lock(cpt);
3703
3704         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3705         if (IS_ERR(lp)) {
3706                 lnet_net_unlock(cpt);
3707                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3708                 return;
3709         }
3710
3711         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3712                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3713
3714         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3715                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3716                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3717                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3718                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3719
3720         lnet_peer_ni_decref_locked(lp);
3721
3722         lnet_net_unlock(cpt);
3723 }
3724
3725 /* Gathering information for userspace. */
3726
3727 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3728                           char aliveness[LNET_MAX_STR_LEN],
3729                           __u32 *cpt_iter, __u32 *refcount,
3730                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3731                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3732                           __u32 *peer_tx_qnob)
3733 {
3734         struct lnet_peer_table          *peer_table;
3735         struct lnet_peer_ni             *lp;
3736         int                             j;
3737         int                             lncpt;
3738         bool                            found = false;
3739
3740         /* get the number of CPTs */
3741         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3742
3743         /* if the cpt number to be examined is >= the number of cpts in
3744          * the system then indicate that there are no more cpts to examin
3745          */
3746         if (*cpt_iter >= lncpt)
3747                 return -ENOENT;
3748
3749         /* get the current table */
3750         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3751         /* if the ptable is NULL then there are no more cpts to examine */
3752         if (peer_table == NULL)
3753                 return -ENOENT;
3754
3755         lnet_net_lock(*cpt_iter);
3756
3757         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3758                 struct list_head *peers = &peer_table->pt_hash[j];
3759
3760                 list_for_each_entry(lp, peers, lpni_hashlist) {
3761                         if (peer_index-- > 0)
3762                                 continue;
3763
3764                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3765                         if (lnet_isrouter(lp) ||
3766                                 lnet_peer_aliveness_enabled(lp))
3767                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3768                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3769
3770                         *nid = lp->lpni_nid;
3771                         *refcount = atomic_read(&lp->lpni_refcount);
3772                         *ni_peer_tx_credits =
3773                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3774                         *peer_tx_credits = lp->lpni_txcredits;
3775                         *peer_rtr_credits = lp->lpni_rtrcredits;
3776                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3777                         *peer_tx_qnob = lp->lpni_txqnob;
3778
3779                         found = true;
3780                 }
3781
3782         }
3783         lnet_net_unlock(*cpt_iter);
3784
3785         *cpt_iter = lncpt;
3786
3787         return found ? 0 : -ENOENT;
3788 }
3789
3790 /* ln_api_mutex is held, which keeps the peer list stable */
3791 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3792 {
3793         struct lnet_ioctl_element_stats *lpni_stats;
3794         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3795         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3796         struct lnet_peer_ni_credit_info *lpni_info;
3797         struct lnet_peer_ni *lpni;
3798         struct lnet_peer *lp;
3799         lnet_nid_t nid;
3800         __u32 size;
3801         int rc;
3802
3803         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3804
3805         if (!lp) {
3806                 rc = -ENOENT;
3807                 goto out;
3808         }
3809
3810         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3811                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3812         size *= lp->lp_nnis;
3813         if (size > cfg->prcfg_size) {
3814                 cfg->prcfg_size = size;
3815                 rc = -E2BIG;
3816                 goto out_lp_decref;
3817         }
3818
3819         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3820         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3821         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3822         cfg->prcfg_count = lp->lp_nnis;
3823         cfg->prcfg_size = size;
3824         cfg->prcfg_state = lp->lp_state;
3825
3826         /* Allocate helper buffers. */
3827         rc = -ENOMEM;
3828         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3829         if (!lpni_info)
3830                 goto out_lp_decref;
3831         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3832         if (!lpni_stats)
3833                 goto out_free_info;
3834         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3835         if (!lpni_msg_stats)
3836                 goto out_free_stats;
3837         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3838         if (!lpni_hstats)
3839                 goto out_free_msg_stats;
3840
3841
3842         lpni = NULL;
3843         rc = -EFAULT;
3844         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3845                 nid = lpni->lpni_nid;
3846                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3847                         goto out_free_hstats;
3848                 bulk += sizeof(nid);
3849
3850                 memset(lpni_info, 0, sizeof(*lpni_info));
3851                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3852                 if (lnet_isrouter(lpni) ||
3853                         lnet_peer_aliveness_enabled(lpni))
3854                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3855                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3856
3857                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3858                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3859                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3860                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3861                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3862                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3863                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3864                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3865                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3866                         goto out_free_hstats;
3867                 bulk += sizeof(*lpni_info);
3868
3869                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3870                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3871                                                             LNET_STATS_TYPE_SEND);
3872                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3873                                                             LNET_STATS_TYPE_RECV);
3874                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3875                                                             LNET_STATS_TYPE_DROP);
3876                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3877                         goto out_free_hstats;
3878                 bulk += sizeof(*lpni_stats);
3879                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3880                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3881                         goto out_free_hstats;
3882                 bulk += sizeof(*lpni_msg_stats);
3883                 lpni_hstats->hlpni_network_timeout =
3884                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3885                 lpni_hstats->hlpni_remote_dropped =
3886                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3887                 lpni_hstats->hlpni_remote_timeout =
3888                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3889                 lpni_hstats->hlpni_remote_error =
3890                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3891                 lpni_hstats->hlpni_health_value =
3892                   atomic_read(&lpni->lpni_healthv);
3893                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3894                         goto out_free_hstats;
3895                 bulk += sizeof(*lpni_hstats);
3896         }
3897         rc = 0;
3898
3899 out_free_hstats:
3900         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3901 out_free_msg_stats:
3902         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3903 out_free_stats:
3904         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3905 out_free_info:
3906         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3907 out_lp_decref:
3908         lnet_peer_decref_locked(lp);
3909 out:
3910         return rc;
3911 }
3912
3913 void
3914 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3915 {
3916         /* the mt could've shutdown and cleaned up the queues */
3917         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3918                 return;
3919
3920         if (list_empty(&lpni->lpni_recovery) &&
3921             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3922                 CDEBUG(D_NET, "lpni %s added to recovery queue. Health = %d\n",
3923                         libcfs_nid2str(lpni->lpni_nid),
3924                         atomic_read(&lpni->lpni_healthv));
3925                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3926                 lnet_peer_ni_addref_locked(lpni);
3927         }
3928 }
3929
3930 /* Call with the ln_api_mutex held */
3931 void
3932 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3933 {
3934         struct lnet_peer_table *ptable;
3935         struct lnet_peer *lp;
3936         struct lnet_peer_net *lpn;
3937         struct lnet_peer_ni *lpni;
3938         int lncpt;
3939         int cpt;
3940
3941         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3942                 return;
3943
3944         if (!all) {
3945                 lnet_net_lock(LNET_LOCK_EX);
3946                 lpni = lnet_find_peer_ni_locked(nid);
3947                 if (!lpni) {
3948                         lnet_net_unlock(LNET_LOCK_EX);
3949                         return;
3950                 }
3951                 atomic_set(&lpni->lpni_healthv, value);
3952                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3953                 lnet_peer_ni_decref_locked(lpni);
3954                 lnet_net_unlock(LNET_LOCK_EX);
3955                 return;
3956         }
3957
3958         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3959
3960         /*
3961          * Walk all the peers and reset the healhv for each one to the
3962          * maximum value.
3963          */
3964         lnet_net_lock(LNET_LOCK_EX);
3965         for (cpt = 0; cpt < lncpt; cpt++) {
3966                 ptable = the_lnet.ln_peer_tables[cpt];
3967                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3968                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3969                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3970                                                     lpni_peer_nis) {
3971                                         atomic_set(&lpni->lpni_healthv, value);
3972                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3973                                 }
3974                         }
3975                 }
3976         }
3977         lnet_net_unlock(LNET_LOCK_EX);
3978 }
3979