lnet/lnet/peer.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/peer.c
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_LNET
  36
  37 #include <linux/sched.h>
  38 #ifdef HAVE_SCHED_HEADERS
  39 #include <linux/sched/signal.h>
  40 #endif
  41 #include <linux/uaccess.h>
  42
  43 #include <lnet/lib-lnet.h>
  44 #include <uapi/linux/lnet/lnet-dlc.h>
  45
  46 /* Value indicating that recovery needs to re-check a peer immediately. */
  47 #define LNET_REDISCOVER_PEER    (1)
  48
  49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
  50
  51 static void
  52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
  53 {
  54         if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
  55                 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
  56                 lnet_peer_ni_decref_locked(lpni);
  57         }
  58 }
  59
  60 void
  61 lnet_peer_net_added(struct lnet_net *net)
  62 {
  63         struct lnet_peer_ni *lpni, *tmp;
  64
  65         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
  66                                  lpni_on_remote_peer_ni_list) {
  67
  68                 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
  69                         lpni->lpni_net = net;
  70
  71                         spin_lock(&lpni->lpni_lock);
  72                         lpni->lpni_txcredits =
  73                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
  74                         lpni->lpni_mintxcredits = lpni->lpni_txcredits;
  75                         lpni->lpni_rtrcredits =
  76                                 lnet_peer_buffer_credits(lpni->lpni_net);
  77                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
  78                         spin_unlock(&lpni->lpni_lock);
  79
  80                         lnet_peer_remove_from_remote_list(lpni);
  81                 }
  82         }
  83 }
  84
  85 static void
  86 lnet_peer_tables_destroy(void)
  87 {
  88         struct lnet_peer_table  *ptable;
  89         struct list_head        *hash;
  90         int                     i;
  91         int                     j;
  92
  93         if (!the_lnet.ln_peer_tables)
  94                 return;
  95
  96         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  97                 hash = ptable->pt_hash;
  98                 if (!hash) /* not intialized */
  99                         break;
 100
 101                 LASSERT(list_empty(&ptable->pt_zombie_list));
 102
 103                 ptable->pt_hash = NULL;
 104                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 105                         LASSERT(list_empty(&hash[j]));
 106
 107                 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
 108         }
 109
 110         cfs_percpt_free(the_lnet.ln_peer_tables);
 111         the_lnet.ln_peer_tables = NULL;
 112 }
 113
 114 int
 115 lnet_peer_tables_create(void)
 116 {
 117         struct lnet_peer_table  *ptable;
 118         struct list_head        *hash;
 119         int                     i;
 120         int                     j;
 121
 122         the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
 123                                                    sizeof(*ptable));
 124         if (the_lnet.ln_peer_tables == NULL) {
 125                 CERROR("Failed to allocate cpu-partition peer tables\n");
 126                 return -ENOMEM;
 127         }
 128
 129         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 130                 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
 131                                  LNET_PEER_HASH_SIZE * sizeof(*hash));
 132                 if (hash == NULL) {
 133                         CERROR("Failed to create peer hash table\n");
 134                         lnet_peer_tables_destroy();
 135                         return -ENOMEM;
 136                 }
 137
 138                 spin_lock_init(&ptable->pt_zombie_lock);
 139                 INIT_LIST_HEAD(&ptable->pt_zombie_list);
 140
 141                 INIT_LIST_HEAD(&ptable->pt_peer_list);
 142
 143                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 144                         INIT_LIST_HEAD(&hash[j]);
 145                 ptable->pt_hash = hash; /* sign of initialization */
 146         }
 147
 148         return 0;
 149 }
 150
 151 static struct lnet_peer_ni *
 152 lnet_peer_ni_alloc(lnet_nid_t nid)
 153 {
 154         struct lnet_peer_ni *lpni;
 155         struct lnet_net *net;
 156         int cpt;
 157
 158         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 159
 160         LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
 161         if (!lpni)
 162                 return NULL;
 163
 164         INIT_LIST_HEAD(&lpni->lpni_txq);
 165         INIT_LIST_HEAD(&lpni->lpni_hashlist);
 166         INIT_LIST_HEAD(&lpni->lpni_peer_nis);
 167         INIT_LIST_HEAD(&lpni->lpni_recovery);
 168         INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
 169         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         if (lnet_peers_start_down())
 262                 lp->lp_alive = false;
 263         else
 264                 lp->lp_alive = true;
 265
 266         /*
 267          * all peers created on a router should have health on
 268          * if it's not already on.
 269          */
 270         if (the_lnet.ln_routing && !lnet_health_sensitivity)
 271                 lp->lp_health_sensitivity = 1;
 272
 273         /*
 274          * Turn off discovery for loopback peer. If you're creating a peer
 275          * for the loopback interface then that was initiated when we
 276          * attempted to send a message over the loopback. There is no need
 277          * to ever use a different interface when sending messages to
 278          * myself.
 279          */
 280         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 281                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 282         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 283
 284         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 285
 286         return lp;
 287 }
 288
 289 void
 290 lnet_destroy_peer_locked(struct lnet_peer *lp)
 291 {
 292         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 293
 294         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 295         LASSERT(lp->lp_rtr_refcount == 0);
 296         LASSERT(list_empty(&lp->lp_peer_nets));
 297         LASSERT(list_empty(&lp->lp_peer_list));
 298         LASSERT(list_empty(&lp->lp_dc_list));
 299
 300         if (lp->lp_data)
 301                 lnet_ping_buffer_decref(lp->lp_data);
 302
 303         /*
 304          * if there are messages still on the pending queue, then make
 305          * sure to queue them on the ln_msg_resend list so they can be
 306          * resent at a later point if the discovery thread is still
 307          * running.
 308          * If the discovery thread has stopped, then the wakeup will be a
 309          * no-op, and it is expected the lnet_shutdown_lndnets() will
 310          * eventually be called, which will traverse this list and
 311          * finalize the messages on the list.
 312          * We can not resend them now because we're holding the cpt lock.
 313          * Releasing the lock can cause an inconsistent state
 314          */
 315         spin_lock(&the_lnet.ln_msg_resend_lock);
 316         spin_lock(&lp->lp_lock);
 317         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 318         spin_unlock(&lp->lp_lock);
 319         spin_unlock(&the_lnet.ln_msg_resend_lock);
 320         wake_up(&the_lnet.ln_dc_waitq);
 321
 322         LIBCFS_FREE(lp, sizeof(*lp));
 323 }
 324
 325 /*
 326  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 327  * that peer_net, detach the peer_net from the peer.
 328  *
 329  * Call with lnet_net_lock/EX held
 330  */
 331 static void
 332 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 333 {
 334         struct lnet_peer_table *ptable;
 335         struct lnet_peer_net *lpn;
 336         struct lnet_peer *lp;
 337
 338         /*
 339          * Belts and suspenders: gracefully handle teardown of a
 340          * partially connected peer_ni.
 341          */
 342         lpn = lpni->lpni_peer_net;
 343
 344         list_del_init(&lpni->lpni_peer_nis);
 345         /*
 346          * If there are no lpni's left, we detach lpn from
 347          * lp_peer_nets, so it cannot be found anymore.
 348          */
 349         if (list_empty(&lpn->lpn_peer_nis))
 350                 list_del_init(&lpn->lpn_peer_nets);
 351
 352         /* Update peer NID count. */
 353         lp = lpn->lpn_peer;
 354         lp->lp_nnis--;
 355
 356         /*
 357          * If there are no more peer nets, make the peer unfindable
 358          * via the peer_tables.
 359          *
 360          * Otherwise, if the peer is DISCOVERED, tell discovery to
 361          * take another look at it. This is a no-op if discovery for
 362          * this peer did the detaching.
 363          */
 364         if (list_empty(&lp->lp_peer_nets)) {
 365                 list_del_init(&lp->lp_peer_list);
 366                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 367                 ptable->pt_peers--;
 368         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 369                 /* Discovery isn't running, nothing to do here. */
 370         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 371                 lnet_peer_queue_for_discovery(lp);
 372                 wake_up(&the_lnet.ln_dc_waitq);
 373         }
 374         CDEBUG(D_NET, "peer %s NID %s\n",
 375                 libcfs_nid2str(lp->lp_primary_nid),
 376                 libcfs_nid2str(lpni->lpni_nid));
 377 }
 378
 379 /* called with lnet_net_lock LNET_LOCK_EX held */
 380 static int
 381 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 382 {
 383         struct lnet_peer_table *ptable = NULL;
 384
 385         /* don't remove a peer_ni if it's also a gateway */
 386         if (lnet_isrouter(lpni) && !force) {
 387                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 388                        libcfs_nid2str(lpni->lpni_nid));
 389                 return -EBUSY;
 390         }
 391
 392         lnet_peer_remove_from_remote_list(lpni);
 393
 394         /* remove peer ni from the hash list. */
 395         list_del_init(&lpni->lpni_hashlist);
 396
 397         /*
 398          * indicate the peer is being deleted so the monitor thread can
 399          * remove it from the recovery queue.
 400          */
 401         spin_lock(&lpni->lpni_lock);
 402         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 403         spin_unlock(&lpni->lpni_lock);
 404
 405         /* decrement the ref count on the peer table */
 406         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 407
 408         /*
 409          * The peer_ni can no longer be found with a lookup. But there
 410          * can be current users, so keep track of it on the zombie
 411          * list until the reference count has gone to zero.
 412          *
 413          * The last reference may be lost in a place where the
 414          * lnet_net_lock locks only a single cpt, and that cpt may not
 415          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 416          * has its own lock.
 417          */
 418         spin_lock(&ptable->pt_zombie_lock);
 419         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 420         ptable->pt_zombies++;
 421         spin_unlock(&ptable->pt_zombie_lock);
 422
 423         /* no need to keep this peer_ni on the hierarchy anymore */
 424         lnet_peer_detach_peer_ni_locked(lpni);
 425
 426         /* remove hashlist reference on peer_ni */
 427         lnet_peer_ni_decref_locked(lpni);
 428
 429         return 0;
 430 }
 431
 432 void lnet_peer_uninit(void)
 433 {
 434         struct lnet_peer_ni *lpni, *tmp;
 435
 436         lnet_net_lock(LNET_LOCK_EX);
 437
 438         /* remove all peer_nis from the remote peer and the hash list */
 439         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 440                                  lpni_on_remote_peer_ni_list)
 441                 lnet_peer_ni_del_locked(lpni, false);
 442
 443         lnet_peer_tables_destroy();
 444
 445         lnet_net_unlock(LNET_LOCK_EX);
 446 }
 447
 448 static int
 449 lnet_peer_del_locked(struct lnet_peer *peer)
 450 {
 451         struct lnet_peer_ni *lpni = NULL, *lpni2;
 452         int rc = 0, rc2 = 0;
 453
 454         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
 455
 456         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 457         while (lpni != NULL) {
 458                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 459                 rc = lnet_peer_ni_del_locked(lpni, false);
 460                 if (rc != 0)
 461                         rc2 = rc;
 462                 lpni = lpni2;
 463         }
 464
 465         return rc2;
 466 }
 467
 468 static int
 469 lnet_peer_del(struct lnet_peer *peer)
 470 {
 471         lnet_net_lock(LNET_LOCK_EX);
 472         lnet_peer_del_locked(peer);
 473         lnet_net_unlock(LNET_LOCK_EX);
 474
 475         return 0;
 476 }
 477
 478 /*
 479  * Delete a NID from a peer. Call with ln_api_mutex held.
 480  *
 481  * Error codes:
 482  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 483  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 484  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 485  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 486  */
 487 static int
 488 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
 489 {
 490         struct lnet_peer_ni *lpni;
 491         lnet_nid_t primary_nid = lp->lp_primary_nid;
 492         int rc = 0;
 493         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 494
 495         if (!(flags & LNET_PEER_CONFIGURED)) {
 496                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 497                         rc = -EPERM;
 498                         goto out;
 499                 }
 500         }
 501         lpni = lnet_find_peer_ni_locked(nid);
 502         if (!lpni) {
 503                 rc = -ENOENT;
 504                 goto out;
 505         }
 506         lnet_peer_ni_decref_locked(lpni);
 507         if (lp != lpni->lpni_peer_net->lpn_peer) {
 508                 rc = -ECHILD;
 509                 goto out;
 510         }
 511
 512         /*
 513          * This function only allows deletion of the primary NID if it
 514          * is the only NID.
 515          */
 516         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
 517                 rc = -EBUSY;
 518                 goto out;
 519         }
 520
 521         lnet_net_lock(LNET_LOCK_EX);
 522
 523         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
 524                 struct lnet_peer_ni *lpni2;
 525                 /* assign the next peer_ni to be the primary */
 526                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 527                 LASSERT(lpni2);
 528                 lp->lp_primary_nid = lpni->lpni_nid;
 529         }
 530         rc = lnet_peer_ni_del_locked(lpni, force);
 531
 532         lnet_net_unlock(LNET_LOCK_EX);
 533
 534 out:
 535         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 536                libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
 537
 538         return rc;
 539 }
 540
 541 static void
 542 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 543                                struct lnet_peer_table *ptable)
 544 {
 545         int                      i;
 546         struct lnet_peer_ni     *next;
 547         struct lnet_peer_ni     *lpni;
 548         struct lnet_peer        *peer;
 549
 550         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 551                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 552                                          lpni_hashlist) {
 553                         if (net != NULL && net != lpni->lpni_net)
 554                                 continue;
 555
 556                         peer = lpni->lpni_peer_net->lpn_peer;
 557                         if (peer->lp_primary_nid != lpni->lpni_nid) {
 558                                 lnet_peer_ni_del_locked(lpni, false);
 559                                 continue;
 560                         }
 561                         /*
 562                          * Removing the primary NID implies removing
 563                          * the entire peer. Advance next beyond any
 564                          * peer_ni that belongs to the same peer.
 565                          */
 566                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 567                                                  lpni_hashlist) {
 568                                 if (next->lpni_peer_net->lpn_peer != peer)
 569                                         break;
 570                         }
 571                         lnet_peer_del_locked(peer);
 572                 }
 573         }
 574 }
 575
 576 static void
 577 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 578 {
 579         int     i = 3;
 580
 581         spin_lock(&ptable->pt_zombie_lock);
 582         while (ptable->pt_zombies) {
 583                 spin_unlock(&ptable->pt_zombie_lock);
 584
 585                 if (is_power_of_2(i)) {
 586                         CDEBUG(D_WARNING,
 587                                "Waiting for %d zombies on peer table\n",
 588                                ptable->pt_zombies);
 589                 }
 590                 schedule_timeout_uninterruptible(cfs_time_seconds(1) >> 1);
 591                 spin_lock(&ptable->pt_zombie_lock);
 592         }
 593         spin_unlock(&ptable->pt_zombie_lock);
 594 }
 595
 596 static void
 597 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 598                                 struct lnet_peer_table *ptable)
 599 {
 600         struct lnet_peer_ni     *lp;
 601         struct lnet_peer_ni     *tmp;
 602         lnet_nid_t              gw_nid;
 603         int                     i;
 604
 605         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 606                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 607                                          lpni_hashlist) {
 608                         if (net != lp->lpni_net)
 609                                 continue;
 610
 611                         if (!lnet_isrouter(lp))
 612                                 continue;
 613
 614                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 615
 616                         lnet_net_unlock(LNET_LOCK_EX);
 617                         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
 618                         lnet_net_lock(LNET_LOCK_EX);
 619                 }
 620         }
 621 }
 622
 623 void
 624 lnet_peer_tables_cleanup(struct lnet_net *net)
 625 {
 626         int i;
 627         struct lnet_peer_table *ptable;
 628
 629         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 630         /* If just deleting the peers for a NI, get rid of any routes these
 631          * peers are gateways for. */
 632         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 633                 lnet_net_lock(LNET_LOCK_EX);
 634                 lnet_peer_table_del_rtrs_locked(net, ptable);
 635                 lnet_net_unlock(LNET_LOCK_EX);
 636         }
 637
 638         /* Start the cleanup process */
 639         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 640                 lnet_net_lock(LNET_LOCK_EX);
 641                 lnet_peer_table_cleanup_locked(net, ptable);
 642                 lnet_net_unlock(LNET_LOCK_EX);
 643         }
 644
 645         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 646                 lnet_peer_ni_finalize_wait(ptable);
 647 }
 648
 649 static struct lnet_peer_ni *
 650 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 651 {
 652         struct list_head        *peers;
 653         struct lnet_peer_ni     *lp;
 654
 655         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 656                 return NULL;
 657
 658         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 659         list_for_each_entry(lp, peers, lpni_hashlist) {
 660                 if (lp->lpni_nid == nid) {
 661                         lnet_peer_ni_addref_locked(lp);
 662                         return lp;
 663                 }
 664         }
 665
 666         return NULL;
 667 }
 668
 669 struct lnet_peer_ni *
 670 lnet_find_peer_ni_locked(lnet_nid_t nid)
 671 {
 672         struct lnet_peer_ni *lpni;
 673         struct lnet_peer_table *ptable;
 674         int cpt;
 675
 676         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 677
 678         ptable = the_lnet.ln_peer_tables[cpt];
 679         lpni = lnet_get_peer_ni_locked(ptable, nid);
 680
 681         return lpni;
 682 }
 683
 684 struct lnet_peer_ni *
 685 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 686 {
 687         struct lnet_peer_net *lpn;
 688         struct lnet_peer_ni *lpni;
 689
 690         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 691         if (!lpn)
 692                 return NULL;
 693
 694         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 695                 if (lpni->lpni_nid == nid)
 696                         return lpni;
 697         }
 698
 699         return NULL;
 700 }
 701
 702 struct lnet_peer *
 703 lnet_find_peer(lnet_nid_t nid)
 704 {
 705         struct lnet_peer_ni *lpni;
 706         struct lnet_peer *lp = NULL;
 707         int cpt;
 708
 709         cpt = lnet_net_lock_current();
 710         lpni = lnet_find_peer_ni_locked(nid);
 711         if (lpni) {
 712                 lp = lpni->lpni_peer_net->lpn_peer;
 713                 lnet_peer_addref_locked(lp);
 714                 lnet_peer_ni_decref_locked(lpni);
 715         }
 716         lnet_net_unlock(cpt);
 717
 718         return lp;
 719 }
 720
 721 struct lnet_peer_net *
 722 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 723 {
 724         struct lnet_peer_net *net;
 725
 726         if (!prev_lpn_id) {
 727                 /* no net id provided return the first net */
 728                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 729                                                struct lnet_peer_net,
 730                                                lpn_peer_nets);
 731
 732                 return net;
 733         }
 734
 735         /* find the net after the one provided */
 736         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 737                 if (net->lpn_net_id == prev_lpn_id) {
 738                         /*
 739                          * if we reached the end of the list loop to the
 740                          * beginning.
 741                          */
 742                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 743                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 744                                                                 struct lnet_peer_net,
 745                                                                 lpn_peer_nets);
 746                         else
 747                                 return list_next_entry(net, lpn_peer_nets);
 748                 }
 749         }
 750
 751         return NULL;
 752 }
 753
 754 struct lnet_peer_ni *
 755 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 756                              struct lnet_peer_net *peer_net,
 757                              struct lnet_peer_ni *prev)
 758 {
 759         struct lnet_peer_ni *lpni;
 760         struct lnet_peer_net *net = peer_net;
 761
 762         if (!prev) {
 763                 if (!net) {
 764                         if (list_empty(&peer->lp_peer_nets))
 765                                 return NULL;
 766
 767                         net = list_entry(peer->lp_peer_nets.next,
 768                                          struct lnet_peer_net,
 769                                          lpn_peer_nets);
 770                 }
 771                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 772                                   lpni_peer_nis);
 773
 774                 return lpni;
 775         }
 776
 777         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 778                 /*
 779                  * if you reached the end of the peer ni list and the peer
 780                  * net is specified then there are no more peer nis in that
 781                  * net.
 782                  */
 783                 if (net)
 784                         return NULL;
 785
 786                 /*
 787                  * we reached the end of this net ni list. move to the
 788                  * next net
 789                  */
 790                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 791                     &peer->lp_peer_nets)
 792                         /* no more nets and no more NIs. */
 793                         return NULL;
 794
 795                 /* get the next net */
 796                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 797                                  struct lnet_peer_net,
 798                                  lpn_peer_nets);
 799                 /* get the ni on it */
 800                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 801                                   lpni_peer_nis);
 802
 803                 return lpni;
 804         }
 805
 806         /* there are more nis left */
 807         lpni = list_entry(prev->lpni_peer_nis.next,
 808                           struct lnet_peer_ni, lpni_peer_nis);
 809
 810         return lpni;
 811 }
 812
 813 /* Call with the ln_api_mutex held */
 814 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 815 {
 816         struct lnet_process_id id;
 817         struct lnet_peer_table *ptable;
 818         struct lnet_peer *lp;
 819         __u32 count = 0;
 820         __u32 size = 0;
 821         int lncpt;
 822         int cpt;
 823         __u32 i;
 824         int rc;
 825
 826         rc = -ESHUTDOWN;
 827         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 828                 goto done;
 829
 830         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 831
 832         /*
 833          * Count the number of peers, and return E2BIG if the buffer
 834          * is too small. We'll also return the desired size.
 835          */
 836         rc = -E2BIG;
 837         for (cpt = 0; cpt < lncpt; cpt++) {
 838                 ptable = the_lnet.ln_peer_tables[cpt];
 839                 count += ptable->pt_peers;
 840         }
 841         size = count * sizeof(*ids);
 842         if (size > *sizep)
 843                 goto done;
 844
 845         /*
 846          * Walk the peer lists and copy out the primary nids.
 847          * This is safe because the peer lists are only modified
 848          * while the ln_api_mutex is held. So we don't need to
 849          * hold the lnet_net_lock as well, and can therefore
 850          * directly call copy_to_user().
 851          */
 852         rc = -EFAULT;
 853         memset(&id, 0, sizeof(id));
 854         id.pid = LNET_PID_LUSTRE;
 855         i = 0;
 856         for (cpt = 0; cpt < lncpt; cpt++) {
 857                 ptable = the_lnet.ln_peer_tables[cpt];
 858                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 859                         if (i >= count)
 860                                 goto done;
 861                         id.nid = lp->lp_primary_nid;
 862                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 863                                 goto done;
 864                         i++;
 865                 }
 866         }
 867         rc = 0;
 868 done:
 869         *countp = count;
 870         *sizep = size;
 871         return rc;
 872 }
 873
 874 /*
 875  * Start pushes to peers that need to be updated for a configuration
 876  * change on this node.
 877  */
 878 void
 879 lnet_push_update_to_peers(int force)
 880 {
 881         struct lnet_peer_table *ptable;
 882         struct lnet_peer *lp;
 883         int lncpt;
 884         int cpt;
 885
 886         lnet_net_lock(LNET_LOCK_EX);
 887         if (lnet_peer_discovery_disabled)
 888                 force = 0;
 889         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 890         for (cpt = 0; cpt < lncpt; cpt++) {
 891                 ptable = the_lnet.ln_peer_tables[cpt];
 892                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 893                         if (force) {
 894                                 spin_lock(&lp->lp_lock);
 895                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 896                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 897                                 spin_unlock(&lp->lp_lock);
 898                         }
 899                         if (lnet_peer_needs_push(lp))
 900                                 lnet_peer_queue_for_discovery(lp);
 901                 }
 902         }
 903         lnet_net_unlock(LNET_LOCK_EX);
 904         wake_up(&the_lnet.ln_dc_waitq);
 905 }
 906
 907 /*
 908  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 909  * this is a preferred point-to-point path. Call with lnet_net_lock in
 910  * shared mmode.
 911  */
 912 bool
 913 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 914 {
 915         int i;
 916
 917         if (lpni->lpni_pref_nnids == 0)
 918                 return false;
 919         if (lpni->lpni_pref_nnids == 1)
 920                 return lpni->lpni_pref.nid == nid;
 921         for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 922                 if (lpni->lpni_pref.nids[i] == nid)
 923                         return true;
 924         }
 925         return false;
 926 }
 927
 928 /*
 929  * Set a single ni as preferred, provided no preferred ni is already
 930  * defined. Only to be used for non-multi-rail peer_ni.
 931  */
 932 int
 933 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 934 {
 935         int rc = 0;
 936
 937         spin_lock(&lpni->lpni_lock);
 938         if (nid == LNET_NID_ANY) {
 939                 rc = -EINVAL;
 940         } else if (lpni->lpni_pref_nnids > 0) {
 941                 rc = -EPERM;
 942         } else if (lpni->lpni_pref_nnids == 0) {
 943                 lpni->lpni_pref.nid = nid;
 944                 lpni->lpni_pref_nnids = 1;
 945                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
 946         }
 947         spin_unlock(&lpni->lpni_lock);
 948
 949         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 950                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
 951         return rc;
 952 }
 953
 954 /*
 955  * Clear the preferred NID from a non-multi-rail peer_ni, provided
 956  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
 957  */
 958 int
 959 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
 960 {
 961         int rc = 0;
 962
 963         spin_lock(&lpni->lpni_lock);
 964         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
 965                 lpni->lpni_pref_nnids = 0;
 966                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 967         } else if (lpni->lpni_pref_nnids == 0) {
 968                 rc = -ENOENT;
 969         } else {
 970                 rc = -EPERM;
 971         }
 972         spin_unlock(&lpni->lpni_lock);
 973
 974         CDEBUG(D_NET, "peer %s: %d\n",
 975                libcfs_nid2str(lpni->lpni_nid), rc);
 976         return rc;
 977 }
 978
 979 /*
 980  * Clear the preferred NIDs from a non-multi-rail peer.
 981  */
 982 void
 983 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
 984 {
 985         struct lnet_peer_ni *lpni = NULL;
 986
 987         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 988                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
 989 }
 990
 991 int
 992 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 993 {
 994         lnet_nid_t *nids = NULL;
 995         lnet_nid_t *oldnids = NULL;
 996         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 997         int size;
 998         int i;
 999         int rc = 0;
1000
1001         if (nid == LNET_NID_ANY) {
1002                 rc = -EINVAL;
1003                 goto out;
1004         }
1005
1006         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1007                 rc = -EEXIST;
1008                 goto out;
1009         }
1010
1011         /* A non-MR node may have only one preferred NI per peer_ni */
1012         if (lpni->lpni_pref_nnids > 0) {
1013                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1014                         rc = -EPERM;
1015                         goto out;
1016                 }
1017         }
1018
1019         if (lpni->lpni_pref_nnids != 0) {
1020                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1021                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1022                 if (!nids) {
1023                         rc = -ENOMEM;
1024                         goto out;
1025                 }
1026                 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1027                         if (lpni->lpni_pref.nids[i] == nid) {
1028                                 LIBCFS_FREE(nids, size);
1029                                 rc = -EEXIST;
1030                                 goto out;
1031                         }
1032                         nids[i] = lpni->lpni_pref.nids[i];
1033                 }
1034                 nids[i] = nid;
1035         }
1036
1037         lnet_net_lock(LNET_LOCK_EX);
1038         spin_lock(&lpni->lpni_lock);
1039         if (lpni->lpni_pref_nnids == 0) {
1040                 lpni->lpni_pref.nid = nid;
1041         } else {
1042                 oldnids = lpni->lpni_pref.nids;
1043                 lpni->lpni_pref.nids = nids;
1044         }
1045         lpni->lpni_pref_nnids++;
1046         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1047         spin_unlock(&lpni->lpni_lock);
1048         lnet_net_unlock(LNET_LOCK_EX);
1049
1050         if (oldnids) {
1051                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1052                 CFS_FREE_PTR_ARRAY(oldnids, size);
1053         }
1054 out:
1055         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1056                 spin_lock(&lpni->lpni_lock);
1057                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1058                 spin_unlock(&lpni->lpni_lock);
1059         }
1060         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1061                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1062         return rc;
1063 }
1064
1065 int
1066 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1067 {
1068         lnet_nid_t *nids = NULL;
1069         lnet_nid_t *oldnids = NULL;
1070         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1071         int size;
1072         int i, j;
1073         int rc = 0;
1074
1075         if (lpni->lpni_pref_nnids == 0) {
1076                 rc = -ENOENT;
1077                 goto out;
1078         }
1079
1080         if (lpni->lpni_pref_nnids == 1) {
1081                 if (lpni->lpni_pref.nid != nid) {
1082                         rc = -ENOENT;
1083                         goto out;
1084                 }
1085         } else if (lpni->lpni_pref_nnids == 2) {
1086                 if (lpni->lpni_pref.nids[0] != nid &&
1087                     lpni->lpni_pref.nids[1] != nid) {
1088                         rc = -ENOENT;
1089                         goto out;
1090                 }
1091         } else {
1092                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1093                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1094                 if (!nids) {
1095                         rc = -ENOMEM;
1096                         goto out;
1097                 }
1098                 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1099                         if (lpni->lpni_pref.nids[i] != nid)
1100                                 continue;
1101                         nids[j++] = lpni->lpni_pref.nids[i];
1102                 }
1103                 /* Check if we actually removed a nid. */
1104                 if (j == lpni->lpni_pref_nnids) {
1105                         LIBCFS_FREE(nids, size);
1106                         rc = -ENOENT;
1107                         goto out;
1108                 }
1109         }
1110
1111         lnet_net_lock(LNET_LOCK_EX);
1112         spin_lock(&lpni->lpni_lock);
1113         if (lpni->lpni_pref_nnids == 1) {
1114                 lpni->lpni_pref.nid = LNET_NID_ANY;
1115         } else if (lpni->lpni_pref_nnids == 2) {
1116                 oldnids = lpni->lpni_pref.nids;
1117                 if (oldnids[0] == nid)
1118                         lpni->lpni_pref.nid = oldnids[1];
1119                 else
1120                         lpni->lpni_pref.nid = oldnids[2];
1121         } else {
1122                 oldnids = lpni->lpni_pref.nids;
1123                 lpni->lpni_pref.nids = nids;
1124         }
1125         lpni->lpni_pref_nnids--;
1126         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1127         spin_unlock(&lpni->lpni_lock);
1128         lnet_net_unlock(LNET_LOCK_EX);
1129
1130         if (oldnids) {
1131                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1132                 CFS_FREE_PTR_ARRAY(oldnids, size);
1133         }
1134 out:
1135         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1136                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1137         return rc;
1138 }
1139
1140 lnet_nid_t
1141 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1142 {
1143         struct lnet_peer_ni *lpni;
1144         lnet_nid_t primary_nid = nid;
1145
1146         lpni = lnet_find_peer_ni_locked(nid);
1147         if (lpni) {
1148                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1149                 lnet_peer_ni_decref_locked(lpni);
1150         }
1151
1152         return primary_nid;
1153 }
1154
1155 bool
1156 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1157 __must_hold(&lp->lp_lock)
1158 {
1159         if (lnet_peer_discovery_disabled)
1160                 return true;
1161
1162         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1163             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1164                 return true;
1165         }
1166
1167         return false;
1168 }
1169
1170 /*
1171  * Peer Discovery
1172  */
1173 bool
1174 lnet_is_discovery_disabled(struct lnet_peer *lp)
1175 {
1176         bool rc = false;
1177
1178         spin_lock(&lp->lp_lock);
1179         rc = lnet_is_discovery_disabled_locked(lp);
1180         spin_unlock(&lp->lp_lock);
1181
1182         return rc;
1183 }
1184
1185 lnet_nid_t
1186 LNetPrimaryNID(lnet_nid_t nid)
1187 {
1188         struct lnet_peer *lp;
1189         struct lnet_peer_ni *lpni;
1190         lnet_nid_t primary_nid = nid;
1191         int rc = 0;
1192         int cpt;
1193
1194         cpt = lnet_net_lock_current();
1195         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1196         if (IS_ERR(lpni)) {
1197                 rc = PTR_ERR(lpni);
1198                 goto out_unlock;
1199         }
1200         lp = lpni->lpni_peer_net->lpn_peer;
1201
1202         while (!lnet_peer_is_uptodate(lp)) {
1203                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1204                 if (rc)
1205                         goto out_decref;
1206                 lp = lpni->lpni_peer_net->lpn_peer;
1207
1208                 /* Only try once if discovery is disabled */
1209                 if (lnet_is_discovery_disabled(lp))
1210                         break;
1211         }
1212         primary_nid = lp->lp_primary_nid;
1213 out_decref:
1214         lnet_peer_ni_decref_locked(lpni);
1215 out_unlock:
1216         lnet_net_unlock(cpt);
1217
1218         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1219                libcfs_nid2str(primary_nid), rc);
1220         return primary_nid;
1221 }
1222 EXPORT_SYMBOL(LNetPrimaryNID);
1223
1224 struct lnet_peer_net *
1225 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1226 {
1227         struct lnet_peer_net *peer_net;
1228         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1229                 if (peer_net->lpn_net_id == net_id)
1230                         return peer_net;
1231         }
1232         return NULL;
1233 }
1234
1235 /*
1236  * Attach a peer_ni to a peer_net and peer. This function assumes
1237  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1238  * may be attached to a different peer, in which case it will be
1239  * properly detached first. The whole operation is done atomically.
1240  *
1241  * Always returns 0.  This is the last function called from functions
1242  * that do return an int, so returning 0 here allows the compiler to
1243  * do a tail call.
1244  */
1245 static int
1246 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1247                                 struct lnet_peer_net *lpn,
1248                                 struct lnet_peer_ni *lpni,
1249                                 unsigned flags)
1250 {
1251         struct lnet_peer_table *ptable;
1252
1253         /* Install the new peer_ni */
1254         lnet_net_lock(LNET_LOCK_EX);
1255         /* Add peer_ni to global peer table hash, if necessary. */
1256         if (list_empty(&lpni->lpni_hashlist)) {
1257                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1258
1259                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1260                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1261                 ptable->pt_version++;
1262                 /* This is the 1st refcount on lpni. */
1263                 atomic_inc(&lpni->lpni_refcount);
1264         }
1265
1266         /* Detach the peer_ni from an existing peer, if necessary. */
1267         if (lpni->lpni_peer_net) {
1268                 LASSERT(lpni->lpni_peer_net != lpn);
1269                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1270                 lnet_peer_detach_peer_ni_locked(lpni);
1271                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1272                 lpni->lpni_peer_net = NULL;
1273         }
1274
1275         /* Add peer_ni to peer_net */
1276         lpni->lpni_peer_net = lpn;
1277         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1278         lnet_update_peer_net_healthv(lpni);
1279         lnet_peer_net_addref_locked(lpn);
1280
1281         /* Add peer_net to peer */
1282         if (!lpn->lpn_peer) {
1283                 lpn->lpn_peer = lp;
1284                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1285                 lnet_peer_addref_locked(lp);
1286         }
1287
1288         /* Add peer to global peer list, if necessary */
1289         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1290         if (list_empty(&lp->lp_peer_list)) {
1291                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1292                 ptable->pt_peers++;
1293         }
1294
1295
1296         /* Update peer state */
1297         spin_lock(&lp->lp_lock);
1298         if (flags & LNET_PEER_CONFIGURED) {
1299                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1300                         lp->lp_state |= LNET_PEER_CONFIGURED;
1301         }
1302         if (flags & LNET_PEER_MULTI_RAIL) {
1303                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1304                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1305                         lnet_peer_clr_non_mr_pref_nids(lp);
1306                 }
1307         }
1308         spin_unlock(&lp->lp_lock);
1309
1310         lp->lp_nnis++;
1311         lnet_net_unlock(LNET_LOCK_EX);
1312
1313         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1314                libcfs_nid2str(lp->lp_primary_nid),
1315                libcfs_nid2str(lpni->lpni_nid), flags);
1316
1317         return 0;
1318 }
1319
1320 /*
1321  * Create a new peer, with nid as its primary nid.
1322  *
1323  * Call with the lnet_api_mutex held.
1324  */
1325 static int
1326 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1327 {
1328         struct lnet_peer *lp;
1329         struct lnet_peer_net *lpn;
1330         struct lnet_peer_ni *lpni;
1331         int rc = 0;
1332
1333         LASSERT(nid != LNET_NID_ANY);
1334
1335         /*
1336          * No need for the lnet_net_lock here, because the
1337          * lnet_api_mutex is held.
1338          */
1339         lpni = lnet_find_peer_ni_locked(nid);
1340         if (lpni) {
1341                 /* A peer with this NID already exists. */
1342                 lp = lpni->lpni_peer_net->lpn_peer;
1343                 lnet_peer_ni_decref_locked(lpni);
1344                 /*
1345                  * This is an error if the peer was configured and the
1346                  * primary NID differs or an attempt is made to change
1347                  * the Multi-Rail flag. Otherwise the assumption is
1348                  * that an existing peer is being modified.
1349                  */
1350                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1351                         if (lp->lp_primary_nid != nid)
1352                                 rc = -EEXIST;
1353                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1354                                 rc = -EPERM;
1355                         goto out;
1356                 }
1357                 /* Delete and recreate as a configured peer. */
1358                 lnet_peer_del(lp);
1359         }
1360
1361         /* Create peer, peer_net, and peer_ni. */
1362         rc = -ENOMEM;
1363         lp = lnet_peer_alloc(nid);
1364         if (!lp)
1365                 goto out;
1366         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1367         if (!lpn)
1368                 goto out_free_lp;
1369         lpni = lnet_peer_ni_alloc(nid);
1370         if (!lpni)
1371                 goto out_free_lpn;
1372
1373         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1374
1375 out_free_lpn:
1376         LIBCFS_FREE(lpn, sizeof(*lpn));
1377 out_free_lp:
1378         LIBCFS_FREE(lp, sizeof(*lp));
1379 out:
1380         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1381                libcfs_nid2str(nid), flags, rc);
1382         return rc;
1383 }
1384
1385 /*
1386  * Add a NID to a peer. Call with ln_api_mutex held.
1387  *
1388  * Error codes:
1389  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1390  *  -EEXIST:   The NID was configured by DLC for a different peer.
1391  *  -ENOMEM:   Out of memory.
1392  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1393  *             non-multi-rail peer.
1394  */
1395 static int
1396 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1397 {
1398         struct lnet_peer_net *lpn;
1399         struct lnet_peer_ni *lpni;
1400         int rc = 0;
1401
1402         LASSERT(lp);
1403         LASSERT(nid != LNET_NID_ANY);
1404
1405         /* A configured peer can only be updated through configuration. */
1406         if (!(flags & LNET_PEER_CONFIGURED)) {
1407                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1408                         rc = -EPERM;
1409                         goto out;
1410                 }
1411         }
1412
1413         /*
1414          * The MULTI_RAIL flag can be set but not cleared, because
1415          * that would leave the peer struct in an invalid state.
1416          */
1417         if (flags & LNET_PEER_MULTI_RAIL) {
1418                 spin_lock(&lp->lp_lock);
1419                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1420                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1421                         lnet_peer_clr_non_mr_pref_nids(lp);
1422                 }
1423                 spin_unlock(&lp->lp_lock);
1424         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1425                 rc = -EPERM;
1426                 goto out;
1427         }
1428
1429         lpni = lnet_find_peer_ni_locked(nid);
1430         if (lpni) {
1431                 /*
1432                  * A peer_ni already exists. This is only a problem if
1433                  * it is not connected to this peer and was configured
1434                  * by DLC.
1435                  */
1436                 lnet_peer_ni_decref_locked(lpni);
1437                 if (lpni->lpni_peer_net->lpn_peer == lp)
1438                         goto out;
1439                 if (lnet_peer_ni_is_configured(lpni)) {
1440                         rc = -EEXIST;
1441                         goto out;
1442                 }
1443                 /* If this is the primary NID, destroy the peer. */
1444                 if (lnet_peer_ni_is_primary(lpni)) {
1445                         struct lnet_peer *rtr_lp =
1446                           lpni->lpni_peer_net->lpn_peer;
1447                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1448                         /*
1449                          * if we're trying to delete a router it means
1450                          * we're moving this peer NI to a new peer so must
1451                          * transfer router properties to the new peer
1452                          */
1453                         if (rtr_refcount > 0) {
1454                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1455                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1456                         }
1457                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1458                         lpni = lnet_peer_ni_alloc(nid);
1459                         if (!lpni) {
1460                                 rc = -ENOMEM;
1461                                 goto out;
1462                         }
1463                 }
1464         } else {
1465                 lpni = lnet_peer_ni_alloc(nid);
1466                 if (!lpni) {
1467                         rc = -ENOMEM;
1468                         goto out;
1469                 }
1470         }
1471
1472         /*
1473          * Get the peer_net. Check that we're not adding a second
1474          * peer_ni on a peer_net of a non-multi-rail peer.
1475          */
1476         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1477         if (!lpn) {
1478                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1479                 if (!lpn) {
1480                         rc = -ENOMEM;
1481                         goto out_free_lpni;
1482                 }
1483         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1484                 rc = -ENOTUNIQ;
1485                 goto out_free_lpni;
1486         }
1487
1488         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1489
1490 out_free_lpni:
1491         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1492         if (!lpni->lpni_peer_net)
1493                 LIBCFS_FREE(lpni, sizeof(*lpni));
1494 out:
1495         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1496                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1497                flags, rc);
1498         return rc;
1499 }
1500
1501 /*
1502  * Update the primary NID of a peer, if possible.
1503  *
1504  * Call with the lnet_api_mutex held.
1505  */
1506 static int
1507 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1508 {
1509         lnet_nid_t old = lp->lp_primary_nid;
1510         int rc = 0;
1511
1512         if (lp->lp_primary_nid == nid)
1513                 goto out;
1514         rc = lnet_peer_add_nid(lp, nid, flags);
1515         if (rc)
1516                 goto out;
1517         lp->lp_primary_nid = nid;
1518 out:
1519         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1520                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1521         return rc;
1522 }
1523
1524 /*
1525  * lpni creation initiated due to traffic either sending or receiving.
1526  */
1527 static int
1528 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1529 {
1530         struct lnet_peer *lp;
1531         struct lnet_peer_net *lpn;
1532         struct lnet_peer_ni *lpni;
1533         unsigned flags = 0;
1534         int rc = 0;
1535
1536         if (nid == LNET_NID_ANY) {
1537                 rc = -EINVAL;
1538                 goto out;
1539         }
1540
1541         /* lnet_net_lock is not needed here because ln_api_lock is held */
1542         lpni = lnet_find_peer_ni_locked(nid);
1543         if (lpni) {
1544                 /*
1545                  * We must have raced with another thread. Since we
1546                  * know next to nothing about a peer_ni created by
1547                  * traffic, we just assume everything is ok and
1548                  * return.
1549                  */
1550                 lnet_peer_ni_decref_locked(lpni);
1551                 goto out;
1552         }
1553
1554         /* Create peer, peer_net, and peer_ni. */
1555         rc = -ENOMEM;
1556         lp = lnet_peer_alloc(nid);
1557         if (!lp)
1558                 goto out;
1559         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1560         if (!lpn)
1561                 goto out_free_lp;
1562         lpni = lnet_peer_ni_alloc(nid);
1563         if (!lpni)
1564                 goto out_free_lpn;
1565         if (pref != LNET_NID_ANY)
1566                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1567
1568         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1569
1570 out_free_lpn:
1571         LIBCFS_FREE(lpn, sizeof(*lpn));
1572 out_free_lp:
1573         LIBCFS_FREE(lp, sizeof(*lp));
1574 out:
1575         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1576         return rc;
1577 }
1578
1579 /*
1580  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1581  *
1582  * This API handles the following combinations:
1583  *   Create a peer with its primary NI if only the prim_nid is provided
1584  *   Add a NID to a peer identified by the prim_nid. The peer identified
1585  *   by the prim_nid must already exist.
1586  *   The peer being created may be non-MR.
1587  *
1588  * The caller must hold ln_api_mutex. This prevents the peer from
1589  * being created/modified/deleted by a different thread.
1590  */
1591 int
1592 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1593 {
1594         struct lnet_peer *lp = NULL;
1595         struct lnet_peer_ni *lpni;
1596         unsigned flags;
1597
1598         /* The prim_nid must always be specified */
1599         if (prim_nid == LNET_NID_ANY)
1600                 return -EINVAL;
1601
1602         flags = LNET_PEER_CONFIGURED;
1603         if (mr)
1604                 flags |= LNET_PEER_MULTI_RAIL;
1605
1606         /*
1607          * If nid isn't specified, we must create a new peer with
1608          * prim_nid as its primary nid.
1609          */
1610         if (nid == LNET_NID_ANY)
1611                 return lnet_peer_add(prim_nid, flags);
1612
1613         /* Look up the prim_nid, which must exist. */
1614         lpni = lnet_find_peer_ni_locked(prim_nid);
1615         if (!lpni)
1616                 return -ENOENT;
1617         lnet_peer_ni_decref_locked(lpni);
1618         lp = lpni->lpni_peer_net->lpn_peer;
1619
1620         /* Peer must have been configured. */
1621         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1622                 CDEBUG(D_NET, "peer %s was not configured\n",
1623                        libcfs_nid2str(prim_nid));
1624                 return -ENOENT;
1625         }
1626
1627         /* Primary NID must match */
1628         if (lp->lp_primary_nid != prim_nid) {
1629                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1630                        libcfs_nid2str(prim_nid),
1631                        libcfs_nid2str(lp->lp_primary_nid));
1632                 return -ENODEV;
1633         }
1634
1635         /* Multi-Rail flag must match. */
1636         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1637                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1638                        libcfs_nid2str(prim_nid));
1639                 return -EPERM;
1640         }
1641
1642         return lnet_peer_add_nid(lp, nid, flags);
1643 }
1644
1645 /*
1646  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1647  *
1648  * This API handles the following combinations:
1649  *   Delete a NI from a peer if both prim_nid and nid are provided.
1650  *   Delete a peer if only prim_nid is provided.
1651  *   Delete a peer if its primary nid is provided.
1652  *
1653  * The caller must hold ln_api_mutex. This prevents the peer from
1654  * being modified/deleted by a different thread.
1655  */
1656 int
1657 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1658 {
1659         struct lnet_peer *lp;
1660         struct lnet_peer_ni *lpni;
1661         unsigned flags;
1662
1663         if (prim_nid == LNET_NID_ANY)
1664                 return -EINVAL;
1665
1666         lpni = lnet_find_peer_ni_locked(prim_nid);
1667         if (!lpni)
1668                 return -ENOENT;
1669         lnet_peer_ni_decref_locked(lpni);
1670         lp = lpni->lpni_peer_net->lpn_peer;
1671
1672         if (prim_nid != lp->lp_primary_nid) {
1673                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1674                        libcfs_nid2str(prim_nid),
1675                        libcfs_nid2str(lp->lp_primary_nid));
1676                 return -ENODEV;
1677         }
1678
1679         lnet_net_lock(LNET_LOCK_EX);
1680         if (lp->lp_rtr_refcount > 0) {
1681                 lnet_net_unlock(LNET_LOCK_EX);
1682                 CERROR("%s is a router. Can not be deleted\n",
1683                        libcfs_nid2str(prim_nid));
1684                 return -EBUSY;
1685         }
1686         lnet_net_unlock(LNET_LOCK_EX);
1687
1688         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1689                 return lnet_peer_del(lp);
1690
1691         flags = LNET_PEER_CONFIGURED;
1692         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1693                 flags |= LNET_PEER_MULTI_RAIL;
1694
1695         return lnet_peer_del_nid(lp, nid, flags);
1696 }
1697
1698 void
1699 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1700 {
1701         struct lnet_peer_table *ptable;
1702         struct lnet_peer_net *lpn;
1703
1704         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1705
1706         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1707         LASSERT(list_empty(&lpni->lpni_txq));
1708         LASSERT(lpni->lpni_txqnob == 0);
1709         LASSERT(list_empty(&lpni->lpni_peer_nis));
1710         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1711
1712         lpn = lpni->lpni_peer_net;
1713         lpni->lpni_peer_net = NULL;
1714         lpni->lpni_net = NULL;
1715
1716         /* remove the peer ni from the zombie list */
1717         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1718         spin_lock(&ptable->pt_zombie_lock);
1719         list_del_init(&lpni->lpni_hashlist);
1720         ptable->pt_zombies--;
1721         spin_unlock(&ptable->pt_zombie_lock);
1722
1723         if (lpni->lpni_pref_nnids > 1)
1724                 CFS_FREE_PTR_ARRAY(lpni->lpni_pref.nids, lpni->lpni_pref_nnids);
1725
1726         LIBCFS_FREE(lpni, sizeof(*lpni));
1727
1728         lnet_peer_net_decref_locked(lpn);
1729 }
1730
1731 struct lnet_peer_ni *
1732 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1733 {
1734         struct lnet_peer_ni *lpni = NULL;
1735         int rc;
1736
1737         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1738                 return ERR_PTR(-ESHUTDOWN);
1739
1740         /*
1741          * find if a peer_ni already exists.
1742          * If so then just return that.
1743          */
1744         lpni = lnet_find_peer_ni_locked(nid);
1745         if (lpni)
1746                 return lpni;
1747
1748         lnet_net_unlock(cpt);
1749
1750         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1751         if (rc) {
1752                 lpni = ERR_PTR(rc);
1753                 goto out_net_relock;
1754         }
1755
1756         lpni = lnet_find_peer_ni_locked(nid);
1757         LASSERT(lpni);
1758
1759 out_net_relock:
1760         lnet_net_lock(cpt);
1761
1762         return lpni;
1763 }
1764
1765 /*
1766  * Get a peer_ni for the given nid, create it if necessary. Takes a
1767  * hold on the peer_ni.
1768  */
1769 struct lnet_peer_ni *
1770 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1771 {
1772         struct lnet_peer_ni *lpni = NULL;
1773         int rc;
1774
1775         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1776                 return ERR_PTR(-ESHUTDOWN);
1777
1778         /*
1779          * find if a peer_ni already exists.
1780          * If so then just return that.
1781          */
1782         lpni = lnet_find_peer_ni_locked(nid);
1783         if (lpni)
1784                 return lpni;
1785
1786         /*
1787          * Slow path:
1788          * use the lnet_api_mutex to serialize the creation of the peer_ni
1789          * and the creation/deletion of the local ni/net. When a local ni is
1790          * created, if there exists a set of peer_nis on that network,
1791          * they need to be traversed and updated. When a local NI is
1792          * deleted, which could result in a network being deleted, then
1793          * all peer nis on that network need to be removed as well.
1794          *
1795          * Creation through traffic should also be serialized with
1796          * creation through DLC.
1797          */
1798         lnet_net_unlock(cpt);
1799         mutex_lock(&the_lnet.ln_api_mutex);
1800         /*
1801          * Shutdown is only set under the ln_api_lock, so a single
1802          * check here is sufficent.
1803          */
1804         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1805                 lpni = ERR_PTR(-ESHUTDOWN);
1806                 goto out_mutex_unlock;
1807         }
1808
1809         rc = lnet_peer_ni_traffic_add(nid, pref);
1810         if (rc) {
1811                 lpni = ERR_PTR(rc);
1812                 goto out_mutex_unlock;
1813         }
1814
1815         lpni = lnet_find_peer_ni_locked(nid);
1816         LASSERT(lpni);
1817
1818 out_mutex_unlock:
1819         mutex_unlock(&the_lnet.ln_api_mutex);
1820         lnet_net_lock(cpt);
1821
1822         /* Lock has been dropped, check again for shutdown. */
1823         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1824                 if (!IS_ERR(lpni))
1825                         lnet_peer_ni_decref_locked(lpni);
1826                 lpni = ERR_PTR(-ESHUTDOWN);
1827         }
1828
1829         return lpni;
1830 }
1831
1832 bool
1833 lnet_peer_gw_discovery(struct lnet_peer *lp)
1834 {
1835         bool rc = false;
1836
1837         spin_lock(&lp->lp_lock);
1838         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1839                 rc = true;
1840         spin_unlock(&lp->lp_lock);
1841
1842         return rc;
1843 }
1844
1845 bool
1846 lnet_peer_is_uptodate(struct lnet_peer *lp)
1847 {
1848         bool rc;
1849
1850         spin_lock(&lp->lp_lock);
1851         rc = lnet_peer_is_uptodate_locked(lp);
1852         spin_unlock(&lp->lp_lock);
1853         return rc;
1854 }
1855
1856 /*
1857  * Is a peer uptodate from the point of view of discovery?
1858  *
1859  * If it is currently being processed, obviously not.
1860  * A forced Ping or Push is also handled by the discovery thread.
1861  *
1862  * Otherwise look at whether the peer needs rediscovering.
1863  */
1864 bool
1865 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1866 __must_hold(&lp->lp_lock)
1867 {
1868         bool rc;
1869
1870         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1871                             LNET_PEER_FORCE_PING |
1872                             LNET_PEER_FORCE_PUSH)) {
1873                 rc = false;
1874         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1875                 rc = false;
1876         } else if (lnet_peer_needs_push(lp)) {
1877                 rc = false;
1878         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1879                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1880                         rc = true;
1881                 else
1882                         rc = false;
1883         } else {
1884                 rc = false;
1885         }
1886
1887         return rc;
1888 }
1889
1890 /*
1891  * Queue a peer for the attention of the discovery thread.  Call with
1892  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1893  * -EALREADY if the peer was already queued.
1894  */
1895 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1896 {
1897         int rc;
1898
1899         spin_lock(&lp->lp_lock);
1900         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1901                 lp->lp_state |= LNET_PEER_DISCOVERING;
1902         spin_unlock(&lp->lp_lock);
1903         if (list_empty(&lp->lp_dc_list)) {
1904                 lnet_peer_addref_locked(lp);
1905                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1906                 wake_up(&the_lnet.ln_dc_waitq);
1907                 rc = 0;
1908         } else {
1909                 rc = -EALREADY;
1910         }
1911
1912         CDEBUG(D_NET, "Queue peer %s: %d\n",
1913                libcfs_nid2str(lp->lp_primary_nid), rc);
1914
1915         return rc;
1916 }
1917
1918 /*
1919  * Discovery of a peer is complete. Wake all waiters on the peer.
1920  * Call with lnet_net_lock/EX held.
1921  */
1922 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1923 {
1924         struct lnet_msg *msg, *tmp;
1925         int rc = 0;
1926         LIST_HEAD(pending_msgs);
1927
1928         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1929                libcfs_nid2str(lp->lp_primary_nid));
1930
1931         list_del_init(&lp->lp_dc_list);
1932         spin_lock(&lp->lp_lock);
1933         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1934         spin_unlock(&lp->lp_lock);
1935         wake_up_all(&lp->lp_dc_waitq);
1936
1937         if (lp->lp_rtr_refcount > 0)
1938                 lnet_router_discovery_complete(lp);
1939
1940         lnet_net_unlock(LNET_LOCK_EX);
1941
1942         /* iterate through all pending messages and send them again */
1943         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1944                 list_del_init(&msg->msg_list);
1945                 if (lp->lp_dc_error) {
1946                         lnet_finalize(msg, lp->lp_dc_error);
1947                         continue;
1948                 }
1949
1950                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1951                        lnet_msgtyp2str(msg->msg_type),
1952                        libcfs_id2str(msg->msg_target));
1953                 rc = lnet_send(msg->msg_src_nid_param, msg,
1954                                msg->msg_rtr_nid_param);
1955                 if (rc < 0) {
1956                         CNETERR("Error sending %s to %s: %d\n",
1957                                lnet_msgtyp2str(msg->msg_type),
1958                                libcfs_id2str(msg->msg_target), rc);
1959                         lnet_finalize(msg, rc);
1960                 }
1961         }
1962         lnet_net_lock(LNET_LOCK_EX);
1963         lnet_peer_decref_locked(lp);
1964 }
1965
1966 /*
1967  * Handle inbound push.
1968  * Like any event handler, called with lnet_res_lock/CPT held.
1969  */
1970 void lnet_peer_push_event(struct lnet_event *ev)
1971 {
1972         struct lnet_ping_buffer *pbuf;
1973         struct lnet_peer *lp;
1974
1975         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start + ev->offset);
1976
1977         /* lnet_find_peer() adds a refcount */
1978         lp = lnet_find_peer(ev->source.nid);
1979         if (!lp) {
1980                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1981                        libcfs_nid2str(ev->initiator.nid),
1982                        libcfs_nid2str(ev->source.nid));
1983                 pbuf->pb_needs_post = true;
1984                 return;
1985         }
1986
1987         /* Ensure peer state remains consistent while we modify it. */
1988         spin_lock(&lp->lp_lock);
1989
1990         /*
1991          * If some kind of error happened the contents of the message
1992          * cannot be used. Clear the NIDS_UPTODATE and set the
1993          * FORCE_PING flag to trigger a ping.
1994          */
1995         if (ev->status) {
1996                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1997                 lp->lp_state |= LNET_PEER_FORCE_PING;
1998                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1999                        ev->status,
2000                        libcfs_nid2str(lp->lp_primary_nid),
2001                        libcfs_nid2str(ev->source.nid));
2002                 goto out;
2003         }
2004
2005         /*
2006          * A push with invalid or corrupted info. Clear the UPTODATE
2007          * flag to trigger a ping.
2008          */
2009         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2010                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2011                 lp->lp_state |= LNET_PEER_FORCE_PING;
2012                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2013                        libcfs_nid2str(lp->lp_primary_nid));
2014                 goto out;
2015         }
2016
2017         /*
2018          * Make sure we'll allocate the correct size ping buffer when
2019          * pinging the peer.
2020          */
2021         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2022                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2023
2024         /*
2025          * A non-Multi-Rail peer is not supposed to be capable of
2026          * sending a push.
2027          */
2028         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2029                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2030                        libcfs_nid2str(lp->lp_primary_nid));
2031                 goto out;
2032         }
2033
2034         /*
2035          * Check the MULTIRAIL flag. Complain if the peer was DLC
2036          * configured without it.
2037          */
2038         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2039                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2040                         CERROR("Push says %s is Multi-Rail, DLC says not\n",
2041                                libcfs_nid2str(lp->lp_primary_nid));
2042                 } else {
2043                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2044                         lnet_peer_clr_non_mr_pref_nids(lp);
2045                 }
2046         }
2047
2048         /*
2049          * The peer may have discovery disabled at its end. Set
2050          * NO_DISCOVERY as appropriate.
2051          */
2052         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2053                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2054                        libcfs_nid2str(lp->lp_primary_nid));
2055                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2056         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2057                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2058                        libcfs_nid2str(lp->lp_primary_nid));
2059                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2060         }
2061
2062         /*
2063          * Check for truncation of the Put message. Clear the
2064          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2065          * and tell discovery to allocate a bigger buffer.
2066          */
2067         if (ev->mlength < ev->rlength) {
2068                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2069                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2070                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2071                 lp->lp_state |= LNET_PEER_FORCE_PING;
2072                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2073                        libcfs_nid2str(lp->lp_primary_nid),
2074                        pbuf->pb_info.pi_nnis);
2075                 goto out;
2076         }
2077
2078         /* always assume new data */
2079         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2080         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2081
2082         /*
2083          * If there is data present that hasn't been processed yet,
2084          * we'll replace it if the Put contained newer data and it
2085          * fits. We're racing with a Ping or earlier Push in this
2086          * case.
2087          */
2088         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2089                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2090                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2091                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2092                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2093                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2094                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2095                               libcfs_nid2str(lp->lp_primary_nid),
2096                               LNET_PING_BUFFER_SEQNO(pbuf),
2097                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2098                 }
2099                 goto out;
2100         }
2101
2102         /*
2103          * Allocate a buffer to copy the data. On a failure we drop
2104          * the Push and set FORCE_PING to force the discovery
2105          * thread to fix the problem by pinging the peer.
2106          */
2107         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2108         if (!lp->lp_data) {
2109                 lp->lp_state |= LNET_PEER_FORCE_PING;
2110                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2111                        libcfs_nid2str(lp->lp_primary_nid),
2112                        LNET_PING_BUFFER_SEQNO(pbuf));
2113                 goto out;
2114         }
2115
2116         /* Success */
2117         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2118                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2119         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2120         CDEBUG(D_NET, "Received Push %s %u\n",
2121                libcfs_nid2str(lp->lp_primary_nid),
2122                LNET_PING_BUFFER_SEQNO(pbuf));
2123
2124 out:
2125         /* We've processed this buffer. It can be reposted */
2126         pbuf->pb_needs_post = true;
2127
2128         /*
2129          * Queue the peer for discovery if not done, force it on the request
2130          * queue and wake the discovery thread if the peer was already queued,
2131          * because its status changed.
2132          */
2133         spin_unlock(&lp->lp_lock);
2134         lnet_net_lock(LNET_LOCK_EX);
2135         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2136                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2137                 wake_up(&the_lnet.ln_dc_waitq);
2138         }
2139         /* Drop refcount from lookup */
2140         lnet_peer_decref_locked(lp);
2141         lnet_net_unlock(LNET_LOCK_EX);
2142 }
2143
2144 /*
2145  * Clear the discovery error state, unless we're already discovering
2146  * this peer, in which case the error is current.
2147  */
2148 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2149 {
2150         spin_lock(&lp->lp_lock);
2151         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2152                 lp->lp_dc_error = 0;
2153         spin_unlock(&lp->lp_lock);
2154 }
2155
2156 /*
2157  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2158  * dropped/retaken within this function. An lnet_peer_ni is passed in
2159  * because discovery could tear down an lnet_peer.
2160  */
2161 int
2162 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2163 {
2164         DEFINE_WAIT(wait);
2165         struct lnet_peer *lp;
2166         int rc = 0;
2167         int count = 0;
2168
2169 again:
2170         lnet_net_unlock(cpt);
2171         lnet_net_lock(LNET_LOCK_EX);
2172         lp = lpni->lpni_peer_net->lpn_peer;
2173         lnet_peer_clear_discovery_error(lp);
2174
2175         /*
2176          * We're willing to be interrupted. The lpni can become a
2177          * zombie if we race with DLC, so we must check for that.
2178          */
2179         for (;;) {
2180                 /* Keep lp alive when the lnet_net_lock is unlocked */
2181                 lnet_peer_addref_locked(lp);
2182                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2183                 if (signal_pending(current))
2184                         break;
2185                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2186                         break;
2187                 /*
2188                  * Don't repeat discovery if discovery is disabled. This is
2189                  * done to ensure we can use discovery as a standard ping as
2190                  * well for backwards compatibility with routers which do not
2191                  * have discovery or have discovery disabled
2192                  */
2193                 if (lnet_is_discovery_disabled(lp) && count > 0)
2194                         break;
2195                 if (lp->lp_dc_error)
2196                         break;
2197                 if (lnet_peer_is_uptodate(lp))
2198                         break;
2199                 lnet_peer_queue_for_discovery(lp);
2200                 count++;
2201                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2202
2203                 /*
2204                  * If caller requested a non-blocking operation then
2205                  * return immediately. Once discovery is complete any
2206                  * pending messages that were stopped due to discovery
2207                  * will be transmitted.
2208                  */
2209                 if (!block)
2210                         break;
2211
2212                 lnet_net_unlock(LNET_LOCK_EX);
2213                 schedule();
2214                 finish_wait(&lp->lp_dc_waitq, &wait);
2215                 lnet_net_lock(LNET_LOCK_EX);
2216                 lnet_peer_decref_locked(lp);
2217                 /* Peer may have changed */
2218                 lp = lpni->lpni_peer_net->lpn_peer;
2219         }
2220         finish_wait(&lp->lp_dc_waitq, &wait);
2221
2222         lnet_net_unlock(LNET_LOCK_EX);
2223         lnet_net_lock(cpt);
2224         lnet_peer_decref_locked(lp);
2225         /*
2226          * The peer may have changed, so re-check and rediscover if that turns
2227          * out to have been the case. The reference count on lp ensured that
2228          * even if it was unlinked from lpni the memory could not be recycled.
2229          * Thus the check below is sufficient to determine whether the peer
2230          * changed. If the peer changed, then lp must not be dereferenced.
2231          */
2232         if (lp != lpni->lpni_peer_net->lpn_peer)
2233                 goto again;
2234
2235         if (signal_pending(current))
2236                 rc = -EINTR;
2237         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2238                 rc = -ESHUTDOWN;
2239         else if (lp->lp_dc_error)
2240                 rc = lp->lp_dc_error;
2241         else if (!block)
2242                 CDEBUG(D_NET, "non-blocking discovery\n");
2243         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2244                 goto again;
2245
2246         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2247                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2248                libcfs_nid2str(lpni->lpni_nid), rc,
2249                (!block) ? "pending discovery" : "discovery complete");
2250
2251         return rc;
2252 }
2253
2254 /* Handle an incoming ack for a push. */
2255 static void
2256 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2257 {
2258         struct lnet_ping_buffer *pbuf;
2259
2260         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2261         spin_lock(&lp->lp_lock);
2262         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2263         lp->lp_push_error = ev->status;
2264         if (ev->status)
2265                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2266         else
2267                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2268         spin_unlock(&lp->lp_lock);
2269
2270         CDEBUG(D_NET, "peer %s ev->status %d\n",
2271                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2272 }
2273
2274 /* Handle a Reply message. This is the reply to a Ping message. */
2275 static void
2276 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2277 {
2278         struct lnet_ping_buffer *pbuf;
2279         int rc;
2280
2281         spin_lock(&lp->lp_lock);
2282
2283         /*
2284          * If some kind of error happened the contents of message
2285          * cannot be used. Set PING_FAILED to trigger a retry.
2286          */
2287         if (ev->status) {
2288                 lp->lp_state |= LNET_PEER_PING_FAILED;
2289                 lp->lp_ping_error = ev->status;
2290                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2291                        ev->status,
2292                        libcfs_nid2str(lp->lp_primary_nid),
2293                        libcfs_nid2str(ev->source.nid));
2294                 goto out;
2295         }
2296
2297         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2298         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2299                 lnet_swap_pinginfo(pbuf);
2300
2301         /*
2302          * A reply with invalid or corrupted info. Set PING_FAILED to
2303          * trigger a retry.
2304          */
2305         rc = lnet_ping_info_validate(&pbuf->pb_info);
2306         if (rc) {
2307                 lp->lp_state |= LNET_PEER_PING_FAILED;
2308                 lp->lp_ping_error = 0;
2309                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2310                        libcfs_nid2str(lp->lp_primary_nid), rc);
2311                 goto out;
2312         }
2313
2314
2315         /*
2316          * The peer may have discovery disabled at its end. Set
2317          * NO_DISCOVERY as appropriate.
2318          */
2319         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2320             !lnet_peer_discovery_disabled) {
2321                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2322                        libcfs_nid2str(lp->lp_primary_nid));
2323                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2324         } else {
2325                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2326                        libcfs_nid2str(lp->lp_primary_nid));
2327                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2328         }
2329
2330         /*
2331          * Update the MULTI_RAIL flag based on the reply. If the peer
2332          * was configured with DLC then the setting should match what
2333          * DLC put in.
2334          */
2335         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2336                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2337                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2338                                libcfs_nid2str(lp->lp_primary_nid), lp);
2339                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2340                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2341                               libcfs_nid2str(lp->lp_primary_nid));
2342                 } else if (lnet_peer_discovery_disabled) {
2343                         CDEBUG(D_NET,
2344                                "peer %s(%p) not MR: DD disabled locally\n",
2345                                libcfs_nid2str(lp->lp_primary_nid), lp);
2346                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2347                         CDEBUG(D_NET,
2348                                "peer %s(%p) not MR: DD disabled remotely\n",
2349                                libcfs_nid2str(lp->lp_primary_nid), lp);
2350                 } else {
2351                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2352                                libcfs_nid2str(lp->lp_primary_nid), lp);
2353                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2354                         lnet_peer_clr_non_mr_pref_nids(lp);
2355                 }
2356         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2357                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2358                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2359                               libcfs_nid2str(lp->lp_primary_nid));
2360                 } else {
2361                         CERROR("Multi-Rail state vanished from %s\n",
2362                                libcfs_nid2str(lp->lp_primary_nid));
2363                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2364                 }
2365         }
2366
2367         /*
2368          * Make sure we'll allocate the correct size ping buffer when
2369          * pinging the peer.
2370          */
2371         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2372                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2373
2374         /*
2375          * Check for truncation of the Reply. Clear PING_SENT and set
2376          * PING_FAILED to trigger a retry.
2377          */
2378         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2379                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2380                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2381                 lp->lp_state |= LNET_PEER_PING_FAILED;
2382                 lp->lp_ping_error = 0;
2383                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2384                        libcfs_nid2str(lp->lp_primary_nid),
2385                        pbuf->pb_info.pi_nnis);
2386                 goto out;
2387         }
2388
2389         /*
2390          * Check the sequence numbers in the reply. These are only
2391          * available if the reply came from a Multi-Rail peer.
2392          */
2393         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2394             pbuf->pb_info.pi_nnis > 1 &&
2395             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2396                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2397                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2398                                 libcfs_nid2str(lp->lp_primary_nid),
2399                                 LNET_PING_BUFFER_SEQNO(pbuf),
2400                                 lp->lp_peer_seqno);
2401
2402                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2403         }
2404
2405         /* We're happy with the state of the data in the buffer. */
2406         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2407                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2408         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2409                 lnet_ping_buffer_decref(lp->lp_data);
2410         else
2411                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2412         lnet_ping_buffer_addref(pbuf);
2413         lp->lp_data = pbuf;
2414 out:
2415         lp->lp_state &= ~LNET_PEER_PING_SENT;
2416         spin_unlock(&lp->lp_lock);
2417
2418         lnet_net_lock(LNET_LOCK_EX);
2419         /*
2420          * If this peer is a gateway, call the routing callback to
2421          * handle the ping reply
2422          */
2423         if (lp->lp_rtr_refcount > 0)
2424                 lnet_router_discovery_ping_reply(lp);
2425         lnet_net_unlock(LNET_LOCK_EX);
2426 }
2427
2428 /*
2429  * Send event handling. Only matters for error cases, where we clean
2430  * up state on the peer and peer_ni that would otherwise be updated in
2431  * the REPLY event handler for a successful Ping, and the ACK event
2432  * handler for a successful Push.
2433  */
2434 static int
2435 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2436 {
2437         int rc = 0;
2438
2439         if (!ev->status)
2440                 goto out;
2441
2442         spin_lock(&lp->lp_lock);
2443         if (ev->msg_type == LNET_MSG_GET) {
2444                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2445                 lp->lp_state |= LNET_PEER_PING_FAILED;
2446                 lp->lp_ping_error = ev->status;
2447         } else { /* ev->msg_type == LNET_MSG_PUT */
2448                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2449                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2450                 lp->lp_push_error = ev->status;
2451         }
2452         spin_unlock(&lp->lp_lock);
2453         rc = LNET_REDISCOVER_PEER;
2454 out:
2455         CDEBUG(D_NET, "%s Send to %s: %d\n",
2456                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2457                 libcfs_nid2str(ev->target.nid), rc);
2458         return rc;
2459 }
2460
2461 /*
2462  * Unlink event handling. This event is only seen if a call to
2463  * LNetMDUnlink() caused the event to be unlinked. If this call was
2464  * made after the event was set up in LNetGet() or LNetPut() then we
2465  * assume the Ping or Push timed out.
2466  */
2467 static void
2468 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2469 {
2470         spin_lock(&lp->lp_lock);
2471         /* We've passed through LNetGet() */
2472         if (lp->lp_state & LNET_PEER_PING_SENT) {
2473                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2474                 lp->lp_state |= LNET_PEER_PING_FAILED;
2475                 lp->lp_ping_error = -ETIMEDOUT;
2476                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2477                         libcfs_nid2str(lp->lp_primary_nid));
2478         }
2479         /* We've passed through LNetPut() */
2480         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2481                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2482                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2483                 lp->lp_push_error = -ETIMEDOUT;
2484                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2485                         libcfs_nid2str(lp->lp_primary_nid));
2486         }
2487         spin_unlock(&lp->lp_lock);
2488 }
2489
2490 /*
2491  * Event handler for the discovery EQ.
2492  *
2493  * Called with lnet_res_lock(cpt) held. The cpt is the
2494  * lnet_cpt_of_cookie() of the md handle cookie.
2495  */
2496 static void lnet_discovery_event_handler(struct lnet_event *event)
2497 {
2498         struct lnet_peer *lp = event->md.user_ptr;
2499         struct lnet_ping_buffer *pbuf;
2500         int rc;
2501
2502         /* discovery needs to take another look */
2503         rc = LNET_REDISCOVER_PEER;
2504
2505         CDEBUG(D_NET, "Received event: %d\n", event->type);
2506
2507         switch (event->type) {
2508         case LNET_EVENT_ACK:
2509                 lnet_discovery_event_ack(lp, event);
2510                 break;
2511         case LNET_EVENT_REPLY:
2512                 lnet_discovery_event_reply(lp, event);
2513                 break;
2514         case LNET_EVENT_SEND:
2515                 /* Only send failure triggers a retry. */
2516                 rc = lnet_discovery_event_send(lp, event);
2517                 break;
2518         case LNET_EVENT_UNLINK:
2519                 /* LNetMDUnlink() was called */
2520                 lnet_discovery_event_unlink(lp, event);
2521                 break;
2522         default:
2523                 /* Invalid events. */
2524                 LBUG();
2525         }
2526         lnet_net_lock(LNET_LOCK_EX);
2527         if (event->unlinked) {
2528                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2529                 lnet_ping_buffer_decref(pbuf);
2530                 lnet_peer_decref_locked(lp);
2531         }
2532
2533         /* put peer back at end of request queue, if discovery not already
2534          * done */
2535         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2536                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2537                 wake_up(&the_lnet.ln_dc_waitq);
2538         }
2539         lnet_net_unlock(LNET_LOCK_EX);
2540 }
2541
2542 /*
2543  * Build a peer from incoming data.
2544  *
2545  * The NIDs in the incoming data are supposed to be structured as follows:
2546  *  - loopback
2547  *  - primary NID
2548  *  - other NIDs in same net
2549  *  - NIDs in second net
2550  *  - NIDs in third net
2551  *  - ...
2552  * This due to the way the list of NIDs in the data is created.
2553  *
2554  * Note that this function will mark the peer uptodate unless an
2555  * ENOMEM is encontered. All other errors are due to a conflict
2556  * between the DLC configuration and what discovery sees. We treat DLC
2557  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2558  * peer from becoming stuck in discovery.
2559  */
2560 static int lnet_peer_merge_data(struct lnet_peer *lp,
2561                                 struct lnet_ping_buffer *pbuf)
2562 {
2563         struct lnet_peer_ni *lpni;
2564         lnet_nid_t *curnis = NULL;
2565         struct lnet_ni_status *addnis = NULL;
2566         lnet_nid_t *delnis = NULL;
2567         unsigned flags;
2568         int ncurnis;
2569         int naddnis;
2570         int ndelnis;
2571         int nnis = 0;
2572         int i;
2573         int j;
2574         int rc;
2575
2576         flags = LNET_PEER_DISCOVERED;
2577         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2578                 flags |= LNET_PEER_MULTI_RAIL;
2579
2580         /*
2581          * Cache the routing feature for the peer; whether it is enabled
2582          * for disabled as reported by the remote peer.
2583          */
2584         spin_lock(&lp->lp_lock);
2585         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2586                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2587         else
2588                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2589         spin_unlock(&lp->lp_lock);
2590
2591         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2592         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2593         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2594         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2595         if (!curnis || !addnis || !delnis) {
2596                 rc = -ENOMEM;
2597                 goto out;
2598         }
2599         ncurnis = 0;
2600         naddnis = 0;
2601         ndelnis = 0;
2602
2603         /* Construct the list of NIDs present in peer. */
2604         lpni = NULL;
2605         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2606                 curnis[ncurnis++] = lpni->lpni_nid;
2607
2608         /*
2609          * Check for NIDs in pbuf not present in curnis[].
2610          * The loop starts at 1 to skip the loopback NID.
2611          */
2612         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2613                 for (j = 0; j < ncurnis; j++)
2614                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2615                                 break;
2616                 if (j == ncurnis)
2617                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2618         }
2619         /*
2620          * Check for NIDs in curnis[] not present in pbuf.
2621          * The nested loop starts at 1 to skip the loopback NID.
2622          *
2623          * But never add the loopback NID to delnis[]: if it is
2624          * present in curnis[] then this peer is for this node.
2625          */
2626         for (i = 0; i < ncurnis; i++) {
2627                 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2628                         continue;
2629                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2630                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2631                                 /*
2632                                  * update the information we cache for the
2633                                  * peer with the latest information we
2634                                  * received
2635                                  */
2636                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2637                                 if (lpni) {
2638                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2639                                         lnet_peer_ni_decref_locked(lpni);
2640                                 }
2641                                 break;
2642                         }
2643                 }
2644                 if (j == pbuf->pb_info.pi_nnis)
2645                         delnis[ndelnis++] = curnis[i];
2646         }
2647
2648         /*
2649          * If we get here and the discovery is disabled then we don't want
2650          * to add or delete any NIs. We just updated the ones we have some
2651          * information on, and call it a day
2652          */
2653         rc = 0;
2654         if (lnet_is_discovery_disabled(lp))
2655                 goto out;
2656
2657         for (i = 0; i < naddnis; i++) {
2658                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2659                 if (rc) {
2660                         CERROR("Error adding NID %s to peer %s: %d\n",
2661                                libcfs_nid2str(addnis[i].ns_nid),
2662                                libcfs_nid2str(lp->lp_primary_nid), rc);
2663                         if (rc == -ENOMEM)
2664                                 goto out;
2665                 }
2666                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2667                 if (lpni) {
2668                         lpni->lpni_ns_status = addnis[i].ns_status;
2669                         lnet_peer_ni_decref_locked(lpni);
2670                 }
2671         }
2672
2673         for (i = 0; i < ndelnis; i++) {
2674                 /*
2675                  * for routers it's okay to delete the primary_nid because
2676                  * the upper layers don't really rely on it. So if we're
2677                  * being told that the router changed its primary_nid
2678                  * then it's okay to delete it.
2679                  */
2680                 if (lp->lp_rtr_refcount > 0)
2681                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2682                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2683                 if (rc) {
2684                         CERROR("Error deleting NID %s from peer %s: %d\n",
2685                                libcfs_nid2str(delnis[i]),
2686                                libcfs_nid2str(lp->lp_primary_nid), rc);
2687                         if (rc == -ENOMEM)
2688                                 goto out;
2689                 }
2690         }
2691         /*
2692          * Errors other than -ENOMEM are due to peers having been
2693          * configured with DLC. Ignore these because DLC overrides
2694          * Discovery.
2695          */
2696         rc = 0;
2697 out:
2698         CFS_FREE_PTR_ARRAY(curnis, nnis);
2699         CFS_FREE_PTR_ARRAY(addnis, nnis);
2700         CFS_FREE_PTR_ARRAY(delnis, nnis);
2701         lnet_ping_buffer_decref(pbuf);
2702         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2703
2704         if (rc) {
2705                 spin_lock(&lp->lp_lock);
2706                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2707                 lp->lp_state |= LNET_PEER_FORCE_PING;
2708                 spin_unlock(&lp->lp_lock);
2709         }
2710         return rc;
2711 }
2712
2713 /*
2714  * The data in pbuf says lp is its primary peer, but the data was
2715  * received by a different peer. Try to update lp with the data.
2716  */
2717 static int
2718 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2719 {
2720         struct lnet_handle_md mdh;
2721
2722         /* Queue lp for discovery, and force it on the request queue. */
2723         lnet_net_lock(LNET_LOCK_EX);
2724         if (lnet_peer_queue_for_discovery(lp))
2725                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2726         lnet_net_unlock(LNET_LOCK_EX);
2727
2728         LNetInvalidateMDHandle(&mdh);
2729
2730         /*
2731          * Decide whether we can move the peer to the DATA_PRESENT state.
2732          *
2733          * We replace stale data for a multi-rail peer, repair PING_FAILED
2734          * status, and preempt FORCE_PING.
2735          *
2736          * If after that we have DATA_PRESENT, we merge it into this peer.
2737          */
2738         spin_lock(&lp->lp_lock);
2739         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2740                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2741                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2742                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2743                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2744                         lnet_ping_buffer_decref(pbuf);
2745                         pbuf = lp->lp_data;
2746                         lp->lp_data = NULL;
2747                 }
2748         }
2749         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2750                 lnet_ping_buffer_decref(lp->lp_data);
2751                 lp->lp_data = NULL;
2752                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2753         }
2754         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2755                 mdh = lp->lp_ping_mdh;
2756                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2757                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2758                 lp->lp_ping_error = 0;
2759         }
2760         if (lp->lp_state & LNET_PEER_FORCE_PING)
2761                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2762         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2763         spin_unlock(&lp->lp_lock);
2764
2765         if (!LNetMDHandleIsInvalid(mdh))
2766                 LNetMDUnlink(mdh);
2767
2768         if (pbuf)
2769                 return lnet_peer_merge_data(lp, pbuf);
2770
2771         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2772         return 0;
2773 }
2774
2775 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2776 {
2777         int i;
2778
2779         for (i = 0; i < pinfo->pi_nnis; i++) {
2780                 if (pinfo->pi_ni[i].ns_nid == nid)
2781                         return true;
2782         }
2783
2784         return false;
2785 }
2786
2787 /*
2788  * Update a peer using the data received.
2789  */
2790 static int lnet_peer_data_present(struct lnet_peer *lp)
2791 __must_hold(&lp->lp_lock)
2792 {
2793         struct lnet_ping_buffer *pbuf;
2794         struct lnet_peer_ni *lpni;
2795         lnet_nid_t nid = LNET_NID_ANY;
2796         unsigned flags;
2797         int rc = 0;
2798
2799         pbuf = lp->lp_data;
2800         lp->lp_data = NULL;
2801         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2802         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2803         spin_unlock(&lp->lp_lock);
2804
2805         /*
2806          * Modifications of peer structures are done while holding the
2807          * ln_api_mutex. A global lock is required because we may be
2808          * modifying multiple peer structures, and a mutex greatly
2809          * simplifies memory management.
2810          *
2811          * The actual changes to the data structures must also protect
2812          * against concurrent lookups, for which the lnet_net_lock in
2813          * LNET_LOCK_EX mode is used.
2814          */
2815         mutex_lock(&the_lnet.ln_api_mutex);
2816         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2817                 rc = -ESHUTDOWN;
2818                 goto out;
2819         }
2820
2821         /*
2822          * If this peer is not on the peer list then it is being torn
2823          * down, and our reference count may be all that is keeping it
2824          * alive. Don't do any work on it.
2825          */
2826         if (list_empty(&lp->lp_peer_list))
2827                 goto out;
2828
2829         flags = LNET_PEER_DISCOVERED;
2830         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2831                 flags |= LNET_PEER_MULTI_RAIL;
2832
2833         /*
2834          * Check whether the primary NID in the message matches the
2835          * primary NID of the peer. If it does, update the peer, if
2836          * it it does not, check whether there is already a peer with
2837          * that primary NID. If no such peer exists, try to update
2838          * the primary NID of the current peer (allowed if it was
2839          * created due to message traffic) and complete the update.
2840          * If the peer did exist, hand off the data to it.
2841          *
2842          * The peer for the loopback interface is a special case: this
2843          * is the peer for the local node, and we want to set its
2844          * primary NID to the correct value here. Moreover, this peer
2845          * can show up with only the loopback NID in the ping buffer.
2846          */
2847         if (pbuf->pb_info.pi_nnis <= 1)
2848                 goto out;
2849         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2850         if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2851                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2852                 if (!rc)
2853                         rc = lnet_peer_merge_data(lp, pbuf);
2854         /*
2855          * if the primary nid of the peer is present in the ping info returned
2856          * from the peer, but it's not the local primary peer we have
2857          * cached and discovery is disabled, then we don't want to update
2858          * our local peer info, by adding or removing NIDs, we just want
2859          * to update the status of the nids that we currently have
2860          * recorded in that peer.
2861          */
2862         } else if (lp->lp_primary_nid == nid ||
2863                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2864                     lnet_is_discovery_disabled(lp))) {
2865                 rc = lnet_peer_merge_data(lp, pbuf);
2866         } else {
2867                 lpni = lnet_find_peer_ni_locked(nid);
2868                 if (!lpni) {
2869                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2870                         if (rc) {
2871                                 CERROR("Primary NID error %s versus %s: %d\n",
2872                                        libcfs_nid2str(lp->lp_primary_nid),
2873                                        libcfs_nid2str(nid), rc);
2874                         } else {
2875                                 rc = lnet_peer_merge_data(lp, pbuf);
2876                         }
2877                 } else {
2878                         struct lnet_peer *new_lp;
2879                         new_lp = lpni->lpni_peer_net->lpn_peer;
2880                         /*
2881                          * if lp has discovery/MR enabled that means new_lp
2882                          * should have discovery/MR enabled as well, since
2883                          * it's the same peer, which we're about to merge
2884                          */
2885                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2886                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2887                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2888                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2889
2890                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2891                         lnet_consolidate_routes_locked(lp, new_lp);
2892                         lnet_peer_ni_decref_locked(lpni);
2893                 }
2894         }
2895 out:
2896         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2897                lp->lp_state);
2898         mutex_unlock(&the_lnet.ln_api_mutex);
2899
2900         spin_lock(&lp->lp_lock);
2901         /* Tell discovery to re-check the peer immediately. */
2902         if (!rc)
2903                 rc = LNET_REDISCOVER_PEER;
2904         return rc;
2905 }
2906
2907 /*
2908  * A ping failed. Clear the PING_FAILED state and set the
2909  * FORCE_PING state, to ensure a retry even if discovery is
2910  * disabled. This avoids being left with incorrect state.
2911  */
2912 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2913 __must_hold(&lp->lp_lock)
2914 {
2915         struct lnet_handle_md mdh;
2916         int rc;
2917
2918         mdh = lp->lp_ping_mdh;
2919         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2920         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2921         lp->lp_state |= LNET_PEER_FORCE_PING;
2922         rc = lp->lp_ping_error;
2923         lp->lp_ping_error = 0;
2924         spin_unlock(&lp->lp_lock);
2925
2926         if (!LNetMDHandleIsInvalid(mdh))
2927                 LNetMDUnlink(mdh);
2928
2929         CDEBUG(D_NET, "peer %s:%d\n",
2930                libcfs_nid2str(lp->lp_primary_nid), rc);
2931
2932         spin_lock(&lp->lp_lock);
2933         return rc ? rc : LNET_REDISCOVER_PEER;
2934 }
2935
2936 /*
2937  * Select NID to send a Ping or Push to.
2938  */
2939 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2940 {
2941         struct lnet_peer_ni *lpni;
2942
2943         /* Look for a direct-connected NID for this peer. */
2944         lpni = NULL;
2945         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2946                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2947                         continue;
2948                 break;
2949         }
2950         if (lpni)
2951                 return lpni->lpni_nid;
2952
2953         /* Look for a routed-connected NID for this peer. */
2954         lpni = NULL;
2955         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2956                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2957                         continue;
2958                 break;
2959         }
2960         if (lpni)
2961                 return lpni->lpni_nid;
2962
2963         return LNET_NID_ANY;
2964 }
2965
2966 /* Active side of ping. */
2967 static int lnet_peer_send_ping(struct lnet_peer *lp)
2968 __must_hold(&lp->lp_lock)
2969 {
2970         lnet_nid_t pnid;
2971         int nnis;
2972         int rc;
2973         int cpt;
2974
2975         lp->lp_state |= LNET_PEER_PING_SENT;
2976         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2977         spin_unlock(&lp->lp_lock);
2978
2979         cpt = lnet_net_lock_current();
2980         /* Refcount for MD. */
2981         lnet_peer_addref_locked(lp);
2982         pnid = lnet_peer_select_nid(lp);
2983         lnet_net_unlock(cpt);
2984
2985         nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2986
2987         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2988                             the_lnet.ln_dc_eq, false);
2989
2990         /*
2991          * if LNetMDBind in lnet_send_ping fails we need to decrement the
2992          * refcount on the peer, otherwise LNetMDUnlink will be called
2993          * which will eventually do that.
2994          */
2995         if (rc > 0) {
2996                 lnet_net_lock(cpt);
2997                 lnet_peer_decref_locked(lp);
2998                 lnet_net_unlock(cpt);
2999                 rc = -rc; /* change the rc to negative value */
3000                 goto fail_error;
3001         } else if (rc < 0) {
3002                 goto fail_error;
3003         }
3004
3005         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3006
3007         spin_lock(&lp->lp_lock);
3008         return 0;
3009
3010 fail_error:
3011         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3012         /*
3013          * The errors that get us here are considered hard errors and
3014          * cause Discovery to terminate. So we clear PING_SENT, but do
3015          * not set either PING_FAILED or FORCE_PING. In fact we need
3016          * to clear PING_FAILED, because the unlink event handler will
3017          * have set it if we called LNetMDUnlink() above.
3018          */
3019         spin_lock(&lp->lp_lock);
3020         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3021         return rc;
3022 }
3023
3024 /*
3025  * This function exists because you cannot call LNetMDUnlink() from an
3026  * event handler.
3027  */
3028 static int lnet_peer_push_failed(struct lnet_peer *lp)
3029 __must_hold(&lp->lp_lock)
3030 {
3031         struct lnet_handle_md mdh;
3032         int rc;
3033
3034         mdh = lp->lp_push_mdh;
3035         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3036         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3037         rc = lp->lp_push_error;
3038         lp->lp_push_error = 0;
3039         spin_unlock(&lp->lp_lock);
3040
3041         if (!LNetMDHandleIsInvalid(mdh))
3042                 LNetMDUnlink(mdh);
3043
3044         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3045         spin_lock(&lp->lp_lock);
3046         return rc ? rc : LNET_REDISCOVER_PEER;
3047 }
3048
3049 /* Active side of push. */
3050 static int lnet_peer_send_push(struct lnet_peer *lp)
3051 __must_hold(&lp->lp_lock)
3052 {
3053         struct lnet_ping_buffer *pbuf;
3054         struct lnet_process_id id;
3055         struct lnet_md md;
3056         int cpt;
3057         int rc;
3058
3059         /* Don't push to a non-multi-rail peer. */
3060         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3061                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3062                 return 0;
3063         }
3064
3065         lp->lp_state |= LNET_PEER_PUSH_SENT;
3066         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3067         spin_unlock(&lp->lp_lock);
3068
3069         cpt = lnet_net_lock_current();
3070         pbuf = the_lnet.ln_ping_target;
3071         lnet_ping_buffer_addref(pbuf);
3072         lnet_net_unlock(cpt);
3073
3074         /* Push source MD */
3075         md.start     = &pbuf->pb_info;
3076         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3077         md.threshold = 2; /* Put/Ack */
3078         md.max_size  = 0;
3079         md.options   = 0;
3080         md.eq_handle = the_lnet.ln_dc_eq;
3081         md.user_ptr  = lp;
3082
3083         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3084         if (rc) {
3085                 lnet_ping_buffer_decref(pbuf);
3086                 CERROR("Can't bind push source MD: %d\n", rc);
3087                 goto fail_error;
3088         }
3089         cpt = lnet_net_lock_current();
3090         /* Refcount for MD. */
3091         lnet_peer_addref_locked(lp);
3092         id.pid = LNET_PID_LUSTRE;
3093         id.nid = lnet_peer_select_nid(lp);
3094         lnet_net_unlock(cpt);
3095
3096         if (id.nid == LNET_NID_ANY) {
3097                 rc = -EHOSTUNREACH;
3098                 goto fail_unlink;
3099         }
3100
3101         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3102                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3103                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3104
3105         if (rc)
3106                 goto fail_unlink;
3107
3108         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3109
3110         spin_lock(&lp->lp_lock);
3111         return 0;
3112
3113 fail_unlink:
3114         LNetMDUnlink(lp->lp_push_mdh);
3115         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3116 fail_error:
3117         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3118         /*
3119          * The errors that get us here are considered hard errors and
3120          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3121          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3122          * because the unlink event handler will have set it if we
3123          * called LNetMDUnlink() above.
3124          */
3125         spin_lock(&lp->lp_lock);
3126         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3127         return rc;
3128 }
3129
3130 /*
3131  * An unrecoverable error was encountered during discovery.
3132  * Set error status in peer and abort discovery.
3133  */
3134 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3135 {
3136         CDEBUG(D_NET, "Discovery error %s: %d\n",
3137                libcfs_nid2str(lp->lp_primary_nid), error);
3138
3139         spin_lock(&lp->lp_lock);
3140         lp->lp_dc_error = error;
3141         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3142         lp->lp_state |= LNET_PEER_REDISCOVER;
3143         spin_unlock(&lp->lp_lock);
3144 }
3145
3146 /*
3147  * Mark the peer as discovered.
3148  */
3149 static int lnet_peer_discovered(struct lnet_peer *lp)
3150 __must_hold(&lp->lp_lock)
3151 {
3152         lp->lp_state |= LNET_PEER_DISCOVERED;
3153         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3154                           LNET_PEER_REDISCOVER);
3155
3156         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3157
3158         return 0;
3159 }
3160
3161
3162 /*
3163  * Discovering this peer is taking too long. Cancel any Ping or Push
3164  * that discovery is waiting on by unlinking the relevant MDs. The
3165  * lnet_discovery_event_handler() will proceed from here and complete
3166  * the cleanup.
3167  */
3168 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3169 {
3170         struct lnet_handle_md ping_mdh;
3171         struct lnet_handle_md push_mdh;
3172
3173         LNetInvalidateMDHandle(&ping_mdh);
3174         LNetInvalidateMDHandle(&push_mdh);
3175
3176         spin_lock(&lp->lp_lock);
3177         if (lp->lp_state & LNET_PEER_PING_SENT) {
3178                 ping_mdh = lp->lp_ping_mdh;
3179                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3180         }
3181         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3182                 push_mdh = lp->lp_push_mdh;
3183                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3184         }
3185         spin_unlock(&lp->lp_lock);
3186
3187         if (!LNetMDHandleIsInvalid(ping_mdh))
3188                 LNetMDUnlink(ping_mdh);
3189         if (!LNetMDHandleIsInvalid(push_mdh))
3190                 LNetMDUnlink(push_mdh);
3191 }
3192
3193 /*
3194  * Wait for work to be queued or some other change that must be
3195  * attended to. Returns non-zero if the discovery thread should shut
3196  * down.
3197  */
3198 static int lnet_peer_discovery_wait_for_work(void)
3199 {
3200         int cpt;
3201         int rc = 0;
3202
3203         DEFINE_WAIT(wait);
3204
3205         cpt = lnet_net_lock_current();
3206         for (;;) {
3207                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3208                                 TASK_INTERRUPTIBLE);
3209                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3210                         break;
3211                 if (lnet_push_target_resize_needed() ||
3212                     the_lnet.ln_push_target->pb_needs_post)
3213                         break;
3214                 if (!list_empty(&the_lnet.ln_dc_request))
3215                         break;
3216                 if (!list_empty(&the_lnet.ln_msg_resend))
3217                         break;
3218                 lnet_net_unlock(cpt);
3219
3220                 /*
3221                  * wakeup max every second to check if there are peers that
3222                  * have been stuck on the working queue for greater than
3223                  * the peer timeout.
3224                  */
3225                 schedule_timeout(cfs_time_seconds(1));
3226                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3227                 cpt = lnet_net_lock_current();
3228         }
3229         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3230
3231         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3232                 rc = -ESHUTDOWN;
3233
3234         lnet_net_unlock(cpt);
3235
3236         CDEBUG(D_NET, "woken: %d\n", rc);
3237
3238         return rc;
3239 }
3240
3241 /*
3242  * Messages that were pending on a destroyed peer will be put on a global
3243  * resend list. The message resend list will be checked by
3244  * the discovery thread when it wakes up, and will resend messages. These
3245  * messages can still be sendable in the case the lpni which was the initial
3246  * cause of the message re-queue was transfered to another peer.
3247  *
3248  * It is possible that LNet could be shutdown while we're iterating
3249  * through the list. lnet_shudown_lndnets() will attempt to access the
3250  * resend list, but will have to wait until the spinlock is released, by
3251  * which time there shouldn't be any more messages on the resend list.
3252  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3253  * for the messages so they can be released. The other case is that
3254  * lnet_shudown_lndnets() can finalize all the messages before this
3255  * function can visit the resend list, in which case this function will be
3256  * a no-op.
3257  */
3258 static void lnet_resend_msgs(void)
3259 {
3260         struct lnet_msg *msg, *tmp;
3261         LIST_HEAD(resend);
3262         int rc;
3263
3264         spin_lock(&the_lnet.ln_msg_resend_lock);
3265         list_splice(&the_lnet.ln_msg_resend, &resend);
3266         spin_unlock(&the_lnet.ln_msg_resend_lock);
3267
3268         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3269                 list_del_init(&msg->msg_list);
3270                 rc = lnet_send(msg->msg_src_nid_param, msg,
3271                                msg->msg_rtr_nid_param);
3272                 if (rc < 0) {
3273                         CNETERR("Error sending %s to %s: %d\n",
3274                                lnet_msgtyp2str(msg->msg_type),
3275                                libcfs_id2str(msg->msg_target), rc);
3276                         lnet_finalize(msg, rc);
3277                 }
3278         }
3279 }
3280
3281 /* The discovery thread. */
3282 static int lnet_peer_discovery(void *arg)
3283 {
3284         struct lnet_peer *lp;
3285         int rc;
3286
3287         wait_for_completion(&the_lnet.ln_started);
3288
3289         CDEBUG(D_NET, "started\n");
3290
3291         for (;;) {
3292                 if (lnet_peer_discovery_wait_for_work())
3293                         break;
3294
3295                 if (lnet_push_target_resize_needed())
3296                         lnet_push_target_resize();
3297                 else if (the_lnet.ln_push_target->pb_needs_post)
3298                         lnet_push_target_post(the_lnet.ln_push_target,
3299                                               &the_lnet.ln_push_target_md);
3300
3301                 lnet_resend_msgs();
3302
3303                 lnet_net_lock(LNET_LOCK_EX);
3304                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3305                         lnet_net_unlock(LNET_LOCK_EX);
3306                         break;
3307                 }
3308
3309                 /*
3310                  * Process all incoming discovery work requests.  When
3311                  * discovery must wait on a peer to change state, it
3312                  * is added to the tail of the ln_dc_working queue. A
3313                  * timestamp keeps track of when the peer was added,
3314                  * so we can time out discovery requests that take too
3315                  * long.
3316                  */
3317                 while (!list_empty(&the_lnet.ln_dc_request)) {
3318                         lp = list_first_entry(&the_lnet.ln_dc_request,
3319                                               struct lnet_peer, lp_dc_list);
3320                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3321                         /*
3322                          * set the time the peer was put on the dc_working
3323                          * queue. It shouldn't remain on the queue
3324                          * forever, in case the GET message (for ping)
3325                          * doesn't get a REPLY or the PUT message (for
3326                          * push) doesn't get an ACK.
3327                          */
3328                         lp->lp_last_queued = ktime_get_real_seconds();
3329                         lnet_net_unlock(LNET_LOCK_EX);
3330
3331                         if (lnet_push_target_resize_needed())
3332                                 lnet_push_target_resize();
3333                         else if (the_lnet.ln_push_target->pb_needs_post)
3334                                 lnet_push_target_post(the_lnet.ln_push_target,
3335                                                       &the_lnet.ln_push_target_md);
3336
3337                         /*
3338                          * Select an action depending on the state of
3339                          * the peer and whether discovery is disabled.
3340                          * The check whether discovery is disabled is
3341                          * done after the code that handles processing
3342                          * for arrived data, cleanup for failures, and
3343                          * forcing a Ping or Push.
3344                          */
3345                         spin_lock(&lp->lp_lock);
3346                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3347                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3348                                 lp->lp_state);
3349                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3350                                 rc = lnet_peer_data_present(lp);
3351                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3352                                 rc = lnet_peer_ping_failed(lp);
3353                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3354                                 rc = lnet_peer_push_failed(lp);
3355                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3356                                 rc = lnet_peer_send_ping(lp);
3357                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3358                                 rc = lnet_peer_send_push(lp);
3359                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3360                                 rc = lnet_peer_send_ping(lp);
3361                         else if (lnet_peer_needs_push(lp))
3362                                 rc = lnet_peer_send_push(lp);
3363                         else
3364                                 rc = lnet_peer_discovered(lp);
3365                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3366                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3367                                 lp->lp_state, rc);
3368                         spin_unlock(&lp->lp_lock);
3369
3370                         lnet_net_lock(LNET_LOCK_EX);
3371                         if (rc == LNET_REDISCOVER_PEER) {
3372                                 list_move(&lp->lp_dc_list,
3373                                           &the_lnet.ln_dc_request);
3374                         } else if (rc) {
3375                                 lnet_peer_discovery_error(lp, rc);
3376                         }
3377                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3378                                 lnet_peer_discovery_complete(lp);
3379                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3380                                 break;
3381                 }
3382
3383                 lnet_net_unlock(LNET_LOCK_EX);
3384         }
3385
3386         CDEBUG(D_NET, "stopping\n");
3387         /*
3388          * Clean up before telling lnet_peer_discovery_stop() that
3389          * we're done. Use wake_up() below to somewhat reduce the
3390          * size of the thundering herd if there are multiple threads
3391          * waiting on discovery of a single peer.
3392          */
3393
3394         /* Queue cleanup 1: stop all pending pings and pushes. */
3395         lnet_net_lock(LNET_LOCK_EX);
3396         while (!list_empty(&the_lnet.ln_dc_working)) {
3397                 lp = list_first_entry(&the_lnet.ln_dc_working,
3398                                       struct lnet_peer, lp_dc_list);
3399                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3400                 lnet_net_unlock(LNET_LOCK_EX);
3401                 lnet_peer_cancel_discovery(lp);
3402                 lnet_net_lock(LNET_LOCK_EX);
3403         }
3404         lnet_net_unlock(LNET_LOCK_EX);
3405
3406         /* Queue cleanup 2: wait for the expired queue to clear. */
3407         while (!list_empty(&the_lnet.ln_dc_expired))
3408                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3409
3410         /* Queue cleanup 3: clear the request queue. */
3411         lnet_net_lock(LNET_LOCK_EX);
3412         while (!list_empty(&the_lnet.ln_dc_request)) {
3413                 lp = list_first_entry(&the_lnet.ln_dc_request,
3414                                       struct lnet_peer, lp_dc_list);
3415                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3416                 lnet_peer_discovery_complete(lp);
3417         }
3418         lnet_net_unlock(LNET_LOCK_EX);
3419
3420         LNetEQFree(the_lnet.ln_dc_eq);
3421         the_lnet.ln_dc_eq = NULL;
3422
3423         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3424         wake_up(&the_lnet.ln_dc_waitq);
3425
3426         CDEBUG(D_NET, "stopped\n");
3427
3428         return 0;
3429 }
3430
3431 /* ln_api_mutex is held on entry. */
3432 int lnet_peer_discovery_start(void)
3433 {
3434         struct task_struct *task;
3435         int rc = 0;
3436
3437         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3438                 return -EALREADY;
3439
3440         the_lnet.ln_dc_eq = LNetEQAlloc(lnet_discovery_event_handler);
3441         if (IS_ERR(the_lnet.ln_dc_eq)) {
3442                 rc = PTR_ERR(the_lnet.ln_dc_eq);
3443                 CERROR("Can't allocate discovery EQ: %d\n", rc);
3444                 return rc;
3445         }
3446
3447         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3448         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3449         if (IS_ERR(task)) {
3450                 rc = PTR_ERR(task);
3451                 CERROR("Can't start peer discovery thread: %d\n", rc);
3452
3453                 LNetEQFree(the_lnet.ln_dc_eq);
3454                 the_lnet.ln_dc_eq = NULL;
3455
3456                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3457         }
3458
3459         CDEBUG(D_NET, "discovery start: %d\n", rc);
3460
3461         return rc;
3462 }
3463
3464 /* ln_api_mutex is held on entry. */
3465 void lnet_peer_discovery_stop(void)
3466 {
3467         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3468                 return;
3469
3470         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3471         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3472
3473         /* In the LNetNIInit() path we may be stopping discovery before it
3474          * entered its work loop
3475          */
3476         if (!completion_done(&the_lnet.ln_started))
3477                 complete(&the_lnet.ln_started);
3478         else
3479                 wake_up(&the_lnet.ln_dc_waitq);
3480
3481         wait_event(the_lnet.ln_dc_waitq,
3482                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3483
3484         LASSERT(list_empty(&the_lnet.ln_dc_request));
3485         LASSERT(list_empty(&the_lnet.ln_dc_working));
3486         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3487
3488         CDEBUG(D_NET, "discovery stopped\n");
3489 }
3490
3491 /* Debugging */
3492
3493 void
3494 lnet_debug_peer(lnet_nid_t nid)
3495 {
3496         char                    *aliveness = "NA";
3497         struct lnet_peer_ni     *lp;
3498         int                     cpt;
3499
3500         cpt = lnet_cpt_of_nid(nid, NULL);
3501         lnet_net_lock(cpt);
3502
3503         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3504         if (IS_ERR(lp)) {
3505                 lnet_net_unlock(cpt);
3506                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3507                 return;
3508         }
3509
3510         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3511                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3512
3513         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3514                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3515                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3516                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3517                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3518
3519         lnet_peer_ni_decref_locked(lp);
3520
3521         lnet_net_unlock(cpt);
3522 }
3523
3524 /* Gathering information for userspace. */
3525
3526 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3527                           char aliveness[LNET_MAX_STR_LEN],
3528                           __u32 *cpt_iter, __u32 *refcount,
3529                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3530                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3531                           __u32 *peer_tx_qnob)
3532 {
3533         struct lnet_peer_table          *peer_table;
3534         struct lnet_peer_ni             *lp;
3535         int                             j;
3536         int                             lncpt;
3537         bool                            found = false;
3538
3539         /* get the number of CPTs */
3540         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3541
3542         /* if the cpt number to be examined is >= the number of cpts in
3543          * the system then indicate that there are no more cpts to examin
3544          */
3545         if (*cpt_iter >= lncpt)
3546                 return -ENOENT;
3547
3548         /* get the current table */
3549         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3550         /* if the ptable is NULL then there are no more cpts to examine */
3551         if (peer_table == NULL)
3552                 return -ENOENT;
3553
3554         lnet_net_lock(*cpt_iter);
3555
3556         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3557                 struct list_head *peers = &peer_table->pt_hash[j];
3558
3559                 list_for_each_entry(lp, peers, lpni_hashlist) {
3560                         if (peer_index-- > 0)
3561                                 continue;
3562
3563                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3564                         if (lnet_isrouter(lp) ||
3565                                 lnet_peer_aliveness_enabled(lp))
3566                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3567                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3568
3569                         *nid = lp->lpni_nid;
3570                         *refcount = atomic_read(&lp->lpni_refcount);
3571                         *ni_peer_tx_credits =
3572                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3573                         *peer_tx_credits = lp->lpni_txcredits;
3574                         *peer_rtr_credits = lp->lpni_rtrcredits;
3575                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3576                         *peer_tx_qnob = lp->lpni_txqnob;
3577
3578                         found = true;
3579                 }
3580
3581         }
3582         lnet_net_unlock(*cpt_iter);
3583
3584         *cpt_iter = lncpt;
3585
3586         return found ? 0 : -ENOENT;
3587 }
3588
3589 /* ln_api_mutex is held, which keeps the peer list stable */
3590 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3591 {
3592         struct lnet_ioctl_element_stats *lpni_stats;
3593         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3594         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3595         struct lnet_peer_ni_credit_info *lpni_info;
3596         struct lnet_peer_ni *lpni;
3597         struct lnet_peer *lp;
3598         lnet_nid_t nid;
3599         __u32 size;
3600         int rc;
3601
3602         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3603
3604         if (!lp) {
3605                 rc = -ENOENT;
3606                 goto out;
3607         }
3608
3609         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3610                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3611         size *= lp->lp_nnis;
3612         if (size > cfg->prcfg_size) {
3613                 cfg->prcfg_size = size;
3614                 rc = -E2BIG;
3615                 goto out_lp_decref;
3616         }
3617
3618         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3619         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3620         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3621         cfg->prcfg_count = lp->lp_nnis;
3622         cfg->prcfg_size = size;
3623         cfg->prcfg_state = lp->lp_state;
3624
3625         /* Allocate helper buffers. */
3626         rc = -ENOMEM;
3627         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3628         if (!lpni_info)
3629                 goto out_lp_decref;
3630         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3631         if (!lpni_stats)
3632                 goto out_free_info;
3633         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3634         if (!lpni_msg_stats)
3635                 goto out_free_stats;
3636         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3637         if (!lpni_hstats)
3638                 goto out_free_msg_stats;
3639
3640
3641         lpni = NULL;
3642         rc = -EFAULT;
3643         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3644                 nid = lpni->lpni_nid;
3645                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3646                         goto out_free_hstats;
3647                 bulk += sizeof(nid);
3648
3649                 memset(lpni_info, 0, sizeof(*lpni_info));
3650                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3651                 if (lnet_isrouter(lpni) ||
3652                         lnet_peer_aliveness_enabled(lpni))
3653                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3654                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3655
3656                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3657                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3658                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3659                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3660                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3661                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3662                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3663                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3664                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3665                         goto out_free_hstats;
3666                 bulk += sizeof(*lpni_info);
3667
3668                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3669                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3670                                                             LNET_STATS_TYPE_SEND);
3671                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3672                                                             LNET_STATS_TYPE_RECV);
3673                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3674                                                             LNET_STATS_TYPE_DROP);
3675                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3676                         goto out_free_hstats;
3677                 bulk += sizeof(*lpni_stats);
3678                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3679                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3680                         goto out_free_hstats;
3681                 bulk += sizeof(*lpni_msg_stats);
3682                 lpni_hstats->hlpni_network_timeout =
3683                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3684                 lpni_hstats->hlpni_remote_dropped =
3685                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3686                 lpni_hstats->hlpni_remote_timeout =
3687                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3688                 lpni_hstats->hlpni_remote_error =
3689                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3690                 lpni_hstats->hlpni_health_value =
3691                   atomic_read(&lpni->lpni_healthv);
3692                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3693                         goto out_free_hstats;
3694                 bulk += sizeof(*lpni_hstats);
3695         }
3696         rc = 0;
3697
3698 out_free_hstats:
3699         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3700 out_free_msg_stats:
3701         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3702 out_free_stats:
3703         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3704 out_free_info:
3705         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3706 out_lp_decref:
3707         lnet_peer_decref_locked(lp);
3708 out:
3709         return rc;
3710 }
3711
3712 void
3713 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3714 {
3715         /* the mt could've shutdown and cleaned up the queues */
3716         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3717                 return;
3718
3719         if (list_empty(&lpni->lpni_recovery) &&
3720             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3721                 CDEBUG(D_NET, "lpni %s added to recovery queue. Health = %d\n",
3722                         libcfs_nid2str(lpni->lpni_nid),
3723                         atomic_read(&lpni->lpni_healthv));
3724                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3725                 lnet_peer_ni_addref_locked(lpni);
3726         }
3727 }
3728
3729 /* Call with the ln_api_mutex held */
3730 void
3731 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3732 {
3733         struct lnet_peer_table *ptable;
3734         struct lnet_peer *lp;
3735         struct lnet_peer_net *lpn;
3736         struct lnet_peer_ni *lpni;
3737         int lncpt;
3738         int cpt;
3739
3740         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3741                 return;
3742
3743         if (!all) {
3744                 lnet_net_lock(LNET_LOCK_EX);
3745                 lpni = lnet_find_peer_ni_locked(nid);
3746                 if (!lpni) {
3747                         lnet_net_unlock(LNET_LOCK_EX);
3748                         return;
3749                 }
3750                 atomic_set(&lpni->lpni_healthv, value);
3751                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3752                 lnet_peer_ni_decref_locked(lpni);
3753                 lnet_net_unlock(LNET_LOCK_EX);
3754                 return;
3755         }
3756
3757         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3758
3759         /*
3760          * Walk all the peers and reset the healhv for each one to the
3761          * maximum value.
3762          */
3763         lnet_net_lock(LNET_LOCK_EX);
3764         for (cpt = 0; cpt < lncpt; cpt++) {
3765                 ptable = the_lnet.ln_peer_tables[cpt];
3766                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3767                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3768                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3769                                                     lpni_peer_nis) {
3770                                         atomic_set(&lpni->lpni_healthv, value);
3771                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3772                                 }
3773                         }
3774                 }
3775         }
3776         lnet_net_unlock(LNET_LOCK_EX);
3777 }
3778