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         lp->lp_disc_src_nid = LNET_NID_ANY;
 262         if (lnet_peers_start_down())
 263                 lp->lp_alive = false;
 264         else
 265                 lp->lp_alive = true;
 266
 267         /*
 268          * all peers created on a router should have health on
 269          * if it's not already on.
 270          */
 271         if (the_lnet.ln_routing && !lnet_health_sensitivity)
 272                 lp->lp_health_sensitivity = 1;
 273
 274         /*
 275          * Turn off discovery for loopback peer. If you're creating a peer
 276          * for the loopback interface then that was initiated when we
 277          * attempted to send a message over the loopback. There is no need
 278          * to ever use a different interface when sending messages to
 279          * myself.
 280          */
 281         if (nid == LNET_NID_LO_0)
 282                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 283         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 284
 285         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 286
 287         return lp;
 288 }
 289
 290 void
 291 lnet_destroy_peer_locked(struct lnet_peer *lp)
 292 {
 293         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 294
 295         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 296         LASSERT(lp->lp_rtr_refcount == 0);
 297         LASSERT(list_empty(&lp->lp_peer_nets));
 298         LASSERT(list_empty(&lp->lp_peer_list));
 299         LASSERT(list_empty(&lp->lp_dc_list));
 300
 301         if (lp->lp_data)
 302                 lnet_ping_buffer_decref(lp->lp_data);
 303
 304         /*
 305          * if there are messages still on the pending queue, then make
 306          * sure to queue them on the ln_msg_resend list so they can be
 307          * resent at a later point if the discovery thread is still
 308          * running.
 309          * If the discovery thread has stopped, then the wakeup will be a
 310          * no-op, and it is expected the lnet_shutdown_lndnets() will
 311          * eventually be called, which will traverse this list and
 312          * finalize the messages on the list.
 313          * We can not resend them now because we're holding the cpt lock.
 314          * Releasing the lock can cause an inconsistent state
 315          */
 316         spin_lock(&the_lnet.ln_msg_resend_lock);
 317         spin_lock(&lp->lp_lock);
 318         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 319         spin_unlock(&lp->lp_lock);
 320         spin_unlock(&the_lnet.ln_msg_resend_lock);
 321         wake_up(&the_lnet.ln_dc_waitq);
 322
 323         LIBCFS_FREE(lp, sizeof(*lp));
 324 }
 325
 326 /*
 327  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 328  * that peer_net, detach the peer_net from the peer.
 329  *
 330  * Call with lnet_net_lock/EX held
 331  */
 332 static void
 333 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 334 {
 335         struct lnet_peer_table *ptable;
 336         struct lnet_peer_net *lpn;
 337         struct lnet_peer *lp;
 338
 339         /*
 340          * Belts and suspenders: gracefully handle teardown of a
 341          * partially connected peer_ni.
 342          */
 343         lpn = lpni->lpni_peer_net;
 344
 345         list_del_init(&lpni->lpni_peer_nis);
 346         /*
 347          * If there are no lpni's left, we detach lpn from
 348          * lp_peer_nets, so it cannot be found anymore.
 349          */
 350         if (list_empty(&lpn->lpn_peer_nis))
 351                 list_del_init(&lpn->lpn_peer_nets);
 352
 353         /* Update peer NID count. */
 354         lp = lpn->lpn_peer;
 355         lp->lp_nnis--;
 356
 357         /*
 358          * If there are no more peer nets, make the peer unfindable
 359          * via the peer_tables.
 360          *
 361          * Otherwise, if the peer is DISCOVERED, tell discovery to
 362          * take another look at it. This is a no-op if discovery for
 363          * this peer did the detaching.
 364          */
 365         if (list_empty(&lp->lp_peer_nets)) {
 366                 list_del_init(&lp->lp_peer_list);
 367                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 368                 ptable->pt_peers--;
 369         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 370                 /* Discovery isn't running, nothing to do here. */
 371         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 372                 lnet_peer_queue_for_discovery(lp);
 373                 wake_up(&the_lnet.ln_dc_waitq);
 374         }
 375         CDEBUG(D_NET, "peer %s NID %s\n",
 376                 libcfs_nid2str(lp->lp_primary_nid),
 377                 libcfs_nid2str(lpni->lpni_nid));
 378 }
 379
 380 /* called with lnet_net_lock LNET_LOCK_EX held */
 381 static int
 382 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 383 {
 384         struct lnet_peer_table *ptable = NULL;
 385
 386         /* don't remove a peer_ni if it's also a gateway */
 387         if (lnet_isrouter(lpni) && !force) {
 388                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 389                        libcfs_nid2str(lpni->lpni_nid));
 390                 return -EBUSY;
 391         }
 392
 393         lnet_peer_remove_from_remote_list(lpni);
 394
 395         /* remove peer ni from the hash list. */
 396         list_del_init(&lpni->lpni_hashlist);
 397
 398         /*
 399          * indicate the peer is being deleted so the monitor thread can
 400          * remove it from the recovery queue.
 401          */
 402         spin_lock(&lpni->lpni_lock);
 403         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 404         spin_unlock(&lpni->lpni_lock);
 405
 406         /* decrement the ref count on the peer table */
 407         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 408
 409         /*
 410          * The peer_ni can no longer be found with a lookup. But there
 411          * can be current users, so keep track of it on the zombie
 412          * list until the reference count has gone to zero.
 413          *
 414          * The last reference may be lost in a place where the
 415          * lnet_net_lock locks only a single cpt, and that cpt may not
 416          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 417          * has its own lock.
 418          */
 419         spin_lock(&ptable->pt_zombie_lock);
 420         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 421         ptable->pt_zombies++;
 422         spin_unlock(&ptable->pt_zombie_lock);
 423
 424         /* no need to keep this peer_ni on the hierarchy anymore */
 425         lnet_peer_detach_peer_ni_locked(lpni);
 426
 427         /* remove hashlist reference on peer_ni */
 428         lnet_peer_ni_decref_locked(lpni);
 429
 430         return 0;
 431 }
 432
 433 void lnet_peer_uninit(void)
 434 {
 435         struct lnet_peer_ni *lpni, *tmp;
 436
 437         lnet_net_lock(LNET_LOCK_EX);
 438
 439         /* remove all peer_nis from the remote peer and the hash list */
 440         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 441                                  lpni_on_remote_peer_ni_list)
 442                 lnet_peer_ni_del_locked(lpni, false);
 443
 444         lnet_peer_tables_destroy();
 445
 446         lnet_net_unlock(LNET_LOCK_EX);
 447 }
 448
 449 static int
 450 lnet_peer_del_locked(struct lnet_peer *peer)
 451 {
 452         struct lnet_peer_ni *lpni = NULL, *lpni2;
 453         int rc = 0, rc2 = 0;
 454
 455         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
 456
 457         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 458         while (lpni != NULL) {
 459                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 460                 rc = lnet_peer_ni_del_locked(lpni, false);
 461                 if (rc != 0)
 462                         rc2 = rc;
 463                 lpni = lpni2;
 464         }
 465
 466         return rc2;
 467 }
 468
 469 static int
 470 lnet_peer_del(struct lnet_peer *peer)
 471 {
 472         lnet_net_lock(LNET_LOCK_EX);
 473         lnet_peer_del_locked(peer);
 474         lnet_net_unlock(LNET_LOCK_EX);
 475
 476         return 0;
 477 }
 478
 479 /*
 480  * Delete a NID from a peer. Call with ln_api_mutex held.
 481  *
 482  * Error codes:
 483  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 484  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 485  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 486  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 487  */
 488 static int
 489 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
 490 {
 491         struct lnet_peer_ni *lpni;
 492         lnet_nid_t primary_nid = lp->lp_primary_nid;
 493         int rc = 0;
 494         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 495
 496         if (!(flags & LNET_PEER_CONFIGURED)) {
 497                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 498                         rc = -EPERM;
 499                         goto out;
 500                 }
 501         }
 502         lpni = lnet_find_peer_ni_locked(nid);
 503         if (!lpni) {
 504                 rc = -ENOENT;
 505                 goto out;
 506         }
 507         lnet_peer_ni_decref_locked(lpni);
 508         if (lp != lpni->lpni_peer_net->lpn_peer) {
 509                 rc = -ECHILD;
 510                 goto out;
 511         }
 512
 513         /*
 514          * This function only allows deletion of the primary NID if it
 515          * is the only NID.
 516          */
 517         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
 518                 rc = -EBUSY;
 519                 goto out;
 520         }
 521
 522         lnet_net_lock(LNET_LOCK_EX);
 523
 524         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
 525                 struct lnet_peer_ni *lpni2;
 526                 /* assign the next peer_ni to be the primary */
 527                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 528                 LASSERT(lpni2);
 529                 lp->lp_primary_nid = lpni2->lpni_nid;
 530         }
 531         rc = lnet_peer_ni_del_locked(lpni, force);
 532
 533         lnet_net_unlock(LNET_LOCK_EX);
 534
 535 out:
 536         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 537                libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
 538
 539         return rc;
 540 }
 541
 542 static void
 543 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 544                                struct lnet_peer_table *ptable)
 545 {
 546         int                      i;
 547         struct lnet_peer_ni     *next;
 548         struct lnet_peer_ni     *lpni;
 549         struct lnet_peer        *peer;
 550
 551         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 552                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 553                                          lpni_hashlist) {
 554                         if (net != NULL && net != lpni->lpni_net)
 555                                 continue;
 556
 557                         peer = lpni->lpni_peer_net->lpn_peer;
 558                         if (peer->lp_primary_nid != lpni->lpni_nid) {
 559                                 lnet_peer_ni_del_locked(lpni, false);
 560                                 continue;
 561                         }
 562                         /*
 563                          * Removing the primary NID implies removing
 564                          * the entire peer. Advance next beyond any
 565                          * peer_ni that belongs to the same peer.
 566                          */
 567                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 568                                                  lpni_hashlist) {
 569                                 if (next->lpni_peer_net->lpn_peer != peer)
 570                                         break;
 571                         }
 572                         lnet_peer_del_locked(peer);
 573                 }
 574         }
 575 }
 576
 577 static void
 578 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 579 {
 580         wait_var_event_warning(&ptable->pt_zombies,
 581                                ptable->pt_zombies == 0,
 582                                "Waiting for %d zombies on peer table\n",
 583                                ptable->pt_zombies);
 584 }
 585
 586 static void
 587 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 588                                 struct lnet_peer_table *ptable)
 589 {
 590         struct lnet_peer_ni     *lp;
 591         struct lnet_peer_ni     *tmp;
 592         lnet_nid_t              gw_nid;
 593         int                     i;
 594
 595         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 596                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 597                                          lpni_hashlist) {
 598                         if (net != lp->lpni_net)
 599                                 continue;
 600
 601                         if (!lnet_isrouter(lp))
 602                                 continue;
 603
 604                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 605
 606                         lnet_net_unlock(LNET_LOCK_EX);
 607                         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
 608                         lnet_net_lock(LNET_LOCK_EX);
 609                 }
 610         }
 611 }
 612
 613 void
 614 lnet_peer_tables_cleanup(struct lnet_net *net)
 615 {
 616         int i;
 617         struct lnet_peer_table *ptable;
 618
 619         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 620         /* If just deleting the peers for a NI, get rid of any routes these
 621          * peers are gateways for. */
 622         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 623                 lnet_net_lock(LNET_LOCK_EX);
 624                 lnet_peer_table_del_rtrs_locked(net, ptable);
 625                 lnet_net_unlock(LNET_LOCK_EX);
 626         }
 627
 628         /* Start the cleanup process */
 629         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 630                 lnet_net_lock(LNET_LOCK_EX);
 631                 lnet_peer_table_cleanup_locked(net, ptable);
 632                 lnet_net_unlock(LNET_LOCK_EX);
 633         }
 634
 635         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 636                 lnet_peer_ni_finalize_wait(ptable);
 637 }
 638
 639 static struct lnet_peer_ni *
 640 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 641 {
 642         struct list_head        *peers;
 643         struct lnet_peer_ni     *lp;
 644
 645         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 646                 return NULL;
 647
 648         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 649         list_for_each_entry(lp, peers, lpni_hashlist) {
 650                 if (lp->lpni_nid == nid) {
 651                         lnet_peer_ni_addref_locked(lp);
 652                         return lp;
 653                 }
 654         }
 655
 656         return NULL;
 657 }
 658
 659 struct lnet_peer_ni *
 660 lnet_find_peer_ni_locked(lnet_nid_t nid)
 661 {
 662         struct lnet_peer_ni *lpni;
 663         struct lnet_peer_table *ptable;
 664         int cpt;
 665
 666         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 667
 668         ptable = the_lnet.ln_peer_tables[cpt];
 669         lpni = lnet_get_peer_ni_locked(ptable, nid);
 670
 671         return lpni;
 672 }
 673
 674 struct lnet_peer_ni *
 675 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 676 {
 677         struct lnet_peer_net *lpn;
 678         struct lnet_peer_ni *lpni;
 679
 680         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 681         if (!lpn)
 682                 return NULL;
 683
 684         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 685                 if (lpni->lpni_nid == nid)
 686                         return lpni;
 687         }
 688
 689         return NULL;
 690 }
 691
 692 struct lnet_peer *
 693 lnet_find_peer(lnet_nid_t nid)
 694 {
 695         struct lnet_peer_ni *lpni;
 696         struct lnet_peer *lp = NULL;
 697         int cpt;
 698
 699         cpt = lnet_net_lock_current();
 700         lpni = lnet_find_peer_ni_locked(nid);
 701         if (lpni) {
 702                 lp = lpni->lpni_peer_net->lpn_peer;
 703                 lnet_peer_addref_locked(lp);
 704                 lnet_peer_ni_decref_locked(lpni);
 705         }
 706         lnet_net_unlock(cpt);
 707
 708         return lp;
 709 }
 710
 711 struct lnet_peer_net *
 712 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 713 {
 714         struct lnet_peer_net *net;
 715
 716         if (!prev_lpn_id) {
 717                 /* no net id provided return the first net */
 718                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 719                                                struct lnet_peer_net,
 720                                                lpn_peer_nets);
 721
 722                 return net;
 723         }
 724
 725         /* find the net after the one provided */
 726         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 727                 if (net->lpn_net_id == prev_lpn_id) {
 728                         /*
 729                          * if we reached the end of the list loop to the
 730                          * beginning.
 731                          */
 732                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 733                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 734                                                                 struct lnet_peer_net,
 735                                                                 lpn_peer_nets);
 736                         else
 737                                 return list_next_entry(net, lpn_peer_nets);
 738                 }
 739         }
 740
 741         return NULL;
 742 }
 743
 744 struct lnet_peer_ni *
 745 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 746                              struct lnet_peer_net *peer_net,
 747                              struct lnet_peer_ni *prev)
 748 {
 749         struct lnet_peer_ni *lpni;
 750         struct lnet_peer_net *net = peer_net;
 751
 752         if (!prev) {
 753                 if (!net) {
 754                         if (list_empty(&peer->lp_peer_nets))
 755                                 return NULL;
 756
 757                         net = list_entry(peer->lp_peer_nets.next,
 758                                          struct lnet_peer_net,
 759                                          lpn_peer_nets);
 760                 }
 761                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 762                                   lpni_peer_nis);
 763
 764                 return lpni;
 765         }
 766
 767         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 768                 /*
 769                  * if you reached the end of the peer ni list and the peer
 770                  * net is specified then there are no more peer nis in that
 771                  * net.
 772                  */
 773                 if (net)
 774                         return NULL;
 775
 776                 /*
 777                  * we reached the end of this net ni list. move to the
 778                  * next net
 779                  */
 780                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 781                     &peer->lp_peer_nets)
 782                         /* no more nets and no more NIs. */
 783                         return NULL;
 784
 785                 /* get the next net */
 786                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 787                                  struct lnet_peer_net,
 788                                  lpn_peer_nets);
 789                 /* get the ni on it */
 790                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 791                                   lpni_peer_nis);
 792
 793                 return lpni;
 794         }
 795
 796         /* there are more nis left */
 797         lpni = list_entry(prev->lpni_peer_nis.next,
 798                           struct lnet_peer_ni, lpni_peer_nis);
 799
 800         return lpni;
 801 }
 802
 803 /* Call with the ln_api_mutex held */
 804 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 805 {
 806         struct lnet_process_id id;
 807         struct lnet_peer_table *ptable;
 808         struct lnet_peer *lp;
 809         __u32 count = 0;
 810         __u32 size = 0;
 811         int lncpt;
 812         int cpt;
 813         __u32 i;
 814         int rc;
 815
 816         rc = -ESHUTDOWN;
 817         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 818                 goto done;
 819
 820         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 821
 822         /*
 823          * Count the number of peers, and return E2BIG if the buffer
 824          * is too small. We'll also return the desired size.
 825          */
 826         rc = -E2BIG;
 827         for (cpt = 0; cpt < lncpt; cpt++) {
 828                 ptable = the_lnet.ln_peer_tables[cpt];
 829                 count += ptable->pt_peers;
 830         }
 831         size = count * sizeof(*ids);
 832         if (size > *sizep)
 833                 goto done;
 834
 835         /*
 836          * Walk the peer lists and copy out the primary nids.
 837          * This is safe because the peer lists are only modified
 838          * while the ln_api_mutex is held. So we don't need to
 839          * hold the lnet_net_lock as well, and can therefore
 840          * directly call copy_to_user().
 841          */
 842         rc = -EFAULT;
 843         memset(&id, 0, sizeof(id));
 844         id.pid = LNET_PID_LUSTRE;
 845         i = 0;
 846         for (cpt = 0; cpt < lncpt; cpt++) {
 847                 ptable = the_lnet.ln_peer_tables[cpt];
 848                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 849                         if (i >= count)
 850                                 goto done;
 851                         id.nid = lp->lp_primary_nid;
 852                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 853                                 goto done;
 854                         i++;
 855                 }
 856         }
 857         rc = 0;
 858 done:
 859         *countp = count;
 860         *sizep = size;
 861         return rc;
 862 }
 863
 864 /*
 865  * Start pushes to peers that need to be updated for a configuration
 866  * change on this node.
 867  */
 868 void
 869 lnet_push_update_to_peers(int force)
 870 {
 871         struct lnet_peer_table *ptable;
 872         struct lnet_peer *lp;
 873         int lncpt;
 874         int cpt;
 875
 876         lnet_net_lock(LNET_LOCK_EX);
 877         if (lnet_peer_discovery_disabled)
 878                 force = 0;
 879         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 880         for (cpt = 0; cpt < lncpt; cpt++) {
 881                 ptable = the_lnet.ln_peer_tables[cpt];
 882                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 883                         if (force) {
 884                                 spin_lock(&lp->lp_lock);
 885                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 886                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 887                                 spin_unlock(&lp->lp_lock);
 888                         }
 889                         if (lnet_peer_needs_push(lp))
 890                                 lnet_peer_queue_for_discovery(lp);
 891                 }
 892         }
 893         lnet_net_unlock(LNET_LOCK_EX);
 894         wake_up(&the_lnet.ln_dc_waitq);
 895 }
 896
 897 /*
 898  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 899  * this is a preferred point-to-point path. Call with lnet_net_lock in
 900  * shared mmode.
 901  */
 902 bool
 903 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 904 {
 905         int i;
 906
 907         if (lpni->lpni_pref_nnids == 0)
 908                 return false;
 909         if (lpni->lpni_pref_nnids == 1)
 910                 return lpni->lpni_pref.nid == nid;
 911         for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 912                 if (lpni->lpni_pref.nids[i] == nid)
 913                         return true;
 914         }
 915         return false;
 916 }
 917
 918 /*
 919  * Set a single ni as preferred, provided no preferred ni is already
 920  * defined. Only to be used for non-multi-rail peer_ni.
 921  */
 922 int
 923 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 924 {
 925         int rc = 0;
 926
 927         spin_lock(&lpni->lpni_lock);
 928         if (nid == LNET_NID_ANY) {
 929                 rc = -EINVAL;
 930         } else if (lpni->lpni_pref_nnids > 0) {
 931                 rc = -EPERM;
 932         } else if (lpni->lpni_pref_nnids == 0) {
 933                 lpni->lpni_pref.nid = nid;
 934                 lpni->lpni_pref_nnids = 1;
 935                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
 936         }
 937         spin_unlock(&lpni->lpni_lock);
 938
 939         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 940                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
 941         return rc;
 942 }
 943
 944 /*
 945  * Clear the preferred NID from a non-multi-rail peer_ni, provided
 946  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
 947  */
 948 int
 949 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
 950 {
 951         int rc = 0;
 952
 953         spin_lock(&lpni->lpni_lock);
 954         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
 955                 lpni->lpni_pref_nnids = 0;
 956                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 957         } else if (lpni->lpni_pref_nnids == 0) {
 958                 rc = -ENOENT;
 959         } else {
 960                 rc = -EPERM;
 961         }
 962         spin_unlock(&lpni->lpni_lock);
 963
 964         CDEBUG(D_NET, "peer %s: %d\n",
 965                libcfs_nid2str(lpni->lpni_nid), rc);
 966         return rc;
 967 }
 968
 969 /*
 970  * Clear the preferred NIDs from a non-multi-rail peer.
 971  */
 972 void
 973 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
 974 {
 975         struct lnet_peer_ni *lpni = NULL;
 976
 977         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 978                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
 979 }
 980
 981 int
 982 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 983 {
 984         lnet_nid_t *nids = NULL;
 985         lnet_nid_t *oldnids = NULL;
 986         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 987         int size;
 988         int i;
 989         int rc = 0;
 990
 991         if (nid == LNET_NID_ANY) {
 992                 rc = -EINVAL;
 993                 goto out;
 994         }
 995
 996         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
 997                 rc = -EEXIST;
 998                 goto out;
 999         }
1000
1001         /* A non-MR node may have only one preferred NI per peer_ni */
1002         if (lpni->lpni_pref_nnids > 0) {
1003                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1004                         rc = -EPERM;
1005                         goto out;
1006                 }
1007         }
1008
1009         if (lpni->lpni_pref_nnids != 0) {
1010                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1011                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1012                 if (!nids) {
1013                         rc = -ENOMEM;
1014                         goto out;
1015                 }
1016                 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1017                         if (lpni->lpni_pref.nids[i] == nid) {
1018                                 LIBCFS_FREE(nids, size);
1019                                 rc = -EEXIST;
1020                                 goto out;
1021                         }
1022                         nids[i] = lpni->lpni_pref.nids[i];
1023                 }
1024                 nids[i] = nid;
1025         }
1026
1027         lnet_net_lock(LNET_LOCK_EX);
1028         spin_lock(&lpni->lpni_lock);
1029         if (lpni->lpni_pref_nnids == 0) {
1030                 lpni->lpni_pref.nid = nid;
1031         } else {
1032                 oldnids = lpni->lpni_pref.nids;
1033                 lpni->lpni_pref.nids = nids;
1034         }
1035         lpni->lpni_pref_nnids++;
1036         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1037         spin_unlock(&lpni->lpni_lock);
1038         lnet_net_unlock(LNET_LOCK_EX);
1039
1040         if (oldnids) {
1041                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1042                 CFS_FREE_PTR_ARRAY(oldnids, size);
1043         }
1044 out:
1045         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1046                 spin_lock(&lpni->lpni_lock);
1047                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1048                 spin_unlock(&lpni->lpni_lock);
1049         }
1050         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1051                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1052         return rc;
1053 }
1054
1055 int
1056 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1057 {
1058         lnet_nid_t *nids = NULL;
1059         lnet_nid_t *oldnids = NULL;
1060         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1061         int size;
1062         int i, j;
1063         int rc = 0;
1064
1065         if (lpni->lpni_pref_nnids == 0) {
1066                 rc = -ENOENT;
1067                 goto out;
1068         }
1069
1070         if (lpni->lpni_pref_nnids == 1) {
1071                 if (lpni->lpni_pref.nid != nid) {
1072                         rc = -ENOENT;
1073                         goto out;
1074                 }
1075         } else if (lpni->lpni_pref_nnids == 2) {
1076                 if (lpni->lpni_pref.nids[0] != nid &&
1077                     lpni->lpni_pref.nids[1] != nid) {
1078                         rc = -ENOENT;
1079                         goto out;
1080                 }
1081         } else {
1082                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1083                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1084                 if (!nids) {
1085                         rc = -ENOMEM;
1086                         goto out;
1087                 }
1088                 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1089                         if (lpni->lpni_pref.nids[i] != nid)
1090                                 continue;
1091                         nids[j++] = lpni->lpni_pref.nids[i];
1092                 }
1093                 /* Check if we actually removed a nid. */
1094                 if (j == lpni->lpni_pref_nnids) {
1095                         LIBCFS_FREE(nids, size);
1096                         rc = -ENOENT;
1097                         goto out;
1098                 }
1099         }
1100
1101         lnet_net_lock(LNET_LOCK_EX);
1102         spin_lock(&lpni->lpni_lock);
1103         if (lpni->lpni_pref_nnids == 1) {
1104                 lpni->lpni_pref.nid = LNET_NID_ANY;
1105         } else if (lpni->lpni_pref_nnids == 2) {
1106                 oldnids = lpni->lpni_pref.nids;
1107                 if (oldnids[0] == nid)
1108                         lpni->lpni_pref.nid = oldnids[1];
1109                 else
1110                         lpni->lpni_pref.nid = oldnids[2];
1111         } else {
1112                 oldnids = lpni->lpni_pref.nids;
1113                 lpni->lpni_pref.nids = nids;
1114         }
1115         lpni->lpni_pref_nnids--;
1116         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1117         spin_unlock(&lpni->lpni_lock);
1118         lnet_net_unlock(LNET_LOCK_EX);
1119
1120         if (oldnids) {
1121                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1122                 CFS_FREE_PTR_ARRAY(oldnids, size);
1123         }
1124 out:
1125         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1126                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1127         return rc;
1128 }
1129
1130 lnet_nid_t
1131 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1132 {
1133         struct lnet_peer_ni *lpni;
1134         lnet_nid_t primary_nid = nid;
1135
1136         lpni = lnet_find_peer_ni_locked(nid);
1137         if (lpni) {
1138                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1139                 lnet_peer_ni_decref_locked(lpni);
1140         }
1141
1142         return primary_nid;
1143 }
1144
1145 bool
1146 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1147 __must_hold(&lp->lp_lock)
1148 {
1149         if (lnet_peer_discovery_disabled)
1150                 return true;
1151
1152         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1153             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1154                 return true;
1155         }
1156
1157         return false;
1158 }
1159
1160 /*
1161  * Peer Discovery
1162  */
1163 bool
1164 lnet_is_discovery_disabled(struct lnet_peer *lp)
1165 {
1166         bool rc = false;
1167
1168         spin_lock(&lp->lp_lock);
1169         rc = lnet_is_discovery_disabled_locked(lp);
1170         spin_unlock(&lp->lp_lock);
1171
1172         return rc;
1173 }
1174
1175 lnet_nid_t
1176 LNetPrimaryNID(lnet_nid_t nid)
1177 {
1178         struct lnet_peer *lp;
1179         struct lnet_peer_ni *lpni;
1180         lnet_nid_t primary_nid = nid;
1181         int rc = 0;
1182         int cpt;
1183
1184         if (nid == LNET_NID_LO_0)
1185                 return LNET_NID_LO_0;
1186
1187         cpt = lnet_net_lock_current();
1188         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1189         if (IS_ERR(lpni)) {
1190                 rc = PTR_ERR(lpni);
1191                 goto out_unlock;
1192         }
1193         lp = lpni->lpni_peer_net->lpn_peer;
1194
1195         while (!lnet_peer_is_uptodate(lp)) {
1196                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1197                 if (rc)
1198                         goto out_decref;
1199                 lp = lpni->lpni_peer_net->lpn_peer;
1200
1201                 /* Only try once if discovery is disabled */
1202                 if (lnet_is_discovery_disabled(lp))
1203                         break;
1204         }
1205         primary_nid = lp->lp_primary_nid;
1206 out_decref:
1207         lnet_peer_ni_decref_locked(lpni);
1208 out_unlock:
1209         lnet_net_unlock(cpt);
1210
1211         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1212                libcfs_nid2str(primary_nid), rc);
1213         return primary_nid;
1214 }
1215 EXPORT_SYMBOL(LNetPrimaryNID);
1216
1217 struct lnet_peer_net *
1218 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1219 {
1220         struct lnet_peer_net *peer_net;
1221         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1222                 if (peer_net->lpn_net_id == net_id)
1223                         return peer_net;
1224         }
1225         return NULL;
1226 }
1227
1228 /*
1229  * Attach a peer_ni to a peer_net and peer. This function assumes
1230  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1231  * may be attached to a different peer, in which case it will be
1232  * properly detached first. The whole operation is done atomically.
1233  *
1234  * Always returns 0.  This is the last function called from functions
1235  * that do return an int, so returning 0 here allows the compiler to
1236  * do a tail call.
1237  */
1238 static int
1239 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1240                                 struct lnet_peer_net *lpn,
1241                                 struct lnet_peer_ni *lpni,
1242                                 unsigned flags)
1243 {
1244         struct lnet_peer_table *ptable;
1245
1246         /* Install the new peer_ni */
1247         lnet_net_lock(LNET_LOCK_EX);
1248         /* Add peer_ni to global peer table hash, if necessary. */
1249         if (list_empty(&lpni->lpni_hashlist)) {
1250                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1251
1252                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1253                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1254                 ptable->pt_version++;
1255                 /* This is the 1st refcount on lpni. */
1256                 atomic_inc(&lpni->lpni_refcount);
1257         }
1258
1259         /* Detach the peer_ni from an existing peer, if necessary. */
1260         if (lpni->lpni_peer_net) {
1261                 LASSERT(lpni->lpni_peer_net != lpn);
1262                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1263                 lnet_peer_detach_peer_ni_locked(lpni);
1264                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1265                 lpni->lpni_peer_net = NULL;
1266         }
1267
1268         /* Add peer_ni to peer_net */
1269         lpni->lpni_peer_net = lpn;
1270         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1271         lnet_update_peer_net_healthv(lpni);
1272         lnet_peer_net_addref_locked(lpn);
1273
1274         /* Add peer_net to peer */
1275         if (!lpn->lpn_peer) {
1276                 lpn->lpn_peer = lp;
1277                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1278                 lnet_peer_addref_locked(lp);
1279         }
1280
1281         /* Add peer to global peer list, if necessary */
1282         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1283         if (list_empty(&lp->lp_peer_list)) {
1284                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1285                 ptable->pt_peers++;
1286         }
1287
1288
1289         /* Update peer state */
1290         spin_lock(&lp->lp_lock);
1291         if (flags & LNET_PEER_CONFIGURED) {
1292                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1293                         lp->lp_state |= LNET_PEER_CONFIGURED;
1294         }
1295         if (flags & LNET_PEER_MULTI_RAIL) {
1296                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1297                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1298                         lnet_peer_clr_non_mr_pref_nids(lp);
1299                 }
1300         }
1301         spin_unlock(&lp->lp_lock);
1302
1303         lp->lp_nnis++;
1304         lnet_net_unlock(LNET_LOCK_EX);
1305
1306         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1307                libcfs_nid2str(lp->lp_primary_nid),
1308                libcfs_nid2str(lpni->lpni_nid), flags);
1309
1310         return 0;
1311 }
1312
1313 /*
1314  * Create a new peer, with nid as its primary nid.
1315  *
1316  * Call with the lnet_api_mutex held.
1317  */
1318 static int
1319 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1320 {
1321         struct lnet_peer *lp;
1322         struct lnet_peer_net *lpn;
1323         struct lnet_peer_ni *lpni;
1324         int rc = 0;
1325
1326         LASSERT(nid != LNET_NID_ANY);
1327
1328         /*
1329          * No need for the lnet_net_lock here, because the
1330          * lnet_api_mutex is held.
1331          */
1332         lpni = lnet_find_peer_ni_locked(nid);
1333         if (lpni) {
1334                 /* A peer with this NID already exists. */
1335                 lp = lpni->lpni_peer_net->lpn_peer;
1336                 lnet_peer_ni_decref_locked(lpni);
1337                 /*
1338                  * This is an error if the peer was configured and the
1339                  * primary NID differs or an attempt is made to change
1340                  * the Multi-Rail flag. Otherwise the assumption is
1341                  * that an existing peer is being modified.
1342                  */
1343                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1344                         if (lp->lp_primary_nid != nid)
1345                                 rc = -EEXIST;
1346                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1347                                 rc = -EPERM;
1348                         goto out;
1349                 }
1350                 /* Delete and recreate as a configured peer. */
1351                 lnet_peer_del(lp);
1352         }
1353
1354         /* Create peer, peer_net, and peer_ni. */
1355         rc = -ENOMEM;
1356         lp = lnet_peer_alloc(nid);
1357         if (!lp)
1358                 goto out;
1359         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1360         if (!lpn)
1361                 goto out_free_lp;
1362         lpni = lnet_peer_ni_alloc(nid);
1363         if (!lpni)
1364                 goto out_free_lpn;
1365
1366         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1367
1368 out_free_lpn:
1369         LIBCFS_FREE(lpn, sizeof(*lpn));
1370 out_free_lp:
1371         LIBCFS_FREE(lp, sizeof(*lp));
1372 out:
1373         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1374                libcfs_nid2str(nid), flags, rc);
1375         return rc;
1376 }
1377
1378 /*
1379  * Add a NID to a peer. Call with ln_api_mutex held.
1380  *
1381  * Error codes:
1382  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1383  *  -EEXIST:   The NID was configured by DLC for a different peer.
1384  *  -ENOMEM:   Out of memory.
1385  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1386  *             non-multi-rail peer.
1387  */
1388 static int
1389 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1390 {
1391         struct lnet_peer_net *lpn;
1392         struct lnet_peer_ni *lpni;
1393         int rc = 0;
1394
1395         LASSERT(lp);
1396         LASSERT(nid != LNET_NID_ANY);
1397
1398         /* A configured peer can only be updated through configuration. */
1399         if (!(flags & LNET_PEER_CONFIGURED)) {
1400                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1401                         rc = -EPERM;
1402                         goto out;
1403                 }
1404         }
1405
1406         /*
1407          * The MULTI_RAIL flag can be set but not cleared, because
1408          * that would leave the peer struct in an invalid state.
1409          */
1410         if (flags & LNET_PEER_MULTI_RAIL) {
1411                 spin_lock(&lp->lp_lock);
1412                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1413                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1414                         lnet_peer_clr_non_mr_pref_nids(lp);
1415                 }
1416                 spin_unlock(&lp->lp_lock);
1417         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1418                 rc = -EPERM;
1419                 goto out;
1420         }
1421
1422         lpni = lnet_find_peer_ni_locked(nid);
1423         if (lpni) {
1424                 /*
1425                  * A peer_ni already exists. This is only a problem if
1426                  * it is not connected to this peer and was configured
1427                  * by DLC.
1428                  */
1429                 lnet_peer_ni_decref_locked(lpni);
1430                 if (lpni->lpni_peer_net->lpn_peer == lp)
1431                         goto out;
1432                 if (lnet_peer_ni_is_configured(lpni)) {
1433                         rc = -EEXIST;
1434                         goto out;
1435                 }
1436                 /* If this is the primary NID, destroy the peer. */
1437                 if (lnet_peer_ni_is_primary(lpni)) {
1438                         struct lnet_peer *rtr_lp =
1439                           lpni->lpni_peer_net->lpn_peer;
1440                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1441                         /*
1442                          * if we're trying to delete a router it means
1443                          * we're moving this peer NI to a new peer so must
1444                          * transfer router properties to the new peer
1445                          */
1446                         if (rtr_refcount > 0) {
1447                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1448                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1449                         }
1450                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1451                         lpni = lnet_peer_ni_alloc(nid);
1452                         if (!lpni) {
1453                                 rc = -ENOMEM;
1454                                 goto out;
1455                         }
1456                 }
1457         } else {
1458                 lpni = lnet_peer_ni_alloc(nid);
1459                 if (!lpni) {
1460                         rc = -ENOMEM;
1461                         goto out;
1462                 }
1463         }
1464
1465         /*
1466          * Get the peer_net. Check that we're not adding a second
1467          * peer_ni on a peer_net of a non-multi-rail peer.
1468          */
1469         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1470         if (!lpn) {
1471                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1472                 if (!lpn) {
1473                         rc = -ENOMEM;
1474                         goto out_free_lpni;
1475                 }
1476         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1477                 rc = -ENOTUNIQ;
1478                 goto out_free_lpni;
1479         }
1480
1481         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1482
1483 out_free_lpni:
1484         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1485         if (!lpni->lpni_peer_net)
1486                 LIBCFS_FREE(lpni, sizeof(*lpni));
1487 out:
1488         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1489                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1490                flags, rc);
1491         return rc;
1492 }
1493
1494 /*
1495  * Update the primary NID of a peer, if possible.
1496  *
1497  * Call with the lnet_api_mutex held.
1498  */
1499 static int
1500 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1501 {
1502         lnet_nid_t old = lp->lp_primary_nid;
1503         int rc = 0;
1504
1505         if (lp->lp_primary_nid == nid)
1506                 goto out;
1507         rc = lnet_peer_add_nid(lp, nid, flags);
1508         if (rc)
1509                 goto out;
1510         lp->lp_primary_nid = nid;
1511 out:
1512         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1513                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1514         return rc;
1515 }
1516
1517 /*
1518  * lpni creation initiated due to traffic either sending or receiving.
1519  */
1520 static int
1521 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1522 {
1523         struct lnet_peer *lp;
1524         struct lnet_peer_net *lpn;
1525         struct lnet_peer_ni *lpni;
1526         unsigned flags = 0;
1527         int rc = 0;
1528
1529         if (nid == LNET_NID_ANY) {
1530                 rc = -EINVAL;
1531                 goto out;
1532         }
1533
1534         /* lnet_net_lock is not needed here because ln_api_lock is held */
1535         lpni = lnet_find_peer_ni_locked(nid);
1536         if (lpni) {
1537                 /*
1538                  * We must have raced with another thread. Since we
1539                  * know next to nothing about a peer_ni created by
1540                  * traffic, we just assume everything is ok and
1541                  * return.
1542                  */
1543                 lnet_peer_ni_decref_locked(lpni);
1544                 goto out;
1545         }
1546
1547         /* Create peer, peer_net, and peer_ni. */
1548         rc = -ENOMEM;
1549         lp = lnet_peer_alloc(nid);
1550         if (!lp)
1551                 goto out;
1552         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1553         if (!lpn)
1554                 goto out_free_lp;
1555         lpni = lnet_peer_ni_alloc(nid);
1556         if (!lpni)
1557                 goto out_free_lpn;
1558         if (pref != LNET_NID_ANY)
1559                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1560
1561         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1562
1563 out_free_lpn:
1564         LIBCFS_FREE(lpn, sizeof(*lpn));
1565 out_free_lp:
1566         LIBCFS_FREE(lp, sizeof(*lp));
1567 out:
1568         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1569         return rc;
1570 }
1571
1572 /*
1573  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1574  *
1575  * This API handles the following combinations:
1576  *   Create a peer with its primary NI if only the prim_nid is provided
1577  *   Add a NID to a peer identified by the prim_nid. The peer identified
1578  *   by the prim_nid must already exist.
1579  *   The peer being created may be non-MR.
1580  *
1581  * The caller must hold ln_api_mutex. This prevents the peer from
1582  * being created/modified/deleted by a different thread.
1583  */
1584 int
1585 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1586 {
1587         struct lnet_peer *lp = NULL;
1588         struct lnet_peer_ni *lpni;
1589         unsigned flags;
1590
1591         /* The prim_nid must always be specified */
1592         if (prim_nid == LNET_NID_ANY)
1593                 return -EINVAL;
1594
1595         flags = LNET_PEER_CONFIGURED;
1596         if (mr)
1597                 flags |= LNET_PEER_MULTI_RAIL;
1598
1599         /*
1600          * If nid isn't specified, we must create a new peer with
1601          * prim_nid as its primary nid.
1602          */
1603         if (nid == LNET_NID_ANY)
1604                 return lnet_peer_add(prim_nid, flags);
1605
1606         /* Look up the prim_nid, which must exist. */
1607         lpni = lnet_find_peer_ni_locked(prim_nid);
1608         if (!lpni)
1609                 return -ENOENT;
1610         lnet_peer_ni_decref_locked(lpni);
1611         lp = lpni->lpni_peer_net->lpn_peer;
1612
1613         /* Peer must have been configured. */
1614         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1615                 CDEBUG(D_NET, "peer %s was not configured\n",
1616                        libcfs_nid2str(prim_nid));
1617                 return -ENOENT;
1618         }
1619
1620         /* Primary NID must match */
1621         if (lp->lp_primary_nid != prim_nid) {
1622                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1623                        libcfs_nid2str(prim_nid),
1624                        libcfs_nid2str(lp->lp_primary_nid));
1625                 return -ENODEV;
1626         }
1627
1628         /* Multi-Rail flag must match. */
1629         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1630                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1631                        libcfs_nid2str(prim_nid));
1632                 return -EPERM;
1633         }
1634
1635         return lnet_peer_add_nid(lp, nid, flags);
1636 }
1637
1638 /*
1639  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1640  *
1641  * This API handles the following combinations:
1642  *   Delete a NI from a peer if both prim_nid and nid are provided.
1643  *   Delete a peer if only prim_nid is provided.
1644  *   Delete a peer if its primary nid is provided.
1645  *
1646  * The caller must hold ln_api_mutex. This prevents the peer from
1647  * being modified/deleted by a different thread.
1648  */
1649 int
1650 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1651 {
1652         struct lnet_peer *lp;
1653         struct lnet_peer_ni *lpni;
1654         unsigned flags;
1655
1656         if (prim_nid == LNET_NID_ANY)
1657                 return -EINVAL;
1658
1659         lpni = lnet_find_peer_ni_locked(prim_nid);
1660         if (!lpni)
1661                 return -ENOENT;
1662         lnet_peer_ni_decref_locked(lpni);
1663         lp = lpni->lpni_peer_net->lpn_peer;
1664
1665         if (prim_nid != lp->lp_primary_nid) {
1666                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1667                        libcfs_nid2str(prim_nid),
1668                        libcfs_nid2str(lp->lp_primary_nid));
1669                 return -ENODEV;
1670         }
1671
1672         lnet_net_lock(LNET_LOCK_EX);
1673         if (lp->lp_rtr_refcount > 0) {
1674                 lnet_net_unlock(LNET_LOCK_EX);
1675                 CERROR("%s is a router. Can not be deleted\n",
1676                        libcfs_nid2str(prim_nid));
1677                 return -EBUSY;
1678         }
1679         lnet_net_unlock(LNET_LOCK_EX);
1680
1681         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1682                 return lnet_peer_del(lp);
1683
1684         flags = LNET_PEER_CONFIGURED;
1685         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1686                 flags |= LNET_PEER_MULTI_RAIL;
1687
1688         return lnet_peer_del_nid(lp, nid, flags);
1689 }
1690
1691 void
1692 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1693 {
1694         struct lnet_peer_table *ptable;
1695         struct lnet_peer_net *lpn;
1696
1697         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1698
1699         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1700         LASSERT(list_empty(&lpni->lpni_txq));
1701         LASSERT(lpni->lpni_txqnob == 0);
1702         LASSERT(list_empty(&lpni->lpni_peer_nis));
1703         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1704
1705         lpn = lpni->lpni_peer_net;
1706         lpni->lpni_peer_net = NULL;
1707         lpni->lpni_net = NULL;
1708
1709         /* remove the peer ni from the zombie list */
1710         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1711         spin_lock(&ptable->pt_zombie_lock);
1712         list_del_init(&lpni->lpni_hashlist);
1713         ptable->pt_zombies--;
1714         spin_unlock(&ptable->pt_zombie_lock);
1715
1716         if (lpni->lpni_pref_nnids > 1)
1717                 CFS_FREE_PTR_ARRAY(lpni->lpni_pref.nids, lpni->lpni_pref_nnids);
1718
1719         LIBCFS_FREE(lpni, sizeof(*lpni));
1720
1721         lnet_peer_net_decref_locked(lpn);
1722 }
1723
1724 struct lnet_peer_ni *
1725 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1726 {
1727         struct lnet_peer_ni *lpni = NULL;
1728         int rc;
1729
1730         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1731                 return ERR_PTR(-ESHUTDOWN);
1732
1733         /*
1734          * find if a peer_ni already exists.
1735          * If so then just return that.
1736          */
1737         lpni = lnet_find_peer_ni_locked(nid);
1738         if (lpni)
1739                 return lpni;
1740
1741         lnet_net_unlock(cpt);
1742
1743         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1744         if (rc) {
1745                 lpni = ERR_PTR(rc);
1746                 goto out_net_relock;
1747         }
1748
1749         lpni = lnet_find_peer_ni_locked(nid);
1750         LASSERT(lpni);
1751
1752 out_net_relock:
1753         lnet_net_lock(cpt);
1754
1755         return lpni;
1756 }
1757
1758 /*
1759  * Get a peer_ni for the given nid, create it if necessary. Takes a
1760  * hold on the peer_ni.
1761  */
1762 struct lnet_peer_ni *
1763 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1764 {
1765         struct lnet_peer_ni *lpni = NULL;
1766         int rc;
1767
1768         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1769                 return ERR_PTR(-ESHUTDOWN);
1770
1771         /*
1772          * find if a peer_ni already exists.
1773          * If so then just return that.
1774          */
1775         lpni = lnet_find_peer_ni_locked(nid);
1776         if (lpni)
1777                 return lpni;
1778
1779         /*
1780          * Slow path:
1781          * use the lnet_api_mutex to serialize the creation of the peer_ni
1782          * and the creation/deletion of the local ni/net. When a local ni is
1783          * created, if there exists a set of peer_nis on that network,
1784          * they need to be traversed and updated. When a local NI is
1785          * deleted, which could result in a network being deleted, then
1786          * all peer nis on that network need to be removed as well.
1787          *
1788          * Creation through traffic should also be serialized with
1789          * creation through DLC.
1790          */
1791         lnet_net_unlock(cpt);
1792         mutex_lock(&the_lnet.ln_api_mutex);
1793         /*
1794          * Shutdown is only set under the ln_api_lock, so a single
1795          * check here is sufficent.
1796          */
1797         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1798                 lpni = ERR_PTR(-ESHUTDOWN);
1799                 goto out_mutex_unlock;
1800         }
1801
1802         rc = lnet_peer_ni_traffic_add(nid, pref);
1803         if (rc) {
1804                 lpni = ERR_PTR(rc);
1805                 goto out_mutex_unlock;
1806         }
1807
1808         lpni = lnet_find_peer_ni_locked(nid);
1809         LASSERT(lpni);
1810
1811 out_mutex_unlock:
1812         mutex_unlock(&the_lnet.ln_api_mutex);
1813         lnet_net_lock(cpt);
1814
1815         /* Lock has been dropped, check again for shutdown. */
1816         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1817                 if (!IS_ERR(lpni))
1818                         lnet_peer_ni_decref_locked(lpni);
1819                 lpni = ERR_PTR(-ESHUTDOWN);
1820         }
1821
1822         return lpni;
1823 }
1824
1825 bool
1826 lnet_peer_gw_discovery(struct lnet_peer *lp)
1827 {
1828         bool rc = false;
1829
1830         spin_lock(&lp->lp_lock);
1831         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1832                 rc = true;
1833         spin_unlock(&lp->lp_lock);
1834
1835         return rc;
1836 }
1837
1838 bool
1839 lnet_peer_is_uptodate(struct lnet_peer *lp)
1840 {
1841         bool rc;
1842
1843         spin_lock(&lp->lp_lock);
1844         rc = lnet_peer_is_uptodate_locked(lp);
1845         spin_unlock(&lp->lp_lock);
1846         return rc;
1847 }
1848
1849 /*
1850  * Is a peer uptodate from the point of view of discovery?
1851  *
1852  * If it is currently being processed, obviously not.
1853  * A forced Ping or Push is also handled by the discovery thread.
1854  *
1855  * Otherwise look at whether the peer needs rediscovering.
1856  */
1857 bool
1858 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1859 __must_hold(&lp->lp_lock)
1860 {
1861         bool rc;
1862
1863         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1864                             LNET_PEER_FORCE_PING |
1865                             LNET_PEER_FORCE_PUSH)) {
1866                 rc = false;
1867         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1868                 rc = false;
1869         } else if (lnet_peer_needs_push(lp)) {
1870                 rc = false;
1871         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1872                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1873                         rc = true;
1874                 else
1875                         rc = false;
1876         } else {
1877                 rc = false;
1878         }
1879
1880         return rc;
1881 }
1882
1883 /*
1884  * Queue a peer for the attention of the discovery thread.  Call with
1885  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1886  * -EALREADY if the peer was already queued.
1887  */
1888 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1889 {
1890         int rc;
1891
1892         spin_lock(&lp->lp_lock);
1893         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1894                 lp->lp_state |= LNET_PEER_DISCOVERING;
1895         spin_unlock(&lp->lp_lock);
1896         if (list_empty(&lp->lp_dc_list)) {
1897                 lnet_peer_addref_locked(lp);
1898                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1899                 wake_up(&the_lnet.ln_dc_waitq);
1900                 rc = 0;
1901         } else {
1902                 rc = -EALREADY;
1903         }
1904
1905         CDEBUG(D_NET, "Queue peer %s: %d\n",
1906                libcfs_nid2str(lp->lp_primary_nid), rc);
1907
1908         return rc;
1909 }
1910
1911 /*
1912  * Discovery of a peer is complete. Wake all waiters on the peer.
1913  * Call with lnet_net_lock/EX held.
1914  */
1915 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1916 {
1917         struct lnet_msg *msg, *tmp;
1918         int rc = 0;
1919         LIST_HEAD(pending_msgs);
1920
1921         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1922                libcfs_nid2str(lp->lp_primary_nid));
1923
1924         list_del_init(&lp->lp_dc_list);
1925         spin_lock(&lp->lp_lock);
1926         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1927         spin_unlock(&lp->lp_lock);
1928         wake_up_all(&lp->lp_dc_waitq);
1929
1930         if (lp->lp_rtr_refcount > 0)
1931                 lnet_router_discovery_complete(lp);
1932
1933         lnet_net_unlock(LNET_LOCK_EX);
1934
1935         /* iterate through all pending messages and send them again */
1936         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1937                 list_del_init(&msg->msg_list);
1938                 if (lp->lp_dc_error) {
1939                         lnet_finalize(msg, lp->lp_dc_error);
1940                         continue;
1941                 }
1942
1943                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1944                        lnet_msgtyp2str(msg->msg_type),
1945                        libcfs_id2str(msg->msg_target));
1946                 rc = lnet_send(msg->msg_src_nid_param, msg,
1947                                msg->msg_rtr_nid_param);
1948                 if (rc < 0) {
1949                         CNETERR("Error sending %s to %s: %d\n",
1950                                lnet_msgtyp2str(msg->msg_type),
1951                                libcfs_id2str(msg->msg_target), rc);
1952                         lnet_finalize(msg, rc);
1953                 }
1954         }
1955         lnet_net_lock(LNET_LOCK_EX);
1956         lnet_peer_decref_locked(lp);
1957 }
1958
1959 /*
1960  * Handle inbound push.
1961  * Like any event handler, called with lnet_res_lock/CPT held.
1962  */
1963 void lnet_peer_push_event(struct lnet_event *ev)
1964 {
1965         struct lnet_ping_buffer *pbuf;
1966         struct lnet_peer *lp;
1967
1968         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
1969
1970         /* lnet_find_peer() adds a refcount */
1971         lp = lnet_find_peer(ev->source.nid);
1972         if (!lp) {
1973                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1974                        libcfs_nid2str(ev->initiator.nid),
1975                        libcfs_nid2str(ev->source.nid));
1976                 pbuf->pb_needs_post = true;
1977                 return;
1978         }
1979
1980         /* Ensure peer state remains consistent while we modify it. */
1981         spin_lock(&lp->lp_lock);
1982
1983         /*
1984          * If some kind of error happened the contents of the message
1985          * cannot be used. Clear the NIDS_UPTODATE and set the
1986          * FORCE_PING flag to trigger a ping.
1987          */
1988         if (ev->status) {
1989                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1990                 lp->lp_state |= LNET_PEER_FORCE_PING;
1991                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1992                        ev->status,
1993                        libcfs_nid2str(lp->lp_primary_nid),
1994                        libcfs_nid2str(ev->source.nid));
1995                 goto out;
1996         }
1997
1998         /*
1999          * A push with invalid or corrupted info. Clear the UPTODATE
2000          * flag to trigger a ping.
2001          */
2002         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2003                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2004                 lp->lp_state |= LNET_PEER_FORCE_PING;
2005                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2006                        libcfs_nid2str(lp->lp_primary_nid));
2007                 goto out;
2008         }
2009
2010         /*
2011          * Make sure we'll allocate the correct size ping buffer when
2012          * pinging the peer.
2013          */
2014         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2015                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2016
2017         /*
2018          * A non-Multi-Rail peer is not supposed to be capable of
2019          * sending a push.
2020          */
2021         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2022                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2023                        libcfs_nid2str(lp->lp_primary_nid));
2024                 goto out;
2025         }
2026
2027         /*
2028          * The peer may have discovery disabled at its end. Set
2029          * NO_DISCOVERY as appropriate.
2030          */
2031         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2032                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2033                        libcfs_nid2str(lp->lp_primary_nid));
2034                 /*
2035                  * If the peer is going from discovery enabled to
2036                  * discovery disabled, we need to reflect that in our
2037                  * representation of the peer.
2038                  */
2039                 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2040                                       LNET_PEER_DISCOVERING)))
2041                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2042                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2043         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2044                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2045                        libcfs_nid2str(lp->lp_primary_nid));
2046                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2047         }
2048
2049         /*
2050          * Update the MULTI_RAIL flag based on the push. If the peer
2051          * was configured with DLC then the setting should match what
2052          * DLC put in.
2053          * NB: We verified above that the MR feature bit is set in pi_features
2054          */
2055         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2056                 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2057                        libcfs_nid2str(lp->lp_primary_nid), lp);
2058         } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2059                 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2060                       libcfs_nid2str(lp->lp_primary_nid));
2061         } else if (lnet_peer_discovery_disabled) {
2062                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2063                        libcfs_nid2str(lp->lp_primary_nid), lp);
2064         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2065                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2066                        libcfs_nid2str(lp->lp_primary_nid), lp);
2067         } else {
2068                 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2069                        libcfs_nid2str(lp->lp_primary_nid), lp);
2070                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2071                 lnet_peer_clr_non_mr_pref_nids(lp);
2072         }
2073
2074         /*
2075          * Check for truncation of the Put message. Clear the
2076          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2077          * and tell discovery to allocate a bigger buffer.
2078          */
2079         if (ev->mlength < ev->rlength) {
2080                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2081                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2082                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2083                 lp->lp_state |= LNET_PEER_FORCE_PING;
2084                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2085                        libcfs_nid2str(lp->lp_primary_nid),
2086                        pbuf->pb_info.pi_nnis);
2087                 goto out;
2088         }
2089
2090         /* always assume new data */
2091         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2092         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2093
2094         /*
2095          * If there is data present that hasn't been processed yet,
2096          * we'll replace it if the Put contained newer data and it
2097          * fits. We're racing with a Ping or earlier Push in this
2098          * case.
2099          */
2100         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2101                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2102                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2103                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2104                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2105                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2106                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2107                               libcfs_nid2str(lp->lp_primary_nid),
2108                               LNET_PING_BUFFER_SEQNO(pbuf),
2109                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2110                 }
2111                 goto out;
2112         }
2113
2114         /*
2115          * Allocate a buffer to copy the data. On a failure we drop
2116          * the Push and set FORCE_PING to force the discovery
2117          * thread to fix the problem by pinging the peer.
2118          */
2119         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2120         if (!lp->lp_data) {
2121                 lp->lp_state |= LNET_PEER_FORCE_PING;
2122                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2123                        libcfs_nid2str(lp->lp_primary_nid),
2124                        LNET_PING_BUFFER_SEQNO(pbuf));
2125                 goto out;
2126         }
2127
2128         /* Success */
2129         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2130                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2131         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2132         CDEBUG(D_NET, "Received Push %s %u\n",
2133                libcfs_nid2str(lp->lp_primary_nid),
2134                LNET_PING_BUFFER_SEQNO(pbuf));
2135
2136 out:
2137         /* We've processed this buffer. It can be reposted */
2138         pbuf->pb_needs_post = true;
2139
2140         /*
2141          * Queue the peer for discovery if not done, force it on the request
2142          * queue and wake the discovery thread if the peer was already queued,
2143          * because its status changed.
2144          */
2145         spin_unlock(&lp->lp_lock);
2146         lnet_net_lock(LNET_LOCK_EX);
2147         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2148                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2149                 wake_up(&the_lnet.ln_dc_waitq);
2150         }
2151         /* Drop refcount from lookup */
2152         lnet_peer_decref_locked(lp);
2153         lnet_net_unlock(LNET_LOCK_EX);
2154 }
2155
2156 /*
2157  * Clear the discovery error state, unless we're already discovering
2158  * this peer, in which case the error is current.
2159  */
2160 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2161 {
2162         spin_lock(&lp->lp_lock);
2163         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2164                 lp->lp_dc_error = 0;
2165         spin_unlock(&lp->lp_lock);
2166 }
2167
2168 /*
2169  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2170  * dropped/retaken within this function. An lnet_peer_ni is passed in
2171  * because discovery could tear down an lnet_peer.
2172  */
2173 int
2174 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2175 {
2176         DEFINE_WAIT(wait);
2177         struct lnet_peer *lp;
2178         int rc = 0;
2179         int count = 0;
2180
2181 again:
2182         lnet_net_unlock(cpt);
2183         lnet_net_lock(LNET_LOCK_EX);
2184         lp = lpni->lpni_peer_net->lpn_peer;
2185         lnet_peer_clear_discovery_error(lp);
2186
2187         /*
2188          * We're willing to be interrupted. The lpni can become a
2189          * zombie if we race with DLC, so we must check for that.
2190          */
2191         for (;;) {
2192                 /* Keep lp alive when the lnet_net_lock is unlocked */
2193                 lnet_peer_addref_locked(lp);
2194                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2195                 if (signal_pending(current))
2196                         break;
2197                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2198                         break;
2199                 /*
2200                  * Don't repeat discovery if discovery is disabled. This is
2201                  * done to ensure we can use discovery as a standard ping as
2202                  * well for backwards compatibility with routers which do not
2203                  * have discovery or have discovery disabled
2204                  */
2205                 if (lnet_is_discovery_disabled(lp) && count > 0)
2206                         break;
2207                 if (lp->lp_dc_error)
2208                         break;
2209                 if (lnet_peer_is_uptodate(lp))
2210                         break;
2211                 lnet_peer_queue_for_discovery(lp);
2212                 count++;
2213                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2214
2215                 /*
2216                  * If caller requested a non-blocking operation then
2217                  * return immediately. Once discovery is complete any
2218                  * pending messages that were stopped due to discovery
2219                  * will be transmitted.
2220                  */
2221                 if (!block)
2222                         break;
2223
2224                 lnet_net_unlock(LNET_LOCK_EX);
2225                 schedule();
2226                 finish_wait(&lp->lp_dc_waitq, &wait);
2227                 lnet_net_lock(LNET_LOCK_EX);
2228                 lnet_peer_decref_locked(lp);
2229                 /* Peer may have changed */
2230                 lp = lpni->lpni_peer_net->lpn_peer;
2231         }
2232         finish_wait(&lp->lp_dc_waitq, &wait);
2233
2234         lnet_net_unlock(LNET_LOCK_EX);
2235         lnet_net_lock(cpt);
2236         lnet_peer_decref_locked(lp);
2237         /*
2238          * The peer may have changed, so re-check and rediscover if that turns
2239          * out to have been the case. The reference count on lp ensured that
2240          * even if it was unlinked from lpni the memory could not be recycled.
2241          * Thus the check below is sufficient to determine whether the peer
2242          * changed. If the peer changed, then lp must not be dereferenced.
2243          */
2244         if (lp != lpni->lpni_peer_net->lpn_peer)
2245                 goto again;
2246
2247         if (signal_pending(current))
2248                 rc = -EINTR;
2249         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2250                 rc = -ESHUTDOWN;
2251         else if (lp->lp_dc_error)
2252                 rc = lp->lp_dc_error;
2253         else if (!block)
2254                 CDEBUG(D_NET, "non-blocking discovery\n");
2255         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2256                 goto again;
2257
2258         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2259                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2260                libcfs_nid2str(lpni->lpni_nid), rc,
2261                (!block) ? "pending discovery" : "discovery complete");
2262
2263         return rc;
2264 }
2265
2266 /* Handle an incoming ack for a push. */
2267 static void
2268 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2269 {
2270         struct lnet_ping_buffer *pbuf;
2271
2272         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2273         spin_lock(&lp->lp_lock);
2274         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2275         lp->lp_push_error = ev->status;
2276         if (ev->status)
2277                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2278         else
2279                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2280         spin_unlock(&lp->lp_lock);
2281
2282         CDEBUG(D_NET, "peer %s ev->status %d\n",
2283                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2284 }
2285
2286 /* Handle a Reply message. This is the reply to a Ping message. */
2287 static void
2288 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2289 {
2290         struct lnet_ping_buffer *pbuf;
2291         int rc;
2292
2293         spin_lock(&lp->lp_lock);
2294
2295         lp->lp_disc_src_nid = ev->target.nid;
2296
2297         /*
2298          * If some kind of error happened the contents of message
2299          * cannot be used. Set PING_FAILED to trigger a retry.
2300          */
2301         if (ev->status) {
2302                 lp->lp_state |= LNET_PEER_PING_FAILED;
2303                 lp->lp_ping_error = ev->status;
2304                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2305                        ev->status,
2306                        libcfs_nid2str(lp->lp_primary_nid),
2307                        libcfs_nid2str(ev->source.nid));
2308                 goto out;
2309         }
2310
2311         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2312         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2313                 lnet_swap_pinginfo(pbuf);
2314
2315         /*
2316          * A reply with invalid or corrupted info. Set PING_FAILED to
2317          * trigger a retry.
2318          */
2319         rc = lnet_ping_info_validate(&pbuf->pb_info);
2320         if (rc) {
2321                 lp->lp_state |= LNET_PEER_PING_FAILED;
2322                 lp->lp_ping_error = 0;
2323                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2324                        libcfs_nid2str(lp->lp_primary_nid), rc);
2325                 goto out;
2326         }
2327
2328
2329         /*
2330          * The peer may have discovery disabled at its end. Set
2331          * NO_DISCOVERY as appropriate.
2332          */
2333         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2334             !lnet_peer_discovery_disabled) {
2335                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2336                        libcfs_nid2str(lp->lp_primary_nid));
2337                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2338         } else {
2339                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2340                        libcfs_nid2str(lp->lp_primary_nid));
2341                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2342         }
2343
2344         /*
2345          * Update the MULTI_RAIL flag based on the reply. If the peer
2346          * was configured with DLC then the setting should match what
2347          * DLC put in.
2348          */
2349         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2350                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2351                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2352                                libcfs_nid2str(lp->lp_primary_nid), lp);
2353                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2354                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2355                               libcfs_nid2str(lp->lp_primary_nid));
2356                 } else if (lnet_peer_discovery_disabled) {
2357                         CDEBUG(D_NET,
2358                                "peer %s(%p) not MR: DD disabled locally\n",
2359                                libcfs_nid2str(lp->lp_primary_nid), lp);
2360                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2361                         CDEBUG(D_NET,
2362                                "peer %s(%p) not MR: DD disabled remotely\n",
2363                                libcfs_nid2str(lp->lp_primary_nid), lp);
2364                 } else {
2365                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2366                                libcfs_nid2str(lp->lp_primary_nid), lp);
2367                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2368                         lnet_peer_clr_non_mr_pref_nids(lp);
2369                 }
2370         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2371                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2372                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2373                               libcfs_nid2str(lp->lp_primary_nid));
2374                 } else {
2375                         CERROR("Multi-Rail state vanished from %s\n",
2376                                libcfs_nid2str(lp->lp_primary_nid));
2377                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2378                 }
2379         }
2380
2381         /*
2382          * Make sure we'll allocate the correct size ping buffer when
2383          * pinging the peer.
2384          */
2385         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2386                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2387
2388         /*
2389          * Check for truncation of the Reply. Clear PING_SENT and set
2390          * PING_FAILED to trigger a retry.
2391          */
2392         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2393                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2394                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2395                 lp->lp_state |= LNET_PEER_PING_FAILED;
2396                 lp->lp_ping_error = 0;
2397                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2398                        libcfs_nid2str(lp->lp_primary_nid),
2399                        pbuf->pb_info.pi_nnis);
2400                 goto out;
2401         }
2402
2403         /*
2404          * Check the sequence numbers in the reply. These are only
2405          * available if the reply came from a Multi-Rail peer.
2406          */
2407         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2408             pbuf->pb_info.pi_nnis > 1 &&
2409             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2410                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2411                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2412                                 libcfs_nid2str(lp->lp_primary_nid),
2413                                 LNET_PING_BUFFER_SEQNO(pbuf),
2414                                 lp->lp_peer_seqno);
2415
2416                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2417         }
2418
2419         /* We're happy with the state of the data in the buffer. */
2420         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2421                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2422         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2423                 lnet_ping_buffer_decref(lp->lp_data);
2424         else
2425                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2426         lnet_ping_buffer_addref(pbuf);
2427         lp->lp_data = pbuf;
2428 out:
2429         lp->lp_state &= ~LNET_PEER_PING_SENT;
2430         spin_unlock(&lp->lp_lock);
2431
2432         lnet_net_lock(LNET_LOCK_EX);
2433         /*
2434          * If this peer is a gateway, call the routing callback to
2435          * handle the ping reply
2436          */
2437         if (lp->lp_rtr_refcount > 0)
2438                 lnet_router_discovery_ping_reply(lp);
2439         lnet_net_unlock(LNET_LOCK_EX);
2440 }
2441
2442 /*
2443  * Send event handling. Only matters for error cases, where we clean
2444  * up state on the peer and peer_ni that would otherwise be updated in
2445  * the REPLY event handler for a successful Ping, and the ACK event
2446  * handler for a successful Push.
2447  */
2448 static int
2449 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2450 {
2451         int rc = 0;
2452
2453         if (!ev->status)
2454                 goto out;
2455
2456         spin_lock(&lp->lp_lock);
2457         if (ev->msg_type == LNET_MSG_GET) {
2458                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2459                 lp->lp_state |= LNET_PEER_PING_FAILED;
2460                 lp->lp_ping_error = ev->status;
2461         } else { /* ev->msg_type == LNET_MSG_PUT */
2462                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2463                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2464                 lp->lp_push_error = ev->status;
2465         }
2466         spin_unlock(&lp->lp_lock);
2467         rc = LNET_REDISCOVER_PEER;
2468 out:
2469         CDEBUG(D_NET, "%s Send to %s: %d\n",
2470                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2471                 libcfs_nid2str(ev->target.nid), rc);
2472         return rc;
2473 }
2474
2475 /*
2476  * Unlink event handling. This event is only seen if a call to
2477  * LNetMDUnlink() caused the event to be unlinked. If this call was
2478  * made after the event was set up in LNetGet() or LNetPut() then we
2479  * assume the Ping or Push timed out.
2480  */
2481 static void
2482 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2483 {
2484         spin_lock(&lp->lp_lock);
2485         /* We've passed through LNetGet() */
2486         if (lp->lp_state & LNET_PEER_PING_SENT) {
2487                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2488                 lp->lp_state |= LNET_PEER_PING_FAILED;
2489                 lp->lp_ping_error = -ETIMEDOUT;
2490                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2491                         libcfs_nid2str(lp->lp_primary_nid));
2492         }
2493         /* We've passed through LNetPut() */
2494         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2495                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2496                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2497                 lp->lp_push_error = -ETIMEDOUT;
2498                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2499                         libcfs_nid2str(lp->lp_primary_nid));
2500         }
2501         spin_unlock(&lp->lp_lock);
2502 }
2503
2504 /*
2505  * Event handler for the discovery EQ.
2506  *
2507  * Called with lnet_res_lock(cpt) held. The cpt is the
2508  * lnet_cpt_of_cookie() of the md handle cookie.
2509  */
2510 static void lnet_discovery_event_handler(struct lnet_event *event)
2511 {
2512         struct lnet_peer *lp = event->md_user_ptr;
2513         struct lnet_ping_buffer *pbuf;
2514         int rc;
2515
2516         /* discovery needs to take another look */
2517         rc = LNET_REDISCOVER_PEER;
2518
2519         CDEBUG(D_NET, "Received event: %d\n", event->type);
2520
2521         switch (event->type) {
2522         case LNET_EVENT_ACK:
2523                 lnet_discovery_event_ack(lp, event);
2524                 break;
2525         case LNET_EVENT_REPLY:
2526                 lnet_discovery_event_reply(lp, event);
2527                 break;
2528         case LNET_EVENT_SEND:
2529                 /* Only send failure triggers a retry. */
2530                 rc = lnet_discovery_event_send(lp, event);
2531                 break;
2532         case LNET_EVENT_UNLINK:
2533                 /* LNetMDUnlink() was called */
2534                 lnet_discovery_event_unlink(lp, event);
2535                 break;
2536         default:
2537                 /* Invalid events. */
2538                 LBUG();
2539         }
2540         lnet_net_lock(LNET_LOCK_EX);
2541         if (event->unlinked) {
2542                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2543                 lnet_ping_buffer_decref(pbuf);
2544                 lnet_peer_decref_locked(lp);
2545         }
2546
2547         /* put peer back at end of request queue, if discovery not already
2548          * done */
2549         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2550                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2551                 wake_up(&the_lnet.ln_dc_waitq);
2552         }
2553         lnet_net_unlock(LNET_LOCK_EX);
2554 }
2555
2556 /*
2557  * Build a peer from incoming data.
2558  *
2559  * The NIDs in the incoming data are supposed to be structured as follows:
2560  *  - loopback
2561  *  - primary NID
2562  *  - other NIDs in same net
2563  *  - NIDs in second net
2564  *  - NIDs in third net
2565  *  - ...
2566  * This due to the way the list of NIDs in the data is created.
2567  *
2568  * Note that this function will mark the peer uptodate unless an
2569  * ENOMEM is encontered. All other errors are due to a conflict
2570  * between the DLC configuration and what discovery sees. We treat DLC
2571  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2572  * peer from becoming stuck in discovery.
2573  */
2574 static int lnet_peer_merge_data(struct lnet_peer *lp,
2575                                 struct lnet_ping_buffer *pbuf)
2576 {
2577         struct lnet_peer_ni *lpni;
2578         lnet_nid_t *curnis = NULL;
2579         struct lnet_ni_status *addnis = NULL;
2580         lnet_nid_t *delnis = NULL;
2581         unsigned flags;
2582         int ncurnis;
2583         int naddnis;
2584         int ndelnis;
2585         int nnis = 0;
2586         int i;
2587         int j;
2588         int rc;
2589
2590         flags = LNET_PEER_DISCOVERED;
2591         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2592                 flags |= LNET_PEER_MULTI_RAIL;
2593
2594         /*
2595          * Cache the routing feature for the peer; whether it is enabled
2596          * for disabled as reported by the remote peer.
2597          */
2598         spin_lock(&lp->lp_lock);
2599         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2600                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2601         else
2602                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2603         spin_unlock(&lp->lp_lock);
2604
2605         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2606         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2607         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2608         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2609         if (!curnis || !addnis || !delnis) {
2610                 rc = -ENOMEM;
2611                 goto out;
2612         }
2613         ncurnis = 0;
2614         naddnis = 0;
2615         ndelnis = 0;
2616
2617         /* Construct the list of NIDs present in peer. */
2618         lpni = NULL;
2619         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2620                 curnis[ncurnis++] = lpni->lpni_nid;
2621
2622         /*
2623          * Check for NIDs in pbuf not present in curnis[].
2624          * The loop starts at 1 to skip the loopback NID.
2625          */
2626         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2627                 for (j = 0; j < ncurnis; j++)
2628                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2629                                 break;
2630                 if (j == ncurnis)
2631                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2632         }
2633         /*
2634          * Check for NIDs in curnis[] not present in pbuf.
2635          * The nested loop starts at 1 to skip the loopback NID.
2636          *
2637          * But never add the loopback NID to delnis[]: if it is
2638          * present in curnis[] then this peer is for this node.
2639          */
2640         for (i = 0; i < ncurnis; i++) {
2641                 if (curnis[i] == LNET_NID_LO_0)
2642                         continue;
2643                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2644                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2645                                 /*
2646                                  * update the information we cache for the
2647                                  * peer with the latest information we
2648                                  * received
2649                                  */
2650                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2651                                 if (lpni) {
2652                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2653                                         lnet_peer_ni_decref_locked(lpni);
2654                                 }
2655                                 break;
2656                         }
2657                 }
2658                 if (j == pbuf->pb_info.pi_nnis)
2659                         delnis[ndelnis++] = curnis[i];
2660         }
2661
2662         /*
2663          * If we get here and the discovery is disabled then we don't want
2664          * to add or delete any NIs. We just updated the ones we have some
2665          * information on, and call it a day
2666          */
2667         rc = 0;
2668         if (lnet_is_discovery_disabled(lp))
2669                 goto out;
2670
2671         for (i = 0; i < naddnis; i++) {
2672                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2673                 if (rc) {
2674                         CERROR("Error adding NID %s to peer %s: %d\n",
2675                                libcfs_nid2str(addnis[i].ns_nid),
2676                                libcfs_nid2str(lp->lp_primary_nid), rc);
2677                         if (rc == -ENOMEM)
2678                                 goto out;
2679                 }
2680                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2681                 if (lpni) {
2682                         lpni->lpni_ns_status = addnis[i].ns_status;
2683                         lnet_peer_ni_decref_locked(lpni);
2684                 }
2685         }
2686
2687         for (i = 0; i < ndelnis; i++) {
2688                 /*
2689                  * for routers it's okay to delete the primary_nid because
2690                  * the upper layers don't really rely on it. So if we're
2691                  * being told that the router changed its primary_nid
2692                  * then it's okay to delete it.
2693                  */
2694                 if (lp->lp_rtr_refcount > 0)
2695                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2696                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2697                 if (rc) {
2698                         CERROR("Error deleting NID %s from peer %s: %d\n",
2699                                libcfs_nid2str(delnis[i]),
2700                                libcfs_nid2str(lp->lp_primary_nid), rc);
2701                         if (rc == -ENOMEM)
2702                                 goto out;
2703                 }
2704         }
2705         /*
2706          * Errors other than -ENOMEM are due to peers having been
2707          * configured with DLC. Ignore these because DLC overrides
2708          * Discovery.
2709          */
2710         rc = 0;
2711 out:
2712         CFS_FREE_PTR_ARRAY(curnis, nnis);
2713         CFS_FREE_PTR_ARRAY(addnis, nnis);
2714         CFS_FREE_PTR_ARRAY(delnis, nnis);
2715         lnet_ping_buffer_decref(pbuf);
2716         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2717
2718         if (rc) {
2719                 spin_lock(&lp->lp_lock);
2720                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2721                 lp->lp_state |= LNET_PEER_FORCE_PING;
2722                 spin_unlock(&lp->lp_lock);
2723         }
2724         return rc;
2725 }
2726
2727 /*
2728  * The data in pbuf says lp is its primary peer, but the data was
2729  * received by a different peer. Try to update lp with the data.
2730  */
2731 static int
2732 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2733 {
2734         struct lnet_handle_md mdh;
2735
2736         /* Queue lp for discovery, and force it on the request queue. */
2737         lnet_net_lock(LNET_LOCK_EX);
2738         if (lnet_peer_queue_for_discovery(lp))
2739                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2740         lnet_net_unlock(LNET_LOCK_EX);
2741
2742         LNetInvalidateMDHandle(&mdh);
2743
2744         /*
2745          * Decide whether we can move the peer to the DATA_PRESENT state.
2746          *
2747          * We replace stale data for a multi-rail peer, repair PING_FAILED
2748          * status, and preempt FORCE_PING.
2749          *
2750          * If after that we have DATA_PRESENT, we merge it into this peer.
2751          */
2752         spin_lock(&lp->lp_lock);
2753         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2754                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2755                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2756                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2757                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2758                         lnet_ping_buffer_decref(pbuf);
2759                         pbuf = lp->lp_data;
2760                         lp->lp_data = NULL;
2761                 }
2762         }
2763         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2764                 lnet_ping_buffer_decref(lp->lp_data);
2765                 lp->lp_data = NULL;
2766                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2767         }
2768         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2769                 mdh = lp->lp_ping_mdh;
2770                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2771                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2772                 lp->lp_ping_error = 0;
2773         }
2774         if (lp->lp_state & LNET_PEER_FORCE_PING)
2775                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2776         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2777         spin_unlock(&lp->lp_lock);
2778
2779         if (!LNetMDHandleIsInvalid(mdh))
2780                 LNetMDUnlink(mdh);
2781
2782         if (pbuf)
2783                 return lnet_peer_merge_data(lp, pbuf);
2784
2785         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2786         return 0;
2787 }
2788
2789 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2790 {
2791         int i;
2792
2793         for (i = 0; i < pinfo->pi_nnis; i++) {
2794                 if (pinfo->pi_ni[i].ns_nid == nid)
2795                         return true;
2796         }
2797
2798         return false;
2799 }
2800
2801 /*
2802  * Update a peer using the data received.
2803  */
2804 static int lnet_peer_data_present(struct lnet_peer *lp)
2805 __must_hold(&lp->lp_lock)
2806 {
2807         struct lnet_ping_buffer *pbuf;
2808         struct lnet_peer_ni *lpni;
2809         lnet_nid_t nid = LNET_NID_ANY;
2810         unsigned flags;
2811         int rc = 0;
2812
2813         pbuf = lp->lp_data;
2814         lp->lp_data = NULL;
2815         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2816         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2817         spin_unlock(&lp->lp_lock);
2818
2819         /*
2820          * Modifications of peer structures are done while holding the
2821          * ln_api_mutex. A global lock is required because we may be
2822          * modifying multiple peer structures, and a mutex greatly
2823          * simplifies memory management.
2824          *
2825          * The actual changes to the data structures must also protect
2826          * against concurrent lookups, for which the lnet_net_lock in
2827          * LNET_LOCK_EX mode is used.
2828          */
2829         mutex_lock(&the_lnet.ln_api_mutex);
2830         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2831                 rc = -ESHUTDOWN;
2832                 goto out;
2833         }
2834
2835         /*
2836          * If this peer is not on the peer list then it is being torn
2837          * down, and our reference count may be all that is keeping it
2838          * alive. Don't do any work on it.
2839          */
2840         if (list_empty(&lp->lp_peer_list))
2841                 goto out;
2842
2843         flags = LNET_PEER_DISCOVERED;
2844         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2845                 flags |= LNET_PEER_MULTI_RAIL;
2846
2847         /*
2848          * Check whether the primary NID in the message matches the
2849          * primary NID of the peer. If it does, update the peer, if
2850          * it it does not, check whether there is already a peer with
2851          * that primary NID. If no such peer exists, try to update
2852          * the primary NID of the current peer (allowed if it was
2853          * created due to message traffic) and complete the update.
2854          * If the peer did exist, hand off the data to it.
2855          *
2856          * The peer for the loopback interface is a special case: this
2857          * is the peer for the local node, and we want to set its
2858          * primary NID to the correct value here. Moreover, this peer
2859          * can show up with only the loopback NID in the ping buffer.
2860          */
2861         if (pbuf->pb_info.pi_nnis <= 1)
2862                 goto out;
2863         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2864         if (lp->lp_primary_nid == LNET_NID_LO_0) {
2865                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2866                 if (!rc)
2867                         rc = lnet_peer_merge_data(lp, pbuf);
2868         /*
2869          * if the primary nid of the peer is present in the ping info returned
2870          * from the peer, but it's not the local primary peer we have
2871          * cached and discovery is disabled, then we don't want to update
2872          * our local peer info, by adding or removing NIDs, we just want
2873          * to update the status of the nids that we currently have
2874          * recorded in that peer.
2875          */
2876         } else if (lp->lp_primary_nid == nid ||
2877                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2878                     lnet_is_discovery_disabled(lp))) {
2879                 rc = lnet_peer_merge_data(lp, pbuf);
2880         } else {
2881                 lpni = lnet_find_peer_ni_locked(nid);
2882                 if (!lpni) {
2883                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2884                         if (rc) {
2885                                 CERROR("Primary NID error %s versus %s: %d\n",
2886                                        libcfs_nid2str(lp->lp_primary_nid),
2887                                        libcfs_nid2str(nid), rc);
2888                         } else {
2889                                 rc = lnet_peer_merge_data(lp, pbuf);
2890                         }
2891                 } else {
2892                         struct lnet_peer *new_lp;
2893                         new_lp = lpni->lpni_peer_net->lpn_peer;
2894                         /*
2895                          * if lp has discovery/MR enabled that means new_lp
2896                          * should have discovery/MR enabled as well, since
2897                          * it's the same peer, which we're about to merge
2898                          */
2899                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2900                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2901                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2902                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2903
2904                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2905                         lnet_consolidate_routes_locked(lp, new_lp);
2906                         lnet_peer_ni_decref_locked(lpni);
2907                 }
2908         }
2909 out:
2910         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2911                lp->lp_state);
2912         mutex_unlock(&the_lnet.ln_api_mutex);
2913
2914         spin_lock(&lp->lp_lock);
2915         /* Tell discovery to re-check the peer immediately. */
2916         if (!rc)
2917                 rc = LNET_REDISCOVER_PEER;
2918         return rc;
2919 }
2920
2921 /*
2922  * A ping failed. Clear the PING_FAILED state and set the
2923  * FORCE_PING state, to ensure a retry even if discovery is
2924  * disabled. This avoids being left with incorrect state.
2925  */
2926 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2927 __must_hold(&lp->lp_lock)
2928 {
2929         struct lnet_handle_md mdh;
2930         int rc;
2931
2932         mdh = lp->lp_ping_mdh;
2933         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2934         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2935         lp->lp_state |= LNET_PEER_FORCE_PING;
2936         rc = lp->lp_ping_error;
2937         lp->lp_ping_error = 0;
2938         spin_unlock(&lp->lp_lock);
2939
2940         if (!LNetMDHandleIsInvalid(mdh))
2941                 LNetMDUnlink(mdh);
2942
2943         CDEBUG(D_NET, "peer %s:%d\n",
2944                libcfs_nid2str(lp->lp_primary_nid), rc);
2945
2946         spin_lock(&lp->lp_lock);
2947         return rc ? rc : LNET_REDISCOVER_PEER;
2948 }
2949
2950 /*
2951  * Select NID to send a Ping or Push to.
2952  */
2953 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2954 {
2955         struct lnet_peer_ni *lpni;
2956
2957         /* Look for a direct-connected NID for this peer. */
2958         lpni = NULL;
2959         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2960                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2961                         continue;
2962                 break;
2963         }
2964         if (lpni)
2965                 return lpni->lpni_nid;
2966
2967         /* Look for a routed-connected NID for this peer. */
2968         lpni = NULL;
2969         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2970                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2971                         continue;
2972                 break;
2973         }
2974         if (lpni)
2975                 return lpni->lpni_nid;
2976
2977         return LNET_NID_ANY;
2978 }
2979
2980 /* Active side of ping. */
2981 static int lnet_peer_send_ping(struct lnet_peer *lp)
2982 __must_hold(&lp->lp_lock)
2983 {
2984         lnet_nid_t pnid;
2985         int nnis;
2986         int rc;
2987         int cpt;
2988
2989         lp->lp_state |= LNET_PEER_PING_SENT;
2990         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2991         spin_unlock(&lp->lp_lock);
2992
2993         cpt = lnet_net_lock_current();
2994         /* Refcount for MD. */
2995         lnet_peer_addref_locked(lp);
2996         pnid = lnet_peer_select_nid(lp);
2997         lnet_net_unlock(cpt);
2998
2999         nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3000
3001         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
3002                             the_lnet.ln_dc_handler, false);
3003
3004         /*
3005          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3006          * refcount on the peer, otherwise LNetMDUnlink will be called
3007          * which will eventually do that.
3008          */
3009         if (rc > 0) {
3010                 lnet_net_lock(cpt);
3011                 lnet_peer_decref_locked(lp);
3012                 lnet_net_unlock(cpt);
3013                 rc = -rc; /* change the rc to negative value */
3014                 goto fail_error;
3015         } else if (rc < 0) {
3016                 goto fail_error;
3017         }
3018
3019         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3020
3021         spin_lock(&lp->lp_lock);
3022         return 0;
3023
3024 fail_error:
3025         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3026         /*
3027          * The errors that get us here are considered hard errors and
3028          * cause Discovery to terminate. So we clear PING_SENT, but do
3029          * not set either PING_FAILED or FORCE_PING. In fact we need
3030          * to clear PING_FAILED, because the unlink event handler will
3031          * have set it if we called LNetMDUnlink() above.
3032          */
3033         spin_lock(&lp->lp_lock);
3034         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3035         return rc;
3036 }
3037
3038 /*
3039  * This function exists because you cannot call LNetMDUnlink() from an
3040  * event handler.
3041  */
3042 static int lnet_peer_push_failed(struct lnet_peer *lp)
3043 __must_hold(&lp->lp_lock)
3044 {
3045         struct lnet_handle_md mdh;
3046         int rc;
3047
3048         mdh = lp->lp_push_mdh;
3049         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3050         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3051         rc = lp->lp_push_error;
3052         lp->lp_push_error = 0;
3053         spin_unlock(&lp->lp_lock);
3054
3055         if (!LNetMDHandleIsInvalid(mdh))
3056                 LNetMDUnlink(mdh);
3057
3058         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3059         spin_lock(&lp->lp_lock);
3060         return rc ? rc : LNET_REDISCOVER_PEER;
3061 }
3062
3063 /* Active side of push. */
3064 static int lnet_peer_send_push(struct lnet_peer *lp)
3065 __must_hold(&lp->lp_lock)
3066 {
3067         struct lnet_ping_buffer *pbuf;
3068         struct lnet_process_id id;
3069         struct lnet_md md;
3070         int cpt;
3071         int rc;
3072
3073         /* Don't push to a non-multi-rail peer. */
3074         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3075                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3076                 return 0;
3077         }
3078
3079         lp->lp_state |= LNET_PEER_PUSH_SENT;
3080         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3081         spin_unlock(&lp->lp_lock);
3082
3083         cpt = lnet_net_lock_current();
3084         pbuf = the_lnet.ln_ping_target;
3085         lnet_ping_buffer_addref(pbuf);
3086         lnet_net_unlock(cpt);
3087
3088         /* Push source MD */
3089         md.start     = &pbuf->pb_info;
3090         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3091         md.threshold = 2; /* Put/Ack */
3092         md.max_size  = 0;
3093         md.options   = 0;
3094         md.handler   = the_lnet.ln_dc_handler;
3095         md.user_ptr  = lp;
3096
3097         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3098         if (rc) {
3099                 lnet_ping_buffer_decref(pbuf);
3100                 CERROR("Can't bind push source MD: %d\n", rc);
3101                 goto fail_error;
3102         }
3103         cpt = lnet_net_lock_current();
3104         /* Refcount for MD. */
3105         lnet_peer_addref_locked(lp);
3106         id.pid = LNET_PID_LUSTRE;
3107         id.nid = lnet_peer_select_nid(lp);
3108         lnet_net_unlock(cpt);
3109
3110         if (id.nid == LNET_NID_ANY) {
3111                 rc = -EHOSTUNREACH;
3112                 goto fail_unlink;
3113         }
3114
3115         rc = LNetPut(lp->lp_disc_src_nid, lp->lp_push_mdh,
3116                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3117                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3118
3119         /*
3120          * reset the discovery nid. There is no need to restrict sending
3121          * from that source, if we call lnet_push_update_to_peers(). It'll
3122          * get set to a specific NID, if we initiate discovery from the
3123          * scratch
3124          */
3125         lp->lp_disc_src_nid = LNET_NID_ANY;
3126
3127         if (rc)
3128                 goto fail_unlink;
3129
3130         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3131
3132         spin_lock(&lp->lp_lock);
3133         return 0;
3134
3135 fail_unlink:
3136         LNetMDUnlink(lp->lp_push_mdh);
3137         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3138 fail_error:
3139         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3140         /*
3141          * The errors that get us here are considered hard errors and
3142          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3143          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3144          * because the unlink event handler will have set it if we
3145          * called LNetMDUnlink() above.
3146          */
3147         spin_lock(&lp->lp_lock);
3148         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3149         return rc;
3150 }
3151
3152 /*
3153  * An unrecoverable error was encountered during discovery.
3154  * Set error status in peer and abort discovery.
3155  */
3156 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3157 {
3158         CDEBUG(D_NET, "Discovery error %s: %d\n",
3159                libcfs_nid2str(lp->lp_primary_nid), error);
3160
3161         spin_lock(&lp->lp_lock);
3162         lp->lp_dc_error = error;
3163         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3164         lp->lp_state |= LNET_PEER_REDISCOVER;
3165         spin_unlock(&lp->lp_lock);
3166 }
3167
3168 /*
3169  * Mark the peer as discovered.
3170  */
3171 static int lnet_peer_discovered(struct lnet_peer *lp)
3172 __must_hold(&lp->lp_lock)
3173 {
3174         lp->lp_state |= LNET_PEER_DISCOVERED;
3175         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3176                           LNET_PEER_REDISCOVER);
3177
3178         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3179
3180         return 0;
3181 }
3182
3183
3184 /*
3185  * Discovering this peer is taking too long. Cancel any Ping or Push
3186  * that discovery is waiting on by unlinking the relevant MDs. The
3187  * lnet_discovery_event_handler() will proceed from here and complete
3188  * the cleanup.
3189  */
3190 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3191 {
3192         struct lnet_handle_md ping_mdh;
3193         struct lnet_handle_md push_mdh;
3194
3195         LNetInvalidateMDHandle(&ping_mdh);
3196         LNetInvalidateMDHandle(&push_mdh);
3197
3198         spin_lock(&lp->lp_lock);
3199         if (lp->lp_state & LNET_PEER_PING_SENT) {
3200                 ping_mdh = lp->lp_ping_mdh;
3201                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3202         }
3203         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3204                 push_mdh = lp->lp_push_mdh;
3205                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3206         }
3207         spin_unlock(&lp->lp_lock);
3208
3209         if (!LNetMDHandleIsInvalid(ping_mdh))
3210                 LNetMDUnlink(ping_mdh);
3211         if (!LNetMDHandleIsInvalid(push_mdh))
3212                 LNetMDUnlink(push_mdh);
3213 }
3214
3215 /*
3216  * Wait for work to be queued or some other change that must be
3217  * attended to. Returns non-zero if the discovery thread should shut
3218  * down.
3219  */
3220 static int lnet_peer_discovery_wait_for_work(void)
3221 {
3222         int cpt;
3223         int rc = 0;
3224
3225         DEFINE_WAIT(wait);
3226
3227         cpt = lnet_net_lock_current();
3228         for (;;) {
3229                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3230                                 TASK_INTERRUPTIBLE);
3231                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3232                         break;
3233                 if (lnet_push_target_resize_needed() ||
3234                     the_lnet.ln_push_target->pb_needs_post)
3235                         break;
3236                 if (!list_empty(&the_lnet.ln_dc_request))
3237                         break;
3238                 if (!list_empty(&the_lnet.ln_msg_resend))
3239                         break;
3240                 lnet_net_unlock(cpt);
3241
3242                 /*
3243                  * wakeup max every second to check if there are peers that
3244                  * have been stuck on the working queue for greater than
3245                  * the peer timeout.
3246                  */
3247                 schedule_timeout(cfs_time_seconds(1));
3248                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3249                 cpt = lnet_net_lock_current();
3250         }
3251         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3252
3253         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3254                 rc = -ESHUTDOWN;
3255
3256         lnet_net_unlock(cpt);
3257
3258         CDEBUG(D_NET, "woken: %d\n", rc);
3259
3260         return rc;
3261 }
3262
3263 /*
3264  * Messages that were pending on a destroyed peer will be put on a global
3265  * resend list. The message resend list will be checked by
3266  * the discovery thread when it wakes up, and will resend messages. These
3267  * messages can still be sendable in the case the lpni which was the initial
3268  * cause of the message re-queue was transfered to another peer.
3269  *
3270  * It is possible that LNet could be shutdown while we're iterating
3271  * through the list. lnet_shudown_lndnets() will attempt to access the
3272  * resend list, but will have to wait until the spinlock is released, by
3273  * which time there shouldn't be any more messages on the resend list.
3274  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3275  * for the messages so they can be released. The other case is that
3276  * lnet_shudown_lndnets() can finalize all the messages before this
3277  * function can visit the resend list, in which case this function will be
3278  * a no-op.
3279  */
3280 static void lnet_resend_msgs(void)
3281 {
3282         struct lnet_msg *msg, *tmp;
3283         LIST_HEAD(resend);
3284         int rc;
3285
3286         spin_lock(&the_lnet.ln_msg_resend_lock);
3287         list_splice(&the_lnet.ln_msg_resend, &resend);
3288         spin_unlock(&the_lnet.ln_msg_resend_lock);
3289
3290         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3291                 list_del_init(&msg->msg_list);
3292                 rc = lnet_send(msg->msg_src_nid_param, msg,
3293                                msg->msg_rtr_nid_param);
3294                 if (rc < 0) {
3295                         CNETERR("Error sending %s to %s: %d\n",
3296                                lnet_msgtyp2str(msg->msg_type),
3297                                libcfs_id2str(msg->msg_target), rc);
3298                         lnet_finalize(msg, rc);
3299                 }
3300         }
3301 }
3302
3303 /* The discovery thread. */
3304 static int lnet_peer_discovery(void *arg)
3305 {
3306         struct lnet_peer *lp;
3307         int rc;
3308
3309         wait_for_completion(&the_lnet.ln_started);
3310
3311         CDEBUG(D_NET, "started\n");
3312
3313         for (;;) {
3314                 if (lnet_peer_discovery_wait_for_work())
3315                         break;
3316
3317                 if (lnet_push_target_resize_needed())
3318                         lnet_push_target_resize();
3319                 else if (the_lnet.ln_push_target->pb_needs_post)
3320                         lnet_push_target_post(the_lnet.ln_push_target,
3321                                               &the_lnet.ln_push_target_md);
3322
3323                 lnet_resend_msgs();
3324
3325                 lnet_net_lock(LNET_LOCK_EX);
3326                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3327                         lnet_net_unlock(LNET_LOCK_EX);
3328                         break;
3329                 }
3330
3331                 /*
3332                  * Process all incoming discovery work requests.  When
3333                  * discovery must wait on a peer to change state, it
3334                  * is added to the tail of the ln_dc_working queue. A
3335                  * timestamp keeps track of when the peer was added,
3336                  * so we can time out discovery requests that take too
3337                  * long.
3338                  */
3339                 while (!list_empty(&the_lnet.ln_dc_request)) {
3340                         lp = list_first_entry(&the_lnet.ln_dc_request,
3341                                               struct lnet_peer, lp_dc_list);
3342                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3343                         /*
3344                          * set the time the peer was put on the dc_working
3345                          * queue. It shouldn't remain on the queue
3346                          * forever, in case the GET message (for ping)
3347                          * doesn't get a REPLY or the PUT message (for
3348                          * push) doesn't get an ACK.
3349                          */
3350                         lp->lp_last_queued = ktime_get_real_seconds();
3351                         lnet_net_unlock(LNET_LOCK_EX);
3352
3353                         if (lnet_push_target_resize_needed())
3354                                 lnet_push_target_resize();
3355                         else if (the_lnet.ln_push_target->pb_needs_post)
3356                                 lnet_push_target_post(the_lnet.ln_push_target,
3357                                                       &the_lnet.ln_push_target_md);
3358
3359                         /*
3360                          * Select an action depending on the state of
3361                          * the peer and whether discovery is disabled.
3362                          * The check whether discovery is disabled is
3363                          * done after the code that handles processing
3364                          * for arrived data, cleanup for failures, and
3365                          * forcing a Ping or Push.
3366                          */
3367                         spin_lock(&lp->lp_lock);
3368                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3369                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3370                                 lp->lp_state);
3371                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3372                                 rc = lnet_peer_data_present(lp);
3373                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3374                                 rc = lnet_peer_ping_failed(lp);
3375                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3376                                 rc = lnet_peer_push_failed(lp);
3377                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3378                                 rc = lnet_peer_send_ping(lp);
3379                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3380                                 rc = lnet_peer_send_push(lp);
3381                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3382                                 rc = lnet_peer_send_ping(lp);
3383                         else if (lnet_peer_needs_push(lp))
3384                                 rc = lnet_peer_send_push(lp);
3385                         else
3386                                 rc = lnet_peer_discovered(lp);
3387                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3388                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3389                                 lp->lp_state, rc);
3390                         spin_unlock(&lp->lp_lock);
3391
3392                         lnet_net_lock(LNET_LOCK_EX);
3393                         if (rc == LNET_REDISCOVER_PEER) {
3394                                 list_move(&lp->lp_dc_list,
3395                                           &the_lnet.ln_dc_request);
3396                         } else if (rc) {
3397                                 lnet_peer_discovery_error(lp, rc);
3398                         }
3399                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3400                                 lnet_peer_discovery_complete(lp);
3401                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3402                                 break;
3403
3404                         if (lp->lp_state & LNET_PEER_MARK_DELETION) {
3405                                 struct list_head rlist;
3406                                 struct lnet_route *route, *tmp;
3407                                 int sensitivity = lp->lp_health_sensitivity;
3408
3409                                 INIT_LIST_HEAD(&rlist);
3410
3411                                 /*
3412                                  * remove the peer from the discovery work
3413                                  * queue if it's on there in preparation
3414                                  * of deleting it.
3415                                  */
3416                                 if (!list_empty(&lp->lp_dc_list))
3417                                         list_del(&lp->lp_dc_list);
3418
3419                                 lnet_net_unlock(LNET_LOCK_EX);
3420
3421                                 mutex_lock(&the_lnet.ln_api_mutex);
3422
3423                                 lnet_net_lock(LNET_LOCK_EX);
3424                                 list_for_each_entry_safe(route, tmp,
3425                                                          &lp->lp_routes,
3426                                                          lr_gwlist)
3427                                         lnet_move_route(route, NULL, &rlist);
3428                                 lnet_net_unlock(LNET_LOCK_EX);
3429
3430                                 /* delete the peer */
3431                                 lnet_peer_del(lp);
3432
3433                                 list_for_each_entry_safe(route, tmp,
3434                                                          &rlist, lr_list) {
3435                                         /* re-add these routes */
3436                                         lnet_add_route(route->lr_net,
3437                                                        route->lr_hops,
3438                                                        route->lr_nid,
3439                                                        route->lr_priority,
3440                                                        sensitivity);
3441                                         LIBCFS_FREE(route, sizeof(*route));
3442                                 }
3443                                 mutex_unlock(&the_lnet.ln_api_mutex);
3444
3445                                 lnet_net_lock(LNET_LOCK_EX);
3446                         }
3447                 }
3448
3449                 lnet_net_unlock(LNET_LOCK_EX);
3450         }
3451
3452         CDEBUG(D_NET, "stopping\n");
3453         /*
3454          * Clean up before telling lnet_peer_discovery_stop() that
3455          * we're done. Use wake_up() below to somewhat reduce the
3456          * size of the thundering herd if there are multiple threads
3457          * waiting on discovery of a single peer.
3458          */
3459
3460         /* Queue cleanup 1: stop all pending pings and pushes. */
3461         lnet_net_lock(LNET_LOCK_EX);
3462         while (!list_empty(&the_lnet.ln_dc_working)) {
3463                 lp = list_first_entry(&the_lnet.ln_dc_working,
3464                                       struct lnet_peer, lp_dc_list);
3465                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3466                 lnet_net_unlock(LNET_LOCK_EX);
3467                 lnet_peer_cancel_discovery(lp);
3468                 lnet_net_lock(LNET_LOCK_EX);
3469         }
3470         lnet_net_unlock(LNET_LOCK_EX);
3471
3472         /* Queue cleanup 2: wait for the expired queue to clear. */
3473         while (!list_empty(&the_lnet.ln_dc_expired))
3474                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3475
3476         /* Queue cleanup 3: clear the request queue. */
3477         lnet_net_lock(LNET_LOCK_EX);
3478         while (!list_empty(&the_lnet.ln_dc_request)) {
3479                 lp = list_first_entry(&the_lnet.ln_dc_request,
3480                                       struct lnet_peer, lp_dc_list);
3481                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3482                 lnet_peer_discovery_complete(lp);
3483         }
3484         lnet_net_unlock(LNET_LOCK_EX);
3485
3486         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3487         the_lnet.ln_dc_handler = NULL;
3488
3489         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3490         wake_up(&the_lnet.ln_dc_waitq);
3491
3492         CDEBUG(D_NET, "stopped\n");
3493
3494         return 0;
3495 }
3496
3497 /* ln_api_mutex is held on entry. */
3498 int lnet_peer_discovery_start(void)
3499 {
3500         struct task_struct *task;
3501         int rc = 0;
3502
3503         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3504                 return -EALREADY;
3505
3506         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3507         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3508         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3509         if (IS_ERR(task)) {
3510                 rc = PTR_ERR(task);
3511                 CERROR("Can't start peer discovery thread: %d\n", rc);
3512
3513                 the_lnet.ln_dc_handler = NULL;
3514
3515                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3516         }
3517
3518         CDEBUG(D_NET, "discovery start: %d\n", rc);
3519
3520         return rc;
3521 }
3522
3523 /* ln_api_mutex is held on entry. */
3524 void lnet_peer_discovery_stop(void)
3525 {
3526         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3527                 return;
3528
3529         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3530         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3531
3532         /* In the LNetNIInit() path we may be stopping discovery before it
3533          * entered its work loop
3534          */
3535         if (!completion_done(&the_lnet.ln_started))
3536                 complete(&the_lnet.ln_started);
3537         else
3538                 wake_up(&the_lnet.ln_dc_waitq);
3539
3540         wait_event(the_lnet.ln_dc_waitq,
3541                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3542
3543         LASSERT(list_empty(&the_lnet.ln_dc_request));
3544         LASSERT(list_empty(&the_lnet.ln_dc_working));
3545         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3546
3547         CDEBUG(D_NET, "discovery stopped\n");
3548 }
3549
3550 /* Debugging */
3551
3552 void
3553 lnet_debug_peer(lnet_nid_t nid)
3554 {
3555         char                    *aliveness = "NA";
3556         struct lnet_peer_ni     *lp;
3557         int                     cpt;
3558
3559         cpt = lnet_cpt_of_nid(nid, NULL);
3560         lnet_net_lock(cpt);
3561
3562         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3563         if (IS_ERR(lp)) {
3564                 lnet_net_unlock(cpt);
3565                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3566                 return;
3567         }
3568
3569         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3570                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3571
3572         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3573                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3574                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3575                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3576                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3577
3578         lnet_peer_ni_decref_locked(lp);
3579
3580         lnet_net_unlock(cpt);
3581 }
3582
3583 /* Gathering information for userspace. */
3584
3585 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3586                           char aliveness[LNET_MAX_STR_LEN],
3587                           __u32 *cpt_iter, __u32 *refcount,
3588                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3589                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3590                           __u32 *peer_tx_qnob)
3591 {
3592         struct lnet_peer_table          *peer_table;
3593         struct lnet_peer_ni             *lp;
3594         int                             j;
3595         int                             lncpt;
3596         bool                            found = false;
3597
3598         /* get the number of CPTs */
3599         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3600
3601         /* if the cpt number to be examined is >= the number of cpts in
3602          * the system then indicate that there are no more cpts to examin
3603          */
3604         if (*cpt_iter >= lncpt)
3605                 return -ENOENT;
3606
3607         /* get the current table */
3608         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3609         /* if the ptable is NULL then there are no more cpts to examine */
3610         if (peer_table == NULL)
3611                 return -ENOENT;
3612
3613         lnet_net_lock(*cpt_iter);
3614
3615         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3616                 struct list_head *peers = &peer_table->pt_hash[j];
3617
3618                 list_for_each_entry(lp, peers, lpni_hashlist) {
3619                         if (peer_index-- > 0)
3620                                 continue;
3621
3622                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3623                         if (lnet_isrouter(lp) ||
3624                                 lnet_peer_aliveness_enabled(lp))
3625                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3626                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3627
3628                         *nid = lp->lpni_nid;
3629                         *refcount = atomic_read(&lp->lpni_refcount);
3630                         *ni_peer_tx_credits =
3631                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3632                         *peer_tx_credits = lp->lpni_txcredits;
3633                         *peer_rtr_credits = lp->lpni_rtrcredits;
3634                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3635                         *peer_tx_qnob = lp->lpni_txqnob;
3636
3637                         found = true;
3638                 }
3639
3640         }
3641         lnet_net_unlock(*cpt_iter);
3642
3643         *cpt_iter = lncpt;
3644
3645         return found ? 0 : -ENOENT;
3646 }
3647
3648 /* ln_api_mutex is held, which keeps the peer list stable */
3649 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3650 {
3651         struct lnet_ioctl_element_stats *lpni_stats;
3652         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3653         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3654         struct lnet_peer_ni_credit_info *lpni_info;
3655         struct lnet_peer_ni *lpni;
3656         struct lnet_peer *lp;
3657         lnet_nid_t nid;
3658         __u32 size;
3659         int rc;
3660
3661         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3662
3663         if (!lp) {
3664                 rc = -ENOENT;
3665                 goto out;
3666         }
3667
3668         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3669                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3670         size *= lp->lp_nnis;
3671         if (size > cfg->prcfg_size) {
3672                 cfg->prcfg_size = size;
3673                 rc = -E2BIG;
3674                 goto out_lp_decref;
3675         }
3676
3677         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3678         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3679         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3680         cfg->prcfg_count = lp->lp_nnis;
3681         cfg->prcfg_size = size;
3682         cfg->prcfg_state = lp->lp_state;
3683
3684         /* Allocate helper buffers. */
3685         rc = -ENOMEM;
3686         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3687         if (!lpni_info)
3688                 goto out_lp_decref;
3689         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3690         if (!lpni_stats)
3691                 goto out_free_info;
3692         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3693         if (!lpni_msg_stats)
3694                 goto out_free_stats;
3695         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3696         if (!lpni_hstats)
3697                 goto out_free_msg_stats;
3698
3699
3700         lpni = NULL;
3701         rc = -EFAULT;
3702         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3703                 nid = lpni->lpni_nid;
3704                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3705                         goto out_free_hstats;
3706                 bulk += sizeof(nid);
3707
3708                 memset(lpni_info, 0, sizeof(*lpni_info));
3709                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3710                 if (lnet_isrouter(lpni) ||
3711                         lnet_peer_aliveness_enabled(lpni))
3712                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3713                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3714
3715                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3716                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3717                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3718                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3719                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3720                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3721                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3722                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3723                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3724                         goto out_free_hstats;
3725                 bulk += sizeof(*lpni_info);
3726
3727                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3728                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3729                                                             LNET_STATS_TYPE_SEND);
3730                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3731                                                             LNET_STATS_TYPE_RECV);
3732                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3733                                                             LNET_STATS_TYPE_DROP);
3734                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3735                         goto out_free_hstats;
3736                 bulk += sizeof(*lpni_stats);
3737                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3738                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3739                         goto out_free_hstats;
3740                 bulk += sizeof(*lpni_msg_stats);
3741                 lpni_hstats->hlpni_network_timeout =
3742                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3743                 lpni_hstats->hlpni_remote_dropped =
3744                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3745                 lpni_hstats->hlpni_remote_timeout =
3746                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3747                 lpni_hstats->hlpni_remote_error =
3748                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3749                 lpni_hstats->hlpni_health_value =
3750                   atomic_read(&lpni->lpni_healthv);
3751                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3752                         goto out_free_hstats;
3753                 bulk += sizeof(*lpni_hstats);
3754         }
3755         rc = 0;
3756
3757 out_free_hstats:
3758         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3759 out_free_msg_stats:
3760         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3761 out_free_stats:
3762         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3763 out_free_info:
3764         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3765 out_lp_decref:
3766         lnet_peer_decref_locked(lp);
3767 out:
3768         return rc;
3769 }
3770
3771 void
3772 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3773 {
3774         /* the mt could've shutdown and cleaned up the queues */
3775         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3776                 return;
3777
3778         if (list_empty(&lpni->lpni_recovery) &&
3779             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3780                 CDEBUG(D_NET, "lpni %s added to recovery queue. Health = %d\n",
3781                         libcfs_nid2str(lpni->lpni_nid),
3782                         atomic_read(&lpni->lpni_healthv));
3783                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3784                 lnet_peer_ni_addref_locked(lpni);
3785         }
3786 }
3787
3788 /* Call with the ln_api_mutex held */
3789 void
3790 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3791 {
3792         struct lnet_peer_table *ptable;
3793         struct lnet_peer *lp;
3794         struct lnet_peer_net *lpn;
3795         struct lnet_peer_ni *lpni;
3796         int lncpt;
3797         int cpt;
3798
3799         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3800                 return;
3801
3802         if (!all) {
3803                 lnet_net_lock(LNET_LOCK_EX);
3804                 lpni = lnet_find_peer_ni_locked(nid);
3805                 if (!lpni) {
3806                         lnet_net_unlock(LNET_LOCK_EX);
3807                         return;
3808                 }
3809                 atomic_set(&lpni->lpni_healthv, value);
3810                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3811                 lnet_peer_ni_decref_locked(lpni);
3812                 lnet_net_unlock(LNET_LOCK_EX);
3813                 return;
3814         }
3815
3816         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3817
3818         /*
3819          * Walk all the peers and reset the healhv for each one to the
3820          * maximum value.
3821          */
3822         lnet_net_lock(LNET_LOCK_EX);
3823         for (cpt = 0; cpt < lncpt; cpt++) {
3824                 ptable = the_lnet.ln_peer_tables[cpt];
3825                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3826                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3827                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3828                                                     lpni_peer_nis) {
3829                                         atomic_set(&lpni->lpni_healthv, value);
3830                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3831                                 }
3832                         }
3833                 }
3834         }
3835         lnet_net_unlock(LNET_LOCK_EX);
3836 }
3837