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    = le64_to_cpu(hdr->dest_nid);
  72                 ev->target.pid    = le32_to_cpu(hdr->dest_pid);
  73                 ev->initiator.nid = LNET_NID_ANY;
  74                 ev->initiator.pid = the_lnet.ln_pid;
  75                 ev->source.nid    = LNET_NID_ANY;
  76                 ev->source.pid    = the_lnet.ln_pid;
  77                 ev->sender        = LNET_NID_ANY;
  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, msg->msg_from);
 394
 395                 lnet_net_lock(cpt);
 396                 /*
 397                  * NB: message is committed for sending, we should return
 398                  * on success because LND will finalize this message later.
 399                  *
 400                  * Also, there is possibility that message is committed for
 401                  * sending and also failed before delivering to LND,
 402                  * i.e: ENOMEM, in that case we can't fall through either
 403                  * because CPT for sending can be different with CPT for
 404                  * receiving, so we should return back to lnet_finalize()
 405                  * to make sure we are locking the correct partition.
 406                  */
 407                 return rc;
 408
 409         } else if (status == 0 &&       /* OK so far */
 410                    (msg->msg_routing && !msg->msg_sending)) {
 411                 /* not forwarded */
 412                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 413                 lnet_net_unlock(cpt);
 414
 415                 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
 416
 417                 lnet_net_lock(cpt);
 418                 /*
 419                  * NB: message is committed for sending, we should return
 420                  * on success because LND will finalize this message later.
 421                  *
 422                  * Also, there is possibility that message is committed for
 423                  * sending and also failed before delivering to LND,
 424                  * i.e: ENOMEM, in that case we can't fall through either:
 425                  * - The rule is message must decommit for sending first if
 426                  *   the it's committed for both sending and receiving
 427                  * - CPT for sending can be different with CPT for receiving,
 428                  *   so we should return back to lnet_finalize() to make
 429                  *   sure we are locking the correct partition.
 430                  */
 431                 return rc;
 432         }
 433
 434         lnet_msg_decommit(msg, cpt, status);
 435         lnet_msg_free(msg);
 436         return 0;
 437 }
 438
 439 static void
 440 lnet_dec_healthv_locked(atomic_t *healthv, int sensitivity)
 441 {
 442         int h = atomic_read(healthv);
 443
 444         if (h < sensitivity) {
 445                 atomic_set(healthv, 0);
 446         } else {
 447                 h -= sensitivity;
 448                 atomic_set(healthv, h);
 449         }
 450 }
 451
 452 static void
 453 lnet_handle_local_failure(struct lnet_ni *local_ni)
 454 {
 455         /*
 456          * the lnet_net_lock(0) is used to protect the addref on the ni
 457          * and the recovery queue.
 458          */
 459         lnet_net_lock(0);
 460         /* the mt could've shutdown and cleaned up the queues */
 461         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 462                 lnet_net_unlock(0);
 463                 return;
 464         }
 465
 466         lnet_dec_healthv_locked(&local_ni->ni_healthv, lnet_health_sensitivity);
 467         /*
 468          * add the NI to the recovery queue if it's not already there
 469          * and it's health value is actually below the maximum. It's
 470          * possible that the sensitivity might be set to 0, and the health
 471          * value will not be reduced. In this case, there is no reason to
 472          * invoke recovery
 473          */
 474         if (list_empty(&local_ni->ni_recovery) &&
 475             atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
 476                 CDEBUG(D_NET, "ni %s added to recovery queue. Health = %d\n",
 477                         libcfs_nid2str(local_ni->ni_nid),
 478                         atomic_read(&local_ni->ni_healthv));
 479                 list_add_tail(&local_ni->ni_recovery,
 480                               &the_lnet.ln_mt_localNIRecovq);
 481                 lnet_ni_addref_locked(local_ni, 0);
 482         }
 483         lnet_net_unlock(0);
 484 }
 485
 486 /* must hold net_lock/0 */
 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          * If there is a health sensitivity in the peer then use that
 495          * instead of the globally set one.
 496          */
 497         lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
 498         if (lp_sensitivity)
 499                 sensitivity = lp_sensitivity;
 500
 501         lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
 502
 503         /* update the peer_net's health value */
 504         lnet_update_peer_net_healthv(lpni);
 505
 506         /*
 507          * add the peer NI to the recovery queue if it's not already there
 508          * and it's health value is actually below the maximum. It's
 509          * possible that the sensitivity might be set to 0, and the health
 510          * value will not be reduced. In this case, there is no reason to
 511          * invoke recovery
 512          */
 513         lnet_peer_ni_add_to_recoveryq_locked(lpni,
 514                                              &the_lnet.ln_mt_peerNIRecovq,
 515                                              ktime_get_seconds());
 516 }
 517
 518 static void
 519 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 520 {
 521         /* lpni could be NULL if we're in the LOLND case */
 522         if (!lpni)
 523                 return;
 524
 525         lnet_net_lock(0);
 526         /* the mt could've shutdown and cleaned up the queues */
 527         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 528                 lnet_net_unlock(0);
 529                 return;
 530         }
 531         lnet_handle_remote_failure_locked(lpni);
 532         lnet_net_unlock(0);
 533 }
 534
 535 static void
 536 lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 537                  enum lnet_msg_hstatus hstatus)
 538 {
 539         struct lnet_counters_health *health;
 540
 541         health = &the_lnet.ln_counters[0]->lct_health;
 542
 543         switch (hstatus) {
 544         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 545                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 546                 health->lch_local_interrupt_count++;
 547                 break;
 548         case LNET_MSG_STATUS_LOCAL_DROPPED:
 549                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 550                 health->lch_local_dropped_count++;
 551                 break;
 552         case LNET_MSG_STATUS_LOCAL_ABORTED:
 553                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 554                 health->lch_local_aborted_count++;
 555                 break;
 556         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 557                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 558                 health->lch_local_no_route_count++;
 559                 break;
 560         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 561                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 562                 health->lch_local_timeout_count++;
 563                 break;
 564         case LNET_MSG_STATUS_LOCAL_ERROR:
 565                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 566                 health->lch_local_error_count++;
 567                 break;
 568         case LNET_MSG_STATUS_REMOTE_DROPPED:
 569                 if (lpni)
 570                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 571                 health->lch_remote_dropped_count++;
 572                 break;
 573         case LNET_MSG_STATUS_REMOTE_ERROR:
 574                 if (lpni)
 575                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 576                 health->lch_remote_error_count++;
 577                 break;
 578         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 579                 if (lpni)
 580                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 581                 health->lch_remote_timeout_count++;
 582                 break;
 583         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 584                 if (lpni)
 585                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 586                 health->lch_network_timeout_count++;
 587                 break;
 588         case LNET_MSG_STATUS_OK:
 589                 break;
 590         default:
 591                 LBUG();
 592         }
 593 }
 594
 595 static void
 596 lnet_resend_msg_locked(struct lnet_msg *msg)
 597 {
 598         msg->msg_retry_count++;
 599
 600         /*
 601          * remove message from the active list and reset it to prepare
 602          * for a resend. Two exceptions to this
 603          *
 604          * 1. the router case. When a message is being routed it is
 605          * committed for rx when received and committed for tx when
 606          * forwarded. We don't want to remove it from the active list, since
 607          * code which handles receiving expects it to remain on the active
 608          * list.
 609          *
 610          * 2. The REPLY case. Reply messages use the same message
 611          * structure for the GET that was received.
 612          */
 613         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 614                 list_del_init(&msg->msg_activelist);
 615                 msg->msg_onactivelist = 0;
 616         }
 617         /*
 618          * The msg_target.nid which was originally set
 619          * when calling LNetGet() or LNetPut() might've
 620          * been overwritten if we're routing this message.
 621          * Call lnet_msg_decommit_tx() to return the credit
 622          * this message consumed. The message will
 623          * consume another credit when it gets resent.
 624          */
 625         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 626         lnet_msg_decommit_tx(msg, -EAGAIN);
 627         msg->msg_sending = 0;
 628         msg->msg_receiving = 0;
 629         msg->msg_target_is_router = 0;
 630
 631         CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
 632                libcfs_nid2str(msg->msg_hdr.src_nid),
 633                libcfs_nid2str(msg->msg_hdr.dest_nid),
 634                lnet_msgtyp2str(msg->msg_type),
 635                lnet_health_error2str(msg->msg_health_status), msg);
 636
 637         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 638
 639         complete(&the_lnet.ln_mt_wait_complete);
 640 }
 641
 642 int
 643 lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
 644                                      struct list_head *containerq,
 645                                      int nworkers, void **workers)
 646 {
 647         int my_slot = -1;
 648         int i;
 649
 650         list_add_tail(&msg->msg_list, containerq);
 651
 652         for (i = 0; i < nworkers; i++) {
 653                 if (workers[i] == current)
 654                         break;
 655
 656                 if (my_slot < 0 && workers[i] == NULL)
 657                         my_slot = i;
 658         }
 659
 660         if (i < nworkers || my_slot < 0)
 661                 return -1;
 662
 663         workers[my_slot] = current;
 664
 665         return my_slot;
 666 }
 667
 668 int
 669 lnet_attempt_msg_resend(struct lnet_msg *msg)
 670 {
 671         struct lnet_msg_container *container;
 672         int my_slot;
 673         int cpt;
 674
 675         /* we can only resend tx_committed messages */
 676         LASSERT(msg->msg_tx_committed);
 677
 678         /* don't resend recovery messages */
 679         if (msg->msg_recovery) {
 680                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 681                         libcfs_nid2str(msg->msg_from),
 682                         libcfs_nid2str(msg->msg_target.nid),
 683                         msg->msg_retry_count);
 684                 return -ENOTRECOVERABLE;
 685         }
 686
 687         /*
 688          * if we explicitly indicated we don't want to resend then just
 689          * return
 690          */
 691         if (msg->msg_no_resend) {
 692                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 693                         libcfs_nid2str(msg->msg_from),
 694                         libcfs_nid2str(msg->msg_target.nid),
 695                         msg->msg_retry_count);
 696                 return -ENOTRECOVERABLE;
 697         }
 698
 699         /* check if the message has exceeded the number of retries */
 700         if (msg->msg_retry_count >= lnet_retry_count) {
 701                 CNETERR("msg %s->%s exceeded retry count %d\n",
 702                         libcfs_nid2str(msg->msg_from),
 703                         libcfs_nid2str(msg->msg_target.nid),
 704                         msg->msg_retry_count);
 705                 return -ENOTRECOVERABLE;
 706         }
 707
 708         cpt = msg->msg_tx_cpt;
 709         lnet_net_lock(cpt);
 710
 711         /* check again under lock */
 712         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 713                 lnet_net_unlock(cpt);
 714                 return -ESHUTDOWN;
 715         }
 716
 717         container = the_lnet.ln_msg_containers[cpt];
 718         my_slot =
 719                 lnet_check_finalize_recursion_locked(msg,
 720                                         &container->msc_resending,
 721                                         container->msc_nfinalizers,
 722                                         container->msc_resenders);
 723
 724         /* enough threads are resending */
 725         if (my_slot == -1) {
 726                 lnet_net_unlock(cpt);
 727                 return 0;
 728         }
 729
 730         while (!list_empty(&container->msc_resending)) {
 731                 msg = list_entry(container->msc_resending.next,
 732                                         struct lnet_msg, msg_list);
 733                 list_del(&msg->msg_list);
 734
 735                 /*
 736                  * resending the message will require us to call
 737                  * lnet_msg_decommit_tx() which will return the credit
 738                  * which this message holds. This could trigger another
 739                  * queued message to be sent. If that message fails and
 740                  * requires a resend we will recurse.
 741                  * But since at this point the slot is taken, the message
 742                  * will be queued in the container and dealt with
 743                  * later. This breaks the recursion.
 744                  */
 745                 lnet_resend_msg_locked(msg);
 746         }
 747
 748         /*
 749          * msc_resenders is an array of process pointers. Each entry holds
 750          * a pointer to the current process operating on the message. An
 751          * array entry is created per CPT. If the array slot is already
 752          * set, then it means that there is a thread on the CPT currently
 753          * resending a message.
 754          * Once the thread finishes clear the slot to enable the thread to
 755          * take on more resend work.
 756          */
 757         container->msc_resenders[my_slot] = NULL;
 758         lnet_net_unlock(cpt);
 759
 760         return 0;
 761 }
 762
 763 /*
 764  * Do a health check on the message:
 765  * return -1 if we're not going to handle the error or
 766  *   if we've reached the maximum number of retries.
 767  *   success case will return -1 as well
 768  * return 0 if it the message is requeued for send
 769  */
 770 static int
 771 lnet_health_check(struct lnet_msg *msg)
 772 {
 773         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 774         struct lnet_peer_ni *lpni;
 775         struct lnet_ni *ni;
 776         bool lo = false;
 777         bool attempt_local_resend;
 778         bool attempt_remote_resend;
 779         bool handle_local_health;
 780         bool handle_remote_health;
 781
 782         /* if we're shutting down no point in handling health. */
 783         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 784                 return -1;
 785
 786         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 787
 788         /*
 789          * if we're sending to the LOLND then the msg_txpeer will not be
 790          * set. So no need to sanity check it.
 791          */
 792         if (msg->msg_tx_committed && msg->msg_txni->ni_nid != LNET_NID_LO_0)
 793                 LASSERT(msg->msg_txpeer);
 794         else if (msg->msg_tx_committed &&
 795                  msg->msg_txni->ni_nid == LNET_NID_LO_0)
 796                 lo = true;
 797
 798         if (hstatus != LNET_MSG_STATUS_OK &&
 799             ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
 800                 return -1;
 801
 802         /*
 803          * always prefer txni/txpeer if they message is committed for both
 804          * directions.
 805          */
 806         if (msg->msg_tx_committed) {
 807                 ni = msg->msg_txni;
 808                 lpni = msg->msg_txpeer;
 809                 attempt_local_resend = attempt_remote_resend = true;
 810         } else {
 811                 ni = msg->msg_rxni;
 812                 lpni = msg->msg_rxpeer;
 813                 attempt_local_resend = attempt_remote_resend = false;
 814         }
 815
 816         /* Don't further decrement the health value if a recovery message
 817          * failed.
 818          */
 819         if (msg->msg_recovery)
 820                 handle_local_health = handle_remote_health = false;
 821         else
 822                 handle_local_health = handle_remote_health = true;
 823
 824         /* For local failures, health/recovery/resends are not needed if I only
 825          * have a single (non-lolnd) interface. NB: pb_nnis includes the lolnd
 826          * interface, so a single-rail node would have pb_nnis == 2.
 827          */
 828         if (the_lnet.ln_ping_target->pb_nnis <= 2) {
 829                 handle_local_health = false;
 830                 attempt_local_resend = false;
 831         }
 832
 833         /* For remote failures, health/recovery/resends are not needed if the
 834          * peer only has a single interface. Special case for routers where we
 835          * rely on health feature to manage route aliveness. NB: unlike pb_nnis
 836          * above, lp_nnis does _not_ include the lolnd, so a single-rail node
 837          * would have lp_nnis == 1.
 838          */
 839         if (lpni && lpni->lpni_peer_net->lpn_peer->lp_nnis <= 1) {
 840                 attempt_remote_resend = false;
 841                 if (!lnet_isrouter(lpni))
 842                         handle_remote_health = false;
 843         }
 844
 845         if (!lo)
 846                 LASSERT(ni && lpni);
 847         else
 848                 LASSERT(ni);
 849
 850         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 851                libcfs_nid2str(ni->ni_nid),
 852                (lo) ? "self" : libcfs_nid2str(lpni->lpni_nid),
 853                lnet_msgtyp2str(msg->msg_type),
 854                lnet_health_error2str(hstatus));
 855
 856         /*
 857          * stats are only incremented for errors so avoid wasting time
 858          * incrementing statistics if there is no error.
 859          */
 860         if (hstatus != LNET_MSG_STATUS_OK) {
 861                 lnet_net_lock(0);
 862                 lnet_incr_hstats(ni, lpni, hstatus);
 863                 lnet_net_unlock(0);
 864         }
 865
 866         switch (hstatus) {
 867         case LNET_MSG_STATUS_OK:
 868                 /*
 869                  * increment the local ni health whether we successfully
 870                  * received or sent a message on it.
 871                  *
 872                  * Ping counts are reset to 0 as appropriate to allow for
 873                  * faster recovery.
 874                  */
 875                 lnet_inc_healthv(&ni->ni_healthv, lnet_health_sensitivity);
 876                 /*
 877                  * It's possible msg_txpeer is NULL in the LOLND
 878                  * case. Only increment the peer's health if we're
 879                  * receiving a message from it. It's the only sure way to
 880                  * know that a remote interface is up.
 881                  * If this interface is part of a router, then take that
 882                  * as indication that the router is fully healthy.
 883                  */
 884                 if (lpni && msg->msg_rx_committed) {
 885                         lpni->lpni_ping_count = 0;
 886                         /*
 887                          * If we're receiving a message from the router or
 888                          * I'm a router, then set that lpni's health to
 889                          * maximum so we can commence communication
 890                          */
 891                         lnet_net_lock(0);
 892                         if (lnet_isrouter(lpni) || the_lnet.ln_routing) {
 893                                 lnet_set_lpni_healthv_locked(lpni,
 894                                         LNET_MAX_HEALTH_VALUE);
 895                         } else {
 896                                 __u32 sensitivity = lpni->lpni_peer_net->
 897                                         lpn_peer->lp_health_sensitivity;
 898
 899                                 lnet_inc_lpni_healthv_locked(lpni,
 900                                         (sensitivity) ? sensitivity :
 901                                         lnet_health_sensitivity);
 902                                 /* This peer NI may have previously aged out
 903                                  * of recovery. Now that we've received a
 904                                  * message from it, we can continue recovery
 905                                  * if its health value is still below the
 906                                  * maximum.
 907                                  */
 908                                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
 909                                                 &the_lnet.ln_mt_peerNIRecovq,
 910                                                 ktime_get_seconds());
 911                         }
 912                         lnet_net_unlock(0);
 913                 }
 914
 915                 /* we can finalize this message */
 916                 return -1;
 917         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 918         case LNET_MSG_STATUS_LOCAL_DROPPED:
 919         case LNET_MSG_STATUS_LOCAL_ABORTED:
 920         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 921         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 922                 if (handle_local_health)
 923                         lnet_handle_local_failure(ni);
 924                 if (attempt_local_resend)
 925                         return lnet_attempt_msg_resend(msg);
 926                 break;
 927         case LNET_MSG_STATUS_LOCAL_ERROR:
 928                 if (handle_local_health)
 929                         lnet_handle_local_failure(ni);
 930                 return -1;
 931         case LNET_MSG_STATUS_REMOTE_DROPPED:
 932                 if (handle_remote_health)
 933                         lnet_handle_remote_failure(lpni);
 934                 if (attempt_remote_resend)
 935                         return lnet_attempt_msg_resend(msg);
 936                 break;
 937         case LNET_MSG_STATUS_REMOTE_ERROR:
 938         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 939                 if (handle_remote_health)
 940                         lnet_handle_remote_failure(lpni);
 941                 return -1;
 942         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 943                 if (handle_remote_health)
 944                         lnet_handle_remote_failure(lpni);
 945                 if (handle_local_health)
 946                         lnet_handle_local_failure(ni);
 947                 return -1;
 948         default:
 949                 LBUG();
 950         }
 951
 952         /* no resend is needed */
 953         return -1;
 954 }
 955
 956 static void
 957 lnet_msg_detach_md(struct lnet_msg *msg, int status)
 958 {
 959         struct lnet_libmd *md = msg->msg_md;
 960         lnet_handler_t handler = NULL;
 961         int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
 962         int unlink;
 963
 964         lnet_res_lock(cpt);
 965         while (md->md_flags & LNET_MD_FLAG_HANDLING)
 966                 /* An event handler is running - wait for it to
 967                  * complete to avoid races.
 968                  */
 969                 lnet_md_wait_handling(md, cpt);
 970
 971         /* Now it's safe to drop my caller's ref */
 972         md->md_refcount--;
 973         LASSERT(md->md_refcount >= 0);
 974
 975         unlink = lnet_md_unlinkable(md);
 976         if (md->md_handler) {
 977                 if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
 978                         msg->msg_ev.status   = -ETIMEDOUT;
 979                         CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
 980                 } else {
 981                         msg->msg_ev.status   = status;
 982                 }
 983                 msg->msg_ev.unlinked = unlink;
 984                 handler = md->md_handler;
 985                 if (!unlink)
 986                         md->md_flags |= LNET_MD_FLAG_HANDLING;
 987         }
 988
 989         if (unlink || (md->md_refcount == 0 &&
 990                        md->md_threshold == LNET_MD_THRESH_INF))
 991                 lnet_detach_rsp_tracker(md, cpt);
 992
 993         msg->msg_md = NULL;
 994         if (unlink)
 995                 lnet_md_unlink(md);
 996
 997         lnet_res_unlock(cpt);
 998
 999         if (handler) {
1000                 handler(&msg->msg_ev);
1001                 if (!unlink) {
1002                         lnet_res_lock(cpt);
1003                         md->md_flags &= ~LNET_MD_FLAG_HANDLING;
1004                         wake_up_var(md);
1005                         lnet_res_unlock(cpt);
1006                 }
1007         }
1008 }
1009
1010 static bool
1011 lnet_is_health_check(struct lnet_msg *msg)
1012 {
1013         bool hc = true;
1014         int status = msg->msg_ev.status;
1015
1016         if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
1017             !msg->msg_onactivelist) {
1018                 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
1019                        msg);
1020                 return false;
1021         }
1022
1023         if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
1024             (msg->msg_rx_committed && !msg->msg_rxpeer)) {
1025                 /* The optimized GET case does not set msg_rxpeer, but status
1026                  * could be zero. Only print the error message if we have a
1027                  * non-zero status.
1028                  */
1029                 if (status)
1030                         CDEBUG(D_NET, "msg %p status %d cannot retry\n", msg,
1031                                status);
1032                 return false;
1033         }
1034
1035         /* Check for status inconsistencies */
1036         if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
1037              (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
1038                 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
1039                        "checking (%d, %d)\n", msg, status,
1040                        msg->msg_health_status);
1041                 hc = false;
1042         }
1043
1044         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
1045                hc, status, msg->msg_health_status);
1046
1047         return hc;
1048 }
1049
1050 char *
1051 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
1052 {
1053         switch (hstatus) {
1054         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
1055                 return "LOCAL_INTERRUPT";
1056         case LNET_MSG_STATUS_LOCAL_DROPPED:
1057                 return "LOCAL_DROPPED";
1058         case LNET_MSG_STATUS_LOCAL_ABORTED:
1059                 return "LOCAL_ABORTED";
1060         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
1061                 return "LOCAL_NO_ROUTE";
1062         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
1063                 return "LOCAL_TIMEOUT";
1064         case LNET_MSG_STATUS_LOCAL_ERROR:
1065                 return "LOCAL_ERROR";
1066         case LNET_MSG_STATUS_REMOTE_DROPPED:
1067                 return "REMOTE_DROPPED";
1068         case LNET_MSG_STATUS_REMOTE_ERROR:
1069                 return "REMOTE_ERROR";
1070         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
1071                 return "REMOTE_TIMEOUT";
1072         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
1073                 return "NETWORK_TIMEOUT";
1074         case LNET_MSG_STATUS_OK:
1075                 return "OK";
1076         default:
1077                 return "<UNKNOWN>";
1078         }
1079 }
1080
1081 bool
1082 lnet_send_error_simulation(struct lnet_msg *msg,
1083                            enum lnet_msg_hstatus *hstatus)
1084 {
1085         if (!msg)
1086                 return false;
1087
1088         if (list_empty(&the_lnet.ln_drop_rules))
1089             return false;
1090
1091         /* match only health rules */
1092         if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
1093                                   hstatus))
1094                 return false;
1095
1096         CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
1097                 libcfs_nid2str(msg->msg_hdr.src_nid),
1098                 libcfs_nid2str(msg->msg_txni->ni_nid),
1099                 libcfs_nid2str(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 (!list_empty(&container->msc_finalizing)) {
1179                 msg = list_entry(container->msc_finalizing.next,
1180                                  struct lnet_msg, msg_list);
1181
1182                 list_del_init(&msg->msg_list);
1183
1184                 /* NB drops and regains the lnet lock if it actually does
1185                  * anything, so my finalizing friends can chomp along too */
1186                 rc = lnet_complete_msg_locked(msg, cpt);
1187                 if (rc != 0)
1188                         break;
1189         }
1190
1191         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1192                 lnet_net_unlock(cpt);
1193                 lnet_delay_rule_check();
1194                 lnet_net_lock(cpt);
1195         }
1196
1197         container->msc_finalizers[my_slot] = NULL;
1198         lnet_net_unlock(cpt);
1199
1200         if (rc != 0)
1201                 goto again;
1202 }
1203 EXPORT_SYMBOL(lnet_finalize);
1204
1205 void
1206 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1207 {
1208         int     count = 0;
1209
1210         if (container->msc_init == 0)
1211                 return;
1212
1213         while (!list_empty(&container->msc_active)) {
1214                 struct lnet_msg *msg;
1215
1216                 msg  = list_entry(container->msc_active.next,
1217                                   struct lnet_msg, msg_activelist);
1218                 LASSERT(msg->msg_onactivelist);
1219                 msg->msg_onactivelist = 0;
1220                 list_del_init(&msg->msg_activelist);
1221                 lnet_msg_free(msg);
1222                 count++;
1223         }
1224
1225         if (count > 0)
1226                 CERROR("%d active msg on exit\n", count);
1227
1228         if (container->msc_finalizers != NULL) {
1229                 CFS_FREE_PTR_ARRAY(container->msc_finalizers,
1230                                    container->msc_nfinalizers);
1231                 container->msc_finalizers = NULL;
1232         }
1233
1234         if (container->msc_resenders != NULL) {
1235                 CFS_FREE_PTR_ARRAY(container->msc_resenders,
1236                                    container->msc_nfinalizers);
1237                 container->msc_resenders = NULL;
1238         }
1239         container->msc_init = 0;
1240 }
1241
1242 int
1243 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1244 {
1245         int rc = 0;
1246
1247         container->msc_init = 1;
1248
1249         INIT_LIST_HEAD(&container->msc_active);
1250         INIT_LIST_HEAD(&container->msc_finalizing);
1251         INIT_LIST_HEAD(&container->msc_resending);
1252
1253         /* number of CPUs */
1254         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1255         if (container->msc_nfinalizers == 0)
1256                 container->msc_nfinalizers = 1;
1257
1258         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1259                          container->msc_nfinalizers *
1260                          sizeof(*container->msc_finalizers));
1261
1262         if (container->msc_finalizers == NULL) {
1263                 CERROR("Failed to allocate message finalizers\n");
1264                 lnet_msg_container_cleanup(container);
1265                 return -ENOMEM;
1266         }
1267
1268         LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
1269                          container->msc_nfinalizers *
1270                          sizeof(*container->msc_resenders));
1271
1272         if (container->msc_resenders == NULL) {
1273                 CERROR("Failed to allocate message resenders\n");
1274                 lnet_msg_container_cleanup(container);
1275                 return -ENOMEM;
1276         }
1277
1278         return rc;
1279 }
1280
1281 void
1282 lnet_msg_containers_destroy(void)
1283 {
1284         struct lnet_msg_container *container;
1285         int     i;
1286
1287         if (the_lnet.ln_msg_containers == NULL)
1288                 return;
1289
1290         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1291                 lnet_msg_container_cleanup(container);
1292
1293         cfs_percpt_free(the_lnet.ln_msg_containers);
1294         the_lnet.ln_msg_containers = NULL;
1295 }
1296
1297 int
1298 lnet_msg_containers_create(void)
1299 {
1300         struct lnet_msg_container *container;
1301         int     rc;
1302         int     i;
1303
1304         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1305                                                       sizeof(*container));
1306
1307         if (the_lnet.ln_msg_containers == NULL) {
1308                 CERROR("Failed to allocate cpu-partition data for network\n");
1309                 return -ENOMEM;
1310         }
1311
1312         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1313                 rc = lnet_msg_container_setup(container, i);
1314                 if (rc != 0) {
1315                         lnet_msg_containers_destroy();
1316                         return rc;
1317                 }
1318         }
1319
1320         return 0;
1321 }