lnet/lnet/lib-msg.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/lib-msg.c
  33  *
  34  * Message decoding, parsing and finalizing routines
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <lnet/lib-lnet.h>
  40
  41 void
  42 lnet_build_unlink_event(struct lnet_libmd *md, struct lnet_event *ev)
  43 {
  44         ENTRY;
  45
  46         memset(ev, 0, sizeof(*ev));
  47
  48         ev->status   = 0;
  49         ev->unlinked = 1;
  50         ev->type     = LNET_EVENT_UNLINK;
  51         lnet_md_deconstruct(md, &ev->md);
  52         lnet_md2handle(&ev->md_handle, md);
  53         EXIT;
  54 }
  55
  56 /*
  57  * Don't need any lock, must be called after lnet_commit_md
  58  */
  59 void
  60 lnet_build_msg_event(struct lnet_msg *msg, enum lnet_event_kind ev_type)
  61 {
  62         struct lnet_hdr *hdr = &msg->msg_hdr;
  63         struct lnet_event *ev = &msg->msg_ev;
  64
  65         LASSERT(!msg->msg_routing);
  66
  67         ev->type = ev_type;
  68         ev->msg_type = msg->msg_type;
  69
  70         if (ev_type == LNET_EVENT_SEND) {
  71                 /* event for active message */
  72                 ev->target.nid    = le64_to_cpu(hdr->dest_nid);
  73                 ev->target.pid    = le32_to_cpu(hdr->dest_pid);
  74                 ev->initiator.nid = LNET_NID_ANY;
  75                 ev->initiator.pid = the_lnet.ln_pid;
  76                 ev->source.nid    = LNET_NID_ANY;
  77                 ev->source.pid    = the_lnet.ln_pid;
  78                 ev->sender        = LNET_NID_ANY;
  79         } else {
  80                 /* event for passive message */
  81                 ev->target.pid    = hdr->dest_pid;
  82                 ev->target.nid    = hdr->dest_nid;
  83                 ev->initiator.pid = hdr->src_pid;
  84                 /* Multi-Rail: resolve src_nid to "primary" peer NID */
  85                 ev->initiator.nid = msg->msg_initiator;
  86                 /* Multi-Rail: track source NID. */
  87                 ev->source.pid    = hdr->src_pid;
  88                 ev->source.nid    = hdr->src_nid;
  89                 ev->rlength       = hdr->payload_length;
  90                 ev->sender        = msg->msg_from;
  91                 ev->mlength       = msg->msg_wanted;
  92                 ev->offset        = msg->msg_offset;
  93         }
  94
  95         switch (ev_type) {
  96         default:
  97                 LBUG();
  98
  99         case LNET_EVENT_PUT: /* passive PUT */
 100                 ev->pt_index   = hdr->msg.put.ptl_index;
 101                 ev->match_bits = hdr->msg.put.match_bits;
 102                 ev->hdr_data   = hdr->msg.put.hdr_data;
 103                 return;
 104
 105         case LNET_EVENT_GET: /* passive GET */
 106                 ev->pt_index   = hdr->msg.get.ptl_index;
 107                 ev->match_bits = hdr->msg.get.match_bits;
 108                 ev->hdr_data   = 0;
 109                 return;
 110
 111         case LNET_EVENT_ACK: /* ACK */
 112                 ev->match_bits = hdr->msg.ack.match_bits;
 113                 ev->mlength    = hdr->msg.ack.mlength;
 114                 return;
 115
 116         case LNET_EVENT_REPLY: /* REPLY */
 117                 return;
 118
 119         case LNET_EVENT_SEND: /* active message */
 120                 if (msg->msg_type == LNET_MSG_PUT) {
 121                         ev->pt_index   = le32_to_cpu(hdr->msg.put.ptl_index);
 122                         ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
 123                         ev->offset     = le32_to_cpu(hdr->msg.put.offset);
 124                         ev->mlength    =
 125                         ev->rlength    = le32_to_cpu(hdr->payload_length);
 126                         ev->hdr_data   = le64_to_cpu(hdr->msg.put.hdr_data);
 127
 128                 } else {
 129                         LASSERT(msg->msg_type == LNET_MSG_GET);
 130                         ev->pt_index   = le32_to_cpu(hdr->msg.get.ptl_index);
 131                         ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
 132                         ev->mlength    =
 133                         ev->rlength    = le32_to_cpu(hdr->msg.get.sink_length);
 134                         ev->offset     = le32_to_cpu(hdr->msg.get.src_offset);
 135                         ev->hdr_data   = 0;
 136                 }
 137                 return;
 138         }
 139 }
 140
 141 void
 142 lnet_msg_commit(struct lnet_msg *msg, int cpt)
 143 {
 144         struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
 145         struct lnet_counters_common *common;
 146         s64 timeout_ns;
 147
 148         /* set the message deadline */
 149         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
 150         msg->msg_deadline = ktime_add_ns(ktime_get(), timeout_ns);
 151
 152         /* routed message can be committed for both receiving and sending */
 153         LASSERT(!msg->msg_tx_committed);
 154
 155         if (msg->msg_sending) {
 156                 LASSERT(!msg->msg_receiving);
 157                 msg->msg_tx_cpt = cpt;
 158                 msg->msg_tx_committed = 1;
 159                 if (msg->msg_rx_committed) { /* routed message REPLY */
 160                         LASSERT(msg->msg_onactivelist);
 161                         return;
 162                 }
 163         } else {
 164                 LASSERT(!msg->msg_sending);
 165                 msg->msg_rx_cpt = cpt;
 166                 msg->msg_rx_committed = 1;
 167         }
 168
 169         LASSERT(!msg->msg_onactivelist);
 170
 171         msg->msg_onactivelist = 1;
 172         list_add_tail(&msg->msg_activelist, &container->msc_active);
 173
 174         common = &the_lnet.ln_counters[cpt]->lct_common;
 175         common->lcc_msgs_alloc++;
 176         if (common->lcc_msgs_alloc > common->lcc_msgs_max)
 177                 common->lcc_msgs_max = common->lcc_msgs_alloc;
 178 }
 179
 180 static void
 181 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
 182 {
 183         struct lnet_counters_common *common;
 184         struct lnet_event *ev = &msg->msg_ev;
 185
 186         LASSERT(msg->msg_tx_committed);
 187         if (status != 0)
 188                 goto out;
 189
 190         common = &(the_lnet.ln_counters[msg->msg_tx_cpt]->lct_common);
 191         switch (ev->type) {
 192         default: /* routed message */
 193                 LASSERT(msg->msg_routing);
 194                 LASSERT(msg->msg_rx_committed);
 195                 LASSERT(ev->type == 0);
 196
 197                 common->lcc_route_length += msg->msg_len;
 198                 common->lcc_route_count++;
 199                 goto incr_stats;
 200
 201         case LNET_EVENT_PUT:
 202                 /* should have been decommitted */
 203                 LASSERT(!msg->msg_rx_committed);
 204                 /* overwritten while sending ACK */
 205                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 206                 msg->msg_type = LNET_MSG_PUT; /* fix type */
 207                 break;
 208
 209         case LNET_EVENT_SEND:
 210                 LASSERT(!msg->msg_rx_committed);
 211                 if (msg->msg_type == LNET_MSG_PUT)
 212                         common->lcc_send_length += msg->msg_len;
 213                 break;
 214
 215         case LNET_EVENT_GET:
 216                 LASSERT(msg->msg_rx_committed);
 217                 /* overwritten while sending reply, we should never be
 218                  * here for optimized GET */
 219                 LASSERT(msg->msg_type == LNET_MSG_REPLY);
 220                 msg->msg_type = LNET_MSG_GET; /* fix type */
 221                 break;
 222         }
 223
 224         common->lcc_send_count++;
 225
 226 incr_stats:
 227         if (msg->msg_txpeer)
 228                 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 229                                 msg->msg_type,
 230                                 LNET_STATS_TYPE_SEND);
 231         if (msg->msg_txni)
 232                 lnet_incr_stats(&msg->msg_txni->ni_stats,
 233                                 msg->msg_type,
 234                                 LNET_STATS_TYPE_SEND);
 235  out:
 236         lnet_return_tx_credits_locked(msg);
 237         msg->msg_tx_committed = 0;
 238 }
 239
 240 static void
 241 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
 242 {
 243         struct lnet_counters_common *common;
 244         struct lnet_event *ev = &msg->msg_ev;
 245
 246         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 247         LASSERT(msg->msg_rx_committed);
 248
 249         if (status != 0)
 250                 goto out;
 251
 252         common = &(the_lnet.ln_counters[msg->msg_rx_cpt]->lct_common);
 253         switch (ev->type) {
 254         default:
 255                 LASSERT(ev->type == 0);
 256                 LASSERT(msg->msg_routing);
 257                 goto incr_stats;
 258
 259         case LNET_EVENT_ACK:
 260                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 261                 break;
 262
 263         case LNET_EVENT_GET:
 264                 /* type is "REPLY" if it's an optimized GET on passive side,
 265                  * because optimized GET will never be committed for sending,
 266                  * so message type wouldn't be changed back to "GET" by
 267                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 268                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 269                         msg->msg_type == LNET_MSG_GET);
 270                 common->lcc_send_length += msg->msg_wanted;
 271                 break;
 272
 273         case LNET_EVENT_PUT:
 274                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 275                 break;
 276
 277         case LNET_EVENT_REPLY:
 278                 /* type is "GET" if it's an optimized GET on active side,
 279                  * see details in lnet_create_reply_msg() */
 280                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 281                         msg->msg_type == LNET_MSG_REPLY);
 282                 break;
 283         }
 284
 285         common->lcc_recv_count++;
 286
 287 incr_stats:
 288         if (msg->msg_rxpeer)
 289                 lnet_incr_stats(&msg->msg_rxpeer->lpni_stats,
 290                                 msg->msg_type,
 291                                 LNET_STATS_TYPE_RECV);
 292         if (msg->msg_rxni)
 293                 lnet_incr_stats(&msg->msg_rxni->ni_stats,
 294                                 msg->msg_type,
 295                                 LNET_STATS_TYPE_RECV);
 296         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 297                 common->lcc_recv_length += msg->msg_wanted;
 298
 299  out:
 300         lnet_return_rx_credits_locked(msg);
 301         msg->msg_rx_committed = 0;
 302 }
 303
 304 void
 305 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
 306 {
 307         int     cpt2 = cpt;
 308
 309         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 310         LASSERT(msg->msg_onactivelist);
 311
 312         if (msg->msg_tx_committed) { /* always decommit for sending first */
 313                 LASSERT(cpt == msg->msg_tx_cpt);
 314                 lnet_msg_decommit_tx(msg, status);
 315         }
 316
 317         if (msg->msg_rx_committed) {
 318                 /* forwarding msg committed for both receiving and sending */
 319                 if (cpt != msg->msg_rx_cpt) {
 320                         lnet_net_unlock(cpt);
 321                         cpt2 = msg->msg_rx_cpt;
 322                         lnet_net_lock(cpt2);
 323                 }
 324                 lnet_msg_decommit_rx(msg, status);
 325         }
 326
 327         list_del(&msg->msg_activelist);
 328         msg->msg_onactivelist = 0;
 329
 330         the_lnet.ln_counters[cpt2]->lct_common.lcc_msgs_alloc--;
 331
 332         if (cpt2 != cpt) {
 333                 lnet_net_unlock(cpt2);
 334                 lnet_net_lock(cpt);
 335         }
 336 }
 337
 338 void
 339 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
 340                    unsigned int offset, unsigned int mlen)
 341 {
 342         /* NB: @offset and @len are only useful for receiving */
 343         /* Here, we attach the MD on lnet_msg and mark it busy and
 344          * decrementing its threshold. Come what may, the lnet_msg "owns"
 345          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 346          * signals completion. */
 347         LASSERT(!msg->msg_routing);
 348
 349         msg->msg_md = md;
 350         if (msg->msg_receiving) { /* committed for receiving */
 351                 msg->msg_offset = offset;
 352                 msg->msg_wanted = mlen;
 353         }
 354
 355         md->md_refcount++;
 356         if (md->md_threshold != LNET_MD_THRESH_INF) {
 357                 LASSERT(md->md_threshold > 0);
 358                 md->md_threshold--;
 359         }
 360
 361         /* build umd in event */
 362         lnet_md2handle(&msg->msg_ev.md_handle, md);
 363         lnet_md_deconstruct(md, &msg->msg_ev.md);
 364 }
 365
 366 static int
 367 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
 368 {
 369         struct lnet_handle_wire ack_wmd;
 370         int                rc;
 371         int                status = msg->msg_ev.status;
 372
 373         LASSERT(msg->msg_onactivelist);
 374
 375         if (status == 0 && msg->msg_ack) {
 376                 /* Only send an ACK if the PUT completed successfully */
 377
 378                 lnet_msg_decommit(msg, cpt, 0);
 379
 380                 msg->msg_ack = 0;
 381                 lnet_net_unlock(cpt);
 382
 383                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 384                 LASSERT(!msg->msg_routing);
 385
 386                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 387
 388                 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
 389
 390                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 391                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 392                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 393
 394                 /* NB: we probably want to use NID of msg::msg_from as 3rd
 395                  * parameter (router NID) if it's routed message */
 396                 rc = lnet_send(msg->msg_ev.target.nid, msg, msg->msg_from);
 397
 398                 lnet_net_lock(cpt);
 399                 /*
 400                  * NB: message is committed for sending, we should return
 401                  * on success because LND will finalize this message later.
 402                  *
 403                  * Also, there is possibility that message is committed for
 404                  * sending and also failed before delivering to LND,
 405                  * i.e: ENOMEM, in that case we can't fall through either
 406                  * because CPT for sending can be different with CPT for
 407                  * receiving, so we should return back to lnet_finalize()
 408                  * to make sure we are locking the correct partition.
 409                  */
 410                 return rc;
 411
 412         } else if (status == 0 &&       /* OK so far */
 413                    (msg->msg_routing && !msg->msg_sending)) {
 414                 /* not forwarded */
 415                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 416                 lnet_net_unlock(cpt);
 417
 418                 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
 419
 420                 lnet_net_lock(cpt);
 421                 /*
 422                  * NB: message is committed for sending, we should return
 423                  * on success because LND will finalize this message later.
 424                  *
 425                  * Also, there is possibility that message is committed for
 426                  * sending and also failed before delivering to LND,
 427                  * i.e: ENOMEM, in that case we can't fall through either:
 428                  * - The rule is message must decommit for sending first if
 429                  *   the it's committed for both sending and receiving
 430                  * - CPT for sending can be different with CPT for receiving,
 431                  *   so we should return back to lnet_finalize() to make
 432                  *   sure we are locking the correct partition.
 433                  */
 434                 return rc;
 435         }
 436
 437         lnet_msg_decommit(msg, cpt, status);
 438         lnet_msg_free(msg);
 439         return 0;
 440 }
 441
 442 static void
 443 lnet_dec_healthv_locked(atomic_t *healthv)
 444 {
 445         int h = atomic_read(healthv);
 446
 447         if (h < lnet_health_sensitivity) {
 448                 atomic_set(healthv, 0);
 449         } else {
 450                 h -= lnet_health_sensitivity;
 451                 atomic_set(healthv, h);
 452         }
 453 }
 454
 455 static void
 456 lnet_handle_local_failure(struct lnet_msg *msg)
 457 {
 458         struct lnet_ni *local_ni;
 459
 460         local_ni = msg->msg_txni;
 461
 462         /*
 463          * the lnet_net_lock(0) is used to protect the addref on the ni
 464          * and the recovery queue.
 465          */
 466         lnet_net_lock(0);
 467         /* the mt could've shutdown and cleaned up the queues */
 468         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 469                 lnet_net_unlock(0);
 470                 return;
 471         }
 472
 473         lnet_dec_healthv_locked(&local_ni->ni_healthv);
 474         /*
 475          * add the NI to the recovery queue if it's not already there
 476          * and it's health value is actually below the maximum. It's
 477          * possible that the sensitivity might be set to 0, and the health
 478          * value will not be reduced. In this case, there is no reason to
 479          * invoke recovery
 480          */
 481         if (list_empty(&local_ni->ni_recovery) &&
 482             atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
 483                 CERROR("ni %s added to recovery queue. Health = %d\n",
 484                         libcfs_nid2str(local_ni->ni_nid),
 485                         atomic_read(&local_ni->ni_healthv));
 486                 list_add_tail(&local_ni->ni_recovery,
 487                               &the_lnet.ln_mt_localNIRecovq);
 488                 lnet_ni_addref_locked(local_ni, 0);
 489         }
 490         lnet_net_unlock(0);
 491 }
 492
 493 void
 494 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
 495 {
 496         /* lpni could be NULL if we're in the LOLND case */
 497         if (!lpni)
 498                 return;
 499
 500         lnet_dec_healthv_locked(&lpni->lpni_healthv);
 501         /*
 502          * add the peer NI to the recovery queue if it's not already there
 503          * and it's health value is actually below the maximum. It's
 504          * possible that the sensitivity might be set to 0, and the health
 505          * value will not be reduced. In this case, there is no reason to
 506          * invoke recovery
 507          */
 508         lnet_peer_ni_add_to_recoveryq_locked(lpni);
 509 }
 510
 511 static void
 512 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 513 {
 514         /* lpni could be NULL if we're in the LOLND case */
 515         if (!lpni)
 516                 return;
 517
 518         lnet_net_lock(0);
 519         /* the mt could've shutdown and cleaned up the queues */
 520         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 521                 lnet_net_unlock(0);
 522                 return;
 523         }
 524         lnet_handle_remote_failure_locked(lpni);
 525         lnet_net_unlock(0);
 526 }
 527
 528 static void
 529 lnet_incr_hstats(struct lnet_msg *msg, enum lnet_msg_hstatus hstatus)
 530 {
 531         struct lnet_ni *ni = msg->msg_txni;
 532         struct lnet_peer_ni *lpni = msg->msg_txpeer;
 533         struct lnet_counters_health *health;
 534
 535         health = &the_lnet.ln_counters[0]->lct_health;
 536
 537         switch (hstatus) {
 538         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 539                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 540                 health->lch_local_interrupt_count++;
 541                 break;
 542         case LNET_MSG_STATUS_LOCAL_DROPPED:
 543                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 544                 health->lch_local_dropped_count++;
 545                 break;
 546         case LNET_MSG_STATUS_LOCAL_ABORTED:
 547                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 548                 health->lch_local_aborted_count++;
 549                 break;
 550         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 551                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 552                 health->lch_local_no_route_count++;
 553                 break;
 554         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 555                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 556                 health->lch_local_timeout_count++;
 557                 break;
 558         case LNET_MSG_STATUS_LOCAL_ERROR:
 559                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 560                 health->lch_local_error_count++;
 561                 break;
 562         case LNET_MSG_STATUS_REMOTE_DROPPED:
 563                 if (lpni)
 564                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 565                 health->lch_remote_dropped_count++;
 566                 break;
 567         case LNET_MSG_STATUS_REMOTE_ERROR:
 568                 if (lpni)
 569                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 570                 health->lch_remote_error_count++;
 571                 break;
 572         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 573                 if (lpni)
 574                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 575                 health->lch_remote_timeout_count++;
 576                 break;
 577         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 578                 if (lpni)
 579                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 580                 health->lch_network_timeout_count++;
 581                 break;
 582         case LNET_MSG_STATUS_OK:
 583                 break;
 584         default:
 585                 LBUG();
 586         }
 587 }
 588
 589 /*
 590  * Do a health check on the message:
 591  * return -1 if we're not going to handle the error or
 592  *   if we've reached the maximum number of retries.
 593  *   success case will return -1 as well
 594  * return 0 if it the message is requeued for send
 595  */
 596 static int
 597 lnet_health_check(struct lnet_msg *msg)
 598 {
 599         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 600         bool lo = false;
 601
 602         /* if we're shutting down no point in handling health. */
 603         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 604                 return -1;
 605
 606         LASSERT(msg->msg_txni);
 607
 608         /*
 609          * if we're sending to the LOLND then the msg_txpeer will not be
 610          * set. So no need to sanity check it.
 611          */
 612         if (LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) != LOLND)
 613                 LASSERT(msg->msg_txpeer);
 614         else
 615                 lo = true;
 616
 617         if (hstatus != LNET_MSG_STATUS_OK &&
 618             ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
 619                 return -1;
 620
 621         /*
 622          * stats are only incremented for errors so avoid wasting time
 623          * incrementing statistics if there is no error.
 624          */
 625         if (hstatus != LNET_MSG_STATUS_OK) {
 626                 lnet_net_lock(0);
 627                 lnet_incr_hstats(msg, hstatus);
 628                 lnet_net_unlock(0);
 629         }
 630
 631         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 632                libcfs_nid2str(msg->msg_txni->ni_nid),
 633                (lo) ? "self" : libcfs_nid2str(msg->msg_txpeer->lpni_nid),
 634                lnet_msgtyp2str(msg->msg_type),
 635                lnet_health_error2str(hstatus));
 636
 637         switch (hstatus) {
 638         case LNET_MSG_STATUS_OK:
 639                 lnet_inc_healthv(&msg->msg_txni->ni_healthv);
 640                 /*
 641                  * It's possible msg_txpeer is NULL in the LOLND
 642                  * case.
 643                  */
 644                 if (msg->msg_txpeer)
 645                         lnet_inc_healthv(&msg->msg_txpeer->lpni_healthv);
 646
 647                 /* we can finalize this message */
 648                 return -1;
 649         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 650         case LNET_MSG_STATUS_LOCAL_DROPPED:
 651         case LNET_MSG_STATUS_LOCAL_ABORTED:
 652         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 653         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 654                 lnet_handle_local_failure(msg);
 655                 /* add to the re-send queue */
 656                 goto resend;
 657
 658         /*
 659          * These errors will not trigger a resend so simply
 660          * finalize the message
 661          */
 662         case LNET_MSG_STATUS_LOCAL_ERROR:
 663                 lnet_handle_local_failure(msg);
 664                 return -1;
 665
 666         /*
 667          * TODO: since the remote dropped the message we can
 668          * attempt a resend safely.
 669          */
 670         case LNET_MSG_STATUS_REMOTE_DROPPED:
 671                 lnet_handle_remote_failure(msg->msg_txpeer);
 672                 goto resend;
 673
 674         case LNET_MSG_STATUS_REMOTE_ERROR:
 675         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 676         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 677                 lnet_handle_remote_failure(msg->msg_txpeer);
 678                 return -1;
 679         default:
 680                 LBUG();
 681         }
 682
 683 resend:
 684         /* don't resend recovery messages */
 685         if (msg->msg_recovery) {
 686                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 687                         libcfs_nid2str(msg->msg_from),
 688                         libcfs_nid2str(msg->msg_target.nid),
 689                         msg->msg_retry_count);
 690                 return -1;
 691         }
 692
 693         /*
 694          * if we explicitly indicated we don't want to resend then just
 695          * return
 696          */
 697         if (msg->msg_no_resend) {
 698                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 699                         libcfs_nid2str(msg->msg_from),
 700                         libcfs_nid2str(msg->msg_target.nid),
 701                         msg->msg_retry_count);
 702                 return -1;
 703         }
 704
 705         /* check if the message has exceeded the number of retries */
 706         if (msg->msg_retry_count >= lnet_retry_count) {
 707                 CNETERR("msg %s->%s exceeded retry count %d\n",
 708                         libcfs_nid2str(msg->msg_from),
 709                         libcfs_nid2str(msg->msg_target.nid),
 710                         msg->msg_retry_count);
 711                 return -1;
 712         }
 713         msg->msg_retry_count++;
 714
 715         lnet_net_lock(msg->msg_tx_cpt);
 716
 717         /* check again under lock */
 718         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 719                 lnet_net_unlock(msg->msg_tx_cpt);
 720                 return -1;
 721         }
 722
 723         /*
 724          * remove message from the active list and reset it in preparation
 725          * for a resend. Two exception to this
 726          *
 727          * 1. the router case, whe a message is committed for rx when
 728          * received, then tx when it is sent. When committed to both tx and
 729          * rx we don't want to remove it from the active list.
 730          *
 731          * 2. The REPLY case since it uses the same msg block for the GET
 732          * that was received.
 733          */
 734         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 735                 list_del_init(&msg->msg_activelist);
 736                 msg->msg_onactivelist = 0;
 737         }
 738         /*
 739          * The msg_target.nid which was originally set
 740          * when calling LNetGet() or LNetPut() might've
 741          * been overwritten if we're routing this message.
 742          * Call lnet_return_tx_credits_locked() to return
 743          * the credit this message consumed. The message will
 744          * consume another credit when it gets resent.
 745          */
 746         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 747         lnet_msg_decommit_tx(msg, -EAGAIN);
 748         msg->msg_sending = 0;
 749         msg->msg_receiving = 0;
 750         msg->msg_target_is_router = 0;
 751
 752         CDEBUG(D_NET, "%s->%s:%s:%s - queuing for resend\n",
 753                libcfs_nid2str(msg->msg_hdr.src_nid),
 754                libcfs_nid2str(msg->msg_hdr.dest_nid),
 755                lnet_msgtyp2str(msg->msg_type),
 756                lnet_health_error2str(hstatus));
 757
 758         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 759         lnet_net_unlock(msg->msg_tx_cpt);
 760
 761         wake_up(&the_lnet.ln_mt_waitq);
 762         return 0;
 763 }
 764
 765 static void
 766 lnet_msg_detach_md(struct lnet_msg *msg, int cpt, int status)
 767 {
 768         struct lnet_libmd *md = msg->msg_md;
 769         int unlink;
 770
 771         /* Now it's safe to drop my caller's ref */
 772         md->md_refcount--;
 773         LASSERT(md->md_refcount >= 0);
 774
 775         unlink = lnet_md_unlinkable(md);
 776         if (md->md_eq != NULL) {
 777                 msg->msg_ev.status   = status;
 778                 msg->msg_ev.unlinked = unlink;
 779                 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
 780         }
 781
 782         if (unlink) {
 783                 lnet_detach_rsp_tracker(md, cpt);
 784                 lnet_md_unlink(md);
 785         }
 786
 787         msg->msg_md = NULL;
 788 }
 789
 790 static bool
 791 lnet_is_health_check(struct lnet_msg *msg)
 792 {
 793         bool hc;
 794         int status = msg->msg_ev.status;
 795
 796         /*
 797          * perform a health check for any message committed for transmit
 798          */
 799         hc = msg->msg_tx_committed;
 800
 801         /* Check for status inconsistencies */
 802         if (hc &&
 803             ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
 804              (status && msg->msg_health_status == LNET_MSG_STATUS_OK))) {
 805                 CERROR("Msg is in inconsistent state, don't perform health "
 806                        "checking (%d, %d)\n", status, msg->msg_health_status);
 807                 hc = false;
 808         }
 809
 810         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
 811                hc, status, msg->msg_health_status);
 812
 813         return hc;
 814 }
 815
 816 char *
 817 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
 818 {
 819         switch (hstatus) {
 820         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 821                 return "LOCAL_INTERRUPT";
 822         case LNET_MSG_STATUS_LOCAL_DROPPED:
 823                 return "LOCAL_DROPPED";
 824         case LNET_MSG_STATUS_LOCAL_ABORTED:
 825                 return "LOCAL_ABORTED";
 826         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 827                 return "LOCAL_NO_ROUTE";
 828         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 829                 return "LOCAL_TIMEOUT";
 830         case LNET_MSG_STATUS_LOCAL_ERROR:
 831                 return "LOCAL_ERROR";
 832         case LNET_MSG_STATUS_REMOTE_DROPPED:
 833                 return "REMOTE_DROPPED";
 834         case LNET_MSG_STATUS_REMOTE_ERROR:
 835                 return "REMOTE_ERROR";
 836         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 837                 return "REMOTE_TIMEOUT";
 838         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 839                 return "NETWORK_TIMEOUT";
 840         case LNET_MSG_STATUS_OK:
 841                 return "OK";
 842         default:
 843                 return "<UNKNOWN>";
 844         }
 845 }
 846
 847 bool
 848 lnet_send_error_simulation(struct lnet_msg *msg,
 849                            enum lnet_msg_hstatus *hstatus)
 850 {
 851         if (!msg)
 852                 return false;
 853
 854         if (list_empty(&the_lnet.ln_drop_rules))
 855             return false;
 856
 857         /* match only health rules */
 858         if (!lnet_drop_rule_match(&msg->msg_hdr, hstatus))
 859                 return false;
 860
 861         CDEBUG(D_NET, "src %s, dst %s: %s simulate health error: %s\n",
 862                 libcfs_nid2str(msg->msg_hdr.src_nid),
 863                 libcfs_nid2str(msg->msg_hdr.dest_nid),
 864                 lnet_msgtyp2str(msg->msg_type),
 865                 lnet_health_error2str(*hstatus));
 866
 867         return true;
 868 }
 869 EXPORT_SYMBOL(lnet_send_error_simulation);
 870
 871 void
 872 lnet_finalize(struct lnet_msg *msg, int status)
 873 {
 874         struct lnet_msg_container *container;
 875         int my_slot;
 876         int cpt;
 877         int rc;
 878         int i;
 879
 880         LASSERT(!in_interrupt());
 881
 882         if (msg == NULL)
 883                 return;
 884
 885         msg->msg_ev.status = status;
 886
 887         if (lnet_is_health_check(msg)) {
 888                 /*
 889                  * Check the health status of the message. If it has one
 890                  * of the errors that we're supposed to handle, and it has
 891                  * not timed out, then
 892                  *      1. Decrement the appropriate health_value
 893                  *      2. queue the message on the resend queue
 894
 895                  * if the message send is success, timed out or failed in the
 896                  * health check for any reason then we'll just finalize the
 897                  * message. Otherwise just return since the message has been
 898                  * put on the resend queue.
 899                  */
 900                 if (!lnet_health_check(msg))
 901                         /* Message is queued for resend */
 902                         return;
 903         }
 904
 905         /*
 906          * We're not going to resend this message so detach its MD and invoke
 907          * the appropriate callbacks
 908          */
 909         if (msg->msg_md != NULL) {
 910                 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
 911                 lnet_res_lock(cpt);
 912                 lnet_msg_detach_md(msg, cpt, status);
 913                 lnet_res_unlock(cpt);
 914         }
 915
 916 again:
 917         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
 918                 /* not committed to network yet */
 919                 LASSERT(!msg->msg_onactivelist);
 920                 lnet_msg_free(msg);
 921                 return;
 922         }
 923
 924         /*
 925          * NB: routed message can be committed for both receiving and sending,
 926          * we should finalize in LIFO order and keep counters correct.
 927          * (finalize sending first then finalize receiving)
 928          */
 929         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
 930         lnet_net_lock(cpt);
 931
 932         container = the_lnet.ln_msg_containers[cpt];
 933         list_add_tail(&msg->msg_list, &container->msc_finalizing);
 934
 935         /* Recursion breaker.  Don't complete the message here if I am (or
 936          * enough other threads are) already completing messages */
 937
 938         my_slot = -1;
 939         for (i = 0; i < container->msc_nfinalizers; i++) {
 940                 if (container->msc_finalizers[i] == current)
 941                         break;
 942
 943                 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
 944                         my_slot = i;
 945         }
 946
 947         if (i < container->msc_nfinalizers || my_slot < 0) {
 948                 lnet_net_unlock(cpt);
 949                 return;
 950         }
 951
 952         container->msc_finalizers[my_slot] = current;
 953
 954         rc = 0;
 955         while (!list_empty(&container->msc_finalizing)) {
 956                 msg = list_entry(container->msc_finalizing.next,
 957                                  struct lnet_msg, msg_list);
 958
 959                 list_del_init(&msg->msg_list);
 960
 961                 /* NB drops and regains the lnet lock if it actually does
 962                  * anything, so my finalizing friends can chomp along too */
 963                 rc = lnet_complete_msg_locked(msg, cpt);
 964                 if (rc != 0)
 965                         break;
 966         }
 967
 968         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
 969                 lnet_net_unlock(cpt);
 970                 lnet_delay_rule_check();
 971                 lnet_net_lock(cpt);
 972         }
 973
 974         container->msc_finalizers[my_slot] = NULL;
 975         lnet_net_unlock(cpt);
 976
 977         if (rc != 0)
 978                 goto again;
 979 }
 980 EXPORT_SYMBOL(lnet_finalize);
 981
 982 void
 983 lnet_msg_container_cleanup(struct lnet_msg_container *container)
 984 {
 985         int     count = 0;
 986
 987         if (container->msc_init == 0)
 988                 return;
 989
 990         while (!list_empty(&container->msc_active)) {
 991                 struct lnet_msg *msg;
 992
 993                 msg  = list_entry(container->msc_active.next,
 994                                   struct lnet_msg, msg_activelist);
 995                 LASSERT(msg->msg_onactivelist);
 996                 msg->msg_onactivelist = 0;
 997                 list_del_init(&msg->msg_activelist);
 998                 lnet_msg_free(msg);
 999                 count++;
1000         }
1001
1002         if (count > 0)
1003                 CERROR("%d active msg on exit\n", count);
1004
1005         if (container->msc_finalizers != NULL) {
1006                 LIBCFS_FREE(container->msc_finalizers,
1007                             container->msc_nfinalizers *
1008                             sizeof(*container->msc_finalizers));
1009                 container->msc_finalizers = NULL;
1010         }
1011         container->msc_init = 0;
1012 }
1013
1014 int
1015 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1016 {
1017         int rc = 0;
1018
1019         container->msc_init = 1;
1020
1021         INIT_LIST_HEAD(&container->msc_active);
1022         INIT_LIST_HEAD(&container->msc_finalizing);
1023
1024         /* number of CPUs */
1025         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1026         if (container->msc_nfinalizers == 0)
1027                 container->msc_nfinalizers = 1;
1028
1029         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1030                          container->msc_nfinalizers *
1031                          sizeof(*container->msc_finalizers));
1032
1033         if (container->msc_finalizers == NULL) {
1034                 CERROR("Failed to allocate message finalizers\n");
1035                 lnet_msg_container_cleanup(container);
1036                 return -ENOMEM;
1037         }
1038
1039         return rc;
1040 }
1041
1042 void
1043 lnet_msg_containers_destroy(void)
1044 {
1045         struct lnet_msg_container *container;
1046         int     i;
1047
1048         if (the_lnet.ln_msg_containers == NULL)
1049                 return;
1050
1051         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1052                 lnet_msg_container_cleanup(container);
1053
1054         cfs_percpt_free(the_lnet.ln_msg_containers);
1055         the_lnet.ln_msg_containers = NULL;
1056 }
1057
1058 int
1059 lnet_msg_containers_create(void)
1060 {
1061         struct lnet_msg_container *container;
1062         int     rc;
1063         int     i;
1064
1065         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1066                                                       sizeof(*container));
1067
1068         if (the_lnet.ln_msg_containers == NULL) {
1069                 CERROR("Failed to allocate cpu-partition data for network\n");
1070                 return -ENOMEM;
1071         }
1072
1073         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1074                 rc = lnet_msg_container_setup(container, i);
1075                 if (rc != 0) {
1076                         lnet_msg_containers_destroy();
1077                         return rc;
1078                 }
1079         }
1080
1081         return 0;
1082 }