lnet/lnet/peer.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/peer.c
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_LNET
  36
  37 #include <linux/sched.h>
  38 #ifdef HAVE_SCHED_HEADERS
  39 #include <linux/sched/signal.h>
  40 #endif
  41 #include <linux/uaccess.h>
  42
  43 #include <lnet/lib-lnet.h>
  44 #include <uapi/linux/lnet/lnet-dlc.h>
  45
  46 /* Value indicating that recovery needs to re-check a peer immediately. */
  47 #define LNET_REDISCOVER_PEER    (1)
  48
  49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
  50
  51 static void
  52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
  53 {
  54         if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
  55                 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
  56                 lnet_peer_ni_decref_locked(lpni);
  57         }
  58 }
  59
  60 void
  61 lnet_peer_net_added(struct lnet_net *net)
  62 {
  63         struct lnet_peer_ni *lpni, *tmp;
  64
  65         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
  66                                  lpni_on_remote_peer_ni_list) {
  67
  68                 if (LNET_NIDNET(lpni->lpni_nid) == net->net_id) {
  69                         lpni->lpni_net = net;
  70
  71                         spin_lock(&lpni->lpni_lock);
  72                         lpni->lpni_txcredits =
  73                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
  74                         lpni->lpni_mintxcredits = lpni->lpni_txcredits;
  75                         lpni->lpni_rtrcredits =
  76                                 lnet_peer_buffer_credits(lpni->lpni_net);
  77                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
  78                         spin_unlock(&lpni->lpni_lock);
  79
  80                         lnet_peer_remove_from_remote_list(lpni);
  81                 }
  82         }
  83 }
  84
  85 static void
  86 lnet_peer_tables_destroy(void)
  87 {
  88         struct lnet_peer_table  *ptable;
  89         struct list_head        *hash;
  90         int                     i;
  91         int                     j;
  92
  93         if (!the_lnet.ln_peer_tables)
  94                 return;
  95
  96         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  97                 hash = ptable->pt_hash;
  98                 if (!hash) /* not intialized */
  99                         break;
 100
 101                 LASSERT(list_empty(&ptable->pt_zombie_list));
 102
 103                 ptable->pt_hash = NULL;
 104                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 105                         LASSERT(list_empty(&hash[j]));
 106
 107                 LIBCFS_FREE(hash, LNET_PEER_HASH_SIZE * sizeof(*hash));
 108         }
 109
 110         cfs_percpt_free(the_lnet.ln_peer_tables);
 111         the_lnet.ln_peer_tables = NULL;
 112 }
 113
 114 int
 115 lnet_peer_tables_create(void)
 116 {
 117         struct lnet_peer_table  *ptable;
 118         struct list_head        *hash;
 119         int                     i;
 120         int                     j;
 121
 122         the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
 123                                                    sizeof(*ptable));
 124         if (the_lnet.ln_peer_tables == NULL) {
 125                 CERROR("Failed to allocate cpu-partition peer tables\n");
 126                 return -ENOMEM;
 127         }
 128
 129         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 130                 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
 131                                  LNET_PEER_HASH_SIZE * sizeof(*hash));
 132                 if (hash == NULL) {
 133                         CERROR("Failed to create peer hash table\n");
 134                         lnet_peer_tables_destroy();
 135                         return -ENOMEM;
 136                 }
 137
 138                 spin_lock_init(&ptable->pt_zombie_lock);
 139                 INIT_LIST_HEAD(&ptable->pt_zombie_list);
 140
 141                 INIT_LIST_HEAD(&ptable->pt_peer_list);
 142
 143                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 144                         INIT_LIST_HEAD(&hash[j]);
 145                 ptable->pt_hash = hash; /* sign of initialization */
 146         }
 147
 148         return 0;
 149 }
 150
 151 static struct lnet_peer_ni *
 152 lnet_peer_ni_alloc(lnet_nid_t nid)
 153 {
 154         struct lnet_peer_ni *lpni;
 155         struct lnet_net *net;
 156         int cpt;
 157
 158         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 159
 160         LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
 161         if (!lpni)
 162                 return NULL;
 163
 164         INIT_LIST_HEAD(&lpni->lpni_txq);
 165         INIT_LIST_HEAD(&lpni->lpni_hashlist);
 166         INIT_LIST_HEAD(&lpni->lpni_peer_nis);
 167         INIT_LIST_HEAD(&lpni->lpni_recovery);
 168         INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
 169         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
 170
 171         spin_lock_init(&lpni->lpni_lock);
 172
 173         if (lnet_peers_start_down())
 174                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 175         else
 176                 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
 177         lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 178         lpni->lpni_nid = nid;
 179         lpni->lpni_cpt = cpt;
 180         atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
 181
 182         net = lnet_get_net_locked(LNET_NIDNET(nid));
 183         lpni->lpni_net = net;
 184         if (net) {
 185                 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
 186                 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
 187                 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
 188                 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 189         } else {
 190                 /*
 191                  * This peer_ni is not on a local network, so we
 192                  * cannot add the credits here. In case the net is
 193                  * added later, add the peer_ni to the remote peer ni
 194                  * list so it can be easily found and revisited.
 195                  */
 196                 /* FIXME: per-net implementation instead? */
 197                 atomic_inc(&lpni->lpni_refcount);
 198                 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
 199                               &the_lnet.ln_remote_peer_ni_list);
 200         }
 201
 202         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
 203
 204         return lpni;
 205 }
 206
 207 static struct lnet_peer_net *
 208 lnet_peer_net_alloc(__u32 net_id)
 209 {
 210         struct lnet_peer_net *lpn;
 211
 212         LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
 213         if (!lpn)
 214                 return NULL;
 215
 216         INIT_LIST_HEAD(&lpn->lpn_peer_nets);
 217         INIT_LIST_HEAD(&lpn->lpn_peer_nis);
 218         lpn->lpn_net_id = net_id;
 219
 220         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 221
 222         return lpn;
 223 }
 224
 225 void
 226 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
 227 {
 228         struct lnet_peer *lp;
 229
 230         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 231
 232         LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
 233         LASSERT(list_empty(&lpn->lpn_peer_nis));
 234         LASSERT(list_empty(&lpn->lpn_peer_nets));
 235         lp = lpn->lpn_peer;
 236         lpn->lpn_peer = NULL;
 237         LIBCFS_FREE(lpn, sizeof(*lpn));
 238
 239         lnet_peer_decref_locked(lp);
 240 }
 241
 242 static struct lnet_peer *
 243 lnet_peer_alloc(lnet_nid_t nid)
 244 {
 245         struct lnet_peer *lp;
 246
 247         LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
 248         if (!lp)
 249                 return NULL;
 250
 251         INIT_LIST_HEAD(&lp->lp_rtrq);
 252         INIT_LIST_HEAD(&lp->lp_routes);
 253         INIT_LIST_HEAD(&lp->lp_peer_list);
 254         INIT_LIST_HEAD(&lp->lp_peer_nets);
 255         INIT_LIST_HEAD(&lp->lp_dc_list);
 256         INIT_LIST_HEAD(&lp->lp_dc_pendq);
 257         INIT_LIST_HEAD(&lp->lp_rtr_list);
 258         init_waitqueue_head(&lp->lp_dc_waitq);
 259         spin_lock_init(&lp->lp_lock);
 260         lp->lp_primary_nid = nid;
 261         if (lnet_peers_start_down())
 262                 lp->lp_alive = false;
 263         else
 264                 lp->lp_alive = true;
 265
 266         /*
 267          * all peers created on a router should have health on
 268          * if it's not already on.
 269          */
 270         if (the_lnet.ln_routing && !lnet_health_sensitivity)
 271                 lp->lp_health_sensitivity = 1;
 272
 273         /*
 274          * Turn off discovery for loopback peer. If you're creating a peer
 275          * for the loopback interface then that was initiated when we
 276          * attempted to send a message over the loopback. There is no need
 277          * to ever use a different interface when sending messages to
 278          * myself.
 279          */
 280         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 281                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 282         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 283
 284         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 285
 286         return lp;
 287 }
 288
 289 void
 290 lnet_destroy_peer_locked(struct lnet_peer *lp)
 291 {
 292         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nid2str(lp->lp_primary_nid));
 293
 294         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 295         LASSERT(lp->lp_rtr_refcount == 0);
 296         LASSERT(list_empty(&lp->lp_peer_nets));
 297         LASSERT(list_empty(&lp->lp_peer_list));
 298         LASSERT(list_empty(&lp->lp_dc_list));
 299
 300         if (lp->lp_data)
 301                 lnet_ping_buffer_decref(lp->lp_data);
 302
 303         /*
 304          * if there are messages still on the pending queue, then make
 305          * sure to queue them on the ln_msg_resend list so they can be
 306          * resent at a later point if the discovery thread is still
 307          * running.
 308          * If the discovery thread has stopped, then the wakeup will be a
 309          * no-op, and it is expected the lnet_shutdown_lndnets() will
 310          * eventually be called, which will traverse this list and
 311          * finalize the messages on the list.
 312          * We can not resend them now because we're holding the cpt lock.
 313          * Releasing the lock can cause an inconsistent state
 314          */
 315         spin_lock(&the_lnet.ln_msg_resend_lock);
 316         spin_lock(&lp->lp_lock);
 317         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 318         spin_unlock(&lp->lp_lock);
 319         spin_unlock(&the_lnet.ln_msg_resend_lock);
 320         wake_up(&the_lnet.ln_dc_waitq);
 321
 322         LIBCFS_FREE(lp, sizeof(*lp));
 323 }
 324
 325 /*
 326  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 327  * that peer_net, detach the peer_net from the peer.
 328  *
 329  * Call with lnet_net_lock/EX held
 330  */
 331 static void
 332 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 333 {
 334         struct lnet_peer_table *ptable;
 335         struct lnet_peer_net *lpn;
 336         struct lnet_peer *lp;
 337
 338         /*
 339          * Belts and suspenders: gracefully handle teardown of a
 340          * partially connected peer_ni.
 341          */
 342         lpn = lpni->lpni_peer_net;
 343
 344         list_del_init(&lpni->lpni_peer_nis);
 345         /*
 346          * If there are no lpni's left, we detach lpn from
 347          * lp_peer_nets, so it cannot be found anymore.
 348          */
 349         if (list_empty(&lpn->lpn_peer_nis))
 350                 list_del_init(&lpn->lpn_peer_nets);
 351
 352         /* Update peer NID count. */
 353         lp = lpn->lpn_peer;
 354         lp->lp_nnis--;
 355
 356         /*
 357          * If there are no more peer nets, make the peer unfindable
 358          * via the peer_tables.
 359          *
 360          * Otherwise, if the peer is DISCOVERED, tell discovery to
 361          * take another look at it. This is a no-op if discovery for
 362          * this peer did the detaching.
 363          */
 364         if (list_empty(&lp->lp_peer_nets)) {
 365                 list_del_init(&lp->lp_peer_list);
 366                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 367                 ptable->pt_peers--;
 368         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 369                 /* Discovery isn't running, nothing to do here. */
 370         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 371                 lnet_peer_queue_for_discovery(lp);
 372                 wake_up(&the_lnet.ln_dc_waitq);
 373         }
 374         CDEBUG(D_NET, "peer %s NID %s\n",
 375                 libcfs_nid2str(lp->lp_primary_nid),
 376                 libcfs_nid2str(lpni->lpni_nid));
 377 }
 378
 379 /* called with lnet_net_lock LNET_LOCK_EX held */
 380 static int
 381 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 382 {
 383         struct lnet_peer_table *ptable = NULL;
 384
 385         /* don't remove a peer_ni if it's also a gateway */
 386         if (lnet_isrouter(lpni) && !force) {
 387                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 388                        libcfs_nid2str(lpni->lpni_nid));
 389                 return -EBUSY;
 390         }
 391
 392         lnet_peer_remove_from_remote_list(lpni);
 393
 394         /* remove peer ni from the hash list. */
 395         list_del_init(&lpni->lpni_hashlist);
 396
 397         /*
 398          * indicate the peer is being deleted so the monitor thread can
 399          * remove it from the recovery queue.
 400          */
 401         spin_lock(&lpni->lpni_lock);
 402         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 403         spin_unlock(&lpni->lpni_lock);
 404
 405         /* decrement the ref count on the peer table */
 406         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 407
 408         /*
 409          * The peer_ni can no longer be found with a lookup. But there
 410          * can be current users, so keep track of it on the zombie
 411          * list until the reference count has gone to zero.
 412          *
 413          * The last reference may be lost in a place where the
 414          * lnet_net_lock locks only a single cpt, and that cpt may not
 415          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 416          * has its own lock.
 417          */
 418         spin_lock(&ptable->pt_zombie_lock);
 419         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 420         ptable->pt_zombies++;
 421         spin_unlock(&ptable->pt_zombie_lock);
 422
 423         /* no need to keep this peer_ni on the hierarchy anymore */
 424         lnet_peer_detach_peer_ni_locked(lpni);
 425
 426         /* remove hashlist reference on peer_ni */
 427         lnet_peer_ni_decref_locked(lpni);
 428
 429         return 0;
 430 }
 431
 432 void lnet_peer_uninit(void)
 433 {
 434         struct lnet_peer_ni *lpni, *tmp;
 435
 436         lnet_net_lock(LNET_LOCK_EX);
 437
 438         /* remove all peer_nis from the remote peer and the hash list */
 439         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 440                                  lpni_on_remote_peer_ni_list)
 441                 lnet_peer_ni_del_locked(lpni, false);
 442
 443         lnet_peer_tables_destroy();
 444
 445         lnet_net_unlock(LNET_LOCK_EX);
 446 }
 447
 448 static int
 449 lnet_peer_del_locked(struct lnet_peer *peer)
 450 {
 451         struct lnet_peer_ni *lpni = NULL, *lpni2;
 452         int rc = 0, rc2 = 0;
 453
 454         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(peer->lp_primary_nid));
 455
 456         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 457         while (lpni != NULL) {
 458                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 459                 rc = lnet_peer_ni_del_locked(lpni, false);
 460                 if (rc != 0)
 461                         rc2 = rc;
 462                 lpni = lpni2;
 463         }
 464
 465         return rc2;
 466 }
 467
 468 static int
 469 lnet_peer_del(struct lnet_peer *peer)
 470 {
 471         lnet_net_lock(LNET_LOCK_EX);
 472         lnet_peer_del_locked(peer);
 473         lnet_net_unlock(LNET_LOCK_EX);
 474
 475         return 0;
 476 }
 477
 478 /*
 479  * Delete a NID from a peer. Call with ln_api_mutex held.
 480  *
 481  * Error codes:
 482  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 483  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 484  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 485  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 486  */
 487 static int
 488 lnet_peer_del_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
 489 {
 490         struct lnet_peer_ni *lpni;
 491         lnet_nid_t primary_nid = lp->lp_primary_nid;
 492         int rc = 0;
 493         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 494
 495         if (!(flags & LNET_PEER_CONFIGURED)) {
 496                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 497                         rc = -EPERM;
 498                         goto out;
 499                 }
 500         }
 501         lpni = lnet_find_peer_ni_locked(nid);
 502         if (!lpni) {
 503                 rc = -ENOENT;
 504                 goto out;
 505         }
 506         lnet_peer_ni_decref_locked(lpni);
 507         if (lp != lpni->lpni_peer_net->lpn_peer) {
 508                 rc = -ECHILD;
 509                 goto out;
 510         }
 511
 512         /*
 513          * This function only allows deletion of the primary NID if it
 514          * is the only NID.
 515          */
 516         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && !force) {
 517                 rc = -EBUSY;
 518                 goto out;
 519         }
 520
 521         lnet_net_lock(LNET_LOCK_EX);
 522
 523         if (nid == lp->lp_primary_nid && lp->lp_nnis != 1 && force) {
 524                 struct lnet_peer_ni *lpni2;
 525                 /* assign the next peer_ni to be the primary */
 526                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 527                 LASSERT(lpni2);
 528                 lp->lp_primary_nid = lpni->lpni_nid;
 529         }
 530         rc = lnet_peer_ni_del_locked(lpni, force);
 531
 532         lnet_net_unlock(LNET_LOCK_EX);
 533
 534 out:
 535         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 536                libcfs_nid2str(primary_nid), libcfs_nid2str(nid), flags, rc);
 537
 538         return rc;
 539 }
 540
 541 static void
 542 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 543                                struct lnet_peer_table *ptable)
 544 {
 545         int                      i;
 546         struct lnet_peer_ni     *next;
 547         struct lnet_peer_ni     *lpni;
 548         struct lnet_peer        *peer;
 549
 550         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 551                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 552                                          lpni_hashlist) {
 553                         if (net != NULL && net != lpni->lpni_net)
 554                                 continue;
 555
 556                         peer = lpni->lpni_peer_net->lpn_peer;
 557                         if (peer->lp_primary_nid != lpni->lpni_nid) {
 558                                 lnet_peer_ni_del_locked(lpni, false);
 559                                 continue;
 560                         }
 561                         /*
 562                          * Removing the primary NID implies removing
 563                          * the entire peer. Advance next beyond any
 564                          * peer_ni that belongs to the same peer.
 565                          */
 566                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 567                                                  lpni_hashlist) {
 568                                 if (next->lpni_peer_net->lpn_peer != peer)
 569                                         break;
 570                         }
 571                         lnet_peer_del_locked(peer);
 572                 }
 573         }
 574 }
 575
 576 static void
 577 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 578 {
 579         int     i = 3;
 580
 581         spin_lock(&ptable->pt_zombie_lock);
 582         while (ptable->pt_zombies) {
 583                 spin_unlock(&ptable->pt_zombie_lock);
 584
 585                 if (is_power_of_2(i)) {
 586                         CDEBUG(D_WARNING,
 587                                "Waiting for %d zombies on peer table\n",
 588                                ptable->pt_zombies);
 589                 }
 590                 set_current_state(TASK_UNINTERRUPTIBLE);
 591                 schedule_timeout(cfs_time_seconds(1) >> 1);
 592                 spin_lock(&ptable->pt_zombie_lock);
 593         }
 594         spin_unlock(&ptable->pt_zombie_lock);
 595 }
 596
 597 static void
 598 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 599                                 struct lnet_peer_table *ptable)
 600 {
 601         struct lnet_peer_ni     *lp;
 602         struct lnet_peer_ni     *tmp;
 603         lnet_nid_t              gw_nid;
 604         int                     i;
 605
 606         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 607                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 608                                          lpni_hashlist) {
 609                         if (net != lp->lpni_net)
 610                                 continue;
 611
 612                         if (!lnet_isrouter(lp))
 613                                 continue;
 614
 615                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 616
 617                         lnet_net_unlock(LNET_LOCK_EX);
 618                         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), gw_nid);
 619                         lnet_net_lock(LNET_LOCK_EX);
 620                 }
 621         }
 622 }
 623
 624 void
 625 lnet_peer_tables_cleanup(struct lnet_net *net)
 626 {
 627         int i;
 628         struct lnet_peer_table *ptable;
 629
 630         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 631         /* If just deleting the peers for a NI, get rid of any routes these
 632          * peers are gateways for. */
 633         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 634                 lnet_net_lock(LNET_LOCK_EX);
 635                 lnet_peer_table_del_rtrs_locked(net, ptable);
 636                 lnet_net_unlock(LNET_LOCK_EX);
 637         }
 638
 639         /* Start the cleanup process */
 640         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 641                 lnet_net_lock(LNET_LOCK_EX);
 642                 lnet_peer_table_cleanup_locked(net, ptable);
 643                 lnet_net_unlock(LNET_LOCK_EX);
 644         }
 645
 646         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 647                 lnet_peer_ni_finalize_wait(ptable);
 648 }
 649
 650 static struct lnet_peer_ni *
 651 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 652 {
 653         struct list_head        *peers;
 654         struct lnet_peer_ni     *lp;
 655
 656         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 657                 return NULL;
 658
 659         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 660         list_for_each_entry(lp, peers, lpni_hashlist) {
 661                 if (lp->lpni_nid == nid) {
 662                         lnet_peer_ni_addref_locked(lp);
 663                         return lp;
 664                 }
 665         }
 666
 667         return NULL;
 668 }
 669
 670 struct lnet_peer_ni *
 671 lnet_find_peer_ni_locked(lnet_nid_t nid)
 672 {
 673         struct lnet_peer_ni *lpni;
 674         struct lnet_peer_table *ptable;
 675         int cpt;
 676
 677         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 678
 679         ptable = the_lnet.ln_peer_tables[cpt];
 680         lpni = lnet_get_peer_ni_locked(ptable, nid);
 681
 682         return lpni;
 683 }
 684
 685 struct lnet_peer_ni *
 686 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 687 {
 688         struct lnet_peer_net *lpn;
 689         struct lnet_peer_ni *lpni;
 690
 691         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 692         if (!lpn)
 693                 return NULL;
 694
 695         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 696                 if (lpni->lpni_nid == nid)
 697                         return lpni;
 698         }
 699
 700         return NULL;
 701 }
 702
 703 struct lnet_peer *
 704 lnet_find_peer(lnet_nid_t nid)
 705 {
 706         struct lnet_peer_ni *lpni;
 707         struct lnet_peer *lp = NULL;
 708         int cpt;
 709
 710         cpt = lnet_net_lock_current();
 711         lpni = lnet_find_peer_ni_locked(nid);
 712         if (lpni) {
 713                 lp = lpni->lpni_peer_net->lpn_peer;
 714                 lnet_peer_addref_locked(lp);
 715                 lnet_peer_ni_decref_locked(lpni);
 716         }
 717         lnet_net_unlock(cpt);
 718
 719         return lp;
 720 }
 721
 722 struct lnet_peer_net *
 723 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 724 {
 725         struct lnet_peer_net *net;
 726
 727         if (!prev_lpn_id) {
 728                 /* no net id provided return the first net */
 729                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 730                                                struct lnet_peer_net,
 731                                                lpn_peer_nets);
 732
 733                 return net;
 734         }
 735
 736         /* find the net after the one provided */
 737         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 738                 if (net->lpn_net_id == prev_lpn_id) {
 739                         /*
 740                          * if we reached the end of the list loop to the
 741                          * beginning.
 742                          */
 743                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 744                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 745                                                                 struct lnet_peer_net,
 746                                                                 lpn_peer_nets);
 747                         else
 748                                 return list_next_entry(net, lpn_peer_nets);
 749                 }
 750         }
 751
 752         return NULL;
 753 }
 754
 755 struct lnet_peer_ni *
 756 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 757                              struct lnet_peer_net *peer_net,
 758                              struct lnet_peer_ni *prev)
 759 {
 760         struct lnet_peer_ni *lpni;
 761         struct lnet_peer_net *net = peer_net;
 762
 763         if (!prev) {
 764                 if (!net) {
 765                         if (list_empty(&peer->lp_peer_nets))
 766                                 return NULL;
 767
 768                         net = list_entry(peer->lp_peer_nets.next,
 769                                          struct lnet_peer_net,
 770                                          lpn_peer_nets);
 771                 }
 772                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 773                                   lpni_peer_nis);
 774
 775                 return lpni;
 776         }
 777
 778         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 779                 /*
 780                  * if you reached the end of the peer ni list and the peer
 781                  * net is specified then there are no more peer nis in that
 782                  * net.
 783                  */
 784                 if (net)
 785                         return NULL;
 786
 787                 /*
 788                  * we reached the end of this net ni list. move to the
 789                  * next net
 790                  */
 791                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 792                     &peer->lp_peer_nets)
 793                         /* no more nets and no more NIs. */
 794                         return NULL;
 795
 796                 /* get the next net */
 797                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 798                                  struct lnet_peer_net,
 799                                  lpn_peer_nets);
 800                 /* get the ni on it */
 801                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 802                                   lpni_peer_nis);
 803
 804                 return lpni;
 805         }
 806
 807         /* there are more nis left */
 808         lpni = list_entry(prev->lpni_peer_nis.next,
 809                           struct lnet_peer_ni, lpni_peer_nis);
 810
 811         return lpni;
 812 }
 813
 814 /* Call with the ln_api_mutex held */
 815 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 816 {
 817         struct lnet_process_id id;
 818         struct lnet_peer_table *ptable;
 819         struct lnet_peer *lp;
 820         __u32 count = 0;
 821         __u32 size = 0;
 822         int lncpt;
 823         int cpt;
 824         __u32 i;
 825         int rc;
 826
 827         rc = -ESHUTDOWN;
 828         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 829                 goto done;
 830
 831         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 832
 833         /*
 834          * Count the number of peers, and return E2BIG if the buffer
 835          * is too small. We'll also return the desired size.
 836          */
 837         rc = -E2BIG;
 838         for (cpt = 0; cpt < lncpt; cpt++) {
 839                 ptable = the_lnet.ln_peer_tables[cpt];
 840                 count += ptable->pt_peers;
 841         }
 842         size = count * sizeof(*ids);
 843         if (size > *sizep)
 844                 goto done;
 845
 846         /*
 847          * Walk the peer lists and copy out the primary nids.
 848          * This is safe because the peer lists are only modified
 849          * while the ln_api_mutex is held. So we don't need to
 850          * hold the lnet_net_lock as well, and can therefore
 851          * directly call copy_to_user().
 852          */
 853         rc = -EFAULT;
 854         memset(&id, 0, sizeof(id));
 855         id.pid = LNET_PID_LUSTRE;
 856         i = 0;
 857         for (cpt = 0; cpt < lncpt; cpt++) {
 858                 ptable = the_lnet.ln_peer_tables[cpt];
 859                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 860                         if (i >= count)
 861                                 goto done;
 862                         id.nid = lp->lp_primary_nid;
 863                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 864                                 goto done;
 865                         i++;
 866                 }
 867         }
 868         rc = 0;
 869 done:
 870         *countp = count;
 871         *sizep = size;
 872         return rc;
 873 }
 874
 875 /*
 876  * Start pushes to peers that need to be updated for a configuration
 877  * change on this node.
 878  */
 879 void
 880 lnet_push_update_to_peers(int force)
 881 {
 882         struct lnet_peer_table *ptable;
 883         struct lnet_peer *lp;
 884         int lncpt;
 885         int cpt;
 886
 887         lnet_net_lock(LNET_LOCK_EX);
 888         if (lnet_peer_discovery_disabled)
 889                 force = 0;
 890         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 891         for (cpt = 0; cpt < lncpt; cpt++) {
 892                 ptable = the_lnet.ln_peer_tables[cpt];
 893                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 894                         if (force) {
 895                                 spin_lock(&lp->lp_lock);
 896                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 897                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 898                                 spin_unlock(&lp->lp_lock);
 899                         }
 900                         if (lnet_peer_needs_push(lp))
 901                                 lnet_peer_queue_for_discovery(lp);
 902                 }
 903         }
 904         lnet_net_unlock(LNET_LOCK_EX);
 905         wake_up(&the_lnet.ln_dc_waitq);
 906 }
 907
 908 /*
 909  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
 910  * this is a preferred point-to-point path. Call with lnet_net_lock in
 911  * shared mmode.
 912  */
 913 bool
 914 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 915 {
 916         int i;
 917
 918         if (lpni->lpni_pref_nnids == 0)
 919                 return false;
 920         if (lpni->lpni_pref_nnids == 1)
 921                 return lpni->lpni_pref.nid == nid;
 922         for (i = 0; i < lpni->lpni_pref_nnids; i++) {
 923                 if (lpni->lpni_pref.nids[i] == nid)
 924                         return true;
 925         }
 926         return false;
 927 }
 928
 929 /*
 930  * Set a single ni as preferred, provided no preferred ni is already
 931  * defined. Only to be used for non-multi-rail peer_ni.
 932  */
 933 int
 934 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 935 {
 936         int rc = 0;
 937
 938         spin_lock(&lpni->lpni_lock);
 939         if (nid == LNET_NID_ANY) {
 940                 rc = -EINVAL;
 941         } else if (lpni->lpni_pref_nnids > 0) {
 942                 rc = -EPERM;
 943         } else if (lpni->lpni_pref_nnids == 0) {
 944                 lpni->lpni_pref.nid = nid;
 945                 lpni->lpni_pref_nnids = 1;
 946                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
 947         }
 948         spin_unlock(&lpni->lpni_lock);
 949
 950         CDEBUG(D_NET, "peer %s nid %s: %d\n",
 951                libcfs_nid2str(lpni->lpni_nid), libcfs_nid2str(nid), rc);
 952         return rc;
 953 }
 954
 955 /*
 956  * Clear the preferred NID from a non-multi-rail peer_ni, provided
 957  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
 958  */
 959 int
 960 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
 961 {
 962         int rc = 0;
 963
 964         spin_lock(&lpni->lpni_lock);
 965         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
 966                 lpni->lpni_pref_nnids = 0;
 967                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
 968         } else if (lpni->lpni_pref_nnids == 0) {
 969                 rc = -ENOENT;
 970         } else {
 971                 rc = -EPERM;
 972         }
 973         spin_unlock(&lpni->lpni_lock);
 974
 975         CDEBUG(D_NET, "peer %s: %d\n",
 976                libcfs_nid2str(lpni->lpni_nid), rc);
 977         return rc;
 978 }
 979
 980 /*
 981  * Clear the preferred NIDs from a non-multi-rail peer.
 982  */
 983 void
 984 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
 985 {
 986         struct lnet_peer_ni *lpni = NULL;
 987
 988         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 989                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
 990 }
 991
 992 int
 993 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
 994 {
 995         lnet_nid_t *nids = NULL;
 996         lnet_nid_t *oldnids = NULL;
 997         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
 998         int size;
 999         int i;
1000         int rc = 0;
1001
1002         if (nid == LNET_NID_ANY) {
1003                 rc = -EINVAL;
1004                 goto out;
1005         }
1006
1007         if (lpni->lpni_pref_nnids == 1 && lpni->lpni_pref.nid == nid) {
1008                 rc = -EEXIST;
1009                 goto out;
1010         }
1011
1012         /* A non-MR node may have only one preferred NI per peer_ni */
1013         if (lpni->lpni_pref_nnids > 0) {
1014                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1015                         rc = -EPERM;
1016                         goto out;
1017                 }
1018         }
1019
1020         if (lpni->lpni_pref_nnids != 0) {
1021                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1022                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1023                 if (!nids) {
1024                         rc = -ENOMEM;
1025                         goto out;
1026                 }
1027                 for (i = 0; i < lpni->lpni_pref_nnids; i++) {
1028                         if (lpni->lpni_pref.nids[i] == nid) {
1029                                 LIBCFS_FREE(nids, size);
1030                                 rc = -EEXIST;
1031                                 goto out;
1032                         }
1033                         nids[i] = lpni->lpni_pref.nids[i];
1034                 }
1035                 nids[i] = nid;
1036         }
1037
1038         lnet_net_lock(LNET_LOCK_EX);
1039         spin_lock(&lpni->lpni_lock);
1040         if (lpni->lpni_pref_nnids == 0) {
1041                 lpni->lpni_pref.nid = nid;
1042         } else {
1043                 oldnids = lpni->lpni_pref.nids;
1044                 lpni->lpni_pref.nids = nids;
1045         }
1046         lpni->lpni_pref_nnids++;
1047         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1048         spin_unlock(&lpni->lpni_lock);
1049         lnet_net_unlock(LNET_LOCK_EX);
1050
1051         if (oldnids) {
1052                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1053                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1054         }
1055 out:
1056         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1057                 spin_lock(&lpni->lpni_lock);
1058                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1059                 spin_unlock(&lpni->lpni_lock);
1060         }
1061         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1062                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1063         return rc;
1064 }
1065
1066 int
1067 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, lnet_nid_t nid)
1068 {
1069         lnet_nid_t *nids = NULL;
1070         lnet_nid_t *oldnids = NULL;
1071         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1072         int size;
1073         int i, j;
1074         int rc = 0;
1075
1076         if (lpni->lpni_pref_nnids == 0) {
1077                 rc = -ENOENT;
1078                 goto out;
1079         }
1080
1081         if (lpni->lpni_pref_nnids == 1) {
1082                 if (lpni->lpni_pref.nid != nid) {
1083                         rc = -ENOENT;
1084                         goto out;
1085                 }
1086         } else if (lpni->lpni_pref_nnids == 2) {
1087                 if (lpni->lpni_pref.nids[0] != nid &&
1088                     lpni->lpni_pref.nids[1] != nid) {
1089                         rc = -ENOENT;
1090                         goto out;
1091                 }
1092         } else {
1093                 size = sizeof(*nids) * (lpni->lpni_pref_nnids - 1);
1094                 LIBCFS_CPT_ALLOC(nids, lnet_cpt_table(), lpni->lpni_cpt, size);
1095                 if (!nids) {
1096                         rc = -ENOMEM;
1097                         goto out;
1098                 }
1099                 for (i = 0, j = 0; i < lpni->lpni_pref_nnids; i++) {
1100                         if (lpni->lpni_pref.nids[i] != nid)
1101                                 continue;
1102                         nids[j++] = lpni->lpni_pref.nids[i];
1103                 }
1104                 /* Check if we actually removed a nid. */
1105                 if (j == lpni->lpni_pref_nnids) {
1106                         LIBCFS_FREE(nids, size);
1107                         rc = -ENOENT;
1108                         goto out;
1109                 }
1110         }
1111
1112         lnet_net_lock(LNET_LOCK_EX);
1113         spin_lock(&lpni->lpni_lock);
1114         if (lpni->lpni_pref_nnids == 1) {
1115                 lpni->lpni_pref.nid = LNET_NID_ANY;
1116         } else if (lpni->lpni_pref_nnids == 2) {
1117                 oldnids = lpni->lpni_pref.nids;
1118                 if (oldnids[0] == nid)
1119                         lpni->lpni_pref.nid = oldnids[1];
1120                 else
1121                         lpni->lpni_pref.nid = oldnids[2];
1122         } else {
1123                 oldnids = lpni->lpni_pref.nids;
1124                 lpni->lpni_pref.nids = nids;
1125         }
1126         lpni->lpni_pref_nnids--;
1127         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1128         spin_unlock(&lpni->lpni_lock);
1129         lnet_net_unlock(LNET_LOCK_EX);
1130
1131         if (oldnids) {
1132                 size = sizeof(*nids) * (lpni->lpni_pref_nnids + 1);
1133                 LIBCFS_FREE(oldnids, sizeof(*oldnids) * size);
1134         }
1135 out:
1136         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1137                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid), rc);
1138         return rc;
1139 }
1140
1141 lnet_nid_t
1142 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1143 {
1144         struct lnet_peer_ni *lpni;
1145         lnet_nid_t primary_nid = nid;
1146
1147         lpni = lnet_find_peer_ni_locked(nid);
1148         if (lpni) {
1149                 primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1150                 lnet_peer_ni_decref_locked(lpni);
1151         }
1152
1153         return primary_nid;
1154 }
1155
1156 bool
1157 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1158 __must_hold(&lp->lp_lock)
1159 {
1160         if (lnet_peer_discovery_disabled)
1161                 return true;
1162
1163         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1164             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1165                 return true;
1166         }
1167
1168         return false;
1169 }
1170
1171 /*
1172  * Peer Discovery
1173  */
1174 bool
1175 lnet_is_discovery_disabled(struct lnet_peer *lp)
1176 {
1177         bool rc = false;
1178
1179         spin_lock(&lp->lp_lock);
1180         rc = lnet_is_discovery_disabled_locked(lp);
1181         spin_unlock(&lp->lp_lock);
1182
1183         return rc;
1184 }
1185
1186 lnet_nid_t
1187 LNetPrimaryNID(lnet_nid_t nid)
1188 {
1189         struct lnet_peer *lp;
1190         struct lnet_peer_ni *lpni;
1191         lnet_nid_t primary_nid = nid;
1192         int rc = 0;
1193         int cpt;
1194
1195         cpt = lnet_net_lock_current();
1196         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1197         if (IS_ERR(lpni)) {
1198                 rc = PTR_ERR(lpni);
1199                 goto out_unlock;
1200         }
1201         lp = lpni->lpni_peer_net->lpn_peer;
1202
1203         while (!lnet_peer_is_uptodate(lp)) {
1204                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1205                 if (rc)
1206                         goto out_decref;
1207                 lp = lpni->lpni_peer_net->lpn_peer;
1208
1209                 /* Only try once if discovery is disabled */
1210                 if (lnet_is_discovery_disabled(lp))
1211                         break;
1212         }
1213         primary_nid = lp->lp_primary_nid;
1214 out_decref:
1215         lnet_peer_ni_decref_locked(lpni);
1216 out_unlock:
1217         lnet_net_unlock(cpt);
1218
1219         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1220                libcfs_nid2str(primary_nid), rc);
1221         return primary_nid;
1222 }
1223 EXPORT_SYMBOL(LNetPrimaryNID);
1224
1225 struct lnet_peer_net *
1226 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1227 {
1228         struct lnet_peer_net *peer_net;
1229         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1230                 if (peer_net->lpn_net_id == net_id)
1231                         return peer_net;
1232         }
1233         return NULL;
1234 }
1235
1236 /*
1237  * Attach a peer_ni to a peer_net and peer. This function assumes
1238  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1239  * may be attached to a different peer, in which case it will be
1240  * properly detached first. The whole operation is done atomically.
1241  *
1242  * Always returns 0.  This is the last function called from functions
1243  * that do return an int, so returning 0 here allows the compiler to
1244  * do a tail call.
1245  */
1246 static int
1247 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1248                                 struct lnet_peer_net *lpn,
1249                                 struct lnet_peer_ni *lpni,
1250                                 unsigned flags)
1251 {
1252         struct lnet_peer_table *ptable;
1253
1254         /* Install the new peer_ni */
1255         lnet_net_lock(LNET_LOCK_EX);
1256         /* Add peer_ni to global peer table hash, if necessary. */
1257         if (list_empty(&lpni->lpni_hashlist)) {
1258                 int hash = lnet_nid2peerhash(lpni->lpni_nid);
1259
1260                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1261                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1262                 ptable->pt_version++;
1263                 /* This is the 1st refcount on lpni. */
1264                 atomic_inc(&lpni->lpni_refcount);
1265         }
1266
1267         /* Detach the peer_ni from an existing peer, if necessary. */
1268         if (lpni->lpni_peer_net) {
1269                 LASSERT(lpni->lpni_peer_net != lpn);
1270                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1271                 lnet_peer_detach_peer_ni_locked(lpni);
1272                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1273                 lpni->lpni_peer_net = NULL;
1274         }
1275
1276         /* Add peer_ni to peer_net */
1277         lpni->lpni_peer_net = lpn;
1278         list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1279         lnet_peer_net_addref_locked(lpn);
1280
1281         /* Add peer_net to peer */
1282         if (!lpn->lpn_peer) {
1283                 lpn->lpn_peer = lp;
1284                 list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1285                 lnet_peer_addref_locked(lp);
1286         }
1287
1288         /* Add peer to global peer list, if necessary */
1289         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1290         if (list_empty(&lp->lp_peer_list)) {
1291                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1292                 ptable->pt_peers++;
1293         }
1294
1295
1296         /* Update peer state */
1297         spin_lock(&lp->lp_lock);
1298         if (flags & LNET_PEER_CONFIGURED) {
1299                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1300                         lp->lp_state |= LNET_PEER_CONFIGURED;
1301         }
1302         if (flags & LNET_PEER_MULTI_RAIL) {
1303                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1304                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1305                         lnet_peer_clr_non_mr_pref_nids(lp);
1306                 }
1307         }
1308         spin_unlock(&lp->lp_lock);
1309
1310         lp->lp_nnis++;
1311         lnet_net_unlock(LNET_LOCK_EX);
1312
1313         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1314                libcfs_nid2str(lp->lp_primary_nid),
1315                libcfs_nid2str(lpni->lpni_nid), flags);
1316
1317         return 0;
1318 }
1319
1320 /*
1321  * Create a new peer, with nid as its primary nid.
1322  *
1323  * Call with the lnet_api_mutex held.
1324  */
1325 static int
1326 lnet_peer_add(lnet_nid_t nid, unsigned flags)
1327 {
1328         struct lnet_peer *lp;
1329         struct lnet_peer_net *lpn;
1330         struct lnet_peer_ni *lpni;
1331         int rc = 0;
1332
1333         LASSERT(nid != LNET_NID_ANY);
1334
1335         /*
1336          * No need for the lnet_net_lock here, because the
1337          * lnet_api_mutex is held.
1338          */
1339         lpni = lnet_find_peer_ni_locked(nid);
1340         if (lpni) {
1341                 /* A peer with this NID already exists. */
1342                 lp = lpni->lpni_peer_net->lpn_peer;
1343                 lnet_peer_ni_decref_locked(lpni);
1344                 /*
1345                  * This is an error if the peer was configured and the
1346                  * primary NID differs or an attempt is made to change
1347                  * the Multi-Rail flag. Otherwise the assumption is
1348                  * that an existing peer is being modified.
1349                  */
1350                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1351                         if (lp->lp_primary_nid != nid)
1352                                 rc = -EEXIST;
1353                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1354                                 rc = -EPERM;
1355                         goto out;
1356                 }
1357                 /* Delete and recreate as a configured peer. */
1358                 lnet_peer_del(lp);
1359         }
1360
1361         /* Create peer, peer_net, and peer_ni. */
1362         rc = -ENOMEM;
1363         lp = lnet_peer_alloc(nid);
1364         if (!lp)
1365                 goto out;
1366         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1367         if (!lpn)
1368                 goto out_free_lp;
1369         lpni = lnet_peer_ni_alloc(nid);
1370         if (!lpni)
1371                 goto out_free_lpn;
1372
1373         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1374
1375 out_free_lpn:
1376         LIBCFS_FREE(lpn, sizeof(*lpn));
1377 out_free_lp:
1378         LIBCFS_FREE(lp, sizeof(*lp));
1379 out:
1380         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1381                libcfs_nid2str(nid), flags, rc);
1382         return rc;
1383 }
1384
1385 /*
1386  * Add a NID to a peer. Call with ln_api_mutex held.
1387  *
1388  * Error codes:
1389  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1390  *  -EEXIST:   The NID was configured by DLC for a different peer.
1391  *  -ENOMEM:   Out of memory.
1392  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1393  *             non-multi-rail peer.
1394  */
1395 static int
1396 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1397 {
1398         struct lnet_peer_net *lpn;
1399         struct lnet_peer_ni *lpni;
1400         int rc = 0;
1401
1402         LASSERT(lp);
1403         LASSERT(nid != LNET_NID_ANY);
1404
1405         /* A configured peer can only be updated through configuration. */
1406         if (!(flags & LNET_PEER_CONFIGURED)) {
1407                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1408                         rc = -EPERM;
1409                         goto out;
1410                 }
1411         }
1412
1413         /*
1414          * The MULTI_RAIL flag can be set but not cleared, because
1415          * that would leave the peer struct in an invalid state.
1416          */
1417         if (flags & LNET_PEER_MULTI_RAIL) {
1418                 spin_lock(&lp->lp_lock);
1419                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1420                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1421                         lnet_peer_clr_non_mr_pref_nids(lp);
1422                 }
1423                 spin_unlock(&lp->lp_lock);
1424         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1425                 rc = -EPERM;
1426                 goto out;
1427         }
1428
1429         lpni = lnet_find_peer_ni_locked(nid);
1430         if (lpni) {
1431                 /*
1432                  * A peer_ni already exists. This is only a problem if
1433                  * it is not connected to this peer and was configured
1434                  * by DLC.
1435                  */
1436                 lnet_peer_ni_decref_locked(lpni);
1437                 if (lpni->lpni_peer_net->lpn_peer == lp)
1438                         goto out;
1439                 if (lnet_peer_ni_is_configured(lpni)) {
1440                         rc = -EEXIST;
1441                         goto out;
1442                 }
1443                 /* If this is the primary NID, destroy the peer. */
1444                 if (lnet_peer_ni_is_primary(lpni)) {
1445                         struct lnet_peer *rtr_lp =
1446                           lpni->lpni_peer_net->lpn_peer;
1447                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1448                         /*
1449                          * if we're trying to delete a router it means
1450                          * we're moving this peer NI to a new peer so must
1451                          * transfer router properties to the new peer
1452                          */
1453                         if (rtr_refcount > 0) {
1454                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1455                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1456                         }
1457                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1458                         lpni = lnet_peer_ni_alloc(nid);
1459                         if (!lpni) {
1460                                 rc = -ENOMEM;
1461                                 goto out;
1462                         }
1463                 }
1464         } else {
1465                 lpni = lnet_peer_ni_alloc(nid);
1466                 if (!lpni) {
1467                         rc = -ENOMEM;
1468                         goto out;
1469                 }
1470         }
1471
1472         /*
1473          * Get the peer_net. Check that we're not adding a second
1474          * peer_ni on a peer_net of a non-multi-rail peer.
1475          */
1476         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
1477         if (!lpn) {
1478                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1479                 if (!lpn) {
1480                         rc = -ENOMEM;
1481                         goto out_free_lpni;
1482                 }
1483         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1484                 rc = -ENOTUNIQ;
1485                 goto out_free_lpni;
1486         }
1487
1488         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1489
1490 out_free_lpni:
1491         /* If the peer_ni was allocated above its peer_net pointer is NULL */
1492         if (!lpni->lpni_peer_net)
1493                 LIBCFS_FREE(lpni, sizeof(*lpni));
1494 out:
1495         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1496                libcfs_nid2str(lp->lp_primary_nid), libcfs_nid2str(nid),
1497                flags, rc);
1498         return rc;
1499 }
1500
1501 /*
1502  * Update the primary NID of a peer, if possible.
1503  *
1504  * Call with the lnet_api_mutex held.
1505  */
1506 static int
1507 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
1508 {
1509         lnet_nid_t old = lp->lp_primary_nid;
1510         int rc = 0;
1511
1512         if (lp->lp_primary_nid == nid)
1513                 goto out;
1514         rc = lnet_peer_add_nid(lp, nid, flags);
1515         if (rc)
1516                 goto out;
1517         lp->lp_primary_nid = nid;
1518 out:
1519         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1520                libcfs_nid2str(old), libcfs_nid2str(nid), rc);
1521         return rc;
1522 }
1523
1524 /*
1525  * lpni creation initiated due to traffic either sending or receiving.
1526  */
1527 static int
1528 lnet_peer_ni_traffic_add(lnet_nid_t nid, lnet_nid_t pref)
1529 {
1530         struct lnet_peer *lp;
1531         struct lnet_peer_net *lpn;
1532         struct lnet_peer_ni *lpni;
1533         unsigned flags = 0;
1534         int rc = 0;
1535
1536         if (nid == LNET_NID_ANY) {
1537                 rc = -EINVAL;
1538                 goto out;
1539         }
1540
1541         /* lnet_net_lock is not needed here because ln_api_lock is held */
1542         lpni = lnet_find_peer_ni_locked(nid);
1543         if (lpni) {
1544                 /*
1545                  * We must have raced with another thread. Since we
1546                  * know next to nothing about a peer_ni created by
1547                  * traffic, we just assume everything is ok and
1548                  * return.
1549                  */
1550                 lnet_peer_ni_decref_locked(lpni);
1551                 goto out;
1552         }
1553
1554         /* Create peer, peer_net, and peer_ni. */
1555         rc = -ENOMEM;
1556         lp = lnet_peer_alloc(nid);
1557         if (!lp)
1558                 goto out;
1559         lpn = lnet_peer_net_alloc(LNET_NIDNET(nid));
1560         if (!lpn)
1561                 goto out_free_lp;
1562         lpni = lnet_peer_ni_alloc(nid);
1563         if (!lpni)
1564                 goto out_free_lpn;
1565         if (pref != LNET_NID_ANY)
1566                 lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1567
1568         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1569
1570 out_free_lpn:
1571         LIBCFS_FREE(lpn, sizeof(*lpn));
1572 out_free_lp:
1573         LIBCFS_FREE(lp, sizeof(*lp));
1574 out:
1575         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(nid), rc);
1576         return rc;
1577 }
1578
1579 /*
1580  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1581  *
1582  * This API handles the following combinations:
1583  *   Create a peer with its primary NI if only the prim_nid is provided
1584  *   Add a NID to a peer identified by the prim_nid. The peer identified
1585  *   by the prim_nid must already exist.
1586  *   The peer being created may be non-MR.
1587  *
1588  * The caller must hold ln_api_mutex. This prevents the peer from
1589  * being created/modified/deleted by a different thread.
1590  */
1591 int
1592 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr)
1593 {
1594         struct lnet_peer *lp = NULL;
1595         struct lnet_peer_ni *lpni;
1596         unsigned flags;
1597
1598         /* The prim_nid must always be specified */
1599         if (prim_nid == LNET_NID_ANY)
1600                 return -EINVAL;
1601
1602         flags = LNET_PEER_CONFIGURED;
1603         if (mr)
1604                 flags |= LNET_PEER_MULTI_RAIL;
1605
1606         /*
1607          * If nid isn't specified, we must create a new peer with
1608          * prim_nid as its primary nid.
1609          */
1610         if (nid == LNET_NID_ANY)
1611                 return lnet_peer_add(prim_nid, flags);
1612
1613         /* Look up the prim_nid, which must exist. */
1614         lpni = lnet_find_peer_ni_locked(prim_nid);
1615         if (!lpni)
1616                 return -ENOENT;
1617         lnet_peer_ni_decref_locked(lpni);
1618         lp = lpni->lpni_peer_net->lpn_peer;
1619
1620         /* Peer must have been configured. */
1621         if (!(lp->lp_state & LNET_PEER_CONFIGURED)) {
1622                 CDEBUG(D_NET, "peer %s was not configured\n",
1623                        libcfs_nid2str(prim_nid));
1624                 return -ENOENT;
1625         }
1626
1627         /* Primary NID must match */
1628         if (lp->lp_primary_nid != prim_nid) {
1629                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1630                        libcfs_nid2str(prim_nid),
1631                        libcfs_nid2str(lp->lp_primary_nid));
1632                 return -ENODEV;
1633         }
1634
1635         /* Multi-Rail flag must match. */
1636         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1637                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1638                        libcfs_nid2str(prim_nid));
1639                 return -EPERM;
1640         }
1641
1642         return lnet_peer_add_nid(lp, nid, flags);
1643 }
1644
1645 /*
1646  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1647  *
1648  * This API handles the following combinations:
1649  *   Delete a NI from a peer if both prim_nid and nid are provided.
1650  *   Delete a peer if only prim_nid is provided.
1651  *   Delete a peer if its primary nid is provided.
1652  *
1653  * The caller must hold ln_api_mutex. This prevents the peer from
1654  * being modified/deleted by a different thread.
1655  */
1656 int
1657 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1658 {
1659         struct lnet_peer *lp;
1660         struct lnet_peer_ni *lpni;
1661         unsigned flags;
1662
1663         if (prim_nid == LNET_NID_ANY)
1664                 return -EINVAL;
1665
1666         lpni = lnet_find_peer_ni_locked(prim_nid);
1667         if (!lpni)
1668                 return -ENOENT;
1669         lnet_peer_ni_decref_locked(lpni);
1670         lp = lpni->lpni_peer_net->lpn_peer;
1671
1672         if (prim_nid != lp->lp_primary_nid) {
1673                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1674                        libcfs_nid2str(prim_nid),
1675                        libcfs_nid2str(lp->lp_primary_nid));
1676                 return -ENODEV;
1677         }
1678
1679         lnet_net_lock(LNET_LOCK_EX);
1680         if (lp->lp_rtr_refcount > 0) {
1681                 lnet_net_unlock(LNET_LOCK_EX);
1682                 CERROR("%s is a router. Can not be deleted\n",
1683                        libcfs_nid2str(prim_nid));
1684                 return -EBUSY;
1685         }
1686         lnet_net_unlock(LNET_LOCK_EX);
1687
1688         if (nid == LNET_NID_ANY || nid == lp->lp_primary_nid)
1689                 return lnet_peer_del(lp);
1690
1691         flags = LNET_PEER_CONFIGURED;
1692         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1693                 flags |= LNET_PEER_MULTI_RAIL;
1694
1695         return lnet_peer_del_nid(lp, nid, flags);
1696 }
1697
1698 void
1699 lnet_destroy_peer_ni_locked(struct lnet_peer_ni *lpni)
1700 {
1701         struct lnet_peer_table *ptable;
1702         struct lnet_peer_net *lpn;
1703
1704         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nid2str(lpni->lpni_nid));
1705
1706         LASSERT(atomic_read(&lpni->lpni_refcount) == 0);
1707         LASSERT(list_empty(&lpni->lpni_txq));
1708         LASSERT(lpni->lpni_txqnob == 0);
1709         LASSERT(list_empty(&lpni->lpni_peer_nis));
1710         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1711
1712         lpn = lpni->lpni_peer_net;
1713         lpni->lpni_peer_net = NULL;
1714         lpni->lpni_net = NULL;
1715
1716         /* remove the peer ni from the zombie list */
1717         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1718         spin_lock(&ptable->pt_zombie_lock);
1719         list_del_init(&lpni->lpni_hashlist);
1720         ptable->pt_zombies--;
1721         spin_unlock(&ptable->pt_zombie_lock);
1722
1723         if (lpni->lpni_pref_nnids > 1) {
1724                 LIBCFS_FREE(lpni->lpni_pref.nids,
1725                         sizeof(*lpni->lpni_pref.nids) * lpni->lpni_pref_nnids);
1726         }
1727         LIBCFS_FREE(lpni, sizeof(*lpni));
1728
1729         lnet_peer_net_decref_locked(lpn);
1730 }
1731
1732 struct lnet_peer_ni *
1733 lnet_nid2peerni_ex(lnet_nid_t nid, int cpt)
1734 {
1735         struct lnet_peer_ni *lpni = NULL;
1736         int rc;
1737
1738         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1739                 return ERR_PTR(-ESHUTDOWN);
1740
1741         /*
1742          * find if a peer_ni already exists.
1743          * If so then just return that.
1744          */
1745         lpni = lnet_find_peer_ni_locked(nid);
1746         if (lpni)
1747                 return lpni;
1748
1749         lnet_net_unlock(cpt);
1750
1751         rc = lnet_peer_ni_traffic_add(nid, LNET_NID_ANY);
1752         if (rc) {
1753                 lpni = ERR_PTR(rc);
1754                 goto out_net_relock;
1755         }
1756
1757         lpni = lnet_find_peer_ni_locked(nid);
1758         LASSERT(lpni);
1759
1760 out_net_relock:
1761         lnet_net_lock(cpt);
1762
1763         return lpni;
1764 }
1765
1766 /*
1767  * Get a peer_ni for the given nid, create it if necessary. Takes a
1768  * hold on the peer_ni.
1769  */
1770 struct lnet_peer_ni *
1771 lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
1772 {
1773         struct lnet_peer_ni *lpni = NULL;
1774         int rc;
1775
1776         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1777                 return ERR_PTR(-ESHUTDOWN);
1778
1779         /*
1780          * find if a peer_ni already exists.
1781          * If so then just return that.
1782          */
1783         lpni = lnet_find_peer_ni_locked(nid);
1784         if (lpni)
1785                 return lpni;
1786
1787         /*
1788          * Slow path:
1789          * use the lnet_api_mutex to serialize the creation of the peer_ni
1790          * and the creation/deletion of the local ni/net. When a local ni is
1791          * created, if there exists a set of peer_nis on that network,
1792          * they need to be traversed and updated. When a local NI is
1793          * deleted, which could result in a network being deleted, then
1794          * all peer nis on that network need to be removed as well.
1795          *
1796          * Creation through traffic should also be serialized with
1797          * creation through DLC.
1798          */
1799         lnet_net_unlock(cpt);
1800         mutex_lock(&the_lnet.ln_api_mutex);
1801         /*
1802          * Shutdown is only set under the ln_api_lock, so a single
1803          * check here is sufficent.
1804          */
1805         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1806                 lpni = ERR_PTR(-ESHUTDOWN);
1807                 goto out_mutex_unlock;
1808         }
1809
1810         rc = lnet_peer_ni_traffic_add(nid, pref);
1811         if (rc) {
1812                 lpni = ERR_PTR(rc);
1813                 goto out_mutex_unlock;
1814         }
1815
1816         lpni = lnet_find_peer_ni_locked(nid);
1817         LASSERT(lpni);
1818
1819 out_mutex_unlock:
1820         mutex_unlock(&the_lnet.ln_api_mutex);
1821         lnet_net_lock(cpt);
1822
1823         /* Lock has been dropped, check again for shutdown. */
1824         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1825                 if (!IS_ERR(lpni))
1826                         lnet_peer_ni_decref_locked(lpni);
1827                 lpni = ERR_PTR(-ESHUTDOWN);
1828         }
1829
1830         return lpni;
1831 }
1832
1833 bool
1834 lnet_peer_gw_discovery(struct lnet_peer *lp)
1835 {
1836         bool rc = false;
1837
1838         spin_lock(&lp->lp_lock);
1839         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
1840                 rc = true;
1841         spin_unlock(&lp->lp_lock);
1842
1843         return rc;
1844 }
1845
1846 bool
1847 lnet_peer_is_uptodate(struct lnet_peer *lp)
1848 {
1849         bool rc;
1850
1851         spin_lock(&lp->lp_lock);
1852         rc = lnet_peer_is_uptodate_locked(lp);
1853         spin_unlock(&lp->lp_lock);
1854         return rc;
1855 }
1856
1857 /*
1858  * Is a peer uptodate from the point of view of discovery?
1859  *
1860  * If it is currently being processed, obviously not.
1861  * A forced Ping or Push is also handled by the discovery thread.
1862  *
1863  * Otherwise look at whether the peer needs rediscovering.
1864  */
1865 bool
1866 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
1867 __must_hold(&lp->lp_lock)
1868 {
1869         bool rc;
1870
1871         if (lp->lp_state & (LNET_PEER_DISCOVERING |
1872                             LNET_PEER_FORCE_PING |
1873                             LNET_PEER_FORCE_PUSH)) {
1874                 rc = false;
1875         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
1876                 rc = false;
1877         } else if (lnet_peer_needs_push(lp)) {
1878                 rc = false;
1879         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
1880                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
1881                         rc = true;
1882                 else
1883                         rc = false;
1884         } else {
1885                 rc = false;
1886         }
1887
1888         return rc;
1889 }
1890
1891 /*
1892  * Queue a peer for the attention of the discovery thread.  Call with
1893  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
1894  * -EALREADY if the peer was already queued.
1895  */
1896 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
1897 {
1898         int rc;
1899
1900         spin_lock(&lp->lp_lock);
1901         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
1902                 lp->lp_state |= LNET_PEER_DISCOVERING;
1903         spin_unlock(&lp->lp_lock);
1904         if (list_empty(&lp->lp_dc_list)) {
1905                 lnet_peer_addref_locked(lp);
1906                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
1907                 wake_up(&the_lnet.ln_dc_waitq);
1908                 rc = 0;
1909         } else {
1910                 rc = -EALREADY;
1911         }
1912
1913         CDEBUG(D_NET, "Queue peer %s: %d\n",
1914                libcfs_nid2str(lp->lp_primary_nid), rc);
1915
1916         return rc;
1917 }
1918
1919 /*
1920  * Discovery of a peer is complete. Wake all waiters on the peer.
1921  * Call with lnet_net_lock/EX held.
1922  */
1923 static void lnet_peer_discovery_complete(struct lnet_peer *lp)
1924 {
1925         struct lnet_msg *msg, *tmp;
1926         int rc = 0;
1927         LIST_HEAD(pending_msgs);
1928
1929         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
1930                libcfs_nid2str(lp->lp_primary_nid));
1931
1932         list_del_init(&lp->lp_dc_list);
1933         spin_lock(&lp->lp_lock);
1934         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
1935         spin_unlock(&lp->lp_lock);
1936         wake_up_all(&lp->lp_dc_waitq);
1937
1938         if (lp->lp_rtr_refcount > 0)
1939                 lnet_router_discovery_complete(lp);
1940
1941         lnet_net_unlock(LNET_LOCK_EX);
1942
1943         /* iterate through all pending messages and send them again */
1944         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
1945                 list_del_init(&msg->msg_list);
1946                 if (lp->lp_dc_error) {
1947                         lnet_finalize(msg, lp->lp_dc_error);
1948                         continue;
1949                 }
1950
1951                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
1952                        lnet_msgtyp2str(msg->msg_type),
1953                        libcfs_id2str(msg->msg_target));
1954                 rc = lnet_send(msg->msg_src_nid_param, msg,
1955                                msg->msg_rtr_nid_param);
1956                 if (rc < 0) {
1957                         CNETERR("Error sending %s to %s: %d\n",
1958                                lnet_msgtyp2str(msg->msg_type),
1959                                libcfs_id2str(msg->msg_target), rc);
1960                         lnet_finalize(msg, rc);
1961                 }
1962         }
1963         lnet_net_lock(LNET_LOCK_EX);
1964         lnet_peer_decref_locked(lp);
1965 }
1966
1967 /*
1968  * Handle inbound push.
1969  * Like any event handler, called with lnet_res_lock/CPT held.
1970  */
1971 void lnet_peer_push_event(struct lnet_event *ev)
1972 {
1973         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1974         struct lnet_peer *lp;
1975
1976         /* lnet_find_peer() adds a refcount */
1977         lp = lnet_find_peer(ev->source.nid);
1978         if (!lp) {
1979                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
1980                        libcfs_nid2str(ev->initiator.nid),
1981                        libcfs_nid2str(ev->source.nid));
1982                 return;
1983         }
1984
1985         /* Ensure peer state remains consistent while we modify it. */
1986         spin_lock(&lp->lp_lock);
1987
1988         /*
1989          * If some kind of error happened the contents of the message
1990          * cannot be used. Clear the NIDS_UPTODATE and set the
1991          * FORCE_PING flag to trigger a ping.
1992          */
1993         if (ev->status) {
1994                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1995                 lp->lp_state |= LNET_PEER_FORCE_PING;
1996                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
1997                        ev->status,
1998                        libcfs_nid2str(lp->lp_primary_nid),
1999                        libcfs_nid2str(ev->source.nid));
2000                 goto out;
2001         }
2002
2003         /*
2004          * A push with invalid or corrupted info. Clear the UPTODATE
2005          * flag to trigger a ping.
2006          */
2007         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2008                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2009                 lp->lp_state |= LNET_PEER_FORCE_PING;
2010                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2011                        libcfs_nid2str(lp->lp_primary_nid));
2012                 goto out;
2013         }
2014
2015         /*
2016          * Make sure we'll allocate the correct size ping buffer when
2017          * pinging the peer.
2018          */
2019         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2020                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2021
2022         /*
2023          * A non-Multi-Rail peer is not supposed to be capable of
2024          * sending a push.
2025          */
2026         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2027                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2028                        libcfs_nid2str(lp->lp_primary_nid));
2029                 goto out;
2030         }
2031
2032         /*
2033          * Check the MULTIRAIL flag. Complain if the peer was DLC
2034          * configured without it.
2035          */
2036         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
2037                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2038                         CERROR("Push says %s is Multi-Rail, DLC says not\n",
2039                                libcfs_nid2str(lp->lp_primary_nid));
2040                 } else {
2041                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2042                         lnet_peer_clr_non_mr_pref_nids(lp);
2043                 }
2044         }
2045
2046         /*
2047          * The peer may have discovery disabled at its end. Set
2048          * NO_DISCOVERY as appropriate.
2049          */
2050         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2051                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2052                        libcfs_nid2str(lp->lp_primary_nid));
2053                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2054         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2055                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2056                        libcfs_nid2str(lp->lp_primary_nid));
2057                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2058         }
2059
2060         /*
2061          * Check for truncation of the Put message. Clear the
2062          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2063          * and tell discovery to allocate a bigger buffer.
2064          */
2065         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2066                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2067                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2068                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2069                 lp->lp_state |= LNET_PEER_FORCE_PING;
2070                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2071                        libcfs_nid2str(lp->lp_primary_nid),
2072                        pbuf->pb_info.pi_nnis);
2073                 goto out;
2074         }
2075
2076         /* always assume new data */
2077         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2078         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2079
2080         /*
2081          * If there is data present that hasn't been processed yet,
2082          * we'll replace it if the Put contained newer data and it
2083          * fits. We're racing with a Ping or earlier Push in this
2084          * case.
2085          */
2086         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2087                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2088                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2089                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2090                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2091                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2092                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2093                               libcfs_nid2str(lp->lp_primary_nid),
2094                               LNET_PING_BUFFER_SEQNO(pbuf),
2095                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2096                 }
2097                 goto out;
2098         }
2099
2100         /*
2101          * Allocate a buffer to copy the data. On a failure we drop
2102          * the Push and set FORCE_PING to force the discovery
2103          * thread to fix the problem by pinging the peer.
2104          */
2105         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2106         if (!lp->lp_data) {
2107                 lp->lp_state |= LNET_PEER_FORCE_PING;
2108                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2109                        libcfs_nid2str(lp->lp_primary_nid),
2110                        LNET_PING_BUFFER_SEQNO(pbuf));
2111                 goto out;
2112         }
2113
2114         /* Success */
2115         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2116                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2117         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2118         CDEBUG(D_NET, "Received Push %s %u\n",
2119                libcfs_nid2str(lp->lp_primary_nid),
2120                LNET_PING_BUFFER_SEQNO(pbuf));
2121
2122 out:
2123         /*
2124          * Queue the peer for discovery if not done, force it on the request
2125          * queue and wake the discovery thread if the peer was already queued,
2126          * because its status changed.
2127          */
2128         spin_unlock(&lp->lp_lock);
2129         lnet_net_lock(LNET_LOCK_EX);
2130         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2131                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2132                 wake_up(&the_lnet.ln_dc_waitq);
2133         }
2134         /* Drop refcount from lookup */
2135         lnet_peer_decref_locked(lp);
2136         lnet_net_unlock(LNET_LOCK_EX);
2137 }
2138
2139 /*
2140  * Clear the discovery error state, unless we're already discovering
2141  * this peer, in which case the error is current.
2142  */
2143 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2144 {
2145         spin_lock(&lp->lp_lock);
2146         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2147                 lp->lp_dc_error = 0;
2148         spin_unlock(&lp->lp_lock);
2149 }
2150
2151 /*
2152  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2153  * dropped/retaken within this function. An lnet_peer_ni is passed in
2154  * because discovery could tear down an lnet_peer.
2155  */
2156 int
2157 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2158 {
2159         DEFINE_WAIT(wait);
2160         struct lnet_peer *lp;
2161         int rc = 0;
2162         int count = 0;
2163
2164 again:
2165         lnet_net_unlock(cpt);
2166         lnet_net_lock(LNET_LOCK_EX);
2167         lp = lpni->lpni_peer_net->lpn_peer;
2168         lnet_peer_clear_discovery_error(lp);
2169
2170         /*
2171          * We're willing to be interrupted. The lpni can become a
2172          * zombie if we race with DLC, so we must check for that.
2173          */
2174         for (;;) {
2175                 /* Keep lp alive when the lnet_net_lock is unlocked */
2176                 lnet_peer_addref_locked(lp);
2177                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2178                 if (signal_pending(current))
2179                         break;
2180                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2181                         break;
2182                 /*
2183                  * Don't repeat discovery if discovery is disabled. This is
2184                  * done to ensure we can use discovery as a standard ping as
2185                  * well for backwards compatibility with routers which do not
2186                  * have discovery or have discovery disabled
2187                  */
2188                 if (lnet_is_discovery_disabled(lp) && count > 0)
2189                         break;
2190                 if (lp->lp_dc_error)
2191                         break;
2192                 if (lnet_peer_is_uptodate(lp))
2193                         break;
2194                 lnet_peer_queue_for_discovery(lp);
2195                 count++;
2196                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2197
2198                 /*
2199                  * If caller requested a non-blocking operation then
2200                  * return immediately. Once discovery is complete any
2201                  * pending messages that were stopped due to discovery
2202                  * will be transmitted.
2203                  */
2204                 if (!block)
2205                         break;
2206
2207                 lnet_net_unlock(LNET_LOCK_EX);
2208                 schedule();
2209                 finish_wait(&lp->lp_dc_waitq, &wait);
2210                 lnet_net_lock(LNET_LOCK_EX);
2211                 lnet_peer_decref_locked(lp);
2212                 /* Peer may have changed */
2213                 lp = lpni->lpni_peer_net->lpn_peer;
2214         }
2215         finish_wait(&lp->lp_dc_waitq, &wait);
2216
2217         lnet_net_unlock(LNET_LOCK_EX);
2218         lnet_net_lock(cpt);
2219         lnet_peer_decref_locked(lp);
2220         /*
2221          * The peer may have changed, so re-check and rediscover if that turns
2222          * out to have been the case. The reference count on lp ensured that
2223          * even if it was unlinked from lpni the memory could not be recycled.
2224          * Thus the check below is sufficient to determine whether the peer
2225          * changed. If the peer changed, then lp must not be dereferenced.
2226          */
2227         if (lp != lpni->lpni_peer_net->lpn_peer)
2228                 goto again;
2229
2230         if (signal_pending(current))
2231                 rc = -EINTR;
2232         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2233                 rc = -ESHUTDOWN;
2234         else if (lp->lp_dc_error)
2235                 rc = lp->lp_dc_error;
2236         else if (!block)
2237                 CDEBUG(D_NET, "non-blocking discovery\n");
2238         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2239                 goto again;
2240
2241         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2242                (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
2243                libcfs_nid2str(lpni->lpni_nid), rc,
2244                (!block) ? "pending discovery" : "discovery complete");
2245
2246         return rc;
2247 }
2248
2249 /* Handle an incoming ack for a push. */
2250 static void
2251 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2252 {
2253         struct lnet_ping_buffer *pbuf;
2254
2255         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2256         spin_lock(&lp->lp_lock);
2257         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2258         lp->lp_push_error = ev->status;
2259         if (ev->status)
2260                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2261         else
2262                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2263         spin_unlock(&lp->lp_lock);
2264
2265         CDEBUG(D_NET, "peer %s ev->status %d\n",
2266                libcfs_nid2str(lp->lp_primary_nid), ev->status);
2267 }
2268
2269 /* Handle a Reply message. This is the reply to a Ping message. */
2270 static void
2271 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2272 {
2273         struct lnet_ping_buffer *pbuf;
2274         int rc;
2275
2276         spin_lock(&lp->lp_lock);
2277
2278         /*
2279          * If some kind of error happened the contents of message
2280          * cannot be used. Set PING_FAILED to trigger a retry.
2281          */
2282         if (ev->status) {
2283                 lp->lp_state |= LNET_PEER_PING_FAILED;
2284                 lp->lp_ping_error = ev->status;
2285                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2286                        ev->status,
2287                        libcfs_nid2str(lp->lp_primary_nid),
2288                        libcfs_nid2str(ev->source.nid));
2289                 goto out;
2290         }
2291
2292         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
2293         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2294                 lnet_swap_pinginfo(pbuf);
2295
2296         /*
2297          * A reply with invalid or corrupted info. Set PING_FAILED to
2298          * trigger a retry.
2299          */
2300         rc = lnet_ping_info_validate(&pbuf->pb_info);
2301         if (rc) {
2302                 lp->lp_state |= LNET_PEER_PING_FAILED;
2303                 lp->lp_ping_error = 0;
2304                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2305                        libcfs_nid2str(lp->lp_primary_nid), rc);
2306                 goto out;
2307         }
2308
2309
2310         /*
2311          * The peer may have discovery disabled at its end. Set
2312          * NO_DISCOVERY as appropriate.
2313          */
2314         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) &&
2315             !lnet_peer_discovery_disabled) {
2316                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2317                        libcfs_nid2str(lp->lp_primary_nid));
2318                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2319         } else {
2320                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2321                        libcfs_nid2str(lp->lp_primary_nid));
2322                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2323         }
2324
2325         /*
2326          * Update the MULTI_RAIL flag based on the reply. If the peer
2327          * was configured with DLC then the setting should match what
2328          * DLC put in.
2329          */
2330         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2331                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2332                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2333                                libcfs_nid2str(lp->lp_primary_nid), lp);
2334                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2335                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2336                               libcfs_nid2str(lp->lp_primary_nid));
2337                 } else if (lnet_peer_discovery_disabled) {
2338                         CDEBUG(D_NET,
2339                                "peer %s(%p) not MR: DD disabled locally\n",
2340                                libcfs_nid2str(lp->lp_primary_nid), lp);
2341                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2342                         CDEBUG(D_NET,
2343                                "peer %s(%p) not MR: DD disabled remotely\n",
2344                                libcfs_nid2str(lp->lp_primary_nid), lp);
2345                 } else {
2346                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2347                                libcfs_nid2str(lp->lp_primary_nid), lp);
2348                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2349                         lnet_peer_clr_non_mr_pref_nids(lp);
2350                 }
2351         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2352                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2353                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2354                               libcfs_nid2str(lp->lp_primary_nid));
2355                 } else {
2356                         CERROR("Multi-Rail state vanished from %s\n",
2357                                libcfs_nid2str(lp->lp_primary_nid));
2358                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2359                 }
2360         }
2361
2362         /*
2363          * Make sure we'll allocate the correct size ping buffer when
2364          * pinging the peer.
2365          */
2366         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2367                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2368
2369         /*
2370          * Check for truncation of the Reply. Clear PING_SENT and set
2371          * PING_FAILED to trigger a retry.
2372          */
2373         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2374                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2375                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2376                 lp->lp_state |= LNET_PEER_PING_FAILED;
2377                 lp->lp_ping_error = 0;
2378                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2379                        libcfs_nid2str(lp->lp_primary_nid),
2380                        pbuf->pb_info.pi_nnis);
2381                 goto out;
2382         }
2383
2384         /*
2385          * Check the sequence numbers in the reply. These are only
2386          * available if the reply came from a Multi-Rail peer.
2387          */
2388         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2389             pbuf->pb_info.pi_nnis > 1 &&
2390             lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
2391                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2392                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2393                                 libcfs_nid2str(lp->lp_primary_nid),
2394                                 LNET_PING_BUFFER_SEQNO(pbuf),
2395                                 lp->lp_peer_seqno);
2396
2397                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2398         }
2399
2400         /* We're happy with the state of the data in the buffer. */
2401         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2402                libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno, lp->lp_state);
2403         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2404                 lnet_ping_buffer_decref(lp->lp_data);
2405         else
2406                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2407         lnet_ping_buffer_addref(pbuf);
2408         lp->lp_data = pbuf;
2409 out:
2410         lp->lp_state &= ~LNET_PEER_PING_SENT;
2411         spin_unlock(&lp->lp_lock);
2412
2413         lnet_net_lock(LNET_LOCK_EX);
2414         /*
2415          * If this peer is a gateway, call the routing callback to
2416          * handle the ping reply
2417          */
2418         if (lp->lp_rtr_refcount > 0)
2419                 lnet_router_discovery_ping_reply(lp);
2420         lnet_net_unlock(LNET_LOCK_EX);
2421 }
2422
2423 /*
2424  * Send event handling. Only matters for error cases, where we clean
2425  * up state on the peer and peer_ni that would otherwise be updated in
2426  * the REPLY event handler for a successful Ping, and the ACK event
2427  * handler for a successful Push.
2428  */
2429 static int
2430 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2431 {
2432         int rc = 0;
2433
2434         if (!ev->status)
2435                 goto out;
2436
2437         spin_lock(&lp->lp_lock);
2438         if (ev->msg_type == LNET_MSG_GET) {
2439                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2440                 lp->lp_state |= LNET_PEER_PING_FAILED;
2441                 lp->lp_ping_error = ev->status;
2442         } else { /* ev->msg_type == LNET_MSG_PUT */
2443                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2444                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2445                 lp->lp_push_error = ev->status;
2446         }
2447         spin_unlock(&lp->lp_lock);
2448         rc = LNET_REDISCOVER_PEER;
2449 out:
2450         CDEBUG(D_NET, "%s Send to %s: %d\n",
2451                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2452                 libcfs_nid2str(ev->target.nid), rc);
2453         return rc;
2454 }
2455
2456 /*
2457  * Unlink event handling. This event is only seen if a call to
2458  * LNetMDUnlink() caused the event to be unlinked. If this call was
2459  * made after the event was set up in LNetGet() or LNetPut() then we
2460  * assume the Ping or Push timed out.
2461  */
2462 static void
2463 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2464 {
2465         spin_lock(&lp->lp_lock);
2466         /* We've passed through LNetGet() */
2467         if (lp->lp_state & LNET_PEER_PING_SENT) {
2468                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2469                 lp->lp_state |= LNET_PEER_PING_FAILED;
2470                 lp->lp_ping_error = -ETIMEDOUT;
2471                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2472                         libcfs_nid2str(lp->lp_primary_nid));
2473         }
2474         /* We've passed through LNetPut() */
2475         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2476                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2477                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2478                 lp->lp_push_error = -ETIMEDOUT;
2479                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2480                         libcfs_nid2str(lp->lp_primary_nid));
2481         }
2482         spin_unlock(&lp->lp_lock);
2483 }
2484
2485 /*
2486  * Event handler for the discovery EQ.
2487  *
2488  * Called with lnet_res_lock(cpt) held. The cpt is the
2489  * lnet_cpt_of_cookie() of the md handle cookie.
2490  */
2491 static void lnet_discovery_event_handler(struct lnet_event *event)
2492 {
2493         struct lnet_peer *lp = event->md.user_ptr;
2494         struct lnet_ping_buffer *pbuf;
2495         int rc;
2496
2497         /* discovery needs to take another look */
2498         rc = LNET_REDISCOVER_PEER;
2499
2500         CDEBUG(D_NET, "Received event: %d\n", event->type);
2501
2502         switch (event->type) {
2503         case LNET_EVENT_ACK:
2504                 lnet_discovery_event_ack(lp, event);
2505                 break;
2506         case LNET_EVENT_REPLY:
2507                 lnet_discovery_event_reply(lp, event);
2508                 break;
2509         case LNET_EVENT_SEND:
2510                 /* Only send failure triggers a retry. */
2511                 rc = lnet_discovery_event_send(lp, event);
2512                 break;
2513         case LNET_EVENT_UNLINK:
2514                 /* LNetMDUnlink() was called */
2515                 lnet_discovery_event_unlink(lp, event);
2516                 break;
2517         default:
2518                 /* Invalid events. */
2519                 LBUG();
2520         }
2521         lnet_net_lock(LNET_LOCK_EX);
2522         if (event->unlinked) {
2523                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
2524                 lnet_ping_buffer_decref(pbuf);
2525                 lnet_peer_decref_locked(lp);
2526         }
2527
2528         /* put peer back at end of request queue, if discovery not already
2529          * done */
2530         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp)) {
2531                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2532                 wake_up(&the_lnet.ln_dc_waitq);
2533         }
2534         lnet_net_unlock(LNET_LOCK_EX);
2535 }
2536
2537 /*
2538  * Build a peer from incoming data.
2539  *
2540  * The NIDs in the incoming data are supposed to be structured as follows:
2541  *  - loopback
2542  *  - primary NID
2543  *  - other NIDs in same net
2544  *  - NIDs in second net
2545  *  - NIDs in third net
2546  *  - ...
2547  * This due to the way the list of NIDs in the data is created.
2548  *
2549  * Note that this function will mark the peer uptodate unless an
2550  * ENOMEM is encontered. All other errors are due to a conflict
2551  * between the DLC configuration and what discovery sees. We treat DLC
2552  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2553  * peer from becoming stuck in discovery.
2554  */
2555 static int lnet_peer_merge_data(struct lnet_peer *lp,
2556                                 struct lnet_ping_buffer *pbuf)
2557 {
2558         struct lnet_peer_ni *lpni;
2559         lnet_nid_t *curnis = NULL;
2560         struct lnet_ni_status *addnis = NULL;
2561         lnet_nid_t *delnis = NULL;
2562         unsigned flags;
2563         int ncurnis;
2564         int naddnis;
2565         int ndelnis;
2566         int nnis = 0;
2567         int i;
2568         int j;
2569         int rc;
2570
2571         flags = LNET_PEER_DISCOVERED;
2572         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2573                 flags |= LNET_PEER_MULTI_RAIL;
2574
2575         /*
2576          * Cache the routing feature for the peer; whether it is enabled
2577          * for disabled as reported by the remote peer.
2578          */
2579         spin_lock(&lp->lp_lock);
2580         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2581                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2582         else
2583                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2584         spin_unlock(&lp->lp_lock);
2585
2586         nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
2587         LIBCFS_ALLOC(curnis, nnis * sizeof(*curnis));
2588         LIBCFS_ALLOC(addnis, nnis * sizeof(*addnis));
2589         LIBCFS_ALLOC(delnis, nnis * sizeof(*delnis));
2590         if (!curnis || !addnis || !delnis) {
2591                 rc = -ENOMEM;
2592                 goto out;
2593         }
2594         ncurnis = 0;
2595         naddnis = 0;
2596         ndelnis = 0;
2597
2598         /* Construct the list of NIDs present in peer. */
2599         lpni = NULL;
2600         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2601                 curnis[ncurnis++] = lpni->lpni_nid;
2602
2603         /*
2604          * Check for NIDs in pbuf not present in curnis[].
2605          * The loop starts at 1 to skip the loopback NID.
2606          */
2607         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2608                 for (j = 0; j < ncurnis; j++)
2609                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2610                                 break;
2611                 if (j == ncurnis)
2612                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2613         }
2614         /*
2615          * Check for NIDs in curnis[] not present in pbuf.
2616          * The nested loop starts at 1 to skip the loopback NID.
2617          *
2618          * But never add the loopback NID to delnis[]: if it is
2619          * present in curnis[] then this peer is for this node.
2620          */
2621         for (i = 0; i < ncurnis; i++) {
2622                 if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
2623                         continue;
2624                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2625                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2626                                 /*
2627                                  * update the information we cache for the
2628                                  * peer with the latest information we
2629                                  * received
2630                                  */
2631                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
2632                                 if (lpni) {
2633                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2634                                         lnet_peer_ni_decref_locked(lpni);
2635                                 }
2636                                 break;
2637                         }
2638                 }
2639                 if (j == pbuf->pb_info.pi_nnis)
2640                         delnis[ndelnis++] = curnis[i];
2641         }
2642
2643         /*
2644          * If we get here and the discovery is disabled then we don't want
2645          * to add or delete any NIs. We just updated the ones we have some
2646          * information on, and call it a day
2647          */
2648         rc = 0;
2649         if (lnet_is_discovery_disabled(lp))
2650                 goto out;
2651
2652         for (i = 0; i < naddnis; i++) {
2653                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
2654                 if (rc) {
2655                         CERROR("Error adding NID %s to peer %s: %d\n",
2656                                libcfs_nid2str(addnis[i].ns_nid),
2657                                libcfs_nid2str(lp->lp_primary_nid), rc);
2658                         if (rc == -ENOMEM)
2659                                 goto out;
2660                 }
2661                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
2662                 if (lpni) {
2663                         lpni->lpni_ns_status = addnis[i].ns_status;
2664                         lnet_peer_ni_decref_locked(lpni);
2665                 }
2666         }
2667
2668         for (i = 0; i < ndelnis; i++) {
2669                 /*
2670                  * for routers it's okay to delete the primary_nid because
2671                  * the upper layers don't really rely on it. So if we're
2672                  * being told that the router changed its primary_nid
2673                  * then it's okay to delete it.
2674                  */
2675                 if (lp->lp_rtr_refcount > 0)
2676                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
2677                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
2678                 if (rc) {
2679                         CERROR("Error deleting NID %s from peer %s: %d\n",
2680                                libcfs_nid2str(delnis[i]),
2681                                libcfs_nid2str(lp->lp_primary_nid), rc);
2682                         if (rc == -ENOMEM)
2683                                 goto out;
2684                 }
2685         }
2686         /*
2687          * Errors other than -ENOMEM are due to peers having been
2688          * configured with DLC. Ignore these because DLC overrides
2689          * Discovery.
2690          */
2691         rc = 0;
2692 out:
2693         LIBCFS_FREE(curnis, nnis * sizeof(*curnis));
2694         LIBCFS_FREE(addnis, nnis * sizeof(*addnis));
2695         LIBCFS_FREE(delnis, nnis * sizeof(*delnis));
2696         lnet_ping_buffer_decref(pbuf);
2697         CDEBUG(D_NET, "peer %s (%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
2698
2699         if (rc) {
2700                 spin_lock(&lp->lp_lock);
2701                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2702                 lp->lp_state |= LNET_PEER_FORCE_PING;
2703                 spin_unlock(&lp->lp_lock);
2704         }
2705         return rc;
2706 }
2707
2708 /*
2709  * The data in pbuf says lp is its primary peer, but the data was
2710  * received by a different peer. Try to update lp with the data.
2711  */
2712 static int
2713 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
2714 {
2715         struct lnet_handle_md mdh;
2716
2717         /* Queue lp for discovery, and force it on the request queue. */
2718         lnet_net_lock(LNET_LOCK_EX);
2719         if (lnet_peer_queue_for_discovery(lp))
2720                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2721         lnet_net_unlock(LNET_LOCK_EX);
2722
2723         LNetInvalidateMDHandle(&mdh);
2724
2725         /*
2726          * Decide whether we can move the peer to the DATA_PRESENT state.
2727          *
2728          * We replace stale data for a multi-rail peer, repair PING_FAILED
2729          * status, and preempt FORCE_PING.
2730          *
2731          * If after that we have DATA_PRESENT, we merge it into this peer.
2732          */
2733         spin_lock(&lp->lp_lock);
2734         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2735                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
2736                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2737                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2738                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2739                         lnet_ping_buffer_decref(pbuf);
2740                         pbuf = lp->lp_data;
2741                         lp->lp_data = NULL;
2742                 }
2743         }
2744         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2745                 lnet_ping_buffer_decref(lp->lp_data);
2746                 lp->lp_data = NULL;
2747                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2748         }
2749         if (lp->lp_state & LNET_PEER_PING_FAILED) {
2750                 mdh = lp->lp_ping_mdh;
2751                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2752                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
2753                 lp->lp_ping_error = 0;
2754         }
2755         if (lp->lp_state & LNET_PEER_FORCE_PING)
2756                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
2757         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2758         spin_unlock(&lp->lp_lock);
2759
2760         if (!LNetMDHandleIsInvalid(mdh))
2761                 LNetMDUnlink(mdh);
2762
2763         if (pbuf)
2764                 return lnet_peer_merge_data(lp, pbuf);
2765
2766         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
2767         return 0;
2768 }
2769
2770 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
2771 {
2772         int i;
2773
2774         for (i = 0; i < pinfo->pi_nnis; i++) {
2775                 if (pinfo->pi_ni[i].ns_nid == nid)
2776                         return true;
2777         }
2778
2779         return false;
2780 }
2781
2782 /*
2783  * Update a peer using the data received.
2784  */
2785 static int lnet_peer_data_present(struct lnet_peer *lp)
2786 __must_hold(&lp->lp_lock)
2787 {
2788         struct lnet_ping_buffer *pbuf;
2789         struct lnet_peer_ni *lpni;
2790         lnet_nid_t nid = LNET_NID_ANY;
2791         unsigned flags;
2792         int rc = 0;
2793
2794         pbuf = lp->lp_data;
2795         lp->lp_data = NULL;
2796         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
2797         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
2798         spin_unlock(&lp->lp_lock);
2799
2800         /*
2801          * Modifications of peer structures are done while holding the
2802          * ln_api_mutex. A global lock is required because we may be
2803          * modifying multiple peer structures, and a mutex greatly
2804          * simplifies memory management.
2805          *
2806          * The actual changes to the data structures must also protect
2807          * against concurrent lookups, for which the lnet_net_lock in
2808          * LNET_LOCK_EX mode is used.
2809          */
2810         mutex_lock(&the_lnet.ln_api_mutex);
2811         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2812                 rc = -ESHUTDOWN;
2813                 goto out;
2814         }
2815
2816         /*
2817          * If this peer is not on the peer list then it is being torn
2818          * down, and our reference count may be all that is keeping it
2819          * alive. Don't do any work on it.
2820          */
2821         if (list_empty(&lp->lp_peer_list))
2822                 goto out;
2823
2824         flags = LNET_PEER_DISCOVERED;
2825         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2826                 flags |= LNET_PEER_MULTI_RAIL;
2827
2828         /*
2829          * Check whether the primary NID in the message matches the
2830          * primary NID of the peer. If it does, update the peer, if
2831          * it it does not, check whether there is already a peer with
2832          * that primary NID. If no such peer exists, try to update
2833          * the primary NID of the current peer (allowed if it was
2834          * created due to message traffic) and complete the update.
2835          * If the peer did exist, hand off the data to it.
2836          *
2837          * The peer for the loopback interface is a special case: this
2838          * is the peer for the local node, and we want to set its
2839          * primary NID to the correct value here. Moreover, this peer
2840          * can show up with only the loopback NID in the ping buffer.
2841          */
2842         if (pbuf->pb_info.pi_nnis <= 1)
2843                 goto out;
2844         nid = pbuf->pb_info.pi_ni[1].ns_nid;
2845         if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
2846                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
2847                 if (!rc)
2848                         rc = lnet_peer_merge_data(lp, pbuf);
2849         /*
2850          * if the primary nid of the peer is present in the ping info returned
2851          * from the peer, but it's not the local primary peer we have
2852          * cached and discovery is disabled, then we don't want to update
2853          * our local peer info, by adding or removing NIDs, we just want
2854          * to update the status of the nids that we currently have
2855          * recorded in that peer.
2856          */
2857         } else if (lp->lp_primary_nid == nid ||
2858                    (lnet_is_nid_in_ping_info(lp->lp_primary_nid, &pbuf->pb_info) &&
2859                     lnet_is_discovery_disabled(lp))) {
2860                 rc = lnet_peer_merge_data(lp, pbuf);
2861         } else {
2862                 lpni = lnet_find_peer_ni_locked(nid);
2863                 if (!lpni) {
2864                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
2865                         if (rc) {
2866                                 CERROR("Primary NID error %s versus %s: %d\n",
2867                                        libcfs_nid2str(lp->lp_primary_nid),
2868                                        libcfs_nid2str(nid), rc);
2869                         } else {
2870                                 rc = lnet_peer_merge_data(lp, pbuf);
2871                         }
2872                 } else {
2873                         struct lnet_peer *new_lp;
2874                         new_lp = lpni->lpni_peer_net->lpn_peer;
2875                         /*
2876                          * if lp has discovery/MR enabled that means new_lp
2877                          * should have discovery/MR enabled as well, since
2878                          * it's the same peer, which we're about to merge
2879                          */
2880                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
2881                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2882                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2883                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
2884
2885                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
2886                         lnet_consolidate_routes_locked(lp, new_lp);
2887                         lnet_peer_ni_decref_locked(lpni);
2888                 }
2889         }
2890 out:
2891         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc,
2892                lp->lp_state);
2893         mutex_unlock(&the_lnet.ln_api_mutex);
2894
2895         spin_lock(&lp->lp_lock);
2896         /* Tell discovery to re-check the peer immediately. */
2897         if (!rc)
2898                 rc = LNET_REDISCOVER_PEER;
2899         return rc;
2900 }
2901
2902 /*
2903  * A ping failed. Clear the PING_FAILED state and set the
2904  * FORCE_PING state, to ensure a retry even if discovery is
2905  * disabled. This avoids being left with incorrect state.
2906  */
2907 static int lnet_peer_ping_failed(struct lnet_peer *lp)
2908 __must_hold(&lp->lp_lock)
2909 {
2910         struct lnet_handle_md mdh;
2911         int rc;
2912
2913         mdh = lp->lp_ping_mdh;
2914         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
2915         lp->lp_state &= ~LNET_PEER_PING_FAILED;
2916         lp->lp_state |= LNET_PEER_FORCE_PING;
2917         rc = lp->lp_ping_error;
2918         lp->lp_ping_error = 0;
2919         spin_unlock(&lp->lp_lock);
2920
2921         if (!LNetMDHandleIsInvalid(mdh))
2922                 LNetMDUnlink(mdh);
2923
2924         CDEBUG(D_NET, "peer %s:%d\n",
2925                libcfs_nid2str(lp->lp_primary_nid), rc);
2926
2927         spin_lock(&lp->lp_lock);
2928         return rc ? rc : LNET_REDISCOVER_PEER;
2929 }
2930
2931 /*
2932  * Select NID to send a Ping or Push to.
2933  */
2934 static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
2935 {
2936         struct lnet_peer_ni *lpni;
2937
2938         /* Look for a direct-connected NID for this peer. */
2939         lpni = NULL;
2940         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2941                 if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
2942                         continue;
2943                 break;
2944         }
2945         if (lpni)
2946                 return lpni->lpni_nid;
2947
2948         /* Look for a routed-connected NID for this peer. */
2949         lpni = NULL;
2950         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
2951                 if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
2952                         continue;
2953                 break;
2954         }
2955         if (lpni)
2956                 return lpni->lpni_nid;
2957
2958         return LNET_NID_ANY;
2959 }
2960
2961 /* Active side of ping. */
2962 static int lnet_peer_send_ping(struct lnet_peer *lp)
2963 __must_hold(&lp->lp_lock)
2964 {
2965         lnet_nid_t pnid;
2966         int nnis;
2967         int rc;
2968         int cpt;
2969
2970         lp->lp_state |= LNET_PEER_PING_SENT;
2971         lp->lp_state &= ~LNET_PEER_FORCE_PING;
2972         spin_unlock(&lp->lp_lock);
2973
2974         cpt = lnet_net_lock_current();
2975         /* Refcount for MD. */
2976         lnet_peer_addref_locked(lp);
2977         pnid = lnet_peer_select_nid(lp);
2978         lnet_net_unlock(cpt);
2979
2980         nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
2981
2982         rc = lnet_send_ping(pnid, &lp->lp_ping_mdh, nnis, lp,
2983                             the_lnet.ln_dc_eqh, false);
2984
2985         /*
2986          * if LNetMDBind in lnet_send_ping fails we need to decrement the
2987          * refcount on the peer, otherwise LNetMDUnlink will be called
2988          * which will eventually do that.
2989          */
2990         if (rc > 0) {
2991                 lnet_net_lock(cpt);
2992                 lnet_peer_decref_locked(lp);
2993                 lnet_net_unlock(cpt);
2994                 rc = -rc; /* change the rc to negative value */
2995                 goto fail_error;
2996         } else if (rc < 0) {
2997                 goto fail_error;
2998         }
2999
3000         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3001
3002         spin_lock(&lp->lp_lock);
3003         return 0;
3004
3005 fail_error:
3006         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
3007         /*
3008          * The errors that get us here are considered hard errors and
3009          * cause Discovery to terminate. So we clear PING_SENT, but do
3010          * not set either PING_FAILED or FORCE_PING. In fact we need
3011          * to clear PING_FAILED, because the unlink event handler will
3012          * have set it if we called LNetMDUnlink() above.
3013          */
3014         spin_lock(&lp->lp_lock);
3015         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3016         return rc;
3017 }
3018
3019 /*
3020  * This function exists because you cannot call LNetMDUnlink() from an
3021  * event handler.
3022  */
3023 static int lnet_peer_push_failed(struct lnet_peer *lp)
3024 __must_hold(&lp->lp_lock)
3025 {
3026         struct lnet_handle_md mdh;
3027         int rc;
3028
3029         mdh = lp->lp_push_mdh;
3030         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3031         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3032         rc = lp->lp_push_error;
3033         lp->lp_push_error = 0;
3034         spin_unlock(&lp->lp_lock);
3035
3036         if (!LNetMDHandleIsInvalid(mdh))
3037                 LNetMDUnlink(mdh);
3038
3039         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3040         spin_lock(&lp->lp_lock);
3041         return rc ? rc : LNET_REDISCOVER_PEER;
3042 }
3043
3044 /* Active side of push. */
3045 static int lnet_peer_send_push(struct lnet_peer *lp)
3046 __must_hold(&lp->lp_lock)
3047 {
3048         struct lnet_ping_buffer *pbuf;
3049         struct lnet_process_id id;
3050         struct lnet_md md;
3051         int cpt;
3052         int rc;
3053
3054         /* Don't push to a non-multi-rail peer. */
3055         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3056                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3057                 return 0;
3058         }
3059
3060         lp->lp_state |= LNET_PEER_PUSH_SENT;
3061         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3062         spin_unlock(&lp->lp_lock);
3063
3064         cpt = lnet_net_lock_current();
3065         pbuf = the_lnet.ln_ping_target;
3066         lnet_ping_buffer_addref(pbuf);
3067         lnet_net_unlock(cpt);
3068
3069         /* Push source MD */
3070         md.start     = &pbuf->pb_info;
3071         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3072         md.threshold = 2; /* Put/Ack */
3073         md.max_size  = 0;
3074         md.options   = 0;
3075         md.eq_handle = the_lnet.ln_dc_eqh;
3076         md.user_ptr  = lp;
3077
3078         rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
3079         if (rc) {
3080                 lnet_ping_buffer_decref(pbuf);
3081                 CERROR("Can't bind push source MD: %d\n", rc);
3082                 goto fail_error;
3083         }
3084         cpt = lnet_net_lock_current();
3085         /* Refcount for MD. */
3086         lnet_peer_addref_locked(lp);
3087         id.pid = LNET_PID_LUSTRE;
3088         id.nid = lnet_peer_select_nid(lp);
3089         lnet_net_unlock(cpt);
3090
3091         if (id.nid == LNET_NID_ANY) {
3092                 rc = -EHOSTUNREACH;
3093                 goto fail_unlink;
3094         }
3095
3096         rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
3097                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3098                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3099
3100         if (rc)
3101                 goto fail_unlink;
3102
3103         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3104
3105         spin_lock(&lp->lp_lock);
3106         return 0;
3107
3108 fail_unlink:
3109         LNetMDUnlink(lp->lp_push_mdh);
3110         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3111 fail_error:
3112         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nid2str(lp->lp_primary_nid), lp, rc);
3113         /*
3114          * The errors that get us here are considered hard errors and
3115          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3116          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3117          * because the unlink event handler will have set it if we
3118          * called LNetMDUnlink() above.
3119          */
3120         spin_lock(&lp->lp_lock);
3121         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3122         return rc;
3123 }
3124
3125 /*
3126  * An unrecoverable error was encountered during discovery.
3127  * Set error status in peer and abort discovery.
3128  */
3129 static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
3130 {
3131         CDEBUG(D_NET, "Discovery error %s: %d\n",
3132                libcfs_nid2str(lp->lp_primary_nid), error);
3133
3134         spin_lock(&lp->lp_lock);
3135         lp->lp_dc_error = error;
3136         lp->lp_state &= ~LNET_PEER_DISCOVERING;
3137         lp->lp_state |= LNET_PEER_REDISCOVER;
3138         spin_unlock(&lp->lp_lock);
3139 }
3140
3141 /*
3142  * Mark the peer as discovered.
3143  */
3144 static int lnet_peer_discovered(struct lnet_peer *lp)
3145 __must_hold(&lp->lp_lock)
3146 {
3147         lp->lp_state |= LNET_PEER_DISCOVERED;
3148         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3149                           LNET_PEER_REDISCOVER);
3150
3151         CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
3152
3153         return 0;
3154 }
3155
3156
3157 /*
3158  * Discovering this peer is taking too long. Cancel any Ping or Push
3159  * that discovery is waiting on by unlinking the relevant MDs. The
3160  * lnet_discovery_event_handler() will proceed from here and complete
3161  * the cleanup.
3162  */
3163 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
3164 {
3165         struct lnet_handle_md ping_mdh;
3166         struct lnet_handle_md push_mdh;
3167
3168         LNetInvalidateMDHandle(&ping_mdh);
3169         LNetInvalidateMDHandle(&push_mdh);
3170
3171         spin_lock(&lp->lp_lock);
3172         if (lp->lp_state & LNET_PEER_PING_SENT) {
3173                 ping_mdh = lp->lp_ping_mdh;
3174                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3175         }
3176         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
3177                 push_mdh = lp->lp_push_mdh;
3178                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
3179         }
3180         spin_unlock(&lp->lp_lock);
3181
3182         if (!LNetMDHandleIsInvalid(ping_mdh))
3183                 LNetMDUnlink(ping_mdh);
3184         if (!LNetMDHandleIsInvalid(push_mdh))
3185                 LNetMDUnlink(push_mdh);
3186 }
3187
3188 /*
3189  * Wait for work to be queued or some other change that must be
3190  * attended to. Returns non-zero if the discovery thread should shut
3191  * down.
3192  */
3193 static int lnet_peer_discovery_wait_for_work(void)
3194 {
3195         int cpt;
3196         int rc = 0;
3197
3198         DEFINE_WAIT(wait);
3199
3200         cpt = lnet_net_lock_current();
3201         for (;;) {
3202                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3203                                 TASK_INTERRUPTIBLE);
3204                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3205                         break;
3206                 if (lnet_push_target_resize_needed())
3207                         break;
3208                 if (!list_empty(&the_lnet.ln_dc_request))
3209                         break;
3210                 if (!list_empty(&the_lnet.ln_msg_resend))
3211                         break;
3212                 lnet_net_unlock(cpt);
3213
3214                 /*
3215                  * wakeup max every second to check if there are peers that
3216                  * have been stuck on the working queue for greater than
3217                  * the peer timeout.
3218                  */
3219                 schedule_timeout(cfs_time_seconds(1));
3220                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3221                 cpt = lnet_net_lock_current();
3222         }
3223         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3224
3225         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3226                 rc = -ESHUTDOWN;
3227
3228         lnet_net_unlock(cpt);
3229
3230         CDEBUG(D_NET, "woken: %d\n", rc);
3231
3232         return rc;
3233 }
3234
3235 /*
3236  * Messages that were pending on a destroyed peer will be put on a global
3237  * resend list. The message resend list will be checked by
3238  * the discovery thread when it wakes up, and will resend messages. These
3239  * messages can still be sendable in the case the lpni which was the initial
3240  * cause of the message re-queue was transfered to another peer.
3241  *
3242  * It is possible that LNet could be shutdown while we're iterating
3243  * through the list. lnet_shudown_lndnets() will attempt to access the
3244  * resend list, but will have to wait until the spinlock is released, by
3245  * which time there shouldn't be any more messages on the resend list.
3246  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3247  * for the messages so they can be released. The other case is that
3248  * lnet_shudown_lndnets() can finalize all the messages before this
3249  * function can visit the resend list, in which case this function will be
3250  * a no-op.
3251  */
3252 static void lnet_resend_msgs(void)
3253 {
3254         struct lnet_msg *msg, *tmp;
3255         LIST_HEAD(resend);
3256         int rc;
3257
3258         spin_lock(&the_lnet.ln_msg_resend_lock);
3259         list_splice(&the_lnet.ln_msg_resend, &resend);
3260         spin_unlock(&the_lnet.ln_msg_resend_lock);
3261
3262         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3263                 list_del_init(&msg->msg_list);
3264                 rc = lnet_send(msg->msg_src_nid_param, msg,
3265                                msg->msg_rtr_nid_param);
3266                 if (rc < 0) {
3267                         CNETERR("Error sending %s to %s: %d\n",
3268                                lnet_msgtyp2str(msg->msg_type),
3269                                libcfs_id2str(msg->msg_target), rc);
3270                         lnet_finalize(msg, rc);
3271                 }
3272         }
3273 }
3274
3275 /* The discovery thread. */
3276 static int lnet_peer_discovery(void *arg)
3277 {
3278         struct lnet_peer *lp;
3279         int rc;
3280
3281         wait_for_completion(&the_lnet.ln_started);
3282
3283         CDEBUG(D_NET, "started\n");
3284         cfs_block_allsigs();
3285
3286         for (;;) {
3287                 if (lnet_peer_discovery_wait_for_work())
3288                         break;
3289
3290                 lnet_resend_msgs();
3291
3292                 if (lnet_push_target_resize_needed())
3293                         lnet_push_target_resize();
3294
3295                 lnet_net_lock(LNET_LOCK_EX);
3296                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3297                         break;
3298
3299                 /*
3300                  * Process all incoming discovery work requests.  When
3301                  * discovery must wait on a peer to change state, it
3302                  * is added to the tail of the ln_dc_working queue. A
3303                  * timestamp keeps track of when the peer was added,
3304                  * so we can time out discovery requests that take too
3305                  * long.
3306                  */
3307                 while (!list_empty(&the_lnet.ln_dc_request)) {
3308                         lp = list_first_entry(&the_lnet.ln_dc_request,
3309                                               struct lnet_peer, lp_dc_list);
3310                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3311                         /*
3312                          * set the time the peer was put on the dc_working
3313                          * queue. It shouldn't remain on the queue
3314                          * forever, in case the GET message (for ping)
3315                          * doesn't get a REPLY or the PUT message (for
3316                          * push) doesn't get an ACK.
3317                          */
3318                         lp->lp_last_queued = ktime_get_real_seconds();
3319                         lnet_net_unlock(LNET_LOCK_EX);
3320
3321                         /*
3322                          * Select an action depending on the state of
3323                          * the peer and whether discovery is disabled.
3324                          * The check whether discovery is disabled is
3325                          * done after the code that handles processing
3326                          * for arrived data, cleanup for failures, and
3327                          * forcing a Ping or Push.
3328                          */
3329                         spin_lock(&lp->lp_lock);
3330                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3331                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3332                                 lp->lp_state);
3333                         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3334                                 rc = lnet_peer_data_present(lp);
3335                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3336                                 rc = lnet_peer_ping_failed(lp);
3337                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3338                                 rc = lnet_peer_push_failed(lp);
3339                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3340                                 rc = lnet_peer_send_ping(lp);
3341                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3342                                 rc = lnet_peer_send_push(lp);
3343                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3344                                 rc = lnet_peer_send_ping(lp);
3345                         else if (lnet_peer_needs_push(lp))
3346                                 rc = lnet_peer_send_push(lp);
3347                         else
3348                                 rc = lnet_peer_discovered(lp);
3349                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3350                                 libcfs_nid2str(lp->lp_primary_nid), lp,
3351                                 lp->lp_state, rc);
3352                         spin_unlock(&lp->lp_lock);
3353
3354                         lnet_net_lock(LNET_LOCK_EX);
3355                         if (rc == LNET_REDISCOVER_PEER) {
3356                                 list_move(&lp->lp_dc_list,
3357                                           &the_lnet.ln_dc_request);
3358                         } else if (rc) {
3359                                 lnet_peer_discovery_error(lp, rc);
3360                         }
3361                         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
3362                                 lnet_peer_discovery_complete(lp);
3363                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3364                                 break;
3365                 }
3366
3367                 lnet_net_unlock(LNET_LOCK_EX);
3368         }
3369
3370         CDEBUG(D_NET, "stopping\n");
3371         /*
3372          * Clean up before telling lnet_peer_discovery_stop() that
3373          * we're done. Use wake_up() below to somewhat reduce the
3374          * size of the thundering herd if there are multiple threads
3375          * waiting on discovery of a single peer.
3376          */
3377
3378         /* Queue cleanup 1: stop all pending pings and pushes. */
3379         lnet_net_lock(LNET_LOCK_EX);
3380         while (!list_empty(&the_lnet.ln_dc_working)) {
3381                 lp = list_first_entry(&the_lnet.ln_dc_working,
3382                                       struct lnet_peer, lp_dc_list);
3383                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3384                 lnet_net_unlock(LNET_LOCK_EX);
3385                 lnet_peer_cancel_discovery(lp);
3386                 lnet_net_lock(LNET_LOCK_EX);
3387         }
3388         lnet_net_unlock(LNET_LOCK_EX);
3389
3390         /* Queue cleanup 2: wait for the expired queue to clear. */
3391         while (!list_empty(&the_lnet.ln_dc_expired))
3392                 schedule_timeout(cfs_time_seconds(1));
3393
3394         /* Queue cleanup 3: clear the request queue. */
3395         lnet_net_lock(LNET_LOCK_EX);
3396         while (!list_empty(&the_lnet.ln_dc_request)) {
3397                 lp = list_first_entry(&the_lnet.ln_dc_request,
3398                                       struct lnet_peer, lp_dc_list);
3399                 lnet_peer_discovery_error(lp, -ESHUTDOWN);
3400                 lnet_peer_discovery_complete(lp);
3401         }
3402         lnet_net_unlock(LNET_LOCK_EX);
3403
3404         LNetEQFree(the_lnet.ln_dc_eqh);
3405         LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3406
3407         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3408         wake_up(&the_lnet.ln_dc_waitq);
3409
3410         CDEBUG(D_NET, "stopped\n");
3411
3412         return 0;
3413 }
3414
3415 /* ln_api_mutex is held on entry. */
3416 int lnet_peer_discovery_start(void)
3417 {
3418         struct task_struct *task;
3419         int rc;
3420
3421         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3422                 return -EALREADY;
3423
3424         rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
3425         if (rc != 0) {
3426                 CERROR("Can't allocate discovery EQ: %d\n", rc);
3427                 return rc;
3428         }
3429
3430         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3431         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3432         if (IS_ERR(task)) {
3433                 rc = PTR_ERR(task);
3434                 CERROR("Can't start peer discovery thread: %d\n", rc);
3435
3436                 LNetEQFree(the_lnet.ln_dc_eqh);
3437                 LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
3438
3439                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3440         }
3441
3442         CDEBUG(D_NET, "discovery start: %d\n", rc);
3443
3444         return rc;
3445 }
3446
3447 /* ln_api_mutex is held on entry. */
3448 void lnet_peer_discovery_stop(void)
3449 {
3450         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3451                 return;
3452
3453         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3454         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3455
3456         /* In the LNetNIInit() path we may be stopping discovery before it
3457          * entered its work loop
3458          */
3459         if (!completion_done(&the_lnet.ln_started))
3460                 complete(&the_lnet.ln_started);
3461         else
3462                 wake_up(&the_lnet.ln_dc_waitq);
3463
3464         wait_event(the_lnet.ln_dc_waitq,
3465                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3466
3467         LASSERT(list_empty(&the_lnet.ln_dc_request));
3468         LASSERT(list_empty(&the_lnet.ln_dc_working));
3469         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3470
3471         CDEBUG(D_NET, "discovery stopped\n");
3472 }
3473
3474 /* Debugging */
3475
3476 void
3477 lnet_debug_peer(lnet_nid_t nid)
3478 {
3479         char                    *aliveness = "NA";
3480         struct lnet_peer_ni     *lp;
3481         int                     cpt;
3482
3483         cpt = lnet_cpt_of_nid(nid, NULL);
3484         lnet_net_lock(cpt);
3485
3486         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3487         if (IS_ERR(lp)) {
3488                 lnet_net_unlock(cpt);
3489                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3490                 return;
3491         }
3492
3493         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3494                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3495
3496         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3497                libcfs_nid2str(lp->lpni_nid), atomic_read(&lp->lpni_refcount),
3498                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3499                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3500                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3501
3502         lnet_peer_ni_decref_locked(lp);
3503
3504         lnet_net_unlock(cpt);
3505 }
3506
3507 /* Gathering information for userspace. */
3508
3509 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3510                           char aliveness[LNET_MAX_STR_LEN],
3511                           __u32 *cpt_iter, __u32 *refcount,
3512                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3513                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3514                           __u32 *peer_tx_qnob)
3515 {
3516         struct lnet_peer_table          *peer_table;
3517         struct lnet_peer_ni             *lp;
3518         int                             j;
3519         int                             lncpt;
3520         bool                            found = false;
3521
3522         /* get the number of CPTs */
3523         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3524
3525         /* if the cpt number to be examined is >= the number of cpts in
3526          * the system then indicate that there are no more cpts to examin
3527          */
3528         if (*cpt_iter >= lncpt)
3529                 return -ENOENT;
3530
3531         /* get the current table */
3532         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3533         /* if the ptable is NULL then there are no more cpts to examine */
3534         if (peer_table == NULL)
3535                 return -ENOENT;
3536
3537         lnet_net_lock(*cpt_iter);
3538
3539         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3540                 struct list_head *peers = &peer_table->pt_hash[j];
3541
3542                 list_for_each_entry(lp, peers, lpni_hashlist) {
3543                         if (peer_index-- > 0)
3544                                 continue;
3545
3546                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3547                         if (lnet_isrouter(lp) ||
3548                                 lnet_peer_aliveness_enabled(lp))
3549                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3550                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3551
3552                         *nid = lp->lpni_nid;
3553                         *refcount = atomic_read(&lp->lpni_refcount);
3554                         *ni_peer_tx_credits =
3555                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3556                         *peer_tx_credits = lp->lpni_txcredits;
3557                         *peer_rtr_credits = lp->lpni_rtrcredits;
3558                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3559                         *peer_tx_qnob = lp->lpni_txqnob;
3560
3561                         found = true;
3562                 }
3563
3564         }
3565         lnet_net_unlock(*cpt_iter);
3566
3567         *cpt_iter = lncpt;
3568
3569         return found ? 0 : -ENOENT;
3570 }
3571
3572 /* ln_api_mutex is held, which keeps the peer list stable */
3573 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
3574 {
3575         struct lnet_ioctl_element_stats *lpni_stats;
3576         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
3577         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
3578         struct lnet_peer_ni_credit_info *lpni_info;
3579         struct lnet_peer_ni *lpni;
3580         struct lnet_peer *lp;
3581         lnet_nid_t nid;
3582         __u32 size;
3583         int rc;
3584
3585         lp = lnet_find_peer(cfg->prcfg_prim_nid);
3586
3587         if (!lp) {
3588                 rc = -ENOENT;
3589                 goto out;
3590         }
3591
3592         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
3593                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
3594         size *= lp->lp_nnis;
3595         if (size > cfg->prcfg_size) {
3596                 cfg->prcfg_size = size;
3597                 rc = -E2BIG;
3598                 goto out_lp_decref;
3599         }
3600
3601         cfg->prcfg_prim_nid = lp->lp_primary_nid;
3602         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
3603         cfg->prcfg_cfg_nid = lp->lp_primary_nid;
3604         cfg->prcfg_count = lp->lp_nnis;
3605         cfg->prcfg_size = size;
3606         cfg->prcfg_state = lp->lp_state;
3607
3608         /* Allocate helper buffers. */
3609         rc = -ENOMEM;
3610         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
3611         if (!lpni_info)
3612                 goto out_lp_decref;
3613         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
3614         if (!lpni_stats)
3615                 goto out_free_info;
3616         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
3617         if (!lpni_msg_stats)
3618                 goto out_free_stats;
3619         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
3620         if (!lpni_hstats)
3621                 goto out_free_msg_stats;
3622
3623
3624         lpni = NULL;
3625         rc = -EFAULT;
3626         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
3627                 nid = lpni->lpni_nid;
3628                 if (copy_to_user(bulk, &nid, sizeof(nid)))
3629                         goto out_free_hstats;
3630                 bulk += sizeof(nid);
3631
3632                 memset(lpni_info, 0, sizeof(*lpni_info));
3633                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
3634                 if (lnet_isrouter(lpni) ||
3635                         lnet_peer_aliveness_enabled(lpni))
3636                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
3637                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
3638
3639                 lpni_info->cr_refcount = atomic_read(&lpni->lpni_refcount);
3640                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
3641                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
3642                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
3643                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
3644                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
3645                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
3646                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
3647                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
3648                         goto out_free_hstats;
3649                 bulk += sizeof(*lpni_info);
3650
3651                 memset(lpni_stats, 0, sizeof(*lpni_stats));
3652                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
3653                                                             LNET_STATS_TYPE_SEND);
3654                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
3655                                                             LNET_STATS_TYPE_RECV);
3656                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
3657                                                             LNET_STATS_TYPE_DROP);
3658                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
3659                         goto out_free_hstats;
3660                 bulk += sizeof(*lpni_stats);
3661                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
3662                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
3663                         goto out_free_hstats;
3664                 bulk += sizeof(*lpni_msg_stats);
3665                 lpni_hstats->hlpni_network_timeout =
3666                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
3667                 lpni_hstats->hlpni_remote_dropped =
3668                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
3669                 lpni_hstats->hlpni_remote_timeout =
3670                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
3671                 lpni_hstats->hlpni_remote_error =
3672                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
3673                 lpni_hstats->hlpni_health_value =
3674                   atomic_read(&lpni->lpni_healthv);
3675                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
3676                         goto out_free_hstats;
3677                 bulk += sizeof(*lpni_hstats);
3678         }
3679         rc = 0;
3680
3681 out_free_hstats:
3682         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
3683 out_free_msg_stats:
3684         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
3685 out_free_stats:
3686         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
3687 out_free_info:
3688         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
3689 out_lp_decref:
3690         lnet_peer_decref_locked(lp);
3691 out:
3692         return rc;
3693 }
3694
3695 void
3696 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni)
3697 {
3698         /* the mt could've shutdown and cleaned up the queues */
3699         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
3700                 return;
3701
3702         if (list_empty(&lpni->lpni_recovery) &&
3703             atomic_read(&lpni->lpni_healthv) < LNET_MAX_HEALTH_VALUE) {
3704                 CERROR("lpni %s added to recovery queue. Health = %d\n",
3705                         libcfs_nid2str(lpni->lpni_nid),
3706                         atomic_read(&lpni->lpni_healthv));
3707                 list_add_tail(&lpni->lpni_recovery, &the_lnet.ln_mt_peerNIRecovq);
3708                 lnet_peer_ni_addref_locked(lpni);
3709         }
3710 }
3711
3712 /* Call with the ln_api_mutex held */
3713 void
3714 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3715 {
3716         struct lnet_peer_table *ptable;
3717         struct lnet_peer *lp;
3718         struct lnet_peer_net *lpn;
3719         struct lnet_peer_ni *lpni;
3720         int lncpt;
3721         int cpt;
3722
3723         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3724                 return;
3725
3726         if (!all) {
3727                 lnet_net_lock(LNET_LOCK_EX);
3728                 lpni = lnet_find_peer_ni_locked(nid);
3729                 if (!lpni) {
3730                         lnet_net_unlock(LNET_LOCK_EX);
3731                         return;
3732                 }
3733                 atomic_set(&lpni->lpni_healthv, value);
3734                 lnet_peer_ni_add_to_recoveryq_locked(lpni);
3735                 lnet_peer_ni_decref_locked(lpni);
3736                 lnet_net_unlock(LNET_LOCK_EX);
3737                 return;
3738         }
3739
3740         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3741
3742         /*
3743          * Walk all the peers and reset the healhv for each one to the
3744          * maximum value.
3745          */
3746         lnet_net_lock(LNET_LOCK_EX);
3747         for (cpt = 0; cpt < lncpt; cpt++) {
3748                 ptable = the_lnet.ln_peer_tables[cpt];
3749                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
3750                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
3751                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
3752                                                     lpni_peer_nis) {
3753                                         atomic_set(&lpni->lpni_healthv, value);
3754                                         lnet_peer_ni_add_to_recoveryq_locked(lpni);
3755                                 }
3756                         }
3757                 }
3758         }
3759         lnet_net_unlock(LNET_LOCK_EX);
3760 }
3761