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