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 __must_hold(&lp->lp_lock)
1156 {
1157         if (lnet_peer_discovery_disabled)
1158                 return true;
1159
1160         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1161             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1162                 return true;
1163         }
1164
1165         return false;
1166 }
1167
1168 /*
1169  * Peer Discovery
1170  */
1171 bool
1172 lnet_is_discovery_disabled(struct lnet_peer *lp)
1173 {
1174         bool rc = false;
1175
1176         spin_lock(&lp->lp_lock);
1177         rc = lnet_is_discovery_disabled_locked(lp);
1178         spin_unlock(&lp->lp_lock);
1179
1180         return rc;
1181 }
1182
1183 lnet_nid_t
1184 LNetPrimaryNID(lnet_nid_t nid)
1185 {
1186         struct lnet_peer *lp;
1187         struct lnet_peer_ni *lpni;
1188         lnet_nid_t primary_nid = nid;
1189         int rc = 0;
1190         int cpt;
1191
1192         cpt = lnet_net_lock_current();
1193         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1194         if (IS_ERR(lpni)) {
1195                 rc = PTR_ERR(lpni);
1196                 goto out_unlock;
1197         }
1198         lp = lpni->lpni_peer_net->lpn_peer;
1199
1200         while (!lnet_peer_is_uptodate(lp)) {
1201                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1202                 if (rc)
1203                         goto out_decref;
1204                 lp = lpni->lpni_peer_net->lpn_peer;
1205
1206                 /* Only try once if discovery is disabled */
1207                 if (lnet_is_discovery_disabled(lp))
1208                         break;
1209         }
1210         primary_nid = lp->lp_primary_nid;
1211 out_decref:
1212         lnet_peer_ni_decref_locked(lpni);
1213 out_unlock:
1214         lnet_net_unlock(cpt);
1215
1216         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1217                libcfs_nid2str(primary_nid), rc);
1218         return primary_nid;
1219 }
1220 EXPORT_SYMBOL(LNetPrimaryNID);
1221
1222 struct lnet_peer_net *
1223 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1224 {
1225         struct lnet_peer_net *peer_net;
1226         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1227                 if (peer_net->lpn_net_id == net_id)
1228                         return peer_net;
1229         }
1230         return NULL;
1231 }
1232
1233 /*
1234  * Attach a peer_ni to a peer_net and peer. This function assumes
1235  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1236  * may be attached to a different peer, in which case it will be
1237  * properly detached first. The whole operation is done atomically.
1238  *
1239  * Always returns 0.  This is the last function called from functions
1240  * that do return an int, so returning 0 here allows the compiler to
1241  * do a tail call.
1242  */
1243 static int
1244 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1245                                 struct lnet_peer_net *lpn,
1246                                 struct lnet_peer_ni *lpni,
1247                                 unsigned flags)
1248 {
1249         struct lnet_peer_table *ptable;
1250
1251         /* Install the new peer_ni */
1252         lnet_net_lock(LNET_LOCK_EX);
1253         /* Add peer_ni to global peer table hash, if necessary. */
1254         if (list_empty(&lpni->lpni_hashlist)) {
1255                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1256
1257                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1258                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1259                 ptable->pt_version++;
1260                 ptable->pt_number++;
1261                 /* This is the 1st refcount on lpni. */
1262                 atomic_inc(&lpni->lpni_refcount);
1263         }
1264
1265         /* Detach the peer_ni from an existing peer, if necessary. */
1266         if (lpni->lpni_peer_net) {
1267                 LASSERT(lpni->lpni_peer_net != lpn);
1268                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1269                 lnet_peer_detach_peer_ni_locked(lpni);
1270                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1271                 lpni->lpni_peer_net = NULL;
1272         }
1273
1274         /* Add peer_ni to peer_net */
1275         lpni->lpni_peer_net = lpn;
1276         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1277         lnet_peer_net_addref_locked(lpn);
1278
1279         /* Add peer_net to peer */
1280         if (!lpn->lpn_peer) {
1281                 lpn->lpn_peer = lp;
1282                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1283                 lnet_peer_addref_locked(lp);
1284         }
1285
1286         /* Add peer to global peer list, if necessary */
1287         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1288         if (list_empty(&lp->lp_peer_list)) {
1289                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1290                 ptable->pt_peers++;
1291         }
1292
1293
1294         /* Update peer state */
1295         spin_lock(&lp->lp_lock);
1296         if (flags & LNET_PEER_CONFIGURED) {
1297                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1298                         lp->lp_state |= LNET_PEER_CONFIGURED;
1299         }
1300         if (flags & LNET_PEER_MULTI_RAIL) {
1301                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1302                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1303                         lnet_peer_clr_non_mr_pref_nids(lp);
1304                 }
1305         }
1306         spin_unlock(&lp->lp_lock);
1307
1308         lp->lp_nnis++;
1309         lnet_net_unlock(LNET_LOCK_EX);
1310
1311         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1312                libcfs_nid2str(lp->lp_primary_nid),
1313                libcfs_nid2str(lpni->lpni_nid), flags);
1314
1315         return 0;
1316 }
1317
1318 /*
1319  * Create a new peer, with nid as its primary nid.
1320  *
1321  * Call with the lnet_api_mutex held.
1322  */
1323 static int
1324 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1325 {
1326         struct lnet_peer *lp;
1327         struct lnet_peer_net *lpn;
1328         struct lnet_peer_ni *lpni;
1329         int rc = 0;
1330
1331         LASSERT(nid != LNET_NID_ANY);
1332
1333         /*
1334          * No need for the lnet_net_lock here, because the
1335          * lnet_api_mutex is held.
1336          */
1337         lpni = lnet_find_peer_ni_locked(nid);
1338         if (lpni) {
1339                 /* A peer with this NID already exists. */
1340                 lp = lpni->lpni_peer_net->lpn_peer;
1341                 lnet_peer_ni_decref_locked(lpni);
1342                 /*
1343                  * This is an error if the peer was configured and the
1344                  * primary NID differs or an attempt is made to change
1345                  * the Multi-Rail flag. Otherwise the assumption is
1346                  * that an existing peer is being modified.
1347                  */
1348                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1349                         if (lp->lp_primary_nid != nid)
1350                                 rc = -EEXIST;
1351                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1352                                 rc = -EPERM;
1353                         goto out;
1354                 }
1355                 /* Delete and recreate as a configured peer. */
1356                 lnet_peer_del(lp);
1357         }
1358
1359         /* Create peer, peer_net, and peer_ni. */
1360         rc = -ENOMEM;
1361         lp = lnet_peer_alloc(nid);
1362         if (!lp)
1363                 goto out;
1364         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1365         if (!lpn)
1366                 goto out_free_lp;
1367         lpni = lnet_peer_ni_alloc(nid);
1368         if (!lpni)
1369                 goto out_free_lpn;
1370
1371         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1372
1373 out_free_lpn:
1374         LIBCFS_FREE(lpn, sizeof(*lpn));
1375 out_free_lp:
1376         LIBCFS_FREE(lp, sizeof(*lp));
1377 out:
1378         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1379                libcfs_nid2str(nid), flags, rc);
1380         return rc;
1381 }
1382
1383 /*
1384  * Add a NID to a peer. Call with ln_api_mutex held.
1385  *
1386  * Error codes:
1387  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1388  *  -EEXIST:   The NID was configured by DLC for a different peer.
1389  *  -ENOMEM:   Out of memory.
1390  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1391  *             non-multi-rail peer.
1392  */
1393 static int
1394 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1395 {
1396         struct lnet_peer_net *lpn;
1397         struct lnet_peer_ni *lpni;
1398         int rc = 0;
1399
1400         LASSERT(lp);
1401         LASSERT(nid != LNET_NID_ANY);
1402
1403         /* A configured peer can only be updated through configuration. */
1404         if (!(flags & LNET_PEER_CONFIGURED)) {
1405                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1406                         rc = -EPERM;
1407                         goto out;
1408                 }
1409         }
1410
1411         /*
1412          * The MULTI_RAIL flag can be set but not cleared, because
1413          * that would leave the peer struct in an invalid state.
1414          */
1415         if (flags & LNET_PEER_MULTI_RAIL) {
1416                 spin_lock(&lp->lp_lock);
1417                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1418                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1419                         lnet_peer_clr_non_mr_pref_nids(lp);
1420                 }
1421                 spin_unlock(&lp->lp_lock);
1422         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1423                 rc = -EPERM;
1424                 goto out;
1425         }
1426
1427         lpni = lnet_find_peer_ni_locked(nid);
1428         if (lpni) {
1429                 /*
1430                  * A peer_ni already exists. This is only a problem if
1431                  * it is not connected to this peer and was configured
1432                  * by DLC.
1433                  */
1434                 lnet_peer_ni_decref_locked(lpni);
1435                 if (lpni->lpni_peer_net->lpn_peer == lp)
1436                         goto out;
1437                 if (lnet_peer_ni_is_configured(lpni)) {
1438                         rc = -EEXIST;
1439                         goto out;
1440                 }
1441                 /* If this is the primary NID, destroy the peer. */
1442                 if (lnet_peer_ni_is_primary(lpni)) {
1443                         struct lnet_peer *rtr_lp =
1444                           lpni->lpni_peer_net->lpn_peer;
1445                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1446                         /*
1447                          * if we're trying to delete a router it means
1448                          * we're moving this peer NI to a new peer so must
1449                          * transfer router properties to the new peer
1450                          */
1451                         if (rtr_refcount > 0) {
1452                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1453                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1454                         }
1455                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1456                         lpni = lnet_peer_ni_alloc(nid);
1457                         if (!lpni) {
1458                                 rc = -ENOMEM;
1459                                 goto out;
1460                         }
1461                 }
1462         } else {
1463                 lpni = lnet_peer_ni_alloc(nid);
1464                 if (!lpni) {
1465                         rc = -ENOMEM;
1466                         goto out;
1467                 }
1468         }
1469
1470         /*
1471          * Get the peer_net. Check that we're not adding a second
1472          * peer_ni on a peer_net of a non-multi-rail peer.
1473          */
1474         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1475         if (!lpn) {
1476                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1477                 if (!lpn) {
1478                         rc = -ENOMEM;
1479                         goto out_free_lpni;
1480                 }
1481         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1482                 rc = -ENOTUNIQ;
1483                 goto out_free_lpni;
1484         }
1485
1486         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1487
1488 out_free_lpni:
1489         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1490         if (!lpni->lpni_peer_net)
1491                 LIBCFS_FREE(lpni, sizeof(*lpni));
1492 out:
1493         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1494                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1495                flags, rc);
1496         return rc;
1497 }
1498
1499 /*
1500  * Update the primary NID of a peer, if possible.
1501  *
1502  * Call with the lnet_api_mutex held.
1503  */
1504 static int
1505 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1506 {
1507         lnet_nid_t old = lp->lp_primary_nid;
1508         int rc = 0;
1509
1510         if (lp->lp_primary_nid == nid)
1511                 goto out;
1512         rc = lnet_peer_add_nid(lp, nid, flags);
1513         if (rc)
1514                 goto out;
1515         lp->lp_primary_nid = nid;
1516 out:
1517         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1518                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1519         return rc;
1520 }
1521
1522 /*
1523  * lpni creation initiated due to traffic either sending or receiving.
1524  */
1525 static int
1526 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1527 {
1528         struct lnet_peer *lp;
1529         struct lnet_peer_net *lpn;
1530         struct lnet_peer_ni *lpni;
1531         /*
1532          * Assume peer is Multi-Rail capable and let discovery find out
1533          * otherwise.
1534          */
1535         unsigned flags = LNET_PEER_MULTI_RAIL;
1536         int rc = 0;
1537
1538         if (nid == LNET_NID_ANY) {
1539                 rc = -EINVAL;
1540                 goto out;
1541         }
1542
1543         /* lnet_net_lock is not needed here because ln_api_lock is held */
1544         lpni = lnet_find_peer_ni_locked(nid);
1545         if (lpni) {
1546                 /*
1547                  * We must have raced with another thread. Since we
1548                  * know next to nothing about a peer_ni created by
1549                  * traffic, we just assume everything is ok and
1550                  * return.
1551                  */
1552                 lnet_peer_ni_decref_locked(lpni);
1553                 goto out;
1554         }
1555
1556         /* Create peer, peer_net, and peer_ni. */
1557         rc = -ENOMEM;
1558         lp = lnet_peer_alloc(nid);
1559         if (!lp)
1560                 goto out;
1561         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1562         if (!lpn)
1563                 goto out_free_lp;
1564         lpni = lnet_peer_ni_alloc(nid);
1565         if (!lpni)
1566                 goto out_free_lpn;
1567         if (pref != LNET_NID_ANY)
1568                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1569
1570         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1571
1572 out_free_lpn:
1573         LIBCFS_FREE(lpn, sizeof(*lpn));
1574 out_free_lp:
1575         LIBCFS_FREE(lp, sizeof(*lp));
1576 out:
1577         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1578         return rc;
1579 }
1580
1581 /*
1582  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1583  *
1584  * This API handles the following combinations:
1585  *   Create a peer with its primary NI if only the prim_nid is provided
1586  *   Add a NID to a peer identified by the prim_nid. The peer identified
1587  *   by the prim_nid must already exist.
1588  *   The peer being created may be non-MR.
1589  *
1590  * The caller must hold ln_api_mutex. This prevents the peer from
1591  * being created/modified/deleted by a different thread.
1592  */
1593 int
1594 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1595 {
1596         struct lnet_peer *lp = NULL;
1597         struct lnet_peer_ni *lpni;
1598         unsigned flags;
1599
1600         /* The prim_nid must always be specified */
1601         if (prim_nid == LNET_NID_ANY)
1602                 return -EINVAL;
1603
1604         flags = LNET_PEER_CONFIGURED;
1605         if (mr)
1606                 flags |= LNET_PEER_MULTI_RAIL;
1607
1608         /*
1609          * If nid isn't specified, we must create a new peer with
1610          * prim_nid as its primary nid.
1611          */
1612         if (nid == LNET_NID_ANY)
1613                 return lnet_peer_add(prim_nid, flags);
1614
1615         /* Look up the prim_nid, which must exist. */
1616         lpni = lnet_find_peer_ni_locked(prim_nid);
1617         if (!lpni)
1618                 return -ENOENT;
1619         lnet_peer_ni_decref_locked(lpni);
1620         lp = lpni->lpni_peer_net->lpn_peer;
1621
1622         /* Peer must have been configured. */
1623         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1624                 CDEBUG(D_NET, "peer %s was not configured\n",
1625                        libcfs_nid2str(prim_nid));
1626                 return -ENOENT;
1627         }
1628
1629         /* Primary NID must match */
1630         if (lp->lp_primary_nid != prim_nid) {
1631                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1632                        libcfs_nid2str(prim_nid),
1633                        libcfs_nid2str(lp->lp_primary_nid));
1634                 return -ENODEV;
1635         }
1636
1637         /* Multi-Rail flag must match. */
1638         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1639                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1640                        libcfs_nid2str(prim_nid));
1641                 return -EPERM;
1642         }
1643
1644         return lnet_peer_add_nid(lp, nid, flags);
1645 }
1646
1647 /*
1648  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1649  *
1650  * This API handles the following combinations:
1651  *   Delete a NI from a peer if both prim_nid and nid are provided.
1652  *   Delete a peer if only prim_nid is provided.
1653  *   Delete a peer if its primary nid is provided.
1654  *
1655  * The caller must hold ln_api_mutex. This prevents the peer from
1656  * being modified/deleted by a different thread.
1657  */
1658 int
1659 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1660 {
1661         struct lnet_peer *lp;
1662         struct lnet_peer_ni *lpni;
1663         unsigned flags;
1664
1665         if (prim_nid == LNET_NID_ANY)
1666                 return -EINVAL;
1667
1668         lpni = lnet_find_peer_ni_locked(prim_nid);
1669         if (!lpni)
1670                 return -ENOENT;
1671         lnet_peer_ni_decref_locked(lpni);
1672         lp = lpni->lpni_peer_net->lpn_peer;
1673
1674         if (prim_nid != lp->lp_primary_nid) {
1675                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1676                        libcfs_nid2str(prim_nid),
1677                        libcfs_nid2str(lp->lp_primary_nid));
1678                 return -ENODEV;
1679         }
1680
1681         lnet_net_lock(LNET_LOCK_EX);
1682         if (lp->lp_rtr_refcount > 0) {
1683                 lnet_net_unlock(LNET_LOCK_EX);
1684                 CERROR("%s is a router. Can not be deleted\n",
1685                        libcfs_nid2str(prim_nid));
1686                 return -EBUSY;
1687         }
1688         lnet_net_unlock(LNET_LOCK_EX);
1689
1690         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1691                 return lnet_peer_del(lp);
1692
1693         flags = LNET_PEER_CONFIGURED;
1694         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1695                 flags |= LNET_PEER_MULTI_RAIL;
1696
1697         return lnet_peer_del_nid(lp, nid, flags);
1698 }
1699
1700 void
1701 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1702 {
1703         struct lnet_peer_table *ptable;
1704         struct lnet_peer_net *lpn;
1705
1706         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1707
1708         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1709         LASSERT(list_empty(&lpni->lpni_txq));
1710         LASSERT(lpni->lpni_txqnob == 0);
1711         LASSERT(list_empty(&lpni->lpni_peer_nis));
1712         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1713
1714         lpn = lpni->lpni_peer_net;
1715         lpni->lpni_peer_net = NULL;
1716         lpni->lpni_net = NULL;
1717
1718         /* remove the peer ni from the zombie list */
1719         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1720         spin_lock(&ptable->pt_zombie_lock);
1721         list_del_init(&lpni->lpni_hashlist);
1722         ptable->pt_zombies--;
1723         spin_unlock(&ptable->pt_zombie_lock);
1724
1725         if (lpni->lpni_pref_nnids > 1) {
1726                 LIBCFS_FREE(lpni->lpni_pref.nids,
1727                         sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1728         }
1729         LIBCFS_FREE(lpni, sizeof(*lpni));
1730
1731         lnet_peer_net_decref_locked(lpn);
1732 }
1733
1734 struct lnet_peer_ni *
1735 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1736 {
1737         struct lnet_peer_ni *lpni = NULL;
1738         int rc;
1739
1740         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1741                 return ERR_PTR(-ESHUTDOWN);
1742
1743         /*
1744          * find if a peer_ni already exists.
1745          * If so then just return that.
1746          */
1747         lpni = lnet_find_peer_ni_locked(nid);
1748         if (lpni)
1749                 return lpni;
1750
1751         lnet_net_unlock(cpt);
1752
1753         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1754         if (rc) {
1755                 lpni = ERR_PTR(rc);
1756                 goto out_net_relock;
1757         }
1758
1759         lpni = lnet_find_peer_ni_locked(nid);
1760         LASSERT(lpni);
1761
1762 out_net_relock:
1763         lnet_net_lock(cpt);
1764
1765         return lpni;
1766 }
1767
1768 /*
1769  * Get a peer_ni for the given nid, create it if necessary. Takes a
1770  * hold on the peer_ni.
1771  */
1772 struct lnet_peer_ni *
1773 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1774 {
1775         struct lnet_peer_ni *lpni = NULL;
1776         int rc;
1777
1778         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1779                 return ERR_PTR(-ESHUTDOWN);
1780
1781         /*
1782          * find if a peer_ni already exists.
1783          * If so then just return that.
1784          */
1785         lpni = lnet_find_peer_ni_locked(nid);
1786         if (lpni)
1787                 return lpni;
1788
1789         /*
1790          * Slow path:
1791          * use the lnet_api_mutex to serialize the creation of the peer_ni
1792          * and the creation/deletion of the local ni/net. When a local ni is
1793          * created, if there exists a set of peer_nis on that network,
1794          * they need to be traversed and updated. When a local NI is
1795          * deleted, which could result in a network being deleted, then
1796          * all peer nis on that network need to be removed as well.
1797          *
1798          * Creation through traffic should also be serialized with
1799          * creation through DLC.
1800          */
1801         lnet_net_unlock(cpt);
1802         mutex_lock(&the_lnet.ln_api_mutex);
1803         /*
1804          * Shutdown is only set under the ln_api_lock, so a single
1805          * check here is sufficent.
1806          */
1807         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1808                 lpni = ERR_PTR(-ESHUTDOWN);
1809                 goto out_mutex_unlock;
1810         }
1811
1812         rc = lnet_peer_ni_traffic_add(nid, pref);
1813         if (rc) {
1814                 lpni = ERR_PTR(rc);
1815                 goto out_mutex_unlock;
1816         }
1817
1818         lpni = lnet_find_peer_ni_locked(nid);
1819         LASSERT(lpni);
1820
1821 out_mutex_unlock:
1822         mutex_unlock(&the_lnet.ln_api_mutex);
1823         lnet_net_lock(cpt);
1824
1825         /* Lock has been dropped, check again for shutdown. */
1826         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1827                 if (!IS_ERR(lpni))
1828                         lnet_peer_ni_decref_locked(lpni);
1829                 lpni = ERR_PTR(-ESHUTDOWN);
1830         }
1831
1832         return lpni;
1833 }
1834
1835 bool
1836 lnet_peer_gw_discovery(struct lnet_peer *lp)
1837 {
1838         bool rc = false;
1839
1840         spin_lock(&lp->lp_lock);
1841         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1842                 rc = true;
1843         spin_unlock(&lp->lp_lock);
1844
1845         return rc;
1846 }
1847
1848 bool
1849 lnet_peer_is_uptodate(struct lnet_peer *lp)
1850 {
1851         bool rc;
1852
1853         spin_lock(&lp->lp_lock);
1854         rc = lnet_peer_is_uptodate_locked(lp);
1855         spin_unlock(&lp->lp_lock);
1856         return rc;
1857 }
1858
1859 /*
1860  * Is a peer uptodate from the point of view of discovery?
1861  *
1862  * If it is currently being processed, obviously not.
1863  * A forced Ping or Push is also handled by the discovery thread.
1864  *
1865  * Otherwise look at whether the peer needs rediscovering.
1866  */
1867 bool
1868 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1869 __must_hold(&lp->lp_lock)
1870 {
1871         bool rc;
1872
1873         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1874                             LNET_PEER_FORCE_PING |
1875                             LNET_PEER_FORCE_PUSH)) {
1876                 rc = false;
1877         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1878                 rc = false;
1879         } else if (lnet_peer_needs_push(lp)) {
1880                 rc = false;
1881         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1882                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1883                         rc = true;
1884                 else
1885                         rc = false;
1886         } else {
1887                 rc = false;
1888         }
1889
1890         return rc;
1891 }
1892
1893 /*
1894  * Queue a peer for the attention of the discovery thread.  Call with
1895  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1896  * -EALREADY if the peer was already queued.
1897  */
1898 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1899 {
1900         int rc;
1901
1902         spin_lock(&lp->lp_lock);
1903         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1904                 lp->lp_state |= LNET_PEER_DISCOVERING;
1905         spin_unlock(&lp->lp_lock);
1906         if (list_empty(&lp->lp_dc_list)) {
1907                 lnet_peer_addref_locked(lp);
1908                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1909                 wake_up(&the_lnet.ln_dc_waitq);
1910                 rc = 0;
1911         } else {
1912                 rc = -EALREADY;
1913         }
1914
1915         CDEBUG(D_NET, "Queue peer %s: %d\n",
1916                libcfs_nid2str(lp->lp_primary_nid), rc);
1917
1918         return rc;
1919 }
1920
1921 /*
1922  * Discovery of a peer is complete. Wake all waiters on the peer.
1923  * Call with lnet_net_lock/EX held.
1924  */
1925 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1926 {
1927         struct lnet_msg *msg, *tmp;
1928         int rc = 0;
1929         struct list_head pending_msgs;
1930
1931         INIT_LIST_HEAD(&pending_msgs);
1932
1933         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1934                libcfs_nid2str(lp->lp_primary_nid));
1935
1936         list_del_init(&lp->lp_dc_list);
1937         spin_lock(&lp->lp_lock);
1938         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1939         spin_unlock(&lp->lp_lock);
1940         wake_up_all(&lp->lp_dc_waitq);
1941
1942         if (lp->lp_rtr_refcount > 0)
1943                 lnet_router_discovery_complete(lp);
1944
1945         lnet_net_unlock(LNET_LOCK_EX);
1946
1947         /* iterate through all pending messages and send them again */
1948         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1949                 list_del_init(&msg->msg_list);
1950                 if (lp->lp_dc_error) {
1951                         lnet_finalize(msg, lp->lp_dc_error);
1952                         continue;
1953                 }
1954
1955                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1956                        lnet_msgtyp2str(msg->msg_type),
1957                        libcfs_id2str(msg->msg_target));
1958                 rc = lnet_send(msg->msg_src_nid_param, msg,
1959                                msg->msg_rtr_nid_param);
1960                 if (rc < 0) {
1961                         CNETERR("Error sending %s to %s: %d\n",
1962                                lnet_msgtyp2str(msg->msg_type),
1963                                libcfs_id2str(msg->msg_target), rc);
1964                         lnet_finalize(msg, rc);
1965                 }
1966         }
1967         lnet_net_lock(LNET_LOCK_EX);
1968         lnet_peer_decref_locked(lp);
1969 }
1970
1971 /*
1972  * Handle inbound push.
1973  * Like any event handler, called with lnet_res_lock/CPT held.
1974  */
1975 void lnet_peer_push_event(struct lnet_event *ev)
1976 {
1977         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1978         struct lnet_peer *lp;
1979
1980         /* lnet_find_peer() adds a refcount */
1981         lp = lnet_find_peer(ev->source.nid);
1982         if (!lp) {
1983                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1984                        libcfs_nid2str(ev->initiator.nid),
1985                        libcfs_nid2str(ev->source.nid));
1986                 return;
1987         }
1988
1989         /* Ensure peer state remains consistent while we modify it. */
1990         spin_lock(&lp->lp_lock);
1991
1992         /*
1993          * If some kind of error happened the contents of the message
1994          * cannot be used. Clear the NIDS_UPTODATE and set the
1995          * FORCE_PING flag to trigger a ping.
1996          */
1997         if (ev->status) {
1998                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1999                 lp->lp_state |= LNET_PEER_FORCE_PING;
2000                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2001                        ev->status,
2002                        libcfs_nid2str(lp->lp_primary_nid),
2003                        libcfs_nid2str(ev->source.nid));
2004                 goto out;
2005         }
2006
2007         /*
2008          * A push with invalid or corrupted info. Clear the UPTODATE
2009          * flag to trigger a ping.
2010          */
2011         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2012                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2013                 lp->lp_state |= LNET_PEER_FORCE_PING;
2014                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2015                        libcfs_nid2str(lp->lp_primary_nid));
2016                 goto out;
2017         }
2018
2019         /*
2020          * Make sure we'll allocate the correct size ping buffer when
2021          * pinging the peer.
2022          */
2023         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2024                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2025
2026         /*
2027          * A non-Multi-Rail peer is not supposed to be capable of
2028          * sending a push.
2029          */
2030         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2031                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2032                        libcfs_nid2str(lp->lp_primary_nid));
2033                 goto out;
2034         }
2035
2036         /*
2037          * Check the MULTIRAIL flag. Complain if the peer was DLC
2038          * configured without it.
2039          */
2040         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2041                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2042                         CERROR("Push says %s is Multi-Rail, DLC says not\n",
2043                                libcfs_nid2str(lp->lp_primary_nid));
2044                 } else {
2045                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2046                         lnet_peer_clr_non_mr_pref_nids(lp);
2047                 }
2048         }
2049
2050         /*
2051          * The peer may have discovery disabled at its end. Set
2052          * NO_DISCOVERY as appropriate.
2053          */
2054         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2055                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2056                        libcfs_nid2str(lp->lp_primary_nid));
2057                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2058         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2059                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2060                        libcfs_nid2str(lp->lp_primary_nid));
2061                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2062         }
2063
2064         /*
2065          * Check for truncation of the Put message. Clear the
2066          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2067          * and tell discovery to allocate a bigger buffer.
2068          */
2069         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2070                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2071                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2072                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2073                 lp->lp_state |= LNET_PEER_FORCE_PING;
2074                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2075                        libcfs_nid2str(lp->lp_primary_nid),
2076                        pbuf->pb_info.pi_nnis);
2077                 goto out;
2078         }
2079
2080         /* always assume new data */
2081         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2082         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2083
2084         /*
2085          * If there is data present that hasn't been processed yet,
2086          * we'll replace it if the Put contained newer data and it
2087          * fits. We're racing with a Ping or earlier Push in this
2088          * case.
2089          */
2090         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2091                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2092                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2093                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2094                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2095                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2096                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2097                               libcfs_nid2str(lp->lp_primary_nid),
2098                               LNET_PING_BUFFER_SEQNO(pbuf),
2099                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2100                 }
2101                 goto out;
2102         }
2103
2104         /*
2105          * Allocate a buffer to copy the data. On a failure we drop
2106          * the Push and set FORCE_PING to force the discovery
2107          * thread to fix the problem by pinging the peer.
2108          */
2109         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2110         if (!lp->lp_data) {
2111                 lp->lp_state |= LNET_PEER_FORCE_PING;
2112                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2113                        libcfs_nid2str(lp->lp_primary_nid),
2114                        LNET_PING_BUFFER_SEQNO(pbuf));
2115                 goto out;
2116         }
2117
2118         /* Success */
2119         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2120                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2121         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2122         CDEBUG(D_NET, "Received Push %s %u\n",
2123                libcfs_nid2str(lp->lp_primary_nid),
2124                LNET_PING_BUFFER_SEQNO(pbuf));
2125
2126 out:
2127         /*
2128          * Queue the peer for discovery if not done, force it on the request
2129          * queue and wake the discovery thread if the peer was already queued,
2130          * because its status changed.
2131          */
2132         spin_unlock(&lp->lp_lock);
2133         lnet_net_lock(LNET_LOCK_EX);
2134         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2135                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2136                 wake_up(&the_lnet.ln_dc_waitq);
2137         }
2138         /* Drop refcount from lookup */
2139         lnet_peer_decref_locked(lp);
2140         lnet_net_unlock(LNET_LOCK_EX);
2141 }
2142
2143 /*
2144  * Clear the discovery error state, unless we're already discovering
2145  * this peer, in which case the error is current.
2146  */
2147 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2148 {
2149         spin_lock(&lp->lp_lock);
2150         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2151                 lp->lp_dc_error = 0;
2152         spin_unlock(&lp->lp_lock);
2153 }
2154
2155 /*
2156  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2157  * dropped/retaken within this function. An lnet_peer_ni is passed in
2158  * because discovery could tear down an lnet_peer.
2159  */
2160 int
2161 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2162 {
2163         DEFINE_WAIT(wait);
2164         struct lnet_peer *lp;
2165         int rc = 0;
2166         int count = 0;
2167
2168 again:
2169         lnet_net_unlock(cpt);
2170         lnet_net_lock(LNET_LOCK_EX);
2171         lp = lpni->lpni_peer_net->lpn_peer;
2172         lnet_peer_clear_discovery_error(lp);
2173
2174         /*
2175          * We're willing to be interrupted. The lpni can become a
2176          * zombie if we race with DLC, so we must check for that.
2177          */
2178         for (;;) {
2179                 /* Keep lp alive when the lnet_net_lock is unlocked */
2180                 lnet_peer_addref_locked(lp);
2181                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2182                 if (signal_pending(current))
2183                         break;
2184                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2185                         break;
2186                 /*
2187                  * Don't repeat discovery if discovery is disabled. This is
2188                  * done to ensure we can use discovery as a standard ping as
2189                  * well for backwards compatibility with routers which do not
2190                  * have discovery or have discovery disabled
2191                  */
2192                 if (lnet_is_discovery_disabled(lp) && count > 0)
2193                         break;
2194                 if (lp->lp_dc_error)
2195                         break;
2196                 if (lnet_peer_is_uptodate(lp))
2197                         break;
2198                 lnet_peer_queue_for_discovery(lp);
2199                 count++;
2200                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2201
2202                 /*
2203                  * If caller requested a non-blocking operation then
2204                  * return immediately. Once discovery is complete any
2205                  * pending messages that were stopped due to discovery
2206                  * will be transmitted.
2207                  */
2208                 if (!block)
2209                         break;
2210
2211                 lnet_net_unlock(LNET_LOCK_EX);
2212                 schedule();
2213                 finish_wait(&lp->lp_dc_waitq, &wait);
2214                 lnet_net_lock(LNET_LOCK_EX);
2215                 lnet_peer_decref_locked(lp);
2216                 /* Peer may have changed */
2217                 lp = lpni->lpni_peer_net->lpn_peer;
2218         }
2219         finish_wait(&lp->lp_dc_waitq, &wait);
2220
2221         lnet_net_unlock(LNET_LOCK_EX);
2222         lnet_net_lock(cpt);
2223         lnet_peer_decref_locked(lp);
2224         /*
2225          * The peer may have changed, so re-check and rediscover if that turns
2226          * out to have been the case. The reference count on lp ensured that
2227          * even if it was unlinked from lpni the memory could not be recycled.
2228          * Thus the check below is sufficient to determine whether the peer
2229          * changed. If the peer changed, then lp must not be dereferenced.
2230          */
2231         if (lp != lpni->lpni_peer_net->lpn_peer)
2232                 goto again;
2233
2234         if (signal_pending(current))
2235                 rc = -EINTR;
2236         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2237                 rc = -ESHUTDOWN;
2238         else if (lp->lp_dc_error)
2239                 rc = lp->lp_dc_error;
2240         else if (!block)
2241                 CDEBUG(D_NET, "non-blocking discovery\n");
2242         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2243                 goto again;
2244
2245         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2246                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2247                libcfs_nid2str(lpni->lpni_nid), rc,
2248                (!block) ? "pending discovery" : "discovery complete");
2249
2250         return rc;
2251 }
2252
2253 /* Handle an incoming ack for a push. */
2254 static void
2255 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2256 {
2257         struct lnet_ping_buffer *pbuf;
2258
2259         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2260         spin_lock(&lp->lp_lock);
2261         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2262         lp->lp_push_error = ev->status;
2263         if (ev->status)
2264                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2265         else
2266                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2267         spin_unlock(&lp->lp_lock);
2268
2269         CDEBUG(D_NET, "peer %s ev->status %d\n",
2270                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2271 }
2272
2273 /* Handle a Reply message. This is the reply to a Ping message. */
2274 static void
2275 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2276 {
2277         struct lnet_ping_buffer *pbuf;
2278         int rc;
2279
2280         spin_lock(&lp->lp_lock);
2281
2282         /*
2283          * If some kind of error happened the contents of message
2284          * cannot be used. Set PING_FAILED to trigger a retry.
2285          */
2286         if (ev->status) {
2287                 lp->lp_state |= LNET_PEER_PING_FAILED;
2288                 lp->lp_ping_error = ev->status;
2289                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2290                        ev->status,
2291                        libcfs_nid2str(lp->lp_primary_nid),
2292                        libcfs_nid2str(ev->source.nid));
2293                 goto out;
2294         }
2295
2296         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2297         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2298                 lnet_swap_pinginfo(pbuf);
2299
2300         /*
2301          * A reply with invalid or corrupted info. Set PING_FAILED to
2302          * trigger a retry.
2303          */
2304         rc = lnet_ping_info_validate(&pbuf->pb_info);
2305         if (rc) {
2306                 lp->lp_state |= LNET_PEER_PING_FAILED;
2307                 lp->lp_ping_error = 0;
2308                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2309                        libcfs_nid2str(lp->lp_primary_nid), rc);
2310                 goto out;
2311         }
2312
2313
2314         /*
2315          * Only enable the multi-rail feature on the peer if both sides of
2316          * the connection have discovery on
2317          */
2318         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2319                 CDEBUG(D_NET, "Peer %s has Multi-Rail feature enabled\n",
2320                        libcfs_nid2str(lp->lp_primary_nid));
2321                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2322         } else {
2323                 CDEBUG(D_NET, "Peer %s has Multi-Rail feature disabled\n",
2324                        libcfs_nid2str(lp->lp_primary_nid));
2325                 lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2326         }
2327
2328         /*
2329          * The peer may have discovery disabled at its end. Set
2330          * NO_DISCOVERY as appropriate.
2331          */
2332         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2333             !lnet_peer_discovery_disabled) {
2334                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2335                        libcfs_nid2str(lp->lp_primary_nid));
2336                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2337         } else {
2338                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2339                        libcfs_nid2str(lp->lp_primary_nid));
2340                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2341         }
2342
2343         /*
2344          * Update the MULTI_RAIL flag based on the reply. If the peer
2345          * was configured with DLC then the setting should match what
2346          * DLC put in.
2347          */
2348         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2349                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2350                         /* Everything's fine */
2351                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2352                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2353                               libcfs_nid2str(lp->lp_primary_nid));
2354                 } else {
2355                         /*
2356                          * if discovery is disabled then we don't want to
2357                          * update the state of the peer. All we'll do is
2358                          * update the peer_nis which were reported back in
2359                          * the initial ping
2360                          */
2361
2362                         if (!lnet_is_discovery_disabled_locked(lp)) {
2363                                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2364                                 lnet_peer_clr_non_mr_pref_nids(lp);
2365                         }
2366                 }
2367         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2368                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2369                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2370                               libcfs_nid2str(lp->lp_primary_nid));
2371                 } else {
2372                         CERROR("Multi-Rail state vanished from %s\n",
2373                                libcfs_nid2str(lp->lp_primary_nid));
2374                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2375                 }
2376         }
2377
2378         /*
2379          * Make sure we'll allocate the correct size ping buffer when
2380          * pinging the peer.
2381          */
2382         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2383                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2384
2385         /*
2386          * Check for truncation of the Reply. Clear PING_SENT and set
2387          * PING_FAILED to trigger a retry.
2388          */
2389         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2390                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2391                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2392                 lp->lp_state |= LNET_PEER_PING_FAILED;
2393                 lp->lp_ping_error = 0;
2394                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2395                        libcfs_nid2str(lp->lp_primary_nid),
2396                        pbuf->pb_info.pi_nnis);
2397                 goto out;
2398         }
2399
2400         /*
2401          * Check the sequence numbers in the reply. These are only
2402          * available if the reply came from a Multi-Rail peer.
2403          */
2404         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2405             pbuf->pb_info.pi_nnis > 1 &&
2406             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2407                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2408                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2409                                 libcfs_nid2str(lp->lp_primary_nid),
2410                                 LNET_PING_BUFFER_SEQNO(pbuf),
2411                                 lp->lp_peer_seqno);
2412
2413                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2414         }
2415
2416         /* We're happy with the state of the data in the buffer. */
2417         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2418                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2419         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2420                 lnet_ping_buffer_decref(lp->lp_data);
2421         else
2422                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2423         lnet_ping_buffer_addref(pbuf);
2424         lp->lp_data = pbuf;
2425 out:
2426         lp->lp_state &= ~LNET_PEER_PING_SENT;
2427         spin_unlock(&lp->lp_lock);
2428
2429         lnet_net_lock(LNET_LOCK_EX);
2430         /*
2431          * If this peer is a gateway, call the routing callback to
2432          * handle the ping reply
2433          */
2434         if (lp->lp_rtr_refcount > 0)
2435                 lnet_router_discovery_ping_reply(lp);
2436         lnet_net_unlock(LNET_LOCK_EX);
2437 }
2438
2439 /*
2440  * Send event handling. Only matters for error cases, where we clean
2441  * up state on the peer and peer_ni that would otherwise be updated in
2442  * the REPLY event handler for a successful Ping, and the ACK event
2443  * handler for a successful Push.
2444  */
2445 static int
2446 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2447 {
2448         int rc = 0;
2449
2450         if (!ev->status)
2451                 goto out;
2452
2453         spin_lock(&lp->lp_lock);
2454         if (ev->msg_type == LNET_MSG_GET) {
2455                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2456                 lp->lp_state |= LNET_PEER_PING_FAILED;
2457                 lp->lp_ping_error = ev->status;
2458         } else { /* ev->msg_type == LNET_MSG_PUT */
2459                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2460                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2461                 lp->lp_push_error = ev->status;
2462         }
2463         spin_unlock(&lp->lp_lock);
2464         rc = LNET_REDISCOVER_PEER;
2465 out:
2466         CDEBUG(D_NET, "%s Send to %s: %d\n",
2467                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2468                 libcfs_nid2str(ev->target.nid), rc);
2469         return rc;
2470 }
2471
2472 /*
2473  * Unlink event handling. This event is only seen if a call to
2474  * LNetMDUnlink() caused the event to be unlinked. If this call was
2475  * made after the event was set up in LNetGet() or LNetPut() then we
2476  * assume the Ping or Push timed out.
2477  */
2478 static void
2479 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2480 {
2481         spin_lock(&lp->lp_lock);
2482         /* We've passed through LNetGet() */
2483         if (lp->lp_state & LNET_PEER_PING_SENT) {
2484                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2485                 lp->lp_state |= LNET_PEER_PING_FAILED;
2486                 lp->lp_ping_error = -ETIMEDOUT;
2487                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2488                         libcfs_nid2str(lp->lp_primary_nid));
2489         }
2490         /* We've passed through LNetPut() */
2491         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2492                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2493                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2494                 lp->lp_push_error = -ETIMEDOUT;
2495                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2496                         libcfs_nid2str(lp->lp_primary_nid));
2497         }
2498         spin_unlock(&lp->lp_lock);
2499 }
2500
2501 /*
2502  * Event handler for the discovery EQ.
2503  *
2504  * Called with lnet_res_lock(cpt) held. The cpt is the
2505  * lnet_cpt_of_cookie() of the md handle cookie.
2506  */
2507 static void lnet_discovery_event_handler(struct lnet_event *event)
2508 {
2509         struct lnet_peer *lp = event->md.user_ptr;
2510         struct lnet_ping_buffer *pbuf;
2511         int rc;
2512
2513         /* discovery needs to take another look */
2514         rc = LNET_REDISCOVER_PEER;
2515
2516         CDEBUG(D_NET, "Received event: %d\n", event->type);
2517
2518         switch (event->type) {
2519         case LNET_EVENT_ACK:
2520                 lnet_discovery_event_ack(lp, event);
2521                 break;
2522         case LNET_EVENT_REPLY:
2523                 lnet_discovery_event_reply(lp, event);
2524                 break;
2525         case LNET_EVENT_SEND:
2526                 /* Only send failure triggers a retry. */
2527                 rc = lnet_discovery_event_send(lp, event);
2528                 break;
2529         case LNET_EVENT_UNLINK:
2530                 /* LNetMDUnlink() was called */
2531                 lnet_discovery_event_unlink(lp, event);
2532                 break;
2533         default:
2534                 /* Invalid events. */
2535                 LBUG();
2536         }
2537         lnet_net_lock(LNET_LOCK_EX);
2538         if (event->unlinked) {
2539                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2540                 lnet_ping_buffer_decref(pbuf);
2541                 lnet_peer_decref_locked(lp);
2542         }
2543
2544         /* put peer back at end of request queue, if discovery not already
2545          * done */
2546         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2547                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2548                 wake_up(&the_lnet.ln_dc_waitq);
2549         }
2550         lnet_net_unlock(LNET_LOCK_EX);
2551 }
2552
2553 /*
2554  * Build a peer from incoming data.
2555  *
2556  * The NIDs in the incoming data are supposed to be structured as follows:
2557  *  - loopback
2558  *  - primary NID
2559  *  - other NIDs in same net
2560  *  - NIDs in second net
2561  *  - NIDs in third net
2562  *  - ...
2563  * This due to the way the list of NIDs in the data is created.
2564  *
2565  * Note that this function will mark the peer uptodate unless an
2566  * ENOMEM is encontered. All other errors are due to a conflict
2567  * between the DLC configuration and what discovery sees. We treat DLC
2568  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2569  * peer from becoming stuck in discovery.
2570  */
2571 static int lnet_peer_merge_data(struct lnet_peer *lp,
2572                                 struct lnet_ping_buffer *pbuf)
2573 {
2574         struct lnet_peer_ni *lpni;
2575         lnet_nid_t *curnis = NULL;
2576         struct lnet_ni_status *addnis = NULL;
2577         lnet_nid_t *delnis = NULL;
2578         unsigned flags;
2579         int ncurnis;
2580         int naddnis;
2581         int ndelnis;
2582         int nnis = 0;
2583         int i;
2584         int j;
2585         int rc;
2586
2587         flags = LNET_PEER_DISCOVERED;
2588         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2589                 flags |= LNET_PEER_MULTI_RAIL;
2590
2591         /*
2592          * Cache the routing feature for the peer; whether it is enabled
2593          * for disabled as reported by the remote peer.
2594          */
2595         spin_lock(&lp->lp_lock);
2596         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2597                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2598         else
2599                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2600         spin_unlock(&lp->lp_lock);
2601
2602         nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2603         LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2604         LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2605         LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2606         if (!curnis || !addnis || !delnis) {
2607                 rc = -ENOMEM;
2608                 goto out;
2609         }
2610         ncurnis = 0;
2611         naddnis = 0;
2612         ndelnis = 0;
2613
2614         /* Construct the list of NIDs present in peer. */
2615         lpni = NULL;
2616         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2617                 curnis[ncurnis++] = lpni->lpni_nid;
2618
2619         /*
2620          * Check for NIDs in pbuf not present in curnis[].
2621          * The loop starts at 1 to skip the loopback NID.
2622          */
2623         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2624                 for (j = 0; j < ncurnis; j++)
2625                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2626                                 break;
2627                 if (j == ncurnis)
2628                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2629         }
2630         /*
2631          * Check for NIDs in curnis[] not present in pbuf.
2632          * The nested loop starts at 1 to skip the loopback NID.
2633          *
2634          * But never add the loopback NID to delnis[]: if it is
2635          * present in curnis[] then this peer is for this node.
2636          */
2637         for (i = 0; i < ncurnis; i++) {
2638                 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2639                         continue;
2640                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2641                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2642                                 /*
2643                                  * update the information we cache for the
2644                                  * peer with the latest information we
2645                                  * received
2646                                  */
2647                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2648                                 if (lpni) {
2649                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2650                                         lnet_peer_ni_decref_locked(lpni);
2651                                 }
2652                                 break;
2653                         }
2654                 }
2655                 if (j == pbuf->pb_info.pi_nnis)
2656                         delnis[ndelnis++] = curnis[i];
2657         }
2658
2659         /*
2660          * If we get here and the discovery is disabled then we don't want
2661          * to add or delete any NIs. We just updated the ones we have some
2662          * information on, and call it a day
2663          */
2664         rc = 0;
2665         if (lnet_is_discovery_disabled(lp))
2666                 goto out;
2667
2668         for (i = 0; i < naddnis; i++) {
2669                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2670                 if (rc) {
2671                         CERROR("Error adding NID %s to peer %s: %d\n",
2672                                libcfs_nid2str(addnis[i].ns_nid),
2673                                libcfs_nid2str(lp->lp_primary_nid), rc);
2674                         if (rc == -ENOMEM)
2675                                 goto out;
2676                 }
2677                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2678                 if (lpni) {
2679                         lpni->lpni_ns_status = addnis[i].ns_status;
2680                         lnet_peer_ni_decref_locked(lpni);
2681                 }
2682         }
2683
2684         for (i = 0; i < ndelnis; i++) {
2685                 /*
2686                  * for routers it's okay to delete the primary_nid because
2687                  * the upper layers don't really rely on it. So if we're
2688                  * being told that the router changed its primary_nid
2689                  * then it's okay to delete it.
2690                  */
2691                 if (lp->lp_rtr_refcount > 0)
2692                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2693                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2694                 if (rc) {
2695                         CERROR("Error deleting NID %s from peer %s: %d\n",
2696                                libcfs_nid2str(delnis[i]),
2697                                libcfs_nid2str(lp->lp_primary_nid), rc);
2698                         if (rc == -ENOMEM)
2699                                 goto out;
2700                 }
2701         }
2702         /*
2703          * Errors other than -ENOMEM are due to peers having been
2704          * configured with DLC. Ignore these because DLC overrides
2705          * Discovery.
2706          */
2707         rc = 0;
2708 out:
2709         LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2710         LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2711         LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2712         lnet_ping_buffer_decref(pbuf);
2713         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2714
2715         if (rc) {
2716                 spin_lock(&lp->lp_lock);
2717                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2718                 lp->lp_state |= LNET_PEER_FORCE_PING;
2719                 spin_unlock(&lp->lp_lock);
2720         }
2721         return rc;
2722 }
2723
2724 /*
2725  * The data in pbuf says lp is its primary peer, but the data was
2726  * received by a different peer. Try to update lp with the data.
2727  */
2728 static int
2729 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2730 {
2731         struct lnet_handle_md mdh;
2732
2733         /* Queue lp for discovery, and force it on the request queue. */
2734         lnet_net_lock(LNET_LOCK_EX);
2735         if (lnet_peer_queue_for_discovery(lp))
2736                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2737         lnet_net_unlock(LNET_LOCK_EX);
2738
2739         LNetInvalidateMDHandle(&mdh);
2740
2741         /*
2742          * Decide whether we can move the peer to the DATA_PRESENT state.
2743          *
2744          * We replace stale data for a multi-rail peer, repair PING_FAILED
2745          * status, and preempt FORCE_PING.
2746          *
2747          * If after that we have DATA_PRESENT, we merge it into this peer.
2748          */
2749         spin_lock(&lp->lp_lock);
2750         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2751                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2752                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2753                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2754                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2755                         lnet_ping_buffer_decref(pbuf);
2756                         pbuf = lp->lp_data;
2757                         lp->lp_data = NULL;
2758                 }
2759         }
2760         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2761                 lnet_ping_buffer_decref(lp->lp_data);
2762                 lp->lp_data = NULL;
2763                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2764         }
2765         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2766                 mdh = lp->lp_ping_mdh;
2767                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2768                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2769                 lp->lp_ping_error = 0;
2770         }
2771         if (lp->lp_state & LNET_PEER_FORCE_PING)
2772                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2773         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2774         spin_unlock(&lp->lp_lock);
2775
2776         if (!LNetMDHandleIsInvalid(mdh))
2777                 LNetMDUnlink(mdh);
2778
2779         if (pbuf)
2780                 return lnet_peer_merge_data(lp, pbuf);
2781
2782         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2783         return 0;
2784 }
2785
2786 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2787 {
2788         int i;
2789
2790         for (i = 0; i < pinfo->pi_nnis; i++) {
2791                 if (pinfo->pi_ni[i].ns_nid == nid)
2792                         return true;
2793         }
2794
2795         return false;
2796 }
2797
2798 /*
2799  * Update a peer using the data received.
2800  */
2801 static int lnet_peer_data_present(struct lnet_peer *lp)
2802 __must_hold(&lp->lp_lock)
2803 {
2804         struct lnet_ping_buffer *pbuf;
2805         struct lnet_peer_ni *lpni;
2806         lnet_nid_t nid = LNET_NID_ANY;
2807         unsigned flags;
2808         int rc = 0;
2809
2810         pbuf = lp->lp_data;
2811         lp->lp_data = NULL;
2812         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2813         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2814         spin_unlock(&lp->lp_lock);
2815
2816         /*
2817          * Modifications of peer structures are done while holding the
2818          * ln_api_mutex. A global lock is required because we may be
2819          * modifying multiple peer structures, and a mutex greatly
2820          * simplifies memory management.
2821          *
2822          * The actual changes to the data structures must also protect
2823          * against concurrent lookups, for which the lnet_net_lock in
2824          * LNET_LOCK_EX mode is used.
2825          */
2826         mutex_lock(&the_lnet.ln_api_mutex);
2827         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2828                 rc = -ESHUTDOWN;
2829                 goto out;
2830         }
2831
2832         /*
2833          * If this peer is not on the peer list then it is being torn
2834          * down, and our reference count may be all that is keeping it
2835          * alive. Don't do any work on it.
2836          */
2837         if (list_empty(&lp->lp_peer_list))
2838                 goto out;
2839
2840         flags = LNET_PEER_DISCOVERED;
2841         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2842                 flags |= LNET_PEER_MULTI_RAIL;
2843
2844         /*
2845          * Check whether the primary NID in the message matches the
2846          * primary NID of the peer. If it does, update the peer, if
2847          * it it does not, check whether there is already a peer with
2848          * that primary NID. If no such peer exists, try to update
2849          * the primary NID of the current peer (allowed if it was
2850          * created due to message traffic) and complete the update.
2851          * If the peer did exist, hand off the data to it.
2852          *
2853          * The peer for the loopback interface is a special case: this
2854          * is the peer for the local node, and we want to set its
2855          * primary NID to the correct value here. Moreover, this peer
2856          * can show up with only the loopback NID in the ping buffer.
2857          */
2858         if (pbuf->pb_info.pi_nnis <= 1)
2859                 goto out;
2860         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2861         if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2862                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2863                 if (!rc)
2864                         rc = lnet_peer_merge_data(lp, pbuf);
2865         /*
2866          * if the primary nid of the peer is present in the ping info returned
2867          * from the peer, but it's not the local primary peer we have
2868          * cached and discovery is disabled, then we don't want to update
2869          * our local peer info, by adding or removing NIDs, we just want
2870          * to update the status of the nids that we currently have
2871          * recorded in that peer.
2872          */
2873         } else if (lp->lp_primary_nid == nid ||
2874                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2875                     lnet_is_discovery_disabled(lp))) {
2876                 rc = lnet_peer_merge_data(lp, pbuf);
2877         } else {
2878                 lpni = lnet_find_peer_ni_locked(nid);
2879                 if (!lpni) {
2880                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2881                         if (rc) {
2882                                 CERROR("Primary NID error %s versus %s: %d\n",
2883                                        libcfs_nid2str(lp->lp_primary_nid),
2884                                        libcfs_nid2str(nid), rc);
2885                         } else {
2886                                 rc = lnet_peer_merge_data(lp, pbuf);
2887                         }
2888                 } else {
2889                         struct lnet_peer *new_lp;
2890                         new_lp = lpni->lpni_peer_net->lpn_peer;
2891                         /*
2892                          * if lp has discovery/MR enabled that means new_lp
2893                          * should have discovery/MR enabled as well, since
2894                          * it's the same peer, which we're about to merge
2895                          */
2896                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2897                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2898                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2899                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2900
2901                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2902                         lnet_consolidate_routes_locked(lp, new_lp);
2903                         lnet_peer_ni_decref_locked(lpni);
2904                 }
2905         }
2906 out:
2907         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2908                lp->lp_state);
2909         mutex_unlock(&the_lnet.ln_api_mutex);
2910
2911         spin_lock(&lp->lp_lock);
2912         /* Tell discovery to re-check the peer immediately. */
2913         if (!rc)
2914                 rc = LNET_REDISCOVER_PEER;
2915         return rc;
2916 }
2917
2918 /*
2919  * A ping failed. Clear the PING_FAILED state and set the
2920  * FORCE_PING state, to ensure a retry even if discovery is
2921  * disabled. This avoids being left with incorrect state.
2922  */
2923 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2924 __must_hold(&lp->lp_lock)
2925 {
2926         struct lnet_handle_md mdh;
2927         int rc;
2928
2929         mdh = lp->lp_ping_mdh;
2930         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2931         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2932         lp->lp_state |= LNET_PEER_FORCE_PING;
2933         rc = lp->lp_ping_error;
2934         lp->lp_ping_error = 0;
2935         spin_unlock(&lp->lp_lock);
2936
2937         if (!LNetMDHandleIsInvalid(mdh))
2938                 LNetMDUnlink(mdh);
2939
2940         CDEBUG(D_NET, "peer %s:%d\n",
2941                libcfs_nid2str(lp->lp_primary_nid), rc);
2942
2943         spin_lock(&lp->lp_lock);
2944         return rc ? rc : LNET_REDISCOVER_PEER;
2945 }
2946
2947 /*
2948  * Select NID to send a Ping or Push to.
2949  */
2950 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2951 {
2952         struct lnet_peer_ni *lpni;
2953
2954         /* Look for a direct-connected NID for this peer. */
2955         lpni = NULL;
2956         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2957                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2958                         continue;
2959                 break;
2960         }
2961         if (lpni)
2962                 return lpni->lpni_nid;
2963
2964         /* Look for a routed-connected NID for this peer. */
2965         lpni = NULL;
2966         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2967                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2968                         continue;
2969                 break;
2970         }
2971         if (lpni)
2972                 return lpni->lpni_nid;
2973
2974         return LNET_NID_ANY;
2975 }
2976
2977 /* Active side of ping. */
2978 static int lnet_peer_send_ping(struct lnet_peer *lp)
2979 __must_hold(&lp->lp_lock)
2980 {
2981         lnet_nid_t pnid;
2982         int nnis;
2983         int rc;
2984         int cpt;
2985
2986         lp->lp_state |= LNET_PEER_PING_SENT;
2987         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2988         spin_unlock(&lp->lp_lock);
2989
2990         cpt = lnet_net_lock_current();
2991         /* Refcount for MD. */
2992         lnet_peer_addref_locked(lp);
2993         pnid = lnet_peer_select_nid(lp);
2994         lnet_net_unlock(cpt);
2995
2996         nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2997
2998         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2999                             the_lnet.ln_dc_eqh, false);
3000
3001         /*
3002          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3003          * refcount on the peer, otherwise LNetMDUnlink will be called
3004          * which will eventually do that.
3005          */
3006         if (rc > 0) {
3007                 lnet_net_lock(cpt);
3008                 lnet_peer_decref_locked(lp);
3009                 lnet_net_unlock(cpt);
3010                 rc = -rc; /* change the rc to negative value */
3011                 goto fail_error;
3012         } else if (rc < 0) {
3013                 goto fail_error;
3014         }
3015
3016         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3017
3018         spin_lock(&lp->lp_lock);
3019         return 0;
3020
3021 fail_error:
3022         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3023         /*
3024          * The errors that get us here are considered hard errors and
3025          * cause Discovery to terminate. So we clear PING_SENT, but do
3026          * not set either PING_FAILED or FORCE_PING. In fact we need
3027          * to clear PING_FAILED, because the unlink event handler will
3028          * have set it if we called LNetMDUnlink() above.
3029          */
3030         spin_lock(&lp->lp_lock);
3031         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3032         return rc;
3033 }
3034
3035 /*
3036  * This function exists because you cannot call LNetMDUnlink() from an
3037  * event handler.
3038  */
3039 static int lnet_peer_push_failed(struct lnet_peer *lp)
3040 __must_hold(&lp->lp_lock)
3041 {
3042         struct lnet_handle_md mdh;
3043         int rc;
3044
3045         mdh = lp->lp_push_mdh;
3046         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3047         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3048         rc = lp->lp_push_error;
3049         lp->lp_push_error = 0;
3050         spin_unlock(&lp->lp_lock);
3051
3052         if (!LNetMDHandleIsInvalid(mdh))
3053                 LNetMDUnlink(mdh);
3054
3055         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3056         spin_lock(&lp->lp_lock);
3057         return rc ? rc : LNET_REDISCOVER_PEER;
3058 }
3059
3060 /* Active side of push. */
3061 static int lnet_peer_send_push(struct lnet_peer *lp)
3062 __must_hold(&lp->lp_lock)
3063 {
3064         struct lnet_ping_buffer *pbuf;
3065         struct lnet_process_id id;
3066         struct lnet_md md;
3067         int cpt;
3068         int rc;
3069
3070         /* Don't push to a non-multi-rail peer. */
3071         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3072                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3073                 return 0;
3074         }
3075
3076         lp->lp_state |= LNET_PEER_PUSH_SENT;
3077         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3078         spin_unlock(&lp->lp_lock);
3079
3080         cpt = lnet_net_lock_current();
3081         pbuf = the_lnet.ln_ping_target;
3082         lnet_ping_buffer_addref(pbuf);
3083         lnet_net_unlock(cpt);
3084
3085         /* Push source MD */
3086         md.start     = &pbuf->pb_info;
3087         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3088         md.threshold = 2; /* Put/Ack */
3089         md.max_size  = 0;
3090         md.options   = 0;
3091         md.eq_handle = the_lnet.ln_dc_eqh;
3092         md.user_ptr  = lp;
3093
3094         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3095         if (rc) {
3096                 lnet_ping_buffer_decref(pbuf);
3097                 CERROR("Can't bind push source MD: %d\n", rc);
3098                 goto fail_error;
3099         }
3100         cpt = lnet_net_lock_current();
3101         /* Refcount for MD. */
3102         lnet_peer_addref_locked(lp);
3103         id.pid = LNET_PID_LUSTRE;
3104         id.nid = lnet_peer_select_nid(lp);
3105         lnet_net_unlock(cpt);
3106
3107         if (id.nid == LNET_NID_ANY) {
3108                 rc = -EHOSTUNREACH;
3109                 goto fail_unlink;
3110         }
3111
3112         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3113                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3114                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3115
3116         if (rc)
3117                 goto fail_unlink;
3118
3119         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3120
3121         spin_lock(&lp->lp_lock);
3122         return 0;
3123
3124 fail_unlink:
3125         LNetMDUnlink(lp->lp_push_mdh);
3126         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3127 fail_error:
3128         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3129         /*
3130          * The errors that get us here are considered hard errors and
3131          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3132          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3133          * because the unlink event handler will have set it if we
3134          * called LNetMDUnlink() above.
3135          */
3136         spin_lock(&lp->lp_lock);
3137         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3138         return rc;
3139 }
3140
3141 /*
3142  * An unrecoverable error was encountered during discovery.
3143  * Set error status in peer and abort discovery.
3144  */
3145 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3146 {
3147         CDEBUG(D_NET, "Discovery error %s: %d\n",
3148                libcfs_nid2str(lp->lp_primary_nid), error);
3149
3150         spin_lock(&lp->lp_lock);
3151         lp->lp_dc_error = error;
3152         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3153         lp->lp_state |= LNET_PEER_REDISCOVER;
3154         spin_unlock(&lp->lp_lock);
3155 }
3156
3157 /*
3158  * Mark the peer as discovered.
3159  */
3160 static int lnet_peer_discovered(struct lnet_peer *lp)
3161 __must_hold(&lp->lp_lock)
3162 {
3163         lp->lp_state |= LNET_PEER_DISCOVERED;
3164         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3165                           LNET_PEER_REDISCOVER);
3166
3167         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3168
3169         return 0;
3170 }
3171
3172
3173 /*
3174  * Discovering this peer is taking too long. Cancel any Ping or Push
3175  * that discovery is waiting on by unlinking the relevant MDs. The
3176  * lnet_discovery_event_handler() will proceed from here and complete
3177  * the cleanup.
3178  */
3179 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3180 {
3181         struct lnet_handle_md ping_mdh;
3182         struct lnet_handle_md push_mdh;
3183
3184         LNetInvalidateMDHandle(&ping_mdh);
3185         LNetInvalidateMDHandle(&push_mdh);
3186
3187         spin_lock(&lp->lp_lock);
3188         if (lp->lp_state & LNET_PEER_PING_SENT) {
3189                 ping_mdh = lp->lp_ping_mdh;
3190                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3191         }
3192         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3193                 push_mdh = lp->lp_push_mdh;
3194                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3195         }
3196         spin_unlock(&lp->lp_lock);
3197
3198         if (!LNetMDHandleIsInvalid(ping_mdh))
3199                 LNetMDUnlink(ping_mdh);
3200         if (!LNetMDHandleIsInvalid(push_mdh))
3201                 LNetMDUnlink(push_mdh);
3202 }
3203
3204 /*
3205  * Wait for work to be queued or some other change that must be
3206  * attended to. Returns non-zero if the discovery thread should shut
3207  * down.
3208  */
3209 static int lnet_peer_discovery_wait_for_work(void)
3210 {
3211         int cpt;
3212         int rc = 0;
3213
3214         DEFINE_WAIT(wait);
3215
3216         cpt = lnet_net_lock_current();
3217         for (;;) {
3218                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3219                                 TASK_INTERRUPTIBLE);
3220                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3221                         break;
3222                 if (lnet_push_target_resize_needed())
3223                         break;
3224                 if (!list_empty(&the_lnet.ln_dc_request))
3225                         break;
3226                 if (!list_empty(&the_lnet.ln_msg_resend))
3227                         break;
3228                 lnet_net_unlock(cpt);
3229
3230                 /*
3231                  * wakeup max every second to check if there are peers that
3232                  * have been stuck on the working queue for greater than
3233                  * the peer timeout.
3234                  */
3235                 schedule_timeout(cfs_time_seconds(1));
3236                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3237                 cpt = lnet_net_lock_current();
3238         }
3239         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3240
3241         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3242                 rc = -ESHUTDOWN;
3243
3244         lnet_net_unlock(cpt);
3245
3246         CDEBUG(D_NET, "woken: %d\n", rc);
3247
3248         return rc;
3249 }
3250
3251 /*
3252  * Messages that were pending on a destroyed peer will be put on a global
3253  * resend list. The message resend list will be checked by
3254  * the discovery thread when it wakes up, and will resend messages. These
3255  * messages can still be sendable in the case the lpni which was the initial
3256  * cause of the message re-queue was transfered to another peer.
3257  *
3258  * It is possible that LNet could be shutdown while we're iterating
3259  * through the list. lnet_shudown_lndnets() will attempt to access the
3260  * resend list, but will have to wait until the spinlock is released, by
3261  * which time there shouldn't be any more messages on the resend list.
3262  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3263  * for the messages so they can be released. The other case is that
3264  * lnet_shudown_lndnets() can finalize all the messages before this
3265  * function can visit the resend list, in which case this function will be
3266  * a no-op.
3267  */
3268 static void lnet_resend_msgs(void)
3269 {
3270         struct lnet_msg *msg, *tmp;
3271         struct list_head resend;
3272         int rc;
3273
3274         INIT_LIST_HEAD(&resend);
3275
3276         spin_lock(&the_lnet.ln_msg_resend_lock);
3277         list_splice(&the_lnet.ln_msg_resend, &resend);
3278         spin_unlock(&the_lnet.ln_msg_resend_lock);
3279
3280         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3281                 list_del_init(&msg->msg_list);
3282                 rc = lnet_send(msg->msg_src_nid_param, msg,
3283                                msg->msg_rtr_nid_param);
3284                 if (rc < 0) {
3285                         CNETERR("Error sending %s to %s: %d\n",
3286                                lnet_msgtyp2str(msg->msg_type),
3287                                libcfs_id2str(msg->msg_target), rc);
3288                         lnet_finalize(msg, rc);
3289                 }
3290         }
3291 }
3292
3293 /* The discovery thread. */
3294 static int lnet_peer_discovery(void *arg)
3295 {
3296         struct lnet_peer *lp;
3297         int rc;
3298
3299         wait_for_completion(&the_lnet.ln_started);
3300
3301         CDEBUG(D_NET, "started\n");
3302         cfs_block_allsigs();
3303
3304         for (;;) {
3305                 if (lnet_peer_discovery_wait_for_work())
3306                         break;
3307
3308                 lnet_resend_msgs();
3309
3310                 if (lnet_push_target_resize_needed())
3311                         lnet_push_target_resize();
3312
3313                 lnet_net_lock(LNET_LOCK_EX);
3314                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3315                         break;
3316
3317                 /*
3318                  * Process all incoming discovery work requests.  When
3319                  * discovery must wait on a peer to change state, it
3320                  * is added to the tail of the ln_dc_working queue. A
3321                  * timestamp keeps track of when the peer was added,
3322                  * so we can time out discovery requests that take too
3323                  * long.
3324                  */
3325                 while (!list_empty(&the_lnet.ln_dc_request)) {
3326                         lp = list_first_entry(&the_lnet.ln_dc_request,
3327                                               struct lnet_peer, lp_dc_list);
3328                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3329                         /*
3330                          * set the time the peer was put on the dc_working
3331                          * queue. It shouldn't remain on the queue
3332                          * forever, in case the GET message (for ping)
3333                          * doesn't get a REPLY or the PUT message (for
3334                          * push) doesn't get an ACK.
3335                          */
3336                         lp->lp_last_queued = ktime_get_real_seconds();
3337                         lnet_net_unlock(LNET_LOCK_EX);
3338
3339                         /*
3340                          * Select an action depending on the state of
3341                          * the peer and whether discovery is disabled.
3342                          * The check whether discovery is disabled is
3343                          * done after the code that handles processing
3344                          * for arrived data, cleanup for failures, and
3345                          * forcing a Ping or Push.
3346                          */
3347                         spin_lock(&lp->lp_lock);
3348                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3349                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3350                                 lp->lp_state);
3351                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3352                                 rc = lnet_peer_data_present(lp);
3353                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3354                                 rc = lnet_peer_ping_failed(lp);
3355                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3356                                 rc = lnet_peer_push_failed(lp);
3357                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3358                                 rc = lnet_peer_send_ping(lp);
3359                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3360                                 rc = lnet_peer_send_push(lp);
3361                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3362                                 rc = lnet_peer_send_ping(lp);
3363                         else if (lnet_peer_needs_push(lp))
3364                                 rc = lnet_peer_send_push(lp);
3365                         else
3366                                 rc = lnet_peer_discovered(lp);
3367                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3368                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3369                                 lp->lp_state, rc);
3370                         spin_unlock(&lp->lp_lock);
3371
3372                         lnet_net_lock(LNET_LOCK_EX);
3373                         if (rc == LNET_REDISCOVER_PEER) {
3374                                 list_move(&lp->lp_dc_list,
3375                                           &the_lnet.ln_dc_request);
3376                         } else if (rc) {
3377                                 lnet_peer_discovery_error(lp, rc);
3378                         }
3379                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3380                                 lnet_peer_discovery_complete(lp);
3381                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3382                                 break;
3383                 }
3384
3385                 lnet_net_unlock(LNET_LOCK_EX);
3386         }
3387
3388         CDEBUG(D_NET, "stopping\n");
3389         /*
3390          * Clean up before telling lnet_peer_discovery_stop() that
3391          * we're done. Use wake_up() below to somewhat reduce the
3392          * size of the thundering herd if there are multiple threads
3393          * waiting on discovery of a single peer.
3394          */
3395
3396         /* Queue cleanup 1: stop all pending pings and pushes. */
3397         lnet_net_lock(LNET_LOCK_EX);
3398         while (!list_empty(&the_lnet.ln_dc_working)) {
3399                 lp = list_first_entry(&the_lnet.ln_dc_working,
3400                                       struct lnet_peer, lp_dc_list);
3401                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3402                 lnet_net_unlock(LNET_LOCK_EX);
3403                 lnet_peer_cancel_discovery(lp);
3404                 lnet_net_lock(LNET_LOCK_EX);
3405         }
3406         lnet_net_unlock(LNET_LOCK_EX);
3407
3408         /* Queue cleanup 2: wait for the expired queue to clear. */
3409         while (!list_empty(&the_lnet.ln_dc_expired))
3410                 schedule_timeout(cfs_time_seconds(1));
3411
3412         /* Queue cleanup 3: clear the request queue. */
3413         lnet_net_lock(LNET_LOCK_EX);
3414         while (!list_empty(&the_lnet.ln_dc_request)) {
3415                 lp = list_first_entry(&the_lnet.ln_dc_request,
3416                                       struct lnet_peer, lp_dc_list);
3417                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3418                 lnet_peer_discovery_complete(lp);
3419         }
3420         lnet_net_unlock(LNET_LOCK_EX);
3421
3422         LNetEQFree(the_lnet.ln_dc_eqh);
3423         LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3424
3425         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3426         wake_up(&the_lnet.ln_dc_waitq);
3427
3428         CDEBUG(D_NET, "stopped\n");
3429
3430         return 0;
3431 }
3432
3433 /* ln_api_mutex is held on entry. */
3434 int lnet_peer_discovery_start(void)
3435 {
3436         struct task_struct *task;
3437         int rc;
3438
3439         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3440                 return -EALREADY;
3441
3442         rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3443         if (rc != 0) {
3444                 CERROR("Can't allocate discovery EQ: %d\n", rc);
3445                 return rc;
3446         }
3447
3448         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3449         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3450         if (IS_ERR(task)) {
3451                 rc = PTR_ERR(task);
3452                 CERROR("Can't start peer discovery thread: %d\n", rc);
3453
3454                 LNetEQFree(the_lnet.ln_dc_eqh);
3455                 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3456
3457                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3458         }
3459
3460         CDEBUG(D_NET, "discovery start: %d\n", rc);
3461
3462         return rc;
3463 }
3464
3465 /* ln_api_mutex is held on entry. */
3466 void lnet_peer_discovery_stop(void)
3467 {
3468         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3469                 return;
3470
3471         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3472         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3473
3474         /* In the LNetNIInit() path we may be stopping discovery before it
3475          * entered its work loop
3476          */
3477         if (!completion_done(&the_lnet.ln_started))
3478                 complete(&the_lnet.ln_started);
3479         else
3480                 wake_up(&the_lnet.ln_dc_waitq);
3481
3482         wait_event(the_lnet.ln_dc_waitq,
3483                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3484
3485         LASSERT(list_empty(&the_lnet.ln_dc_request));
3486         LASSERT(list_empty(&the_lnet.ln_dc_working));
3487         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3488
3489         CDEBUG(D_NET, "discovery stopped\n");
3490 }
3491
3492 /* Debugging */
3493
3494 void
3495 lnet_debug_peer(lnet_nid_t nid)
3496 {
3497         char                    *aliveness = "NA";
3498         struct lnet_peer_ni     *lp;
3499         int                     cpt;
3500
3501         cpt = lnet_cpt_of_nid(nid, NULL);
3502         lnet_net_lock(cpt);
3503
3504         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3505         if (IS_ERR(lp)) {
3506                 lnet_net_unlock(cpt);
3507                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3508                 return;
3509         }
3510
3511         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3512                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3513
3514         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3515                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3516                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3517                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3518                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3519
3520         lnet_peer_ni_decref_locked(lp);
3521
3522         lnet_net_unlock(cpt);
3523 }
3524
3525 /* Gathering information for userspace. */
3526
3527 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3528                           char aliveness[LNET_MAX_STR_LEN],
3529                           __u32 *cpt_iter, __u32 *refcount,
3530                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3531                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3532                           __u32 *peer_tx_qnob)
3533 {
3534         struct lnet_peer_table          *peer_table;
3535         struct lnet_peer_ni             *lp;
3536         int                             j;
3537         int                             lncpt;
3538         bool                            found = false;
3539
3540         /* get the number of CPTs */
3541         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3542
3543         /* if the cpt number to be examined is >= the number of cpts in
3544          * the system then indicate that there are no more cpts to examin
3545          */
3546         if (*cpt_iter >= lncpt)
3547                 return -ENOENT;
3548
3549         /* get the current table */
3550         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3551         /* if the ptable is NULL then there are no more cpts to examine */
3552         if (peer_table == NULL)
3553                 return -ENOENT;
3554
3555         lnet_net_lock(*cpt_iter);
3556
3557         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3558                 struct list_head *peers = &peer_table->pt_hash[j];
3559
3560                 list_for_each_entry(lp, peers, lpni_hashlist) {
3561                         if (peer_index-- > 0)
3562                                 continue;
3563
3564                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3565                         if (lnet_isrouter(lp) ||
3566                                 lnet_peer_aliveness_enabled(lp))
3567                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3568                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3569
3570                         *nid = lp->lpni_nid;
3571                         *refcount = atomic_read(&lp->lpni_refcount);
3572                         *ni_peer_tx_credits =
3573                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3574                         *peer_tx_credits = lp->lpni_txcredits;
3575                         *peer_rtr_credits = lp->lpni_rtrcredits;
3576                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3577                         *peer_tx_qnob = lp->lpni_txqnob;
3578
3579                         found = true;
3580                 }
3581
3582         }
3583         lnet_net_unlock(*cpt_iter);
3584
3585         *cpt_iter = lncpt;
3586
3587         return found ? 0 : -ENOENT;
3588 }
3589
3590 /* ln_api_mutex is held, which keeps the peer list stable */
3591 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3592 {
3593         struct lnet_ioctl_element_stats *lpni_stats;
3594         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3595         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3596         struct lnet_peer_ni_credit_info *lpni_info;
3597         struct lnet_peer_ni *lpni;
3598         struct lnet_peer *lp;
3599         lnet_nid_t nid;
3600         __u32 size;
3601         int rc;
3602
3603         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3604
3605         if (!lp) {
3606                 rc = -ENOENT;
3607                 goto out;
3608         }
3609
3610         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3611                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3612         size *= lp->lp_nnis;
3613         if (size > cfg->prcfg_size) {
3614                 cfg->prcfg_size = size;
3615                 rc = -E2BIG;
3616                 goto out_lp_decref;
3617         }
3618
3619         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3620         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3621         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3622         cfg->prcfg_count = lp->lp_nnis;
3623         cfg->prcfg_size = size;
3624         cfg->prcfg_state = lp->lp_state;
3625
3626         /* Allocate helper buffers. */
3627         rc = -ENOMEM;
3628         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3629         if (!lpni_info)
3630                 goto out_lp_decref;
3631         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3632         if (!lpni_stats)
3633                 goto out_free_info;
3634         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3635         if (!lpni_msg_stats)
3636                 goto out_free_stats;
3637         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3638         if (!lpni_hstats)
3639                 goto out_free_msg_stats;
3640
3641
3642         lpni = NULL;
3643         rc = -EFAULT;
3644         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3645                 nid = lpni->lpni_nid;
3646                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3647                         goto out_free_hstats;
3648                 bulk += sizeof(nid);
3649
3650                 memset(lpni_info, 0, sizeof(*lpni_info));
3651                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3652                 if (lnet_isrouter(lpni) ||
3653                         lnet_peer_aliveness_enabled(lpni))
3654                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3655                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3656
3657                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3658                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3659                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3660                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3661                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3662                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3663                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3664                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3665                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3666                         goto out_free_hstats;
3667                 bulk += sizeof(*lpni_info);
3668
3669                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3670                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3671                                                             LNET_STATS_TYPE_SEND);
3672                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3673                                                             LNET_STATS_TYPE_RECV);
3674                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3675                                                             LNET_STATS_TYPE_DROP);
3676                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3677                         goto out_free_hstats;
3678                 bulk += sizeof(*lpni_stats);
3679                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3680                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3681                         goto out_free_hstats;
3682                 bulk += sizeof(*lpni_msg_stats);
3683                 lpni_hstats->hlpni_network_timeout =
3684                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3685                 lpni_hstats->hlpni_remote_dropped =
3686                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3687                 lpni_hstats->hlpni_remote_timeout =
3688                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3689                 lpni_hstats->hlpni_remote_error =
3690                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3691                 lpni_hstats->hlpni_health_value =
3692                   atomic_read(&lpni->lpni_healthv);
3693                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3694                         goto out_free_hstats;
3695                 bulk += sizeof(*lpni_hstats);
3696         }
3697         rc = 0;
3698
3699 out_free_hstats:
3700         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3701 out_free_msg_stats:
3702         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3703 out_free_stats:
3704         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3705 out_free_info:
3706         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3707 out_lp_decref:
3708         lnet_peer_decref_locked(lp);
3709 out:
3710         return rc;
3711 }
3712
3713 void
3714 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3715 {
3716         /* the mt could've shutdown and cleaned up the queues */
3717         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3718                 return;
3719
3720         if (list_empty(&lpni->lpni_recovery) &&
3721             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3722                 CERROR("lpni %s added to recovery queue. Health = %d\n",
3723                         libcfs_nid2str(lpni->lpni_nid),
3724                         atomic_read(&lpni->lpni_healthv));
3725                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3726                 lnet_peer_ni_addref_locked(lpni);
3727         }
3728 }
3729
3730 /* Call with the ln_api_mutex held */
3731 void
3732 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3733 {
3734         struct lnet_peer_table *ptable;
3735         struct lnet_peer *lp;
3736         struct lnet_peer_net *lpn;
3737         struct lnet_peer_ni *lpni;
3738         int lncpt;
3739         int cpt;
3740
3741         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3742                 return;
3743
3744         if (!all) {
3745                 lnet_net_lock(LNET_LOCK_EX);
3746                 lpni = lnet_find_peer_ni_locked(nid);
3747                 if (!lpni) {
3748                         lnet_net_unlock(LNET_LOCK_EX);
3749                         return;
3750                 }
3751                 atomic_set(&lpni->lpni_healthv, value);
3752                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3753                 lnet_peer_ni_decref_locked(lpni);
3754                 lnet_net_unlock(LNET_LOCK_EX);
3755                 return;
3756         }
3757
3758         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3759
3760         /*
3761          * Walk all the peers and reset the healhv for each one to the
3762          * maximum value.
3763          */
3764         lnet_net_lock(LNET_LOCK_EX);
3765         for (cpt = 0; cpt < lncpt; cpt++) {
3766                 ptable = the_lnet.ln_peer_tables[cpt];
3767                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3768                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3769                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3770                                                     lpni_peer_nis) {
3771                                         atomic_set(&lpni->lpni_healthv, value);
3772                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3773                                 }
3774                         }
3775                 }
3776         }
3777         lnet_net_unlock(LNET_LOCK_EX);
3778 }
3779