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 (nid == LNET_NID_LO_0)
 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 = lpni2->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         wait_var_event_warning(&ptable->pt_zombies,
 580                                ptable->pt_zombies == 0,
 581                                "Waiting for %d zombies on peer table\n",
 582                                ptable->pt_zombies);
 583 }
 584
 585 static void
 586 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 587                                 struct lnet_peer_table *ptable)
 588 {
 589         struct lnet_peer_ni     *lp;
 590         struct lnet_peer_ni     *tmp;
 591         lnet_nid_t              gw_nid;
 592         int                     i;
 593
 594         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 595                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 596                                          lpni_hashlist) {
 597                         if (net != lp->lpni_net)
 598                                 continue;
 599
 600                         if (!lnet_isrouter(lp))
 601                                 continue;
 602
 603                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 604
 605                         lnet_net_unlock(LNET_LOCK_EX);
 606                         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
 607                         lnet_net_lock(LNET_LOCK_EX);
 608                 }
 609         }
 610 }
 611
 612 void
 613 lnet_peer_tables_cleanup(struct lnet_net *net)
 614 {
 615         int i;
 616         struct lnet_peer_table *ptable;
 617
 618         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 619         /* If just deleting the peers for a NI, get rid of any routes these
 620          * peers are gateways for. */
 621         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 622                 lnet_net_lock(LNET_LOCK_EX);
 623                 lnet_peer_table_del_rtrs_locked(net, ptable);
 624                 lnet_net_unlock(LNET_LOCK_EX);
 625         }
 626
 627         /* Start the cleanup process */
 628         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 629                 lnet_net_lock(LNET_LOCK_EX);
 630                 lnet_peer_table_cleanup_locked(net, ptable);
 631                 lnet_net_unlock(LNET_LOCK_EX);
 632         }
 633
 634         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 635                 lnet_peer_ni_finalize_wait(ptable);
 636 }
 637
 638 static struct lnet_peer_ni *
 639 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 640 {
 641         struct list_head        *peers;
 642         struct lnet_peer_ni     *lp;
 643
 644         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 645                 return NULL;
 646
 647         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 648         list_for_each_entry(lp, peers, lpni_hashlist) {
 649                 if (lp->lpni_nid == nid) {
 650                         lnet_peer_ni_addref_locked(lp);
 651                         return lp;
 652                 }
 653         }
 654
 655         return NULL;
 656 }
 657
 658 struct lnet_peer_ni *
 659 lnet_find_peer_ni_locked(lnet_nid_t nid)
 660 {
 661         struct lnet_peer_ni *lpni;
 662         struct lnet_peer_table *ptable;
 663         int cpt;
 664
 665         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 666
 667         ptable = the_lnet.ln_peer_tables[cpt];
 668         lpni = lnet_get_peer_ni_locked(ptable, nid);
 669
 670         return lpni;
 671 }
 672
 673 struct lnet_peer_ni *
 674 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 675 {
 676         struct lnet_peer_net *lpn;
 677         struct lnet_peer_ni *lpni;
 678
 679         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 680         if (!lpn)
 681                 return NULL;
 682
 683         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 684                 if (lpni->lpni_nid == nid)
 685                         return lpni;
 686         }
 687
 688         return NULL;
 689 }
 690
 691 struct lnet_peer *
 692 lnet_find_peer(lnet_nid_t nid)
 693 {
 694         struct lnet_peer_ni *lpni;
 695         struct lnet_peer *lp = NULL;
 696         int cpt;
 697
 698         cpt = lnet_net_lock_current();
 699         lpni = lnet_find_peer_ni_locked(nid);
 700         if (lpni) {
 701                 lp = lpni->lpni_peer_net->lpn_peer;
 702                 lnet_peer_addref_locked(lp);
 703                 lnet_peer_ni_decref_locked(lpni);
 704         }
 705         lnet_net_unlock(cpt);
 706
 707         return lp;
 708 }
 709
 710 struct lnet_peer_net *
 711 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 712 {
 713         struct lnet_peer_net *net;
 714
 715         if (!prev_lpn_id) {
 716                 /* no net id provided return the first net */
 717                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 718                                                struct lnet_peer_net,
 719                                                lpn_peer_nets);
 720
 721                 return net;
 722         }
 723
 724         /* find the net after the one provided */
 725         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 726                 if (net->lpn_net_id == prev_lpn_id) {
 727                         /*
 728                          * if we reached the end of the list loop to the
 729                          * beginning.
 730                          */
 731                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 732                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 733                                                                 struct lnet_peer_net,
 734                                                                 lpn_peer_nets);
 735                         else
 736                                 return list_next_entry(net, lpn_peer_nets);
 737                 }
 738         }
 739
 740         return NULL;
 741 }
 742
 743 struct lnet_peer_ni *
 744 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 745                              struct lnet_peer_net *peer_net,
 746                              struct lnet_peer_ni *prev)
 747 {
 748         struct lnet_peer_ni *lpni;
 749         struct lnet_peer_net *net = peer_net;
 750
 751         if (!prev) {
 752                 if (!net) {
 753                         if (list_empty(&peer->lp_peer_nets))
 754                                 return NULL;
 755
 756                         net = list_entry(peer->lp_peer_nets.next,
 757                                          struct lnet_peer_net,
 758                                          lpn_peer_nets);
 759                 }
 760                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 761                                   lpni_peer_nis);
 762
 763                 return lpni;
 764         }
 765
 766         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 767                 /*
 768                  * if you reached the end of the peer ni list and the peer
 769                  * net is specified then there are no more peer nis in that
 770                  * net.
 771                  */
 772                 if (net)
 773                         return NULL;
 774
 775                 /*
 776                  * we reached the end of this net ni list. move to the
 777                  * next net
 778                  */
 779                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 780                     &peer->lp_peer_nets)
 781                         /* no more nets and no more NIs. */
 782                         return NULL;
 783
 784                 /* get the next net */
 785                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 786                                  struct lnet_peer_net,
 787                                  lpn_peer_nets);
 788                 /* get the ni on it */
 789                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 790                                   lpni_peer_nis);
 791
 792                 return lpni;
 793         }
 794
 795         /* there are more nis left */
 796         lpni = list_entry(prev->lpni_peer_nis.next,
 797                           struct lnet_peer_ni, lpni_peer_nis);
 798
 799         return lpni;
 800 }
 801
 802 /* Call with the ln_api_mutex held */
 803 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 804 {
 805         struct lnet_process_id id;
 806         struct lnet_peer_table *ptable;
 807         struct lnet_peer *lp;
 808         __u32 count = 0;
 809         __u32 size = 0;
 810         int lncpt;
 811         int cpt;
 812         __u32 i;
 813         int rc;
 814
 815         rc = -ESHUTDOWN;
 816         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 817                 goto done;
 818
 819         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 820
 821         /*
 822          * Count the number of peers, and return E2BIG if the buffer
 823          * is too small. We'll also return the desired size.
 824          */
 825         rc = -E2BIG;
 826         for (cpt = 0; cpt < lncpt; cpt++) {
 827                 ptable = the_lnet.ln_peer_tables[cpt];
 828                 count += ptable->pt_peers;
 829         }
 830         size = count * sizeof(*ids);
 831         if (size > *sizep)
 832                 goto done;
 833
 834         /*
 835          * Walk the peer lists and copy out the primary nids.
 836          * This is safe because the peer lists are only modified
 837          * while the ln_api_mutex is held. So we don't need to
 838          * hold the lnet_net_lock as well, and can therefore
 839          * directly call copy_to_user().
 840          */
 841         rc = -EFAULT;
 842         memset(&id, 0, sizeof(id));
 843         id.pid = LNET_PID_LUSTRE;
 844         i = 0;
 845         for (cpt = 0; cpt < lncpt; cpt++) {
 846                 ptable = the_lnet.ln_peer_tables[cpt];
 847                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 848                         if (i >= count)
 849                                 goto done;
 850                         id.nid = lp->lp_primary_nid;
 851                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 852                                 goto done;
 853                         i++;
 854                 }
 855         }
 856         rc = 0;
 857 done:
 858         *countp = count;
 859         *sizep = size;
 860         return rc;
 861 }
 862
 863 /*
 864  * Start pushes to peers that need to be updated for a configuration
 865  * change on this node.
 866  */
 867 void
 868 lnet_push_update_to_peers(int force)
 869 {
 870         struct lnet_peer_table *ptable;
 871         struct lnet_peer *lp;
 872         int lncpt;
 873         int cpt;
 874
 875         lnet_net_lock(LNET_LOCK_EX);
 876         if (lnet_peer_discovery_disabled)
 877                 force = 0;
 878         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 879         for (cpt = 0; cpt < lncpt; cpt++) {
 880                 ptable = the_lnet.ln_peer_tables[cpt];
 881                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 882                         if (force) {
 883                                 spin_lock(&lp->lp_lock);
 884                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 885                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 886                                 spin_unlock(&lp->lp_lock);
 887                         }
 888                         if (lnet_peer_needs_push(lp))
 889                                 lnet_peer_queue_for_discovery(lp);
 890                 }
 891         }
 892         lnet_net_unlock(LNET_LOCK_EX);
 893         wake_up(&the_lnet.ln_dc_waitq);
 894 }
 895
 896 /*
 897  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 898  * this is a preferred point-to-point path. Call with lnet_net_lock in
 899  * shared mmode.
 900  */
 901 bool
 902 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 903 {
 904         int i;
 905
 906         if (lpni->lpni_pref_nnids == 0)
 907                 return false;
 908         if (lpni->lpni_pref_nnids == 1)
 909                 return lpni->lpni_pref.nid == nid;
 910         for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 911                 if (lpni->lpni_pref.nids[i] == nid)
 912                         return true;
 913         }
 914         return false;
 915 }
 916
 917 /*
 918  * Set a single ni as preferred, provided no preferred ni is already
 919  * defined. Only to be used for non-multi-rail peer_ni.
 920  */
 921 int
 922 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 923 {
 924         int rc = 0;
 925
 926         spin_lock(&lpni->lpni_lock);
 927         if (nid == LNET_NID_ANY) {
 928                 rc = -EINVAL;
 929         } else if (lpni->lpni_pref_nnids > 0) {
 930                 rc = -EPERM;
 931         } else if (lpni->lpni_pref_nnids == 0) {
 932                 lpni->lpni_pref.nid = nid;
 933                 lpni->lpni_pref_nnids = 1;
 934                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
 935         }
 936         spin_unlock(&lpni->lpni_lock);
 937
 938         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 939                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
 940         return rc;
 941 }
 942
 943 /*
 944  * Clear the preferred NID from a non-multi-rail peer_ni, provided
 945  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
 946  */
 947 int
 948 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
 949 {
 950         int rc = 0;
 951
 952         spin_lock(&lpni->lpni_lock);
 953         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
 954                 lpni->lpni_pref_nnids = 0;
 955                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 956         } else if (lpni->lpni_pref_nnids == 0) {
 957                 rc = -ENOENT;
 958         } else {
 959                 rc = -EPERM;
 960         }
 961         spin_unlock(&lpni->lpni_lock);
 962
 963         CDEBUG(D_NET, "peer %s: %d\n",
 964                libcfs_nid2str(lpni->lpni_nid), rc);
 965         return rc;
 966 }
 967
 968 /*
 969  * Clear the preferred NIDs from a non-multi-rail peer.
 970  */
 971 void
 972 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
 973 {
 974         struct lnet_peer_ni *lpni = NULL;
 975
 976         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 977                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
 978 }
 979
 980 int
 981 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 982 {
 983         lnet_nid_t *nids = NULL;
 984         lnet_nid_t *oldnids = NULL;
 985         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 986         int size;
 987         int i;
 988         int rc = 0;
 989
 990         if (nid == LNET_NID_ANY) {
 991                 rc = -EINVAL;
 992                 goto out;
 993         }
 994
 995         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
 996                 rc = -EEXIST;
 997                 goto out;
 998         }
 999
1000         /* A non-MR node may have only one preferred NI per peer_ni */
1001         if (lpni->lpni_pref_nnids > 0) {
1002                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1003                         rc = -EPERM;
1004                         goto out;
1005                 }
1006         }
1007
1008         if (lpni->lpni_pref_nnids != 0) {
1009                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1010                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1011                 if (!nids) {
1012                         rc = -ENOMEM;
1013                         goto out;
1014                 }
1015                 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1016                         if (lpni->lpni_pref.nids[i] == nid) {
1017                                 LIBCFS_FREE(nids, size);
1018                                 rc = -EEXIST;
1019                                 goto out;
1020                         }
1021                         nids[i] = lpni->lpni_pref.nids[i];
1022                 }
1023                 nids[i] = nid;
1024         }
1025
1026         lnet_net_lock(LNET_LOCK_EX);
1027         spin_lock(&lpni->lpni_lock);
1028         if (lpni->lpni_pref_nnids == 0) {
1029                 lpni->lpni_pref.nid = nid;
1030         } else {
1031                 oldnids = lpni->lpni_pref.nids;
1032                 lpni->lpni_pref.nids = nids;
1033         }
1034         lpni->lpni_pref_nnids++;
1035         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1036         spin_unlock(&lpni->lpni_lock);
1037         lnet_net_unlock(LNET_LOCK_EX);
1038
1039         if (oldnids) {
1040                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1041                 CFS_FREE_PTR_ARRAY(oldnids, size);
1042         }
1043 out:
1044         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1045                 spin_lock(&lpni->lpni_lock);
1046                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1047                 spin_unlock(&lpni->lpni_lock);
1048         }
1049         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1050                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1051         return rc;
1052 }
1053
1054 int
1055 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1056 {
1057         lnet_nid_t *nids = NULL;
1058         lnet_nid_t *oldnids = NULL;
1059         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1060         int size;
1061         int i, j;
1062         int rc = 0;
1063
1064         if (lpni->lpni_pref_nnids == 0) {
1065                 rc = -ENOENT;
1066                 goto out;
1067         }
1068
1069         if (lpni->lpni_pref_nnids == 1) {
1070                 if (lpni->lpni_pref.nid != nid) {
1071                         rc = -ENOENT;
1072                         goto out;
1073                 }
1074         } else if (lpni->lpni_pref_nnids == 2) {
1075                 if (lpni->lpni_pref.nids[0] != nid &&
1076                     lpni->lpni_pref.nids[1] != nid) {
1077                         rc = -ENOENT;
1078                         goto out;
1079                 }
1080         } else {
1081                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1082                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1083                 if (!nids) {
1084                         rc = -ENOMEM;
1085                         goto out;
1086                 }
1087                 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1088                         if (lpni->lpni_pref.nids[i] != nid)
1089                                 continue;
1090                         nids[j++] = lpni->lpni_pref.nids[i];
1091                 }
1092                 /* Check if we actually removed a nid. */
1093                 if (j == lpni->lpni_pref_nnids) {
1094                         LIBCFS_FREE(nids, size);
1095                         rc = -ENOENT;
1096                         goto out;
1097                 }
1098         }
1099
1100         lnet_net_lock(LNET_LOCK_EX);
1101         spin_lock(&lpni->lpni_lock);
1102         if (lpni->lpni_pref_nnids == 1) {
1103                 lpni->lpni_pref.nid = LNET_NID_ANY;
1104         } else if (lpni->lpni_pref_nnids == 2) {
1105                 oldnids = lpni->lpni_pref.nids;
1106                 if (oldnids[0] == nid)
1107                         lpni->lpni_pref.nid = oldnids[1];
1108                 else
1109                         lpni->lpni_pref.nid = oldnids[2];
1110         } else {
1111                 oldnids = lpni->lpni_pref.nids;
1112                 lpni->lpni_pref.nids = nids;
1113         }
1114         lpni->lpni_pref_nnids--;
1115         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1116         spin_unlock(&lpni->lpni_lock);
1117         lnet_net_unlock(LNET_LOCK_EX);
1118
1119         if (oldnids) {
1120                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1121                 CFS_FREE_PTR_ARRAY(oldnids, size);
1122         }
1123 out:
1124         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1125                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1126         return rc;
1127 }
1128
1129 lnet_nid_t
1130 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1131 {
1132         struct lnet_peer_ni *lpni;
1133         lnet_nid_t primary_nid = nid;
1134
1135         lpni = lnet_find_peer_ni_locked(nid);
1136         if (lpni) {
1137                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1138                 lnet_peer_ni_decref_locked(lpni);
1139         }
1140
1141         return primary_nid;
1142 }
1143
1144 bool
1145 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1146 __must_hold(&lp->lp_lock)
1147 {
1148         if (lnet_peer_discovery_disabled)
1149                 return true;
1150
1151         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1152             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1153                 return true;
1154         }
1155
1156         return false;
1157 }
1158
1159 /*
1160  * Peer Discovery
1161  */
1162 bool
1163 lnet_is_discovery_disabled(struct lnet_peer *lp)
1164 {
1165         bool rc = false;
1166
1167         spin_lock(&lp->lp_lock);
1168         rc = lnet_is_discovery_disabled_locked(lp);
1169         spin_unlock(&lp->lp_lock);
1170
1171         return rc;
1172 }
1173
1174 lnet_nid_t
1175 LNetPrimaryNID(lnet_nid_t nid)
1176 {
1177         struct lnet_peer *lp;
1178         struct lnet_peer_ni *lpni;
1179         lnet_nid_t primary_nid = nid;
1180         int rc = 0;
1181         int cpt;
1182
1183         if (nid == LNET_NID_LO_0)
1184                 return LNET_NID_LO_0;
1185
1186         cpt = lnet_net_lock_current();
1187         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1188         if (IS_ERR(lpni)) {
1189                 rc = PTR_ERR(lpni);
1190                 goto out_unlock;
1191         }
1192         lp = lpni->lpni_peer_net->lpn_peer;
1193
1194         while (!lnet_peer_is_uptodate(lp)) {
1195                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1196                 if (rc)
1197                         goto out_decref;
1198                 lp = lpni->lpni_peer_net->lpn_peer;
1199
1200                 /* Only try once if discovery is disabled */
1201                 if (lnet_is_discovery_disabled(lp))
1202                         break;
1203         }
1204         primary_nid = lp->lp_primary_nid;
1205 out_decref:
1206         lnet_peer_ni_decref_locked(lpni);
1207 out_unlock:
1208         lnet_net_unlock(cpt);
1209
1210         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1211                libcfs_nid2str(primary_nid), rc);
1212         return primary_nid;
1213 }
1214 EXPORT_SYMBOL(LNetPrimaryNID);
1215
1216 struct lnet_peer_net *
1217 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1218 {
1219         struct lnet_peer_net *peer_net;
1220         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1221                 if (peer_net->lpn_net_id == net_id)
1222                         return peer_net;
1223         }
1224         return NULL;
1225 }
1226
1227 /*
1228  * Attach a peer_ni to a peer_net and peer. This function assumes
1229  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1230  * may be attached to a different peer, in which case it will be
1231  * properly detached first. The whole operation is done atomically.
1232  *
1233  * Always returns 0.  This is the last function called from functions
1234  * that do return an int, so returning 0 here allows the compiler to
1235  * do a tail call.
1236  */
1237 static int
1238 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1239                                 struct lnet_peer_net *lpn,
1240                                 struct lnet_peer_ni *lpni,
1241                                 unsigned flags)
1242 {
1243         struct lnet_peer_table *ptable;
1244
1245         /* Install the new peer_ni */
1246         lnet_net_lock(LNET_LOCK_EX);
1247         /* Add peer_ni to global peer table hash, if necessary. */
1248         if (list_empty(&lpni->lpni_hashlist)) {
1249                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1250
1251                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1252                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1253                 ptable->pt_version++;
1254                 /* This is the 1st refcount on lpni. */
1255                 atomic_inc(&lpni->lpni_refcount);
1256         }
1257
1258         /* Detach the peer_ni from an existing peer, if necessary. */
1259         if (lpni->lpni_peer_net) {
1260                 LASSERT(lpni->lpni_peer_net != lpn);
1261                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1262                 lnet_peer_detach_peer_ni_locked(lpni);
1263                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1264                 lpni->lpni_peer_net = NULL;
1265         }
1266
1267         /* Add peer_ni to peer_net */
1268         lpni->lpni_peer_net = lpn;
1269         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1270         lnet_update_peer_net_healthv(lpni);
1271         lnet_peer_net_addref_locked(lpn);
1272
1273         /* Add peer_net to peer */
1274         if (!lpn->lpn_peer) {
1275                 lpn->lpn_peer = lp;
1276                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1277                 lnet_peer_addref_locked(lp);
1278         }
1279
1280         /* Add peer to global peer list, if necessary */
1281         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1282         if (list_empty(&lp->lp_peer_list)) {
1283                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1284                 ptable->pt_peers++;
1285         }
1286
1287
1288         /* Update peer state */
1289         spin_lock(&lp->lp_lock);
1290         if (flags & LNET_PEER_CONFIGURED) {
1291                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1292                         lp->lp_state |= LNET_PEER_CONFIGURED;
1293         }
1294         if (flags & LNET_PEER_MULTI_RAIL) {
1295                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1296                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1297                         lnet_peer_clr_non_mr_pref_nids(lp);
1298                 }
1299         }
1300         spin_unlock(&lp->lp_lock);
1301
1302         lp->lp_nnis++;
1303         lnet_net_unlock(LNET_LOCK_EX);
1304
1305         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1306                libcfs_nid2str(lp->lp_primary_nid),
1307                libcfs_nid2str(lpni->lpni_nid), flags);
1308
1309         return 0;
1310 }
1311
1312 /*
1313  * Create a new peer, with nid as its primary nid.
1314  *
1315  * Call with the lnet_api_mutex held.
1316  */
1317 static int
1318 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1319 {
1320         struct lnet_peer *lp;
1321         struct lnet_peer_net *lpn;
1322         struct lnet_peer_ni *lpni;
1323         int rc = 0;
1324
1325         LASSERT(nid != LNET_NID_ANY);
1326
1327         /*
1328          * No need for the lnet_net_lock here, because the
1329          * lnet_api_mutex is held.
1330          */
1331         lpni = lnet_find_peer_ni_locked(nid);
1332         if (lpni) {
1333                 /* A peer with this NID already exists. */
1334                 lp = lpni->lpni_peer_net->lpn_peer;
1335                 lnet_peer_ni_decref_locked(lpni);
1336                 /*
1337                  * This is an error if the peer was configured and the
1338                  * primary NID differs or an attempt is made to change
1339                  * the Multi-Rail flag. Otherwise the assumption is
1340                  * that an existing peer is being modified.
1341                  */
1342                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1343                         if (lp->lp_primary_nid != nid)
1344                                 rc = -EEXIST;
1345                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1346                                 rc = -EPERM;
1347                         goto out;
1348                 }
1349                 /* Delete and recreate as a configured peer. */
1350                 lnet_peer_del(lp);
1351         }
1352
1353         /* Create peer, peer_net, and peer_ni. */
1354         rc = -ENOMEM;
1355         lp = lnet_peer_alloc(nid);
1356         if (!lp)
1357                 goto out;
1358         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1359         if (!lpn)
1360                 goto out_free_lp;
1361         lpni = lnet_peer_ni_alloc(nid);
1362         if (!lpni)
1363                 goto out_free_lpn;
1364
1365         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1366
1367 out_free_lpn:
1368         LIBCFS_FREE(lpn, sizeof(*lpn));
1369 out_free_lp:
1370         LIBCFS_FREE(lp, sizeof(*lp));
1371 out:
1372         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1373                libcfs_nid2str(nid), flags, rc);
1374         return rc;
1375 }
1376
1377 /*
1378  * Add a NID to a peer. Call with ln_api_mutex held.
1379  *
1380  * Error codes:
1381  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1382  *  -EEXIST:   The NID was configured by DLC for a different peer.
1383  *  -ENOMEM:   Out of memory.
1384  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1385  *             non-multi-rail peer.
1386  */
1387 static int
1388 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1389 {
1390         struct lnet_peer_net *lpn;
1391         struct lnet_peer_ni *lpni;
1392         int rc = 0;
1393
1394         LASSERT(lp);
1395         LASSERT(nid != LNET_NID_ANY);
1396
1397         /* A configured peer can only be updated through configuration. */
1398         if (!(flags & LNET_PEER_CONFIGURED)) {
1399                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1400                         rc = -EPERM;
1401                         goto out;
1402                 }
1403         }
1404
1405         /*
1406          * The MULTI_RAIL flag can be set but not cleared, because
1407          * that would leave the peer struct in an invalid state.
1408          */
1409         if (flags & LNET_PEER_MULTI_RAIL) {
1410                 spin_lock(&lp->lp_lock);
1411                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1412                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1413                         lnet_peer_clr_non_mr_pref_nids(lp);
1414                 }
1415                 spin_unlock(&lp->lp_lock);
1416         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1417                 rc = -EPERM;
1418                 goto out;
1419         }
1420
1421         lpni = lnet_find_peer_ni_locked(nid);
1422         if (lpni) {
1423                 /*
1424                  * A peer_ni already exists. This is only a problem if
1425                  * it is not connected to this peer and was configured
1426                  * by DLC.
1427                  */
1428                 lnet_peer_ni_decref_locked(lpni);
1429                 if (lpni->lpni_peer_net->lpn_peer == lp)
1430                         goto out;
1431                 if (lnet_peer_ni_is_configured(lpni)) {
1432                         rc = -EEXIST;
1433                         goto out;
1434                 }
1435                 /* If this is the primary NID, destroy the peer. */
1436                 if (lnet_peer_ni_is_primary(lpni)) {
1437                         struct lnet_peer *rtr_lp =
1438                           lpni->lpni_peer_net->lpn_peer;
1439                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1440                         /*
1441                          * if we're trying to delete a router it means
1442                          * we're moving this peer NI to a new peer so must
1443                          * transfer router properties to the new peer
1444                          */
1445                         if (rtr_refcount > 0) {
1446                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1447                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1448                         }
1449                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1450                         lpni = lnet_peer_ni_alloc(nid);
1451                         if (!lpni) {
1452                                 rc = -ENOMEM;
1453                                 goto out;
1454                         }
1455                 }
1456         } else {
1457                 lpni = lnet_peer_ni_alloc(nid);
1458                 if (!lpni) {
1459                         rc = -ENOMEM;
1460                         goto out;
1461                 }
1462         }
1463
1464         /*
1465          * Get the peer_net. Check that we're not adding a second
1466          * peer_ni on a peer_net of a non-multi-rail peer.
1467          */
1468         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1469         if (!lpn) {
1470                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1471                 if (!lpn) {
1472                         rc = -ENOMEM;
1473                         goto out_free_lpni;
1474                 }
1475         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1476                 rc = -ENOTUNIQ;
1477                 goto out_free_lpni;
1478         }
1479
1480         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1481
1482 out_free_lpni:
1483         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1484         if (!lpni->lpni_peer_net)
1485                 LIBCFS_FREE(lpni, sizeof(*lpni));
1486 out:
1487         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1488                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1489                flags, rc);
1490         return rc;
1491 }
1492
1493 /*
1494  * Update the primary NID of a peer, if possible.
1495  *
1496  * Call with the lnet_api_mutex held.
1497  */
1498 static int
1499 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1500 {
1501         lnet_nid_t old = lp->lp_primary_nid;
1502         int rc = 0;
1503
1504         if (lp->lp_primary_nid == nid)
1505                 goto out;
1506         rc = lnet_peer_add_nid(lp, nid, flags);
1507         if (rc)
1508                 goto out;
1509         lp->lp_primary_nid = nid;
1510 out:
1511         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1512                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1513         return rc;
1514 }
1515
1516 /*
1517  * lpni creation initiated due to traffic either sending or receiving.
1518  */
1519 static int
1520 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1521 {
1522         struct lnet_peer *lp;
1523         struct lnet_peer_net *lpn;
1524         struct lnet_peer_ni *lpni;
1525         unsigned flags = 0;
1526         int rc = 0;
1527
1528         if (nid == LNET_NID_ANY) {
1529                 rc = -EINVAL;
1530                 goto out;
1531         }
1532
1533         /* lnet_net_lock is not needed here because ln_api_lock is held */
1534         lpni = lnet_find_peer_ni_locked(nid);
1535         if (lpni) {
1536                 /*
1537                  * We must have raced with another thread. Since we
1538                  * know next to nothing about a peer_ni created by
1539                  * traffic, we just assume everything is ok and
1540                  * return.
1541                  */
1542                 lnet_peer_ni_decref_locked(lpni);
1543                 goto out;
1544         }
1545
1546         /* Create peer, peer_net, and peer_ni. */
1547         rc = -ENOMEM;
1548         lp = lnet_peer_alloc(nid);
1549         if (!lp)
1550                 goto out;
1551         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1552         if (!lpn)
1553                 goto out_free_lp;
1554         lpni = lnet_peer_ni_alloc(nid);
1555         if (!lpni)
1556                 goto out_free_lpn;
1557         if (pref != LNET_NID_ANY)
1558                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1559
1560         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1561
1562 out_free_lpn:
1563         LIBCFS_FREE(lpn, sizeof(*lpn));
1564 out_free_lp:
1565         LIBCFS_FREE(lp, sizeof(*lp));
1566 out:
1567         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1568         return rc;
1569 }
1570
1571 /*
1572  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1573  *
1574  * This API handles the following combinations:
1575  *   Create a peer with its primary NI if only the prim_nid is provided
1576  *   Add a NID to a peer identified by the prim_nid. The peer identified
1577  *   by the prim_nid must already exist.
1578  *   The peer being created may be non-MR.
1579  *
1580  * The caller must hold ln_api_mutex. This prevents the peer from
1581  * being created/modified/deleted by a different thread.
1582  */
1583 int
1584 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1585 {
1586         struct lnet_peer *lp = NULL;
1587         struct lnet_peer_ni *lpni;
1588         unsigned flags;
1589
1590         /* The prim_nid must always be specified */
1591         if (prim_nid == LNET_NID_ANY)
1592                 return -EINVAL;
1593
1594         flags = LNET_PEER_CONFIGURED;
1595         if (mr)
1596                 flags |= LNET_PEER_MULTI_RAIL;
1597
1598         /*
1599          * If nid isn't specified, we must create a new peer with
1600          * prim_nid as its primary nid.
1601          */
1602         if (nid == LNET_NID_ANY)
1603                 return lnet_peer_add(prim_nid, flags);
1604
1605         /* Look up the prim_nid, which must exist. */
1606         lpni = lnet_find_peer_ni_locked(prim_nid);
1607         if (!lpni)
1608                 return -ENOENT;
1609         lnet_peer_ni_decref_locked(lpni);
1610         lp = lpni->lpni_peer_net->lpn_peer;
1611
1612         /* Peer must have been configured. */
1613         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1614                 CDEBUG(D_NET, "peer %s was not configured\n",
1615                        libcfs_nid2str(prim_nid));
1616                 return -ENOENT;
1617         }
1618
1619         /* Primary NID must match */
1620         if (lp->lp_primary_nid != prim_nid) {
1621                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1622                        libcfs_nid2str(prim_nid),
1623                        libcfs_nid2str(lp->lp_primary_nid));
1624                 return -ENODEV;
1625         }
1626
1627         /* Multi-Rail flag must match. */
1628         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1629                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1630                        libcfs_nid2str(prim_nid));
1631                 return -EPERM;
1632         }
1633
1634         return lnet_peer_add_nid(lp, nid, flags);
1635 }
1636
1637 /*
1638  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1639  *
1640  * This API handles the following combinations:
1641  *   Delete a NI from a peer if both prim_nid and nid are provided.
1642  *   Delete a peer if only prim_nid is provided.
1643  *   Delete a peer if its primary nid is provided.
1644  *
1645  * The caller must hold ln_api_mutex. This prevents the peer from
1646  * being modified/deleted by a different thread.
1647  */
1648 int
1649 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1650 {
1651         struct lnet_peer *lp;
1652         struct lnet_peer_ni *lpni;
1653         unsigned flags;
1654
1655         if (prim_nid == LNET_NID_ANY)
1656                 return -EINVAL;
1657
1658         lpni = lnet_find_peer_ni_locked(prim_nid);
1659         if (!lpni)
1660                 return -ENOENT;
1661         lnet_peer_ni_decref_locked(lpni);
1662         lp = lpni->lpni_peer_net->lpn_peer;
1663
1664         if (prim_nid != lp->lp_primary_nid) {
1665                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1666                        libcfs_nid2str(prim_nid),
1667                        libcfs_nid2str(lp->lp_primary_nid));
1668                 return -ENODEV;
1669         }
1670
1671         lnet_net_lock(LNET_LOCK_EX);
1672         if (lp->lp_rtr_refcount > 0) {
1673                 lnet_net_unlock(LNET_LOCK_EX);
1674                 CERROR("%s is a router. Can not be deleted\n",
1675                        libcfs_nid2str(prim_nid));
1676                 return -EBUSY;
1677         }
1678         lnet_net_unlock(LNET_LOCK_EX);
1679
1680         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1681                 return lnet_peer_del(lp);
1682
1683         flags = LNET_PEER_CONFIGURED;
1684         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1685                 flags |= LNET_PEER_MULTI_RAIL;
1686
1687         return lnet_peer_del_nid(lp, nid, flags);
1688 }
1689
1690 void
1691 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1692 {
1693         struct lnet_peer_table *ptable;
1694         struct lnet_peer_net *lpn;
1695
1696         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1697
1698         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1699         LASSERT(list_empty(&lpni->lpni_txq));
1700         LASSERT(lpni->lpni_txqnob == 0);
1701         LASSERT(list_empty(&lpni->lpni_peer_nis));
1702         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1703
1704         lpn = lpni->lpni_peer_net;
1705         lpni->lpni_peer_net = NULL;
1706         lpni->lpni_net = NULL;
1707
1708         /* remove the peer ni from the zombie list */
1709         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1710         spin_lock(&ptable->pt_zombie_lock);
1711         list_del_init(&lpni->lpni_hashlist);
1712         ptable->pt_zombies--;
1713         spin_unlock(&ptable->pt_zombie_lock);
1714
1715         if (lpni->lpni_pref_nnids > 1)
1716                 CFS_FREE_PTR_ARRAY(lpni->lpni_pref.nids, lpni->lpni_pref_nnids);
1717
1718         LIBCFS_FREE(lpni, sizeof(*lpni));
1719
1720         lnet_peer_net_decref_locked(lpn);
1721 }
1722
1723 struct lnet_peer_ni *
1724 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1725 {
1726         struct lnet_peer_ni *lpni = NULL;
1727         int rc;
1728
1729         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1730                 return ERR_PTR(-ESHUTDOWN);
1731
1732         /*
1733          * find if a peer_ni already exists.
1734          * If so then just return that.
1735          */
1736         lpni = lnet_find_peer_ni_locked(nid);
1737         if (lpni)
1738                 return lpni;
1739
1740         lnet_net_unlock(cpt);
1741
1742         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1743         if (rc) {
1744                 lpni = ERR_PTR(rc);
1745                 goto out_net_relock;
1746         }
1747
1748         lpni = lnet_find_peer_ni_locked(nid);
1749         LASSERT(lpni);
1750
1751 out_net_relock:
1752         lnet_net_lock(cpt);
1753
1754         return lpni;
1755 }
1756
1757 /*
1758  * Get a peer_ni for the given nid, create it if necessary. Takes a
1759  * hold on the peer_ni.
1760  */
1761 struct lnet_peer_ni *
1762 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1763 {
1764         struct lnet_peer_ni *lpni = NULL;
1765         int rc;
1766
1767         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1768                 return ERR_PTR(-ESHUTDOWN);
1769
1770         /*
1771          * find if a peer_ni already exists.
1772          * If so then just return that.
1773          */
1774         lpni = lnet_find_peer_ni_locked(nid);
1775         if (lpni)
1776                 return lpni;
1777
1778         /*
1779          * Slow path:
1780          * use the lnet_api_mutex to serialize the creation of the peer_ni
1781          * and the creation/deletion of the local ni/net. When a local ni is
1782          * created, if there exists a set of peer_nis on that network,
1783          * they need to be traversed and updated. When a local NI is
1784          * deleted, which could result in a network being deleted, then
1785          * all peer nis on that network need to be removed as well.
1786          *
1787          * Creation through traffic should also be serialized with
1788          * creation through DLC.
1789          */
1790         lnet_net_unlock(cpt);
1791         mutex_lock(&the_lnet.ln_api_mutex);
1792         /*
1793          * Shutdown is only set under the ln_api_lock, so a single
1794          * check here is sufficent.
1795          */
1796         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1797                 lpni = ERR_PTR(-ESHUTDOWN);
1798                 goto out_mutex_unlock;
1799         }
1800
1801         rc = lnet_peer_ni_traffic_add(nid, pref);
1802         if (rc) {
1803                 lpni = ERR_PTR(rc);
1804                 goto out_mutex_unlock;
1805         }
1806
1807         lpni = lnet_find_peer_ni_locked(nid);
1808         LASSERT(lpni);
1809
1810 out_mutex_unlock:
1811         mutex_unlock(&the_lnet.ln_api_mutex);
1812         lnet_net_lock(cpt);
1813
1814         /* Lock has been dropped, check again for shutdown. */
1815         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1816                 if (!IS_ERR(lpni))
1817                         lnet_peer_ni_decref_locked(lpni);
1818                 lpni = ERR_PTR(-ESHUTDOWN);
1819         }
1820
1821         return lpni;
1822 }
1823
1824 bool
1825 lnet_peer_gw_discovery(struct lnet_peer *lp)
1826 {
1827         bool rc = false;
1828
1829         spin_lock(&lp->lp_lock);
1830         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1831                 rc = true;
1832         spin_unlock(&lp->lp_lock);
1833
1834         return rc;
1835 }
1836
1837 bool
1838 lnet_peer_is_uptodate(struct lnet_peer *lp)
1839 {
1840         bool rc;
1841
1842         spin_lock(&lp->lp_lock);
1843         rc = lnet_peer_is_uptodate_locked(lp);
1844         spin_unlock(&lp->lp_lock);
1845         return rc;
1846 }
1847
1848 /*
1849  * Is a peer uptodate from the point of view of discovery?
1850  *
1851  * If it is currently being processed, obviously not.
1852  * A forced Ping or Push is also handled by the discovery thread.
1853  *
1854  * Otherwise look at whether the peer needs rediscovering.
1855  */
1856 bool
1857 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1858 __must_hold(&lp->lp_lock)
1859 {
1860         bool rc;
1861
1862         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1863                             LNET_PEER_FORCE_PING |
1864                             LNET_PEER_FORCE_PUSH)) {
1865                 rc = false;
1866         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1867                 rc = false;
1868         } else if (lnet_peer_needs_push(lp)) {
1869                 rc = false;
1870         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1871                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1872                         rc = true;
1873                 else
1874                         rc = false;
1875         } else {
1876                 rc = false;
1877         }
1878
1879         return rc;
1880 }
1881
1882 /*
1883  * Queue a peer for the attention of the discovery thread.  Call with
1884  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1885  * -EALREADY if the peer was already queued.
1886  */
1887 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1888 {
1889         int rc;
1890
1891         spin_lock(&lp->lp_lock);
1892         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1893                 lp->lp_state |= LNET_PEER_DISCOVERING;
1894         spin_unlock(&lp->lp_lock);
1895         if (list_empty(&lp->lp_dc_list)) {
1896                 lnet_peer_addref_locked(lp);
1897                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1898                 wake_up(&the_lnet.ln_dc_waitq);
1899                 rc = 0;
1900         } else {
1901                 rc = -EALREADY;
1902         }
1903
1904         CDEBUG(D_NET, "Queue peer %s: %d\n",
1905                libcfs_nid2str(lp->lp_primary_nid), rc);
1906
1907         return rc;
1908 }
1909
1910 /*
1911  * Discovery of a peer is complete. Wake all waiters on the peer.
1912  * Call with lnet_net_lock/EX held.
1913  */
1914 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1915 {
1916         struct lnet_msg *msg, *tmp;
1917         int rc = 0;
1918         LIST_HEAD(pending_msgs);
1919
1920         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1921                libcfs_nid2str(lp->lp_primary_nid));
1922
1923         list_del_init(&lp->lp_dc_list);
1924         spin_lock(&lp->lp_lock);
1925         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1926         spin_unlock(&lp->lp_lock);
1927         wake_up_all(&lp->lp_dc_waitq);
1928
1929         if (lp->lp_rtr_refcount > 0)
1930                 lnet_router_discovery_complete(lp);
1931
1932         lnet_net_unlock(LNET_LOCK_EX);
1933
1934         /* iterate through all pending messages and send them again */
1935         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1936                 list_del_init(&msg->msg_list);
1937                 if (lp->lp_dc_error) {
1938                         lnet_finalize(msg, lp->lp_dc_error);
1939                         continue;
1940                 }
1941
1942                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1943                        lnet_msgtyp2str(msg->msg_type),
1944                        libcfs_id2str(msg->msg_target));
1945                 rc = lnet_send(msg->msg_src_nid_param, msg,
1946                                msg->msg_rtr_nid_param);
1947                 if (rc < 0) {
1948                         CNETERR("Error sending %s to %s: %d\n",
1949                                lnet_msgtyp2str(msg->msg_type),
1950                                libcfs_id2str(msg->msg_target), rc);
1951                         lnet_finalize(msg, rc);
1952                 }
1953         }
1954         lnet_net_lock(LNET_LOCK_EX);
1955         lnet_peer_decref_locked(lp);
1956 }
1957
1958 /*
1959  * Handle inbound push.
1960  * Like any event handler, called with lnet_res_lock/CPT held.
1961  */
1962 void lnet_peer_push_event(struct lnet_event *ev)
1963 {
1964         struct lnet_ping_buffer *pbuf;
1965         struct lnet_peer *lp;
1966
1967         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
1968
1969         /* lnet_find_peer() adds a refcount */
1970         lp = lnet_find_peer(ev->source.nid);
1971         if (!lp) {
1972                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1973                        libcfs_nid2str(ev->initiator.nid),
1974                        libcfs_nid2str(ev->source.nid));
1975                 pbuf->pb_needs_post = true;
1976                 return;
1977         }
1978
1979         /* Ensure peer state remains consistent while we modify it. */
1980         spin_lock(&lp->lp_lock);
1981
1982         /*
1983          * If some kind of error happened the contents of the message
1984          * cannot be used. Clear the NIDS_UPTODATE and set the
1985          * FORCE_PING flag to trigger a ping.
1986          */
1987         if (ev->status) {
1988                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1989                 lp->lp_state |= LNET_PEER_FORCE_PING;
1990                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1991                        ev->status,
1992                        libcfs_nid2str(lp->lp_primary_nid),
1993                        libcfs_nid2str(ev->source.nid));
1994                 goto out;
1995         }
1996
1997         /*
1998          * A push with invalid or corrupted info. Clear the UPTODATE
1999          * flag to trigger a ping.
2000          */
2001         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2002                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2003                 lp->lp_state |= LNET_PEER_FORCE_PING;
2004                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2005                        libcfs_nid2str(lp->lp_primary_nid));
2006                 goto out;
2007         }
2008
2009         /*
2010          * Make sure we'll allocate the correct size ping buffer when
2011          * pinging the peer.
2012          */
2013         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2014                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2015
2016         /*
2017          * A non-Multi-Rail peer is not supposed to be capable of
2018          * sending a push.
2019          */
2020         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2021                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2022                        libcfs_nid2str(lp->lp_primary_nid));
2023                 goto out;
2024         }
2025
2026         /*
2027          * The peer may have discovery disabled at its end. Set
2028          * NO_DISCOVERY as appropriate.
2029          */
2030         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2031                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2032                        libcfs_nid2str(lp->lp_primary_nid));
2033                 /*
2034                  * If the peer is going from discovery enabled to
2035                  * discovery disabled, we need to reflect that in our
2036                  * representation of the peer.
2037                  */
2038                 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2039                                       LNET_PEER_DISCOVERING)))
2040                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2041                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2042         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2043                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2044                        libcfs_nid2str(lp->lp_primary_nid));
2045                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2046         }
2047
2048         /*
2049          * Update the MULTI_RAIL flag based on the push. If the peer
2050          * was configured with DLC then the setting should match what
2051          * DLC put in.
2052          * NB: We verified above that the MR feature bit is set in pi_features
2053          */
2054         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2055                 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2056                        libcfs_nid2str(lp->lp_primary_nid), lp);
2057         } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2058                 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2059                       libcfs_nid2str(lp->lp_primary_nid));
2060         } else if (lnet_peer_discovery_disabled) {
2061                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2062                        libcfs_nid2str(lp->lp_primary_nid), lp);
2063         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2064                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2065                        libcfs_nid2str(lp->lp_primary_nid), lp);
2066         } else {
2067                 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2068                        libcfs_nid2str(lp->lp_primary_nid), lp);
2069                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2070                 lnet_peer_clr_non_mr_pref_nids(lp);
2071         }
2072
2073         /*
2074          * Check for truncation of the Put message. Clear the
2075          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2076          * and tell discovery to allocate a bigger buffer.
2077          */
2078         if (ev->mlength < ev->rlength) {
2079                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2080                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2081                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2082                 lp->lp_state |= LNET_PEER_FORCE_PING;
2083                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2084                        libcfs_nid2str(lp->lp_primary_nid),
2085                        pbuf->pb_info.pi_nnis);
2086                 goto out;
2087         }
2088
2089         /* always assume new data */
2090         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2091         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2092
2093         /*
2094          * If there is data present that hasn't been processed yet,
2095          * we'll replace it if the Put contained newer data and it
2096          * fits. We're racing with a Ping or earlier Push in this
2097          * case.
2098          */
2099         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2100                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2101                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2102                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2103                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2104                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2105                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2106                               libcfs_nid2str(lp->lp_primary_nid),
2107                               LNET_PING_BUFFER_SEQNO(pbuf),
2108                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2109                 }
2110                 goto out;
2111         }
2112
2113         /*
2114          * Allocate a buffer to copy the data. On a failure we drop
2115          * the Push and set FORCE_PING to force the discovery
2116          * thread to fix the problem by pinging the peer.
2117          */
2118         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2119         if (!lp->lp_data) {
2120                 lp->lp_state |= LNET_PEER_FORCE_PING;
2121                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2122                        libcfs_nid2str(lp->lp_primary_nid),
2123                        LNET_PING_BUFFER_SEQNO(pbuf));
2124                 goto out;
2125         }
2126
2127         /* Success */
2128         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2129                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2130         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2131         CDEBUG(D_NET, "Received Push %s %u\n",
2132                libcfs_nid2str(lp->lp_primary_nid),
2133                LNET_PING_BUFFER_SEQNO(pbuf));
2134
2135 out:
2136         /* We've processed this buffer. It can be reposted */
2137         pbuf->pb_needs_post = true;
2138
2139         /*
2140          * Queue the peer for discovery if not done, force it on the request
2141          * queue and wake the discovery thread if the peer was already queued,
2142          * because its status changed.
2143          */
2144         spin_unlock(&lp->lp_lock);
2145         lnet_net_lock(LNET_LOCK_EX);
2146         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2147                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2148                 wake_up(&the_lnet.ln_dc_waitq);
2149         }
2150         /* Drop refcount from lookup */
2151         lnet_peer_decref_locked(lp);
2152         lnet_net_unlock(LNET_LOCK_EX);
2153 }
2154
2155 /*
2156  * Clear the discovery error state, unless we're already discovering
2157  * this peer, in which case the error is current.
2158  */
2159 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2160 {
2161         spin_lock(&lp->lp_lock);
2162         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2163                 lp->lp_dc_error = 0;
2164         spin_unlock(&lp->lp_lock);
2165 }
2166
2167 /*
2168  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2169  * dropped/retaken within this function. An lnet_peer_ni is passed in
2170  * because discovery could tear down an lnet_peer.
2171  */
2172 int
2173 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2174 {
2175         DEFINE_WAIT(wait);
2176         struct lnet_peer *lp;
2177         int rc = 0;
2178         int count = 0;
2179
2180 again:
2181         lnet_net_unlock(cpt);
2182         lnet_net_lock(LNET_LOCK_EX);
2183         lp = lpni->lpni_peer_net->lpn_peer;
2184         lnet_peer_clear_discovery_error(lp);
2185
2186         /*
2187          * We're willing to be interrupted. The lpni can become a
2188          * zombie if we race with DLC, so we must check for that.
2189          */
2190         for (;;) {
2191                 /* Keep lp alive when the lnet_net_lock is unlocked */
2192                 lnet_peer_addref_locked(lp);
2193                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2194                 if (signal_pending(current))
2195                         break;
2196                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2197                         break;
2198                 /*
2199                  * Don't repeat discovery if discovery is disabled. This is
2200                  * done to ensure we can use discovery as a standard ping as
2201                  * well for backwards compatibility with routers which do not
2202                  * have discovery or have discovery disabled
2203                  */
2204                 if (lnet_is_discovery_disabled(lp) && count > 0)
2205                         break;
2206                 if (lp->lp_dc_error)
2207                         break;
2208                 if (lnet_peer_is_uptodate(lp))
2209                         break;
2210                 lnet_peer_queue_for_discovery(lp);
2211                 count++;
2212                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2213
2214                 /*
2215                  * If caller requested a non-blocking operation then
2216                  * return immediately. Once discovery is complete any
2217                  * pending messages that were stopped due to discovery
2218                  * will be transmitted.
2219                  */
2220                 if (!block)
2221                         break;
2222
2223                 lnet_net_unlock(LNET_LOCK_EX);
2224                 schedule();
2225                 finish_wait(&lp->lp_dc_waitq, &wait);
2226                 lnet_net_lock(LNET_LOCK_EX);
2227                 lnet_peer_decref_locked(lp);
2228                 /* Peer may have changed */
2229                 lp = lpni->lpni_peer_net->lpn_peer;
2230         }
2231         finish_wait(&lp->lp_dc_waitq, &wait);
2232
2233         lnet_net_unlock(LNET_LOCK_EX);
2234         lnet_net_lock(cpt);
2235         lnet_peer_decref_locked(lp);
2236         /*
2237          * The peer may have changed, so re-check and rediscover if that turns
2238          * out to have been the case. The reference count on lp ensured that
2239          * even if it was unlinked from lpni the memory could not be recycled.
2240          * Thus the check below is sufficient to determine whether the peer
2241          * changed. If the peer changed, then lp must not be dereferenced.
2242          */
2243         if (lp != lpni->lpni_peer_net->lpn_peer)
2244                 goto again;
2245
2246         if (signal_pending(current))
2247                 rc = -EINTR;
2248         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2249                 rc = -ESHUTDOWN;
2250         else if (lp->lp_dc_error)
2251                 rc = lp->lp_dc_error;
2252         else if (!block)
2253                 CDEBUG(D_NET, "non-blocking discovery\n");
2254         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2255                 goto again;
2256
2257         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2258                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2259                libcfs_nid2str(lpni->lpni_nid), rc,
2260                (!block) ? "pending discovery" : "discovery complete");
2261
2262         return rc;
2263 }
2264
2265 /* Handle an incoming ack for a push. */
2266 static void
2267 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2268 {
2269         struct lnet_ping_buffer *pbuf;
2270
2271         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2272         spin_lock(&lp->lp_lock);
2273         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2274         lp->lp_push_error = ev->status;
2275         if (ev->status)
2276                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2277         else
2278                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2279         spin_unlock(&lp->lp_lock);
2280
2281         CDEBUG(D_NET, "peer %s ev->status %d\n",
2282                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2283 }
2284
2285 /* Handle a Reply message. This is the reply to a Ping message. */
2286 static void
2287 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2288 {
2289         struct lnet_ping_buffer *pbuf;
2290         int rc;
2291
2292         spin_lock(&lp->lp_lock);
2293
2294         /*
2295          * If some kind of error happened the contents of message
2296          * cannot be used. Set PING_FAILED to trigger a retry.
2297          */
2298         if (ev->status) {
2299                 lp->lp_state |= LNET_PEER_PING_FAILED;
2300                 lp->lp_ping_error = ev->status;
2301                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2302                        ev->status,
2303                        libcfs_nid2str(lp->lp_primary_nid),
2304                        libcfs_nid2str(ev->source.nid));
2305                 goto out;
2306         }
2307
2308         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2309         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2310                 lnet_swap_pinginfo(pbuf);
2311
2312         /*
2313          * A reply with invalid or corrupted info. Set PING_FAILED to
2314          * trigger a retry.
2315          */
2316         rc = lnet_ping_info_validate(&pbuf->pb_info);
2317         if (rc) {
2318                 lp->lp_state |= LNET_PEER_PING_FAILED;
2319                 lp->lp_ping_error = 0;
2320                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2321                        libcfs_nid2str(lp->lp_primary_nid), rc);
2322                 goto out;
2323         }
2324
2325
2326         /*
2327          * The peer may have discovery disabled at its end. Set
2328          * NO_DISCOVERY as appropriate.
2329          */
2330         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2331             !lnet_peer_discovery_disabled) {
2332                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2333                        libcfs_nid2str(lp->lp_primary_nid));
2334                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2335         } else {
2336                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2337                        libcfs_nid2str(lp->lp_primary_nid));
2338                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2339         }
2340
2341         /*
2342          * Update the MULTI_RAIL flag based on the reply. If the peer
2343          * was configured with DLC then the setting should match what
2344          * DLC put in.
2345          */
2346         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2347                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2348                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2349                                libcfs_nid2str(lp->lp_primary_nid), lp);
2350                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2351                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2352                               libcfs_nid2str(lp->lp_primary_nid));
2353                 } else if (lnet_peer_discovery_disabled) {
2354                         CDEBUG(D_NET,
2355                                "peer %s(%p) not MR: DD disabled locally\n",
2356                                libcfs_nid2str(lp->lp_primary_nid), lp);
2357                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2358                         CDEBUG(D_NET,
2359                                "peer %s(%p) not MR: DD disabled remotely\n",
2360                                libcfs_nid2str(lp->lp_primary_nid), lp);
2361                 } else {
2362                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2363                                libcfs_nid2str(lp->lp_primary_nid), lp);
2364                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2365                         lnet_peer_clr_non_mr_pref_nids(lp);
2366                 }
2367         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2368                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2369                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2370                               libcfs_nid2str(lp->lp_primary_nid));
2371                 } else {
2372                         CERROR("Multi-Rail state vanished from %s\n",
2373                                libcfs_nid2str(lp->lp_primary_nid));
2374                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2375                 }
2376         }
2377
2378         /*
2379          * Make sure we'll allocate the correct size ping buffer when
2380          * pinging the peer.
2381          */
2382         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2383                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2384
2385         /*
2386          * Check for truncation of the Reply. Clear PING_SENT and set
2387          * PING_FAILED to trigger a retry.
2388          */
2389         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2390                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2391                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2392                 lp->lp_state |= LNET_PEER_PING_FAILED;
2393                 lp->lp_ping_error = 0;
2394                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2395                        libcfs_nid2str(lp->lp_primary_nid),
2396                        pbuf->pb_info.pi_nnis);
2397                 goto out;
2398         }
2399
2400         /*
2401          * Check the sequence numbers in the reply. These are only
2402          * available if the reply came from a Multi-Rail peer.
2403          */
2404         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2405             pbuf->pb_info.pi_nnis > 1 &&
2406             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2407                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2408                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2409                                 libcfs_nid2str(lp->lp_primary_nid),
2410                                 LNET_PING_BUFFER_SEQNO(pbuf),
2411                                 lp->lp_peer_seqno);
2412
2413                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2414         }
2415
2416         /* We're happy with the state of the data in the buffer. */
2417         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2418                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2419         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2420                 lnet_ping_buffer_decref(lp->lp_data);
2421         else
2422                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2423         lnet_ping_buffer_addref(pbuf);
2424         lp->lp_data = pbuf;
2425 out:
2426         lp->lp_state &= ~LNET_PEER_PING_SENT;
2427         spin_unlock(&lp->lp_lock);
2428
2429         lnet_net_lock(LNET_LOCK_EX);
2430         /*
2431          * If this peer is a gateway, call the routing callback to
2432          * handle the ping reply
2433          */
2434         if (lp->lp_rtr_refcount > 0)
2435                 lnet_router_discovery_ping_reply(lp);
2436         lnet_net_unlock(LNET_LOCK_EX);
2437 }
2438
2439 /*
2440  * Send event handling. Only matters for error cases, where we clean
2441  * up state on the peer and peer_ni that would otherwise be updated in
2442  * the REPLY event handler for a successful Ping, and the ACK event
2443  * handler for a successful Push.
2444  */
2445 static int
2446 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2447 {
2448         int rc = 0;
2449
2450         if (!ev->status)
2451                 goto out;
2452
2453         spin_lock(&lp->lp_lock);
2454         if (ev->msg_type == LNET_MSG_GET) {
2455                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2456                 lp->lp_state |= LNET_PEER_PING_FAILED;
2457                 lp->lp_ping_error = ev->status;
2458         } else { /* ev->msg_type == LNET_MSG_PUT */
2459                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2460                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2461                 lp->lp_push_error = ev->status;
2462         }
2463         spin_unlock(&lp->lp_lock);
2464         rc = LNET_REDISCOVER_PEER;
2465 out:
2466         CDEBUG(D_NET, "%s Send to %s: %d\n",
2467                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2468                 libcfs_nid2str(ev->target.nid), rc);
2469         return rc;
2470 }
2471
2472 /*
2473  * Unlink event handling. This event is only seen if a call to
2474  * LNetMDUnlink() caused the event to be unlinked. If this call was
2475  * made after the event was set up in LNetGet() or LNetPut() then we
2476  * assume the Ping or Push timed out.
2477  */
2478 static void
2479 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2480 {
2481         spin_lock(&lp->lp_lock);
2482         /* We've passed through LNetGet() */
2483         if (lp->lp_state & LNET_PEER_PING_SENT) {
2484                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2485                 lp->lp_state |= LNET_PEER_PING_FAILED;
2486                 lp->lp_ping_error = -ETIMEDOUT;
2487                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2488                         libcfs_nid2str(lp->lp_primary_nid));
2489         }
2490         /* We've passed through LNetPut() */
2491         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2492                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2493                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2494                 lp->lp_push_error = -ETIMEDOUT;
2495                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2496                         libcfs_nid2str(lp->lp_primary_nid));
2497         }
2498         spin_unlock(&lp->lp_lock);
2499 }
2500
2501 /*
2502  * Event handler for the discovery EQ.
2503  *
2504  * Called with lnet_res_lock(cpt) held. The cpt is the
2505  * lnet_cpt_of_cookie() of the md handle cookie.
2506  */
2507 static void lnet_discovery_event_handler(struct lnet_event *event)
2508 {
2509         struct lnet_peer *lp = event->md_user_ptr;
2510         struct lnet_ping_buffer *pbuf;
2511         int rc;
2512
2513         /* discovery needs to take another look */
2514         rc = LNET_REDISCOVER_PEER;
2515
2516         CDEBUG(D_NET, "Received event: %d\n", event->type);
2517
2518         switch (event->type) {
2519         case LNET_EVENT_ACK:
2520                 lnet_discovery_event_ack(lp, event);
2521                 break;
2522         case LNET_EVENT_REPLY:
2523                 lnet_discovery_event_reply(lp, event);
2524                 break;
2525         case LNET_EVENT_SEND:
2526                 /* Only send failure triggers a retry. */
2527                 rc = lnet_discovery_event_send(lp, event);
2528                 break;
2529         case LNET_EVENT_UNLINK:
2530                 /* LNetMDUnlink() was called */
2531                 lnet_discovery_event_unlink(lp, event);
2532                 break;
2533         default:
2534                 /* Invalid events. */
2535                 LBUG();
2536         }
2537         lnet_net_lock(LNET_LOCK_EX);
2538         if (event->unlinked) {
2539                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2540                 lnet_ping_buffer_decref(pbuf);
2541                 lnet_peer_decref_locked(lp);
2542         }
2543
2544         /* put peer back at end of request queue, if discovery not already
2545          * done */
2546         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2547                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2548                 wake_up(&the_lnet.ln_dc_waitq);
2549         }
2550         lnet_net_unlock(LNET_LOCK_EX);
2551 }
2552
2553 /*
2554  * Build a peer from incoming data.
2555  *
2556  * The NIDs in the incoming data are supposed to be structured as follows:
2557  *  - loopback
2558  *  - primary NID
2559  *  - other NIDs in same net
2560  *  - NIDs in second net
2561  *  - NIDs in third net
2562  *  - ...
2563  * This due to the way the list of NIDs in the data is created.
2564  *
2565  * Note that this function will mark the peer uptodate unless an
2566  * ENOMEM is encontered. All other errors are due to a conflict
2567  * between the DLC configuration and what discovery sees. We treat DLC
2568  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2569  * peer from becoming stuck in discovery.
2570  */
2571 static int lnet_peer_merge_data(struct lnet_peer *lp,
2572                                 struct lnet_ping_buffer *pbuf)
2573 {
2574         struct lnet_peer_ni *lpni;
2575         lnet_nid_t *curnis = NULL;
2576         struct lnet_ni_status *addnis = NULL;
2577         lnet_nid_t *delnis = NULL;
2578         unsigned flags;
2579         int ncurnis;
2580         int naddnis;
2581         int ndelnis;
2582         int nnis = 0;
2583         int i;
2584         int j;
2585         int rc;
2586
2587         flags = LNET_PEER_DISCOVERED;
2588         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2589                 flags |= LNET_PEER_MULTI_RAIL;
2590
2591         /*
2592          * Cache the routing feature for the peer; whether it is enabled
2593          * for disabled as reported by the remote peer.
2594          */
2595         spin_lock(&lp->lp_lock);
2596         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2597                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2598         else
2599                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2600         spin_unlock(&lp->lp_lock);
2601
2602         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2603         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2604         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2605         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2606         if (!curnis || !addnis || !delnis) {
2607                 rc = -ENOMEM;
2608                 goto out;
2609         }
2610         ncurnis = 0;
2611         naddnis = 0;
2612         ndelnis = 0;
2613
2614         /* Construct the list of NIDs present in peer. */
2615         lpni = NULL;
2616         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2617                 curnis[ncurnis++] = lpni->lpni_nid;
2618
2619         /*
2620          * Check for NIDs in pbuf not present in curnis[].
2621          * The loop starts at 1 to skip the loopback NID.
2622          */
2623         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2624                 for (j = 0; j < ncurnis; j++)
2625                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2626                                 break;
2627                 if (j == ncurnis)
2628                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2629         }
2630         /*
2631          * Check for NIDs in curnis[] not present in pbuf.
2632          * The nested loop starts at 1 to skip the loopback NID.
2633          *
2634          * But never add the loopback NID to delnis[]: if it is
2635          * present in curnis[] then this peer is for this node.
2636          */
2637         for (i = 0; i < ncurnis; i++) {
2638                 if (curnis[i] == LNET_NID_LO_0)
2639                         continue;
2640                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2641                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2642                                 /*
2643                                  * update the information we cache for the
2644                                  * peer with the latest information we
2645                                  * received
2646                                  */
2647                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2648                                 if (lpni) {
2649                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2650                                         lnet_peer_ni_decref_locked(lpni);
2651                                 }
2652                                 break;
2653                         }
2654                 }
2655                 if (j == pbuf->pb_info.pi_nnis)
2656                         delnis[ndelnis++] = curnis[i];
2657         }
2658
2659         /*
2660          * If we get here and the discovery is disabled then we don't want
2661          * to add or delete any NIs. We just updated the ones we have some
2662          * information on, and call it a day
2663          */
2664         rc = 0;
2665         if (lnet_is_discovery_disabled(lp))
2666                 goto out;
2667
2668         for (i = 0; i < naddnis; i++) {
2669                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2670                 if (rc) {
2671                         CERROR("Error adding NID %s to peer %s: %d\n",
2672                                libcfs_nid2str(addnis[i].ns_nid),
2673                                libcfs_nid2str(lp->lp_primary_nid), rc);
2674                         if (rc == -ENOMEM)
2675                                 goto out;
2676                 }
2677                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2678                 if (lpni) {
2679                         lpni->lpni_ns_status = addnis[i].ns_status;
2680                         lnet_peer_ni_decref_locked(lpni);
2681                 }
2682         }
2683
2684         for (i = 0; i < ndelnis; i++) {
2685                 /*
2686                  * for routers it's okay to delete the primary_nid because
2687                  * the upper layers don't really rely on it. So if we're
2688                  * being told that the router changed its primary_nid
2689                  * then it's okay to delete it.
2690                  */
2691                 if (lp->lp_rtr_refcount > 0)
2692                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2693                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2694                 if (rc) {
2695                         CERROR("Error deleting NID %s from peer %s: %d\n",
2696                                libcfs_nid2str(delnis[i]),
2697                                libcfs_nid2str(lp->lp_primary_nid), rc);
2698                         if (rc == -ENOMEM)
2699                                 goto out;
2700                 }
2701         }
2702         /*
2703          * Errors other than -ENOMEM are due to peers having been
2704          * configured with DLC. Ignore these because DLC overrides
2705          * Discovery.
2706          */
2707         rc = 0;
2708 out:
2709         CFS_FREE_PTR_ARRAY(curnis, nnis);
2710         CFS_FREE_PTR_ARRAY(addnis, nnis);
2711         CFS_FREE_PTR_ARRAY(delnis, nnis);
2712         lnet_ping_buffer_decref(pbuf);
2713         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2714
2715         if (rc) {
2716                 spin_lock(&lp->lp_lock);
2717                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2718                 lp->lp_state |= LNET_PEER_FORCE_PING;
2719                 spin_unlock(&lp->lp_lock);
2720         }
2721         return rc;
2722 }
2723
2724 /*
2725  * The data in pbuf says lp is its primary peer, but the data was
2726  * received by a different peer. Try to update lp with the data.
2727  */
2728 static int
2729 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2730 {
2731         struct lnet_handle_md mdh;
2732
2733         /* Queue lp for discovery, and force it on the request queue. */
2734         lnet_net_lock(LNET_LOCK_EX);
2735         if (lnet_peer_queue_for_discovery(lp))
2736                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2737         lnet_net_unlock(LNET_LOCK_EX);
2738
2739         LNetInvalidateMDHandle(&mdh);
2740
2741         /*
2742          * Decide whether we can move the peer to the DATA_PRESENT state.
2743          *
2744          * We replace stale data for a multi-rail peer, repair PING_FAILED
2745          * status, and preempt FORCE_PING.
2746          *
2747          * If after that we have DATA_PRESENT, we merge it into this peer.
2748          */
2749         spin_lock(&lp->lp_lock);
2750         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2751                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2752                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2753                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2754                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2755                         lnet_ping_buffer_decref(pbuf);
2756                         pbuf = lp->lp_data;
2757                         lp->lp_data = NULL;
2758                 }
2759         }
2760         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2761                 lnet_ping_buffer_decref(lp->lp_data);
2762                 lp->lp_data = NULL;
2763                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2764         }
2765         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2766                 mdh = lp->lp_ping_mdh;
2767                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2768                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2769                 lp->lp_ping_error = 0;
2770         }
2771         if (lp->lp_state & LNET_PEER_FORCE_PING)
2772                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2773         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2774         spin_unlock(&lp->lp_lock);
2775
2776         if (!LNetMDHandleIsInvalid(mdh))
2777                 LNetMDUnlink(mdh);
2778
2779         if (pbuf)
2780                 return lnet_peer_merge_data(lp, pbuf);
2781
2782         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2783         return 0;
2784 }
2785
2786 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2787 {
2788         int i;
2789
2790         for (i = 0; i < pinfo->pi_nnis; i++) {
2791                 if (pinfo->pi_ni[i].ns_nid == nid)
2792                         return true;
2793         }
2794
2795         return false;
2796 }
2797
2798 /*
2799  * Update a peer using the data received.
2800  */
2801 static int lnet_peer_data_present(struct lnet_peer *lp)
2802 __must_hold(&lp->lp_lock)
2803 {
2804         struct lnet_ping_buffer *pbuf;
2805         struct lnet_peer_ni *lpni;
2806         lnet_nid_t nid = LNET_NID_ANY;
2807         unsigned flags;
2808         int rc = 0;
2809
2810         pbuf = lp->lp_data;
2811         lp->lp_data = NULL;
2812         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2813         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2814         spin_unlock(&lp->lp_lock);
2815
2816         /*
2817          * Modifications of peer structures are done while holding the
2818          * ln_api_mutex. A global lock is required because we may be
2819          * modifying multiple peer structures, and a mutex greatly
2820          * simplifies memory management.
2821          *
2822          * The actual changes to the data structures must also protect
2823          * against concurrent lookups, for which the lnet_net_lock in
2824          * LNET_LOCK_EX mode is used.
2825          */
2826         mutex_lock(&the_lnet.ln_api_mutex);
2827         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2828                 rc = -ESHUTDOWN;
2829                 goto out;
2830         }
2831
2832         /*
2833          * If this peer is not on the peer list then it is being torn
2834          * down, and our reference count may be all that is keeping it
2835          * alive. Don't do any work on it.
2836          */
2837         if (list_empty(&lp->lp_peer_list))
2838                 goto out;
2839
2840         flags = LNET_PEER_DISCOVERED;
2841         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2842                 flags |= LNET_PEER_MULTI_RAIL;
2843
2844         /*
2845          * Check whether the primary NID in the message matches the
2846          * primary NID of the peer. If it does, update the peer, if
2847          * it it does not, check whether there is already a peer with
2848          * that primary NID. If no such peer exists, try to update
2849          * the primary NID of the current peer (allowed if it was
2850          * created due to message traffic) and complete the update.
2851          * If the peer did exist, hand off the data to it.
2852          *
2853          * The peer for the loopback interface is a special case: this
2854          * is the peer for the local node, and we want to set its
2855          * primary NID to the correct value here. Moreover, this peer
2856          * can show up with only the loopback NID in the ping buffer.
2857          */
2858         if (pbuf->pb_info.pi_nnis <= 1)
2859                 goto out;
2860         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2861         if (lp->lp_primary_nid == LNET_NID_LO_0) {
2862                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2863                 if (!rc)
2864                         rc = lnet_peer_merge_data(lp, pbuf);
2865         /*
2866          * if the primary nid of the peer is present in the ping info returned
2867          * from the peer, but it's not the local primary peer we have
2868          * cached and discovery is disabled, then we don't want to update
2869          * our local peer info, by adding or removing NIDs, we just want
2870          * to update the status of the nids that we currently have
2871          * recorded in that peer.
2872          */
2873         } else if (lp->lp_primary_nid == nid ||
2874                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2875                     lnet_is_discovery_disabled(lp))) {
2876                 rc = lnet_peer_merge_data(lp, pbuf);
2877         } else {
2878                 lpni = lnet_find_peer_ni_locked(nid);
2879                 if (!lpni) {
2880                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2881                         if (rc) {
2882                                 CERROR("Primary NID error %s versus %s: %d\n",
2883                                        libcfs_nid2str(lp->lp_primary_nid),
2884                                        libcfs_nid2str(nid), rc);
2885                         } else {
2886                                 rc = lnet_peer_merge_data(lp, pbuf);
2887                         }
2888                 } else {
2889                         struct lnet_peer *new_lp;
2890                         new_lp = lpni->lpni_peer_net->lpn_peer;
2891                         /*
2892                          * if lp has discovery/MR enabled that means new_lp
2893                          * should have discovery/MR enabled as well, since
2894                          * it's the same peer, which we're about to merge
2895                          */
2896                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2897                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2898                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2899                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2900
2901                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2902                         lnet_consolidate_routes_locked(lp, new_lp);
2903                         lnet_peer_ni_decref_locked(lpni);
2904                 }
2905         }
2906 out:
2907         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2908                lp->lp_state);
2909         mutex_unlock(&the_lnet.ln_api_mutex);
2910
2911         spin_lock(&lp->lp_lock);
2912         /* Tell discovery to re-check the peer immediately. */
2913         if (!rc)
2914                 rc = LNET_REDISCOVER_PEER;
2915         return rc;
2916 }
2917
2918 /*
2919  * A ping failed. Clear the PING_FAILED state and set the
2920  * FORCE_PING state, to ensure a retry even if discovery is
2921  * disabled. This avoids being left with incorrect state.
2922  */
2923 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2924 __must_hold(&lp->lp_lock)
2925 {
2926         struct lnet_handle_md mdh;
2927         int rc;
2928
2929         mdh = lp->lp_ping_mdh;
2930         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2931         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2932         lp->lp_state |= LNET_PEER_FORCE_PING;
2933         rc = lp->lp_ping_error;
2934         lp->lp_ping_error = 0;
2935         spin_unlock(&lp->lp_lock);
2936
2937         if (!LNetMDHandleIsInvalid(mdh))
2938                 LNetMDUnlink(mdh);
2939
2940         CDEBUG(D_NET, "peer %s:%d\n",
2941                libcfs_nid2str(lp->lp_primary_nid), rc);
2942
2943         spin_lock(&lp->lp_lock);
2944         return rc ? rc : LNET_REDISCOVER_PEER;
2945 }
2946
2947 /*
2948  * Select NID to send a Ping or Push to.
2949  */
2950 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2951 {
2952         struct lnet_peer_ni *lpni;
2953
2954         /* Look for a direct-connected NID for this peer. */
2955         lpni = NULL;
2956         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2957                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2958                         continue;
2959                 break;
2960         }
2961         if (lpni)
2962                 return lpni->lpni_nid;
2963
2964         /* Look for a routed-connected NID for this peer. */
2965         lpni = NULL;
2966         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2967                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2968                         continue;
2969                 break;
2970         }
2971         if (lpni)
2972                 return lpni->lpni_nid;
2973
2974         return LNET_NID_ANY;
2975 }
2976
2977 /* Active side of ping. */
2978 static int lnet_peer_send_ping(struct lnet_peer *lp)
2979 __must_hold(&lp->lp_lock)
2980 {
2981         lnet_nid_t pnid;
2982         int nnis;
2983         int rc;
2984         int cpt;
2985
2986         lp->lp_state |= LNET_PEER_PING_SENT;
2987         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2988         spin_unlock(&lp->lp_lock);
2989
2990         cpt = lnet_net_lock_current();
2991         /* Refcount for MD. */
2992         lnet_peer_addref_locked(lp);
2993         pnid = lnet_peer_select_nid(lp);
2994         lnet_net_unlock(cpt);
2995
2996         nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2997
2998         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2999                             the_lnet.ln_dc_handler, false);
3000
3001         /*
3002          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3003          * refcount on the peer, otherwise LNetMDUnlink will be called
3004          * which will eventually do that.
3005          */
3006         if (rc > 0) {
3007                 lnet_net_lock(cpt);
3008                 lnet_peer_decref_locked(lp);
3009                 lnet_net_unlock(cpt);
3010                 rc = -rc; /* change the rc to negative value */
3011                 goto fail_error;
3012         } else if (rc < 0) {
3013                 goto fail_error;
3014         }
3015
3016         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3017
3018         spin_lock(&lp->lp_lock);
3019         return 0;
3020
3021 fail_error:
3022         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3023         /*
3024          * The errors that get us here are considered hard errors and
3025          * cause Discovery to terminate. So we clear PING_SENT, but do
3026          * not set either PING_FAILED or FORCE_PING. In fact we need
3027          * to clear PING_FAILED, because the unlink event handler will
3028          * have set it if we called LNetMDUnlink() above.
3029          */
3030         spin_lock(&lp->lp_lock);
3031         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3032         return rc;
3033 }
3034
3035 /*
3036  * This function exists because you cannot call LNetMDUnlink() from an
3037  * event handler.
3038  */
3039 static int lnet_peer_push_failed(struct lnet_peer *lp)
3040 __must_hold(&lp->lp_lock)
3041 {
3042         struct lnet_handle_md mdh;
3043         int rc;
3044
3045         mdh = lp->lp_push_mdh;
3046         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3047         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3048         rc = lp->lp_push_error;
3049         lp->lp_push_error = 0;
3050         spin_unlock(&lp->lp_lock);
3051
3052         if (!LNetMDHandleIsInvalid(mdh))
3053                 LNetMDUnlink(mdh);
3054
3055         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3056         spin_lock(&lp->lp_lock);
3057         return rc ? rc : LNET_REDISCOVER_PEER;
3058 }
3059
3060 /* Active side of push. */
3061 static int lnet_peer_send_push(struct lnet_peer *lp)
3062 __must_hold(&lp->lp_lock)
3063 {
3064         struct lnet_ping_buffer *pbuf;
3065         struct lnet_process_id id;
3066         struct lnet_md md;
3067         int cpt;
3068         int rc;
3069
3070         /* Don't push to a non-multi-rail peer. */
3071         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3072                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3073                 return 0;
3074         }
3075
3076         lp->lp_state |= LNET_PEER_PUSH_SENT;
3077         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3078         spin_unlock(&lp->lp_lock);
3079
3080         cpt = lnet_net_lock_current();
3081         pbuf = the_lnet.ln_ping_target;
3082         lnet_ping_buffer_addref(pbuf);
3083         lnet_net_unlock(cpt);
3084
3085         /* Push source MD */
3086         md.start     = &pbuf->pb_info;
3087         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3088         md.threshold = 2; /* Put/Ack */
3089         md.max_size  = 0;
3090         md.options   = 0;
3091         md.handler   = the_lnet.ln_dc_handler;
3092         md.user_ptr  = lp;
3093
3094         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3095         if (rc) {
3096                 lnet_ping_buffer_decref(pbuf);
3097                 CERROR("Can't bind push source MD: %d\n", rc);
3098                 goto fail_error;
3099         }
3100         cpt = lnet_net_lock_current();
3101         /* Refcount for MD. */
3102         lnet_peer_addref_locked(lp);
3103         id.pid = LNET_PID_LUSTRE;
3104         id.nid = lnet_peer_select_nid(lp);
3105         lnet_net_unlock(cpt);
3106
3107         if (id.nid == LNET_NID_ANY) {
3108                 rc = -EHOSTUNREACH;
3109                 goto fail_unlink;
3110         }
3111
3112         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3113                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3114                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3115
3116         if (rc)
3117                 goto fail_unlink;
3118
3119         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3120
3121         spin_lock(&lp->lp_lock);
3122         return 0;
3123
3124 fail_unlink:
3125         LNetMDUnlink(lp->lp_push_mdh);
3126         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3127 fail_error:
3128         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3129         /*
3130          * The errors that get us here are considered hard errors and
3131          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3132          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3133          * because the unlink event handler will have set it if we
3134          * called LNetMDUnlink() above.
3135          */
3136         spin_lock(&lp->lp_lock);
3137         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3138         return rc;
3139 }
3140
3141 /*
3142  * An unrecoverable error was encountered during discovery.
3143  * Set error status in peer and abort discovery.
3144  */
3145 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3146 {
3147         CDEBUG(D_NET, "Discovery error %s: %d\n",
3148                libcfs_nid2str(lp->lp_primary_nid), error);
3149
3150         spin_lock(&lp->lp_lock);
3151         lp->lp_dc_error = error;
3152         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3153         lp->lp_state |= LNET_PEER_REDISCOVER;
3154         spin_unlock(&lp->lp_lock);
3155 }
3156
3157 /*
3158  * Mark the peer as discovered.
3159  */
3160 static int lnet_peer_discovered(struct lnet_peer *lp)
3161 __must_hold(&lp->lp_lock)
3162 {
3163         lp->lp_state |= LNET_PEER_DISCOVERED;
3164         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3165                           LNET_PEER_REDISCOVER);
3166
3167         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3168
3169         return 0;
3170 }
3171
3172
3173 /*
3174  * Discovering this peer is taking too long. Cancel any Ping or Push
3175  * that discovery is waiting on by unlinking the relevant MDs. The
3176  * lnet_discovery_event_handler() will proceed from here and complete
3177  * the cleanup.
3178  */
3179 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3180 {
3181         struct lnet_handle_md ping_mdh;
3182         struct lnet_handle_md push_mdh;
3183
3184         LNetInvalidateMDHandle(&ping_mdh);
3185         LNetInvalidateMDHandle(&push_mdh);
3186
3187         spin_lock(&lp->lp_lock);
3188         if (lp->lp_state & LNET_PEER_PING_SENT) {
3189                 ping_mdh = lp->lp_ping_mdh;
3190                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3191         }
3192         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3193                 push_mdh = lp->lp_push_mdh;
3194                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3195         }
3196         spin_unlock(&lp->lp_lock);
3197
3198         if (!LNetMDHandleIsInvalid(ping_mdh))
3199                 LNetMDUnlink(ping_mdh);
3200         if (!LNetMDHandleIsInvalid(push_mdh))
3201                 LNetMDUnlink(push_mdh);
3202 }
3203
3204 /*
3205  * Wait for work to be queued or some other change that must be
3206  * attended to. Returns non-zero if the discovery thread should shut
3207  * down.
3208  */
3209 static int lnet_peer_discovery_wait_for_work(void)
3210 {
3211         int cpt;
3212         int rc = 0;
3213
3214         DEFINE_WAIT(wait);
3215
3216         cpt = lnet_net_lock_current();
3217         for (;;) {
3218                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3219                                 TASK_INTERRUPTIBLE);
3220                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3221                         break;
3222                 if (lnet_push_target_resize_needed() ||
3223                     the_lnet.ln_push_target->pb_needs_post)
3224                         break;
3225                 if (!list_empty(&the_lnet.ln_dc_request))
3226                         break;
3227                 if (!list_empty(&the_lnet.ln_msg_resend))
3228                         break;
3229                 lnet_net_unlock(cpt);
3230
3231                 /*
3232                  * wakeup max every second to check if there are peers that
3233                  * have been stuck on the working queue for greater than
3234                  * the peer timeout.
3235                  */
3236                 schedule_timeout(cfs_time_seconds(1));
3237                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3238                 cpt = lnet_net_lock_current();
3239         }
3240         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3241
3242         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3243                 rc = -ESHUTDOWN;
3244
3245         lnet_net_unlock(cpt);
3246
3247         CDEBUG(D_NET, "woken: %d\n", rc);
3248
3249         return rc;
3250 }
3251
3252 /*
3253  * Messages that were pending on a destroyed peer will be put on a global
3254  * resend list. The message resend list will be checked by
3255  * the discovery thread when it wakes up, and will resend messages. These
3256  * messages can still be sendable in the case the lpni which was the initial
3257  * cause of the message re-queue was transfered to another peer.
3258  *
3259  * It is possible that LNet could be shutdown while we're iterating
3260  * through the list. lnet_shudown_lndnets() will attempt to access the
3261  * resend list, but will have to wait until the spinlock is released, by
3262  * which time there shouldn't be any more messages on the resend list.
3263  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3264  * for the messages so they can be released. The other case is that
3265  * lnet_shudown_lndnets() can finalize all the messages before this
3266  * function can visit the resend list, in which case this function will be
3267  * a no-op.
3268  */
3269 static void lnet_resend_msgs(void)
3270 {
3271         struct lnet_msg *msg, *tmp;
3272         LIST_HEAD(resend);
3273         int rc;
3274
3275         spin_lock(&the_lnet.ln_msg_resend_lock);
3276         list_splice(&the_lnet.ln_msg_resend, &resend);
3277         spin_unlock(&the_lnet.ln_msg_resend_lock);
3278
3279         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3280                 list_del_init(&msg->msg_list);
3281                 rc = lnet_send(msg->msg_src_nid_param, msg,
3282                                msg->msg_rtr_nid_param);
3283                 if (rc < 0) {
3284                         CNETERR("Error sending %s to %s: %d\n",
3285                                lnet_msgtyp2str(msg->msg_type),
3286                                libcfs_id2str(msg->msg_target), rc);
3287                         lnet_finalize(msg, rc);
3288                 }
3289         }
3290 }
3291
3292 /* The discovery thread. */
3293 static int lnet_peer_discovery(void *arg)
3294 {
3295         struct lnet_peer *lp;
3296         int rc;
3297
3298         wait_for_completion(&the_lnet.ln_started);
3299
3300         CDEBUG(D_NET, "started\n");
3301
3302         for (;;) {
3303                 if (lnet_peer_discovery_wait_for_work())
3304                         break;
3305
3306                 if (lnet_push_target_resize_needed())
3307                         lnet_push_target_resize();
3308                 else if (the_lnet.ln_push_target->pb_needs_post)
3309                         lnet_push_target_post(the_lnet.ln_push_target,
3310                                               &the_lnet.ln_push_target_md);
3311
3312                 lnet_resend_msgs();
3313
3314                 lnet_net_lock(LNET_LOCK_EX);
3315                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3316                         lnet_net_unlock(LNET_LOCK_EX);
3317                         break;
3318                 }
3319
3320                 /*
3321                  * Process all incoming discovery work requests.  When
3322                  * discovery must wait on a peer to change state, it
3323                  * is added to the tail of the ln_dc_working queue. A
3324                  * timestamp keeps track of when the peer was added,
3325                  * so we can time out discovery requests that take too
3326                  * long.
3327                  */
3328                 while (!list_empty(&the_lnet.ln_dc_request)) {
3329                         lp = list_first_entry(&the_lnet.ln_dc_request,
3330                                               struct lnet_peer, lp_dc_list);
3331                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3332                         /*
3333                          * set the time the peer was put on the dc_working
3334                          * queue. It shouldn't remain on the queue
3335                          * forever, in case the GET message (for ping)
3336                          * doesn't get a REPLY or the PUT message (for
3337                          * push) doesn't get an ACK.
3338                          */
3339                         lp->lp_last_queued = ktime_get_real_seconds();
3340                         lnet_net_unlock(LNET_LOCK_EX);
3341
3342                         if (lnet_push_target_resize_needed())
3343                                 lnet_push_target_resize();
3344                         else if (the_lnet.ln_push_target->pb_needs_post)
3345                                 lnet_push_target_post(the_lnet.ln_push_target,
3346                                                       &the_lnet.ln_push_target_md);
3347
3348                         /*
3349                          * Select an action depending on the state of
3350                          * the peer and whether discovery is disabled.
3351                          * The check whether discovery is disabled is
3352                          * done after the code that handles processing
3353                          * for arrived data, cleanup for failures, and
3354                          * forcing a Ping or Push.
3355                          */
3356                         spin_lock(&lp->lp_lock);
3357                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3358                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3359                                 lp->lp_state);
3360                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3361                                 rc = lnet_peer_data_present(lp);
3362                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3363                                 rc = lnet_peer_ping_failed(lp);
3364                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3365                                 rc = lnet_peer_push_failed(lp);
3366                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3367                                 rc = lnet_peer_send_ping(lp);
3368                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3369                                 rc = lnet_peer_send_push(lp);
3370                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3371                                 rc = lnet_peer_send_ping(lp);
3372                         else if (lnet_peer_needs_push(lp))
3373                                 rc = lnet_peer_send_push(lp);
3374                         else
3375                                 rc = lnet_peer_discovered(lp);
3376                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3377                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3378                                 lp->lp_state, rc);
3379                         spin_unlock(&lp->lp_lock);
3380
3381                         lnet_net_lock(LNET_LOCK_EX);
3382                         if (rc == LNET_REDISCOVER_PEER) {
3383                                 list_move(&lp->lp_dc_list,
3384                                           &the_lnet.ln_dc_request);
3385                         } else if (rc) {
3386                                 lnet_peer_discovery_error(lp, rc);
3387                         }
3388                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3389                                 lnet_peer_discovery_complete(lp);
3390                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3391                                 break;
3392
3393                         if (lp->lp_state & LNET_PEER_MARK_DELETION) {
3394                                 struct list_head rlist;
3395                                 struct lnet_route *route, *tmp;
3396                                 int sensitivity = lp->lp_health_sensitivity;
3397
3398                                 INIT_LIST_HEAD(&rlist);
3399
3400                                 /*
3401                                  * remove the peer from the discovery work
3402                                  * queue if it's on there in preparation
3403                                  * of deleting it.
3404                                  */
3405                                 if (!list_empty(&lp->lp_dc_list))
3406                                         list_del(&lp->lp_dc_list);
3407
3408                                 lnet_net_unlock(LNET_LOCK_EX);
3409
3410                                 mutex_lock(&the_lnet.ln_api_mutex);
3411
3412                                 lnet_net_lock(LNET_LOCK_EX);
3413                                 list_for_each_entry_safe(route, tmp,
3414                                                          &lp->lp_routes,
3415                                                          lr_gwlist)
3416                                         lnet_move_route(route, NULL, &rlist);
3417                                 lnet_net_unlock(LNET_LOCK_EX);
3418
3419                                 /* delete the peer */
3420                                 lnet_peer_del(lp);
3421
3422                                 list_for_each_entry_safe(route, tmp,
3423                                                          &rlist, lr_list) {
3424                                         /* re-add these routes */
3425                                         lnet_add_route(route->lr_net,
3426                                                        route->lr_hops,
3427                                                        route->lr_nid,
3428                                                        route->lr_priority,
3429                                                        sensitivity);
3430                                         LIBCFS_FREE(route, sizeof(*route));
3431                                 }
3432                                 mutex_unlock(&the_lnet.ln_api_mutex);
3433
3434                                 lnet_net_lock(LNET_LOCK_EX);
3435                         }
3436                 }
3437
3438                 lnet_net_unlock(LNET_LOCK_EX);
3439         }
3440
3441         CDEBUG(D_NET, "stopping\n");
3442         /*
3443          * Clean up before telling lnet_peer_discovery_stop() that
3444          * we're done. Use wake_up() below to somewhat reduce the
3445          * size of the thundering herd if there are multiple threads
3446          * waiting on discovery of a single peer.
3447          */
3448
3449         /* Queue cleanup 1: stop all pending pings and pushes. */
3450         lnet_net_lock(LNET_LOCK_EX);
3451         while (!list_empty(&the_lnet.ln_dc_working)) {
3452                 lp = list_first_entry(&the_lnet.ln_dc_working,
3453                                       struct lnet_peer, lp_dc_list);
3454                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3455                 lnet_net_unlock(LNET_LOCK_EX);
3456                 lnet_peer_cancel_discovery(lp);
3457                 lnet_net_lock(LNET_LOCK_EX);
3458         }
3459         lnet_net_unlock(LNET_LOCK_EX);
3460
3461         /* Queue cleanup 2: wait for the expired queue to clear. */
3462         while (!list_empty(&the_lnet.ln_dc_expired))
3463                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3464
3465         /* Queue cleanup 3: clear the request queue. */
3466         lnet_net_lock(LNET_LOCK_EX);
3467         while (!list_empty(&the_lnet.ln_dc_request)) {
3468                 lp = list_first_entry(&the_lnet.ln_dc_request,
3469                                       struct lnet_peer, lp_dc_list);
3470                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3471                 lnet_peer_discovery_complete(lp);
3472         }
3473         lnet_net_unlock(LNET_LOCK_EX);
3474
3475         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3476         the_lnet.ln_dc_handler = NULL;
3477
3478         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3479         wake_up(&the_lnet.ln_dc_waitq);
3480
3481         CDEBUG(D_NET, "stopped\n");
3482
3483         return 0;
3484 }
3485
3486 /* ln_api_mutex is held on entry. */
3487 int lnet_peer_discovery_start(void)
3488 {
3489         struct task_struct *task;
3490         int rc = 0;
3491
3492         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3493                 return -EALREADY;
3494
3495         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3496         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3497         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3498         if (IS_ERR(task)) {
3499                 rc = PTR_ERR(task);
3500                 CERROR("Can't start peer discovery thread: %d\n", rc);
3501
3502                 the_lnet.ln_dc_handler = NULL;
3503
3504                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3505         }
3506
3507         CDEBUG(D_NET, "discovery start: %d\n", rc);
3508
3509         return rc;
3510 }
3511
3512 /* ln_api_mutex is held on entry. */
3513 void lnet_peer_discovery_stop(void)
3514 {
3515         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3516                 return;
3517
3518         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3519         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3520
3521         /* In the LNetNIInit() path we may be stopping discovery before it
3522          * entered its work loop
3523          */
3524         if (!completion_done(&the_lnet.ln_started))
3525                 complete(&the_lnet.ln_started);
3526         else
3527                 wake_up(&the_lnet.ln_dc_waitq);
3528
3529         wait_event(the_lnet.ln_dc_waitq,
3530                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3531
3532         LASSERT(list_empty(&the_lnet.ln_dc_request));
3533         LASSERT(list_empty(&the_lnet.ln_dc_working));
3534         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3535
3536         CDEBUG(D_NET, "discovery stopped\n");
3537 }
3538
3539 /* Debugging */
3540
3541 void
3542 lnet_debug_peer(lnet_nid_t nid)
3543 {
3544         char                    *aliveness = "NA";
3545         struct lnet_peer_ni     *lp;
3546         int                     cpt;
3547
3548         cpt = lnet_cpt_of_nid(nid, NULL);
3549         lnet_net_lock(cpt);
3550
3551         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3552         if (IS_ERR(lp)) {
3553                 lnet_net_unlock(cpt);
3554                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3555                 return;
3556         }
3557
3558         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3559                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3560
3561         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3562                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3563                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3564                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3565                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3566
3567         lnet_peer_ni_decref_locked(lp);
3568
3569         lnet_net_unlock(cpt);
3570 }
3571
3572 /* Gathering information for userspace. */
3573
3574 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3575                           char aliveness[LNET_MAX_STR_LEN],
3576                           __u32 *cpt_iter, __u32 *refcount,
3577                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3578                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3579                           __u32 *peer_tx_qnob)
3580 {
3581         struct lnet_peer_table          *peer_table;
3582         struct lnet_peer_ni             *lp;
3583         int                             j;
3584         int                             lncpt;
3585         bool                            found = false;
3586
3587         /* get the number of CPTs */
3588         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3589
3590         /* if the cpt number to be examined is >= the number of cpts in
3591          * the system then indicate that there are no more cpts to examin
3592          */
3593         if (*cpt_iter >= lncpt)
3594                 return -ENOENT;
3595
3596         /* get the current table */
3597         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3598         /* if the ptable is NULL then there are no more cpts to examine */
3599         if (peer_table == NULL)
3600                 return -ENOENT;
3601
3602         lnet_net_lock(*cpt_iter);
3603
3604         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3605                 struct list_head *peers = &peer_table->pt_hash[j];
3606
3607                 list_for_each_entry(lp, peers, lpni_hashlist) {
3608                         if (peer_index-- > 0)
3609                                 continue;
3610
3611                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3612                         if (lnet_isrouter(lp) ||
3613                                 lnet_peer_aliveness_enabled(lp))
3614                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3615                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3616
3617                         *nid = lp->lpni_nid;
3618                         *refcount = atomic_read(&lp->lpni_refcount);
3619                         *ni_peer_tx_credits =
3620                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3621                         *peer_tx_credits = lp->lpni_txcredits;
3622                         *peer_rtr_credits = lp->lpni_rtrcredits;
3623                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3624                         *peer_tx_qnob = lp->lpni_txqnob;
3625
3626                         found = true;
3627                 }
3628
3629         }
3630         lnet_net_unlock(*cpt_iter);
3631
3632         *cpt_iter = lncpt;
3633
3634         return found ? 0 : -ENOENT;
3635 }
3636
3637 /* ln_api_mutex is held, which keeps the peer list stable */
3638 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3639 {
3640         struct lnet_ioctl_element_stats *lpni_stats;
3641         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3642         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3643         struct lnet_peer_ni_credit_info *lpni_info;
3644         struct lnet_peer_ni *lpni;
3645         struct lnet_peer *lp;
3646         lnet_nid_t nid;
3647         __u32 size;
3648         int rc;
3649
3650         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3651
3652         if (!lp) {
3653                 rc = -ENOENT;
3654                 goto out;
3655         }
3656
3657         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3658                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3659         size *= lp->lp_nnis;
3660         if (size > cfg->prcfg_size) {
3661                 cfg->prcfg_size = size;
3662                 rc = -E2BIG;
3663                 goto out_lp_decref;
3664         }
3665
3666         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3667         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3668         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3669         cfg->prcfg_count = lp->lp_nnis;
3670         cfg->prcfg_size = size;
3671         cfg->prcfg_state = lp->lp_state;
3672
3673         /* Allocate helper buffers. */
3674         rc = -ENOMEM;
3675         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3676         if (!lpni_info)
3677                 goto out_lp_decref;
3678         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3679         if (!lpni_stats)
3680                 goto out_free_info;
3681         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3682         if (!lpni_msg_stats)
3683                 goto out_free_stats;
3684         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3685         if (!lpni_hstats)
3686                 goto out_free_msg_stats;
3687
3688
3689         lpni = NULL;
3690         rc = -EFAULT;
3691         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3692                 nid = lpni->lpni_nid;
3693                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3694                         goto out_free_hstats;
3695                 bulk += sizeof(nid);
3696
3697                 memset(lpni_info, 0, sizeof(*lpni_info));
3698                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3699                 if (lnet_isrouter(lpni) ||
3700                         lnet_peer_aliveness_enabled(lpni))
3701                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3702                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3703
3704                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3705                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3706                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3707                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3708                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3709                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3710                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3711                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3712                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3713                         goto out_free_hstats;
3714                 bulk += sizeof(*lpni_info);
3715
3716                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3717                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3718                                                             LNET_STATS_TYPE_SEND);
3719                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3720                                                             LNET_STATS_TYPE_RECV);
3721                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3722                                                             LNET_STATS_TYPE_DROP);
3723                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3724                         goto out_free_hstats;
3725                 bulk += sizeof(*lpni_stats);
3726                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3727                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3728                         goto out_free_hstats;
3729                 bulk += sizeof(*lpni_msg_stats);
3730                 lpni_hstats->hlpni_network_timeout =
3731                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3732                 lpni_hstats->hlpni_remote_dropped =
3733                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3734                 lpni_hstats->hlpni_remote_timeout =
3735                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3736                 lpni_hstats->hlpni_remote_error =
3737                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3738                 lpni_hstats->hlpni_health_value =
3739                   atomic_read(&lpni->lpni_healthv);
3740                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3741                         goto out_free_hstats;
3742                 bulk += sizeof(*lpni_hstats);
3743         }
3744         rc = 0;
3745
3746 out_free_hstats:
3747         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3748 out_free_msg_stats:
3749         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3750 out_free_stats:
3751         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3752 out_free_info:
3753         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3754 out_lp_decref:
3755         lnet_peer_decref_locked(lp);
3756 out:
3757         return rc;
3758 }
3759
3760 void
3761 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3762 {
3763         /* the mt could've shutdown and cleaned up the queues */
3764         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3765                 return;
3766
3767         if (list_empty(&lpni->lpni_recovery) &&
3768             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3769                 CDEBUG(D_NET, "lpni %s added to recovery queue. Health = %d\n",
3770                         libcfs_nid2str(lpni->lpni_nid),
3771                         atomic_read(&lpni->lpni_healthv));
3772                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3773                 lnet_peer_ni_addref_locked(lpni);
3774         }
3775 }
3776
3777 /* Call with the ln_api_mutex held */
3778 void
3779 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3780 {
3781         struct lnet_peer_table *ptable;
3782         struct lnet_peer *lp;
3783         struct lnet_peer_net *lpn;
3784         struct lnet_peer_ni *lpni;
3785         int lncpt;
3786         int cpt;
3787
3788         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3789                 return;
3790
3791         if (!all) {
3792                 lnet_net_lock(LNET_LOCK_EX);
3793                 lpni = lnet_find_peer_ni_locked(nid);
3794                 if (!lpni) {
3795                         lnet_net_unlock(LNET_LOCK_EX);
3796                         return;
3797                 }
3798                 atomic_set(&lpni->lpni_healthv, value);
3799                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3800                 lnet_peer_ni_decref_locked(lpni);
3801                 lnet_net_unlock(LNET_LOCK_EX);
3802                 return;
3803         }
3804
3805         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3806
3807         /*
3808          * Walk all the peers and reset the healhv for each one to the
3809          * maximum value.
3810          */
3811         lnet_net_lock(LNET_LOCK_EX);
3812         for (cpt = 0; cpt < lncpt; cpt++) {
3813                 ptable = the_lnet.ln_peer_tables[cpt];
3814                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3815                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3816                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3817                                                     lpni_peer_nis) {
3818                                         atomic_set(&lpni->lpni_healthv, value);
3819                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3820                                 }
3821                         }
3822                 }
3823         }
3824         lnet_net_unlock(LNET_LOCK_EX);
3825 }
3826