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                 LIBCFS_FREE(hash, LNET_PEER_HASH_SIZE * sizeof(*hash));
 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         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
 170
 171         spin_lock_init(&lpni->lpni_lock);
 172
 173         if (lnet_peers_start_down())
 174                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 175         else
 176                 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
 177         lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 178         lpni->lpni_nid = nid;
 179         lpni->lpni_cpt = cpt;
 180         atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
 181
 182         net = lnet_get_net_locked(LNET_NIDNET(nid));
 183         lpni->lpni_net = net;
 184         if (net) {
 185                 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
 186                 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
 187                 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
 188                 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 189         } else {
 190                 /*
 191                  * This peer_ni is not on a local network, so we
 192                  * cannot add the credits here. In case the net is
 193                  * added later, add the peer_ni to the remote peer ni
 194                  * list so it can be easily found and revisited.
 195                  */
 196                 /* FIXME: per-net implementation instead? */
 197                 atomic_inc(&lpni->lpni_refcount);
 198                 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
 199                               &the_lnet.ln_remote_peer_ni_list);
 200         }
 201
 202         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
 203
 204         return lpni;
 205 }
 206
 207 static struct lnet_peer_net *
 208 lnet_peer_net_alloc(__u32 net_id)
 209 {
 210         struct lnet_peer_net *lpn;
 211
 212         LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
 213         if (!lpn)
 214                 return NULL;
 215
 216         INIT_LIST_HEAD(&lpn->lpn_peer_nets);
 217         INIT_LIST_HEAD(&lpn->lpn_peer_nis);
 218         lpn->lpn_net_id = net_id;
 219
 220         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 221
 222         return lpn;
 223 }
 224
 225 void
 226 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
 227 {
 228         struct lnet_peer *lp;
 229
 230         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 231
 232         LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
 233         LASSERT(list_empty(&lpn->lpn_peer_nis));
 234         LASSERT(list_empty(&lpn->lpn_peer_nets));
 235         lp = lpn->lpn_peer;
 236         lpn->lpn_peer = NULL;
 237         LIBCFS_FREE(lpn, sizeof(*lpn));
 238
 239         lnet_peer_decref_locked(lp);
 240 }
 241
 242 static struct lnet_peer *
 243 lnet_peer_alloc(lnet_nid_t nid)
 244 {
 245         struct lnet_peer *lp;
 246
 247         LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
 248         if (!lp)
 249                 return NULL;
 250
 251         INIT_LIST_HEAD(&lp->lp_rtrq);
 252         INIT_LIST_HEAD(&lp->lp_routes);
 253         INIT_LIST_HEAD(&lp->lp_peer_list);
 254         INIT_LIST_HEAD(&lp->lp_peer_nets);
 255         INIT_LIST_HEAD(&lp->lp_dc_list);
 256         INIT_LIST_HEAD(&lp->lp_dc_pendq);
 257         INIT_LIST_HEAD(&lp->lp_rtr_list);
 258         init_waitqueue_head(&lp->lp_dc_waitq);
 259         spin_lock_init(&lp->lp_lock);
 260         lp->lp_primary_nid = nid;
 261         /*
 262          * Turn off discovery for loopback peer. If you're creating a peer
 263          * for the loopback interface then that was initiated when we
 264          * attempted to send a message over the loopback. There is no need
 265          * to ever use a different interface when sending messages to
 266          * myself.
 267          */
 268         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 269                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 270         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 271
 272         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 273
 274         return lp;
 275 }
 276
 277 void
 278 lnet_destroy_peer_locked(struct lnet_peer *lp)
 279 {
 280         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 281
 282         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 283         LASSERT(lp->lp_rtr_refcount == 0);
 284         LASSERT(list_empty(&lp->lp_peer_nets));
 285         LASSERT(list_empty(&lp->lp_peer_list));
 286         LASSERT(list_empty(&lp->lp_dc_list));
 287
 288         if (lp->lp_data)
 289                 lnet_ping_buffer_decref(lp->lp_data);
 290
 291         /*
 292          * if there are messages still on the pending queue, then make
 293          * sure to queue them on the ln_msg_resend list so they can be
 294          * resent at a later point if the discovery thread is still
 295          * running.
 296          * If the discovery thread has stopped, then the wakeup will be a
 297          * no-op, and it is expected the lnet_shutdown_lndnets() will
 298          * eventually be called, which will traverse this list and
 299          * finalize the messages on the list.
 300          * We can not resend them now because we're holding the cpt lock.
 301          * Releasing the lock can cause an inconsistent state
 302          */
 303         spin_lock(&the_lnet.ln_msg_resend_lock);
 304         spin_lock(&lp->lp_lock);
 305         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 306         spin_unlock(&lp->lp_lock);
 307         spin_unlock(&the_lnet.ln_msg_resend_lock);
 308         wake_up(&the_lnet.ln_dc_waitq);
 309
 310         LIBCFS_FREE(lp, sizeof(*lp));
 311 }
 312
 313 /*
 314  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 315  * that peer_net, detach the peer_net from the peer.
 316  *
 317  * Call with lnet_net_lock/EX held
 318  */
 319 static void
 320 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 321 {
 322         struct lnet_peer_table *ptable;
 323         struct lnet_peer_net *lpn;
 324         struct lnet_peer *lp;
 325
 326         /*
 327          * Belts and suspenders: gracefully handle teardown of a
 328          * partially connected peer_ni.
 329          */
 330         lpn = lpni->lpni_peer_net;
 331
 332         list_del_init(&lpni->lpni_peer_nis);
 333         /*
 334          * If there are no lpni's left, we detach lpn from
 335          * lp_peer_nets, so it cannot be found anymore.
 336          */
 337         if (list_empty(&lpn->lpn_peer_nis))
 338                 list_del_init(&lpn->lpn_peer_nets);
 339
 340         /* Update peer NID count. */
 341         lp = lpn->lpn_peer;
 342         lp->lp_nnis--;
 343
 344         /*
 345          * If there are no more peer nets, make the peer unfindable
 346          * via the peer_tables.
 347          *
 348          * Otherwise, if the peer is DISCOVERED, tell discovery to
 349          * take another look at it. This is a no-op if discovery for
 350          * this peer did the detaching.
 351          */
 352         if (list_empty(&lp->lp_peer_nets)) {
 353                 list_del_init(&lp->lp_peer_list);
 354                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 355                 ptable->pt_peers--;
 356         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 357                 /* Discovery isn't running, nothing to do here. */
 358         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 359                 lnet_peer_queue_for_discovery(lp);
 360                 wake_up(&the_lnet.ln_dc_waitq);
 361         }
 362         CDEBUG(D_NET, "peer %s NID %s\n",
 363                 libcfs_nid2str(lp->lp_primary_nid),
 364                 libcfs_nid2str(lpni->lpni_nid));
 365 }
 366
 367 /* called with lnet_net_lock LNET_LOCK_EX held */
 368 static int
 369 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 370 {
 371         struct lnet_peer_table *ptable = NULL;
 372
 373         /* don't remove a peer_ni if it's also a gateway */
 374         if (lnet_isrouter(lpni) && !force) {
 375                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 376                        libcfs_nid2str(lpni->lpni_nid));
 377                 return -EBUSY;
 378         }
 379
 380         lnet_peer_remove_from_remote_list(lpni);
 381
 382         /* remove peer ni from the hash list. */
 383         list_del_init(&lpni->lpni_hashlist);
 384
 385         /*
 386          * indicate the peer is being deleted so the monitor thread can
 387          * remove it from the recovery queue.
 388          */
 389         spin_lock(&lpni->lpni_lock);
 390         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 391         spin_unlock(&lpni->lpni_lock);
 392
 393         /* decrement the ref count on the peer table */
 394         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 395         LASSERT(ptable->pt_number > 0);
 396         ptable->pt_number--;
 397
 398         /*
 399          * The peer_ni can no longer be found with a lookup. But there
 400          * can be current users, so keep track of it on the zombie
 401          * list until the reference count has gone to zero.
 402          *
 403          * The last reference may be lost in a place where the
 404          * lnet_net_lock locks only a single cpt, and that cpt may not
 405          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 406          * has its own lock.
 407          */
 408         spin_lock(&ptable->pt_zombie_lock);
 409         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 410         ptable->pt_zombies++;
 411         spin_unlock(&ptable->pt_zombie_lock);
 412
 413         /* no need to keep this peer_ni on the hierarchy anymore */
 414         lnet_peer_detach_peer_ni_locked(lpni);
 415
 416         /* remove hashlist reference on peer_ni */
 417         lnet_peer_ni_decref_locked(lpni);
 418
 419         return 0;
 420 }
 421
 422 void lnet_peer_uninit(void)
 423 {
 424         struct lnet_peer_ni *lpni, *tmp;
 425
 426         lnet_net_lock(LNET_LOCK_EX);
 427
 428         /* remove all peer_nis from the remote peer and the hash list */
 429         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 430                                  lpni_on_remote_peer_ni_list)
 431                 lnet_peer_ni_del_locked(lpni, false);
 432
 433         lnet_peer_tables_destroy();
 434
 435         lnet_net_unlock(LNET_LOCK_EX);
 436 }
 437
 438 static int
 439 lnet_peer_del_locked(struct lnet_peer *peer)
 440 {
 441         struct lnet_peer_ni *lpni = NULL, *lpni2;
 442         int rc = 0, rc2 = 0;
 443
 444         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
 445
 446         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 447         while (lpni != NULL) {
 448                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 449                 rc = lnet_peer_ni_del_locked(lpni, false);
 450                 if (rc != 0)
 451                         rc2 = rc;
 452                 lpni = lpni2;
 453         }
 454
 455         return rc2;
 456 }
 457
 458 static int
 459 lnet_peer_del(struct lnet_peer *peer)
 460 {
 461         lnet_net_lock(LNET_LOCK_EX);
 462         lnet_peer_del_locked(peer);
 463         lnet_net_unlock(LNET_LOCK_EX);
 464
 465         return 0;
 466 }
 467
 468 /*
 469  * Delete a NID from a peer. Call with ln_api_mutex held.
 470  *
 471  * Error codes:
 472  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 473  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 474  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 475  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 476  */
 477 static int
 478 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
 479 {
 480         struct lnet_peer_ni *lpni;
 481         lnet_nid_t primary_nid = lp->lp_primary_nid;
 482         int rc = 0;
 483         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 484
 485         if (!(flags & LNET_PEER_CONFIGURED)) {
 486                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 487                         rc = -EPERM;
 488                         goto out;
 489                 }
 490         }
 491         lpni = lnet_find_peer_ni_locked(nid);
 492         if (!lpni) {
 493                 rc = -ENOENT;
 494                 goto out;
 495         }
 496         lnet_peer_ni_decref_locked(lpni);
 497         if (lp != lpni->lpni_peer_net->lpn_peer) {
 498                 rc = -ECHILD;
 499                 goto out;
 500         }
 501
 502         /*
 503          * This function only allows deletion of the primary NID if it
 504          * is the only NID.
 505          */
 506         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
 507                 rc = -EBUSY;
 508                 goto out;
 509         }
 510
 511         lnet_net_lock(LNET_LOCK_EX);
 512
 513         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
 514                 struct lnet_peer_ni *lpni2;
 515                 /* assign the next peer_ni to be the primary */
 516                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 517                 LASSERT(lpni2);
 518                 lp->lp_primary_nid = lpni->lpni_nid;
 519         }
 520         rc = lnet_peer_ni_del_locked(lpni, force);
 521
 522         lnet_net_unlock(LNET_LOCK_EX);
 523
 524 out:
 525         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 526                libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
 527
 528         return rc;
 529 }
 530
 531 static void
 532 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 533                                struct lnet_peer_table *ptable)
 534 {
 535         int                      i;
 536         struct lnet_peer_ni     *next;
 537         struct lnet_peer_ni     *lpni;
 538         struct lnet_peer        *peer;
 539
 540         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 541                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 542                                          lpni_hashlist) {
 543                         if (net != NULL && net != lpni->lpni_net)
 544                                 continue;
 545
 546                         peer = lpni->lpni_peer_net->lpn_peer;
 547                         if (peer->lp_primary_nid != lpni->lpni_nid) {
 548                                 lnet_peer_ni_del_locked(lpni, false);
 549                                 continue;
 550                         }
 551                         /*
 552                          * Removing the primary NID implies removing
 553                          * the entire peer. Advance next beyond any
 554                          * peer_ni that belongs to the same peer.
 555                          */
 556                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 557                                                  lpni_hashlist) {
 558                                 if (next->lpni_peer_net->lpn_peer != peer)
 559                                         break;
 560                         }
 561                         lnet_peer_del_locked(peer);
 562                 }
 563         }
 564 }
 565
 566 static void
 567 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 568 {
 569         int     i = 3;
 570
 571         spin_lock(&ptable->pt_zombie_lock);
 572         while (ptable->pt_zombies) {
 573                 spin_unlock(&ptable->pt_zombie_lock);
 574
 575                 if (is_power_of_2(i)) {
 576                         CDEBUG(D_WARNING,
 577                                "Waiting for %d zombies on peer table\n",
 578                                ptable->pt_zombies);
 579                 }
 580                 set_current_state(TASK_UNINTERRUPTIBLE);
 581                 schedule_timeout(cfs_time_seconds(1) >> 1);
 582                 spin_lock(&ptable->pt_zombie_lock);
 583         }
 584         spin_unlock(&ptable->pt_zombie_lock);
 585 }
 586
 587 static void
 588 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 589                                 struct lnet_peer_table *ptable)
 590 {
 591         struct lnet_peer_ni     *lp;
 592         struct lnet_peer_ni     *tmp;
 593         lnet_nid_t              gw_nid;
 594         int                     i;
 595
 596         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 597                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 598                                          lpni_hashlist) {
 599                         if (net != lp->lpni_net)
 600                                 continue;
 601
 602                         if (!lnet_isrouter(lp))
 603                                 continue;
 604
 605                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 606
 607                         lnet_net_unlock(LNET_LOCK_EX);
 608                         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
 609                         lnet_net_lock(LNET_LOCK_EX);
 610                 }
 611         }
 612 }
 613
 614 void
 615 lnet_peer_tables_cleanup(struct lnet_net *net)
 616 {
 617         int i;
 618         struct lnet_peer_table *ptable;
 619
 620         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 621         /* If just deleting the peers for a NI, get rid of any routes these
 622          * peers are gateways for. */
 623         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 624                 lnet_net_lock(LNET_LOCK_EX);
 625                 lnet_peer_table_del_rtrs_locked(net, ptable);
 626                 lnet_net_unlock(LNET_LOCK_EX);
 627         }
 628
 629         /* Start the cleanup process */
 630         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 631                 lnet_net_lock(LNET_LOCK_EX);
 632                 lnet_peer_table_cleanup_locked(net, ptable);
 633                 lnet_net_unlock(LNET_LOCK_EX);
 634         }
 635
 636         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 637                 lnet_peer_ni_finalize_wait(ptable);
 638 }
 639
 640 static struct lnet_peer_ni *
 641 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 642 {
 643         struct list_head        *peers;
 644         struct lnet_peer_ni     *lp;
 645
 646         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 647
 648         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 649         list_for_each_entry(lp, peers, lpni_hashlist) {
 650                 if (lp->lpni_nid == nid) {
 651                         lnet_peer_ni_addref_locked(lp);
 652                         return lp;
 653                 }
 654         }
 655
 656         return NULL;
 657 }
 658
 659 struct lnet_peer_ni *
 660 lnet_find_peer_ni_locked(lnet_nid_t nid)
 661 {
 662         struct lnet_peer_ni *lpni;
 663         struct lnet_peer_table *ptable;
 664         int cpt;
 665
 666         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 667
 668         ptable = the_lnet.ln_peer_tables[cpt];
 669         lpni = lnet_get_peer_ni_locked(ptable, nid);
 670
 671         return lpni;
 672 }
 673
 674 struct lnet_peer_ni *
 675 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 676 {
 677         struct lnet_peer_net *lpn;
 678         struct lnet_peer_ni *lpni;
 679
 680         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 681         if (!lpn)
 682                 return NULL;
 683
 684         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 685                 if (lpni->lpni_nid == nid)
 686                         return lpni;
 687         }
 688
 689         return NULL;
 690 }
 691
 692 struct lnet_peer *
 693 lnet_find_peer(lnet_nid_t nid)
 694 {
 695         struct lnet_peer_ni *lpni;
 696         struct lnet_peer *lp = NULL;
 697         int cpt;
 698
 699         cpt = lnet_net_lock_current();
 700         lpni = lnet_find_peer_ni_locked(nid);
 701         if (lpni) {
 702                 lp = lpni->lpni_peer_net->lpn_peer;
 703                 lnet_peer_addref_locked(lp);
 704                 lnet_peer_ni_decref_locked(lpni);
 705         }
 706         lnet_net_unlock(cpt);
 707
 708         return lp;
 709 }
 710
 711 struct lnet_peer_ni *
 712 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 713                              struct lnet_peer_net *peer_net,
 714                              struct lnet_peer_ni *prev)
 715 {
 716         struct lnet_peer_ni *lpni;
 717         struct lnet_peer_net *net = peer_net;
 718
 719         if (!prev) {
 720                 if (!net) {
 721                         if (list_empty(&peer->lp_peer_nets))
 722                                 return NULL;
 723
 724                         net = list_entry(peer->lp_peer_nets.next,
 725                                          struct lnet_peer_net,
 726                                          lpn_peer_nets);
 727                 }
 728                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 729                                   lpni_peer_nis);
 730
 731                 return lpni;
 732         }
 733
 734         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 735                 /*
 736                  * if you reached the end of the peer ni list and the peer
 737                  * net is specified then there are no more peer nis in that
 738                  * net.
 739                  */
 740                 if (net)
 741                         return NULL;
 742
 743                 /*
 744                  * we reached the end of this net ni list. move to the
 745                  * next net
 746                  */
 747                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 748                     &peer->lp_peer_nets)
 749                         /* no more nets and no more NIs. */
 750                         return NULL;
 751
 752                 /* get the next net */
 753                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 754                                  struct lnet_peer_net,
 755                                  lpn_peer_nets);
 756                 /* get the ni on it */
 757                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 758                                   lpni_peer_nis);
 759
 760                 return lpni;
 761         }
 762
 763         /* there are more nis left */
 764         lpni = list_entry(prev->lpni_peer_nis.next,
 765                           struct lnet_peer_ni, lpni_peer_nis);
 766
 767         return lpni;
 768 }
 769
 770 /* Call with the ln_api_mutex held */
 771 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 772 {
 773         struct lnet_process_id id;
 774         struct lnet_peer_table *ptable;
 775         struct lnet_peer *lp;
 776         __u32 count = 0;
 777         __u32 size = 0;
 778         int lncpt;
 779         int cpt;
 780         __u32 i;
 781         int rc;
 782
 783         rc = -ESHUTDOWN;
 784         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 785                 goto done;
 786
 787         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 788
 789         /*
 790          * Count the number of peers, and return E2BIG if the buffer
 791          * is too small. We'll also return the desired size.
 792          */
 793         rc = -E2BIG;
 794         for (cpt = 0; cpt < lncpt; cpt++) {
 795                 ptable = the_lnet.ln_peer_tables[cpt];
 796                 count += ptable->pt_peers;
 797         }
 798         size = count * sizeof(*ids);
 799         if (size > *sizep)
 800                 goto done;
 801
 802         /*
 803          * Walk the peer lists and copy out the primary nids.
 804          * This is safe because the peer lists are only modified
 805          * while the ln_api_mutex is held. So we don't need to
 806          * hold the lnet_net_lock as well, and can therefore
 807          * directly call copy_to_user().
 808          */
 809         rc = -EFAULT;
 810         memset(&id, 0, sizeof(id));
 811         id.pid = LNET_PID_LUSTRE;
 812         i = 0;
 813         for (cpt = 0; cpt < lncpt; cpt++) {
 814                 ptable = the_lnet.ln_peer_tables[cpt];
 815                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 816                         if (i >= count)
 817                                 goto done;
 818                         id.nid = lp->lp_primary_nid;
 819                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 820                                 goto done;
 821                         i++;
 822                 }
 823         }
 824         rc = 0;
 825 done:
 826         *countp = count;
 827         *sizep = size;
 828         return rc;
 829 }
 830
 831 /*
 832  * Start pushes to peers that need to be updated for a configuration
 833  * change on this node.
 834  */
 835 void
 836 lnet_push_update_to_peers(int force)
 837 {
 838         struct lnet_peer_table *ptable;
 839         struct lnet_peer *lp;
 840         int lncpt;
 841         int cpt;
 842
 843         lnet_net_lock(LNET_LOCK_EX);
 844         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 845         for (cpt = 0; cpt < lncpt; cpt++) {
 846                 ptable = the_lnet.ln_peer_tables[cpt];
 847                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 848                         if (force) {
 849                                 spin_lock(&lp->lp_lock);
 850                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 851                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 852                                 spin_unlock(&lp->lp_lock);
 853                         }
 854                         if (lnet_peer_needs_push(lp))
 855                                 lnet_peer_queue_for_discovery(lp);
 856                 }
 857         }
 858         lnet_net_unlock(LNET_LOCK_EX);
 859         wake_up(&the_lnet.ln_dc_waitq);
 860 }
 861
 862 /*
 863  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 864  * this is a preferred point-to-point path. Call with lnet_net_lock in
 865  * shared mmode.
 866  */
 867 bool
 868 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 869 {
 870         int i;
 871
 872         if (lpni->lpni_pref_nnids == 0)
 873                 return false;
 874         if (lpni->lpni_pref_nnids == 1)
 875                 return lpni->lpni_pref.nid == nid;
 876         for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 877                 if (lpni->lpni_pref.nids[i] == nid)
 878                         return true;
 879         }
 880         return false;
 881 }
 882
 883 /*
 884  * Set a single ni as preferred, provided no preferred ni is already
 885  * defined. Only to be used for non-multi-rail peer_ni.
 886  */
 887 int
 888 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 889 {
 890         int rc = 0;
 891
 892         spin_lock(&lpni->lpni_lock);
 893         if (nid == LNET_NID_ANY) {
 894                 rc = -EINVAL;
 895         } else if (lpni->lpni_pref_nnids > 0) {
 896                 rc = -EPERM;
 897         } else if (lpni->lpni_pref_nnids == 0) {
 898                 lpni->lpni_pref.nid = nid;
 899                 lpni->lpni_pref_nnids = 1;
 900                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
 901         }
 902         spin_unlock(&lpni->lpni_lock);
 903
 904         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 905                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
 906         return rc;
 907 }
 908
 909 /*
 910  * Clear the preferred NID from a non-multi-rail peer_ni, provided
 911  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
 912  */
 913 int
 914 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
 915 {
 916         int rc = 0;
 917
 918         spin_lock(&lpni->lpni_lock);
 919         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
 920                 lpni->lpni_pref_nnids = 0;
 921                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 922         } else if (lpni->lpni_pref_nnids == 0) {
 923                 rc = -ENOENT;
 924         } else {
 925                 rc = -EPERM;
 926         }
 927         spin_unlock(&lpni->lpni_lock);
 928
 929         CDEBUG(D_NET, "peer %s: %d\n",
 930                libcfs_nid2str(lpni->lpni_nid), rc);
 931         return rc;
 932 }
 933
 934 /*
 935  * Clear the preferred NIDs from a non-multi-rail peer.
 936  */
 937 void
 938 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
 939 {
 940         struct lnet_peer_ni *lpni = NULL;
 941
 942         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 943                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
 944 }
 945
 946 int
 947 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 948 {
 949         lnet_nid_t *nids = NULL;
 950         lnet_nid_t *oldnids = NULL;
 951         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 952         int size;
 953         int i;
 954         int rc = 0;
 955
 956         if (nid == LNET_NID_ANY) {
 957                 rc = -EINVAL;
 958                 goto out;
 959         }
 960
 961         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
 962                 rc = -EEXIST;
 963                 goto out;
 964         }
 965
 966         /* A non-MR node may have only one preferred NI per peer_ni */
 967         if (lpni->lpni_pref_nnids > 0) {
 968                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
 969                         rc = -EPERM;
 970                         goto out;
 971                 }
 972         }
 973
 974         if (lpni->lpni_pref_nnids != 0) {
 975                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
 976                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
 977                 if (!nids) {
 978                         rc = -ENOMEM;
 979                         goto out;
 980                 }
 981                 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 982                         if (lpni->lpni_pref.nids[i] == nid) {
 983                                 LIBCFS_FREE(nids, size);
 984                                 rc = -EEXIST;
 985                                 goto out;
 986                         }
 987                         nids[i] = lpni->lpni_pref.nids[i];
 988                 }
 989                 nids[i] = nid;
 990         }
 991
 992         lnet_net_lock(LNET_LOCK_EX);
 993         spin_lock(&lpni->lpni_lock);
 994         if (lpni->lpni_pref_nnids == 0) {
 995                 lpni->lpni_pref.nid = nid;
 996         } else {
 997                 oldnids = lpni->lpni_pref.nids;
 998                 lpni->lpni_pref.nids = nids;
 999         }
1000         lpni->lpni_pref_nnids++;
1001         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1002         spin_unlock(&lpni->lpni_lock);
1003         lnet_net_unlock(LNET_LOCK_EX);
1004
1005         if (oldnids) {
1006                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1007                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1008         }
1009 out:
1010         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1011                 spin_lock(&lpni->lpni_lock);
1012                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1013                 spin_unlock(&lpni->lpni_lock);
1014         }
1015         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1016                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1017         return rc;
1018 }
1019
1020 int
1021 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1022 {
1023         lnet_nid_t *nids = NULL;
1024         lnet_nid_t *oldnids = NULL;
1025         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1026         int size;
1027         int i, j;
1028         int rc = 0;
1029
1030         if (lpni->lpni_pref_nnids == 0) {
1031                 rc = -ENOENT;
1032                 goto out;
1033         }
1034
1035         if (lpni->lpni_pref_nnids == 1) {
1036                 if (lpni->lpni_pref.nid != nid) {
1037                         rc = -ENOENT;
1038                         goto out;
1039                 }
1040         } else if (lpni->lpni_pref_nnids == 2) {
1041                 if (lpni->lpni_pref.nids[0] != nid &&
1042                     lpni->lpni_pref.nids[1] != nid) {
1043                         rc = -ENOENT;
1044                         goto out;
1045                 }
1046         } else {
1047                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1048                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1049                 if (!nids) {
1050                         rc = -ENOMEM;
1051                         goto out;
1052                 }
1053                 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1054                         if (lpni->lpni_pref.nids[i] != nid)
1055                                 continue;
1056                         nids[j++] = lpni->lpni_pref.nids[i];
1057                 }
1058                 /* Check if we actually removed a nid. */
1059                 if (j == lpni->lpni_pref_nnids) {
1060                         LIBCFS_FREE(nids, size);
1061                         rc = -ENOENT;
1062                         goto out;
1063                 }
1064         }
1065
1066         lnet_net_lock(LNET_LOCK_EX);
1067         spin_lock(&lpni->lpni_lock);
1068         if (lpni->lpni_pref_nnids == 1) {
1069                 lpni->lpni_pref.nid = LNET_NID_ANY;
1070         } else if (lpni->lpni_pref_nnids == 2) {
1071                 oldnids = lpni->lpni_pref.nids;
1072                 if (oldnids[0] == nid)
1073                         lpni->lpni_pref.nid = oldnids[1];
1074                 else
1075                         lpni->lpni_pref.nid = oldnids[2];
1076         } else {
1077                 oldnids = lpni->lpni_pref.nids;
1078                 lpni->lpni_pref.nids = nids;
1079         }
1080         lpni->lpni_pref_nnids--;
1081         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1082         spin_unlock(&lpni->lpni_lock);
1083         lnet_net_unlock(LNET_LOCK_EX);
1084
1085         if (oldnids) {
1086                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1087                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1088         }
1089 out:
1090         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1091                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1092         return rc;
1093 }
1094
1095 lnet_nid_t
1096 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1097 {
1098         struct lnet_peer_ni *lpni;
1099         lnet_nid_t primary_nid = nid;
1100
1101         lpni = lnet_find_peer_ni_locked(nid);
1102         if (lpni) {
1103                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1104                 lnet_peer_ni_decref_locked(lpni);
1105         }
1106
1107         return primary_nid;
1108 }
1109
1110 lnet_nid_t
1111 LNetPrimaryNID(lnet_nid_t nid)
1112 {
1113         struct lnet_peer *lp;
1114         struct lnet_peer_ni *lpni;
1115         lnet_nid_t primary_nid = nid;
1116         int rc = 0;
1117         int cpt;
1118
1119         cpt = lnet_net_lock_current();
1120         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1121         if (IS_ERR(lpni)) {
1122                 rc = PTR_ERR(lpni);
1123                 goto out_unlock;
1124         }
1125         lp = lpni->lpni_peer_net->lpn_peer;
1126         while (!lnet_peer_is_uptodate(lp)) {
1127                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1128                 if (rc)
1129                         goto out_decref;
1130                 lp = lpni->lpni_peer_net->lpn_peer;
1131         }
1132         primary_nid = lp->lp_primary_nid;
1133 out_decref:
1134         lnet_peer_ni_decref_locked(lpni);
1135 out_unlock:
1136         lnet_net_unlock(cpt);
1137
1138         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1139                libcfs_nid2str(primary_nid), rc);
1140         return primary_nid;
1141 }
1142 EXPORT_SYMBOL(LNetPrimaryNID);
1143
1144 struct lnet_peer_net *
1145 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1146 {
1147         struct lnet_peer_net *peer_net;
1148         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1149                 if (peer_net->lpn_net_id == net_id)
1150                         return peer_net;
1151         }
1152         return NULL;
1153 }
1154
1155 /*
1156  * Attach a peer_ni to a peer_net and peer. This function assumes
1157  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1158  * may be attached to a different peer, in which case it will be
1159  * properly detached first. The whole operation is done atomically.
1160  *
1161  * Always returns 0.  This is the last function called from functions
1162  * that do return an int, so returning 0 here allows the compiler to
1163  * do a tail call.
1164  */
1165 static int
1166 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1167                                 struct lnet_peer_net *lpn,
1168                                 struct lnet_peer_ni *lpni,
1169                                 unsigned flags)
1170 {
1171         struct lnet_peer_table *ptable;
1172
1173         /* Install the new peer_ni */
1174         lnet_net_lock(LNET_LOCK_EX);
1175         /* Add peer_ni to global peer table hash, if necessary. */
1176         if (list_empty(&lpni->lpni_hashlist)) {
1177                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1178
1179                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1180                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1181                 ptable->pt_version++;
1182                 ptable->pt_number++;
1183                 /* This is the 1st refcount on lpni. */
1184                 atomic_inc(&lpni->lpni_refcount);
1185         }
1186
1187         /* Detach the peer_ni from an existing peer, if necessary. */
1188         if (lpni->lpni_peer_net) {
1189                 LASSERT(lpni->lpni_peer_net != lpn);
1190                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1191                 lnet_peer_detach_peer_ni_locked(lpni);
1192                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1193                 lpni->lpni_peer_net = NULL;
1194         }
1195
1196         /* Add peer_ni to peer_net */
1197         lpni->lpni_peer_net = lpn;
1198         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1199         lnet_peer_net_addref_locked(lpn);
1200
1201         /* Add peer_net to peer */
1202         if (!lpn->lpn_peer) {
1203                 lpn->lpn_peer = lp;
1204                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1205                 lnet_peer_addref_locked(lp);
1206         }
1207
1208         /* Add peer to global peer list, if necessary */
1209         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1210         if (list_empty(&lp->lp_peer_list)) {
1211                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1212                 ptable->pt_peers++;
1213         }
1214
1215
1216         /* Update peer state */
1217         spin_lock(&lp->lp_lock);
1218         if (flags & LNET_PEER_CONFIGURED) {
1219                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1220                         lp->lp_state |= LNET_PEER_CONFIGURED;
1221         }
1222         if (flags & LNET_PEER_MULTI_RAIL) {
1223                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1224                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1225                         lnet_peer_clr_non_mr_pref_nids(lp);
1226                 }
1227         }
1228         spin_unlock(&lp->lp_lock);
1229
1230         lp->lp_nnis++;
1231         lnet_net_unlock(LNET_LOCK_EX);
1232
1233         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1234                libcfs_nid2str(lp->lp_primary_nid),
1235                libcfs_nid2str(lpni->lpni_nid), flags);
1236
1237         return 0;
1238 }
1239
1240 /*
1241  * Create a new peer, with nid as its primary nid.
1242  *
1243  * Call with the lnet_api_mutex held.
1244  */
1245 static int
1246 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1247 {
1248         struct lnet_peer *lp;
1249         struct lnet_peer_net *lpn;
1250         struct lnet_peer_ni *lpni;
1251         int rc = 0;
1252
1253         LASSERT(nid != LNET_NID_ANY);
1254
1255         /*
1256          * No need for the lnet_net_lock here, because the
1257          * lnet_api_mutex is held.
1258          */
1259         lpni = lnet_find_peer_ni_locked(nid);
1260         if (lpni) {
1261                 /* A peer with this NID already exists. */
1262                 lp = lpni->lpni_peer_net->lpn_peer;
1263                 lnet_peer_ni_decref_locked(lpni);
1264                 /*
1265                  * This is an error if the peer was configured and the
1266                  * primary NID differs or an attempt is made to change
1267                  * the Multi-Rail flag. Otherwise the assumption is
1268                  * that an existing peer is being modified.
1269                  */
1270                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1271                         if (lp->lp_primary_nid != nid)
1272                                 rc = -EEXIST;
1273                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1274                                 rc = -EPERM;
1275                         goto out;
1276                 }
1277                 /* Delete and recreate as a configured peer. */
1278                 lnet_peer_del(lp);
1279         }
1280
1281         /* Create peer, peer_net, and peer_ni. */
1282         rc = -ENOMEM;
1283         lp = lnet_peer_alloc(nid);
1284         if (!lp)
1285                 goto out;
1286         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1287         if (!lpn)
1288                 goto out_free_lp;
1289         lpni = lnet_peer_ni_alloc(nid);
1290         if (!lpni)
1291                 goto out_free_lpn;
1292
1293         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1294
1295 out_free_lpn:
1296         LIBCFS_FREE(lpn, sizeof(*lpn));
1297 out_free_lp:
1298         LIBCFS_FREE(lp, sizeof(*lp));
1299 out:
1300         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1301                libcfs_nid2str(nid), flags, rc);
1302         return rc;
1303 }
1304
1305 /*
1306  * Add a NID to a peer. Call with ln_api_mutex held.
1307  *
1308  * Error codes:
1309  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1310  *  -EEXIST:   The NID was configured by DLC for a different peer.
1311  *  -ENOMEM:   Out of memory.
1312  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1313  *             non-multi-rail peer.
1314  */
1315 static int
1316 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1317 {
1318         struct lnet_peer_net *lpn;
1319         struct lnet_peer_ni *lpni;
1320         int rc = 0;
1321
1322         LASSERT(lp);
1323         LASSERT(nid != LNET_NID_ANY);
1324
1325         /* A configured peer can only be updated through configuration. */
1326         if (!(flags & LNET_PEER_CONFIGURED)) {
1327                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1328                         rc = -EPERM;
1329                         goto out;
1330                 }
1331         }
1332
1333         /*
1334          * The MULTI_RAIL flag can be set but not cleared, because
1335          * that would leave the peer struct in an invalid state.
1336          */
1337         if (flags & LNET_PEER_MULTI_RAIL) {
1338                 spin_lock(&lp->lp_lock);
1339                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1340                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1341                         lnet_peer_clr_non_mr_pref_nids(lp);
1342                 }
1343                 spin_unlock(&lp->lp_lock);
1344         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1345                 rc = -EPERM;
1346                 goto out;
1347         }
1348
1349         lpni = lnet_find_peer_ni_locked(nid);
1350         if (lpni) {
1351                 /*
1352                  * A peer_ni already exists. This is only a problem if
1353                  * it is not connected to this peer and was configured
1354                  * by DLC.
1355                  */
1356                 lnet_peer_ni_decref_locked(lpni);
1357                 if (lpni->lpni_peer_net->lpn_peer == lp)
1358                         goto out;
1359                 if (lnet_peer_ni_is_configured(lpni)) {
1360                         rc = -EEXIST;
1361                         goto out;
1362                 }
1363                 /* If this is the primary NID, destroy the peer. */
1364                 if (lnet_peer_ni_is_primary(lpni)) {
1365                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1366                         lpni = lnet_peer_ni_alloc(nid);
1367                         if (!lpni) {
1368                                 rc = -ENOMEM;
1369                                 goto out;
1370                         }
1371                 }
1372         } else {
1373                 lpni = lnet_peer_ni_alloc(nid);
1374                 if (!lpni) {
1375                         rc = -ENOMEM;
1376                         goto out;
1377                 }
1378         }
1379
1380         /*
1381          * Get the peer_net. Check that we're not adding a second
1382          * peer_ni on a peer_net of a non-multi-rail peer.
1383          */
1384         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1385         if (!lpn) {
1386                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1387                 if (!lpn) {
1388                         rc = -ENOMEM;
1389                         goto out_free_lpni;
1390                 }
1391         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1392                 rc = -ENOTUNIQ;
1393                 goto out_free_lpni;
1394         }
1395
1396         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1397
1398 out_free_lpni:
1399         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1400         if (!lpni->lpni_peer_net)
1401                 LIBCFS_FREE(lpni, sizeof(*lpni));
1402 out:
1403         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1404                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1405                flags, rc);
1406         return rc;
1407 }
1408
1409 /*
1410  * Update the primary NID of a peer, if possible.
1411  *
1412  * Call with the lnet_api_mutex held.
1413  */
1414 static int
1415 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1416 {
1417         lnet_nid_t old = lp->lp_primary_nid;
1418         int rc = 0;
1419
1420         if (lp->lp_primary_nid == nid)
1421                 goto out;
1422         rc = lnet_peer_add_nid(lp, nid, flags);
1423         if (rc)
1424                 goto out;
1425         lp->lp_primary_nid = nid;
1426 out:
1427         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1428                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1429         return rc;
1430 }
1431
1432 /*
1433  * lpni creation initiated due to traffic either sending or receiving.
1434  */
1435 static int
1436 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1437 {
1438         struct lnet_peer *lp;
1439         struct lnet_peer_net *lpn;
1440         struct lnet_peer_ni *lpni;
1441         unsigned flags = 0;
1442         int rc = 0;
1443
1444         if (nid == LNET_NID_ANY) {
1445                 rc = -EINVAL;
1446                 goto out;
1447         }
1448
1449         /* lnet_net_lock is not needed here because ln_api_lock is held */
1450         lpni = lnet_find_peer_ni_locked(nid);
1451         if (lpni) {
1452                 /*
1453                  * We must have raced with another thread. Since we
1454                  * know next to nothing about a peer_ni created by
1455                  * traffic, we just assume everything is ok and
1456                  * return.
1457                  */
1458                 lnet_peer_ni_decref_locked(lpni);
1459                 goto out;
1460         }
1461
1462         /* Create peer, peer_net, and peer_ni. */
1463         rc = -ENOMEM;
1464         lp = lnet_peer_alloc(nid);
1465         if (!lp)
1466                 goto out;
1467         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1468         if (!lpn)
1469                 goto out_free_lp;
1470         lpni = lnet_peer_ni_alloc(nid);
1471         if (!lpni)
1472                 goto out_free_lpn;
1473         if (pref != LNET_NID_ANY)
1474                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1475
1476         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1477
1478 out_free_lpn:
1479         LIBCFS_FREE(lpn, sizeof(*lpn));
1480 out_free_lp:
1481         LIBCFS_FREE(lp, sizeof(*lp));
1482 out:
1483         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1484         return rc;
1485 }
1486
1487 /*
1488  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1489  *
1490  * This API handles the following combinations:
1491  *   Create a peer with its primary NI if only the prim_nid is provided
1492  *   Add a NID to a peer identified by the prim_nid. The peer identified
1493  *   by the prim_nid must already exist.
1494  *   The peer being created may be non-MR.
1495  *
1496  * The caller must hold ln_api_mutex. This prevents the peer from
1497  * being created/modified/deleted by a different thread.
1498  */
1499 int
1500 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1501 {
1502         struct lnet_peer *lp = NULL;
1503         struct lnet_peer_ni *lpni;
1504         unsigned flags;
1505
1506         /* The prim_nid must always be specified */
1507         if (prim_nid == LNET_NID_ANY)
1508                 return -EINVAL;
1509
1510         flags = LNET_PEER_CONFIGURED;
1511         if (mr)
1512                 flags |= LNET_PEER_MULTI_RAIL;
1513
1514         /*
1515          * If nid isn't specified, we must create a new peer with
1516          * prim_nid as its primary nid.
1517          */
1518         if (nid == LNET_NID_ANY)
1519                 return lnet_peer_add(prim_nid, flags);
1520
1521         /* Look up the prim_nid, which must exist. */
1522         lpni = lnet_find_peer_ni_locked(prim_nid);
1523         if (!lpni)
1524                 return -ENOENT;
1525         lnet_peer_ni_decref_locked(lpni);
1526         lp = lpni->lpni_peer_net->lpn_peer;
1527
1528         /* Peer must have been configured. */
1529         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1530                 CDEBUG(D_NET, "peer %s was not configured\n",
1531                        libcfs_nid2str(prim_nid));
1532                 return -ENOENT;
1533         }
1534
1535         /* Primary NID must match */
1536         if (lp->lp_primary_nid != prim_nid) {
1537                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1538                        libcfs_nid2str(prim_nid),
1539                        libcfs_nid2str(lp->lp_primary_nid));
1540                 return -ENODEV;
1541         }
1542
1543         /* Multi-Rail flag must match. */
1544         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1545                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1546                        libcfs_nid2str(prim_nid));
1547                 return -EPERM;
1548         }
1549
1550         return lnet_peer_add_nid(lp, nid, flags);
1551 }
1552
1553 /*
1554  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1555  *
1556  * This API handles the following combinations:
1557  *   Delete a NI from a peer if both prim_nid and nid are provided.
1558  *   Delete a peer if only prim_nid is provided.
1559  *   Delete a peer if its primary nid is provided.
1560  *
1561  * The caller must hold ln_api_mutex. This prevents the peer from
1562  * being modified/deleted by a different thread.
1563  */
1564 int
1565 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1566 {
1567         struct lnet_peer *lp;
1568         struct lnet_peer_ni *lpni;
1569         unsigned flags;
1570
1571         if (prim_nid == LNET_NID_ANY)
1572                 return -EINVAL;
1573
1574         lpni = lnet_find_peer_ni_locked(prim_nid);
1575         if (!lpni)
1576                 return -ENOENT;
1577         lnet_peer_ni_decref_locked(lpni);
1578         lp = lpni->lpni_peer_net->lpn_peer;
1579
1580         if (prim_nid != lp->lp_primary_nid) {
1581                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1582                        libcfs_nid2str(prim_nid),
1583                        libcfs_nid2str(lp->lp_primary_nid));
1584                 return -ENODEV;
1585         }
1586
1587         lnet_net_lock(LNET_LOCK_EX);
1588         if (lp->lp_rtr_refcount > 0) {
1589                 lnet_net_unlock(LNET_LOCK_EX);
1590                 CERROR("%s is a router. Can not be deleted\n",
1591                        libcfs_nid2str(prim_nid));
1592                 return -EBUSY;
1593         }
1594         lnet_net_unlock(LNET_LOCK_EX);
1595
1596         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1597                 return lnet_peer_del(lp);
1598
1599         flags = LNET_PEER_CONFIGURED;
1600         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1601                 flags |= LNET_PEER_MULTI_RAIL;
1602
1603         return lnet_peer_del_nid(lp, nid, flags);
1604 }
1605
1606 void
1607 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1608 {
1609         struct lnet_peer_table *ptable;
1610         struct lnet_peer_net *lpn;
1611
1612         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1613
1614         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1615         LASSERT(list_empty(&lpni->lpni_txq));
1616         LASSERT(lpni->lpni_txqnob == 0);
1617         LASSERT(list_empty(&lpni->lpni_peer_nis));
1618         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1619
1620         lpn = lpni->lpni_peer_net;
1621         lpni->lpni_peer_net = NULL;
1622         lpni->lpni_net = NULL;
1623
1624         /* remove the peer ni from the zombie list */
1625         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1626         spin_lock(&ptable->pt_zombie_lock);
1627         list_del_init(&lpni->lpni_hashlist);
1628         ptable->pt_zombies--;
1629         spin_unlock(&ptable->pt_zombie_lock);
1630
1631         if (lpni->lpni_pref_nnids > 1) {
1632                 LIBCFS_FREE(lpni->lpni_pref.nids,
1633                         sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1634         }
1635         LIBCFS_FREE(lpni, sizeof(*lpni));
1636
1637         lnet_peer_net_decref_locked(lpn);
1638 }
1639
1640 struct lnet_peer_ni *
1641 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1642 {
1643         struct lnet_peer_ni *lpni = NULL;
1644         int rc;
1645
1646         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1647                 return ERR_PTR(-ESHUTDOWN);
1648
1649         /*
1650          * find if a peer_ni already exists.
1651          * If so then just return that.
1652          */
1653         lpni = lnet_find_peer_ni_locked(nid);
1654         if (lpni)
1655                 return lpni;
1656
1657         lnet_net_unlock(cpt);
1658
1659         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1660         if (rc) {
1661                 lpni = ERR_PTR(rc);
1662                 goto out_net_relock;
1663         }
1664
1665         lpni = lnet_find_peer_ni_locked(nid);
1666         LASSERT(lpni);
1667
1668 out_net_relock:
1669         lnet_net_lock(cpt);
1670
1671         return lpni;
1672 }
1673
1674 /*
1675  * Get a peer_ni for the given nid, create it if necessary. Takes a
1676  * hold on the peer_ni.
1677  */
1678 struct lnet_peer_ni *
1679 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1680 {
1681         struct lnet_peer_ni *lpni = NULL;
1682         int rc;
1683
1684         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1685                 return ERR_PTR(-ESHUTDOWN);
1686
1687         /*
1688          * find if a peer_ni already exists.
1689          * If so then just return that.
1690          */
1691         lpni = lnet_find_peer_ni_locked(nid);
1692         if (lpni)
1693                 return lpni;
1694
1695         /*
1696          * Slow path:
1697          * use the lnet_api_mutex to serialize the creation of the peer_ni
1698          * and the creation/deletion of the local ni/net. When a local ni is
1699          * created, if there exists a set of peer_nis on that network,
1700          * they need to be traversed and updated. When a local NI is
1701          * deleted, which could result in a network being deleted, then
1702          * all peer nis on that network need to be removed as well.
1703          *
1704          * Creation through traffic should also be serialized with
1705          * creation through DLC.
1706          */
1707         lnet_net_unlock(cpt);
1708         mutex_lock(&the_lnet.ln_api_mutex);
1709         /*
1710          * Shutdown is only set under the ln_api_lock, so a single
1711          * check here is sufficent.
1712          */
1713         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1714                 lpni = ERR_PTR(-ESHUTDOWN);
1715                 goto out_mutex_unlock;
1716         }
1717
1718         rc = lnet_peer_ni_traffic_add(nid, pref);
1719         if (rc) {
1720                 lpni = ERR_PTR(rc);
1721                 goto out_mutex_unlock;
1722         }
1723
1724         lpni = lnet_find_peer_ni_locked(nid);
1725         LASSERT(lpni);
1726
1727 out_mutex_unlock:
1728         mutex_unlock(&the_lnet.ln_api_mutex);
1729         lnet_net_lock(cpt);
1730
1731         /* Lock has been dropped, check again for shutdown. */
1732         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1733                 if (!IS_ERR(lpni))
1734                         lnet_peer_ni_decref_locked(lpni);
1735                 lpni = ERR_PTR(-ESHUTDOWN);
1736         }
1737
1738         return lpni;
1739 }
1740
1741 /*
1742  * Peer Discovery
1743  */
1744
1745 bool
1746 lnet_peer_gw_discovery(struct lnet_peer *lp)
1747 {
1748         bool rc = false;
1749
1750         spin_lock(&lp->lp_lock);
1751         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1752                 rc = true;
1753         spin_unlock(&lp->lp_lock);
1754
1755         return rc;
1756 }
1757
1758 /*
1759  * Is a peer uptodate from the point of view of discovery?
1760  *
1761  * If it is currently being processed, obviously not.
1762  * A forced Ping or Push is also handled by the discovery thread.
1763  *
1764  * Otherwise look at whether the peer needs rediscovering.
1765  */
1766 bool
1767 lnet_peer_is_uptodate(struct lnet_peer *lp)
1768 {
1769         bool rc;
1770
1771         spin_lock(&lp->lp_lock);
1772         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1773                             LNET_PEER_FORCE_PING |
1774                             LNET_PEER_FORCE_PUSH)) {
1775                 rc = false;
1776         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
1777                 rc = true;
1778         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1779                 if (lnet_peer_discovery_disabled)
1780                         rc = true;
1781                 else
1782                         rc = false;
1783         } else if (lnet_peer_needs_push(lp)) {
1784                 rc = false;
1785         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1786                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1787                         rc = true;
1788                 else
1789                         rc = false;
1790         } else {
1791                 rc = false;
1792         }
1793         spin_unlock(&lp->lp_lock);
1794
1795         return rc;
1796 }
1797
1798 /*
1799  * Queue a peer for the attention of the discovery thread.  Call with
1800  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1801  * -EALREADY if the peer was already queued.
1802  */
1803 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1804 {
1805         int rc;
1806
1807         spin_lock(&lp->lp_lock);
1808         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1809                 lp->lp_state |= LNET_PEER_DISCOVERING;
1810         spin_unlock(&lp->lp_lock);
1811         if (list_empty(&lp->lp_dc_list)) {
1812                 lnet_peer_addref_locked(lp);
1813                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1814                 wake_up(&the_lnet.ln_dc_waitq);
1815                 rc = 0;
1816         } else {
1817                 rc = -EALREADY;
1818         }
1819
1820         CDEBUG(D_NET, "Queue peer %s: %d\n",
1821                libcfs_nid2str(lp->lp_primary_nid), rc);
1822
1823         return rc;
1824 }
1825
1826 /*
1827  * Discovery of a peer is complete. Wake all waiters on the peer.
1828  * Call with lnet_net_lock/EX held.
1829  */
1830 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1831 {
1832         struct lnet_msg *msg, *tmp;
1833         int rc = 0;
1834         struct list_head pending_msgs;
1835
1836         INIT_LIST_HEAD(&pending_msgs);
1837
1838         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1839                libcfs_nid2str(lp->lp_primary_nid));
1840
1841         list_del_init(&lp->lp_dc_list);
1842         spin_lock(&lp->lp_lock);
1843         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1844         spin_unlock(&lp->lp_lock);
1845         wake_up_all(&lp->lp_dc_waitq);
1846
1847         if (lp->lp_rtr_refcount > 0)
1848                 lnet_router_discovery_complete(lp);
1849
1850         lnet_net_unlock(LNET_LOCK_EX);
1851
1852         /* iterate through all pending messages and send them again */
1853         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1854                 list_del_init(&msg->msg_list);
1855                 if (lp->lp_dc_error) {
1856                         lnet_finalize(msg, lp->lp_dc_error);
1857                         continue;
1858                 }
1859
1860                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1861                        lnet_msgtyp2str(msg->msg_type),
1862                        libcfs_id2str(msg->msg_target));
1863                 rc = lnet_send(msg->msg_src_nid_param, msg,
1864                                msg->msg_rtr_nid_param);
1865                 if (rc < 0) {
1866                         CNETERR("Error sending %s to %s: %d\n",
1867                                lnet_msgtyp2str(msg->msg_type),
1868                                libcfs_id2str(msg->msg_target), rc);
1869                         lnet_finalize(msg, rc);
1870                 }
1871         }
1872         lnet_net_lock(LNET_LOCK_EX);
1873         lnet_peer_decref_locked(lp);
1874 }
1875
1876 /*
1877  * Handle inbound push.
1878  * Like any event handler, called with lnet_res_lock/CPT held.
1879  */
1880 void lnet_peer_push_event(struct lnet_event *ev)
1881 {
1882         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1883         struct lnet_peer *lp;
1884
1885         /* lnet_find_peer() adds a refcount */
1886         lp = lnet_find_peer(ev->source.nid);
1887         if (!lp) {
1888                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1889                        libcfs_nid2str(ev->initiator.nid),
1890                        libcfs_nid2str(ev->source.nid));
1891                 return;
1892         }
1893
1894         /* Ensure peer state remains consistent while we modify it. */
1895         spin_lock(&lp->lp_lock);
1896
1897         /*
1898          * If some kind of error happened the contents of the message
1899          * cannot be used. Clear the NIDS_UPTODATE and set the
1900          * FORCE_PING flag to trigger a ping.
1901          */
1902         if (ev->status) {
1903                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1904                 lp->lp_state |= LNET_PEER_FORCE_PING;
1905                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1906                        ev->status,
1907                        libcfs_nid2str(lp->lp_primary_nid),
1908                        libcfs_nid2str(ev->source.nid));
1909                 goto out;
1910         }
1911
1912         /*
1913          * A push with invalid or corrupted info. Clear the UPTODATE
1914          * flag to trigger a ping.
1915          */
1916         if (lnet_ping_info_validate(&pbuf->pb_info)) {
1917                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1918                 lp->lp_state |= LNET_PEER_FORCE_PING;
1919                 CDEBUG(D_NET, "Corrupted Push from %s\n",
1920                        libcfs_nid2str(lp->lp_primary_nid));
1921                 goto out;
1922         }
1923
1924         /*
1925          * Make sure we'll allocate the correct size ping buffer when
1926          * pinging the peer.
1927          */
1928         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
1929                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
1930
1931         /*
1932          * A non-Multi-Rail peer is not supposed to be capable of
1933          * sending a push.
1934          */
1935         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
1936                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
1937                        libcfs_nid2str(lp->lp_primary_nid));
1938                 goto out;
1939         }
1940
1941         /*
1942          * Check the MULTIRAIL flag. Complain if the peer was DLC
1943          * configured without it.
1944          */
1945         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1946                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1947                         CERROR("Push says %s is Multi-Rail, DLC says not\n",
1948                                libcfs_nid2str(lp->lp_primary_nid));
1949                 } else {
1950                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1951                         lnet_peer_clr_non_mr_pref_nids(lp);
1952                 }
1953         }
1954
1955         /*
1956          * The peer may have discovery disabled at its end. Set
1957          * NO_DISCOVERY as appropriate.
1958          */
1959         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
1960                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
1961                        libcfs_nid2str(lp->lp_primary_nid));
1962                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
1963         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
1964                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
1965                        libcfs_nid2str(lp->lp_primary_nid));
1966                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
1967         }
1968
1969         /*
1970          * Check for truncation of the Put message. Clear the
1971          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
1972          * and tell discovery to allocate a bigger buffer.
1973          */
1974         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
1975                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
1976                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
1977                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1978                 lp->lp_state |= LNET_PEER_FORCE_PING;
1979                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
1980                        libcfs_nid2str(lp->lp_primary_nid),
1981                        pbuf->pb_info.pi_nnis);
1982                 goto out;
1983         }
1984
1985         /*
1986          * Check whether the Put data is stale. Stale data can just be
1987          * dropped.
1988          */
1989         if (pbuf->pb_info.pi_nnis > 1 &&
1990             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid &&
1991             LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno) {
1992                 CDEBUG(D_NET, "Stale Push from %s: got %u have %u\n",
1993                        libcfs_nid2str(lp->lp_primary_nid),
1994                        LNET_PING_BUFFER_SEQNO(pbuf),
1995                        lp->lp_peer_seqno);
1996                 goto out;
1997         }
1998
1999         /*
2000          * Check whether the Put data is new, in which case we clear
2001          * the UPTODATE flag and prepare to process it.
2002          *
2003          * If the Put data is current, and the peer is UPTODATE then
2004          * we assome everything is all right and drop the data as
2005          * stale.
2006          */
2007         if (LNET_PING_BUFFER_SEQNO(pbuf) > lp->lp_peer_seqno) {
2008                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2009                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2010         } else if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
2011                 CDEBUG(D_NET, "Stale Push from %s: got %u have %u\n",
2012                        libcfs_nid2str(lp->lp_primary_nid),
2013                        LNET_PING_BUFFER_SEQNO(pbuf),
2014                        lp->lp_peer_seqno);
2015                 goto out;
2016         }
2017
2018         /*
2019          * If there is data present that hasn't been processed yet,
2020          * we'll replace it if the Put contained newer data and it
2021          * fits. We're racing with a Ping or earlier Push in this
2022          * case.
2023          */
2024         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2025                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2026                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2027                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2028                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2029                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2030                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2031                               libcfs_nid2str(lp->lp_primary_nid),
2032                               LNET_PING_BUFFER_SEQNO(pbuf),
2033                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2034                 }
2035                 goto out;
2036         }
2037
2038         /*
2039          * Allocate a buffer to copy the data. On a failure we drop
2040          * the Push and set FORCE_PING to force the discovery
2041          * thread to fix the problem by pinging the peer.
2042          */
2043         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2044         if (!lp->lp_data) {
2045                 lp->lp_state |= LNET_PEER_FORCE_PING;
2046                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2047                        libcfs_nid2str(lp->lp_primary_nid),
2048                        LNET_PING_BUFFER_SEQNO(pbuf));
2049                 goto out;
2050         }
2051
2052         /* Success */
2053         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2054                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2055         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2056         CDEBUG(D_NET, "Received Push %s %u\n",
2057                libcfs_nid2str(lp->lp_primary_nid),
2058                LNET_PING_BUFFER_SEQNO(pbuf));
2059
2060 out:
2061         /*
2062          * Queue the peer for discovery if not done, force it on the request
2063          * queue and wake the discovery thread if the peer was already queued,
2064          * because its status changed.
2065          */
2066         spin_unlock(&lp->lp_lock);
2067         lnet_net_lock(LNET_LOCK_EX);
2068         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2069                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2070                 wake_up(&the_lnet.ln_dc_waitq);
2071         }
2072         /* Drop refcount from lookup */
2073         lnet_peer_decref_locked(lp);
2074         lnet_net_unlock(LNET_LOCK_EX);
2075 }
2076
2077 /*
2078  * Clear the discovery error state, unless we're already discovering
2079  * this peer, in which case the error is current.
2080  */
2081 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2082 {
2083         spin_lock(&lp->lp_lock);
2084         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2085                 lp->lp_dc_error = 0;
2086         spin_unlock(&lp->lp_lock);
2087 }
2088
2089 /*
2090  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2091  * dropped/retaken within this function. An lnet_peer_ni is passed in
2092  * because discovery could tear down an lnet_peer.
2093  */
2094 int
2095 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2096 {
2097         DEFINE_WAIT(wait);
2098         struct lnet_peer *lp;
2099         int rc = 0;
2100
2101 again:
2102         lnet_net_unlock(cpt);
2103         lnet_net_lock(LNET_LOCK_EX);
2104         lp = lpni->lpni_peer_net->lpn_peer;
2105         lnet_peer_clear_discovery_error(lp);
2106
2107         /*
2108          * We're willing to be interrupted. The lpni can become a
2109          * zombie if we race with DLC, so we must check for that.
2110          */
2111         for (;;) {
2112                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2113                 if (signal_pending(current))
2114                         break;
2115                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2116                         break;
2117                 if (lp->lp_dc_error)
2118                         break;
2119                 if (lnet_peer_is_uptodate(lp))
2120                         break;
2121                 lnet_peer_queue_for_discovery(lp);
2122                 /*
2123                  * if caller requested a non-blocking operation then
2124                  * return immediately. Once discovery is complete then the
2125                  * peer ref will be decremented and any pending messages
2126                  * that were stopped due to discovery will be transmitted.
2127                  */
2128                 if (!block)
2129                         break;
2130
2131                 lnet_peer_addref_locked(lp);
2132                 lnet_net_unlock(LNET_LOCK_EX);
2133                 schedule();
2134                 finish_wait(&lp->lp_dc_waitq, &wait);
2135                 lnet_net_lock(LNET_LOCK_EX);
2136                 lnet_peer_decref_locked(lp);
2137                 /* Peer may have changed */
2138                 lp = lpni->lpni_peer_net->lpn_peer;
2139         }
2140         finish_wait(&lp->lp_dc_waitq, &wait);
2141
2142         lnet_net_unlock(LNET_LOCK_EX);
2143         lnet_net_lock(cpt);
2144
2145         /*
2146          * If the peer has changed after we've discovered the older peer,
2147          * then we need to discovery the new peer to make sure the
2148          * interface information is up to date
2149          */
2150         if (lp != lpni->lpni_peer_net->lpn_peer)
2151                 goto again;
2152
2153         if (signal_pending(current))
2154                 rc = -EINTR;
2155         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2156                 rc = -ESHUTDOWN;
2157         else if (lp->lp_dc_error)
2158                 rc = lp->lp_dc_error;
2159         else if (!block)
2160                 CDEBUG(D_NET, "non-blocking discovery\n");
2161         else if (!lnet_peer_is_uptodate(lp))
2162                 goto again;
2163
2164         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2165                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2166                libcfs_nid2str(lpni->lpni_nid), rc,
2167                (!block) ? "pending discovery" : "discovery complete");
2168
2169         return rc;
2170 }
2171
2172 /* Handle an incoming ack for a push. */
2173 static void
2174 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2175 {
2176         struct lnet_ping_buffer *pbuf;
2177
2178         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2179         spin_lock(&lp->lp_lock);
2180         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2181         lp->lp_push_error = ev->status;
2182         if (ev->status)
2183                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2184         else
2185                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2186         spin_unlock(&lp->lp_lock);
2187
2188         CDEBUG(D_NET, "peer %s ev->status %d\n",
2189                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2190 }
2191
2192 /* Handle a Reply message. This is the reply to a Ping message. */
2193 static void
2194 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2195 {
2196         struct lnet_ping_buffer *pbuf;
2197         int rc;
2198
2199         spin_lock(&lp->lp_lock);
2200
2201         /*
2202          * If some kind of error happened the contents of message
2203          * cannot be used. Set PING_FAILED to trigger a retry.
2204          */
2205         if (ev->status) {
2206                 lp->lp_state |= LNET_PEER_PING_FAILED;
2207                 lp->lp_ping_error = ev->status;
2208                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2209                        ev->status,
2210                        libcfs_nid2str(lp->lp_primary_nid),
2211                        libcfs_nid2str(ev->source.nid));
2212                 goto out;
2213         }
2214
2215         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2216         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2217                 lnet_swap_pinginfo(pbuf);
2218
2219         /*
2220          * A reply with invalid or corrupted info. Set PING_FAILED to
2221          * trigger a retry.
2222          */
2223         rc = lnet_ping_info_validate(&pbuf->pb_info);
2224         if (rc) {
2225                 lp->lp_state |= LNET_PEER_PING_FAILED;
2226                 lp->lp_ping_error = 0;
2227                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2228                        libcfs_nid2str(lp->lp_primary_nid), rc);
2229                 goto out;
2230         }
2231
2232         /*
2233          * Update the MULTI_RAIL flag based on the reply. If the peer
2234          * was configured with DLC then the setting should match what
2235          * DLC put in.
2236          */
2237         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2238                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2239                         /* Everything's fine */
2240                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2241                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2242                               libcfs_nid2str(lp->lp_primary_nid));
2243                 } else {
2244                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2245                         lnet_peer_clr_non_mr_pref_nids(lp);
2246                 }
2247         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2248                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2249                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2250                               libcfs_nid2str(lp->lp_primary_nid));
2251                 } else {
2252                         CERROR("Multi-Rail state vanished from %s\n",
2253                                libcfs_nid2str(lp->lp_primary_nid));
2254                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2255                 }
2256         }
2257
2258         /*
2259          * Make sure we'll allocate the correct size ping buffer when
2260          * pinging the peer.
2261          */
2262         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2263                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2264
2265         /*
2266          * The peer may have discovery disabled at its end. Set
2267          * NO_DISCOVERY as appropriate.
2268          */
2269         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2270                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2271                        libcfs_nid2str(lp->lp_primary_nid));
2272                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2273         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2274                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2275                        libcfs_nid2str(lp->lp_primary_nid));
2276                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2277         }
2278
2279         /*
2280          * Check for truncation of the Reply. Clear PING_SENT and set
2281          * PING_FAILED to trigger a retry.
2282          */
2283         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2284                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2285                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2286                 lp->lp_state |= LNET_PEER_PING_FAILED;
2287                 lp->lp_ping_error = 0;
2288                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2289                        libcfs_nid2str(lp->lp_primary_nid),
2290                        pbuf->pb_info.pi_nnis);
2291                 goto out;
2292         }
2293
2294         /*
2295          * Check the sequence numbers in the reply. These are only
2296          * available if the reply came from a Multi-Rail peer.
2297          */
2298         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2299             pbuf->pb_info.pi_nnis > 1 &&
2300             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2301                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno) {
2302                         CDEBUG(D_NET, "Stale Reply from %s: got %u have %u\n",
2303                                 libcfs_nid2str(lp->lp_primary_nid),
2304                                 LNET_PING_BUFFER_SEQNO(pbuf),
2305                                 lp->lp_peer_seqno);
2306                         goto out;
2307                 }
2308
2309                 if (LNET_PING_BUFFER_SEQNO(pbuf) > lp->lp_peer_seqno)
2310                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2311         }
2312
2313         /* We're happy with the state of the data in the buffer. */
2314         CDEBUG(D_NET, "peer %s data present %u\n",
2315                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno);
2316         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2317                 lnet_ping_buffer_decref(lp->lp_data);
2318         else
2319                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2320         lnet_ping_buffer_addref(pbuf);
2321         lp->lp_data = pbuf;
2322 out:
2323         lp->lp_state &= ~LNET_PEER_PING_SENT;
2324         spin_unlock(&lp->lp_lock);
2325 }
2326
2327 /*
2328  * Send event handling. Only matters for error cases, where we clean
2329  * up state on the peer and peer_ni that would otherwise be updated in
2330  * the REPLY event handler for a successful Ping, and the ACK event
2331  * handler for a successful Push.
2332  */
2333 static int
2334 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2335 {
2336         int rc = 0;
2337
2338         if (!ev->status)
2339                 goto out;
2340
2341         spin_lock(&lp->lp_lock);
2342         if (ev->msg_type == LNET_MSG_GET) {
2343                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2344                 lp->lp_state |= LNET_PEER_PING_FAILED;
2345                 lp->lp_ping_error = ev->status;
2346         } else { /* ev->msg_type == LNET_MSG_PUT */
2347                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2348                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2349                 lp->lp_push_error = ev->status;
2350         }
2351         spin_unlock(&lp->lp_lock);
2352         rc = LNET_REDISCOVER_PEER;
2353 out:
2354         CDEBUG(D_NET, "%s Send to %s: %d\n",
2355                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2356                 libcfs_nid2str(ev->target.nid), rc);
2357         return rc;
2358 }
2359
2360 /*
2361  * Unlink event handling. This event is only seen if a call to
2362  * LNetMDUnlink() caused the event to be unlinked. If this call was
2363  * made after the event was set up in LNetGet() or LNetPut() then we
2364  * assume the Ping or Push timed out.
2365  */
2366 static void
2367 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2368 {
2369         spin_lock(&lp->lp_lock);
2370         /* We've passed through LNetGet() */
2371         if (lp->lp_state & LNET_PEER_PING_SENT) {
2372                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2373                 lp->lp_state |= LNET_PEER_PING_FAILED;
2374                 lp->lp_ping_error = -ETIMEDOUT;
2375                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2376                         libcfs_nid2str(lp->lp_primary_nid));
2377         }
2378         /* We've passed through LNetPut() */
2379         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2380                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2381                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2382                 lp->lp_push_error = -ETIMEDOUT;
2383                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2384                         libcfs_nid2str(lp->lp_primary_nid));
2385         }
2386         spin_unlock(&lp->lp_lock);
2387 }
2388
2389 /*
2390  * Event handler for the discovery EQ.
2391  *
2392  * Called with lnet_res_lock(cpt) held. The cpt is the
2393  * lnet_cpt_of_cookie() of the md handle cookie.
2394  */
2395 static void lnet_discovery_event_handler(struct lnet_event *event)
2396 {
2397         struct lnet_peer *lp = event->md.user_ptr;
2398         struct lnet_ping_buffer *pbuf;
2399         int rc;
2400
2401         /* discovery needs to take another look */
2402         rc = LNET_REDISCOVER_PEER;
2403
2404         CDEBUG(D_NET, "Received event: %d\n", event->type);
2405
2406         switch (event->type) {
2407         case LNET_EVENT_ACK:
2408                 lnet_discovery_event_ack(lp, event);
2409                 break;
2410         case LNET_EVENT_REPLY:
2411                 lnet_discovery_event_reply(lp, event);
2412                 break;
2413         case LNET_EVENT_SEND:
2414                 /* Only send failure triggers a retry. */
2415                 rc = lnet_discovery_event_send(lp, event);
2416                 break;
2417         case LNET_EVENT_UNLINK:
2418                 /* LNetMDUnlink() was called */
2419                 lnet_discovery_event_unlink(lp, event);
2420                 break;
2421         default:
2422                 /* Invalid events. */
2423                 LBUG();
2424         }
2425         lnet_net_lock(LNET_LOCK_EX);
2426         if (event->unlinked) {
2427                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2428                 lnet_ping_buffer_decref(pbuf);
2429                 lnet_peer_decref_locked(lp);
2430         }
2431
2432         /* put peer back at end of request queue, if discovery not already
2433          * done */
2434         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2435                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2436                 wake_up(&the_lnet.ln_dc_waitq);
2437         }
2438         lnet_net_unlock(LNET_LOCK_EX);
2439 }
2440
2441 /*
2442  * Build a peer from incoming data.
2443  *
2444  * The NIDs in the incoming data are supposed to be structured as follows:
2445  *  - loopback
2446  *  - primary NID
2447  *  - other NIDs in same net
2448  *  - NIDs in second net
2449  *  - NIDs in third net
2450  *  - ...
2451  * This due to the way the list of NIDs in the data is created.
2452  *
2453  * Note that this function will mark the peer uptodate unless an
2454  * ENOMEM is encontered. All other errors are due to a conflict
2455  * between the DLC configuration and what discovery sees. We treat DLC
2456  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2457  * peer from becoming stuck in discovery.
2458  */
2459 static int lnet_peer_merge_data(struct lnet_peer *lp,
2460                                 struct lnet_ping_buffer *pbuf)
2461 {
2462         struct lnet_peer_ni *lpni;
2463         lnet_nid_t *curnis = NULL;
2464         struct lnet_ni_status *addnis = NULL;
2465         lnet_nid_t *delnis = NULL;
2466         unsigned flags;
2467         int ncurnis;
2468         int naddnis;
2469         int ndelnis;
2470         int nnis = 0;
2471         int i;
2472         int j;
2473         int rc;
2474
2475         flags = LNET_PEER_DISCOVERED;
2476         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2477                 flags |= LNET_PEER_MULTI_RAIL;
2478
2479         /*
2480          * Cache the routing feature for the peer; whether it is enabled
2481          * for disabled as reported by the remote peer.
2482          */
2483         spin_lock(&lp->lp_lock);
2484         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2485                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2486         else
2487                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2488         spin_unlock(&lp->lp_lock);
2489
2490         nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2491         LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2492         LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2493         LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2494         if (!curnis || !addnis || !delnis) {
2495                 rc = -ENOMEM;
2496                 goto out;
2497         }
2498         ncurnis = 0;
2499         naddnis = 0;
2500         ndelnis = 0;
2501
2502         /* Construct the list of NIDs present in peer. */
2503         lpni = NULL;
2504         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2505                 curnis[ncurnis++] = lpni->lpni_nid;
2506
2507         /*
2508          * Check for NIDs in pbuf not present in curnis[].
2509          * The loop starts at 1 to skip the loopback NID.
2510          */
2511         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2512                 for (j = 0; j < ncurnis; j++)
2513                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2514                                 break;
2515                 if (j == ncurnis)
2516                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2517         }
2518         /*
2519          * Check for NIDs in curnis[] not present in pbuf.
2520          * The nested loop starts at 1 to skip the loopback NID.
2521          *
2522          * But never add the loopback NID to delnis[]: if it is
2523          * present in curnis[] then this peer is for this node.
2524          */
2525         for (i = 0; i < ncurnis; i++) {
2526                 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2527                         continue;
2528                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2529                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2530                                 /*
2531                                  * update the information we cache for the
2532                                  * peer with the latest information we
2533                                  * received
2534                                  */
2535                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2536                                 if (lpni) {
2537                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2538                                         lnet_peer_ni_decref_locked(lpni);
2539                                 }
2540                                 break;
2541                         }
2542                 }
2543                 if (j == pbuf->pb_info.pi_nnis)
2544                         delnis[ndelnis++] = curnis[i];
2545         }
2546
2547         for (i = 0; i < naddnis; i++) {
2548                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2549                 if (rc) {
2550                         CERROR("Error adding NID %s to peer %s: %d\n",
2551                                libcfs_nid2str(addnis[i].ns_nid),
2552                                libcfs_nid2str(lp->lp_primary_nid), rc);
2553                         if (rc == -ENOMEM)
2554                                 goto out;
2555                 }
2556                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2557                 if (lpni) {
2558                         lpni->lpni_ns_status = addnis[i].ns_status;
2559                         lnet_peer_ni_decref_locked(lpni);
2560                 }
2561         }
2562
2563         for (i = 0; i < ndelnis; i++) {
2564                 /*
2565                  * for routers it's okay to delete the primary_nid because
2566                  * the upper layers don't really rely on it. So if we're
2567                  * being told that the router changed its primary_nid
2568                  * then it's okay to delete it.
2569                  */
2570                 if (lp->lp_rtr_refcount > 0)
2571                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2572                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2573                 if (rc) {
2574                         CERROR("Error deleting NID %s from peer %s: %d\n",
2575                                libcfs_nid2str(delnis[i]),
2576                                libcfs_nid2str(lp->lp_primary_nid), rc);
2577                         if (rc == -ENOMEM)
2578                                 goto out;
2579                 }
2580         }
2581         /*
2582          * Errors other than -ENOMEM are due to peers having been
2583          * configured with DLC. Ignore these because DLC overrides
2584          * Discovery.
2585          */
2586         rc = 0;
2587 out:
2588         LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2589         LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2590         LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2591         lnet_ping_buffer_decref(pbuf);
2592         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2593
2594         if (rc) {
2595                 spin_lock(&lp->lp_lock);
2596                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2597                 lp->lp_state |= LNET_PEER_FORCE_PING;
2598                 spin_unlock(&lp->lp_lock);
2599         }
2600         return rc;
2601 }
2602
2603 /*
2604  * The data in pbuf says lp is its primary peer, but the data was
2605  * received by a different peer. Try to update lp with the data.
2606  */
2607 static int
2608 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2609 {
2610         struct lnet_handle_md mdh;
2611
2612         /* Queue lp for discovery, and force it on the request queue. */
2613         lnet_net_lock(LNET_LOCK_EX);
2614         if (lnet_peer_queue_for_discovery(lp))
2615                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2616         lnet_net_unlock(LNET_LOCK_EX);
2617
2618         LNetInvalidateMDHandle(&mdh);
2619
2620         /*
2621          * Decide whether we can move the peer to the DATA_PRESENT state.
2622          *
2623          * We replace stale data for a multi-rail peer, repair PING_FAILED
2624          * status, and preempt FORCE_PING.
2625          *
2626          * If after that we have DATA_PRESENT, we merge it into this peer.
2627          */
2628         spin_lock(&lp->lp_lock);
2629         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2630                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2631                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2632                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2633                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2634                         lnet_ping_buffer_decref(pbuf);
2635                         pbuf = lp->lp_data;
2636                         lp->lp_data = NULL;
2637                 }
2638         }
2639         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2640                 lnet_ping_buffer_decref(lp->lp_data);
2641                 lp->lp_data = NULL;
2642                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2643         }
2644         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2645                 mdh = lp->lp_ping_mdh;
2646                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2647                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2648                 lp->lp_ping_error = 0;
2649         }
2650         if (lp->lp_state & LNET_PEER_FORCE_PING)
2651                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2652         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2653         spin_unlock(&lp->lp_lock);
2654
2655         if (!LNetMDHandleIsInvalid(mdh))
2656                 LNetMDUnlink(mdh);
2657
2658         if (pbuf)
2659                 return lnet_peer_merge_data(lp, pbuf);
2660
2661         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2662         return 0;
2663 }
2664
2665 /*
2666  * Update a peer using the data received.
2667  */
2668 static int lnet_peer_data_present(struct lnet_peer *lp)
2669 __must_hold(&lp->lp_lock)
2670 {
2671         struct lnet_ping_buffer *pbuf;
2672         struct lnet_peer_ni *lpni;
2673         lnet_nid_t nid = LNET_NID_ANY;
2674         unsigned flags;
2675         int rc = 0;
2676
2677         pbuf = lp->lp_data;
2678         lp->lp_data = NULL;
2679         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2680         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2681         spin_unlock(&lp->lp_lock);
2682
2683         /*
2684          * Modifications of peer structures are done while holding the
2685          * ln_api_mutex. A global lock is required because we may be
2686          * modifying multiple peer structures, and a mutex greatly
2687          * simplifies memory management.
2688          *
2689          * The actual changes to the data structures must also protect
2690          * against concurrent lookups, for which the lnet_net_lock in
2691          * LNET_LOCK_EX mode is used.
2692          */
2693         mutex_lock(&the_lnet.ln_api_mutex);
2694         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2695                 rc = -ESHUTDOWN;
2696                 goto out;
2697         }
2698
2699         /*
2700          * If this peer is not on the peer list then it is being torn
2701          * down, and our reference count may be all that is keeping it
2702          * alive. Don't do any work on it.
2703          */
2704         if (list_empty(&lp->lp_peer_list))
2705                 goto out;
2706
2707         flags = LNET_PEER_DISCOVERED;
2708         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2709                 flags |= LNET_PEER_MULTI_RAIL;
2710
2711         /*
2712          * Check whether the primary NID in the message matches the
2713          * primary NID of the peer. If it does, update the peer, if
2714          * it it does not, check whether there is already a peer with
2715          * that primary NID. If no such peer exists, try to update
2716          * the primary NID of the current peer (allowed if it was
2717          * created due to message traffic) and complete the update.
2718          * If the peer did exist, hand off the data to it.
2719          *
2720          * The peer for the loopback interface is a special case: this
2721          * is the peer for the local node, and we want to set its
2722          * primary NID to the correct value here. Moreover, this peer
2723          * can show up with only the loopback NID in the ping buffer.
2724          */
2725         if (pbuf->pb_info.pi_nnis <= 1)
2726                 goto out;
2727         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2728         if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2729                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2730                 if (!rc)
2731                         rc = lnet_peer_merge_data(lp, pbuf);
2732         } else if (lp->lp_primary_nid == nid) {
2733                 rc = lnet_peer_merge_data(lp, pbuf);
2734         } else {
2735                 lpni = lnet_find_peer_ni_locked(nid);
2736                 if (!lpni) {
2737                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2738                         if (rc) {
2739                                 CERROR("Primary NID error %s versus %s: %d\n",
2740                                        libcfs_nid2str(lp->lp_primary_nid),
2741                                        libcfs_nid2str(nid), rc);
2742                         } else {
2743                                 rc = lnet_peer_merge_data(lp, pbuf);
2744                         }
2745                 } else {
2746                         struct lnet_peer *new_lp;
2747                         new_lp = lpni->lpni_peer_net->lpn_peer;
2748                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2749                         lnet_consolidate_routes_locked(lp, new_lp);
2750                         lnet_peer_ni_decref_locked(lpni);
2751                 }
2752         }
2753 out:
2754         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2755         mutex_unlock(&the_lnet.ln_api_mutex);
2756
2757         spin_lock(&lp->lp_lock);
2758         /* Tell discovery to re-check the peer immediately. */
2759         if (!rc)
2760                 rc = LNET_REDISCOVER_PEER;
2761         return rc;
2762 }
2763
2764 /*
2765  * A ping failed. Clear the PING_FAILED state and set the
2766  * FORCE_PING state, to ensure a retry even if discovery is
2767  * disabled. This avoids being left with incorrect state.
2768  */
2769 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2770 __must_hold(&lp->lp_lock)
2771 {
2772         struct lnet_handle_md mdh;
2773         int rc;
2774
2775         mdh = lp->lp_ping_mdh;
2776         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2777         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2778         lp->lp_state |= LNET_PEER_FORCE_PING;
2779         rc = lp->lp_ping_error;
2780         lp->lp_ping_error = 0;
2781         spin_unlock(&lp->lp_lock);
2782
2783         if (!LNetMDHandleIsInvalid(mdh))
2784                 LNetMDUnlink(mdh);
2785
2786         CDEBUG(D_NET, "peer %s:%d\n",
2787                libcfs_nid2str(lp->lp_primary_nid), rc);
2788
2789         spin_lock(&lp->lp_lock);
2790         return rc ? rc : LNET_REDISCOVER_PEER;
2791 }
2792
2793 /*
2794  * Select NID to send a Ping or Push to.
2795  */
2796 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2797 {
2798         struct lnet_peer_ni *lpni;
2799
2800         /* Look for a direct-connected NID for this peer. */
2801         lpni = NULL;
2802         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2803                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2804                         continue;
2805                 break;
2806         }
2807         if (lpni)
2808                 return lpni->lpni_nid;
2809
2810         /* Look for a routed-connected NID for this peer. */
2811         lpni = NULL;
2812         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2813                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2814                         continue;
2815                 break;
2816         }
2817         if (lpni)
2818                 return lpni->lpni_nid;
2819
2820         return LNET_NID_ANY;
2821 }
2822
2823 /* Active side of ping. */
2824 static int lnet_peer_send_ping(struct lnet_peer *lp)
2825 __must_hold(&lp->lp_lock)
2826 {
2827         lnet_nid_t pnid;
2828         int nnis;
2829         int rc;
2830         int cpt;
2831
2832         lp->lp_state |= LNET_PEER_PING_SENT;
2833         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2834         spin_unlock(&lp->lp_lock);
2835
2836         cpt = lnet_net_lock_current();
2837         /* Refcount for MD. */
2838         lnet_peer_addref_locked(lp);
2839         pnid = lnet_peer_select_nid(lp);
2840         lnet_net_unlock(cpt);
2841
2842         nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2843
2844         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2845                             the_lnet.ln_dc_eqh, false);
2846
2847         /*
2848          * if LNetMDBind in lnet_send_ping fails we need to decrement the
2849          * refcount on the peer, otherwise LNetMDUnlink will be called
2850          * which will eventually do that.
2851          */
2852         if (rc > 0) {
2853                 lnet_net_lock(cpt);
2854                 lnet_peer_decref_locked(lp);
2855                 lnet_net_unlock(cpt);
2856                 rc = -rc; /* change the rc to negative value */
2857                 goto fail_error;
2858         } else if (rc < 0) {
2859                 goto fail_error;
2860         }
2861
2862         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2863
2864         spin_lock(&lp->lp_lock);
2865         return 0;
2866
2867 fail_error:
2868         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2869         /*
2870          * The errors that get us here are considered hard errors and
2871          * cause Discovery to terminate. So we clear PING_SENT, but do
2872          * not set either PING_FAILED or FORCE_PING. In fact we need
2873          * to clear PING_FAILED, because the unlink event handler will
2874          * have set it if we called LNetMDUnlink() above.
2875          */
2876         spin_lock(&lp->lp_lock);
2877         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
2878         return rc;
2879 }
2880
2881 /*
2882  * This function exists because you cannot call LNetMDUnlink() from an
2883  * event handler.
2884  */
2885 static int lnet_peer_push_failed(struct lnet_peer *lp)
2886 __must_hold(&lp->lp_lock)
2887 {
2888         struct lnet_handle_md mdh;
2889         int rc;
2890
2891         mdh = lp->lp_push_mdh;
2892         LNetInvalidateMDHandle(&lp->lp_push_mdh);
2893         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
2894         rc = lp->lp_push_error;
2895         lp->lp_push_error = 0;
2896         spin_unlock(&lp->lp_lock);
2897
2898         if (!LNetMDHandleIsInvalid(mdh))
2899                 LNetMDUnlink(mdh);
2900
2901         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2902         spin_lock(&lp->lp_lock);
2903         return rc ? rc : LNET_REDISCOVER_PEER;
2904 }
2905
2906 /* Active side of push. */
2907 static int lnet_peer_send_push(struct lnet_peer *lp)
2908 __must_hold(&lp->lp_lock)
2909 {
2910         struct lnet_ping_buffer *pbuf;
2911         struct lnet_process_id id;
2912         struct lnet_md md;
2913         int cpt;
2914         int rc;
2915
2916         /* Don't push to a non-multi-rail peer. */
2917         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2918                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
2919                 return 0;
2920         }
2921
2922         lp->lp_state |= LNET_PEER_PUSH_SENT;
2923         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
2924         spin_unlock(&lp->lp_lock);
2925
2926         cpt = lnet_net_lock_current();
2927         pbuf = the_lnet.ln_ping_target;
2928         lnet_ping_buffer_addref(pbuf);
2929         lnet_net_unlock(cpt);
2930
2931         /* Push source MD */
2932         md.start     = &pbuf->pb_info;
2933         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
2934         md.threshold = 2; /* Put/Ack */
2935         md.max_size  = 0;
2936         md.options   = 0;
2937         md.eq_handle = the_lnet.ln_dc_eqh;
2938         md.user_ptr  = lp;
2939
2940         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
2941         if (rc) {
2942                 lnet_ping_buffer_decref(pbuf);
2943                 CERROR("Can't bind push source MD: %d\n", rc);
2944                 goto fail_error;
2945         }
2946         cpt = lnet_net_lock_current();
2947         /* Refcount for MD. */
2948         lnet_peer_addref_locked(lp);
2949         id.pid = LNET_PID_LUSTRE;
2950         id.nid = lnet_peer_select_nid(lp);
2951         lnet_net_unlock(cpt);
2952
2953         if (id.nid == LNET_NID_ANY) {
2954                 rc = -EHOSTUNREACH;
2955                 goto fail_unlink;
2956         }
2957
2958         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
2959                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
2960                      LNET_PROTO_PING_MATCHBITS, 0, 0);
2961
2962         if (rc)
2963                 goto fail_unlink;
2964
2965         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2966
2967         spin_lock(&lp->lp_lock);
2968         return 0;
2969
2970 fail_unlink:
2971         LNetMDUnlink(lp->lp_push_mdh);
2972         LNetInvalidateMDHandle(&lp->lp_push_mdh);
2973 fail_error:
2974         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2975         /*
2976          * The errors that get us here are considered hard errors and
2977          * cause Discovery to terminate. So we clear PUSH_SENT, but do
2978          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
2979          * because the unlink event handler will have set it if we
2980          * called LNetMDUnlink() above.
2981          */
2982         spin_lock(&lp->lp_lock);
2983         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
2984         return rc;
2985 }
2986
2987 /*
2988  * An unrecoverable error was encountered during discovery.
2989  * Set error status in peer and abort discovery.
2990  */
2991 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
2992 {
2993         CDEBUG(D_NET, "Discovery error %s: %d\n",
2994                libcfs_nid2str(lp->lp_primary_nid), error);
2995
2996         spin_lock(&lp->lp_lock);
2997         lp->lp_dc_error = error;
2998         lp->lp_state &= ~LNET_PEER_DISCOVERING;
2999         lp->lp_state |= LNET_PEER_REDISCOVER;
3000         spin_unlock(&lp->lp_lock);
3001 }
3002
3003 /*
3004  * Mark the peer as discovered.
3005  */
3006 static int lnet_peer_discovered(struct lnet_peer *lp)
3007 __must_hold(&lp->lp_lock)
3008 {
3009         lp->lp_state |= LNET_PEER_DISCOVERED;
3010         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3011                           LNET_PEER_REDISCOVER);
3012
3013         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3014
3015         return 0;
3016 }
3017
3018 /*
3019  * Mark the peer as to be rediscovered.
3020  */
3021 static int lnet_peer_rediscover(struct lnet_peer *lp)
3022 __must_hold(&lp->lp_lock)
3023 {
3024         lp->lp_state |= LNET_PEER_REDISCOVER;
3025         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3026
3027         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3028
3029         return 0;
3030 }
3031
3032 /*
3033  * Discovering this peer is taking too long. Cancel any Ping or Push
3034  * that discovery is waiting on by unlinking the relevant MDs. The
3035  * lnet_discovery_event_handler() will proceed from here and complete
3036  * the cleanup.
3037  */
3038 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3039 {
3040         struct lnet_handle_md ping_mdh;
3041         struct lnet_handle_md push_mdh;
3042
3043         LNetInvalidateMDHandle(&ping_mdh);
3044         LNetInvalidateMDHandle(&push_mdh);
3045
3046         spin_lock(&lp->lp_lock);
3047         if (lp->lp_state & LNET_PEER_PING_SENT) {
3048                 ping_mdh = lp->lp_ping_mdh;
3049                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3050         }
3051         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3052                 push_mdh = lp->lp_push_mdh;
3053                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3054         }
3055         spin_unlock(&lp->lp_lock);
3056
3057         if (!LNetMDHandleIsInvalid(ping_mdh))
3058                 LNetMDUnlink(ping_mdh);
3059         if (!LNetMDHandleIsInvalid(push_mdh))
3060                 LNetMDUnlink(push_mdh);
3061 }
3062
3063 /*
3064  * Wait for work to be queued or some other change that must be
3065  * attended to. Returns non-zero if the discovery thread should shut
3066  * down.
3067  */
3068 static int lnet_peer_discovery_wait_for_work(void)
3069 {
3070         int cpt;
3071         int rc = 0;
3072
3073         DEFINE_WAIT(wait);
3074
3075         cpt = lnet_net_lock_current();
3076         for (;;) {
3077                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3078                                 TASK_INTERRUPTIBLE);
3079                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3080                         break;
3081                 if (lnet_push_target_resize_needed())
3082                         break;
3083                 if (!list_empty(&the_lnet.ln_dc_request))
3084                         break;
3085                 if (!list_empty(&the_lnet.ln_msg_resend))
3086                         break;
3087                 lnet_net_unlock(cpt);
3088
3089                 /*
3090                  * wakeup max every second to check if there are peers that
3091                  * have been stuck on the working queue for greater than
3092                  * the peer timeout.
3093                  */
3094                 schedule_timeout(cfs_time_seconds(1));
3095                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3096                 cpt = lnet_net_lock_current();
3097         }
3098         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3099
3100         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3101                 rc = -ESHUTDOWN;
3102
3103         lnet_net_unlock(cpt);
3104
3105         CDEBUG(D_NET, "woken: %d\n", rc);
3106
3107         return rc;
3108 }
3109
3110 /*
3111  * Messages that were pending on a destroyed peer will be put on a global
3112  * resend list. The message resend list will be checked by
3113  * the discovery thread when it wakes up, and will resend messages. These
3114  * messages can still be sendable in the case the lpni which was the initial
3115  * cause of the message re-queue was transfered to another peer.
3116  *
3117  * It is possible that LNet could be shutdown while we're iterating
3118  * through the list. lnet_shudown_lndnets() will attempt to access the
3119  * resend list, but will have to wait until the spinlock is released, by
3120  * which time there shouldn't be any more messages on the resend list.
3121  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3122  * for the messages so they can be released. The other case is that
3123  * lnet_shudown_lndnets() can finalize all the messages before this
3124  * function can visit the resend list, in which case this function will be
3125  * a no-op.
3126  */
3127 static void lnet_resend_msgs(void)
3128 {
3129         struct lnet_msg *msg, *tmp;
3130         struct list_head resend;
3131         int rc;
3132
3133         INIT_LIST_HEAD(&resend);
3134
3135         spin_lock(&the_lnet.ln_msg_resend_lock);
3136         list_splice(&the_lnet.ln_msg_resend, &resend);
3137         spin_unlock(&the_lnet.ln_msg_resend_lock);
3138
3139         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3140                 list_del_init(&msg->msg_list);
3141                 rc = lnet_send(msg->msg_src_nid_param, msg,
3142                                msg->msg_rtr_nid_param);
3143                 if (rc < 0) {
3144                         CNETERR("Error sending %s to %s: %d\n",
3145                                lnet_msgtyp2str(msg->msg_type),
3146                                libcfs_id2str(msg->msg_target), rc);
3147                         lnet_finalize(msg, rc);
3148                 }
3149         }
3150 }
3151
3152 /* The discovery thread. */
3153 static int lnet_peer_discovery(void *arg)
3154 {
3155         struct lnet_peer *lp;
3156         int rc;
3157
3158         CDEBUG(D_NET, "started\n");
3159         cfs_block_allsigs();
3160
3161         for (;;) {
3162                 if (lnet_peer_discovery_wait_for_work())
3163                         break;
3164
3165                 lnet_resend_msgs();
3166
3167                 if (lnet_push_target_resize_needed())
3168                         lnet_push_target_resize();
3169
3170                 lnet_net_lock(LNET_LOCK_EX);
3171                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3172                         break;
3173
3174                 /*
3175                  * Process all incoming discovery work requests.  When
3176                  * discovery must wait on a peer to change state, it
3177                  * is added to the tail of the ln_dc_working queue. A
3178                  * timestamp keeps track of when the peer was added,
3179                  * so we can time out discovery requests that take too
3180                  * long.
3181                  */
3182                 while (!list_empty(&the_lnet.ln_dc_request)) {
3183                         lp = list_first_entry(&the_lnet.ln_dc_request,
3184                                               struct lnet_peer, lp_dc_list);
3185                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3186                         /*
3187                          * set the time the peer was put on the dc_working
3188                          * queue. It shouldn't remain on the queue
3189                          * forever, in case the GET message (for ping)
3190                          * doesn't get a REPLY or the PUT message (for
3191                          * push) doesn't get an ACK.
3192                          */
3193                         lp->lp_last_queued = ktime_get_real_seconds();
3194                         lnet_net_unlock(LNET_LOCK_EX);
3195
3196                         /*
3197                          * Select an action depending on the state of
3198                          * the peer and whether discovery is disabled.
3199                          * The check whether discovery is disabled is
3200                          * done after the code that handles processing
3201                          * for arrived data, cleanup for failures, and
3202                          * forcing a Ping or Push.
3203                          */
3204                         spin_lock(&lp->lp_lock);
3205                         CDEBUG(D_NET, "peer %s state %#x\n",
3206                                 libcfs_nid2str(lp->lp_primary_nid),
3207                                 lp->lp_state);
3208                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3209                                 rc = lnet_peer_data_present(lp);
3210                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3211                                 rc = lnet_peer_ping_failed(lp);
3212                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3213                                 rc = lnet_peer_push_failed(lp);
3214                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3215                                 rc = lnet_peer_send_ping(lp);
3216                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3217                                 rc = lnet_peer_send_push(lp);
3218                         else if (lnet_peer_discovery_disabled)
3219                                 rc = lnet_peer_rediscover(lp);
3220                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3221                                 rc = lnet_peer_send_ping(lp);
3222                         else if (lnet_peer_needs_push(lp))
3223                                 rc = lnet_peer_send_push(lp);
3224                         else
3225                                 rc = lnet_peer_discovered(lp);
3226                         CDEBUG(D_NET, "peer %s state %#x rc %d\n",
3227                                 libcfs_nid2str(lp->lp_primary_nid),
3228                                 lp->lp_state, rc);
3229                         spin_unlock(&lp->lp_lock);
3230
3231                         lnet_net_lock(LNET_LOCK_EX);
3232                         if (rc == LNET_REDISCOVER_PEER) {
3233                                 list_move(&lp->lp_dc_list,
3234                                           &the_lnet.ln_dc_request);
3235                         } else if (rc) {
3236                                 lnet_peer_discovery_error(lp, rc);
3237                         }
3238                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3239                                 lnet_peer_discovery_complete(lp);
3240                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3241                                 break;
3242                 }
3243
3244                 lnet_net_unlock(LNET_LOCK_EX);
3245         }
3246
3247         CDEBUG(D_NET, "stopping\n");
3248         /*
3249          * Clean up before telling lnet_peer_discovery_stop() that
3250          * we're done. Use wake_up() below to somewhat reduce the
3251          * size of the thundering herd if there are multiple threads
3252          * waiting on discovery of a single peer.
3253          */
3254
3255         /* Queue cleanup 1: stop all pending pings and pushes. */
3256         lnet_net_lock(LNET_LOCK_EX);
3257         while (!list_empty(&the_lnet.ln_dc_working)) {
3258                 lp = list_first_entry(&the_lnet.ln_dc_working,
3259                                       struct lnet_peer, lp_dc_list);
3260                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3261                 lnet_net_unlock(LNET_LOCK_EX);
3262                 lnet_peer_cancel_discovery(lp);
3263                 lnet_net_lock(LNET_LOCK_EX);
3264         }
3265         lnet_net_unlock(LNET_LOCK_EX);
3266
3267         /* Queue cleanup 2: wait for the expired queue to clear. */
3268         while (!list_empty(&the_lnet.ln_dc_expired))
3269                 schedule_timeout(cfs_time_seconds(1));
3270
3271         /* Queue cleanup 3: clear the request queue. */
3272         lnet_net_lock(LNET_LOCK_EX);
3273         while (!list_empty(&the_lnet.ln_dc_request)) {
3274                 lp = list_first_entry(&the_lnet.ln_dc_request,
3275                                       struct lnet_peer, lp_dc_list);
3276                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3277                 lnet_peer_discovery_complete(lp);
3278         }
3279         lnet_net_unlock(LNET_LOCK_EX);
3280
3281         LNetEQFree(the_lnet.ln_dc_eqh);
3282         LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3283
3284         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3285         wake_up(&the_lnet.ln_dc_waitq);
3286
3287         CDEBUG(D_NET, "stopped\n");
3288
3289         return 0;
3290 }
3291
3292 /* ln_api_mutex is held on entry. */
3293 int lnet_peer_discovery_start(void)
3294 {
3295         struct task_struct *task;
3296         int rc;
3297
3298         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3299                 return -EALREADY;
3300
3301         rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3302         if (rc != 0) {
3303                 CERROR("Can't allocate discovery EQ: %d\n", rc);
3304                 return rc;
3305         }
3306
3307         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3308         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3309         if (IS_ERR(task)) {
3310                 rc = PTR_ERR(task);
3311                 CERROR("Can't start peer discovery thread: %d\n", rc);
3312
3313                 LNetEQFree(the_lnet.ln_dc_eqh);
3314                 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3315
3316                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3317         }
3318
3319         CDEBUG(D_NET, "discovery start: %d\n", rc);
3320
3321         return rc;
3322 }
3323
3324 /* ln_api_mutex is held on entry. */
3325 void lnet_peer_discovery_stop(void)
3326 {
3327         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3328                 return;
3329
3330         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3331         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3332         wake_up(&the_lnet.ln_dc_waitq);
3333
3334         wait_event(the_lnet.ln_dc_waitq,
3335                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3336
3337         LASSERT(list_empty(&the_lnet.ln_dc_request));
3338         LASSERT(list_empty(&the_lnet.ln_dc_working));
3339         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3340
3341         CDEBUG(D_NET, "discovery stopped\n");
3342 }
3343
3344 /* Debugging */
3345
3346 void
3347 lnet_debug_peer(lnet_nid_t nid)
3348 {
3349         char                    *aliveness = "NA";
3350         struct lnet_peer_ni     *lp;
3351         int                     cpt;
3352
3353         cpt = lnet_cpt_of_nid(nid, NULL);
3354         lnet_net_lock(cpt);
3355
3356         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3357         if (IS_ERR(lp)) {
3358                 lnet_net_unlock(cpt);
3359                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3360                 return;
3361         }
3362
3363         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3364                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3365
3366         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3367                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3368                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3369                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3370                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3371
3372         lnet_peer_ni_decref_locked(lp);
3373
3374         lnet_net_unlock(cpt);
3375 }
3376
3377 /* Gathering information for userspace. */
3378
3379 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3380                           char aliveness[LNET_MAX_STR_LEN],
3381                           __u32 *cpt_iter, __u32 *refcount,
3382                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3383                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3384                           __u32 *peer_tx_qnob)
3385 {
3386         struct lnet_peer_table          *peer_table;
3387         struct lnet_peer_ni             *lp;
3388         int                             j;
3389         int                             lncpt;
3390         bool                            found = false;
3391
3392         /* get the number of CPTs */
3393         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3394
3395         /* if the cpt number to be examined is >= the number of cpts in
3396          * the system then indicate that there are no more cpts to examin
3397          */
3398         if (*cpt_iter >= lncpt)
3399                 return -ENOENT;
3400
3401         /* get the current table */
3402         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3403         /* if the ptable is NULL then there are no more cpts to examine */
3404         if (peer_table == NULL)
3405                 return -ENOENT;
3406
3407         lnet_net_lock(*cpt_iter);
3408
3409         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3410                 struct list_head *peers = &peer_table->pt_hash[j];
3411
3412                 list_for_each_entry(lp, peers, lpni_hashlist) {
3413                         if (peer_index-- > 0)
3414                                 continue;
3415
3416                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3417                         if (lnet_isrouter(lp) ||
3418                                 lnet_peer_aliveness_enabled(lp))
3419                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3420                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3421
3422                         *nid = lp->lpni_nid;
3423                         *refcount = atomic_read(&lp->lpni_refcount);
3424                         *ni_peer_tx_credits =
3425                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3426                         *peer_tx_credits = lp->lpni_txcredits;
3427                         *peer_rtr_credits = lp->lpni_rtrcredits;
3428                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3429                         *peer_tx_qnob = lp->lpni_txqnob;
3430
3431                         found = true;
3432                 }
3433
3434         }
3435         lnet_net_unlock(*cpt_iter);
3436
3437         *cpt_iter = lncpt;
3438
3439         return found ? 0 : -ENOENT;
3440 }
3441
3442 /* ln_api_mutex is held, which keeps the peer list stable */
3443 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3444 {
3445         struct lnet_ioctl_element_stats *lpni_stats;
3446         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3447         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3448         struct lnet_peer_ni_credit_info *lpni_info;
3449         struct lnet_peer_ni *lpni;
3450         struct lnet_peer *lp;
3451         lnet_nid_t nid;
3452         __u32 size;
3453         int rc;
3454
3455         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3456
3457         if (!lp) {
3458                 rc = -ENOENT;
3459                 goto out;
3460         }
3461
3462         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3463                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3464         size *= lp->lp_nnis;
3465         if (size > cfg->prcfg_size) {
3466                 cfg->prcfg_size = size;
3467                 rc = -E2BIG;
3468                 goto out_lp_decref;
3469         }
3470
3471         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3472         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3473         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3474         cfg->prcfg_count = lp->lp_nnis;
3475         cfg->prcfg_size = size;
3476         cfg->prcfg_state = lp->lp_state;
3477
3478         /* Allocate helper buffers. */
3479         rc = -ENOMEM;
3480         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3481         if (!lpni_info)
3482                 goto out_lp_decref;
3483         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3484         if (!lpni_stats)
3485                 goto out_free_info;
3486         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3487         if (!lpni_msg_stats)
3488                 goto out_free_stats;
3489         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3490         if (!lpni_hstats)
3491                 goto out_free_msg_stats;
3492
3493
3494         lpni = NULL;
3495         rc = -EFAULT;
3496         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3497                 nid = lpni->lpni_nid;
3498                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3499                         goto out_free_hstats;
3500                 bulk += sizeof(nid);
3501
3502                 memset(lpni_info, 0, sizeof(*lpni_info));
3503                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3504                 if (lnet_isrouter(lpni) ||
3505                         lnet_peer_aliveness_enabled(lpni))
3506                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3507                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3508
3509                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3510                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3511                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3512                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3513                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3514                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3515                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3516                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3517                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3518                         goto out_free_hstats;
3519                 bulk += sizeof(*lpni_info);
3520
3521                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3522                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3523                                                             LNET_STATS_TYPE_SEND);
3524                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3525                                                             LNET_STATS_TYPE_RECV);
3526                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3527                                                             LNET_STATS_TYPE_DROP);
3528                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3529                         goto out_free_hstats;
3530                 bulk += sizeof(*lpni_stats);
3531                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3532                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3533                         goto out_free_hstats;
3534                 bulk += sizeof(*lpni_msg_stats);
3535                 lpni_hstats->hlpni_network_timeout =
3536                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3537                 lpni_hstats->hlpni_remote_dropped =
3538                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3539                 lpni_hstats->hlpni_remote_timeout =
3540                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3541                 lpni_hstats->hlpni_remote_error =
3542                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3543                 lpni_hstats->hlpni_health_value =
3544                   atomic_read(&lpni->lpni_healthv);
3545                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3546                         goto out_free_hstats;
3547                 bulk += sizeof(*lpni_hstats);
3548         }
3549         rc = 0;
3550
3551 out_free_hstats:
3552         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3553 out_free_msg_stats:
3554         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3555 out_free_stats:
3556         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3557 out_free_info:
3558         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3559 out_lp_decref:
3560         lnet_peer_decref_locked(lp);
3561 out:
3562         return rc;
3563 }
3564
3565 void
3566 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3567 {
3568         /* the mt could've shutdown and cleaned up the queues */
3569         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3570                 return;
3571
3572         if (list_empty(&lpni->lpni_recovery) &&
3573             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3574                 CERROR("lpni %s added to recovery queue. Health = %d\n",
3575                         libcfs_nid2str(lpni->lpni_nid),
3576                         atomic_read(&lpni->lpni_healthv));
3577                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3578                 lnet_peer_ni_addref_locked(lpni);
3579         }
3580 }
3581
3582 /* Call with the ln_api_mutex held */
3583 void
3584 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3585 {
3586         struct lnet_peer_table *ptable;
3587         struct lnet_peer *lp;
3588         struct lnet_peer_net *lpn;
3589         struct lnet_peer_ni *lpni;
3590         int lncpt;
3591         int cpt;
3592
3593         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3594                 return;
3595
3596         if (!all) {
3597                 lnet_net_lock(LNET_LOCK_EX);
3598                 lpni = lnet_find_peer_ni_locked(nid);
3599                 if (!lpni) {
3600                         lnet_net_unlock(LNET_LOCK_EX);
3601                         return;
3602                 }
3603                 atomic_set(&lpni->lpni_healthv, value);
3604                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3605                 lnet_peer_ni_decref_locked(lpni);
3606                 lnet_net_unlock(LNET_LOCK_EX);
3607                 return;
3608         }
3609
3610         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3611
3612         /*
3613          * Walk all the peers and reset the healhv for each one to the
3614          * maximum value.
3615          */
3616         lnet_net_lock(LNET_LOCK_EX);
3617         for (cpt = 0; cpt < lncpt; cpt++) {
3618                 ptable = the_lnet.ln_peer_tables[cpt];
3619                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3620                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3621                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3622                                                     lpni_peer_nis) {
3623                                         atomic_set(&lpni->lpni_healthv, value);
3624                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3625                                 }
3626                         }
3627                 }
3628         }
3629         lnet_net_unlock(LNET_LOCK_EX);
3630 }
3631