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                         struct lnet_peer *rtr_lp =
1366                           lpni->lpni_peer_net->lpn_peer;
1367                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1368                         /*
1369                          * if we're trying to delete a router it means
1370                          * we're moving this peer NI to a new peer so must
1371                          * transfer router properties to the new peer
1372                          */
1373                         if (rtr_refcount > 0) {
1374                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1375                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1376                         }
1377                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1378                         lpni = lnet_peer_ni_alloc(nid);
1379                         if (!lpni) {
1380                                 rc = -ENOMEM;
1381                                 goto out;
1382                         }
1383                 }
1384         } else {
1385                 lpni = lnet_peer_ni_alloc(nid);
1386                 if (!lpni) {
1387                         rc = -ENOMEM;
1388                         goto out;
1389                 }
1390         }
1391
1392         /*
1393          * Get the peer_net. Check that we're not adding a second
1394          * peer_ni on a peer_net of a non-multi-rail peer.
1395          */
1396         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1397         if (!lpn) {
1398                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1399                 if (!lpn) {
1400                         rc = -ENOMEM;
1401                         goto out_free_lpni;
1402                 }
1403         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1404                 rc = -ENOTUNIQ;
1405                 goto out_free_lpni;
1406         }
1407
1408         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1409
1410 out_free_lpni:
1411         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1412         if (!lpni->lpni_peer_net)
1413                 LIBCFS_FREE(lpni, sizeof(*lpni));
1414 out:
1415         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1416                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1417                flags, rc);
1418         return rc;
1419 }
1420
1421 /*
1422  * Update the primary NID of a peer, if possible.
1423  *
1424  * Call with the lnet_api_mutex held.
1425  */
1426 static int
1427 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1428 {
1429         lnet_nid_t old = lp->lp_primary_nid;
1430         int rc = 0;
1431
1432         if (lp->lp_primary_nid == nid)
1433                 goto out;
1434         rc = lnet_peer_add_nid(lp, nid, flags);
1435         if (rc)
1436                 goto out;
1437         lp->lp_primary_nid = nid;
1438 out:
1439         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1440                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1441         return rc;
1442 }
1443
1444 /*
1445  * lpni creation initiated due to traffic either sending or receiving.
1446  */
1447 static int
1448 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1449 {
1450         struct lnet_peer *lp;
1451         struct lnet_peer_net *lpn;
1452         struct lnet_peer_ni *lpni;
1453         /*
1454          * Assume peer is Multi-Rail capable and let discovery find out
1455          * otherwise.
1456          */
1457         unsigned flags = LNET_PEER_MULTI_RAIL;
1458         int rc = 0;
1459
1460         if (nid == LNET_NID_ANY) {
1461                 rc = -EINVAL;
1462                 goto out;
1463         }
1464
1465         /* lnet_net_lock is not needed here because ln_api_lock is held */
1466         lpni = lnet_find_peer_ni_locked(nid);
1467         if (lpni) {
1468                 /*
1469                  * We must have raced with another thread. Since we
1470                  * know next to nothing about a peer_ni created by
1471                  * traffic, we just assume everything is ok and
1472                  * return.
1473                  */
1474                 lnet_peer_ni_decref_locked(lpni);
1475                 goto out;
1476         }
1477
1478         /* Create peer, peer_net, and peer_ni. */
1479         rc = -ENOMEM;
1480         lp = lnet_peer_alloc(nid);
1481         if (!lp)
1482                 goto out;
1483         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1484         if (!lpn)
1485                 goto out_free_lp;
1486         lpni = lnet_peer_ni_alloc(nid);
1487         if (!lpni)
1488                 goto out_free_lpn;
1489         if (pref != LNET_NID_ANY)
1490                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1491
1492         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1493
1494 out_free_lpn:
1495         LIBCFS_FREE(lpn, sizeof(*lpn));
1496 out_free_lp:
1497         LIBCFS_FREE(lp, sizeof(*lp));
1498 out:
1499         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1500         return rc;
1501 }
1502
1503 /*
1504  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1505  *
1506  * This API handles the following combinations:
1507  *   Create a peer with its primary NI if only the prim_nid is provided
1508  *   Add a NID to a peer identified by the prim_nid. The peer identified
1509  *   by the prim_nid must already exist.
1510  *   The peer being created may be non-MR.
1511  *
1512  * The caller must hold ln_api_mutex. This prevents the peer from
1513  * being created/modified/deleted by a different thread.
1514  */
1515 int
1516 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1517 {
1518         struct lnet_peer *lp = NULL;
1519         struct lnet_peer_ni *lpni;
1520         unsigned flags;
1521
1522         /* The prim_nid must always be specified */
1523         if (prim_nid == LNET_NID_ANY)
1524                 return -EINVAL;
1525
1526         flags = LNET_PEER_CONFIGURED;
1527         if (mr)
1528                 flags |= LNET_PEER_MULTI_RAIL;
1529
1530         /*
1531          * If nid isn't specified, we must create a new peer with
1532          * prim_nid as its primary nid.
1533          */
1534         if (nid == LNET_NID_ANY)
1535                 return lnet_peer_add(prim_nid, flags);
1536
1537         /* Look up the prim_nid, which must exist. */
1538         lpni = lnet_find_peer_ni_locked(prim_nid);
1539         if (!lpni)
1540                 return -ENOENT;
1541         lnet_peer_ni_decref_locked(lpni);
1542         lp = lpni->lpni_peer_net->lpn_peer;
1543
1544         /* Peer must have been configured. */
1545         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1546                 CDEBUG(D_NET, "peer %s was not configured\n",
1547                        libcfs_nid2str(prim_nid));
1548                 return -ENOENT;
1549         }
1550
1551         /* Primary NID must match */
1552         if (lp->lp_primary_nid != prim_nid) {
1553                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1554                        libcfs_nid2str(prim_nid),
1555                        libcfs_nid2str(lp->lp_primary_nid));
1556                 return -ENODEV;
1557         }
1558
1559         /* Multi-Rail flag must match. */
1560         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1561                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1562                        libcfs_nid2str(prim_nid));
1563                 return -EPERM;
1564         }
1565
1566         return lnet_peer_add_nid(lp, nid, flags);
1567 }
1568
1569 /*
1570  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1571  *
1572  * This API handles the following combinations:
1573  *   Delete a NI from a peer if both prim_nid and nid are provided.
1574  *   Delete a peer if only prim_nid is provided.
1575  *   Delete a peer if its primary nid is provided.
1576  *
1577  * The caller must hold ln_api_mutex. This prevents the peer from
1578  * being modified/deleted by a different thread.
1579  */
1580 int
1581 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1582 {
1583         struct lnet_peer *lp;
1584         struct lnet_peer_ni *lpni;
1585         unsigned flags;
1586
1587         if (prim_nid == LNET_NID_ANY)
1588                 return -EINVAL;
1589
1590         lpni = lnet_find_peer_ni_locked(prim_nid);
1591         if (!lpni)
1592                 return -ENOENT;
1593         lnet_peer_ni_decref_locked(lpni);
1594         lp = lpni->lpni_peer_net->lpn_peer;
1595
1596         if (prim_nid != lp->lp_primary_nid) {
1597                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1598                        libcfs_nid2str(prim_nid),
1599                        libcfs_nid2str(lp->lp_primary_nid));
1600                 return -ENODEV;
1601         }
1602
1603         lnet_net_lock(LNET_LOCK_EX);
1604         if (lp->lp_rtr_refcount > 0) {
1605                 lnet_net_unlock(LNET_LOCK_EX);
1606                 CERROR("%s is a router. Can not be deleted\n",
1607                        libcfs_nid2str(prim_nid));
1608                 return -EBUSY;
1609         }
1610         lnet_net_unlock(LNET_LOCK_EX);
1611
1612         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1613                 return lnet_peer_del(lp);
1614
1615         flags = LNET_PEER_CONFIGURED;
1616         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1617                 flags |= LNET_PEER_MULTI_RAIL;
1618
1619         return lnet_peer_del_nid(lp, nid, flags);
1620 }
1621
1622 void
1623 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1624 {
1625         struct lnet_peer_table *ptable;
1626         struct lnet_peer_net *lpn;
1627
1628         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1629
1630         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1631         LASSERT(list_empty(&lpni->lpni_txq));
1632         LASSERT(lpni->lpni_txqnob == 0);
1633         LASSERT(list_empty(&lpni->lpni_peer_nis));
1634         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1635
1636         lpn = lpni->lpni_peer_net;
1637         lpni->lpni_peer_net = NULL;
1638         lpni->lpni_net = NULL;
1639
1640         /* remove the peer ni from the zombie list */
1641         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1642         spin_lock(&ptable->pt_zombie_lock);
1643         list_del_init(&lpni->lpni_hashlist);
1644         ptable->pt_zombies--;
1645         spin_unlock(&ptable->pt_zombie_lock);
1646
1647         if (lpni->lpni_pref_nnids > 1) {
1648                 LIBCFS_FREE(lpni->lpni_pref.nids,
1649                         sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1650         }
1651         LIBCFS_FREE(lpni, sizeof(*lpni));
1652
1653         lnet_peer_net_decref_locked(lpn);
1654 }
1655
1656 struct lnet_peer_ni *
1657 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1658 {
1659         struct lnet_peer_ni *lpni = NULL;
1660         int rc;
1661
1662         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1663                 return ERR_PTR(-ESHUTDOWN);
1664
1665         /*
1666          * find if a peer_ni already exists.
1667          * If so then just return that.
1668          */
1669         lpni = lnet_find_peer_ni_locked(nid);
1670         if (lpni)
1671                 return lpni;
1672
1673         lnet_net_unlock(cpt);
1674
1675         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1676         if (rc) {
1677                 lpni = ERR_PTR(rc);
1678                 goto out_net_relock;
1679         }
1680
1681         lpni = lnet_find_peer_ni_locked(nid);
1682         LASSERT(lpni);
1683
1684 out_net_relock:
1685         lnet_net_lock(cpt);
1686
1687         return lpni;
1688 }
1689
1690 /*
1691  * Get a peer_ni for the given nid, create it if necessary. Takes a
1692  * hold on the peer_ni.
1693  */
1694 struct lnet_peer_ni *
1695 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1696 {
1697         struct lnet_peer_ni *lpni = NULL;
1698         int rc;
1699
1700         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1701                 return ERR_PTR(-ESHUTDOWN);
1702
1703         /*
1704          * find if a peer_ni already exists.
1705          * If so then just return that.
1706          */
1707         lpni = lnet_find_peer_ni_locked(nid);
1708         if (lpni)
1709                 return lpni;
1710
1711         /*
1712          * Slow path:
1713          * use the lnet_api_mutex to serialize the creation of the peer_ni
1714          * and the creation/deletion of the local ni/net. When a local ni is
1715          * created, if there exists a set of peer_nis on that network,
1716          * they need to be traversed and updated. When a local NI is
1717          * deleted, which could result in a network being deleted, then
1718          * all peer nis on that network need to be removed as well.
1719          *
1720          * Creation through traffic should also be serialized with
1721          * creation through DLC.
1722          */
1723         lnet_net_unlock(cpt);
1724         mutex_lock(&the_lnet.ln_api_mutex);
1725         /*
1726          * Shutdown is only set under the ln_api_lock, so a single
1727          * check here is sufficent.
1728          */
1729         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1730                 lpni = ERR_PTR(-ESHUTDOWN);
1731                 goto out_mutex_unlock;
1732         }
1733
1734         rc = lnet_peer_ni_traffic_add(nid, pref);
1735         if (rc) {
1736                 lpni = ERR_PTR(rc);
1737                 goto out_mutex_unlock;
1738         }
1739
1740         lpni = lnet_find_peer_ni_locked(nid);
1741         LASSERT(lpni);
1742
1743 out_mutex_unlock:
1744         mutex_unlock(&the_lnet.ln_api_mutex);
1745         lnet_net_lock(cpt);
1746
1747         /* Lock has been dropped, check again for shutdown. */
1748         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1749                 if (!IS_ERR(lpni))
1750                         lnet_peer_ni_decref_locked(lpni);
1751                 lpni = ERR_PTR(-ESHUTDOWN);
1752         }
1753
1754         return lpni;
1755 }
1756
1757 bool
1758 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1759 {
1760         if (lnet_peer_discovery_disabled)
1761                 return true;
1762
1763         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1764             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1765                 return true;
1766         }
1767
1768         return false;
1769 }
1770
1771 /*
1772  * Peer Discovery
1773  */
1774 bool
1775 lnet_is_discovery_disabled(struct lnet_peer *lp)
1776 {
1777         bool rc = false;
1778
1779         spin_lock(&lp->lp_lock);
1780         rc = lnet_is_discovery_disabled_locked(lp);
1781         spin_unlock(&lp->lp_lock);
1782
1783         return rc;
1784 }
1785
1786 bool
1787 lnet_peer_gw_discovery(struct lnet_peer *lp)
1788 {
1789         bool rc = false;
1790
1791         spin_lock(&lp->lp_lock);
1792         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1793                 rc = true;
1794         spin_unlock(&lp->lp_lock);
1795
1796         return rc;
1797 }
1798
1799 /*
1800  * Is a peer uptodate from the point of view of discovery?
1801  *
1802  * If it is currently being processed, obviously not.
1803  * A forced Ping or Push is also handled by the discovery thread.
1804  *
1805  * Otherwise look at whether the peer needs rediscovering.
1806  */
1807 bool
1808 lnet_peer_is_uptodate(struct lnet_peer *lp)
1809 {
1810         bool rc;
1811
1812         spin_lock(&lp->lp_lock);
1813         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1814                             LNET_PEER_FORCE_PING |
1815                             LNET_PEER_FORCE_PUSH)) {
1816                 rc = false;
1817         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1818                 rc = false;
1819         } else if (lnet_peer_needs_push(lp)) {
1820                 rc = false;
1821         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1822                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1823                         rc = true;
1824                 else
1825                         rc = false;
1826         } else {
1827                 rc = false;
1828         }
1829         spin_unlock(&lp->lp_lock);
1830
1831         return rc;
1832 }
1833
1834 /*
1835  * Queue a peer for the attention of the discovery thread.  Call with
1836  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1837  * -EALREADY if the peer was already queued.
1838  */
1839 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1840 {
1841         int rc;
1842
1843         spin_lock(&lp->lp_lock);
1844         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1845                 lp->lp_state |= LNET_PEER_DISCOVERING;
1846         spin_unlock(&lp->lp_lock);
1847         if (list_empty(&lp->lp_dc_list)) {
1848                 lnet_peer_addref_locked(lp);
1849                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1850                 wake_up(&the_lnet.ln_dc_waitq);
1851                 rc = 0;
1852         } else {
1853                 rc = -EALREADY;
1854         }
1855
1856         CDEBUG(D_NET, "Queue peer %s: %d\n",
1857                libcfs_nid2str(lp->lp_primary_nid), rc);
1858
1859         return rc;
1860 }
1861
1862 /*
1863  * Discovery of a peer is complete. Wake all waiters on the peer.
1864  * Call with lnet_net_lock/EX held.
1865  */
1866 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1867 {
1868         struct lnet_msg *msg, *tmp;
1869         int rc = 0;
1870         struct list_head pending_msgs;
1871
1872         INIT_LIST_HEAD(&pending_msgs);
1873
1874         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1875                libcfs_nid2str(lp->lp_primary_nid));
1876
1877         list_del_init(&lp->lp_dc_list);
1878         spin_lock(&lp->lp_lock);
1879         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1880         spin_unlock(&lp->lp_lock);
1881         wake_up_all(&lp->lp_dc_waitq);
1882
1883         if (lp->lp_rtr_refcount > 0)
1884                 lnet_router_discovery_complete(lp);
1885
1886         lnet_net_unlock(LNET_LOCK_EX);
1887
1888         /* iterate through all pending messages and send them again */
1889         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1890                 list_del_init(&msg->msg_list);
1891                 if (lp->lp_dc_error) {
1892                         lnet_finalize(msg, lp->lp_dc_error);
1893                         continue;
1894                 }
1895
1896                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1897                        lnet_msgtyp2str(msg->msg_type),
1898                        libcfs_id2str(msg->msg_target));
1899                 rc = lnet_send(msg->msg_src_nid_param, msg,
1900                                msg->msg_rtr_nid_param);
1901                 if (rc < 0) {
1902                         CNETERR("Error sending %s to %s: %d\n",
1903                                lnet_msgtyp2str(msg->msg_type),
1904                                libcfs_id2str(msg->msg_target), rc);
1905                         lnet_finalize(msg, rc);
1906                 }
1907         }
1908         lnet_net_lock(LNET_LOCK_EX);
1909         lnet_peer_decref_locked(lp);
1910 }
1911
1912 /*
1913  * Handle inbound push.
1914  * Like any event handler, called with lnet_res_lock/CPT held.
1915  */
1916 void lnet_peer_push_event(struct lnet_event *ev)
1917 {
1918         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1919         struct lnet_peer *lp;
1920
1921         /* lnet_find_peer() adds a refcount */
1922         lp = lnet_find_peer(ev->source.nid);
1923         if (!lp) {
1924                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1925                        libcfs_nid2str(ev->initiator.nid),
1926                        libcfs_nid2str(ev->source.nid));
1927                 return;
1928         }
1929
1930         /* Ensure peer state remains consistent while we modify it. */
1931         spin_lock(&lp->lp_lock);
1932
1933         /*
1934          * If some kind of error happened the contents of the message
1935          * cannot be used. Clear the NIDS_UPTODATE and set the
1936          * FORCE_PING flag to trigger a ping.
1937          */
1938         if (ev->status) {
1939                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1940                 lp->lp_state |= LNET_PEER_FORCE_PING;
1941                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1942                        ev->status,
1943                        libcfs_nid2str(lp->lp_primary_nid),
1944                        libcfs_nid2str(ev->source.nid));
1945                 goto out;
1946         }
1947
1948         /*
1949          * A push with invalid or corrupted info. Clear the UPTODATE
1950          * flag to trigger a ping.
1951          */
1952         if (lnet_ping_info_validate(&pbuf->pb_info)) {
1953                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1954                 lp->lp_state |= LNET_PEER_FORCE_PING;
1955                 CDEBUG(D_NET, "Corrupted Push from %s\n",
1956                        libcfs_nid2str(lp->lp_primary_nid));
1957                 goto out;
1958         }
1959
1960         /*
1961          * Make sure we'll allocate the correct size ping buffer when
1962          * pinging the peer.
1963          */
1964         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
1965                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
1966
1967         /*
1968          * A non-Multi-Rail peer is not supposed to be capable of
1969          * sending a push.
1970          */
1971         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
1972                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
1973                        libcfs_nid2str(lp->lp_primary_nid));
1974                 goto out;
1975         }
1976
1977         /*
1978          * Check the MULTIRAIL flag. Complain if the peer was DLC
1979          * configured without it.
1980          */
1981         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1982                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1983                         CERROR("Push says %s is Multi-Rail, DLC says not\n",
1984                                libcfs_nid2str(lp->lp_primary_nid));
1985                 } else {
1986                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1987                         lnet_peer_clr_non_mr_pref_nids(lp);
1988                 }
1989         }
1990
1991         /*
1992          * The peer may have discovery disabled at its end. Set
1993          * NO_DISCOVERY as appropriate.
1994          */
1995         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
1996                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
1997                        libcfs_nid2str(lp->lp_primary_nid));
1998                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
1999         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2000                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2001                        libcfs_nid2str(lp->lp_primary_nid));
2002                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2003         }
2004
2005         /*
2006          * Check for truncation of the Put message. Clear the
2007          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2008          * and tell discovery to allocate a bigger buffer.
2009          */
2010         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2011                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2012                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2013                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2014                 lp->lp_state |= LNET_PEER_FORCE_PING;
2015                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2016                        libcfs_nid2str(lp->lp_primary_nid),
2017                        pbuf->pb_info.pi_nnis);
2018                 goto out;
2019         }
2020
2021         /* always assume new data */
2022         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2023         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2024
2025         /*
2026          * If there is data present that hasn't been processed yet,
2027          * we'll replace it if the Put contained newer data and it
2028          * fits. We're racing with a Ping or earlier Push in this
2029          * case.
2030          */
2031         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2032                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2033                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2034                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2035                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2036                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2037                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2038                               libcfs_nid2str(lp->lp_primary_nid),
2039                               LNET_PING_BUFFER_SEQNO(pbuf),
2040                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2041                 }
2042                 goto out;
2043         }
2044
2045         /*
2046          * Allocate a buffer to copy the data. On a failure we drop
2047          * the Push and set FORCE_PING to force the discovery
2048          * thread to fix the problem by pinging the peer.
2049          */
2050         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2051         if (!lp->lp_data) {
2052                 lp->lp_state |= LNET_PEER_FORCE_PING;
2053                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2054                        libcfs_nid2str(lp->lp_primary_nid),
2055                        LNET_PING_BUFFER_SEQNO(pbuf));
2056                 goto out;
2057         }
2058
2059         /* Success */
2060         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2061                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2062         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2063         CDEBUG(D_NET, "Received Push %s %u\n",
2064                libcfs_nid2str(lp->lp_primary_nid),
2065                LNET_PING_BUFFER_SEQNO(pbuf));
2066
2067 out:
2068         /*
2069          * Queue the peer for discovery if not done, force it on the request
2070          * queue and wake the discovery thread if the peer was already queued,
2071          * because its status changed.
2072          */
2073         spin_unlock(&lp->lp_lock);
2074         lnet_net_lock(LNET_LOCK_EX);
2075         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2076                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2077                 wake_up(&the_lnet.ln_dc_waitq);
2078         }
2079         /* Drop refcount from lookup */
2080         lnet_peer_decref_locked(lp);
2081         lnet_net_unlock(LNET_LOCK_EX);
2082 }
2083
2084 /*
2085  * Clear the discovery error state, unless we're already discovering
2086  * this peer, in which case the error is current.
2087  */
2088 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2089 {
2090         spin_lock(&lp->lp_lock);
2091         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2092                 lp->lp_dc_error = 0;
2093         spin_unlock(&lp->lp_lock);
2094 }
2095
2096 /*
2097  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2098  * dropped/retaken within this function. An lnet_peer_ni is passed in
2099  * because discovery could tear down an lnet_peer.
2100  */
2101 int
2102 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2103 {
2104         DEFINE_WAIT(wait);
2105         struct lnet_peer *lp;
2106         int rc = 0;
2107
2108 again:
2109         lnet_net_unlock(cpt);
2110         lnet_net_lock(LNET_LOCK_EX);
2111         lp = lpni->lpni_peer_net->lpn_peer;
2112         lnet_peer_clear_discovery_error(lp);
2113
2114         /*
2115          * We're willing to be interrupted. The lpni can become a
2116          * zombie if we race with DLC, so we must check for that.
2117          */
2118         for (;;) {
2119                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2120                 if (signal_pending(current))
2121                         break;
2122                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2123                         break;
2124                 if (lp->lp_dc_error)
2125                         break;
2126                 if (lnet_peer_is_uptodate(lp))
2127                         break;
2128                 lnet_peer_queue_for_discovery(lp);
2129
2130                 if (lnet_is_discovery_disabled(lp))
2131                         break;
2132                 /*
2133                  * if caller requested a non-blocking operation then
2134                  * return immediately. Once discovery is complete then the
2135                  * peer ref will be decremented and any pending messages
2136                  * that were stopped due to discovery will be transmitted.
2137                  */
2138                 if (!block)
2139                         break;
2140
2141                 lnet_peer_addref_locked(lp);
2142                 lnet_net_unlock(LNET_LOCK_EX);
2143                 schedule();
2144                 finish_wait(&lp->lp_dc_waitq, &wait);
2145                 lnet_net_lock(LNET_LOCK_EX);
2146                 lnet_peer_decref_locked(lp);
2147                 /* Peer may have changed */
2148                 lp = lpni->lpni_peer_net->lpn_peer;
2149         }
2150         finish_wait(&lp->lp_dc_waitq, &wait);
2151
2152         lnet_net_unlock(LNET_LOCK_EX);
2153         lnet_net_lock(cpt);
2154
2155         /*
2156          * If the peer has changed after we've discovered the older peer,
2157          * then we need to discovery the new peer to make sure the
2158          * interface information is up to date
2159          */
2160         if (lp != lpni->lpni_peer_net->lpn_peer)
2161                 goto again;
2162
2163         if (signal_pending(current))
2164                 rc = -EINTR;
2165         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2166                 rc = -ESHUTDOWN;
2167         else if (lp->lp_dc_error)
2168                 rc = lp->lp_dc_error;
2169         else if (!block)
2170                 CDEBUG(D_NET, "non-blocking discovery\n");
2171         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2172                 goto again;
2173
2174         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2175                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2176                libcfs_nid2str(lpni->lpni_nid), rc,
2177                (!block) ? "pending discovery" : "discovery complete");
2178
2179         return rc;
2180 }
2181
2182 /* Handle an incoming ack for a push. */
2183 static void
2184 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2185 {
2186         struct lnet_ping_buffer *pbuf;
2187
2188         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2189         spin_lock(&lp->lp_lock);
2190         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2191         lp->lp_push_error = ev->status;
2192         if (ev->status)
2193                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2194         else
2195                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2196         spin_unlock(&lp->lp_lock);
2197
2198         CDEBUG(D_NET, "peer %s ev->status %d\n",
2199                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2200 }
2201
2202 /* Handle a Reply message. This is the reply to a Ping message. */
2203 static void
2204 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2205 {
2206         struct lnet_ping_buffer *pbuf;
2207         int rc;
2208
2209         spin_lock(&lp->lp_lock);
2210
2211         /*
2212          * If some kind of error happened the contents of message
2213          * cannot be used. Set PING_FAILED to trigger a retry.
2214          */
2215         if (ev->status) {
2216                 lp->lp_state |= LNET_PEER_PING_FAILED;
2217                 lp->lp_ping_error = ev->status;
2218                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2219                        ev->status,
2220                        libcfs_nid2str(lp->lp_primary_nid),
2221                        libcfs_nid2str(ev->source.nid));
2222                 goto out;
2223         }
2224
2225         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2226         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2227                 lnet_swap_pinginfo(pbuf);
2228
2229         /*
2230          * A reply with invalid or corrupted info. Set PING_FAILED to
2231          * trigger a retry.
2232          */
2233         rc = lnet_ping_info_validate(&pbuf->pb_info);
2234         if (rc) {
2235                 lp->lp_state |= LNET_PEER_PING_FAILED;
2236                 lp->lp_ping_error = 0;
2237                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2238                        libcfs_nid2str(lp->lp_primary_nid), rc);
2239                 goto out;
2240         }
2241
2242
2243         /*
2244          * Only enable the multi-rail feature on the peer if both sides of
2245          * the connection have discovery on
2246          */
2247         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2248                 CDEBUG(D_NET, "Peer %s has Multi-Rail feature enabled\n",
2249                        libcfs_nid2str(lp->lp_primary_nid));
2250                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2251         } else {
2252                 CDEBUG(D_NET, "Peer %s has Multi-Rail feature disabled\n",
2253                        libcfs_nid2str(lp->lp_primary_nid));
2254                 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2255         }
2256
2257         /*
2258          * The peer may have discovery disabled at its end. Set
2259          * NO_DISCOVERY as appropriate.
2260          */
2261         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2262             !lnet_peer_discovery_disabled) {
2263                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2264                        libcfs_nid2str(lp->lp_primary_nid));
2265                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2266         } else {
2267                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2268                        libcfs_nid2str(lp->lp_primary_nid));
2269                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2270         }
2271
2272         /*
2273          * Update the MULTI_RAIL flag based on the reply. If the peer
2274          * was configured with DLC then the setting should match what
2275          * DLC put in.
2276          */
2277         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2278                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2279                         /* Everything's fine */
2280                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2281                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2282                               libcfs_nid2str(lp->lp_primary_nid));
2283                 } else {
2284                         /*
2285                          * if discovery is disabled then we don't want to
2286                          * update the state of the peer. All we'll do is
2287                          * update the peer_nis which were reported back in
2288                          * the initial ping
2289                          */
2290
2291                         if (!lnet_is_discovery_disabled_locked(lp)) {
2292                                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2293                                 lnet_peer_clr_non_mr_pref_nids(lp);
2294                         }
2295                 }
2296         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2297                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2298                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2299                               libcfs_nid2str(lp->lp_primary_nid));
2300                 } else {
2301                         CERROR("Multi-Rail state vanished from %s\n",
2302                                libcfs_nid2str(lp->lp_primary_nid));
2303                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2304                 }
2305         }
2306
2307         /*
2308          * Make sure we'll allocate the correct size ping buffer when
2309          * pinging the peer.
2310          */
2311         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2312                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2313
2314         /*
2315          * Check for truncation of the Reply. Clear PING_SENT and set
2316          * PING_FAILED to trigger a retry.
2317          */
2318         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2319                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2320                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2321                 lp->lp_state |= LNET_PEER_PING_FAILED;
2322                 lp->lp_ping_error = 0;
2323                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2324                        libcfs_nid2str(lp->lp_primary_nid),
2325                        pbuf->pb_info.pi_nnis);
2326                 goto out;
2327         }
2328
2329         /*
2330          * Check the sequence numbers in the reply. These are only
2331          * available if the reply came from a Multi-Rail peer.
2332          */
2333         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2334             pbuf->pb_info.pi_nnis > 1 &&
2335             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2336                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2337                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2338                                 libcfs_nid2str(lp->lp_primary_nid),
2339                                 LNET_PING_BUFFER_SEQNO(pbuf),
2340                                 lp->lp_peer_seqno);
2341
2342                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2343         }
2344
2345         /* We're happy with the state of the data in the buffer. */
2346         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2347                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2348         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2349                 lnet_ping_buffer_decref(lp->lp_data);
2350         else
2351                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2352         lnet_ping_buffer_addref(pbuf);
2353         lp->lp_data = pbuf;
2354 out:
2355         lp->lp_state &= ~LNET_PEER_PING_SENT;
2356         spin_unlock(&lp->lp_lock);
2357 }
2358
2359 /*
2360  * Send event handling. Only matters for error cases, where we clean
2361  * up state on the peer and peer_ni that would otherwise be updated in
2362  * the REPLY event handler for a successful Ping, and the ACK event
2363  * handler for a successful Push.
2364  */
2365 static int
2366 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2367 {
2368         int rc = 0;
2369
2370         if (!ev->status)
2371                 goto out;
2372
2373         spin_lock(&lp->lp_lock);
2374         if (ev->msg_type == LNET_MSG_GET) {
2375                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2376                 lp->lp_state |= LNET_PEER_PING_FAILED;
2377                 lp->lp_ping_error = ev->status;
2378         } else { /* ev->msg_type == LNET_MSG_PUT */
2379                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2380                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2381                 lp->lp_push_error = ev->status;
2382         }
2383         spin_unlock(&lp->lp_lock);
2384         rc = LNET_REDISCOVER_PEER;
2385 out:
2386         CDEBUG(D_NET, "%s Send to %s: %d\n",
2387                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2388                 libcfs_nid2str(ev->target.nid), rc);
2389         return rc;
2390 }
2391
2392 /*
2393  * Unlink event handling. This event is only seen if a call to
2394  * LNetMDUnlink() caused the event to be unlinked. If this call was
2395  * made after the event was set up in LNetGet() or LNetPut() then we
2396  * assume the Ping or Push timed out.
2397  */
2398 static void
2399 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2400 {
2401         spin_lock(&lp->lp_lock);
2402         /* We've passed through LNetGet() */
2403         if (lp->lp_state & LNET_PEER_PING_SENT) {
2404                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2405                 lp->lp_state |= LNET_PEER_PING_FAILED;
2406                 lp->lp_ping_error = -ETIMEDOUT;
2407                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2408                         libcfs_nid2str(lp->lp_primary_nid));
2409         }
2410         /* We've passed through LNetPut() */
2411         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2412                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2413                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2414                 lp->lp_push_error = -ETIMEDOUT;
2415                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2416                         libcfs_nid2str(lp->lp_primary_nid));
2417         }
2418         spin_unlock(&lp->lp_lock);
2419 }
2420
2421 /*
2422  * Event handler for the discovery EQ.
2423  *
2424  * Called with lnet_res_lock(cpt) held. The cpt is the
2425  * lnet_cpt_of_cookie() of the md handle cookie.
2426  */
2427 static void lnet_discovery_event_handler(struct lnet_event *event)
2428 {
2429         struct lnet_peer *lp = event->md.user_ptr;
2430         struct lnet_ping_buffer *pbuf;
2431         int rc;
2432
2433         /* discovery needs to take another look */
2434         rc = LNET_REDISCOVER_PEER;
2435
2436         CDEBUG(D_NET, "Received event: %d\n", event->type);
2437
2438         switch (event->type) {
2439         case LNET_EVENT_ACK:
2440                 lnet_discovery_event_ack(lp, event);
2441                 break;
2442         case LNET_EVENT_REPLY:
2443                 lnet_discovery_event_reply(lp, event);
2444                 break;
2445         case LNET_EVENT_SEND:
2446                 /* Only send failure triggers a retry. */
2447                 rc = lnet_discovery_event_send(lp, event);
2448                 break;
2449         case LNET_EVENT_UNLINK:
2450                 /* LNetMDUnlink() was called */
2451                 lnet_discovery_event_unlink(lp, event);
2452                 break;
2453         default:
2454                 /* Invalid events. */
2455                 LBUG();
2456         }
2457         lnet_net_lock(LNET_LOCK_EX);
2458         if (event->unlinked) {
2459                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2460                 lnet_ping_buffer_decref(pbuf);
2461                 lnet_peer_decref_locked(lp);
2462         }
2463
2464         /* put peer back at end of request queue, if discovery not already
2465          * done */
2466         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2467                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2468                 wake_up(&the_lnet.ln_dc_waitq);
2469         }
2470         lnet_net_unlock(LNET_LOCK_EX);
2471 }
2472
2473 /*
2474  * Build a peer from incoming data.
2475  *
2476  * The NIDs in the incoming data are supposed to be structured as follows:
2477  *  - loopback
2478  *  - primary NID
2479  *  - other NIDs in same net
2480  *  - NIDs in second net
2481  *  - NIDs in third net
2482  *  - ...
2483  * This due to the way the list of NIDs in the data is created.
2484  *
2485  * Note that this function will mark the peer uptodate unless an
2486  * ENOMEM is encontered. All other errors are due to a conflict
2487  * between the DLC configuration and what discovery sees. We treat DLC
2488  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2489  * peer from becoming stuck in discovery.
2490  */
2491 static int lnet_peer_merge_data(struct lnet_peer *lp,
2492                                 struct lnet_ping_buffer *pbuf)
2493 {
2494         struct lnet_peer_ni *lpni;
2495         lnet_nid_t *curnis = NULL;
2496         struct lnet_ni_status *addnis = NULL;
2497         lnet_nid_t *delnis = NULL;
2498         unsigned flags;
2499         int ncurnis;
2500         int naddnis;
2501         int ndelnis;
2502         int nnis = 0;
2503         int i;
2504         int j;
2505         int rc;
2506
2507         flags = LNET_PEER_DISCOVERED;
2508         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2509                 flags |= LNET_PEER_MULTI_RAIL;
2510
2511         /*
2512          * Cache the routing feature for the peer; whether it is enabled
2513          * for disabled as reported by the remote peer.
2514          */
2515         spin_lock(&lp->lp_lock);
2516         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2517                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2518         else
2519                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2520         spin_unlock(&lp->lp_lock);
2521
2522         nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2523         LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2524         LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2525         LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2526         if (!curnis || !addnis || !delnis) {
2527                 rc = -ENOMEM;
2528                 goto out;
2529         }
2530         ncurnis = 0;
2531         naddnis = 0;
2532         ndelnis = 0;
2533
2534         /* Construct the list of NIDs present in peer. */
2535         lpni = NULL;
2536         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2537                 curnis[ncurnis++] = lpni->lpni_nid;
2538
2539         /*
2540          * Check for NIDs in pbuf not present in curnis[].
2541          * The loop starts at 1 to skip the loopback NID.
2542          */
2543         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2544                 for (j = 0; j < ncurnis; j++)
2545                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2546                                 break;
2547                 if (j == ncurnis)
2548                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2549         }
2550         /*
2551          * Check for NIDs in curnis[] not present in pbuf.
2552          * The nested loop starts at 1 to skip the loopback NID.
2553          *
2554          * But never add the loopback NID to delnis[]: if it is
2555          * present in curnis[] then this peer is for this node.
2556          */
2557         for (i = 0; i < ncurnis; i++) {
2558                 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2559                         continue;
2560                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2561                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2562                                 /*
2563                                  * update the information we cache for the
2564                                  * peer with the latest information we
2565                                  * received
2566                                  */
2567                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2568                                 if (lpni) {
2569                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2570                                         lnet_peer_ni_decref_locked(lpni);
2571                                 }
2572                                 break;
2573                         }
2574                 }
2575                 if (j == pbuf->pb_info.pi_nnis)
2576                         delnis[ndelnis++] = curnis[i];
2577         }
2578
2579         /*
2580          * If we get here and the discovery is disabled then we don't want
2581          * to add or delete any NIs. We just updated the ones we have some
2582          * information on, and call it a day
2583          */
2584         rc = 0;
2585         if (lnet_is_discovery_disabled(lp))
2586                 goto out;
2587
2588         for (i = 0; i < naddnis; i++) {
2589                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2590                 if (rc) {
2591                         CERROR("Error adding NID %s to peer %s: %d\n",
2592                                libcfs_nid2str(addnis[i].ns_nid),
2593                                libcfs_nid2str(lp->lp_primary_nid), rc);
2594                         if (rc == -ENOMEM)
2595                                 goto out;
2596                 }
2597                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2598                 if (lpni) {
2599                         lpni->lpni_ns_status = addnis[i].ns_status;
2600                         lnet_peer_ni_decref_locked(lpni);
2601                 }
2602         }
2603
2604         for (i = 0; i < ndelnis; i++) {
2605                 /*
2606                  * for routers it's okay to delete the primary_nid because
2607                  * the upper layers don't really rely on it. So if we're
2608                  * being told that the router changed its primary_nid
2609                  * then it's okay to delete it.
2610                  */
2611                 if (lp->lp_rtr_refcount > 0)
2612                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2613                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2614                 if (rc) {
2615                         CERROR("Error deleting NID %s from peer %s: %d\n",
2616                                libcfs_nid2str(delnis[i]),
2617                                libcfs_nid2str(lp->lp_primary_nid), rc);
2618                         if (rc == -ENOMEM)
2619                                 goto out;
2620                 }
2621         }
2622         /*
2623          * Errors other than -ENOMEM are due to peers having been
2624          * configured with DLC. Ignore these because DLC overrides
2625          * Discovery.
2626          */
2627         rc = 0;
2628 out:
2629         LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2630         LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2631         LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2632         lnet_ping_buffer_decref(pbuf);
2633         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2634
2635         if (rc) {
2636                 spin_lock(&lp->lp_lock);
2637                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2638                 lp->lp_state |= LNET_PEER_FORCE_PING;
2639                 spin_unlock(&lp->lp_lock);
2640         }
2641         return rc;
2642 }
2643
2644 /*
2645  * The data in pbuf says lp is its primary peer, but the data was
2646  * received by a different peer. Try to update lp with the data.
2647  */
2648 static int
2649 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2650 {
2651         struct lnet_handle_md mdh;
2652
2653         /* Queue lp for discovery, and force it on the request queue. */
2654         lnet_net_lock(LNET_LOCK_EX);
2655         if (lnet_peer_queue_for_discovery(lp))
2656                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2657         lnet_net_unlock(LNET_LOCK_EX);
2658
2659         LNetInvalidateMDHandle(&mdh);
2660
2661         /*
2662          * Decide whether we can move the peer to the DATA_PRESENT state.
2663          *
2664          * We replace stale data for a multi-rail peer, repair PING_FAILED
2665          * status, and preempt FORCE_PING.
2666          *
2667          * If after that we have DATA_PRESENT, we merge it into this peer.
2668          */
2669         spin_lock(&lp->lp_lock);
2670         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2671                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2672                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2673                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2674                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2675                         lnet_ping_buffer_decref(pbuf);
2676                         pbuf = lp->lp_data;
2677                         lp->lp_data = NULL;
2678                 }
2679         }
2680         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2681                 lnet_ping_buffer_decref(lp->lp_data);
2682                 lp->lp_data = NULL;
2683                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2684         }
2685         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2686                 mdh = lp->lp_ping_mdh;
2687                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2688                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2689                 lp->lp_ping_error = 0;
2690         }
2691         if (lp->lp_state & LNET_PEER_FORCE_PING)
2692                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2693         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2694         spin_unlock(&lp->lp_lock);
2695
2696         if (!LNetMDHandleIsInvalid(mdh))
2697                 LNetMDUnlink(mdh);
2698
2699         if (pbuf)
2700                 return lnet_peer_merge_data(lp, pbuf);
2701
2702         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2703         return 0;
2704 }
2705
2706 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2707 {
2708         int i;
2709
2710         for (i = 0; i < pinfo->pi_nnis; i++) {
2711                 if (pinfo->pi_ni[i].ns_nid == nid)
2712                         return true;
2713         }
2714
2715         return false;
2716 }
2717
2718 /*
2719  * Update a peer using the data received.
2720  */
2721 static int lnet_peer_data_present(struct lnet_peer *lp)
2722 __must_hold(&lp->lp_lock)
2723 {
2724         struct lnet_ping_buffer *pbuf;
2725         struct lnet_peer_ni *lpni;
2726         lnet_nid_t nid = LNET_NID_ANY;
2727         unsigned flags;
2728         int rc = 0;
2729
2730         pbuf = lp->lp_data;
2731         lp->lp_data = NULL;
2732         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2733         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2734         spin_unlock(&lp->lp_lock);
2735
2736         /*
2737          * Modifications of peer structures are done while holding the
2738          * ln_api_mutex. A global lock is required because we may be
2739          * modifying multiple peer structures, and a mutex greatly
2740          * simplifies memory management.
2741          *
2742          * The actual changes to the data structures must also protect
2743          * against concurrent lookups, for which the lnet_net_lock in
2744          * LNET_LOCK_EX mode is used.
2745          */
2746         mutex_lock(&the_lnet.ln_api_mutex);
2747         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2748                 rc = -ESHUTDOWN;
2749                 goto out;
2750         }
2751
2752         /*
2753          * If this peer is not on the peer list then it is being torn
2754          * down, and our reference count may be all that is keeping it
2755          * alive. Don't do any work on it.
2756          */
2757         if (list_empty(&lp->lp_peer_list))
2758                 goto out;
2759
2760         flags = LNET_PEER_DISCOVERED;
2761         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2762                 flags |= LNET_PEER_MULTI_RAIL;
2763
2764         /*
2765          * Check whether the primary NID in the message matches the
2766          * primary NID of the peer. If it does, update the peer, if
2767          * it it does not, check whether there is already a peer with
2768          * that primary NID. If no such peer exists, try to update
2769          * the primary NID of the current peer (allowed if it was
2770          * created due to message traffic) and complete the update.
2771          * If the peer did exist, hand off the data to it.
2772          *
2773          * The peer for the loopback interface is a special case: this
2774          * is the peer for the local node, and we want to set its
2775          * primary NID to the correct value here. Moreover, this peer
2776          * can show up with only the loopback NID in the ping buffer.
2777          */
2778         if (pbuf->pb_info.pi_nnis <= 1)
2779                 goto out;
2780         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2781         if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2782                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2783                 if (!rc)
2784                         rc = lnet_peer_merge_data(lp, pbuf);
2785         /*
2786          * if the primary nid of the peer is present in the ping info returned
2787          * from the peer, but it's not the local primary peer we have
2788          * cached and discovery is disabled, then we don't want to update
2789          * our local peer info, by adding or removing NIDs, we just want
2790          * to update the status of the nids that we currently have
2791          * recorded in that peer.
2792          */
2793         } else if (lp->lp_primary_nid == nid ||
2794                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2795                     lnet_is_discovery_disabled(lp))) {
2796                 rc = lnet_peer_merge_data(lp, pbuf);
2797         } else {
2798                 lpni = lnet_find_peer_ni_locked(nid);
2799                 if (!lpni) {
2800                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2801                         if (rc) {
2802                                 CERROR("Primary NID error %s versus %s: %d\n",
2803                                        libcfs_nid2str(lp->lp_primary_nid),
2804                                        libcfs_nid2str(nid), rc);
2805                         } else {
2806                                 rc = lnet_peer_merge_data(lp, pbuf);
2807                         }
2808                 } else {
2809                         struct lnet_peer *new_lp;
2810                         new_lp = lpni->lpni_peer_net->lpn_peer;
2811                         /*
2812                          * if lp has discovery/MR enabled that means new_lp
2813                          * should have discovery/MR enabled as well, since
2814                          * it's the same peer, which we're about to merge
2815                          */
2816                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2817                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2818                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2819                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2820
2821                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2822                         lnet_consolidate_routes_locked(lp, new_lp);
2823                         lnet_peer_ni_decref_locked(lpni);
2824                 }
2825         }
2826 out:
2827         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2828                lp->lp_state);
2829         mutex_unlock(&the_lnet.ln_api_mutex);
2830
2831         spin_lock(&lp->lp_lock);
2832         /* Tell discovery to re-check the peer immediately. */
2833         if (!rc)
2834                 rc = LNET_REDISCOVER_PEER;
2835         return rc;
2836 }
2837
2838 /*
2839  * A ping failed. Clear the PING_FAILED state and set the
2840  * FORCE_PING state, to ensure a retry even if discovery is
2841  * disabled. This avoids being left with incorrect state.
2842  */
2843 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2844 __must_hold(&lp->lp_lock)
2845 {
2846         struct lnet_handle_md mdh;
2847         int rc;
2848
2849         mdh = lp->lp_ping_mdh;
2850         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2851         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2852         lp->lp_state |= LNET_PEER_FORCE_PING;
2853         rc = lp->lp_ping_error;
2854         lp->lp_ping_error = 0;
2855         spin_unlock(&lp->lp_lock);
2856
2857         if (!LNetMDHandleIsInvalid(mdh))
2858                 LNetMDUnlink(mdh);
2859
2860         CDEBUG(D_NET, "peer %s:%d\n",
2861                libcfs_nid2str(lp->lp_primary_nid), rc);
2862
2863         spin_lock(&lp->lp_lock);
2864         return rc ? rc : LNET_REDISCOVER_PEER;
2865 }
2866
2867 /*
2868  * Select NID to send a Ping or Push to.
2869  */
2870 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2871 {
2872         struct lnet_peer_ni *lpni;
2873
2874         /* Look for a direct-connected NID for this peer. */
2875         lpni = NULL;
2876         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2877                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2878                         continue;
2879                 break;
2880         }
2881         if (lpni)
2882                 return lpni->lpni_nid;
2883
2884         /* Look for a routed-connected NID for this peer. */
2885         lpni = NULL;
2886         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2887                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2888                         continue;
2889                 break;
2890         }
2891         if (lpni)
2892                 return lpni->lpni_nid;
2893
2894         return LNET_NID_ANY;
2895 }
2896
2897 /* Active side of ping. */
2898 static int lnet_peer_send_ping(struct lnet_peer *lp)
2899 __must_hold(&lp->lp_lock)
2900 {
2901         lnet_nid_t pnid;
2902         int nnis;
2903         int rc;
2904         int cpt;
2905
2906         lp->lp_state |= LNET_PEER_PING_SENT;
2907         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2908         spin_unlock(&lp->lp_lock);
2909
2910         cpt = lnet_net_lock_current();
2911         /* Refcount for MD. */
2912         lnet_peer_addref_locked(lp);
2913         pnid = lnet_peer_select_nid(lp);
2914         lnet_net_unlock(cpt);
2915
2916         nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2917
2918         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2919                             the_lnet.ln_dc_eqh, false);
2920
2921         /*
2922          * if LNetMDBind in lnet_send_ping fails we need to decrement the
2923          * refcount on the peer, otherwise LNetMDUnlink will be called
2924          * which will eventually do that.
2925          */
2926         if (rc > 0) {
2927                 lnet_net_lock(cpt);
2928                 lnet_peer_decref_locked(lp);
2929                 lnet_net_unlock(cpt);
2930                 rc = -rc; /* change the rc to negative value */
2931                 goto fail_error;
2932         } else if (rc < 0) {
2933                 goto fail_error;
2934         }
2935
2936         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2937
2938         spin_lock(&lp->lp_lock);
2939         return 0;
2940
2941 fail_error:
2942         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
2943         /*
2944          * The errors that get us here are considered hard errors and
2945          * cause Discovery to terminate. So we clear PING_SENT, but do
2946          * not set either PING_FAILED or FORCE_PING. In fact we need
2947          * to clear PING_FAILED, because the unlink event handler will
2948          * have set it if we called LNetMDUnlink() above.
2949          */
2950         spin_lock(&lp->lp_lock);
2951         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
2952         return rc;
2953 }
2954
2955 /*
2956  * This function exists because you cannot call LNetMDUnlink() from an
2957  * event handler.
2958  */
2959 static int lnet_peer_push_failed(struct lnet_peer *lp)
2960 __must_hold(&lp->lp_lock)
2961 {
2962         struct lnet_handle_md mdh;
2963         int rc;
2964
2965         mdh = lp->lp_push_mdh;
2966         LNetInvalidateMDHandle(&lp->lp_push_mdh);
2967         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
2968         rc = lp->lp_push_error;
2969         lp->lp_push_error = 0;
2970         spin_unlock(&lp->lp_lock);
2971
2972         if (!LNetMDHandleIsInvalid(mdh))
2973                 LNetMDUnlink(mdh);
2974
2975         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2976         spin_lock(&lp->lp_lock);
2977         return rc ? rc : LNET_REDISCOVER_PEER;
2978 }
2979
2980 /* Active side of push. */
2981 static int lnet_peer_send_push(struct lnet_peer *lp)
2982 __must_hold(&lp->lp_lock)
2983 {
2984         struct lnet_ping_buffer *pbuf;
2985         struct lnet_process_id id;
2986         struct lnet_md md;
2987         int cpt;
2988         int rc;
2989
2990         /* Don't push to a non-multi-rail peer. */
2991         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2992                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
2993                 return 0;
2994         }
2995
2996         lp->lp_state |= LNET_PEER_PUSH_SENT;
2997         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
2998         spin_unlock(&lp->lp_lock);
2999
3000         cpt = lnet_net_lock_current();
3001         pbuf = the_lnet.ln_ping_target;
3002         lnet_ping_buffer_addref(pbuf);
3003         lnet_net_unlock(cpt);
3004
3005         /* Push source MD */
3006         md.start     = &pbuf->pb_info;
3007         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3008         md.threshold = 2; /* Put/Ack */
3009         md.max_size  = 0;
3010         md.options   = 0;
3011         md.eq_handle = the_lnet.ln_dc_eqh;
3012         md.user_ptr  = lp;
3013
3014         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3015         if (rc) {
3016                 lnet_ping_buffer_decref(pbuf);
3017                 CERROR("Can't bind push source MD: %d\n", rc);
3018                 goto fail_error;
3019         }
3020         cpt = lnet_net_lock_current();
3021         /* Refcount for MD. */
3022         lnet_peer_addref_locked(lp);
3023         id.pid = LNET_PID_LUSTRE;
3024         id.nid = lnet_peer_select_nid(lp);
3025         lnet_net_unlock(cpt);
3026
3027         if (id.nid == LNET_NID_ANY) {
3028                 rc = -EHOSTUNREACH;
3029                 goto fail_unlink;
3030         }
3031
3032         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3033                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3034                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3035
3036         if (rc)
3037                 goto fail_unlink;
3038
3039         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3040
3041         spin_lock(&lp->lp_lock);
3042         return 0;
3043
3044 fail_unlink:
3045         LNetMDUnlink(lp->lp_push_mdh);
3046         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3047 fail_error:
3048         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3049         /*
3050          * The errors that get us here are considered hard errors and
3051          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3052          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3053          * because the unlink event handler will have set it if we
3054          * called LNetMDUnlink() above.
3055          */
3056         spin_lock(&lp->lp_lock);
3057         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3058         return rc;
3059 }
3060
3061 /*
3062  * An unrecoverable error was encountered during discovery.
3063  * Set error status in peer and abort discovery.
3064  */
3065 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3066 {
3067         CDEBUG(D_NET, "Discovery error %s: %d\n",
3068                libcfs_nid2str(lp->lp_primary_nid), error);
3069
3070         spin_lock(&lp->lp_lock);
3071         lp->lp_dc_error = error;
3072         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3073         lp->lp_state |= LNET_PEER_REDISCOVER;
3074         spin_unlock(&lp->lp_lock);
3075 }
3076
3077 /*
3078  * Mark the peer as discovered.
3079  */
3080 static int lnet_peer_discovered(struct lnet_peer *lp)
3081 __must_hold(&lp->lp_lock)
3082 {
3083         lp->lp_state |= LNET_PEER_DISCOVERED;
3084         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3085                           LNET_PEER_REDISCOVER);
3086
3087         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3088
3089         return 0;
3090 }
3091
3092
3093 /*
3094  * Discovering this peer is taking too long. Cancel any Ping or Push
3095  * that discovery is waiting on by unlinking the relevant MDs. The
3096  * lnet_discovery_event_handler() will proceed from here and complete
3097  * the cleanup.
3098  */
3099 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3100 {
3101         struct lnet_handle_md ping_mdh;
3102         struct lnet_handle_md push_mdh;
3103
3104         LNetInvalidateMDHandle(&ping_mdh);
3105         LNetInvalidateMDHandle(&push_mdh);
3106
3107         spin_lock(&lp->lp_lock);
3108         if (lp->lp_state & LNET_PEER_PING_SENT) {
3109                 ping_mdh = lp->lp_ping_mdh;
3110                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3111         }
3112         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3113                 push_mdh = lp->lp_push_mdh;
3114                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3115         }
3116         spin_unlock(&lp->lp_lock);
3117
3118         if (!LNetMDHandleIsInvalid(ping_mdh))
3119                 LNetMDUnlink(ping_mdh);
3120         if (!LNetMDHandleIsInvalid(push_mdh))
3121                 LNetMDUnlink(push_mdh);
3122 }
3123
3124 /*
3125  * Wait for work to be queued or some other change that must be
3126  * attended to. Returns non-zero if the discovery thread should shut
3127  * down.
3128  */
3129 static int lnet_peer_discovery_wait_for_work(void)
3130 {
3131         int cpt;
3132         int rc = 0;
3133
3134         DEFINE_WAIT(wait);
3135
3136         cpt = lnet_net_lock_current();
3137         for (;;) {
3138                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3139                                 TASK_INTERRUPTIBLE);
3140                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3141                         break;
3142                 if (lnet_push_target_resize_needed())
3143                         break;
3144                 if (!list_empty(&the_lnet.ln_dc_request))
3145                         break;
3146                 if (!list_empty(&the_lnet.ln_msg_resend))
3147                         break;
3148                 lnet_net_unlock(cpt);
3149
3150                 /*
3151                  * wakeup max every second to check if there are peers that
3152                  * have been stuck on the working queue for greater than
3153                  * the peer timeout.
3154                  */
3155                 schedule_timeout(cfs_time_seconds(1));
3156                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3157                 cpt = lnet_net_lock_current();
3158         }
3159         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3160
3161         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3162                 rc = -ESHUTDOWN;
3163
3164         lnet_net_unlock(cpt);
3165
3166         CDEBUG(D_NET, "woken: %d\n", rc);
3167
3168         return rc;
3169 }
3170
3171 /*
3172  * Messages that were pending on a destroyed peer will be put on a global
3173  * resend list. The message resend list will be checked by
3174  * the discovery thread when it wakes up, and will resend messages. These
3175  * messages can still be sendable in the case the lpni which was the initial
3176  * cause of the message re-queue was transfered to another peer.
3177  *
3178  * It is possible that LNet could be shutdown while we're iterating
3179  * through the list. lnet_shudown_lndnets() will attempt to access the
3180  * resend list, but will have to wait until the spinlock is released, by
3181  * which time there shouldn't be any more messages on the resend list.
3182  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3183  * for the messages so they can be released. The other case is that
3184  * lnet_shudown_lndnets() can finalize all the messages before this
3185  * function can visit the resend list, in which case this function will be
3186  * a no-op.
3187  */
3188 static void lnet_resend_msgs(void)
3189 {
3190         struct lnet_msg *msg, *tmp;
3191         struct list_head resend;
3192         int rc;
3193
3194         INIT_LIST_HEAD(&resend);
3195
3196         spin_lock(&the_lnet.ln_msg_resend_lock);
3197         list_splice(&the_lnet.ln_msg_resend, &resend);
3198         spin_unlock(&the_lnet.ln_msg_resend_lock);
3199
3200         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3201                 list_del_init(&msg->msg_list);
3202                 rc = lnet_send(msg->msg_src_nid_param, msg,
3203                                msg->msg_rtr_nid_param);
3204                 if (rc < 0) {
3205                         CNETERR("Error sending %s to %s: %d\n",
3206                                lnet_msgtyp2str(msg->msg_type),
3207                                libcfs_id2str(msg->msg_target), rc);
3208                         lnet_finalize(msg, rc);
3209                 }
3210         }
3211 }
3212
3213 /* The discovery thread. */
3214 static int lnet_peer_discovery(void *arg)
3215 {
3216         struct lnet_peer *lp;
3217         int rc;
3218
3219         CDEBUG(D_NET, "started\n");
3220         cfs_block_allsigs();
3221
3222         for (;;) {
3223                 if (lnet_peer_discovery_wait_for_work())
3224                         break;
3225
3226                 lnet_resend_msgs();
3227
3228                 if (lnet_push_target_resize_needed())
3229                         lnet_push_target_resize();
3230
3231                 lnet_net_lock(LNET_LOCK_EX);
3232                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3233                         break;
3234
3235                 /*
3236                  * Process all incoming discovery work requests.  When
3237                  * discovery must wait on a peer to change state, it
3238                  * is added to the tail of the ln_dc_working queue. A
3239                  * timestamp keeps track of when the peer was added,
3240                  * so we can time out discovery requests that take too
3241                  * long.
3242                  */
3243                 while (!list_empty(&the_lnet.ln_dc_request)) {
3244                         lp = list_first_entry(&the_lnet.ln_dc_request,
3245                                               struct lnet_peer, lp_dc_list);
3246                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3247                         /*
3248                          * set the time the peer was put on the dc_working
3249                          * queue. It shouldn't remain on the queue
3250                          * forever, in case the GET message (for ping)
3251                          * doesn't get a REPLY or the PUT message (for
3252                          * push) doesn't get an ACK.
3253                          */
3254                         lp->lp_last_queued = ktime_get_real_seconds();
3255                         lnet_net_unlock(LNET_LOCK_EX);
3256
3257                         /*
3258                          * Select an action depending on the state of
3259                          * the peer and whether discovery is disabled.
3260                          * The check whether discovery is disabled is
3261                          * done after the code that handles processing
3262                          * for arrived data, cleanup for failures, and
3263                          * forcing a Ping or Push.
3264                          */
3265                         spin_lock(&lp->lp_lock);
3266                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3267                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3268                                 lp->lp_state);
3269                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3270                                 rc = lnet_peer_data_present(lp);
3271                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3272                                 rc = lnet_peer_ping_failed(lp);
3273                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3274                                 rc = lnet_peer_push_failed(lp);
3275                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3276                                 rc = lnet_peer_send_ping(lp);
3277                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3278                                 rc = lnet_peer_send_push(lp);
3279                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3280                                 rc = lnet_peer_send_ping(lp);
3281                         else if (lnet_peer_needs_push(lp))
3282                                 rc = lnet_peer_send_push(lp);
3283                         else
3284                                 rc = lnet_peer_discovered(lp);
3285                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3286                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3287                                 lp->lp_state, rc);
3288                         spin_unlock(&lp->lp_lock);
3289
3290                         lnet_net_lock(LNET_LOCK_EX);
3291                         if (rc == LNET_REDISCOVER_PEER) {
3292                                 list_move(&lp->lp_dc_list,
3293                                           &the_lnet.ln_dc_request);
3294                         } else if (rc) {
3295                                 lnet_peer_discovery_error(lp, rc);
3296                         }
3297                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3298                                 lnet_peer_discovery_complete(lp);
3299                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3300                                 break;
3301                 }
3302
3303                 lnet_net_unlock(LNET_LOCK_EX);
3304         }
3305
3306         CDEBUG(D_NET, "stopping\n");
3307         /*
3308          * Clean up before telling lnet_peer_discovery_stop() that
3309          * we're done. Use wake_up() below to somewhat reduce the
3310          * size of the thundering herd if there are multiple threads
3311          * waiting on discovery of a single peer.
3312          */
3313
3314         /* Queue cleanup 1: stop all pending pings and pushes. */
3315         lnet_net_lock(LNET_LOCK_EX);
3316         while (!list_empty(&the_lnet.ln_dc_working)) {
3317                 lp = list_first_entry(&the_lnet.ln_dc_working,
3318                                       struct lnet_peer, lp_dc_list);
3319                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3320                 lnet_net_unlock(LNET_LOCK_EX);
3321                 lnet_peer_cancel_discovery(lp);
3322                 lnet_net_lock(LNET_LOCK_EX);
3323         }
3324         lnet_net_unlock(LNET_LOCK_EX);
3325
3326         /* Queue cleanup 2: wait for the expired queue to clear. */
3327         while (!list_empty(&the_lnet.ln_dc_expired))
3328                 schedule_timeout(cfs_time_seconds(1));
3329
3330         /* Queue cleanup 3: clear the request queue. */
3331         lnet_net_lock(LNET_LOCK_EX);
3332         while (!list_empty(&the_lnet.ln_dc_request)) {
3333                 lp = list_first_entry(&the_lnet.ln_dc_request,
3334                                       struct lnet_peer, lp_dc_list);
3335                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3336                 lnet_peer_discovery_complete(lp);
3337         }
3338         lnet_net_unlock(LNET_LOCK_EX);
3339
3340         LNetEQFree(the_lnet.ln_dc_eqh);
3341         LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3342
3343         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3344         wake_up(&the_lnet.ln_dc_waitq);
3345
3346         CDEBUG(D_NET, "stopped\n");
3347
3348         return 0;
3349 }
3350
3351 /* ln_api_mutex is held on entry. */
3352 int lnet_peer_discovery_start(void)
3353 {
3354         struct task_struct *task;
3355         int rc;
3356
3357         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3358                 return -EALREADY;
3359
3360         rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3361         if (rc != 0) {
3362                 CERROR("Can't allocate discovery EQ: %d\n", rc);
3363                 return rc;
3364         }
3365
3366         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3367         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3368         if (IS_ERR(task)) {
3369                 rc = PTR_ERR(task);
3370                 CERROR("Can't start peer discovery thread: %d\n", rc);
3371
3372                 LNetEQFree(the_lnet.ln_dc_eqh);
3373                 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3374
3375                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3376         }
3377
3378         CDEBUG(D_NET, "discovery start: %d\n", rc);
3379
3380         return rc;
3381 }
3382
3383 /* ln_api_mutex is held on entry. */
3384 void lnet_peer_discovery_stop(void)
3385 {
3386         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3387                 return;
3388
3389         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3390         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3391         wake_up(&the_lnet.ln_dc_waitq);
3392
3393         wait_event(the_lnet.ln_dc_waitq,
3394                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3395
3396         LASSERT(list_empty(&the_lnet.ln_dc_request));
3397         LASSERT(list_empty(&the_lnet.ln_dc_working));
3398         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3399
3400         CDEBUG(D_NET, "discovery stopped\n");
3401 }
3402
3403 /* Debugging */
3404
3405 void
3406 lnet_debug_peer(lnet_nid_t nid)
3407 {
3408         char                    *aliveness = "NA";
3409         struct lnet_peer_ni     *lp;
3410         int                     cpt;
3411
3412         cpt = lnet_cpt_of_nid(nid, NULL);
3413         lnet_net_lock(cpt);
3414
3415         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3416         if (IS_ERR(lp)) {
3417                 lnet_net_unlock(cpt);
3418                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3419                 return;
3420         }
3421
3422         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3423                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3424
3425         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3426                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3427                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3428                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3429                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3430
3431         lnet_peer_ni_decref_locked(lp);
3432
3433         lnet_net_unlock(cpt);
3434 }
3435
3436 /* Gathering information for userspace. */
3437
3438 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3439                           char aliveness[LNET_MAX_STR_LEN],
3440                           __u32 *cpt_iter, __u32 *refcount,
3441                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3442                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3443                           __u32 *peer_tx_qnob)
3444 {
3445         struct lnet_peer_table          *peer_table;
3446         struct lnet_peer_ni             *lp;
3447         int                             j;
3448         int                             lncpt;
3449         bool                            found = false;
3450
3451         /* get the number of CPTs */
3452         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3453
3454         /* if the cpt number to be examined is >= the number of cpts in
3455          * the system then indicate that there are no more cpts to examin
3456          */
3457         if (*cpt_iter >= lncpt)
3458                 return -ENOENT;
3459
3460         /* get the current table */
3461         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3462         /* if the ptable is NULL then there are no more cpts to examine */
3463         if (peer_table == NULL)
3464                 return -ENOENT;
3465
3466         lnet_net_lock(*cpt_iter);
3467
3468         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3469                 struct list_head *peers = &peer_table->pt_hash[j];
3470
3471                 list_for_each_entry(lp, peers, lpni_hashlist) {
3472                         if (peer_index-- > 0)
3473                                 continue;
3474
3475                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3476                         if (lnet_isrouter(lp) ||
3477                                 lnet_peer_aliveness_enabled(lp))
3478                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3479                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3480
3481                         *nid = lp->lpni_nid;
3482                         *refcount = atomic_read(&lp->lpni_refcount);
3483                         *ni_peer_tx_credits =
3484                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3485                         *peer_tx_credits = lp->lpni_txcredits;
3486                         *peer_rtr_credits = lp->lpni_rtrcredits;
3487                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3488                         *peer_tx_qnob = lp->lpni_txqnob;
3489
3490                         found = true;
3491                 }
3492
3493         }
3494         lnet_net_unlock(*cpt_iter);
3495
3496         *cpt_iter = lncpt;
3497
3498         return found ? 0 : -ENOENT;
3499 }
3500
3501 /* ln_api_mutex is held, which keeps the peer list stable */
3502 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3503 {
3504         struct lnet_ioctl_element_stats *lpni_stats;
3505         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3506         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3507         struct lnet_peer_ni_credit_info *lpni_info;
3508         struct lnet_peer_ni *lpni;
3509         struct lnet_peer *lp;
3510         lnet_nid_t nid;
3511         __u32 size;
3512         int rc;
3513
3514         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3515
3516         if (!lp) {
3517                 rc = -ENOENT;
3518                 goto out;
3519         }
3520
3521         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3522                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3523         size *= lp->lp_nnis;
3524         if (size > cfg->prcfg_size) {
3525                 cfg->prcfg_size = size;
3526                 rc = -E2BIG;
3527                 goto out_lp_decref;
3528         }
3529
3530         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3531         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3532         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3533         cfg->prcfg_count = lp->lp_nnis;
3534         cfg->prcfg_size = size;
3535         cfg->prcfg_state = lp->lp_state;
3536
3537         /* Allocate helper buffers. */
3538         rc = -ENOMEM;
3539         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3540         if (!lpni_info)
3541                 goto out_lp_decref;
3542         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3543         if (!lpni_stats)
3544                 goto out_free_info;
3545         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3546         if (!lpni_msg_stats)
3547                 goto out_free_stats;
3548         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3549         if (!lpni_hstats)
3550                 goto out_free_msg_stats;
3551
3552
3553         lpni = NULL;
3554         rc = -EFAULT;
3555         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3556                 nid = lpni->lpni_nid;
3557                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3558                         goto out_free_hstats;
3559                 bulk += sizeof(nid);
3560
3561                 memset(lpni_info, 0, sizeof(*lpni_info));
3562                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3563                 if (lnet_isrouter(lpni) ||
3564                         lnet_peer_aliveness_enabled(lpni))
3565                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3566                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3567
3568                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3569                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3570                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3571                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3572                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3573                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3574                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3575                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3576                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3577                         goto out_free_hstats;
3578                 bulk += sizeof(*lpni_info);
3579
3580                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3581                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3582                                                             LNET_STATS_TYPE_SEND);
3583                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3584                                                             LNET_STATS_TYPE_RECV);
3585                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3586                                                             LNET_STATS_TYPE_DROP);
3587                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3588                         goto out_free_hstats;
3589                 bulk += sizeof(*lpni_stats);
3590                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3591                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3592                         goto out_free_hstats;
3593                 bulk += sizeof(*lpni_msg_stats);
3594                 lpni_hstats->hlpni_network_timeout =
3595                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3596                 lpni_hstats->hlpni_remote_dropped =
3597                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3598                 lpni_hstats->hlpni_remote_timeout =
3599                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3600                 lpni_hstats->hlpni_remote_error =
3601                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3602                 lpni_hstats->hlpni_health_value =
3603                   atomic_read(&lpni->lpni_healthv);
3604                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3605                         goto out_free_hstats;
3606                 bulk += sizeof(*lpni_hstats);
3607         }
3608         rc = 0;
3609
3610 out_free_hstats:
3611         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3612 out_free_msg_stats:
3613         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3614 out_free_stats:
3615         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3616 out_free_info:
3617         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3618 out_lp_decref:
3619         lnet_peer_decref_locked(lp);
3620 out:
3621         return rc;
3622 }
3623
3624 void
3625 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3626 {
3627         /* the mt could've shutdown and cleaned up the queues */
3628         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3629                 return;
3630
3631         if (list_empty(&lpni->lpni_recovery) &&
3632             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3633                 CERROR("lpni %s added to recovery queue. Health = %d\n",
3634                         libcfs_nid2str(lpni->lpni_nid),
3635                         atomic_read(&lpni->lpni_healthv));
3636                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3637                 lnet_peer_ni_addref_locked(lpni);
3638         }
3639 }
3640
3641 /* Call with the ln_api_mutex held */
3642 void
3643 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3644 {
3645         struct lnet_peer_table *ptable;
3646         struct lnet_peer *lp;
3647         struct lnet_peer_net *lpn;
3648         struct lnet_peer_ni *lpni;
3649         int lncpt;
3650         int cpt;
3651
3652         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3653                 return;
3654
3655         if (!all) {
3656                 lnet_net_lock(LNET_LOCK_EX);
3657                 lpni = lnet_find_peer_ni_locked(nid);
3658                 if (!lpni) {
3659                         lnet_net_unlock(LNET_LOCK_EX);
3660                         return;
3661                 }
3662                 atomic_set(&lpni->lpni_healthv, value);
3663                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3664                 lnet_peer_ni_decref_locked(lpni);
3665                 lnet_net_unlock(LNET_LOCK_EX);
3666                 return;
3667         }
3668
3669         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3670
3671         /*
3672          * Walk all the peers and reset the healhv for each one to the
3673          * maximum value.
3674          */
3675         lnet_net_lock(LNET_LOCK_EX);
3676         for (cpt = 0; cpt < lncpt; cpt++) {
3677                 ptable = the_lnet.ln_peer_tables[cpt];
3678                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3679                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3680                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3681                                                     lpni_peer_nis) {
3682                                         atomic_set(&lpni->lpni_healthv, value);
3683                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3684                                 }
3685                         }
3686                 }
3687         }
3688         lnet_net_unlock(LNET_LOCK_EX);
3689 }
3690