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         unsafe_memcpy(&lp->lp_data->pb_info, &pbuf->pb_info, infobytes,
2444                       FLEXIBLE_OBJECT);
2445         lp->lp_state |= LNET_PEER_DATA_PRESENT;
2446         CDEBUG(D_NET, "Received Push %s %u\n",
2447                libcfs_nidstr(&lp->lp_primary_nid),
2448                LNET_PING_BUFFER_SEQNO(pbuf));
2449
2450 out:
2451         /* We've processed this buffer. It can be reposted */
2452         pbuf->pb_needs_post = true;
2453
2454         /*
2455          * Queue the peer for discovery if not done, force it on the request
2456          * queue and wake the discovery thread if the peer was already queued,
2457          * because its status changed.
2458          */
2459         spin_unlock(&lp->lp_lock);
2460         lnet_net_lock(LNET_LOCK_EX);
2461         if (!lnet_peer_is_uptodate(lp) && lnet_peer_queue_for_discovery(lp)) {
2462                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2463                 wake_up(&the_lnet.ln_dc_waitq);
2464         }
2465         /* Drop refcount from lookup */
2466         lnet_peer_decref_locked(lp);
2467         lnet_net_unlock(LNET_LOCK_EX);
2468 }
2469
2470 /*
2471  * Clear the discovery error state, unless we're already discovering
2472  * this peer, in which case the error is current.
2473  */
2474 static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
2475 {
2476         spin_lock(&lp->lp_lock);
2477         if (!(lp->lp_state & LNET_PEER_DISCOVERING))
2478                 lp->lp_dc_error = 0;
2479         spin_unlock(&lp->lp_lock);
2480 }
2481
2482 /*
2483  * Peer discovery slow path. The ln_api_mutex is held on entry, and
2484  * dropped/retaken within this function. An lnet_peer_ni is passed in
2485  * because discovery could tear down an lnet_peer.
2486  */
2487 int
2488 lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
2489 {
2490         DEFINE_WAIT(wait);
2491         struct lnet_peer *lp;
2492         int rc = 0;
2493         int count = 0;
2494
2495 again:
2496         lnet_net_unlock(cpt);
2497         lnet_net_lock(LNET_LOCK_EX);
2498         lp = lpni->lpni_peer_net->lpn_peer;
2499         lnet_peer_clear_discovery_error(lp);
2500
2501         /*
2502          * We're willing to be interrupted. The lpni can become a
2503          * zombie if we race with DLC, so we must check for that.
2504          */
2505         for (;;) {
2506                 /* Keep lp alive when the lnet_net_lock is unlocked */
2507                 lnet_peer_addref_locked(lp);
2508                 prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
2509                 if (signal_pending(current))
2510                         break;
2511                 if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2512                         break;
2513                 /*
2514                  * Don't repeat discovery if discovery is disabled. This is
2515                  * done to ensure we can use discovery as a standard ping as
2516                  * well for backwards compatibility with routers which do not
2517                  * have discovery or have discovery disabled
2518                  */
2519                 if (lnet_is_discovery_disabled(lp) && count > 0)
2520                         break;
2521                 if (lp->lp_dc_error)
2522                         break;
2523                 if (lnet_peer_is_uptodate(lp))
2524                         break;
2525                 if (lp->lp_state & LNET_PEER_MARK_DELETED)
2526                         break;
2527                 lnet_peer_queue_for_discovery(lp);
2528                 count++;
2529                 CDEBUG(D_NET, "Discovery attempt # %d\n", count);
2530
2531                 /*
2532                  * If caller requested a non-blocking operation then
2533                  * return immediately. Once discovery is complete any
2534                  * pending messages that were stopped due to discovery
2535                  * will be transmitted.
2536                  */
2537                 if (!block)
2538                         break;
2539
2540                 lnet_net_unlock(LNET_LOCK_EX);
2541                 schedule();
2542                 finish_wait(&lp->lp_dc_waitq, &wait);
2543                 lnet_net_lock(LNET_LOCK_EX);
2544                 lnet_peer_decref_locked(lp);
2545                 /* Peer may have changed */
2546                 lp = lpni->lpni_peer_net->lpn_peer;
2547         }
2548         finish_wait(&lp->lp_dc_waitq, &wait);
2549
2550         lnet_net_unlock(LNET_LOCK_EX);
2551         lnet_net_lock(cpt);
2552         lnet_peer_decref_locked(lp);
2553         /*
2554          * The peer may have changed, so re-check and rediscover if that turns
2555          * out to have been the case. The reference count on lp ensured that
2556          * even if it was unlinked from lpni the memory could not be recycled.
2557          * Thus the check below is sufficient to determine whether the peer
2558          * changed. If the peer changed, then lp must not be dereferenced.
2559          */
2560         if (lp != lpni->lpni_peer_net->lpn_peer)
2561                 goto again;
2562
2563         if (signal_pending(current))
2564                 rc = -EINTR;
2565         else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
2566                 rc = -ESHUTDOWN;
2567         else if (lp->lp_dc_error)
2568                 rc = lp->lp_dc_error;
2569         else if (!block)
2570                 CDEBUG(D_NET, "non-blocking discovery\n");
2571         else if (!lnet_peer_is_uptodate(lp) &&
2572                  !(lnet_is_discovery_disabled(lp) ||
2573                    (lp->lp_state & LNET_PEER_MARK_DELETED)))
2574                 goto again;
2575
2576         CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
2577                (lp ? libcfs_nidstr(&lp->lp_primary_nid) : "(none)"),
2578                libcfs_nidstr(&lpni->lpni_nid), rc,
2579                (!block) ? "pending discovery" : "discovery complete");
2580
2581         return rc;
2582 }
2583
2584 /* Handle an incoming ack for a push. */
2585 static void
2586 lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
2587 {
2588         struct lnet_ping_buffer *pbuf;
2589
2590         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2591         spin_lock(&lp->lp_lock);
2592         lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2593         lp->lp_push_error = ev->status;
2594         if (ev->status)
2595                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2596         else
2597                 lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2598         spin_unlock(&lp->lp_lock);
2599
2600         CDEBUG(D_NET, "peer %s ev->status %d\n",
2601                libcfs_nidstr(&lp->lp_primary_nid), ev->status);
2602 }
2603
2604 static bool find_primary(struct lnet_nid *nid,
2605                          struct lnet_ping_buffer *pbuf)
2606 {
2607         struct lnet_ping_info *pi = &pbuf->pb_info;
2608         struct lnet_ping_iter piter;
2609         __u32 *stp;
2610
2611         if (pi->pi_features & LNET_PING_FEAT_PRIMARY_LARGE) {
2612                 /* First large nid is primary */
2613                 for (stp = ping_iter_first(&piter, pbuf, nid);
2614                      stp;
2615                      stp = ping_iter_next(&piter, nid)) {
2616                         if (nid_is_nid4(nid))
2617                                 continue;
2618                         /* nid has already been copied in */
2619                         return true;
2620                 }
2621                 /* no large nids ... weird ... ignore the flag
2622                  * and use first nid.
2623                  */
2624         }
2625         /* pi_nids[1] is primary */
2626         if (pi->pi_nnis < 2)
2627                 return false;
2628         lnet_nid4_to_nid(pbuf->pb_info.pi_ni[1].ns_nid, nid);
2629         return true;
2630 }
2631
2632 /* Handle a Reply message. This is the reply to a Ping message. */
2633 static void
2634 lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
2635 {
2636         struct lnet_ping_buffer *pbuf;
2637         struct lnet_nid primary;
2638         int infobytes;
2639         int rc;
2640         bool ping_feat_disc;
2641
2642         spin_lock(&lp->lp_lock);
2643
2644         lp->lp_disc_src_nid = ev->target.nid;
2645         lp->lp_disc_dst_nid = ev->source.nid;
2646
2647         /*
2648          * If some kind of error happened the contents of message
2649          * cannot be used. Set PING_FAILED to trigger a retry.
2650          */
2651         if (ev->status) {
2652                 lp->lp_state |= LNET_PEER_PING_FAILED;
2653                 lp->lp_ping_error = ev->status;
2654                 CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
2655                        ev->status,
2656                        libcfs_nidstr(&lp->lp_primary_nid),
2657                        libcfs_nidstr(&ev->source.nid));
2658                 goto out;
2659         }
2660
2661         pbuf = LNET_PING_INFO_TO_BUFFER(ev->md_start);
2662         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2663                 lnet_swap_pinginfo(pbuf);
2664
2665         /*
2666          * A reply with invalid or corrupted info. Set PING_FAILED to
2667          * trigger a retry.
2668          */
2669         rc = lnet_ping_info_validate(&pbuf->pb_info);
2670         if (rc) {
2671                 lp->lp_state |= LNET_PEER_PING_FAILED;
2672                 lp->lp_ping_error = 0;
2673                 CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
2674                        libcfs_nidstr(&lp->lp_primary_nid), rc);
2675                 goto out;
2676         }
2677
2678         /*
2679          * The peer may have discovery disabled at its end. Set
2680          * NO_DISCOVERY as appropriate.
2681          */
2682         ping_feat_disc = pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY;
2683         if (!ping_feat_disc || lnet_peer_discovery_disabled) {
2684                 CDEBUG(D_NET, "Peer %s has discovery %s, local discovery %s\n",
2685                        libcfs_nidstr(&lp->lp_primary_nid),
2686                        ping_feat_disc ? "enabled" : "disabled",
2687                        lnet_peer_discovery_disabled ? "disabled" : "enabled");
2688
2689                 /* Detect whether this peer has toggled discovery from on to
2690                  * off and whether we can delete and re-create the peer. Peers
2691                  * that were manually configured cannot be deleted by discovery.
2692                  * We need to delete this peer and re-create it if the peer was
2693                  * not configured manually, is currently considered DD capable,
2694                  * and either:
2695                  * 1. We've already discovered the peer (the peer has toggled
2696                  *    the discovery feature from on to off), or
2697                  * 2. The peer is considered MR, but it was not user configured
2698                  *    (this was a "temporary" peer created via the kernel APIs
2699                  *     that we're discovering for the first time)
2700                  */
2701                 if (!(lp->lp_state & (LNET_PEER_CONFIGURED |
2702                                       LNET_PEER_NO_DISCOVERY)) &&
2703                     (lp->lp_state & (LNET_PEER_DISCOVERED |
2704                                      LNET_PEER_MULTI_RAIL))) {
2705                         CDEBUG(D_NET, "Marking %s:0x%x for deletion\n",
2706                                libcfs_nidstr(&lp->lp_primary_nid),
2707                                lp->lp_state);
2708                         lp->lp_state |= LNET_PEER_MARK_DELETION;
2709                 }
2710                 lp->lp_state |= LNET_PEER_NO_DISCOVERY;
2711         } else {
2712                 CDEBUG(D_NET, "Peer %s has discovery enabled\n",
2713                        libcfs_nidstr(&lp->lp_primary_nid));
2714                 lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
2715         }
2716
2717         /*
2718          * Update the MULTI_RAIL flag based on the reply. If the peer
2719          * was configured with DLC then the setting should match what
2720          * DLC put in.
2721          */
2722         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
2723                 if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2724                         CDEBUG(D_NET, "peer %s(%p) is MR\n",
2725                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2726                 } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
2727                         CWARN("Reply says %s is Multi-Rail, DLC says not\n",
2728                               libcfs_nidstr(&lp->lp_primary_nid));
2729                 } else if (lnet_peer_discovery_disabled) {
2730                         CDEBUG(D_NET,
2731                                "peer %s(%p) not MR: DD disabled locally\n",
2732                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2733                 } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
2734                         CDEBUG(D_NET,
2735                                "peer %s(%p) not MR: DD disabled remotely\n",
2736                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2737                 } else {
2738                         CDEBUG(D_NET, "peer %s(%p) is MR capable\n",
2739                                libcfs_nidstr(&lp->lp_primary_nid), lp);
2740                         lp->lp_state |= LNET_PEER_MULTI_RAIL;
2741                         lnet_peer_clr_non_mr_pref_nids(lp);
2742                 }
2743         } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
2744                 if (lp->lp_state & LNET_PEER_CONFIGURED) {
2745                         CWARN("DLC says %s is Multi-Rail, Reply says not\n",
2746                               libcfs_nidstr(&lp->lp_primary_nid));
2747                 } else {
2748                         CERROR("Multi-Rail state vanished from %s\n",
2749                                libcfs_nidstr(&lp->lp_primary_nid));
2750                         lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
2751                 }
2752         }
2753
2754         infobytes = lnet_ping_info_size(&pbuf->pb_info);
2755         /*
2756          * Make sure we'll allocate the correct size ping buffer when
2757          * pinging the peer.
2758          */
2759         if (lp->lp_data_bytes < infobytes)
2760                 lp->lp_data_bytes = infobytes;
2761
2762         /* Check for truncation of the Reply. Clear PING_SENT and set
2763          * PING_FAILED to trigger a retry.
2764          */
2765         if (pbuf->pb_nbytes < infobytes) {
2766                 if (the_lnet.ln_push_target_nbytes < infobytes)
2767                         the_lnet.ln_push_target_nbytes = infobytes;
2768                 lp->lp_state |= LNET_PEER_PING_FAILED;
2769                 lp->lp_ping_error = 0;
2770                 CDEBUG(D_NET, "Truncated Reply from %s (%d bytes)\n",
2771                        libcfs_nidstr(&lp->lp_primary_nid), infobytes);
2772                 goto out;
2773         }
2774
2775         /*
2776          * Check the sequence numbers in the reply. These are only
2777          * available if the reply came from a Multi-Rail peer.
2778          */
2779         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
2780             find_primary(&primary, pbuf) &&
2781             nid_same(&lp->lp_primary_nid, &primary)) {
2782                 if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno)
2783                         CDEBUG(D_NET, "peer %s: seq# got %u have %u. peer rebooted?\n",
2784                                 libcfs_nidstr(&lp->lp_primary_nid),
2785                                 LNET_PING_BUFFER_SEQNO(pbuf),
2786                                 lp->lp_peer_seqno);
2787
2788                 lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
2789         }
2790
2791         /* We're happy with the state of the data in the buffer. */
2792         CDEBUG(D_NET, "peer %s data present %u. state = 0x%x\n",
2793                libcfs_nidstr(&lp->lp_primary_nid), lp->lp_peer_seqno,
2794                lp->lp_state);
2795         if (lp->lp_state & LNET_PEER_DATA_PRESENT)
2796                 lnet_ping_buffer_decref(lp->lp_data);
2797         else
2798                 lp->lp_state |= LNET_PEER_DATA_PRESENT;
2799         lnet_ping_buffer_addref(pbuf);
2800         lp->lp_data = pbuf;
2801 out:
2802         lp->lp_state &= ~LNET_PEER_PING_SENT;
2803         spin_unlock(&lp->lp_lock);
2804 }
2805
2806 /*
2807  * Send event handling. Only matters for error cases, where we clean
2808  * up state on the peer and peer_ni that would otherwise be updated in
2809  * the REPLY event handler for a successful Ping, and the ACK event
2810  * handler for a successful Push.
2811  */
2812 static int
2813 lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
2814 {
2815         int rc = 0;
2816
2817         if (!ev->status)
2818                 goto out;
2819
2820         spin_lock(&lp->lp_lock);
2821         if (ev->msg_type == LNET_MSG_GET) {
2822                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2823                 lp->lp_state |= LNET_PEER_PING_FAILED;
2824                 lp->lp_ping_error = ev->status;
2825         } else { /* ev->msg_type == LNET_MSG_PUT */
2826                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2827                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2828                 lp->lp_push_error = ev->status;
2829         }
2830         spin_unlock(&lp->lp_lock);
2831         rc = LNET_REDISCOVER_PEER;
2832 out:
2833         CDEBUG(D_NET, "%s Send to %s: %d\n",
2834                 (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
2835                 libcfs_nidstr(&ev->target.nid), rc);
2836         return rc;
2837 }
2838
2839 /*
2840  * Unlink event handling. This event is only seen if a call to
2841  * LNetMDUnlink() caused the event to be unlinked. If this call was
2842  * made after the event was set up in LNetGet() or LNetPut() then we
2843  * assume the Ping or Push timed out.
2844  */
2845 static void
2846 lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
2847 {
2848         spin_lock(&lp->lp_lock);
2849         /* We've passed through LNetGet() */
2850         if (lp->lp_state & LNET_PEER_PING_SENT) {
2851                 lp->lp_state &= ~LNET_PEER_PING_SENT;
2852                 lp->lp_state |= LNET_PEER_PING_FAILED;
2853                 lp->lp_ping_error = -ETIMEDOUT;
2854                 CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
2855                         libcfs_nidstr(&lp->lp_primary_nid));
2856         }
2857         /* We've passed through LNetPut() */
2858         if (lp->lp_state & LNET_PEER_PUSH_SENT) {
2859                 lp->lp_state &= ~LNET_PEER_PUSH_SENT;
2860                 lp->lp_state |= LNET_PEER_PUSH_FAILED;
2861                 lp->lp_push_error = -ETIMEDOUT;
2862                 CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
2863                         libcfs_nidstr(&lp->lp_primary_nid));
2864         }
2865         spin_unlock(&lp->lp_lock);
2866 }
2867
2868 /*
2869  * Event handler for the discovery EQ.
2870  *
2871  * Called with lnet_res_lock(cpt) held. The cpt is the
2872  * lnet_cpt_of_cookie() of the md handle cookie.
2873  */
2874 static void lnet_discovery_event_handler(struct lnet_event *event)
2875 {
2876         struct lnet_peer *lp = event->md_user_ptr;
2877         struct lnet_ping_buffer *pbuf;
2878         int rc;
2879
2880         /* discovery needs to take another look */
2881         rc = LNET_REDISCOVER_PEER;
2882
2883         CDEBUG(D_NET, "Received event: %d\n", event->type);
2884
2885         switch (event->type) {
2886         case LNET_EVENT_ACK:
2887                 lnet_discovery_event_ack(lp, event);
2888                 break;
2889         case LNET_EVENT_REPLY:
2890                 lnet_discovery_event_reply(lp, event);
2891                 break;
2892         case LNET_EVENT_SEND:
2893                 /* Only send failure triggers a retry. */
2894                 rc = lnet_discovery_event_send(lp, event);
2895                 break;
2896         case LNET_EVENT_UNLINK:
2897                 /* LNetMDUnlink() was called */
2898                 lnet_discovery_event_unlink(lp, event);
2899                 break;
2900         default:
2901                 /* Invalid events. */
2902                 LBUG();
2903         }
2904         lnet_net_lock(LNET_LOCK_EX);
2905         if (event->unlinked) {
2906                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
2907                 lnet_ping_buffer_decref(pbuf);
2908                 lnet_peer_decref_locked(lp);
2909         }
2910
2911         /* put peer back at end of request queue, if discovery not already
2912          * done */
2913         if (rc == LNET_REDISCOVER_PEER && !lnet_peer_is_uptodate(lp) &&
2914             lnet_peer_queue_for_discovery(lp)) {
2915                 list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
2916                 wake_up(&the_lnet.ln_dc_waitq);
2917         }
2918         lnet_net_unlock(LNET_LOCK_EX);
2919 }
2920
2921 u32 *ping_iter_first(struct lnet_ping_iter *pi,
2922                      struct lnet_ping_buffer *pbuf,
2923                      struct lnet_nid *nid)
2924 {
2925         pi->pinfo = &pbuf->pb_info;
2926         pi->pos = &pbuf->pb_info.pi_ni;
2927         pi->end = (void *)pi->pinfo +
2928                   min_t(int, pbuf->pb_nbytes,
2929                         lnet_ping_info_size(pi->pinfo));
2930         /* lnet_ping_info_validiate ensures there will be one
2931          * lnet_ni_status at the start
2932          */
2933         if (nid)
2934                 lnet_nid4_to_nid(pbuf->pb_info.pi_ni[0].ns_nid, nid);
2935
2936         pi->pos += sizeof(struct lnet_ni_status);
2937         return &pbuf->pb_info.pi_ni[0].ns_status;
2938 }
2939
2940 u32 *ping_iter_next(struct lnet_ping_iter *pi, struct lnet_nid *nid)
2941 {
2942         int off = offsetof(struct lnet_ping_info, pi_ni[pi->pinfo->pi_nnis]);
2943
2944         if (pi->pos < ((void *)pi->pinfo + off)) {
2945                 struct lnet_ni_status *ns = pi->pos;
2946
2947                 pi->pos = ns + 1;
2948                 if (pi->pos > pi->end)
2949                         return NULL;
2950                 if (nid)
2951                         lnet_nid4_to_nid(ns->ns_nid, nid);
2952                 return &ns->ns_status;
2953         }
2954
2955         while (pi->pinfo->pi_features & LNET_PING_FEAT_LARGE_ADDR) {
2956                 struct lnet_ni_large_status *lns = pi->pos;
2957
2958                 if (pi->pos + 8 > pi->end)
2959                         /* Not safe to examine next */
2960                         return NULL;
2961                 pi->pos = lnet_ping_sts_next(lns);
2962                 if (pi->pos > pi->end)
2963                         return NULL;
2964                 if (NID_BYTES(&lns->ns_nid) > sizeof(struct lnet_nid))
2965                         continue;
2966                 if (nid)
2967                         *nid = lns->ns_nid;
2968                 return &lns->ns_status;
2969         }
2970         return NULL;
2971 }
2972
2973 int ping_info_count_entries(struct lnet_ping_buffer *pbuf)
2974 {
2975         struct lnet_ping_iter pi;
2976         u32 *st;
2977         int nnis = 0;
2978
2979         for (st = ping_iter_first(&pi, pbuf, NULL); st;
2980              st = ping_iter_next(&pi, NULL))
2981                 nnis += 1;
2982
2983         return nnis;
2984 }
2985
2986 /*
2987  * Build a peer from incoming data.
2988  *
2989  * The NIDs in the incoming data are supposed to be structured as follows:
2990  *  - loopback
2991  *  - primary NID
2992  *  - other NIDs in same net
2993  *  - NIDs in second net
2994  *  - NIDs in third net
2995  *  - ...
2996  * This due to the way the list of NIDs in the data is created.
2997  *
2998  * Note that this function will mark the peer uptodate unless an
2999  * ENOMEM is encontered. All other errors are due to a conflict
3000  * between the DLC configuration and what discovery sees. We treat DLC
3001  * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
3002  * peer from becoming stuck in discovery.
3003  */
3004 static int lnet_peer_merge_data(struct lnet_peer *lp,
3005                                 struct lnet_ping_buffer *pbuf)
3006 {
3007         struct lnet_peer_net *lpn;
3008         struct lnet_peer_ni *lpni;
3009         struct lnet_nid *curnis = NULL;
3010         struct lnet_ni_large_status *addnis = NULL;
3011         struct lnet_nid *delnis = NULL;
3012         struct lnet_ping_iter pi;
3013         struct lnet_nid nid;
3014         u32 *stp;
3015         struct lnet_nid primary = {};
3016         bool want_large_primary;
3017         unsigned int flags;
3018         int ncurnis;
3019         int naddnis;
3020         int ndelnis;
3021         int nnis = 0;
3022         int i;
3023         int j;
3024         int rc;
3025
3026         flags = LNET_PEER_DISCOVERED;
3027         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3028                 flags |= LNET_PEER_MULTI_RAIL;
3029
3030         /*
3031          * Cache the routing feature for the peer; whether it is enabled
3032          * for disabled as reported by the remote peer.
3033          */
3034         spin_lock(&lp->lp_lock);
3035         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED))
3036                 lp->lp_state |= LNET_PEER_ROUTER_ENABLED;
3037         else
3038                 lp->lp_state &= ~LNET_PEER_ROUTER_ENABLED;
3039         spin_unlock(&lp->lp_lock);
3040
3041         nnis = ping_info_count_entries(pbuf);
3042         nnis = max_t(int, lp->lp_nnis, nnis);
3043         CFS_ALLOC_PTR_ARRAY(curnis, nnis);
3044         CFS_ALLOC_PTR_ARRAY(addnis, nnis);
3045         CFS_ALLOC_PTR_ARRAY(delnis, nnis);
3046         if (!curnis || !addnis || !delnis) {
3047                 rc = -ENOMEM;
3048                 goto out;
3049         }
3050         ncurnis = 0;
3051         naddnis = 0;
3052         ndelnis = 0;
3053
3054         /* Construct the list of NIDs present in peer. */
3055         lpni = NULL;
3056         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
3057                 curnis[ncurnis++] = lpni->lpni_nid;
3058
3059         /* Check for NIDs in pbuf not present in curnis[].
3060          * Skip the first, which is loop-back.  Take second as
3061          * primary, unless a large primary is found.
3062          */
3063         ping_iter_first(&pi, pbuf, NULL);
3064         stp = ping_iter_next(&pi, &nid);
3065         if (stp)
3066                 primary = nid;
3067         want_large_primary = (pbuf->pb_info.pi_features &
3068                               LNET_PING_FEAT_PRIMARY_LARGE);
3069         for (; stp; stp = ping_iter_next(&pi, &nid)) {
3070                 for (j = 0; j < ncurnis; j++)
3071                         if (nid_same(&nid, &curnis[j]))
3072                                 break;
3073                 if (j == ncurnis) {
3074                         addnis[naddnis].ns_nid = nid;
3075                         addnis[naddnis].ns_status = *stp;
3076                         naddnis += 1;
3077                 }
3078                 if (want_large_primary && nid.nid_size) {
3079                         primary = nid;
3080                         want_large_primary = false;
3081                 }
3082         }
3083         /*
3084          * Check for NIDs in curnis[] not present in pbuf.
3085          * The nested loop starts at 1 to skip the loopback NID.
3086          *
3087          * But never add the loopback NID to delnis[]: if it is
3088          * present in curnis[] then this peer is for this node.
3089          */
3090         for (i = 0; i < ncurnis; i++) {
3091                 if (nid_is_lo0(&curnis[i]))
3092                         continue;
3093                 ping_iter_first(&pi, pbuf, NULL);
3094                 while ((stp = ping_iter_next(&pi, &nid)) != NULL) {
3095                         if (nid_same(&curnis[i], &nid)) {
3096                                 /*
3097                                  * update the information we cache for the
3098                                  * peer with the latest information we
3099                                  * received
3100                                  */
3101                                 lpni = lnet_peer_ni_find_locked(&curnis[i]);
3102                                 if (lpni) {
3103                                         lpni->lpni_ns_status = *stp;
3104                                         lnet_peer_ni_decref_locked(lpni);
3105                                 }
3106                                 break;
3107                         }
3108                 }
3109                 if (!stp)
3110                         delnis[ndelnis++] = curnis[i];
3111         }
3112
3113         /*
3114          * If we get here and the discovery is disabled then we don't want
3115          * to add or delete any NIs. We just updated the ones we have some
3116          * information on, and call it a day
3117          */
3118         rc = 0;
3119         if (lnet_is_discovery_disabled(lp))
3120                 goto out;
3121
3122         for (i = 0; i < naddnis; i++) {
3123                 rc = lnet_peer_add_nid(lp, &addnis[i].ns_nid, flags);
3124                 if (rc) {
3125                         CERROR("Error adding NID %s to peer %s: %d\n",
3126                                libcfs_nidstr(&addnis[i].ns_nid),
3127                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3128                         if (rc == -ENOMEM)
3129                                 goto out;
3130                 }
3131                 lpni = lnet_peer_ni_find_locked(&addnis[i].ns_nid);
3132                 if (lpni) {
3133                         lpni->lpni_ns_status = addnis[i].ns_status;
3134                         lnet_peer_ni_decref_locked(lpni);
3135                 }
3136         }
3137
3138         for (i = 0; i < ndelnis; i++) {
3139                 /*
3140                  * for routers it's okay to delete the primary_nid because
3141                  * the upper layers don't really rely on it. So if we're
3142                  * being told that the router changed its primary_nid
3143                  * then it's okay to delete it.
3144                  */
3145                 if (lp->lp_rtr_refcount > 0)
3146                         flags |= LNET_PEER_RTR_NI_FORCE_DEL;
3147                 rc = lnet_peer_del_nid(lp, &delnis[i], flags);
3148                 if (rc) {
3149                         CERROR("Error deleting NID %s from peer %s: %d\n",
3150                                libcfs_nidstr(&delnis[i]),
3151                                libcfs_nidstr(&lp->lp_primary_nid), rc);
3152                         if (rc == -ENOMEM)
3153                                 goto out;
3154                 }
3155         }
3156
3157         /* The peer net for the primary NID should be the first entry in the
3158          * peer's lp_peer_nets list, and the peer NI for the primary NID should
3159          * be the first entry in its peer net's lpn_peer_nis list.
3160          */
3161         find_primary(&nid, pbuf);
3162         lpni = lnet_peer_ni_find_locked(&nid);
3163         if (!lpni) {
3164                 CERROR("Internal error: Failed to lookup peer NI for primary NID: %s\n",
3165                        libcfs_nidstr(&nid));
3166                 goto out;
3167         }
3168
3169         lnet_peer_ni_decref_locked(lpni);
3170
3171         lpn = lpni->lpni_peer_net;
3172         if (lpn->lpn_peer_nets.prev != &lp->lp_peer_nets)
3173                 list_move(&lpn->lpn_peer_nets, &lp->lp_peer_nets);
3174
3175         if (lpni->lpni_peer_nis.prev != &lpni->lpni_peer_net->lpn_peer_nis)
3176                 list_move(&lpni->lpni_peer_nis,
3177                           &lpni->lpni_peer_net->lpn_peer_nis);
3178
3179         /*
3180          * Errors other than -ENOMEM are due to peers having been
3181          * configured with DLC. Ignore these because DLC overrides
3182          * Discovery.
3183          */
3184         rc = 0;
3185 out:
3186         /* If this peer is a gateway, invoke the routing callback to update
3187          * the associated route status
3188          */
3189         if (lp->lp_rtr_refcount > 0)
3190                 lnet_router_discovery_ping_reply(lp, pbuf);
3191
3192         CFS_FREE_PTR_ARRAY(curnis, nnis);
3193         CFS_FREE_PTR_ARRAY(addnis, nnis);
3194         CFS_FREE_PTR_ARRAY(delnis, nnis);
3195         lnet_ping_buffer_decref(pbuf);
3196         CDEBUG(D_NET, "peer %s (%p): %d\n",
3197                libcfs_nidstr(&lp->lp_primary_nid), lp, rc);
3198
3199         if (rc) {
3200                 spin_lock(&lp->lp_lock);
3201                 lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
3202                 lp->lp_state |= LNET_PEER_FORCE_PING;
3203                 spin_unlock(&lp->lp_lock);
3204         }
3205         return rc;
3206 }
3207
3208 /*
3209  * The data in pbuf says lp is its primary peer, but the data was
3210  * received by a different peer. Try to update lp with the data.
3211  */
3212 static int
3213 lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
3214 {
3215         struct lnet_handle_md mdh;
3216
3217         /* Queue lp for discovery, and force it on the request queue. */
3218         lnet_net_lock(LNET_LOCK_EX);
3219         if (lnet_peer_queue_for_discovery(lp))
3220                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
3221         lnet_net_unlock(LNET_LOCK_EX);
3222
3223         LNetInvalidateMDHandle(&mdh);
3224
3225         /*
3226          * Decide whether we can move the peer to the DATA_PRESENT state.
3227          *
3228          * We replace stale data for a multi-rail peer, repair PING_FAILED
3229          * status, and preempt FORCE_PING.
3230          *
3231          * If after that we have DATA_PRESENT, we merge it into this peer.
3232          */
3233         spin_lock(&lp->lp_lock);
3234         if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
3235                 if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
3236                         lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
3237                 } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3238                         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3239                         lnet_ping_buffer_decref(pbuf);
3240                         pbuf = lp->lp_data;
3241                         lp->lp_data = NULL;
3242                 }
3243         }
3244         if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
3245                 lnet_ping_buffer_decref(lp->lp_data);
3246                 lp->lp_data = NULL;
3247                 lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3248         }
3249         if (lp->lp_state & LNET_PEER_PING_FAILED) {
3250                 mdh = lp->lp_ping_mdh;
3251                 LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3252                 lp->lp_state &= ~LNET_PEER_PING_FAILED;
3253                 lp->lp_ping_error = 0;
3254         }
3255         if (lp->lp_state & LNET_PEER_FORCE_PING)
3256                 lp->lp_state &= ~LNET_PEER_FORCE_PING;
3257         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3258         spin_unlock(&lp->lp_lock);
3259
3260         if (!LNetMDHandleIsInvalid(mdh))
3261                 LNetMDUnlink(mdh);
3262
3263         if (pbuf)
3264                 return lnet_peer_merge_data(lp, pbuf);
3265
3266         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3267         return 0;
3268 }
3269
3270 static bool lnet_is_nid_in_ping_info(struct lnet_nid *nid,
3271                                      struct lnet_ping_buffer *pbuf)
3272 {
3273         struct lnet_ping_iter pi;
3274         struct lnet_nid pnid;
3275         u32 *st;
3276
3277         for (st = ping_iter_first(&pi, pbuf, &pnid);
3278              st;
3279              st = ping_iter_next(&pi, &pnid))
3280                 if (nid_same(nid, &pnid))
3281                         return true;
3282         return false;
3283 }
3284
3285 /* Delete a peer that has been marked for deletion. NB: when this peer was added
3286  * to the discovery queue a reference was taken that will prevent the peer from
3287  * actually being freed by this function. After this function exits the
3288  * discovery thread should call lnet_peer_discovery_complete() which will
3289  * drop that reference as well as wake any waiters that may also be holding a
3290  * ref on the peer
3291  */
3292 static int lnet_peer_deletion(struct lnet_peer *lp)
3293 __must_hold(&lp->lp_lock)
3294 {
3295         struct list_head rlist;
3296         struct lnet_route *route, *tmp;
3297         int sensitivity = lp->lp_health_sensitivity;
3298         int rc = 0;
3299
3300         INIT_LIST_HEAD(&rlist);
3301
3302         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3303                libcfs_nidstr(&lp->lp_primary_nid), lp, lp->lp_state);
3304
3305         /* no-op if lnet_peer_del() has already been called on this peer */
3306         if (lp->lp_state & LNET_PEER_MARK_DELETED)
3307                 goto clear_discovering;
3308
3309         spin_unlock(&lp->lp_lock);
3310
3311         mutex_lock(&the_lnet.ln_api_mutex);
3312         if (the_lnet.ln_state != LNET_STATE_RUNNING ||
3313             the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
3314                 mutex_unlock(&the_lnet.ln_api_mutex);
3315                 spin_lock(&lp->lp_lock);
3316                 rc = -ESHUTDOWN;
3317                 goto clear_discovering;
3318         }
3319
3320         lnet_peer_cancel_discovery(lp);
3321         lnet_net_lock(LNET_LOCK_EX);
3322         list_for_each_entry_safe(route, tmp,
3323                                  &lp->lp_routes,
3324                                  lr_gwlist)
3325                 lnet_move_route(route, NULL, &rlist);
3326
3327         /* lnet_peer_del_locked() deletes all the peer NIs owned by this peer */
3328         rc = lnet_peer_del_locked(lp);
3329         if (rc)
3330                 CNETERR("Internal error: Unable to delete peer %s rc %d\n",
3331                         libcfs_nidstr(&lp->lp_primary_nid), rc);
3332
3333         lnet_net_unlock(LNET_LOCK_EX);
3334
3335         list_for_each_entry_safe(route, tmp,
3336                                  &rlist, lr_list) {
3337                 /* re-add these routes */
3338                 lnet_add_route(route->lr_net,
3339                                route->lr_hops,
3340                                &route->lr_nid,
3341                                route->lr_priority,
3342                                sensitivity);
3343                 LIBCFS_FREE(route, sizeof(*route));
3344         }
3345
3346         mutex_unlock(&the_lnet.ln_api_mutex);
3347
3348         spin_lock(&lp->lp_lock);
3349
3350         rc = 0;
3351
3352 clear_discovering:
3353         lp->lp_state &= ~(LNET_PEER_DISCOVERING | LNET_PEER_FORCE_PING |
3354                           LNET_PEER_FORCE_PUSH);
3355
3356         return rc;
3357 }
3358
3359 /*
3360  * Update a peer using the data received.
3361  */
3362 static int lnet_peer_data_present(struct lnet_peer *lp)
3363 __must_hold(&lp->lp_lock)
3364 {
3365         struct lnet_ping_buffer *pbuf;
3366         struct lnet_peer_ni *lpni;
3367         struct lnet_nid nid;
3368         unsigned int flags;
3369         int rc = 0;
3370
3371         pbuf = lp->lp_data;
3372         lp->lp_data = NULL;
3373         lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
3374         lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
3375         spin_unlock(&lp->lp_lock);
3376
3377         /*
3378          * Modifications of peer structures are done while holding the
3379          * ln_api_mutex. A global lock is required because we may be
3380          * modifying multiple peer structures, and a mutex greatly
3381          * simplifies memory management.
3382          *
3383          * The actual changes to the data structures must also protect
3384          * against concurrent lookups, for which the lnet_net_lock in
3385          * LNET_LOCK_EX mode is used.
3386          */
3387         mutex_lock(&the_lnet.ln_api_mutex);
3388         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3389                 rc = -ESHUTDOWN;
3390                 goto out;
3391         }
3392
3393         /*
3394          * If this peer is not on the peer list then it is being torn
3395          * down, and our reference count may be all that is keeping it
3396          * alive. Don't do any work on it.
3397          */
3398         if (list_empty(&lp->lp_peer_list)) {
3399                 lnet_ping_buffer_decref(pbuf);
3400                 goto out;
3401         }
3402
3403         flags = LNET_PEER_DISCOVERED;
3404         if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
3405                 flags |= LNET_PEER_MULTI_RAIL;
3406
3407         /*
3408          * Check whether the primary NID in the message matches the
3409          * primary NID of the peer. If it does, update the peer, if
3410          * it it does not, check whether there is already a peer with
3411          * that primary NID. If no such peer exists, try to update
3412          * the primary NID of the current peer (allowed if it was
3413          * created due to message traffic) and complete the update.
3414          * If the peer did exist, hand off the data to it.
3415          *
3416          * The peer for the loopback interface is a special case: this
3417          * is the peer for the local node, and we want to set its
3418          * primary NID to the correct value here. Moreover, this peer
3419          * can show up with only the loopback NID in the ping buffer.
3420          */
3421         if (!find_primary(&nid, pbuf)) {
3422                 lnet_ping_buffer_decref(pbuf);
3423                 goto out;
3424         }
3425         if (nid_is_lo0(&lp->lp_primary_nid)) {
3426                 rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3427                 if (rc)
3428                         lnet_ping_buffer_decref(pbuf);
3429                 else
3430                         rc = lnet_peer_merge_data(lp, pbuf);
3431         /*
3432          * if the primary nid of the peer is present in the ping info returned
3433          * from the peer, but it's not the local primary peer we have
3434          * cached and discovery is disabled, then we don't want to update
3435          * our local peer info, by adding or removing NIDs, we just want
3436          * to update the status of the nids that we currently have
3437          * recorded in that peer.
3438          */
3439         } else if (nid_same(&lp->lp_primary_nid, &nid) ||
3440                    (lnet_is_nid_in_ping_info(&lp->lp_primary_nid, pbuf) &&
3441                     lnet_is_discovery_disabled(lp))) {
3442                 rc = lnet_peer_merge_data(lp, pbuf);
3443         } else {
3444                 lpni = lnet_peer_ni_find_locked(&nid);
3445                 if (!lpni || lp == lpni->lpni_peer_net->lpn_peer) {
3446                         rc = lnet_peer_set_primary_nid(lp, &nid, flags);
3447                         if (rc) {
3448                                 CERROR("Primary NID error %s versus %s: %d\n",
3449                                        libcfs_nidstr(&lp->lp_primary_nid),
3450                                        libcfs_nidstr(&nid), rc);
3451                                 lnet_ping_buffer_decref(pbuf);
3452                         } else {
3453                                 rc = lnet_peer_merge_data(lp, pbuf);
3454                         }
3455                         if (lpni)
3456                                 lnet_peer_ni_decref_locked(lpni);
3457                 } else {
3458                         struct lnet_peer *new_lp;
3459                         new_lp = lpni->lpni_peer_net->lpn_peer;
3460                         /*
3461                          * if lp has discovery/MR enabled that means new_lp
3462                          * should have discovery/MR enabled as well, since
3463                          * it's the same peer, which we're about to merge
3464                          */
3465                         spin_lock(&lp->lp_lock);
3466                         spin_lock(&new_lp->lp_lock);
3467                         if (!(lp->lp_state & LNET_PEER_NO_DISCOVERY))
3468                                 new_lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
3469                         if (lp->lp_state & LNET_PEER_MULTI_RAIL)
3470                                 new_lp->lp_state |= LNET_PEER_MULTI_RAIL;
3471                         /* If we're processing a ping reply then we may be
3472                          * about to send a push to the peer that we ping'd.
3473                          * Since the ping reply that we're processing was
3474                          * received by lp, we need to set the discovery source
3475                          * NID for new_lp to the NID stored in lp.
3476                          */
3477                         if (!LNET_NID_IS_ANY(&lp->lp_disc_src_nid)) {
3478                                 new_lp->lp_disc_src_nid = lp->lp_disc_src_nid;
3479                                 new_lp->lp_disc_dst_nid = lp->lp_disc_dst_nid;
3480                         }
3481                         spin_unlock(&new_lp->lp_lock);
3482                         spin_unlock(&lp->lp_lock);
3483
3484                         rc = lnet_peer_set_primary_data(new_lp, pbuf);
3485                         lnet_consolidate_routes_locked(lp, new_lp);
3486                         lnet_peer_ni_decref_locked(lpni);
3487                 }
3488         }
3489 out:
3490         CDEBUG(D_NET, "peer %s(%p): %d. state = 0x%x\n",
3491                libcfs_nidstr(&lp->lp_primary_nid), lp, rc,
3492                lp->lp_state);
3493         mutex_unlock(&the_lnet.ln_api_mutex);
3494
3495         spin_lock(&lp->lp_lock);
3496         /* Tell discovery to re-check the peer immediately. */
3497         if (!rc)
3498                 rc = LNET_REDISCOVER_PEER;
3499         return rc;
3500 }
3501
3502 /*
3503  * A ping failed. Clear the PING_FAILED state and set the
3504  * FORCE_PING state, to ensure a retry even if discovery is
3505  * disabled. This avoids being left with incorrect state.
3506  */
3507 static int lnet_peer_ping_failed(struct lnet_peer *lp)
3508 __must_hold(&lp->lp_lock)
3509 {
3510         struct lnet_handle_md mdh;
3511         int rc;
3512
3513         mdh = lp->lp_ping_mdh;
3514         LNetInvalidateMDHandle(&lp->lp_ping_mdh);
3515         lp->lp_state &= ~LNET_PEER_PING_FAILED;
3516         lp->lp_state |= LNET_PEER_FORCE_PING;
3517         rc = lp->lp_ping_error;
3518         lp->lp_ping_error = 0;
3519         spin_unlock(&lp->lp_lock);
3520
3521         if (!LNetMDHandleIsInvalid(mdh))
3522                 LNetMDUnlink(mdh);
3523
3524         CDEBUG(D_NET, "peer %s:%d\n",
3525                libcfs_nidstr(&lp->lp_primary_nid), rc);
3526
3527         spin_lock(&lp->lp_lock);
3528         return rc ? rc : LNET_REDISCOVER_PEER;
3529 }
3530
3531 /* Active side of ping. */
3532 static int lnet_peer_send_ping(struct lnet_peer *lp)
3533 __must_hold(&lp->lp_lock)
3534 {
3535         int bytes;
3536         int rc;
3537         int cpt;
3538
3539         lp->lp_state |= LNET_PEER_PING_SENT;
3540         lp->lp_state &= ~LNET_PEER_FORCE_PING;
3541         spin_unlock(&lp->lp_lock);
3542
3543         cpt = lnet_net_lock_current();
3544         /* Refcount for MD. */
3545         lnet_peer_addref_locked(lp);
3546         lnet_net_unlock(cpt);
3547
3548         bytes = max_t(int, lp->lp_data_bytes, LNET_PING_INFO_MIN_SIZE);
3549
3550         rc = lnet_send_ping(&lp->lp_primary_nid, &lp->lp_ping_mdh, bytes, lp,
3551                             the_lnet.ln_dc_handler, false);
3552         /* if LNetMDBind in lnet_send_ping fails we need to decrement the
3553          * refcount on the peer, otherwise LNetMDUnlink will be called
3554          * which will eventually do that.
3555          */
3556         if (rc > 0) {
3557                 lnet_net_lock(cpt);
3558                 lnet_peer_decref_locked(lp);
3559                 lnet_net_unlock(cpt);
3560                 rc = -rc; /* change the rc to negative value */
3561                 goto fail_error;
3562         } else if (rc < 0) {
3563                 goto fail_error;
3564         }
3565
3566         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3567
3568         spin_lock(&lp->lp_lock);
3569         return 0;
3570
3571 fail_error:
3572         CDEBUG(D_NET, "peer %s: %d\n", libcfs_nidstr(&lp->lp_primary_nid), rc);
3573         /*
3574          * The errors that get us here are considered hard errors and
3575          * cause Discovery to terminate. So we clear PING_SENT, but do
3576          * not set either PING_FAILED or FORCE_PING. In fact we need
3577          * to clear PING_FAILED, because the unlink event handler will
3578          * have set it if we called LNetMDUnlink() above.
3579          */
3580         spin_lock(&lp->lp_lock);
3581         lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
3582         return rc;
3583 }
3584
3585 /*
3586  * This function exists because you cannot call LNetMDUnlink() from an
3587  * event handler.
3588  */
3589 static int lnet_peer_push_failed(struct lnet_peer *lp)
3590 __must_hold(&lp->lp_lock)
3591 {
3592         struct lnet_handle_md mdh;
3593         int rc;
3594
3595         mdh = lp->lp_push_mdh;
3596         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3597         lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
3598         rc = lp->lp_push_error;
3599         lp->lp_push_error = 0;
3600         spin_unlock(&lp->lp_lock);
3601
3602         if (!LNetMDHandleIsInvalid(mdh))
3603                 LNetMDUnlink(mdh);
3604
3605         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3606         spin_lock(&lp->lp_lock);
3607         return rc ? rc : LNET_REDISCOVER_PEER;
3608 }
3609
3610 /*
3611  * Mark the peer as discovered.
3612  */
3613 static int lnet_peer_discovered(struct lnet_peer *lp)
3614 __must_hold(&lp->lp_lock)
3615 {
3616         lp->lp_state |= LNET_PEER_DISCOVERED;
3617         lp->lp_state &= ~(LNET_PEER_DISCOVERING |
3618                           LNET_PEER_REDISCOVER);
3619
3620         lp->lp_dc_error = 0;
3621
3622         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3623
3624         return 0;
3625 }
3626
3627 /* Active side of push. */
3628 static int lnet_peer_send_push(struct lnet_peer *lp)
3629 __must_hold(&lp->lp_lock)
3630 {
3631         struct lnet_ping_buffer *pbuf;
3632         struct lnet_processid id;
3633         struct lnet_md md;
3634         int cpt;
3635         int rc;
3636
3637         /* Don't push to a non-multi-rail peer. */
3638         if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
3639                 lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3640                 /* if peer's NIDs are uptodate then peer is discovered */
3641                 if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
3642                         rc = lnet_peer_discovered(lp);
3643                         return rc;
3644                 }
3645
3646                 return 0;
3647         }
3648
3649         lp->lp_state |= LNET_PEER_PUSH_SENT;
3650         lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
3651         spin_unlock(&lp->lp_lock);
3652
3653         cpt = lnet_net_lock_current();
3654         pbuf = the_lnet.ln_ping_target;
3655         lnet_ping_buffer_addref(pbuf);
3656         lnet_net_unlock(cpt);
3657
3658         /* Push source MD */
3659         md.start     = &pbuf->pb_info;
3660         md.length    = pbuf->pb_nbytes;
3661         md.threshold = 2; /* Put/Ack */
3662         md.max_size  = 0;
3663         md.options   = LNET_MD_TRACK_RESPONSE;
3664         md.handler   = the_lnet.ln_dc_handler;
3665         md.user_ptr  = lp;
3666
3667         rc = LNetMDBind(&md, LNET_UNLINK, &lp->lp_push_mdh);
3668         if (rc) {
3669                 lnet_ping_buffer_decref(pbuf);
3670                 CERROR("Can't bind push source MD: %d\n", rc);
3671                 goto fail_error;
3672         }
3673
3674         cpt = lnet_net_lock_current();
3675         /* Refcount for MD. */
3676         lnet_peer_addref_locked(lp);
3677         id.pid = LNET_PID_LUSTRE;
3678         if (!LNET_NID_IS_ANY(&lp->lp_disc_dst_nid))
3679                 id.nid = lp->lp_disc_dst_nid;
3680         else
3681                 id.nid = lp->lp_primary_nid;
3682         lnet_net_unlock(cpt);
3683
3684         rc = LNetPut(&lp->lp_disc_src_nid, lp->lp_push_mdh,
3685                      LNET_ACK_REQ, &id, LNET_RESERVED_PORTAL,
3686                      LNET_PROTO_PING_MATCHBITS, 0, 0);
3687
3688         /*
3689          * reset the discovery nid. There is no need to restrict sending
3690          * from that source, if we call lnet_push_update_to_peers(). It'll
3691          * get set to a specific NID, if we initiate discovery from the
3692          * scratch
3693          */
3694         lp->lp_disc_src_nid = LNET_ANY_NID;
3695         lp->lp_disc_dst_nid = LNET_ANY_NID;
3696
3697         if (rc)
3698                 goto fail_unlink;
3699
3700         CDEBUG(D_NET, "peer %s\n", libcfs_nidstr(&lp->lp_primary_nid));
3701
3702         spin_lock(&lp->lp_lock);
3703         return 0;
3704
3705 fail_unlink:
3706         LNetMDUnlink(lp->lp_push_mdh);
3707         LNetInvalidateMDHandle(&lp->lp_push_mdh);
3708 fail_error:
3709         CDEBUG(D_NET, "peer %s(%p): %d\n", libcfs_nidstr(&lp->lp_primary_nid),
3710                lp, rc);
3711         /*
3712          * The errors that get us here are considered hard errors and
3713          * cause Discovery to terminate. So we clear PUSH_SENT, but do
3714          * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
3715          * because the unlink event handler will have set it if we
3716          * called LNetMDUnlink() above.
3717          */
3718         spin_lock(&lp->lp_lock);
3719         lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
3720         return rc;
3721 }
3722
3723 /*
3724  * Wait for work to be queued or some other change that must be
3725  * attended to. Returns non-zero if the discovery thread should shut
3726  * down.
3727  */
3728 static int lnet_peer_discovery_wait_for_work(void)
3729 {
3730         int cpt;
3731         int rc = 0;
3732
3733         DEFINE_WAIT(wait);
3734
3735         cpt = lnet_net_lock_current();
3736         for (;;) {
3737                 prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
3738                                 TASK_INTERRUPTIBLE);
3739                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3740                         break;
3741                 if (lnet_push_target_resize_needed() ||
3742                     the_lnet.ln_push_target->pb_needs_post)
3743                         break;
3744                 if (!list_empty(&the_lnet.ln_dc_request))
3745                         break;
3746                 if (!list_empty(&the_lnet.ln_msg_resend))
3747                         break;
3748                 lnet_net_unlock(cpt);
3749
3750                 /*
3751                  * wakeup max every second to check if there are peers that
3752                  * have been stuck on the working queue for greater than
3753                  * the peer timeout.
3754                  */
3755                 schedule_timeout(cfs_time_seconds(1));
3756                 finish_wait(&the_lnet.ln_dc_waitq, &wait);
3757                 cpt = lnet_net_lock_current();
3758         }
3759         finish_wait(&the_lnet.ln_dc_waitq, &wait);
3760
3761         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3762                 rc = -ESHUTDOWN;
3763
3764         lnet_net_unlock(cpt);
3765
3766         CDEBUG(D_NET, "woken: %d\n", rc);
3767
3768         return rc;
3769 }
3770
3771 /*
3772  * Messages that were pending on a destroyed peer will be put on a global
3773  * resend list. The message resend list will be checked by
3774  * the discovery thread when it wakes up, and will resend messages. These
3775  * messages can still be sendable in the case the lpni which was the initial
3776  * cause of the message re-queue was transfered to another peer.
3777  *
3778  * It is possible that LNet could be shutdown while we're iterating
3779  * through the list. lnet_shudown_lndnets() will attempt to access the
3780  * resend list, but will have to wait until the spinlock is released, by
3781  * which time there shouldn't be any more messages on the resend list.
3782  * During shutdown lnet_send() will fail and lnet_finalize() will be called
3783  * for the messages so they can be released. The other case is that
3784  * lnet_shudown_lndnets() can finalize all the messages before this
3785  * function can visit the resend list, in which case this function will be
3786  * a no-op.
3787  */
3788 static void lnet_resend_msgs(void)
3789 {
3790         struct lnet_msg *msg, *tmp;
3791         LIST_HEAD(resend);
3792         int rc;
3793
3794         spin_lock(&the_lnet.ln_msg_resend_lock);
3795         list_splice(&the_lnet.ln_msg_resend, &resend);
3796         spin_unlock(&the_lnet.ln_msg_resend_lock);
3797
3798         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
3799                 list_del_init(&msg->msg_list);
3800                 rc = lnet_send(&msg->msg_src_nid_param, msg,
3801                                &msg->msg_rtr_nid_param);
3802                 if (rc < 0) {
3803                         CNETERR("Error sending %s to %s: %d\n",
3804                                lnet_msgtyp2str(msg->msg_type),
3805                                libcfs_idstr(&msg->msg_target), rc);
3806                         lnet_finalize(msg, rc);
3807                 }
3808         }
3809 }
3810
3811 /* The discovery thread. */
3812 static int lnet_peer_discovery(void *arg)
3813 {
3814         struct lnet_peer *lp;
3815         int rc;
3816
3817         wait_for_completion(&the_lnet.ln_started);
3818
3819         CDEBUG(D_NET, "started\n");
3820
3821         for (;;) {
3822                 if (lnet_peer_discovery_wait_for_work())
3823                         break;
3824
3825                 if (lnet_push_target_resize_needed())
3826                         lnet_push_target_resize();
3827                 else if (the_lnet.ln_push_target->pb_needs_post)
3828                         lnet_push_target_post(the_lnet.ln_push_target,
3829                                               &the_lnet.ln_push_target_md);
3830
3831                 lnet_resend_msgs();
3832
3833                 lnet_net_lock(LNET_LOCK_EX);
3834                 if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING) {
3835                         lnet_net_unlock(LNET_LOCK_EX);
3836                         break;
3837                 }
3838
3839                 /*
3840                  * Process all incoming discovery work requests.  When
3841                  * discovery must wait on a peer to change state, it
3842                  * is added to the tail of the ln_dc_working queue. A
3843                  * timestamp keeps track of when the peer was added,
3844                  * so we can time out discovery requests that take too
3845                  * long.
3846                  */
3847                 while (!list_empty(&the_lnet.ln_dc_request)) {
3848                         lp = list_first_entry(&the_lnet.ln_dc_request,
3849                                               struct lnet_peer, lp_dc_list);
3850                         list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
3851                         /*
3852                          * set the time the peer was put on the dc_working
3853                          * queue. It shouldn't remain on the queue
3854                          * forever, in case the GET message (for ping)
3855                          * doesn't get a REPLY or the PUT message (for
3856                          * push) doesn't get an ACK.
3857                          */
3858                         lp->lp_last_queued = ktime_get_real_seconds();
3859                         lnet_net_unlock(LNET_LOCK_EX);
3860
3861                         if (lnet_push_target_resize_needed())
3862                                 lnet_push_target_resize();
3863                         else if (the_lnet.ln_push_target->pb_needs_post)
3864                                 lnet_push_target_post(the_lnet.ln_push_target,
3865                                                       &the_lnet.ln_push_target_md);
3866
3867                         /*
3868                          * Select an action depending on the state of
3869                          * the peer and whether discovery is disabled.
3870                          * The check whether discovery is disabled is
3871                          * done after the code that handles processing
3872                          * for arrived data, cleanup for failures, and
3873                          * forcing a Ping or Push.
3874                          */
3875                         spin_lock(&lp->lp_lock);
3876                         CDEBUG(D_NET, "peer %s(%p) state %#x\n",
3877                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3878                                 lp->lp_state);
3879                         if (lp->lp_state & (LNET_PEER_MARK_DELETION |
3880                                             LNET_PEER_MARK_DELETED))
3881                                 rc = lnet_peer_deletion(lp);
3882                         else if (lp->lp_state & LNET_PEER_DATA_PRESENT)
3883                                 rc = lnet_peer_data_present(lp);
3884                         else if (lp->lp_state & LNET_PEER_PING_FAILED)
3885                                 rc = lnet_peer_ping_failed(lp);
3886                         else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
3887                                 rc = lnet_peer_push_failed(lp);
3888                         else if (lp->lp_state & LNET_PEER_FORCE_PING)
3889                                 rc = lnet_peer_send_ping(lp);
3890                         else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
3891                                 rc = lnet_peer_send_push(lp);
3892                         else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
3893                                 rc = lnet_peer_send_ping(lp);
3894                         else if (lnet_peer_needs_push(lp))
3895                                 rc = lnet_peer_send_push(lp);
3896                         else
3897                                 rc = lnet_peer_discovered(lp);
3898                         CDEBUG(D_NET, "peer %s(%p) state %#x rc %d\n",
3899                                 libcfs_nidstr(&lp->lp_primary_nid), lp,
3900                                 lp->lp_state, rc);
3901
3902                         if (rc == LNET_REDISCOVER_PEER) {
3903                                 spin_unlock(&lp->lp_lock);
3904                                 lnet_net_lock(LNET_LOCK_EX);
3905                                 list_move(&lp->lp_dc_list,
3906                                           &the_lnet.ln_dc_request);
3907                         } else if (rc ||
3908                                    !(lp->lp_state & LNET_PEER_DISCOVERING)) {
3909                                 spin_unlock(&lp->lp_lock);
3910                                 lnet_net_lock(LNET_LOCK_EX);
3911                                 lnet_peer_discovery_complete(lp, rc);
3912                         } else {
3913                                 spin_unlock(&lp->lp_lock);
3914                                 lnet_net_lock(LNET_LOCK_EX);
3915                         }
3916
3917                         if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
3918                                 break;
3919
3920                 }
3921
3922                 lnet_net_unlock(LNET_LOCK_EX);
3923         }
3924
3925         CDEBUG(D_NET, "stopping\n");
3926         /*
3927          * Clean up before telling lnet_peer_discovery_stop() that
3928          * we're done. Use wake_up() below to somewhat reduce the
3929          * size of the thundering herd if there are multiple threads
3930          * waiting on discovery of a single peer.
3931          */
3932
3933         /* Queue cleanup 1: stop all pending pings and pushes. */
3934         lnet_net_lock(LNET_LOCK_EX);
3935         while (!list_empty(&the_lnet.ln_dc_working)) {
3936                 lp = list_first_entry(&the_lnet.ln_dc_working,
3937                                       struct lnet_peer, lp_dc_list);
3938                 list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
3939                 lnet_net_unlock(LNET_LOCK_EX);
3940                 lnet_peer_cancel_discovery(lp);
3941                 lnet_net_lock(LNET_LOCK_EX);
3942         }
3943         lnet_net_unlock(LNET_LOCK_EX);
3944
3945         /* Queue cleanup 2: wait for the expired queue to clear. */
3946         while (!list_empty(&the_lnet.ln_dc_expired))
3947                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
3948
3949         /* Queue cleanup 3: clear the request queue. */
3950         lnet_net_lock(LNET_LOCK_EX);
3951         while (!list_empty(&the_lnet.ln_dc_request)) {
3952                 lp = list_first_entry(&the_lnet.ln_dc_request,
3953                                       struct lnet_peer, lp_dc_list);
3954                 lnet_peer_discovery_complete(lp, -ESHUTDOWN);
3955         }
3956         lnet_net_unlock(LNET_LOCK_EX);
3957
3958         lnet_assert_handler_unused(the_lnet.ln_dc_handler);
3959         the_lnet.ln_dc_handler = NULL;
3960
3961         the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3962         wake_up(&the_lnet.ln_dc_waitq);
3963
3964         CDEBUG(D_NET, "stopped\n");
3965
3966         return 0;
3967 }
3968
3969 /* ln_api_mutex is held on entry. */
3970 int lnet_peer_discovery_start(void)
3971 {
3972         struct task_struct *task;
3973         int rc = 0;
3974
3975         if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
3976                 return -EALREADY;
3977
3978         the_lnet.ln_dc_handler = lnet_discovery_event_handler;
3979         the_lnet.ln_dc_state = LNET_DC_STATE_RUNNING;
3980         task = kthread_run(lnet_peer_discovery, NULL, "lnet_discovery");
3981         if (IS_ERR(task)) {
3982                 rc = PTR_ERR(task);
3983                 CERROR("Can't start peer discovery thread: %d\n", rc);
3984
3985                 the_lnet.ln_dc_handler = NULL;
3986
3987                 the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
3988         }
3989
3990         CDEBUG(D_NET, "discovery start: %d\n", rc);
3991
3992         return rc;
3993 }
3994
3995 /* ln_api_mutex is held on entry. */
3996 void lnet_peer_discovery_stop(void)
3997 {
3998         if (the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN)
3999                 return;
4000
4001         LASSERT(the_lnet.ln_dc_state == LNET_DC_STATE_RUNNING);
4002         the_lnet.ln_dc_state = LNET_DC_STATE_STOPPING;
4003
4004         /* In the LNetNIInit() path we may be stopping discovery before it
4005          * entered its work loop
4006          */
4007         if (!completion_done(&the_lnet.ln_started))
4008                 complete(&the_lnet.ln_started);
4009         else
4010                 wake_up(&the_lnet.ln_dc_waitq);
4011
4012         mutex_unlock(&the_lnet.ln_api_mutex);
4013         wait_event(the_lnet.ln_dc_waitq,
4014                    the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
4015         mutex_lock(&the_lnet.ln_api_mutex);
4016
4017         LASSERT(list_empty(&the_lnet.ln_dc_request));
4018         LASSERT(list_empty(&the_lnet.ln_dc_working));
4019         LASSERT(list_empty(&the_lnet.ln_dc_expired));
4020
4021         CDEBUG(D_NET, "discovery stopped\n");
4022 }
4023
4024 /* Debugging */
4025
4026 void
4027 lnet_debug_peer(struct lnet_nid *nid)
4028 {
4029         char                    *aliveness = "NA";
4030         struct lnet_peer_ni     *lp;
4031         int                     cpt;
4032
4033         cpt = lnet_nid2cpt(nid, NULL);
4034         lnet_net_lock(cpt);
4035
4036         lp = lnet_peerni_by_nid_locked(nid, NULL, cpt);
4037         if (IS_ERR(lp)) {
4038                 lnet_net_unlock(cpt);
4039                 CDEBUG(D_WARNING, "No peer %s\n", libcfs_nidstr(nid));
4040                 return;
4041         }
4042
4043         if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
4044                 aliveness = (lnet_is_peer_ni_alive(lp)) ? "up" : "down";
4045
4046         CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
4047                libcfs_nidstr(&lp->lpni_nid), kref_read(&lp->lpni_kref),
4048                aliveness, lp->lpni_net->net_tunables.lct_peer_tx_credits,
4049                lp->lpni_rtrcredits, lp->lpni_minrtrcredits,
4050                lp->lpni_txcredits, lp->lpni_mintxcredits, lp->lpni_txqnob);
4051
4052         lnet_peer_ni_decref_locked(lp);
4053
4054         lnet_net_unlock(cpt);
4055 }
4056
4057 /* Gathering information for userspace. */
4058
4059 int lnet_get_peer_ni_info(__u32 peer_index, __u64 *nid,
4060                           char aliveness[LNET_MAX_STR_LEN],
4061                           __u32 *cpt_iter, __u32 *refcount,
4062                           __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
4063                           __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
4064                           __u32 *peer_tx_qnob)
4065 {
4066         struct lnet_peer_table          *peer_table;
4067         struct lnet_peer_ni             *lp;
4068         int                             j;
4069         int                             lncpt;
4070         bool                            found = false;
4071
4072         /* get the number of CPTs */
4073         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4074
4075         /* if the cpt number to be examined is >= the number of cpts in
4076          * the system then indicate that there are no more cpts to examin
4077          */
4078         if (*cpt_iter >= lncpt)
4079                 return -ENOENT;
4080
4081         /* get the current table */
4082         peer_table = the_lnet.ln_peer_tables[*cpt_iter];
4083         /* if the ptable is NULL then there are no more cpts to examine */
4084         if (peer_table == NULL)
4085                 return -ENOENT;
4086
4087         lnet_net_lock(*cpt_iter);
4088
4089         for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
4090                 struct list_head *peers = &peer_table->pt_hash[j];
4091
4092                 list_for_each_entry(lp, peers, lpni_hashlist) {
4093                         if (!nid_is_nid4(&lp->lpni_nid))
4094                                 continue;
4095                         if (peer_index-- > 0)
4096                                 continue;
4097
4098                         snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
4099                         if (lnet_isrouter(lp) ||
4100                                 lnet_peer_aliveness_enabled(lp))
4101                                 snprintf(aliveness, LNET_MAX_STR_LEN,
4102                                          lnet_is_peer_ni_alive(lp) ? "up" : "down");
4103
4104                         *nid = lnet_nid_to_nid4(&lp->lpni_nid);
4105                         *refcount = kref_read(&lp->lpni_kref);
4106                         *ni_peer_tx_credits =
4107                                 lp->lpni_net->net_tunables.lct_peer_tx_credits;
4108                         *peer_tx_credits = lp->lpni_txcredits;
4109                         *peer_rtr_credits = lp->lpni_rtrcredits;
4110                         *peer_min_rtr_credits = lp->lpni_mintxcredits;
4111                         *peer_tx_qnob = lp->lpni_txqnob;
4112
4113                         found = true;
4114                 }
4115
4116         }
4117         lnet_net_unlock(*cpt_iter);
4118
4119         *cpt_iter = lncpt;
4120
4121         return found ? 0 : -ENOENT;
4122 }
4123
4124 /* ln_api_mutex is held, which keeps the peer list stable */
4125 int lnet_get_peer_info(struct lnet_ioctl_peer_cfg *cfg, void __user *bulk)
4126 {
4127         struct lnet_ioctl_element_stats *lpni_stats;
4128         struct lnet_ioctl_element_msg_stats *lpni_msg_stats;
4129         struct lnet_ioctl_peer_ni_hstats *lpni_hstats;
4130         struct lnet_peer_ni_credit_info *lpni_info;
4131         struct lnet_peer_ni *lpni;
4132         struct lnet_peer *lp;
4133         lnet_nid_t nid4;
4134         struct lnet_nid nid;
4135         __u32 size;
4136         int rc;
4137
4138         lnet_nid4_to_nid(cfg->prcfg_prim_nid, &nid);
4139         lp = lnet_find_peer(&nid);
4140
4141         if (!lp) {
4142                 rc = -ENOENT;
4143                 goto out;
4144         }
4145
4146         size = sizeof(nid4) + sizeof(*lpni_info) + sizeof(*lpni_stats)
4147                 + sizeof(*lpni_msg_stats) + sizeof(*lpni_hstats);
4148         size *= lp->lp_nnis;
4149         if (size > cfg->prcfg_size) {
4150                 cfg->prcfg_size = size;
4151                 rc = -E2BIG;
4152                 goto out_lp_decref;
4153         }
4154
4155         cfg->prcfg_prim_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4156         cfg->prcfg_mr = lnet_peer_is_multi_rail(lp);
4157         cfg->prcfg_cfg_nid = lnet_nid_to_nid4(&lp->lp_primary_nid);
4158         cfg->prcfg_count = lp->lp_nnis;
4159         cfg->prcfg_size = size;
4160         cfg->prcfg_state = lp->lp_state;
4161
4162         /* Allocate helper buffers. */
4163         rc = -ENOMEM;
4164         LIBCFS_ALLOC(lpni_info, sizeof(*lpni_info));
4165         if (!lpni_info)
4166                 goto out_lp_decref;
4167         LIBCFS_ALLOC(lpni_stats, sizeof(*lpni_stats));
4168         if (!lpni_stats)
4169                 goto out_free_info;
4170         LIBCFS_ALLOC(lpni_msg_stats, sizeof(*lpni_msg_stats));
4171         if (!lpni_msg_stats)
4172                 goto out_free_stats;
4173         LIBCFS_ALLOC(lpni_hstats, sizeof(*lpni_hstats));
4174         if (!lpni_hstats)
4175                 goto out_free_msg_stats;
4176
4177
4178         lpni = NULL;
4179         rc = -EFAULT;
4180         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
4181                 if (!nid_is_nid4(&lpni->lpni_nid))
4182                         continue;
4183                 nid4 = lnet_nid_to_nid4(&lpni->lpni_nid);
4184                 if (copy_to_user(bulk, &nid4, sizeof(nid4)))
4185                         goto out_free_hstats;
4186                 bulk += sizeof(nid4);
4187
4188                 memset(lpni_info, 0, sizeof(*lpni_info));
4189                 snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN, "NA");
4190                 if (lnet_isrouter(lpni) ||
4191                         lnet_peer_aliveness_enabled(lpni))
4192                         snprintf(lpni_info->cr_aliveness, LNET_MAX_STR_LEN,
4193                                 lnet_is_peer_ni_alive(lpni) ? "up" : "down");
4194
4195                 lpni_info->cr_refcount = kref_read(&lpni->lpni_kref);
4196                 lpni_info->cr_ni_peer_tx_credits = (lpni->lpni_net != NULL) ?
4197                         lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0;
4198                 lpni_info->cr_peer_tx_credits = lpni->lpni_txcredits;
4199                 lpni_info->cr_peer_rtr_credits = lpni->lpni_rtrcredits;
4200                 lpni_info->cr_peer_min_rtr_credits = lpni->lpni_minrtrcredits;
4201                 lpni_info->cr_peer_min_tx_credits = lpni->lpni_mintxcredits;
4202                 lpni_info->cr_peer_tx_qnob = lpni->lpni_txqnob;
4203                 if (copy_to_user(bulk, lpni_info, sizeof(*lpni_info)))
4204                         goto out_free_hstats;
4205                 bulk += sizeof(*lpni_info);
4206
4207                 memset(lpni_stats, 0, sizeof(*lpni_stats));
4208                 lpni_stats->iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
4209                                                             LNET_STATS_TYPE_SEND);
4210                 lpni_stats->iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
4211                                                             LNET_STATS_TYPE_RECV);
4212                 lpni_stats->iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
4213                                                             LNET_STATS_TYPE_DROP);
4214                 if (copy_to_user(bulk, lpni_stats, sizeof(*lpni_stats)))
4215                         goto out_free_hstats;
4216                 bulk += sizeof(*lpni_stats);
4217                 lnet_usr_translate_stats(lpni_msg_stats, &lpni->lpni_stats);
4218                 if (copy_to_user(bulk, lpni_msg_stats, sizeof(*lpni_msg_stats)))
4219                         goto out_free_hstats;
4220                 bulk += sizeof(*lpni_msg_stats);
4221                 lpni_hstats->hlpni_network_timeout =
4222                   atomic_read(&lpni->lpni_hstats.hlt_network_timeout);
4223                 lpni_hstats->hlpni_remote_dropped =
4224                   atomic_read(&lpni->lpni_hstats.hlt_remote_dropped);
4225                 lpni_hstats->hlpni_remote_timeout =
4226                   atomic_read(&lpni->lpni_hstats.hlt_remote_timeout);
4227                 lpni_hstats->hlpni_remote_error =
4228                   atomic_read(&lpni->lpni_hstats.hlt_remote_error);
4229                 lpni_hstats->hlpni_health_value =
4230                   atomic_read(&lpni->lpni_healthv);
4231                 lpni_hstats->hlpni_ping_count = lpni->lpni_ping_count;
4232                 lpni_hstats->hlpni_next_ping = lpni->lpni_next_ping;
4233                 if (copy_to_user(bulk, lpni_hstats, sizeof(*lpni_hstats)))
4234                         goto out_free_hstats;
4235                 bulk += sizeof(*lpni_hstats);
4236         }
4237         rc = 0;
4238
4239 out_free_hstats:
4240         LIBCFS_FREE(lpni_hstats, sizeof(*lpni_hstats));
4241 out_free_msg_stats:
4242         LIBCFS_FREE(lpni_msg_stats, sizeof(*lpni_msg_stats));
4243 out_free_stats:
4244         LIBCFS_FREE(lpni_stats, sizeof(*lpni_stats));
4245 out_free_info:
4246         LIBCFS_FREE(lpni_info, sizeof(*lpni_info));
4247 out_lp_decref:
4248         lnet_peer_decref_locked(lp);
4249 out:
4250         return rc;
4251 }
4252
4253 /* must hold net_lock/0 */
4254 void
4255 lnet_peer_ni_add_to_recoveryq_locked(struct lnet_peer_ni *lpni,
4256                                      struct list_head *recovery_queue,
4257                                      time64_t now)
4258 {
4259         /* the mt could've shutdown and cleaned up the queues */
4260         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
4261                 return;
4262
4263         if (!list_empty(&lpni->lpni_recovery))
4264                 return;
4265
4266         if (atomic_read(&lpni->lpni_healthv) == LNET_MAX_HEALTH_VALUE)
4267                 return;
4268
4269         if (!lpni->lpni_last_alive) {
4270                 CDEBUG(D_NET,
4271                        "lpni %s(%p) not eligible for recovery last alive %lld\n",
4272                        libcfs_nidstr(&lpni->lpni_nid), lpni,
4273                        lpni->lpni_last_alive);
4274                 return;
4275         }
4276
4277         if (lnet_recovery_limit &&
4278             now > lpni->lpni_last_alive + lnet_recovery_limit) {
4279                 CDEBUG(D_NET, "lpni %s aged out last alive %lld\n",
4280                        libcfs_nidstr(&lpni->lpni_nid),
4281                        lpni->lpni_last_alive);
4282                 /* Reset the ping count so that if this peer NI is added back to
4283                  * the recovery queue we will send the first ping right away.
4284                  */
4285                 lpni->lpni_ping_count = 0;
4286                 return;
4287         }
4288
4289         /* This peer NI is going on the recovery queue, so take a ref on it */
4290         lnet_peer_ni_addref_locked(lpni);
4291
4292         lnet_peer_ni_set_next_ping(lpni, now);
4293
4294         CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld last alive: %lld health: %d\n",
4295                libcfs_nidstr(&lpni->lpni_nid),
4296                lpni->lpni_ping_count,
4297                lpni->lpni_next_ping,
4298                lpni->lpni_last_alive,
4299                atomic_read(&lpni->lpni_healthv));
4300
4301         list_add_tail(&lpni->lpni_recovery, recovery_queue);
4302 }
4303
4304 /* Call with the ln_api_mutex held */
4305 void
4306 lnet_peer_ni_set_healthv(lnet_nid_t nid4, int value, bool all)
4307 {
4308         struct lnet_peer_table *ptable;
4309         struct lnet_peer *lp;
4310         struct lnet_peer_net *lpn;
4311         struct lnet_peer_ni *lpni;
4312         struct lnet_nid nid;
4313         int lncpt;
4314         int cpt;
4315         time64_t now;
4316
4317         if (the_lnet.ln_state != LNET_STATE_RUNNING)
4318                 return;
4319
4320         lnet_nid4_to_nid(nid4, &nid);
4321         now = ktime_get_seconds();
4322
4323         if (!all) {
4324                 lnet_net_lock(LNET_LOCK_EX);
4325                 lpni = lnet_peer_ni_find_locked(&nid);
4326                 if (!lpni) {
4327                         lnet_net_unlock(LNET_LOCK_EX);
4328                         return;
4329                 }
4330                 lnet_set_lpni_healthv_locked(lpni, value);
4331                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
4332                                              &the_lnet.ln_mt_peerNIRecovq, now);
4333                 lnet_peer_ni_decref_locked(lpni);
4334                 lnet_net_unlock(LNET_LOCK_EX);
4335                 return;
4336         }
4337
4338         lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
4339
4340         /*
4341          * Walk all the peers and reset the health value for each one to the
4342          * specified value.
4343          */
4344         lnet_net_lock(LNET_LOCK_EX);
4345         for (cpt = 0; cpt < lncpt; cpt++) {
4346                 ptable = the_lnet.ln_peer_tables[cpt];
4347                 list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
4348                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
4349                                 list_for_each_entry(lpni, &lpn->lpn_peer_nis,
4350                                                     lpni_peer_nis) {
4351                                         lnet_set_lpni_healthv_locked(lpni,
4352                                                                      value);
4353                                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
4354                                              &the_lnet.ln_mt_peerNIRecovq, now);
4355                                 }
4356                         }
4357                 }
4358         }
4359         lnet_net_unlock(LNET_LOCK_EX);
4360 }
4361