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 *counters = the_lnet.ln_counters[cpt];
 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         counters->msgs_alloc++;
 175         if (counters->msgs_alloc > counters->msgs_max)
 176                 counters->msgs_max = counters->msgs_alloc;
 177 }
 178
 179 static void
 180 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
 181 {
 182         struct lnet_counters *counters;
 183         struct lnet_event *ev = &msg->msg_ev;
 184
 185         LASSERT(msg->msg_tx_committed);
 186         if (status != 0)
 187                 goto out;
 188
 189         counters = the_lnet.ln_counters[msg->msg_tx_cpt];
 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                 counters->route_length += msg->msg_len;
 197                 counters->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                         counters->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         counters->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 *counters;
 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         counters = the_lnet.ln_counters[msg->msg_rx_cpt];
 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                 counters->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         counters->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                 counters->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]->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.md);
 363 }
 364
 365 void
 366 lnet_msg_detach_md(struct lnet_msg *msg, int status)
 367 {
 368         struct lnet_libmd *md = msg->msg_md;
 369         int unlink;
 370
 371         /* Now it's safe to drop my caller's ref */
 372         md->md_refcount--;
 373         LASSERT(md->md_refcount >= 0);
 374
 375         unlink = lnet_md_unlinkable(md);
 376         if (md->md_eq != NULL) {
 377                 msg->msg_ev.status   = status;
 378                 msg->msg_ev.unlinked = unlink;
 379                 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
 380         }
 381
 382         if (unlink)
 383                 lnet_md_unlink(md);
 384
 385         msg->msg_md = NULL;
 386 }
 387
 388 static int
 389 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
 390 {
 391         struct lnet_handle_wire ack_wmd;
 392         int                rc;
 393         int                status = msg->msg_ev.status;
 394
 395         LASSERT(msg->msg_onactivelist);
 396
 397         if (status == 0 && msg->msg_ack) {
 398                 /* Only send an ACK if the PUT completed successfully */
 399
 400                 lnet_msg_decommit(msg, cpt, 0);
 401
 402                 msg->msg_ack = 0;
 403                 lnet_net_unlock(cpt);
 404
 405                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 406                 LASSERT(!msg->msg_routing);
 407
 408                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 409
 410                 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
 411
 412                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 413                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 414                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 415
 416                 /* NB: we probably want to use NID of msg::msg_from as 3rd
 417                  * parameter (router NID) if it's routed message */
 418                 rc = lnet_send(msg->msg_ev.target.nid, 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                  * because CPT for sending can be different with CPT for
 429                  * receiving, so we should return back to lnet_finalize()
 430                  * to make sure we are locking the correct partition.
 431                  */
 432                 return rc;
 433
 434         } else if (status == 0 &&       /* OK so far */
 435                    (msg->msg_routing && !msg->msg_sending)) {
 436                 /* not forwarded */
 437                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 438                 lnet_net_unlock(cpt);
 439
 440                 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
 441
 442                 lnet_net_lock(cpt);
 443                 /*
 444                  * NB: message is committed for sending, we should return
 445                  * on success because LND will finalize this message later.
 446                  *
 447                  * Also, there is possibility that message is committed for
 448                  * sending and also failed before delivering to LND,
 449                  * i.e: ENOMEM, in that case we can't fall through either:
 450                  * - The rule is message must decommit for sending first if
 451                  *   the it's committed for both sending and receiving
 452                  * - CPT for sending can be different with CPT for receiving,
 453                  *   so we should return back to lnet_finalize() to make
 454                  *   sure we are locking the correct partition.
 455                  */
 456                 return rc;
 457         }
 458
 459         lnet_msg_decommit(msg, cpt, status);
 460         lnet_msg_free(msg);
 461         return 0;
 462 }
 463
 464 static void
 465 lnet_dec_healthv_locked(atomic_t *healthv)
 466 {
 467         int h = atomic_read(healthv);
 468
 469         if (h < lnet_health_sensitivity) {
 470                 atomic_set(healthv, 0);
 471         } else {
 472                 h -= lnet_health_sensitivity;
 473                 atomic_set(healthv, h);
 474         }
 475 }
 476
 477 static inline void
 478 lnet_inc_healthv(atomic_t *healthv)
 479 {
 480         atomic_add_unless(healthv, 1, LNET_MAX_HEALTH_VALUE);
 481 }
 482
 483 static void
 484 lnet_handle_local_failure(struct lnet_msg *msg)
 485 {
 486         struct lnet_ni *local_ni;
 487
 488         local_ni = msg->msg_txni;
 489
 490         /*
 491          * the lnet_net_lock(0) is used to protect the addref on the ni
 492          * and the recovery queue.
 493          */
 494         lnet_net_lock(0);
 495         /* the mt could've shutdown and cleaned up the queues */
 496         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 497                 lnet_net_unlock(0);
 498                 return;
 499         }
 500
 501         lnet_dec_healthv_locked(&local_ni->ni_healthv);
 502         /*
 503          * add the NI to the recovery queue if it's not already there
 504          * and it's health value is actually below the maximum. It's
 505          * possible that the sensitivity might be set to 0, and the health
 506          * value will not be reduced. In this case, there is no reason to
 507          * invoke recovery
 508          */
 509         if (list_empty(&local_ni->ni_recovery) &&
 510             atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
 511                 CERROR("ni %s added to recovery queue. Health = %d\n",
 512                         libcfs_nid2str(local_ni->ni_nid),
 513                         atomic_read(&local_ni->ni_healthv));
 514                 list_add_tail(&local_ni->ni_recovery,
 515                               &the_lnet.ln_mt_localNIRecovq);
 516                 lnet_ni_addref_locked(local_ni, 0);
 517         }
 518         lnet_net_unlock(0);
 519 }
 520
 521 /*
 522  * Do a health check on the message:
 523  * return -1 if we're not going to handle the error
 524  *   success case will return -1 as well
 525  * return 0 if it the message is requeued for send
 526  */
 527 static int
 528 lnet_health_check(struct lnet_msg *msg)
 529 {
 530         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 531
 532         /* TODO: lnet_incr_hstats(hstatus); */
 533
 534         LASSERT(msg->msg_txni);
 535
 536         if (hstatus != LNET_MSG_STATUS_OK &&
 537             ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
 538                 return -1;
 539
 540         /* if we're shutting down no point in handling health. */
 541         if (the_lnet.ln_state != LNET_STATE_RUNNING)
 542                 return -1;
 543
 544         switch (hstatus) {
 545         case LNET_MSG_STATUS_OK:
 546                 lnet_inc_healthv(&msg->msg_txni->ni_healthv);
 547                 /* we can finalize this message */
 548                 return -1;
 549         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 550         case LNET_MSG_STATUS_LOCAL_DROPPED:
 551         case LNET_MSG_STATUS_LOCAL_ABORTED:
 552         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 553         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 554                 lnet_handle_local_failure(msg);
 555                 /* add to the re-send queue */
 556                 goto resend;
 557
 558         /*
 559                 * TODO: since the remote dropped the message we can
 560                 * attempt a resend safely.
 561                 */
 562         case LNET_MSG_STATUS_REMOTE_DROPPED:
 563         break;
 564
 565         /*
 566                 * These errors will not trigger a resend so simply
 567                 * finalize the message
 568                 */
 569         case LNET_MSG_STATUS_LOCAL_ERROR:
 570                 lnet_handle_local_failure(msg);
 571                 return -1;
 572         case LNET_MSG_STATUS_REMOTE_ERROR:
 573         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 574         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 575                 return -1;
 576         }
 577
 578 resend:
 579         /* don't resend recovery messages */
 580         if (msg->msg_recovery)
 581                 return -1;
 582
 583         /*
 584          * if we explicitly indicated we don't want to resend then just
 585          * return
 586          */
 587         if (msg->msg_no_resend)
 588                 return -1;
 589
 590         lnet_net_lock(msg->msg_tx_cpt);
 591
 592         /*
 593          * remove message from the active list and reset it in preparation
 594          * for a resend. Two exception to this
 595          *
 596          * 1. the router case, whe a message is committed for rx when
 597          * received, then tx when it is sent. When committed to both tx and
 598          * rx we don't want to remove it from the active list.
 599          *
 600          * 2. The REPLY case since it uses the same msg block for the GET
 601          * that was received.
 602          */
 603         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 604                 list_del_init(&msg->msg_activelist);
 605                 msg->msg_onactivelist = 0;
 606         }
 607         /*
 608          * The msg_target.nid which was originally set
 609          * when calling LNetGet() or LNetPut() might've
 610          * been overwritten if we're routing this message.
 611          * Call lnet_return_tx_credits_locked() to return
 612          * the credit this message consumed. The message will
 613          * consume another credit when it gets resent.
 614          */
 615         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 616         lnet_msg_decommit_tx(msg, -EAGAIN);
 617         msg->msg_sending = 0;
 618         msg->msg_receiving = 0;
 619         msg->msg_target_is_router = 0;
 620
 621         CDEBUG(D_NET, "%s->%s:%s:%s - queuing for resend\n",
 622                libcfs_nid2str(msg->msg_hdr.src_nid),
 623                libcfs_nid2str(msg->msg_hdr.dest_nid),
 624                lnet_msgtyp2str(msg->msg_type),
 625                lnet_health_error2str(hstatus));
 626
 627         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 628         lnet_net_unlock(msg->msg_tx_cpt);
 629
 630         wake_up(&the_lnet.ln_mt_waitq);
 631         return 0;
 632 }
 633
 634 static void
 635 lnet_detach_md(struct lnet_msg *msg, int status)
 636 {
 637         int cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
 638
 639         lnet_res_lock(cpt);
 640         lnet_msg_detach_md(msg, status);
 641         lnet_res_unlock(cpt);
 642 }
 643
 644 static bool
 645 lnet_is_health_check(struct lnet_msg *msg)
 646 {
 647         bool hc;
 648         int status = msg->msg_ev.status;
 649
 650         /*
 651          * perform a health check for any message committed for transmit
 652          */
 653         hc = msg->msg_tx_committed;
 654
 655         /* Check for status inconsistencies */
 656         if (hc &&
 657             ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
 658              (status && msg->msg_health_status == LNET_MSG_STATUS_OK))) {
 659                 CERROR("Msg is in inconsistent state, don't perform health "
 660                        "checking (%d, %d)\n", status, msg->msg_health_status);
 661                 hc = false;
 662         }
 663
 664         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
 665                hc, status, msg->msg_health_status);
 666
 667         return hc;
 668 }
 669
 670 char *
 671 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
 672 {
 673         switch (hstatus) {
 674         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 675                 return "LOCAL_INTERRUPT";
 676         case LNET_MSG_STATUS_LOCAL_DROPPED:
 677                 return "LOCAL_DROPPED";
 678         case LNET_MSG_STATUS_LOCAL_ABORTED:
 679                 return "LOCAL_ABORTED";
 680         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 681                 return "LOCAL_NO_ROUTE";
 682         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 683                 return "LOCAL_TIMEOUT";
 684         case LNET_MSG_STATUS_LOCAL_ERROR:
 685                 return "LOCAL_ERROR";
 686         case LNET_MSG_STATUS_REMOTE_DROPPED:
 687                 return "REMOTE_DROPPED";
 688         case LNET_MSG_STATUS_REMOTE_ERROR:
 689                 return "REMOTE_ERROR";
 690         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 691                 return "REMOTE_TIMEOUT";
 692         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 693                 return "NETWORK_TIMEOUT";
 694         case LNET_MSG_STATUS_OK:
 695                 return "OK";
 696         default:
 697                 return "<UNKNOWN>";
 698         }
 699 }
 700
 701 void
 702 lnet_finalize(struct lnet_msg *msg, int status)
 703 {
 704         struct lnet_msg_container *container;
 705         int my_slot;
 706         int cpt;
 707         int rc;
 708         int i;
 709         bool hc;
 710
 711         LASSERT(!in_interrupt());
 712
 713         if (msg == NULL)
 714                 return;
 715
 716         msg->msg_ev.status = status;
 717
 718         /* if the message is successfully sent, no need to keep the MD around */
 719         if (msg->msg_md != NULL && !status)
 720                 lnet_detach_md(msg, status);
 721
 722 again:
 723         hc = lnet_is_health_check(msg);
 724
 725         /*
 726          * the MD would've been detached from the message if it was
 727          * successfully sent. However, if it wasn't successfully sent the
 728          * MD would be around. And since we recalculate whether to
 729          * health check or not, it's possible that we change our minds and
 730          * we don't want to health check this message. In this case also
 731          * free the MD.
 732          *
 733          * If the message is successful we're going to
 734          * go through the lnet_health_check() function, but that'll just
 735          * increment the appropriate health value and return.
 736          */
 737         if (msg->msg_md != NULL && !hc)
 738                 lnet_detach_md(msg, status);
 739
 740         rc = 0;
 741         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
 742                 /* not committed to network yet */
 743                 LASSERT(!msg->msg_onactivelist);
 744                 lnet_msg_free(msg);
 745                 return;
 746         }
 747
 748         if (hc) {
 749                 /*
 750                  * Check the health status of the message. If it has one
 751                  * of the errors that we're supposed to handle, and it has
 752                  * not timed out, then
 753                  *      1. Decrement the appropriate health_value
 754                  *      2. queue the message on the resend queue
 755
 756                  * if the message send is success, timed out or failed in the
 757                  * health check for any reason then we'll just finalize the
 758                  * message. Otherwise just return since the message has been
 759                  * put on the resend queue.
 760                  */
 761                 if (!lnet_health_check(msg))
 762                         return;
 763
 764                 /*
 765                  * if we get here then we need to clean up the md because we're
 766                  * finalizing the message.
 767                 */
 768                 if (msg->msg_md != NULL)
 769                         lnet_detach_md(msg, status);
 770         }
 771
 772         /*
 773          * NB: routed message can be committed for both receiving and sending,
 774          * we should finalize in LIFO order and keep counters correct.
 775          * (finalize sending first then finalize receiving)
 776          */
 777         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
 778         lnet_net_lock(cpt);
 779
 780         container = the_lnet.ln_msg_containers[cpt];
 781         list_add_tail(&msg->msg_list, &container->msc_finalizing);
 782
 783         /* Recursion breaker.  Don't complete the message here if I am (or
 784          * enough other threads are) already completing messages */
 785
 786         my_slot = -1;
 787         for (i = 0; i < container->msc_nfinalizers; i++) {
 788                 if (container->msc_finalizers[i] == current)
 789                         break;
 790
 791                 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
 792                         my_slot = i;
 793         }
 794
 795         if (i < container->msc_nfinalizers || my_slot < 0) {
 796                 lnet_net_unlock(cpt);
 797                 return;
 798         }
 799
 800         container->msc_finalizers[my_slot] = current;
 801
 802         while (!list_empty(&container->msc_finalizing)) {
 803                 msg = list_entry(container->msc_finalizing.next,
 804                                  struct lnet_msg, msg_list);
 805
 806                 list_del_init(&msg->msg_list);
 807
 808                 /* NB drops and regains the lnet lock if it actually does
 809                  * anything, so my finalizing friends can chomp along too */
 810                 rc = lnet_complete_msg_locked(msg, cpt);
 811                 if (rc != 0)
 812                         break;
 813         }
 814
 815         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
 816                 lnet_net_unlock(cpt);
 817                 lnet_delay_rule_check();
 818                 lnet_net_lock(cpt);
 819         }
 820
 821         container->msc_finalizers[my_slot] = NULL;
 822         lnet_net_unlock(cpt);
 823
 824         if (rc != 0)
 825                 goto again;
 826 }
 827 EXPORT_SYMBOL(lnet_finalize);
 828
 829 void
 830 lnet_msg_container_cleanup(struct lnet_msg_container *container)
 831 {
 832         int     count = 0;
 833
 834         if (container->msc_init == 0)
 835                 return;
 836
 837         while (!list_empty(&container->msc_active)) {
 838                 struct lnet_msg *msg;
 839
 840                 msg  = list_entry(container->msc_active.next,
 841                                   struct lnet_msg, msg_activelist);
 842                 LASSERT(msg->msg_onactivelist);
 843                 msg->msg_onactivelist = 0;
 844                 list_del_init(&msg->msg_activelist);
 845                 lnet_msg_free(msg);
 846                 count++;
 847         }
 848
 849         if (count > 0)
 850                 CERROR("%d active msg on exit\n", count);
 851
 852         if (container->msc_finalizers != NULL) {
 853                 LIBCFS_FREE(container->msc_finalizers,
 854                             container->msc_nfinalizers *
 855                             sizeof(*container->msc_finalizers));
 856                 container->msc_finalizers = NULL;
 857         }
 858         container->msc_init = 0;
 859 }
 860
 861 int
 862 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
 863 {
 864         int rc = 0;
 865
 866         container->msc_init = 1;
 867
 868         INIT_LIST_HEAD(&container->msc_active);
 869         INIT_LIST_HEAD(&container->msc_finalizing);
 870
 871         /* number of CPUs */
 872         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
 873         if (container->msc_nfinalizers == 0)
 874                 container->msc_nfinalizers = 1;
 875
 876         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
 877                          container->msc_nfinalizers *
 878                          sizeof(*container->msc_finalizers));
 879
 880         if (container->msc_finalizers == NULL) {
 881                 CERROR("Failed to allocate message finalizers\n");
 882                 lnet_msg_container_cleanup(container);
 883                 return -ENOMEM;
 884         }
 885
 886         return rc;
 887 }
 888
 889 void
 890 lnet_msg_containers_destroy(void)
 891 {
 892         struct lnet_msg_container *container;
 893         int     i;
 894
 895         if (the_lnet.ln_msg_containers == NULL)
 896                 return;
 897
 898         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
 899                 lnet_msg_container_cleanup(container);
 900
 901         cfs_percpt_free(the_lnet.ln_msg_containers);
 902         the_lnet.ln_msg_containers = NULL;
 903 }
 904
 905 int
 906 lnet_msg_containers_create(void)
 907 {
 908         struct lnet_msg_container *container;
 909         int     rc;
 910         int     i;
 911
 912         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
 913                                                       sizeof(*container));
 914
 915         if (the_lnet.ln_msg_containers == NULL) {
 916                 CERROR("Failed to allocate cpu-partition data for network\n");
 917                 return -ENOMEM;
 918         }
 919
 920         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
 921                 rc = lnet_msg_container_setup(container, i);
 922                 if (rc != 0) {
 923                         lnet_msg_containers_destroy();
 924                         return rc;
 925                 }
 926         }
 927
 928         return 0;
 929 }