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, lnet_nid_t nid4, unsigned int flags)
 533 {
 534         struct lnet_peer_ni *lpni;
 535         struct lnet_nid primary_nid = lp->lp_primary_nid;
 536         struct lnet_nid nid;
 537         int rc = 0;
 538         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 539
 540         lnet_nid4_to_nid(nid4, &nid);
 541         if (!(flags & LNET_PEER_CONFIGURED)) {
 542                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 543                         rc = -EPERM;
 544                         goto out;
 545                 }
 546         }
 547
 548         lpni = lnet_peer_ni_find_locked(&nid);
 549         if (!lpni) {
 550                 rc = -ENOENT;
 551                 goto out;
 552         }
 553         lnet_peer_ni_decref_locked(lpni);
 554         if (lp != lpni->lpni_peer_net->lpn_peer) {
 555                 rc = -ECHILD;
 556                 goto out;
 557         }
 558
 559         /*
 560          * This function only allows deletion of the primary NID if it
 561          * is the only NID.
 562          */
 563         if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && !force) {
 564                 rc = -EBUSY;
 565                 goto out;
 566         }
 567
 568         lnet_net_lock(LNET_LOCK_EX);
 569
 570         if (nid_same(&nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && force) {
 571                 struct lnet_peer_ni *lpni2;
 572                 /* assign the next peer_ni to be the primary */
 573                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 574                 LASSERT(lpni2);
 575                 lp->lp_primary_nid = lpni2->lpni_nid;
 576         }
 577         rc = lnet_peer_ni_del_locked(lpni, force);
 578
 579         lnet_net_unlock(LNET_LOCK_EX);
 580
 581 out:
 582         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 583                libcfs_nidstr(&primary_nid), libcfs_nidstr(&nid),
 584                flags, rc);
 585
 586         return rc;
 587 }
 588
 589 static void
 590 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 591                                struct lnet_peer_table *ptable)
 592 {
 593         int                      i;
 594         struct lnet_peer_ni     *next;
 595         struct lnet_peer_ni     *lpni;
 596         struct lnet_peer        *peer;
 597
 598         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 599                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 600                                          lpni_hashlist) {
 601                         if (net != NULL && net != lpni->lpni_net)
 602                                 continue;
 603
 604                         peer = lpni->lpni_peer_net->lpn_peer;
 605                         if (!nid_same(&peer->lp_primary_nid,
 606                                        &lpni->lpni_nid)) {
 607                                 lnet_peer_ni_del_locked(lpni, false);
 608                                 continue;
 609                         }
 610                         /*
 611                          * Removing the primary NID implies removing
 612                          * the entire peer. Advance next beyond any
 613                          * peer_ni that belongs to the same peer.
 614                          */
 615                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 616                                                  lpni_hashlist) {
 617                                 if (next->lpni_peer_net->lpn_peer != peer)
 618                                         break;
 619                         }
 620                         lnet_peer_del_locked(peer);
 621                 }
 622         }
 623 }
 624
 625 static void
 626 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 627 {
 628         wait_var_event_warning(&ptable->pt_zombies,
 629                                ptable->pt_zombies == 0,
 630                                "Waiting for %d zombies on peer table\n",
 631                                ptable->pt_zombies);
 632 }
 633
 634 static void
 635 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 636                                 struct lnet_peer_table *ptable)
 637 {
 638         struct lnet_peer_ni     *lp;
 639         struct lnet_peer_ni     *tmp;
 640         lnet_nid_t              gw_nid;
 641         int                     i;
 642
 643         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 644                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 645                                          lpni_hashlist) {
 646                         if (net != lp->lpni_net)
 647                                 continue;
 648
 649                         if (!lnet_isrouter(lp))
 650                                 continue;
 651
 652                         /* FIXME handle large-addr nid */
 653                         gw_nid = lnet_nid_to_nid4(
 654                                 &lp->lpni_peer_net->lpn_peer->lp_primary_nid);
 655
 656                         lnet_net_unlock(LNET_LOCK_EX);
 657                         lnet_del_route(LNET_NET_ANY, gw_nid);
 658                         lnet_net_lock(LNET_LOCK_EX);
 659                 }
 660         }
 661 }
 662
 663 void
 664 lnet_peer_tables_cleanup(struct lnet_net *net)
 665 {
 666         int i;
 667         struct lnet_peer_table *ptable;
 668
 669         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 670         /* If just deleting the peers for a NI, get rid of any routes these
 671          * peers are gateways for. */
 672         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 673                 lnet_net_lock(LNET_LOCK_EX);
 674                 lnet_peer_table_del_rtrs_locked(net, ptable);
 675                 lnet_net_unlock(LNET_LOCK_EX);
 676         }
 677
 678         /* Start the cleanup process */
 679         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 680                 lnet_net_lock(LNET_LOCK_EX);
 681                 lnet_peer_table_cleanup_locked(net, ptable);
 682                 lnet_net_unlock(LNET_LOCK_EX);
 683         }
 684
 685         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 686                 lnet_peer_ni_finalize_wait(ptable);
 687 }
 688
 689 static struct lnet_peer_ni *
 690 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, struct lnet_nid *nid)
 691 {
 692         struct list_head        *peers;
 693         struct lnet_peer_ni     *lp;
 694
 695         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 696                 return NULL;
 697
 698         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 699         list_for_each_entry(lp, peers, lpni_hashlist) {
 700                 if (nid_same(&lp->lpni_nid, nid)) {
 701                         lnet_peer_ni_addref_locked(lp);
 702                         return lp;
 703                 }
 704         }
 705
 706         return NULL;
 707 }
 708
 709 struct lnet_peer_ni *
 710 lnet_find_peer_ni_locked(lnet_nid_t nid4)
 711 {
 712         struct lnet_peer_ni *lpni;
 713         struct lnet_peer_table *ptable;
 714         int cpt;
 715         struct lnet_nid nid;
 716
 717         lnet_nid4_to_nid(nid4, &nid);
 718
 719         cpt = lnet_nid_cpt_hash(&nid, LNET_CPT_NUMBER);
 720
 721         ptable = the_lnet.ln_peer_tables[cpt];
 722         lpni = lnet_get_peer_ni_locked(ptable, &nid);
 723
 724         return lpni;
 725 }
 726
 727 struct lnet_peer_ni *
 728 lnet_peer_ni_find_locked(struct lnet_nid *nid)
 729 {
 730         struct lnet_peer_ni *lpni;
 731         struct lnet_peer_table *ptable;
 732         int cpt;
 733
 734         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 735
 736         ptable = the_lnet.ln_peer_tables[cpt];
 737         lpni = lnet_get_peer_ni_locked(ptable, nid);
 738
 739         return lpni;
 740 }
 741
 742 struct lnet_peer_ni *
 743 lnet_peer_get_ni_locked(struct lnet_peer *lp, lnet_nid_t nid)
 744 {
 745         struct lnet_peer_net *lpn;
 746         struct lnet_peer_ni *lpni;
 747
 748         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid));
 749         if (!lpn)
 750                 return NULL;
 751
 752         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 753                 if (lnet_nid_to_nid4(&lpni->lpni_nid) == nid)
 754                         return lpni;
 755         }
 756
 757         return NULL;
 758 }
 759
 760 struct lnet_peer_ni *
 761 lnet_peer_ni_get_locked(struct lnet_peer *lp, struct lnet_nid *nid)
 762 {
 763         struct lnet_peer_net *lpn;
 764         struct lnet_peer_ni *lpni;
 765
 766         lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
 767         if (!lpn)
 768                 return NULL;
 769
 770         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 771                 if (nid_same(&lpni->lpni_nid, nid))
 772                         return lpni;
 773         }
 774
 775         return NULL;
 776 }
 777
 778 struct lnet_peer *
 779 lnet_find_peer(lnet_nid_t nid)
 780 {
 781         struct lnet_peer_ni *lpni;
 782         struct lnet_peer *lp = NULL;
 783         int cpt;
 784
 785         cpt = lnet_net_lock_current();
 786         lpni = lnet_find_peer_ni_locked(nid);
 787         if (lpni) {
 788                 lp = lpni->lpni_peer_net->lpn_peer;
 789                 lnet_peer_addref_locked(lp);
 790                 lnet_peer_ni_decref_locked(lpni);
 791         }
 792         lnet_net_unlock(cpt);
 793
 794         return lp;
 795 }
 796
 797 struct lnet_peer_net *
 798 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 799 {
 800         struct lnet_peer_net *net;
 801
 802         if (!prev_lpn_id) {
 803                 /* no net id provided return the first net */
 804                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 805                                                struct lnet_peer_net,
 806                                                lpn_peer_nets);
 807
 808                 return net;
 809         }
 810
 811         /* find the net after the one provided */
 812         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 813                 if (net->lpn_net_id == prev_lpn_id) {
 814                         /*
 815                          * if we reached the end of the list loop to the
 816                          * beginning.
 817                          */
 818                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 819                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 820                                                                 struct lnet_peer_net,
 821                                                                 lpn_peer_nets);
 822                         else
 823                                 return list_next_entry(net, lpn_peer_nets);
 824                 }
 825         }
 826
 827         return NULL;
 828 }
 829
 830 struct lnet_peer_ni *
 831 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 832                              struct lnet_peer_net *peer_net,
 833                              struct lnet_peer_ni *prev)
 834 {
 835         struct lnet_peer_ni *lpni;
 836         struct lnet_peer_net *net = peer_net;
 837
 838         if (!prev) {
 839                 if (!net) {
 840                         if (list_empty(&peer->lp_peer_nets))
 841                                 return NULL;
 842
 843                         net = list_entry(peer->lp_peer_nets.next,
 844                                          struct lnet_peer_net,
 845                                          lpn_peer_nets);
 846                 }
 847                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 848                                   lpni_peer_nis);
 849
 850                 return lpni;
 851         }
 852
 853         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 854                 /*
 855                  * if you reached the end of the peer ni list and the peer
 856                  * net is specified then there are no more peer nis in that
 857                  * net.
 858                  */
 859                 if (net)
 860                         return NULL;
 861
 862                 /*
 863                  * we reached the end of this net ni list. move to the
 864                  * next net
 865                  */
 866                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 867                     &peer->lp_peer_nets)
 868                         /* no more nets and no more NIs. */
 869                         return NULL;
 870
 871                 /* get the next net */
 872                 net = list_entry(prev->lpni_peer_net->lpn_peer_nets.next,
 873                                  struct lnet_peer_net,
 874                                  lpn_peer_nets);
 875                 /* get the ni on it */
 876                 lpni = list_entry(net->lpn_peer_nis.next, struct lnet_peer_ni,
 877                                   lpni_peer_nis);
 878
 879                 return lpni;
 880         }
 881
 882         /* there are more nis left */
 883         lpni = list_entry(prev->lpni_peer_nis.next,
 884                           struct lnet_peer_ni, lpni_peer_nis);
 885
 886         return lpni;
 887 }
 888
 889 /* Call with the ln_api_mutex held */
 890 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 891 {
 892         struct lnet_process_id id;
 893         struct lnet_peer_table *ptable;
 894         struct lnet_peer *lp;
 895         __u32 count = 0;
 896         __u32 size = 0;
 897         int lncpt;
 898         int cpt;
 899         __u32 i;
 900         int rc;
 901
 902         rc = -ESHUTDOWN;
 903         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 904                 goto done;
 905
 906         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 907
 908         /*
 909          * Count the number of peers, and return E2BIG if the buffer
 910          * is too small. We'll also return the desired size.
 911          */
 912         rc = -E2BIG;
 913         for (cpt = 0; cpt < lncpt; cpt++) {
 914                 ptable = the_lnet.ln_peer_tables[cpt];
 915                 count += ptable->pt_peers;
 916         }
 917         size = count * sizeof(*ids);
 918         if (size > *sizep)
 919                 goto done;
 920
 921         /*
 922          * Walk the peer lists and copy out the primary nids.
 923          * This is safe because the peer lists are only modified
 924          * while the ln_api_mutex is held. So we don't need to
 925          * hold the lnet_net_lock as well, and can therefore
 926          * directly call copy_to_user().
 927          */
 928         rc = -EFAULT;
 929         memset(&id, 0, sizeof(id));
 930         id.pid = LNET_PID_LUSTRE;
 931         i = 0;
 932         for (cpt = 0; cpt < lncpt; cpt++) {
 933                 ptable = the_lnet.ln_peer_tables[cpt];
 934                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 935                         if (!nid_is_nid4(&lp->lp_primary_nid))
 936                                 continue;
 937                         if (i >= count)
 938                                 goto done;
 939                         id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
 940                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 941                                 goto done;
 942                         i++;
 943                 }
 944         }
 945         rc = 0;
 946 done:
 947         *countp = count;
 948         *sizep = size;
 949         return rc;
 950 }
 951
 952 /*
 953  * Start pushes to peers that need to be updated for a configuration
 954  * change on this node.
 955  */
 956 void
 957 lnet_push_update_to_peers(int force)
 958 {
 959         struct lnet_peer_table *ptable;
 960         struct lnet_peer *lp;
 961         int lncpt;
 962         int cpt;
 963
 964         lnet_net_lock(LNET_LOCK_EX);
 965         if (lnet_peer_discovery_disabled)
 966                 force = 0;
 967         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 968         for (cpt = 0; cpt < lncpt; cpt++) {
 969                 ptable = the_lnet.ln_peer_tables[cpt];
 970                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 971                         if (force) {
 972                                 spin_lock(&lp->lp_lock);
 973                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 974                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 975                                 spin_unlock(&lp->lp_lock);
 976                         }
 977                         if (lnet_peer_needs_push(lp))
 978                                 lnet_peer_queue_for_discovery(lp);
 979                 }
 980         }
 981         lnet_net_unlock(LNET_LOCK_EX);
 982         wake_up(&the_lnet.ln_dc_waitq);
 983 }
 984
 985 /* find the NID in the preferred gateways for the remote peer
 986  * return:
 987  *      false: list is not empty and NID is not preferred
 988  *      false: list is empty
 989  *      true: nid is found in the list
 990  */
 991 bool
 992 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
 993                              struct lnet_nid *gw_nid)
 994 {
 995         struct lnet_nid_list *ne;
 996
 997         CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
 998                libcfs_nidstr(&lpni->lpni_nid),
 999                list_empty(&lpni->lpni_rtr_pref_nids));
1000
1001         if (list_empty(&lpni->lpni_rtr_pref_nids))
1002                 return false;
1003
1004         /* iterate through all the preferred NIDs and see if any of them
1005          * matches the provided gw_nid
1006          */
1007         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1008                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
1009                        libcfs_nidstr(&ne->nl_nid),
1010                        libcfs_nidstr(gw_nid));
1011                 if (nid_same(&ne->nl_nid, gw_nid))
1012                         return true;
1013         }
1014
1015         return false;
1016 }
1017
1018 void
1019 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
1020 {
1021         struct list_head zombies;
1022         struct lnet_nid_list *ne;
1023         struct lnet_nid_list *tmp;
1024         int cpt = lpni->lpni_cpt;
1025
1026         INIT_LIST_HEAD(&zombies);
1027
1028         lnet_net_lock(cpt);
1029         list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
1030         lnet_net_unlock(cpt);
1031
1032         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1033                 list_del(&ne->nl_list);
1034                 LIBCFS_FREE(ne, sizeof(*ne));
1035         }
1036 }
1037
1038 int
1039 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
1040                        struct lnet_nid *gw_nid)
1041 {
1042         int cpt = lpni->lpni_cpt;
1043         struct lnet_nid_list *ne = NULL;
1044
1045         /* This function is called with api_mutex held. When the api_mutex
1046          * is held the list can not be modified, as it is only modified as
1047          * a result of applying a UDSP and that happens under api_mutex
1048          * lock.
1049          */
1050         __must_hold(&the_lnet.ln_api_mutex);
1051
1052         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1053                 if (nid_same(&ne->nl_nid, gw_nid))
1054                         return -EEXIST;
1055         }
1056
1057         LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1058         if (!ne)
1059                 return -ENOMEM;
1060
1061         ne->nl_nid = *gw_nid;
1062
1063         /* Lock the cpt to protect against addition and checks in the
1064          * selection algorithm
1065          */
1066         lnet_net_lock(cpt);
1067         list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1068         lnet_net_unlock(cpt);
1069
1070         return 0;
1071 }
1072
1073 /*
1074  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1075  * this is a preferred point-to-point path. Call with lnet_net_lock in
1076  * shared mmode.
1077  */
1078 bool
1079 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1080 {
1081         struct lnet_nid_list *ne;
1082
1083         if (lpni->lpni_pref_nnids == 0)
1084                 return false;
1085         if (lpni->lpni_pref_nnids == 1)
1086                 return nid_same(&lpni->lpni_pref.nid, nid);
1087         list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1088                 if (nid_same(&ne->nl_nid, nid))
1089                         return true;
1090         }
1091         return false;
1092 }
1093
1094 /*
1095  * Set a single ni as preferred, provided no preferred ni is already
1096  * defined. Only to be used for non-multi-rail peer_ni.
1097  */
1098 int
1099 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni,
1100                                   struct lnet_nid *nid)
1101 {
1102         int rc = 0;
1103
1104         if (!nid)
1105                 return -EINVAL;
1106         spin_lock(&lpni->lpni_lock);
1107         if (LNET_NID_IS_ANY(nid)) {
1108                 rc = -EINVAL;
1109         } else if (lpni->lpni_pref_nnids > 0) {
1110                 rc = -EPERM;
1111         } else if (lpni->lpni_pref_nnids == 0) {
1112                 lpni->lpni_pref.nid = *nid;
1113                 lpni->lpni_pref_nnids = 1;
1114                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1115         }
1116         spin_unlock(&lpni->lpni_lock);
1117
1118         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1119                libcfs_nidstr(&lpni->lpni_nid), libcfs_nidstr(nid), rc);
1120         return rc;
1121 }
1122
1123 /*
1124  * Clear the preferred NID from a non-multi-rail peer_ni, provided
1125  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1126  */
1127 int
1128 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1129 {
1130         int rc = 0;
1131
1132         spin_lock(&lpni->lpni_lock);
1133         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1134                 lpni->lpni_pref_nnids = 0;
1135                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1136         } else if (lpni->lpni_pref_nnids == 0) {
1137                 rc = -ENOENT;
1138         } else {
1139                 rc = -EPERM;
1140         }
1141         spin_unlock(&lpni->lpni_lock);
1142
1143         CDEBUG(D_NET, "peer %s: %d\n",
1144                libcfs_nidstr(&lpni->lpni_nid), rc);
1145         return rc;
1146 }
1147
1148 void
1149 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1150 {
1151         lpni->lpni_sel_priority = priority;
1152 }
1153
1154 /*
1155  * Clear the preferred NIDs from a non-multi-rail peer.
1156  */
1157 void
1158 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1159 {
1160         struct lnet_peer_ni *lpni = NULL;
1161
1162         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1163                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1164 }
1165
1166 int
1167 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1168 {
1169         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1170         struct lnet_nid_list *ne1 = NULL;
1171         struct lnet_nid_list *ne2 = NULL;
1172         struct lnet_nid *tmp_nid = NULL;
1173         int rc = 0;
1174
1175         if (LNET_NID_IS_ANY(nid)) {
1176                 rc = -EINVAL;
1177                 goto out;
1178         }
1179
1180         if (lpni->lpni_pref_nnids == 1 &&
1181             nid_same(&lpni->lpni_pref.nid, nid)) {
1182                 rc = -EEXIST;
1183                 goto out;
1184         }
1185
1186         /* A non-MR node may have only one preferred NI per peer_ni */
1187         if (lpni->lpni_pref_nnids > 0 &&
1188             !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1189                 rc = -EPERM;
1190                 goto out;
1191         }
1192
1193         /* add the new preferred nid to the list of preferred nids */
1194         if (lpni->lpni_pref_nnids != 0) {
1195                 size_t alloc_size = sizeof(*ne1);
1196
1197                 if (lpni->lpni_pref_nnids == 1) {
1198                         tmp_nid = &lpni->lpni_pref.nid;
1199                         INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1200                 }
1201
1202                 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1203                         if (nid_same(&ne1->nl_nid, nid)) {
1204                                 rc = -EEXIST;
1205                                 goto out;
1206                         }
1207                 }
1208
1209                 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1210                                  alloc_size);
1211                 if (!ne1) {
1212                         rc = -ENOMEM;
1213                         goto out;
1214                 }
1215
1216                 /* move the originally stored nid to the list */
1217                 if (lpni->lpni_pref_nnids == 1) {
1218                         LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1219                                 lpni->lpni_cpt, alloc_size);
1220                         if (!ne2) {
1221                                 rc = -ENOMEM;
1222                                 goto out;
1223                         }
1224                         INIT_LIST_HEAD(&ne2->nl_list);
1225                         ne2->nl_nid = *tmp_nid;
1226                 }
1227                 ne1->nl_nid = *nid;
1228         }
1229
1230         lnet_net_lock(LNET_LOCK_EX);
1231         spin_lock(&lpni->lpni_lock);
1232         if (lpni->lpni_pref_nnids == 0) {
1233                 lpni->lpni_pref.nid = *nid;
1234         } else {
1235                 if (ne2)
1236                         list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1237                 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1238         }
1239         lpni->lpni_pref_nnids++;
1240         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1241         spin_unlock(&lpni->lpni_lock);
1242         lnet_net_unlock(LNET_LOCK_EX);
1243
1244 out:
1245         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1246                 spin_lock(&lpni->lpni_lock);
1247                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1248                 spin_unlock(&lpni->lpni_lock);
1249         }
1250         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1251                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1252         return rc;
1253 }
1254
1255 int
1256 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1257 {
1258         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1259         struct lnet_nid_list *ne = NULL;
1260         int rc = 0;
1261
1262         if (lpni->lpni_pref_nnids == 0) {
1263                 rc = -ENOENT;
1264                 goto out;
1265         }
1266
1267         if (lpni->lpni_pref_nnids == 1) {
1268                 if (!nid_same(&lpni->lpni_pref.nid, nid)) {
1269                         rc = -ENOENT;
1270                         goto out;
1271                 }
1272         } else {
1273                 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1274                         if (nid_same(&ne->nl_nid, nid))
1275                                 goto remove_nid_entry;
1276                 }
1277                 rc = -ENOENT;
1278                 ne = NULL;
1279                 goto out;
1280         }
1281
1282 remove_nid_entry:
1283         lnet_net_lock(LNET_LOCK_EX);
1284         spin_lock(&lpni->lpni_lock);
1285         if (lpni->lpni_pref_nnids == 1)
1286                 lpni->lpni_pref.nid = LNET_ANY_NID;
1287         else {
1288                 list_del_init(&ne->nl_list);
1289                 if (lpni->lpni_pref_nnids == 2) {
1290                         struct lnet_nid_list *ne, *tmp;
1291
1292                         list_for_each_entry_safe(ne, tmp,
1293                                                  &lpni->lpni_pref.nids,
1294                                                  nl_list) {
1295                                 lpni->lpni_pref.nid = ne->nl_nid;
1296                                 list_del_init(&ne->nl_list);
1297                                 LIBCFS_FREE(ne, sizeof(*ne));
1298                         }
1299                 }
1300         }
1301         lpni->lpni_pref_nnids--;
1302         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1303         spin_unlock(&lpni->lpni_lock);
1304         lnet_net_unlock(LNET_LOCK_EX);
1305
1306         if (ne)
1307                 LIBCFS_FREE(ne, sizeof(*ne));
1308 out:
1309         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1310                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1311         return rc;
1312 }
1313
1314 void
1315 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1316 {
1317         struct list_head zombies;
1318         struct lnet_nid_list *ne;
1319         struct lnet_nid_list *tmp;
1320
1321         INIT_LIST_HEAD(&zombies);
1322
1323         lnet_net_lock(LNET_LOCK_EX);
1324         if (lpni->lpni_pref_nnids == 1)
1325                 lpni->lpni_pref.nid = LNET_ANY_NID;
1326         else if (lpni->lpni_pref_nnids > 1)
1327                 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1328         lpni->lpni_pref_nnids = 0;
1329         lnet_net_unlock(LNET_LOCK_EX);
1330
1331         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1332                 list_del_init(&ne->nl_list);
1333                 LIBCFS_FREE(ne, sizeof(*ne));
1334         }
1335 }
1336
1337 lnet_nid_t
1338 lnet_peer_primary_nid_locked(lnet_nid_t nid)
1339 {
1340         /* FIXME handle large-addr nid */
1341         struct lnet_peer_ni *lpni;
1342         lnet_nid_t primary_nid = nid;
1343
1344         lpni = lnet_find_peer_ni_locked(nid);
1345         if (lpni) {
1346                 primary_nid = lnet_nid_to_nid4(
1347                         &lpni->lpni_peer_net->lpn_peer->lp_primary_nid);
1348                 lnet_peer_ni_decref_locked(lpni);
1349         }
1350
1351         return primary_nid;
1352 }
1353
1354 bool
1355 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1356 __must_hold(&lp->lp_lock)
1357 {
1358         if (lnet_peer_discovery_disabled)
1359                 return true;
1360
1361         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1362             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1363                 return true;
1364         }
1365
1366         return false;
1367 }
1368
1369 /*
1370  * Peer Discovery
1371  */
1372 bool
1373 lnet_is_discovery_disabled(struct lnet_peer *lp)
1374 {
1375         bool rc = false;
1376
1377         spin_lock(&lp->lp_lock);
1378         rc = lnet_is_discovery_disabled_locked(lp);
1379         spin_unlock(&lp->lp_lock);
1380
1381         return rc;
1382 }
1383
1384 int
1385 LNetAddPeer(lnet_nid_t *nids, __u32 num_nids)
1386 {
1387         lnet_nid_t pnid = 0;
1388         bool mr;
1389         int i, rc;
1390
1391         if (!nids || num_nids < 1)
1392                 return -EINVAL;
1393
1394         rc = LNetNIInit(LNET_PID_ANY);
1395         if (rc < 0)
1396                 return rc;
1397
1398         mutex_lock(&the_lnet.ln_api_mutex);
1399
1400         mr = lnet_peer_discovery_disabled == 0;
1401
1402         rc = 0;
1403         for (i = 0; i < num_nids; i++) {
1404                 if (nids[i] == LNET_NID_LO_0)
1405                         continue;
1406
1407                 if (!pnid) {
1408                         pnid = nids[i];
1409                         rc = lnet_add_peer_ni(pnid, LNET_NID_ANY, mr, true);
1410                 } else if (lnet_peer_discovery_disabled) {
1411                         rc = lnet_add_peer_ni(nids[i], LNET_NID_ANY, mr, true);
1412                 } else {
1413                         rc = lnet_add_peer_ni(pnid, nids[i], mr, true);
1414                 }
1415
1416                 if (rc && rc != -EEXIST)
1417                         goto unlock;
1418         }
1419
1420 unlock:
1421         mutex_unlock(&the_lnet.ln_api_mutex);
1422
1423         LNetNIFini();
1424
1425         return rc == -EEXIST ? 0 : rc;
1426 }
1427 EXPORT_SYMBOL(LNetAddPeer);
1428
1429 /* FIXME support large-addr nid */
1430 lnet_nid_t
1431 LNetPrimaryNID(lnet_nid_t nid)
1432 {
1433         struct lnet_peer *lp;
1434         struct lnet_peer_ni *lpni;
1435         lnet_nid_t primary_nid = nid;
1436         int rc = 0;
1437         int cpt;
1438
1439         if (nid == LNET_NID_LO_0)
1440                 return LNET_NID_LO_0;
1441
1442         cpt = lnet_net_lock_current();
1443         lpni = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
1444         if (IS_ERR(lpni)) {
1445                 rc = PTR_ERR(lpni);
1446                 goto out_unlock;
1447         }
1448         lp = lpni->lpni_peer_net->lpn_peer;
1449
1450         /* If discovery is disabled locally then we needn't bother running
1451          * discovery here because discovery will not modify whatever
1452          * primary NID is currently set for this peer. If the specified peer is
1453          * down then this discovery can introduce long delays into the mount
1454          * process, so skip it if it isn't necessary.
1455          */
1456         while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1457                 spin_lock(&lp->lp_lock);
1458                 /* force a full discovery cycle */
1459                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1460                 spin_unlock(&lp->lp_lock);
1461
1462                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1463                 if (rc)
1464                         goto out_decref;
1465                 /* The lpni (or lp) for this NID may have changed and our ref is
1466                  * the only thing keeping the old one around. Release the ref
1467                  * and lookup the lpni again
1468                  */
1469                 lnet_peer_ni_decref_locked(lpni);
1470                 lpni = lnet_find_peer_ni_locked(nid);
1471                 if (!lpni) {
1472                         rc = -ENOENT;
1473                         goto out_unlock;
1474                 }
1475                 lp = lpni->lpni_peer_net->lpn_peer;
1476
1477                 /* If we find that the peer has discovery disabled then we will
1478                  * not modify whatever primary NID is currently set for this
1479                  * peer. Thus, we can break out of this loop even if the peer
1480                  * is not fully up to date.
1481                  */
1482                 if (lnet_is_discovery_disabled(lp))
1483                         break;
1484         }
1485         primary_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
1486 out_decref:
1487         lnet_peer_ni_decref_locked(lpni);
1488 out_unlock:
1489         lnet_net_unlock(cpt);
1490
1491         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nid2str(nid),
1492                libcfs_nid2str(primary_nid), rc);
1493         return primary_nid;
1494 }
1495 EXPORT_SYMBOL(LNetPrimaryNID);
1496
1497 struct lnet_peer_net *
1498 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1499 {
1500         struct lnet_peer_net *peer_net;
1501         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1502                 if (peer_net->lpn_net_id == net_id)
1503                         return peer_net;
1504         }
1505         return NULL;
1506 }
1507
1508 /*
1509  * Attach a peer_ni to a peer_net and peer. This function assumes
1510  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1511  * may be attached to a different peer, in which case it will be
1512  * properly detached first. The whole operation is done atomically.
1513  *
1514  * This function consumes the reference on lpni and Always returns 0.
1515  * This is the last function called from functions that do return an
1516  * int, so returning 0 here allows the compiler to do a tail call.
1517  */
1518 static int
1519 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1520                                 struct lnet_peer_net *lpn,
1521                                 struct lnet_peer_ni *lpni,
1522                                 unsigned flags)
1523 {
1524         struct lnet_peer_table *ptable;
1525         bool new_lpn = false;
1526         int rc;
1527
1528         /* Install the new peer_ni */
1529         lnet_net_lock(LNET_LOCK_EX);
1530         /* Add peer_ni to global peer table hash, if necessary. */
1531         if (list_empty(&lpni->lpni_hashlist)) {
1532                 int hash = lnet_nid2peerhash(&lpni->lpni_nid);
1533
1534                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1535                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1536                 ptable->pt_version++;
1537                 lnet_peer_ni_addref_locked(lpni);
1538         }
1539
1540         /* Detach the peer_ni from an existing peer, if necessary. */
1541         if (lpni->lpni_peer_net) {
1542                 LASSERT(lpni->lpni_peer_net != lpn);
1543                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1544                 lnet_peer_detach_peer_ni_locked(lpni);
1545                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1546                 lpni->lpni_peer_net = NULL;
1547         }
1548
1549         /* Add peer_ni to peer_net */
1550         lpni->lpni_peer_net = lpn;
1551         if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1552                 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1553         else
1554                 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1555         lnet_update_peer_net_healthv(lpni);
1556         lnet_peer_net_addref_locked(lpn);
1557
1558         /* Add peer_net to peer */
1559         if (!lpn->lpn_peer) {
1560                 new_lpn = true;
1561                 lpn->lpn_peer = lp;
1562                 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1563                         list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1564                 else
1565                         list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1566                 lnet_peer_addref_locked(lp);
1567         }
1568
1569         /* Add peer to global peer list, if necessary */
1570         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1571         if (list_empty(&lp->lp_peer_list)) {
1572                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1573                 ptable->pt_peers++;
1574         }
1575
1576
1577         /* Update peer state */
1578         spin_lock(&lp->lp_lock);
1579         if (flags & LNET_PEER_CONFIGURED) {
1580                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1581                         lp->lp_state |= LNET_PEER_CONFIGURED;
1582         }
1583         if (flags & LNET_PEER_MULTI_RAIL) {
1584                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1585                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1586                         lnet_peer_clr_non_mr_pref_nids(lp);
1587                 }
1588         }
1589         spin_unlock(&lp->lp_lock);
1590
1591         lp->lp_nnis++;
1592
1593         /* apply UDSPs */
1594         if (new_lpn) {
1595                 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1596                 if (rc)
1597                         CERROR("Failed to apply UDSPs on lpn %s\n",
1598                                libcfs_net2str(lpn->lpn_net_id));
1599         }
1600         rc = lnet_udsp_apply_policies_on_lpni(lpni);
1601         if (rc)
1602                 CERROR("Failed to apply UDSPs on lpni %s\n",
1603                        libcfs_nidstr(&lpni->lpni_nid));
1604
1605         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1606                libcfs_nidstr(&lp->lp_primary_nid),
1607                libcfs_nidstr(&lpni->lpni_nid), flags);
1608         lnet_peer_ni_decref_locked(lpni);
1609         lnet_net_unlock(LNET_LOCK_EX);
1610
1611         return 0;
1612 }
1613
1614 /*
1615  * Create a new peer, with nid as its primary nid.
1616  *
1617  * Call with the lnet_api_mutex held.
1618  */
1619 static int
1620 lnet_peer_add(lnet_nid_t nid4, unsigned int flags)
1621 {
1622         struct lnet_nid nid;
1623         struct lnet_peer *lp;
1624         struct lnet_peer_net *lpn;
1625         struct lnet_peer_ni *lpni;
1626         int rc = 0;
1627
1628         LASSERT(nid4 != LNET_NID_ANY);
1629
1630         /*
1631          * No need for the lnet_net_lock here, because the
1632          * lnet_api_mutex is held.
1633          */
1634         lpni = lnet_find_peer_ni_locked(nid4);
1635         if (lpni) {
1636                 /* A peer with this NID already exists. */
1637                 lp = lpni->lpni_peer_net->lpn_peer;
1638                 lnet_peer_ni_decref_locked(lpni);
1639                 /*
1640                  * This is an error if the peer was configured and the
1641                  * primary NID differs or an attempt is made to change
1642                  * the Multi-Rail flag. Otherwise the assumption is
1643                  * that an existing peer is being modified.
1644                  */
1645                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1646                         if (lnet_nid_to_nid4(&lp->lp_primary_nid) != nid4)
1647                                 rc = -EEXIST;
1648                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1649                                 rc = -EPERM;
1650                         goto out;
1651                 } else if (!(flags & LNET_PEER_CONFIGURED)) {
1652                         if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid4) {
1653                                 rc = -EEXIST;
1654                                 goto out;
1655                         }
1656                 }
1657                 /* Delete and recreate as a configured peer. */
1658                 rc = lnet_peer_del(lp);
1659                 if (rc)
1660                         goto out;
1661         }
1662
1663         /* Create peer, peer_net, and peer_ni. */
1664         rc = -ENOMEM;
1665         lnet_nid4_to_nid(nid4, &nid);
1666         lp = lnet_peer_alloc(&nid);
1667         if (!lp)
1668                 goto out;
1669         lpn = lnet_peer_net_alloc(LNET_NID_NET(&nid));
1670         if (!lpn)
1671                 goto out_free_lp;
1672         lpni = lnet_peer_ni_alloc(&nid);
1673         if (!lpni)
1674                 goto out_free_lpn;
1675
1676         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1677
1678 out_free_lpn:
1679         LIBCFS_FREE(lpn, sizeof(*lpn));
1680 out_free_lp:
1681         LIBCFS_FREE(lp, sizeof(*lp));
1682 out:
1683         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1684                libcfs_nid2str(nid4), flags, rc);
1685         return rc;
1686 }
1687
1688 /*
1689  * Add a NID to a peer. Call with ln_api_mutex held.
1690  *
1691  * Error codes:
1692  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1693  *  -EEXIST:   The NID was configured by DLC for a different peer.
1694  *  -ENOMEM:   Out of memory.
1695  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1696  *             non-multi-rail peer.
1697  */
1698 static int
1699 lnet_peer_add_nid(struct lnet_peer *lp, lnet_nid_t nid4, unsigned int flags)
1700 {
1701         struct lnet_peer_net *lpn;
1702         struct lnet_peer_ni *lpni;
1703         struct lnet_nid nid;
1704         int rc = 0;
1705
1706         LASSERT(lp);
1707         LASSERT(nid4 != LNET_NID_ANY);
1708
1709         lnet_nid4_to_nid(nid4, &nid);
1710
1711         /* A configured peer can only be updated through configuration. */
1712         if (!(flags & LNET_PEER_CONFIGURED)) {
1713                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1714                         rc = -EPERM;
1715                         goto out;
1716                 }
1717         }
1718
1719         /*
1720          * The MULTI_RAIL flag can be set but not cleared, because
1721          * that would leave the peer struct in an invalid state.
1722          */
1723         if (flags & LNET_PEER_MULTI_RAIL) {
1724                 spin_lock(&lp->lp_lock);
1725                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1726                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1727                         lnet_peer_clr_non_mr_pref_nids(lp);
1728                 }
1729                 spin_unlock(&lp->lp_lock);
1730         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1731                 rc = -EPERM;
1732                 goto out;
1733         }
1734
1735         lpni = lnet_find_peer_ni_locked(nid4);
1736         if (lpni) {
1737                 /*
1738                  * A peer_ni already exists. This is only a problem if
1739                  * it is not connected to this peer and was configured
1740                  * by DLC.
1741                  */
1742                 if (lpni->lpni_peer_net->lpn_peer == lp)
1743                         goto out_free_lpni;
1744                 if (lnet_peer_ni_is_configured(lpni)) {
1745                         rc = -EEXIST;
1746                         goto out_free_lpni;
1747                 }
1748                 /* If this is the primary NID, destroy the peer. */
1749                 if (lnet_peer_ni_is_primary(lpni)) {
1750                         struct lnet_peer *rtr_lp =
1751                                 lpni->lpni_peer_net->lpn_peer;
1752                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1753                         /*
1754                          * if we're trying to delete a router it means
1755                          * we're moving this peer NI to a new peer so must
1756                          * transfer router properties to the new peer
1757                          */
1758                         if (rtr_refcount > 0) {
1759                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1760                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1761                         }
1762                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1763                         lnet_peer_ni_decref_locked(lpni);
1764                         lpni = lnet_peer_ni_alloc(&nid);
1765                         if (!lpni) {
1766                                 rc = -ENOMEM;
1767                                 goto out_free_lpni;
1768                         }
1769                 }
1770         } else {
1771                 lpni = lnet_peer_ni_alloc(&nid);
1772                 if (!lpni) {
1773                         rc = -ENOMEM;
1774                         goto out_free_lpni;
1775                 }
1776         }
1777
1778         /*
1779          * Get the peer_net. Check that we're not adding a second
1780          * peer_ni on a peer_net of a non-multi-rail peer.
1781          */
1782         lpn = lnet_peer_get_net_locked(lp, LNET_NIDNET(nid4));
1783         if (!lpn) {
1784                 lpn = lnet_peer_net_alloc(LNET_NIDNET(nid4));
1785                 if (!lpn) {
1786                         rc = -ENOMEM;
1787                         goto out_free_lpni;
1788                 }
1789         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1790                 rc = -ENOTUNIQ;
1791                 goto out_free_lpni;
1792         }
1793
1794         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1795
1796 out_free_lpni:
1797         lnet_peer_ni_decref_locked(lpni);
1798 out:
1799         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1800                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nid2str(nid4),
1801                flags, rc);
1802         return rc;
1803 }
1804
1805 /*
1806  * Update the primary NID of a peer, if possible.
1807  *
1808  * Call with the lnet_api_mutex held.
1809  */
1810 static int
1811 lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid,
1812                           unsigned int flags)
1813 {
1814         struct lnet_nid old = lp->lp_primary_nid;
1815         int rc = 0;
1816
1817         if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid)
1818                 goto out;
1819
1820         lnet_nid4_to_nid(nid, &lp->lp_primary_nid);
1821
1822         rc = lnet_peer_add_nid(lp, nid, flags);
1823         if (rc) {
1824                 lp->lp_primary_nid = old;
1825                 goto out;
1826         }
1827 out:
1828         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1829                libcfs_nidstr(&old), libcfs_nid2str(nid), rc);
1830
1831         return rc;
1832 }
1833
1834 /*
1835  * lpni creation initiated due to traffic either sending or receiving.
1836  */
1837 static int
1838 lnet_peer_ni_traffic_add(struct lnet_nid *nid, struct lnet_nid *pref)
1839 {
1840         struct lnet_peer *lp;
1841         struct lnet_peer_net *lpn;
1842         struct lnet_peer_ni *lpni;
1843         unsigned flags = 0;
1844         int rc = 0;
1845
1846         if (LNET_NID_IS_ANY(nid)) {
1847                 rc = -EINVAL;
1848                 goto out;
1849         }
1850
1851         /* lnet_net_lock is not needed here because ln_api_lock is held */
1852         lpni = lnet_peer_ni_find_locked(nid);
1853         if (lpni) {
1854                 /*
1855                  * We must have raced with another thread. Since we
1856                  * know next to nothing about a peer_ni created by
1857                  * traffic, we just assume everything is ok and
1858                  * return.
1859                  */
1860                 lnet_peer_ni_decref_locked(lpni);
1861                 goto out;
1862         }
1863
1864         /* Create peer, peer_net, and peer_ni. */
1865         rc = -ENOMEM;
1866         lp = lnet_peer_alloc(nid);
1867         if (!lp)
1868                 goto out;
1869         lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1870         if (!lpn)
1871                 goto out_free_lp;
1872         lpni = lnet_peer_ni_alloc(nid);
1873         if (!lpni)
1874                 goto out_free_lpn;
1875         lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1876
1877         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1878
1879 out_free_lpn:
1880         LIBCFS_FREE(lpn, sizeof(*lpn));
1881 out_free_lp:
1882         LIBCFS_FREE(lp, sizeof(*lp));
1883 out:
1884         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(nid), rc);
1885         return rc;
1886 }
1887
1888 /*
1889  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1890  *
1891  * This API handles the following combinations:
1892  *   Create a peer with its primary NI if only the prim_nid is provided
1893  *   Add a NID to a peer identified by the prim_nid. The peer identified
1894  *   by the prim_nid must already exist.
1895  *   The peer being created may be non-MR.
1896  *
1897  * The caller must hold ln_api_mutex. This prevents the peer from
1898  * being created/modified/deleted by a different thread.
1899  */
1900 int
1901 lnet_add_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid, bool mr, bool temp)
1902 {
1903         struct lnet_peer *lp = NULL;
1904         struct lnet_peer_ni *lpni;
1905         unsigned int flags = 0;
1906
1907         /* The prim_nid must always be specified */
1908         if (prim_nid == LNET_NID_ANY)
1909                 return -EINVAL;
1910
1911         if (!temp)
1912                 flags = LNET_PEER_CONFIGURED;
1913
1914         if (mr)
1915                 flags |= LNET_PEER_MULTI_RAIL;
1916
1917         /*
1918          * If nid isn't specified, we must create a new peer with
1919          * prim_nid as its primary nid.
1920          */
1921         if (nid == LNET_NID_ANY)
1922                 return lnet_peer_add(prim_nid, flags);
1923
1924         /* Look up the prim_nid, which must exist. */
1925         lpni = lnet_find_peer_ni_locked(prim_nid);
1926         if (!lpni)
1927                 return -ENOENT;
1928         lnet_peer_ni_decref_locked(lpni);
1929         lp = lpni->lpni_peer_net->lpn_peer;
1930
1931         /* Peer must have been configured. */
1932         if (!temp && !(lp->lp_state & LNET_PEER_CONFIGURED)) {
1933                 CDEBUG(D_NET, "peer %s was not configured\n",
1934                        libcfs_nid2str(prim_nid));
1935                 return -ENOENT;
1936         }
1937
1938         /* Primary NID must match */
1939         if (lnet_nid_to_nid4(&lp->lp_primary_nid) != prim_nid) {
1940                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1941                        libcfs_nid2str(prim_nid),
1942                        libcfs_nidstr(&lp->lp_primary_nid));
1943                 return -ENODEV;
1944         }
1945
1946         /* Multi-Rail flag must match. */
1947         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1948                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1949                        libcfs_nid2str(prim_nid));
1950                 return -EPERM;
1951         }
1952
1953         return lnet_peer_add_nid(lp, nid, flags);
1954 }
1955
1956 /*
1957  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1958  *
1959  * This API handles the following combinations:
1960  *   Delete a NI from a peer if both prim_nid and nid are provided.
1961  *   Delete a peer if only prim_nid is provided.
1962  *   Delete a peer if its primary nid is provided.
1963  *
1964  * The caller must hold ln_api_mutex. This prevents the peer from
1965  * being modified/deleted by a different thread.
1966  */
1967 int
1968 lnet_del_peer_ni(lnet_nid_t prim_nid, lnet_nid_t nid)
1969 {
1970         struct lnet_peer *lp;
1971         struct lnet_peer_ni *lpni;
1972         unsigned flags;
1973
1974         if (prim_nid == LNET_NID_ANY)
1975                 return -EINVAL;
1976
1977         lpni = lnet_find_peer_ni_locked(prim_nid);
1978         if (!lpni)
1979                 return -ENOENT;
1980         lnet_peer_ni_decref_locked(lpni);
1981         lp = lpni->lpni_peer_net->lpn_peer;
1982
1983         if (prim_nid != lnet_nid_to_nid4(&lp->lp_primary_nid)) {
1984                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1985                        libcfs_nid2str(prim_nid),
1986                        libcfs_nidstr(&lp->lp_primary_nid));
1987                 return -ENODEV;
1988         }
1989
1990         lnet_net_lock(LNET_LOCK_EX);
1991         if (lp->lp_rtr_refcount > 0) {
1992                 lnet_net_unlock(LNET_LOCK_EX);
1993                 CERROR("%s is a router. Can not be deleted\n",
1994                        libcfs_nid2str(prim_nid));
1995                 return -EBUSY;
1996         }
1997         lnet_net_unlock(LNET_LOCK_EX);
1998
1999         if (nid == LNET_NID_ANY || nid == lnet_nid_to_nid4(&lp->lp_primary_nid))
2000                 return lnet_peer_del(lp);
2001
2002         flags = LNET_PEER_CONFIGURED;
2003         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
2004                 flags |= LNET_PEER_MULTI_RAIL;
2005
2006         return lnet_peer_del_nid(lp, nid, flags);
2007 }
2008
2009 void
2010 lnet_destroy_peer_ni_locked(struct kref *ref)
2011 {
2012         struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
2013                                                  lpni_kref);
2014         struct lnet_peer_table *ptable;
2015         struct lnet_peer_net *lpn;
2016
2017         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
2018
2019         LASSERT(kref_read(&lpni->lpni_kref) == 0);
2020         LASSERT(list_empty(&lpni->lpni_txq));
2021         LASSERT(lpni->lpni_txqnob == 0);
2022         LASSERT(list_empty(&lpni->lpni_peer_nis));
2023         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
2024
2025         lpn = lpni->lpni_peer_net;
2026         lpni->lpni_peer_net = NULL;
2027         lpni->lpni_net = NULL;
2028
2029         if (!list_empty(&lpni->lpni_hashlist)) {
2030                 /* remove the peer ni from the zombie list */
2031                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
2032                 spin_lock(&ptable->pt_zombie_lock);
2033                 list_del_init(&lpni->lpni_hashlist);
2034                 ptable->pt_zombies--;
2035                 spin_unlock(&ptable->pt_zombie_lock);
2036         }
2037
2038         if (lpni->lpni_pref_nnids > 1) {
2039                 struct lnet_nid_list *ne, *tmp;
2040
2041                 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
2042                                          nl_list) {
2043                         list_del_init(&ne->nl_list);
2044                         LIBCFS_FREE(ne, sizeof(*ne));
2045                 }
2046         }
2047         LIBCFS_FREE(lpni, sizeof(*lpni));
2048
2049         if (lpn)
2050                 lnet_peer_net_decref_locked(lpn);
2051 }
2052
2053 struct lnet_peer_ni *
2054 lnet_nid2peerni_ex(struct lnet_nid *nid, int cpt)
2055 {
2056         struct lnet_peer_ni *lpni = NULL;
2057         int rc;
2058
2059         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2060                 return ERR_PTR(-ESHUTDOWN);
2061
2062         /*
2063          * find if a peer_ni already exists.
2064          * If so then just return that.
2065          */
2066         lpni = lnet_peer_ni_find_locked(nid);
2067         if (lpni)
2068                 return lpni;
2069
2070         lnet_net_unlock(cpt);
2071
2072         rc = lnet_peer_ni_traffic_add(nid, NULL);
2073         if (rc) {
2074                 lpni = ERR_PTR(rc);
2075                 goto out_net_relock;
2076         }
2077
2078         lpni = lnet_peer_ni_find_locked(nid);
2079         LASSERT(lpni);
2080
2081 out_net_relock:
2082         lnet_net_lock(cpt);
2083
2084         return lpni;
2085 }
2086
2087 /*
2088  * Get a peer_ni for the given nid, create it if necessary. Takes a
2089  * hold on the peer_ni.
2090  */
2091 struct lnet_peer_ni *
2092 lnet_peerni_by_nid_locked(struct lnet_nid *nid,
2093                         struct lnet_nid *pref, int cpt)
2094 {
2095         struct lnet_peer_ni *lpni = NULL;
2096         int rc;
2097
2098         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2099                 return ERR_PTR(-ESHUTDOWN);
2100
2101         /*
2102          * find if a peer_ni already exists.
2103          * If so then just return that.
2104          */
2105         lpni = lnet_peer_ni_find_locked(nid);
2106         if (lpni)
2107                 return lpni;
2108
2109         /*
2110          * Slow path:
2111          * use the lnet_api_mutex to serialize the creation of the peer_ni
2112          * and the creation/deletion of the local ni/net. When a local ni is
2113          * created, if there exists a set of peer_nis on that network,
2114          * they need to be traversed and updated. When a local NI is
2115          * deleted, which could result in a network being deleted, then
2116          * all peer nis on that network need to be removed as well.
2117          *
2118          * Creation through traffic should also be serialized with
2119          * creation through DLC.
2120          */
2121         lnet_net_unlock(cpt);
2122         mutex_lock(&the_lnet.ln_api_mutex);
2123         /*
2124          * Shutdown is only set under the ln_api_lock, so a single
2125          * check here is sufficent.
2126          */
2127         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2128                 lpni = ERR_PTR(-ESHUTDOWN);
2129                 goto out_mutex_unlock;
2130         }
2131
2132         rc = lnet_peer_ni_traffic_add(nid, pref);
2133         if (rc) {
2134                 lpni = ERR_PTR(rc);
2135                 goto out_mutex_unlock;
2136         }
2137
2138         lpni = lnet_peer_ni_find_locked(nid);
2139         LASSERT(lpni);
2140
2141 out_mutex_unlock:
2142         mutex_unlock(&the_lnet.ln_api_mutex);
2143         lnet_net_lock(cpt);
2144
2145         /* Lock has been dropped, check again for shutdown. */
2146         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2147                 if (!IS_ERR(lpni))
2148                         lnet_peer_ni_decref_locked(lpni);
2149                 lpni = ERR_PTR(-ESHUTDOWN);
2150         }
2151
2152         return lpni;
2153 }
2154
2155 struct lnet_peer_ni *
2156 lnet_nid2peerni_locked(lnet_nid_t nid4, lnet_nid_t pref4, int cpt)
2157 {
2158         struct lnet_nid nid, pref;
2159
2160         lnet_nid4_to_nid(nid4, &nid);
2161         lnet_nid4_to_nid(pref4, &pref);
2162         if (pref4 == LNET_NID_ANY)
2163                 return lnet_peerni_by_nid_locked(&nid, NULL, cpt);
2164         else
2165                 return lnet_peerni_by_nid_locked(&nid, &pref, cpt);
2166 }
2167
2168 bool
2169 lnet_peer_gw_discovery(struct lnet_peer *lp)
2170 {
2171         bool rc = false;
2172
2173         spin_lock(&lp->lp_lock);
2174         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2175                 rc = true;
2176         spin_unlock(&lp->lp_lock);
2177
2178         return rc;
2179 }
2180
2181 bool
2182 lnet_peer_is_uptodate(struct lnet_peer *lp)
2183 {
2184         bool rc;
2185
2186         spin_lock(&lp->lp_lock);
2187         rc = lnet_peer_is_uptodate_locked(lp);
2188         spin_unlock(&lp->lp_lock);
2189         return rc;
2190 }
2191
2192 /*
2193  * Is a peer uptodate from the point of view of discovery?
2194  *
2195  * If it is currently being processed, obviously not.
2196  * A forced Ping or Push is also handled by the discovery thread.
2197  *
2198  * Otherwise look at whether the peer needs rediscovering.
2199  */
2200 bool
2201 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2202 __must_hold(&lp->lp_lock)
2203 {
2204         bool rc;
2205
2206         if (lp->lp_state & (LNET_PEER_DISCOVERING |
2207                             LNET_PEER_FORCE_PING |
2208                             LNET_PEER_FORCE_PUSH)) {
2209                 rc = false;
2210         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2211                 rc = false;
2212         } else if (lnet_peer_needs_push(lp)) {
2213                 rc = false;
2214         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2215                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2216                         rc = true;
2217                 else
2218                         rc = false;
2219         } else {
2220                 rc = false;
2221         }
2222
2223         return rc;
2224 }
2225
2226 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2227 void
2228 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2229 {
2230         /* The discovery thread holds net_lock/EX and lp_lock when it splices
2231          * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2232          * when adding to the list and queuing the peer to ensure that we do not
2233          * strand any messages on the lp_dc_pendq. This scheme ensures the
2234          * message will be resent even if the peer is already being discovered.
2235          * Therefore we needn't check the return value of
2236          * lnet_peer_queue_for_discovery(lp).
2237          */
2238         lnet_net_lock(LNET_LOCK_EX);
2239         spin_lock(&lp->lp_lock);
2240         list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2241         spin_unlock(&lp->lp_lock);
2242         lnet_peer_queue_for_discovery(lp);
2243         lnet_net_unlock(LNET_LOCK_EX);
2244 }
2245
2246 /*
2247  * Queue a peer for the attention of the discovery thread.  Call with
2248  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2249  * -EALREADY if the peer was already queued.
2250  */
2251 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2252 {
2253         int rc;
2254
2255         spin_lock(&lp->lp_lock);
2256         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2257                 lp->lp_state |= LNET_PEER_DISCOVERING;
2258         spin_unlock(&lp->lp_lock);
2259         if (list_empty(&lp->lp_dc_list)) {
2260                 lnet_peer_addref_locked(lp);
2261                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2262                 wake_up(&the_lnet.ln_dc_waitq);
2263                 rc = 0;
2264         } else {
2265                 rc = -EALREADY;
2266         }
2267
2268         CDEBUG(D_NET, "Queue peer %s: %d\n",
2269                libcfs_nidstr(&lp->lp_primary_nid), rc);
2270
2271         return rc;
2272 }
2273
2274 /*
2275  * Discovery of a peer is complete. Wake all waiters on the peer.
2276  * Call with lnet_net_lock/EX held.
2277  */
2278 static void lnet_peer_discovery_complete(struct lnet_peer *lp, int dc_error)
2279 {
2280         struct lnet_msg *msg, *tmp;
2281         int rc = 0;
2282         LIST_HEAD(pending_msgs);
2283
2284         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2285                libcfs_nidstr(&lp->lp_primary_nid));
2286
2287         list_del_init(&lp->lp_dc_list);
2288         spin_lock(&lp->lp_lock);
2289         if (dc_error) {
2290                 lp->lp_dc_error = dc_error;
2291                 lp->lp_state &= ~LNET_PEER_DISCOVERING;
2292                 lp->lp_state |= LNET_PEER_REDISCOVER;
2293         }
2294         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2295         spin_unlock(&lp->lp_lock);
2296         wake_up(&lp->lp_dc_waitq);
2297
2298         if (lp->lp_rtr_refcount > 0)
2299                 lnet_router_discovery_complete(lp);
2300
2301         lnet_net_unlock(LNET_LOCK_EX);
2302
2303         /* iterate through all pending messages and send them again */
2304         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2305                 list_del_init(&msg->msg_list);
2306                 if (dc_error) {
2307                         lnet_finalize(msg, dc_error);
2308                         continue;
2309                 }
2310
2311                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2312                        lnet_msgtyp2str(msg->msg_type),
2313                        libcfs_id2str(msg->msg_target));
2314                 rc = lnet_send(msg->msg_src_nid_param, msg,
2315                                msg->msg_rtr_nid_param);
2316                 if (rc < 0) {
2317                         CNETERR("Error sending %s to %s: %d\n",
2318                                lnet_msgtyp2str(msg->msg_type),
2319                                libcfs_id2str(msg->msg_target), rc);
2320                         lnet_finalize(msg, rc);
2321                 }
2322         }
2323         lnet_net_lock(LNET_LOCK_EX);
2324         lnet_peer_decref_locked(lp);
2325 }
2326
2327 /*
2328  * Handle inbound push.
2329  * Like any event handler, called with lnet_res_lock/CPT held.
2330  */
2331 void lnet_peer_push_event(struct lnet_event *ev)
2332 {
2333         struct lnet_ping_buffer *pbuf;
2334         struct lnet_peer *lp;
2335
2336         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2337
2338         /* lnet_find_peer() adds a refcount */
2339         lp = lnet_find_peer(ev->source.nid);
2340         if (!lp) {
2341                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2342                        libcfs_nid2str(ev->initiator.nid),
2343                        libcfs_nid2str(ev->source.nid));
2344                 pbuf->pb_needs_post = true;
2345                 return;
2346         }
2347
2348         /* Ensure peer state remains consistent while we modify it. */
2349         spin_lock(&lp->lp_lock);
2350
2351         /*
2352          * If some kind of error happened the contents of the message
2353          * cannot be used. Clear the NIDS_UPTODATE and set the
2354          * FORCE_PING flag to trigger a ping.
2355          */
2356         if (ev->status) {
2357                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2358                 lp->lp_state |= LNET_PEER_FORCE_PING;
2359                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2360                        ev->status,
2361                        libcfs_nidstr(&lp->lp_primary_nid),
2362                        libcfs_nid2str(ev->source.nid));
2363                 goto out;
2364         }
2365
2366         /*
2367          * A push with invalid or corrupted info. Clear the UPTODATE
2368          * flag to trigger a ping.
2369          */
2370         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2371                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2372                 lp->lp_state |= LNET_PEER_FORCE_PING;
2373                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2374                        libcfs_nidstr(&lp->lp_primary_nid));
2375                 goto out;
2376         }
2377
2378         /*
2379          * Make sure we'll allocate the correct size ping buffer when
2380          * pinging the peer.
2381          */
2382         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2383                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2384
2385         /*
2386          * A non-Multi-Rail peer is not supposed to be capable of
2387          * sending a push.
2388          */
2389         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2390                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2391                        libcfs_nidstr(&lp->lp_primary_nid));
2392                 goto out;
2393         }
2394
2395         /*
2396          * The peer may have discovery disabled at its end. Set
2397          * NO_DISCOVERY as appropriate.
2398          */
2399         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2400                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2401                        libcfs_nidstr(&lp->lp_primary_nid));
2402                 /*
2403                  * Mark the peer for deletion if we already know about it
2404                  * and it's going from discovery set to no discovery set
2405                  */
2406                 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2407                                       LNET_PEER_DISCOVERING)) &&
2408                      lp->lp_state & LNET_PEER_DISCOVERED) {
2409                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2410                                libcfs_nidstr(&lp->lp_primary_nid),
2411                                lp->lp_state);
2412                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2413                 }
2414                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2415         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2416                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2417                        libcfs_nidstr(&lp->lp_primary_nid));
2418                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2419         }
2420
2421         /*
2422          * Update the MULTI_RAIL flag based on the push. If the peer
2423          * was configured with DLC then the setting should match what
2424          * DLC put in.
2425          * NB: We verified above that the MR feature bit is set in pi_features
2426          */
2427         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2428                 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2429                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2430         } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2431                 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2432                       libcfs_nidstr(&lp->lp_primary_nid));
2433         } else if (lnet_peer_discovery_disabled) {
2434                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2435                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2436         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2437                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2438                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2439         } else {
2440                 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2441                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2442                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2443                 lnet_peer_clr_non_mr_pref_nids(lp);
2444         }
2445
2446         /*
2447          * Check for truncation of the Put message. Clear the
2448          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2449          * and tell discovery to allocate a bigger buffer.
2450          */
2451         if (ev->mlength < ev->rlength) {
2452                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2453                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2454                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2455                 lp->lp_state |= LNET_PEER_FORCE_PING;
2456                 CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
2457                        libcfs_nidstr(&lp->lp_primary_nid),
2458                        pbuf->pb_info.pi_nnis);
2459                 goto out;
2460         }
2461
2462         /* always assume new data */
2463         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2464         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2465
2466         /*
2467          * If there is data present that hasn't been processed yet,
2468          * we'll replace it if the Put contained newer data and it
2469          * fits. We're racing with a Ping or earlier Push in this
2470          * case.
2471          */
2472         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2473                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2474                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2475                     pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
2476                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2477                                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2478                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2479                               libcfs_nidstr(&lp->lp_primary_nid),
2480                               LNET_PING_BUFFER_SEQNO(pbuf),
2481                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2482                 }
2483                 goto out;
2484         }
2485
2486         /*
2487          * Allocate a buffer to copy the data. On a failure we drop
2488          * the Push and set FORCE_PING to force the discovery
2489          * thread to fix the problem by pinging the peer.
2490          */
2491         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
2492         if (!lp->lp_data) {
2493                 lp->lp_state |= LNET_PEER_FORCE_PING;
2494                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2495                        libcfs_nidstr(&lp->lp_primary_nid),
2496                        LNET_PING_BUFFER_SEQNO(pbuf));
2497                 goto out;
2498         }
2499
2500         /* Success */
2501         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2502                LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
2503         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2504         CDEBUG(D_NET, "Received Push %s %u\n",
2505                libcfs_nidstr(&lp->lp_primary_nid),
2506                LNET_PING_BUFFER_SEQNO(pbuf));
2507
2508 out:
2509         /* We've processed this buffer. It can be reposted */
2510         pbuf->pb_needs_post = true;
2511
2512         /*
2513          * Queue the peer for discovery if not done, force it on the request
2514          * queue and wake the discovery thread if the peer was already queued,
2515          * because its status changed.
2516          */
2517         spin_unlock(&lp->lp_lock);
2518         lnet_net_lock(LNET_LOCK_EX);
2519         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2520                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2521                 wake_up(&the_lnet.ln_dc_waitq);
2522         }
2523         /* Drop refcount from lookup */
2524         lnet_peer_decref_locked(lp);
2525         lnet_net_unlock(LNET_LOCK_EX);
2526 }
2527
2528 /*
2529  * Clear the discovery error state, unless we're already discovering
2530  * this peer, in which case the error is current.
2531  */
2532 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2533 {
2534         spin_lock(&lp->lp_lock);
2535         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2536                 lp->lp_dc_error = 0;
2537         spin_unlock(&lp->lp_lock);
2538 }
2539
2540 /*
2541  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2542  * dropped/retaken within this function. An lnet_peer_ni is passed in
2543  * because discovery could tear down an lnet_peer.
2544  */
2545 int
2546 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2547 {
2548         DEFINE_WAIT(wait);
2549         struct lnet_peer *lp;
2550         int rc = 0;
2551         int count = 0;
2552
2553 again:
2554         lnet_net_unlock(cpt);
2555         lnet_net_lock(LNET_LOCK_EX);
2556         lp = lpni->lpni_peer_net->lpn_peer;
2557         lnet_peer_clear_discovery_error(lp);
2558
2559         /*
2560          * We're willing to be interrupted. The lpni can become a
2561          * zombie if we race with DLC, so we must check for that.
2562          */
2563         for (;;) {
2564                 /* Keep lp alive when the lnet_net_lock is unlocked */
2565                 lnet_peer_addref_locked(lp);
2566                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2567                 if (signal_pending(current))
2568                         break;
2569                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2570                         break;
2571                 /*
2572                  * Don't repeat discovery if discovery is disabled. This is
2573                  * done to ensure we can use discovery as a standard ping as
2574                  * well for backwards compatibility with routers which do not
2575                  * have discovery or have discovery disabled
2576                  */
2577                 if (lnet_is_discovery_disabled(lp) && count > 0)
2578                         break;
2579                 if (lp->lp_dc_error)
2580                         break;
2581                 if (lnet_peer_is_uptodate(lp))
2582                         break;
2583                 lnet_peer_queue_for_discovery(lp);
2584                 count++;
2585                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2586
2587                 /*
2588                  * If caller requested a non-blocking operation then
2589                  * return immediately. Once discovery is complete any
2590                  * pending messages that were stopped due to discovery
2591                  * will be transmitted.
2592                  */
2593                 if (!block)
2594                         break;
2595
2596                 lnet_net_unlock(LNET_LOCK_EX);
2597                 schedule();
2598                 finish_wait(&lp->lp_dc_waitq, &wait);
2599                 lnet_net_lock(LNET_LOCK_EX);
2600                 lnet_peer_decref_locked(lp);
2601                 /* Peer may have changed */
2602                 lp = lpni->lpni_peer_net->lpn_peer;
2603         }
2604         finish_wait(&lp->lp_dc_waitq, &wait);
2605
2606         lnet_net_unlock(LNET_LOCK_EX);
2607         lnet_net_lock(cpt);
2608         lnet_peer_decref_locked(lp);
2609         /*
2610          * The peer may have changed, so re-check and rediscover if that turns
2611          * out to have been the case. The reference count on lp ensured that
2612          * even if it was unlinked from lpni the memory could not be recycled.
2613          * Thus the check below is sufficient to determine whether the peer
2614          * changed. If the peer changed, then lp must not be dereferenced.
2615          */
2616         if (lp != lpni->lpni_peer_net->lpn_peer)
2617                 goto again;
2618
2619         if (signal_pending(current))
2620                 rc = -EINTR;
2621         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2622                 rc = -ESHUTDOWN;
2623         else if (lp->lp_dc_error)
2624                 rc = lp->lp_dc_error;
2625         else if (!block)
2626                 CDEBUG(D_NET, "non-blocking discovery\n");
2627         else if (!lnet_peer_is_uptodate(lp) && !lnet_is_discovery_disabled(lp))
2628                 goto again;
2629
2630         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2631                (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2632                libcfs_nidstr(&lpni->lpni_nid), rc,
2633                (!block) ? "pending discovery" : "discovery complete");
2634
2635         return rc;
2636 }
2637
2638 /* Handle an incoming ack for a push. */
2639 static void
2640 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2641 {
2642         struct lnet_ping_buffer *pbuf;
2643
2644         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2645         spin_lock(&lp->lp_lock);
2646         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2647         lp->lp_push_error = ev->status;
2648         if (ev->status)
2649                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2650         else
2651                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2652         spin_unlock(&lp->lp_lock);
2653
2654         CDEBUG(D_NET, "peer %s ev->status %d\n",
2655                libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2656 }
2657
2658 /* Handle a Reply message. This is the reply to a Ping message. */
2659 static void
2660 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2661 {
2662         struct lnet_ping_buffer *pbuf;
2663         int rc;
2664
2665         spin_lock(&lp->lp_lock);
2666
2667         lnet_nid4_to_nid(ev->target.nid, &lp->lp_disc_src_nid);
2668         lnet_nid4_to_nid(ev->source.nid, &lp->lp_disc_dst_nid);
2669
2670         /*
2671          * If some kind of error happened the contents of message
2672          * cannot be used. Set PING_FAILED to trigger a retry.
2673          */
2674         if (ev->status) {
2675                 lp->lp_state |= LNET_PEER_PING_FAILED;
2676                 lp->lp_ping_error = ev->status;
2677                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2678                        ev->status,
2679                        libcfs_nidstr(&lp->lp_primary_nid),
2680                        libcfs_nid2str(ev->source.nid));
2681                 goto out;
2682         }
2683
2684         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2685         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2686                 lnet_swap_pinginfo(pbuf);
2687
2688         /*
2689          * A reply with invalid or corrupted info. Set PING_FAILED to
2690          * trigger a retry.
2691          */
2692         rc = lnet_ping_info_validate(&pbuf->pb_info);
2693         if (rc) {
2694                 lp->lp_state |= LNET_PEER_PING_FAILED;
2695                 lp->lp_ping_error = 0;
2696                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2697                        libcfs_nidstr(&lp->lp_primary_nid), rc);
2698                 goto out;
2699         }
2700
2701         /*
2702          * The peer may have discovery disabled at its end. Set
2703          * NO_DISCOVERY as appropriate.
2704          */
2705         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY) ||
2706             lnet_peer_discovery_disabled) {
2707                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2708                        libcfs_nidstr(&lp->lp_primary_nid));
2709
2710                 /* Detect whether this peer has toggled discovery from on to
2711                  * off and whether we can delete and re-create the peer. Peers
2712                  * that were manually configured cannot be deleted by discovery.
2713                  * We need to delete this peer and re-create it if the peer was
2714                  * not configured manually, is currently considered DD capable,
2715                  * and either:
2716                  * 1. We've already discovered the peer (the peer has toggled
2717                  *    the discovery feature from on to off), or
2718                  * 2. The peer is considered MR, but it was not user configured
2719                  *    (this was a "temporary" peer created via the kernel APIs
2720                  *     that we're discovering for the first time)
2721                  */
2722                 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2723                                       LNET_PEER_NO_DISCOVERY)) &&
2724                     (lp->lp_state & (LNET_PEER_DISCOVERED |
2725                                      LNET_PEER_MULTI_RAIL))) {
2726                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2727                                libcfs_nidstr(&lp->lp_primary_nid),
2728                                lp->lp_state);
2729                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2730                 }
2731                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2732         } else {
2733                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2734                        libcfs_nidstr(&lp->lp_primary_nid));
2735                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2736         }
2737
2738         /*
2739          * Update the MULTI_RAIL flag based on the reply. If the peer
2740          * was configured with DLC then the setting should match what
2741          * DLC put in.
2742          */
2743         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2744                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2745                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2746                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2747                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2748                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2749                               libcfs_nidstr(&lp->lp_primary_nid));
2750                 } else if (lnet_peer_discovery_disabled) {
2751                         CDEBUG(D_NET,
2752                                "peer %s(%p) not MR: DD disabled locally\n",
2753                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2754                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2755                         CDEBUG(D_NET,
2756                                "peer %s(%p) not MR: DD disabled remotely\n",
2757                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2758                 } else {
2759                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2760                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2761                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2762                         lnet_peer_clr_non_mr_pref_nids(lp);
2763                 }
2764         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2765                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2766                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2767                               libcfs_nidstr(&lp->lp_primary_nid));
2768                 } else {
2769                         CERROR("Multi-Rail state vanished from %s\n",
2770                                libcfs_nidstr(&lp->lp_primary_nid));
2771                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2772                 }
2773         }
2774
2775         /*
2776          * Make sure we'll allocate the correct size ping buffer when
2777          * pinging the peer.
2778          */
2779         if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
2780                 lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
2781
2782         /*
2783          * Check for truncation of the Reply. Clear PING_SENT and set
2784          * PING_FAILED to trigger a retry.
2785          */
2786         if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
2787                 if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
2788                         the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
2789                 lp->lp_state |= LNET_PEER_PING_FAILED;
2790                 lp->lp_ping_error = 0;
2791                 CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
2792                        libcfs_nidstr(&lp->lp_primary_nid),
2793                        pbuf->pb_info.pi_nnis);
2794                 goto out;
2795         }
2796
2797         /*
2798          * Check the sequence numbers in the reply. These are only
2799          * available if the reply came from a Multi-Rail peer.
2800          */
2801         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2802             pbuf->pb_info.pi_nnis > 1 &&
2803             lnet_nid_to_nid4(&lp->lp_primary_nid) ==
2804             pbuf->pb_info.pi_ni[1].ns_nid) {
2805                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2806                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2807                                 libcfs_nidstr(&lp->lp_primary_nid),
2808                                 LNET_PING_BUFFER_SEQNO(pbuf),
2809                                 lp->lp_peer_seqno);
2810
2811                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2812         }
2813
2814         /* We're happy with the state of the data in the buffer. */
2815         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2816                libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2817                lp->lp_state);
2818         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2819                 lnet_ping_buffer_decref(lp->lp_data);
2820         else
2821                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2822         lnet_ping_buffer_addref(pbuf);
2823         lp->lp_data = pbuf;
2824 out:
2825         lp->lp_state &= ~LNET_PEER_PING_SENT;
2826         spin_unlock(&lp->lp_lock);
2827
2828         lnet_net_lock(LNET_LOCK_EX);
2829         /*
2830          * If this peer is a gateway, call the routing callback to
2831          * handle the ping reply
2832          */
2833         if (lp->lp_rtr_refcount > 0)
2834                 lnet_router_discovery_ping_reply(lp);
2835         lnet_net_unlock(LNET_LOCK_EX);
2836 }
2837
2838 /*
2839  * Send event handling. Only matters for error cases, where we clean
2840  * up state on the peer and peer_ni that would otherwise be updated in
2841  * the REPLY event handler for a successful Ping, and the ACK event
2842  * handler for a successful Push.
2843  */
2844 static int
2845 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2846 {
2847         int rc = 0;
2848
2849         if (!ev->status)
2850                 goto out;
2851
2852         spin_lock(&lp->lp_lock);
2853         if (ev->msg_type == LNET_MSG_GET) {
2854                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2855                 lp->lp_state |= LNET_PEER_PING_FAILED;
2856                 lp->lp_ping_error = ev->status;
2857         } else { /* ev->msg_type == LNET_MSG_PUT */
2858                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2859                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2860                 lp->lp_push_error = ev->status;
2861         }
2862         spin_unlock(&lp->lp_lock);
2863         rc = LNET_REDISCOVER_PEER;
2864 out:
2865         CDEBUG(D_NET, "%s Send to %s: %d\n",
2866                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2867                 libcfs_nid2str(ev->target.nid), rc);
2868         return rc;
2869 }
2870
2871 /*
2872  * Unlink event handling. This event is only seen if a call to
2873  * LNetMDUnlink() caused the event to be unlinked. If this call was
2874  * made after the event was set up in LNetGet() or LNetPut() then we
2875  * assume the Ping or Push timed out.
2876  */
2877 static void
2878 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2879 {
2880         spin_lock(&lp->lp_lock);
2881         /* We've passed through LNetGet() */
2882         if (lp->lp_state & LNET_PEER_PING_SENT) {
2883                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2884                 lp->lp_state |= LNET_PEER_PING_FAILED;
2885                 lp->lp_ping_error = -ETIMEDOUT;
2886                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2887                         libcfs_nidstr(&lp->lp_primary_nid));
2888         }
2889         /* We've passed through LNetPut() */
2890         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2891                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2892                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2893                 lp->lp_push_error = -ETIMEDOUT;
2894                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2895                         libcfs_nidstr(&lp->lp_primary_nid));
2896         }
2897         spin_unlock(&lp->lp_lock);
2898 }
2899
2900 /*
2901  * Event handler for the discovery EQ.
2902  *
2903  * Called with lnet_res_lock(cpt) held. The cpt is the
2904  * lnet_cpt_of_cookie() of the md handle cookie.
2905  */
2906 static void lnet_discovery_event_handler(struct lnet_event *event)
2907 {
2908         struct lnet_peer *lp = event->md_user_ptr;
2909         struct lnet_ping_buffer *pbuf;
2910         int rc;
2911
2912         /* discovery needs to take another look */
2913         rc = LNET_REDISCOVER_PEER;
2914
2915         CDEBUG(D_NET, "Received event: %d\n", event->type);
2916
2917         switch (event->type) {
2918         case LNET_EVENT_ACK:
2919                 lnet_discovery_event_ack(lp, event);
2920                 break;
2921         case LNET_EVENT_REPLY:
2922                 lnet_discovery_event_reply(lp, event);
2923                 break;
2924         case LNET_EVENT_SEND:
2925                 /* Only send failure triggers a retry. */
2926                 rc = lnet_discovery_event_send(lp, event);
2927                 break;
2928         case LNET_EVENT_UNLINK:
2929                 /* LNetMDUnlink() was called */
2930                 lnet_discovery_event_unlink(lp, event);
2931                 break;
2932         default:
2933                 /* Invalid events. */
2934                 LBUG();
2935         }
2936         lnet_net_lock(LNET_LOCK_EX);
2937         if (event->unlinked) {
2938                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2939                 lnet_ping_buffer_decref(pbuf);
2940                 lnet_peer_decref_locked(lp);
2941         }
2942
2943         /* put peer back at end of request queue, if discovery not already
2944          * done */
2945         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2946             lnet_peer_queue_for_discovery(lp)) {
2947                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2948                 wake_up(&the_lnet.ln_dc_waitq);
2949         }
2950         lnet_net_unlock(LNET_LOCK_EX);
2951 }
2952
2953 /*
2954  * Build a peer from incoming data.
2955  *
2956  * The NIDs in the incoming data are supposed to be structured as follows:
2957  *  - loopback
2958  *  - primary NID
2959  *  - other NIDs in same net
2960  *  - NIDs in second net
2961  *  - NIDs in third net
2962  *  - ...
2963  * This due to the way the list of NIDs in the data is created.
2964  *
2965  * Note that this function will mark the peer uptodate unless an
2966  * ENOMEM is encontered. All other errors are due to a conflict
2967  * between the DLC configuration and what discovery sees. We treat DLC
2968  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2969  * peer from becoming stuck in discovery.
2970  */
2971 static int lnet_peer_merge_data(struct lnet_peer *lp,
2972                                 struct lnet_ping_buffer *pbuf)
2973 {
2974         struct lnet_peer_net *lpn;
2975         struct lnet_peer_ni *lpni;
2976         lnet_nid_t *curnis = NULL;
2977         struct lnet_ni_status *addnis = NULL;
2978         lnet_nid_t *delnis = NULL;
2979         unsigned flags;
2980         int ncurnis;
2981         int naddnis;
2982         int ndelnis;
2983         int nnis = 0;
2984         int i;
2985         int j;
2986         int rc;
2987
2988         flags = LNET_PEER_DISCOVERED;
2989         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2990                 flags |= LNET_PEER_MULTI_RAIL;
2991
2992         /*
2993          * Cache the routing feature for the peer; whether it is enabled
2994          * for disabled as reported by the remote peer.
2995          */
2996         spin_lock(&lp->lp_lock);
2997         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2998                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2999         else
3000                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
3001         spin_unlock(&lp->lp_lock);
3002
3003         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
3004         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
3005         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
3006         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
3007         if (!curnis || !addnis || !delnis) {
3008                 rc = -ENOMEM;
3009                 goto out;
3010         }
3011         ncurnis = 0;
3012         naddnis = 0;
3013         ndelnis = 0;
3014
3015         /* Construct the list of NIDs present in peer. */
3016         lpni = NULL;
3017         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
3018                 curnis[ncurnis++] = lnet_nid_to_nid4(&lpni->lpni_nid);
3019
3020         /*
3021          * Check for NIDs in pbuf not present in curnis[].
3022          * The loop starts at 1 to skip the loopback NID.
3023          */
3024         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
3025                 for (j = 0; j < ncurnis; j++)
3026                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
3027                                 break;
3028                 if (j == ncurnis)
3029                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
3030         }
3031         /*
3032          * Check for NIDs in curnis[] not present in pbuf.
3033          * The nested loop starts at 1 to skip the loopback NID.
3034          *
3035          * But never add the loopback NID to delnis[]: if it is
3036          * present in curnis[] then this peer is for this node.
3037          */
3038         for (i = 0; i < ncurnis; i++) {
3039                 if (curnis[i] == LNET_NID_LO_0)
3040                         continue;
3041                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
3042                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
3043                                 /*
3044                                  * update the information we cache for the
3045                                  * peer with the latest information we
3046                                  * received
3047                                  */
3048                                 lpni = lnet_find_peer_ni_locked(curnis[i]);
3049                                 if (lpni) {
3050                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
3051                                         lnet_peer_ni_decref_locked(lpni);
3052                                 }
3053                                 break;
3054                         }
3055                 }
3056                 if (j == pbuf->pb_info.pi_nnis)
3057                         delnis[ndelnis++] = curnis[i];
3058         }
3059
3060         /*
3061          * If we get here and the discovery is disabled then we don't want
3062          * to add or delete any NIs. We just updated the ones we have some
3063          * information on, and call it a day
3064          */
3065         rc = 0;
3066         if (lnet_is_discovery_disabled(lp))
3067                 goto out;
3068
3069         for (i = 0; i < naddnis; i++) {
3070                 rc = lnet_peer_add_nid(lp, addnis[i].ns_nid, flags);
3071                 if (rc) {
3072                         CERROR("Error adding NID %s to peer %s: %d\n",
3073                                libcfs_nid2str(addnis[i].ns_nid),
3074                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3075                         if (rc == -ENOMEM)
3076                                 goto out;
3077                 }
3078                 lpni = lnet_find_peer_ni_locked(addnis[i].ns_nid);
3079                 if (lpni) {
3080                         lpni->lpni_ns_status = addnis[i].ns_status;
3081                         lnet_peer_ni_decref_locked(lpni);
3082                 }
3083         }
3084
3085         for (i = 0; i < ndelnis; i++) {
3086                 /*
3087                  * for routers it's okay to delete the primary_nid because
3088                  * the upper layers don't really rely on it. So if we're
3089                  * being told that the router changed its primary_nid
3090                  * then it's okay to delete it.
3091                  */
3092                 if (lp->lp_rtr_refcount > 0)
3093                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3094                 rc = lnet_peer_del_nid(lp, delnis[i], flags);
3095                 if (rc) {
3096                         CERROR("Error deleting NID %s from peer %s: %d\n",
3097                                libcfs_nid2str(delnis[i]),
3098                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3099                         if (rc == -ENOMEM)
3100                                 goto out;
3101                 }
3102         }
3103
3104         /* The peer net for the primary NID should be the first entry in the
3105          * peer's lp_peer_nets list, and the peer NI for the primary NID should
3106          * be the first entry in its peer net's lpn_peer_nis list.
3107          */
3108         lpni = lnet_find_peer_ni_locked(pbuf->pb_info.pi_ni[1].ns_nid);
3109         if (!lpni) {
3110                 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3111                        libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
3112                 goto out;
3113         }
3114
3115         lnet_peer_ni_decref_locked(lpni);
3116
3117         lpn = lpni->lpni_peer_net;
3118         if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3119                 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3120
3121         if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3122                 list_move(&lpni->lpni_peer_nis,
3123                           &lpni->lpni_peer_net->lpn_peer_nis);
3124
3125         /*
3126          * Errors other than -ENOMEM are due to peers having been
3127          * configured with DLC. Ignore these because DLC overrides
3128          * Discovery.
3129          */
3130         rc = 0;
3131 out:
3132         CFS_FREE_PTR_ARRAY(curnis, nnis);
3133         CFS_FREE_PTR_ARRAY(addnis, nnis);
3134         CFS_FREE_PTR_ARRAY(delnis, nnis);
3135         lnet_ping_buffer_decref(pbuf);
3136         CDEBUG(D_NET, "peer %s (%p): %d\n",
3137                libcfs_nidstr(&lp->lp_primary_nid), lp, rc);
3138
3139         if (rc) {
3140                 spin_lock(&lp->lp_lock);
3141                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
3142                 lp->lp_state |= LNET_PEER_FORCE_PING;
3143                 spin_unlock(&lp->lp_lock);
3144         }
3145         return rc;
3146 }
3147
3148 /*
3149  * The data in pbuf says lp is its primary peer, but the data was
3150  * received by a different peer. Try to update lp with the data.
3151  */
3152 static int
3153 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
3154 {
3155         struct lnet_handle_md mdh;
3156
3157         /* Queue lp for discovery, and force it on the request queue. */
3158         lnet_net_lock(LNET_LOCK_EX);
3159         if (lnet_peer_queue_for_discovery(lp))
3160                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3161         lnet_net_unlock(LNET_LOCK_EX);
3162
3163         LNetInvalidateMDHandle(&mdh);
3164
3165         /*
3166          * Decide whether we can move the peer to the DATA_PRESENT state.
3167          *
3168          * We replace stale data for a multi-rail peer, repair PING_FAILED
3169          * status, and preempt FORCE_PING.
3170          *
3171          * If after that we have DATA_PRESENT, we merge it into this peer.
3172          */
3173         spin_lock(&lp->lp_lock);
3174         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3175                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3176                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3177                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3178                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3179                         lnet_ping_buffer_decref(pbuf);
3180                         pbuf = lp->lp_data;
3181                         lp->lp_data = NULL;
3182                 }
3183         }
3184         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3185                 lnet_ping_buffer_decref(lp->lp_data);
3186                 lp->lp_data = NULL;
3187                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3188         }
3189         if (lp->lp_state & LNET_PEER_PING_FAILED) {
3190                 mdh = lp->lp_ping_mdh;
3191                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3192                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3193                 lp->lp_ping_error = 0;
3194         }
3195         if (lp->lp_state & LNET_PEER_FORCE_PING)
3196                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3197         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3198         spin_unlock(&lp->lp_lock);
3199
3200         if (!LNetMDHandleIsInvalid(mdh))
3201                 LNetMDUnlink(mdh);
3202
3203         if (pbuf)
3204                 return lnet_peer_merge_data(lp, pbuf);
3205
3206         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3207         return 0;
3208 }
3209
3210 static bool lnet_is_nid_in_ping_info(lnet_nid_t nid, struct lnet_ping_info *pinfo)
3211 {
3212         int i;
3213
3214         for (i = 0; i < pinfo->pi_nnis; i++) {
3215                 if (pinfo->pi_ni[i].ns_nid == nid)
3216                         return true;
3217         }
3218
3219         return false;
3220 }
3221
3222 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3223  * to the discovery queue a reference was taken that will prevent the peer from
3224  * actually being freed by this function. After this function exits the
3225  * discovery thread should call lnet_peer_discovery_complete() which will
3226  * drop that reference as well as wake any waiters that may also be holding a
3227  * ref on the peer
3228  */
3229 static int lnet_peer_deletion(struct lnet_peer *lp)
3230 __must_hold(&lp->lp_lock)
3231 {
3232         struct list_head rlist;
3233         struct lnet_route *route, *tmp;
3234         int sensitivity = lp->lp_health_sensitivity;
3235         int rc;
3236
3237         INIT_LIST_HEAD(&rlist);
3238
3239         lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3240                           LNET_PEER_FORCE_PUSH);
3241         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3242                libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3243
3244         /* no-op if lnet_peer_del() has already been called on this peer */
3245         if (lp->lp_state & LNET_PEER_MARK_DELETED)
3246                 return 0;
3247
3248         if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
3249                 return -ESHUTDOWN;
3250
3251         spin_unlock(&lp->lp_lock);
3252
3253         mutex_lock(&the_lnet.ln_api_mutex);
3254
3255         lnet_net_lock(LNET_LOCK_EX);
3256         /* remove the peer from the discovery work
3257          * queue if it's on there in preparation
3258          * of deleting it.
3259          */
3260         if (!list_empty(&lp->lp_dc_list))
3261                 list_del_init(&lp->lp_dc_list);
3262         list_for_each_entry_safe(route, tmp,
3263                                  &lp->lp_routes,
3264                                  lr_gwlist)
3265                 lnet_move_route(route, NULL, &rlist);
3266         lnet_net_unlock(LNET_LOCK_EX);
3267
3268         /* lnet_peer_del() deletes all the peer NIs owned by this peer */
3269         rc = lnet_peer_del(lp);
3270         if (rc)
3271                 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3272                         libcfs_nidstr(&lp->lp_primary_nid), rc);
3273
3274         list_for_each_entry_safe(route, tmp,
3275                                  &rlist, lr_list) {
3276                 /* re-add these routes */
3277                 lnet_add_route(route->lr_net,
3278                                route->lr_hops,
3279                                &route->lr_nid,
3280                                route->lr_priority,
3281                                sensitivity);
3282                 LIBCFS_FREE(route, sizeof(*route));
3283         }
3284
3285         mutex_unlock(&the_lnet.ln_api_mutex);
3286
3287         spin_lock(&lp->lp_lock);
3288
3289         return 0;
3290 }
3291
3292 /*
3293  * Update a peer using the data received.
3294  */
3295 static int lnet_peer_data_present(struct lnet_peer *lp)
3296 __must_hold(&lp->lp_lock)
3297 {
3298         struct lnet_ping_buffer *pbuf;
3299         struct lnet_peer_ni *lpni;
3300         lnet_nid_t nid = LNET_NID_ANY;
3301         unsigned flags;
3302         int rc = 0;
3303
3304         pbuf = lp->lp_data;
3305         lp->lp_data = NULL;
3306         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3307         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3308         spin_unlock(&lp->lp_lock);
3309
3310         /*
3311          * Modifications of peer structures are done while holding the
3312          * ln_api_mutex. A global lock is required because we may be
3313          * modifying multiple peer structures, and a mutex greatly
3314          * simplifies memory management.
3315          *
3316          * The actual changes to the data structures must also protect
3317          * against concurrent lookups, for which the lnet_net_lock in
3318          * LNET_LOCK_EX mode is used.
3319          */
3320         mutex_lock(&the_lnet.ln_api_mutex);
3321         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3322                 rc = -ESHUTDOWN;
3323                 goto out;
3324         }
3325
3326         /*
3327          * If this peer is not on the peer list then it is being torn
3328          * down, and our reference count may be all that is keeping it
3329          * alive. Don't do any work on it.
3330          */
3331         if (list_empty(&lp->lp_peer_list))
3332                 goto out;
3333
3334         flags = LNET_PEER_DISCOVERED;
3335         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3336                 flags |= LNET_PEER_MULTI_RAIL;
3337
3338         /*
3339          * Check whether the primary NID in the message matches the
3340          * primary NID of the peer. If it does, update the peer, if
3341          * it it does not, check whether there is already a peer with
3342          * that primary NID. If no such peer exists, try to update
3343          * the primary NID of the current peer (allowed if it was
3344          * created due to message traffic) and complete the update.
3345          * If the peer did exist, hand off the data to it.
3346          *
3347          * The peer for the loopback interface is a special case: this
3348          * is the peer for the local node, and we want to set its
3349          * primary NID to the correct value here. Moreover, this peer
3350          * can show up with only the loopback NID in the ping buffer.
3351          */
3352         if (pbuf->pb_info.pi_nnis <= 1)
3353                 goto out;
3354         nid = pbuf->pb_info.pi_ni[1].ns_nid;
3355         if (nid_is_lo0(&lp->lp_primary_nid)) {
3356                 rc = lnet_peer_set_primary_nid(lp, nid, flags);
3357                 if (!rc)
3358                         rc = lnet_peer_merge_data(lp, pbuf);
3359         /*
3360          * if the primary nid of the peer is present in the ping info returned
3361          * from the peer, but it's not the local primary peer we have
3362          * cached and discovery is disabled, then we don't want to update
3363          * our local peer info, by adding or removing NIDs, we just want
3364          * to update the status of the nids that we currently have
3365          * recorded in that peer.
3366          */
3367         } else if (lnet_nid_to_nid4(&lp->lp_primary_nid) == nid ||
3368                    (lnet_is_nid_in_ping_info(lnet_nid_to_nid4(&lp->lp_primary_nid),
3369                                              &pbuf->pb_info) &&
3370                     lnet_is_discovery_disabled(lp))) {
3371                 rc = lnet_peer_merge_data(lp, pbuf);
3372         } else {
3373                 lpni = lnet_find_peer_ni_locked(nid);
3374                 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3375                         rc = lnet_peer_set_primary_nid(lp, nid, flags);
3376                         if (rc) {
3377                                 CERROR("Primary NID error %s versus %s: %d\n",
3378                                        libcfs_nidstr(&lp->lp_primary_nid),
3379                                        libcfs_nid2str(nid), rc);
3380                         } else {
3381                                 rc = lnet_peer_merge_data(lp, pbuf);
3382                         }
3383                         if (lpni)
3384                                 lnet_peer_ni_decref_locked(lpni);
3385                 } else {
3386                         struct lnet_peer *new_lp;
3387                         new_lp = lpni->lpni_peer_net->lpn_peer;
3388                         /*
3389                          * if lp has discovery/MR enabled that means new_lp
3390                          * should have discovery/MR enabled as well, since
3391                          * it's the same peer, which we're about to merge
3392                          */
3393                         spin_lock(&lp->lp_lock);
3394                         spin_lock(&new_lp->lp_lock);
3395                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3396                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3397                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3398                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3399                         /* If we're processing a ping reply then we may be
3400                          * about to send a push to the peer that we ping'd.
3401                          * Since the ping reply that we're processing was
3402                          * received by lp, we need to set the discovery source
3403                          * NID for new_lp to the NID stored in lp.
3404                          */
3405                         if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3406                                 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3407                                 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3408                         }
3409                         spin_unlock(&new_lp->lp_lock);
3410                         spin_unlock(&lp->lp_lock);
3411
3412                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
3413                         lnet_consolidate_routes_locked(lp, new_lp);
3414                         lnet_peer_ni_decref_locked(lpni);
3415                 }
3416         }
3417 out:
3418         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3419                libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3420                lp->lp_state);
3421         mutex_unlock(&the_lnet.ln_api_mutex);
3422
3423         spin_lock(&lp->lp_lock);
3424         /* Tell discovery to re-check the peer immediately. */
3425         if (!rc)
3426                 rc = LNET_REDISCOVER_PEER;
3427         return rc;
3428 }
3429
3430 /*
3431  * A ping failed. Clear the PING_FAILED state and set the
3432  * FORCE_PING state, to ensure a retry even if discovery is
3433  * disabled. This avoids being left with incorrect state.
3434  */
3435 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3436 __must_hold(&lp->lp_lock)
3437 {
3438         struct lnet_handle_md mdh;
3439         int rc;
3440
3441         mdh = lp->lp_ping_mdh;
3442         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3443         lp->lp_state &= ~LNET_PEER_PING_FAILED;
3444         lp->lp_state |= LNET_PEER_FORCE_PING;
3445         rc = lp->lp_ping_error;
3446         lp->lp_ping_error = 0;
3447         spin_unlock(&lp->lp_lock);
3448
3449         if (!LNetMDHandleIsInvalid(mdh))
3450                 LNetMDUnlink(mdh);
3451
3452         CDEBUG(D_NET, "peer %s:%d\n",
3453                libcfs_nidstr(&lp->lp_primary_nid), rc);
3454
3455         spin_lock(&lp->lp_lock);
3456         return rc ? rc : LNET_REDISCOVER_PEER;
3457 }
3458
3459 /* Active side of ping. */
3460 static int lnet_peer_send_ping(struct lnet_peer *lp)
3461 __must_hold(&lp->lp_lock)
3462 {
3463         int nnis;
3464         int rc;
3465         int cpt;
3466
3467         lp->lp_state |= LNET_PEER_PING_SENT;
3468         lp->lp_state &= ~LNET_PEER_FORCE_PING;
3469         spin_unlock(&lp->lp_lock);
3470
3471         cpt = lnet_net_lock_current();
3472         /* Refcount for MD. */
3473         lnet_peer_addref_locked(lp);
3474         lnet_net_unlock(cpt);
3475
3476         nnis = max(lp->lp_data_nnis, LNET_INTERFACES_MIN);
3477
3478         rc = lnet_send_ping(lnet_nid_to_nid4(&lp->lp_primary_nid),
3479                             &lp->lp_ping_mdh, nnis, lp,
3480                             the_lnet.ln_dc_handler, false);
3481
3482         /*
3483          * if LNetMDBind in lnet_send_ping fails we need to decrement the
3484          * refcount on the peer, otherwise LNetMDUnlink will be called
3485          * which will eventually do that.
3486          */
3487         if (rc > 0) {
3488                 lnet_net_lock(cpt);
3489                 lnet_peer_decref_locked(lp);
3490                 lnet_net_unlock(cpt);
3491                 rc = -rc; /* change the rc to negative value */
3492                 goto fail_error;
3493         } else if (rc < 0) {
3494                 goto fail_error;
3495         }
3496
3497         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3498
3499         spin_lock(&lp->lp_lock);
3500         return 0;
3501
3502 fail_error:
3503         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3504         /*
3505          * The errors that get us here are considered hard errors and
3506          * cause Discovery to terminate. So we clear PING_SENT, but do
3507          * not set either PING_FAILED or FORCE_PING. In fact we need
3508          * to clear PING_FAILED, because the unlink event handler will
3509          * have set it if we called LNetMDUnlink() above.
3510          */
3511         spin_lock(&lp->lp_lock);
3512         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3513         return rc;
3514 }
3515
3516 /*
3517  * This function exists because you cannot call LNetMDUnlink() from an
3518  * event handler.
3519  */
3520 static int lnet_peer_push_failed(struct lnet_peer *lp)
3521 __must_hold(&lp->lp_lock)
3522 {
3523         struct lnet_handle_md mdh;
3524         int rc;
3525
3526         mdh = lp->lp_push_mdh;
3527         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3528         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3529         rc = lp->lp_push_error;
3530         lp->lp_push_error = 0;
3531         spin_unlock(&lp->lp_lock);
3532
3533         if (!LNetMDHandleIsInvalid(mdh))
3534                 LNetMDUnlink(mdh);
3535
3536         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3537         spin_lock(&lp->lp_lock);
3538         return rc ? rc : LNET_REDISCOVER_PEER;
3539 }
3540
3541 /*
3542  * Mark the peer as discovered.
3543  */
3544 static int lnet_peer_discovered(struct lnet_peer *lp)
3545 __must_hold(&lp->lp_lock)
3546 {
3547         lp->lp_state |= LNET_PEER_DISCOVERED;
3548         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3549                           LNET_PEER_REDISCOVER);
3550
3551         lp->lp_dc_error = 0;
3552
3553         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3554
3555         return 0;
3556 }
3557
3558 /* Active side of push. */
3559 static int lnet_peer_send_push(struct lnet_peer *lp)
3560 __must_hold(&lp->lp_lock)
3561 {
3562         struct lnet_ping_buffer *pbuf;
3563         struct lnet_process_id id;
3564         struct lnet_md md;
3565         int cpt;
3566         int rc;
3567
3568         /* Don't push to a non-multi-rail peer. */
3569         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3570                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3571                 /* if peer's NIDs are uptodate then peer is discovered */
3572                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3573                         rc = lnet_peer_discovered(lp);
3574                         return rc;
3575                 }
3576
3577                 return 0;
3578         }
3579
3580         lp->lp_state |= LNET_PEER_PUSH_SENT;
3581         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3582         spin_unlock(&lp->lp_lock);
3583
3584         cpt = lnet_net_lock_current();
3585         pbuf = the_lnet.ln_ping_target;
3586         lnet_ping_buffer_addref(pbuf);
3587         lnet_net_unlock(cpt);
3588
3589         /* Push source MD */
3590         md.start     = &pbuf->pb_info;
3591         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
3592         md.threshold = 2; /* Put/Ack */
3593         md.max_size  = 0;
3594         md.options   = LNET_MD_TRACK_RESPONSE;
3595         md.handler   = the_lnet.ln_dc_handler;
3596         md.user_ptr  = lp;
3597
3598         rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3599         if (rc) {
3600                 lnet_ping_buffer_decref(pbuf);
3601                 CERROR("Can't bind push source MD: %d\n", rc);
3602                 goto fail_error;
3603         }
3604
3605         cpt = lnet_net_lock_current();
3606         /* Refcount for MD. */
3607         lnet_peer_addref_locked(lp);
3608         id.pid = LNET_PID_LUSTRE;
3609         if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3610                 id.nid = lnet_nid_to_nid4(&lp->lp_disc_dst_nid);
3611         else
3612                 id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
3613         lnet_net_unlock(cpt);
3614
3615         rc = LNetPut(lnet_nid_to_nid4(&lp->lp_disc_src_nid), lp->lp_push_mdh,
3616                      LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
3617                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3618
3619         /*
3620          * reset the discovery nid. There is no need to restrict sending
3621          * from that source, if we call lnet_push_update_to_peers(). It'll
3622          * get set to a specific NID, if we initiate discovery from the
3623          * scratch
3624          */
3625         lp->lp_disc_src_nid = LNET_ANY_NID;
3626         lp->lp_disc_dst_nid = LNET_ANY_NID;
3627
3628         if (rc)
3629                 goto fail_unlink;
3630
3631         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3632
3633         spin_lock(&lp->lp_lock);
3634         return 0;
3635
3636 fail_unlink:
3637         LNetMDUnlink(lp->lp_push_mdh);
3638         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3639 fail_error:
3640         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3641                lp, rc);
3642         /*
3643          * The errors that get us here are considered hard errors and
3644          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3645          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3646          * because the unlink event handler will have set it if we
3647          * called LNetMDUnlink() above.
3648          */
3649         spin_lock(&lp->lp_lock);
3650         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3651         return rc;
3652 }
3653
3654 /*
3655  * Wait for work to be queued or some other change that must be
3656  * attended to. Returns non-zero if the discovery thread should shut
3657  * down.
3658  */
3659 static int lnet_peer_discovery_wait_for_work(void)
3660 {
3661         int cpt;
3662         int rc = 0;
3663
3664         DEFINE_WAIT(wait);
3665
3666         cpt = lnet_net_lock_current();
3667         for (;;) {
3668                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3669                                 TASK_INTERRUPTIBLE);
3670                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3671                         break;
3672                 if (lnet_push_target_resize_needed() ||
3673                     the_lnet.ln_push_target->pb_needs_post)
3674                         break;
3675                 if (!list_empty(&the_lnet.ln_dc_request))
3676                         break;
3677                 if (!list_empty(&the_lnet.ln_msg_resend))
3678                         break;
3679                 lnet_net_unlock(cpt);
3680
3681                 /*
3682                  * wakeup max every second to check if there are peers that
3683                  * have been stuck on the working queue for greater than
3684                  * the peer timeout.
3685                  */
3686                 schedule_timeout(cfs_time_seconds(1));
3687                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3688                 cpt = lnet_net_lock_current();
3689         }
3690         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3691
3692         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3693                 rc = -ESHUTDOWN;
3694
3695         lnet_net_unlock(cpt);
3696
3697         CDEBUG(D_NET, "woken: %d\n", rc);
3698
3699         return rc;
3700 }
3701
3702 /*
3703  * Messages that were pending on a destroyed peer will be put on a global
3704  * resend list. The message resend list will be checked by
3705  * the discovery thread when it wakes up, and will resend messages. These
3706  * messages can still be sendable in the case the lpni which was the initial
3707  * cause of the message re-queue was transfered to another peer.
3708  *
3709  * It is possible that LNet could be shutdown while we're iterating
3710  * through the list. lnet_shudown_lndnets() will attempt to access the
3711  * resend list, but will have to wait until the spinlock is released, by
3712  * which time there shouldn't be any more messages on the resend list.
3713  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3714  * for the messages so they can be released. The other case is that
3715  * lnet_shudown_lndnets() can finalize all the messages before this
3716  * function can visit the resend list, in which case this function will be
3717  * a no-op.
3718  */
3719 static void lnet_resend_msgs(void)
3720 {
3721         struct lnet_msg *msg, *tmp;
3722         LIST_HEAD(resend);
3723         int rc;
3724
3725         spin_lock(&the_lnet.ln_msg_resend_lock);
3726         list_splice(&the_lnet.ln_msg_resend, &resend);
3727         spin_unlock(&the_lnet.ln_msg_resend_lock);
3728
3729         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3730                 list_del_init(&msg->msg_list);
3731                 rc = lnet_send(msg->msg_src_nid_param, msg,
3732                                msg->msg_rtr_nid_param);
3733                 if (rc < 0) {
3734                         CNETERR("Error sending %s to %s: %d\n",
3735                                lnet_msgtyp2str(msg->msg_type),
3736                                libcfs_id2str(msg->msg_target), rc);
3737                         lnet_finalize(msg, rc);
3738                 }
3739         }
3740 }
3741
3742 /* The discovery thread. */
3743 static int lnet_peer_discovery(void *arg)
3744 {
3745         struct lnet_peer *lp;
3746         int rc;
3747
3748         wait_for_completion(&the_lnet.ln_started);
3749
3750         CDEBUG(D_NET, "started\n");
3751
3752         for (;;) {
3753                 if (lnet_peer_discovery_wait_for_work())
3754                         break;
3755
3756                 if (lnet_push_target_resize_needed())
3757                         lnet_push_target_resize();
3758                 else if (the_lnet.ln_push_target->pb_needs_post)
3759                         lnet_push_target_post(the_lnet.ln_push_target,
3760                                               &the_lnet.ln_push_target_md);
3761
3762                 lnet_resend_msgs();
3763
3764                 lnet_net_lock(LNET_LOCK_EX);
3765                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3766                         lnet_net_unlock(LNET_LOCK_EX);
3767                         break;
3768                 }
3769
3770                 /*
3771                  * Process all incoming discovery work requests.  When
3772                  * discovery must wait on a peer to change state, it
3773                  * is added to the tail of the ln_dc_working queue. A
3774                  * timestamp keeps track of when the peer was added,
3775                  * so we can time out discovery requests that take too
3776                  * long.
3777                  */
3778                 while (!list_empty(&the_lnet.ln_dc_request)) {
3779                         lp = list_first_entry(&the_lnet.ln_dc_request,
3780                                               struct lnet_peer, lp_dc_list);
3781                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3782                         /*
3783                          * set the time the peer was put on the dc_working
3784                          * queue. It shouldn't remain on the queue
3785                          * forever, in case the GET message (for ping)
3786                          * doesn't get a REPLY or the PUT message (for
3787                          * push) doesn't get an ACK.
3788                          */
3789                         lp->lp_last_queued = ktime_get_real_seconds();
3790                         lnet_net_unlock(LNET_LOCK_EX);
3791
3792                         if (lnet_push_target_resize_needed())
3793                                 lnet_push_target_resize();
3794                         else if (the_lnet.ln_push_target->pb_needs_post)
3795                                 lnet_push_target_post(the_lnet.ln_push_target,
3796                                                       &the_lnet.ln_push_target_md);
3797
3798                         /*
3799                          * Select an action depending on the state of
3800                          * the peer and whether discovery is disabled.
3801                          * The check whether discovery is disabled is
3802                          * done after the code that handles processing
3803                          * for arrived data, cleanup for failures, and
3804                          * forcing a Ping or Push.
3805                          */
3806                         spin_lock(&lp->lp_lock);
3807                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3808                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3809                                 lp->lp_state);
3810                         if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3811                                             LNET_PEER_MARK_DELETED))
3812                                 rc = lnet_peer_deletion(lp);
3813                         else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3814                                 rc = lnet_peer_data_present(lp);
3815                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3816                                 rc = lnet_peer_ping_failed(lp);
3817                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3818                                 rc = lnet_peer_push_failed(lp);
3819                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3820                                 rc = lnet_peer_send_ping(lp);
3821                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3822                                 rc = lnet_peer_send_push(lp);
3823                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3824                                 rc = lnet_peer_send_ping(lp);
3825                         else if (lnet_peer_needs_push(lp))
3826                                 rc = lnet_peer_send_push(lp);
3827                         else
3828                                 rc = lnet_peer_discovered(lp);
3829                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3830                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3831                                 lp->lp_state, rc);
3832
3833                         if (rc == LNET_REDISCOVER_PEER) {
3834                                 spin_unlock(&lp->lp_lock);
3835                                 lnet_net_lock(LNET_LOCK_EX);
3836                                 list_move(&lp->lp_dc_list,
3837                                           &the_lnet.ln_dc_request);
3838                         } else if (rc ||
3839                                    !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3840                                 spin_unlock(&lp->lp_lock);
3841                                 lnet_net_lock(LNET_LOCK_EX);
3842                                 lnet_peer_discovery_complete(lp, rc);
3843                         } else {
3844                                 spin_unlock(&lp->lp_lock);
3845                                 lnet_net_lock(LNET_LOCK_EX);
3846                         }
3847
3848                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3849                                 break;
3850
3851                 }
3852
3853                 lnet_net_unlock(LNET_LOCK_EX);
3854         }
3855
3856         CDEBUG(D_NET, "stopping\n");
3857         /*
3858          * Clean up before telling lnet_peer_discovery_stop() that
3859          * we're done. Use wake_up() below to somewhat reduce the
3860          * size of the thundering herd if there are multiple threads
3861          * waiting on discovery of a single peer.
3862          */
3863
3864         /* Queue cleanup 1: stop all pending pings and pushes. */
3865         lnet_net_lock(LNET_LOCK_EX);
3866         while (!list_empty(&the_lnet.ln_dc_working)) {
3867                 lp = list_first_entry(&the_lnet.ln_dc_working,
3868                                       struct lnet_peer, lp_dc_list);
3869                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3870                 lnet_net_unlock(LNET_LOCK_EX);
3871                 lnet_peer_cancel_discovery(lp);
3872                 lnet_net_lock(LNET_LOCK_EX);
3873         }
3874         lnet_net_unlock(LNET_LOCK_EX);
3875
3876         /* Queue cleanup 2: wait for the expired queue to clear. */
3877         while (!list_empty(&the_lnet.ln_dc_expired))
3878                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3879
3880         /* Queue cleanup 3: clear the request queue. */
3881         lnet_net_lock(LNET_LOCK_EX);
3882         while (!list_empty(&the_lnet.ln_dc_request)) {
3883                 lp = list_first_entry(&the_lnet.ln_dc_request,
3884                                       struct lnet_peer, lp_dc_list);
3885                 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3886         }
3887         lnet_net_unlock(LNET_LOCK_EX);
3888
3889         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3890         the_lnet.ln_dc_handler = NULL;
3891
3892         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3893         wake_up(&the_lnet.ln_dc_waitq);
3894
3895         CDEBUG(D_NET, "stopped\n");
3896
3897         return 0;
3898 }
3899
3900 /* ln_api_mutex is held on entry. */
3901 int lnet_peer_discovery_start(void)
3902 {
3903         struct task_struct *task;
3904         int rc = 0;
3905
3906         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3907                 return -EALREADY;
3908
3909         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3910         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3911         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3912         if (IS_ERR(task)) {
3913                 rc = PTR_ERR(task);
3914                 CERROR("Can't start peer discovery thread: %d\n", rc);
3915
3916                 the_lnet.ln_dc_handler = NULL;
3917
3918                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3919         }
3920
3921         CDEBUG(D_NET, "discovery start: %d\n", rc);
3922
3923         return rc;
3924 }
3925
3926 /* ln_api_mutex is held on entry. */
3927 void lnet_peer_discovery_stop(void)
3928 {
3929         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3930                 return;
3931
3932         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3933         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3934
3935         /* In the LNetNIInit() path we may be stopping discovery before it
3936          * entered its work loop
3937          */
3938         if (!completion_done(&the_lnet.ln_started))
3939                 complete(&the_lnet.ln_started);
3940         else
3941                 wake_up(&the_lnet.ln_dc_waitq);
3942
3943         wait_event(the_lnet.ln_dc_waitq,
3944                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3945
3946         LASSERT(list_empty(&the_lnet.ln_dc_request));
3947         LASSERT(list_empty(&the_lnet.ln_dc_working));
3948         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3949
3950         CDEBUG(D_NET, "discovery stopped\n");
3951 }
3952
3953 /* Debugging */
3954
3955 void
3956 lnet_debug_peer(lnet_nid_t nid)
3957 {
3958         char                    *aliveness = "NA";
3959         struct lnet_peer_ni     *lp;
3960         int                     cpt;
3961
3962         cpt = lnet_cpt_of_nid(nid, NULL);
3963         lnet_net_lock(cpt);
3964
3965         lp = lnet_nid2peerni_locked(nid, LNET_NID_ANY, cpt);
3966         if (IS_ERR(lp)) {
3967                 lnet_net_unlock(cpt);
3968                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
3969                 return;
3970         }
3971
3972         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3973                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3974
3975         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3976                libcfs_nidstr(&lp->lpni_nid), kref_read(&lp->lpni_kref),
3977                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3978                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3979                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3980
3981         lnet_peer_ni_decref_locked(lp);
3982
3983         lnet_net_unlock(cpt);
3984 }
3985
3986 /* Gathering information for userspace. */
3987
3988 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
3989                           char aliveness[LNET_MAX_STR_LEN],
3990                           __u32 *cpt_iter, __u32 *refcount,
3991                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
3992                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
3993                           __u32 *peer_tx_qnob)
3994 {
3995         struct lnet_peer_table          *peer_table;
3996         struct lnet_peer_ni             *lp;
3997         int                             j;
3998         int                             lncpt;
3999         bool                            found = false;
4000
4001         /* get the number of CPTs */
4002         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4003
4004         /* if the cpt number to be examined is >= the number of cpts in
4005          * the system then indicate that there are no more cpts to examin
4006          */
4007         if (*cpt_iter >= lncpt)
4008                 return -ENOENT;
4009
4010         /* get the current table */
4011         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
4012         /* if the ptable is NULL then there are no more cpts to examine */
4013         if (peer_table == NULL)
4014                 return -ENOENT;
4015
4016         lnet_net_lock(*cpt_iter);
4017
4018         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
4019                 struct list_head *peers = &peer_table->pt_hash[j];
4020
4021                 list_for_each_entry(lp, peers, lpni_hashlist) {
4022                         if (!nid_is_nid4(&lp->lpni_nid))
4023                                 continue;
4024                         if (peer_index-- > 0)
4025                                 continue;
4026
4027                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
4028                         if (lnet_isrouter(lp) ||
4029                                 lnet_peer_aliveness_enabled(lp))
4030                                 snprintf(aliveness, LNET_MAX_STR_LEN,
4031                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
4032
4033                         *nid = lnet_nid_to_nid4(&lp->lpni_nid);
4034                         *refcount = kref_read(&lp->lpni_kref);
4035                         *ni_peer_tx_credits =
4036                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
4037                         *peer_tx_credits = lp->lpni_txcredits;
4038                         *peer_rtr_credits = lp->lpni_rtrcredits;
4039                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
4040                         *peer_tx_qnob = lp->lpni_txqnob;
4041
4042                         found = true;
4043                 }
4044
4045         }
4046         lnet_net_unlock(*cpt_iter);
4047
4048         *cpt_iter = lncpt;
4049
4050         return found ? 0 : -ENOENT;
4051 }
4052
4053 /* ln_api_mutex is held, which keeps the peer list stable */
4054 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
4055 {
4056         struct lnet_ioctl_element_stats *lpni_stats;
4057         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
4058         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
4059         struct lnet_peer_ni_credit_info *lpni_info;
4060         struct lnet_peer_ni *lpni;
4061         struct lnet_peer *lp;
4062         lnet_nid_t nid;
4063         __u32 size;
4064         int rc;
4065
4066         lp = lnet_find_peer(cfg->prcfg_prim_nid);
4067
4068         if (!lp) {
4069                 rc = -ENOENT;
4070                 goto out;
4071         }
4072
4073         size = sizeof(nid) + sizeof(*lpni_info) + sizeof(*lpni_stats)
4074                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
4075         size *= lp->lp_nnis;
4076         if (size > cfg->prcfg_size) {
4077                 cfg->prcfg_size = size;
4078                 rc = -E2BIG;
4079                 goto out_lp_decref;
4080         }
4081
4082         cfg->prcfg_prim_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4083         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
4084         cfg->prcfg_cfg_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4085         cfg->prcfg_count = lp->lp_nnis;
4086         cfg->prcfg_size = size;
4087         cfg->prcfg_state = lp->lp_state;
4088
4089         /* Allocate helper buffers. */
4090         rc = -ENOMEM;
4091         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
4092         if (!lpni_info)
4093                 goto out_lp_decref;
4094         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
4095         if (!lpni_stats)
4096                 goto out_free_info;
4097         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
4098         if (!lpni_msg_stats)
4099                 goto out_free_stats;
4100         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
4101         if (!lpni_hstats)
4102                 goto out_free_msg_stats;
4103
4104
4105         lpni = NULL;
4106         rc = -EFAULT;
4107         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
4108                 if (!nid_is_nid4(&lpni->lpni_nid))
4109                         continue;
4110                 nid = lnet_nid_to_nid4(&lpni->lpni_nid);
4111                 if (copy_to_user(bulk, &nid, sizeof(nid)))
4112                         goto out_free_hstats;
4113                 bulk += sizeof(nid);
4114
4115                 memset(lpni_info, 0, sizeof(*lpni_info));
4116                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
4117                 if (lnet_isrouter(lpni) ||
4118                         lnet_peer_aliveness_enabled(lpni))
4119                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
4120                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
4121
4122                 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
4123                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
4124                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
4125                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
4126                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
4127                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
4128                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
4129                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4130                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4131                         goto out_free_hstats;
4132                 bulk += sizeof(*lpni_info);
4133
4134                 memset(lpni_stats, 0, sizeof(*lpni_stats));
4135                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4136                                                             LNET_STATS_TYPE_SEND);
4137                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4138                                                             LNET_STATS_TYPE_RECV);
4139                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4140                                                             LNET_STATS_TYPE_DROP);
4141                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4142                         goto out_free_hstats;
4143                 bulk += sizeof(*lpni_stats);
4144                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4145                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4146                         goto out_free_hstats;
4147                 bulk += sizeof(*lpni_msg_stats);
4148                 lpni_hstats->hlpni_network_timeout =
4149                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4150                 lpni_hstats->hlpni_remote_dropped =
4151                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4152                 lpni_hstats->hlpni_remote_timeout =
4153                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4154                 lpni_hstats->hlpni_remote_error =
4155                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4156                 lpni_hstats->hlpni_health_value =
4157                   atomic_read(&lpni->lpni_healthv);
4158                 lpni_hstats->hlpni_ping_count = lpni->lpni_ping_count;
4159                 lpni_hstats->hlpni_next_ping = lpni->lpni_next_ping;
4160                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4161                         goto out_free_hstats;
4162                 bulk += sizeof(*lpni_hstats);
4163         }
4164         rc = 0;
4165
4166 out_free_hstats:
4167         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4168 out_free_msg_stats:
4169         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4170 out_free_stats:
4171         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4172 out_free_info:
4173         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4174 out_lp_decref:
4175         lnet_peer_decref_locked(lp);
4176 out:
4177         return rc;
4178 }
4179
4180 /* must hold net_lock/0 */
4181 void
4182 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4183                                      struct list_head *recovery_queue,
4184                                      time64_t now)
4185 {
4186         /* the mt could've shutdown and cleaned up the queues */
4187         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4188                 return;
4189
4190         if (!list_empty(&lpni->lpni_recovery))
4191                 return;
4192
4193         if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4194                 return;
4195
4196         if (!lpni->lpni_last_alive) {
4197                 CDEBUG(D_NET,
4198                        "lpni %s(%p) not eligible for recovery last alive %lld\n",
4199                        libcfs_nidstr(&lpni->lpni_nid), lpni,
4200                        lpni->lpni_last_alive);
4201                 return;
4202         }
4203
4204         if (lnet_recovery_limit &&
4205             now > lpni->lpni_last_alive + lnet_recovery_limit) {
4206                 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4207                        libcfs_nidstr(&lpni->lpni_nid),
4208                        lpni->lpni_last_alive);
4209                 /* Reset the ping count so that if this peer NI is added back to
4210                  * the recovery queue we will send the first ping right away.
4211                  */
4212                 lpni->lpni_ping_count = 0;
4213                 return;
4214         }
4215
4216         /* This peer NI is going on the recovery queue, so take a ref on it */
4217         lnet_peer_ni_addref_locked(lpni);
4218
4219         lnet_peer_ni_set_next_ping(lpni, now);
4220
4221         CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4222                libcfs_nidstr(&lpni->lpni_nid),
4223                lpni->lpni_ping_count,
4224                lpni->lpni_next_ping,
4225                lpni->lpni_last_alive,
4226                atomic_read(&lpni->lpni_healthv));
4227
4228         list_add_tail(&lpni->lpni_recovery, recovery_queue);
4229 }
4230
4231 /* Call with the ln_api_mutex held */
4232 void
4233 lnet_peer_ni_set_healthv(lnet_nid_t nid, int value, bool all)
4234 {
4235         struct lnet_peer_table *ptable;
4236         struct lnet_peer *lp;
4237         struct lnet_peer_net *lpn;
4238         struct lnet_peer_ni *lpni;
4239         int lncpt;
4240         int cpt;
4241         time64_t now;
4242
4243         if (the_lnet.ln_state != LNET_STATE_RUNNING)
4244                 return;
4245
4246         now = ktime_get_seconds();
4247
4248         if (!all) {
4249                 lnet_net_lock(LNET_LOCK_EX);
4250                 lpni = lnet_find_peer_ni_locked(nid);
4251                 if (!lpni) {
4252                         lnet_net_unlock(LNET_LOCK_EX);
4253                         return;
4254                 }
4255                 lnet_set_lpni_healthv_locked(lpni, value);
4256                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4257                                              &the_lnet.ln_mt_peerNIRecovq, now);
4258                 lnet_peer_ni_decref_locked(lpni);
4259                 lnet_net_unlock(LNET_LOCK_EX);
4260                 return;
4261         }
4262
4263         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4264
4265         /*
4266          * Walk all the peers and reset the health value for each one to the
4267          * specified value.
4268          */
4269         lnet_net_lock(LNET_LOCK_EX);
4270         for (cpt = 0; cpt < lncpt; cpt++) {
4271                 ptable = the_lnet.ln_peer_tables[cpt];
4272                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4273                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4274                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4275                                                     lpni_peer_nis) {
4276                                         lnet_set_lpni_healthv_locked(lpni,
4277                                                                      value);
4278                                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
4279                                              &the_lnet.ln_mt_peerNIRecovq, now);
4280                                 }
4281                         }
4282                 }
4283         }
4284         lnet_net_unlock(LNET_LOCK_EX);
4285 }
4286