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  *
  31  * lnet/lnet/peer.c
  32  */
  33
  34 #define DEBUG_SUBSYSTEM S_LNET
  35
  36 #include <linux/sched.h>
  37 #ifdef HAVE_SCHED_HEADERS
  38 #include <linux/sched/signal.h>
  39 #endif
  40 #include <linux/uaccess.h>
  41
  42 #include <lnet/udsp.h>
  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_NID_NET(&lpni->lpni_nid) == net->net_id) {
  69                         lpni->lpni_net = net;
  70
  71                         spin_lock(&lpni->lpni_lock);
  72                         lpni->lpni_txcredits =
  73                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
  74                         lpni->lpni_mintxcredits = lpni->lpni_txcredits;
  75                         lpni->lpni_rtrcredits =
  76                                 lnet_peer_buffer_credits(lpni->lpni_net);
  77                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
  78                         spin_unlock(&lpni->lpni_lock);
  79
  80                         lnet_peer_remove_from_remote_list(lpni);
  81                 }
  82         }
  83 }
  84
  85 static void
  86 lnet_peer_tables_destroy(void)
  87 {
  88         struct lnet_peer_table  *ptable;
  89         struct list_head        *hash;
  90         int                     i;
  91         int                     j;
  92
  93         if (!the_lnet.ln_peer_tables)
  94                 return;
  95
  96         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  97                 hash = ptable->pt_hash;
  98                 if (!hash) /* not intialized */
  99                         break;
 100
 101                 LASSERT(list_empty(&ptable->pt_zombie_list));
 102
 103                 ptable->pt_hash = NULL;
 104                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 105                         LASSERT(list_empty(&hash[j]));
 106
 107                 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
 108         }
 109
 110         cfs_percpt_free(the_lnet.ln_peer_tables);
 111         the_lnet.ln_peer_tables = NULL;
 112 }
 113
 114 int
 115 lnet_peer_tables_create(void)
 116 {
 117         struct lnet_peer_table  *ptable;
 118         struct list_head        *hash;
 119         int                     i;
 120         int                     j;
 121
 122         the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
 123                                                    sizeof(*ptable));
 124         if (the_lnet.ln_peer_tables == NULL) {
 125                 CERROR("Failed to allocate cpu-partition peer tables\n");
 126                 return -ENOMEM;
 127         }
 128
 129         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 130                 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
 131                                  LNET_PEER_HASH_SIZE * sizeof(*hash));
 132                 if (hash == NULL) {
 133                         CERROR("Failed to create peer hash table\n");
 134                         lnet_peer_tables_destroy();
 135                         return -ENOMEM;
 136                 }
 137
 138                 spin_lock_init(&ptable->pt_zombie_lock);
 139                 INIT_LIST_HEAD(&ptable->pt_zombie_list);
 140
 141                 INIT_LIST_HEAD(&ptable->pt_peer_list);
 142
 143                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 144                         INIT_LIST_HEAD(&hash[j]);
 145                 ptable->pt_hash = hash; /* sign of initialization */
 146         }
 147
 148         return 0;
 149 }
 150
 151 static struct lnet_peer_ni *
 152 lnet_peer_ni_alloc(struct lnet_nid *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         INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
 170         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
 171         kref_init(&lpni->lpni_kref);
 172         lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
 173
 174         spin_lock_init(&lpni->lpni_lock);
 175
 176         if (lnet_peers_start_down())
 177                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 178         else
 179                 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
 180         lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 181         lpni->lpni_nid = *nid;
 182         lpni->lpni_cpt = cpt;
 183         atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
 184
 185         net = lnet_get_net_locked(LNET_NID_NET(nid));
 186         lpni->lpni_net = net;
 187         if (net) {
 188                 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
 189                 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
 190                 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
 191                 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 192         } else {
 193                 /*
 194                  * This peer_ni is not on a local network, so we
 195                  * cannot add the credits here. In case the net is
 196                  * added later, add the peer_ni to the remote peer ni
 197                  * list so it can be easily found and revisited.
 198                  */
 199                 /* FIXME: per-net implementation instead? */
 200                 lnet_peer_ni_addref_locked(lpni);
 201                 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
 202                               &the_lnet.ln_remote_peer_ni_list);
 203         }
 204
 205         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
 206
 207         return lpni;
 208 }
 209
 210 static struct lnet_peer_net *
 211 lnet_peer_net_alloc(__u32 net_id)
 212 {
 213         struct lnet_peer_net *lpn;
 214
 215         LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
 216         if (!lpn)
 217                 return NULL;
 218
 219         INIT_LIST_HEAD(&lpn->lpn_peer_nets);
 220         INIT_LIST_HEAD(&lpn->lpn_peer_nis);
 221         lpn->lpn_net_id = net_id;
 222         lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
 223
 224         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 225
 226         return lpn;
 227 }
 228
 229 void
 230 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
 231 {
 232         struct lnet_peer *lp;
 233
 234         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 235
 236         LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
 237         LASSERT(list_empty(&lpn->lpn_peer_nis));
 238         LASSERT(list_empty(&lpn->lpn_peer_nets));
 239         lp = lpn->lpn_peer;
 240         lpn->lpn_peer = NULL;
 241         LIBCFS_FREE(lpn, sizeof(*lpn));
 242
 243         lnet_peer_decref_locked(lp);
 244 }
 245
 246 static struct lnet_peer *
 247 lnet_peer_alloc(struct lnet_nid *nid)
 248 {
 249         struct lnet_peer *lp;
 250
 251         LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
 252         if (!lp)
 253                 return NULL;
 254
 255         INIT_LIST_HEAD(&lp->lp_rtrq);
 256         INIT_LIST_HEAD(&lp->lp_routes);
 257         INIT_LIST_HEAD(&lp->lp_peer_list);
 258         INIT_LIST_HEAD(&lp->lp_peer_nets);
 259         INIT_LIST_HEAD(&lp->lp_dc_list);
 260         INIT_LIST_HEAD(&lp->lp_dc_pendq);
 261         INIT_LIST_HEAD(&lp->lp_rtr_list);
 262         init_waitqueue_head(&lp->lp_dc_waitq);
 263         spin_lock_init(&lp->lp_lock);
 264         lp->lp_primary_nid = *nid;
 265         lp->lp_disc_src_nid = LNET_ANY_NID;
 266         lp->lp_disc_dst_nid = LNET_ANY_NID;
 267         if (lnet_peers_start_down())
 268                 lp->lp_alive = false;
 269         else
 270                 lp->lp_alive = true;
 271
 272         /*
 273          * all peers created on a router should have health on
 274          * if it's not already on.
 275          */
 276         if (the_lnet.ln_routing && !lnet_health_sensitivity)
 277                 lp->lp_health_sensitivity = 1;
 278
 279         /*
 280          * Turn off discovery for loopback peer. If you're creating a peer
 281          * for the loopback interface then that was initiated when we
 282          * attempted to send a message over the loopback. There is no need
 283          * to ever use a different interface when sending messages to
 284          * myself.
 285          */
 286         if (nid_is_lo0(nid))
 287                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 288         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 289
 290         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
 291
 292         return lp;
 293 }
 294
 295 void
 296 lnet_destroy_peer_locked(struct lnet_peer *lp)
 297 {
 298         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
 299
 300         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 301         LASSERT(lp->lp_rtr_refcount == 0);
 302         LASSERT(list_empty(&lp->lp_peer_nets));
 303         LASSERT(list_empty(&lp->lp_peer_list));
 304         LASSERT(list_empty(&lp->lp_dc_list));
 305
 306         if (lp->lp_data)
 307                 lnet_ping_buffer_decref(lp->lp_data);
 308
 309         /*
 310          * if there are messages still on the pending queue, then make
 311          * sure to queue them on the ln_msg_resend list so they can be
 312          * resent at a later point if the discovery thread is still
 313          * running.
 314          * If the discovery thread has stopped, then the wakeup will be a
 315          * no-op, and it is expected the lnet_shutdown_lndnets() will
 316          * eventually be called, which will traverse this list and
 317          * finalize the messages on the list.
 318          * We can not resend them now because we're holding the cpt lock.
 319          * Releasing the lock can cause an inconsistent state
 320          */
 321         spin_lock(&the_lnet.ln_msg_resend_lock);
 322         spin_lock(&lp->lp_lock);
 323         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 324         spin_unlock(&lp->lp_lock);
 325         spin_unlock(&the_lnet.ln_msg_resend_lock);
 326         wake_up(&the_lnet.ln_dc_waitq);
 327
 328         LIBCFS_FREE(lp, sizeof(*lp));
 329 }
 330
 331 /*
 332  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 333  * that peer_net, detach the peer_net from the peer.
 334  *
 335  * Call with lnet_net_lock/EX held
 336  */
 337 static void
 338 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 339 {
 340         struct lnet_peer_table *ptable;
 341         struct lnet_peer_net *lpn;
 342         struct lnet_peer *lp;
 343
 344         /*
 345          * Belts and suspenders: gracefully handle teardown of a
 346          * partially connected peer_ni.
 347          */
 348         lpn = lpni->lpni_peer_net;
 349
 350         list_del_init(&lpni->lpni_peer_nis);
 351         /*
 352          * If there are no lpni's left, we detach lpn from
 353          * lp_peer_nets, so it cannot be found anymore.
 354          */
 355         if (list_empty(&lpn->lpn_peer_nis))
 356                 list_del_init(&lpn->lpn_peer_nets);
 357
 358         /* Update peer NID count. */
 359         lp = lpn->lpn_peer;
 360         lp->lp_nnis--;
 361
 362         /*
 363          * If there are no more peer nets, make the peer unfindable
 364          * via the peer_tables.
 365          *
 366          * Otherwise, if the peer is DISCOVERED, tell discovery to
 367          * take another look at it. This is a no-op if discovery for
 368          * this peer did the detaching.
 369          */
 370         if (list_empty(&lp->lp_peer_nets)) {
 371                 list_del_init(&lp->lp_peer_list);
 372                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 373                 ptable->pt_peers--;
 374         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 375                 /* Discovery isn't running, nothing to do here. */
 376         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 377                 lnet_peer_queue_for_discovery(lp);
 378                 wake_up(&the_lnet.ln_dc_waitq);
 379         }
 380         CDEBUG(D_NET, "peer %s NID %s\n",
 381                 libcfs_nidstr(&lp->lp_primary_nid),
 382                 libcfs_nidstr(&lpni->lpni_nid));
 383 }
 384
 385 /* called with lnet_net_lock LNET_LOCK_EX held */
 386 static int
 387 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 388 {
 389         struct lnet_peer_table *ptable = NULL;
 390
 391         /* don't remove a peer_ni if it's also a gateway */
 392         if (lnet_isrouter(lpni) && !force) {
 393                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 394                        libcfs_nidstr(&lpni->lpni_nid));
 395                 return -EBUSY;
 396         }
 397
 398         lnet_peer_remove_from_remote_list(lpni);
 399
 400         /* remove peer ni from the hash list. */
 401         list_del_init(&lpni->lpni_hashlist);
 402
 403         /*
 404          * indicate the peer is being deleted so the monitor thread can
 405          * remove it from the recovery queue.
 406          */
 407         spin_lock(&lpni->lpni_lock);
 408         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 409         spin_unlock(&lpni->lpni_lock);
 410
 411         /* decrement the ref count on the peer table */
 412         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 413
 414         /*
 415          * The peer_ni can no longer be found with a lookup. But there
 416          * can be current users, so keep track of it on the zombie
 417          * list until the reference count has gone to zero.
 418          *
 419          * The last reference may be lost in a place where the
 420          * lnet_net_lock locks only a single cpt, and that cpt may not
 421          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 422          * has its own lock.
 423          */
 424         spin_lock(&ptable->pt_zombie_lock);
 425         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 426         ptable->pt_zombies++;
 427         spin_unlock(&ptable->pt_zombie_lock);
 428
 429         /* no need to keep this peer_ni on the hierarchy anymore */
 430         lnet_peer_detach_peer_ni_locked(lpni);
 431
 432         /* remove hashlist reference on peer_ni */
 433         lnet_peer_ni_decref_locked(lpni);
 434
 435         return 0;
 436 }
 437
 438 void lnet_peer_uninit(void)
 439 {
 440         struct lnet_peer_ni *lpni, *tmp;
 441
 442         lnet_net_lock(LNET_LOCK_EX);
 443
 444         /* remove all peer_nis from the remote peer and the hash list */
 445         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 446                                  lpni_on_remote_peer_ni_list)
 447                 lnet_peer_ni_del_locked(lpni, false);
 448
 449         lnet_peer_tables_destroy();
 450
 451         lnet_net_unlock(LNET_LOCK_EX);
 452 }
 453
 454 static int
 455 lnet_peer_del_locked(struct lnet_peer *peer)
 456 {
 457         struct lnet_peer_ni *lpni = NULL, *lpni2;
 458         int rc = 0, rc2 = 0;
 459
 460         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&peer->lp_primary_nid));
 461
 462         spin_lock(&peer->lp_lock);
 463         peer->lp_state |= LNET_PEER_MARK_DELETED;
 464         spin_unlock(&peer->lp_lock);
 465
 466         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 467         while (lpni != NULL) {
 468                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 469                 rc = lnet_peer_ni_del_locked(lpni, false);
 470                 if (rc != 0)
 471                         rc2 = rc;
 472                 lpni = lpni2;
 473         }
 474
 475         return rc2;
 476 }
 477
 478 /*
 479  * Discovering this peer is taking too long. Cancel any Ping or Push
 480  * that discovery is waiting on by unlinking the relevant MDs. The
 481  * lnet_discovery_event_handler() will proceed from here and complete
 482  * the cleanup.
 483  */
 484 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
 485 {
 486         struct lnet_handle_md ping_mdh;
 487         struct lnet_handle_md push_mdh;
 488
 489         LNetInvalidateMDHandle(&ping_mdh);
 490         LNetInvalidateMDHandle(&push_mdh);
 491
 492         spin_lock(&lp->lp_lock);
 493         if (lp->lp_state & LNET_PEER_PING_SENT) {
 494                 ping_mdh = lp->lp_ping_mdh;
 495                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
 496         }
 497         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
 498                 push_mdh = lp->lp_push_mdh;
 499                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
 500         }
 501         spin_unlock(&lp->lp_lock);
 502
 503         if (!LNetMDHandleIsInvalid(ping_mdh))
 504                 LNetMDUnlink(ping_mdh);
 505         if (!LNetMDHandleIsInvalid(push_mdh))
 506                 LNetMDUnlink(push_mdh);
 507 }
 508
 509 static int
 510 lnet_peer_del(struct lnet_peer *peer)
 511 {
 512         int rc;
 513
 514         lnet_peer_cancel_discovery(peer);
 515         lnet_net_lock(LNET_LOCK_EX);
 516         rc = lnet_peer_del_locked(peer);
 517         lnet_net_unlock(LNET_LOCK_EX);
 518
 519         return rc;
 520 }
 521
 522 /*
 523  * Delete a NID from a peer. Call with ln_api_mutex held.
 524  *
 525  * Error codes:
 526  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 527  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 528  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 529  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 530  */
 531 static int
 532 lnet_peer_del_nid(struct lnet_peer *lp, struct lnet_nid *nid,
 533                   unsigned int flags)
 534 {
 535         struct lnet_peer_ni *lpni;
 536         struct lnet_nid primary_nid = lp->lp_primary_nid;
 537         int rc = 0;
 538         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 539
 540         if (!(flags & LNET_PEER_CONFIGURED)) {
 541                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 542                         rc = -EPERM;
 543                         goto out;
 544                 }
 545         }
 546
 547         lpni = lnet_peer_ni_find_locked(nid);
 548         if (!lpni) {
 549                 rc = -ENOENT;
 550                 goto out;
 551         }
 552         lnet_peer_ni_decref_locked(lpni);
 553         if (lp != lpni->lpni_peer_net->lpn_peer) {
 554                 rc = -ECHILD;
 555                 goto out;
 556         }
 557
 558         /*
 559          * This function only allows deletion of the primary NID if it
 560          * is the only NID.
 561          */
 562         if (nid_same(nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && !force) {
 563                 rc = -EBUSY;
 564                 goto out;
 565         }
 566
 567         lnet_net_lock(LNET_LOCK_EX);
 568
 569         if (nid_same(nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && force) {
 570                 struct lnet_peer_ni *lpni2;
 571                 /* assign the next peer_ni to be the primary */
 572                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 573                 LASSERT(lpni2);
 574                 lp->lp_primary_nid = lpni2->lpni_nid;
 575         }
 576         rc = lnet_peer_ni_del_locked(lpni, force);
 577
 578         lnet_net_unlock(LNET_LOCK_EX);
 579
 580 out:
 581         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 582                libcfs_nidstr(&primary_nid), libcfs_nidstr(nid),
 583                flags, rc);
 584
 585         return rc;
 586 }
 587
 588 static void
 589 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 590                                struct lnet_peer_table *ptable)
 591 {
 592         int                      i;
 593         struct lnet_peer_ni     *next;
 594         struct lnet_peer_ni     *lpni;
 595         struct lnet_peer        *peer;
 596
 597         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 598                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 599                                          lpni_hashlist) {
 600                         if (net != NULL && net != lpni->lpni_net)
 601                                 continue;
 602
 603                         peer = lpni->lpni_peer_net->lpn_peer;
 604                         if (!nid_same(&peer->lp_primary_nid,
 605                                        &lpni->lpni_nid)) {
 606                                 lnet_peer_ni_del_locked(lpni, false);
 607                                 continue;
 608                         }
 609                         /*
 610                          * Removing the primary NID implies removing
 611                          * the entire peer. Advance next beyond any
 612                          * peer_ni that belongs to the same peer.
 613                          */
 614                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 615                                                  lpni_hashlist) {
 616                                 if (next->lpni_peer_net->lpn_peer != peer)
 617                                         break;
 618                         }
 619                         lnet_peer_del_locked(peer);
 620                 }
 621         }
 622 }
 623
 624 static void
 625 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 626 {
 627         wait_var_event_warning(&ptable->pt_zombies,
 628                                ptable->pt_zombies == 0,
 629                                "Waiting for %d zombies on peer table\n",
 630                                ptable->pt_zombies);
 631 }
 632
 633 static void
 634 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 635                                 struct lnet_peer_table *ptable)
 636 {
 637         struct lnet_peer_ni     *lp;
 638         struct lnet_peer_ni     *tmp;
 639         struct lnet_nid         gw_nid;
 640         int                     i;
 641
 642         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 643                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 644                                          lpni_hashlist) {
 645                         if (net != lp->lpni_net)
 646                                 continue;
 647
 648                         if (!lnet_isrouter(lp))
 649                                 continue;
 650
 651                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 652
 653                         lnet_net_unlock(LNET_LOCK_EX);
 654                         lnet_del_route(LNET_NET_ANY, &gw_nid);
 655                         lnet_net_lock(LNET_LOCK_EX);
 656                 }
 657         }
 658 }
 659
 660 void
 661 lnet_peer_tables_cleanup(struct lnet_net *net)
 662 {
 663         int i;
 664         struct lnet_peer_table *ptable;
 665
 666         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 667         /* If just deleting the peers for a NI, get rid of any routes these
 668          * peers are gateways for. */
 669         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 670                 lnet_net_lock(LNET_LOCK_EX);
 671                 lnet_peer_table_del_rtrs_locked(net, ptable);
 672                 lnet_net_unlock(LNET_LOCK_EX);
 673         }
 674
 675         /* Start the cleanup process */
 676         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 677                 lnet_net_lock(LNET_LOCK_EX);
 678                 lnet_peer_table_cleanup_locked(net, ptable);
 679                 lnet_net_unlock(LNET_LOCK_EX);
 680         }
 681
 682         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 683                 lnet_peer_ni_finalize_wait(ptable);
 684 }
 685
 686 static struct lnet_peer_ni *
 687 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, struct lnet_nid *nid)
 688 {
 689         struct list_head        *peers;
 690         struct lnet_peer_ni     *lp;
 691
 692         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 693                 return NULL;
 694
 695         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 696         list_for_each_entry(lp, peers, lpni_hashlist) {
 697                 if (nid_same(&lp->lpni_nid, nid)) {
 698                         lnet_peer_ni_addref_locked(lp);
 699                         return lp;
 700                 }
 701         }
 702
 703         return NULL;
 704 }
 705
 706 struct lnet_peer_ni *
 707 lnet_peer_ni_find_locked(struct lnet_nid *nid)
 708 {
 709         struct lnet_peer_ni *lpni;
 710         struct lnet_peer_table *ptable;
 711         int cpt;
 712
 713         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 714
 715         ptable = the_lnet.ln_peer_tables[cpt];
 716         lpni = lnet_get_peer_ni_locked(ptable, nid);
 717
 718         return lpni;
 719 }
 720
 721 struct lnet_peer_ni *
 722 lnet_peer_ni_get_locked(struct lnet_peer *lp, struct lnet_nid *nid)
 723 {
 724         struct lnet_peer_net *lpn;
 725         struct lnet_peer_ni *lpni;
 726
 727         lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
 728         if (!lpn)
 729                 return NULL;
 730
 731         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 732                 if (nid_same(&lpni->lpni_nid, nid))
 733                         return lpni;
 734         }
 735
 736         return NULL;
 737 }
 738
 739 struct lnet_peer *
 740 lnet_find_peer(struct lnet_nid *nid)
 741 {
 742         struct lnet_peer_ni *lpni;
 743         struct lnet_peer *lp = NULL;
 744         int cpt;
 745
 746         cpt = lnet_net_lock_current();
 747         lpni = lnet_peer_ni_find_locked(nid);
 748         if (lpni) {
 749                 lp = lpni->lpni_peer_net->lpn_peer;
 750                 lnet_peer_addref_locked(lp);
 751                 lnet_peer_ni_decref_locked(lpni);
 752         }
 753         lnet_net_unlock(cpt);
 754
 755         return lp;
 756 }
 757
 758 struct lnet_peer_net *
 759 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 760 {
 761         struct lnet_peer_net *net;
 762
 763         if (!prev_lpn_id) {
 764                 /* no net id provided return the first net */
 765                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 766                                                struct lnet_peer_net,
 767                                                lpn_peer_nets);
 768
 769                 return net;
 770         }
 771
 772         /* find the net after the one provided */
 773         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 774                 if (net->lpn_net_id == prev_lpn_id) {
 775                         /*
 776                          * if we reached the end of the list loop to the
 777                          * beginning.
 778                          */
 779                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 780                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 781                                                                 struct lnet_peer_net,
 782                                                                 lpn_peer_nets);
 783                         else
 784                                 return list_next_entry(net, lpn_peer_nets);
 785                 }
 786         }
 787
 788         return NULL;
 789 }
 790
 791 struct lnet_peer_ni *
 792 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 793                              struct lnet_peer_net *peer_net,
 794                              struct lnet_peer_ni *prev)
 795 {
 796         struct lnet_peer_ni *lpni;
 797         struct lnet_peer_net *net = peer_net;
 798
 799         if (!prev) {
 800                 if (!net) {
 801                         if (list_empty(&peer->lp_peer_nets))
 802                                 return NULL;
 803
 804                         net = list_entry(peer->lp_peer_nets.next,
 805                                          struct lnet_peer_net,
 806                                          lpn_peer_nets);
 807                 }
 808                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 809                                   lpni_peer_nis);
 810
 811                 return lpni;
 812         }
 813
 814         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 815                 /*
 816                  * if you reached the end of the peer ni list and the peer
 817                  * net is specified then there are no more peer nis in that
 818                  * net.
 819                  */
 820                 if (net)
 821                         return NULL;
 822
 823                 /*
 824                  * we reached the end of this net ni list. move to the
 825                  * next net
 826                  */
 827                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 828                     &peer->lp_peer_nets)
 829                         /* no more nets and no more NIs. */
 830                         return NULL;
 831
 832                 /* get the next net */
 833                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 834                                  struct lnet_peer_net,
 835                                  lpn_peer_nets);
 836                 /* get the ni on it */
 837                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 838                                   lpni_peer_nis);
 839
 840                 return lpni;
 841         }
 842
 843         /* there are more nis left */
 844         lpni = list_entry(prev->lpni_peer_nis.next,
 845                           struct lnet_peer_ni, lpni_peer_nis);
 846
 847         return lpni;
 848 }
 849
 850 /* Call with the ln_api_mutex held */
 851 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 852 {
 853         struct lnet_process_id id;
 854         struct lnet_peer_table *ptable;
 855         struct lnet_peer *lp;
 856         __u32 count = 0;
 857         __u32 size = 0;
 858         int lncpt;
 859         int cpt;
 860         __u32 i;
 861         int rc;
 862
 863         rc = -ESHUTDOWN;
 864         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 865                 goto done;
 866
 867         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 868
 869         /*
 870          * Count the number of peers, and return E2BIG if the buffer
 871          * is too small. We'll also return the desired size.
 872          */
 873         rc = -E2BIG;
 874         for (cpt = 0; cpt < lncpt; cpt++) {
 875                 ptable = the_lnet.ln_peer_tables[cpt];
 876                 count += ptable->pt_peers;
 877         }
 878         size = count * sizeof(*ids);
 879         if (size > *sizep)
 880                 goto done;
 881
 882         /*
 883          * Walk the peer lists and copy out the primary nids.
 884          * This is safe because the peer lists are only modified
 885          * while the ln_api_mutex is held. So we don't need to
 886          * hold the lnet_net_lock as well, and can therefore
 887          * directly call copy_to_user().
 888          */
 889         rc = -EFAULT;
 890         memset(&id, 0, sizeof(id));
 891         id.pid = LNET_PID_LUSTRE;
 892         i = 0;
 893         for (cpt = 0; cpt < lncpt; cpt++) {
 894                 ptable = the_lnet.ln_peer_tables[cpt];
 895                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 896                         if (!nid_is_nid4(&lp->lp_primary_nid))
 897                                 continue;
 898                         if (i >= count)
 899                                 goto done;
 900                         id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
 901                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 902                                 goto done;
 903                         i++;
 904                 }
 905         }
 906         rc = 0;
 907 done:
 908         *countp = count;
 909         *sizep = size;
 910         return rc;
 911 }
 912
 913 /*
 914  * Start pushes to peers that need to be updated for a configuration
 915  * change on this node.
 916  */
 917 void
 918 lnet_push_update_to_peers(int force)
 919 {
 920         struct lnet_peer_table *ptable;
 921         struct lnet_peer *lp;
 922         int lncpt;
 923         int cpt;
 924
 925         lnet_net_lock(LNET_LOCK_EX);
 926         if (lnet_peer_discovery_disabled)
 927                 force = 0;
 928         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 929         for (cpt = 0; cpt < lncpt; cpt++) {
 930                 ptable = the_lnet.ln_peer_tables[cpt];
 931                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 932                         if (force) {
 933                                 spin_lock(&lp->lp_lock);
 934                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 935                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 936                                 spin_unlock(&lp->lp_lock);
 937                         }
 938                         if (lnet_peer_needs_push(lp))
 939                                 lnet_peer_queue_for_discovery(lp);
 940                 }
 941         }
 942         lnet_net_unlock(LNET_LOCK_EX);
 943         wake_up(&the_lnet.ln_dc_waitq);
 944 }
 945
 946 /* find the NID in the preferred gateways for the remote peer
 947  * return:
 948  *      false: list is not empty and NID is not preferred
 949  *      false: list is empty
 950  *      true: nid is found in the list
 951  */
 952 bool
 953 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
 954                              struct lnet_nid *gw_nid)
 955 {
 956         struct lnet_nid_list *ne;
 957
 958         CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
 959                libcfs_nidstr(&lpni->lpni_nid),
 960                list_empty(&lpni->lpni_rtr_pref_nids));
 961
 962         if (list_empty(&lpni->lpni_rtr_pref_nids))
 963                 return false;
 964
 965         /* iterate through all the preferred NIDs and see if any of them
 966          * matches the provided gw_nid
 967          */
 968         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
 969                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
 970                        libcfs_nidstr(&ne->nl_nid),
 971                        libcfs_nidstr(gw_nid));
 972                 if (nid_same(&ne->nl_nid, gw_nid))
 973                         return true;
 974         }
 975
 976         return false;
 977 }
 978
 979 void
 980 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
 981 {
 982         struct list_head zombies;
 983         struct lnet_nid_list *ne;
 984         struct lnet_nid_list *tmp;
 985         int cpt = lpni->lpni_cpt;
 986
 987         INIT_LIST_HEAD(&zombies);
 988
 989         lnet_net_lock(cpt);
 990         list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
 991         lnet_net_unlock(cpt);
 992
 993         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
 994                 list_del(&ne->nl_list);
 995                 LIBCFS_FREE(ne, sizeof(*ne));
 996         }
 997 }
 998
 999 int
1000 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
1001                        struct lnet_nid *gw_nid)
1002 {
1003         int cpt = lpni->lpni_cpt;
1004         struct lnet_nid_list *ne = NULL;
1005
1006         /* This function is called with api_mutex held. When the api_mutex
1007          * is held the list can not be modified, as it is only modified as
1008          * a result of applying a UDSP and that happens under api_mutex
1009          * lock.
1010          */
1011         __must_hold(&the_lnet.ln_api_mutex);
1012
1013         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1014                 if (nid_same(&ne->nl_nid, gw_nid))
1015                         return -EEXIST;
1016         }
1017
1018         LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1019         if (!ne)
1020                 return -ENOMEM;
1021
1022         ne->nl_nid = *gw_nid;
1023
1024         /* Lock the cpt to protect against addition and checks in the
1025          * selection algorithm
1026          */
1027         lnet_net_lock(cpt);
1028         list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1029         lnet_net_unlock(cpt);
1030
1031         return 0;
1032 }
1033
1034 /*
1035  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1036  * this is a preferred point-to-point path. Call with lnet_net_lock in
1037  * shared mmode.
1038  */
1039 bool
1040 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1041 {
1042         struct lnet_nid_list *ne;
1043
1044         if (lpni->lpni_pref_nnids == 0)
1045                 return false;
1046         if (lpni->lpni_pref_nnids == 1)
1047                 return nid_same(&lpni->lpni_pref.nid, nid);
1048         list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1049                 if (nid_same(&ne->nl_nid, nid))
1050                         return true;
1051         }
1052         return false;
1053 }
1054
1055 /*
1056  * Set a single ni as preferred, provided no preferred ni is already
1057  * defined. Only to be used for non-multi-rail peer_ni.
1058  */
1059 int
1060 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni,
1061                                   struct lnet_nid *nid)
1062 {
1063         int rc = 0;
1064
1065         if (!nid)
1066                 return -EINVAL;
1067         spin_lock(&lpni->lpni_lock);
1068         if (LNET_NID_IS_ANY(nid)) {
1069                 rc = -EINVAL;
1070         } else if (lpni->lpni_pref_nnids > 0) {
1071                 rc = -EPERM;
1072         } else if (lpni->lpni_pref_nnids == 0) {
1073                 lpni->lpni_pref.nid = *nid;
1074                 lpni->lpni_pref_nnids = 1;
1075                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1076         }
1077         spin_unlock(&lpni->lpni_lock);
1078
1079         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1080                libcfs_nidstr(&lpni->lpni_nid), libcfs_nidstr(nid), rc);
1081         return rc;
1082 }
1083
1084 /*
1085  * Clear the preferred NID from a non-multi-rail peer_ni, provided
1086  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1087  */
1088 int
1089 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1090 {
1091         int rc = 0;
1092
1093         spin_lock(&lpni->lpni_lock);
1094         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1095                 lpni->lpni_pref_nnids = 0;
1096                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1097         } else if (lpni->lpni_pref_nnids == 0) {
1098                 rc = -ENOENT;
1099         } else {
1100                 rc = -EPERM;
1101         }
1102         spin_unlock(&lpni->lpni_lock);
1103
1104         CDEBUG(D_NET, "peer %s: %d\n",
1105                libcfs_nidstr(&lpni->lpni_nid), rc);
1106         return rc;
1107 }
1108
1109 void
1110 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1111 {
1112         lpni->lpni_sel_priority = priority;
1113 }
1114
1115 /*
1116  * Clear the preferred NIDs from a non-multi-rail peer.
1117  */
1118 void
1119 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1120 {
1121         struct lnet_peer_ni *lpni = NULL;
1122
1123         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1124                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1125 }
1126
1127 int
1128 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1129 {
1130         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1131         struct lnet_nid_list *ne1 = NULL;
1132         struct lnet_nid_list *ne2 = NULL;
1133         struct lnet_nid *tmp_nid = NULL;
1134         int rc = 0;
1135
1136         if (LNET_NID_IS_ANY(nid)) {
1137                 rc = -EINVAL;
1138                 goto out;
1139         }
1140
1141         if (lpni->lpni_pref_nnids == 1 &&
1142             nid_same(&lpni->lpni_pref.nid, nid)) {
1143                 rc = -EEXIST;
1144                 goto out;
1145         }
1146
1147         /* A non-MR node may have only one preferred NI per peer_ni */
1148         if (lpni->lpni_pref_nnids > 0 &&
1149             !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1150                 rc = -EPERM;
1151                 goto out;
1152         }
1153
1154         /* add the new preferred nid to the list of preferred nids */
1155         if (lpni->lpni_pref_nnids != 0) {
1156                 size_t alloc_size = sizeof(*ne1);
1157
1158                 if (lpni->lpni_pref_nnids == 1) {
1159                         tmp_nid = &lpni->lpni_pref.nid;
1160                         INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1161                 }
1162
1163                 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1164                         if (nid_same(&ne1->nl_nid, nid)) {
1165                                 rc = -EEXIST;
1166                                 goto out;
1167                         }
1168                 }
1169
1170                 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1171                                  alloc_size);
1172                 if (!ne1) {
1173                         rc = -ENOMEM;
1174                         goto out;
1175                 }
1176
1177                 /* move the originally stored nid to the list */
1178                 if (lpni->lpni_pref_nnids == 1) {
1179                         LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1180                                 lpni->lpni_cpt, alloc_size);
1181                         if (!ne2) {
1182                                 rc = -ENOMEM;
1183                                 goto out;
1184                         }
1185                         INIT_LIST_HEAD(&ne2->nl_list);
1186                         ne2->nl_nid = *tmp_nid;
1187                 }
1188                 ne1->nl_nid = *nid;
1189         }
1190
1191         lnet_net_lock(LNET_LOCK_EX);
1192         spin_lock(&lpni->lpni_lock);
1193         if (lpni->lpni_pref_nnids == 0) {
1194                 lpni->lpni_pref.nid = *nid;
1195         } else {
1196                 if (ne2)
1197                         list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1198                 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1199         }
1200         lpni->lpni_pref_nnids++;
1201         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1202         spin_unlock(&lpni->lpni_lock);
1203         lnet_net_unlock(LNET_LOCK_EX);
1204
1205 out:
1206         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1207                 spin_lock(&lpni->lpni_lock);
1208                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1209                 spin_unlock(&lpni->lpni_lock);
1210         }
1211         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1212                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1213         return rc;
1214 }
1215
1216 int
1217 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1218 {
1219         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1220         struct lnet_nid_list *ne = NULL;
1221         int rc = 0;
1222
1223         if (lpni->lpni_pref_nnids == 0) {
1224                 rc = -ENOENT;
1225                 goto out;
1226         }
1227
1228         if (lpni->lpni_pref_nnids == 1) {
1229                 if (!nid_same(&lpni->lpni_pref.nid, nid)) {
1230                         rc = -ENOENT;
1231                         goto out;
1232                 }
1233         } else {
1234                 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1235                         if (nid_same(&ne->nl_nid, nid))
1236                                 goto remove_nid_entry;
1237                 }
1238                 rc = -ENOENT;
1239                 ne = NULL;
1240                 goto out;
1241         }
1242
1243 remove_nid_entry:
1244         lnet_net_lock(LNET_LOCK_EX);
1245         spin_lock(&lpni->lpni_lock);
1246         if (lpni->lpni_pref_nnids == 1)
1247                 lpni->lpni_pref.nid = LNET_ANY_NID;
1248         else {
1249                 list_del_init(&ne->nl_list);
1250                 if (lpni->lpni_pref_nnids == 2) {
1251                         struct lnet_nid_list *ne, *tmp;
1252
1253                         list_for_each_entry_safe(ne, tmp,
1254                                                  &lpni->lpni_pref.nids,
1255                                                  nl_list) {
1256                                 lpni->lpni_pref.nid = ne->nl_nid;
1257                                 list_del_init(&ne->nl_list);
1258                                 LIBCFS_FREE(ne, sizeof(*ne));
1259                         }
1260                 }
1261         }
1262         lpni->lpni_pref_nnids--;
1263         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1264         spin_unlock(&lpni->lpni_lock);
1265         lnet_net_unlock(LNET_LOCK_EX);
1266
1267         if (ne)
1268                 LIBCFS_FREE(ne, sizeof(*ne));
1269 out:
1270         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1271                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1272         return rc;
1273 }
1274
1275 void
1276 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1277 {
1278         struct list_head zombies;
1279         struct lnet_nid_list *ne;
1280         struct lnet_nid_list *tmp;
1281
1282         INIT_LIST_HEAD(&zombies);
1283
1284         lnet_net_lock(LNET_LOCK_EX);
1285         if (lpni->lpni_pref_nnids == 1)
1286                 lpni->lpni_pref.nid = LNET_ANY_NID;
1287         else if (lpni->lpni_pref_nnids > 1)
1288                 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1289         lpni->lpni_pref_nnids = 0;
1290         lnet_net_unlock(LNET_LOCK_EX);
1291
1292         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1293                 list_del_init(&ne->nl_list);
1294                 LIBCFS_FREE(ne, sizeof(*ne));
1295         }
1296 }
1297
1298 void
1299 lnet_peer_primary_nid_locked(struct lnet_nid *nid, struct lnet_nid *result)
1300 {
1301         struct lnet_peer_ni *lpni;
1302
1303         *result = *nid;
1304         lpni = lnet_peer_ni_find_locked(nid);
1305         if (lpni) {
1306                 *result = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1307                 lnet_peer_ni_decref_locked(lpni);
1308         }
1309 }
1310
1311 bool
1312 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1313 __must_hold(&lp->lp_lock)
1314 {
1315         if (lnet_peer_discovery_disabled)
1316                 return true;
1317
1318         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1319             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1320                 return true;
1321         }
1322
1323         return false;
1324 }
1325
1326 /*
1327  * Peer Discovery
1328  */
1329 bool
1330 lnet_is_discovery_disabled(struct lnet_peer *lp)
1331 {
1332         bool rc = false;
1333
1334         spin_lock(&lp->lp_lock);
1335         rc = lnet_is_discovery_disabled_locked(lp);
1336         spin_unlock(&lp->lp_lock);
1337
1338         return rc;
1339 }
1340
1341 int
1342 LNetAddPeer(lnet_nid_t *nids, __u32 num_nids)
1343 {
1344         struct lnet_nid pnid = LNET_ANY_NID;
1345         bool mr;
1346         int i, rc;
1347
1348         if (!nids || num_nids < 1)
1349                 return -EINVAL;
1350
1351         rc = LNetNIInit(LNET_PID_ANY);
1352         if (rc < 0)
1353                 return rc;
1354
1355         mutex_lock(&the_lnet.ln_api_mutex);
1356
1357         mr = lnet_peer_discovery_disabled == 0;
1358
1359         rc = 0;
1360         for (i = 0; i < num_nids; i++) {
1361                 struct lnet_nid nid;
1362
1363                 if (nids[i] == LNET_NID_LO_0)
1364                         continue;
1365
1366                 lnet_nid4_to_nid(nids[i], &nid);
1367                 if (LNET_NID_IS_ANY(&pnid)) {
1368                         lnet_nid4_to_nid(nids[i], &pnid);
1369                         rc = lnet_add_peer_ni(&pnid, &LNET_ANY_NID, mr, true);
1370                 } else if (lnet_peer_discovery_disabled) {
1371                         lnet_nid4_to_nid(nids[i], &nid);
1372                         rc = lnet_add_peer_ni(&nid, &LNET_ANY_NID, mr, true);
1373                 } else {
1374                         lnet_nid4_to_nid(nids[i], &nid);
1375                         rc = lnet_add_peer_ni(&pnid, &nid, mr, true);
1376                 }
1377
1378                 if (rc && rc != -EEXIST)
1379                         goto unlock;
1380         }
1381
1382 unlock:
1383         mutex_unlock(&the_lnet.ln_api_mutex);
1384
1385         LNetNIFini();
1386
1387         return rc == -EEXIST ? 0 : rc;
1388 }
1389 EXPORT_SYMBOL(LNetAddPeer);
1390
1391 void LNetPrimaryNID(struct lnet_nid *nid)
1392 {
1393         struct lnet_peer *lp;
1394         struct lnet_peer_ni *lpni;
1395         struct lnet_nid orig;
1396         int rc = 0;
1397         int cpt;
1398
1399         if (!nid || nid_is_lo0(nid))
1400                 return;
1401         orig = *nid;
1402
1403         cpt = lnet_net_lock_current();
1404         lpni = lnet_peerni_by_nid_locked(nid, NULL, cpt);
1405         if (IS_ERR(lpni)) {
1406                 rc = PTR_ERR(lpni);
1407                 goto out_unlock;
1408         }
1409         lp = lpni->lpni_peer_net->lpn_peer;
1410
1411         /* If discovery is disabled locally then we needn't bother running
1412          * discovery here because discovery will not modify whatever
1413          * primary NID is currently set for this peer. If the specified peer is
1414          * down then this discovery can introduce long delays into the mount
1415          * process, so skip it if it isn't necessary.
1416          */
1417         while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1418                 spin_lock(&lp->lp_lock);
1419                 /* force a full discovery cycle */
1420                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1421                 spin_unlock(&lp->lp_lock);
1422
1423                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1424                 if (rc)
1425                         goto out_decref;
1426                 /* The lpni (or lp) for this NID may have changed and our ref is
1427                  * the only thing keeping the old one around. Release the ref
1428                  * and lookup the lpni again
1429                  */
1430                 lnet_peer_ni_decref_locked(lpni);
1431                 lpni = lnet_peer_ni_find_locked(nid);
1432                 if (!lpni) {
1433                         rc = -ENOENT;
1434                         goto out_unlock;
1435                 }
1436                 lp = lpni->lpni_peer_net->lpn_peer;
1437
1438                 /* If we find that the peer has discovery disabled then we will
1439                  * not modify whatever primary NID is currently set for this
1440                  * peer. Thus, we can break out of this loop even if the peer
1441                  * is not fully up to date.
1442                  */
1443                 if (lnet_is_discovery_disabled(lp))
1444                         break;
1445         }
1446         *nid = lp->lp_primary_nid;
1447 out_decref:
1448         lnet_peer_ni_decref_locked(lpni);
1449 out_unlock:
1450         lnet_net_unlock(cpt);
1451
1452         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nidstr(&orig),
1453                libcfs_nidstr(nid), rc);
1454 }
1455 EXPORT_SYMBOL(LNetPrimaryNID);
1456
1457 struct lnet_peer_net *
1458 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1459 {
1460         struct lnet_peer_net *peer_net;
1461         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1462                 if (peer_net->lpn_net_id == net_id)
1463                         return peer_net;
1464         }
1465         return NULL;
1466 }
1467
1468 /*
1469  * Attach a peer_ni to a peer_net and peer. This function assumes
1470  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1471  * may be attached to a different peer, in which case it will be
1472  * properly detached first. The whole operation is done atomically.
1473  *
1474  * This function consumes the reference on lpni and Always returns 0.
1475  * This is the last function called from functions that do return an
1476  * int, so returning 0 here allows the compiler to do a tail call.
1477  */
1478 static int
1479 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1480                                 struct lnet_peer_net *lpn,
1481                                 struct lnet_peer_ni *lpni,
1482                                 unsigned flags)
1483 {
1484         struct lnet_peer_table *ptable;
1485         bool new_lpn = false;
1486         int rc;
1487
1488         /* Install the new peer_ni */
1489         lnet_net_lock(LNET_LOCK_EX);
1490         /* Add peer_ni to global peer table hash, if necessary. */
1491         if (list_empty(&lpni->lpni_hashlist)) {
1492                 int hash = lnet_nid2peerhash(&lpni->lpni_nid);
1493
1494                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1495                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1496                 ptable->pt_version++;
1497                 lnet_peer_ni_addref_locked(lpni);
1498         }
1499
1500         /* Detach the peer_ni from an existing peer, if necessary. */
1501         if (lpni->lpni_peer_net) {
1502                 LASSERT(lpni->lpni_peer_net != lpn);
1503                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1504                 lnet_peer_detach_peer_ni_locked(lpni);
1505                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1506                 lpni->lpni_peer_net = NULL;
1507         }
1508
1509         /* Add peer_ni to peer_net */
1510         lpni->lpni_peer_net = lpn;
1511         if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1512                 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1513         else
1514                 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1515         lnet_update_peer_net_healthv(lpni);
1516         lnet_peer_net_addref_locked(lpn);
1517
1518         /* Add peer_net to peer */
1519         if (!lpn->lpn_peer) {
1520                 new_lpn = true;
1521                 lpn->lpn_peer = lp;
1522                 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1523                         list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1524                 else
1525                         list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1526                 lnet_peer_addref_locked(lp);
1527         }
1528
1529         /* Add peer to global peer list, if necessary */
1530         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1531         if (list_empty(&lp->lp_peer_list)) {
1532                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1533                 ptable->pt_peers++;
1534         }
1535
1536
1537         /* Update peer state */
1538         spin_lock(&lp->lp_lock);
1539         if (flags & LNET_PEER_CONFIGURED) {
1540                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1541                         lp->lp_state |= LNET_PEER_CONFIGURED;
1542         }
1543         if (flags & LNET_PEER_MULTI_RAIL) {
1544                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1545                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1546                         lnet_peer_clr_non_mr_pref_nids(lp);
1547                 }
1548         }
1549         spin_unlock(&lp->lp_lock);
1550
1551         lp->lp_nnis++;
1552
1553         /* apply UDSPs */
1554         if (new_lpn) {
1555                 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1556                 if (rc)
1557                         CERROR("Failed to apply UDSPs on lpn %s\n",
1558                                libcfs_net2str(lpn->lpn_net_id));
1559         }
1560         rc = lnet_udsp_apply_policies_on_lpni(lpni);
1561         if (rc)
1562                 CERROR("Failed to apply UDSPs on lpni %s\n",
1563                        libcfs_nidstr(&lpni->lpni_nid));
1564
1565         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1566                libcfs_nidstr(&lp->lp_primary_nid),
1567                libcfs_nidstr(&lpni->lpni_nid), flags);
1568         lnet_peer_ni_decref_locked(lpni);
1569         lnet_net_unlock(LNET_LOCK_EX);
1570
1571         return 0;
1572 }
1573
1574 /*
1575  * Create a new peer, with nid as its primary nid.
1576  *
1577  * Call with the lnet_api_mutex held.
1578  */
1579 static int
1580 lnet_peer_add(struct lnet_nid *nid, unsigned int flags)
1581 {
1582         struct lnet_peer *lp;
1583         struct lnet_peer_net *lpn;
1584         struct lnet_peer_ni *lpni;
1585         int rc = 0;
1586
1587         LASSERT(nid);
1588
1589         /*
1590          * No need for the lnet_net_lock here, because the
1591          * lnet_api_mutex is held.
1592          */
1593         lpni = lnet_peer_ni_find_locked(nid);
1594         if (lpni) {
1595                 /* A peer with this NID already exists. */
1596                 lp = lpni->lpni_peer_net->lpn_peer;
1597                 lnet_peer_ni_decref_locked(lpni);
1598                 /*
1599                  * This is an error if the peer was configured and the
1600                  * primary NID differs or an attempt is made to change
1601                  * the Multi-Rail flag. Otherwise the assumption is
1602                  * that an existing peer is being modified.
1603                  */
1604                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1605                         if (!nid_same(&lp->lp_primary_nid, nid))
1606                                 rc = -EEXIST;
1607                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1608                                 rc = -EPERM;
1609                         goto out;
1610                 } else if (!(flags & LNET_PEER_CONFIGURED)) {
1611                         if (nid_same(&lp->lp_primary_nid, nid)) {
1612                                 rc = -EEXIST;
1613                                 goto out;
1614                         }
1615                 }
1616                 /* Delete and recreate as a configured peer. */
1617                 rc = lnet_peer_del(lp);
1618                 if (rc)
1619                         goto out;
1620         }
1621
1622         /* Create peer, peer_net, and peer_ni. */
1623         rc = -ENOMEM;
1624         lp = lnet_peer_alloc(nid);
1625         if (!lp)
1626                 goto out;
1627         lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1628         if (!lpn)
1629                 goto out_free_lp;
1630         lpni = lnet_peer_ni_alloc(nid);
1631         if (!lpni)
1632                 goto out_free_lpn;
1633
1634         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1635
1636 out_free_lpn:
1637         LIBCFS_FREE(lpn, sizeof(*lpn));
1638 out_free_lp:
1639         LIBCFS_FREE(lp, sizeof(*lp));
1640 out:
1641         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1642                libcfs_nidstr(nid), flags, rc);
1643         return rc;
1644 }
1645
1646 /*
1647  * Add a NID to a peer. Call with ln_api_mutex held.
1648  *
1649  * Error codes:
1650  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1651  *  -EEXIST:   The NID was configured by DLC for a different peer.
1652  *  -ENOMEM:   Out of memory.
1653  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1654  *             non-multi-rail peer.
1655  */
1656 static int
1657 lnet_peer_add_nid(struct lnet_peer *lp, struct lnet_nid *nid,
1658                   unsigned int flags)
1659 {
1660         struct lnet_peer_net *lpn;
1661         struct lnet_peer_ni *lpni;
1662         int rc = 0;
1663
1664         LASSERT(lp);
1665         LASSERT(nid);
1666
1667         /* A configured peer can only be updated through configuration. */
1668         if (!(flags & LNET_PEER_CONFIGURED)) {
1669                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1670                         rc = -EPERM;
1671                         goto out;
1672                 }
1673         }
1674
1675         /*
1676          * The MULTI_RAIL flag can be set but not cleared, because
1677          * that would leave the peer struct in an invalid state.
1678          */
1679         if (flags & LNET_PEER_MULTI_RAIL) {
1680                 spin_lock(&lp->lp_lock);
1681                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1682                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1683                         lnet_peer_clr_non_mr_pref_nids(lp);
1684                 }
1685                 spin_unlock(&lp->lp_lock);
1686         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1687                 rc = -EPERM;
1688                 goto out;
1689         }
1690
1691         lpni = lnet_peer_ni_find_locked(nid);
1692         if (lpni) {
1693                 /*
1694                  * A peer_ni already exists. This is only a problem if
1695                  * it is not connected to this peer and was configured
1696                  * by DLC.
1697                  */
1698                 if (lpni->lpni_peer_net->lpn_peer == lp)
1699                         goto out_free_lpni;
1700                 if (lnet_peer_ni_is_configured(lpni)) {
1701                         rc = -EEXIST;
1702                         goto out_free_lpni;
1703                 }
1704                 /* If this is the primary NID, destroy the peer. */
1705                 if (lnet_peer_ni_is_primary(lpni)) {
1706                         struct lnet_peer *rtr_lp =
1707                                 lpni->lpni_peer_net->lpn_peer;
1708                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1709                         /*
1710                          * if we're trying to delete a router it means
1711                          * we're moving this peer NI to a new peer so must
1712                          * transfer router properties to the new peer
1713                          */
1714                         if (rtr_refcount > 0) {
1715                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1716                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1717                         }
1718                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1719                         lnet_peer_ni_decref_locked(lpni);
1720                         lpni = lnet_peer_ni_alloc(nid);
1721                         if (!lpni) {
1722                                 rc = -ENOMEM;
1723                                 goto out_free_lpni;
1724                         }
1725                 }
1726         } else {
1727                 lpni = lnet_peer_ni_alloc(nid);
1728                 if (!lpni) {
1729                         rc = -ENOMEM;
1730                         goto out_free_lpni;
1731                 }
1732         }
1733
1734         /*
1735          * Get the peer_net. Check that we're not adding a second
1736          * peer_ni on a peer_net of a non-multi-rail peer.
1737          */
1738         lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
1739         if (!lpn) {
1740                 lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1741                 if (!lpn) {
1742                         rc = -ENOMEM;
1743                         goto out_free_lpni;
1744                 }
1745         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1746                 rc = -ENOTUNIQ;
1747                 goto out_free_lpni;
1748         }
1749
1750         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1751
1752 out_free_lpni:
1753         lnet_peer_ni_decref_locked(lpni);
1754 out:
1755         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1756                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid),
1757                flags, rc);
1758         return rc;
1759 }
1760
1761 /*
1762  * Update the primary NID of a peer, if possible.
1763  *
1764  * Call with the lnet_api_mutex held.
1765  */
1766 static int
1767 lnet_peer_set_primary_nid(struct lnet_peer *lp, struct lnet_nid *nid,
1768                           unsigned int flags)
1769 {
1770         struct lnet_nid old = lp->lp_primary_nid;
1771         int rc = 0;
1772
1773         if (nid_same(&lp->lp_primary_nid, nid))
1774                 goto out;
1775
1776         lp->lp_primary_nid = *nid;
1777
1778         rc = lnet_peer_add_nid(lp, nid, flags);
1779         if (rc) {
1780                 lp->lp_primary_nid = old;
1781                 goto out;
1782         }
1783 out:
1784         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1785                libcfs_nidstr(&old), libcfs_nidstr(nid), rc);
1786
1787         return rc;
1788 }
1789
1790 /*
1791  * lpni creation initiated due to traffic either sending or receiving.
1792  * Callers must hold ln_api_mutex
1793  * Ref taken on lnet_peer_ni returned by this function
1794  */
1795 static struct lnet_peer_ni *
1796 lnet_peer_ni_traffic_add(struct lnet_nid *nid, struct lnet_nid *pref)
1797 __must_hold(&the_lnet.ln_api_mutex)
1798 {
1799         struct lnet_peer *lp = NULL;
1800         struct lnet_peer_net *lpn = NULL;
1801         struct lnet_peer_ni *lpni;
1802         unsigned flags = 0;
1803         int rc = 0;
1804
1805         if (LNET_NID_IS_ANY(nid)) {
1806                 rc = -EINVAL;
1807                 goto out_err;
1808         }
1809
1810         /* lnet_net_lock is not needed here because ln_api_lock is held */
1811         lpni = lnet_peer_ni_find_locked(nid);
1812         if (lpni) {
1813                 /*
1814                  * We must have raced with another thread. Since we
1815                  * know next to nothing about a peer_ni created by
1816                  * traffic, we just assume everything is ok and
1817                  * return.
1818                  */
1819                 goto out;
1820         }
1821
1822         /* Create peer, peer_net, and peer_ni. */
1823         rc = -ENOMEM;
1824         lp = lnet_peer_alloc(nid);
1825         if (!lp)
1826                 goto out_err;
1827         lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1828         if (!lpn)
1829                 goto out_err;
1830         lpni = lnet_peer_ni_alloc(nid);
1831         if (!lpni)
1832                 goto out_err;
1833         lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1834
1835         /* lnet_peer_attach_peer_ni() always returns 0 */
1836         rc = lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1837
1838         lnet_peer_ni_addref_locked(lpni);
1839
1840 out_err:
1841         if (rc) {
1842                 if (lpn)
1843                         LIBCFS_FREE(lpn, sizeof(*lpn));
1844                 if (lp)
1845                         LIBCFS_FREE(lp, sizeof(*lp));
1846                 lpni = ERR_PTR(rc);
1847         }
1848 out:
1849         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(nid), rc);
1850         return lpni;
1851 }
1852
1853 /*
1854  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1855  *
1856  * This API handles the following combinations:
1857  *   Create a peer with its primary NI if only the prim_nid is provided
1858  *   Add a NID to a peer identified by the prim_nid. The peer identified
1859  *   by the prim_nid must already exist.
1860  *   The peer being created may be non-MR.
1861  *
1862  * The caller must hold ln_api_mutex. This prevents the peer from
1863  * being created/modified/deleted by a different thread.
1864  */
1865 int
1866 lnet_add_peer_ni(struct lnet_nid *prim_nid, struct lnet_nid *nid, bool mr,
1867                  bool temp)
1868 {
1869         struct lnet_peer *lp = NULL;
1870         struct lnet_peer_ni *lpni;
1871         unsigned int flags = 0;
1872
1873         /* The prim_nid must always be specified */
1874         if (LNET_NID_IS_ANY(prim_nid))
1875                 return -EINVAL;
1876
1877         if (!temp)
1878                 flags = LNET_PEER_CONFIGURED;
1879
1880         if (mr)
1881                 flags |= LNET_PEER_MULTI_RAIL;
1882
1883         /*
1884          * If nid isn't specified, we must create a new peer with
1885          * prim_nid as its primary nid.
1886          */
1887         if (LNET_NID_IS_ANY(nid))
1888                 return lnet_peer_add(prim_nid, flags);
1889
1890         /* Look up the prim_nid, which must exist. */
1891         lpni = lnet_peer_ni_find_locked(prim_nid);
1892         if (!lpni)
1893                 return -ENOENT;
1894         lnet_peer_ni_decref_locked(lpni);
1895         lp = lpni->lpni_peer_net->lpn_peer;
1896
1897         /* Peer must have been configured. */
1898         if (!temp && !(lp->lp_state & LNET_PEER_CONFIGURED)) {
1899                 CDEBUG(D_NET, "peer %s was not configured\n",
1900                        libcfs_nidstr(prim_nid));
1901                 return -ENOENT;
1902         }
1903
1904         /* Primary NID must match */
1905         if (!nid_same(&lp->lp_primary_nid, prim_nid)) {
1906                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1907                        libcfs_nidstr(prim_nid),
1908                        libcfs_nidstr(&lp->lp_primary_nid));
1909                 return -ENODEV;
1910         }
1911
1912         /* Multi-Rail flag must match. */
1913         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1914                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1915                        libcfs_nidstr(prim_nid));
1916                 return -EPERM;
1917         }
1918
1919         return lnet_peer_add_nid(lp, nid, flags);
1920 }
1921
1922 /*
1923  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1924  *
1925  * This API handles the following combinations:
1926  *   Delete a NI from a peer if both prim_nid and nid are provided.
1927  *   Delete a peer if only prim_nid is provided.
1928  *   Delete a peer if its primary nid is provided.
1929  *
1930  * The caller must hold ln_api_mutex. This prevents the peer from
1931  * being modified/deleted by a different thread.
1932  */
1933 int
1934 lnet_del_peer_ni(struct lnet_nid *prim_nid, struct lnet_nid *nid)
1935 {
1936         struct lnet_peer *lp;
1937         struct lnet_peer_ni *lpni;
1938         unsigned int flags;
1939
1940         if (!prim_nid || LNET_NID_IS_ANY(prim_nid))
1941                 return -EINVAL;
1942
1943         lpni = lnet_peer_ni_find_locked(prim_nid);
1944         if (!lpni)
1945                 return -ENOENT;
1946         lnet_peer_ni_decref_locked(lpni);
1947         lp = lpni->lpni_peer_net->lpn_peer;
1948
1949         if (!nid_same(prim_nid, &lp->lp_primary_nid)) {
1950                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1951                        libcfs_nidstr(prim_nid),
1952                        libcfs_nidstr(&lp->lp_primary_nid));
1953                 return -ENODEV;
1954         }
1955
1956         lnet_net_lock(LNET_LOCK_EX);
1957         if (lp->lp_rtr_refcount > 0) {
1958                 lnet_net_unlock(LNET_LOCK_EX);
1959                 CERROR("%s is a router. Can not be deleted\n",
1960                        libcfs_nidstr(prim_nid));
1961                 return -EBUSY;
1962         }
1963         lnet_net_unlock(LNET_LOCK_EX);
1964
1965         if (LNET_NID_IS_ANY(nid) || nid_same(nid, &lp->lp_primary_nid))
1966                 return lnet_peer_del(lp);
1967
1968         flags = LNET_PEER_CONFIGURED;
1969         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1970                 flags |= LNET_PEER_MULTI_RAIL;
1971
1972         return lnet_peer_del_nid(lp, nid, flags);
1973 }
1974
1975 void
1976 lnet_destroy_peer_ni_locked(struct kref *ref)
1977 {
1978         struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
1979                                                  lpni_kref);
1980         struct lnet_peer_table *ptable;
1981         struct lnet_peer_net *lpn;
1982
1983         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
1984
1985         LASSERT(kref_read(&lpni->lpni_kref) == 0);
1986         LASSERT(list_empty(&lpni->lpni_txq));
1987         LASSERT(lpni->lpni_txqnob == 0);
1988         LASSERT(list_empty(&lpni->lpni_peer_nis));
1989         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1990
1991         lpn = lpni->lpni_peer_net;
1992         lpni->lpni_peer_net = NULL;
1993         lpni->lpni_net = NULL;
1994
1995         if (!list_empty(&lpni->lpni_hashlist)) {
1996                 /* remove the peer ni from the zombie list */
1997                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1998                 spin_lock(&ptable->pt_zombie_lock);
1999                 list_del_init(&lpni->lpni_hashlist);
2000                 ptable->pt_zombies--;
2001                 spin_unlock(&ptable->pt_zombie_lock);
2002         }
2003
2004         if (lpni->lpni_pref_nnids > 1) {
2005                 struct lnet_nid_list *ne, *tmp;
2006
2007                 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
2008                                          nl_list) {
2009                         list_del_init(&ne->nl_list);
2010                         LIBCFS_FREE(ne, sizeof(*ne));
2011                 }
2012         }
2013         LIBCFS_FREE(lpni, sizeof(*lpni));
2014
2015         if (lpn)
2016                 lnet_peer_net_decref_locked(lpn);
2017 }
2018
2019 struct lnet_peer_ni *
2020 lnet_nid2peerni_ex(struct lnet_nid *nid)
2021 __must_hold(&the_lnet.ln_api_mutex)
2022 {
2023         struct lnet_peer_ni *lpni = NULL;
2024
2025         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2026                 return ERR_PTR(-ESHUTDOWN);
2027
2028         /*
2029          * find if a peer_ni already exists.
2030          * If so then just return that.
2031          */
2032         lpni = lnet_peer_ni_find_locked(nid);
2033         if (lpni)
2034                 return lpni;
2035
2036         lnet_net_unlock(LNET_LOCK_EX);
2037
2038         lpni = lnet_peer_ni_traffic_add(nid, NULL);
2039
2040         lnet_net_lock(LNET_LOCK_EX);
2041
2042         return lpni;
2043 }
2044
2045 /*
2046  * Get a peer_ni for the given nid, create it if necessary. Takes a
2047  * hold on the peer_ni.
2048  */
2049 struct lnet_peer_ni *
2050 lnet_peerni_by_nid_locked(struct lnet_nid *nid,
2051                         struct lnet_nid *pref, int cpt)
2052 {
2053         struct lnet_peer_ni *lpni = NULL;
2054
2055         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2056                 return ERR_PTR(-ESHUTDOWN);
2057
2058         /*
2059          * find if a peer_ni already exists.
2060          * If so then just return that.
2061          */
2062         lpni = lnet_peer_ni_find_locked(nid);
2063         if (lpni)
2064                 return lpni;
2065
2066         /*
2067          * Slow path:
2068          * use the lnet_api_mutex to serialize the creation of the peer_ni
2069          * and the creation/deletion of the local ni/net. When a local ni is
2070          * created, if there exists a set of peer_nis on that network,
2071          * they need to be traversed and updated. When a local NI is
2072          * deleted, which could result in a network being deleted, then
2073          * all peer nis on that network need to be removed as well.
2074          *
2075          * Creation through traffic should also be serialized with
2076          * creation through DLC.
2077          */
2078         lnet_net_unlock(cpt);
2079         mutex_lock(&the_lnet.ln_api_mutex);
2080         /*
2081          * the_lnet.ln_state is only modified under the ln_api_lock, so a single
2082          * check here is sufficent.
2083          */
2084         if (the_lnet.ln_state == LNET_STATE_RUNNING)
2085                 lpni = lnet_peer_ni_traffic_add(nid, pref);
2086
2087         mutex_unlock(&the_lnet.ln_api_mutex);
2088         lnet_net_lock(cpt);
2089
2090         /* Lock has been dropped, check again for shutdown. */
2091         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2092                 if (!IS_ERR_OR_NULL(lpni))
2093                         lnet_peer_ni_decref_locked(lpni);
2094                 lpni = ERR_PTR(-ESHUTDOWN);
2095         }
2096
2097         return lpni;
2098 }
2099
2100 bool
2101 lnet_peer_gw_discovery(struct lnet_peer *lp)
2102 {
2103         bool rc = false;
2104
2105         spin_lock(&lp->lp_lock);
2106         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2107                 rc = true;
2108         spin_unlock(&lp->lp_lock);
2109
2110         return rc;
2111 }
2112
2113 bool
2114 lnet_peer_is_uptodate(struct lnet_peer *lp)
2115 {
2116         bool rc;
2117
2118         spin_lock(&lp->lp_lock);
2119         rc = lnet_peer_is_uptodate_locked(lp);
2120         spin_unlock(&lp->lp_lock);
2121         return rc;
2122 }
2123
2124 /*
2125  * Is a peer uptodate from the point of view of discovery?
2126  *
2127  * If it is currently being processed, obviously not.
2128  * A forced Ping or Push is also handled by the discovery thread.
2129  *
2130  * Otherwise look at whether the peer needs rediscovering.
2131  */
2132 bool
2133 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2134 __must_hold(&lp->lp_lock)
2135 {
2136         bool rc;
2137
2138         if (lp->lp_state & (LNET_PEER_DISCOVERING |
2139                             LNET_PEER_FORCE_PING |
2140                             LNET_PEER_FORCE_PUSH)) {
2141                 rc = false;
2142         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2143                 rc = false;
2144         } else if (lnet_peer_needs_push(lp)) {
2145                 rc = false;
2146         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2147                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2148                         rc = true;
2149                 else
2150                         rc = false;
2151         } else {
2152                 rc = false;
2153         }
2154
2155         return rc;
2156 }
2157
2158 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2159 void
2160 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2161 {
2162         /* The discovery thread holds net_lock/EX and lp_lock when it splices
2163          * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2164          * when adding to the list and queuing the peer to ensure that we do not
2165          * strand any messages on the lp_dc_pendq. This scheme ensures the
2166          * message will be resent even if the peer is already being discovered.
2167          * Therefore we needn't check the return value of
2168          * lnet_peer_queue_for_discovery(lp).
2169          */
2170         lnet_net_lock(LNET_LOCK_EX);
2171         spin_lock(&lp->lp_lock);
2172         list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2173         spin_unlock(&lp->lp_lock);
2174         lnet_peer_queue_for_discovery(lp);
2175         lnet_net_unlock(LNET_LOCK_EX);
2176 }
2177
2178 /*
2179  * Queue a peer for the attention of the discovery thread.  Call with
2180  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2181  * -EALREADY if the peer was already queued.
2182  */
2183 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2184 {
2185         int rc;
2186
2187         spin_lock(&lp->lp_lock);
2188         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2189                 lp->lp_state |= LNET_PEER_DISCOVERING;
2190         spin_unlock(&lp->lp_lock);
2191         if (list_empty(&lp->lp_dc_list)) {
2192                 lnet_peer_addref_locked(lp);
2193                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2194                 wake_up(&the_lnet.ln_dc_waitq);
2195                 rc = 0;
2196         } else {
2197                 rc = -EALREADY;
2198         }
2199
2200         CDEBUG(D_NET, "Queue peer %s: %d\n",
2201                libcfs_nidstr(&lp->lp_primary_nid), rc);
2202
2203         return rc;
2204 }
2205
2206 /*
2207  * Discovery of a peer is complete. Wake all waiters on the peer.
2208  * Call with lnet_net_lock/EX held.
2209  */
2210 static void lnet_peer_discovery_complete(struct lnet_peer *lp, int dc_error)
2211 {
2212         struct lnet_msg *msg, *tmp;
2213         int rc = 0;
2214         LIST_HEAD(pending_msgs);
2215
2216         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2217                libcfs_nidstr(&lp->lp_primary_nid));
2218
2219         list_del_init(&lp->lp_dc_list);
2220         spin_lock(&lp->lp_lock);
2221         if (dc_error) {
2222                 lp->lp_dc_error = dc_error;
2223                 lp->lp_state &= ~LNET_PEER_DISCOVERING;
2224                 lp->lp_state |= LNET_PEER_REDISCOVER;
2225         }
2226         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2227         spin_unlock(&lp->lp_lock);
2228         wake_up(&lp->lp_dc_waitq);
2229
2230         if (lp->lp_rtr_refcount > 0)
2231                 lnet_router_discovery_complete(lp);
2232
2233         lnet_net_unlock(LNET_LOCK_EX);
2234
2235         /* iterate through all pending messages and send them again */
2236         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2237                 list_del_init(&msg->msg_list);
2238                 if (dc_error) {
2239                         lnet_finalize(msg, dc_error);
2240                         continue;
2241                 }
2242
2243                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2244                        lnet_msgtyp2str(msg->msg_type),
2245                        libcfs_idstr(&msg->msg_target));
2246                 rc = lnet_send(&msg->msg_src_nid_param, msg,
2247                                &msg->msg_rtr_nid_param);
2248                 if (rc < 0) {
2249                         CNETERR("Error sending %s to %s: %d\n",
2250                                lnet_msgtyp2str(msg->msg_type),
2251                                libcfs_idstr(&msg->msg_target), rc);
2252                         lnet_finalize(msg, rc);
2253                 }
2254         }
2255         lnet_net_lock(LNET_LOCK_EX);
2256         lnet_peer_decref_locked(lp);
2257 }
2258
2259 /*
2260  * Handle inbound push.
2261  * Like any event handler, called with lnet_res_lock/CPT held.
2262  */
2263 void lnet_peer_push_event(struct lnet_event *ev)
2264 {
2265         struct lnet_ping_buffer *pbuf;
2266         struct lnet_peer *lp;
2267
2268         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2269
2270         /* lnet_find_peer() adds a refcount */
2271         lp = lnet_find_peer(&ev->source.nid);
2272         if (!lp) {
2273                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2274                        libcfs_nidstr(&ev->initiator.nid),
2275                        libcfs_nidstr(&ev->source.nid));
2276                 pbuf->pb_needs_post = true;
2277                 return;
2278         }
2279
2280         /* Ensure peer state remains consistent while we modify it. */
2281         spin_lock(&lp->lp_lock);
2282
2283         /*
2284          * If some kind of error happened the contents of the message
2285          * cannot be used. Clear the NIDS_UPTODATE and set the
2286          * FORCE_PING flag to trigger a ping.
2287          */
2288         if (ev->status) {
2289                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2290                 lp->lp_state |= LNET_PEER_FORCE_PING;
2291                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2292                        ev->status,
2293                        libcfs_nidstr(&lp->lp_primary_nid),
2294                        libcfs_nidstr(&ev->source.nid));
2295                 goto out;
2296         }
2297
2298         /*
2299          * A push with invalid or corrupted info. Clear the UPTODATE
2300          * flag to trigger a ping.
2301          */
2302         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2303                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2304                 lp->lp_state |= LNET_PEER_FORCE_PING;
2305                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2306                        libcfs_nidstr(&lp->lp_primary_nid));
2307                 goto out;
2308         }
2309
2310         /*
2311          * Make sure we'll allocate the correct size ping buffer when
2312          * pinging the peer.
2313          */
2314         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2315                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2316
2317         /*
2318          * A non-Multi-Rail peer is not supposed to be capable of
2319          * sending a push.
2320          */
2321         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2322                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2323                        libcfs_nidstr(&lp->lp_primary_nid));
2324                 goto out;
2325         }
2326
2327         /*
2328          * The peer may have discovery disabled at its end. Set
2329          * NO_DISCOVERY as appropriate.
2330          */
2331         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2332                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2333                        libcfs_nidstr(&lp->lp_primary_nid));
2334                 /*
2335                  * Mark the peer for deletion if we already know about it
2336                  * and it's going from discovery set to no discovery set
2337                  */
2338                 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2339                                       LNET_PEER_DISCOVERING)) &&
2340                      lp->lp_state & LNET_PEER_DISCOVERED) {
2341                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2342                                libcfs_nidstr(&lp->lp_primary_nid),
2343                                lp->lp_state);
2344                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2345                 }
2346                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2347         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2348                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2349                        libcfs_nidstr(&lp->lp_primary_nid));
2350                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2351         }
2352
2353         /*
2354          * Update the MULTI_RAIL flag based on the push. If the peer
2355          * was configured with DLC then the setting should match what
2356          * DLC put in.
2357          * NB: We verified above that the MR feature bit is set in pi_features
2358          */
2359         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2360                 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2361                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2362         } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2363                 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2364                       libcfs_nidstr(&lp->lp_primary_nid));
2365         } else if (lnet_peer_discovery_disabled) {
2366                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2367                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2368         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2369                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2370                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2371         } else {
2372                 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2373                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2374                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2375                 lnet_peer_clr_non_mr_pref_nids(lp);
2376         }
2377
2378         /*
2379          * Check for truncation of the Put message. Clear the
2380          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2381          * and tell discovery to allocate a bigger buffer.
2382          */
2383         if (ev->mlength < ev->rlength) {
2384                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2385                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2386                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2387                 lp->lp_state |= LNET_PEER_FORCE_PING;
2388                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2389                        libcfs_nidstr(&lp->lp_primary_nid),
2390                        pbuf->pb_info.pi_nnis);
2391                 goto out;
2392         }
2393
2394         /* always assume new data */
2395         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2396         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2397
2398         /*
2399          * If there is data present that hasn't been processed yet,
2400          * we'll replace it if the Put contained newer data and it
2401          * fits. We're racing with a Ping or earlier Push in this
2402          * case.
2403          */
2404         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2405                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2406                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2407                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2408                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2409                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2410                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2411                               libcfs_nidstr(&lp->lp_primary_nid),
2412                               LNET_PING_BUFFER_SEQNO(pbuf),
2413                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2414                 }
2415                 goto out;
2416         }
2417
2418         /*
2419          * Allocate a buffer to copy the data. On a failure we drop
2420          * the Push and set FORCE_PING to force the discovery
2421          * thread to fix the problem by pinging the peer.
2422          */
2423         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2424         if (!lp->lp_data) {
2425                 lp->lp_state |= LNET_PEER_FORCE_PING;
2426                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2427                        libcfs_nidstr(&lp->lp_primary_nid),
2428                        LNET_PING_BUFFER_SEQNO(pbuf));
2429                 goto out;
2430         }
2431
2432         /* Success */
2433         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2434                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2435         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2436         CDEBUG(D_NET, "Received Push %s %u\n",
2437                libcfs_nidstr(&lp->lp_primary_nid),
2438                LNET_PING_BUFFER_SEQNO(pbuf));
2439
2440 out:
2441         /* We've processed this buffer. It can be reposted */
2442         pbuf->pb_needs_post = true;
2443
2444         /*
2445          * Queue the peer for discovery if not done, force it on the request
2446          * queue and wake the discovery thread if the peer was already queued,
2447          * because its status changed.
2448          */
2449         spin_unlock(&lp->lp_lock);
2450         lnet_net_lock(LNET_LOCK_EX);
2451         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2452                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2453                 wake_up(&the_lnet.ln_dc_waitq);
2454         }
2455         /* Drop refcount from lookup */
2456         lnet_peer_decref_locked(lp);
2457         lnet_net_unlock(LNET_LOCK_EX);
2458 }
2459
2460 /*
2461  * Clear the discovery error state, unless we're already discovering
2462  * this peer, in which case the error is current.
2463  */
2464 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2465 {
2466         spin_lock(&lp->lp_lock);
2467         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2468                 lp->lp_dc_error = 0;
2469         spin_unlock(&lp->lp_lock);
2470 }
2471
2472 /*
2473  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2474  * dropped/retaken within this function. An lnet_peer_ni is passed in
2475  * because discovery could tear down an lnet_peer.
2476  */
2477 int
2478 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2479 {
2480         DEFINE_WAIT(wait);
2481         struct lnet_peer *lp;
2482         int rc = 0;
2483         int count = 0;
2484
2485 again:
2486         lnet_net_unlock(cpt);
2487         lnet_net_lock(LNET_LOCK_EX);
2488         lp = lpni->lpni_peer_net->lpn_peer;
2489         lnet_peer_clear_discovery_error(lp);
2490
2491         /*
2492          * We're willing to be interrupted. The lpni can become a
2493          * zombie if we race with DLC, so we must check for that.
2494          */
2495         for (;;) {
2496                 /* Keep lp alive when the lnet_net_lock is unlocked */
2497                 lnet_peer_addref_locked(lp);
2498                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2499                 if (signal_pending(current))
2500                         break;
2501                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2502                         break;
2503                 /*
2504                  * Don't repeat discovery if discovery is disabled. This is
2505                  * done to ensure we can use discovery as a standard ping as
2506                  * well for backwards compatibility with routers which do not
2507                  * have discovery or have discovery disabled
2508                  */
2509                 if (lnet_is_discovery_disabled(lp) && count > 0)
2510                         break;
2511                 if (lp->lp_dc_error)
2512                         break;
2513                 if (lnet_peer_is_uptodate(lp))
2514                         break;
2515                 if (lp->lp_state & LNET_PEER_MARK_DELETED)
2516                         break;
2517                 lnet_peer_queue_for_discovery(lp);
2518                 count++;
2519                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2520
2521                 /*
2522                  * If caller requested a non-blocking operation then
2523                  * return immediately. Once discovery is complete any
2524                  * pending messages that were stopped due to discovery
2525                  * will be transmitted.
2526                  */
2527                 if (!block)
2528                         break;
2529
2530                 lnet_net_unlock(LNET_LOCK_EX);
2531                 schedule();
2532                 finish_wait(&lp->lp_dc_waitq, &wait);
2533                 lnet_net_lock(LNET_LOCK_EX);
2534                 lnet_peer_decref_locked(lp);
2535                 /* Peer may have changed */
2536                 lp = lpni->lpni_peer_net->lpn_peer;
2537         }
2538         finish_wait(&lp->lp_dc_waitq, &wait);
2539
2540         lnet_net_unlock(LNET_LOCK_EX);
2541         lnet_net_lock(cpt);
2542         lnet_peer_decref_locked(lp);
2543         /*
2544          * The peer may have changed, so re-check and rediscover if that turns
2545          * out to have been the case. The reference count on lp ensured that
2546          * even if it was unlinked from lpni the memory could not be recycled.
2547          * Thus the check below is sufficient to determine whether the peer
2548          * changed. If the peer changed, then lp must not be dereferenced.
2549          */
2550         if (lp != lpni->lpni_peer_net->lpn_peer)
2551                 goto again;
2552
2553         if (signal_pending(current))
2554                 rc = -EINTR;
2555         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2556                 rc = -ESHUTDOWN;
2557         else if (lp->lp_dc_error)
2558                 rc = lp->lp_dc_error;
2559         else if (!block)
2560                 CDEBUG(D_NET, "non-blocking discovery\n");
2561         else if (!lnet_peer_is_uptodate(lp) &&
2562                  !(lnet_is_discovery_disabled(lp) ||
2563                    (lp->lp_state & LNET_PEER_MARK_DELETED)))
2564                 goto again;
2565
2566         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2567                (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2568                libcfs_nidstr(&lpni->lpni_nid), rc,
2569                (!block) ? "pending discovery" : "discovery complete");
2570
2571         return rc;
2572 }
2573
2574 /* Handle an incoming ack for a push. */
2575 static void
2576 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2577 {
2578         struct lnet_ping_buffer *pbuf;
2579
2580         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2581         spin_lock(&lp->lp_lock);
2582         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2583         lp->lp_push_error = ev->status;
2584         if (ev->status)
2585                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2586         else
2587                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2588         spin_unlock(&lp->lp_lock);
2589
2590         CDEBUG(D_NET, "peer %s ev->status %d\n",
2591                libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2592 }
2593
2594 /* Handle a Reply message. This is the reply to a Ping message. */
2595 static void
2596 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2597 {
2598         struct lnet_ping_buffer *pbuf;
2599         int rc;
2600
2601         spin_lock(&lp->lp_lock);
2602
2603         lp->lp_disc_src_nid = ev->target.nid;
2604         lp->lp_disc_dst_nid = ev->source.nid;
2605
2606         /*
2607          * If some kind of error happened the contents of message
2608          * cannot be used. Set PING_FAILED to trigger a retry.
2609          */
2610         if (ev->status) {
2611                 lp->lp_state |= LNET_PEER_PING_FAILED;
2612                 lp->lp_ping_error = ev->status;
2613                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2614                        ev->status,
2615                        libcfs_nidstr(&lp->lp_primary_nid),
2616                        libcfs_nidstr(&ev->source.nid));
2617                 goto out;
2618         }
2619
2620         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2621         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2622                 lnet_swap_pinginfo(pbuf);
2623
2624         /*
2625          * A reply with invalid or corrupted info. Set PING_FAILED to
2626          * trigger a retry.
2627          */
2628         rc = lnet_ping_info_validate(&pbuf->pb_info);
2629         if (rc) {
2630                 lp->lp_state |= LNET_PEER_PING_FAILED;
2631                 lp->lp_ping_error = 0;
2632                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2633                        libcfs_nidstr(&lp->lp_primary_nid), rc);
2634                 goto out;
2635         }
2636
2637         /*
2638          * The peer may have discovery disabled at its end. Set
2639          * NO_DISCOVERY as appropriate.
2640          */
2641         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) ||
2642             lnet_peer_discovery_disabled) {
2643                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2644                        libcfs_nidstr(&lp->lp_primary_nid));
2645
2646                 /* Detect whether this peer has toggled discovery from on to
2647                  * off and whether we can delete and re-create the peer. Peers
2648                  * that were manually configured cannot be deleted by discovery.
2649                  * We need to delete this peer and re-create it if the peer was
2650                  * not configured manually, is currently considered DD capable,
2651                  * and either:
2652                  * 1. We've already discovered the peer (the peer has toggled
2653                  *    the discovery feature from on to off), or
2654                  * 2. The peer is considered MR, but it was not user configured
2655                  *    (this was a "temporary" peer created via the kernel APIs
2656                  *     that we're discovering for the first time)
2657                  */
2658                 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2659                                       LNET_PEER_NO_DISCOVERY)) &&
2660                     (lp->lp_state & (LNET_PEER_DISCOVERED |
2661                                      LNET_PEER_MULTI_RAIL))) {
2662                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2663                                libcfs_nidstr(&lp->lp_primary_nid),
2664                                lp->lp_state);
2665                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2666                 }
2667                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2668         } else {
2669                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2670                        libcfs_nidstr(&lp->lp_primary_nid));
2671                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2672         }
2673
2674         /*
2675          * Update the MULTI_RAIL flag based on the reply. If the peer
2676          * was configured with DLC then the setting should match what
2677          * DLC put in.
2678          */
2679         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2680                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2681                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2682                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2683                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2684                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2685                               libcfs_nidstr(&lp->lp_primary_nid));
2686                 } else if (lnet_peer_discovery_disabled) {
2687                         CDEBUG(D_NET,
2688                                "peer %s(%p) not MR: DD disabled locally\n",
2689                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2690                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2691                         CDEBUG(D_NET,
2692                                "peer %s(%p) not MR: DD disabled remotely\n",
2693                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2694                 } else {
2695                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2696                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2697                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2698                         lnet_peer_clr_non_mr_pref_nids(lp);
2699                 }
2700         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2701                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2702                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2703                               libcfs_nidstr(&lp->lp_primary_nid));
2704                 } else {
2705                         CERROR("Multi-Rail state vanished from %s\n",
2706                                libcfs_nidstr(&lp->lp_primary_nid));
2707                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2708                 }
2709         }
2710
2711         /*
2712          * Make sure we'll allocate the correct size ping buffer when
2713          * pinging the peer.
2714          */
2715         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2716                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2717
2718         /*
2719          * Check for truncation of the Reply. Clear PING_SENT and set
2720          * PING_FAILED to trigger a retry.
2721          */
2722         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2723                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2724                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2725                 lp->lp_state |= LNET_PEER_PING_FAILED;
2726                 lp->lp_ping_error = 0;
2727                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2728                        libcfs_nidstr(&lp->lp_primary_nid),
2729                        pbuf->pb_info.pi_nnis);
2730                 goto out;
2731         }
2732
2733         /*
2734          * Check the sequence numbers in the reply. These are only
2735          * available if the reply came from a Multi-Rail peer.
2736          */
2737         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2738             pbuf->pb_info.pi_nnis > 1 &&
2739             lnet_nid_to_nid4(&lp->lp_primary_nid) ==
2740             pbuf->pb_info.pi_ni[1].ns_nid) {
2741                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2742                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2743                                 libcfs_nidstr(&lp->lp_primary_nid),
2744                                 LNET_PING_BUFFER_SEQNO(pbuf),
2745                                 lp->lp_peer_seqno);
2746
2747                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2748         }
2749
2750         /* We're happy with the state of the data in the buffer. */
2751         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2752                libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2753                lp->lp_state);
2754         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2755                 lnet_ping_buffer_decref(lp->lp_data);
2756         else
2757                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2758         lnet_ping_buffer_addref(pbuf);
2759         lp->lp_data = pbuf;
2760 out:
2761         lp->lp_state &= ~LNET_PEER_PING_SENT;
2762         spin_unlock(&lp->lp_lock);
2763
2764         lnet_net_lock(LNET_LOCK_EX);
2765         /*
2766          * If this peer is a gateway, call the routing callback to
2767          * handle the ping reply
2768          */
2769         if (lp->lp_rtr_refcount > 0)
2770                 lnet_router_discovery_ping_reply(lp);
2771         lnet_net_unlock(LNET_LOCK_EX);
2772 }
2773
2774 /*
2775  * Send event handling. Only matters for error cases, where we clean
2776  * up state on the peer and peer_ni that would otherwise be updated in
2777  * the REPLY event handler for a successful Ping, and the ACK event
2778  * handler for a successful Push.
2779  */
2780 static int
2781 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2782 {
2783         int rc = 0;
2784
2785         if (!ev->status)
2786                 goto out;
2787
2788         spin_lock(&lp->lp_lock);
2789         if (ev->msg_type == LNET_MSG_GET) {
2790                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2791                 lp->lp_state |= LNET_PEER_PING_FAILED;
2792                 lp->lp_ping_error = ev->status;
2793         } else { /* ev->msg_type == LNET_MSG_PUT */
2794                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2795                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2796                 lp->lp_push_error = ev->status;
2797         }
2798         spin_unlock(&lp->lp_lock);
2799         rc = LNET_REDISCOVER_PEER;
2800 out:
2801         CDEBUG(D_NET, "%s Send to %s: %d\n",
2802                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2803                 libcfs_nidstr(&ev->target.nid), rc);
2804         return rc;
2805 }
2806
2807 /*
2808  * Unlink event handling. This event is only seen if a call to
2809  * LNetMDUnlink() caused the event to be unlinked. If this call was
2810  * made after the event was set up in LNetGet() or LNetPut() then we
2811  * assume the Ping or Push timed out.
2812  */
2813 static void
2814 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2815 {
2816         spin_lock(&lp->lp_lock);
2817         /* We've passed through LNetGet() */
2818         if (lp->lp_state & LNET_PEER_PING_SENT) {
2819                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2820                 lp->lp_state |= LNET_PEER_PING_FAILED;
2821                 lp->lp_ping_error = -ETIMEDOUT;
2822                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2823                         libcfs_nidstr(&lp->lp_primary_nid));
2824         }
2825         /* We've passed through LNetPut() */
2826         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2827                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2828                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2829                 lp->lp_push_error = -ETIMEDOUT;
2830                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2831                         libcfs_nidstr(&lp->lp_primary_nid));
2832         }
2833         spin_unlock(&lp->lp_lock);
2834 }
2835
2836 /*
2837  * Event handler for the discovery EQ.
2838  *
2839  * Called with lnet_res_lock(cpt) held. The cpt is the
2840  * lnet_cpt_of_cookie() of the md handle cookie.
2841  */
2842 static void lnet_discovery_event_handler(struct lnet_event *event)
2843 {
2844         struct lnet_peer *lp = event->md_user_ptr;
2845         struct lnet_ping_buffer *pbuf;
2846         int rc;
2847
2848         /* discovery needs to take another look */
2849         rc = LNET_REDISCOVER_PEER;
2850
2851         CDEBUG(D_NET, "Received event: %d\n", event->type);
2852
2853         switch (event->type) {
2854         case LNET_EVENT_ACK:
2855                 lnet_discovery_event_ack(lp, event);
2856                 break;
2857         case LNET_EVENT_REPLY:
2858                 lnet_discovery_event_reply(lp, event);
2859                 break;
2860         case LNET_EVENT_SEND:
2861                 /* Only send failure triggers a retry. */
2862                 rc = lnet_discovery_event_send(lp, event);
2863                 break;
2864         case LNET_EVENT_UNLINK:
2865                 /* LNetMDUnlink() was called */
2866                 lnet_discovery_event_unlink(lp, event);
2867                 break;
2868         default:
2869                 /* Invalid events. */
2870                 LBUG();
2871         }
2872         lnet_net_lock(LNET_LOCK_EX);
2873         if (event->unlinked) {
2874                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2875                 lnet_ping_buffer_decref(pbuf);
2876                 lnet_peer_decref_locked(lp);
2877         }
2878
2879         /* put peer back at end of request queue, if discovery not already
2880          * done */
2881         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2882             lnet_peer_queue_for_discovery(lp)) {
2883                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2884                 wake_up(&the_lnet.ln_dc_waitq);
2885         }
2886         lnet_net_unlock(LNET_LOCK_EX);
2887 }
2888
2889 /*
2890  * Build a peer from incoming data.
2891  *
2892  * The NIDs in the incoming data are supposed to be structured as follows:
2893  *  - loopback
2894  *  - primary NID
2895  *  - other NIDs in same net
2896  *  - NIDs in second net
2897  *  - NIDs in third net
2898  *  - ...
2899  * This due to the way the list of NIDs in the data is created.
2900  *
2901  * Note that this function will mark the peer uptodate unless an
2902  * ENOMEM is encontered. All other errors are due to a conflict
2903  * between the DLC configuration and what discovery sees. We treat DLC
2904  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2905  * peer from becoming stuck in discovery.
2906  */
2907 static int lnet_peer_merge_data(struct lnet_peer *lp,
2908                                 struct lnet_ping_buffer *pbuf)
2909 {
2910         struct lnet_peer_net *lpn;
2911         struct lnet_peer_ni *lpni;
2912         lnet_nid_t *curnis = NULL;
2913         struct lnet_ni_status *addnis = NULL;
2914         lnet_nid_t *delnis = NULL;
2915         struct lnet_nid nid;
2916         unsigned int flags;
2917         int ncurnis;
2918         int naddnis;
2919         int ndelnis;
2920         int nnis = 0;
2921         int i;
2922         int j;
2923         int rc;
2924
2925         flags = LNET_PEER_DISCOVERED;
2926         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2927                 flags |= LNET_PEER_MULTI_RAIL;
2928
2929         /*
2930          * Cache the routing feature for the peer; whether it is enabled
2931          * for disabled as reported by the remote peer.
2932          */
2933         spin_lock(&lp->lp_lock);
2934         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2935                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2936         else
2937                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2938         spin_unlock(&lp->lp_lock);
2939
2940         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2941         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2942         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2943         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2944         if (!curnis || !addnis || !delnis) {
2945                 rc = -ENOMEM;
2946                 goto out;
2947         }
2948         ncurnis = 0;
2949         naddnis = 0;
2950         ndelnis = 0;
2951
2952         /* Construct the list of NIDs present in peer. */
2953         lpni = NULL;
2954         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
2955                 curnis[ncurnis++] = lnet_nid_to_nid4(&lpni->lpni_nid);
2956
2957         /*
2958          * Check for NIDs in pbuf not present in curnis[].
2959          * The loop starts at 1 to skip the loopback NID.
2960          */
2961         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
2962                 for (j = 0; j < ncurnis; j++)
2963                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
2964                                 break;
2965                 if (j == ncurnis)
2966                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
2967         }
2968         /*
2969          * Check for NIDs in curnis[] not present in pbuf.
2970          * The nested loop starts at 1 to skip the loopback NID.
2971          *
2972          * But never add the loopback NID to delnis[]: if it is
2973          * present in curnis[] then this peer is for this node.
2974          */
2975         for (i = 0; i < ncurnis; i++) {
2976                 if (curnis[i] == LNET_NID_LO_0)
2977                         continue;
2978                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
2979                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
2980                                 /*
2981                                  * update the information we cache for the
2982                                  * peer with the latest information we
2983                                  * received
2984                                  */
2985                                 lnet_nid4_to_nid(curnis[i], &nid);
2986                                 lpni = lnet_peer_ni_find_locked(&nid);
2987                                 if (lpni) {
2988                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
2989                                         lnet_peer_ni_decref_locked(lpni);
2990                                 }
2991                                 break;
2992                         }
2993                 }
2994                 if (j == pbuf->pb_info.pi_nnis)
2995                         delnis[ndelnis++] = curnis[i];
2996         }
2997
2998         /*
2999          * If we get here and the discovery is disabled then we don't want
3000          * to add or delete any NIs. We just updated the ones we have some
3001          * information on, and call it a day
3002          */
3003         rc = 0;
3004         if (lnet_is_discovery_disabled(lp))
3005                 goto out;
3006
3007         for (i = 0; i < naddnis; i++) {
3008                 lnet_nid4_to_nid(addnis[i].ns_nid, &nid);
3009                 rc = lnet_peer_add_nid(lp, &nid, flags);
3010                 if (rc) {
3011                         CERROR("Error adding NID %s to peer %s: %d\n",
3012                                libcfs_nid2str(addnis[i].ns_nid),
3013                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3014                         if (rc == -ENOMEM)
3015                                 goto out;
3016                 }
3017                 lpni = lnet_peer_ni_find_locked(&nid);
3018                 if (lpni) {
3019                         lpni->lpni_ns_status = addnis[i].ns_status;
3020                         lnet_peer_ni_decref_locked(lpni);
3021                 }
3022         }
3023
3024         for (i = 0; i < ndelnis; i++) {
3025                 /*
3026                  * for routers it's okay to delete the primary_nid because
3027                  * the upper layers don't really rely on it. So if we're
3028                  * being told that the router changed its primary_nid
3029                  * then it's okay to delete it.
3030                  */
3031                 lnet_nid4_to_nid(delnis[i], &nid);
3032                 if (lp->lp_rtr_refcount > 0)
3033                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3034                 rc = lnet_peer_del_nid(lp, &nid, flags);
3035                 if (rc) {
3036                         CERROR("Error deleting NID %s from peer %s: %d\n",
3037                                libcfs_nid2str(delnis[i]),
3038                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3039                         if (rc == -ENOMEM)
3040                                 goto out;
3041                 }
3042         }
3043
3044         /* The peer net for the primary NID should be the first entry in the
3045          * peer's lp_peer_nets list, and the peer NI for the primary NID should
3046          * be the first entry in its peer net's lpn_peer_nis list.
3047          */
3048         lnet_nid4_to_nid(pbuf->pb_info.pi_ni[1].ns_nid, &nid);
3049         lpni = lnet_peer_ni_find_locked(&nid);
3050         if (!lpni) {
3051                 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3052                        libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
3053                 goto out;
3054         }
3055
3056         lnet_peer_ni_decref_locked(lpni);
3057
3058         lpn = lpni->lpni_peer_net;
3059         if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3060                 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3061
3062         if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3063                 list_move(&lpni->lpni_peer_nis,
3064                           &lpni->lpni_peer_net->lpn_peer_nis);
3065
3066         /*
3067          * Errors other than -ENOMEM are due to peers having been
3068          * configured with DLC. Ignore these because DLC overrides
3069          * Discovery.
3070          */
3071         rc = 0;
3072 out:
3073         CFS_FREE_PTR_ARRAY(curnis, nnis);
3074         CFS_FREE_PTR_ARRAY(addnis, nnis);
3075         CFS_FREE_PTR_ARRAY(delnis, nnis);
3076         lnet_ping_buffer_decref(pbuf);
3077         CDEBUG(D_NET, "peer %s (%p): %d\n",
3078                libcfs_nidstr(&lp->lp_primary_nid), lp, rc);
3079
3080         if (rc) {
3081                 spin_lock(&lp->lp_lock);
3082                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
3083                 lp->lp_state |= LNET_PEER_FORCE_PING;
3084                 spin_unlock(&lp->lp_lock);
3085         }
3086         return rc;
3087 }
3088
3089 /*
3090  * The data in pbuf says lp is its primary peer, but the data was
3091  * received by a different peer. Try to update lp with the data.
3092  */
3093 static int
3094 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
3095 {
3096         struct lnet_handle_md mdh;
3097
3098         /* Queue lp for discovery, and force it on the request queue. */
3099         lnet_net_lock(LNET_LOCK_EX);
3100         if (lnet_peer_queue_for_discovery(lp))
3101                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3102         lnet_net_unlock(LNET_LOCK_EX);
3103
3104         LNetInvalidateMDHandle(&mdh);
3105
3106         /*
3107          * Decide whether we can move the peer to the DATA_PRESENT state.
3108          *
3109          * We replace stale data for a multi-rail peer, repair PING_FAILED
3110          * status, and preempt FORCE_PING.
3111          *
3112          * If after that we have DATA_PRESENT, we merge it into this peer.
3113          */
3114         spin_lock(&lp->lp_lock);
3115         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3116                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3117                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3118                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3119                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3120                         lnet_ping_buffer_decref(pbuf);
3121                         pbuf = lp->lp_data;
3122                         lp->lp_data = NULL;
3123                 }
3124         }
3125         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3126                 lnet_ping_buffer_decref(lp->lp_data);
3127                 lp->lp_data = NULL;
3128                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3129         }
3130         if (lp->lp_state & LNET_PEER_PING_FAILED) {
3131                 mdh = lp->lp_ping_mdh;
3132                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3133                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3134                 lp->lp_ping_error = 0;
3135         }
3136         if (lp->lp_state & LNET_PEER_FORCE_PING)
3137                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3138         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3139         spin_unlock(&lp->lp_lock);
3140
3141         if (!LNetMDHandleIsInvalid(mdh))
3142                 LNetMDUnlink(mdh);
3143
3144         if (pbuf)
3145                 return lnet_peer_merge_data(lp, pbuf);
3146
3147         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3148         return 0;
3149 }
3150
3151 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
3152 {
3153         int i;
3154
3155         for (i = 0; i < pinfo->pi_nnis; i++) {
3156                 if (pinfo->pi_ni[i].ns_nid == nid)
3157                         return true;
3158         }
3159
3160         return false;
3161 }
3162
3163 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3164  * to the discovery queue a reference was taken that will prevent the peer from
3165  * actually being freed by this function. After this function exits the
3166  * discovery thread should call lnet_peer_discovery_complete() which will
3167  * drop that reference as well as wake any waiters that may also be holding a
3168  * ref on the peer
3169  */
3170 static int lnet_peer_deletion(struct lnet_peer *lp)
3171 __must_hold(&lp->lp_lock)
3172 {
3173         struct list_head rlist;
3174         struct lnet_route *route, *tmp;
3175         int sensitivity = lp->lp_health_sensitivity;
3176         int rc;
3177
3178         INIT_LIST_HEAD(&rlist);
3179
3180         lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3181                           LNET_PEER_FORCE_PUSH);
3182         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3183                libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3184
3185         /* no-op if lnet_peer_del() has already been called on this peer */
3186         if (lp->lp_state & LNET_PEER_MARK_DELETED)
3187                 return 0;
3188
3189         spin_unlock(&lp->lp_lock);
3190
3191         mutex_lock(&the_lnet.ln_api_mutex);
3192         if (the_lnet.ln_state != LNET_STATE_RUNNING ||
3193             the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
3194                 mutex_unlock(&the_lnet.ln_api_mutex);
3195                 spin_lock(&lp->lp_lock);
3196                 return -ESHUTDOWN;
3197         }
3198
3199         lnet_net_lock(LNET_LOCK_EX);
3200         /* remove the peer from the discovery work
3201          * queue if it's on there in preparation
3202          * of deleting it.
3203          */
3204         if (!list_empty(&lp->lp_dc_list))
3205                 list_del_init(&lp->lp_dc_list);
3206         list_for_each_entry_safe(route, tmp,
3207                                  &lp->lp_routes,
3208                                  lr_gwlist)
3209                 lnet_move_route(route, NULL, &rlist);
3210         lnet_net_unlock(LNET_LOCK_EX);
3211
3212         /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3213         rc = lnet_peer_del(lp);
3214         if (rc)
3215                 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3216                         libcfs_nidstr(&lp->lp_primary_nid), rc);
3217
3218         list_for_each_entry_safe(route, tmp,
3219                                  &rlist, lr_list) {
3220                 /* re-add these routes */
3221                 lnet_add_route(route->lr_net,
3222                                route->lr_hops,
3223                                &route->lr_nid,
3224                                route->lr_priority,
3225                                sensitivity);
3226                 LIBCFS_FREE(route, sizeof(*route));
3227         }
3228
3229         mutex_unlock(&the_lnet.ln_api_mutex);
3230
3231         spin_lock(&lp->lp_lock);
3232
3233         return 0;
3234 }
3235
3236 /*
3237  * Update a peer using the data received.
3238  */
3239 static int lnet_peer_data_present(struct lnet_peer *lp)
3240 __must_hold(&lp->lp_lock)
3241 {
3242         struct lnet_ping_buffer *pbuf;
3243         struct lnet_peer_ni *lpni;
3244         struct lnet_nid nid;
3245         unsigned int flags;
3246         int rc = 0;
3247
3248         pbuf = lp->lp_data;
3249         lp->lp_data = NULL;
3250         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3251         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3252         spin_unlock(&lp->lp_lock);
3253
3254         /*
3255          * Modifications of peer structures are done while holding the
3256          * ln_api_mutex. A global lock is required because we may be
3257          * modifying multiple peer structures, and a mutex greatly
3258          * simplifies memory management.
3259          *
3260          * The actual changes to the data structures must also protect
3261          * against concurrent lookups, for which the lnet_net_lock in
3262          * LNET_LOCK_EX mode is used.
3263          */
3264         mutex_lock(&the_lnet.ln_api_mutex);
3265         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3266                 rc = -ESHUTDOWN;
3267                 goto out;
3268         }
3269
3270         /*
3271          * If this peer is not on the peer list then it is being torn
3272          * down, and our reference count may be all that is keeping it
3273          * alive. Don't do any work on it.
3274          */
3275         if (list_empty(&lp->lp_peer_list)) {
3276                 lnet_ping_buffer_decref(pbuf);
3277                 goto out;
3278         }
3279
3280         flags = LNET_PEER_DISCOVERED;
3281         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3282                 flags |= LNET_PEER_MULTI_RAIL;
3283
3284         /*
3285          * Check whether the primary NID in the message matches the
3286          * primary NID of the peer. If it does, update the peer, if
3287          * it it does not, check whether there is already a peer with
3288          * that primary NID. If no such peer exists, try to update
3289          * the primary NID of the current peer (allowed if it was
3290          * created due to message traffic) and complete the update.
3291          * If the peer did exist, hand off the data to it.
3292          *
3293          * The peer for the loopback interface is a special case: this
3294          * is the peer for the local node, and we want to set its
3295          * primary NID to the correct value here. Moreover, this peer
3296          * can show up with only the loopback NID in the ping buffer.
3297          */
3298         if (pbuf->pb_info.pi_nnis <= 1) {
3299                 lnet_ping_buffer_decref(pbuf);
3300                 goto out;
3301         }
3302         lnet_nid4_to_nid(pbuf->pb_info.pi_ni[1].ns_nid, &nid);
3303         if (nid_is_lo0(&lp->lp_primary_nid)) {
3304                 rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3305                 if (rc)
3306                         lnet_ping_buffer_decref(pbuf);
3307                 else
3308                         rc = lnet_peer_merge_data(lp, pbuf);
3309         /*
3310          * if the primary nid of the peer is present in the ping info returned
3311          * from the peer, but it's not the local primary peer we have
3312          * cached and discovery is disabled, then we don't want to update
3313          * our local peer info, by adding or removing NIDs, we just want
3314          * to update the status of the nids that we currently have
3315          * recorded in that peer.
3316          */
3317         } else if (nid_same(&lp->lp_primary_nid, &nid) ||
3318                    (lnet_is_nid_in_ping_info(lnet_nid_to_nid4(&lp->lp_primary_nid),
3319                                              &pbuf->pb_info) &&
3320                     lnet_is_discovery_disabled(lp))) {
3321                 rc = lnet_peer_merge_data(lp, pbuf);
3322         } else {
3323                 lpni = lnet_peer_ni_find_locked(&nid);
3324                 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3325                         rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3326                         if (rc) {
3327                                 CERROR("Primary NID error %s versus %s: %d\n",
3328                                        libcfs_nidstr(&lp->lp_primary_nid),
3329                                        libcfs_nidstr(&nid), rc);
3330                                 lnet_ping_buffer_decref(pbuf);
3331                         } else {
3332                                 rc = lnet_peer_merge_data(lp, pbuf);
3333                         }
3334                         if (lpni)
3335                                 lnet_peer_ni_decref_locked(lpni);
3336                 } else {
3337                         struct lnet_peer *new_lp;
3338                         new_lp = lpni->lpni_peer_net->lpn_peer;
3339                         /*
3340                          * if lp has discovery/MR enabled that means new_lp
3341                          * should have discovery/MR enabled as well, since
3342                          * it's the same peer, which we're about to merge
3343                          */
3344                         spin_lock(&lp->lp_lock);
3345                         spin_lock(&new_lp->lp_lock);
3346                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3347                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3348                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3349                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3350                         /* If we're processing a ping reply then we may be
3351                          * about to send a push to the peer that we ping'd.
3352                          * Since the ping reply that we're processing was
3353                          * received by lp, we need to set the discovery source
3354                          * NID for new_lp to the NID stored in lp.
3355                          */
3356                         if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3357                                 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3358                                 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3359                         }
3360                         spin_unlock(&new_lp->lp_lock);
3361                         spin_unlock(&lp->lp_lock);
3362
3363                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
3364                         lnet_consolidate_routes_locked(lp, new_lp);
3365                         lnet_peer_ni_decref_locked(lpni);
3366                 }
3367         }
3368 out:
3369         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3370                libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3371                lp->lp_state);
3372         mutex_unlock(&the_lnet.ln_api_mutex);
3373
3374         spin_lock(&lp->lp_lock);
3375         /* Tell discovery to re-check the peer immediately. */
3376         if (!rc)
3377                 rc = LNET_REDISCOVER_PEER;
3378         return rc;
3379 }
3380
3381 /*
3382  * A ping failed. Clear the PING_FAILED state and set the
3383  * FORCE_PING state, to ensure a retry even if discovery is
3384  * disabled. This avoids being left with incorrect state.
3385  */
3386 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3387 __must_hold(&lp->lp_lock)
3388 {
3389         struct lnet_handle_md mdh;
3390         int rc;
3391
3392         mdh = lp->lp_ping_mdh;
3393         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3394         lp->lp_state &= ~LNET_PEER_PING_FAILED;
3395         lp->lp_state |= LNET_PEER_FORCE_PING;
3396         rc = lp->lp_ping_error;
3397         lp->lp_ping_error = 0;
3398         spin_unlock(&lp->lp_lock);
3399
3400         if (!LNetMDHandleIsInvalid(mdh))
3401                 LNetMDUnlink(mdh);
3402
3403         CDEBUG(D_NET, "peer %s:%d\n",
3404                libcfs_nidstr(&lp->lp_primary_nid), rc);
3405
3406         spin_lock(&lp->lp_lock);
3407         return rc ? rc : LNET_REDISCOVER_PEER;
3408 }
3409
3410 /* Active side of ping. */
3411 static int lnet_peer_send_ping(struct lnet_peer *lp)
3412 __must_hold(&lp->lp_lock)
3413 {
3414         int nnis;
3415         int rc;
3416         int cpt;
3417
3418         lp->lp_state |= LNET_PEER_PING_SENT;
3419         lp->lp_state &= ~LNET_PEER_FORCE_PING;
3420         spin_unlock(&lp->lp_lock);
3421
3422         cpt = lnet_net_lock_current();
3423         /* Refcount for MD. */
3424         lnet_peer_addref_locked(lp);
3425         lnet_net_unlock(cpt);
3426
3427         nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3428
3429         rc = lnet_send_ping(&lp->lp_primary_nid, &lp->lp_ping_mdh, nnis, lp,
3430                             the_lnet.ln_dc_handler, false);
3431
3432         /*
3433          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3434          * refcount on the peer, otherwise LNetMDUnlink will be called
3435          * which will eventually do that.
3436          */
3437         if (rc > 0) {
3438                 lnet_net_lock(cpt);
3439                 lnet_peer_decref_locked(lp);
3440                 lnet_net_unlock(cpt);
3441                 rc = -rc; /* change the rc to negative value */
3442                 goto fail_error;
3443         } else if (rc < 0) {
3444                 goto fail_error;
3445         }
3446
3447         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3448
3449         spin_lock(&lp->lp_lock);
3450         return 0;
3451
3452 fail_error:
3453         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3454         /*
3455          * The errors that get us here are considered hard errors and
3456          * cause Discovery to terminate. So we clear PING_SENT, but do
3457          * not set either PING_FAILED or FORCE_PING. In fact we need
3458          * to clear PING_FAILED, because the unlink event handler will
3459          * have set it if we called LNetMDUnlink() above.
3460          */
3461         spin_lock(&lp->lp_lock);
3462         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3463         return rc;
3464 }
3465
3466 /*
3467  * This function exists because you cannot call LNetMDUnlink() from an
3468  * event handler.
3469  */
3470 static int lnet_peer_push_failed(struct lnet_peer *lp)
3471 __must_hold(&lp->lp_lock)
3472 {
3473         struct lnet_handle_md mdh;
3474         int rc;
3475
3476         mdh = lp->lp_push_mdh;
3477         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3478         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3479         rc = lp->lp_push_error;
3480         lp->lp_push_error = 0;
3481         spin_unlock(&lp->lp_lock);
3482
3483         if (!LNetMDHandleIsInvalid(mdh))
3484                 LNetMDUnlink(mdh);
3485
3486         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3487         spin_lock(&lp->lp_lock);
3488         return rc ? rc : LNET_REDISCOVER_PEER;
3489 }
3490
3491 /*
3492  * Mark the peer as discovered.
3493  */
3494 static int lnet_peer_discovered(struct lnet_peer *lp)
3495 __must_hold(&lp->lp_lock)
3496 {
3497         lp->lp_state |= LNET_PEER_DISCOVERED;
3498         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3499                           LNET_PEER_REDISCOVER);
3500
3501         lp->lp_dc_error = 0;
3502
3503         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3504
3505         return 0;
3506 }
3507
3508 /* Active side of push. */
3509 static int lnet_peer_send_push(struct lnet_peer *lp)
3510 __must_hold(&lp->lp_lock)
3511 {
3512         struct lnet_ping_buffer *pbuf;
3513         struct lnet_processid id;
3514         struct lnet_md md;
3515         int cpt;
3516         int rc;
3517
3518         /* Don't push to a non-multi-rail peer. */
3519         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3520                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3521                 /* if peer's NIDs are uptodate then peer is discovered */
3522                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3523                         rc = lnet_peer_discovered(lp);
3524                         return rc;
3525                 }
3526
3527                 return 0;
3528         }
3529
3530         lp->lp_state |= LNET_PEER_PUSH_SENT;
3531         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3532         spin_unlock(&lp->lp_lock);
3533
3534         cpt = lnet_net_lock_current();
3535         pbuf = the_lnet.ln_ping_target;
3536         lnet_ping_buffer_addref(pbuf);
3537         lnet_net_unlock(cpt);
3538
3539         /* Push source MD */
3540         md.start     = &pbuf->pb_info;
3541         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3542         md.threshold = 2; /* Put/Ack */
3543         md.max_size  = 0;
3544         md.options   = LNET_MD_TRACK_RESPONSE;
3545         md.handler   = the_lnet.ln_dc_handler;
3546         md.user_ptr  = lp;
3547
3548         rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3549         if (rc) {
3550                 lnet_ping_buffer_decref(pbuf);
3551                 CERROR("Can't bind push source MD: %d\n", rc);
3552                 goto fail_error;
3553         }
3554
3555         cpt = lnet_net_lock_current();
3556         /* Refcount for MD. */
3557         lnet_peer_addref_locked(lp);
3558         id.pid = LNET_PID_LUSTRE;
3559         if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3560                 id.nid = lp->lp_disc_dst_nid;
3561         else
3562                 id.nid = lp->lp_primary_nid;
3563         lnet_net_unlock(cpt);
3564
3565         rc = LNetPut(&lp->lp_disc_src_nid, lp->lp_push_mdh,
3566                      LNET_ACK_REQ, &id, LNET_RESERVED_PORTAL,
3567                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3568
3569         /*
3570          * reset the discovery nid. There is no need to restrict sending
3571          * from that source, if we call lnet_push_update_to_peers(). It'll
3572          * get set to a specific NID, if we initiate discovery from the
3573          * scratch
3574          */
3575         lp->lp_disc_src_nid = LNET_ANY_NID;
3576         lp->lp_disc_dst_nid = LNET_ANY_NID;
3577
3578         if (rc)
3579                 goto fail_unlink;
3580
3581         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3582
3583         spin_lock(&lp->lp_lock);
3584         return 0;
3585
3586 fail_unlink:
3587         LNetMDUnlink(lp->lp_push_mdh);
3588         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3589 fail_error:
3590         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3591                lp, rc);
3592         /*
3593          * The errors that get us here are considered hard errors and
3594          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3595          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3596          * because the unlink event handler will have set it if we
3597          * called LNetMDUnlink() above.
3598          */
3599         spin_lock(&lp->lp_lock);
3600         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3601         return rc;
3602 }
3603
3604 /*
3605  * Wait for work to be queued or some other change that must be
3606  * attended to. Returns non-zero if the discovery thread should shut
3607  * down.
3608  */
3609 static int lnet_peer_discovery_wait_for_work(void)
3610 {
3611         int cpt;
3612         int rc = 0;
3613
3614         DEFINE_WAIT(wait);
3615
3616         cpt = lnet_net_lock_current();
3617         for (;;) {
3618                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3619                                 TASK_INTERRUPTIBLE);
3620                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3621                         break;
3622                 if (lnet_push_target_resize_needed() ||
3623                     the_lnet.ln_push_target->pb_needs_post)
3624                         break;
3625                 if (!list_empty(&the_lnet.ln_dc_request))
3626                         break;
3627                 if (!list_empty(&the_lnet.ln_msg_resend))
3628                         break;
3629                 lnet_net_unlock(cpt);
3630
3631                 /*
3632                  * wakeup max every second to check if there are peers that
3633                  * have been stuck on the working queue for greater than
3634                  * the peer timeout.
3635                  */
3636                 schedule_timeout(cfs_time_seconds(1));
3637                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3638                 cpt = lnet_net_lock_current();
3639         }
3640         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3641
3642         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3643                 rc = -ESHUTDOWN;
3644
3645         lnet_net_unlock(cpt);
3646
3647         CDEBUG(D_NET, "woken: %d\n", rc);
3648
3649         return rc;
3650 }
3651
3652 /*
3653  * Messages that were pending on a destroyed peer will be put on a global
3654  * resend list. The message resend list will be checked by
3655  * the discovery thread when it wakes up, and will resend messages. These
3656  * messages can still be sendable in the case the lpni which was the initial
3657  * cause of the message re-queue was transfered to another peer.
3658  *
3659  * It is possible that LNet could be shutdown while we're iterating
3660  * through the list. lnet_shudown_lndnets() will attempt to access the
3661  * resend list, but will have to wait until the spinlock is released, by
3662  * which time there shouldn't be any more messages on the resend list.
3663  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3664  * for the messages so they can be released. The other case is that
3665  * lnet_shudown_lndnets() can finalize all the messages before this
3666  * function can visit the resend list, in which case this function will be
3667  * a no-op.
3668  */
3669 static void lnet_resend_msgs(void)
3670 {
3671         struct lnet_msg *msg, *tmp;
3672         LIST_HEAD(resend);
3673         int rc;
3674
3675         spin_lock(&the_lnet.ln_msg_resend_lock);
3676         list_splice(&the_lnet.ln_msg_resend, &resend);
3677         spin_unlock(&the_lnet.ln_msg_resend_lock);
3678
3679         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3680                 list_del_init(&msg->msg_list);
3681                 rc = lnet_send(&msg->msg_src_nid_param, msg,
3682                                &msg->msg_rtr_nid_param);
3683                 if (rc < 0) {
3684                         CNETERR("Error sending %s to %s: %d\n",
3685                                lnet_msgtyp2str(msg->msg_type),
3686                                libcfs_idstr(&msg->msg_target), rc);
3687                         lnet_finalize(msg, rc);
3688                 }
3689         }
3690 }
3691
3692 /* The discovery thread. */
3693 static int lnet_peer_discovery(void *arg)
3694 {
3695         struct lnet_peer *lp;
3696         int rc;
3697
3698         wait_for_completion(&the_lnet.ln_started);
3699
3700         CDEBUG(D_NET, "started\n");
3701
3702         for (;;) {
3703                 if (lnet_peer_discovery_wait_for_work())
3704                         break;
3705
3706                 if (lnet_push_target_resize_needed())
3707                         lnet_push_target_resize();
3708                 else if (the_lnet.ln_push_target->pb_needs_post)
3709                         lnet_push_target_post(the_lnet.ln_push_target,
3710                                               &the_lnet.ln_push_target_md);
3711
3712                 lnet_resend_msgs();
3713
3714                 lnet_net_lock(LNET_LOCK_EX);
3715                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3716                         lnet_net_unlock(LNET_LOCK_EX);
3717                         break;
3718                 }
3719
3720                 /*
3721                  * Process all incoming discovery work requests.  When
3722                  * discovery must wait on a peer to change state, it
3723                  * is added to the tail of the ln_dc_working queue. A
3724                  * timestamp keeps track of when the peer was added,
3725                  * so we can time out discovery requests that take too
3726                  * long.
3727                  */
3728                 while (!list_empty(&the_lnet.ln_dc_request)) {
3729                         lp = list_first_entry(&the_lnet.ln_dc_request,
3730                                               struct lnet_peer, lp_dc_list);
3731                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3732                         /*
3733                          * set the time the peer was put on the dc_working
3734                          * queue. It shouldn't remain on the queue
3735                          * forever, in case the GET message (for ping)
3736                          * doesn't get a REPLY or the PUT message (for
3737                          * push) doesn't get an ACK.
3738                          */
3739                         lp->lp_last_queued = ktime_get_real_seconds();
3740                         lnet_net_unlock(LNET_LOCK_EX);
3741
3742                         if (lnet_push_target_resize_needed())
3743                                 lnet_push_target_resize();
3744                         else if (the_lnet.ln_push_target->pb_needs_post)
3745                                 lnet_push_target_post(the_lnet.ln_push_target,
3746                                                       &the_lnet.ln_push_target_md);
3747
3748                         /*
3749                          * Select an action depending on the state of
3750                          * the peer and whether discovery is disabled.
3751                          * The check whether discovery is disabled is
3752                          * done after the code that handles processing
3753                          * for arrived data, cleanup for failures, and
3754                          * forcing a Ping or Push.
3755                          */
3756                         spin_lock(&lp->lp_lock);
3757                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3758                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3759                                 lp->lp_state);
3760                         if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3761                                             LNET_PEER_MARK_DELETED))
3762                                 rc = lnet_peer_deletion(lp);
3763                         else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3764                                 rc = lnet_peer_data_present(lp);
3765                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3766                                 rc = lnet_peer_ping_failed(lp);
3767                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3768                                 rc = lnet_peer_push_failed(lp);
3769                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3770                                 rc = lnet_peer_send_ping(lp);
3771                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3772                                 rc = lnet_peer_send_push(lp);
3773                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3774                                 rc = lnet_peer_send_ping(lp);
3775                         else if (lnet_peer_needs_push(lp))
3776                                 rc = lnet_peer_send_push(lp);
3777                         else
3778                                 rc = lnet_peer_discovered(lp);
3779                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3780                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3781                                 lp->lp_state, rc);
3782
3783                         if (rc == LNET_REDISCOVER_PEER) {
3784                                 spin_unlock(&lp->lp_lock);
3785                                 lnet_net_lock(LNET_LOCK_EX);
3786                                 list_move(&lp->lp_dc_list,
3787                                           &the_lnet.ln_dc_request);
3788                         } else if (rc ||
3789                                    !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3790                                 spin_unlock(&lp->lp_lock);
3791                                 lnet_net_lock(LNET_LOCK_EX);
3792                                 lnet_peer_discovery_complete(lp, rc);
3793                         } else {
3794                                 spin_unlock(&lp->lp_lock);
3795                                 lnet_net_lock(LNET_LOCK_EX);
3796                         }
3797
3798                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3799                                 break;
3800
3801                 }
3802
3803                 lnet_net_unlock(LNET_LOCK_EX);
3804         }
3805
3806         CDEBUG(D_NET, "stopping\n");
3807         /*
3808          * Clean up before telling lnet_peer_discovery_stop() that
3809          * we're done. Use wake_up() below to somewhat reduce the
3810          * size of the thundering herd if there are multiple threads
3811          * waiting on discovery of a single peer.
3812          */
3813
3814         /* Queue cleanup 1: stop all pending pings and pushes. */
3815         lnet_net_lock(LNET_LOCK_EX);
3816         while (!list_empty(&the_lnet.ln_dc_working)) {
3817                 lp = list_first_entry(&the_lnet.ln_dc_working,
3818                                       struct lnet_peer, lp_dc_list);
3819                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3820                 lnet_net_unlock(LNET_LOCK_EX);
3821                 lnet_peer_cancel_discovery(lp);
3822                 lnet_net_lock(LNET_LOCK_EX);
3823         }
3824         lnet_net_unlock(LNET_LOCK_EX);
3825
3826         /* Queue cleanup 2: wait for the expired queue to clear. */
3827         while (!list_empty(&the_lnet.ln_dc_expired))
3828                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3829
3830         /* Queue cleanup 3: clear the request queue. */
3831         lnet_net_lock(LNET_LOCK_EX);
3832         while (!list_empty(&the_lnet.ln_dc_request)) {
3833                 lp = list_first_entry(&the_lnet.ln_dc_request,
3834                                       struct lnet_peer, lp_dc_list);
3835                 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3836         }
3837         lnet_net_unlock(LNET_LOCK_EX);
3838
3839         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3840         the_lnet.ln_dc_handler = NULL;
3841
3842         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3843         wake_up(&the_lnet.ln_dc_waitq);
3844
3845         CDEBUG(D_NET, "stopped\n");
3846
3847         return 0;
3848 }
3849
3850 /* ln_api_mutex is held on entry. */
3851 int lnet_peer_discovery_start(void)
3852 {
3853         struct task_struct *task;
3854         int rc = 0;
3855
3856         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3857                 return -EALREADY;
3858
3859         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3860         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3861         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3862         if (IS_ERR(task)) {
3863                 rc = PTR_ERR(task);
3864                 CERROR("Can't start peer discovery thread: %d\n", rc);
3865
3866                 the_lnet.ln_dc_handler = NULL;
3867
3868                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3869         }
3870
3871         CDEBUG(D_NET, "discovery start: %d\n", rc);
3872
3873         return rc;
3874 }
3875
3876 /* ln_api_mutex is held on entry. */
3877 void lnet_peer_discovery_stop(void)
3878 {
3879         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3880                 return;
3881
3882         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3883         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3884
3885         /* In the LNetNIInit() path we may be stopping discovery before it
3886          * entered its work loop
3887          */
3888         if (!completion_done(&the_lnet.ln_started))
3889                 complete(&the_lnet.ln_started);
3890         else
3891                 wake_up(&the_lnet.ln_dc_waitq);
3892
3893         mutex_unlock(&the_lnet.ln_api_mutex);
3894         wait_event(the_lnet.ln_dc_waitq,
3895                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3896         mutex_lock(&the_lnet.ln_api_mutex);
3897
3898         LASSERT(list_empty(&the_lnet.ln_dc_request));
3899         LASSERT(list_empty(&the_lnet.ln_dc_working));
3900         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3901
3902         CDEBUG(D_NET, "discovery stopped\n");
3903 }
3904
3905 /* Debugging */
3906
3907 void
3908 lnet_debug_peer(lnet_nid_t nid4)
3909 {
3910         char                    *aliveness = "NA";
3911         struct lnet_peer_ni     *lp;
3912         int                     cpt;
3913         struct lnet_nid nid;
3914
3915         lnet_nid4_to_nid(nid4, &nid);
3916         cpt = lnet_nid2cpt(&nid, NULL);
3917         lnet_net_lock(cpt);
3918
3919         lp = lnet_peerni_by_nid_locked(&nid, NULL, cpt);
3920         if (IS_ERR(lp)) {
3921                 lnet_net_unlock(cpt);
3922                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nidstr(&nid));
3923                 return;
3924         }
3925
3926         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3927                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3928
3929         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3930                libcfs_nidstr(&lp->lpni_nid), kref_read(&lp->lpni_kref),
3931                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3932                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3933                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3934
3935         lnet_peer_ni_decref_locked(lp);
3936
3937         lnet_net_unlock(cpt);
3938 }
3939
3940 /* Gathering information for userspace. */
3941
3942 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3943                           char aliveness[LNET_MAX_STR_LEN],
3944                           __u32 *cpt_iter, __u32 *refcount,
3945                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3946                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3947                           __u32 *peer_tx_qnob)
3948 {
3949         struct lnet_peer_table          *peer_table;
3950         struct lnet_peer_ni             *lp;
3951         int                             j;
3952         int                             lncpt;
3953         bool                            found = false;
3954
3955         /* get the number of CPTs */
3956         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
3957
3958         /* if the cpt number to be examined is >= the number of cpts in
3959          * the system then indicate that there are no more cpts to examin
3960          */
3961         if (*cpt_iter >= lncpt)
3962                 return -ENOENT;
3963
3964         /* get the current table */
3965         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
3966         /* if the ptable is NULL then there are no more cpts to examine */
3967         if (peer_table == NULL)
3968                 return -ENOENT;
3969
3970         lnet_net_lock(*cpt_iter);
3971
3972         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
3973                 struct list_head *peers = &peer_table->pt_hash[j];
3974
3975                 list_for_each_entry(lp, peers, lpni_hashlist) {
3976                         if (!nid_is_nid4(&lp->lpni_nid))
3977                                 continue;
3978                         if (peer_index-- > 0)
3979                                 continue;
3980
3981                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
3982                         if (lnet_isrouter(lp) ||
3983                                 lnet_peer_aliveness_enabled(lp))
3984                                 snprintf(aliveness, LNET_MAX_STR_LEN,
3985                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
3986
3987                         *nid = lnet_nid_to_nid4(&lp->lpni_nid);
3988                         *refcount = kref_read(&lp->lpni_kref);
3989                         *ni_peer_tx_credits =
3990                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
3991                         *peer_tx_credits = lp->lpni_txcredits;
3992                         *peer_rtr_credits = lp->lpni_rtrcredits;
3993                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
3994                         *peer_tx_qnob = lp->lpni_txqnob;
3995
3996                         found = true;
3997                 }
3998
3999         }
4000         lnet_net_unlock(*cpt_iter);
4001
4002         *cpt_iter = lncpt;
4003
4004         return found ? 0 : -ENOENT;
4005 }
4006
4007 /* ln_api_mutex is held, which keeps the peer list stable */
4008 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
4009 {
4010         struct lnet_ioctl_element_stats *lpni_stats;
4011         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
4012         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
4013         struct lnet_peer_ni_credit_info *lpni_info;
4014         struct lnet_peer_ni *lpni;
4015         struct lnet_peer *lp;
4016         lnet_nid_t nid4;
4017         struct lnet_nid nid;
4018         __u32 size;
4019         int rc;
4020
4021         lnet_nid4_to_nid(cfg->prcfg_prim_nid, &nid);
4022         lp = lnet_find_peer(&nid);
4023
4024         if (!lp) {
4025                 rc = -ENOENT;
4026                 goto out;
4027         }
4028
4029         size = sizeof(nid4) + sizeof(*lpni_info) + sizeof(*lpni_stats)
4030                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
4031         size *= lp->lp_nnis;
4032         if (size > cfg->prcfg_size) {
4033                 cfg->prcfg_size = size;
4034                 rc = -E2BIG;
4035                 goto out_lp_decref;
4036         }
4037
4038         cfg->prcfg_prim_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4039         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
4040         cfg->prcfg_cfg_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4041         cfg->prcfg_count = lp->lp_nnis;
4042         cfg->prcfg_size = size;
4043         cfg->prcfg_state = lp->lp_state;
4044
4045         /* Allocate helper buffers. */
4046         rc = -ENOMEM;
4047         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
4048         if (!lpni_info)
4049                 goto out_lp_decref;
4050         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
4051         if (!lpni_stats)
4052                 goto out_free_info;
4053         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
4054         if (!lpni_msg_stats)
4055                 goto out_free_stats;
4056         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
4057         if (!lpni_hstats)
4058                 goto out_free_msg_stats;
4059
4060
4061         lpni = NULL;
4062         rc = -EFAULT;
4063         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
4064                 if (!nid_is_nid4(&lpni->lpni_nid))
4065                         continue;
4066                 nid4 = lnet_nid_to_nid4(&lpni->lpni_nid);
4067                 if (copy_to_user(bulk, &nid4, sizeof(nid4)))
4068                         goto out_free_hstats;
4069                 bulk += sizeof(nid4);
4070
4071                 memset(lpni_info, 0, sizeof(*lpni_info));
4072                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
4073                 if (lnet_isrouter(lpni) ||
4074                         lnet_peer_aliveness_enabled(lpni))
4075                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
4076                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
4077
4078                 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
4079                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
4080                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
4081                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
4082                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
4083                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
4084                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
4085                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4086                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4087                         goto out_free_hstats;
4088                 bulk += sizeof(*lpni_info);
4089
4090                 memset(lpni_stats, 0, sizeof(*lpni_stats));
4091                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4092                                                             LNET_STATS_TYPE_SEND);
4093                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4094                                                             LNET_STATS_TYPE_RECV);
4095                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4096                                                             LNET_STATS_TYPE_DROP);
4097                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4098                         goto out_free_hstats;
4099                 bulk += sizeof(*lpni_stats);
4100                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4101                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4102                         goto out_free_hstats;
4103                 bulk += sizeof(*lpni_msg_stats);
4104                 lpni_hstats->hlpni_network_timeout =
4105                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4106                 lpni_hstats->hlpni_remote_dropped =
4107                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4108                 lpni_hstats->hlpni_remote_timeout =
4109                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4110                 lpni_hstats->hlpni_remote_error =
4111                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4112                 lpni_hstats->hlpni_health_value =
4113                   atomic_read(&lpni->lpni_healthv);
4114                 lpni_hstats->hlpni_ping_count = lpni->lpni_ping_count;
4115                 lpni_hstats->hlpni_next_ping = lpni->lpni_next_ping;
4116                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4117                         goto out_free_hstats;
4118                 bulk += sizeof(*lpni_hstats);
4119         }
4120         rc = 0;
4121
4122 out_free_hstats:
4123         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4124 out_free_msg_stats:
4125         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4126 out_free_stats:
4127         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4128 out_free_info:
4129         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4130 out_lp_decref:
4131         lnet_peer_decref_locked(lp);
4132 out:
4133         return rc;
4134 }
4135
4136 /* must hold net_lock/0 */
4137 void
4138 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4139                                      struct list_head *recovery_queue,
4140                                      time64_t now)
4141 {
4142         /* the mt could've shutdown and cleaned up the queues */
4143         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4144                 return;
4145
4146         if (!list_empty(&lpni->lpni_recovery))
4147                 return;
4148
4149         if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4150                 return;
4151
4152         if (!lpni->lpni_last_alive) {
4153                 CDEBUG(D_NET,
4154                        "lpni %s(%p) not eligible for recovery last alive %lld\n",
4155                        libcfs_nidstr(&lpni->lpni_nid), lpni,
4156                        lpni->lpni_last_alive);
4157                 return;
4158         }
4159
4160         if (lnet_recovery_limit &&
4161             now > lpni->lpni_last_alive + lnet_recovery_limit) {
4162                 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4163                        libcfs_nidstr(&lpni->lpni_nid),
4164                        lpni->lpni_last_alive);
4165                 /* Reset the ping count so that if this peer NI is added back to
4166                  * the recovery queue we will send the first ping right away.
4167                  */
4168                 lpni->lpni_ping_count = 0;
4169                 return;
4170         }
4171
4172         /* This peer NI is going on the recovery queue, so take a ref on it */
4173         lnet_peer_ni_addref_locked(lpni);
4174
4175         lnet_peer_ni_set_next_ping(lpni, now);
4176
4177         CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4178                libcfs_nidstr(&lpni->lpni_nid),
4179                lpni->lpni_ping_count,
4180                lpni->lpni_next_ping,
4181                lpni->lpni_last_alive,
4182                atomic_read(&lpni->lpni_healthv));
4183
4184         list_add_tail(&lpni->lpni_recovery, recovery_queue);
4185 }
4186
4187 /* Call with the ln_api_mutex held */
4188 void
4189 lnet_peer_ni_set_healthv(lnet_nid_t nid4, int value, bool all)
4190 {
4191         struct lnet_peer_table *ptable;
4192         struct lnet_peer *lp;
4193         struct lnet_peer_net *lpn;
4194         struct lnet_peer_ni *lpni;
4195         struct lnet_nid nid;
4196         int lncpt;
4197         int cpt;
4198         time64_t now;
4199
4200         if (the_lnet.ln_state != LNET_STATE_RUNNING)
4201                 return;
4202
4203         lnet_nid4_to_nid(nid4, &nid);
4204         now = ktime_get_seconds();
4205
4206         if (!all) {
4207                 lnet_net_lock(LNET_LOCK_EX);
4208                 lpni = lnet_peer_ni_find_locked(&nid);
4209                 if (!lpni) {
4210                         lnet_net_unlock(LNET_LOCK_EX);
4211                         return;
4212                 }
4213                 lnet_set_lpni_healthv_locked(lpni, value);
4214                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4215                                              &the_lnet.ln_mt_peerNIRecovq, now);
4216                 lnet_peer_ni_decref_locked(lpni);
4217                 lnet_net_unlock(LNET_LOCK_EX);
4218                 return;
4219         }
4220
4221         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4222
4223         /*
4224          * Walk all the peers and reset the health value for each one to the
4225          * specified value.
4226          */
4227         lnet_net_lock(LNET_LOCK_EX);
4228         for (cpt = 0; cpt < lncpt; cpt++) {
4229                 ptable = the_lnet.ln_peer_tables[cpt];
4230                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4231                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4232                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4233                                                     lpni_peer_nis) {
4234                                         lnet_set_lpni_healthv_locked(lpni,
4235                                                                      value);
4236                                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
4237                                              &the_lnet.ln_mt_peerNIRecovq, now);
4238                                 }
4239                         }
4240                 }
4241         }
4242         lnet_net_unlock(LNET_LOCK_EX);
4243 }
4244