lnet/lnet/lib-msg.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lnet/lnet/lib-msg.c
  32  *
  33  * Message decoding, parsing and finalizing routines
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_LNET
  37
  38 #include <lnet/lib-lnet.h>
  39
  40 void
  41 lnet_build_unlink_event(struct lnet_libmd *md, struct lnet_event *ev)
  42 {
  43         ENTRY;
  44
  45         memset(ev, 0, sizeof(*ev));
  46
  47         ev->status   = 0;
  48         ev->unlinked = 1;
  49         ev->type     = LNET_EVENT_UNLINK;
  50         lnet_md_deconstruct(md, ev);
  51         lnet_md2handle(&ev->md_handle, md);
  52         EXIT;
  53 }
  54
  55 /*
  56  * Don't need any lock, must be called after lnet_commit_md
  57  */
  58 void
  59 lnet_build_msg_event(struct lnet_msg *msg, enum lnet_event_kind ev_type)
  60 {
  61         struct lnet_hdr *hdr = &msg->msg_hdr;
  62         struct lnet_event *ev = &msg->msg_ev;
  63
  64         LASSERT(!msg->msg_routing);
  65
  66         ev->type = ev_type;
  67         ev->msg_type = msg->msg_type;
  68
  69         if (ev_type == LNET_EVENT_SEND) {
  70                 /* event for active message */
  71                 ev->target.nid    = hdr->dest_nid;
  72                 ev->target.pid    = hdr->dest_pid;
  73                 ev->initiator.nid = LNET_ANY_NID;
  74                 ev->initiator.pid = the_lnet.ln_pid;
  75                 ev->source.nid    = LNET_ANY_NID;
  76                 ev->source.pid    = the_lnet.ln_pid;
  77                 ev->sender        = LNET_ANY_NID;
  78         } else {
  79                 /* event for passive message */
  80                 ev->target.pid    = hdr->dest_pid;
  81                 ev->target.nid    = hdr->dest_nid;
  82                 ev->initiator.pid = hdr->src_pid;
  83                 /* Multi-Rail: resolve src_nid to "primary" peer NID */
  84                 ev->initiator.nid = msg->msg_initiator;
  85                 /* Multi-Rail: track source NID. */
  86                 ev->source.pid    = hdr->src_pid;
  87                 ev->source.nid    = hdr->src_nid;
  88                 ev->rlength       = hdr->payload_length;
  89                 ev->sender        = msg->msg_from;
  90                 ev->mlength       = msg->msg_wanted;
  91                 ev->offset        = msg->msg_offset;
  92         }
  93
  94         switch (ev_type) {
  95         default:
  96                 LBUG();
  97
  98         case LNET_EVENT_PUT: /* passive PUT */
  99                 ev->pt_index   = hdr->msg.put.ptl_index;
 100                 ev->match_bits = hdr->msg.put.match_bits;
 101                 ev->hdr_data   = hdr->msg.put.hdr_data;
 102                 return;
 103
 104         case LNET_EVENT_GET: /* passive GET */
 105                 ev->pt_index   = hdr->msg.get.ptl_index;
 106                 ev->match_bits = hdr->msg.get.match_bits;
 107                 ev->hdr_data   = 0;
 108                 return;
 109
 110         case LNET_EVENT_ACK: /* ACK */
 111                 ev->match_bits = hdr->msg.ack.match_bits;
 112                 ev->mlength    = hdr->msg.ack.mlength;
 113                 return;
 114
 115         case LNET_EVENT_REPLY: /* REPLY */
 116                 return;
 117
 118         case LNET_EVENT_SEND: /* active message */
 119                 if (msg->msg_type == LNET_MSG_PUT) {
 120                         ev->pt_index   = le32_to_cpu(hdr->msg.put.ptl_index);
 121                         ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
 122                         ev->offset     = le32_to_cpu(hdr->msg.put.offset);
 123                         ev->mlength    =
 124                         ev->rlength    = le32_to_cpu(hdr->payload_length);
 125                         ev->hdr_data   = le64_to_cpu(hdr->msg.put.hdr_data);
 126
 127                 } else {
 128                         LASSERT(msg->msg_type == LNET_MSG_GET);
 129                         ev->pt_index   = le32_to_cpu(hdr->msg.get.ptl_index);
 130                         ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
 131                         ev->mlength    =
 132                         ev->rlength    = le32_to_cpu(hdr->msg.get.sink_length);
 133                         ev->offset     = le32_to_cpu(hdr->msg.get.src_offset);
 134                         ev->hdr_data   = 0;
 135                 }
 136                 return;
 137         }
 138 }
 139
 140 void
 141 lnet_msg_commit(struct lnet_msg *msg, int cpt)
 142 {
 143         struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
 144         struct lnet_counters_common *common;
 145         s64 timeout_ns;
 146
 147         /* set the message deadline */
 148         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
 149         msg->msg_deadline = ktime_add_ns(ktime_get(), timeout_ns);
 150
 151         /* routed message can be committed for both receiving and sending */
 152         LASSERT(!msg->msg_tx_committed);
 153
 154         if (msg->msg_sending) {
 155                 LASSERT(!msg->msg_receiving);
 156                 msg->msg_tx_cpt = cpt;
 157                 msg->msg_tx_committed = 1;
 158                 if (msg->msg_rx_committed) { /* routed message REPLY */
 159                         LASSERT(msg->msg_onactivelist);
 160                         return;
 161                 }
 162         } else {
 163                 LASSERT(!msg->msg_sending);
 164                 msg->msg_rx_cpt = cpt;
 165                 msg->msg_rx_committed = 1;
 166         }
 167
 168         LASSERT(!msg->msg_onactivelist);
 169
 170         msg->msg_onactivelist = 1;
 171         list_add_tail(&msg->msg_activelist, &container->msc_active);
 172
 173         common = &the_lnet.ln_counters[cpt]->lct_common;
 174         common->lcc_msgs_alloc++;
 175         if (common->lcc_msgs_alloc > common->lcc_msgs_max)
 176                 common->lcc_msgs_max = common->lcc_msgs_alloc;
 177 }
 178
 179 static void
 180 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
 181 {
 182         struct lnet_counters_common *common;
 183         struct lnet_event *ev = &msg->msg_ev;
 184
 185         LASSERT(msg->msg_tx_committed);
 186         if (status != 0)
 187                 goto out;
 188
 189         common = &(the_lnet.ln_counters[msg->msg_tx_cpt]->lct_common);
 190         switch (ev->type) {
 191         default: /* routed message */
 192                 LASSERT(msg->msg_routing);
 193                 LASSERT(msg->msg_rx_committed);
 194                 LASSERT(ev->type == 0);
 195
 196                 common->lcc_route_length += msg->msg_len;
 197                 common->lcc_route_count++;
 198                 goto incr_stats;
 199
 200         case LNET_EVENT_PUT:
 201                 /* should have been decommitted */
 202                 LASSERT(!msg->msg_rx_committed);
 203                 /* overwritten while sending ACK */
 204                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 205                 msg->msg_type = LNET_MSG_PUT; /* fix type */
 206                 break;
 207
 208         case LNET_EVENT_SEND:
 209                 LASSERT(!msg->msg_rx_committed);
 210                 if (msg->msg_type == LNET_MSG_PUT)
 211                         common->lcc_send_length += msg->msg_len;
 212                 break;
 213
 214         case LNET_EVENT_GET:
 215                 LASSERT(msg->msg_rx_committed);
 216                 /* overwritten while sending reply, we should never be
 217                  * here for optimized GET */
 218                 LASSERT(msg->msg_type == LNET_MSG_REPLY);
 219                 msg->msg_type = LNET_MSG_GET; /* fix type */
 220                 break;
 221         }
 222
 223         common->lcc_send_count++;
 224
 225 incr_stats:
 226         if (msg->msg_txpeer)
 227                 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 228                                 msg->msg_type,
 229                                 LNET_STATS_TYPE_SEND);
 230         if (msg->msg_txni)
 231                 lnet_incr_stats(&msg->msg_txni->ni_stats,
 232                                 msg->msg_type,
 233                                 LNET_STATS_TYPE_SEND);
 234  out:
 235         lnet_return_tx_credits_locked(msg);
 236         msg->msg_tx_committed = 0;
 237 }
 238
 239 static void
 240 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
 241 {
 242         struct lnet_counters_common *common;
 243         struct lnet_event *ev = &msg->msg_ev;
 244
 245         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 246         LASSERT(msg->msg_rx_committed);
 247
 248         if (status != 0)
 249                 goto out;
 250
 251         common = &(the_lnet.ln_counters[msg->msg_rx_cpt]->lct_common);
 252         switch (ev->type) {
 253         default:
 254                 LASSERT(ev->type == 0);
 255                 LASSERT(msg->msg_routing);
 256                 goto incr_stats;
 257
 258         case LNET_EVENT_ACK:
 259                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 260                 break;
 261
 262         case LNET_EVENT_GET:
 263                 /* type is "REPLY" if it's an optimized GET on passive side,
 264                  * because optimized GET will never be committed for sending,
 265                  * so message type wouldn't be changed back to "GET" by
 266                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 267                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 268                         msg->msg_type == LNET_MSG_GET);
 269                 common->lcc_send_length += msg->msg_wanted;
 270                 break;
 271
 272         case LNET_EVENT_PUT:
 273                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 274                 break;
 275
 276         case LNET_EVENT_REPLY:
 277                 /* type is "GET" if it's an optimized GET on active side,
 278                  * see details in lnet_create_reply_msg() */
 279                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 280                         msg->msg_type == LNET_MSG_REPLY);
 281                 break;
 282         }
 283
 284         common->lcc_recv_count++;
 285
 286 incr_stats:
 287         if (msg->msg_rxpeer)
 288                 lnet_incr_stats(&msg->msg_rxpeer->lpni_stats,
 289                                 msg->msg_type,
 290                                 LNET_STATS_TYPE_RECV);
 291         if (msg->msg_rxni)
 292                 lnet_incr_stats(&msg->msg_rxni->ni_stats,
 293                                 msg->msg_type,
 294                                 LNET_STATS_TYPE_RECV);
 295         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 296                 common->lcc_recv_length += msg->msg_wanted;
 297
 298  out:
 299         lnet_return_rx_credits_locked(msg);
 300         msg->msg_rx_committed = 0;
 301 }
 302
 303 void
 304 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
 305 {
 306         int     cpt2 = cpt;
 307
 308         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 309         LASSERT(msg->msg_onactivelist);
 310
 311         if (msg->msg_tx_committed) { /* always decommit for sending first */
 312                 LASSERT(cpt == msg->msg_tx_cpt);
 313                 lnet_msg_decommit_tx(msg, status);
 314         }
 315
 316         if (msg->msg_rx_committed) {
 317                 /* forwarding msg committed for both receiving and sending */
 318                 if (cpt != msg->msg_rx_cpt) {
 319                         lnet_net_unlock(cpt);
 320                         cpt2 = msg->msg_rx_cpt;
 321                         lnet_net_lock(cpt2);
 322                 }
 323                 lnet_msg_decommit_rx(msg, status);
 324         }
 325
 326         list_del(&msg->msg_activelist);
 327         msg->msg_onactivelist = 0;
 328
 329         the_lnet.ln_counters[cpt2]->lct_common.lcc_msgs_alloc--;
 330
 331         if (cpt2 != cpt) {
 332                 lnet_net_unlock(cpt2);
 333                 lnet_net_lock(cpt);
 334         }
 335 }
 336
 337 void
 338 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
 339                    unsigned int offset, unsigned int mlen)
 340 {
 341         /* NB: @offset and @len are only useful for receiving */
 342         /* Here, we attach the MD on lnet_msg and mark it busy and
 343          * decrementing its threshold. Come what may, the lnet_msg "owns"
 344          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 345          * signals completion. */
 346         LASSERT(!msg->msg_routing);
 347
 348         msg->msg_md = md;
 349         if (msg->msg_receiving) { /* committed for receiving */
 350                 msg->msg_offset = offset;
 351                 msg->msg_wanted = mlen;
 352         }
 353
 354         md->md_refcount++;
 355         if (md->md_threshold != LNET_MD_THRESH_INF) {
 356                 LASSERT(md->md_threshold > 0);
 357                 md->md_threshold--;
 358         }
 359
 360         /* build umd in event */
 361         lnet_md2handle(&msg->msg_ev.md_handle, md);
 362         lnet_md_deconstruct(md, &msg->msg_ev);
 363 }
 364
 365 static int
 366 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
 367 {
 368         struct lnet_handle_wire ack_wmd;
 369         int                rc;
 370         int                status = msg->msg_ev.status;
 371
 372         LASSERT(msg->msg_onactivelist);
 373
 374         if (status == 0 && msg->msg_ack) {
 375                 /* Only send an ACK if the PUT completed successfully */
 376
 377                 lnet_msg_decommit(msg, cpt, 0);
 378
 379                 msg->msg_ack = 0;
 380                 lnet_net_unlock(cpt);
 381
 382                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 383                 LASSERT(!msg->msg_routing);
 384
 385                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 386
 387                 lnet_prep_send(msg, LNET_MSG_ACK,
 388                                lnet_pid_to_pid4(&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,
 395                                &msg->msg_from);
 396
 397                 lnet_net_lock(cpt);
 398                 /*
 399                  * NB: message is committed for sending, we should return
 400                  * on success because LND will finalize this message later.
 401                  *
 402                  * Also, there is possibility that message is committed for
 403                  * sending and also failed before delivering to LND,
 404                  * i.e: ENOMEM, in that case we can't fall through either
 405                  * because CPT for sending can be different with CPT for
 406                  * receiving, so we should return back to lnet_finalize()
 407                  * to make sure we are locking the correct partition.
 408                  */
 409                 return rc;
 410
 411         } else if (status == 0 &&       /* OK so far */
 412                    (msg->msg_routing && !msg->msg_sending)) {
 413                 /* not forwarded */
 414                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 415                 lnet_net_unlock(cpt);
 416
 417                 rc = lnet_send(NULL, msg, NULL);
 418
 419                 lnet_net_lock(cpt);
 420                 /*
 421                  * NB: message is committed for sending, we should return
 422                  * on success because LND will finalize this message later.
 423                  *
 424                  * Also, there is possibility that message is committed for
 425                  * sending and also failed before delivering to LND,
 426                  * i.e: ENOMEM, in that case we can't fall through either:
 427                  * - The rule is message must decommit for sending first if
 428                  *   the it's committed for both sending and receiving
 429                  * - CPT for sending can be different with CPT for receiving,
 430                  *   so we should return back to lnet_finalize() to make
 431                  *   sure we are locking the correct partition.
 432                  */
 433                 return rc;
 434         }
 435
 436         lnet_msg_decommit(msg, cpt, status);
 437         lnet_msg_free(msg);
 438         return 0;
 439 }
 440
 441 static void
 442 lnet_dec_healthv_locked(atomic_t *healthv, int sensitivity)
 443 {
 444         int h = atomic_read(healthv);
 445
 446         if (h < sensitivity) {
 447                 atomic_set(healthv, 0);
 448         } else {
 449                 h -= sensitivity;
 450                 atomic_set(healthv, h);
 451         }
 452 }
 453
 454 /* must hold net_lock/0 */
 455 void
 456 lnet_ni_add_to_recoveryq_locked(struct lnet_ni *ni,
 457                                 struct list_head *recovery_queue, time64_t now)
 458 {
 459         if (!list_empty(&ni->ni_recovery))
 460                 return;
 461
 462         if (atomic_read(&ni->ni_healthv) == LNET_MAX_HEALTH_VALUE)
 463                 return;
 464
 465         /* This NI is going on the recovery queue, so take a ref on it */
 466         lnet_ni_addref_locked(ni, 0);
 467
 468         lnet_ni_set_next_ping(ni, now);
 469
 470         CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld health :%d\n",
 471                libcfs_nidstr(&ni->ni_nid),
 472                ni->ni_ping_count,
 473                ni->ni_next_ping,
 474                atomic_read(&ni->ni_healthv));
 475
 476         list_add_tail(&ni->ni_recovery, recovery_queue);
 477 }
 478
 479 static void
 480 lnet_handle_local_failure(struct lnet_ni *local_ni)
 481 {
 482         /*
 483          * the lnet_net_lock(0) is used to protect the addref on the ni
 484          * and the recovery queue.
 485          */
 486         lnet_net_lock(0);
 487         /* the mt could've shutdown and cleaned up the queues */
 488         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 489                 lnet_net_unlock(0);
 490                 return;
 491         }
 492
 493         lnet_dec_healthv_locked(&local_ni->ni_healthv, lnet_health_sensitivity);
 494         lnet_ni_add_to_recoveryq_locked(local_ni, &the_lnet.ln_mt_localNIRecovq,
 495                                         ktime_get_seconds());
 496         lnet_net_unlock(0);
 497 }
 498
 499 /* must hold net_lock/0 */
 500 void
 501 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
 502 {
 503         __u32 sensitivity = lnet_health_sensitivity;
 504         __u32 lp_sensitivity;
 505
 506         /*
 507          * If there is a health sensitivity in the peer then use that
 508          * instead of the globally set one.
 509          */
 510         lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
 511         if (lp_sensitivity)
 512                 sensitivity = lp_sensitivity;
 513
 514         lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
 515
 516         /* update the peer_net's health value */
 517         lnet_update_peer_net_healthv(lpni);
 518
 519         /*
 520          * add the peer NI to the recovery queue if it's not already there
 521          * and it's health value is actually below the maximum. It's
 522          * possible that the sensitivity might be set to 0, and the health
 523          * value will not be reduced. In this case, there is no reason to
 524          * invoke recovery
 525          */
 526         lnet_peer_ni_add_to_recoveryq_locked(lpni,
 527                                              &the_lnet.ln_mt_peerNIRecovq,
 528                                              ktime_get_seconds());
 529 }
 530
 531 static void
 532 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 533 {
 534         /* lpni could be NULL if we're in the LOLND case */
 535         if (!lpni)
 536                 return;
 537
 538         lnet_net_lock(0);
 539         /* the mt could've shutdown and cleaned up the queues */
 540         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 541                 lnet_net_unlock(0);
 542                 return;
 543         }
 544         lnet_handle_remote_failure_locked(lpni);
 545         lnet_net_unlock(0);
 546 }
 547
 548 static void
 549 lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 550                  enum lnet_msg_hstatus hstatus)
 551 {
 552         struct lnet_counters_health *health;
 553
 554         health = &the_lnet.ln_counters[0]->lct_health;
 555
 556         switch (hstatus) {
 557         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 558                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 559                 health->lch_local_interrupt_count++;
 560                 break;
 561         case LNET_MSG_STATUS_LOCAL_DROPPED:
 562                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 563                 health->lch_local_dropped_count++;
 564                 break;
 565         case LNET_MSG_STATUS_LOCAL_ABORTED:
 566                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 567                 health->lch_local_aborted_count++;
 568                 break;
 569         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 570                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 571                 health->lch_local_no_route_count++;
 572                 break;
 573         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 574                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 575                 health->lch_local_timeout_count++;
 576                 break;
 577         case LNET_MSG_STATUS_LOCAL_ERROR:
 578                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 579                 health->lch_local_error_count++;
 580                 break;
 581         case LNET_MSG_STATUS_REMOTE_DROPPED:
 582                 if (lpni)
 583                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 584                 health->lch_remote_dropped_count++;
 585                 break;
 586         case LNET_MSG_STATUS_REMOTE_ERROR:
 587                 if (lpni)
 588                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 589                 health->lch_remote_error_count++;
 590                 break;
 591         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 592                 if (lpni)
 593                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 594                 health->lch_remote_timeout_count++;
 595                 break;
 596         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 597                 if (lpni)
 598                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 599                 health->lch_network_timeout_count++;
 600                 break;
 601         case LNET_MSG_STATUS_OK:
 602                 break;
 603         default:
 604                 LBUG();
 605         }
 606 }
 607
 608 static void
 609 lnet_resend_msg_locked(struct lnet_msg *msg)
 610 {
 611         msg->msg_retry_count++;
 612
 613         /*
 614          * remove message from the active list and reset it to prepare
 615          * for a resend. Two exceptions to this
 616          *
 617          * 1. the router case. When a message is being routed it is
 618          * committed for rx when received and committed for tx when
 619          * forwarded. We don't want to remove it from the active list, since
 620          * code which handles receiving expects it to remain on the active
 621          * list.
 622          *
 623          * 2. The REPLY case. Reply messages use the same message
 624          * structure for the GET that was received.
 625          */
 626         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 627                 list_del_init(&msg->msg_activelist);
 628                 msg->msg_onactivelist = 0;
 629         }
 630         /*
 631          * The msg_target.nid which was originally set
 632          * when calling LNetGet() or LNetPut() might've
 633          * been overwritten if we're routing this message.
 634          * Call lnet_msg_decommit_tx() to return the credit
 635          * this message consumed. The message will
 636          * consume another credit when it gets resent.
 637          */
 638         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 639         lnet_msg_decommit_tx(msg, -EAGAIN);
 640         msg->msg_sending = 0;
 641         msg->msg_receiving = 0;
 642         msg->msg_target_is_router = 0;
 643
 644         CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
 645                libcfs_nidstr(&msg->msg_hdr.src_nid),
 646                libcfs_nidstr(&msg->msg_hdr.dest_nid),
 647                lnet_msgtyp2str(msg->msg_type),
 648                lnet_health_error2str(msg->msg_health_status), msg);
 649
 650         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 651
 652         complete(&the_lnet.ln_mt_wait_complete);
 653 }
 654
 655 int
 656 lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
 657                                      struct list_head *containerq,
 658                                      int nworkers, void **workers)
 659 {
 660         int my_slot = -1;
 661         int i;
 662
 663         list_add_tail(&msg->msg_list, containerq);
 664
 665         for (i = 0; i < nworkers; i++) {
 666                 if (workers[i] == current)
 667                         break;
 668
 669                 if (my_slot < 0 && workers[i] == NULL)
 670                         my_slot = i;
 671         }
 672
 673         if (i < nworkers || my_slot < 0)
 674                 return -1;
 675
 676         workers[my_slot] = current;
 677
 678         return my_slot;
 679 }
 680
 681 int
 682 lnet_attempt_msg_resend(struct lnet_msg *msg)
 683 {
 684         struct lnet_msg_container *container;
 685         int my_slot;
 686         int cpt;
 687
 688         /* we can only resend tx_committed messages */
 689         LASSERT(msg->msg_tx_committed);
 690
 691         /* don't resend recovery messages */
 692         if (msg->msg_recovery) {
 693                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 694                         libcfs_nidstr(&msg->msg_from),
 695                         libcfs_nidstr(&msg->msg_target.nid),
 696                         msg->msg_retry_count);
 697                 return -ENOTRECOVERABLE;
 698         }
 699
 700         /*
 701          * if we explicitly indicated we don't want to resend then just
 702          * return
 703          */
 704         if (msg->msg_no_resend) {
 705                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 706                         libcfs_nidstr(&msg->msg_from),
 707                         libcfs_nidstr(&msg->msg_target.nid),
 708                         msg->msg_retry_count);
 709                 return -ENOTRECOVERABLE;
 710         }
 711
 712         /* check if the message has exceeded the number of retries */
 713         if (msg->msg_retry_count >= lnet_retry_count) {
 714                 CNETERR("msg %s->%s exceeded retry count %d\n",
 715                         libcfs_nidstr(&msg->msg_from),
 716                         libcfs_nidstr(&msg->msg_target.nid),
 717                         msg->msg_retry_count);
 718                 return -ENOTRECOVERABLE;
 719         }
 720
 721         cpt = msg->msg_tx_cpt;
 722         lnet_net_lock(cpt);
 723
 724         /* check again under lock */
 725         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 726                 lnet_net_unlock(cpt);
 727                 return -ESHUTDOWN;
 728         }
 729
 730         container = the_lnet.ln_msg_containers[cpt];
 731         my_slot =
 732                 lnet_check_finalize_recursion_locked(msg,
 733                                         &container->msc_resending,
 734                                         container->msc_nfinalizers,
 735                                         container->msc_resenders);
 736
 737         /* enough threads are resending */
 738         if (my_slot == -1) {
 739                 lnet_net_unlock(cpt);
 740                 return 0;
 741         }
 742
 743         while (!list_empty(&container->msc_resending)) {
 744                 msg = list_entry(container->msc_resending.next,
 745                                         struct lnet_msg, msg_list);
 746                 list_del(&msg->msg_list);
 747
 748                 /*
 749                  * resending the message will require us to call
 750                  * lnet_msg_decommit_tx() which will return the credit
 751                  * which this message holds. This could trigger another
 752                  * queued message to be sent. If that message fails and
 753                  * requires a resend we will recurse.
 754                  * But since at this point the slot is taken, the message
 755                  * will be queued in the container and dealt with
 756                  * later. This breaks the recursion.
 757                  */
 758                 lnet_resend_msg_locked(msg);
 759         }
 760
 761         /*
 762          * msc_resenders is an array of process pointers. Each entry holds
 763          * a pointer to the current process operating on the message. An
 764          * array entry is created per CPT. If the array slot is already
 765          * set, then it means that there is a thread on the CPT currently
 766          * resending a message.
 767          * Once the thread finishes clear the slot to enable the thread to
 768          * take on more resend work.
 769          */
 770         container->msc_resenders[my_slot] = NULL;
 771         lnet_net_unlock(cpt);
 772
 773         return 0;
 774 }
 775
 776 /*
 777  * Do a health check on the message:
 778  * return -1 if we're not going to handle the error or
 779  *   if we've reached the maximum number of retries.
 780  *   success case will return -1 as well
 781  * return 0 if it the message is requeued for send
 782  */
 783 static int
 784 lnet_health_check(struct lnet_msg *msg)
 785 {
 786         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 787         struct lnet_peer_ni *lpni;
 788         struct lnet_ni *ni;
 789         bool lo = false;
 790         bool attempt_local_resend;
 791         bool attempt_remote_resend;
 792         bool handle_local_health;
 793         bool handle_remote_health;
 794
 795         /* if we're shutting down no point in handling health. */
 796         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 797                 return -1;
 798
 799         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 800
 801         /*
 802          * if we're sending to the LOLND then the msg_txpeer will not be
 803          * set. So no need to sanity check it.
 804          */
 805         if (msg->msg_tx_committed &&
 806             !nid_is_lo0(&msg->msg_txni->ni_nid))
 807                 LASSERT(msg->msg_txpeer);
 808         else if (msg->msg_tx_committed &&
 809                  nid_is_lo0(&msg->msg_txni->ni_nid))
 810                 lo = true;
 811
 812         if (hstatus != LNET_MSG_STATUS_OK &&
 813             ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
 814                 return -1;
 815
 816         /*
 817          * always prefer txni/txpeer if they message is committed for both
 818          * directions.
 819          */
 820         if (msg->msg_tx_committed) {
 821                 ni = msg->msg_txni;
 822                 lpni = msg->msg_txpeer;
 823                 attempt_local_resend = attempt_remote_resend = true;
 824         } else {
 825                 ni = msg->msg_rxni;
 826                 lpni = msg->msg_rxpeer;
 827                 attempt_local_resend = attempt_remote_resend = false;
 828         }
 829
 830         if (!lo)
 831                 LASSERT(ni && lpni);
 832         else
 833                 LASSERT(ni);
 834
 835         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 836                libcfs_nidstr(&ni->ni_nid),
 837                (lo) ? "self" : libcfs_nidstr(&lpni->lpni_nid),
 838                lnet_msgtyp2str(msg->msg_type),
 839                lnet_health_error2str(hstatus));
 840
 841         /*
 842          * stats are only incremented for errors so avoid wasting time
 843          * incrementing statistics if there is no error. Similarly, whether to
 844          * update health values or perform resends is only applicable for
 845          * messages with a health status != OK.
 846          */
 847         if (hstatus != LNET_MSG_STATUS_OK) {
 848                 /* Don't further decrement the health value if a recovery
 849                  * message failed.
 850                  */
 851                 if (msg->msg_recovery)
 852                         handle_local_health = handle_remote_health = false;
 853                 else
 854                         handle_local_health = handle_remote_health = true;
 855
 856                 /* For local failures, health/recovery/resends are not needed if
 857                  * I only have a single (non-lolnd) interface. NB: pb_nnis
 858                  * includes the lolnd interface, so a single-rail node would
 859                  * have pb_nnis == 2.
 860                  */
 861                 if (the_lnet.ln_ping_target->pb_nnis <= 2) {
 862                         handle_local_health = false;
 863                         attempt_local_resend = false;
 864                 }
 865
 866                 lnet_net_lock(0);
 867                 lnet_incr_hstats(ni, lpni, hstatus);
 868                 /* For remote failures, health/recovery/resends are not needed
 869                  * if the peer only has a single interface. Special case for
 870                  * routers where we rely on health feature to manage route
 871                  * aliveness. NB: unlike pb_nnis above, lp_nnis does _not_
 872                  * include the lolnd, so a single-rail node would have
 873                  * lp_nnis == 1.
 874                  */
 875                 if (lpni && lpni->lpni_peer_net &&
 876                     lpni->lpni_peer_net->lpn_peer &&
 877                     lpni->lpni_peer_net->lpn_peer->lp_nnis <= 1) {
 878                         attempt_remote_resend = false;
 879                         if (!lnet_isrouter(lpni))
 880                                 handle_remote_health = false;
 881                 }
 882                 lnet_net_unlock(0);
 883         }
 884
 885         switch (hstatus) {
 886         case LNET_MSG_STATUS_OK:
 887                 /*
 888                  * increment the local ni health whether we successfully
 889                  * received or sent a message on it.
 890                  *
 891                  * Ping counts are reset to 0 as appropriate to allow for
 892                  * faster recovery.
 893                  */
 894                 lnet_inc_healthv(&ni->ni_healthv, lnet_health_sensitivity);
 895                 /*
 896                  * It's possible msg_txpeer is NULL in the LOLND
 897                  * case. Only increment the peer's health if we're
 898                  * receiving a message from it. It's the only sure way to
 899                  * know that a remote interface is up.
 900                  * If this interface is part of a router, then take that
 901                  * as indication that the router is fully healthy.
 902                  */
 903                 if (lpni && msg->msg_rx_committed) {
 904                         lnet_net_lock(0);
 905                         lpni->lpni_ping_count = 0;
 906                         ni->ni_ping_count = 0;
 907                         /*
 908                          * If we're receiving a message from the router or
 909                          * I'm a router, then set that lpni's health to
 910                          * maximum so we can commence communication
 911                          */
 912                         if (lnet_isrouter(lpni) || the_lnet.ln_routing) {
 913                                 lnet_set_lpni_healthv_locked(lpni,
 914                                         LNET_MAX_HEALTH_VALUE);
 915                         } else {
 916                                 __u32 sensitivity = lpni->lpni_peer_net->
 917                                         lpn_peer->lp_health_sensitivity;
 918
 919                                 lnet_inc_lpni_healthv_locked(lpni,
 920                                         (sensitivity) ? sensitivity :
 921                                         lnet_health_sensitivity);
 922                                 /* This peer NI may have previously aged out
 923                                  * of recovery. Now that we've received a
 924                                  * message from it, we can continue recovery
 925                                  * if its health value is still below the
 926                                  * maximum.
 927                                  */
 928                                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
 929                                                 &the_lnet.ln_mt_peerNIRecovq,
 930                                                 ktime_get_seconds());
 931                         }
 932                         lnet_net_unlock(0);
 933                 }
 934
 935                 /* we can finalize this message */
 936                 return -1;
 937         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 938         case LNET_MSG_STATUS_LOCAL_DROPPED:
 939         case LNET_MSG_STATUS_LOCAL_ABORTED:
 940         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 941         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 942                 if (handle_local_health)
 943                         lnet_handle_local_failure(ni);
 944                 if (attempt_local_resend)
 945                         return lnet_attempt_msg_resend(msg);
 946                 break;
 947         case LNET_MSG_STATUS_LOCAL_ERROR:
 948                 if (handle_local_health)
 949                         lnet_handle_local_failure(ni);
 950                 return -1;
 951         case LNET_MSG_STATUS_REMOTE_DROPPED:
 952                 if (handle_remote_health)
 953                         lnet_handle_remote_failure(lpni);
 954                 if (attempt_remote_resend)
 955                         return lnet_attempt_msg_resend(msg);
 956                 break;
 957         case LNET_MSG_STATUS_REMOTE_ERROR:
 958         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 959                 if (handle_remote_health)
 960                         lnet_handle_remote_failure(lpni);
 961                 return -1;
 962         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 963                 if (handle_remote_health)
 964                         lnet_handle_remote_failure(lpni);
 965                 if (handle_local_health)
 966                         lnet_handle_local_failure(ni);
 967                 return -1;
 968         default:
 969                 LBUG();
 970         }
 971
 972         /* no resend is needed */
 973         return -1;
 974 }
 975
 976 static void
 977 lnet_msg_detach_md(struct lnet_msg *msg, int status)
 978 {
 979         struct lnet_libmd *md = msg->msg_md;
 980         lnet_handler_t handler = NULL;
 981         int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
 982         int unlink;
 983
 984         lnet_res_lock(cpt);
 985         while (md->md_flags & LNET_MD_FLAG_HANDLING)
 986                 /* An event handler is running - wait for it to
 987                  * complete to avoid races.
 988                  */
 989                 lnet_md_wait_handling(md, cpt);
 990
 991         /* Now it's safe to drop my caller's ref */
 992         md->md_refcount--;
 993         LASSERT(md->md_refcount >= 0);
 994
 995         unlink = lnet_md_unlinkable(md);
 996         if (md->md_handler) {
 997                 if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
 998                         msg->msg_ev.status   = -ETIMEDOUT;
 999                         CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
1000                 } else {
1001                         msg->msg_ev.status   = status;
1002                 }
1003                 msg->msg_ev.unlinked = unlink;
1004                 handler = md->md_handler;
1005                 if (!unlink)
1006                         md->md_flags |= LNET_MD_FLAG_HANDLING;
1007         }
1008
1009         if (unlink || (md->md_refcount == 0 &&
1010                        md->md_threshold == LNET_MD_THRESH_INF))
1011                 lnet_detach_rsp_tracker(md, cpt);
1012
1013         msg->msg_md = NULL;
1014         if (unlink)
1015                 lnet_md_unlink(md);
1016
1017         lnet_res_unlock(cpt);
1018
1019         if (handler) {
1020                 handler(&msg->msg_ev);
1021                 if (!unlink) {
1022                         lnet_res_lock(cpt);
1023                         md->md_flags &= ~LNET_MD_FLAG_HANDLING;
1024                         wake_up_var(md);
1025                         lnet_res_unlock(cpt);
1026                 }
1027         }
1028 }
1029
1030 static bool
1031 lnet_is_health_check(struct lnet_msg *msg)
1032 {
1033         bool hc = true;
1034         int status = msg->msg_ev.status;
1035
1036         if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
1037             !msg->msg_onactivelist) {
1038                 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
1039                        msg);
1040                 return false;
1041         }
1042
1043         if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
1044             (msg->msg_rx_committed && !msg->msg_rxpeer)) {
1045                 /* The optimized GET case does not set msg_rxpeer, but status
1046                  * could be zero. Only print the error message if we have a
1047                  * non-zero status.
1048                  */
1049                 if (status)
1050                         CDEBUG(D_NET, "msg %p status %d cannot retry\n", msg,
1051                                status);
1052                 return false;
1053         }
1054
1055         /* Check for status inconsistencies */
1056         if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
1057              (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
1058                 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
1059                        "checking (%d, %d)\n", msg, status,
1060                        msg->msg_health_status);
1061                 hc = false;
1062         }
1063
1064         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
1065                hc, status, msg->msg_health_status);
1066
1067         return hc;
1068 }
1069
1070 char *
1071 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
1072 {
1073         switch (hstatus) {
1074         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
1075                 return "LOCAL_INTERRUPT";
1076         case LNET_MSG_STATUS_LOCAL_DROPPED:
1077                 return "LOCAL_DROPPED";
1078         case LNET_MSG_STATUS_LOCAL_ABORTED:
1079                 return "LOCAL_ABORTED";
1080         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
1081                 return "LOCAL_NO_ROUTE";
1082         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
1083                 return "LOCAL_TIMEOUT";
1084         case LNET_MSG_STATUS_LOCAL_ERROR:
1085                 return "LOCAL_ERROR";
1086         case LNET_MSG_STATUS_REMOTE_DROPPED:
1087                 return "REMOTE_DROPPED";
1088         case LNET_MSG_STATUS_REMOTE_ERROR:
1089                 return "REMOTE_ERROR";
1090         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
1091                 return "REMOTE_TIMEOUT";
1092         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
1093                 return "NETWORK_TIMEOUT";
1094         case LNET_MSG_STATUS_OK:
1095                 return "OK";
1096         default:
1097                 return "<UNKNOWN>";
1098         }
1099 }
1100
1101 bool
1102 lnet_send_error_simulation(struct lnet_msg *msg,
1103                            enum lnet_msg_hstatus *hstatus)
1104 {
1105         if (!msg)
1106                 return false;
1107
1108         if (list_empty(&the_lnet.ln_drop_rules))
1109             return false;
1110
1111         /* match only health rules */
1112         if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
1113                                   hstatus))
1114                 return false;
1115
1116         CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
1117                 libcfs_nidstr(&msg->msg_hdr.src_nid),
1118                 libcfs_nidstr(&msg->msg_txni->ni_nid),
1119                 libcfs_nidstr(&msg->msg_hdr.dest_nid),
1120                 lnet_msgtyp2str(msg->msg_type),
1121                 lnet_health_error2str(*hstatus));
1122
1123         return true;
1124 }
1125 EXPORT_SYMBOL(lnet_send_error_simulation);
1126
1127 void
1128 lnet_finalize(struct lnet_msg *msg, int status)
1129 {
1130         struct lnet_msg_container *container;
1131         int my_slot;
1132         int cpt;
1133         int rc;
1134
1135         LASSERT(!in_interrupt());
1136
1137         if (msg == NULL)
1138                 return;
1139
1140         msg->msg_ev.status = status;
1141
1142         if (lnet_is_health_check(msg)) {
1143                 /*
1144                  * Check the health status of the message. If it has one
1145                  * of the errors that we're supposed to handle, and it has
1146                  * not timed out, then
1147                  *      1. Decrement the appropriate health_value
1148                  *      2. queue the message on the resend queue
1149
1150                  * if the message send is success, timed out or failed in the
1151                  * health check for any reason then we'll just finalize the
1152                  * message. Otherwise just return since the message has been
1153                  * put on the resend queue.
1154                  */
1155                 if (!lnet_health_check(msg))
1156                         return;
1157         }
1158
1159         /*
1160          * We're not going to resend this message so detach its MD and invoke
1161          * the appropriate callbacks
1162          */
1163         if (msg->msg_md != NULL)
1164                 lnet_msg_detach_md(msg, status);
1165
1166 again:
1167         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
1168                 /* not committed to network yet */
1169                 LASSERT(!msg->msg_onactivelist);
1170                 lnet_msg_free(msg);
1171                 return;
1172         }
1173
1174         /*
1175          * NB: routed message can be committed for both receiving and sending,
1176          * we should finalize in LIFO order and keep counters correct.
1177          * (finalize sending first then finalize receiving)
1178          */
1179         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
1180         lnet_net_lock(cpt);
1181
1182         container = the_lnet.ln_msg_containers[cpt];
1183
1184         /* Recursion breaker.  Don't complete the message here if I am (or
1185          * enough other threads are) already completing messages */
1186         my_slot = lnet_check_finalize_recursion_locked(msg,
1187                                                 &container->msc_finalizing,
1188                                                 container->msc_nfinalizers,
1189                                                 container->msc_finalizers);
1190
1191         /* enough threads are resending */
1192         if (my_slot == -1) {
1193                 lnet_net_unlock(cpt);
1194                 return;
1195         }
1196
1197         rc = 0;
1198         while (!list_empty(&container->msc_finalizing)) {
1199                 msg = list_entry(container->msc_finalizing.next,
1200                                  struct lnet_msg, msg_list);
1201
1202                 list_del_init(&msg->msg_list);
1203
1204                 /* NB drops and regains the lnet lock if it actually does
1205                  * anything, so my finalizing friends can chomp along too */
1206                 rc = lnet_complete_msg_locked(msg, cpt);
1207                 if (rc != 0)
1208                         break;
1209         }
1210
1211         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1212                 lnet_net_unlock(cpt);
1213                 lnet_delay_rule_check();
1214                 lnet_net_lock(cpt);
1215         }
1216
1217         container->msc_finalizers[my_slot] = NULL;
1218         lnet_net_unlock(cpt);
1219
1220         if (rc != 0)
1221                 goto again;
1222 }
1223 EXPORT_SYMBOL(lnet_finalize);
1224
1225 void
1226 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1227 {
1228         int     count = 0;
1229
1230         if (container->msc_init == 0)
1231                 return;
1232
1233         while (!list_empty(&container->msc_active)) {
1234                 struct lnet_msg *msg;
1235
1236                 msg  = list_entry(container->msc_active.next,
1237                                   struct lnet_msg, msg_activelist);
1238                 LASSERT(msg->msg_onactivelist);
1239                 msg->msg_onactivelist = 0;
1240                 list_del_init(&msg->msg_activelist);
1241                 lnet_msg_free(msg);
1242                 count++;
1243         }
1244
1245         if (count > 0)
1246                 CERROR("%d active msg on exit\n", count);
1247
1248         if (container->msc_finalizers != NULL) {
1249                 CFS_FREE_PTR_ARRAY(container->msc_finalizers,
1250                                    container->msc_nfinalizers);
1251                 container->msc_finalizers = NULL;
1252         }
1253
1254         if (container->msc_resenders != NULL) {
1255                 CFS_FREE_PTR_ARRAY(container->msc_resenders,
1256                                    container->msc_nfinalizers);
1257                 container->msc_resenders = NULL;
1258         }
1259         container->msc_init = 0;
1260 }
1261
1262 int
1263 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1264 {
1265         int rc = 0;
1266
1267         container->msc_init = 1;
1268
1269         INIT_LIST_HEAD(&container->msc_active);
1270         INIT_LIST_HEAD(&container->msc_finalizing);
1271         INIT_LIST_HEAD(&container->msc_resending);
1272
1273         /* number of CPUs */
1274         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1275         if (container->msc_nfinalizers == 0)
1276                 container->msc_nfinalizers = 1;
1277
1278         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1279                          container->msc_nfinalizers *
1280                          sizeof(*container->msc_finalizers));
1281
1282         if (container->msc_finalizers == NULL) {
1283                 CERROR("Failed to allocate message finalizers\n");
1284                 lnet_msg_container_cleanup(container);
1285                 return -ENOMEM;
1286         }
1287
1288         LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
1289                          container->msc_nfinalizers *
1290                          sizeof(*container->msc_resenders));
1291
1292         if (container->msc_resenders == NULL) {
1293                 CERROR("Failed to allocate message resenders\n");
1294                 lnet_msg_container_cleanup(container);
1295                 return -ENOMEM;
1296         }
1297
1298         return rc;
1299 }
1300
1301 void
1302 lnet_msg_containers_destroy(void)
1303 {
1304         struct lnet_msg_container *container;
1305         int     i;
1306
1307         if (the_lnet.ln_msg_containers == NULL)
1308                 return;
1309
1310         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1311                 lnet_msg_container_cleanup(container);
1312
1313         cfs_percpt_free(the_lnet.ln_msg_containers);
1314         the_lnet.ln_msg_containers = NULL;
1315 }
1316
1317 int
1318 lnet_msg_containers_create(void)
1319 {
1320         struct lnet_msg_container *container;
1321         int     rc;
1322         int     i;
1323
1324         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1325                                                       sizeof(*container));
1326
1327         if (the_lnet.ln_msg_containers == NULL) {
1328                 CERROR("Failed to allocate cpu-partition data for network\n");
1329                 return -ENOMEM;
1330         }
1331
1332         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1333                 rc = lnet_msg_container_setup(container, i);
1334                 if (rc != 0) {
1335                         lnet_msg_containers_destroy();
1336                         return rc;
1337                 }
1338         }
1339
1340         return 0;
1341 }