lnet/lnet/peer.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lnet/lnet/peer.c
  32  */
  33
  34 #define DEBUG_SUBSYSTEM S_LNET
  35
  36 #include <linux/sched.h>
  37 #ifdef HAVE_SCHED_HEADERS
  38 #include <linux/sched/signal.h>
  39 #endif
  40 #include <linux/uaccess.h>
  41
  42 #include <lnet/udsp.h>
  43 #include <lnet/lib-lnet.h>
  44 #include <uapi/linux/lnet/lnet-dlc.h>
  45
  46 /* Value indicating that recovery needs to re-check a peer immediately. */
  47 #define LNET_REDISCOVER_PEER    (1)
  48
  49 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
  50
  51 static void
  52 lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
  53 {
  54         if (!list_empty(&lpni->lpni_on_remote_peer_ni_list)) {
  55                 list_del_init(&lpni->lpni_on_remote_peer_ni_list);
  56                 lnet_peer_ni_decref_locked(lpni);
  57         }
  58 }
  59
  60 void
  61 lnet_peer_net_added(struct lnet_net *net)
  62 {
  63         struct lnet_peer_ni *lpni, *tmp;
  64
  65         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
  66                                  lpni_on_remote_peer_ni_list) {
  67
  68                 if (LNET_NID_NET(&lpni->lpni_nid) == net->net_id) {
  69                         lpni->lpni_net = net;
  70
  71                         spin_lock(&lpni->lpni_lock);
  72                         lpni->lpni_txcredits =
  73                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits;
  74                         lpni->lpni_mintxcredits = lpni->lpni_txcredits;
  75                         lpni->lpni_rtrcredits =
  76                                 lnet_peer_buffer_credits(lpni->lpni_net);
  77                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
  78                         spin_unlock(&lpni->lpni_lock);
  79
  80                         lnet_peer_remove_from_remote_list(lpni);
  81                 }
  82         }
  83 }
  84
  85 static void
  86 lnet_peer_tables_destroy(void)
  87 {
  88         struct lnet_peer_table  *ptable;
  89         struct list_head        *hash;
  90         int                     i;
  91         int                     j;
  92
  93         if (!the_lnet.ln_peer_tables)
  94                 return;
  95
  96         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  97                 hash = ptable->pt_hash;
  98                 if (!hash) /* not intialized */
  99                         break;
 100
 101                 LASSERT(list_empty(&ptable->pt_zombie_list));
 102
 103                 ptable->pt_hash = NULL;
 104                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 105                         LASSERT(list_empty(&hash[j]));
 106
 107                 CFS_FREE_PTR_ARRAY(hash, LNET_PEER_HASH_SIZE);
 108         }
 109
 110         cfs_percpt_free(the_lnet.ln_peer_tables);
 111         the_lnet.ln_peer_tables = NULL;
 112 }
 113
 114 int
 115 lnet_peer_tables_create(void)
 116 {
 117         struct lnet_peer_table  *ptable;
 118         struct list_head        *hash;
 119         int                     i;
 120         int                     j;
 121
 122         the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
 123                                                    sizeof(*ptable));
 124         if (the_lnet.ln_peer_tables == NULL) {
 125                 CERROR("Failed to allocate cpu-partition peer tables\n");
 126                 return -ENOMEM;
 127         }
 128
 129         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 130                 LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
 131                                  LNET_PEER_HASH_SIZE * sizeof(*hash));
 132                 if (hash == NULL) {
 133                         CERROR("Failed to create peer hash table\n");
 134                         lnet_peer_tables_destroy();
 135                         return -ENOMEM;
 136                 }
 137
 138                 spin_lock_init(&ptable->pt_zombie_lock);
 139                 INIT_LIST_HEAD(&ptable->pt_zombie_list);
 140
 141                 INIT_LIST_HEAD(&ptable->pt_peer_list);
 142
 143                 for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
 144                         INIT_LIST_HEAD(&hash[j]);
 145                 ptable->pt_hash = hash; /* sign of initialization */
 146         }
 147
 148         return 0;
 149 }
 150
 151 static struct lnet_peer_ni *
 152 lnet_peer_ni_alloc(struct lnet_nid *nid)
 153 {
 154         struct lnet_peer_ni *lpni;
 155         struct lnet_net *net;
 156         int cpt;
 157
 158         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 159
 160         LIBCFS_CPT_ALLOC(lpni, lnet_cpt_table(), cpt, sizeof(*lpni));
 161         if (!lpni)
 162                 return NULL;
 163
 164         INIT_LIST_HEAD(&lpni->lpni_txq);
 165         INIT_LIST_HEAD(&lpni->lpni_hashlist);
 166         INIT_LIST_HEAD(&lpni->lpni_peer_nis);
 167         INIT_LIST_HEAD(&lpni->lpni_recovery);
 168         INIT_LIST_HEAD(&lpni->lpni_on_remote_peer_ni_list);
 169         INIT_LIST_HEAD(&lpni->lpni_rtr_pref_nids);
 170         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
 171         kref_init(&lpni->lpni_kref);
 172         lpni->lpni_sel_priority = LNET_MAX_SELECTION_PRIORITY;
 173
 174         spin_lock_init(&lpni->lpni_lock);
 175
 176         if (lnet_peers_start_down())
 177                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 178         else
 179                 lpni->lpni_ns_status = LNET_NI_STATUS_UP;
 180         lpni->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 181         lpni->lpni_nid = *nid;
 182         lpni->lpni_cpt = cpt;
 183         atomic_set(&lpni->lpni_healthv, LNET_MAX_HEALTH_VALUE);
 184
 185         net = lnet_get_net_locked(LNET_NID_NET(nid));
 186         lpni->lpni_net = net;
 187         if (net) {
 188                 lpni->lpni_txcredits = net->net_tunables.lct_peer_tx_credits;
 189                 lpni->lpni_mintxcredits = lpni->lpni_txcredits;
 190                 lpni->lpni_rtrcredits = lnet_peer_buffer_credits(net);
 191                 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 192         } else {
 193                 /*
 194                  * This peer_ni is not on a local network, so we
 195                  * cannot add the credits here. In case the net is
 196                  * added later, add the peer_ni to the remote peer ni
 197                  * list so it can be easily found and revisited.
 198                  */
 199                 /* FIXME: per-net implementation instead? */
 200                 lnet_peer_ni_addref_locked(lpni);
 201                 list_add_tail(&lpni->lpni_on_remote_peer_ni_list,
 202                               &the_lnet.ln_remote_peer_ni_list);
 203         }
 204
 205         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
 206
 207         return lpni;
 208 }
 209
 210 static struct lnet_peer_net *
 211 lnet_peer_net_alloc(__u32 net_id)
 212 {
 213         struct lnet_peer_net *lpn;
 214
 215         LIBCFS_CPT_ALLOC(lpn, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lpn));
 216         if (!lpn)
 217                 return NULL;
 218
 219         INIT_LIST_HEAD(&lpn->lpn_peer_nets);
 220         INIT_LIST_HEAD(&lpn->lpn_peer_nis);
 221         lpn->lpn_net_id = net_id;
 222         lpn->lpn_sel_priority = LNET_MAX_SELECTION_PRIORITY;
 223
 224         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 225
 226         return lpn;
 227 }
 228
 229 void
 230 lnet_destroy_peer_net_locked(struct lnet_peer_net *lpn)
 231 {
 232         struct lnet_peer *lp;
 233
 234         CDEBUG(D_NET, "%p net %s\n", lpn, libcfs_net2str(lpn->lpn_net_id));
 235
 236         LASSERT(atomic_read(&lpn->lpn_refcount) == 0);
 237         LASSERT(list_empty(&lpn->lpn_peer_nis));
 238         LASSERT(list_empty(&lpn->lpn_peer_nets));
 239         lp = lpn->lpn_peer;
 240         lpn->lpn_peer = NULL;
 241         LIBCFS_FREE(lpn, sizeof(*lpn));
 242
 243         lnet_peer_decref_locked(lp);
 244 }
 245
 246 static struct lnet_peer *
 247 lnet_peer_alloc(struct lnet_nid *nid)
 248 {
 249         struct lnet_peer *lp;
 250
 251         LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), CFS_CPT_ANY, sizeof(*lp));
 252         if (!lp)
 253                 return NULL;
 254
 255         INIT_LIST_HEAD(&lp->lp_rtrq);
 256         INIT_LIST_HEAD(&lp->lp_routes);
 257         INIT_LIST_HEAD(&lp->lp_peer_list);
 258         INIT_LIST_HEAD(&lp->lp_peer_nets);
 259         INIT_LIST_HEAD(&lp->lp_dc_list);
 260         INIT_LIST_HEAD(&lp->lp_dc_pendq);
 261         INIT_LIST_HEAD(&lp->lp_rtr_list);
 262         init_waitqueue_head(&lp->lp_dc_waitq);
 263         spin_lock_init(&lp->lp_lock);
 264         lp->lp_primary_nid = *nid;
 265         lp->lp_disc_src_nid = LNET_ANY_NID;
 266         lp->lp_disc_dst_nid = LNET_ANY_NID;
 267         if (lnet_peers_start_down())
 268                 lp->lp_alive = false;
 269         else
 270                 lp->lp_alive = true;
 271
 272         /*
 273          * all peers created on a router should have health on
 274          * if it's not already on.
 275          */
 276         if (the_lnet.ln_routing && !lnet_health_sensitivity)
 277                 lp->lp_health_sensitivity = 1;
 278
 279         /*
 280          * Turn off discovery for loopback peer. If you're creating a peer
 281          * for the loopback interface then that was initiated when we
 282          * attempted to send a message over the loopback. There is no need
 283          * to ever use a different interface when sending messages to
 284          * myself.
 285          */
 286         if (nid_is_lo0(nid))
 287                 lp->lp_state = LNET_PEER_NO_DISCOVERY;
 288         lp->lp_cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 289
 290         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
 291
 292         return lp;
 293 }
 294
 295 void
 296 lnet_destroy_peer_locked(struct lnet_peer *lp)
 297 {
 298         CDEBUG(D_NET, "%p nid %s\n", lp, libcfs_nidstr(&lp->lp_primary_nid));
 299
 300         LASSERT(atomic_read(&lp->lp_refcount) == 0);
 301         LASSERT(lp->lp_rtr_refcount == 0);
 302         LASSERT(list_empty(&lp->lp_peer_nets));
 303         LASSERT(list_empty(&lp->lp_peer_list));
 304         LASSERT(list_empty(&lp->lp_dc_list));
 305
 306         if (lp->lp_data)
 307                 lnet_ping_buffer_decref(lp->lp_data);
 308
 309         /*
 310          * if there are messages still on the pending queue, then make
 311          * sure to queue them on the ln_msg_resend list so they can be
 312          * resent at a later point if the discovery thread is still
 313          * running.
 314          * If the discovery thread has stopped, then the wakeup will be a
 315          * no-op, and it is expected the lnet_shutdown_lndnets() will
 316          * eventually be called, which will traverse this list and
 317          * finalize the messages on the list.
 318          * We can not resend them now because we're holding the cpt lock.
 319          * Releasing the lock can cause an inconsistent state
 320          */
 321         spin_lock(&the_lnet.ln_msg_resend_lock);
 322         spin_lock(&lp->lp_lock);
 323         list_splice(&lp->lp_dc_pendq, &the_lnet.ln_msg_resend);
 324         spin_unlock(&lp->lp_lock);
 325         spin_unlock(&the_lnet.ln_msg_resend_lock);
 326         wake_up(&the_lnet.ln_dc_waitq);
 327
 328         LIBCFS_FREE(lp, sizeof(*lp));
 329 }
 330
 331 /*
 332  * Detach a peer_ni from its peer_net. If this was the last peer_ni on
 333  * that peer_net, detach the peer_net from the peer.
 334  *
 335  * Call with lnet_net_lock/EX held
 336  */
 337 static void
 338 lnet_peer_detach_peer_ni_locked(struct lnet_peer_ni *lpni)
 339 {
 340         struct lnet_peer_table *ptable;
 341         struct lnet_peer_net *lpn;
 342         struct lnet_peer *lp;
 343
 344         /*
 345          * Belts and suspenders: gracefully handle teardown of a
 346          * partially connected peer_ni.
 347          */
 348         lpn = lpni->lpni_peer_net;
 349
 350         list_del_init(&lpni->lpni_peer_nis);
 351         /*
 352          * If there are no lpni's left, we detach lpn from
 353          * lp_peer_nets, so it cannot be found anymore.
 354          */
 355         if (list_empty(&lpn->lpn_peer_nis))
 356                 list_del_init(&lpn->lpn_peer_nets);
 357
 358         /* Update peer NID count. */
 359         lp = lpn->lpn_peer;
 360         lp->lp_nnis--;
 361
 362         /*
 363          * If there are no more peer nets, make the peer unfindable
 364          * via the peer_tables.
 365          *
 366          * Otherwise, if the peer is DISCOVERED, tell discovery to
 367          * take another look at it. This is a no-op if discovery for
 368          * this peer did the detaching.
 369          */
 370         if (list_empty(&lp->lp_peer_nets)) {
 371                 list_del_init(&lp->lp_peer_list);
 372                 ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 373                 ptable->pt_peers--;
 374         } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
 375                 /* Discovery isn't running, nothing to do here. */
 376         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
 377                 lnet_peer_queue_for_discovery(lp);
 378                 wake_up(&the_lnet.ln_dc_waitq);
 379         }
 380         CDEBUG(D_NET, "peer %s NID %s\n",
 381                 libcfs_nidstr(&lp->lp_primary_nid),
 382                 libcfs_nidstr(&lpni->lpni_nid));
 383 }
 384
 385 /* called with lnet_net_lock LNET_LOCK_EX held */
 386 static int
 387 lnet_peer_ni_del_locked(struct lnet_peer_ni *lpni, bool force)
 388 {
 389         struct lnet_peer_table *ptable = NULL;
 390
 391         /* don't remove a peer_ni if it's also a gateway */
 392         if (lnet_isrouter(lpni) && !force) {
 393                 CERROR("Peer NI %s is a gateway. Can not delete it\n",
 394                        libcfs_nidstr(&lpni->lpni_nid));
 395                 return -EBUSY;
 396         }
 397
 398         lnet_peer_remove_from_remote_list(lpni);
 399
 400         /* remove peer ni from the hash list. */
 401         list_del_init(&lpni->lpni_hashlist);
 402
 403         /*
 404          * indicate the peer is being deleted so the monitor thread can
 405          * remove it from the recovery queue.
 406          */
 407         spin_lock(&lpni->lpni_lock);
 408         lpni->lpni_state |= LNET_PEER_NI_DELETING;
 409         spin_unlock(&lpni->lpni_lock);
 410
 411         /* decrement the ref count on the peer table */
 412         ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
 413
 414         /*
 415          * The peer_ni can no longer be found with a lookup. But there
 416          * can be current users, so keep track of it on the zombie
 417          * list until the reference count has gone to zero.
 418          *
 419          * The last reference may be lost in a place where the
 420          * lnet_net_lock locks only a single cpt, and that cpt may not
 421          * be lpni->lpni_cpt. So the zombie list of lnet_peer_table
 422          * has its own lock.
 423          */
 424         spin_lock(&ptable->pt_zombie_lock);
 425         list_add(&lpni->lpni_hashlist, &ptable->pt_zombie_list);
 426         ptable->pt_zombies++;
 427         spin_unlock(&ptable->pt_zombie_lock);
 428
 429         /* no need to keep this peer_ni on the hierarchy anymore */
 430         lnet_peer_detach_peer_ni_locked(lpni);
 431
 432         /* remove hashlist reference on peer_ni */
 433         lnet_peer_ni_decref_locked(lpni);
 434
 435         return 0;
 436 }
 437
 438 void lnet_peer_uninit(void)
 439 {
 440         struct lnet_peer_ni *lpni, *tmp;
 441
 442         lnet_net_lock(LNET_LOCK_EX);
 443
 444         /* remove all peer_nis from the remote peer and the hash list */
 445         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_remote_peer_ni_list,
 446                                  lpni_on_remote_peer_ni_list)
 447                 lnet_peer_ni_del_locked(lpni, false);
 448
 449         lnet_peer_tables_destroy();
 450
 451         lnet_net_unlock(LNET_LOCK_EX);
 452 }
 453
 454 static int
 455 lnet_peer_del_locked(struct lnet_peer *peer)
 456 {
 457         struct lnet_peer_ni *lpni = NULL, *lpni2;
 458         int rc = 0, rc2 = 0;
 459
 460         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&peer->lp_primary_nid));
 461
 462         spin_lock(&peer->lp_lock);
 463         peer->lp_state |= LNET_PEER_MARK_DELETED;
 464         spin_unlock(&peer->lp_lock);
 465
 466         lpni = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 467         while (lpni != NULL) {
 468                 lpni2 = lnet_get_next_peer_ni_locked(peer, NULL, lpni);
 469                 rc = lnet_peer_ni_del_locked(lpni, false);
 470                 if (rc != 0)
 471                         rc2 = rc;
 472                 lpni = lpni2;
 473         }
 474
 475         return rc2;
 476 }
 477
 478 /*
 479  * Discovering this peer is taking too long. Cancel any Ping or Push
 480  * that discovery is waiting on by unlinking the relevant MDs. The
 481  * lnet_discovery_event_handler() will proceed from here and complete
 482  * the cleanup.
 483  */
 484 static void lnet_peer_cancel_discovery(struct lnet_peer *lp)
 485 {
 486         struct lnet_handle_md ping_mdh;
 487         struct lnet_handle_md push_mdh;
 488
 489         LNetInvalidateMDHandle(&ping_mdh);
 490         LNetInvalidateMDHandle(&push_mdh);
 491
 492         spin_lock(&lp->lp_lock);
 493         if (lp->lp_state & LNET_PEER_PING_SENT) {
 494                 ping_mdh = lp->lp_ping_mdh;
 495                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
 496         }
 497         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
 498                 push_mdh = lp->lp_push_mdh;
 499                 LNetInvalidateMDHandle(&lp->lp_push_mdh);
 500         }
 501         spin_unlock(&lp->lp_lock);
 502
 503         if (!LNetMDHandleIsInvalid(ping_mdh))
 504                 LNetMDUnlink(ping_mdh);
 505         if (!LNetMDHandleIsInvalid(push_mdh))
 506                 LNetMDUnlink(push_mdh);
 507 }
 508
 509 static int
 510 lnet_peer_del(struct lnet_peer *peer)
 511 {
 512         int rc;
 513
 514         lnet_peer_cancel_discovery(peer);
 515         lnet_net_lock(LNET_LOCK_EX);
 516         rc = lnet_peer_del_locked(peer);
 517         lnet_net_unlock(LNET_LOCK_EX);
 518
 519         return rc;
 520 }
 521
 522 /*
 523  * Delete a NID from a peer. Call with ln_api_mutex held.
 524  *
 525  * Error codes:
 526  *  -EPERM:  Non-DLC deletion from DLC-configured peer.
 527  *  -ENOENT: No lnet_peer_ni corresponding to the nid.
 528  *  -ECHILD: The lnet_peer_ni isn't connected to the peer.
 529  *  -EBUSY:  The lnet_peer_ni is the primary, and not the only peer_ni.
 530  */
 531 static int
 532 lnet_peer_del_nid(struct lnet_peer *lp, struct lnet_nid *nid,
 533                   unsigned int flags)
 534 {
 535         struct lnet_peer_ni *lpni;
 536         struct lnet_nid primary_nid = lp->lp_primary_nid;
 537         int rc = 0;
 538         bool force = (flags & LNET_PEER_RTR_NI_FORCE_DEL) ? true : false;
 539
 540         if (!(flags & LNET_PEER_CONFIGURED)) {
 541                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
 542                         rc = -EPERM;
 543                         goto out;
 544                 }
 545         }
 546
 547         lpni = lnet_peer_ni_find_locked(nid);
 548         if (!lpni) {
 549                 rc = -ENOENT;
 550                 goto out;
 551         }
 552         lnet_peer_ni_decref_locked(lpni);
 553         if (lp != lpni->lpni_peer_net->lpn_peer) {
 554                 rc = -ECHILD;
 555                 goto out;
 556         }
 557
 558         /*
 559          * This function only allows deletion of the primary NID if it
 560          * is the only NID.
 561          */
 562         if (nid_same(nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && !force) {
 563                 rc = -EBUSY;
 564                 goto out;
 565         }
 566
 567         lnet_net_lock(LNET_LOCK_EX);
 568
 569         if (nid_same(nid, &lp->lp_primary_nid) && lp->lp_nnis != 1 && force) {
 570                 struct lnet_peer_ni *lpni2;
 571                 /* assign the next peer_ni to be the primary */
 572                 lpni2 = lnet_get_next_peer_ni_locked(lp, NULL, lpni);
 573                 LASSERT(lpni2);
 574                 lp->lp_primary_nid = lpni2->lpni_nid;
 575         }
 576         rc = lnet_peer_ni_del_locked(lpni, force);
 577
 578         lnet_net_unlock(LNET_LOCK_EX);
 579
 580 out:
 581         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
 582                libcfs_nidstr(&primary_nid), libcfs_nidstr(nid),
 583                flags, rc);
 584
 585         return rc;
 586 }
 587
 588 static void
 589 lnet_peer_table_cleanup_locked(struct lnet_net *net,
 590                                struct lnet_peer_table *ptable)
 591 {
 592         int                      i;
 593         struct lnet_peer_ni     *next;
 594         struct lnet_peer_ni     *lpni;
 595         struct lnet_peer        *peer;
 596
 597         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 598                 list_for_each_entry_safe(lpni, next, &ptable->pt_hash[i],
 599                                          lpni_hashlist) {
 600                         if (net != NULL && net != lpni->lpni_net)
 601                                 continue;
 602
 603                         peer = lpni->lpni_peer_net->lpn_peer;
 604                         if (!nid_same(&peer->lp_primary_nid,
 605                                        &lpni->lpni_nid)) {
 606                                 lnet_peer_ni_del_locked(lpni, false);
 607                                 continue;
 608                         }
 609                         /*
 610                          * Removing the primary NID implies removing
 611                          * the entire peer. Advance next beyond any
 612                          * peer_ni that belongs to the same peer.
 613                          */
 614                         list_for_each_entry_from(next, &ptable->pt_hash[i],
 615                                                  lpni_hashlist) {
 616                                 if (next->lpni_peer_net->lpn_peer != peer)
 617                                         break;
 618                         }
 619                         lnet_peer_del_locked(peer);
 620                 }
 621         }
 622 }
 623
 624 static void
 625 lnet_peer_ni_finalize_wait(struct lnet_peer_table *ptable)
 626 {
 627         wait_var_event_warning(&ptable->pt_zombies,
 628                                ptable->pt_zombies == 0,
 629                                "Waiting for %d zombies on peer table\n",
 630                                ptable->pt_zombies);
 631 }
 632
 633 static void
 634 lnet_peer_table_del_rtrs_locked(struct lnet_net *net,
 635                                 struct lnet_peer_table *ptable)
 636 {
 637         struct lnet_peer_ni     *lp;
 638         struct lnet_peer_ni     *tmp;
 639         struct lnet_nid         gw_nid;
 640         int                     i;
 641
 642         for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 643                 list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 644                                          lpni_hashlist) {
 645                         if (net != lp->lpni_net)
 646                                 continue;
 647
 648                         if (!lnet_isrouter(lp))
 649                                 continue;
 650
 651                         gw_nid = lp->lpni_peer_net->lpn_peer->lp_primary_nid;
 652
 653                         lnet_net_unlock(LNET_LOCK_EX);
 654                         lnet_del_route(LNET_NET_ANY, &gw_nid);
 655                         lnet_net_lock(LNET_LOCK_EX);
 656                 }
 657         }
 658 }
 659
 660 void
 661 lnet_peer_tables_cleanup(struct lnet_net *net)
 662 {
 663         int i;
 664         struct lnet_peer_table *ptable;
 665
 666         LASSERT(the_lnet.ln_state != LNET_STATE_SHUTDOWN || net != NULL);
 667         /* If just deleting the peers for a NI, get rid of any routes these
 668          * peers are gateways for. */
 669         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 670                 lnet_net_lock(LNET_LOCK_EX);
 671                 lnet_peer_table_del_rtrs_locked(net, ptable);
 672                 lnet_net_unlock(LNET_LOCK_EX);
 673         }
 674
 675         /* Start the cleanup process */
 676         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 677                 lnet_net_lock(LNET_LOCK_EX);
 678                 lnet_peer_table_cleanup_locked(net, ptable);
 679                 lnet_net_unlock(LNET_LOCK_EX);
 680         }
 681
 682         cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables)
 683                 lnet_peer_ni_finalize_wait(ptable);
 684 }
 685
 686 static struct lnet_peer_ni *
 687 lnet_get_peer_ni_locked(struct lnet_peer_table *ptable, struct lnet_nid *nid)
 688 {
 689         struct list_head        *peers;
 690         struct lnet_peer_ni     *lp;
 691
 692         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 693                 return NULL;
 694
 695         peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 696         list_for_each_entry(lp, peers, lpni_hashlist) {
 697                 if (nid_same(&lp->lpni_nid, nid)) {
 698                         lnet_peer_ni_addref_locked(lp);
 699                         return lp;
 700                 }
 701         }
 702
 703         return NULL;
 704 }
 705
 706 struct lnet_peer_ni *
 707 lnet_peer_ni_find_locked(struct lnet_nid *nid)
 708 {
 709         struct lnet_peer_ni *lpni;
 710         struct lnet_peer_table *ptable;
 711         int cpt;
 712
 713         cpt = lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 714
 715         ptable = the_lnet.ln_peer_tables[cpt];
 716         lpni = lnet_get_peer_ni_locked(ptable, nid);
 717
 718         return lpni;
 719 }
 720
 721 struct lnet_peer_ni *
 722 lnet_peer_ni_get_locked(struct lnet_peer *lp, struct lnet_nid *nid)
 723 {
 724         struct lnet_peer_net *lpn;
 725         struct lnet_peer_ni *lpni;
 726
 727         lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
 728         if (!lpn)
 729                 return NULL;
 730
 731         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 732                 if (nid_same(&lpni->lpni_nid, nid))
 733                         return lpni;
 734         }
 735
 736         return NULL;
 737 }
 738
 739 struct lnet_peer *
 740 lnet_find_peer(struct lnet_nid *nid)
 741 {
 742         struct lnet_peer_ni *lpni;
 743         struct lnet_peer *lp = NULL;
 744         int cpt;
 745
 746         cpt = lnet_net_lock_current();
 747         lpni = lnet_peer_ni_find_locked(nid);
 748         if (lpni) {
 749                 lp = lpni->lpni_peer_net->lpn_peer;
 750                 lnet_peer_addref_locked(lp);
 751                 lnet_peer_ni_decref_locked(lpni);
 752         }
 753         lnet_net_unlock(cpt);
 754
 755         return lp;
 756 }
 757
 758 struct lnet_peer_net *
 759 lnet_get_next_peer_net_locked(struct lnet_peer *lp, __u32 prev_lpn_id)
 760 {
 761         struct lnet_peer_net *net;
 762
 763         if (!prev_lpn_id) {
 764                 /* no net id provided return the first net */
 765                 net = list_first_entry_or_null(&lp->lp_peer_nets,
 766                                                struct lnet_peer_net,
 767                                                lpn_peer_nets);
 768
 769                 return net;
 770         }
 771
 772         /* find the net after the one provided */
 773         list_for_each_entry(net, &lp->lp_peer_nets, lpn_peer_nets) {
 774                 if (net->lpn_net_id == prev_lpn_id) {
 775                         /*
 776                          * if we reached the end of the list loop to the
 777                          * beginning.
 778                          */
 779                         if (net->lpn_peer_nets.next == &lp->lp_peer_nets)
 780                                 return list_first_entry_or_null(&lp->lp_peer_nets,
 781                                                                 struct lnet_peer_net,
 782                                                                 lpn_peer_nets);
 783                         else
 784                                 return list_next_entry(net, lpn_peer_nets);
 785                 }
 786         }
 787
 788         return NULL;
 789 }
 790
 791 struct lnet_peer_ni *
 792 lnet_get_next_peer_ni_locked(struct lnet_peer *peer,
 793                              struct lnet_peer_net *peer_net,
 794                              struct lnet_peer_ni *prev)
 795 {
 796         struct lnet_peer_ni *lpni;
 797         struct lnet_peer_net *net = peer_net;
 798
 799         if (!prev) {
 800                 if (!net) {
 801                         if (list_empty(&peer->lp_peer_nets))
 802                                 return NULL;
 803
 804                         net = list_first_entry(&peer->lp_peer_nets,
 805                                                struct lnet_peer_net,
 806                                                lpn_peer_nets);
 807                 }
 808                 lpni = list_first_entry(&net->lpn_peer_nis, struct lnet_peer_ni,
 809                                         lpni_peer_nis);
 810
 811                 return lpni;
 812         }
 813
 814         if (prev->lpni_peer_nis.next == &prev->lpni_peer_net->lpn_peer_nis) {
 815                 /*
 816                  * if you reached the end of the peer ni list and the peer
 817                  * net is specified then there are no more peer nis in that
 818                  * net.
 819                  */
 820                 if (net)
 821                         return NULL;
 822
 823                 /*
 824                  * we reached the end of this net ni list. move to the
 825                  * next net
 826                  */
 827                 if (prev->lpni_peer_net->lpn_peer_nets.next ==
 828                     &peer->lp_peer_nets)
 829                         /* no more nets and no more NIs. */
 830                         return NULL;
 831
 832                 /* get the next net */
 833                 net = list_first_entry(&prev->lpni_peer_net->lpn_peer_nets,
 834                                        struct lnet_peer_net,
 835                                        lpn_peer_nets);
 836                 /* get the ni on it */
 837                 lpni = list_first_entry(&net->lpn_peer_nis, struct lnet_peer_ni,
 838                                         lpni_peer_nis);
 839
 840                 return lpni;
 841         }
 842
 843         /* there are more nis left */
 844         lpni = list_first_entry(&prev->lpni_peer_nis,
 845                                 struct lnet_peer_ni, lpni_peer_nis);
 846
 847         return lpni;
 848 }
 849
 850 /* Call with the ln_api_mutex held */
 851 int lnet_get_peer_list(u32 *countp, u32 *sizep, struct lnet_process_id __user *ids)
 852 {
 853         struct lnet_process_id id;
 854         struct lnet_peer_table *ptable;
 855         struct lnet_peer *lp;
 856         __u32 count = 0;
 857         __u32 size = 0;
 858         int lncpt;
 859         int cpt;
 860         __u32 i;
 861         int rc;
 862
 863         rc = -ESHUTDOWN;
 864         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 865                 goto done;
 866
 867         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 868
 869         /*
 870          * Count the number of peers, and return E2BIG if the buffer
 871          * is too small. We'll also return the desired size.
 872          */
 873         rc = -E2BIG;
 874         for (cpt = 0; cpt < lncpt; cpt++) {
 875                 ptable = the_lnet.ln_peer_tables[cpt];
 876                 count += ptable->pt_peers;
 877         }
 878         size = count * sizeof(*ids);
 879         if (size > *sizep)
 880                 goto done;
 881
 882         /*
 883          * Walk the peer lists and copy out the primary nids.
 884          * This is safe because the peer lists are only modified
 885          * while the ln_api_mutex is held. So we don't need to
 886          * hold the lnet_net_lock as well, and can therefore
 887          * directly call copy_to_user().
 888          */
 889         rc = -EFAULT;
 890         memset(&id, 0, sizeof(id));
 891         id.pid = LNET_PID_LUSTRE;
 892         i = 0;
 893         for (cpt = 0; cpt < lncpt; cpt++) {
 894                 ptable = the_lnet.ln_peer_tables[cpt];
 895                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 896                         if (!nid_is_nid4(&lp->lp_primary_nid))
 897                                 continue;
 898                         if (i >= count)
 899                                 goto done;
 900                         id.nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
 901                         if (copy_to_user(&ids[i], &id, sizeof(id)))
 902                                 goto done;
 903                         i++;
 904                 }
 905         }
 906         rc = 0;
 907 done:
 908         *countp = count;
 909         *sizep = size;
 910         return rc;
 911 }
 912
 913 /*
 914  * Start pushes to peers that need to be updated for a configuration
 915  * change on this node.
 916  */
 917 void
 918 lnet_push_update_to_peers(int force)
 919 {
 920         struct lnet_peer_table *ptable;
 921         struct lnet_peer *lp;
 922         int lncpt;
 923         int cpt;
 924
 925         lnet_net_lock(LNET_LOCK_EX);
 926         if (lnet_peer_discovery_disabled)
 927                 force = 0;
 928         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 929         for (cpt = 0; cpt < lncpt; cpt++) {
 930                 ptable = the_lnet.ln_peer_tables[cpt];
 931                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
 932                         if (force) {
 933                                 spin_lock(&lp->lp_lock);
 934                                 if (lp->lp_state & LNET_PEER_MULTI_RAIL)
 935                                         lp->lp_state |= LNET_PEER_FORCE_PUSH;
 936                                 spin_unlock(&lp->lp_lock);
 937                         }
 938                         if (lnet_peer_needs_push(lp))
 939                                 lnet_peer_queue_for_discovery(lp);
 940                 }
 941         }
 942         lnet_net_unlock(LNET_LOCK_EX);
 943         wake_up(&the_lnet.ln_dc_waitq);
 944 }
 945
 946 /* find the NID in the preferred gateways for the remote peer
 947  * return:
 948  *      false: list is not empty and NID is not preferred
 949  *      false: list is empty
 950  *      true: nid is found in the list
 951  */
 952 bool
 953 lnet_peer_is_pref_rtr_locked(struct lnet_peer_ni *lpni,
 954                              struct lnet_nid *gw_nid)
 955 {
 956         struct lnet_nid_list *ne;
 957
 958         CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
 959                libcfs_nidstr(&lpni->lpni_nid),
 960                list_empty(&lpni->lpni_rtr_pref_nids));
 961
 962         if (list_empty(&lpni->lpni_rtr_pref_nids))
 963                 return false;
 964
 965         /* iterate through all the preferred NIDs and see if any of them
 966          * matches the provided gw_nid
 967          */
 968         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
 969                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
 970                        libcfs_nidstr(&ne->nl_nid),
 971                        libcfs_nidstr(gw_nid));
 972                 if (nid_same(&ne->nl_nid, gw_nid))
 973                         return true;
 974         }
 975
 976         return false;
 977 }
 978
 979 void
 980 lnet_peer_clr_pref_rtrs(struct lnet_peer_ni *lpni)
 981 {
 982         struct list_head zombies;
 983         struct lnet_nid_list *ne;
 984         struct lnet_nid_list *tmp;
 985         int cpt = lpni->lpni_cpt;
 986
 987         INIT_LIST_HEAD(&zombies);
 988
 989         lnet_net_lock(cpt);
 990         list_splice_init(&lpni->lpni_rtr_pref_nids, &zombies);
 991         lnet_net_unlock(cpt);
 992
 993         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
 994                 list_del(&ne->nl_list);
 995                 LIBCFS_FREE(ne, sizeof(*ne));
 996         }
 997 }
 998
 999 int
1000 lnet_peer_add_pref_rtr(struct lnet_peer_ni *lpni,
1001                        struct lnet_nid *gw_nid)
1002 {
1003         int cpt = lpni->lpni_cpt;
1004         struct lnet_nid_list *ne = NULL;
1005
1006         /* This function is called with api_mutex held. When the api_mutex
1007          * is held the list can not be modified, as it is only modified as
1008          * a result of applying a UDSP and that happens under api_mutex
1009          * lock.
1010          */
1011         __must_hold(&the_lnet.ln_api_mutex);
1012
1013         list_for_each_entry(ne, &lpni->lpni_rtr_pref_nids, nl_list) {
1014                 if (nid_same(&ne->nl_nid, gw_nid))
1015                         return -EEXIST;
1016         }
1017
1018         LIBCFS_CPT_ALLOC(ne, lnet_cpt_table(), cpt, sizeof(*ne));
1019         if (!ne)
1020                 return -ENOMEM;
1021
1022         ne->nl_nid = *gw_nid;
1023
1024         /* Lock the cpt to protect against addition and checks in the
1025          * selection algorithm
1026          */
1027         lnet_net_lock(cpt);
1028         list_add(&ne->nl_list, &lpni->lpni_rtr_pref_nids);
1029         lnet_net_unlock(cpt);
1030
1031         return 0;
1032 }
1033
1034 /*
1035  * Test whether a ni is a preferred ni for this peer_ni, e.g, whether
1036  * this is a preferred point-to-point path. Call with lnet_net_lock in
1037  * shared mmode.
1038  */
1039 bool
1040 lnet_peer_is_pref_nid_locked(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1041 {
1042         struct lnet_nid_list *ne;
1043
1044         if (lpni->lpni_pref_nnids == 0)
1045                 return false;
1046         if (lpni->lpni_pref_nnids == 1)
1047                 return nid_same(&lpni->lpni_pref.nid, nid);
1048         list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1049                 if (nid_same(&ne->nl_nid, nid))
1050                         return true;
1051         }
1052         return false;
1053 }
1054
1055 /*
1056  * Set a single ni as preferred, provided no preferred ni is already
1057  * defined. Only to be used for non-multi-rail peer_ni.
1058  */
1059 int
1060 lnet_peer_ni_set_non_mr_pref_nid(struct lnet_peer_ni *lpni,
1061                                   struct lnet_nid *nid)
1062 {
1063         int rc = 0;
1064
1065         if (!nid)
1066                 return -EINVAL;
1067         spin_lock(&lpni->lpni_lock);
1068         if (LNET_NID_IS_ANY(nid)) {
1069                 rc = -EINVAL;
1070         } else if (lpni->lpni_pref_nnids > 0) {
1071                 rc = -EPERM;
1072         } else if (lpni->lpni_pref_nnids == 0) {
1073                 lpni->lpni_pref.nid = *nid;
1074                 lpni->lpni_pref_nnids = 1;
1075                 lpni->lpni_state |= LNET_PEER_NI_NON_MR_PREF;
1076         }
1077         spin_unlock(&lpni->lpni_lock);
1078
1079         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1080                libcfs_nidstr(&lpni->lpni_nid), libcfs_nidstr(nid), rc);
1081         return rc;
1082 }
1083
1084 /*
1085  * Clear the preferred NID from a non-multi-rail peer_ni, provided
1086  * this preference was set by lnet_peer_ni_set_non_mr_pref_nid().
1087  */
1088 int
1089 lnet_peer_ni_clr_non_mr_pref_nid(struct lnet_peer_ni *lpni)
1090 {
1091         int rc = 0;
1092
1093         spin_lock(&lpni->lpni_lock);
1094         if (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF) {
1095                 lpni->lpni_pref_nnids = 0;
1096                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1097         } else if (lpni->lpni_pref_nnids == 0) {
1098                 rc = -ENOENT;
1099         } else {
1100                 rc = -EPERM;
1101         }
1102         spin_unlock(&lpni->lpni_lock);
1103
1104         CDEBUG(D_NET, "peer %s: %d\n",
1105                libcfs_nidstr(&lpni->lpni_nid), rc);
1106         return rc;
1107 }
1108
1109 void
1110 lnet_peer_ni_set_selection_priority(struct lnet_peer_ni *lpni, __u32 priority)
1111 {
1112         lpni->lpni_sel_priority = priority;
1113 }
1114
1115 /*
1116  * Clear the preferred NIDs from a non-multi-rail peer.
1117  */
1118 void
1119 lnet_peer_clr_non_mr_pref_nids(struct lnet_peer *lp)
1120 {
1121         struct lnet_peer_ni *lpni = NULL;
1122
1123         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
1124                 lnet_peer_ni_clr_non_mr_pref_nid(lpni);
1125 }
1126
1127 int
1128 lnet_peer_add_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1129 {
1130         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1131         struct lnet_nid_list *ne1 = NULL;
1132         struct lnet_nid_list *ne2 = NULL;
1133         struct lnet_nid *tmp_nid = NULL;
1134         int rc = 0;
1135
1136         if (LNET_NID_IS_ANY(nid)) {
1137                 rc = -EINVAL;
1138                 goto out;
1139         }
1140
1141         if (lpni->lpni_pref_nnids == 1 &&
1142             nid_same(&lpni->lpni_pref.nid, nid)) {
1143                 rc = -EEXIST;
1144                 goto out;
1145         }
1146
1147         /* A non-MR node may have only one preferred NI per peer_ni */
1148         if (lpni->lpni_pref_nnids > 0 &&
1149             !(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1150                 rc = -EPERM;
1151                 goto out;
1152         }
1153
1154         /* add the new preferred nid to the list of preferred nids */
1155         if (lpni->lpni_pref_nnids != 0) {
1156                 size_t alloc_size = sizeof(*ne1);
1157
1158                 if (lpni->lpni_pref_nnids == 1) {
1159                         tmp_nid = &lpni->lpni_pref.nid;
1160                         INIT_LIST_HEAD(&lpni->lpni_pref.nids);
1161                 }
1162
1163                 list_for_each_entry(ne1, &lpni->lpni_pref.nids, nl_list) {
1164                         if (nid_same(&ne1->nl_nid, nid)) {
1165                                 rc = -EEXIST;
1166                                 goto out;
1167                         }
1168                 }
1169
1170                 LIBCFS_CPT_ALLOC(ne1, lnet_cpt_table(), lpni->lpni_cpt,
1171                                  alloc_size);
1172                 if (!ne1) {
1173                         rc = -ENOMEM;
1174                         goto out;
1175                 }
1176
1177                 /* move the originally stored nid to the list */
1178                 if (lpni->lpni_pref_nnids == 1) {
1179                         LIBCFS_CPT_ALLOC(ne2, lnet_cpt_table(),
1180                                 lpni->lpni_cpt, alloc_size);
1181                         if (!ne2) {
1182                                 rc = -ENOMEM;
1183                                 goto out;
1184                         }
1185                         INIT_LIST_HEAD(&ne2->nl_list);
1186                         ne2->nl_nid = *tmp_nid;
1187                 }
1188                 ne1->nl_nid = *nid;
1189         }
1190
1191         lnet_net_lock(LNET_LOCK_EX);
1192         spin_lock(&lpni->lpni_lock);
1193         if (lpni->lpni_pref_nnids == 0) {
1194                 lpni->lpni_pref.nid = *nid;
1195         } else {
1196                 if (ne2)
1197                         list_add_tail(&ne2->nl_list, &lpni->lpni_pref.nids);
1198                 list_add_tail(&ne1->nl_list, &lpni->lpni_pref.nids);
1199         }
1200         lpni->lpni_pref_nnids++;
1201         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1202         spin_unlock(&lpni->lpni_lock);
1203         lnet_net_unlock(LNET_LOCK_EX);
1204
1205 out:
1206         if (rc == -EEXIST && (lpni->lpni_state & LNET_PEER_NI_NON_MR_PREF)) {
1207                 spin_lock(&lpni->lpni_lock);
1208                 lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1209                 spin_unlock(&lpni->lpni_lock);
1210         }
1211         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1212                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1213         return rc;
1214 }
1215
1216 int
1217 lnet_peer_del_pref_nid(struct lnet_peer_ni *lpni, struct lnet_nid *nid)
1218 {
1219         struct lnet_peer *lp = lpni->lpni_peer_net->lpn_peer;
1220         struct lnet_nid_list *ne = NULL;
1221         int rc = 0;
1222
1223         if (lpni->lpni_pref_nnids == 0) {
1224                 rc = -ENOENT;
1225                 goto out;
1226         }
1227
1228         if (lpni->lpni_pref_nnids == 1) {
1229                 if (!nid_same(&lpni->lpni_pref.nid, nid)) {
1230                         rc = -ENOENT;
1231                         goto out;
1232                 }
1233         } else {
1234                 list_for_each_entry(ne, &lpni->lpni_pref.nids, nl_list) {
1235                         if (nid_same(&ne->nl_nid, nid))
1236                                 goto remove_nid_entry;
1237                 }
1238                 rc = -ENOENT;
1239                 ne = NULL;
1240                 goto out;
1241         }
1242
1243 remove_nid_entry:
1244         lnet_net_lock(LNET_LOCK_EX);
1245         spin_lock(&lpni->lpni_lock);
1246         if (lpni->lpni_pref_nnids == 1)
1247                 lpni->lpni_pref.nid = LNET_ANY_NID;
1248         else {
1249                 list_del_init(&ne->nl_list);
1250                 if (lpni->lpni_pref_nnids == 2) {
1251                         struct lnet_nid_list *ne, *tmp;
1252
1253                         list_for_each_entry_safe(ne, tmp,
1254                                                  &lpni->lpni_pref.nids,
1255                                                  nl_list) {
1256                                 lpni->lpni_pref.nid = ne->nl_nid;
1257                                 list_del_init(&ne->nl_list);
1258                                 LIBCFS_FREE(ne, sizeof(*ne));
1259                         }
1260                 }
1261         }
1262         lpni->lpni_pref_nnids--;
1263         lpni->lpni_state &= ~LNET_PEER_NI_NON_MR_PREF;
1264         spin_unlock(&lpni->lpni_lock);
1265         lnet_net_unlock(LNET_LOCK_EX);
1266
1267         if (ne)
1268                 LIBCFS_FREE(ne, sizeof(*ne));
1269 out:
1270         CDEBUG(D_NET, "peer %s nid %s: %d\n",
1271                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid), rc);
1272         return rc;
1273 }
1274
1275 void
1276 lnet_peer_clr_pref_nids(struct lnet_peer_ni *lpni)
1277 {
1278         struct list_head zombies;
1279         struct lnet_nid_list *ne;
1280         struct lnet_nid_list *tmp;
1281
1282         INIT_LIST_HEAD(&zombies);
1283
1284         lnet_net_lock(LNET_LOCK_EX);
1285         if (lpni->lpni_pref_nnids == 1)
1286                 lpni->lpni_pref.nid = LNET_ANY_NID;
1287         else if (lpni->lpni_pref_nnids > 1)
1288                 list_splice_init(&lpni->lpni_pref.nids, &zombies);
1289         lpni->lpni_pref_nnids = 0;
1290         lnet_net_unlock(LNET_LOCK_EX);
1291
1292         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1293                 list_del_init(&ne->nl_list);
1294                 LIBCFS_FREE(ne, sizeof(*ne));
1295         }
1296 }
1297
1298 void
1299 lnet_peer_primary_nid_locked(struct lnet_nid *nid, struct lnet_nid *result)
1300 {
1301         struct lnet_peer_ni *lpni;
1302
1303         *result = *nid;
1304         lpni = lnet_peer_ni_find_locked(nid);
1305         if (lpni) {
1306                 *result = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
1307                 lnet_peer_ni_decref_locked(lpni);
1308         }
1309 }
1310
1311 bool
1312 lnet_is_discovery_disabled_locked(struct lnet_peer *lp)
1313 __must_hold(&lp->lp_lock)
1314 {
1315         if (lnet_peer_discovery_disabled)
1316                 return true;
1317
1318         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL) ||
1319             (lp->lp_state & LNET_PEER_NO_DISCOVERY)) {
1320                 return true;
1321         }
1322
1323         return false;
1324 }
1325
1326 /*
1327  * Peer Discovery
1328  */
1329 bool
1330 lnet_is_discovery_disabled(struct lnet_peer *lp)
1331 {
1332         bool rc = false;
1333
1334         spin_lock(&lp->lp_lock);
1335         rc = lnet_is_discovery_disabled_locked(lp);
1336         spin_unlock(&lp->lp_lock);
1337
1338         return rc;
1339 }
1340
1341 int
1342 LNetAddPeer(lnet_nid_t *nids, __u32 num_nids)
1343 {
1344         struct lnet_nid pnid = LNET_ANY_NID;
1345         bool mr;
1346         int i, rc;
1347
1348         if (!nids || num_nids < 1)
1349                 return -EINVAL;
1350
1351         rc = LNetNIInit(LNET_PID_ANY);
1352         if (rc < 0)
1353                 return rc;
1354
1355         mutex_lock(&the_lnet.ln_api_mutex);
1356
1357         mr = lnet_peer_discovery_disabled == 0;
1358
1359         rc = 0;
1360         for (i = 0; i < num_nids; i++) {
1361                 struct lnet_nid nid;
1362
1363                 if (nids[i] == LNET_NID_LO_0)
1364                         continue;
1365
1366                 lnet_nid4_to_nid(nids[i], &nid);
1367                 if (LNET_NID_IS_ANY(&pnid)) {
1368                         lnet_nid4_to_nid(nids[i], &pnid);
1369                         rc = lnet_add_peer_ni(&pnid, &LNET_ANY_NID, mr, true);
1370                 } else if (lnet_peer_discovery_disabled) {
1371                         lnet_nid4_to_nid(nids[i], &nid);
1372                         rc = lnet_add_peer_ni(&nid, &LNET_ANY_NID, mr, true);
1373                 } else {
1374                         lnet_nid4_to_nid(nids[i], &nid);
1375                         rc = lnet_add_peer_ni(&pnid, &nid, mr, true);
1376                 }
1377
1378                 if (rc && rc != -EEXIST)
1379                         goto unlock;
1380         }
1381
1382 unlock:
1383         mutex_unlock(&the_lnet.ln_api_mutex);
1384
1385         LNetNIFini();
1386
1387         return rc == -EEXIST ? 0 : rc;
1388 }
1389 EXPORT_SYMBOL(LNetAddPeer);
1390
1391 void LNetPrimaryNID(struct lnet_nid *nid)
1392 {
1393         struct lnet_peer *lp;
1394         struct lnet_peer_ni *lpni;
1395         struct lnet_nid orig;
1396         int rc = 0;
1397         int cpt;
1398
1399         if (!nid || nid_is_lo0(nid))
1400                 return;
1401         orig = *nid;
1402
1403         cpt = lnet_net_lock_current();
1404         lpni = lnet_peerni_by_nid_locked(nid, NULL, cpt);
1405         if (IS_ERR(lpni)) {
1406                 rc = PTR_ERR(lpni);
1407                 goto out_unlock;
1408         }
1409         lp = lpni->lpni_peer_net->lpn_peer;
1410
1411         /* If discovery is disabled locally then we needn't bother running
1412          * discovery here because discovery will not modify whatever
1413          * primary NID is currently set for this peer. If the specified peer is
1414          * down then this discovery can introduce long delays into the mount
1415          * process, so skip it if it isn't necessary.
1416          */
1417         while (!lnet_peer_discovery_disabled && !lnet_peer_is_uptodate(lp)) {
1418                 spin_lock(&lp->lp_lock);
1419                 /* force a full discovery cycle */
1420                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
1421                 spin_unlock(&lp->lp_lock);
1422
1423                 rc = lnet_discover_peer_locked(lpni, cpt, true);
1424                 if (rc)
1425                         goto out_decref;
1426                 /* The lpni (or lp) for this NID may have changed and our ref is
1427                  * the only thing keeping the old one around. Release the ref
1428                  * and lookup the lpni again
1429                  */
1430                 lnet_peer_ni_decref_locked(lpni);
1431                 lpni = lnet_peer_ni_find_locked(nid);
1432                 if (!lpni) {
1433                         rc = -ENOENT;
1434                         goto out_unlock;
1435                 }
1436                 lp = lpni->lpni_peer_net->lpn_peer;
1437
1438                 /* If we find that the peer has discovery disabled then we will
1439                  * not modify whatever primary NID is currently set for this
1440                  * peer. Thus, we can break out of this loop even if the peer
1441                  * is not fully up to date.
1442                  */
1443                 if (lnet_is_discovery_disabled(lp))
1444                         break;
1445         }
1446         *nid = lp->lp_primary_nid;
1447 out_decref:
1448         lnet_peer_ni_decref_locked(lpni);
1449 out_unlock:
1450         lnet_net_unlock(cpt);
1451
1452         CDEBUG(D_NET, "NID %s primary NID %s rc %d\n", libcfs_nidstr(&orig),
1453                libcfs_nidstr(nid), rc);
1454 }
1455 EXPORT_SYMBOL(LNetPrimaryNID);
1456
1457 struct lnet_peer_net *
1458 lnet_peer_get_net_locked(struct lnet_peer *peer, __u32 net_id)
1459 {
1460         struct lnet_peer_net *peer_net;
1461         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1462                 if (peer_net->lpn_net_id == net_id)
1463                         return peer_net;
1464         }
1465         return NULL;
1466 }
1467
1468 /*
1469  * Attach a peer_ni to a peer_net and peer. This function assumes
1470  * peer_ni is not already attached to the peer_net/peer. The peer_ni
1471  * may be attached to a different peer, in which case it will be
1472  * properly detached first. The whole operation is done atomically.
1473  *
1474  * This function consumes the reference on lpni and Always returns 0.
1475  * This is the last function called from functions that do return an
1476  * int, so returning 0 here allows the compiler to do a tail call.
1477  */
1478 static int
1479 lnet_peer_attach_peer_ni(struct lnet_peer *lp,
1480                                 struct lnet_peer_net *lpn,
1481                                 struct lnet_peer_ni *lpni,
1482                                 unsigned flags)
1483 {
1484         struct lnet_peer_table *ptable;
1485         bool new_lpn = false;
1486         int rc;
1487
1488         /* Install the new peer_ni */
1489         lnet_net_lock(LNET_LOCK_EX);
1490         /* Add peer_ni to global peer table hash, if necessary. */
1491         if (list_empty(&lpni->lpni_hashlist)) {
1492                 int hash = lnet_nid2peerhash(&lpni->lpni_nid);
1493
1494                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
1495                 list_add_tail(&lpni->lpni_hashlist, &ptable->pt_hash[hash]);
1496                 ptable->pt_version++;
1497                 lnet_peer_ni_addref_locked(lpni);
1498         }
1499
1500         /* Detach the peer_ni from an existing peer, if necessary. */
1501         if (lpni->lpni_peer_net) {
1502                 LASSERT(lpni->lpni_peer_net != lpn);
1503                 LASSERT(lpni->lpni_peer_net->lpn_peer != lp);
1504                 lnet_peer_detach_peer_ni_locked(lpni);
1505                 lnet_peer_net_decref_locked(lpni->lpni_peer_net);
1506                 lpni->lpni_peer_net = NULL;
1507         }
1508
1509         /* Add peer_ni to peer_net */
1510         lpni->lpni_peer_net = lpn;
1511         if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1512                 list_add(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1513         else
1514                 list_add_tail(&lpni->lpni_peer_nis, &lpn->lpn_peer_nis);
1515         lnet_update_peer_net_healthv(lpni);
1516         lnet_peer_net_addref_locked(lpn);
1517
1518         /* Add peer_net to peer */
1519         if (!lpn->lpn_peer) {
1520                 new_lpn = true;
1521                 lpn->lpn_peer = lp;
1522                 if (nid_same(&lp->lp_primary_nid, &lpni->lpni_nid))
1523                         list_add(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1524                 else
1525                         list_add_tail(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
1526                 lnet_peer_addref_locked(lp);
1527         }
1528
1529         /* Add peer to global peer list, if necessary */
1530         ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
1531         if (list_empty(&lp->lp_peer_list)) {
1532                 list_add_tail(&lp->lp_peer_list, &ptable->pt_peer_list);
1533                 ptable->pt_peers++;
1534         }
1535
1536
1537         /* Update peer state */
1538         spin_lock(&lp->lp_lock);
1539         if (flags & LNET_PEER_CONFIGURED) {
1540                 if (!(lp->lp_state & LNET_PEER_CONFIGURED))
1541                         lp->lp_state |= LNET_PEER_CONFIGURED;
1542         }
1543         if (flags & LNET_PEER_MULTI_RAIL) {
1544                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1545                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1546                         lnet_peer_clr_non_mr_pref_nids(lp);
1547                 }
1548         }
1549         spin_unlock(&lp->lp_lock);
1550
1551         lp->lp_nnis++;
1552
1553         /* apply UDSPs */
1554         if (new_lpn) {
1555                 rc = lnet_udsp_apply_policies_on_lpn(lpn);
1556                 if (rc)
1557                         CERROR("Failed to apply UDSPs on lpn %s\n",
1558                                libcfs_net2str(lpn->lpn_net_id));
1559         }
1560         rc = lnet_udsp_apply_policies_on_lpni(lpni);
1561         if (rc)
1562                 CERROR("Failed to apply UDSPs on lpni %s\n",
1563                        libcfs_nidstr(&lpni->lpni_nid));
1564
1565         CDEBUG(D_NET, "peer %s NID %s flags %#x\n",
1566                libcfs_nidstr(&lp->lp_primary_nid),
1567                libcfs_nidstr(&lpni->lpni_nid), flags);
1568         lnet_peer_ni_decref_locked(lpni);
1569         lnet_net_unlock(LNET_LOCK_EX);
1570
1571         return 0;
1572 }
1573
1574 /*
1575  * Create a new peer, with nid as its primary nid.
1576  *
1577  * Call with the lnet_api_mutex held.
1578  */
1579 static int
1580 lnet_peer_add(struct lnet_nid *nid, unsigned int flags)
1581 {
1582         struct lnet_peer *lp;
1583         struct lnet_peer_net *lpn;
1584         struct lnet_peer_ni *lpni;
1585         int rc = 0;
1586
1587         LASSERT(nid);
1588
1589         /*
1590          * No need for the lnet_net_lock here, because the
1591          * lnet_api_mutex is held.
1592          */
1593         lpni = lnet_peer_ni_find_locked(nid);
1594         if (lpni) {
1595                 /* A peer with this NID already exists. */
1596                 lp = lpni->lpni_peer_net->lpn_peer;
1597                 lnet_peer_ni_decref_locked(lpni);
1598                 /*
1599                  * This is an error if the peer was configured and the
1600                  * primary NID differs or an attempt is made to change
1601                  * the Multi-Rail flag. Otherwise the assumption is
1602                  * that an existing peer is being modified.
1603                  */
1604                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1605                         if (!nid_same(&lp->lp_primary_nid, nid))
1606                                 rc = -EEXIST;
1607                         else if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL)
1608                                 rc = -EPERM;
1609                         goto out;
1610                 } else if (!(flags & LNET_PEER_CONFIGURED)) {
1611                         if (nid_same(&lp->lp_primary_nid, nid)) {
1612                                 rc = -EEXIST;
1613                                 goto out;
1614                         }
1615                 }
1616                 /* Delete and recreate as a configured peer. */
1617                 rc = lnet_peer_del(lp);
1618                 if (rc)
1619                         goto out;
1620         }
1621
1622         /* Create peer, peer_net, and peer_ni. */
1623         rc = -ENOMEM;
1624         lp = lnet_peer_alloc(nid);
1625         if (!lp)
1626                 goto out;
1627         lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1628         if (!lpn)
1629                 goto out_free_lp;
1630         lpni = lnet_peer_ni_alloc(nid);
1631         if (!lpni)
1632                 goto out_free_lpn;
1633
1634         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1635
1636 out_free_lpn:
1637         LIBCFS_FREE(lpn, sizeof(*lpn));
1638 out_free_lp:
1639         LIBCFS_FREE(lp, sizeof(*lp));
1640 out:
1641         CDEBUG(D_NET, "peer %s NID flags %#x: %d\n",
1642                libcfs_nidstr(nid), flags, rc);
1643         return rc;
1644 }
1645
1646 /*
1647  * Add a NID to a peer. Call with ln_api_mutex held.
1648  *
1649  * Error codes:
1650  *  -EPERM:    Non-DLC addition to a DLC-configured peer.
1651  *  -EEXIST:   The NID was configured by DLC for a different peer.
1652  *  -ENOMEM:   Out of memory.
1653  *  -ENOTUNIQ: Adding a second peer NID on a single network on a
1654  *             non-multi-rail peer.
1655  */
1656 static int
1657 lnet_peer_add_nid(struct lnet_peer *lp, struct lnet_nid *nid,
1658                   unsigned int flags)
1659 {
1660         struct lnet_peer_net *lpn;
1661         struct lnet_peer_ni *lpni;
1662         int rc = 0;
1663
1664         LASSERT(lp);
1665         LASSERT(nid);
1666
1667         /* A configured peer can only be updated through configuration. */
1668         if (!(flags & LNET_PEER_CONFIGURED)) {
1669                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
1670                         rc = -EPERM;
1671                         goto out;
1672                 }
1673         }
1674
1675         /*
1676          * The MULTI_RAIL flag can be set but not cleared, because
1677          * that would leave the peer struct in an invalid state.
1678          */
1679         if (flags & LNET_PEER_MULTI_RAIL) {
1680                 spin_lock(&lp->lp_lock);
1681                 if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1682                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
1683                         lnet_peer_clr_non_mr_pref_nids(lp);
1684                 }
1685                 spin_unlock(&lp->lp_lock);
1686         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
1687                 rc = -EPERM;
1688                 goto out;
1689         }
1690
1691         lpni = lnet_peer_ni_find_locked(nid);
1692         if (lpni) {
1693                 /*
1694                  * A peer_ni already exists. This is only a problem if
1695                  * it is not connected to this peer and was configured
1696                  * by DLC.
1697                  */
1698                 if (lpni->lpni_peer_net->lpn_peer == lp)
1699                         goto out_free_lpni;
1700                 if (lnet_peer_ni_is_configured(lpni)) {
1701                         rc = -EEXIST;
1702                         goto out_free_lpni;
1703                 }
1704                 /* If this is the primary NID, destroy the peer. */
1705                 if (lnet_peer_ni_is_primary(lpni)) {
1706                         struct lnet_peer *rtr_lp =
1707                                 lpni->lpni_peer_net->lpn_peer;
1708                         int rtr_refcount = rtr_lp->lp_rtr_refcount;
1709                         /*
1710                          * if we're trying to delete a router it means
1711                          * we're moving this peer NI to a new peer so must
1712                          * transfer router properties to the new peer
1713                          */
1714                         if (rtr_refcount > 0) {
1715                                 flags |= LNET_PEER_RTR_NI_FORCE_DEL;
1716                                 lnet_rtr_transfer_to_peer(rtr_lp, lp);
1717                         }
1718                         lnet_peer_del(lpni->lpni_peer_net->lpn_peer);
1719                         lnet_peer_ni_decref_locked(lpni);
1720                         lpni = lnet_peer_ni_alloc(nid);
1721                         if (!lpni) {
1722                                 rc = -ENOMEM;
1723                                 goto out_free_lpni;
1724                         }
1725                 }
1726         } else {
1727                 lpni = lnet_peer_ni_alloc(nid);
1728                 if (!lpni) {
1729                         rc = -ENOMEM;
1730                         goto out_free_lpni;
1731                 }
1732         }
1733
1734         /*
1735          * Get the peer_net. Check that we're not adding a second
1736          * peer_ni on a peer_net of a non-multi-rail peer.
1737          */
1738         lpn = lnet_peer_get_net_locked(lp, LNET_NID_NET(nid));
1739         if (!lpn) {
1740                 lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1741                 if (!lpn) {
1742                         rc = -ENOMEM;
1743                         goto out_free_lpni;
1744                 }
1745         } else if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
1746                 rc = -ENOTUNIQ;
1747                 goto out_free_lpni;
1748         }
1749
1750         return lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1751
1752 out_free_lpni:
1753         lnet_peer_ni_decref_locked(lpni);
1754 out:
1755         CDEBUG(D_NET, "peer %s NID %s flags %#x: %d\n",
1756                libcfs_nidstr(&lp->lp_primary_nid), libcfs_nidstr(nid),
1757                flags, rc);
1758         return rc;
1759 }
1760
1761 /*
1762  * Update the primary NID of a peer, if possible.
1763  *
1764  * Call with the lnet_api_mutex held.
1765  */
1766 static int
1767 lnet_peer_set_primary_nid(struct lnet_peer *lp, struct lnet_nid *nid,
1768                           unsigned int flags)
1769 {
1770         struct lnet_nid old = lp->lp_primary_nid;
1771         int rc = 0;
1772
1773         if (nid_same(&lp->lp_primary_nid, nid))
1774                 goto out;
1775
1776         lp->lp_primary_nid = *nid;
1777
1778         rc = lnet_peer_add_nid(lp, nid, flags);
1779         if (rc) {
1780                 lp->lp_primary_nid = old;
1781                 goto out;
1782         }
1783 out:
1784         CDEBUG(D_NET, "peer %s NID %s: %d\n",
1785                libcfs_nidstr(&old), libcfs_nidstr(nid), rc);
1786
1787         return rc;
1788 }
1789
1790 /*
1791  * lpni creation initiated due to traffic either sending or receiving.
1792  * Callers must hold ln_api_mutex
1793  * Ref taken on lnet_peer_ni returned by this function
1794  */
1795 static struct lnet_peer_ni *
1796 lnet_peer_ni_traffic_add(struct lnet_nid *nid, struct lnet_nid *pref)
1797 __must_hold(&the_lnet.ln_api_mutex)
1798 {
1799         struct lnet_peer *lp = NULL;
1800         struct lnet_peer_net *lpn = NULL;
1801         struct lnet_peer_ni *lpni;
1802         unsigned flags = 0;
1803         int rc = 0;
1804
1805         if (LNET_NID_IS_ANY(nid)) {
1806                 rc = -EINVAL;
1807                 goto out_err;
1808         }
1809
1810         /* lnet_net_lock is not needed here because ln_api_lock is held */
1811         lpni = lnet_peer_ni_find_locked(nid);
1812         if (lpni) {
1813                 /*
1814                  * We must have raced with another thread. Since we
1815                  * know next to nothing about a peer_ni created by
1816                  * traffic, we just assume everything is ok and
1817                  * return.
1818                  */
1819                 goto out;
1820         }
1821
1822         /* Create peer, peer_net, and peer_ni. */
1823         rc = -ENOMEM;
1824         lp = lnet_peer_alloc(nid);
1825         if (!lp)
1826                 goto out_err;
1827         lpn = lnet_peer_net_alloc(LNET_NID_NET(nid));
1828         if (!lpn)
1829                 goto out_err;
1830         lpni = lnet_peer_ni_alloc(nid);
1831         if (!lpni)
1832                 goto out_err;
1833         lnet_peer_ni_set_non_mr_pref_nid(lpni, pref);
1834
1835         /* lnet_peer_attach_peer_ni() always returns 0 */
1836         rc = lnet_peer_attach_peer_ni(lp, lpn, lpni, flags);
1837
1838         lnet_peer_ni_addref_locked(lpni);
1839
1840 out_err:
1841         if (rc) {
1842                 if (lpn)
1843                         LIBCFS_FREE(lpn, sizeof(*lpn));
1844                 if (lp)
1845                         LIBCFS_FREE(lp, sizeof(*lp));
1846                 lpni = ERR_PTR(rc);
1847         }
1848 out:
1849         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(nid), rc);
1850         return lpni;
1851 }
1852
1853 /*
1854  * Implementation of IOC_LIBCFS_ADD_PEER_NI.
1855  *
1856  * This API handles the following combinations:
1857  *   Create a peer with its primary NI if only the prim_nid is provided
1858  *   Add a NID to a peer identified by the prim_nid. The peer identified
1859  *   by the prim_nid must already exist.
1860  *   The peer being created may be non-MR.
1861  *
1862  * The caller must hold ln_api_mutex. This prevents the peer from
1863  * being created/modified/deleted by a different thread.
1864  */
1865 int
1866 lnet_add_peer_ni(struct lnet_nid *prim_nid, struct lnet_nid *nid, bool mr,
1867                  bool temp)
1868 __must_hold(&the_lnet.ln_api_mutex)
1869 {
1870         struct lnet_peer *lp = NULL;
1871         struct lnet_peer_ni *lpni;
1872         unsigned int flags = 0;
1873
1874         /* The prim_nid must always be specified */
1875         if (LNET_NID_IS_ANY(prim_nid))
1876                 return -EINVAL;
1877
1878         if (!temp)
1879                 flags = LNET_PEER_CONFIGURED;
1880
1881         if (mr)
1882                 flags |= LNET_PEER_MULTI_RAIL;
1883
1884         /*
1885          * If nid isn't specified, we must create a new peer with
1886          * prim_nid as its primary nid.
1887          */
1888         if (LNET_NID_IS_ANY(nid))
1889                 return lnet_peer_add(prim_nid, flags);
1890
1891         /* Look up the prim_nid, which must exist. */
1892         lpni = lnet_peer_ni_find_locked(prim_nid);
1893         if (!lpni)
1894                 return -ENOENT;
1895         lnet_peer_ni_decref_locked(lpni);
1896         lp = lpni->lpni_peer_net->lpn_peer;
1897
1898         /* Peer must have been configured. */
1899         if (!temp && !(lp->lp_state & LNET_PEER_CONFIGURED)) {
1900                 CDEBUG(D_NET, "peer %s was not configured\n",
1901                        libcfs_nidstr(prim_nid));
1902                 return -ENOENT;
1903         }
1904
1905         /* Primary NID must match */
1906         if (!nid_same(&lp->lp_primary_nid, prim_nid)) {
1907                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1908                        libcfs_nidstr(prim_nid),
1909                        libcfs_nidstr(&lp->lp_primary_nid));
1910                 return -ENODEV;
1911         }
1912
1913         /* Multi-Rail flag must match. */
1914         if ((lp->lp_state ^ flags) & LNET_PEER_MULTI_RAIL) {
1915                 CDEBUG(D_NET, "multi-rail state mismatch for peer %s\n",
1916                        libcfs_nidstr(prim_nid));
1917                 return -EPERM;
1918         }
1919
1920         if (temp && lnet_peer_is_uptodate(lp)) {
1921                 CDEBUG(D_NET,
1922                        "Don't add temporary peer NI for uptodate peer %s\n",
1923                        libcfs_nidstr(&lp->lp_primary_nid));
1924                 return -EINVAL;
1925         }
1926
1927         return lnet_peer_add_nid(lp, nid, flags);
1928 }
1929
1930 /*
1931  * Implementation of IOC_LIBCFS_DEL_PEER_NI.
1932  *
1933  * This API handles the following combinations:
1934  *   Delete a NI from a peer if both prim_nid and nid are provided.
1935  *   Delete a peer if only prim_nid is provided.
1936  *   Delete a peer if its primary nid is provided.
1937  *
1938  * The caller must hold ln_api_mutex. This prevents the peer from
1939  * being modified/deleted by a different thread.
1940  */
1941 int
1942 lnet_del_peer_ni(struct lnet_nid *prim_nid, struct lnet_nid *nid)
1943 {
1944         struct lnet_peer *lp;
1945         struct lnet_peer_ni *lpni;
1946         unsigned int flags;
1947
1948         if (!prim_nid || LNET_NID_IS_ANY(prim_nid))
1949                 return -EINVAL;
1950
1951         lpni = lnet_peer_ni_find_locked(prim_nid);
1952         if (!lpni)
1953                 return -ENOENT;
1954         lnet_peer_ni_decref_locked(lpni);
1955         lp = lpni->lpni_peer_net->lpn_peer;
1956
1957         if (!nid_same(prim_nid, &lp->lp_primary_nid)) {
1958                 CDEBUG(D_NET, "prim_nid %s is not primary for peer %s\n",
1959                        libcfs_nidstr(prim_nid),
1960                        libcfs_nidstr(&lp->lp_primary_nid));
1961                 return -ENODEV;
1962         }
1963
1964         lnet_net_lock(LNET_LOCK_EX);
1965         if (lp->lp_rtr_refcount > 0) {
1966                 lnet_net_unlock(LNET_LOCK_EX);
1967                 CERROR("%s is a router. Can not be deleted\n",
1968                        libcfs_nidstr(prim_nid));
1969                 return -EBUSY;
1970         }
1971         lnet_net_unlock(LNET_LOCK_EX);
1972
1973         if (LNET_NID_IS_ANY(nid) || nid_same(nid, &lp->lp_primary_nid))
1974                 return lnet_peer_del(lp);
1975
1976         flags = LNET_PEER_CONFIGURED;
1977         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
1978                 flags |= LNET_PEER_MULTI_RAIL;
1979
1980         return lnet_peer_del_nid(lp, nid, flags);
1981 }
1982
1983 void
1984 lnet_destroy_peer_ni_locked(struct kref *ref)
1985 {
1986         struct lnet_peer_ni *lpni = container_of(ref, struct lnet_peer_ni,
1987                                                  lpni_kref);
1988         struct lnet_peer_table *ptable;
1989         struct lnet_peer_net *lpn;
1990
1991         CDEBUG(D_NET, "%p nid %s\n", lpni, libcfs_nidstr(&lpni->lpni_nid));
1992
1993         LASSERT(kref_read(&lpni->lpni_kref) == 0);
1994         LASSERT(list_empty(&lpni->lpni_txq));
1995         LASSERT(lpni->lpni_txqnob == 0);
1996         LASSERT(list_empty(&lpni->lpni_peer_nis));
1997         LASSERT(list_empty(&lpni->lpni_on_remote_peer_ni_list));
1998
1999         lpn = lpni->lpni_peer_net;
2000         lpni->lpni_peer_net = NULL;
2001         lpni->lpni_net = NULL;
2002
2003         if (!list_empty(&lpni->lpni_hashlist)) {
2004                 /* remove the peer ni from the zombie list */
2005                 ptable = the_lnet.ln_peer_tables[lpni->lpni_cpt];
2006                 spin_lock(&ptable->pt_zombie_lock);
2007                 list_del_init(&lpni->lpni_hashlist);
2008                 ptable->pt_zombies--;
2009                 spin_unlock(&ptable->pt_zombie_lock);
2010         }
2011
2012         if (lpni->lpni_pref_nnids > 1) {
2013                 struct lnet_nid_list *ne, *tmp;
2014
2015                 list_for_each_entry_safe(ne, tmp, &lpni->lpni_pref.nids,
2016                                          nl_list) {
2017                         list_del_init(&ne->nl_list);
2018                         LIBCFS_FREE(ne, sizeof(*ne));
2019                 }
2020         }
2021         LIBCFS_FREE(lpni, sizeof(*lpni));
2022
2023         if (lpn)
2024                 lnet_peer_net_decref_locked(lpn);
2025 }
2026
2027 struct lnet_peer_ni *
2028 lnet_nid2peerni_ex(struct lnet_nid *nid)
2029 __must_hold(&the_lnet.ln_api_mutex)
2030 {
2031         struct lnet_peer_ni *lpni = NULL;
2032
2033         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2034                 return ERR_PTR(-ESHUTDOWN);
2035
2036         /*
2037          * find if a peer_ni already exists.
2038          * If so then just return that.
2039          */
2040         lpni = lnet_peer_ni_find_locked(nid);
2041         if (lpni)
2042                 return lpni;
2043
2044         lnet_net_unlock(LNET_LOCK_EX);
2045
2046         lpni = lnet_peer_ni_traffic_add(nid, NULL);
2047
2048         lnet_net_lock(LNET_LOCK_EX);
2049
2050         return lpni;
2051 }
2052
2053 /*
2054  * Get a peer_ni for the given nid, create it if necessary. Takes a
2055  * hold on the peer_ni.
2056  */
2057 struct lnet_peer_ni *
2058 lnet_peerni_by_nid_locked(struct lnet_nid *nid,
2059                         struct lnet_nid *pref, int cpt)
2060 {
2061         struct lnet_peer_ni *lpni = NULL;
2062
2063         if (the_lnet.ln_state != LNET_STATE_RUNNING)
2064                 return ERR_PTR(-ESHUTDOWN);
2065
2066         /*
2067          * find if a peer_ni already exists.
2068          * If so then just return that.
2069          */
2070         lpni = lnet_peer_ni_find_locked(nid);
2071         if (lpni)
2072                 return lpni;
2073
2074         /*
2075          * Slow path:
2076          * use the lnet_api_mutex to serialize the creation of the peer_ni
2077          * and the creation/deletion of the local ni/net. When a local ni is
2078          * created, if there exists a set of peer_nis on that network,
2079          * they need to be traversed and updated. When a local NI is
2080          * deleted, which could result in a network being deleted, then
2081          * all peer nis on that network need to be removed as well.
2082          *
2083          * Creation through traffic should also be serialized with
2084          * creation through DLC.
2085          */
2086         lnet_net_unlock(cpt);
2087         mutex_lock(&the_lnet.ln_api_mutex);
2088         /*
2089          * the_lnet.ln_state is only modified under the ln_api_lock, so a single
2090          * check here is sufficent.
2091          */
2092         if (the_lnet.ln_state == LNET_STATE_RUNNING)
2093                 lpni = lnet_peer_ni_traffic_add(nid, pref);
2094
2095         mutex_unlock(&the_lnet.ln_api_mutex);
2096         lnet_net_lock(cpt);
2097
2098         /* Lock has been dropped, check again for shutdown. */
2099         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
2100                 if (!IS_ERR_OR_NULL(lpni))
2101                         lnet_peer_ni_decref_locked(lpni);
2102                 lpni = ERR_PTR(-ESHUTDOWN);
2103         }
2104
2105         return lpni;
2106 }
2107
2108 bool
2109 lnet_peer_gw_discovery(struct lnet_peer *lp)
2110 {
2111         bool rc = false;
2112
2113         spin_lock(&lp->lp_lock);
2114         if (lp->lp_state & LNET_PEER_RTR_DISCOVERY)
2115                 rc = true;
2116         spin_unlock(&lp->lp_lock);
2117
2118         return rc;
2119 }
2120
2121 bool
2122 lnet_peer_is_uptodate(struct lnet_peer *lp)
2123 {
2124         bool rc;
2125
2126         spin_lock(&lp->lp_lock);
2127         rc = lnet_peer_is_uptodate_locked(lp);
2128         spin_unlock(&lp->lp_lock);
2129         return rc;
2130 }
2131
2132 /*
2133  * Is a peer uptodate from the point of view of discovery?
2134  *
2135  * If it is currently being processed, obviously not.
2136  * A forced Ping or Push is also handled by the discovery thread.
2137  *
2138  * Otherwise look at whether the peer needs rediscovering.
2139  */
2140 bool
2141 lnet_peer_is_uptodate_locked(struct lnet_peer *lp)
2142 __must_hold(&lp->lp_lock)
2143 {
2144         bool rc;
2145
2146         if (lp->lp_state & (LNET_PEER_DISCOVERING |
2147                             LNET_PEER_FORCE_PING |
2148                             LNET_PEER_FORCE_PUSH)) {
2149                 rc = false;
2150         } else if (lp->lp_state & LNET_PEER_REDISCOVER) {
2151                 rc = false;
2152         } else if (lnet_peer_needs_push(lp)) {
2153                 rc = false;
2154         } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
2155                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
2156                         rc = true;
2157                 else
2158                         rc = false;
2159         } else {
2160                 rc = false;
2161         }
2162
2163         return rc;
2164 }
2165
2166 /* Add the message to the peer's lp_dc_pendq and queue the peer for discovery */
2167 void
2168 lnet_peer_queue_message(struct lnet_peer *lp, struct lnet_msg *msg)
2169 {
2170         /* The discovery thread holds net_lock/EX and lp_lock when it splices
2171          * the lp_dc_pendq onto a local list for resending. Thus, we do the same
2172          * when adding to the list and queuing the peer to ensure that we do not
2173          * strand any messages on the lp_dc_pendq. This scheme ensures the
2174          * message will be resent even if the peer is already being discovered.
2175          * Therefore we needn't check the return value of
2176          * lnet_peer_queue_for_discovery(lp).
2177          */
2178         lnet_net_lock(LNET_LOCK_EX);
2179         spin_lock(&lp->lp_lock);
2180         list_add_tail(&msg->msg_list, &lp->lp_dc_pendq);
2181         spin_unlock(&lp->lp_lock);
2182         lnet_peer_queue_for_discovery(lp);
2183         lnet_net_unlock(LNET_LOCK_EX);
2184 }
2185
2186 /*
2187  * Queue a peer for the attention of the discovery thread.  Call with
2188  * lnet_net_lock/EX held. Returns 0 if the peer was queued, and
2189  * -EALREADY if the peer was already queued.
2190  */
2191 static int lnet_peer_queue_for_discovery(struct lnet_peer *lp)
2192 {
2193         int rc;
2194
2195         spin_lock(&lp->lp_lock);
2196         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2197                 lp->lp_state |= LNET_PEER_DISCOVERING;
2198         spin_unlock(&lp->lp_lock);
2199         if (list_empty(&lp->lp_dc_list)) {
2200                 lnet_peer_addref_locked(lp);
2201                 list_add_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2202                 wake_up(&the_lnet.ln_dc_waitq);
2203                 rc = 0;
2204         } else {
2205                 rc = -EALREADY;
2206         }
2207
2208         CDEBUG(D_NET, "Queue peer %s: %d\n",
2209                libcfs_nidstr(&lp->lp_primary_nid), rc);
2210
2211         return rc;
2212 }
2213
2214 /*
2215  * Discovery of a peer is complete. Wake all waiters on the peer.
2216  * Call with lnet_net_lock/EX held.
2217  */
2218 static void lnet_peer_discovery_complete(struct lnet_peer *lp, int dc_error)
2219 {
2220         struct lnet_msg *msg, *tmp;
2221         int rc = 0;
2222         LIST_HEAD(pending_msgs);
2223
2224         CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
2225                libcfs_nidstr(&lp->lp_primary_nid));
2226
2227         spin_lock(&lp->lp_lock);
2228         /* Our caller dropped lp_lock which may have allowed another thread to
2229          * set LNET_PEER_DISCOVERING, or it may be set if dc_error is non-zero.
2230          * Ensure it is cleared.
2231          */
2232         lp->lp_state &= ~LNET_PEER_DISCOVERING;
2233         if (dc_error) {
2234                 lp->lp_dc_error = dc_error;
2235                 lp->lp_state |= LNET_PEER_REDISCOVER;
2236         }
2237         list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
2238         spin_unlock(&lp->lp_lock);
2239         list_del_init(&lp->lp_dc_list);
2240         wake_up(&lp->lp_dc_waitq);
2241
2242         if (lp->lp_rtr_refcount > 0)
2243                 lnet_router_discovery_complete(lp);
2244
2245         lnet_net_unlock(LNET_LOCK_EX);
2246
2247         /* iterate through all pending messages and send them again */
2248         list_for_each_entry_safe(msg, tmp, &pending_msgs, msg_list) {
2249                 list_del_init(&msg->msg_list);
2250                 if (dc_error) {
2251                         lnet_finalize(msg, dc_error);
2252                         continue;
2253                 }
2254
2255                 CDEBUG(D_NET, "sending pending message %s to target %s\n",
2256                        lnet_msgtyp2str(msg->msg_type),
2257                        libcfs_idstr(&msg->msg_target));
2258                 rc = lnet_send(&msg->msg_src_nid_param, msg,
2259                                &msg->msg_rtr_nid_param);
2260                 if (rc < 0) {
2261                         CNETERR("Error sending %s to %s: %d\n",
2262                                lnet_msgtyp2str(msg->msg_type),
2263                                libcfs_idstr(&msg->msg_target), rc);
2264                         lnet_finalize(msg, rc);
2265                 }
2266         }
2267         lnet_net_lock(LNET_LOCK_EX);
2268         lnet_peer_decref_locked(lp);
2269 }
2270
2271 /*
2272  * Handle inbound push.
2273  * Like any event handler, called with lnet_res_lock/CPT held.
2274  */
2275 void lnet_peer_push_event(struct lnet_event *ev)
2276 {
2277         struct lnet_ping_buffer *pbuf;
2278         struct lnet_peer *lp;
2279         int infobytes;
2280
2281         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start + ev->offset);
2282
2283         /* lnet_find_peer() adds a refcount */
2284         lp = lnet_find_peer(&ev->source.nid);
2285         if (!lp) {
2286                 CDEBUG(D_NET, "Push Put from unknown %s (source %s). Ignoring...\n",
2287                        libcfs_nidstr(&ev->initiator.nid),
2288                        libcfs_nidstr(&ev->source.nid));
2289                 pbuf->pb_needs_post = true;
2290                 return;
2291         }
2292
2293         /* Ensure peer state remains consistent while we modify it. */
2294         spin_lock(&lp->lp_lock);
2295
2296         /*
2297          * If some kind of error happened the contents of the message
2298          * cannot be used. Clear the NIDS_UPTODATE and set the
2299          * FORCE_PING flag to trigger a ping.
2300          */
2301         if (ev->status) {
2302                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2303                 lp->lp_state |= LNET_PEER_FORCE_PING;
2304                 CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
2305                        ev->status,
2306                        libcfs_nidstr(&lp->lp_primary_nid),
2307                        libcfs_nidstr(&ev->source.nid));
2308                 goto out;
2309         }
2310
2311         /*
2312          * A push with invalid or corrupted info. Clear the UPTODATE
2313          * flag to trigger a ping.
2314          */
2315         if (lnet_ping_info_validate(&pbuf->pb_info)) {
2316                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2317                 lp->lp_state |= LNET_PEER_FORCE_PING;
2318                 CDEBUG(D_NET, "Corrupted Push from %s\n",
2319                        libcfs_nidstr(&lp->lp_primary_nid));
2320                 goto out;
2321         }
2322
2323         /* Make sure we'll allocate the correct size ping buffer when
2324          * pinging the peer.
2325          */
2326         infobytes = lnet_ping_info_size(&pbuf->pb_info);
2327         if (lp->lp_data_bytes < infobytes)
2328                 lp->lp_data_bytes = infobytes;
2329
2330         /*
2331          * A non-Multi-Rail peer is not supposed to be capable of
2332          * sending a push.
2333          */
2334         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
2335                 CERROR("Push from non-Multi-Rail peer %s dropped\n",
2336                        libcfs_nidstr(&lp->lp_primary_nid));
2337                 goto out;
2338         }
2339
2340         /*
2341          * The peer may have discovery disabled at its end. Set
2342          * NO_DISCOVERY as appropriate.
2343          */
2344         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
2345                 CDEBUG(D_NET, "Peer %s has discovery disabled\n",
2346                        libcfs_nidstr(&lp->lp_primary_nid));
2347                 /*
2348                  * Mark the peer for deletion if we already know about it
2349                  * and it's going from discovery set to no discovery set
2350                  */
2351                 if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
2352                                       LNET_PEER_DISCOVERING)) &&
2353                      lp->lp_state & LNET_PEER_DISCOVERED) {
2354                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2355                                libcfs_nidstr(&lp->lp_primary_nid),
2356                                lp->lp_state);
2357                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2358                 }
2359                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2360         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2361                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2362                        libcfs_nidstr(&lp->lp_primary_nid));
2363                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2364         }
2365
2366         /*
2367          * Update the MULTI_RAIL flag based on the push. If the peer
2368          * was configured with DLC then the setting should match what
2369          * DLC put in.
2370          * NB: We verified above that the MR feature bit is set in pi_features
2371          */
2372         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2373                 CDEBUG(D_NET, "peer %s(%p) is MR\n",
2374                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2375         } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2376                 CWARN("Push says %s is Multi-Rail, DLC says not\n",
2377                       libcfs_nidstr(&lp->lp_primary_nid));
2378         } else if (lnet_peer_discovery_disabled) {
2379                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled locally\n",
2380                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2381         } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2382                 CDEBUG(D_NET, "peer %s(%p) not MR: DD disabled remotely\n",
2383                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2384         } else {
2385                 CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2386                        libcfs_nidstr(&lp->lp_primary_nid), lp);
2387                 lp->lp_state |= LNET_PEER_MULTI_RAIL;
2388                 lnet_peer_clr_non_mr_pref_nids(lp);
2389         }
2390
2391         /* Check for truncation of the Put message. Clear the
2392          * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
2393          * and tell discovery to allocate a bigger buffer.
2394          */
2395         if (ev->mlength < ev->rlength) {
2396                 if (the_lnet.ln_push_target_nbytes < infobytes)
2397                         the_lnet.ln_push_target_nbytes = infobytes;
2398                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2399                 lp->lp_state |= LNET_PEER_FORCE_PING;
2400                 CDEBUG(D_NET, "Truncated Push from %s (%d bytes)\n",
2401                        libcfs_nidstr(&lp->lp_primary_nid), infobytes);
2402                 goto out;
2403         }
2404
2405         /* always assume new data */
2406         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2407         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
2408
2409         /* If there is data present that hasn't been processed yet,
2410          * we'll replace it if the Put contained newer data and it
2411          * fits. We're racing with a Ping or earlier Push in this
2412          * case.
2413          */
2414         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
2415                 if (LNET_PING_BUFFER_SEQNO(pbuf) >
2416                         LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
2417                     infobytes <= lp->lp_data->pb_nbytes) {
2418                         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
2419                                infobytes);
2420                         CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
2421                               libcfs_nidstr(&lp->lp_primary_nid),
2422                               LNET_PING_BUFFER_SEQNO(pbuf),
2423                               LNET_PING_BUFFER_SEQNO(lp->lp_data));
2424                 }
2425                 goto out;
2426         }
2427
2428         /*
2429          * Allocate a buffer to copy the data. On a failure we drop
2430          * the Push and set FORCE_PING to force the discovery
2431          * thread to fix the problem by pinging the peer.
2432          */
2433         lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_bytes, GFP_ATOMIC);
2434         if (!lp->lp_data) {
2435                 lp->lp_state |= LNET_PEER_FORCE_PING;
2436                 CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
2437                        libcfs_nidstr(&lp->lp_primary_nid),
2438                        LNET_PING_BUFFER_SEQNO(pbuf));
2439                 goto out;
2440         }
2441
2442         /* Success */
2443         memcpy(&lp->lp_data->pb_info, &pbuf->pb_info, infobytes);
2444         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2445         CDEBUG(D_NET, "Received Push %s %u\n",
2446                libcfs_nidstr(&lp->lp_primary_nid),
2447                LNET_PING_BUFFER_SEQNO(pbuf));
2448
2449 out:
2450         /* We've processed this buffer. It can be reposted */
2451         pbuf->pb_needs_post = true;
2452
2453         /*
2454          * Queue the peer for discovery if not done, force it on the request
2455          * queue and wake the discovery thread if the peer was already queued,
2456          * because its status changed.
2457          */
2458         spin_unlock(&lp->lp_lock);
2459         lnet_net_lock(LNET_LOCK_EX);
2460         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2461                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2462                 wake_up(&the_lnet.ln_dc_waitq);
2463         }
2464         /* Drop refcount from lookup */
2465         lnet_peer_decref_locked(lp);
2466         lnet_net_unlock(LNET_LOCK_EX);
2467 }
2468
2469 /*
2470  * Clear the discovery error state, unless we're already discovering
2471  * this peer, in which case the error is current.
2472  */
2473 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2474 {
2475         spin_lock(&lp->lp_lock);
2476         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2477                 lp->lp_dc_error = 0;
2478         spin_unlock(&lp->lp_lock);
2479 }
2480
2481 /*
2482  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2483  * dropped/retaken within this function. An lnet_peer_ni is passed in
2484  * because discovery could tear down an lnet_peer.
2485  */
2486 int
2487 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2488 {
2489         DEFINE_WAIT(wait);
2490         struct lnet_peer *lp;
2491         int rc = 0;
2492         int count = 0;
2493
2494 again:
2495         lnet_net_unlock(cpt);
2496         lnet_net_lock(LNET_LOCK_EX);
2497         lp = lpni->lpni_peer_net->lpn_peer;
2498         lnet_peer_clear_discovery_error(lp);
2499
2500         /*
2501          * We're willing to be interrupted. The lpni can become a
2502          * zombie if we race with DLC, so we must check for that.
2503          */
2504         for (;;) {
2505                 /* Keep lp alive when the lnet_net_lock is unlocked */
2506                 lnet_peer_addref_locked(lp);
2507                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2508                 if (signal_pending(current))
2509                         break;
2510                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2511                         break;
2512                 /*
2513                  * Don't repeat discovery if discovery is disabled. This is
2514                  * done to ensure we can use discovery as a standard ping as
2515                  * well for backwards compatibility with routers which do not
2516                  * have discovery or have discovery disabled
2517                  */
2518                 if (lnet_is_discovery_disabled(lp) && count > 0)
2519                         break;
2520                 if (lp->lp_dc_error)
2521                         break;
2522                 if (lnet_peer_is_uptodate(lp))
2523                         break;
2524                 if (lp->lp_state & LNET_PEER_MARK_DELETED)
2525                         break;
2526                 lnet_peer_queue_for_discovery(lp);
2527                 count++;
2528                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2529
2530                 /*
2531                  * If caller requested a non-blocking operation then
2532                  * return immediately. Once discovery is complete any
2533                  * pending messages that were stopped due to discovery
2534                  * will be transmitted.
2535                  */
2536                 if (!block)
2537                         break;
2538
2539                 lnet_net_unlock(LNET_LOCK_EX);
2540                 schedule();
2541                 finish_wait(&lp->lp_dc_waitq, &wait);
2542                 lnet_net_lock(LNET_LOCK_EX);
2543                 lnet_peer_decref_locked(lp);
2544                 /* Peer may have changed */
2545                 lp = lpni->lpni_peer_net->lpn_peer;
2546         }
2547         finish_wait(&lp->lp_dc_waitq, &wait);
2548
2549         lnet_net_unlock(LNET_LOCK_EX);
2550         lnet_net_lock(cpt);
2551         lnet_peer_decref_locked(lp);
2552         /*
2553          * The peer may have changed, so re-check and rediscover if that turns
2554          * out to have been the case. The reference count on lp ensured that
2555          * even if it was unlinked from lpni the memory could not be recycled.
2556          * Thus the check below is sufficient to determine whether the peer
2557          * changed. If the peer changed, then lp must not be dereferenced.
2558          */
2559         if (lp != lpni->lpni_peer_net->lpn_peer)
2560                 goto again;
2561
2562         if (signal_pending(current))
2563                 rc = -EINTR;
2564         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2565                 rc = -ESHUTDOWN;
2566         else if (lp->lp_dc_error)
2567                 rc = lp->lp_dc_error;
2568         else if (!block)
2569                 CDEBUG(D_NET, "non-blocking discovery\n");
2570         else if (!lnet_peer_is_uptodate(lp) &&
2571                  !(lnet_is_discovery_disabled(lp) ||
2572                    (lp->lp_state & LNET_PEER_MARK_DELETED)))
2573                 goto again;
2574
2575         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2576                (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2577                libcfs_nidstr(&lpni->lpni_nid), rc,
2578                (!block) ? "pending discovery" : "discovery complete");
2579
2580         return rc;
2581 }
2582
2583 /* Handle an incoming ack for a push. */
2584 static void
2585 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2586 {
2587         struct lnet_ping_buffer *pbuf;
2588
2589         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2590         spin_lock(&lp->lp_lock);
2591         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2592         lp->lp_push_error = ev->status;
2593         if (ev->status)
2594                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2595         else
2596                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2597         spin_unlock(&lp->lp_lock);
2598
2599         CDEBUG(D_NET, "peer %s ev->status %d\n",
2600                libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2601 }
2602
2603 /* Handle a Reply message. This is the reply to a Ping message. */
2604 static void
2605 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2606 {
2607         struct lnet_ping_buffer *pbuf;
2608         int infobytes;
2609         int rc;
2610         bool ping_feat_disc;
2611
2612         spin_lock(&lp->lp_lock);
2613
2614         lp->lp_disc_src_nid = ev->target.nid;
2615         lp->lp_disc_dst_nid = ev->source.nid;
2616
2617         /*
2618          * If some kind of error happened the contents of message
2619          * cannot be used. Set PING_FAILED to trigger a retry.
2620          */
2621         if (ev->status) {
2622                 lp->lp_state |= LNET_PEER_PING_FAILED;
2623                 lp->lp_ping_error = ev->status;
2624                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2625                        ev->status,
2626                        libcfs_nidstr(&lp->lp_primary_nid),
2627                        libcfs_nidstr(&ev->source.nid));
2628                 goto out;
2629         }
2630
2631         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2632         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2633                 lnet_swap_pinginfo(pbuf);
2634
2635         /*
2636          * A reply with invalid or corrupted info. Set PING_FAILED to
2637          * trigger a retry.
2638          */
2639         rc = lnet_ping_info_validate(&pbuf->pb_info);
2640         if (rc) {
2641                 lp->lp_state |= LNET_PEER_PING_FAILED;
2642                 lp->lp_ping_error = 0;
2643                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2644                        libcfs_nidstr(&lp->lp_primary_nid), rc);
2645                 goto out;
2646         }
2647
2648         /*
2649          * The peer may have discovery disabled at its end. Set
2650          * NO_DISCOVERY as appropriate.
2651          */
2652         ping_feat_disc = pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY;
2653         if (!ping_feat_disc || lnet_peer_discovery_disabled) {
2654                 CDEBUG(D_NET, "Peer %s has discovery %s, local discovery %s\n",
2655                        libcfs_nidstr(&lp->lp_primary_nid),
2656                        ping_feat_disc ? "enabled" : "disabled",
2657                        lnet_peer_discovery_disabled ? "disabled" : "enabled");
2658
2659                 /* Detect whether this peer has toggled discovery from on to
2660                  * off and whether we can delete and re-create the peer. Peers
2661                  * that were manually configured cannot be deleted by discovery.
2662                  * We need to delete this peer and re-create it if the peer was
2663                  * not configured manually, is currently considered DD capable,
2664                  * and either:
2665                  * 1. We've already discovered the peer (the peer has toggled
2666                  *    the discovery feature from on to off), or
2667                  * 2. The peer is considered MR, but it was not user configured
2668                  *    (this was a "temporary" peer created via the kernel APIs
2669                  *     that we're discovering for the first time)
2670                  */
2671                 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2672                                       LNET_PEER_NO_DISCOVERY)) &&
2673                     (lp->lp_state & (LNET_PEER_DISCOVERED |
2674                                      LNET_PEER_MULTI_RAIL))) {
2675                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2676                                libcfs_nidstr(&lp->lp_primary_nid),
2677                                lp->lp_state);
2678                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2679                 }
2680                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2681         } else {
2682                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2683                        libcfs_nidstr(&lp->lp_primary_nid));
2684                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2685         }
2686
2687         /*
2688          * Update the MULTI_RAIL flag based on the reply. If the peer
2689          * was configured with DLC then the setting should match what
2690          * DLC put in.
2691          */
2692         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2693                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2694                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2695                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2696                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2697                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2698                               libcfs_nidstr(&lp->lp_primary_nid));
2699                 } else if (lnet_peer_discovery_disabled) {
2700                         CDEBUG(D_NET,
2701                                "peer %s(%p) not MR: DD disabled locally\n",
2702                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2703                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2704                         CDEBUG(D_NET,
2705                                "peer %s(%p) not MR: DD disabled remotely\n",
2706                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2707                 } else {
2708                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2709                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2710                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2711                         lnet_peer_clr_non_mr_pref_nids(lp);
2712                 }
2713         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2714                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2715                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2716                               libcfs_nidstr(&lp->lp_primary_nid));
2717                 } else {
2718                         CERROR("Multi-Rail state vanished from %s\n",
2719                                libcfs_nidstr(&lp->lp_primary_nid));
2720                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2721                 }
2722         }
2723
2724         infobytes = lnet_ping_info_size(&pbuf->pb_info);
2725         /*
2726          * Make sure we'll allocate the correct size ping buffer when
2727          * pinging the peer.
2728          */
2729         if (lp->lp_data_bytes < infobytes)
2730                 lp->lp_data_bytes = infobytes;
2731
2732         /* Check for truncation of the Reply. Clear PING_SENT and set
2733          * PING_FAILED to trigger a retry.
2734          */
2735         if (pbuf->pb_nbytes < infobytes) {
2736                 if (the_lnet.ln_push_target_nbytes < infobytes)
2737                         the_lnet.ln_push_target_nbytes = infobytes;
2738                 lp->lp_state |= LNET_PEER_PING_FAILED;
2739                 lp->lp_ping_error = 0;
2740                 CDEBUG(D_NET, "Truncated Reply from %s (%d bytes)\n",
2741                        libcfs_nidstr(&lp->lp_primary_nid), infobytes);
2742                 goto out;
2743         }
2744
2745         /*
2746          * Check the sequence numbers in the reply. These are only
2747          * available if the reply came from a Multi-Rail peer.
2748          */
2749         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2750             pbuf->pb_info.pi_nnis > 1 &&
2751             lnet_nid_to_nid4(&lp->lp_primary_nid) ==
2752             pbuf->pb_info.pi_ni[1].ns_nid) {
2753                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2754                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2755                                 libcfs_nidstr(&lp->lp_primary_nid),
2756                                 LNET_PING_BUFFER_SEQNO(pbuf),
2757                                 lp->lp_peer_seqno);
2758
2759                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2760         }
2761
2762         /* We're happy with the state of the data in the buffer. */
2763         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2764                libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2765                lp->lp_state);
2766         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2767                 lnet_ping_buffer_decref(lp->lp_data);
2768         else
2769                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2770         lnet_ping_buffer_addref(pbuf);
2771         lp->lp_data = pbuf;
2772 out:
2773         lp->lp_state &= ~LNET_PEER_PING_SENT;
2774         spin_unlock(&lp->lp_lock);
2775 }
2776
2777 /*
2778  * Send event handling. Only matters for error cases, where we clean
2779  * up state on the peer and peer_ni that would otherwise be updated in
2780  * the REPLY event handler for a successful Ping, and the ACK event
2781  * handler for a successful Push.
2782  */
2783 static int
2784 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2785 {
2786         int rc = 0;
2787
2788         if (!ev->status)
2789                 goto out;
2790
2791         spin_lock(&lp->lp_lock);
2792         if (ev->msg_type == LNET_MSG_GET) {
2793                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2794                 lp->lp_state |= LNET_PEER_PING_FAILED;
2795                 lp->lp_ping_error = ev->status;
2796         } else { /* ev->msg_type == LNET_MSG_PUT */
2797                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2798                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2799                 lp->lp_push_error = ev->status;
2800         }
2801         spin_unlock(&lp->lp_lock);
2802         rc = LNET_REDISCOVER_PEER;
2803 out:
2804         CDEBUG(D_NET, "%s Send to %s: %d\n",
2805                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2806                 libcfs_nidstr(&ev->target.nid), rc);
2807         return rc;
2808 }
2809
2810 /*
2811  * Unlink event handling. This event is only seen if a call to
2812  * LNetMDUnlink() caused the event to be unlinked. If this call was
2813  * made after the event was set up in LNetGet() or LNetPut() then we
2814  * assume the Ping or Push timed out.
2815  */
2816 static void
2817 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2818 {
2819         spin_lock(&lp->lp_lock);
2820         /* We've passed through LNetGet() */
2821         if (lp->lp_state & LNET_PEER_PING_SENT) {
2822                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2823                 lp->lp_state |= LNET_PEER_PING_FAILED;
2824                 lp->lp_ping_error = -ETIMEDOUT;
2825                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2826                         libcfs_nidstr(&lp->lp_primary_nid));
2827         }
2828         /* We've passed through LNetPut() */
2829         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2830                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2831                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2832                 lp->lp_push_error = -ETIMEDOUT;
2833                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2834                         libcfs_nidstr(&lp->lp_primary_nid));
2835         }
2836         spin_unlock(&lp->lp_lock);
2837 }
2838
2839 /*
2840  * Event handler for the discovery EQ.
2841  *
2842  * Called with lnet_res_lock(cpt) held. The cpt is the
2843  * lnet_cpt_of_cookie() of the md handle cookie.
2844  */
2845 static void lnet_discovery_event_handler(struct lnet_event *event)
2846 {
2847         struct lnet_peer *lp = event->md_user_ptr;
2848         struct lnet_ping_buffer *pbuf;
2849         int rc;
2850
2851         /* discovery needs to take another look */
2852         rc = LNET_REDISCOVER_PEER;
2853
2854         CDEBUG(D_NET, "Received event: %d\n", event->type);
2855
2856         switch (event->type) {
2857         case LNET_EVENT_ACK:
2858                 lnet_discovery_event_ack(lp, event);
2859                 break;
2860         case LNET_EVENT_REPLY:
2861                 lnet_discovery_event_reply(lp, event);
2862                 break;
2863         case LNET_EVENT_SEND:
2864                 /* Only send failure triggers a retry. */
2865                 rc = lnet_discovery_event_send(lp, event);
2866                 break;
2867         case LNET_EVENT_UNLINK:
2868                 /* LNetMDUnlink() was called */
2869                 lnet_discovery_event_unlink(lp, event);
2870                 break;
2871         default:
2872                 /* Invalid events. */
2873                 LBUG();
2874         }
2875         lnet_net_lock(LNET_LOCK_EX);
2876         if (event->unlinked) {
2877                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2878                 lnet_ping_buffer_decref(pbuf);
2879                 lnet_peer_decref_locked(lp);
2880         }
2881
2882         /* put peer back at end of request queue, if discovery not already
2883          * done */
2884         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2885             lnet_peer_queue_for_discovery(lp)) {
2886                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2887                 wake_up(&the_lnet.ln_dc_waitq);
2888         }
2889         lnet_net_unlock(LNET_LOCK_EX);
2890 }
2891
2892 u32 *ping_iter_first(struct lnet_ping_iter *pi,
2893                      struct lnet_ping_buffer *pbuf,
2894                      struct lnet_nid *nid)
2895 {
2896         pi->pinfo = &pbuf->pb_info;
2897         pi->pos = &pbuf->pb_info.pi_ni;
2898         pi->end = (void *)pi->pinfo +
2899                   min_t(int, pbuf->pb_nbytes,
2900                         lnet_ping_info_size(pi->pinfo));
2901         /* lnet_ping_info_validiate ensures there will be one
2902          * lnet_ni_status at the start
2903          */
2904         if (nid)
2905                 lnet_nid4_to_nid(pbuf->pb_info.pi_ni[0].ns_nid, nid);
2906         return &pbuf->pb_info.pi_ni[0].ns_status;
2907 }
2908
2909 u32 *ping_iter_next(struct lnet_ping_iter *pi, struct lnet_nid *nid)
2910 {
2911         int off = offsetof(struct lnet_ping_info, pi_ni[pi->pinfo->pi_nnis]);
2912
2913         if (pi->pos < ((void *)pi->pinfo + off)) {
2914                 struct lnet_ni_status *ns = pi->pos;
2915
2916                 pi->pos = ns + 1;
2917                 if (pi->pos > pi->end)
2918                         return NULL;
2919                 if (nid)
2920                         lnet_nid4_to_nid(ns->ns_nid, nid);
2921                 return &ns->ns_status;
2922         }
2923
2924         while (pi->pinfo->pi_features & LNET_PING_FEAT_LARGE_ADDR) {
2925                 struct lnet_ni_large_status *lns = pi->pos;
2926
2927                 if (pi->pos + 8 > pi->end)
2928                         /* Not safe to examine next */
2929                         return NULL;
2930                 pi->pos = lnet_ping_sts_next(lns);
2931                 if (pi->pos > pi->end)
2932                         return NULL;
2933                 if (NID_BYTES(&lns->ns_nid) > sizeof(struct lnet_nid))
2934                         continue;
2935                 if (nid)
2936                         *nid = lns->ns_nid;
2937                 return &lns->ns_status;
2938         }
2939         return NULL;
2940 }
2941
2942 /*
2943  * Build a peer from incoming data.
2944  *
2945  * The NIDs in the incoming data are supposed to be structured as follows:
2946  *  - loopback
2947  *  - primary NID
2948  *  - other NIDs in same net
2949  *  - NIDs in second net
2950  *  - NIDs in third net
2951  *  - ...
2952  * This due to the way the list of NIDs in the data is created.
2953  *
2954  * Note that this function will mark the peer uptodate unless an
2955  * ENOMEM is encontered. All other errors are due to a conflict
2956  * between the DLC configuration and what discovery sees. We treat DLC
2957  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
2958  * peer from becoming stuck in discovery.
2959  */
2960 static int lnet_peer_merge_data(struct lnet_peer *lp,
2961                                 struct lnet_ping_buffer *pbuf)
2962 {
2963         struct lnet_peer_net *lpn;
2964         struct lnet_peer_ni *lpni;
2965         lnet_nid_t *curnis = NULL;
2966         struct lnet_ni_status *addnis = NULL;
2967         lnet_nid_t *delnis = NULL;
2968         struct lnet_nid nid;
2969         unsigned int flags;
2970         int ncurnis;
2971         int naddnis;
2972         int ndelnis;
2973         int nnis = 0;
2974         int i;
2975         int j;
2976         int rc;
2977
2978         flags = LNET_PEER_DISCOVERED;
2979         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
2980                 flags |= LNET_PEER_MULTI_RAIL;
2981
2982         /*
2983          * Cache the routing feature for the peer; whether it is enabled
2984          * for disabled as reported by the remote peer.
2985          */
2986         spin_lock(&lp->lp_lock);
2987         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
2988                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
2989         else
2990                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
2991         spin_unlock(&lp->lp_lock);
2992
2993         nnis = max_t(int, lp->lp_nnis, pbuf->pb_info.pi_nnis);
2994         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
2995         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
2996         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
2997         if (!curnis || !addnis || !delnis) {
2998                 rc = -ENOMEM;
2999                 goto out;
3000         }
3001         ncurnis = 0;
3002         naddnis = 0;
3003         ndelnis = 0;
3004
3005         /* Construct the list of NIDs present in peer. */
3006         lpni = NULL;
3007         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
3008                 curnis[ncurnis++] = lnet_nid_to_nid4(&lpni->lpni_nid);
3009
3010         /*
3011          * Check for NIDs in pbuf not present in curnis[].
3012          * The loop starts at 1 to skip the loopback NID.
3013          */
3014         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
3015                 for (j = 0; j < ncurnis; j++)
3016                         if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
3017                                 break;
3018                 if (j == ncurnis)
3019                         addnis[naddnis++] = pbuf->pb_info.pi_ni[i];
3020         }
3021         /*
3022          * Check for NIDs in curnis[] not present in pbuf.
3023          * The nested loop starts at 1 to skip the loopback NID.
3024          *
3025          * But never add the loopback NID to delnis[]: if it is
3026          * present in curnis[] then this peer is for this node.
3027          */
3028         for (i = 0; i < ncurnis; i++) {
3029                 if (curnis[i] == LNET_NID_LO_0)
3030                         continue;
3031                 for (j = 1; j < pbuf->pb_info.pi_nnis; j++) {
3032                         if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid) {
3033                                 /*
3034                                  * update the information we cache for the
3035                                  * peer with the latest information we
3036                                  * received
3037                                  */
3038                                 lnet_nid4_to_nid(curnis[i], &nid);
3039                                 lpni = lnet_peer_ni_find_locked(&nid);
3040                                 if (lpni) {
3041                                         lpni->lpni_ns_status = pbuf->pb_info.pi_ni[j].ns_status;
3042                                         lnet_peer_ni_decref_locked(lpni);
3043                                 }
3044                                 break;
3045                         }
3046                 }
3047                 if (j == pbuf->pb_info.pi_nnis)
3048                         delnis[ndelnis++] = curnis[i];
3049         }
3050
3051         /*
3052          * If we get here and the discovery is disabled then we don't want
3053          * to add or delete any NIs. We just updated the ones we have some
3054          * information on, and call it a day
3055          */
3056         rc = 0;
3057         if (lnet_is_discovery_disabled(lp))
3058                 goto out;
3059
3060         for (i = 0; i < naddnis; i++) {
3061                 lnet_nid4_to_nid(addnis[i].ns_nid, &nid);
3062                 rc = lnet_peer_add_nid(lp, &nid, flags);
3063                 if (rc) {
3064                         CERROR("Error adding NID %s to peer %s: %d\n",
3065                                libcfs_nid2str(addnis[i].ns_nid),
3066                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3067                         if (rc == -ENOMEM)
3068                                 goto out;
3069                 }
3070                 lpni = lnet_peer_ni_find_locked(&nid);
3071                 if (lpni) {
3072                         lpni->lpni_ns_status = addnis[i].ns_status;
3073                         lnet_peer_ni_decref_locked(lpni);
3074                 }
3075         }
3076
3077         for (i = 0; i < ndelnis; i++) {
3078                 /*
3079                  * for routers it's okay to delete the primary_nid because
3080                  * the upper layers don't really rely on it. So if we're
3081                  * being told that the router changed its primary_nid
3082                  * then it's okay to delete it.
3083                  */
3084                 lnet_nid4_to_nid(delnis[i], &nid);
3085                 if (lp->lp_rtr_refcount > 0)
3086                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3087                 rc = lnet_peer_del_nid(lp, &nid, flags);
3088                 if (rc) {
3089                         CERROR("Error deleting NID %s from peer %s: %d\n",
3090                                libcfs_nid2str(delnis[i]),
3091                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3092                         if (rc == -ENOMEM)
3093                                 goto out;
3094                 }
3095         }
3096
3097         /* The peer net for the primary NID should be the first entry in the
3098          * peer's lp_peer_nets list, and the peer NI for the primary NID should
3099          * be the first entry in its peer net's lpn_peer_nis list.
3100          */
3101         lnet_nid4_to_nid(pbuf->pb_info.pi_ni[1].ns_nid, &nid);
3102         lpni = lnet_peer_ni_find_locked(&nid);
3103         if (!lpni) {
3104                 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3105                        libcfs_nid2str(pbuf->pb_info.pi_ni[1].ns_nid));
3106                 goto out;
3107         }
3108
3109         lnet_peer_ni_decref_locked(lpni);
3110
3111         lpn = lpni->lpni_peer_net;
3112         if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3113                 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3114
3115         if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3116                 list_move(&lpni->lpni_peer_nis,
3117                           &lpni->lpni_peer_net->lpn_peer_nis);
3118
3119         /*
3120          * Errors other than -ENOMEM are due to peers having been
3121          * configured with DLC. Ignore these because DLC overrides
3122          * Discovery.
3123          */
3124         rc = 0;
3125 out:
3126         /* If this peer is a gateway, invoke the routing callback to update
3127          * the associated route status
3128          */
3129         if (lp->lp_rtr_refcount > 0)
3130                 lnet_router_discovery_ping_reply(lp, pbuf);
3131
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(struct lnet_nid *nid,
3211                                      struct lnet_ping_buffer *pbuf)
3212 {
3213         struct lnet_ping_iter pi;
3214         struct lnet_nid pnid;
3215         u32 *st;
3216
3217         for (st = ping_iter_first(&pi, pbuf, &pnid);
3218              st;
3219              st = ping_iter_next(&pi, &pnid))
3220                 if (nid_same(nid, &pnid))
3221                         return true;
3222         return false;
3223 }
3224
3225 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3226  * to the discovery queue a reference was taken that will prevent the peer from
3227  * actually being freed by this function. After this function exits the
3228  * discovery thread should call lnet_peer_discovery_complete() which will
3229  * drop that reference as well as wake any waiters that may also be holding a
3230  * ref on the peer
3231  */
3232 static int lnet_peer_deletion(struct lnet_peer *lp)
3233 __must_hold(&lp->lp_lock)
3234 {
3235         struct list_head rlist;
3236         struct lnet_route *route, *tmp;
3237         int sensitivity = lp->lp_health_sensitivity;
3238         int rc = 0;
3239
3240         INIT_LIST_HEAD(&rlist);
3241
3242         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3243                libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3244
3245         /* no-op if lnet_peer_del() has already been called on this peer */
3246         if (lp->lp_state & LNET_PEER_MARK_DELETED)
3247                 goto clear_discovering;
3248
3249         spin_unlock(&lp->lp_lock);
3250
3251         mutex_lock(&the_lnet.ln_api_mutex);
3252         if (the_lnet.ln_state != LNET_STATE_RUNNING ||
3253             the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
3254                 mutex_unlock(&the_lnet.ln_api_mutex);
3255                 spin_lock(&lp->lp_lock);
3256                 rc = -ESHUTDOWN;
3257                 goto clear_discovering;
3258         }
3259
3260         lnet_peer_cancel_discovery(lp);
3261         lnet_net_lock(LNET_LOCK_EX);
3262         list_for_each_entry_safe(route, tmp,
3263                                  &lp->lp_routes,
3264                                  lr_gwlist)
3265                 lnet_move_route(route, NULL, &rlist);
3266
3267         /* lnet_peer_del_locked() deletes all the peer NIs owned by this peer */
3268         rc = lnet_peer_del_locked(lp);
3269         if (rc)
3270                 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3271                         libcfs_nidstr(&lp->lp_primary_nid), rc);
3272
3273         lnet_net_unlock(LNET_LOCK_EX);
3274
3275         list_for_each_entry_safe(route, tmp,
3276                                  &rlist, lr_list) {
3277                 /* re-add these routes */
3278                 lnet_add_route(route->lr_net,
3279                                route->lr_hops,
3280                                &route->lr_nid,
3281                                route->lr_priority,
3282                                sensitivity);
3283                 LIBCFS_FREE(route, sizeof(*route));
3284         }
3285
3286         mutex_unlock(&the_lnet.ln_api_mutex);
3287
3288         spin_lock(&lp->lp_lock);
3289
3290         rc = 0;
3291
3292 clear_discovering:
3293         lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3294                           LNET_PEER_FORCE_PUSH);
3295
3296         return rc;
3297 }
3298
3299 /*
3300  * Update a peer using the data received.
3301  */
3302 static int lnet_peer_data_present(struct lnet_peer *lp)
3303 __must_hold(&lp->lp_lock)
3304 {
3305         struct lnet_ping_buffer *pbuf;
3306         struct lnet_peer_ni *lpni;
3307         struct lnet_nid nid;
3308         unsigned int flags;
3309         int rc = 0;
3310
3311         pbuf = lp->lp_data;
3312         lp->lp_data = NULL;
3313         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3314         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3315         spin_unlock(&lp->lp_lock);
3316
3317         /*
3318          * Modifications of peer structures are done while holding the
3319          * ln_api_mutex. A global lock is required because we may be
3320          * modifying multiple peer structures, and a mutex greatly
3321          * simplifies memory management.
3322          *
3323          * The actual changes to the data structures must also protect
3324          * against concurrent lookups, for which the lnet_net_lock in
3325          * LNET_LOCK_EX mode is used.
3326          */
3327         mutex_lock(&the_lnet.ln_api_mutex);
3328         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3329                 rc = -ESHUTDOWN;
3330                 goto out;
3331         }
3332
3333         /*
3334          * If this peer is not on the peer list then it is being torn
3335          * down, and our reference count may be all that is keeping it
3336          * alive. Don't do any work on it.
3337          */
3338         if (list_empty(&lp->lp_peer_list)) {
3339                 lnet_ping_buffer_decref(pbuf);
3340                 goto out;
3341         }
3342
3343         flags = LNET_PEER_DISCOVERED;
3344         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3345                 flags |= LNET_PEER_MULTI_RAIL;
3346
3347         /*
3348          * Check whether the primary NID in the message matches the
3349          * primary NID of the peer. If it does, update the peer, if
3350          * it it does not, check whether there is already a peer with
3351          * that primary NID. If no such peer exists, try to update
3352          * the primary NID of the current peer (allowed if it was
3353          * created due to message traffic) and complete the update.
3354          * If the peer did exist, hand off the data to it.
3355          *
3356          * The peer for the loopback interface is a special case: this
3357          * is the peer for the local node, and we want to set its
3358          * primary NID to the correct value here. Moreover, this peer
3359          * can show up with only the loopback NID in the ping buffer.
3360          */
3361         if (pbuf->pb_info.pi_nnis <= 1) {
3362                 lnet_ping_buffer_decref(pbuf);
3363                 goto out;
3364         }
3365         lnet_nid4_to_nid(pbuf->pb_info.pi_ni[1].ns_nid, &nid);
3366         if (nid_is_lo0(&lp->lp_primary_nid)) {
3367                 rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3368                 if (rc)
3369                         lnet_ping_buffer_decref(pbuf);
3370                 else
3371                         rc = lnet_peer_merge_data(lp, pbuf);
3372         /*
3373          * if the primary nid of the peer is present in the ping info returned
3374          * from the peer, but it's not the local primary peer we have
3375          * cached and discovery is disabled, then we don't want to update
3376          * our local peer info, by adding or removing NIDs, we just want
3377          * to update the status of the nids that we currently have
3378          * recorded in that peer.
3379          */
3380         } else if (nid_same(&lp->lp_primary_nid, &nid) ||
3381                    (lnet_is_nid_in_ping_info(&lp->lp_primary_nid, pbuf) &&
3382                     lnet_is_discovery_disabled(lp))) {
3383                 rc = lnet_peer_merge_data(lp, pbuf);
3384         } else {
3385                 lpni = lnet_peer_ni_find_locked(&nid);
3386                 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3387                         rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3388                         if (rc) {
3389                                 CERROR("Primary NID error %s versus %s: %d\n",
3390                                        libcfs_nidstr(&lp->lp_primary_nid),
3391                                        libcfs_nidstr(&nid), rc);
3392                                 lnet_ping_buffer_decref(pbuf);
3393                         } else {
3394                                 rc = lnet_peer_merge_data(lp, pbuf);
3395                         }
3396                         if (lpni)
3397                                 lnet_peer_ni_decref_locked(lpni);
3398                 } else {
3399                         struct lnet_peer *new_lp;
3400                         new_lp = lpni->lpni_peer_net->lpn_peer;
3401                         /*
3402                          * if lp has discovery/MR enabled that means new_lp
3403                          * should have discovery/MR enabled as well, since
3404                          * it's the same peer, which we're about to merge
3405                          */
3406                         spin_lock(&lp->lp_lock);
3407                         spin_lock(&new_lp->lp_lock);
3408                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3409                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3410                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3411                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3412                         /* If we're processing a ping reply then we may be
3413                          * about to send a push to the peer that we ping'd.
3414                          * Since the ping reply that we're processing was
3415                          * received by lp, we need to set the discovery source
3416                          * NID for new_lp to the NID stored in lp.
3417                          */
3418                         if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3419                                 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3420                                 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3421                         }
3422                         spin_unlock(&new_lp->lp_lock);
3423                         spin_unlock(&lp->lp_lock);
3424
3425                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
3426                         lnet_consolidate_routes_locked(lp, new_lp);
3427                         lnet_peer_ni_decref_locked(lpni);
3428                 }
3429         }
3430 out:
3431         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3432                libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3433                lp->lp_state);
3434         mutex_unlock(&the_lnet.ln_api_mutex);
3435
3436         spin_lock(&lp->lp_lock);
3437         /* Tell discovery to re-check the peer immediately. */
3438         if (!rc)
3439                 rc = LNET_REDISCOVER_PEER;
3440         return rc;
3441 }
3442
3443 /*
3444  * A ping failed. Clear the PING_FAILED state and set the
3445  * FORCE_PING state, to ensure a retry even if discovery is
3446  * disabled. This avoids being left with incorrect state.
3447  */
3448 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3449 __must_hold(&lp->lp_lock)
3450 {
3451         struct lnet_handle_md mdh;
3452         int rc;
3453
3454         mdh = lp->lp_ping_mdh;
3455         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3456         lp->lp_state &= ~LNET_PEER_PING_FAILED;
3457         lp->lp_state |= LNET_PEER_FORCE_PING;
3458         rc = lp->lp_ping_error;
3459         lp->lp_ping_error = 0;
3460         spin_unlock(&lp->lp_lock);
3461
3462         if (!LNetMDHandleIsInvalid(mdh))
3463                 LNetMDUnlink(mdh);
3464
3465         CDEBUG(D_NET, "peer %s:%d\n",
3466                libcfs_nidstr(&lp->lp_primary_nid), rc);
3467
3468         spin_lock(&lp->lp_lock);
3469         return rc ? rc : LNET_REDISCOVER_PEER;
3470 }
3471
3472 /* Active side of ping. */
3473 static int lnet_peer_send_ping(struct lnet_peer *lp)
3474 __must_hold(&lp->lp_lock)
3475 {
3476         int bytes;
3477         int rc;
3478         int cpt;
3479
3480         lp->lp_state |= LNET_PEER_PING_SENT;
3481         lp->lp_state &= ~LNET_PEER_FORCE_PING;
3482         spin_unlock(&lp->lp_lock);
3483
3484         cpt = lnet_net_lock_current();
3485         /* Refcount for MD. */
3486         lnet_peer_addref_locked(lp);
3487         lnet_net_unlock(cpt);
3488
3489         bytes = max_t(int, lp->lp_data_bytes, LNET_PING_INFO_MIN_SIZE);
3490
3491         rc = lnet_send_ping(&lp->lp_primary_nid, &lp->lp_ping_mdh, bytes, lp,
3492                             the_lnet.ln_dc_handler, false);
3493         /* if LNetMDBind in lnet_send_ping fails we need to decrement the
3494          * refcount on the peer, otherwise LNetMDUnlink will be called
3495          * which will eventually do that.
3496          */
3497         if (rc > 0) {
3498                 lnet_net_lock(cpt);
3499                 lnet_peer_decref_locked(lp);
3500                 lnet_net_unlock(cpt);
3501                 rc = -rc; /* change the rc to negative value */
3502                 goto fail_error;
3503         } else if (rc < 0) {
3504                 goto fail_error;
3505         }
3506
3507         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3508
3509         spin_lock(&lp->lp_lock);
3510         return 0;
3511
3512 fail_error:
3513         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3514         /*
3515          * The errors that get us here are considered hard errors and
3516          * cause Discovery to terminate. So we clear PING_SENT, but do
3517          * not set either PING_FAILED or FORCE_PING. In fact we need
3518          * to clear PING_FAILED, because the unlink event handler will
3519          * have set it if we called LNetMDUnlink() above.
3520          */
3521         spin_lock(&lp->lp_lock);
3522         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3523         return rc;
3524 }
3525
3526 /*
3527  * This function exists because you cannot call LNetMDUnlink() from an
3528  * event handler.
3529  */
3530 static int lnet_peer_push_failed(struct lnet_peer *lp)
3531 __must_hold(&lp->lp_lock)
3532 {
3533         struct lnet_handle_md mdh;
3534         int rc;
3535
3536         mdh = lp->lp_push_mdh;
3537         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3538         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3539         rc = lp->lp_push_error;
3540         lp->lp_push_error = 0;
3541         spin_unlock(&lp->lp_lock);
3542
3543         if (!LNetMDHandleIsInvalid(mdh))
3544                 LNetMDUnlink(mdh);
3545
3546         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3547         spin_lock(&lp->lp_lock);
3548         return rc ? rc : LNET_REDISCOVER_PEER;
3549 }
3550
3551 /*
3552  * Mark the peer as discovered.
3553  */
3554 static int lnet_peer_discovered(struct lnet_peer *lp)
3555 __must_hold(&lp->lp_lock)
3556 {
3557         lp->lp_state |= LNET_PEER_DISCOVERED;
3558         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3559                           LNET_PEER_REDISCOVER);
3560
3561         lp->lp_dc_error = 0;
3562
3563         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3564
3565         return 0;
3566 }
3567
3568 /* Active side of push. */
3569 static int lnet_peer_send_push(struct lnet_peer *lp)
3570 __must_hold(&lp->lp_lock)
3571 {
3572         struct lnet_ping_buffer *pbuf;
3573         struct lnet_processid id;
3574         struct lnet_md md;
3575         int cpt;
3576         int rc;
3577
3578         /* Don't push to a non-multi-rail peer. */
3579         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3580                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3581                 /* if peer's NIDs are uptodate then peer is discovered */
3582                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3583                         rc = lnet_peer_discovered(lp);
3584                         return rc;
3585                 }
3586
3587                 return 0;
3588         }
3589
3590         lp->lp_state |= LNET_PEER_PUSH_SENT;
3591         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3592         spin_unlock(&lp->lp_lock);
3593
3594         cpt = lnet_net_lock_current();
3595         pbuf = the_lnet.ln_ping_target;
3596         lnet_ping_buffer_addref(pbuf);
3597         lnet_net_unlock(cpt);
3598
3599         /* Push source MD */
3600         md.start     = &pbuf->pb_info;
3601         md.length    = pbuf->pb_nbytes;
3602         md.threshold = 2; /* Put/Ack */
3603         md.max_size  = 0;
3604         md.options   = LNET_MD_TRACK_RESPONSE;
3605         md.handler   = the_lnet.ln_dc_handler;
3606         md.user_ptr  = lp;
3607
3608         rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3609         if (rc) {
3610                 lnet_ping_buffer_decref(pbuf);
3611                 CERROR("Can't bind push source MD: %d\n", rc);
3612                 goto fail_error;
3613         }
3614
3615         cpt = lnet_net_lock_current();
3616         /* Refcount for MD. */
3617         lnet_peer_addref_locked(lp);
3618         id.pid = LNET_PID_LUSTRE;
3619         if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3620                 id.nid = lp->lp_disc_dst_nid;
3621         else
3622                 id.nid = lp->lp_primary_nid;
3623         lnet_net_unlock(cpt);
3624
3625         rc = LNetPut(&lp->lp_disc_src_nid, lp->lp_push_mdh,
3626                      LNET_ACK_REQ, &id, LNET_RESERVED_PORTAL,
3627                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3628
3629         /*
3630          * reset the discovery nid. There is no need to restrict sending
3631          * from that source, if we call lnet_push_update_to_peers(). It'll
3632          * get set to a specific NID, if we initiate discovery from the
3633          * scratch
3634          */
3635         lp->lp_disc_src_nid = LNET_ANY_NID;
3636         lp->lp_disc_dst_nid = LNET_ANY_NID;
3637
3638         if (rc)
3639                 goto fail_unlink;
3640
3641         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3642
3643         spin_lock(&lp->lp_lock);
3644         return 0;
3645
3646 fail_unlink:
3647         LNetMDUnlink(lp->lp_push_mdh);
3648         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3649 fail_error:
3650         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3651                lp, rc);
3652         /*
3653          * The errors that get us here are considered hard errors and
3654          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3655          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3656          * because the unlink event handler will have set it if we
3657          * called LNetMDUnlink() above.
3658          */
3659         spin_lock(&lp->lp_lock);
3660         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3661         return rc;
3662 }
3663
3664 /*
3665  * Wait for work to be queued or some other change that must be
3666  * attended to. Returns non-zero if the discovery thread should shut
3667  * down.
3668  */
3669 static int lnet_peer_discovery_wait_for_work(void)
3670 {
3671         int cpt;
3672         int rc = 0;
3673
3674         DEFINE_WAIT(wait);
3675
3676         cpt = lnet_net_lock_current();
3677         for (;;) {
3678                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3679                                 TASK_INTERRUPTIBLE);
3680                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3681                         break;
3682                 if (lnet_push_target_resize_needed() ||
3683                     the_lnet.ln_push_target->pb_needs_post)
3684                         break;
3685                 if (!list_empty(&the_lnet.ln_dc_request))
3686                         break;
3687                 if (!list_empty(&the_lnet.ln_msg_resend))
3688                         break;
3689                 lnet_net_unlock(cpt);
3690
3691                 /*
3692                  * wakeup max every second to check if there are peers that
3693                  * have been stuck on the working queue for greater than
3694                  * the peer timeout.
3695                  */
3696                 schedule_timeout(cfs_time_seconds(1));
3697                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3698                 cpt = lnet_net_lock_current();
3699         }
3700         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3701
3702         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3703                 rc = -ESHUTDOWN;
3704
3705         lnet_net_unlock(cpt);
3706
3707         CDEBUG(D_NET, "woken: %d\n", rc);
3708
3709         return rc;
3710 }
3711
3712 /*
3713  * Messages that were pending on a destroyed peer will be put on a global
3714  * resend list. The message resend list will be checked by
3715  * the discovery thread when it wakes up, and will resend messages. These
3716  * messages can still be sendable in the case the lpni which was the initial
3717  * cause of the message re-queue was transfered to another peer.
3718  *
3719  * It is possible that LNet could be shutdown while we're iterating
3720  * through the list. lnet_shudown_lndnets() will attempt to access the
3721  * resend list, but will have to wait until the spinlock is released, by
3722  * which time there shouldn't be any more messages on the resend list.
3723  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3724  * for the messages so they can be released. The other case is that
3725  * lnet_shudown_lndnets() can finalize all the messages before this
3726  * function can visit the resend list, in which case this function will be
3727  * a no-op.
3728  */
3729 static void lnet_resend_msgs(void)
3730 {
3731         struct lnet_msg *msg, *tmp;
3732         LIST_HEAD(resend);
3733         int rc;
3734
3735         spin_lock(&the_lnet.ln_msg_resend_lock);
3736         list_splice(&the_lnet.ln_msg_resend, &resend);
3737         spin_unlock(&the_lnet.ln_msg_resend_lock);
3738
3739         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3740                 list_del_init(&msg->msg_list);
3741                 rc = lnet_send(&msg->msg_src_nid_param, msg,
3742                                &msg->msg_rtr_nid_param);
3743                 if (rc < 0) {
3744                         CNETERR("Error sending %s to %s: %d\n",
3745                                lnet_msgtyp2str(msg->msg_type),
3746                                libcfs_idstr(&msg->msg_target), rc);
3747                         lnet_finalize(msg, rc);
3748                 }
3749         }
3750 }
3751
3752 /* The discovery thread. */
3753 static int lnet_peer_discovery(void *arg)
3754 {
3755         struct lnet_peer *lp;
3756         int rc;
3757
3758         wait_for_completion(&the_lnet.ln_started);
3759
3760         CDEBUG(D_NET, "started\n");
3761
3762         for (;;) {
3763                 if (lnet_peer_discovery_wait_for_work())
3764                         break;
3765
3766                 if (lnet_push_target_resize_needed())
3767                         lnet_push_target_resize();
3768                 else if (the_lnet.ln_push_target->pb_needs_post)
3769                         lnet_push_target_post(the_lnet.ln_push_target,
3770                                               &the_lnet.ln_push_target_md);
3771
3772                 lnet_resend_msgs();
3773
3774                 lnet_net_lock(LNET_LOCK_EX);
3775                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3776                         lnet_net_unlock(LNET_LOCK_EX);
3777                         break;
3778                 }
3779
3780                 /*
3781                  * Process all incoming discovery work requests.  When
3782                  * discovery must wait on a peer to change state, it
3783                  * is added to the tail of the ln_dc_working queue. A
3784                  * timestamp keeps track of when the peer was added,
3785                  * so we can time out discovery requests that take too
3786                  * long.
3787                  */
3788                 while (!list_empty(&the_lnet.ln_dc_request)) {
3789                         lp = list_first_entry(&the_lnet.ln_dc_request,
3790                                               struct lnet_peer, lp_dc_list);
3791                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3792                         /*
3793                          * set the time the peer was put on the dc_working
3794                          * queue. It shouldn't remain on the queue
3795                          * forever, in case the GET message (for ping)
3796                          * doesn't get a REPLY or the PUT message (for
3797                          * push) doesn't get an ACK.
3798                          */
3799                         lp->lp_last_queued = ktime_get_real_seconds();
3800                         lnet_net_unlock(LNET_LOCK_EX);
3801
3802                         if (lnet_push_target_resize_needed())
3803                                 lnet_push_target_resize();
3804                         else if (the_lnet.ln_push_target->pb_needs_post)
3805                                 lnet_push_target_post(the_lnet.ln_push_target,
3806                                                       &the_lnet.ln_push_target_md);
3807
3808                         /*
3809                          * Select an action depending on the state of
3810                          * the peer and whether discovery is disabled.
3811                          * The check whether discovery is disabled is
3812                          * done after the code that handles processing
3813                          * for arrived data, cleanup for failures, and
3814                          * forcing a Ping or Push.
3815                          */
3816                         spin_lock(&lp->lp_lock);
3817                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3818                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3819                                 lp->lp_state);
3820                         if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3821                                             LNET_PEER_MARK_DELETED))
3822                                 rc = lnet_peer_deletion(lp);
3823                         else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3824                                 rc = lnet_peer_data_present(lp);
3825                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3826                                 rc = lnet_peer_ping_failed(lp);
3827                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3828                                 rc = lnet_peer_push_failed(lp);
3829                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3830                                 rc = lnet_peer_send_ping(lp);
3831                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3832                                 rc = lnet_peer_send_push(lp);
3833                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3834                                 rc = lnet_peer_send_ping(lp);
3835                         else if (lnet_peer_needs_push(lp))
3836                                 rc = lnet_peer_send_push(lp);
3837                         else
3838                                 rc = lnet_peer_discovered(lp);
3839                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3840                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3841                                 lp->lp_state, rc);
3842
3843                         if (rc == LNET_REDISCOVER_PEER) {
3844                                 spin_unlock(&lp->lp_lock);
3845                                 lnet_net_lock(LNET_LOCK_EX);
3846                                 list_move(&lp->lp_dc_list,
3847                                           &the_lnet.ln_dc_request);
3848                         } else if (rc ||
3849                                    !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3850                                 spin_unlock(&lp->lp_lock);
3851                                 lnet_net_lock(LNET_LOCK_EX);
3852                                 lnet_peer_discovery_complete(lp, rc);
3853                         } else {
3854                                 spin_unlock(&lp->lp_lock);
3855                                 lnet_net_lock(LNET_LOCK_EX);
3856                         }
3857
3858                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3859                                 break;
3860
3861                 }
3862
3863                 lnet_net_unlock(LNET_LOCK_EX);
3864         }
3865
3866         CDEBUG(D_NET, "stopping\n");
3867         /*
3868          * Clean up before telling lnet_peer_discovery_stop() that
3869          * we're done. Use wake_up() below to somewhat reduce the
3870          * size of the thundering herd if there are multiple threads
3871          * waiting on discovery of a single peer.
3872          */
3873
3874         /* Queue cleanup 1: stop all pending pings and pushes. */
3875         lnet_net_lock(LNET_LOCK_EX);
3876         while (!list_empty(&the_lnet.ln_dc_working)) {
3877                 lp = list_first_entry(&the_lnet.ln_dc_working,
3878                                       struct lnet_peer, lp_dc_list);
3879                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3880                 lnet_net_unlock(LNET_LOCK_EX);
3881                 lnet_peer_cancel_discovery(lp);
3882                 lnet_net_lock(LNET_LOCK_EX);
3883         }
3884         lnet_net_unlock(LNET_LOCK_EX);
3885
3886         /* Queue cleanup 2: wait for the expired queue to clear. */
3887         while (!list_empty(&the_lnet.ln_dc_expired))
3888                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3889
3890         /* Queue cleanup 3: clear the request queue. */
3891         lnet_net_lock(LNET_LOCK_EX);
3892         while (!list_empty(&the_lnet.ln_dc_request)) {
3893                 lp = list_first_entry(&the_lnet.ln_dc_request,
3894                                       struct lnet_peer, lp_dc_list);
3895                 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3896         }
3897         lnet_net_unlock(LNET_LOCK_EX);
3898
3899         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3900         the_lnet.ln_dc_handler = NULL;
3901
3902         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3903         wake_up(&the_lnet.ln_dc_waitq);
3904
3905         CDEBUG(D_NET, "stopped\n");
3906
3907         return 0;
3908 }
3909
3910 /* ln_api_mutex is held on entry. */
3911 int lnet_peer_discovery_start(void)
3912 {
3913         struct task_struct *task;
3914         int rc = 0;
3915
3916         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3917                 return -EALREADY;
3918
3919         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3920         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3921         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3922         if (IS_ERR(task)) {
3923                 rc = PTR_ERR(task);
3924                 CERROR("Can't start peer discovery thread: %d\n", rc);
3925
3926                 the_lnet.ln_dc_handler = NULL;
3927
3928                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3929         }
3930
3931         CDEBUG(D_NET, "discovery start: %d\n", rc);
3932
3933         return rc;
3934 }
3935
3936 /* ln_api_mutex is held on entry. */
3937 void lnet_peer_discovery_stop(void)
3938 {
3939         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3940                 return;
3941
3942         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
3943         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
3944
3945         /* In the LNetNIInit() path we may be stopping discovery before it
3946          * entered its work loop
3947          */
3948         if (!completion_done(&the_lnet.ln_started))
3949                 complete(&the_lnet.ln_started);
3950         else
3951                 wake_up(&the_lnet.ln_dc_waitq);
3952
3953         mutex_unlock(&the_lnet.ln_api_mutex);
3954         wait_event(the_lnet.ln_dc_waitq,
3955                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
3956         mutex_lock(&the_lnet.ln_api_mutex);
3957
3958         LASSERT(list_empty(&the_lnet.ln_dc_request));
3959         LASSERT(list_empty(&the_lnet.ln_dc_working));
3960         LASSERT(list_empty(&the_lnet.ln_dc_expired));
3961
3962         CDEBUG(D_NET, "discovery stopped\n");
3963 }
3964
3965 /* Debugging */
3966
3967 void
3968 lnet_debug_peer(lnet_nid_t nid4)
3969 {
3970         char                    *aliveness = "NA";
3971         struct lnet_peer_ni     *lp;
3972         int                     cpt;
3973         struct lnet_nid nid;
3974
3975         lnet_nid4_to_nid(nid4, &nid);
3976         cpt = lnet_nid2cpt(&nid, NULL);
3977         lnet_net_lock(cpt);
3978
3979         lp = lnet_peerni_by_nid_locked(&nid, NULL, cpt);
3980         if (IS_ERR(lp)) {
3981                 lnet_net_unlock(cpt);
3982                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nidstr(&nid));
3983                 return;
3984         }
3985
3986         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
3987                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
3988
3989         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
3990                libcfs_nidstr(&lp->lpni_nid), kref_read(&lp->lpni_kref),
3991                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
3992                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
3993                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
3994
3995         lnet_peer_ni_decref_locked(lp);
3996
3997         lnet_net_unlock(cpt);
3998 }
3999
4000 /* Gathering information for userspace. */
4001
4002 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
4003                           char aliveness[LNET_MAX_STR_LEN],
4004                           __u32 *cpt_iter, __u32 *refcount,
4005                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
4006                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
4007                           __u32 *peer_tx_qnob)
4008 {
4009         struct lnet_peer_table          *peer_table;
4010         struct lnet_peer_ni             *lp;
4011         int                             j;
4012         int                             lncpt;
4013         bool                            found = false;
4014
4015         /* get the number of CPTs */
4016         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4017
4018         /* if the cpt number to be examined is >= the number of cpts in
4019          * the system then indicate that there are no more cpts to examin
4020          */
4021         if (*cpt_iter >= lncpt)
4022                 return -ENOENT;
4023
4024         /* get the current table */
4025         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
4026         /* if the ptable is NULL then there are no more cpts to examine */
4027         if (peer_table == NULL)
4028                 return -ENOENT;
4029
4030         lnet_net_lock(*cpt_iter);
4031
4032         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
4033                 struct list_head *peers = &peer_table->pt_hash[j];
4034
4035                 list_for_each_entry(lp, peers, lpni_hashlist) {
4036                         if (!nid_is_nid4(&lp->lpni_nid))
4037                                 continue;
4038                         if (peer_index-- > 0)
4039                                 continue;
4040
4041                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
4042                         if (lnet_isrouter(lp) ||
4043                                 lnet_peer_aliveness_enabled(lp))
4044                                 snprintf(aliveness, LNET_MAX_STR_LEN,
4045                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
4046
4047                         *nid = lnet_nid_to_nid4(&lp->lpni_nid);
4048                         *refcount = kref_read(&lp->lpni_kref);
4049                         *ni_peer_tx_credits =
4050                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
4051                         *peer_tx_credits = lp->lpni_txcredits;
4052                         *peer_rtr_credits = lp->lpni_rtrcredits;
4053                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
4054                         *peer_tx_qnob = lp->lpni_txqnob;
4055
4056                         found = true;
4057                 }
4058
4059         }
4060         lnet_net_unlock(*cpt_iter);
4061
4062         *cpt_iter = lncpt;
4063
4064         return found ? 0 : -ENOENT;
4065 }
4066
4067 /* ln_api_mutex is held, which keeps the peer list stable */
4068 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
4069 {
4070         struct lnet_ioctl_element_stats *lpni_stats;
4071         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
4072         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
4073         struct lnet_peer_ni_credit_info *lpni_info;
4074         struct lnet_peer_ni *lpni;
4075         struct lnet_peer *lp;
4076         lnet_nid_t nid4;
4077         struct lnet_nid nid;
4078         __u32 size;
4079         int rc;
4080
4081         lnet_nid4_to_nid(cfg->prcfg_prim_nid, &nid);
4082         lp = lnet_find_peer(&nid);
4083
4084         if (!lp) {
4085                 rc = -ENOENT;
4086                 goto out;
4087         }
4088
4089         size = sizeof(nid4) + sizeof(*lpni_info) + sizeof(*lpni_stats)
4090                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
4091         size *= lp->lp_nnis;
4092         if (size > cfg->prcfg_size) {
4093                 cfg->prcfg_size = size;
4094                 rc = -E2BIG;
4095                 goto out_lp_decref;
4096         }
4097
4098         cfg->prcfg_prim_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4099         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
4100         cfg->prcfg_cfg_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4101         cfg->prcfg_count = lp->lp_nnis;
4102         cfg->prcfg_size = size;
4103         cfg->prcfg_state = lp->lp_state;
4104
4105         /* Allocate helper buffers. */
4106         rc = -ENOMEM;
4107         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
4108         if (!lpni_info)
4109                 goto out_lp_decref;
4110         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
4111         if (!lpni_stats)
4112                 goto out_free_info;
4113         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
4114         if (!lpni_msg_stats)
4115                 goto out_free_stats;
4116         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
4117         if (!lpni_hstats)
4118                 goto out_free_msg_stats;
4119
4120
4121         lpni = NULL;
4122         rc = -EFAULT;
4123         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
4124                 if (!nid_is_nid4(&lpni->lpni_nid))
4125                         continue;
4126                 nid4 = lnet_nid_to_nid4(&lpni->lpni_nid);
4127                 if (copy_to_user(bulk, &nid4, sizeof(nid4)))
4128                         goto out_free_hstats;
4129                 bulk += sizeof(nid4);
4130
4131                 memset(lpni_info, 0, sizeof(*lpni_info));
4132                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
4133                 if (lnet_isrouter(lpni) ||
4134                         lnet_peer_aliveness_enabled(lpni))
4135                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
4136                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
4137
4138                 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
4139                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
4140                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
4141                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
4142                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
4143                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
4144                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
4145                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4146                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4147                         goto out_free_hstats;
4148                 bulk += sizeof(*lpni_info);
4149
4150                 memset(lpni_stats, 0, sizeof(*lpni_stats));
4151                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4152                                                             LNET_STATS_TYPE_SEND);
4153                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4154                                                             LNET_STATS_TYPE_RECV);
4155                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4156                                                             LNET_STATS_TYPE_DROP);
4157                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4158                         goto out_free_hstats;
4159                 bulk += sizeof(*lpni_stats);
4160                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4161                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4162                         goto out_free_hstats;
4163                 bulk += sizeof(*lpni_msg_stats);
4164                 lpni_hstats->hlpni_network_timeout =
4165                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4166                 lpni_hstats->hlpni_remote_dropped =
4167                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4168                 lpni_hstats->hlpni_remote_timeout =
4169                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4170                 lpni_hstats->hlpni_remote_error =
4171                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4172                 lpni_hstats->hlpni_health_value =
4173                   atomic_read(&lpni->lpni_healthv);
4174                 lpni_hstats->hlpni_ping_count = lpni->lpni_ping_count;
4175                 lpni_hstats->hlpni_next_ping = lpni->lpni_next_ping;
4176                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4177                         goto out_free_hstats;
4178                 bulk += sizeof(*lpni_hstats);
4179         }
4180         rc = 0;
4181
4182 out_free_hstats:
4183         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4184 out_free_msg_stats:
4185         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4186 out_free_stats:
4187         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4188 out_free_info:
4189         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4190 out_lp_decref:
4191         lnet_peer_decref_locked(lp);
4192 out:
4193         return rc;
4194 }
4195
4196 /* must hold net_lock/0 */
4197 void
4198 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4199                                      struct list_head *recovery_queue,
4200                                      time64_t now)
4201 {
4202         /* the mt could've shutdown and cleaned up the queues */
4203         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4204                 return;
4205
4206         if (!list_empty(&lpni->lpni_recovery))
4207                 return;
4208
4209         if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4210                 return;
4211
4212         if (!lpni->lpni_last_alive) {
4213                 CDEBUG(D_NET,
4214                        "lpni %s(%p) not eligible for recovery last alive %lld\n",
4215                        libcfs_nidstr(&lpni->lpni_nid), lpni,
4216                        lpni->lpni_last_alive);
4217                 return;
4218         }
4219
4220         if (lnet_recovery_limit &&
4221             now > lpni->lpni_last_alive + lnet_recovery_limit) {
4222                 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4223                        libcfs_nidstr(&lpni->lpni_nid),
4224                        lpni->lpni_last_alive);
4225                 /* Reset the ping count so that if this peer NI is added back to
4226                  * the recovery queue we will send the first ping right away.
4227                  */
4228                 lpni->lpni_ping_count = 0;
4229                 return;
4230         }
4231
4232         /* This peer NI is going on the recovery queue, so take a ref on it */
4233         lnet_peer_ni_addref_locked(lpni);
4234
4235         lnet_peer_ni_set_next_ping(lpni, now);
4236
4237         CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4238                libcfs_nidstr(&lpni->lpni_nid),
4239                lpni->lpni_ping_count,
4240                lpni->lpni_next_ping,
4241                lpni->lpni_last_alive,
4242                atomic_read(&lpni->lpni_healthv));
4243
4244         list_add_tail(&lpni->lpni_recovery, recovery_queue);
4245 }
4246
4247 /* Call with the ln_api_mutex held */
4248 void
4249 lnet_peer_ni_set_healthv(lnet_nid_t nid4, int value, bool all)
4250 {
4251         struct lnet_peer_table *ptable;
4252         struct lnet_peer *lp;
4253         struct lnet_peer_net *lpn;
4254         struct lnet_peer_ni *lpni;
4255         struct lnet_nid nid;
4256         int lncpt;
4257         int cpt;
4258         time64_t now;
4259
4260         if (the_lnet.ln_state != LNET_STATE_RUNNING)
4261                 return;
4262
4263         lnet_nid4_to_nid(nid4, &nid);
4264         now = ktime_get_seconds();
4265
4266         if (!all) {
4267                 lnet_net_lock(LNET_LOCK_EX);
4268                 lpni = lnet_peer_ni_find_locked(&nid);
4269                 if (!lpni) {
4270                         lnet_net_unlock(LNET_LOCK_EX);
4271                         return;
4272                 }
4273                 lnet_set_lpni_healthv_locked(lpni, value);
4274                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4275                                              &the_lnet.ln_mt_peerNIRecovq, now);
4276                 lnet_peer_ni_decref_locked(lpni);
4277                 lnet_net_unlock(LNET_LOCK_EX);
4278                 return;
4279         }
4280
4281         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4282
4283         /*
4284          * Walk all the peers and reset the health value for each one to the
4285          * specified value.
4286          */
4287         lnet_net_lock(LNET_LOCK_EX);
4288         for (cpt = 0; cpt < lncpt; cpt++) {
4289                 ptable = the_lnet.ln_peer_tables[cpt];
4290                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4291                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4292                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4293                                                     lpni_peer_nis) {
4294                                         lnet_set_lpni_healthv_locked(lpni,
4295                                                                      value);
4296                                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
4297                                              &the_lnet.ln_mt_peerNIRecovq, now);
4298                                 }
4299                         }
4300                 }
4301         }
4302         lnet_net_unlock(LNET_LOCK_EX);
4303 }
4304