lnet/lnet/lib-msg.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) 2003, 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/lib-msg.c
  32  *
  33  * Message decoding, parsing and finalizing routines
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_LNET
  37
  38 #include <lnet/lib-lnet.h>
  39
  40 void
  41 lnet_build_unlink_event(struct lnet_libmd *md, struct lnet_event *ev)
  42 {
  43         ENTRY;
  44
  45         memset(ev, 0, sizeof(*ev));
  46
  47         ev->status   = 0;
  48         ev->unlinked = 1;
  49         ev->type     = LNET_EVENT_UNLINK;
  50         lnet_md_deconstruct(md, ev);
  51         lnet_md2handle(&ev->md_handle, md);
  52         EXIT;
  53 }
  54
  55 /*
  56  * Don't need any lock, must be called after lnet_commit_md
  57  */
  58 void
  59 lnet_build_msg_event(struct lnet_msg *msg, enum lnet_event_kind ev_type)
  60 {
  61         struct lnet_hdr *hdr = &msg->msg_hdr;
  62         struct lnet_event *ev = &msg->msg_ev;
  63
  64         LASSERT(!msg->msg_routing);
  65
  66         ev->type = ev_type;
  67         ev->msg_type = msg->msg_type;
  68
  69         if (ev_type == LNET_EVENT_SEND) {
  70                 /* event for active message */
  71                 ev->target.nid    = hdr->dest_nid;
  72                 ev->target.pid    = hdr->dest_pid;
  73                 ev->initiator.nid = LNET_ANY_NID;
  74                 ev->initiator.pid = the_lnet.ln_pid;
  75                 ev->source.nid    = LNET_ANY_NID;
  76                 ev->source.pid    = the_lnet.ln_pid;
  77                 ev->sender        = LNET_ANY_NID;
  78         } else {
  79                 /* event for passive message */
  80                 ev->target.pid    = hdr->dest_pid;
  81                 ev->target.nid    = hdr->dest_nid;
  82                 ev->initiator.pid = hdr->src_pid;
  83                 /* Multi-Rail: resolve src_nid to "primary" peer NID */
  84                 ev->initiator.nid = msg->msg_initiator;
  85                 /* Multi-Rail: track source NID. */
  86                 ev->source.pid    = hdr->src_pid;
  87                 ev->source.nid    = hdr->src_nid;
  88                 ev->rlength       = hdr->payload_length;
  89                 ev->sender        = msg->msg_from;
  90                 ev->mlength       = msg->msg_wanted;
  91                 ev->offset        = msg->msg_offset;
  92         }
  93
  94         switch (ev_type) {
  95         default:
  96                 LBUG();
  97
  98         case LNET_EVENT_PUT: /* passive PUT */
  99                 ev->pt_index   = hdr->msg.put.ptl_index;
 100                 ev->match_bits = hdr->msg.put.match_bits;
 101                 ev->hdr_data   = hdr->msg.put.hdr_data;
 102                 return;
 103
 104         case LNET_EVENT_GET: /* passive GET */
 105                 ev->pt_index   = hdr->msg.get.ptl_index;
 106                 ev->match_bits = hdr->msg.get.match_bits;
 107                 ev->hdr_data   = 0;
 108                 return;
 109
 110         case LNET_EVENT_ACK: /* ACK */
 111                 ev->match_bits = hdr->msg.ack.match_bits;
 112                 ev->mlength    = hdr->msg.ack.mlength;
 113                 return;
 114
 115         case LNET_EVENT_REPLY: /* REPLY */
 116                 return;
 117
 118         case LNET_EVENT_SEND: /* active message */
 119                 if (msg->msg_type == LNET_MSG_PUT) {
 120                         ev->pt_index   = le32_to_cpu(hdr->msg.put.ptl_index);
 121                         ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
 122                         ev->offset     = le32_to_cpu(hdr->msg.put.offset);
 123                         ev->mlength    =
 124                         ev->rlength    = le32_to_cpu(hdr->payload_length);
 125                         ev->hdr_data   = le64_to_cpu(hdr->msg.put.hdr_data);
 126
 127                 } else {
 128                         LASSERT(msg->msg_type == LNET_MSG_GET);
 129                         ev->pt_index   = le32_to_cpu(hdr->msg.get.ptl_index);
 130                         ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
 131                         ev->mlength    =
 132                         ev->rlength    = le32_to_cpu(hdr->msg.get.sink_length);
 133                         ev->offset     = le32_to_cpu(hdr->msg.get.src_offset);
 134                         ev->hdr_data   = 0;
 135                 }
 136                 return;
 137         }
 138 }
 139
 140 void
 141 lnet_msg_commit(struct lnet_msg *msg, int cpt)
 142 {
 143         struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
 144         struct lnet_counters_common *common;
 145         s64 timeout_ns;
 146
 147         /* set the message deadline */
 148         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
 149         msg->msg_deadline = ktime_add_ns(ktime_get(), timeout_ns);
 150
 151         /* routed message can be committed for both receiving and sending */
 152         LASSERT(!msg->msg_tx_committed);
 153
 154         if (msg->msg_sending) {
 155                 LASSERT(!msg->msg_receiving);
 156                 msg->msg_tx_cpt = cpt;
 157                 msg->msg_tx_committed = 1;
 158                 if (msg->msg_rx_committed) { /* routed message REPLY */
 159                         LASSERT(msg->msg_onactivelist);
 160                         return;
 161                 }
 162         } else {
 163                 LASSERT(!msg->msg_sending);
 164                 msg->msg_rx_cpt = cpt;
 165                 msg->msg_rx_committed = 1;
 166         }
 167
 168         LASSERT(!msg->msg_onactivelist);
 169
 170         msg->msg_onactivelist = 1;
 171         list_add_tail(&msg->msg_activelist, &container->msc_active);
 172
 173         common = &the_lnet.ln_counters[cpt]->lct_common;
 174         common->lcc_msgs_alloc++;
 175         if (common->lcc_msgs_alloc > common->lcc_msgs_max)
 176                 common->lcc_msgs_max = common->lcc_msgs_alloc;
 177 }
 178
 179 static void
 180 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
 181 {
 182         struct lnet_counters_common *common;
 183         struct lnet_event *ev = &msg->msg_ev;
 184
 185         LASSERT(msg->msg_tx_committed);
 186         if (status != 0)
 187                 goto out;
 188
 189         common = &(the_lnet.ln_counters[msg->msg_tx_cpt]->lct_common);
 190         switch (ev->type) {
 191         default: /* routed message */
 192                 LASSERT(msg->msg_routing);
 193                 LASSERT(msg->msg_rx_committed);
 194                 LASSERT(ev->type == 0);
 195
 196                 common->lcc_route_length += msg->msg_len;
 197                 common->lcc_route_count++;
 198                 goto incr_stats;
 199
 200         case LNET_EVENT_PUT:
 201                 /* should have been decommitted */
 202                 LASSERT(!msg->msg_rx_committed);
 203                 /* overwritten while sending ACK */
 204                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 205                 msg->msg_type = LNET_MSG_PUT; /* fix type */
 206                 break;
 207
 208         case LNET_EVENT_SEND:
 209                 LASSERT(!msg->msg_rx_committed);
 210                 if (msg->msg_type == LNET_MSG_PUT)
 211                         common->lcc_send_length += msg->msg_len;
 212                 break;
 213
 214         case LNET_EVENT_GET:
 215                 LASSERT(msg->msg_rx_committed);
 216                 /* overwritten while sending reply, we should never be
 217                  * here for optimized GET */
 218                 LASSERT(msg->msg_type == LNET_MSG_REPLY);
 219                 msg->msg_type = LNET_MSG_GET; /* fix type */
 220                 break;
 221         }
 222
 223         common->lcc_send_count++;
 224
 225 incr_stats:
 226         if (msg->msg_txpeer)
 227                 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 228                                 msg->msg_type,
 229                                 LNET_STATS_TYPE_SEND);
 230         if (msg->msg_txni)
 231                 lnet_incr_stats(&msg->msg_txni->ni_stats,
 232                                 msg->msg_type,
 233                                 LNET_STATS_TYPE_SEND);
 234  out:
 235         lnet_return_tx_credits_locked(msg);
 236         msg->msg_tx_committed = 0;
 237 }
 238
 239 static void
 240 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
 241 {
 242         struct lnet_counters_common *common;
 243         struct lnet_event *ev = &msg->msg_ev;
 244
 245         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 246         LASSERT(msg->msg_rx_committed);
 247
 248         if (status != 0)
 249                 goto out;
 250
 251         common = &(the_lnet.ln_counters[msg->msg_rx_cpt]->lct_common);
 252         switch (ev->type) {
 253         default:
 254                 LASSERT(ev->type == 0);
 255                 LASSERT(msg->msg_routing);
 256                 goto incr_stats;
 257
 258         case LNET_EVENT_ACK:
 259                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 260                 break;
 261
 262         case LNET_EVENT_GET:
 263                 /* type is "REPLY" if it's an optimized GET on passive side,
 264                  * because optimized GET will never be committed for sending,
 265                  * so message type wouldn't be changed back to "GET" by
 266                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 267                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 268                         msg->msg_type == LNET_MSG_GET);
 269                 common->lcc_send_length += msg->msg_wanted;
 270                 break;
 271
 272         case LNET_EVENT_PUT:
 273                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 274                 break;
 275
 276         case LNET_EVENT_REPLY:
 277                 /* type is "GET" if it's an optimized GET on active side,
 278                  * see details in lnet_create_reply_msg() */
 279                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 280                         msg->msg_type == LNET_MSG_REPLY);
 281                 break;
 282         }
 283
 284         common->lcc_recv_count++;
 285
 286 incr_stats:
 287         if (msg->msg_rxpeer)
 288                 lnet_incr_stats(&msg->msg_rxpeer->lpni_stats,
 289                                 msg->msg_type,
 290                                 LNET_STATS_TYPE_RECV);
 291         if (msg->msg_rxni)
 292                 lnet_incr_stats(&msg->msg_rxni->ni_stats,
 293                                 msg->msg_type,
 294                                 LNET_STATS_TYPE_RECV);
 295         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 296                 common->lcc_recv_length += msg->msg_wanted;
 297
 298  out:
 299         lnet_return_rx_credits_locked(msg);
 300         msg->msg_rx_committed = 0;
 301 }
 302
 303 void
 304 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
 305 {
 306         int     cpt2 = cpt;
 307
 308         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 309         LASSERT(msg->msg_onactivelist);
 310
 311         if (msg->msg_tx_committed) { /* always decommit for sending first */
 312                 LASSERT(cpt == msg->msg_tx_cpt);
 313                 lnet_msg_decommit_tx(msg, status);
 314         }
 315
 316         if (msg->msg_rx_committed) {
 317                 /* forwarding msg committed for both receiving and sending */
 318                 if (cpt != msg->msg_rx_cpt) {
 319                         lnet_net_unlock(cpt);
 320                         cpt2 = msg->msg_rx_cpt;
 321                         lnet_net_lock(cpt2);
 322                 }
 323                 lnet_msg_decommit_rx(msg, status);
 324         }
 325
 326         list_del(&msg->msg_activelist);
 327         msg->msg_onactivelist = 0;
 328
 329         the_lnet.ln_counters[cpt2]->lct_common.lcc_msgs_alloc--;
 330
 331         if (cpt2 != cpt) {
 332                 lnet_net_unlock(cpt2);
 333                 lnet_net_lock(cpt);
 334         }
 335 }
 336
 337 void
 338 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
 339                    unsigned int offset, unsigned int mlen)
 340 {
 341         /* NB: @offset and @len are only useful for receiving */
 342         /* Here, we attach the MD on lnet_msg and mark it busy and
 343          * decrementing its threshold. Come what may, the lnet_msg "owns"
 344          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 345          * signals completion. */
 346         LASSERT(!msg->msg_routing);
 347
 348         msg->msg_md = md;
 349         if (msg->msg_receiving) { /* committed for receiving */
 350                 msg->msg_offset = offset;
 351                 msg->msg_wanted = mlen;
 352         }
 353
 354         md->md_refcount++;
 355         if (md->md_threshold != LNET_MD_THRESH_INF) {
 356                 LASSERT(md->md_threshold > 0);
 357                 md->md_threshold--;
 358         }
 359
 360         /* build umd in event */
 361         lnet_md2handle(&msg->msg_ev.md_handle, md);
 362         lnet_md_deconstruct(md, &msg->msg_ev);
 363 }
 364
 365 static int
 366 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
 367 {
 368         struct lnet_handle_wire ack_wmd;
 369         int                rc;
 370         int                status = msg->msg_ev.status;
 371
 372         LASSERT(msg->msg_onactivelist);
 373
 374         if (status == 0 && msg->msg_ack) {
 375                 /* Only send an ACK if the PUT completed successfully */
 376
 377                 lnet_msg_decommit(msg, cpt, 0);
 378
 379                 msg->msg_ack = 0;
 380                 lnet_net_unlock(cpt);
 381
 382                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 383                 LASSERT(!msg->msg_routing);
 384
 385                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 386
 387                 lnet_prep_send(msg, LNET_MSG_ACK, &msg->msg_ev.source, 0, 0);
 388
 389                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 390                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 391                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 392
 393                 rc = lnet_send(&msg->msg_ev.target.nid, msg,
 394                                &msg->msg_from);
 395
 396                 lnet_net_lock(cpt);
 397                 /*
 398                  * NB: message is committed for sending, we should return
 399                  * on success because LND will finalize this message later.
 400                  *
 401                  * Also, there is possibility that message is committed for
 402                  * sending and also failed before delivering to LND,
 403                  * i.e: ENOMEM, in that case we can't fall through either
 404                  * because CPT for sending can be different with CPT for
 405                  * receiving, so we should return back to lnet_finalize()
 406                  * to make sure we are locking the correct partition.
 407                  */
 408                 return rc;
 409
 410         } else if (status == 0 &&       /* OK so far */
 411                    (msg->msg_routing && !msg->msg_sending)) {
 412                 /* not forwarded */
 413                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 414                 lnet_net_unlock(cpt);
 415
 416                 rc = lnet_send(NULL, msg, NULL);
 417
 418                 lnet_net_lock(cpt);
 419                 /*
 420                  * NB: message is committed for sending, we should return
 421                  * on success because LND will finalize this message later.
 422                  *
 423                  * Also, there is possibility that message is committed for
 424                  * sending and also failed before delivering to LND,
 425                  * i.e: ENOMEM, in that case we can't fall through either:
 426                  * - The rule is message must decommit for sending first if
 427                  *   the it's committed for both sending and receiving
 428                  * - CPT for sending can be different with CPT for receiving,
 429                  *   so we should return back to lnet_finalize() to make
 430                  *   sure we are locking the correct partition.
 431                  */
 432                 return rc;
 433         }
 434
 435         lnet_msg_decommit(msg, cpt, status);
 436         lnet_msg_free(msg);
 437         return 0;
 438 }
 439
 440 /* must hold net_lock/0 */
 441 void
 442 lnet_ni_add_to_recoveryq_locked(struct lnet_ni *ni,
 443                                 struct list_head *recovery_queue, time64_t now)
 444 {
 445         if (!list_empty(&ni->ni_recovery))
 446                 return;
 447
 448         if (atomic_read(&ni->ni_healthv) == LNET_MAX_HEALTH_VALUE)
 449                 return;
 450
 451         /* This NI is going on the recovery queue, so take a ref on it */
 452         lnet_ni_addref_locked(ni, 0);
 453
 454         lnet_ni_set_next_ping(ni, now);
 455
 456         CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld health :%d\n",
 457                libcfs_nidstr(&ni->ni_nid),
 458                ni->ni_ping_count,
 459                ni->ni_next_ping,
 460                atomic_read(&ni->ni_healthv));
 461
 462         list_add_tail(&ni->ni_recovery, recovery_queue);
 463 }
 464
 465 static void
 466 lnet_handle_local_failure(struct lnet_ni *local_ni)
 467 {
 468         /*
 469          * the lnet_net_lock(0) is used to protect the addref on the ni
 470          * and the recovery queue.
 471          */
 472         lnet_net_lock(0);
 473         /* the mt could've shutdown and cleaned up the queues */
 474         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 475                 lnet_net_unlock(0);
 476                 return;
 477         }
 478
 479         lnet_dec_healthv_locked(&local_ni->ni_healthv, lnet_health_sensitivity);
 480         lnet_ni_add_to_recoveryq_locked(local_ni, &the_lnet.ln_mt_localNIRecovq,
 481                                         ktime_get_seconds());
 482         lnet_net_unlock(0);
 483 }
 484
 485 /* must hold net_lock/0 */
 486 void
 487 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
 488 {
 489         lnet_dec_lpni_healthv_locked(lpni);
 490
 491         /*
 492          * add the peer NI to the recovery queue if it's not already there
 493          * and it's health value is actually below the maximum. It's
 494          * possible that the sensitivity might be set to 0, and the health
 495          * value will not be reduced. In this case, there is no reason to
 496          * invoke recovery
 497          */
 498         lnet_peer_ni_add_to_recoveryq_locked(lpni,
 499                                              &the_lnet.ln_mt_peerNIRecovq,
 500                                              ktime_get_seconds());
 501 }
 502
 503 static void
 504 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 505 {
 506         /* lpni could be NULL if we're in the LOLND case */
 507         if (!lpni)
 508                 return;
 509
 510         lnet_net_lock(0);
 511         /* the mt could've shutdown and cleaned up the queues */
 512         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 513                 lnet_net_unlock(0);
 514                 return;
 515         }
 516         lnet_handle_remote_failure_locked(lpni);
 517         lnet_net_unlock(0);
 518 }
 519
 520 static void
 521 lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 522                  enum lnet_msg_hstatus hstatus)
 523 {
 524         struct lnet_counters_health *health;
 525
 526         health = &the_lnet.ln_counters[0]->lct_health;
 527
 528         switch (hstatus) {
 529         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 530                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 531                 health->lch_local_interrupt_count++;
 532                 break;
 533         case LNET_MSG_STATUS_LOCAL_DROPPED:
 534                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 535                 health->lch_local_dropped_count++;
 536                 break;
 537         case LNET_MSG_STATUS_LOCAL_ABORTED:
 538                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 539                 health->lch_local_aborted_count++;
 540                 break;
 541         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 542                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 543                 health->lch_local_no_route_count++;
 544                 break;
 545         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 546                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 547                 health->lch_local_timeout_count++;
 548                 break;
 549         case LNET_MSG_STATUS_LOCAL_ERROR:
 550                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 551                 health->lch_local_error_count++;
 552                 break;
 553         case LNET_MSG_STATUS_REMOTE_DROPPED:
 554                 if (lpni)
 555                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 556                 health->lch_remote_dropped_count++;
 557                 break;
 558         case LNET_MSG_STATUS_REMOTE_ERROR:
 559                 if (lpni)
 560                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 561                 health->lch_remote_error_count++;
 562                 break;
 563         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 564                 if (lpni)
 565                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 566                 health->lch_remote_timeout_count++;
 567                 break;
 568         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 569                 if (lpni)
 570                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 571                 health->lch_network_timeout_count++;
 572                 break;
 573         case LNET_MSG_STATUS_OK:
 574                 break;
 575         default:
 576                 LBUG();
 577         }
 578 }
 579
 580 static void
 581 lnet_resend_msg_locked(struct lnet_msg *msg)
 582 {
 583         msg->msg_retry_count++;
 584
 585         /*
 586          * remove message from the active list and reset it to prepare
 587          * for a resend. Two exceptions to this
 588          *
 589          * 1. the router case. When a message is being routed it is
 590          * committed for rx when received and committed for tx when
 591          * forwarded. We don't want to remove it from the active list, since
 592          * code which handles receiving expects it to remain on the active
 593          * list.
 594          *
 595          * 2. The REPLY case. Reply messages use the same message
 596          * structure for the GET that was received.
 597          */
 598         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 599                 list_del_init(&msg->msg_activelist);
 600                 msg->msg_onactivelist = 0;
 601         }
 602         /*
 603          * The msg_target.nid which was originally set
 604          * when calling LNetGet() or LNetPut() might've
 605          * been overwritten if we're routing this message.
 606          * Call lnet_msg_decommit_tx() to return the credit
 607          * this message consumed. The message will
 608          * consume another credit when it gets resent.
 609          */
 610         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 611         lnet_msg_decommit_tx(msg, -EAGAIN);
 612         msg->msg_sending = 0;
 613         msg->msg_receiving = 0;
 614         msg->msg_target_is_router = 0;
 615
 616         CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
 617                libcfs_nidstr(&msg->msg_hdr.src_nid),
 618                libcfs_nidstr(&msg->msg_hdr.dest_nid),
 619                lnet_msgtyp2str(msg->msg_type),
 620                lnet_health_error2str(msg->msg_health_status), msg);
 621
 622         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 623
 624         complete(&the_lnet.ln_mt_wait_complete);
 625 }
 626
 627 static int
 628 lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
 629                                      struct list_head *containerq,
 630                                      int nworkers, void **workers)
 631 {
 632         int my_slot = -1;
 633         int i;
 634
 635         list_add_tail(&msg->msg_list, containerq);
 636
 637         for (i = 0; i < nworkers; i++) {
 638                 if (workers[i] == current)
 639                         break;
 640
 641                 if (my_slot < 0 && workers[i] == NULL)
 642                         my_slot = i;
 643         }
 644
 645         if (i < nworkers || my_slot < 0)
 646                 return -1;
 647
 648         workers[my_slot] = current;
 649
 650         return my_slot;
 651 }
 652
 653 static int
 654 lnet_attempt_msg_resend(struct lnet_msg *msg)
 655 {
 656         struct lnet_msg_container *container;
 657         int my_slot;
 658         int cpt;
 659
 660         /* we can only resend tx_committed messages */
 661         LASSERT(msg->msg_tx_committed);
 662
 663         /* don't resend recovery messages */
 664         if (msg->msg_recovery) {
 665                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 666                         libcfs_nidstr(&msg->msg_from),
 667                         libcfs_nidstr(&msg->msg_target.nid),
 668                         msg->msg_retry_count);
 669                 return -ENOTRECOVERABLE;
 670         }
 671
 672         /*
 673          * if we explicitly indicated we don't want to resend then just
 674          * return
 675          */
 676         if (msg->msg_no_resend) {
 677                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 678                         libcfs_nidstr(&msg->msg_from),
 679                         libcfs_nidstr(&msg->msg_target.nid),
 680                         msg->msg_retry_count);
 681                 return -ENOTRECOVERABLE;
 682         }
 683
 684         /* check if the message has exceeded the number of retries */
 685         if (msg->msg_retry_count >= lnet_retry_count) {
 686                 CNETERR("msg %s->%s exceeded retry count %d\n",
 687                         libcfs_nidstr(&msg->msg_from),
 688                         libcfs_nidstr(&msg->msg_target.nid),
 689                         msg->msg_retry_count);
 690                 return -ENOTRECOVERABLE;
 691         }
 692
 693         cpt = msg->msg_tx_cpt;
 694         lnet_net_lock(cpt);
 695
 696         /* check again under lock */
 697         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 698                 lnet_net_unlock(cpt);
 699                 return -ESHUTDOWN;
 700         }
 701
 702         container = the_lnet.ln_msg_containers[cpt];
 703         my_slot =
 704                 lnet_check_finalize_recursion_locked(msg,
 705                                         &container->msc_resending,
 706                                         container->msc_nfinalizers,
 707                                         container->msc_resenders);
 708
 709         /* enough threads are resending */
 710         if (my_slot == -1) {
 711                 lnet_net_unlock(cpt);
 712                 return 0;
 713         }
 714
 715         while ((msg = list_first_entry_or_null(&container->msc_resending,
 716                                                struct lnet_msg,
 717                                                msg_list)) != NULL) {
 718                 list_del(&msg->msg_list);
 719
 720                 /*
 721                  * resending the message will require us to call
 722                  * lnet_msg_decommit_tx() which will return the credit
 723                  * which this message holds. This could trigger another
 724                  * queued message to be sent. If that message fails and
 725                  * requires a resend we will recurse.
 726                  * But since at this point the slot is taken, the message
 727                  * will be queued in the container and dealt with
 728                  * later. This breaks the recursion.
 729                  */
 730                 lnet_resend_msg_locked(msg);
 731         }
 732
 733         /*
 734          * msc_resenders is an array of process pointers. Each entry holds
 735          * a pointer to the current process operating on the message. An
 736          * array entry is created per CPT. If the array slot is already
 737          * set, then it means that there is a thread on the CPT currently
 738          * resending a message.
 739          * Once the thread finishes clear the slot to enable the thread to
 740          * take on more resend work.
 741          */
 742         container->msc_resenders[my_slot] = NULL;
 743         lnet_net_unlock(cpt);
 744
 745         return 0;
 746 }
 747
 748 /*
 749  * Do a health check on the message:
 750  * return -1 if we're not going to handle the error or
 751  *   if we've reached the maximum number of retries.
 752  *   success case will return -1 as well
 753  * return 0 if it the message is requeued for send
 754  */
 755 static int
 756 lnet_health_check(struct lnet_msg *msg)
 757 {
 758         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 759         struct lnet_peer_ni *lpni;
 760         struct lnet_ni *ni;
 761         bool lo = false;
 762         bool attempt_local_resend;
 763         bool attempt_remote_resend;
 764         bool handle_local_health;
 765         bool handle_remote_health;
 766         ktime_t now;
 767
 768         /* if we're shutting down no point in handling health. */
 769         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 770                 return -1;
 771
 772         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 773
 774         /*
 775          * if we're sending to the LOLND then the msg_txpeer will not be
 776          * set. So no need to sanity check it.
 777          */
 778         if (msg->msg_tx_committed &&
 779             !nid_is_lo0(&msg->msg_txni->ni_nid))
 780                 LASSERT(msg->msg_txpeer);
 781         else if (msg->msg_tx_committed &&
 782                  nid_is_lo0(&msg->msg_txni->ni_nid))
 783                 lo = true;
 784
 785         /*
 786          * always prefer txni/txpeer if they message is committed for both
 787          * directions.
 788          */
 789         if (msg->msg_tx_committed) {
 790                 ni = msg->msg_txni;
 791                 lpni = msg->msg_txpeer;
 792                 attempt_local_resend = attempt_remote_resend = true;
 793         } else {
 794                 ni = msg->msg_rxni;
 795                 lpni = msg->msg_rxpeer;
 796                 attempt_local_resend = attempt_remote_resend = false;
 797         }
 798
 799         if (!lo)
 800                 LASSERT(ni && lpni);
 801         else
 802                 LASSERT(ni);
 803
 804         now = ktime_get();
 805         if (ktime_after(now, msg->msg_deadline)) {
 806                 s64 time = ktime_to_ns(ktime_sub(now, msg->msg_deadline));
 807
 808                 atomic64_add(time, &the_lnet.ln_late_msg_nsecs);
 809                 atomic_inc(&the_lnet.ln_late_msg_count);
 810
 811                 if (hstatus != LNET_MSG_STATUS_OK)
 812                         return -1;
 813         }
 814
 815         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 816                libcfs_nidstr(&ni->ni_nid),
 817                (lo) ? "self" : libcfs_nidstr(&lpni->lpni_nid),
 818                lnet_msgtyp2str(msg->msg_type),
 819                lnet_health_error2str(hstatus));
 820
 821         /*
 822          * stats are only incremented for errors so avoid wasting time
 823          * incrementing statistics if there is no error. Similarly, whether to
 824          * update health values or perform resends is only applicable for
 825          * messages with a health status != OK.
 826          */
 827         if (hstatus != LNET_MSG_STATUS_OK) {
 828                 struct lnet_ping_info *pi;
 829
 830                 /* Don't further decrement the health value if a recovery
 831                  * message failed.
 832                  */
 833                 if (msg->msg_recovery)
 834                         handle_local_health = handle_remote_health = false;
 835                 else
 836                         handle_local_health = handle_remote_health = true;
 837
 838                 /* For local failures, health/recovery/resends are not needed if
 839                  * I only have a single (non-lolnd) interface.
 840                  */
 841                 pi = &the_lnet.ln_ping_target->pb_info;
 842                 if (lnet_ping_at_least_two_entries(pi)) {
 843                         handle_local_health = false;
 844                         attempt_local_resend = false;
 845                 }
 846
 847                 lnet_net_lock(0);
 848                 lnet_incr_hstats(ni, lpni, hstatus);
 849                 /* For remote failures, health/recovery/resends are not needed
 850                  * if the peer only has a single interface. Special case for
 851                  * routers where we rely on health feature to manage route
 852                  * aliveness. NB: lp_nnis does _not_ include the lolnd, so a
 853                  * single-rail node would have lp_nnis == 1.
 854                  */
 855                 if (lpni && lpni->lpni_peer_net &&
 856                     lpni->lpni_peer_net->lpn_peer &&
 857                     lpni->lpni_peer_net->lpn_peer->lp_nnis <= 1) {
 858                         attempt_remote_resend = false;
 859                         if (!lnet_isrouter(lpni))
 860                                 handle_remote_health = false;
 861                 }
 862                 /* Do not put my interfaces into peer NI recovery. They should
 863                  * be handled with local NI recovery.
 864                  */
 865                 if (handle_remote_health && lpni &&
 866                     lnet_nid_to_ni_locked(&lpni->lpni_nid, 0))
 867                         handle_remote_health = false;
 868                 lnet_net_unlock(0);
 869         }
 870
 871         switch (hstatus) {
 872         case LNET_MSG_STATUS_OK:
 873                 /*
 874                  * increment the local ni health whether we successfully
 875                  * received or sent a message on it.
 876                  *
 877                  * Ping counts are reset to 0 as appropriate to allow for
 878                  * faster recovery.
 879                  */
 880                 lnet_inc_healthv(&ni->ni_healthv, lnet_health_sensitivity);
 881                 /*
 882                  * It's possible msg_txpeer is NULL in the LOLND
 883                  * case. Only increment the peer's health if we're
 884                  * receiving a message from it. It's the only sure way to
 885                  * know that a remote interface is up.
 886                  * If this interface is part of a router, then take that
 887                  * as indication that the router is fully healthy.
 888                  */
 889                 if (lpni && msg->msg_rx_committed) {
 890                         lnet_net_lock(0);
 891                         lpni->lpni_ping_count = 0;
 892                         ni->ni_ping_count = 0;
 893                         /*
 894                          * If we're receiving a message from the router or
 895                          * I'm a router, then set that lpni's health to
 896                          * maximum so we can commence communication
 897                          */
 898                         if (lnet_isrouter(lpni) || the_lnet.ln_routing) {
 899                                 lnet_set_lpni_healthv_locked(lpni,
 900                                         LNET_MAX_HEALTH_VALUE);
 901                         } else {
 902                                 lnet_inc_lpni_healthv_locked(lpni);
 903                                 /* This peer NI may have previously aged out
 904                                  * of recovery. Now that we've received a
 905                                  * message from it, we can continue recovery
 906                                  * if its health value is still below the
 907                                  * maximum.
 908                                  */
 909                                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
 910                                                 &the_lnet.ln_mt_peerNIRecovq,
 911                                                 ktime_get_seconds());
 912                         }
 913                         lnet_net_unlock(0);
 914                 }
 915
 916                 /* we can finalize this message */
 917                 return -1;
 918         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 919         case LNET_MSG_STATUS_LOCAL_DROPPED:
 920         case LNET_MSG_STATUS_LOCAL_ABORTED:
 921         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 922         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 923                 if (handle_local_health)
 924                         lnet_handle_local_failure(ni);
 925                 if (attempt_local_resend)
 926                         return lnet_attempt_msg_resend(msg);
 927                 break;
 928         case LNET_MSG_STATUS_LOCAL_ERROR:
 929                 if (handle_local_health)
 930                         lnet_handle_local_failure(ni);
 931                 return -1;
 932         case LNET_MSG_STATUS_REMOTE_DROPPED:
 933                 if (handle_remote_health)
 934                         lnet_handle_remote_failure(lpni);
 935                 if (attempt_remote_resend)
 936                         return lnet_attempt_msg_resend(msg);
 937                 break;
 938         case LNET_MSG_STATUS_REMOTE_ERROR:
 939         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 940                 if (handle_remote_health)
 941                         lnet_handle_remote_failure(lpni);
 942                 return -1;
 943         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 944                 if (handle_remote_health)
 945                         lnet_handle_remote_failure(lpni);
 946                 if (handle_local_health)
 947                         lnet_handle_local_failure(ni);
 948                 return -1;
 949         default:
 950                 LBUG();
 951         }
 952
 953         /* no resend is needed */
 954         return -1;
 955 }
 956
 957 static void
 958 lnet_msg_detach_md(struct lnet_msg *msg, int status)
 959 {
 960         struct lnet_libmd *md = msg->msg_md;
 961         lnet_handler_t handler = NULL;
 962         int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
 963         int unlink;
 964
 965         lnet_res_lock(cpt);
 966         while (md->md_flags & LNET_MD_FLAG_HANDLING)
 967                 /* An event handler is running - wait for it to
 968                  * complete to avoid races.
 969                  */
 970                 lnet_md_wait_handling(md, cpt);
 971
 972         /* Now it's safe to drop my caller's ref */
 973         md->md_refcount--;
 974         LASSERT(md->md_refcount >= 0);
 975
 976         unlink = lnet_md_unlinkable(md);
 977         if (md->md_handler) {
 978                 if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
 979                         msg->msg_ev.status   = -ETIMEDOUT;
 980                         CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
 981                 } else {
 982                         msg->msg_ev.status   = status;
 983                 }
 984                 msg->msg_ev.unlinked = unlink;
 985                 handler = md->md_handler;
 986                 if (!unlink)
 987                         md->md_flags |= LNET_MD_FLAG_HANDLING;
 988         }
 989
 990         if (unlink || (md->md_refcount == 0 &&
 991                        md->md_threshold == LNET_MD_THRESH_INF))
 992                 lnet_detach_rsp_tracker(md, cpt);
 993
 994         msg->msg_md = NULL;
 995         if (unlink)
 996                 lnet_md_unlink(md);
 997
 998         lnet_res_unlock(cpt);
 999
1000         if (handler) {
1001                 handler(&msg->msg_ev);
1002                 if (!unlink) {
1003                         lnet_res_lock(cpt);
1004                         md->md_flags &= ~LNET_MD_FLAG_HANDLING;
1005                         wake_up_var(md);
1006                         lnet_res_unlock(cpt);
1007                 }
1008         }
1009 }
1010
1011 static bool
1012 lnet_is_health_check(struct lnet_msg *msg)
1013 {
1014         bool hc = true;
1015         int status = msg->msg_ev.status;
1016
1017         if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
1018             !msg->msg_onactivelist) {
1019                 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
1020                        msg);
1021                 return false;
1022         }
1023
1024         if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
1025             (msg->msg_rx_committed && !msg->msg_rxpeer)) {
1026                 /* The optimized GET case does not set msg_rxpeer, but status
1027                  * could be zero. Only print the error message if we have a
1028                  * non-zero status.
1029                  */
1030                 if (status)
1031                         CDEBUG(D_NET, "msg %p status %d cannot retry\n", msg,
1032                                status);
1033                 return false;
1034         }
1035
1036         /* Check for status inconsistencies */
1037         if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
1038              (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
1039                 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
1040                        "checking (%d, %d)\n", msg, status,
1041                        msg->msg_health_status);
1042                 hc = false;
1043         }
1044
1045         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
1046                hc, status, msg->msg_health_status);
1047
1048         return hc;
1049 }
1050
1051 char *
1052 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
1053 {
1054         switch (hstatus) {
1055         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
1056                 return "LOCAL_INTERRUPT";
1057         case LNET_MSG_STATUS_LOCAL_DROPPED:
1058                 return "LOCAL_DROPPED";
1059         case LNET_MSG_STATUS_LOCAL_ABORTED:
1060                 return "LOCAL_ABORTED";
1061         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
1062                 return "LOCAL_NO_ROUTE";
1063         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
1064                 return "LOCAL_TIMEOUT";
1065         case LNET_MSG_STATUS_LOCAL_ERROR:
1066                 return "LOCAL_ERROR";
1067         case LNET_MSG_STATUS_REMOTE_DROPPED:
1068                 return "REMOTE_DROPPED";
1069         case LNET_MSG_STATUS_REMOTE_ERROR:
1070                 return "REMOTE_ERROR";
1071         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
1072                 return "REMOTE_TIMEOUT";
1073         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
1074                 return "NETWORK_TIMEOUT";
1075         case LNET_MSG_STATUS_OK:
1076                 return "OK";
1077         default:
1078                 return "<UNKNOWN>";
1079         }
1080 }
1081
1082 bool
1083 lnet_send_error_simulation(struct lnet_msg *msg,
1084                            enum lnet_msg_hstatus *hstatus)
1085 {
1086         if (!msg)
1087                 return false;
1088
1089         if (list_empty(&the_lnet.ln_drop_rules))
1090             return false;
1091
1092         /* match only health rules */
1093         if (!lnet_drop_rule_match(&msg->msg_hdr, NULL, hstatus))
1094                 return false;
1095
1096         CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
1097                 libcfs_nidstr(&msg->msg_hdr.src_nid),
1098                 libcfs_nidstr(&msg->msg_txni->ni_nid),
1099                 libcfs_nidstr(&msg->msg_hdr.dest_nid),
1100                 lnet_msgtyp2str(msg->msg_type),
1101                 lnet_health_error2str(*hstatus));
1102
1103         return true;
1104 }
1105 EXPORT_SYMBOL(lnet_send_error_simulation);
1106
1107 void
1108 lnet_finalize(struct lnet_msg *msg, int status)
1109 {
1110         struct lnet_msg_container *container;
1111         int my_slot;
1112         int cpt;
1113         int rc;
1114
1115         LASSERT(!in_interrupt());
1116
1117         if (msg == NULL)
1118                 return;
1119
1120         msg->msg_ev.status = status;
1121
1122         if (lnet_is_health_check(msg)) {
1123                 /*
1124                  * Check the health status of the message. If it has one
1125                  * of the errors that we're supposed to handle, and it has
1126                  * not timed out, then
1127                  *      1. Decrement the appropriate health_value
1128                  *      2. queue the message on the resend queue
1129
1130                  * if the message send is success, timed out or failed in the
1131                  * health check for any reason then we'll just finalize the
1132                  * message. Otherwise just return since the message has been
1133                  * put on the resend queue.
1134                  */
1135                 if (!lnet_health_check(msg))
1136                         return;
1137         }
1138
1139         /*
1140          * We're not going to resend this message so detach its MD and invoke
1141          * the appropriate callbacks
1142          */
1143         if (msg->msg_md != NULL)
1144                 lnet_msg_detach_md(msg, status);
1145
1146 again:
1147         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
1148                 /* not committed to network yet */
1149                 LASSERT(!msg->msg_onactivelist);
1150                 lnet_msg_free(msg);
1151                 return;
1152         }
1153
1154         /*
1155          * NB: routed message can be committed for both receiving and sending,
1156          * we should finalize in LIFO order and keep counters correct.
1157          * (finalize sending first then finalize receiving)
1158          */
1159         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
1160         lnet_net_lock(cpt);
1161
1162         container = the_lnet.ln_msg_containers[cpt];
1163
1164         /* Recursion breaker.  Don't complete the message here if I am (or
1165          * enough other threads are) already completing messages */
1166         my_slot = lnet_check_finalize_recursion_locked(msg,
1167                                                 &container->msc_finalizing,
1168                                                 container->msc_nfinalizers,
1169                                                 container->msc_finalizers);
1170
1171         /* enough threads are resending */
1172         if (my_slot == -1) {
1173                 lnet_net_unlock(cpt);
1174                 return;
1175         }
1176
1177         rc = 0;
1178         while ((msg = list_first_entry_or_null(&container->msc_finalizing,
1179                                                struct lnet_msg,
1180                                                msg_list)) != NULL) {
1181                 list_del_init(&msg->msg_list);
1182
1183                 /* NB drops and regains the lnet lock if it actually does
1184                  * anything, so my finalizing friends can chomp along too */
1185                 rc = lnet_complete_msg_locked(msg, cpt);
1186                 if (rc != 0)
1187                         break;
1188         }
1189
1190         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1191                 lnet_net_unlock(cpt);
1192                 lnet_delay_rule_check();
1193                 lnet_net_lock(cpt);
1194         }
1195
1196         container->msc_finalizers[my_slot] = NULL;
1197         lnet_net_unlock(cpt);
1198
1199         if (rc != 0)
1200                 goto again;
1201 }
1202 EXPORT_SYMBOL(lnet_finalize);
1203
1204 void
1205 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1206 {
1207         struct lnet_msg *msg;
1208         int count = 0;
1209
1210         if (container->msc_init == 0)
1211                 return;
1212
1213         while ((msg = list_first_entry_or_null(&container->msc_active,
1214                                                struct lnet_msg,
1215                                                msg_activelist)) != NULL) {
1216                 LASSERT(msg->msg_onactivelist);
1217                 msg->msg_onactivelist = 0;
1218                 list_del_init(&msg->msg_activelist);
1219                 lnet_msg_free(msg);
1220                 count++;
1221         }
1222
1223         if (count > 0)
1224                 CERROR("%d active msg on exit\n", count);
1225
1226         if (container->msc_finalizers != NULL) {
1227                 CFS_FREE_PTR_ARRAY(container->msc_finalizers,
1228                                    container->msc_nfinalizers);
1229                 container->msc_finalizers = NULL;
1230         }
1231
1232         if (container->msc_resenders != NULL) {
1233                 CFS_FREE_PTR_ARRAY(container->msc_resenders,
1234                                    container->msc_nfinalizers);
1235                 container->msc_resenders = NULL;
1236         }
1237         container->msc_init = 0;
1238 }
1239
1240 int
1241 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1242 {
1243         int rc = 0;
1244
1245         container->msc_init = 1;
1246
1247         INIT_LIST_HEAD(&container->msc_active);
1248         INIT_LIST_HEAD(&container->msc_finalizing);
1249         INIT_LIST_HEAD(&container->msc_resending);
1250
1251         /* number of CPUs */
1252         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1253         if (container->msc_nfinalizers == 0)
1254                 container->msc_nfinalizers = 1;
1255
1256         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1257                          container->msc_nfinalizers *
1258                          sizeof(*container->msc_finalizers));
1259
1260         if (container->msc_finalizers == NULL) {
1261                 CERROR("Failed to allocate message finalizers\n");
1262                 lnet_msg_container_cleanup(container);
1263                 return -ENOMEM;
1264         }
1265
1266         LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
1267                          container->msc_nfinalizers *
1268                          sizeof(*container->msc_resenders));
1269
1270         if (container->msc_resenders == NULL) {
1271                 CERROR("Failed to allocate message resenders\n");
1272                 lnet_msg_container_cleanup(container);
1273                 return -ENOMEM;
1274         }
1275
1276         return rc;
1277 }
1278
1279 void
1280 lnet_msg_containers_destroy(void)
1281 {
1282         struct lnet_msg_container *container;
1283         int     i;
1284
1285         if (the_lnet.ln_msg_containers == NULL)
1286                 return;
1287
1288         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1289                 lnet_msg_container_cleanup(container);
1290
1291         cfs_percpt_free(the_lnet.ln_msg_containers);
1292         the_lnet.ln_msg_containers = NULL;
1293 }
1294
1295 int
1296 lnet_msg_containers_create(void)
1297 {
1298         struct lnet_msg_container *container;
1299         int     rc;
1300         int     i;
1301
1302         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1303                                                       sizeof(*container));
1304
1305         if (the_lnet.ln_msg_containers == NULL) {
1306                 CERROR("Failed to allocate cpu-partition data for network\n");
1307                 return -ENOMEM;
1308         }
1309
1310         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1311                 rc = lnet_msg_container_setup(container, i);
1312                 if (rc != 0) {
1313                         lnet_msg_containers_destroy();
1314                         return rc;
1315                 }
1316         }
1317
1318         return 0;
1319 }