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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/lib-msg.c
  33  *
  34  * Message decoding, parsing and finalizing routines
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <lnet/lib-lnet.h>
  40
  41 void
  42 lnet_build_unlink_event(struct lnet_libmd *md, struct lnet_event *ev)
  43 {
  44         ENTRY;
  45
  46         memset(ev, 0, sizeof(*ev));
  47
  48         ev->status   = 0;
  49         ev->unlinked = 1;
  50         ev->type     = LNET_EVENT_UNLINK;
  51         lnet_md_deconstruct(md, ev);
  52         lnet_md2handle(&ev->md_handle, md);
  53         EXIT;
  54 }
  55
  56 /*
  57  * Don't need any lock, must be called after lnet_commit_md
  58  */
  59 void
  60 lnet_build_msg_event(struct lnet_msg *msg, enum lnet_event_kind ev_type)
  61 {
  62         struct lnet_hdr *hdr = &msg->msg_hdr;
  63         struct lnet_event *ev = &msg->msg_ev;
  64
  65         LASSERT(!msg->msg_routing);
  66
  67         ev->type = ev_type;
  68         ev->msg_type = msg->msg_type;
  69
  70         if (ev_type == LNET_EVENT_SEND) {
  71                 /* event for active message */
  72                 ev->target.nid    = le64_to_cpu(hdr->dest_nid);
  73                 ev->target.pid    = le32_to_cpu(hdr->dest_pid);
  74                 ev->initiator.nid = LNET_NID_ANY;
  75                 ev->initiator.pid = the_lnet.ln_pid;
  76                 ev->source.nid    = LNET_NID_ANY;
  77                 ev->source.pid    = the_lnet.ln_pid;
  78                 ev->sender        = LNET_NID_ANY;
  79         } else {
  80                 /* event for passive message */
  81                 ev->target.pid    = hdr->dest_pid;
  82                 ev->target.nid    = hdr->dest_nid;
  83                 ev->initiator.pid = hdr->src_pid;
  84                 /* Multi-Rail: resolve src_nid to "primary" peer NID */
  85                 ev->initiator.nid = msg->msg_initiator;
  86                 /* Multi-Rail: track source NID. */
  87                 ev->source.pid    = hdr->src_pid;
  88                 ev->source.nid    = hdr->src_nid;
  89                 ev->rlength       = hdr->payload_length;
  90                 ev->sender        = msg->msg_from;
  91                 ev->mlength       = msg->msg_wanted;
  92                 ev->offset        = msg->msg_offset;
  93         }
  94
  95         switch (ev_type) {
  96         default:
  97                 LBUG();
  98
  99         case LNET_EVENT_PUT: /* passive PUT */
 100                 ev->pt_index   = hdr->msg.put.ptl_index;
 101                 ev->match_bits = hdr->msg.put.match_bits;
 102                 ev->hdr_data   = hdr->msg.put.hdr_data;
 103                 return;
 104
 105         case LNET_EVENT_GET: /* passive GET */
 106                 ev->pt_index   = hdr->msg.get.ptl_index;
 107                 ev->match_bits = hdr->msg.get.match_bits;
 108                 ev->hdr_data   = 0;
 109                 return;
 110
 111         case LNET_EVENT_ACK: /* ACK */
 112                 ev->match_bits = hdr->msg.ack.match_bits;
 113                 ev->mlength    = hdr->msg.ack.mlength;
 114                 return;
 115
 116         case LNET_EVENT_REPLY: /* REPLY */
 117                 return;
 118
 119         case LNET_EVENT_SEND: /* active message */
 120                 if (msg->msg_type == LNET_MSG_PUT) {
 121                         ev->pt_index   = le32_to_cpu(hdr->msg.put.ptl_index);
 122                         ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
 123                         ev->offset     = le32_to_cpu(hdr->msg.put.offset);
 124                         ev->mlength    =
 125                         ev->rlength    = le32_to_cpu(hdr->payload_length);
 126                         ev->hdr_data   = le64_to_cpu(hdr->msg.put.hdr_data);
 127
 128                 } else {
 129                         LASSERT(msg->msg_type == LNET_MSG_GET);
 130                         ev->pt_index   = le32_to_cpu(hdr->msg.get.ptl_index);
 131                         ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
 132                         ev->mlength    =
 133                         ev->rlength    = le32_to_cpu(hdr->msg.get.sink_length);
 134                         ev->offset     = le32_to_cpu(hdr->msg.get.src_offset);
 135                         ev->hdr_data   = 0;
 136                 }
 137                 return;
 138         }
 139 }
 140
 141 void
 142 lnet_msg_commit(struct lnet_msg *msg, int cpt)
 143 {
 144         struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
 145         struct lnet_counters_common *common;
 146         s64 timeout_ns;
 147
 148         /* set the message deadline */
 149         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
 150         msg->msg_deadline = ktime_add_ns(ktime_get(), timeout_ns);
 151
 152         /* routed message can be committed for both receiving and sending */
 153         LASSERT(!msg->msg_tx_committed);
 154
 155         if (msg->msg_sending) {
 156                 LASSERT(!msg->msg_receiving);
 157                 msg->msg_tx_cpt = cpt;
 158                 msg->msg_tx_committed = 1;
 159                 if (msg->msg_rx_committed) { /* routed message REPLY */
 160                         LASSERT(msg->msg_onactivelist);
 161                         return;
 162                 }
 163         } else {
 164                 LASSERT(!msg->msg_sending);
 165                 msg->msg_rx_cpt = cpt;
 166                 msg->msg_rx_committed = 1;
 167         }
 168
 169         LASSERT(!msg->msg_onactivelist);
 170
 171         msg->msg_onactivelist = 1;
 172         list_add_tail(&msg->msg_activelist, &container->msc_active);
 173
 174         common = &the_lnet.ln_counters[cpt]->lct_common;
 175         common->lcc_msgs_alloc++;
 176         if (common->lcc_msgs_alloc > common->lcc_msgs_max)
 177                 common->lcc_msgs_max = common->lcc_msgs_alloc;
 178 }
 179
 180 static void
 181 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
 182 {
 183         struct lnet_counters_common *common;
 184         struct lnet_event *ev = &msg->msg_ev;
 185
 186         LASSERT(msg->msg_tx_committed);
 187         if (status != 0)
 188                 goto out;
 189
 190         common = &(the_lnet.ln_counters[msg->msg_tx_cpt]->lct_common);
 191         switch (ev->type) {
 192         default: /* routed message */
 193                 LASSERT(msg->msg_routing);
 194                 LASSERT(msg->msg_rx_committed);
 195                 LASSERT(ev->type == 0);
 196
 197                 common->lcc_route_length += msg->msg_len;
 198                 common->lcc_route_count++;
 199                 goto incr_stats;
 200
 201         case LNET_EVENT_PUT:
 202                 /* should have been decommitted */
 203                 LASSERT(!msg->msg_rx_committed);
 204                 /* overwritten while sending ACK */
 205                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 206                 msg->msg_type = LNET_MSG_PUT; /* fix type */
 207                 break;
 208
 209         case LNET_EVENT_SEND:
 210                 LASSERT(!msg->msg_rx_committed);
 211                 if (msg->msg_type == LNET_MSG_PUT)
 212                         common->lcc_send_length += msg->msg_len;
 213                 break;
 214
 215         case LNET_EVENT_GET:
 216                 LASSERT(msg->msg_rx_committed);
 217                 /* overwritten while sending reply, we should never be
 218                  * here for optimized GET */
 219                 LASSERT(msg->msg_type == LNET_MSG_REPLY);
 220                 msg->msg_type = LNET_MSG_GET; /* fix type */
 221                 break;
 222         }
 223
 224         common->lcc_send_count++;
 225
 226 incr_stats:
 227         if (msg->msg_txpeer)
 228                 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 229                                 msg->msg_type,
 230                                 LNET_STATS_TYPE_SEND);
 231         if (msg->msg_txni)
 232                 lnet_incr_stats(&msg->msg_txni->ni_stats,
 233                                 msg->msg_type,
 234                                 LNET_STATS_TYPE_SEND);
 235  out:
 236         lnet_return_tx_credits_locked(msg);
 237         msg->msg_tx_committed = 0;
 238 }
 239
 240 static void
 241 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
 242 {
 243         struct lnet_counters_common *common;
 244         struct lnet_event *ev = &msg->msg_ev;
 245
 246         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 247         LASSERT(msg->msg_rx_committed);
 248
 249         if (status != 0)
 250                 goto out;
 251
 252         common = &(the_lnet.ln_counters[msg->msg_rx_cpt]->lct_common);
 253         switch (ev->type) {
 254         default:
 255                 LASSERT(ev->type == 0);
 256                 LASSERT(msg->msg_routing);
 257                 goto incr_stats;
 258
 259         case LNET_EVENT_ACK:
 260                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 261                 break;
 262
 263         case LNET_EVENT_GET:
 264                 /* type is "REPLY" if it's an optimized GET on passive side,
 265                  * because optimized GET will never be committed for sending,
 266                  * so message type wouldn't be changed back to "GET" by
 267                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 268                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 269                         msg->msg_type == LNET_MSG_GET);
 270                 common->lcc_send_length += msg->msg_wanted;
 271                 break;
 272
 273         case LNET_EVENT_PUT:
 274                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 275                 break;
 276
 277         case LNET_EVENT_REPLY:
 278                 /* type is "GET" if it's an optimized GET on active side,
 279                  * see details in lnet_create_reply_msg() */
 280                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 281                         msg->msg_type == LNET_MSG_REPLY);
 282                 break;
 283         }
 284
 285         common->lcc_recv_count++;
 286
 287 incr_stats:
 288         if (msg->msg_rxpeer)
 289                 lnet_incr_stats(&msg->msg_rxpeer->lpni_stats,
 290                                 msg->msg_type,
 291                                 LNET_STATS_TYPE_RECV);
 292         if (msg->msg_rxni)
 293                 lnet_incr_stats(&msg->msg_rxni->ni_stats,
 294                                 msg->msg_type,
 295                                 LNET_STATS_TYPE_RECV);
 296         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 297                 common->lcc_recv_length += msg->msg_wanted;
 298
 299  out:
 300         lnet_return_rx_credits_locked(msg);
 301         msg->msg_rx_committed = 0;
 302 }
 303
 304 void
 305 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
 306 {
 307         int     cpt2 = cpt;
 308
 309         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 310         LASSERT(msg->msg_onactivelist);
 311
 312         if (msg->msg_tx_committed) { /* always decommit for sending first */
 313                 LASSERT(cpt == msg->msg_tx_cpt);
 314                 lnet_msg_decommit_tx(msg, status);
 315         }
 316
 317         if (msg->msg_rx_committed) {
 318                 /* forwarding msg committed for both receiving and sending */
 319                 if (cpt != msg->msg_rx_cpt) {
 320                         lnet_net_unlock(cpt);
 321                         cpt2 = msg->msg_rx_cpt;
 322                         lnet_net_lock(cpt2);
 323                 }
 324                 lnet_msg_decommit_rx(msg, status);
 325         }
 326
 327         list_del(&msg->msg_activelist);
 328         msg->msg_onactivelist = 0;
 329
 330         the_lnet.ln_counters[cpt2]->lct_common.lcc_msgs_alloc--;
 331
 332         if (cpt2 != cpt) {
 333                 lnet_net_unlock(cpt2);
 334                 lnet_net_lock(cpt);
 335         }
 336 }
 337
 338 void
 339 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
 340                    unsigned int offset, unsigned int mlen)
 341 {
 342         /* NB: @offset and @len are only useful for receiving */
 343         /* Here, we attach the MD on lnet_msg and mark it busy and
 344          * decrementing its threshold. Come what may, the lnet_msg "owns"
 345          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 346          * signals completion. */
 347         LASSERT(!msg->msg_routing);
 348
 349         msg->msg_md = md;
 350         if (msg->msg_receiving) { /* committed for receiving */
 351                 msg->msg_offset = offset;
 352                 msg->msg_wanted = mlen;
 353         }
 354
 355         md->md_refcount++;
 356         if (md->md_threshold != LNET_MD_THRESH_INF) {
 357                 LASSERT(md->md_threshold > 0);
 358                 md->md_threshold--;
 359         }
 360
 361         /* build umd in event */
 362         lnet_md2handle(&msg->msg_ev.md_handle, md);
 363         lnet_md_deconstruct(md, &msg->msg_ev);
 364 }
 365
 366 static int
 367 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
 368 {
 369         struct lnet_handle_wire ack_wmd;
 370         int                rc;
 371         int                status = msg->msg_ev.status;
 372
 373         LASSERT(msg->msg_onactivelist);
 374
 375         if (status == 0 && msg->msg_ack) {
 376                 /* Only send an ACK if the PUT completed successfully */
 377
 378                 lnet_msg_decommit(msg, cpt, 0);
 379
 380                 msg->msg_ack = 0;
 381                 lnet_net_unlock(cpt);
 382
 383                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 384                 LASSERT(!msg->msg_routing);
 385
 386                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 387
 388                 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
 389
 390                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 391                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 392                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 393
 394                 rc = lnet_send(msg->msg_ev.target.nid, msg, 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(LNET_NID_ANY, msg, LNET_NID_ANY);
 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 static void
 441 lnet_dec_healthv_locked(atomic_t *healthv, int sensitivity)
 442 {
 443         int h = atomic_read(healthv);
 444
 445         if (h < sensitivity) {
 446                 atomic_set(healthv, 0);
 447         } else {
 448                 h -= sensitivity;
 449                 atomic_set(healthv, h);
 450         }
 451 }
 452
 453 static void
 454 lnet_handle_local_failure(struct lnet_ni *local_ni)
 455 {
 456         /*
 457          * the lnet_net_lock(0) is used to protect the addref on the ni
 458          * and the recovery queue.
 459          */
 460         lnet_net_lock(0);
 461         /* the mt could've shutdown and cleaned up the queues */
 462         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 463                 lnet_net_unlock(0);
 464                 return;
 465         }
 466
 467         lnet_dec_healthv_locked(&local_ni->ni_healthv, lnet_health_sensitivity);
 468         /*
 469          * add the NI to the recovery queue if it's not already there
 470          * and it's health value is actually below the maximum. It's
 471          * possible that the sensitivity might be set to 0, and the health
 472          * value will not be reduced. In this case, there is no reason to
 473          * invoke recovery
 474          */
 475         if (list_empty(&local_ni->ni_recovery) &&
 476             atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
 477                 CDEBUG(D_NET, "ni %s added to recovery queue. Health = %d\n",
 478                         libcfs_nid2str(local_ni->ni_nid),
 479                         atomic_read(&local_ni->ni_healthv));
 480                 list_add_tail(&local_ni->ni_recovery,
 481                               &the_lnet.ln_mt_localNIRecovq);
 482                 lnet_ni_addref_locked(local_ni, 0);
 483         }
 484         lnet_net_unlock(0);
 485 }
 486
 487 void
 488 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
 489 {
 490         __u32 sensitivity = lnet_health_sensitivity;
 491         __u32 lp_sensitivity;
 492
 493         /*
 494          * NO-OP if:
 495          * 1. lpni could be NULL if we're in the LOLND case
 496          * 2. this is a recovery message
 497          */
 498         if (!lpni)
 499                 return;
 500
 501         /*
 502          * If there is a health sensitivity in the peer then use that
 503          * instead of the globally set one.
 504          */
 505         lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
 506         if (lp_sensitivity)
 507                 sensitivity = lp_sensitivity;
 508
 509         lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
 510
 511         /* update the peer_net's health value */
 512         lnet_update_peer_net_healthv(lpni);
 513
 514         /*
 515          * add the peer NI to the recovery queue if it's not already there
 516          * and it's health value is actually below the maximum. It's
 517          * possible that the sensitivity might be set to 0, and the health
 518          * value will not be reduced. In this case, there is no reason to
 519          * invoke recovery
 520          */
 521         lnet_peer_ni_add_to_recoveryq_locked(lpni);
 522 }
 523
 524 static void
 525 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 526 {
 527         /* lpni could be NULL if we're in the LOLND case */
 528         if (!lpni)
 529                 return;
 530
 531         lnet_net_lock(0);
 532         /* the mt could've shutdown and cleaned up the queues */
 533         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 534                 lnet_net_unlock(0);
 535                 return;
 536         }
 537         lnet_handle_remote_failure_locked(lpni);
 538         lnet_net_unlock(0);
 539 }
 540
 541 static void
 542 lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 543                  enum lnet_msg_hstatus hstatus)
 544 {
 545         struct lnet_counters_health *health;
 546
 547         health = &the_lnet.ln_counters[0]->lct_health;
 548
 549         switch (hstatus) {
 550         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 551                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 552                 health->lch_local_interrupt_count++;
 553                 break;
 554         case LNET_MSG_STATUS_LOCAL_DROPPED:
 555                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 556                 health->lch_local_dropped_count++;
 557                 break;
 558         case LNET_MSG_STATUS_LOCAL_ABORTED:
 559                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 560                 health->lch_local_aborted_count++;
 561                 break;
 562         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 563                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 564                 health->lch_local_no_route_count++;
 565                 break;
 566         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 567                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 568                 health->lch_local_timeout_count++;
 569                 break;
 570         case LNET_MSG_STATUS_LOCAL_ERROR:
 571                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 572                 health->lch_local_error_count++;
 573                 break;
 574         case LNET_MSG_STATUS_REMOTE_DROPPED:
 575                 if (lpni)
 576                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 577                 health->lch_remote_dropped_count++;
 578                 break;
 579         case LNET_MSG_STATUS_REMOTE_ERROR:
 580                 if (lpni)
 581                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 582                 health->lch_remote_error_count++;
 583                 break;
 584         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 585                 if (lpni)
 586                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 587                 health->lch_remote_timeout_count++;
 588                 break;
 589         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 590                 if (lpni)
 591                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 592                 health->lch_network_timeout_count++;
 593                 break;
 594         case LNET_MSG_STATUS_OK:
 595                 break;
 596         default:
 597                 LBUG();
 598         }
 599 }
 600
 601 static void
 602 lnet_resend_msg_locked(struct lnet_msg *msg)
 603 {
 604         msg->msg_retry_count++;
 605
 606         /*
 607          * remove message from the active list and reset it to prepare
 608          * for a resend. Two exceptions to this
 609          *
 610          * 1. the router case. When a message is being routed it is
 611          * committed for rx when received and committed for tx when
 612          * forwarded. We don't want to remove it from the active list, since
 613          * code which handles receiving expects it to remain on the active
 614          * list.
 615          *
 616          * 2. The REPLY case. Reply messages use the same message
 617          * structure for the GET that was received.
 618          */
 619         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 620                 list_del_init(&msg->msg_activelist);
 621                 msg->msg_onactivelist = 0;
 622         }
 623         /*
 624          * The msg_target.nid which was originally set
 625          * when calling LNetGet() or LNetPut() might've
 626          * been overwritten if we're routing this message.
 627          * Call lnet_msg_decommit_tx() to return the credit
 628          * this message consumed. The message will
 629          * consume another credit when it gets resent.
 630          */
 631         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 632         lnet_msg_decommit_tx(msg, -EAGAIN);
 633         msg->msg_sending = 0;
 634         msg->msg_receiving = 0;
 635         msg->msg_target_is_router = 0;
 636
 637         CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
 638                libcfs_nid2str(msg->msg_hdr.src_nid),
 639                libcfs_nid2str(msg->msg_hdr.dest_nid),
 640                lnet_msgtyp2str(msg->msg_type),
 641                lnet_health_error2str(msg->msg_health_status), msg);
 642
 643         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 644
 645         complete(&the_lnet.ln_mt_wait_complete);
 646 }
 647
 648 int
 649 lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
 650                                      struct list_head *containerq,
 651                                      int nworkers, void **workers)
 652 {
 653         int my_slot = -1;
 654         int i;
 655
 656         list_add_tail(&msg->msg_list, containerq);
 657
 658         for (i = 0; i < nworkers; i++) {
 659                 if (workers[i] == current)
 660                         break;
 661
 662                 if (my_slot < 0 && workers[i] == NULL)
 663                         my_slot = i;
 664         }
 665
 666         if (i < nworkers || my_slot < 0)
 667                 return -1;
 668
 669         workers[my_slot] = current;
 670
 671         return my_slot;
 672 }
 673
 674 int
 675 lnet_attempt_msg_resend(struct lnet_msg *msg)
 676 {
 677         struct lnet_msg_container *container;
 678         int my_slot;
 679         int cpt;
 680
 681         /* we can only resend tx_committed messages */
 682         LASSERT(msg->msg_tx_committed);
 683
 684         /* don't resend recovery messages */
 685         if (msg->msg_recovery) {
 686                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 687                         libcfs_nid2str(msg->msg_from),
 688                         libcfs_nid2str(msg->msg_target.nid),
 689                         msg->msg_retry_count);
 690                 return -ENOTRECOVERABLE;
 691         }
 692
 693         /*
 694          * if we explicitly indicated we don't want to resend then just
 695          * return
 696          */
 697         if (msg->msg_no_resend) {
 698                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 699                         libcfs_nid2str(msg->msg_from),
 700                         libcfs_nid2str(msg->msg_target.nid),
 701                         msg->msg_retry_count);
 702                 return -ENOTRECOVERABLE;
 703         }
 704
 705         /* check if the message has exceeded the number of retries */
 706         if (msg->msg_retry_count >= lnet_retry_count) {
 707                 CNETERR("msg %s->%s exceeded retry count %d\n",
 708                         libcfs_nid2str(msg->msg_from),
 709                         libcfs_nid2str(msg->msg_target.nid),
 710                         msg->msg_retry_count);
 711                 return -ENOTRECOVERABLE;
 712         }
 713
 714         cpt = msg->msg_tx_cpt;
 715         lnet_net_lock(cpt);
 716
 717         /* check again under lock */
 718         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 719                 lnet_net_unlock(cpt);
 720                 return -ESHUTDOWN;
 721         }
 722
 723         container = the_lnet.ln_msg_containers[cpt];
 724         my_slot =
 725                 lnet_check_finalize_recursion_locked(msg,
 726                                         &container->msc_resending,
 727                                         container->msc_nfinalizers,
 728                                         container->msc_resenders);
 729
 730         /* enough threads are resending */
 731         if (my_slot == -1) {
 732                 lnet_net_unlock(cpt);
 733                 return 0;
 734         }
 735
 736         while (!list_empty(&container->msc_resending)) {
 737                 msg = list_entry(container->msc_resending.next,
 738                                         struct lnet_msg, msg_list);
 739                 list_del(&msg->msg_list);
 740
 741                 /*
 742                  * resending the message will require us to call
 743                  * lnet_msg_decommit_tx() which will return the credit
 744                  * which this message holds. This could trigger another
 745                  * queued message to be sent. If that message fails and
 746                  * requires a resend we will recurse.
 747                  * But since at this point the slot is taken, the message
 748                  * will be queued in the container and dealt with
 749                  * later. This breaks the recursion.
 750                  */
 751                 lnet_resend_msg_locked(msg);
 752         }
 753
 754         /*
 755          * msc_resenders is an array of process pointers. Each entry holds
 756          * a pointer to the current process operating on the message. An
 757          * array entry is created per CPT. If the array slot is already
 758          * set, then it means that there is a thread on the CPT currently
 759          * resending a message.
 760          * Once the thread finishes clear the slot to enable the thread to
 761          * take on more resend work.
 762          */
 763         container->msc_resenders[my_slot] = NULL;
 764         lnet_net_unlock(cpt);
 765
 766         return 0;
 767 }
 768
 769 /*
 770  * Do a health check on the message:
 771  * return -1 if we're not going to handle the error or
 772  *   if we've reached the maximum number of retries.
 773  *   success case will return -1 as well
 774  * return 0 if it the message is requeued for send
 775  */
 776 static int
 777 lnet_health_check(struct lnet_msg *msg)
 778 {
 779         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 780         struct lnet_peer_ni *lpni;
 781         struct lnet_ni *ni;
 782         bool lo = false;
 783         bool attempt_local_resend;
 784         bool attempt_remote_resend;
 785         bool handle_local_health;
 786         bool handle_remote_health;
 787
 788         /* if we're shutting down no point in handling health. */
 789         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 790                 return -1;
 791
 792         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 793
 794         /*
 795          * if we're sending to the LOLND then the msg_txpeer will not be
 796          * set. So no need to sanity check it.
 797          */
 798         if (msg->msg_tx_committed && msg->msg_txni->ni_nid != LNET_NID_LO_0)
 799                 LASSERT(msg->msg_txpeer);
 800         else if (msg->msg_tx_committed &&
 801                  msg->msg_txni->ni_nid == LNET_NID_LO_0)
 802                 lo = true;
 803
 804         if (hstatus != LNET_MSG_STATUS_OK &&
 805             ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
 806                 return -1;
 807
 808         /*
 809          * always prefer txni/txpeer if they message is committed for both
 810          * directions.
 811          */
 812         if (msg->msg_tx_committed) {
 813                 ni = msg->msg_txni;
 814                 lpni = msg->msg_txpeer;
 815                 attempt_local_resend = attempt_remote_resend = true;
 816         } else {
 817                 ni = msg->msg_rxni;
 818                 lpni = msg->msg_rxpeer;
 819                 attempt_local_resend = attempt_remote_resend = false;
 820         }
 821
 822         /* Don't further decrement the health value if a recovery message
 823          * failed.
 824          */
 825         if (msg->msg_recovery)
 826                 handle_local_health = handle_remote_health = false;
 827         else
 828                 handle_local_health = handle_remote_health = true;
 829
 830         /* For local failures, health/recovery/resends are not needed if I only
 831          * have a single (non-lolnd) interface. NB: pb_nnis includes the lolnd
 832          * interface, so a single-rail node would have pb_nnis == 2.
 833          */
 834         if (the_lnet.ln_ping_target->pb_nnis <= 2) {
 835                 handle_local_health = false;
 836                 attempt_local_resend = false;
 837         }
 838
 839         /* For remote failures, health/recovery/resends are not needed if the
 840          * peer only has a single interface. Special case for routers where we
 841          * rely on health feature to manage route aliveness. NB: unlike pb_nnis
 842          * above, lp_nnis does _not_ include the lolnd, so a single-rail node
 843          * would have lp_nnis == 1.
 844          */
 845         if (lpni && lpni->lpni_peer_net->lpn_peer->lp_nnis <= 1) {
 846                 attempt_remote_resend = false;
 847                 if (!lnet_isrouter(lpni))
 848                         handle_remote_health = false;
 849         }
 850
 851         if (!lo)
 852                 LASSERT(ni && lpni);
 853         else
 854                 LASSERT(ni);
 855
 856         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 857                libcfs_nid2str(ni->ni_nid),
 858                (lo) ? "self" : libcfs_nid2str(lpni->lpni_nid),
 859                lnet_msgtyp2str(msg->msg_type),
 860                lnet_health_error2str(hstatus));
 861
 862         /*
 863          * stats are only incremented for errors so avoid wasting time
 864          * incrementing statistics if there is no error.
 865          */
 866         if (hstatus != LNET_MSG_STATUS_OK) {
 867                 lnet_net_lock(0);
 868                 lnet_incr_hstats(ni, lpni, hstatus);
 869                 lnet_net_unlock(0);
 870         }
 871
 872         switch (hstatus) {
 873         case LNET_MSG_STATUS_OK:
 874                 /*
 875                  * increment the local ni health weather we successfully
 876                  * received or sent a message on it.
 877                  */
 878                 lnet_inc_healthv(&ni->ni_healthv, lnet_health_sensitivity);
 879                 /*
 880                  * It's possible msg_txpeer is NULL in the LOLND
 881                  * case. Only increment the peer's health if we're
 882                  * receiving a message from it. It's the only sure way to
 883                  * know that a remote interface is up.
 884                  * If this interface is part of a router, then take that
 885                  * as indication that the router is fully healthy.
 886                  */
 887                 if (lpni && msg->msg_rx_committed) {
 888                         /*
 889                          * If we're receiving a message from the router or
 890                          * I'm a router, then set that lpni's health to
 891                          * maximum so we can commence communication
 892                          */
 893                         lnet_net_lock(0);
 894                         if (lnet_isrouter(lpni) || the_lnet.ln_routing) {
 895                                 lnet_set_lpni_healthv_locked(lpni,
 896                                         LNET_MAX_HEALTH_VALUE);
 897                         } else {
 898                                 __u32 sensitivity = lpni->lpni_peer_net->
 899                                         lpn_peer->lp_health_sensitivity;
 900
 901                                 lnet_inc_lpni_healthv_locked(lpni,
 902                                         (sensitivity) ? sensitivity :
 903                                         lnet_health_sensitivity);
 904                         }
 905                         lnet_net_unlock(0);
 906                 }
 907
 908                 /* we can finalize this message */
 909                 return -1;
 910         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 911         case LNET_MSG_STATUS_LOCAL_DROPPED:
 912         case LNET_MSG_STATUS_LOCAL_ABORTED:
 913         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 914         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 915                 if (handle_local_health)
 916                         lnet_handle_local_failure(ni);
 917                 if (attempt_local_resend)
 918                         return lnet_attempt_msg_resend(msg);
 919                 break;
 920         case LNET_MSG_STATUS_LOCAL_ERROR:
 921                 if (handle_local_health)
 922                         lnet_handle_local_failure(ni);
 923                 return -1;
 924         case LNET_MSG_STATUS_REMOTE_DROPPED:
 925                 if (handle_remote_health)
 926                         lnet_handle_remote_failure(lpni);
 927                 if (attempt_remote_resend)
 928                         return lnet_attempt_msg_resend(msg);
 929                 break;
 930         case LNET_MSG_STATUS_REMOTE_ERROR:
 931         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 932                 if (handle_remote_health)
 933                         lnet_handle_remote_failure(lpni);
 934                 return -1;
 935         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 936                 if (handle_remote_health)
 937                         lnet_handle_remote_failure(lpni);
 938                 if (handle_local_health)
 939                         lnet_handle_local_failure(ni);
 940                 return -1;
 941         default:
 942                 LBUG();
 943         }
 944
 945         /* no resend is needed */
 946         return -1;
 947 }
 948
 949 static void
 950 lnet_msg_detach_md(struct lnet_msg *msg, int status)
 951 {
 952         struct lnet_libmd *md = msg->msg_md;
 953         lnet_handler_t handler = NULL;
 954         int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
 955         int unlink;
 956
 957         lnet_res_lock(cpt);
 958         while (md->md_flags & LNET_MD_FLAG_HANDLING)
 959                 /* An event handler is running - wait for it to
 960                  * complete to avoid races.
 961                  */
 962                 lnet_md_wait_handling(md, cpt);
 963
 964         /* Now it's safe to drop my caller's ref */
 965         md->md_refcount--;
 966         LASSERT(md->md_refcount >= 0);
 967
 968         unlink = lnet_md_unlinkable(md);
 969         if (md->md_handler) {
 970                 if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
 971                         msg->msg_ev.status   = -ETIMEDOUT;
 972                         CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
 973                 } else {
 974                         msg->msg_ev.status   = status;
 975                 }
 976                 msg->msg_ev.unlinked = unlink;
 977                 handler = md->md_handler;
 978                 if (!unlink)
 979                         md->md_flags |= LNET_MD_FLAG_HANDLING;
 980         }
 981
 982         if (unlink || (md->md_refcount == 0 &&
 983                        md->md_threshold == LNET_MD_THRESH_INF))
 984                 lnet_detach_rsp_tracker(md, cpt);
 985
 986         msg->msg_md = NULL;
 987         if (unlink)
 988                 lnet_md_unlink(md);
 989
 990         lnet_res_unlock(cpt);
 991
 992         if (handler) {
 993                 handler(&msg->msg_ev);
 994                 if (!unlink) {
 995                         lnet_res_lock(cpt);
 996                         md->md_flags &= ~LNET_MD_FLAG_HANDLING;
 997                         wake_up_var(md);
 998                         lnet_res_unlock(cpt);
 999                 }
1000         }
1001 }
1002
1003 static bool
1004 lnet_is_health_check(struct lnet_msg *msg)
1005 {
1006         bool hc = true;
1007         int status = msg->msg_ev.status;
1008
1009         if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
1010             !msg->msg_onactivelist) {
1011                 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
1012                        msg);
1013                 return false;
1014         }
1015
1016         if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
1017             (msg->msg_rx_committed && !msg->msg_rxpeer)) {
1018                 /* The optimized GET case does not set msg_rxpeer, but status
1019                  * could be zero. Only print the error message if we have a
1020                  * non-zero status.
1021                  */
1022                 if (status)
1023                         CDEBUG(D_NET, "msg %p status %d cannot retry\n", msg,
1024                                status);
1025                 return false;
1026         }
1027
1028         /* Check for status inconsistencies */
1029         if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
1030              (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
1031                 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
1032                        "checking (%d, %d)\n", msg, status,
1033                        msg->msg_health_status);
1034                 hc = false;
1035         }
1036
1037         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
1038                hc, status, msg->msg_health_status);
1039
1040         return hc;
1041 }
1042
1043 char *
1044 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
1045 {
1046         switch (hstatus) {
1047         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
1048                 return "LOCAL_INTERRUPT";
1049         case LNET_MSG_STATUS_LOCAL_DROPPED:
1050                 return "LOCAL_DROPPED";
1051         case LNET_MSG_STATUS_LOCAL_ABORTED:
1052                 return "LOCAL_ABORTED";
1053         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
1054                 return "LOCAL_NO_ROUTE";
1055         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
1056                 return "LOCAL_TIMEOUT";
1057         case LNET_MSG_STATUS_LOCAL_ERROR:
1058                 return "LOCAL_ERROR";
1059         case LNET_MSG_STATUS_REMOTE_DROPPED:
1060                 return "REMOTE_DROPPED";
1061         case LNET_MSG_STATUS_REMOTE_ERROR:
1062                 return "REMOTE_ERROR";
1063         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
1064                 return "REMOTE_TIMEOUT";
1065         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
1066                 return "NETWORK_TIMEOUT";
1067         case LNET_MSG_STATUS_OK:
1068                 return "OK";
1069         default:
1070                 return "<UNKNOWN>";
1071         }
1072 }
1073
1074 bool
1075 lnet_send_error_simulation(struct lnet_msg *msg,
1076                            enum lnet_msg_hstatus *hstatus)
1077 {
1078         if (!msg)
1079                 return false;
1080
1081         if (list_empty(&the_lnet.ln_drop_rules))
1082             return false;
1083
1084         /* match only health rules */
1085         if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
1086                                   hstatus))
1087                 return false;
1088
1089         CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
1090                 libcfs_nid2str(msg->msg_hdr.src_nid),
1091                 libcfs_nid2str(msg->msg_txni->ni_nid),
1092                 libcfs_nid2str(msg->msg_hdr.dest_nid),
1093                 lnet_msgtyp2str(msg->msg_type),
1094                 lnet_health_error2str(*hstatus));
1095
1096         return true;
1097 }
1098 EXPORT_SYMBOL(lnet_send_error_simulation);
1099
1100 void
1101 lnet_finalize(struct lnet_msg *msg, int status)
1102 {
1103         struct lnet_msg_container *container;
1104         int my_slot;
1105         int cpt;
1106         int rc;
1107
1108         LASSERT(!in_interrupt());
1109
1110         if (msg == NULL)
1111                 return;
1112
1113         msg->msg_ev.status = status;
1114
1115         if (lnet_is_health_check(msg)) {
1116                 /*
1117                  * Check the health status of the message. If it has one
1118                  * of the errors that we're supposed to handle, and it has
1119                  * not timed out, then
1120                  *      1. Decrement the appropriate health_value
1121                  *      2. queue the message on the resend queue
1122
1123                  * if the message send is success, timed out or failed in the
1124                  * health check for any reason then we'll just finalize the
1125                  * message. Otherwise just return since the message has been
1126                  * put on the resend queue.
1127                  */
1128                 if (!lnet_health_check(msg))
1129                         return;
1130         }
1131
1132         /*
1133          * We're not going to resend this message so detach its MD and invoke
1134          * the appropriate callbacks
1135          */
1136         if (msg->msg_md != NULL)
1137                 lnet_msg_detach_md(msg, status);
1138
1139 again:
1140         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
1141                 /* not committed to network yet */
1142                 LASSERT(!msg->msg_onactivelist);
1143                 lnet_msg_free(msg);
1144                 return;
1145         }
1146
1147         /*
1148          * NB: routed message can be committed for both receiving and sending,
1149          * we should finalize in LIFO order and keep counters correct.
1150          * (finalize sending first then finalize receiving)
1151          */
1152         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
1153         lnet_net_lock(cpt);
1154
1155         container = the_lnet.ln_msg_containers[cpt];
1156
1157         /* Recursion breaker.  Don't complete the message here if I am (or
1158          * enough other threads are) already completing messages */
1159         my_slot = lnet_check_finalize_recursion_locked(msg,
1160                                                 &container->msc_finalizing,
1161                                                 container->msc_nfinalizers,
1162                                                 container->msc_finalizers);
1163
1164         /* enough threads are resending */
1165         if (my_slot == -1) {
1166                 lnet_net_unlock(cpt);
1167                 return;
1168         }
1169
1170         rc = 0;
1171         while (!list_empty(&container->msc_finalizing)) {
1172                 msg = list_entry(container->msc_finalizing.next,
1173                                  struct lnet_msg, msg_list);
1174
1175                 list_del_init(&msg->msg_list);
1176
1177                 /* NB drops and regains the lnet lock if it actually does
1178                  * anything, so my finalizing friends can chomp along too */
1179                 rc = lnet_complete_msg_locked(msg, cpt);
1180                 if (rc != 0)
1181                         break;
1182         }
1183
1184         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1185                 lnet_net_unlock(cpt);
1186                 lnet_delay_rule_check();
1187                 lnet_net_lock(cpt);
1188         }
1189
1190         container->msc_finalizers[my_slot] = NULL;
1191         lnet_net_unlock(cpt);
1192
1193         if (rc != 0)
1194                 goto again;
1195 }
1196 EXPORT_SYMBOL(lnet_finalize);
1197
1198 void
1199 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1200 {
1201         int     count = 0;
1202
1203         if (container->msc_init == 0)
1204                 return;
1205
1206         while (!list_empty(&container->msc_active)) {
1207                 struct lnet_msg *msg;
1208
1209                 msg  = list_entry(container->msc_active.next,
1210                                   struct lnet_msg, msg_activelist);
1211                 LASSERT(msg->msg_onactivelist);
1212                 msg->msg_onactivelist = 0;
1213                 list_del_init(&msg->msg_activelist);
1214                 lnet_msg_free(msg);
1215                 count++;
1216         }
1217
1218         if (count > 0)
1219                 CERROR("%d active msg on exit\n", count);
1220
1221         if (container->msc_finalizers != NULL) {
1222                 CFS_FREE_PTR_ARRAY(container->msc_finalizers,
1223                                    container->msc_nfinalizers);
1224                 container->msc_finalizers = NULL;
1225         }
1226
1227         if (container->msc_resenders != NULL) {
1228                 CFS_FREE_PTR_ARRAY(container->msc_resenders,
1229                                    container->msc_nfinalizers);
1230                 container->msc_resenders = NULL;
1231         }
1232         container->msc_init = 0;
1233 }
1234
1235 int
1236 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1237 {
1238         int rc = 0;
1239
1240         container->msc_init = 1;
1241
1242         INIT_LIST_HEAD(&container->msc_active);
1243         INIT_LIST_HEAD(&container->msc_finalizing);
1244         INIT_LIST_HEAD(&container->msc_resending);
1245
1246         /* number of CPUs */
1247         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1248         if (container->msc_nfinalizers == 0)
1249                 container->msc_nfinalizers = 1;
1250
1251         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1252                          container->msc_nfinalizers *
1253                          sizeof(*container->msc_finalizers));
1254
1255         if (container->msc_finalizers == NULL) {
1256                 CERROR("Failed to allocate message finalizers\n");
1257                 lnet_msg_container_cleanup(container);
1258                 return -ENOMEM;
1259         }
1260
1261         LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
1262                          container->msc_nfinalizers *
1263                          sizeof(*container->msc_resenders));
1264
1265         if (container->msc_resenders == NULL) {
1266                 CERROR("Failed to allocate message resenders\n");
1267                 lnet_msg_container_cleanup(container);
1268                 return -ENOMEM;
1269         }
1270
1271         return rc;
1272 }
1273
1274 void
1275 lnet_msg_containers_destroy(void)
1276 {
1277         struct lnet_msg_container *container;
1278         int     i;
1279
1280         if (the_lnet.ln_msg_containers == NULL)
1281                 return;
1282
1283         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1284                 lnet_msg_container_cleanup(container);
1285
1286         cfs_percpt_free(the_lnet.ln_msg_containers);
1287         the_lnet.ln_msg_containers = NULL;
1288 }
1289
1290 int
1291 lnet_msg_containers_create(void)
1292 {
1293         struct lnet_msg_container *container;
1294         int     rc;
1295         int     i;
1296
1297         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1298                                                       sizeof(*container));
1299
1300         if (the_lnet.ln_msg_containers == NULL) {
1301                 CERROR("Failed to allocate cpu-partition data for network\n");
1302                 return -ENOMEM;
1303         }
1304
1305         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1306                 rc = lnet_msg_container_setup(container, i);
1307                 if (rc != 0) {
1308                         lnet_msg_containers_destroy();
1309                         return rc;
1310                 }
1311         }
1312
1313         return 0;
1314 }