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         /*
 263          * all peers created on a router should have health on
 264          * if it's not already on.
 265          */
 266         if (the_lnet.ln_routing && !lnet_health_sensitivity)
 267                 lp->lp_health_sensitivity = 1;
 268
 269         /*
 270          * Turn off discovery for loopback peer. If you're creating a peer
 271          * for the loopback interface then that was initiated when we
 272          * attempted to send a message over the loopback. There is no need
 273          * to ever use a different interface when sending messages to
 274          * myself.
 275          */
 276         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 277                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 278         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 279
 280         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 281
 282         return lp;
 283 }
 284
 285 void
 286 lnet_destroy_peer_locked(struct lnet_peer *lp)
 287 {
 288         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 289
 290         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 291         LASSERT(lp->lp_rtr_refcount == 0);
 292         LASSERT(list_empty(&lp->lp_peer_nets));
 293         LASSERT(list_empty(&lp->lp_peer_list));
 294         LASSERT(list_empty(&lp->lp_dc_list));
 295
 296         if (lp->lp_data)
 297                 lnet_ping_buffer_decref(lp->lp_data);
 298
 299         /*
 300          * if there are messages still on the pending queue, then make
 301          * sure to queue them on the ln_msg_resend list so they can be
 302          * resent at a later point if the discovery thread is still
 303          * running.
 304          * If the discovery thread has stopped, then the wakeup will be a
 305          * no-op, and it is expected the lnet_shutdown_lndnets() will
 306          * eventually be called, which will traverse this list and
 307          * finalize the messages on the list.
 308          * We can not resend them now because we're holding the cpt lock.
 309          * Releasing the lock can cause an inconsistent state
 310          */
 311         spin_lock(&the_lnet.ln_msg_resend_lock);
 312         spin_lock(&lp->lp_lock);
 313         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 314         spin_unlock(&lp->lp_lock);
 315         spin_unlock(&the_lnet.ln_msg_resend_lock);
 316         wake_up(&the_lnet.ln_dc_waitq);
 317
 318         LIBCFS_FREE(lp, sizeof(*lp));
 319 }
 320
 321 /*
 322  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 323  * that peer_net, detach the peer_net from the peer.
 324  *
 325  * Call with lnet_net_lock/EX held
 326  */
 327 static void
 328 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 329 {
 330         struct lnet_peer_table *ptable;
 331         struct lnet_peer_net *lpn;
 332         struct lnet_peer *lp;
 333
 334         /*
 335          * Belts and suspenders: gracefully handle teardown of a
 336          * partially connected peer_ni.
 337          */
 338         lpn = lpni->lpni_peer_net;
 339
 340         list_del_init(&lpni->lpni_peer_nis);
 341         /*
 342          * If there are no lpni's left, we detach lpn from
 343          * lp_peer_nets, so it cannot be found anymore.
 344          */
 345         if (list_empty(&lpn->lpn_peer_nis))
 346                 list_del_init(&lpn->lpn_peer_nets);
 347
 348         /* Update peer NID count. */
 349         lp = lpn->lpn_peer;
 350         lp->lp_nnis--;
 351
 352         /*
 353          * If there are no more peer nets, make the peer unfindable
 354          * via the peer_tables.
 355          *
 356          * Otherwise, if the peer is DISCOVERED, tell discovery to
 357          * take another look at it. This is a no-op if discovery for
 358          * this peer did the detaching.
 359          */
 360         if (list_empty(&lp->lp_peer_nets)) {
 361                 list_del_init(&lp->lp_peer_list);
 362                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 363                 ptable->pt_peers--;
 364         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 365                 /* Discovery isn't running, nothing to do here. */
 366         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 367                 lnet_peer_queue_for_discovery(lp);
 368                 wake_up(&the_lnet.ln_dc_waitq);
 369         }
 370         CDEBUG(D_NET, "peer %s NID %s\n",
 371                 libcfs_nid2str(lp->lp_primary_nid),
 372                 libcfs_nid2str(lpni->lpni_nid));
 373 }
 374
 375 /* called with lnet_net_lock LNET_LOCK_EX held */
 376 static int
 377 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 378 {
 379         struct lnet_peer_table *ptable = NULL;
 380
 381         /* don't remove a peer_ni if it's also a gateway */
 382         if (lnet_isrouter(lpni) && !force) {
 383                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 384                        libcfs_nid2str(lpni->lpni_nid));
 385                 return -EBUSY;
 386         }
 387
 388         lnet_peer_remove_from_remote_list(lpni);
 389
 390         /* remove peer ni from the hash list. */
 391         list_del_init(&lpni->lpni_hashlist);
 392
 393         /*
 394          * indicate the peer is being deleted so the monitor thread can
 395          * remove it from the recovery queue.
 396          */
 397         spin_lock(&lpni->lpni_lock);
 398         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 399         spin_unlock(&lpni->lpni_lock);
 400
 401         /* decrement the ref count on the peer table */
 402         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 403         LASSERT(ptable->pt_number > 0);
 404         ptable->pt_number--;
 405
 406         /*
 407          * The peer_ni can no longer be found with a lookup. But there
 408          * can be current users, so keep track of it on the zombie
 409          * list until the reference count has gone to zero.
 410          *
 411          * The last reference may be lost in a place where the
 412          * lnet_net_lock locks only a single cpt, and that cpt may not
 413          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 414          * has its own lock.
 415          */
 416         spin_lock(&ptable->pt_zombie_lock);
 417         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 418         ptable->pt_zombies++;
 419         spin_unlock(&ptable->pt_zombie_lock);
 420
 421         /* no need to keep this peer_ni on the hierarchy anymore */
 422         lnet_peer_detach_peer_ni_locked(lpni);
 423
 424         /* remove hashlist reference on peer_ni */
 425         lnet_peer_ni_decref_locked(lpni);
 426
 427         return 0;
 428 }
 429
 430 void lnet_peer_uninit(void)
 431 {
 432         struct lnet_peer_ni *lpni, *tmp;
 433
 434         lnet_net_lock(LNET_LOCK_EX);
 435
 436         /* remove all peer_nis from the remote peer and the hash list */
 437         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 438                                  lpni_on_remote_peer_ni_list)
 439                 lnet_peer_ni_del_locked(lpni, false);
 440
 441         lnet_peer_tables_destroy();
 442
 443         lnet_net_unlock(LNET_LOCK_EX);
 444 }
 445
 446 static int
 447 lnet_peer_del_locked(struct lnet_peer *peer)
 448 {
 449         struct lnet_peer_ni *lpni = NULL, *lpni2;
 450         int rc = 0, rc2 = 0;
 451
 452         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
 453
 454         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 455         while (lpni != NULL) {
 456                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 457                 rc = lnet_peer_ni_del_locked(lpni, false);
 458                 if (rc != 0)
 459                         rc2 = rc;
 460                 lpni = lpni2;
 461         }
 462
 463         return rc2;
 464 }
 465
 466 static int
 467 lnet_peer_del(struct lnet_peer *peer)
 468 {
 469         lnet_net_lock(LNET_LOCK_EX);
 470         lnet_peer_del_locked(peer);
 471         lnet_net_unlock(LNET_LOCK_EX);
 472
 473         return 0;
 474 }
 475
 476 /*
 477  * Delete a NID from a peer. Call with ln_api_mutex held.
 478  *
 479  * Error codes:
 480  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 481  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 482  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 483  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 484  */
 485 static int
 486 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
 487 {
 488         struct lnet_peer_ni *lpni;
 489         lnet_nid_t primary_nid = lp->lp_primary_nid;
 490         int rc = 0;
 491         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 492
 493         if (!(flags & LNET_PEER_CONFIGURED)) {
 494                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 495                         rc = -EPERM;
 496                         goto out;
 497                 }
 498         }
 499         lpni = lnet_find_peer_ni_locked(nid);
 500         if (!lpni) {
 501                 rc = -ENOENT;
 502                 goto out;
 503         }
 504         lnet_peer_ni_decref_locked(lpni);
 505         if (lp != lpni->lpni_peer_net->lpn_peer) {
 506                 rc = -ECHILD;
 507                 goto out;
 508         }
 509
 510         /*
 511          * This function only allows deletion of the primary NID if it
 512          * is the only NID.
 513          */
 514         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
 515                 rc = -EBUSY;
 516                 goto out;
 517         }
 518
 519         lnet_net_lock(LNET_LOCK_EX);
 520
 521         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
 522                 struct lnet_peer_ni *lpni2;
 523                 /* assign the next peer_ni to be the primary */
 524                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 525                 LASSERT(lpni2);
 526                 lp->lp_primary_nid = lpni->lpni_nid;
 527         }
 528         rc = lnet_peer_ni_del_locked(lpni, force);
 529
 530         lnet_net_unlock(LNET_LOCK_EX);
 531
 532 out:
 533         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 534                libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
 535
 536         return rc;
 537 }
 538
 539 static void
 540 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 541                                struct lnet_peer_table *ptable)
 542 {
 543         int                      i;
 544         struct lnet_peer_ni     *next;
 545         struct lnet_peer_ni     *lpni;
 546         struct lnet_peer        *peer;
 547
 548         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 549                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 550                                          lpni_hashlist) {
 551                         if (net != NULL && net != lpni->lpni_net)
 552                                 continue;
 553
 554                         peer = lpni->lpni_peer_net->lpn_peer;
 555                         if (peer->lp_primary_nid != lpni->lpni_nid) {
 556                                 lnet_peer_ni_del_locked(lpni, false);
 557                                 continue;
 558                         }
 559                         /*
 560                          * Removing the primary NID implies removing
 561                          * the entire peer. Advance next beyond any
 562                          * peer_ni that belongs to the same peer.
 563                          */
 564                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 565                                                  lpni_hashlist) {
 566                                 if (next->lpni_peer_net->lpn_peer != peer)
 567                                         break;
 568                         }
 569                         lnet_peer_del_locked(peer);
 570                 }
 571         }
 572 }
 573
 574 static void
 575 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 576 {
 577         int     i = 3;
 578
 579         spin_lock(&ptable->pt_zombie_lock);
 580         while (ptable->pt_zombies) {
 581                 spin_unlock(&ptable->pt_zombie_lock);
 582
 583                 if (is_power_of_2(i)) {
 584                         CDEBUG(D_WARNING,
 585                                "Waiting for %d zombies on peer table\n",
 586                                ptable->pt_zombies);
 587                 }
 588                 set_current_state(TASK_UNINTERRUPTIBLE);
 589                 schedule_timeout(cfs_time_seconds(1) >> 1);
 590                 spin_lock(&ptable->pt_zombie_lock);
 591         }
 592         spin_unlock(&ptable->pt_zombie_lock);
 593 }
 594
 595 static void
 596 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 597                                 struct lnet_peer_table *ptable)
 598 {
 599         struct lnet_peer_ni     *lp;
 600         struct lnet_peer_ni     *tmp;
 601         lnet_nid_t              gw_nid;
 602         int                     i;
 603
 604         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 605                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 606                                          lpni_hashlist) {
 607                         if (net != lp->lpni_net)
 608                                 continue;
 609
 610                         if (!lnet_isrouter(lp))
 611                                 continue;
 612
 613                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 614
 615                         lnet_net_unlock(LNET_LOCK_EX);
 616                         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
 617                         lnet_net_lock(LNET_LOCK_EX);
 618                 }
 619         }
 620 }
 621
 622 void
 623 lnet_peer_tables_cleanup(struct lnet_net *net)
 624 {
 625         int i;
 626         struct lnet_peer_table *ptable;
 627
 628         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 629         /* If just deleting the peers for a NI, get rid of any routes these
 630          * peers are gateways for. */
 631         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 632                 lnet_net_lock(LNET_LOCK_EX);
 633                 lnet_peer_table_del_rtrs_locked(net, ptable);
 634                 lnet_net_unlock(LNET_LOCK_EX);
 635         }
 636
 637         /* Start the cleanup process */
 638         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 639                 lnet_net_lock(LNET_LOCK_EX);
 640                 lnet_peer_table_cleanup_locked(net, ptable);
 641                 lnet_net_unlock(LNET_LOCK_EX);
 642         }
 643
 644         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 645                 lnet_peer_ni_finalize_wait(ptable);
 646 }
 647
 648 static struct lnet_peer_ni *
 649 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 650 {
 651         struct list_head        *peers;
 652         struct lnet_peer_ni     *lp;
 653
 654         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 655
 656         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 657         list_for_each_entry(lp, peers, lpni_hashlist) {
 658                 if (lp->lpni_nid == nid) {
 659                         lnet_peer_ni_addref_locked(lp);
 660                         return lp;
 661                 }
 662         }
 663
 664         return NULL;
 665 }
 666
 667 struct lnet_peer_ni *
 668 lnet_find_peer_ni_locked(lnet_nid_t nid)
 669 {
 670         struct lnet_peer_ni *lpni;
 671         struct lnet_peer_table *ptable;
 672         int cpt;
 673
 674         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 675
 676         ptable = the_lnet.ln_peer_tables[cpt];
 677         lpni = lnet_get_peer_ni_locked(ptable, nid);
 678
 679         return lpni;
 680 }
 681
 682 struct lnet_peer_ni *
 683 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 684 {
 685         struct lnet_peer_net *lpn;
 686         struct lnet_peer_ni *lpni;
 687
 688         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 689         if (!lpn)
 690                 return NULL;
 691
 692         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 693                 if (lpni->lpni_nid == nid)
 694                         return lpni;
 695         }
 696
 697         return NULL;
 698 }
 699
 700 struct lnet_peer *
 701 lnet_find_peer(lnet_nid_t nid)
 702 {
 703         struct lnet_peer_ni *lpni;
 704         struct lnet_peer *lp = NULL;
 705         int cpt;
 706
 707         cpt = lnet_net_lock_current();
 708         lpni = lnet_find_peer_ni_locked(nid);
 709         if (lpni) {
 710                 lp = lpni->lpni_peer_net->lpn_peer;
 711                 lnet_peer_addref_locked(lp);
 712                 lnet_peer_ni_decref_locked(lpni);
 713         }
 714         lnet_net_unlock(cpt);
 715
 716         return lp;
 717 }
 718
 719 struct lnet_peer_net *
 720 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 721 {
 722         struct lnet_peer_net *net;
 723
 724         if (!prev_lpn_id) {
 725                 /* no net id provided return the first net */
 726                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 727                                                struct lnet_peer_net,
 728                                                lpn_peer_nets);
 729
 730                 return net;
 731         }
 732
 733         /* find the net after the one provided */
 734         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 735                 if (net->lpn_net_id == prev_lpn_id) {
 736                         /*
 737                          * if we reached the end of the list loop to the
 738                          * beginning.
 739                          */
 740                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 741                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 742                                                                 struct lnet_peer_net,
 743                                                                 lpn_peer_nets);
 744                         else
 745                                 return list_next_entry(net, lpn_peer_nets);
 746                 }
 747         }
 748
 749         return NULL;
 750 }
 751
 752 struct lnet_peer_ni *
 753 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 754                              struct lnet_peer_net *peer_net,
 755                              struct lnet_peer_ni *prev)
 756 {
 757         struct lnet_peer_ni *lpni;
 758         struct lnet_peer_net *net = peer_net;
 759
 760         if (!prev) {
 761                 if (!net) {
 762                         if (list_empty(&peer->lp_peer_nets))
 763                                 return NULL;
 764
 765                         net = list_entry(peer->lp_peer_nets.next,
 766                                          struct lnet_peer_net,
 767                                          lpn_peer_nets);
 768                 }
 769                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 770                                   lpni_peer_nis);
 771
 772                 return lpni;
 773         }
 774
 775         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 776                 /*
 777                  * if you reached the end of the peer ni list and the peer
 778                  * net is specified then there are no more peer nis in that
 779                  * net.
 780                  */
 781                 if (net)
 782                         return NULL;
 783
 784                 /*
 785                  * we reached the end of this net ni list. move to the
 786                  * next net
 787                  */
 788                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 789                     &peer->lp_peer_nets)
 790                         /* no more nets and no more NIs. */
 791                         return NULL;
 792
 793                 /* get the next net */
 794                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 795                                  struct lnet_peer_net,
 796                                  lpn_peer_nets);
 797                 /* get the ni on it */
 798                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 799                                   lpni_peer_nis);
 800
 801                 return lpni;
 802         }
 803
 804         /* there are more nis left */
 805         lpni = list_entry(prev->lpni_peer_nis.next,
 806                           struct lnet_peer_ni, lpni_peer_nis);
 807
 808         return lpni;
 809 }
 810
 811 /* Call with the ln_api_mutex held */
 812 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 813 {
 814         struct lnet_process_id id;
 815         struct lnet_peer_table *ptable;
 816         struct lnet_peer *lp;
 817         __u32 count = 0;
 818         __u32 size = 0;
 819         int lncpt;
 820         int cpt;
 821         __u32 i;
 822         int rc;
 823
 824         rc = -ESHUTDOWN;
 825         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 826                 goto done;
 827
 828         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 829
 830         /*
 831          * Count the number of peers, and return E2BIG if the buffer
 832          * is too small. We'll also return the desired size.
 833          */
 834         rc = -E2BIG;
 835         for (cpt = 0; cpt < lncpt; cpt++) {
 836                 ptable = the_lnet.ln_peer_tables[cpt];
 837                 count += ptable->pt_peers;
 838         }
 839         size = count * sizeof(*ids);
 840         if (size > *sizep)
 841                 goto done;
 842
 843         /*
 844          * Walk the peer lists and copy out the primary nids.
 845          * This is safe because the peer lists are only modified
 846          * while the ln_api_mutex is held. So we don't need to
 847          * hold the lnet_net_lock as well, and can therefore
 848          * directly call copy_to_user().
 849          */
 850         rc = -EFAULT;
 851         memset(&id, 0, sizeof(id));
 852         id.pid = LNET_PID_LUSTRE;
 853         i = 0;
 854         for (cpt = 0; cpt < lncpt; cpt++) {
 855                 ptable = the_lnet.ln_peer_tables[cpt];
 856                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 857                         if (i >= count)
 858                                 goto done;
 859                         id.nid = lp->lp_primary_nid;
 860                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 861                                 goto done;
 862                         i++;
 863                 }
 864         }
 865         rc = 0;
 866 done:
 867         *countp = count;
 868         *sizep = size;
 869         return rc;
 870 }
 871
 872 /*
 873  * Start pushes to peers that need to be updated for a configuration
 874  * change on this node.
 875  */
 876 void
 877 lnet_push_update_to_peers(int force)
 878 {
 879         struct lnet_peer_table *ptable;
 880         struct lnet_peer *lp;
 881         int lncpt;
 882         int cpt;
 883
 884         lnet_net_lock(LNET_LOCK_EX);
 885         if (lnet_peer_discovery_disabled)
 886                 force = 0;
 887         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 888         for (cpt = 0; cpt < lncpt; cpt++) {
 889                 ptable = the_lnet.ln_peer_tables[cpt];
 890                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 891                         if (force) {
 892                                 spin_lock(&lp->lp_lock);
 893                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 894                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 895                                 spin_unlock(&lp->lp_lock);
 896                         }
 897                         if (lnet_peer_needs_push(lp))
 898                                 lnet_peer_queue_for_discovery(lp);
 899                 }
 900         }
 901         lnet_net_unlock(LNET_LOCK_EX);
 902         wake_up(&the_lnet.ln_dc_waitq);
 903 }
 904
 905 /*
 906  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 907  * this is a preferred point-to-point path. Call with lnet_net_lock in
 908  * shared mmode.
 909  */
 910 bool
 911 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 912 {
 913         int i;
 914
 915         if (lpni->lpni_pref_nnids == 0)
 916                 return false;
 917         if (lpni->lpni_pref_nnids == 1)
 918                 return lpni->lpni_pref.nid == nid;
 919         for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 920                 if (lpni->lpni_pref.nids[i] == nid)
 921                         return true;
 922         }
 923         return false;
 924 }
 925
 926 /*
 927  * Set a single ni as preferred, provided no preferred ni is already
 928  * defined. Only to be used for non-multi-rail peer_ni.
 929  */
 930 int
 931 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 932 {
 933         int rc = 0;
 934
 935         spin_lock(&lpni->lpni_lock);
 936         if (nid == LNET_NID_ANY) {
 937                 rc = -EINVAL;
 938         } else if (lpni->lpni_pref_nnids > 0) {
 939                 rc = -EPERM;
 940         } else if (lpni->lpni_pref_nnids == 0) {
 941                 lpni->lpni_pref.nid = nid;
 942                 lpni->lpni_pref_nnids = 1;
 943                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
 944         }
 945         spin_unlock(&lpni->lpni_lock);
 946
 947         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 948                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
 949         return rc;
 950 }
 951
 952 /*
 953  * Clear the preferred NID from a non-multi-rail peer_ni, provided
 954  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
 955  */
 956 int
 957 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
 958 {
 959         int rc = 0;
 960
 961         spin_lock(&lpni->lpni_lock);
 962         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
 963                 lpni->lpni_pref_nnids = 0;
 964                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 965         } else if (lpni->lpni_pref_nnids == 0) {
 966                 rc = -ENOENT;
 967         } else {
 968                 rc = -EPERM;
 969         }
 970         spin_unlock(&lpni->lpni_lock);
 971
 972         CDEBUG(D_NET, "peer %s: %d\n",
 973                libcfs_nid2str(lpni->lpni_nid), rc);
 974         return rc;
 975 }
 976
 977 /*
 978  * Clear the preferred NIDs from a non-multi-rail peer.
 979  */
 980 void
 981 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
 982 {
 983         struct lnet_peer_ni *lpni = NULL;
 984
 985         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 986                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
 987 }
 988
 989 int
 990 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 991 {
 992         lnet_nid_t *nids = NULL;
 993         lnet_nid_t *oldnids = NULL;
 994         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 995         int size;
 996         int i;
 997         int rc = 0;
 998
 999         if (nid == LNET_NID_ANY) {
1000                 rc = -EINVAL;
1001                 goto out;
1002         }
1003
1004         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1005                 rc = -EEXIST;
1006                 goto out;
1007         }
1008
1009         /* A non-MR node may have only one preferred NI per peer_ni */
1010         if (lpni->lpni_pref_nnids > 0) {
1011                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1012                         rc = -EPERM;
1013                         goto out;
1014                 }
1015         }
1016
1017         if (lpni->lpni_pref_nnids != 0) {
1018                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1019                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1020                 if (!nids) {
1021                         rc = -ENOMEM;
1022                         goto out;
1023                 }
1024                 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1025                         if (lpni->lpni_pref.nids[i] == nid) {
1026                                 LIBCFS_FREE(nids, size);
1027                                 rc = -EEXIST;
1028                                 goto out;
1029                         }
1030                         nids[i] = lpni->lpni_pref.nids[i];
1031                 }
1032                 nids[i] = nid;
1033         }
1034
1035         lnet_net_lock(LNET_LOCK_EX);
1036         spin_lock(&lpni->lpni_lock);
1037         if (lpni->lpni_pref_nnids == 0) {
1038                 lpni->lpni_pref.nid = nid;
1039         } else {
1040                 oldnids = lpni->lpni_pref.nids;
1041                 lpni->lpni_pref.nids = nids;
1042         }
1043         lpni->lpni_pref_nnids++;
1044         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1045         spin_unlock(&lpni->lpni_lock);
1046         lnet_net_unlock(LNET_LOCK_EX);
1047
1048         if (oldnids) {
1049                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1050                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1051         }
1052 out:
1053         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1054                 spin_lock(&lpni->lpni_lock);
1055                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1056                 spin_unlock(&lpni->lpni_lock);
1057         }
1058         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1059                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1060         return rc;
1061 }
1062
1063 int
1064 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1065 {
1066         lnet_nid_t *nids = NULL;
1067         lnet_nid_t *oldnids = NULL;
1068         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1069         int size;
1070         int i, j;
1071         int rc = 0;
1072
1073         if (lpni->lpni_pref_nnids == 0) {
1074                 rc = -ENOENT;
1075                 goto out;
1076         }
1077
1078         if (lpni->lpni_pref_nnids == 1) {
1079                 if (lpni->lpni_pref.nid != nid) {
1080                         rc = -ENOENT;
1081                         goto out;
1082                 }
1083         } else if (lpni->lpni_pref_nnids == 2) {
1084                 if (lpni->lpni_pref.nids[0] != nid &&
1085                     lpni->lpni_pref.nids[1] != nid) {
1086                         rc = -ENOENT;
1087                         goto out;
1088                 }
1089         } else {
1090                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1091                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1092                 if (!nids) {
1093                         rc = -ENOMEM;
1094                         goto out;
1095                 }
1096                 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1097                         if (lpni->lpni_pref.nids[i] != nid)
1098                                 continue;
1099                         nids[j++] = lpni->lpni_pref.nids[i];
1100                 }
1101                 /* Check if we actually removed a nid. */
1102                 if (j == lpni->lpni_pref_nnids) {
1103                         LIBCFS_FREE(nids, size);
1104                         rc = -ENOENT;
1105                         goto out;
1106                 }
1107         }
1108
1109         lnet_net_lock(LNET_LOCK_EX);
1110         spin_lock(&lpni->lpni_lock);
1111         if (lpni->lpni_pref_nnids == 1) {
1112                 lpni->lpni_pref.nid = LNET_NID_ANY;
1113         } else if (lpni->lpni_pref_nnids == 2) {
1114                 oldnids = lpni->lpni_pref.nids;
1115                 if (oldnids[0] == nid)
1116                         lpni->lpni_pref.nid = oldnids[1];
1117                 else
1118                         lpni->lpni_pref.nid = oldnids[2];
1119         } else {
1120                 oldnids = lpni->lpni_pref.nids;
1121                 lpni->lpni_pref.nids = nids;
1122         }
1123         lpni->lpni_pref_nnids--;
1124         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1125         spin_unlock(&lpni->lpni_lock);
1126         lnet_net_unlock(LNET_LOCK_EX);
1127
1128         if (oldnids) {
1129                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1130                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1131         }
1132 out:
1133         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1134                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1135         return rc;
1136 }
1137
1138 lnet_nid_t
1139 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1140 {
1141         struct lnet_peer_ni *lpni;
1142         lnet_nid_t primary_nid = nid;
1143
1144         lpni = lnet_find_peer_ni_locked(nid);
1145         if (lpni) {
1146                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1147                 lnet_peer_ni_decref_locked(lpni);
1148         }
1149
1150         return primary_nid;
1151 }
1152
1153 bool
1154 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1155 {
1156         if (lnet_peer_discovery_disabled)
1157                 return true;
1158
1159         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1160             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1161                 return true;
1162         }
1163
1164         return false;
1165 }
1166
1167 /*
1168  * Peer Discovery
1169  */
1170 bool
1171 lnet_is_discovery_disabled(struct lnet_peer *lp)
1172 {
1173         bool rc = false;
1174
1175         spin_lock(&lp->lp_lock);
1176         rc = lnet_is_discovery_disabled_locked(lp);
1177         spin_unlock(&lp->lp_lock);
1178
1179         return rc;
1180 }
1181
1182 lnet_nid_t
1183 LNetPrimaryNID(lnet_nid_t nid)
1184 {
1185         struct lnet_peer *lp;
1186         struct lnet_peer_ni *lpni;
1187         lnet_nid_t primary_nid = nid;
1188         int rc = 0;
1189         int cpt;
1190
1191         cpt = lnet_net_lock_current();
1192         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1193         if (IS_ERR(lpni)) {
1194                 rc = PTR_ERR(lpni);
1195                 goto out_unlock;
1196         }
1197         lp = lpni->lpni_peer_net->lpn_peer;
1198
1199         while (!lnet_peer_is_uptodate(lp)) {
1200                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1201                 if (rc)
1202                         goto out_decref;
1203                 lp = lpni->lpni_peer_net->lpn_peer;
1204
1205                 /* Only try once if discovery is disabled */
1206                 if (lnet_is_discovery_disabled(lp))
1207                         break;
1208         }
1209         primary_nid = lp->lp_primary_nid;
1210 out_decref:
1211         lnet_peer_ni_decref_locked(lpni);
1212 out_unlock:
1213         lnet_net_unlock(cpt);
1214
1215         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1216                libcfs_nid2str(primary_nid), rc);
1217         return primary_nid;
1218 }
1219 EXPORT_SYMBOL(LNetPrimaryNID);
1220
1221 struct lnet_peer_net *
1222 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1223 {
1224         struct lnet_peer_net *peer_net;
1225         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1226                 if (peer_net->lpn_net_id == net_id)
1227                         return peer_net;
1228         }
1229         return NULL;
1230 }
1231
1232 /*
1233  * Attach a peer_ni to a peer_net and peer. This function assumes
1234  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1235  * may be attached to a different peer, in which case it will be
1236  * properly detached first. The whole operation is done atomically.
1237  *
1238  * Always returns 0.  This is the last function called from functions
1239  * that do return an int, so returning 0 here allows the compiler to
1240  * do a tail call.
1241  */
1242 static int
1243 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1244                                 struct lnet_peer_net *lpn,
1245                                 struct lnet_peer_ni *lpni,
1246                                 unsigned flags)
1247 {
1248         struct lnet_peer_table *ptable;
1249
1250         /* Install the new peer_ni */
1251         lnet_net_lock(LNET_LOCK_EX);
1252         /* Add peer_ni to global peer table hash, if necessary. */
1253         if (list_empty(&lpni->lpni_hashlist)) {
1254                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1255
1256                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1257                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1258                 ptable->pt_version++;
1259                 ptable->pt_number++;
1260                 /* This is the 1st refcount on lpni. */
1261                 atomic_inc(&lpni->lpni_refcount);
1262         }
1263
1264         /* Detach the peer_ni from an existing peer, if necessary. */
1265         if (lpni->lpni_peer_net) {
1266                 LASSERT(lpni->lpni_peer_net != lpn);
1267                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1268                 lnet_peer_detach_peer_ni_locked(lpni);
1269                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1270                 lpni->lpni_peer_net = NULL;
1271         }
1272
1273         /* Add peer_ni to peer_net */
1274         lpni->lpni_peer_net = lpn;
1275         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1276         lnet_peer_net_addref_locked(lpn);
1277
1278         /* Add peer_net to peer */
1279         if (!lpn->lpn_peer) {
1280                 lpn->lpn_peer = lp;
1281                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1282                 lnet_peer_addref_locked(lp);
1283         }
1284
1285         /* Add peer to global peer list, if necessary */
1286         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1287         if (list_empty(&lp->lp_peer_list)) {
1288                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1289                 ptable->pt_peers++;
1290         }
1291
1292
1293         /* Update peer state */
1294         spin_lock(&lp->lp_lock);
1295         if (flags & LNET_PEER_CONFIGURED) {
1296                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1297                         lp->lp_state |= LNET_PEER_CONFIGURED;
1298         }
1299         if (flags & LNET_PEER_MULTI_RAIL) {
1300                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1301                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1302                         lnet_peer_clr_non_mr_pref_nids(lp);
1303                 }
1304         }
1305         spin_unlock(&lp->lp_lock);
1306
1307         lp->lp_nnis++;
1308         lnet_net_unlock(LNET_LOCK_EX);
1309
1310         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1311                libcfs_nid2str(lp->lp_primary_nid),
1312                libcfs_nid2str(lpni->lpni_nid), flags);
1313
1314         return 0;
1315 }
1316
1317 /*
1318  * Create a new peer, with nid as its primary nid.
1319  *
1320  * Call with the lnet_api_mutex held.
1321  */
1322 static int
1323 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1324 {
1325         struct lnet_peer *lp;
1326         struct lnet_peer_net *lpn;
1327         struct lnet_peer_ni *lpni;
1328         int rc = 0;
1329
1330         LASSERT(nid != LNET_NID_ANY);
1331
1332         /*
1333          * No need for the lnet_net_lock here, because the
1334          * lnet_api_mutex is held.
1335          */
1336         lpni = lnet_find_peer_ni_locked(nid);
1337         if (lpni) {
1338                 /* A peer with this NID already exists. */
1339                 lp = lpni->lpni_peer_net->lpn_peer;
1340                 lnet_peer_ni_decref_locked(lpni);
1341                 /*
1342                  * This is an error if the peer was configured and the
1343                  * primary NID differs or an attempt is made to change
1344                  * the Multi-Rail flag. Otherwise the assumption is
1345                  * that an existing peer is being modified.
1346                  */
1347                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1348                         if (lp->lp_primary_nid != nid)
1349                                 rc = -EEXIST;
1350                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1351                                 rc = -EPERM;
1352                         goto out;
1353                 }
1354                 /* Delete and recreate as a configured peer. */
1355                 lnet_peer_del(lp);
1356         }
1357
1358         /* Create peer, peer_net, and peer_ni. */
1359         rc = -ENOMEM;
1360         lp = lnet_peer_alloc(nid);
1361         if (!lp)
1362                 goto out;
1363         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1364         if (!lpn)
1365                 goto out_free_lp;
1366         lpni = lnet_peer_ni_alloc(nid);
1367         if (!lpni)
1368                 goto out_free_lpn;
1369
1370         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1371
1372 out_free_lpn:
1373         LIBCFS_FREE(lpn, sizeof(*lpn));
1374 out_free_lp:
1375         LIBCFS_FREE(lp, sizeof(*lp));
1376 out:
1377         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1378                libcfs_nid2str(nid), flags, rc);
1379         return rc;
1380 }
1381
1382 /*
1383  * Add a NID to a peer. Call with ln_api_mutex held.
1384  *
1385  * Error codes:
1386  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1387  *  -EEXIST:   The NID was configured by DLC for a different peer.
1388  *  -ENOMEM:   Out of memory.
1389  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1390  *             non-multi-rail peer.
1391  */
1392 static int
1393 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1394 {
1395         struct lnet_peer_net *lpn;
1396         struct lnet_peer_ni *lpni;
1397         int rc = 0;
1398
1399         LASSERT(lp);
1400         LASSERT(nid != LNET_NID_ANY);
1401
1402         /* A configured peer can only be updated through configuration. */
1403         if (!(flags & LNET_PEER_CONFIGURED)) {
1404                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1405                         rc = -EPERM;
1406                         goto out;
1407                 }
1408         }
1409
1410         /*
1411          * The MULTI_RAIL flag can be set but not cleared, because
1412          * that would leave the peer struct in an invalid state.
1413          */
1414         if (flags & LNET_PEER_MULTI_RAIL) {
1415                 spin_lock(&lp->lp_lock);
1416                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1417                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1418                         lnet_peer_clr_non_mr_pref_nids(lp);
1419                 }
1420                 spin_unlock(&lp->lp_lock);
1421         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1422                 rc = -EPERM;
1423                 goto out;
1424         }
1425
1426         lpni = lnet_find_peer_ni_locked(nid);
1427         if (lpni) {
1428                 /*
1429                  * A peer_ni already exists. This is only a problem if
1430                  * it is not connected to this peer and was configured
1431                  * by DLC.
1432                  */
1433                 lnet_peer_ni_decref_locked(lpni);
1434                 if (lpni->lpni_peer_net->lpn_peer == lp)
1435                         goto out;
1436                 if (lnet_peer_ni_is_configured(lpni)) {
1437                         rc = -EEXIST;
1438                         goto out;
1439                 }
1440                 /* If this is the primary NID, destroy the peer. */
1441                 if (lnet_peer_ni_is_primary(lpni)) {
1442                         struct lnet_peer *rtr_lp =
1443                           lpni->lpni_peer_net->lpn_peer;
1444                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1445                         /*
1446                          * if we're trying to delete a router it means
1447                          * we're moving this peer NI to a new peer so must
1448                          * transfer router properties to the new peer
1449                          */
1450                         if (rtr_refcount > 0) {
1451                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1452                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1453                         }
1454                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1455                         lpni = lnet_peer_ni_alloc(nid);
1456                         if (!lpni) {
1457                                 rc = -ENOMEM;
1458                                 goto out;
1459                         }
1460                 }
1461         } else {
1462                 lpni = lnet_peer_ni_alloc(nid);
1463                 if (!lpni) {
1464                         rc = -ENOMEM;
1465                         goto out;
1466                 }
1467         }
1468
1469         /*
1470          * Get the peer_net. Check that we're not adding a second
1471          * peer_ni on a peer_net of a non-multi-rail peer.
1472          */
1473         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1474         if (!lpn) {
1475                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1476                 if (!lpn) {
1477                         rc = -ENOMEM;
1478                         goto out_free_lpni;
1479                 }
1480         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1481                 rc = -ENOTUNIQ;
1482                 goto out_free_lpni;
1483         }
1484
1485         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1486
1487 out_free_lpni:
1488         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1489         if (!lpni->lpni_peer_net)
1490                 LIBCFS_FREE(lpni, sizeof(*lpni));
1491 out:
1492         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1493                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1494                flags, rc);
1495         return rc;
1496 }
1497
1498 /*
1499  * Update the primary NID of a peer, if possible.
1500  *
1501  * Call with the lnet_api_mutex held.
1502  */
1503 static int
1504 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1505 {
1506         lnet_nid_t old = lp->lp_primary_nid;
1507         int rc = 0;
1508
1509         if (lp->lp_primary_nid == nid)
1510                 goto out;
1511         rc = lnet_peer_add_nid(lp, nid, flags);
1512         if (rc)
1513                 goto out;
1514         lp->lp_primary_nid = nid;
1515 out:
1516         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1517                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1518         return rc;
1519 }
1520
1521 /*
1522  * lpni creation initiated due to traffic either sending or receiving.
1523  */
1524 static int
1525 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1526 {
1527         struct lnet_peer *lp;
1528         struct lnet_peer_net *lpn;
1529         struct lnet_peer_ni *lpni;
1530         /*
1531          * Assume peer is Multi-Rail capable and let discovery find out
1532          * otherwise.
1533          */
1534         unsigned flags = LNET_PEER_MULTI_RAIL;
1535         int rc = 0;
1536
1537         if (nid == LNET_NID_ANY) {
1538                 rc = -EINVAL;
1539                 goto out;
1540         }
1541
1542         /* lnet_net_lock is not needed here because ln_api_lock is held */
1543         lpni = lnet_find_peer_ni_locked(nid);
1544         if (lpni) {
1545                 /*
1546                  * We must have raced with another thread. Since we
1547                  * know next to nothing about a peer_ni created by
1548                  * traffic, we just assume everything is ok and
1549                  * return.
1550                  */
1551                 lnet_peer_ni_decref_locked(lpni);
1552                 goto out;
1553         }
1554
1555         /* Create peer, peer_net, and peer_ni. */
1556         rc = -ENOMEM;
1557         lp = lnet_peer_alloc(nid);
1558         if (!lp)
1559                 goto out;
1560         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1561         if (!lpn)
1562                 goto out_free_lp;
1563         lpni = lnet_peer_ni_alloc(nid);
1564         if (!lpni)
1565                 goto out_free_lpn;
1566         if (pref != LNET_NID_ANY)
1567                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1568
1569         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1570
1571 out_free_lpn:
1572         LIBCFS_FREE(lpn, sizeof(*lpn));
1573 out_free_lp:
1574         LIBCFS_FREE(lp, sizeof(*lp));
1575 out:
1576         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1577         return rc;
1578 }
1579
1580 /*
1581  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1582  *
1583  * This API handles the following combinations:
1584  *   Create a peer with its primary NI if only the prim_nid is provided
1585  *   Add a NID to a peer identified by the prim_nid. The peer identified
1586  *   by the prim_nid must already exist.
1587  *   The peer being created may be non-MR.
1588  *
1589  * The caller must hold ln_api_mutex. This prevents the peer from
1590  * being created/modified/deleted by a different thread.
1591  */
1592 int
1593 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1594 {
1595         struct lnet_peer *lp = NULL;
1596         struct lnet_peer_ni *lpni;
1597         unsigned flags;
1598
1599         /* The prim_nid must always be specified */
1600         if (prim_nid == LNET_NID_ANY)
1601                 return -EINVAL;
1602
1603         flags = LNET_PEER_CONFIGURED;
1604         if (mr)
1605                 flags |= LNET_PEER_MULTI_RAIL;
1606
1607         /*
1608          * If nid isn't specified, we must create a new peer with
1609          * prim_nid as its primary nid.
1610          */
1611         if (nid == LNET_NID_ANY)
1612                 return lnet_peer_add(prim_nid, flags);
1613
1614         /* Look up the prim_nid, which must exist. */
1615         lpni = lnet_find_peer_ni_locked(prim_nid);
1616         if (!lpni)
1617                 return -ENOENT;
1618         lnet_peer_ni_decref_locked(lpni);
1619         lp = lpni->lpni_peer_net->lpn_peer;
1620
1621         /* Peer must have been configured. */
1622         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1623                 CDEBUG(D_NET, "peer %s was not configured\n",
1624                        libcfs_nid2str(prim_nid));
1625                 return -ENOENT;
1626         }
1627
1628         /* Primary NID must match */
1629         if (lp->lp_primary_nid != prim_nid) {
1630                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1631                        libcfs_nid2str(prim_nid),
1632                        libcfs_nid2str(lp->lp_primary_nid));
1633                 return -ENODEV;
1634         }
1635
1636         /* Multi-Rail flag must match. */
1637         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1638                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1639                        libcfs_nid2str(prim_nid));
1640                 return -EPERM;
1641         }
1642
1643         return lnet_peer_add_nid(lp, nid, flags);
1644 }
1645
1646 /*
1647  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1648  *
1649  * This API handles the following combinations:
1650  *   Delete a NI from a peer if both prim_nid and nid are provided.
1651  *   Delete a peer if only prim_nid is provided.
1652  *   Delete a peer if its primary nid is provided.
1653  *
1654  * The caller must hold ln_api_mutex. This prevents the peer from
1655  * being modified/deleted by a different thread.
1656  */
1657 int
1658 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1659 {
1660         struct lnet_peer *lp;
1661         struct lnet_peer_ni *lpni;
1662         unsigned flags;
1663
1664         if (prim_nid == LNET_NID_ANY)
1665                 return -EINVAL;
1666
1667         lpni = lnet_find_peer_ni_locked(prim_nid);
1668         if (!lpni)
1669                 return -ENOENT;
1670         lnet_peer_ni_decref_locked(lpni);
1671         lp = lpni->lpni_peer_net->lpn_peer;
1672
1673         if (prim_nid != lp->lp_primary_nid) {
1674                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1675                        libcfs_nid2str(prim_nid),
1676                        libcfs_nid2str(lp->lp_primary_nid));
1677                 return -ENODEV;
1678         }
1679
1680         lnet_net_lock(LNET_LOCK_EX);
1681         if (lp->lp_rtr_refcount > 0) {
1682                 lnet_net_unlock(LNET_LOCK_EX);
1683                 CERROR("%s is a router. Can not be deleted\n",
1684                        libcfs_nid2str(prim_nid));
1685                 return -EBUSY;
1686         }
1687         lnet_net_unlock(LNET_LOCK_EX);
1688
1689         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1690                 return lnet_peer_del(lp);
1691
1692         flags = LNET_PEER_CONFIGURED;
1693         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1694                 flags |= LNET_PEER_MULTI_RAIL;
1695
1696         return lnet_peer_del_nid(lp, nid, flags);
1697 }
1698
1699 void
1700 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1701 {
1702         struct lnet_peer_table *ptable;
1703         struct lnet_peer_net *lpn;
1704
1705         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1706
1707         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1708         LASSERT(list_empty(&lpni->lpni_txq));
1709         LASSERT(lpni->lpni_txqnob == 0);
1710         LASSERT(list_empty(&lpni->lpni_peer_nis));
1711         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1712
1713         lpn = lpni->lpni_peer_net;
1714         lpni->lpni_peer_net = NULL;
1715         lpni->lpni_net = NULL;
1716
1717         /* remove the peer ni from the zombie list */
1718         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1719         spin_lock(&ptable->pt_zombie_lock);
1720         list_del_init(&lpni->lpni_hashlist);
1721         ptable->pt_zombies--;
1722         spin_unlock(&ptable->pt_zombie_lock);
1723
1724         if (lpni->lpni_pref_nnids > 1) {
1725                 LIBCFS_FREE(lpni->lpni_pref.nids,
1726                         sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1727         }
1728         LIBCFS_FREE(lpni, sizeof(*lpni));
1729
1730         lnet_peer_net_decref_locked(lpn);
1731 }
1732
1733 struct lnet_peer_ni *
1734 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1735 {
1736         struct lnet_peer_ni *lpni = NULL;
1737         int rc;
1738
1739         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1740                 return ERR_PTR(-ESHUTDOWN);
1741
1742         /*
1743          * find if a peer_ni already exists.
1744          * If so then just return that.
1745          */
1746         lpni = lnet_find_peer_ni_locked(nid);
1747         if (lpni)
1748                 return lpni;
1749
1750         lnet_net_unlock(cpt);
1751
1752         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1753         if (rc) {
1754                 lpni = ERR_PTR(rc);
1755                 goto out_net_relock;
1756         }
1757
1758         lpni = lnet_find_peer_ni_locked(nid);
1759         LASSERT(lpni);
1760
1761 out_net_relock:
1762         lnet_net_lock(cpt);
1763
1764         return lpni;
1765 }
1766
1767 /*
1768  * Get a peer_ni for the given nid, create it if necessary. Takes a
1769  * hold on the peer_ni.
1770  */
1771 struct lnet_peer_ni *
1772 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1773 {
1774         struct lnet_peer_ni *lpni = NULL;
1775         int rc;
1776
1777         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1778                 return ERR_PTR(-ESHUTDOWN);
1779
1780         /*
1781          * find if a peer_ni already exists.
1782          * If so then just return that.
1783          */
1784         lpni = lnet_find_peer_ni_locked(nid);
1785         if (lpni)
1786                 return lpni;
1787
1788         /*
1789          * Slow path:
1790          * use the lnet_api_mutex to serialize the creation of the peer_ni
1791          * and the creation/deletion of the local ni/net. When a local ni is
1792          * created, if there exists a set of peer_nis on that network,
1793          * they need to be traversed and updated. When a local NI is
1794          * deleted, which could result in a network being deleted, then
1795          * all peer nis on that network need to be removed as well.
1796          *
1797          * Creation through traffic should also be serialized with
1798          * creation through DLC.
1799          */
1800         lnet_net_unlock(cpt);
1801         mutex_lock(&the_lnet.ln_api_mutex);
1802         /*
1803          * Shutdown is only set under the ln_api_lock, so a single
1804          * check here is sufficent.
1805          */
1806         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1807                 lpni = ERR_PTR(-ESHUTDOWN);
1808                 goto out_mutex_unlock;
1809         }
1810
1811         rc = lnet_peer_ni_traffic_add(nid, pref);
1812         if (rc) {
1813                 lpni = ERR_PTR(rc);
1814                 goto out_mutex_unlock;
1815         }
1816
1817         lpni = lnet_find_peer_ni_locked(nid);
1818         LASSERT(lpni);
1819
1820 out_mutex_unlock:
1821         mutex_unlock(&the_lnet.ln_api_mutex);
1822         lnet_net_lock(cpt);
1823
1824         /* Lock has been dropped, check again for shutdown. */
1825         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1826                 if (!IS_ERR(lpni))
1827                         lnet_peer_ni_decref_locked(lpni);
1828                 lpni = ERR_PTR(-ESHUTDOWN);
1829         }
1830
1831         return lpni;
1832 }
1833
1834 bool
1835 lnet_peer_gw_discovery(struct lnet_peer *lp)
1836 {
1837         bool rc = false;
1838
1839         spin_lock(&lp->lp_lock);
1840         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1841                 rc = true;
1842         spin_unlock(&lp->lp_lock);
1843
1844         return rc;
1845 }
1846
1847 bool
1848 lnet_peer_is_uptodate(struct lnet_peer *lp)
1849 {
1850         bool rc;
1851
1852         spin_lock(&lp->lp_lock);
1853         rc = lnet_peer_is_uptodate_locked(lp);
1854         spin_unlock(&lp->lp_lock);
1855         return rc;
1856 }
1857
1858 /*
1859  * Is a peer uptodate from the point of view of discovery?
1860  *
1861  * If it is currently being processed, obviously not.
1862  * A forced Ping or Push is also handled by the discovery thread.
1863  *
1864  * Otherwise look at whether the peer needs rediscovering.
1865  */
1866 bool
1867 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1868 __must_hold(&lp->lp_lock)
1869 {
1870         bool rc;
1871
1872         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1873                             LNET_PEER_FORCE_PING |
1874                             LNET_PEER_FORCE_PUSH)) {
1875                 rc = false;
1876         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1877                 rc = false;
1878         } else if (lnet_peer_needs_push(lp)) {
1879                 rc = false;
1880         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1881                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1882                         rc = true;
1883                 else
1884                         rc = false;
1885         } else {
1886                 rc = false;
1887         }
1888
1889         return rc;
1890 }
1891
1892 /*
1893  * Queue a peer for the attention of the discovery thread.  Call with
1894  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1895  * -EALREADY if the peer was already queued.
1896  */
1897 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1898 {
1899         int rc;
1900
1901         spin_lock(&lp->lp_lock);
1902         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1903                 lp->lp_state |= LNET_PEER_DISCOVERING;
1904         spin_unlock(&lp->lp_lock);
1905         if (list_empty(&lp->lp_dc_list)) {
1906                 lnet_peer_addref_locked(lp);
1907                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1908                 wake_up(&the_lnet.ln_dc_waitq);
1909                 rc = 0;
1910         } else {
1911                 rc = -EALREADY;
1912         }
1913
1914         CDEBUG(D_NET, "Queue peer %s: %d\n",
1915                libcfs_nid2str(lp->lp_primary_nid), rc);
1916
1917         return rc;
1918 }
1919
1920 /*
1921  * Discovery of a peer is complete. Wake all waiters on the peer.
1922  * Call with lnet_net_lock/EX held.
1923  */
1924 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1925 {
1926         struct lnet_msg *msg, *tmp;
1927         int rc = 0;
1928         struct list_head pending_msgs;
1929
1930         INIT_LIST_HEAD(&pending_msgs);
1931
1932         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1933                libcfs_nid2str(lp->lp_primary_nid));
1934
1935         list_del_init(&lp->lp_dc_list);
1936         spin_lock(&lp->lp_lock);
1937         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1938         spin_unlock(&lp->lp_lock);
1939         wake_up_all(&lp->lp_dc_waitq);
1940
1941         if (lp->lp_rtr_refcount > 0)
1942                 lnet_router_discovery_complete(lp);
1943
1944         lnet_net_unlock(LNET_LOCK_EX);
1945
1946         /* iterate through all pending messages and send them again */
1947         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1948                 list_del_init(&msg->msg_list);
1949                 if (lp->lp_dc_error) {
1950                         lnet_finalize(msg, lp->lp_dc_error);
1951                         continue;
1952                 }
1953
1954                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1955                        lnet_msgtyp2str(msg->msg_type),
1956                        libcfs_id2str(msg->msg_target));
1957                 rc = lnet_send(msg->msg_src_nid_param, msg,
1958                                msg->msg_rtr_nid_param);
1959                 if (rc < 0) {
1960                         CNETERR("Error sending %s to %s: %d\n",
1961                                lnet_msgtyp2str(msg->msg_type),
1962                                libcfs_id2str(msg->msg_target), rc);
1963                         lnet_finalize(msg, rc);
1964                 }
1965         }
1966         lnet_net_lock(LNET_LOCK_EX);
1967         lnet_peer_decref_locked(lp);
1968 }
1969
1970 /*
1971  * Handle inbound push.
1972  * Like any event handler, called with lnet_res_lock/CPT held.
1973  */
1974 void lnet_peer_push_event(struct lnet_event *ev)
1975 {
1976         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1977         struct lnet_peer *lp;
1978
1979         /* lnet_find_peer() adds a refcount */
1980         lp = lnet_find_peer(ev->source.nid);
1981         if (!lp) {
1982                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1983                        libcfs_nid2str(ev->initiator.nid),
1984                        libcfs_nid2str(ev->source.nid));
1985                 return;
1986         }
1987
1988         /* Ensure peer state remains consistent while we modify it. */
1989         spin_lock(&lp->lp_lock);
1990
1991         /*
1992          * If some kind of error happened the contents of the message
1993          * cannot be used. Clear the NIDS_UPTODATE and set the
1994          * FORCE_PING flag to trigger a ping.
1995          */
1996         if (ev->status) {
1997                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1998                 lp->lp_state |= LNET_PEER_FORCE_PING;
1999                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2000                        ev->status,
2001                        libcfs_nid2str(lp->lp_primary_nid),
2002                        libcfs_nid2str(ev->source.nid));
2003                 goto out;
2004         }
2005
2006         /*
2007          * A push with invalid or corrupted info. Clear the UPTODATE
2008          * flag to trigger a ping.
2009          */
2010         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2011                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2012                 lp->lp_state |= LNET_PEER_FORCE_PING;
2013                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2014                        libcfs_nid2str(lp->lp_primary_nid));
2015                 goto out;
2016         }
2017
2018         /*
2019          * Make sure we'll allocate the correct size ping buffer when
2020          * pinging the peer.
2021          */
2022         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2023                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2024
2025         /*
2026          * A non-Multi-Rail peer is not supposed to be capable of
2027          * sending a push.
2028          */
2029         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2030                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2031                        libcfs_nid2str(lp->lp_primary_nid));
2032                 goto out;
2033         }
2034
2035         /*
2036          * Check the MULTIRAIL flag. Complain if the peer was DLC
2037          * configured without it.
2038          */
2039         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2040                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2041                         CERROR("Push says %s is Multi-Rail, DLC says not\n",
2042                                libcfs_nid2str(lp->lp_primary_nid));
2043                 } else {
2044                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2045                         lnet_peer_clr_non_mr_pref_nids(lp);
2046                 }
2047         }
2048
2049         /*
2050          * The peer may have discovery disabled at its end. Set
2051          * NO_DISCOVERY as appropriate.
2052          */
2053         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2054                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2055                        libcfs_nid2str(lp->lp_primary_nid));
2056                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2057         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2058                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2059                        libcfs_nid2str(lp->lp_primary_nid));
2060                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2061         }
2062
2063         /*
2064          * Check for truncation of the Put message. Clear the
2065          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2066          * and tell discovery to allocate a bigger buffer.
2067          */
2068         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2069                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2070                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2071                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2072                 lp->lp_state |= LNET_PEER_FORCE_PING;
2073                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2074                        libcfs_nid2str(lp->lp_primary_nid),
2075                        pbuf->pb_info.pi_nnis);
2076                 goto out;
2077         }
2078
2079         /* always assume new data */
2080         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2081         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2082
2083         /*
2084          * If there is data present that hasn't been processed yet,
2085          * we'll replace it if the Put contained newer data and it
2086          * fits. We're racing with a Ping or earlier Push in this
2087          * case.
2088          */
2089         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2090                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2091                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2092                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2093                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2094                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2095                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2096                               libcfs_nid2str(lp->lp_primary_nid),
2097                               LNET_PING_BUFFER_SEQNO(pbuf),
2098                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2099                 }
2100                 goto out;
2101         }
2102
2103         /*
2104          * Allocate a buffer to copy the data. On a failure we drop
2105          * the Push and set FORCE_PING to force the discovery
2106          * thread to fix the problem by pinging the peer.
2107          */
2108         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2109         if (!lp->lp_data) {
2110                 lp->lp_state |= LNET_PEER_FORCE_PING;
2111                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2112                        libcfs_nid2str(lp->lp_primary_nid),
2113                        LNET_PING_BUFFER_SEQNO(pbuf));
2114                 goto out;
2115         }
2116
2117         /* Success */
2118         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2119                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2120         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2121         CDEBUG(D_NET, "Received Push %s %u\n",
2122                libcfs_nid2str(lp->lp_primary_nid),
2123                LNET_PING_BUFFER_SEQNO(pbuf));
2124
2125 out:
2126         /*
2127          * Queue the peer for discovery if not done, force it on the request
2128          * queue and wake the discovery thread if the peer was already queued,
2129          * because its status changed.
2130          */
2131         spin_unlock(&lp->lp_lock);
2132         lnet_net_lock(LNET_LOCK_EX);
2133         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2134                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2135                 wake_up(&the_lnet.ln_dc_waitq);
2136         }
2137         /* Drop refcount from lookup */
2138         lnet_peer_decref_locked(lp);
2139         lnet_net_unlock(LNET_LOCK_EX);
2140 }
2141
2142 /*
2143  * Clear the discovery error state, unless we're already discovering
2144  * this peer, in which case the error is current.
2145  */
2146 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2147 {
2148         spin_lock(&lp->lp_lock);
2149         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2150                 lp->lp_dc_error = 0;
2151         spin_unlock(&lp->lp_lock);
2152 }
2153
2154 /*
2155  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2156  * dropped/retaken within this function. An lnet_peer_ni is passed in
2157  * because discovery could tear down an lnet_peer.
2158  */
2159 int
2160 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2161 {
2162         DEFINE_WAIT(wait);
2163         struct lnet_peer *lp;
2164         int rc = 0;
2165         int count = 0;
2166
2167 again:
2168         lnet_net_unlock(cpt);
2169         lnet_net_lock(LNET_LOCK_EX);
2170         lp = lpni->lpni_peer_net->lpn_peer;
2171         lnet_peer_clear_discovery_error(lp);
2172
2173         /*
2174          * We're willing to be interrupted. The lpni can become a
2175          * zombie if we race with DLC, so we must check for that.
2176          */
2177         for (;;) {
2178                 /* Keep lp alive when the lnet_net_lock is unlocked */
2179                 lnet_peer_addref_locked(lp);
2180                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2181                 if (signal_pending(current))
2182                         break;
2183                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2184                         break;
2185                 /*
2186                  * Don't repeat discovery if discovery is disabled. This is
2187                  * done to ensure we can use discovery as a standard ping as
2188                  * well for backwards compatibility with routers which do not
2189                  * have discovery or have discovery disabled
2190                  */
2191                 if (lnet_is_discovery_disabled(lp) && count > 0)
2192                         break;
2193                 if (lp->lp_dc_error)
2194                         break;
2195                 if (lnet_peer_is_uptodate(lp))
2196                         break;
2197                 lnet_peer_queue_for_discovery(lp);
2198                 count++;
2199                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2200
2201                 /*
2202                  * If caller requested a non-blocking operation then
2203                  * return immediately. Once discovery is complete any
2204                  * pending messages that were stopped due to discovery
2205                  * will be transmitted.
2206                  */
2207                 if (!block)
2208                         break;
2209
2210                 lnet_net_unlock(LNET_LOCK_EX);
2211                 schedule();
2212                 finish_wait(&lp->lp_dc_waitq, &wait);
2213                 lnet_net_lock(LNET_LOCK_EX);
2214                 lnet_peer_decref_locked(lp);
2215                 /* Peer may have changed */
2216                 lp = lpni->lpni_peer_net->lpn_peer;
2217         }
2218         finish_wait(&lp->lp_dc_waitq, &wait);
2219
2220         lnet_net_unlock(LNET_LOCK_EX);
2221         lnet_net_lock(cpt);
2222         lnet_peer_decref_locked(lp);
2223         /*
2224          * The peer may have changed, so re-check and rediscover if that turns
2225          * out to have been the case. The reference count on lp ensured that
2226          * even if it was unlinked from lpni the memory could not be recycled.
2227          * Thus the check below is sufficient to determine whether the peer
2228          * changed. If the peer changed, then lp must not be dereferenced.
2229          */
2230         if (lp != lpni->lpni_peer_net->lpn_peer)
2231                 goto again;
2232
2233         if (signal_pending(current))
2234                 rc = -EINTR;
2235         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2236                 rc = -ESHUTDOWN;
2237         else if (lp->lp_dc_error)
2238                 rc = lp->lp_dc_error;
2239         else if (!block)
2240                 CDEBUG(D_NET, "non-blocking discovery\n");
2241         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2242                 goto again;
2243
2244         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2245                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2246                libcfs_nid2str(lpni->lpni_nid), rc,
2247                (!block) ? "pending discovery" : "discovery complete");
2248
2249         return rc;
2250 }
2251
2252 /* Handle an incoming ack for a push. */
2253 static void
2254 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2255 {
2256         struct lnet_ping_buffer *pbuf;
2257
2258         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2259         spin_lock(&lp->lp_lock);
2260         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2261         lp->lp_push_error = ev->status;
2262         if (ev->status)
2263                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2264         else
2265                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2266         spin_unlock(&lp->lp_lock);
2267
2268         CDEBUG(D_NET, "peer %s ev->status %d\n",
2269                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2270 }
2271
2272 /* Handle a Reply message. This is the reply to a Ping message. */
2273 static void
2274 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2275 {
2276         struct lnet_ping_buffer *pbuf;
2277         int rc;
2278
2279         spin_lock(&lp->lp_lock);
2280
2281         /*
2282          * If some kind of error happened the contents of message
2283          * cannot be used. Set PING_FAILED to trigger a retry.
2284          */
2285         if (ev->status) {
2286                 lp->lp_state |= LNET_PEER_PING_FAILED;
2287                 lp->lp_ping_error = ev->status;
2288                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2289                        ev->status,
2290                        libcfs_nid2str(lp->lp_primary_nid),
2291                        libcfs_nid2str(ev->source.nid));
2292                 goto out;
2293         }
2294
2295         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2296         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2297                 lnet_swap_pinginfo(pbuf);
2298
2299         /*
2300          * A reply with invalid or corrupted info. Set PING_FAILED to
2301          * trigger a retry.
2302          */
2303         rc = lnet_ping_info_validate(&pbuf->pb_info);
2304         if (rc) {
2305                 lp->lp_state |= LNET_PEER_PING_FAILED;
2306                 lp->lp_ping_error = 0;
2307                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2308                        libcfs_nid2str(lp->lp_primary_nid), rc);
2309                 goto out;
2310         }
2311
2312
2313         /*
2314          * Only enable the multi-rail feature on the peer if both sides of
2315          * the connection have discovery on
2316          */
2317         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2318                 CDEBUG(D_NET, "Peer %s has Multi-Rail feature enabled\n",
2319                        libcfs_nid2str(lp->lp_primary_nid));
2320                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2321         } else {
2322                 CDEBUG(D_NET, "Peer %s has Multi-Rail feature disabled\n",
2323                        libcfs_nid2str(lp->lp_primary_nid));
2324                 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2325         }
2326
2327         /*
2328          * The peer may have discovery disabled at its end. Set
2329          * NO_DISCOVERY as appropriate.
2330          */
2331         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2332             !lnet_peer_discovery_disabled) {
2333                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2334                        libcfs_nid2str(lp->lp_primary_nid));
2335                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2336         } else {
2337                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2338                        libcfs_nid2str(lp->lp_primary_nid));
2339                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2340         }
2341
2342         /*
2343          * Update the MULTI_RAIL flag based on the reply. If the peer
2344          * was configured with DLC then the setting should match what
2345          * DLC put in.
2346          */
2347         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2348                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2349                         /* Everything's fine */
2350                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2351                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2352                               libcfs_nid2str(lp->lp_primary_nid));
2353                 } else {
2354                         /*
2355                          * if discovery is disabled then we don't want to
2356                          * update the state of the peer. All we'll do is
2357                          * update the peer_nis which were reported back in
2358                          * the initial ping
2359                          */
2360
2361                         if (!lnet_is_discovery_disabled_locked(lp)) {
2362                                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2363                                 lnet_peer_clr_non_mr_pref_nids(lp);
2364                         }
2365                 }
2366         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2367                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2368                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2369                               libcfs_nid2str(lp->lp_primary_nid));
2370                 } else {
2371                         CERROR("Multi-Rail state vanished from %s\n",
2372                                libcfs_nid2str(lp->lp_primary_nid));
2373                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2374                 }
2375         }
2376
2377         /*
2378          * Make sure we'll allocate the correct size ping buffer when
2379          * pinging the peer.
2380          */
2381         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2382                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2383
2384         /*
2385          * Check for truncation of the Reply. Clear PING_SENT and set
2386          * PING_FAILED to trigger a retry.
2387          */
2388         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2389                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2390                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2391                 lp->lp_state |= LNET_PEER_PING_FAILED;
2392                 lp->lp_ping_error = 0;
2393                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2394                        libcfs_nid2str(lp->lp_primary_nid),
2395                        pbuf->pb_info.pi_nnis);
2396                 goto out;
2397         }
2398
2399         /*
2400          * Check the sequence numbers in the reply. These are only
2401          * available if the reply came from a Multi-Rail peer.
2402          */
2403         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2404             pbuf->pb_info.pi_nnis > 1 &&
2405             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2406                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2407                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2408                                 libcfs_nid2str(lp->lp_primary_nid),
2409                                 LNET_PING_BUFFER_SEQNO(pbuf),
2410                                 lp->lp_peer_seqno);
2411
2412                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2413         }
2414
2415         /* We're happy with the state of the data in the buffer. */
2416         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2417                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2418         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2419                 lnet_ping_buffer_decref(lp->lp_data);
2420         else
2421                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2422         lnet_ping_buffer_addref(pbuf);
2423         lp->lp_data = pbuf;
2424 out:
2425         lp->lp_state &= ~LNET_PEER_PING_SENT;
2426         spin_unlock(&lp->lp_lock);
2427
2428         lnet_net_lock(LNET_LOCK_EX);
2429         /*
2430          * If this peer is a gateway, call the routing callback to
2431          * handle the ping reply
2432          */
2433         if (lp->lp_rtr_refcount > 0)
2434                 lnet_router_discovery_ping_reply(lp);
2435         lnet_net_unlock(LNET_LOCK_EX);
2436 }
2437
2438 /*
2439  * Send event handling. Only matters for error cases, where we clean
2440  * up state on the peer and peer_ni that would otherwise be updated in
2441  * the REPLY event handler for a successful Ping, and the ACK event
2442  * handler for a successful Push.
2443  */
2444 static int
2445 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2446 {
2447         int rc = 0;
2448
2449         if (!ev->status)
2450                 goto out;
2451
2452         spin_lock(&lp->lp_lock);
2453         if (ev->msg_type == LNET_MSG_GET) {
2454                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2455                 lp->lp_state |= LNET_PEER_PING_FAILED;
2456                 lp->lp_ping_error = ev->status;
2457         } else { /* ev->msg_type == LNET_MSG_PUT */
2458                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2459                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2460                 lp->lp_push_error = ev->status;
2461         }
2462         spin_unlock(&lp->lp_lock);
2463         rc = LNET_REDISCOVER_PEER;
2464 out:
2465         CDEBUG(D_NET, "%s Send to %s: %d\n",
2466                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2467                 libcfs_nid2str(ev->target.nid), rc);
2468         return rc;
2469 }
2470
2471 /*
2472  * Unlink event handling. This event is only seen if a call to
2473  * LNetMDUnlink() caused the event to be unlinked. If this call was
2474  * made after the event was set up in LNetGet() or LNetPut() then we
2475  * assume the Ping or Push timed out.
2476  */
2477 static void
2478 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2479 {
2480         spin_lock(&lp->lp_lock);
2481         /* We've passed through LNetGet() */
2482         if (lp->lp_state & LNET_PEER_PING_SENT) {
2483                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2484                 lp->lp_state |= LNET_PEER_PING_FAILED;
2485                 lp->lp_ping_error = -ETIMEDOUT;
2486                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2487                         libcfs_nid2str(lp->lp_primary_nid));
2488         }
2489         /* We've passed through LNetPut() */
2490         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2491                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2492                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2493                 lp->lp_push_error = -ETIMEDOUT;
2494                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2495                         libcfs_nid2str(lp->lp_primary_nid));
2496         }
2497         spin_unlock(&lp->lp_lock);
2498 }
2499
2500 /*
2501  * Event handler for the discovery EQ.
2502  *
2503  * Called with lnet_res_lock(cpt) held. The cpt is the
2504  * lnet_cpt_of_cookie() of the md handle cookie.
2505  */
2506 static void lnet_discovery_event_handler(struct lnet_event *event)
2507 {
2508         struct lnet_peer *lp = event->md.user_ptr;
2509         struct lnet_ping_buffer *pbuf;
2510         int rc;
2511
2512         /* discovery needs to take another look */
2513         rc = LNET_REDISCOVER_PEER;
2514
2515         CDEBUG(D_NET, "Received event: %d\n", event->type);
2516
2517         switch (event->type) {
2518         case LNET_EVENT_ACK:
2519                 lnet_discovery_event_ack(lp, event);
2520                 break;
2521         case LNET_EVENT_REPLY:
2522                 lnet_discovery_event_reply(lp, event);
2523                 break;
2524         case LNET_EVENT_SEND:
2525                 /* Only send failure triggers a retry. */
2526                 rc = lnet_discovery_event_send(lp, event);
2527                 break;
2528         case LNET_EVENT_UNLINK:
2529                 /* LNetMDUnlink() was called */
2530                 lnet_discovery_event_unlink(lp, event);
2531                 break;
2532         default:
2533                 /* Invalid events. */
2534                 LBUG();
2535         }
2536         lnet_net_lock(LNET_LOCK_EX);
2537         if (event->unlinked) {
2538                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2539                 lnet_ping_buffer_decref(pbuf);
2540                 lnet_peer_decref_locked(lp);
2541         }
2542
2543         /* put peer back at end of request queue, if discovery not already
2544          * done */
2545         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2546                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2547                 wake_up(&the_lnet.ln_dc_waitq);
2548         }
2549         lnet_net_unlock(LNET_LOCK_EX);
2550 }
2551
2552 /*
2553  * Build a peer from incoming data.
2554  *
2555  * The NIDs in the incoming data are supposed to be structured as follows:
2556  *  - loopback
2557  *  - primary NID
2558  *  - other NIDs in same net
2559  *  - NIDs in second net
2560  *  - NIDs in third net
2561  *  - ...
2562  * This due to the way the list of NIDs in the data is created.
2563  *
2564  * Note that this function will mark the peer uptodate unless an
2565  * ENOMEM is encontered. All other errors are due to a conflict
2566  * between the DLC configuration and what discovery sees. We treat DLC
2567  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2568  * peer from becoming stuck in discovery.
2569  */
2570 static int lnet_peer_merge_data(struct lnet_peer *lp,
2571                                 struct lnet_ping_buffer *pbuf)
2572 {
2573         struct lnet_peer_ni *lpni;
2574         lnet_nid_t *curnis = NULL;
2575         struct lnet_ni_status *addnis = NULL;
2576         lnet_nid_t *delnis = NULL;
2577         unsigned flags;
2578         int ncurnis;
2579         int naddnis;
2580         int ndelnis;
2581         int nnis = 0;
2582         int i;
2583         int j;
2584         int rc;
2585
2586         flags = LNET_PEER_DISCOVERED;
2587         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2588                 flags |= LNET_PEER_MULTI_RAIL;
2589
2590         /*
2591          * Cache the routing feature for the peer; whether it is enabled
2592          * for disabled as reported by the remote peer.
2593          */
2594         spin_lock(&lp->lp_lock);
2595         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2596                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2597         else
2598                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2599         spin_unlock(&lp->lp_lock);
2600
2601         nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2602         LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2603         LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2604         LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2605         if (!curnis || !addnis || !delnis) {
2606                 rc = -ENOMEM;
2607                 goto out;
2608         }
2609         ncurnis = 0;
2610         naddnis = 0;
2611         ndelnis = 0;
2612
2613         /* Construct the list of NIDs present in peer. */
2614         lpni = NULL;
2615         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2616                 curnis[ncurnis++] = lpni->lpni_nid;
2617
2618         /*
2619          * Check for NIDs in pbuf not present in curnis[].
2620          * The loop starts at 1 to skip the loopback NID.
2621          */
2622         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2623                 for (j = 0; j < ncurnis; j++)
2624                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2625                                 break;
2626                 if (j == ncurnis)
2627                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2628         }
2629         /*
2630          * Check for NIDs in curnis[] not present in pbuf.
2631          * The nested loop starts at 1 to skip the loopback NID.
2632          *
2633          * But never add the loopback NID to delnis[]: if it is
2634          * present in curnis[] then this peer is for this node.
2635          */
2636         for (i = 0; i < ncurnis; i++) {
2637                 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2638                         continue;
2639                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2640                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2641                                 /*
2642                                  * update the information we cache for the
2643                                  * peer with the latest information we
2644                                  * received
2645                                  */
2646                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2647                                 if (lpni) {
2648                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2649                                         lnet_peer_ni_decref_locked(lpni);
2650                                 }
2651                                 break;
2652                         }
2653                 }
2654                 if (j == pbuf->pb_info.pi_nnis)
2655                         delnis[ndelnis++] = curnis[i];
2656         }
2657
2658         /*
2659          * If we get here and the discovery is disabled then we don't want
2660          * to add or delete any NIs. We just updated the ones we have some
2661          * information on, and call it a day
2662          */
2663         rc = 0;
2664         if (lnet_is_discovery_disabled(lp))
2665                 goto out;
2666
2667         for (i = 0; i < naddnis; i++) {
2668                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2669                 if (rc) {
2670                         CERROR("Error adding NID %s to peer %s: %d\n",
2671                                libcfs_nid2str(addnis[i].ns_nid),
2672                                libcfs_nid2str(lp->lp_primary_nid), rc);
2673                         if (rc == -ENOMEM)
2674                                 goto out;
2675                 }
2676                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2677                 if (lpni) {
2678                         lpni->lpni_ns_status = addnis[i].ns_status;
2679                         lnet_peer_ni_decref_locked(lpni);
2680                 }
2681         }
2682
2683         for (i = 0; i < ndelnis; i++) {
2684                 /*
2685                  * for routers it's okay to delete the primary_nid because
2686                  * the upper layers don't really rely on it. So if we're
2687                  * being told that the router changed its primary_nid
2688                  * then it's okay to delete it.
2689                  */
2690                 if (lp->lp_rtr_refcount > 0)
2691                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2692                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2693                 if (rc) {
2694                         CERROR("Error deleting NID %s from peer %s: %d\n",
2695                                libcfs_nid2str(delnis[i]),
2696                                libcfs_nid2str(lp->lp_primary_nid), rc);
2697                         if (rc == -ENOMEM)
2698                                 goto out;
2699                 }
2700         }
2701         /*
2702          * Errors other than -ENOMEM are due to peers having been
2703          * configured with DLC. Ignore these because DLC overrides
2704          * Discovery.
2705          */
2706         rc = 0;
2707 out:
2708         LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2709         LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2710         LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2711         lnet_ping_buffer_decref(pbuf);
2712         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2713
2714         if (rc) {
2715                 spin_lock(&lp->lp_lock);
2716                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2717                 lp->lp_state |= LNET_PEER_FORCE_PING;
2718                 spin_unlock(&lp->lp_lock);
2719         }
2720         return rc;
2721 }
2722
2723 /*
2724  * The data in pbuf says lp is its primary peer, but the data was
2725  * received by a different peer. Try to update lp with the data.
2726  */
2727 static int
2728 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2729 {
2730         struct lnet_handle_md mdh;
2731
2732         /* Queue lp for discovery, and force it on the request queue. */
2733         lnet_net_lock(LNET_LOCK_EX);
2734         if (lnet_peer_queue_for_discovery(lp))
2735                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2736         lnet_net_unlock(LNET_LOCK_EX);
2737
2738         LNetInvalidateMDHandle(&mdh);
2739
2740         /*
2741          * Decide whether we can move the peer to the DATA_PRESENT state.
2742          *
2743          * We replace stale data for a multi-rail peer, repair PING_FAILED
2744          * status, and preempt FORCE_PING.
2745          *
2746          * If after that we have DATA_PRESENT, we merge it into this peer.
2747          */
2748         spin_lock(&lp->lp_lock);
2749         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2750                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2751                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2752                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2753                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2754                         lnet_ping_buffer_decref(pbuf);
2755                         pbuf = lp->lp_data;
2756                         lp->lp_data = NULL;
2757                 }
2758         }
2759         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2760                 lnet_ping_buffer_decref(lp->lp_data);
2761                 lp->lp_data = NULL;
2762                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2763         }
2764         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2765                 mdh = lp->lp_ping_mdh;
2766                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2767                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2768                 lp->lp_ping_error = 0;
2769         }
2770         if (lp->lp_state & LNET_PEER_FORCE_PING)
2771                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2772         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2773         spin_unlock(&lp->lp_lock);
2774
2775         if (!LNetMDHandleIsInvalid(mdh))
2776                 LNetMDUnlink(mdh);
2777
2778         if (pbuf)
2779                 return lnet_peer_merge_data(lp, pbuf);
2780
2781         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2782         return 0;
2783 }
2784
2785 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2786 {
2787         int i;
2788
2789         for (i = 0; i < pinfo->pi_nnis; i++) {
2790                 if (pinfo->pi_ni[i].ns_nid == nid)
2791                         return true;
2792         }
2793
2794         return false;
2795 }
2796
2797 /*
2798  * Update a peer using the data received.
2799  */
2800 static int lnet_peer_data_present(struct lnet_peer *lp)
2801 __must_hold(&lp->lp_lock)
2802 {
2803         struct lnet_ping_buffer *pbuf;
2804         struct lnet_peer_ni *lpni;
2805         lnet_nid_t nid = LNET_NID_ANY;
2806         unsigned flags;
2807         int rc = 0;
2808
2809         pbuf = lp->lp_data;
2810         lp->lp_data = NULL;
2811         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2812         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2813         spin_unlock(&lp->lp_lock);
2814
2815         /*
2816          * Modifications of peer structures are done while holding the
2817          * ln_api_mutex. A global lock is required because we may be
2818          * modifying multiple peer structures, and a mutex greatly
2819          * simplifies memory management.
2820          *
2821          * The actual changes to the data structures must also protect
2822          * against concurrent lookups, for which the lnet_net_lock in
2823          * LNET_LOCK_EX mode is used.
2824          */
2825         mutex_lock(&the_lnet.ln_api_mutex);
2826         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2827                 rc = -ESHUTDOWN;
2828                 goto out;
2829         }
2830
2831         /*
2832          * If this peer is not on the peer list then it is being torn
2833          * down, and our reference count may be all that is keeping it
2834          * alive. Don't do any work on it.
2835          */
2836         if (list_empty(&lp->lp_peer_list))
2837                 goto out;
2838
2839         flags = LNET_PEER_DISCOVERED;
2840         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2841                 flags |= LNET_PEER_MULTI_RAIL;
2842
2843         /*
2844          * Check whether the primary NID in the message matches the
2845          * primary NID of the peer. If it does, update the peer, if
2846          * it it does not, check whether there is already a peer with
2847          * that primary NID. If no such peer exists, try to update
2848          * the primary NID of the current peer (allowed if it was
2849          * created due to message traffic) and complete the update.
2850          * If the peer did exist, hand off the data to it.
2851          *
2852          * The peer for the loopback interface is a special case: this
2853          * is the peer for the local node, and we want to set its
2854          * primary NID to the correct value here. Moreover, this peer
2855          * can show up with only the loopback NID in the ping buffer.
2856          */
2857         if (pbuf->pb_info.pi_nnis <= 1)
2858                 goto out;
2859         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2860         if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2861                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2862                 if (!rc)
2863                         rc = lnet_peer_merge_data(lp, pbuf);
2864         /*
2865          * if the primary nid of the peer is present in the ping info returned
2866          * from the peer, but it's not the local primary peer we have
2867          * cached and discovery is disabled, then we don't want to update
2868          * our local peer info, by adding or removing NIDs, we just want
2869          * to update the status of the nids that we currently have
2870          * recorded in that peer.
2871          */
2872         } else if (lp->lp_primary_nid == nid ||
2873                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2874                     lnet_is_discovery_disabled(lp))) {
2875                 rc = lnet_peer_merge_data(lp, pbuf);
2876         } else {
2877                 lpni = lnet_find_peer_ni_locked(nid);
2878                 if (!lpni) {
2879                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2880                         if (rc) {
2881                                 CERROR("Primary NID error %s versus %s: %d\n",
2882                                        libcfs_nid2str(lp->lp_primary_nid),
2883                                        libcfs_nid2str(nid), rc);
2884                         } else {
2885                                 rc = lnet_peer_merge_data(lp, pbuf);
2886                         }
2887                 } else {
2888                         struct lnet_peer *new_lp;
2889                         new_lp = lpni->lpni_peer_net->lpn_peer;
2890                         /*
2891                          * if lp has discovery/MR enabled that means new_lp
2892                          * should have discovery/MR enabled as well, since
2893                          * it's the same peer, which we're about to merge
2894                          */
2895                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2896                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2897                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2898                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2899
2900                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2901                         lnet_consolidate_routes_locked(lp, new_lp);
2902                         lnet_peer_ni_decref_locked(lpni);
2903                 }
2904         }
2905 out:
2906         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2907                lp->lp_state);
2908         mutex_unlock(&the_lnet.ln_api_mutex);
2909
2910         spin_lock(&lp->lp_lock);
2911         /* Tell discovery to re-check the peer immediately. */
2912         if (!rc)
2913                 rc = LNET_REDISCOVER_PEER;
2914         return rc;
2915 }
2916
2917 /*
2918  * A ping failed. Clear the PING_FAILED state and set the
2919  * FORCE_PING state, to ensure a retry even if discovery is
2920  * disabled. This avoids being left with incorrect state.
2921  */
2922 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2923 __must_hold(&lp->lp_lock)
2924 {
2925         struct lnet_handle_md mdh;
2926         int rc;
2927
2928         mdh = lp->lp_ping_mdh;
2929         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2930         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2931         lp->lp_state |= LNET_PEER_FORCE_PING;
2932         rc = lp->lp_ping_error;
2933         lp->lp_ping_error = 0;
2934         spin_unlock(&lp->lp_lock);
2935
2936         if (!LNetMDHandleIsInvalid(mdh))
2937                 LNetMDUnlink(mdh);
2938
2939         CDEBUG(D_NET, "peer %s:%d\n",
2940                libcfs_nid2str(lp->lp_primary_nid), rc);
2941
2942         spin_lock(&lp->lp_lock);
2943         return rc ? rc : LNET_REDISCOVER_PEER;
2944 }
2945
2946 /*
2947  * Select NID to send a Ping or Push to.
2948  */
2949 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2950 {
2951         struct lnet_peer_ni *lpni;
2952
2953         /* Look for a direct-connected NID for this peer. */
2954         lpni = NULL;
2955         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2956                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2957                         continue;
2958                 break;
2959         }
2960         if (lpni)
2961                 return lpni->lpni_nid;
2962
2963         /* Look for a routed-connected NID for this peer. */
2964         lpni = NULL;
2965         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2966                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2967                         continue;
2968                 break;
2969         }
2970         if (lpni)
2971                 return lpni->lpni_nid;
2972
2973         return LNET_NID_ANY;
2974 }
2975
2976 /* Active side of ping. */
2977 static int lnet_peer_send_ping(struct lnet_peer *lp)
2978 __must_hold(&lp->lp_lock)
2979 {
2980         lnet_nid_t pnid;
2981         int nnis;
2982         int rc;
2983         int cpt;
2984
2985         lp->lp_state |= LNET_PEER_PING_SENT;
2986         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2987         spin_unlock(&lp->lp_lock);
2988
2989         cpt = lnet_net_lock_current();
2990         /* Refcount for MD. */
2991         lnet_peer_addref_locked(lp);
2992         pnid = lnet_peer_select_nid(lp);
2993         lnet_net_unlock(cpt);
2994
2995         nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2996
2997         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2998                             the_lnet.ln_dc_eqh, false);
2999
3000         /*
3001          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3002          * refcount on the peer, otherwise LNetMDUnlink will be called
3003          * which will eventually do that.
3004          */
3005         if (rc > 0) {
3006                 lnet_net_lock(cpt);
3007                 lnet_peer_decref_locked(lp);
3008                 lnet_net_unlock(cpt);
3009                 rc = -rc; /* change the rc to negative value */
3010                 goto fail_error;
3011         } else if (rc < 0) {
3012                 goto fail_error;
3013         }
3014
3015         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3016
3017         spin_lock(&lp->lp_lock);
3018         return 0;
3019
3020 fail_error:
3021         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3022         /*
3023          * The errors that get us here are considered hard errors and
3024          * cause Discovery to terminate. So we clear PING_SENT, but do
3025          * not set either PING_FAILED or FORCE_PING. In fact we need
3026          * to clear PING_FAILED, because the unlink event handler will
3027          * have set it if we called LNetMDUnlink() above.
3028          */
3029         spin_lock(&lp->lp_lock);
3030         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3031         return rc;
3032 }
3033
3034 /*
3035  * This function exists because you cannot call LNetMDUnlink() from an
3036  * event handler.
3037  */
3038 static int lnet_peer_push_failed(struct lnet_peer *lp)
3039 __must_hold(&lp->lp_lock)
3040 {
3041         struct lnet_handle_md mdh;
3042         int rc;
3043
3044         mdh = lp->lp_push_mdh;
3045         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3046         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3047         rc = lp->lp_push_error;
3048         lp->lp_push_error = 0;
3049         spin_unlock(&lp->lp_lock);
3050
3051         if (!LNetMDHandleIsInvalid(mdh))
3052                 LNetMDUnlink(mdh);
3053
3054         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3055         spin_lock(&lp->lp_lock);
3056         return rc ? rc : LNET_REDISCOVER_PEER;
3057 }
3058
3059 /* Active side of push. */
3060 static int lnet_peer_send_push(struct lnet_peer *lp)
3061 __must_hold(&lp->lp_lock)
3062 {
3063         struct lnet_ping_buffer *pbuf;
3064         struct lnet_process_id id;
3065         struct lnet_md md;
3066         int cpt;
3067         int rc;
3068
3069         /* Don't push to a non-multi-rail peer. */
3070         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3071                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3072                 return 0;
3073         }
3074
3075         lp->lp_state |= LNET_PEER_PUSH_SENT;
3076         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3077         spin_unlock(&lp->lp_lock);
3078
3079         cpt = lnet_net_lock_current();
3080         pbuf = the_lnet.ln_ping_target;
3081         lnet_ping_buffer_addref(pbuf);
3082         lnet_net_unlock(cpt);
3083
3084         /* Push source MD */
3085         md.start     = &pbuf->pb_info;
3086         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3087         md.threshold = 2; /* Put/Ack */
3088         md.max_size  = 0;
3089         md.options   = 0;
3090         md.eq_handle = the_lnet.ln_dc_eqh;
3091         md.user_ptr  = lp;
3092
3093         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3094         if (rc) {
3095                 lnet_ping_buffer_decref(pbuf);
3096                 CERROR("Can't bind push source MD: %d\n", rc);
3097                 goto fail_error;
3098         }
3099         cpt = lnet_net_lock_current();
3100         /* Refcount for MD. */
3101         lnet_peer_addref_locked(lp);
3102         id.pid = LNET_PID_LUSTRE;
3103         id.nid = lnet_peer_select_nid(lp);
3104         lnet_net_unlock(cpt);
3105
3106         if (id.nid == LNET_NID_ANY) {
3107                 rc = -EHOSTUNREACH;
3108                 goto fail_unlink;
3109         }
3110
3111         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3112                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3113                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3114
3115         if (rc)
3116                 goto fail_unlink;
3117
3118         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3119
3120         spin_lock(&lp->lp_lock);
3121         return 0;
3122
3123 fail_unlink:
3124         LNetMDUnlink(lp->lp_push_mdh);
3125         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3126 fail_error:
3127         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3128         /*
3129          * The errors that get us here are considered hard errors and
3130          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3131          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3132          * because the unlink event handler will have set it if we
3133          * called LNetMDUnlink() above.
3134          */
3135         spin_lock(&lp->lp_lock);
3136         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3137         return rc;
3138 }
3139
3140 /*
3141  * An unrecoverable error was encountered during discovery.
3142  * Set error status in peer and abort discovery.
3143  */
3144 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3145 {
3146         CDEBUG(D_NET, "Discovery error %s: %d\n",
3147                libcfs_nid2str(lp->lp_primary_nid), error);
3148
3149         spin_lock(&lp->lp_lock);
3150         lp->lp_dc_error = error;
3151         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3152         lp->lp_state |= LNET_PEER_REDISCOVER;
3153         spin_unlock(&lp->lp_lock);
3154 }
3155
3156 /*
3157  * Mark the peer as discovered.
3158  */
3159 static int lnet_peer_discovered(struct lnet_peer *lp)
3160 __must_hold(&lp->lp_lock)
3161 {
3162         lp->lp_state |= LNET_PEER_DISCOVERED;
3163         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3164                           LNET_PEER_REDISCOVER);
3165
3166         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3167
3168         return 0;
3169 }
3170
3171
3172 /*
3173  * Discovering this peer is taking too long. Cancel any Ping or Push
3174  * that discovery is waiting on by unlinking the relevant MDs. The
3175  * lnet_discovery_event_handler() will proceed from here and complete
3176  * the cleanup.
3177  */
3178 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3179 {
3180         struct lnet_handle_md ping_mdh;
3181         struct lnet_handle_md push_mdh;
3182
3183         LNetInvalidateMDHandle(&ping_mdh);
3184         LNetInvalidateMDHandle(&push_mdh);
3185
3186         spin_lock(&lp->lp_lock);
3187         if (lp->lp_state & LNET_PEER_PING_SENT) {
3188                 ping_mdh = lp->lp_ping_mdh;
3189                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3190         }
3191         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3192                 push_mdh = lp->lp_push_mdh;
3193                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3194         }
3195         spin_unlock(&lp->lp_lock);
3196
3197         if (!LNetMDHandleIsInvalid(ping_mdh))
3198                 LNetMDUnlink(ping_mdh);
3199         if (!LNetMDHandleIsInvalid(push_mdh))
3200                 LNetMDUnlink(push_mdh);
3201 }
3202
3203 /*
3204  * Wait for work to be queued or some other change that must be
3205  * attended to. Returns non-zero if the discovery thread should shut
3206  * down.
3207  */
3208 static int lnet_peer_discovery_wait_for_work(void)
3209 {
3210         int cpt;
3211         int rc = 0;
3212
3213         DEFINE_WAIT(wait);
3214
3215         cpt = lnet_net_lock_current();
3216         for (;;) {
3217                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3218                                 TASK_INTERRUPTIBLE);
3219                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3220                         break;
3221                 if (lnet_push_target_resize_needed())
3222                         break;
3223                 if (!list_empty(&the_lnet.ln_dc_request))
3224                         break;
3225                 if (!list_empty(&the_lnet.ln_msg_resend))
3226                         break;
3227                 lnet_net_unlock(cpt);
3228
3229                 /*
3230                  * wakeup max every second to check if there are peers that
3231                  * have been stuck on the working queue for greater than
3232                  * the peer timeout.
3233                  */
3234                 schedule_timeout(cfs_time_seconds(1));
3235                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3236                 cpt = lnet_net_lock_current();
3237         }
3238         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3239
3240         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3241                 rc = -ESHUTDOWN;
3242
3243         lnet_net_unlock(cpt);
3244
3245         CDEBUG(D_NET, "woken: %d\n", rc);
3246
3247         return rc;
3248 }
3249
3250 /*
3251  * Messages that were pending on a destroyed peer will be put on a global
3252  * resend list. The message resend list will be checked by
3253  * the discovery thread when it wakes up, and will resend messages. These
3254  * messages can still be sendable in the case the lpni which was the initial
3255  * cause of the message re-queue was transfered to another peer.
3256  *
3257  * It is possible that LNet could be shutdown while we're iterating
3258  * through the list. lnet_shudown_lndnets() will attempt to access the
3259  * resend list, but will have to wait until the spinlock is released, by
3260  * which time there shouldn't be any more messages on the resend list.
3261  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3262  * for the messages so they can be released. The other case is that
3263  * lnet_shudown_lndnets() can finalize all the messages before this
3264  * function can visit the resend list, in which case this function will be
3265  * a no-op.
3266  */
3267 static void lnet_resend_msgs(void)
3268 {
3269         struct lnet_msg *msg, *tmp;
3270         struct list_head resend;
3271         int rc;
3272
3273         INIT_LIST_HEAD(&resend);
3274
3275         spin_lock(&the_lnet.ln_msg_resend_lock);
3276         list_splice(&the_lnet.ln_msg_resend, &resend);
3277         spin_unlock(&the_lnet.ln_msg_resend_lock);
3278
3279         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3280                 list_del_init(&msg->msg_list);
3281                 rc = lnet_send(msg->msg_src_nid_param, msg,
3282                                msg->msg_rtr_nid_param);
3283                 if (rc < 0) {
3284                         CNETERR("Error sending %s to %s: %d\n",
3285                                lnet_msgtyp2str(msg->msg_type),
3286                                libcfs_id2str(msg->msg_target), rc);
3287                         lnet_finalize(msg, rc);
3288                 }
3289         }
3290 }
3291
3292 /* The discovery thread. */
3293 static int lnet_peer_discovery(void *arg)
3294 {
3295         struct lnet_peer *lp;
3296         int rc;
3297
3298         wait_for_completion(&the_lnet.ln_started);
3299
3300         CDEBUG(D_NET, "started\n");
3301         cfs_block_allsigs();
3302
3303         for (;;) {
3304                 if (lnet_peer_discovery_wait_for_work())
3305                         break;
3306
3307                 lnet_resend_msgs();
3308
3309                 if (lnet_push_target_resize_needed())
3310                         lnet_push_target_resize();
3311
3312                 lnet_net_lock(LNET_LOCK_EX);
3313                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3314                         break;
3315
3316                 /*
3317                  * Process all incoming discovery work requests.  When
3318                  * discovery must wait on a peer to change state, it
3319                  * is added to the tail of the ln_dc_working queue. A
3320                  * timestamp keeps track of when the peer was added,
3321                  * so we can time out discovery requests that take too
3322                  * long.
3323                  */
3324                 while (!list_empty(&the_lnet.ln_dc_request)) {
3325                         lp = list_first_entry(&the_lnet.ln_dc_request,
3326                                               struct lnet_peer, lp_dc_list);
3327                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3328                         /*
3329                          * set the time the peer was put on the dc_working
3330                          * queue. It shouldn't remain on the queue
3331                          * forever, in case the GET message (for ping)
3332                          * doesn't get a REPLY or the PUT message (for
3333                          * push) doesn't get an ACK.
3334                          */
3335                         lp->lp_last_queued = ktime_get_real_seconds();
3336                         lnet_net_unlock(LNET_LOCK_EX);
3337
3338                         /*
3339                          * Select an action depending on the state of
3340                          * the peer and whether discovery is disabled.
3341                          * The check whether discovery is disabled is
3342                          * done after the code that handles processing
3343                          * for arrived data, cleanup for failures, and
3344                          * forcing a Ping or Push.
3345                          */
3346                         spin_lock(&lp->lp_lock);
3347                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3348                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3349                                 lp->lp_state);
3350                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3351                                 rc = lnet_peer_data_present(lp);
3352                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3353                                 rc = lnet_peer_ping_failed(lp);
3354                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3355                                 rc = lnet_peer_push_failed(lp);
3356                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3357                                 rc = lnet_peer_send_ping(lp);
3358                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3359                                 rc = lnet_peer_send_push(lp);
3360                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3361                                 rc = lnet_peer_send_ping(lp);
3362                         else if (lnet_peer_needs_push(lp))
3363                                 rc = lnet_peer_send_push(lp);
3364                         else
3365                                 rc = lnet_peer_discovered(lp);
3366                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3367                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3368                                 lp->lp_state, rc);
3369                         spin_unlock(&lp->lp_lock);
3370
3371                         lnet_net_lock(LNET_LOCK_EX);
3372                         if (rc == LNET_REDISCOVER_PEER) {
3373                                 list_move(&lp->lp_dc_list,
3374                                           &the_lnet.ln_dc_request);
3375                         } else if (rc) {
3376                                 lnet_peer_discovery_error(lp, rc);
3377                         }
3378                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3379                                 lnet_peer_discovery_complete(lp);
3380                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3381                                 break;
3382                 }
3383
3384                 lnet_net_unlock(LNET_LOCK_EX);
3385         }
3386
3387         CDEBUG(D_NET, "stopping\n");
3388         /*
3389          * Clean up before telling lnet_peer_discovery_stop() that
3390          * we're done. Use wake_up() below to somewhat reduce the
3391          * size of the thundering herd if there are multiple threads
3392          * waiting on discovery of a single peer.
3393          */
3394
3395         /* Queue cleanup 1: stop all pending pings and pushes. */
3396         lnet_net_lock(LNET_LOCK_EX);
3397         while (!list_empty(&the_lnet.ln_dc_working)) {
3398                 lp = list_first_entry(&the_lnet.ln_dc_working,
3399                                       struct lnet_peer, lp_dc_list);
3400                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3401                 lnet_net_unlock(LNET_LOCK_EX);
3402                 lnet_peer_cancel_discovery(lp);
3403                 lnet_net_lock(LNET_LOCK_EX);
3404         }
3405         lnet_net_unlock(LNET_LOCK_EX);
3406
3407         /* Queue cleanup 2: wait for the expired queue to clear. */
3408         while (!list_empty(&the_lnet.ln_dc_expired))
3409                 schedule_timeout(cfs_time_seconds(1));
3410
3411         /* Queue cleanup 3: clear the request queue. */
3412         lnet_net_lock(LNET_LOCK_EX);
3413         while (!list_empty(&the_lnet.ln_dc_request)) {
3414                 lp = list_first_entry(&the_lnet.ln_dc_request,
3415                                       struct lnet_peer, lp_dc_list);
3416                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3417                 lnet_peer_discovery_complete(lp);
3418         }
3419         lnet_net_unlock(LNET_LOCK_EX);
3420
3421         LNetEQFree(the_lnet.ln_dc_eqh);
3422         LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3423
3424         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3425         wake_up(&the_lnet.ln_dc_waitq);
3426
3427         CDEBUG(D_NET, "stopped\n");
3428
3429         return 0;
3430 }
3431
3432 /* ln_api_mutex is held on entry. */
3433 int lnet_peer_discovery_start(void)
3434 {
3435         struct task_struct *task;
3436         int rc;
3437
3438         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3439                 return -EALREADY;
3440
3441         rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3442         if (rc != 0) {
3443                 CERROR("Can't allocate discovery EQ: %d\n", rc);
3444                 return rc;
3445         }
3446
3447         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3448         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3449         if (IS_ERR(task)) {
3450                 rc = PTR_ERR(task);
3451                 CERROR("Can't start peer discovery thread: %d\n", rc);
3452
3453                 LNetEQFree(the_lnet.ln_dc_eqh);
3454                 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3455
3456                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3457         }
3458
3459         CDEBUG(D_NET, "discovery start: %d\n", rc);
3460
3461         return rc;
3462 }
3463
3464 /* ln_api_mutex is held on entry. */
3465 void lnet_peer_discovery_stop(void)
3466 {
3467         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3468                 return;
3469
3470         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3471         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3472
3473         /* In the LNetNIInit() path we may be stopping discovery before it
3474          * entered its work loop
3475          */
3476         if (!completion_done(&the_lnet.ln_started))
3477                 complete(&the_lnet.ln_started);
3478         else
3479                 wake_up(&the_lnet.ln_dc_waitq);
3480
3481         wait_event(the_lnet.ln_dc_waitq,
3482                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3483
3484         LASSERT(list_empty(&the_lnet.ln_dc_request));
3485         LASSERT(list_empty(&the_lnet.ln_dc_working));
3486         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3487
3488         CDEBUG(D_NET, "discovery stopped\n");
3489 }
3490
3491 /* Debugging */
3492
3493 void
3494 lnet_debug_peer(lnet_nid_t nid)
3495 {
3496         char                    *aliveness = "NA";
3497         struct lnet_peer_ni     *lp;
3498         int                     cpt;
3499
3500         cpt = lnet_cpt_of_nid(nid, NULL);
3501         lnet_net_lock(cpt);
3502
3503         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3504         if (IS_ERR(lp)) {
3505                 lnet_net_unlock(cpt);
3506                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3507                 return;
3508         }
3509
3510         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3511                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3512
3513         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3514                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3515                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3516                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3517                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3518
3519         lnet_peer_ni_decref_locked(lp);
3520
3521         lnet_net_unlock(cpt);
3522 }
3523
3524 /* Gathering information for userspace. */
3525
3526 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3527                           char aliveness[LNET_MAX_STR_LEN],
3528                           __u32 *cpt_iter, __u32 *refcount,
3529                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3530                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3531                           __u32 *peer_tx_qnob)
3532 {
3533         struct lnet_peer_table          *peer_table;
3534         struct lnet_peer_ni             *lp;
3535         int                             j;
3536         int                             lncpt;
3537         bool                            found = false;
3538
3539         /* get the number of CPTs */
3540         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3541
3542         /* if the cpt number to be examined is >= the number of cpts in
3543          * the system then indicate that there are no more cpts to examin
3544          */
3545         if (*cpt_iter >= lncpt)
3546                 return -ENOENT;
3547
3548         /* get the current table */
3549         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3550         /* if the ptable is NULL then there are no more cpts to examine */
3551         if (peer_table == NULL)
3552                 return -ENOENT;
3553
3554         lnet_net_lock(*cpt_iter);
3555
3556         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3557                 struct list_head *peers = &peer_table->pt_hash[j];
3558
3559                 list_for_each_entry(lp, peers, lpni_hashlist) {
3560                         if (peer_index-- > 0)
3561                                 continue;
3562
3563                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3564                         if (lnet_isrouter(lp) ||
3565                                 lnet_peer_aliveness_enabled(lp))
3566                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3567                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3568
3569                         *nid = lp->lpni_nid;
3570                         *refcount = atomic_read(&lp->lpni_refcount);
3571                         *ni_peer_tx_credits =
3572                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3573                         *peer_tx_credits = lp->lpni_txcredits;
3574                         *peer_rtr_credits = lp->lpni_rtrcredits;
3575                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3576                         *peer_tx_qnob = lp->lpni_txqnob;
3577
3578                         found = true;
3579                 }
3580
3581         }
3582         lnet_net_unlock(*cpt_iter);
3583
3584         *cpt_iter = lncpt;
3585
3586         return found ? 0 : -ENOENT;
3587 }
3588
3589 /* ln_api_mutex is held, which keeps the peer list stable */
3590 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3591 {
3592         struct lnet_ioctl_element_stats *lpni_stats;
3593         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3594         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3595         struct lnet_peer_ni_credit_info *lpni_info;
3596         struct lnet_peer_ni *lpni;
3597         struct lnet_peer *lp;
3598         lnet_nid_t nid;
3599         __u32 size;
3600         int rc;
3601
3602         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3603
3604         if (!lp) {
3605                 rc = -ENOENT;
3606                 goto out;
3607         }
3608
3609         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3610                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3611         size *= lp->lp_nnis;
3612         if (size > cfg->prcfg_size) {
3613                 cfg->prcfg_size = size;
3614                 rc = -E2BIG;
3615                 goto out_lp_decref;
3616         }
3617
3618         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3619         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3620         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3621         cfg->prcfg_count = lp->lp_nnis;
3622         cfg->prcfg_size = size;
3623         cfg->prcfg_state = lp->lp_state;
3624
3625         /* Allocate helper buffers. */
3626         rc = -ENOMEM;
3627         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3628         if (!lpni_info)
3629                 goto out_lp_decref;
3630         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3631         if (!lpni_stats)
3632                 goto out_free_info;
3633         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3634         if (!lpni_msg_stats)
3635                 goto out_free_stats;
3636         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3637         if (!lpni_hstats)
3638                 goto out_free_msg_stats;
3639
3640
3641         lpni = NULL;
3642         rc = -EFAULT;
3643         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3644                 nid = lpni->lpni_nid;
3645                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3646                         goto out_free_hstats;
3647                 bulk += sizeof(nid);
3648
3649                 memset(lpni_info, 0, sizeof(*lpni_info));
3650                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3651                 if (lnet_isrouter(lpni) ||
3652                         lnet_peer_aliveness_enabled(lpni))
3653                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3654                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3655
3656                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3657                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3658                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3659                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3660                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3661                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3662                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3663                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3664                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3665                         goto out_free_hstats;
3666                 bulk += sizeof(*lpni_info);
3667
3668                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3669                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3670                                                             LNET_STATS_TYPE_SEND);
3671                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3672                                                             LNET_STATS_TYPE_RECV);
3673                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3674                                                             LNET_STATS_TYPE_DROP);
3675                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3676                         goto out_free_hstats;
3677                 bulk += sizeof(*lpni_stats);
3678                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3679                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3680                         goto out_free_hstats;
3681                 bulk += sizeof(*lpni_msg_stats);
3682                 lpni_hstats->hlpni_network_timeout =
3683                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3684                 lpni_hstats->hlpni_remote_dropped =
3685                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3686                 lpni_hstats->hlpni_remote_timeout =
3687                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3688                 lpni_hstats->hlpni_remote_error =
3689                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3690                 lpni_hstats->hlpni_health_value =
3691                   atomic_read(&lpni->lpni_healthv);
3692                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3693                         goto out_free_hstats;
3694                 bulk += sizeof(*lpni_hstats);
3695         }
3696         rc = 0;
3697
3698 out_free_hstats:
3699         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3700 out_free_msg_stats:
3701         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3702 out_free_stats:
3703         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3704 out_free_info:
3705         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3706 out_lp_decref:
3707         lnet_peer_decref_locked(lp);
3708 out:
3709         return rc;
3710 }
3711
3712 void
3713 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3714 {
3715         /* the mt could've shutdown and cleaned up the queues */
3716         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3717                 return;
3718
3719         if (list_empty(&lpni->lpni_recovery) &&
3720             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3721                 CERROR("lpni %s added to recovery queue. Health = %d\n",
3722                         libcfs_nid2str(lpni->lpni_nid),
3723                         atomic_read(&lpni->lpni_healthv));
3724                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3725                 lnet_peer_ni_addref_locked(lpni);
3726         }
3727 }
3728
3729 /* Call with the ln_api_mutex held */
3730 void
3731 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3732 {
3733         struct lnet_peer_table *ptable;
3734         struct lnet_peer *lp;
3735         struct lnet_peer_net *lpn;
3736         struct lnet_peer_ni *lpni;
3737         int lncpt;
3738         int cpt;
3739
3740         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3741                 return;
3742
3743         if (!all) {
3744                 lnet_net_lock(LNET_LOCK_EX);
3745                 lpni = lnet_find_peer_ni_locked(nid);
3746                 if (!lpni) {
3747                         lnet_net_unlock(LNET_LOCK_EX);
3748                         return;
3749                 }
3750                 atomic_set(&lpni->lpni_healthv, value);
3751                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3752                 lnet_peer_ni_decref_locked(lpni);
3753                 lnet_net_unlock(LNET_LOCK_EX);
3754                 return;
3755         }
3756
3757         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3758
3759         /*
3760          * Walk all the peers and reset the healhv for each one to the
3761          * maximum value.
3762          */
3763         lnet_net_lock(LNET_LOCK_EX);
3764         for (cpt = 0; cpt < lncpt; cpt++) {
3765                 ptable = the_lnet.ln_peer_tables[cpt];
3766                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3767                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3768                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3769                                                     lpni_peer_nis) {
3770                                         atomic_set(&lpni->lpni_healthv, value);
3771                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3772                                 }
3773                         }
3774                 }
3775         }
3776         lnet_net_unlock(LNET_LOCK_EX);
3777 }
3778