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_ni *local_ni)
 457 {
 458         /*
 459          * the lnet_net_lock(0) is used to protect the addref on the ni
 460          * and the recovery queue.
 461          */
 462         lnet_net_lock(0);
 463         /* the mt could've shutdown and cleaned up the queues */
 464         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 465                 lnet_net_unlock(0);
 466                 return;
 467         }
 468
 469         lnet_dec_healthv_locked(&local_ni->ni_healthv);
 470         /*
 471          * add the NI to the recovery queue if it's not already there
 472          * and it's health value is actually below the maximum. It's
 473          * possible that the sensitivity might be set to 0, and the health
 474          * value will not be reduced. In this case, there is no reason to
 475          * invoke recovery
 476          */
 477         if (list_empty(&local_ni->ni_recovery) &&
 478             atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
 479                 CERROR("ni %s added to recovery queue. Health = %d\n",
 480                         libcfs_nid2str(local_ni->ni_nid),
 481                         atomic_read(&local_ni->ni_healthv));
 482                 list_add_tail(&local_ni->ni_recovery,
 483                               &the_lnet.ln_mt_localNIRecovq);
 484                 lnet_ni_addref_locked(local_ni, 0);
 485         }
 486         lnet_net_unlock(0);
 487 }
 488
 489 void
 490 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
 491 {
 492         /* lpni could be NULL if we're in the LOLND case */
 493         if (!lpni)
 494                 return;
 495
 496         lnet_dec_healthv_locked(&lpni->lpni_healthv);
 497         /*
 498          * add the peer NI to the recovery queue if it's not already there
 499          * and it's health value is actually below the maximum. It's
 500          * possible that the sensitivity might be set to 0, and the health
 501          * value will not be reduced. In this case, there is no reason to
 502          * invoke recovery
 503          */
 504         lnet_peer_ni_add_to_recoveryq_locked(lpni);
 505 }
 506
 507 static void
 508 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 509 {
 510         /* lpni could be NULL if we're in the LOLND case */
 511         if (!lpni)
 512                 return;
 513
 514         lnet_net_lock(0);
 515         /* the mt could've shutdown and cleaned up the queues */
 516         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 517                 lnet_net_unlock(0);
 518                 return;
 519         }
 520         lnet_handle_remote_failure_locked(lpni);
 521         lnet_net_unlock(0);
 522 }
 523
 524 static void
 525 lnet_incr_hstats(struct lnet_msg *msg, enum lnet_msg_hstatus hstatus)
 526 {
 527         struct lnet_ni *ni = msg->msg_txni;
 528         struct lnet_peer_ni *lpni = msg->msg_txpeer;
 529         struct lnet_counters_health *health;
 530
 531         health = &the_lnet.ln_counters[0]->lct_health;
 532
 533         switch (hstatus) {
 534         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 535                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 536                 health->lch_local_interrupt_count++;
 537                 break;
 538         case LNET_MSG_STATUS_LOCAL_DROPPED:
 539                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 540                 health->lch_local_dropped_count++;
 541                 break;
 542         case LNET_MSG_STATUS_LOCAL_ABORTED:
 543                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 544                 health->lch_local_aborted_count++;
 545                 break;
 546         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 547                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 548                 health->lch_local_no_route_count++;
 549                 break;
 550         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 551                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 552                 health->lch_local_timeout_count++;
 553                 break;
 554         case LNET_MSG_STATUS_LOCAL_ERROR:
 555                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 556                 health->lch_local_error_count++;
 557                 break;
 558         case LNET_MSG_STATUS_REMOTE_DROPPED:
 559                 if (lpni)
 560                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 561                 health->lch_remote_dropped_count++;
 562                 break;
 563         case LNET_MSG_STATUS_REMOTE_ERROR:
 564                 if (lpni)
 565                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 566                 health->lch_remote_error_count++;
 567                 break;
 568         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 569                 if (lpni)
 570                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 571                 health->lch_remote_timeout_count++;
 572                 break;
 573         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 574                 if (lpni)
 575                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 576                 health->lch_network_timeout_count++;
 577                 break;
 578         case LNET_MSG_STATUS_OK:
 579                 break;
 580         default:
 581                 LBUG();
 582         }
 583 }
 584
 585 /*
 586  * Do a health check on the message:
 587  * return -1 if we're not going to handle the error or
 588  *   if we've reached the maximum number of retries.
 589  *   success case will return -1 as well
 590  * return 0 if it the message is requeued for send
 591  */
 592 static int
 593 lnet_health_check(struct lnet_msg *msg)
 594 {
 595         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 596         bool lo = false;
 597         struct lnet_ni *ni;
 598         struct lnet_peer_ni *lpni;
 599
 600         /* if we're shutting down no point in handling health. */
 601         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 602                 return -1;
 603
 604         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 605
 606         /*
 607          * if we're sending to the LOLND then the msg_txpeer will not be
 608          * set. So no need to sanity check it.
 609          */
 610         if (msg->msg_tx_committed &&
 611             LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) != LOLND)
 612                 LASSERT(msg->msg_txpeer);
 613         else if (msg->msg_tx_committed &&
 614                  LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) == LOLND)
 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         /*
 632          * always prefer txni/txpeer if they message is committed for both
 633          * directions.
 634          */
 635         if (msg->msg_tx_committed) {
 636                 ni = msg->msg_txni;
 637                 lpni = msg->msg_txpeer;
 638         } else {
 639                 ni = msg->msg_rxni;
 640                 lpni = msg->msg_rxpeer;
 641         }
 642
 643         if (!lo)
 644                 LASSERT(ni && lpni);
 645         else
 646                 LASSERT(ni);
 647
 648         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 649                libcfs_nid2str(ni->ni_nid),
 650                (lo) ? "self" : libcfs_nid2str(lpni->lpni_nid),
 651                lnet_msgtyp2str(msg->msg_type),
 652                lnet_health_error2str(hstatus));
 653
 654         switch (hstatus) {
 655         case LNET_MSG_STATUS_OK:
 656                 /*
 657                  * increment the local ni health weather we successfully
 658                  * received or sent a message on it.
 659                  */
 660                 lnet_inc_healthv(&ni->ni_healthv);
 661                 /*
 662                  * It's possible msg_txpeer is NULL in the LOLND
 663                  * case. Only increment the peer's health if we're
 664                  * receiving a message from it. It's the only sure way to
 665                  * know that a remote interface is up.
 666                  * If this interface is part of a router, then take that
 667                  * as indication that the router is fully healthy.
 668                  */
 669                 if (lpni && msg->msg_rx_committed) {
 670                         /*
 671                          * If we're receiving a message from the router or
 672                          * I'm a router, then set that lpni's health to
 673                          * maximum so we can commence communication
 674                          */
 675                         if (lnet_isrouter(lpni) || the_lnet.ln_routing)
 676                                 lnet_set_healthv(&lpni->lpni_healthv,
 677                                                  LNET_MAX_HEALTH_VALUE);
 678                         else
 679                                 lnet_inc_healthv(&lpni->lpni_healthv);
 680                 }
 681
 682                 /* we can finalize this message */
 683                 return -1;
 684         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 685         case LNET_MSG_STATUS_LOCAL_DROPPED:
 686         case LNET_MSG_STATUS_LOCAL_ABORTED:
 687         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 688         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 689                 lnet_handle_local_failure(ni);
 690                 if (msg->msg_tx_committed)
 691                         /* add to the re-send queue */
 692                         goto resend;
 693                 break;
 694
 695         /*
 696          * These errors will not trigger a resend so simply
 697          * finalize the message
 698          */
 699         case LNET_MSG_STATUS_LOCAL_ERROR:
 700                 lnet_handle_local_failure(ni);
 701                 return -1;
 702
 703         /*
 704          * TODO: since the remote dropped the message we can
 705          * attempt a resend safely.
 706          */
 707         case LNET_MSG_STATUS_REMOTE_DROPPED:
 708                 lnet_handle_remote_failure(lpni);
 709                 if (msg->msg_tx_committed)
 710                         goto resend;
 711                 break;
 712
 713         case LNET_MSG_STATUS_REMOTE_ERROR:
 714         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 715         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 716                 lnet_handle_remote_failure(lpni);
 717                 return -1;
 718         default:
 719                 LBUG();
 720         }
 721
 722 resend:
 723         /* we can only resend tx_committed messages */
 724         LASSERT(msg->msg_tx_committed);
 725
 726         /* don't resend recovery messages */
 727         if (msg->msg_recovery) {
 728                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 729                         libcfs_nid2str(msg->msg_from),
 730                         libcfs_nid2str(msg->msg_target.nid),
 731                         msg->msg_retry_count);
 732                 return -1;
 733         }
 734
 735         /*
 736          * if we explicitly indicated we don't want to resend then just
 737          * return
 738          */
 739         if (msg->msg_no_resend) {
 740                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 741                         libcfs_nid2str(msg->msg_from),
 742                         libcfs_nid2str(msg->msg_target.nid),
 743                         msg->msg_retry_count);
 744                 return -1;
 745         }
 746
 747         /* check if the message has exceeded the number of retries */
 748         if (msg->msg_retry_count >= lnet_retry_count) {
 749                 CNETERR("msg %s->%s exceeded retry count %d\n",
 750                         libcfs_nid2str(msg->msg_from),
 751                         libcfs_nid2str(msg->msg_target.nid),
 752                         msg->msg_retry_count);
 753                 return -1;
 754         }
 755         msg->msg_retry_count++;
 756
 757         lnet_net_lock(msg->msg_tx_cpt);
 758
 759         /* check again under lock */
 760         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 761                 lnet_net_unlock(msg->msg_tx_cpt);
 762                 return -1;
 763         }
 764
 765         /*
 766          * remove message from the active list and reset it in preparation
 767          * for a resend. Two exception to this
 768          *
 769          * 1. the router case, whe a message is committed for rx when
 770          * received, then tx when it is sent. When committed to both tx and
 771          * rx we don't want to remove it from the active list.
 772          *
 773          * 2. The REPLY case since it uses the same msg block for the GET
 774          * that was received.
 775          */
 776         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 777                 list_del_init(&msg->msg_activelist);
 778                 msg->msg_onactivelist = 0;
 779         }
 780         /*
 781          * The msg_target.nid which was originally set
 782          * when calling LNetGet() or LNetPut() might've
 783          * been overwritten if we're routing this message.
 784          * Call lnet_return_tx_credits_locked() to return
 785          * the credit this message consumed. The message will
 786          * consume another credit when it gets resent.
 787          */
 788         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 789         lnet_msg_decommit_tx(msg, -EAGAIN);
 790         msg->msg_sending = 0;
 791         msg->msg_receiving = 0;
 792         msg->msg_target_is_router = 0;
 793
 794         CDEBUG(D_NET, "%s->%s:%s:%s - queuing for resend\n",
 795                libcfs_nid2str(msg->msg_hdr.src_nid),
 796                libcfs_nid2str(msg->msg_hdr.dest_nid),
 797                lnet_msgtyp2str(msg->msg_type),
 798                lnet_health_error2str(hstatus));
 799
 800         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 801         lnet_net_unlock(msg->msg_tx_cpt);
 802
 803         wake_up(&the_lnet.ln_mt_waitq);
 804         return 0;
 805 }
 806
 807 static void
 808 lnet_msg_detach_md(struct lnet_msg *msg, int cpt, int status)
 809 {
 810         struct lnet_libmd *md = msg->msg_md;
 811         int unlink;
 812
 813         /* Now it's safe to drop my caller's ref */
 814         md->md_refcount--;
 815         LASSERT(md->md_refcount >= 0);
 816
 817         unlink = lnet_md_unlinkable(md);
 818         if (md->md_eq != NULL) {
 819                 msg->msg_ev.status   = status;
 820                 msg->msg_ev.unlinked = unlink;
 821                 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
 822         }
 823
 824         if (unlink) {
 825                 lnet_detach_rsp_tracker(md, cpt);
 826                 lnet_md_unlink(md);
 827         }
 828
 829         msg->msg_md = NULL;
 830 }
 831
 832 static bool
 833 lnet_is_health_check(struct lnet_msg *msg)
 834 {
 835         bool hc = true;
 836         int status = msg->msg_ev.status;
 837
 838         if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
 839             !msg->msg_onactivelist) {
 840                 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
 841                        msg);
 842                 return false;
 843         }
 844
 845         if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
 846             (msg->msg_rx_committed && !msg->msg_rxpeer)) {
 847                 CDEBUG(D_NET, "msg %p failed too early to retry and send\n",
 848                        msg);
 849                 return false;
 850         }
 851
 852         /* Check for status inconsistencies */
 853         if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
 854              (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
 855                 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
 856                        "checking (%d, %d)\n", msg, status,
 857                        msg->msg_health_status);
 858                 hc = false;
 859         }
 860
 861         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
 862                hc, status, msg->msg_health_status);
 863
 864         return hc;
 865 }
 866
 867 char *
 868 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
 869 {
 870         switch (hstatus) {
 871         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 872                 return "LOCAL_INTERRUPT";
 873         case LNET_MSG_STATUS_LOCAL_DROPPED:
 874                 return "LOCAL_DROPPED";
 875         case LNET_MSG_STATUS_LOCAL_ABORTED:
 876                 return "LOCAL_ABORTED";
 877         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 878                 return "LOCAL_NO_ROUTE";
 879         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 880                 return "LOCAL_TIMEOUT";
 881         case LNET_MSG_STATUS_LOCAL_ERROR:
 882                 return "LOCAL_ERROR";
 883         case LNET_MSG_STATUS_REMOTE_DROPPED:
 884                 return "REMOTE_DROPPED";
 885         case LNET_MSG_STATUS_REMOTE_ERROR:
 886                 return "REMOTE_ERROR";
 887         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 888                 return "REMOTE_TIMEOUT";
 889         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 890                 return "NETWORK_TIMEOUT";
 891         case LNET_MSG_STATUS_OK:
 892                 return "OK";
 893         default:
 894                 return "<UNKNOWN>";
 895         }
 896 }
 897
 898 bool
 899 lnet_send_error_simulation(struct lnet_msg *msg,
 900                            enum lnet_msg_hstatus *hstatus)
 901 {
 902         if (!msg)
 903                 return false;
 904
 905         if (list_empty(&the_lnet.ln_drop_rules))
 906             return false;
 907
 908         /* match only health rules */
 909         if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
 910                                   hstatus))
 911                 return false;
 912
 913         CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
 914                 libcfs_nid2str(msg->msg_hdr.src_nid),
 915                 libcfs_nid2str(msg->msg_txni->ni_nid),
 916                 libcfs_nid2str(msg->msg_hdr.dest_nid),
 917                 lnet_msgtyp2str(msg->msg_type),
 918                 lnet_health_error2str(*hstatus));
 919
 920         return true;
 921 }
 922 EXPORT_SYMBOL(lnet_send_error_simulation);
 923
 924 void
 925 lnet_finalize(struct lnet_msg *msg, int status)
 926 {
 927         struct lnet_msg_container *container;
 928         int my_slot;
 929         int cpt;
 930         int rc;
 931         int i;
 932
 933         LASSERT(!in_interrupt());
 934
 935         if (msg == NULL)
 936                 return;
 937
 938         msg->msg_ev.status = status;
 939
 940         if (lnet_is_health_check(msg)) {
 941                 /*
 942                  * Check the health status of the message. If it has one
 943                  * of the errors that we're supposed to handle, and it has
 944                  * not timed out, then
 945                  *      1. Decrement the appropriate health_value
 946                  *      2. queue the message on the resend queue
 947
 948                  * if the message send is success, timed out or failed in the
 949                  * health check for any reason then we'll just finalize the
 950                  * message. Otherwise just return since the message has been
 951                  * put on the resend queue.
 952                  */
 953                 if (!lnet_health_check(msg))
 954                         /* Message is queued for resend */
 955                         return;
 956         }
 957
 958         /*
 959          * We're not going to resend this message so detach its MD and invoke
 960          * the appropriate callbacks
 961          */
 962         if (msg->msg_md != NULL) {
 963                 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
 964                 lnet_res_lock(cpt);
 965                 lnet_msg_detach_md(msg, cpt, status);
 966                 lnet_res_unlock(cpt);
 967         }
 968
 969 again:
 970         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
 971                 /* not committed to network yet */
 972                 LASSERT(!msg->msg_onactivelist);
 973                 lnet_msg_free(msg);
 974                 return;
 975         }
 976
 977         /*
 978          * NB: routed message can be committed for both receiving and sending,
 979          * we should finalize in LIFO order and keep counters correct.
 980          * (finalize sending first then finalize receiving)
 981          */
 982         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
 983         lnet_net_lock(cpt);
 984
 985         container = the_lnet.ln_msg_containers[cpt];
 986         list_add_tail(&msg->msg_list, &container->msc_finalizing);
 987
 988         /* Recursion breaker.  Don't complete the message here if I am (or
 989          * enough other threads are) already completing messages */
 990
 991         my_slot = -1;
 992         for (i = 0; i < container->msc_nfinalizers; i++) {
 993                 if (container->msc_finalizers[i] == current)
 994                         break;
 995
 996                 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
 997                         my_slot = i;
 998         }
 999
1000         if (i < container->msc_nfinalizers || my_slot < 0) {
1001                 lnet_net_unlock(cpt);
1002                 return;
1003         }
1004
1005         container->msc_finalizers[my_slot] = current;
1006
1007         rc = 0;
1008         while (!list_empty(&container->msc_finalizing)) {
1009                 msg = list_entry(container->msc_finalizing.next,
1010                                  struct lnet_msg, msg_list);
1011
1012                 list_del_init(&msg->msg_list);
1013
1014                 /* NB drops and regains the lnet lock if it actually does
1015                  * anything, so my finalizing friends can chomp along too */
1016                 rc = lnet_complete_msg_locked(msg, cpt);
1017                 if (rc != 0)
1018                         break;
1019         }
1020
1021         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1022                 lnet_net_unlock(cpt);
1023                 lnet_delay_rule_check();
1024                 lnet_net_lock(cpt);
1025         }
1026
1027         container->msc_finalizers[my_slot] = NULL;
1028         lnet_net_unlock(cpt);
1029
1030         if (rc != 0)
1031                 goto again;
1032 }
1033 EXPORT_SYMBOL(lnet_finalize);
1034
1035 void
1036 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1037 {
1038         int     count = 0;
1039
1040         if (container->msc_init == 0)
1041                 return;
1042
1043         while (!list_empty(&container->msc_active)) {
1044                 struct lnet_msg *msg;
1045
1046                 msg  = list_entry(container->msc_active.next,
1047                                   struct lnet_msg, msg_activelist);
1048                 LASSERT(msg->msg_onactivelist);
1049                 msg->msg_onactivelist = 0;
1050                 list_del_init(&msg->msg_activelist);
1051                 lnet_msg_free(msg);
1052                 count++;
1053         }
1054
1055         if (count > 0)
1056                 CERROR("%d active msg on exit\n", count);
1057
1058         if (container->msc_finalizers != NULL) {
1059                 LIBCFS_FREE(container->msc_finalizers,
1060                             container->msc_nfinalizers *
1061                             sizeof(*container->msc_finalizers));
1062                 container->msc_finalizers = NULL;
1063         }
1064         container->msc_init = 0;
1065 }
1066
1067 int
1068 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1069 {
1070         int rc = 0;
1071
1072         container->msc_init = 1;
1073
1074         INIT_LIST_HEAD(&container->msc_active);
1075         INIT_LIST_HEAD(&container->msc_finalizing);
1076
1077         /* number of CPUs */
1078         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1079         if (container->msc_nfinalizers == 0)
1080                 container->msc_nfinalizers = 1;
1081
1082         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1083                          container->msc_nfinalizers *
1084                          sizeof(*container->msc_finalizers));
1085
1086         if (container->msc_finalizers == NULL) {
1087                 CERROR("Failed to allocate message finalizers\n");
1088                 lnet_msg_container_cleanup(container);
1089                 return -ENOMEM;
1090         }
1091
1092         return rc;
1093 }
1094
1095 void
1096 lnet_msg_containers_destroy(void)
1097 {
1098         struct lnet_msg_container *container;
1099         int     i;
1100
1101         if (the_lnet.ln_msg_containers == NULL)
1102                 return;
1103
1104         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1105                 lnet_msg_container_cleanup(container);
1106
1107         cfs_percpt_free(the_lnet.ln_msg_containers);
1108         the_lnet.ln_msg_containers = NULL;
1109 }
1110
1111 int
1112 lnet_msg_containers_create(void)
1113 {
1114         struct lnet_msg_container *container;
1115         int     rc;
1116         int     i;
1117
1118         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1119                                                       sizeof(*container));
1120
1121         if (the_lnet.ln_msg_containers == NULL) {
1122                 CERROR("Failed to allocate cpu-partition data for network\n");
1123                 return -ENOMEM;
1124         }
1125
1126         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1127                 rc = lnet_msg_container_setup(container, i);
1128                 if (rc != 0) {
1129                         lnet_msg_containers_destroy();
1130                         return rc;
1131                 }
1132         }
1133
1134         return 0;
1135 }