lnet/lnet/lib-msg.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lnet/lnet/lib-msg.c
  32  *
  33  * Message decoding, parsing and finalizing routines
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_LNET
  37
  38 #include <lnet/lib-lnet.h>
  39
  40 void
  41 lnet_build_unlink_event(struct lnet_libmd *md, struct lnet_event *ev)
  42 {
  43         ENTRY;
  44
  45         memset(ev, 0, sizeof(*ev));
  46
  47         ev->status   = 0;
  48         ev->unlinked = 1;
  49         ev->type     = LNET_EVENT_UNLINK;
  50         lnet_md_deconstruct(md, ev);
  51         lnet_md2handle(&ev->md_handle, md);
  52         EXIT;
  53 }
  54
  55 /*
  56  * Don't need any lock, must be called after lnet_commit_md
  57  */
  58 void
  59 lnet_build_msg_event(struct lnet_msg *msg, enum lnet_event_kind ev_type)
  60 {
  61         struct lnet_hdr *hdr = &msg->msg_hdr;
  62         struct lnet_event *ev = &msg->msg_ev;
  63
  64         LASSERT(!msg->msg_routing);
  65
  66         ev->type = ev_type;
  67         ev->msg_type = msg->msg_type;
  68
  69         if (ev_type == LNET_EVENT_SEND) {
  70                 /* event for active message */
  71                 ev->target.nid    = hdr->dest_nid;
  72                 ev->target.pid    = hdr->dest_pid;
  73                 ev->initiator.nid = LNET_ANY_NID;
  74                 ev->initiator.pid = the_lnet.ln_pid;
  75                 ev->source.nid    = LNET_ANY_NID;
  76                 ev->source.pid    = the_lnet.ln_pid;
  77                 ev->sender        = LNET_ANY_NID;
  78         } else {
  79                 /* event for passive message */
  80                 ev->target.pid    = hdr->dest_pid;
  81                 ev->target.nid    = hdr->dest_nid;
  82                 ev->initiator.pid = hdr->src_pid;
  83                 /* Multi-Rail: resolve src_nid to "primary" peer NID */
  84                 ev->initiator.nid = msg->msg_initiator;
  85                 /* Multi-Rail: track source NID. */
  86                 ev->source.pid    = hdr->src_pid;
  87                 ev->source.nid    = hdr->src_nid;
  88                 ev->rlength       = hdr->payload_length;
  89                 ev->sender        = msg->msg_from;
  90                 ev->mlength       = msg->msg_wanted;
  91                 ev->offset        = msg->msg_offset;
  92         }
  93
  94         switch (ev_type) {
  95         default:
  96                 LBUG();
  97
  98         case LNET_EVENT_PUT: /* passive PUT */
  99                 ev->pt_index   = hdr->msg.put.ptl_index;
 100                 ev->match_bits = hdr->msg.put.match_bits;
 101                 ev->hdr_data   = hdr->msg.put.hdr_data;
 102                 return;
 103
 104         case LNET_EVENT_GET: /* passive GET */
 105                 ev->pt_index   = hdr->msg.get.ptl_index;
 106                 ev->match_bits = hdr->msg.get.match_bits;
 107                 ev->hdr_data   = 0;
 108                 return;
 109
 110         case LNET_EVENT_ACK: /* ACK */
 111                 ev->match_bits = hdr->msg.ack.match_bits;
 112                 ev->mlength    = hdr->msg.ack.mlength;
 113                 return;
 114
 115         case LNET_EVENT_REPLY: /* REPLY */
 116                 return;
 117
 118         case LNET_EVENT_SEND: /* active message */
 119                 if (msg->msg_type == LNET_MSG_PUT) {
 120                         ev->pt_index   = le32_to_cpu(hdr->msg.put.ptl_index);
 121                         ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
 122                         ev->offset     = le32_to_cpu(hdr->msg.put.offset);
 123                         ev->mlength    =
 124                         ev->rlength    = le32_to_cpu(hdr->payload_length);
 125                         ev->hdr_data   = le64_to_cpu(hdr->msg.put.hdr_data);
 126
 127                 } else {
 128                         LASSERT(msg->msg_type == LNET_MSG_GET);
 129                         ev->pt_index   = le32_to_cpu(hdr->msg.get.ptl_index);
 130                         ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
 131                         ev->mlength    =
 132                         ev->rlength    = le32_to_cpu(hdr->msg.get.sink_length);
 133                         ev->offset     = le32_to_cpu(hdr->msg.get.src_offset);
 134                         ev->hdr_data   = 0;
 135                 }
 136                 return;
 137         }
 138 }
 139
 140 void
 141 lnet_msg_commit(struct lnet_msg *msg, int cpt)
 142 {
 143         struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
 144         struct lnet_counters_common *common;
 145         s64 timeout_ns;
 146
 147         /* set the message deadline */
 148         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
 149         msg->msg_deadline = ktime_add_ns(ktime_get(), timeout_ns);
 150
 151         /* routed message can be committed for both receiving and sending */
 152         LASSERT(!msg->msg_tx_committed);
 153
 154         if (msg->msg_sending) {
 155                 LASSERT(!msg->msg_receiving);
 156                 msg->msg_tx_cpt = cpt;
 157                 msg->msg_tx_committed = 1;
 158                 if (msg->msg_rx_committed) { /* routed message REPLY */
 159                         LASSERT(msg->msg_onactivelist);
 160                         return;
 161                 }
 162         } else {
 163                 LASSERT(!msg->msg_sending);
 164                 msg->msg_rx_cpt = cpt;
 165                 msg->msg_rx_committed = 1;
 166         }
 167
 168         LASSERT(!msg->msg_onactivelist);
 169
 170         msg->msg_onactivelist = 1;
 171         list_add_tail(&msg->msg_activelist, &container->msc_active);
 172
 173         common = &the_lnet.ln_counters[cpt]->lct_common;
 174         common->lcc_msgs_alloc++;
 175         if (common->lcc_msgs_alloc > common->lcc_msgs_max)
 176                 common->lcc_msgs_max = common->lcc_msgs_alloc;
 177 }
 178
 179 static void
 180 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
 181 {
 182         struct lnet_counters_common *common;
 183         struct lnet_event *ev = &msg->msg_ev;
 184
 185         LASSERT(msg->msg_tx_committed);
 186         if (status != 0)
 187                 goto out;
 188
 189         common = &(the_lnet.ln_counters[msg->msg_tx_cpt]->lct_common);
 190         switch (ev->type) {
 191         default: /* routed message */
 192                 LASSERT(msg->msg_routing);
 193                 LASSERT(msg->msg_rx_committed);
 194                 LASSERT(ev->type == 0);
 195
 196                 common->lcc_route_length += msg->msg_len;
 197                 common->lcc_route_count++;
 198                 goto incr_stats;
 199
 200         case LNET_EVENT_PUT:
 201                 /* should have been decommitted */
 202                 LASSERT(!msg->msg_rx_committed);
 203                 /* overwritten while sending ACK */
 204                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 205                 msg->msg_type = LNET_MSG_PUT; /* fix type */
 206                 break;
 207
 208         case LNET_EVENT_SEND:
 209                 LASSERT(!msg->msg_rx_committed);
 210                 if (msg->msg_type == LNET_MSG_PUT)
 211                         common->lcc_send_length += msg->msg_len;
 212                 break;
 213
 214         case LNET_EVENT_GET:
 215                 LASSERT(msg->msg_rx_committed);
 216                 /* overwritten while sending reply, we should never be
 217                  * here for optimized GET */
 218                 LASSERT(msg->msg_type == LNET_MSG_REPLY);
 219                 msg->msg_type = LNET_MSG_GET; /* fix type */
 220                 break;
 221         }
 222
 223         common->lcc_send_count++;
 224
 225 incr_stats:
 226         if (msg->msg_txpeer)
 227                 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 228                                 msg->msg_type,
 229                                 LNET_STATS_TYPE_SEND);
 230         if (msg->msg_txni)
 231                 lnet_incr_stats(&msg->msg_txni->ni_stats,
 232                                 msg->msg_type,
 233                                 LNET_STATS_TYPE_SEND);
 234  out:
 235         lnet_return_tx_credits_locked(msg);
 236         msg->msg_tx_committed = 0;
 237 }
 238
 239 static void
 240 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
 241 {
 242         struct lnet_counters_common *common;
 243         struct lnet_event *ev = &msg->msg_ev;
 244
 245         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 246         LASSERT(msg->msg_rx_committed);
 247
 248         if (status != 0)
 249                 goto out;
 250
 251         common = &(the_lnet.ln_counters[msg->msg_rx_cpt]->lct_common);
 252         switch (ev->type) {
 253         default:
 254                 LASSERT(ev->type == 0);
 255                 LASSERT(msg->msg_routing);
 256                 goto incr_stats;
 257
 258         case LNET_EVENT_ACK:
 259                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 260                 break;
 261
 262         case LNET_EVENT_GET:
 263                 /* type is "REPLY" if it's an optimized GET on passive side,
 264                  * because optimized GET will never be committed for sending,
 265                  * so message type wouldn't be changed back to "GET" by
 266                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 267                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 268                         msg->msg_type == LNET_MSG_GET);
 269                 common->lcc_send_length += msg->msg_wanted;
 270                 break;
 271
 272         case LNET_EVENT_PUT:
 273                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 274                 break;
 275
 276         case LNET_EVENT_REPLY:
 277                 /* type is "GET" if it's an optimized GET on active side,
 278                  * see details in lnet_create_reply_msg() */
 279                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 280                         msg->msg_type == LNET_MSG_REPLY);
 281                 break;
 282         }
 283
 284         common->lcc_recv_count++;
 285
 286 incr_stats:
 287         if (msg->msg_rxpeer)
 288                 lnet_incr_stats(&msg->msg_rxpeer->lpni_stats,
 289                                 msg->msg_type,
 290                                 LNET_STATS_TYPE_RECV);
 291         if (msg->msg_rxni)
 292                 lnet_incr_stats(&msg->msg_rxni->ni_stats,
 293                                 msg->msg_type,
 294                                 LNET_STATS_TYPE_RECV);
 295         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 296                 common->lcc_recv_length += msg->msg_wanted;
 297
 298  out:
 299         lnet_return_rx_credits_locked(msg);
 300         msg->msg_rx_committed = 0;
 301 }
 302
 303 void
 304 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
 305 {
 306         int     cpt2 = cpt;
 307
 308         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 309         LASSERT(msg->msg_onactivelist);
 310
 311         if (msg->msg_tx_committed) { /* always decommit for sending first */
 312                 LASSERT(cpt == msg->msg_tx_cpt);
 313                 lnet_msg_decommit_tx(msg, status);
 314         }
 315
 316         if (msg->msg_rx_committed) {
 317                 /* forwarding msg committed for both receiving and sending */
 318                 if (cpt != msg->msg_rx_cpt) {
 319                         lnet_net_unlock(cpt);
 320                         cpt2 = msg->msg_rx_cpt;
 321                         lnet_net_lock(cpt2);
 322                 }
 323                 lnet_msg_decommit_rx(msg, status);
 324         }
 325
 326         list_del(&msg->msg_activelist);
 327         msg->msg_onactivelist = 0;
 328
 329         the_lnet.ln_counters[cpt2]->lct_common.lcc_msgs_alloc--;
 330
 331         if (cpt2 != cpt) {
 332                 lnet_net_unlock(cpt2);
 333                 lnet_net_lock(cpt);
 334         }
 335 }
 336
 337 void
 338 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
 339                    unsigned int offset, unsigned int mlen)
 340 {
 341         /* NB: @offset and @len are only useful for receiving */
 342         /* Here, we attach the MD on lnet_msg and mark it busy and
 343          * decrementing its threshold. Come what may, the lnet_msg "owns"
 344          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 345          * signals completion. */
 346         LASSERT(!msg->msg_routing);
 347
 348         msg->msg_md = md;
 349         if (msg->msg_receiving) { /* committed for receiving */
 350                 msg->msg_offset = offset;
 351                 msg->msg_wanted = mlen;
 352         }
 353
 354         md->md_refcount++;
 355         if (md->md_threshold != LNET_MD_THRESH_INF) {
 356                 LASSERT(md->md_threshold > 0);
 357                 md->md_threshold--;
 358         }
 359
 360         /* build umd in event */
 361         lnet_md2handle(&msg->msg_ev.md_handle, md);
 362         lnet_md_deconstruct(md, &msg->msg_ev);
 363 }
 364
 365 static int
 366 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
 367 {
 368         struct lnet_handle_wire ack_wmd;
 369         int                rc;
 370         int                status = msg->msg_ev.status;
 371
 372         LASSERT(msg->msg_onactivelist);
 373
 374         if (status == 0 && msg->msg_ack) {
 375                 /* Only send an ACK if the PUT completed successfully */
 376
 377                 lnet_msg_decommit(msg, cpt, 0);
 378
 379                 msg->msg_ack = 0;
 380                 lnet_net_unlock(cpt);
 381
 382                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 383                 LASSERT(!msg->msg_routing);
 384
 385                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 386
 387                 lnet_prep_send(msg, LNET_MSG_ACK, &msg->msg_ev.source, 0, 0);
 388
 389                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 390                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 391                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 392
 393                 rc = lnet_send(&msg->msg_ev.target.nid, msg,
 394                                &msg->msg_from);
 395
 396                 lnet_net_lock(cpt);
 397                 /*
 398                  * NB: message is committed for sending, we should return
 399                  * on success because LND will finalize this message later.
 400                  *
 401                  * Also, there is possibility that message is committed for
 402                  * sending and also failed before delivering to LND,
 403                  * i.e: ENOMEM, in that case we can't fall through either
 404                  * because CPT for sending can be different with CPT for
 405                  * receiving, so we should return back to lnet_finalize()
 406                  * to make sure we are locking the correct partition.
 407                  */
 408                 return rc;
 409
 410         } else if (status == 0 &&       /* OK so far */
 411                    (msg->msg_routing && !msg->msg_sending)) {
 412                 /* not forwarded */
 413                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 414                 lnet_net_unlock(cpt);
 415
 416                 rc = lnet_send(NULL, msg, NULL);
 417
 418                 lnet_net_lock(cpt);
 419                 /*
 420                  * NB: message is committed for sending, we should return
 421                  * on success because LND will finalize this message later.
 422                  *
 423                  * Also, there is possibility that message is committed for
 424                  * sending and also failed before delivering to LND,
 425                  * i.e: ENOMEM, in that case we can't fall through either:
 426                  * - The rule is message must decommit for sending first if
 427                  *   the it's committed for both sending and receiving
 428                  * - CPT for sending can be different with CPT for receiving,
 429                  *   so we should return back to lnet_finalize() to make
 430                  *   sure we are locking the correct partition.
 431                  */
 432                 return rc;
 433         }
 434
 435         lnet_msg_decommit(msg, cpt, status);
 436         lnet_msg_free(msg);
 437         return 0;
 438 }
 439
 440 static void
 441 lnet_dec_healthv_locked(atomic_t *healthv, int sensitivity)
 442 {
 443         int h = atomic_read(healthv);
 444
 445         if (h < sensitivity) {
 446                 atomic_set(healthv, 0);
 447         } else {
 448                 h -= sensitivity;
 449                 atomic_set(healthv, h);
 450         }
 451 }
 452
 453 /* must hold net_lock/0 */
 454 void
 455 lnet_ni_add_to_recoveryq_locked(struct lnet_ni *ni,
 456                                 struct list_head *recovery_queue, time64_t now)
 457 {
 458         if (!list_empty(&ni->ni_recovery))
 459                 return;
 460
 461         if (atomic_read(&ni->ni_healthv) == LNET_MAX_HEALTH_VALUE)
 462                 return;
 463
 464         /* This NI is going on the recovery queue, so take a ref on it */
 465         lnet_ni_addref_locked(ni, 0);
 466
 467         lnet_ni_set_next_ping(ni, now);
 468
 469         CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld health :%d\n",
 470                libcfs_nidstr(&ni->ni_nid),
 471                ni->ni_ping_count,
 472                ni->ni_next_ping,
 473                atomic_read(&ni->ni_healthv));
 474
 475         list_add_tail(&ni->ni_recovery, recovery_queue);
 476 }
 477
 478 static void
 479 lnet_handle_local_failure(struct lnet_ni *local_ni)
 480 {
 481         /*
 482          * the lnet_net_lock(0) is used to protect the addref on the ni
 483          * and the recovery queue.
 484          */
 485         lnet_net_lock(0);
 486         /* the mt could've shutdown and cleaned up the queues */
 487         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 488                 lnet_net_unlock(0);
 489                 return;
 490         }
 491
 492         lnet_dec_healthv_locked(&local_ni->ni_healthv, lnet_health_sensitivity);
 493         lnet_ni_add_to_recoveryq_locked(local_ni, &the_lnet.ln_mt_localNIRecovq,
 494                                         ktime_get_seconds());
 495         lnet_net_unlock(0);
 496 }
 497
 498 /* must hold net_lock/0 */
 499 void
 500 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
 501 {
 502         __u32 sensitivity = lnet_health_sensitivity;
 503         __u32 lp_sensitivity;
 504
 505         /*
 506          * If there is a health sensitivity in the peer then use that
 507          * instead of the globally set one.
 508          */
 509         lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
 510         if (lp_sensitivity)
 511                 sensitivity = lp_sensitivity;
 512
 513         lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
 514
 515         /* update the peer_net's health value */
 516         lnet_update_peer_net_healthv(lpni);
 517
 518         /*
 519          * add the peer NI to the recovery queue if it's not already there
 520          * and it's health value is actually below the maximum. It's
 521          * possible that the sensitivity might be set to 0, and the health
 522          * value will not be reduced. In this case, there is no reason to
 523          * invoke recovery
 524          */
 525         lnet_peer_ni_add_to_recoveryq_locked(lpni,
 526                                              &the_lnet.ln_mt_peerNIRecovq,
 527                                              ktime_get_seconds());
 528 }
 529
 530 static void
 531 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 532 {
 533         /* lpni could be NULL if we're in the LOLND case */
 534         if (!lpni)
 535                 return;
 536
 537         lnet_net_lock(0);
 538         /* the mt could've shutdown and cleaned up the queues */
 539         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 540                 lnet_net_unlock(0);
 541                 return;
 542         }
 543         lnet_handle_remote_failure_locked(lpni);
 544         lnet_net_unlock(0);
 545 }
 546
 547 static void
 548 lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 549                  enum lnet_msg_hstatus hstatus)
 550 {
 551         struct lnet_counters_health *health;
 552
 553         health = &the_lnet.ln_counters[0]->lct_health;
 554
 555         switch (hstatus) {
 556         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 557                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 558                 health->lch_local_interrupt_count++;
 559                 break;
 560         case LNET_MSG_STATUS_LOCAL_DROPPED:
 561                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 562                 health->lch_local_dropped_count++;
 563                 break;
 564         case LNET_MSG_STATUS_LOCAL_ABORTED:
 565                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 566                 health->lch_local_aborted_count++;
 567                 break;
 568         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 569                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 570                 health->lch_local_no_route_count++;
 571                 break;
 572         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 573                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 574                 health->lch_local_timeout_count++;
 575                 break;
 576         case LNET_MSG_STATUS_LOCAL_ERROR:
 577                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 578                 health->lch_local_error_count++;
 579                 break;
 580         case LNET_MSG_STATUS_REMOTE_DROPPED:
 581                 if (lpni)
 582                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 583                 health->lch_remote_dropped_count++;
 584                 break;
 585         case LNET_MSG_STATUS_REMOTE_ERROR:
 586                 if (lpni)
 587                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 588                 health->lch_remote_error_count++;
 589                 break;
 590         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 591                 if (lpni)
 592                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 593                 health->lch_remote_timeout_count++;
 594                 break;
 595         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 596                 if (lpni)
 597                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 598                 health->lch_network_timeout_count++;
 599                 break;
 600         case LNET_MSG_STATUS_OK:
 601                 break;
 602         default:
 603                 LBUG();
 604         }
 605 }
 606
 607 static void
 608 lnet_resend_msg_locked(struct lnet_msg *msg)
 609 {
 610         msg->msg_retry_count++;
 611
 612         /*
 613          * remove message from the active list and reset it to prepare
 614          * for a resend. Two exceptions to this
 615          *
 616          * 1. the router case. When a message is being routed it is
 617          * committed for rx when received and committed for tx when
 618          * forwarded. We don't want to remove it from the active list, since
 619          * code which handles receiving expects it to remain on the active
 620          * list.
 621          *
 622          * 2. The REPLY case. Reply messages use the same message
 623          * structure for the GET that was received.
 624          */
 625         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 626                 list_del_init(&msg->msg_activelist);
 627                 msg->msg_onactivelist = 0;
 628         }
 629         /*
 630          * The msg_target.nid which was originally set
 631          * when calling LNetGet() or LNetPut() might've
 632          * been overwritten if we're routing this message.
 633          * Call lnet_msg_decommit_tx() to return the credit
 634          * this message consumed. The message will
 635          * consume another credit when it gets resent.
 636          */
 637         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 638         lnet_msg_decommit_tx(msg, -EAGAIN);
 639         msg->msg_sending = 0;
 640         msg->msg_receiving = 0;
 641         msg->msg_target_is_router = 0;
 642
 643         CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
 644                libcfs_nidstr(&msg->msg_hdr.src_nid),
 645                libcfs_nidstr(&msg->msg_hdr.dest_nid),
 646                lnet_msgtyp2str(msg->msg_type),
 647                lnet_health_error2str(msg->msg_health_status), msg);
 648
 649         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 650
 651         complete(&the_lnet.ln_mt_wait_complete);
 652 }
 653
 654 int
 655 lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
 656                                      struct list_head *containerq,
 657                                      int nworkers, void **workers)
 658 {
 659         int my_slot = -1;
 660         int i;
 661
 662         list_add_tail(&msg->msg_list, containerq);
 663
 664         for (i = 0; i < nworkers; i++) {
 665                 if (workers[i] == current)
 666                         break;
 667
 668                 if (my_slot < 0 && workers[i] == NULL)
 669                         my_slot = i;
 670         }
 671
 672         if (i < nworkers || my_slot < 0)
 673                 return -1;
 674
 675         workers[my_slot] = current;
 676
 677         return my_slot;
 678 }
 679
 680 int
 681 lnet_attempt_msg_resend(struct lnet_msg *msg)
 682 {
 683         struct lnet_msg_container *container;
 684         int my_slot;
 685         int cpt;
 686
 687         /* we can only resend tx_committed messages */
 688         LASSERT(msg->msg_tx_committed);
 689
 690         /* don't resend recovery messages */
 691         if (msg->msg_recovery) {
 692                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 693                         libcfs_nidstr(&msg->msg_from),
 694                         libcfs_nidstr(&msg->msg_target.nid),
 695                         msg->msg_retry_count);
 696                 return -ENOTRECOVERABLE;
 697         }
 698
 699         /*
 700          * if we explicitly indicated we don't want to resend then just
 701          * return
 702          */
 703         if (msg->msg_no_resend) {
 704                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 705                         libcfs_nidstr(&msg->msg_from),
 706                         libcfs_nidstr(&msg->msg_target.nid),
 707                         msg->msg_retry_count);
 708                 return -ENOTRECOVERABLE;
 709         }
 710
 711         /* check if the message has exceeded the number of retries */
 712         if (msg->msg_retry_count >= lnet_retry_count) {
 713                 CNETERR("msg %s->%s exceeded retry count %d\n",
 714                         libcfs_nidstr(&msg->msg_from),
 715                         libcfs_nidstr(&msg->msg_target.nid),
 716                         msg->msg_retry_count);
 717                 return -ENOTRECOVERABLE;
 718         }
 719
 720         cpt = msg->msg_tx_cpt;
 721         lnet_net_lock(cpt);
 722
 723         /* check again under lock */
 724         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 725                 lnet_net_unlock(cpt);
 726                 return -ESHUTDOWN;
 727         }
 728
 729         container = the_lnet.ln_msg_containers[cpt];
 730         my_slot =
 731                 lnet_check_finalize_recursion_locked(msg,
 732                                         &container->msc_resending,
 733                                         container->msc_nfinalizers,
 734                                         container->msc_resenders);
 735
 736         /* enough threads are resending */
 737         if (my_slot == -1) {
 738                 lnet_net_unlock(cpt);
 739                 return 0;
 740         }
 741
 742         while (!list_empty(&container->msc_resending)) {
 743                 msg = list_entry(container->msc_resending.next,
 744                                         struct lnet_msg, msg_list);
 745                 list_del(&msg->msg_list);
 746
 747                 /*
 748                  * resending the message will require us to call
 749                  * lnet_msg_decommit_tx() which will return the credit
 750                  * which this message holds. This could trigger another
 751                  * queued message to be sent. If that message fails and
 752                  * requires a resend we will recurse.
 753                  * But since at this point the slot is taken, the message
 754                  * will be queued in the container and dealt with
 755                  * later. This breaks the recursion.
 756                  */
 757                 lnet_resend_msg_locked(msg);
 758         }
 759
 760         /*
 761          * msc_resenders is an array of process pointers. Each entry holds
 762          * a pointer to the current process operating on the message. An
 763          * array entry is created per CPT. If the array slot is already
 764          * set, then it means that there is a thread on the CPT currently
 765          * resending a message.
 766          * Once the thread finishes clear the slot to enable the thread to
 767          * take on more resend work.
 768          */
 769         container->msc_resenders[my_slot] = NULL;
 770         lnet_net_unlock(cpt);
 771
 772         return 0;
 773 }
 774
 775 /*
 776  * Do a health check on the message:
 777  * return -1 if we're not going to handle the error or
 778  *   if we've reached the maximum number of retries.
 779  *   success case will return -1 as well
 780  * return 0 if it the message is requeued for send
 781  */
 782 static int
 783 lnet_health_check(struct lnet_msg *msg)
 784 {
 785         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 786         struct lnet_peer_ni *lpni;
 787         struct lnet_ni *ni;
 788         bool lo = false;
 789         bool attempt_local_resend;
 790         bool attempt_remote_resend;
 791         bool handle_local_health;
 792         bool handle_remote_health;
 793
 794         /* if we're shutting down no point in handling health. */
 795         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 796                 return -1;
 797
 798         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 799
 800         /*
 801          * if we're sending to the LOLND then the msg_txpeer will not be
 802          * set. So no need to sanity check it.
 803          */
 804         if (msg->msg_tx_committed &&
 805             !nid_is_lo0(&msg->msg_txni->ni_nid))
 806                 LASSERT(msg->msg_txpeer);
 807         else if (msg->msg_tx_committed &&
 808                  nid_is_lo0(&msg->msg_txni->ni_nid))
 809                 lo = true;
 810
 811         if (hstatus != LNET_MSG_STATUS_OK &&
 812             ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
 813                 return -1;
 814
 815         /*
 816          * always prefer txni/txpeer if they message is committed for both
 817          * directions.
 818          */
 819         if (msg->msg_tx_committed) {
 820                 ni = msg->msg_txni;
 821                 lpni = msg->msg_txpeer;
 822                 attempt_local_resend = attempt_remote_resend = true;
 823         } else {
 824                 ni = msg->msg_rxni;
 825                 lpni = msg->msg_rxpeer;
 826                 attempt_local_resend = attempt_remote_resend = false;
 827         }
 828
 829         if (!lo)
 830                 LASSERT(ni && lpni);
 831         else
 832                 LASSERT(ni);
 833
 834         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 835                libcfs_nidstr(&ni->ni_nid),
 836                (lo) ? "self" : libcfs_nidstr(&lpni->lpni_nid),
 837                lnet_msgtyp2str(msg->msg_type),
 838                lnet_health_error2str(hstatus));
 839
 840         /*
 841          * stats are only incremented for errors so avoid wasting time
 842          * incrementing statistics if there is no error. Similarly, whether to
 843          * update health values or perform resends is only applicable for
 844          * messages with a health status != OK.
 845          */
 846         if (hstatus != LNET_MSG_STATUS_OK) {
 847                 /* Don't further decrement the health value if a recovery
 848                  * message failed.
 849                  */
 850                 if (msg->msg_recovery)
 851                         handle_local_health = handle_remote_health = false;
 852                 else
 853                         handle_local_health = handle_remote_health = true;
 854
 855                 /* For local failures, health/recovery/resends are not needed if
 856                  * I only have a single (non-lolnd) interface. NB: pb_nnis
 857                  * includes the lolnd interface, so a single-rail node would
 858                  * have pb_nnis == 2.
 859                  */
 860                 if (the_lnet.ln_ping_target->pb_nnis <= 2) {
 861                         handle_local_health = false;
 862                         attempt_local_resend = false;
 863                 }
 864
 865                 lnet_net_lock(0);
 866                 lnet_incr_hstats(ni, lpni, hstatus);
 867                 /* For remote failures, health/recovery/resends are not needed
 868                  * if the peer only has a single interface. Special case for
 869                  * routers where we rely on health feature to manage route
 870                  * aliveness. NB: unlike pb_nnis above, lp_nnis does _not_
 871                  * include the lolnd, so a single-rail node would have
 872                  * lp_nnis == 1.
 873                  */
 874                 if (lpni && lpni->lpni_peer_net &&
 875                     lpni->lpni_peer_net->lpn_peer &&
 876                     lpni->lpni_peer_net->lpn_peer->lp_nnis <= 1) {
 877                         attempt_remote_resend = false;
 878                         if (!lnet_isrouter(lpni))
 879                                 handle_remote_health = false;
 880                 }
 881                 lnet_net_unlock(0);
 882         }
 883
 884         switch (hstatus) {
 885         case LNET_MSG_STATUS_OK:
 886                 /*
 887                  * increment the local ni health whether we successfully
 888                  * received or sent a message on it.
 889                  *
 890                  * Ping counts are reset to 0 as appropriate to allow for
 891                  * faster recovery.
 892                  */
 893                 lnet_inc_healthv(&ni->ni_healthv, lnet_health_sensitivity);
 894                 /*
 895                  * It's possible msg_txpeer is NULL in the LOLND
 896                  * case. Only increment the peer's health if we're
 897                  * receiving a message from it. It's the only sure way to
 898                  * know that a remote interface is up.
 899                  * If this interface is part of a router, then take that
 900                  * as indication that the router is fully healthy.
 901                  */
 902                 if (lpni && msg->msg_rx_committed) {
 903                         lnet_net_lock(0);
 904                         lpni->lpni_ping_count = 0;
 905                         ni->ni_ping_count = 0;
 906                         /*
 907                          * If we're receiving a message from the router or
 908                          * I'm a router, then set that lpni's health to
 909                          * maximum so we can commence communication
 910                          */
 911                         if (lnet_isrouter(lpni) || the_lnet.ln_routing) {
 912                                 lnet_set_lpni_healthv_locked(lpni,
 913                                         LNET_MAX_HEALTH_VALUE);
 914                         } else {
 915                                 __u32 sensitivity = lpni->lpni_peer_net->
 916                                         lpn_peer->lp_health_sensitivity;
 917
 918                                 lnet_inc_lpni_healthv_locked(lpni,
 919                                         (sensitivity) ? sensitivity :
 920                                         lnet_health_sensitivity);
 921                                 /* This peer NI may have previously aged out
 922                                  * of recovery. Now that we've received a
 923                                  * message from it, we can continue recovery
 924                                  * if its health value is still below the
 925                                  * maximum.
 926                                  */
 927                                 lnet_peer_ni_add_to_recoveryq_locked(lpni,
 928                                                 &the_lnet.ln_mt_peerNIRecovq,
 929                                                 ktime_get_seconds());
 930                         }
 931                         lnet_net_unlock(0);
 932                 }
 933
 934                 /* we can finalize this message */
 935                 return -1;
 936         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 937         case LNET_MSG_STATUS_LOCAL_DROPPED:
 938         case LNET_MSG_STATUS_LOCAL_ABORTED:
 939         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 940         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 941                 if (handle_local_health)
 942                         lnet_handle_local_failure(ni);
 943                 if (attempt_local_resend)
 944                         return lnet_attempt_msg_resend(msg);
 945                 break;
 946         case LNET_MSG_STATUS_LOCAL_ERROR:
 947                 if (handle_local_health)
 948                         lnet_handle_local_failure(ni);
 949                 return -1;
 950         case LNET_MSG_STATUS_REMOTE_DROPPED:
 951                 if (handle_remote_health)
 952                         lnet_handle_remote_failure(lpni);
 953                 if (attempt_remote_resend)
 954                         return lnet_attempt_msg_resend(msg);
 955                 break;
 956         case LNET_MSG_STATUS_REMOTE_ERROR:
 957         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 958                 if (handle_remote_health)
 959                         lnet_handle_remote_failure(lpni);
 960                 return -1;
 961         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 962                 if (handle_remote_health)
 963                         lnet_handle_remote_failure(lpni);
 964                 if (handle_local_health)
 965                         lnet_handle_local_failure(ni);
 966                 return -1;
 967         default:
 968                 LBUG();
 969         }
 970
 971         /* no resend is needed */
 972         return -1;
 973 }
 974
 975 static void
 976 lnet_msg_detach_md(struct lnet_msg *msg, int status)
 977 {
 978         struct lnet_libmd *md = msg->msg_md;
 979         lnet_handler_t handler = NULL;
 980         int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
 981         int unlink;
 982
 983         lnet_res_lock(cpt);
 984         while (md->md_flags & LNET_MD_FLAG_HANDLING)
 985                 /* An event handler is running - wait for it to
 986                  * complete to avoid races.
 987                  */
 988                 lnet_md_wait_handling(md, cpt);
 989
 990         /* Now it's safe to drop my caller's ref */
 991         md->md_refcount--;
 992         LASSERT(md->md_refcount >= 0);
 993
 994         unlink = lnet_md_unlinkable(md);
 995         if (md->md_handler) {
 996                 if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
 997                         msg->msg_ev.status   = -ETIMEDOUT;
 998                         CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
 999                 } else {
1000                         msg->msg_ev.status   = status;
1001                 }
1002                 msg->msg_ev.unlinked = unlink;
1003                 handler = md->md_handler;
1004                 if (!unlink)
1005                         md->md_flags |= LNET_MD_FLAG_HANDLING;
1006         }
1007
1008         if (unlink || (md->md_refcount == 0 &&
1009                        md->md_threshold == LNET_MD_THRESH_INF))
1010                 lnet_detach_rsp_tracker(md, cpt);
1011
1012         msg->msg_md = NULL;
1013         if (unlink)
1014                 lnet_md_unlink(md);
1015
1016         lnet_res_unlock(cpt);
1017
1018         if (handler) {
1019                 handler(&msg->msg_ev);
1020                 if (!unlink) {
1021                         lnet_res_lock(cpt);
1022                         md->md_flags &= ~LNET_MD_FLAG_HANDLING;
1023                         wake_up_var(md);
1024                         lnet_res_unlock(cpt);
1025                 }
1026         }
1027 }
1028
1029 static bool
1030 lnet_is_health_check(struct lnet_msg *msg)
1031 {
1032         bool hc = true;
1033         int status = msg->msg_ev.status;
1034
1035         if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
1036             !msg->msg_onactivelist) {
1037                 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
1038                        msg);
1039                 return false;
1040         }
1041
1042         if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
1043             (msg->msg_rx_committed && !msg->msg_rxpeer)) {
1044                 /* The optimized GET case does not set msg_rxpeer, but status
1045                  * could be zero. Only print the error message if we have a
1046                  * non-zero status.
1047                  */
1048                 if (status)
1049                         CDEBUG(D_NET, "msg %p status %d cannot retry\n", msg,
1050                                status);
1051                 return false;
1052         }
1053
1054         /* Check for status inconsistencies */
1055         if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
1056              (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
1057                 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
1058                        "checking (%d, %d)\n", msg, status,
1059                        msg->msg_health_status);
1060                 hc = false;
1061         }
1062
1063         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
1064                hc, status, msg->msg_health_status);
1065
1066         return hc;
1067 }
1068
1069 char *
1070 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
1071 {
1072         switch (hstatus) {
1073         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
1074                 return "LOCAL_INTERRUPT";
1075         case LNET_MSG_STATUS_LOCAL_DROPPED:
1076                 return "LOCAL_DROPPED";
1077         case LNET_MSG_STATUS_LOCAL_ABORTED:
1078                 return "LOCAL_ABORTED";
1079         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
1080                 return "LOCAL_NO_ROUTE";
1081         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
1082                 return "LOCAL_TIMEOUT";
1083         case LNET_MSG_STATUS_LOCAL_ERROR:
1084                 return "LOCAL_ERROR";
1085         case LNET_MSG_STATUS_REMOTE_DROPPED:
1086                 return "REMOTE_DROPPED";
1087         case LNET_MSG_STATUS_REMOTE_ERROR:
1088                 return "REMOTE_ERROR";
1089         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
1090                 return "REMOTE_TIMEOUT";
1091         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
1092                 return "NETWORK_TIMEOUT";
1093         case LNET_MSG_STATUS_OK:
1094                 return "OK";
1095         default:
1096                 return "<UNKNOWN>";
1097         }
1098 }
1099
1100 bool
1101 lnet_send_error_simulation(struct lnet_msg *msg,
1102                            enum lnet_msg_hstatus *hstatus)
1103 {
1104         if (!msg)
1105                 return false;
1106
1107         if (list_empty(&the_lnet.ln_drop_rules))
1108             return false;
1109
1110         /* match only health rules */
1111         if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
1112                                   hstatus))
1113                 return false;
1114
1115         CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
1116                 libcfs_nidstr(&msg->msg_hdr.src_nid),
1117                 libcfs_nidstr(&msg->msg_txni->ni_nid),
1118                 libcfs_nidstr(&msg->msg_hdr.dest_nid),
1119                 lnet_msgtyp2str(msg->msg_type),
1120                 lnet_health_error2str(*hstatus));
1121
1122         return true;
1123 }
1124 EXPORT_SYMBOL(lnet_send_error_simulation);
1125
1126 void
1127 lnet_finalize(struct lnet_msg *msg, int status)
1128 {
1129         struct lnet_msg_container *container;
1130         int my_slot;
1131         int cpt;
1132         int rc;
1133
1134         LASSERT(!in_interrupt());
1135
1136         if (msg == NULL)
1137                 return;
1138
1139         msg->msg_ev.status = status;
1140
1141         if (lnet_is_health_check(msg)) {
1142                 /*
1143                  * Check the health status of the message. If it has one
1144                  * of the errors that we're supposed to handle, and it has
1145                  * not timed out, then
1146                  *      1. Decrement the appropriate health_value
1147                  *      2. queue the message on the resend queue
1148
1149                  * if the message send is success, timed out or failed in the
1150                  * health check for any reason then we'll just finalize the
1151                  * message. Otherwise just return since the message has been
1152                  * put on the resend queue.
1153                  */
1154                 if (!lnet_health_check(msg))
1155                         return;
1156         }
1157
1158         /*
1159          * We're not going to resend this message so detach its MD and invoke
1160          * the appropriate callbacks
1161          */
1162         if (msg->msg_md != NULL)
1163                 lnet_msg_detach_md(msg, status);
1164
1165 again:
1166         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
1167                 /* not committed to network yet */
1168                 LASSERT(!msg->msg_onactivelist);
1169                 lnet_msg_free(msg);
1170                 return;
1171         }
1172
1173         /*
1174          * NB: routed message can be committed for both receiving and sending,
1175          * we should finalize in LIFO order and keep counters correct.
1176          * (finalize sending first then finalize receiving)
1177          */
1178         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
1179         lnet_net_lock(cpt);
1180
1181         container = the_lnet.ln_msg_containers[cpt];
1182
1183         /* Recursion breaker.  Don't complete the message here if I am (or
1184          * enough other threads are) already completing messages */
1185         my_slot = lnet_check_finalize_recursion_locked(msg,
1186                                                 &container->msc_finalizing,
1187                                                 container->msc_nfinalizers,
1188                                                 container->msc_finalizers);
1189
1190         /* enough threads are resending */
1191         if (my_slot == -1) {
1192                 lnet_net_unlock(cpt);
1193                 return;
1194         }
1195
1196         rc = 0;
1197         while (!list_empty(&container->msc_finalizing)) {
1198                 msg = list_entry(container->msc_finalizing.next,
1199                                  struct lnet_msg, msg_list);
1200
1201                 list_del_init(&msg->msg_list);
1202
1203                 /* NB drops and regains the lnet lock if it actually does
1204                  * anything, so my finalizing friends can chomp along too */
1205                 rc = lnet_complete_msg_locked(msg, cpt);
1206                 if (rc != 0)
1207                         break;
1208         }
1209
1210         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1211                 lnet_net_unlock(cpt);
1212                 lnet_delay_rule_check();
1213                 lnet_net_lock(cpt);
1214         }
1215
1216         container->msc_finalizers[my_slot] = NULL;
1217         lnet_net_unlock(cpt);
1218
1219         if (rc != 0)
1220                 goto again;
1221 }
1222 EXPORT_SYMBOL(lnet_finalize);
1223
1224 void
1225 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1226 {
1227         int     count = 0;
1228
1229         if (container->msc_init == 0)
1230                 return;
1231
1232         while (!list_empty(&container->msc_active)) {
1233                 struct lnet_msg *msg;
1234
1235                 msg  = list_entry(container->msc_active.next,
1236                                   struct lnet_msg, msg_activelist);
1237                 LASSERT(msg->msg_onactivelist);
1238                 msg->msg_onactivelist = 0;
1239                 list_del_init(&msg->msg_activelist);
1240                 lnet_msg_free(msg);
1241                 count++;
1242         }
1243
1244         if (count > 0)
1245                 CERROR("%d active msg on exit\n", count);
1246
1247         if (container->msc_finalizers != NULL) {
1248                 CFS_FREE_PTR_ARRAY(container->msc_finalizers,
1249                                    container->msc_nfinalizers);
1250                 container->msc_finalizers = NULL;
1251         }
1252
1253         if (container->msc_resenders != NULL) {
1254                 CFS_FREE_PTR_ARRAY(container->msc_resenders,
1255                                    container->msc_nfinalizers);
1256                 container->msc_resenders = NULL;
1257         }
1258         container->msc_init = 0;
1259 }
1260
1261 int
1262 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1263 {
1264         int rc = 0;
1265
1266         container->msc_init = 1;
1267
1268         INIT_LIST_HEAD(&container->msc_active);
1269         INIT_LIST_HEAD(&container->msc_finalizing);
1270         INIT_LIST_HEAD(&container->msc_resending);
1271
1272         /* number of CPUs */
1273         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1274         if (container->msc_nfinalizers == 0)
1275                 container->msc_nfinalizers = 1;
1276
1277         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1278                          container->msc_nfinalizers *
1279                          sizeof(*container->msc_finalizers));
1280
1281         if (container->msc_finalizers == NULL) {
1282                 CERROR("Failed to allocate message finalizers\n");
1283                 lnet_msg_container_cleanup(container);
1284                 return -ENOMEM;
1285         }
1286
1287         LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
1288                          container->msc_nfinalizers *
1289                          sizeof(*container->msc_resenders));
1290
1291         if (container->msc_resenders == NULL) {
1292                 CERROR("Failed to allocate message resenders\n");
1293                 lnet_msg_container_cleanup(container);
1294                 return -ENOMEM;
1295         }
1296
1297         return rc;
1298 }
1299
1300 void
1301 lnet_msg_containers_destroy(void)
1302 {
1303         struct lnet_msg_container *container;
1304         int     i;
1305
1306         if (the_lnet.ln_msg_containers == NULL)
1307                 return;
1308
1309         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1310                 lnet_msg_container_cleanup(container);
1311
1312         cfs_percpt_free(the_lnet.ln_msg_containers);
1313         the_lnet.ln_msg_containers = NULL;
1314 }
1315
1316 int
1317 lnet_msg_containers_create(void)
1318 {
1319         struct lnet_msg_container *container;
1320         int     rc;
1321         int     i;
1322
1323         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1324                                                       sizeof(*container));
1325
1326         if (the_lnet.ln_msg_containers == NULL) {
1327                 CERROR("Failed to allocate cpu-partition data for network\n");
1328                 return -ENOMEM;
1329         }
1330
1331         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1332                 rc = lnet_msg_container_setup(container, i);
1333                 if (rc != 0) {
1334                         lnet_msg_containers_destroy();
1335                         return rc;
1336                 }
1337         }
1338
1339         return 0;
1340 }