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, int sensitivity)
 444 {
 445         int h = atomic_read(healthv);
 446
 447         if (h < sensitivity) {
 448                 atomic_set(healthv, 0);
 449         } else {
 450                 h -= 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, lnet_health_sensitivity);
 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         __u32 sensitivity = lnet_health_sensitivity;
 493         __u32 lp_sensitivity;
 494
 495         /* lpni could be NULL if we're in the LOLND case */
 496         if (!lpni)
 497                 return;
 498
 499         /*
 500          * If there is a health sensitivity in the peer then use that
 501          * instead of the globally set one.
 502          */
 503         lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
 504         if (lp_sensitivity)
 505                 sensitivity = lp_sensitivity;
 506
 507         lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
 508         /*
 509          * add the peer NI to the recovery queue if it's not already there
 510          * and it's health value is actually below the maximum. It's
 511          * possible that the sensitivity might be set to 0, and the health
 512          * value will not be reduced. In this case, there is no reason to
 513          * invoke recovery
 514          */
 515         lnet_peer_ni_add_to_recoveryq_locked(lpni);
 516 }
 517
 518 static void
 519 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
 520 {
 521         /* lpni could be NULL if we're in the LOLND case */
 522         if (!lpni)
 523                 return;
 524
 525         lnet_net_lock(0);
 526         /* the mt could've shutdown and cleaned up the queues */
 527         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 528                 lnet_net_unlock(0);
 529                 return;
 530         }
 531         lnet_handle_remote_failure_locked(lpni);
 532         lnet_net_unlock(0);
 533 }
 534
 535 static void
 536 lnet_incr_hstats(struct lnet_msg *msg, enum lnet_msg_hstatus hstatus)
 537 {
 538         struct lnet_ni *ni = msg->msg_txni;
 539         struct lnet_peer_ni *lpni = msg->msg_txpeer;
 540         struct lnet_counters_health *health;
 541
 542         health = &the_lnet.ln_counters[0]->lct_health;
 543
 544         switch (hstatus) {
 545         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 546                 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
 547                 health->lch_local_interrupt_count++;
 548                 break;
 549         case LNET_MSG_STATUS_LOCAL_DROPPED:
 550                 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
 551                 health->lch_local_dropped_count++;
 552                 break;
 553         case LNET_MSG_STATUS_LOCAL_ABORTED:
 554                 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
 555                 health->lch_local_aborted_count++;
 556                 break;
 557         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 558                 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
 559                 health->lch_local_no_route_count++;
 560                 break;
 561         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 562                 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
 563                 health->lch_local_timeout_count++;
 564                 break;
 565         case LNET_MSG_STATUS_LOCAL_ERROR:
 566                 atomic_inc(&ni->ni_hstats.hlt_local_error);
 567                 health->lch_local_error_count++;
 568                 break;
 569         case LNET_MSG_STATUS_REMOTE_DROPPED:
 570                 if (lpni)
 571                         atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
 572                 health->lch_remote_dropped_count++;
 573                 break;
 574         case LNET_MSG_STATUS_REMOTE_ERROR:
 575                 if (lpni)
 576                         atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
 577                 health->lch_remote_error_count++;
 578                 break;
 579         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 580                 if (lpni)
 581                         atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
 582                 health->lch_remote_timeout_count++;
 583                 break;
 584         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 585                 if (lpni)
 586                         atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
 587                 health->lch_network_timeout_count++;
 588                 break;
 589         case LNET_MSG_STATUS_OK:
 590                 break;
 591         default:
 592                 LBUG();
 593         }
 594 }
 595
 596 static void
 597 lnet_resend_msg_locked(struct lnet_msg *msg)
 598 {
 599         msg->msg_retry_count++;
 600
 601         /*
 602          * remove message from the active list and reset it to prepare
 603          * for a resend. Two exceptions to this
 604          *
 605          * 1. the router case. When a message is being routed it is
 606          * committed for rx when received and committed for tx when
 607          * forwarded. We don't want to remove it from the active list, since
 608          * code which handles receiving expects it to remain on the active
 609          * list.
 610          *
 611          * 2. The REPLY case. Reply messages use the same message
 612          * structure for the GET that was received.
 613          */
 614         if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
 615                 list_del_init(&msg->msg_activelist);
 616                 msg->msg_onactivelist = 0;
 617         }
 618         /*
 619          * The msg_target.nid which was originally set
 620          * when calling LNetGet() or LNetPut() might've
 621          * been overwritten if we're routing this message.
 622          * Call lnet_msg_decommit_tx() to return the credit
 623          * this message consumed. The message will
 624          * consume another credit when it gets resent.
 625          */
 626         msg->msg_target.nid = msg->msg_hdr.dest_nid;
 627         lnet_msg_decommit_tx(msg, -EAGAIN);
 628         msg->msg_sending = 0;
 629         msg->msg_receiving = 0;
 630         msg->msg_target_is_router = 0;
 631
 632         CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
 633                libcfs_nid2str(msg->msg_hdr.src_nid),
 634                libcfs_nid2str(msg->msg_hdr.dest_nid),
 635                lnet_msgtyp2str(msg->msg_type),
 636                lnet_health_error2str(msg->msg_health_status), msg);
 637
 638         list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
 639
 640         wake_up(&the_lnet.ln_mt_waitq);
 641 }
 642
 643 int
 644 lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
 645                                      struct list_head *containerq,
 646                                      int nworkers, void **workers)
 647 {
 648         int my_slot = -1;
 649         int i;
 650
 651         list_add_tail(&msg->msg_list, containerq);
 652
 653         for (i = 0; i < nworkers; i++) {
 654                 if (workers[i] == current)
 655                         break;
 656
 657                 if (my_slot < 0 && workers[i] == NULL)
 658                         my_slot = i;
 659         }
 660
 661         if (i < nworkers || my_slot < 0)
 662                 return -1;
 663
 664         workers[my_slot] = current;
 665
 666         return my_slot;
 667 }
 668
 669 int
 670 lnet_attempt_msg_resend(struct lnet_msg *msg)
 671 {
 672         struct lnet_msg_container *container;
 673         int my_slot;
 674         int cpt;
 675
 676         /* we can only resend tx_committed messages */
 677         LASSERT(msg->msg_tx_committed);
 678
 679         /* don't resend recovery messages */
 680         if (msg->msg_recovery) {
 681                 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
 682                         libcfs_nid2str(msg->msg_from),
 683                         libcfs_nid2str(msg->msg_target.nid),
 684                         msg->msg_retry_count);
 685                 return -ENOTRECOVERABLE;
 686         }
 687
 688         /*
 689          * if we explicitly indicated we don't want to resend then just
 690          * return
 691          */
 692         if (msg->msg_no_resend) {
 693                 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
 694                         libcfs_nid2str(msg->msg_from),
 695                         libcfs_nid2str(msg->msg_target.nid),
 696                         msg->msg_retry_count);
 697                 return -ENOTRECOVERABLE;
 698         }
 699
 700         /* check if the message has exceeded the number of retries */
 701         if (msg->msg_retry_count >= lnet_retry_count) {
 702                 CNETERR("msg %s->%s exceeded retry count %d\n",
 703                         libcfs_nid2str(msg->msg_from),
 704                         libcfs_nid2str(msg->msg_target.nid),
 705                         msg->msg_retry_count);
 706                 return -ENOTRECOVERABLE;
 707         }
 708
 709         cpt = msg->msg_tx_cpt;
 710         lnet_net_lock(cpt);
 711
 712         /* check again under lock */
 713         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
 714                 lnet_net_unlock(cpt);
 715                 return -ESHUTDOWN;
 716         }
 717
 718         container = the_lnet.ln_msg_containers[cpt];
 719         my_slot =
 720                 lnet_check_finalize_recursion_locked(msg,
 721                                         &container->msc_resending,
 722                                         container->msc_nfinalizers,
 723                                         container->msc_resenders);
 724
 725         /* enough threads are resending */
 726         if (my_slot == -1) {
 727                 lnet_net_unlock(cpt);
 728                 return 0;
 729         }
 730
 731         while (!list_empty(&container->msc_resending)) {
 732                 msg = list_entry(container->msc_resending.next,
 733                                         struct lnet_msg, msg_list);
 734                 list_del(&msg->msg_list);
 735
 736                 /*
 737                  * resending the message will require us to call
 738                  * lnet_msg_decommit_tx() which will return the credit
 739                  * which this message holds. This could trigger another
 740                  * queued message to be sent. If that message fails and
 741                  * requires a resend we will recurse.
 742                  * But since at this point the slot is taken, the message
 743                  * will be queued in the container and dealt with
 744                  * later. This breaks the recursion.
 745                  */
 746                 lnet_resend_msg_locked(msg);
 747         }
 748
 749         /*
 750          * msc_resenders is an array of process pointers. Each entry holds
 751          * a pointer to the current process operating on the message. An
 752          * array entry is created per CPT. If the array slot is already
 753          * set, then it means that there is a thread on the CPT currently
 754          * resending a message.
 755          * Once the thread finishes clear the slot to enable the thread to
 756          * take on more resend work.
 757          */
 758         container->msc_resenders[my_slot] = NULL;
 759         lnet_net_unlock(cpt);
 760
 761         return 0;
 762 }
 763
 764 /*
 765  * Do a health check on the message:
 766  * return -1 if we're not going to handle the error or
 767  *   if we've reached the maximum number of retries.
 768  *   success case will return -1 as well
 769  * return 0 if it the message is requeued for send
 770  */
 771 static int
 772 lnet_health_check(struct lnet_msg *msg)
 773 {
 774         enum lnet_msg_hstatus hstatus = msg->msg_health_status;
 775         struct lnet_peer_ni *lpni;
 776         struct lnet_ni *ni;
 777         bool lo = false;
 778
 779         /* if we're shutting down no point in handling health. */
 780         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 781                 return -1;
 782
 783         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 784
 785         /*
 786          * if we're sending to the LOLND then the msg_txpeer will not be
 787          * set. So no need to sanity check it.
 788          */
 789         if (msg->msg_tx_committed &&
 790             LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) != LOLND)
 791                 LASSERT(msg->msg_txpeer);
 792         else if (msg->msg_tx_committed &&
 793                  LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) == LOLND)
 794                 lo = true;
 795
 796         if (hstatus != LNET_MSG_STATUS_OK &&
 797             ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
 798                 return -1;
 799
 800         /*
 801          * stats are only incremented for errors so avoid wasting time
 802          * incrementing statistics if there is no error.
 803          */
 804         if (hstatus != LNET_MSG_STATUS_OK) {
 805                 lnet_net_lock(0);
 806                 lnet_incr_hstats(msg, hstatus);
 807                 lnet_net_unlock(0);
 808         }
 809
 810         /*
 811          * always prefer txni/txpeer if they message is committed for both
 812          * directions.
 813          */
 814         if (msg->msg_tx_committed) {
 815                 ni = msg->msg_txni;
 816                 lpni = msg->msg_txpeer;
 817         } else {
 818                 ni = msg->msg_rxni;
 819                 lpni = msg->msg_rxpeer;
 820         }
 821
 822         if (!lo)
 823                 LASSERT(ni && lpni);
 824         else
 825                 LASSERT(ni);
 826
 827         CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
 828                libcfs_nid2str(ni->ni_nid),
 829                (lo) ? "self" : libcfs_nid2str(lpni->lpni_nid),
 830                lnet_msgtyp2str(msg->msg_type),
 831                lnet_health_error2str(hstatus));
 832
 833         switch (hstatus) {
 834         case LNET_MSG_STATUS_OK:
 835                 /*
 836                  * increment the local ni health weather we successfully
 837                  * received or sent a message on it.
 838                  */
 839                 lnet_inc_healthv(&ni->ni_healthv);
 840                 /*
 841                  * It's possible msg_txpeer is NULL in the LOLND
 842                  * case. Only increment the peer's health if we're
 843                  * receiving a message from it. It's the only sure way to
 844                  * know that a remote interface is up.
 845                  * If this interface is part of a router, then take that
 846                  * as indication that the router is fully healthy.
 847                  */
 848                 if (lpni && msg->msg_rx_committed) {
 849                         /*
 850                          * If we're receiving a message from the router or
 851                          * I'm a router, then set that lpni's health to
 852                          * maximum so we can commence communication
 853                          */
 854                         if (lnet_isrouter(lpni) || the_lnet.ln_routing)
 855                                 lnet_set_healthv(&lpni->lpni_healthv,
 856                                                  LNET_MAX_HEALTH_VALUE);
 857                         else
 858                                 lnet_inc_healthv(&lpni->lpni_healthv);
 859                 }
 860
 861                 /* we can finalize this message */
 862                 return -1;
 863         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 864         case LNET_MSG_STATUS_LOCAL_DROPPED:
 865         case LNET_MSG_STATUS_LOCAL_ABORTED:
 866         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 867         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 868                 lnet_handle_local_failure(ni);
 869                 if (msg->msg_tx_committed)
 870                         /* add to the re-send queue */
 871                         return lnet_attempt_msg_resend(msg);
 872                 break;
 873
 874         /*
 875          * These errors will not trigger a resend so simply
 876          * finalize the message
 877          */
 878         case LNET_MSG_STATUS_LOCAL_ERROR:
 879                 lnet_handle_local_failure(ni);
 880                 return -1;
 881
 882         /*
 883          * TODO: since the remote dropped the message we can
 884          * attempt a resend safely.
 885          */
 886         case LNET_MSG_STATUS_REMOTE_DROPPED:
 887                 lnet_handle_remote_failure(lpni);
 888                 if (msg->msg_tx_committed)
 889                         return lnet_attempt_msg_resend(msg);
 890                 break;
 891
 892         case LNET_MSG_STATUS_REMOTE_ERROR:
 893         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 894         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
 895                 lnet_handle_remote_failure(lpni);
 896                 return -1;
 897         default:
 898                 LBUG();
 899         }
 900
 901         /* no resend is needed */
 902         return -1;
 903 }
 904
 905 static void
 906 lnet_msg_detach_md(struct lnet_msg *msg, int cpt, int status)
 907 {
 908         struct lnet_libmd *md = msg->msg_md;
 909         int unlink;
 910
 911         /* Now it's safe to drop my caller's ref */
 912         md->md_refcount--;
 913         LASSERT(md->md_refcount >= 0);
 914
 915         unlink = lnet_md_unlinkable(md);
 916         if (md->md_eq != NULL) {
 917                 if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
 918                         msg->msg_ev.status   = -ETIMEDOUT;
 919                         CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
 920                 } else {
 921                         msg->msg_ev.status   = status;
 922                 }
 923                 msg->msg_ev.unlinked = unlink;
 924                 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
 925         }
 926
 927         if (unlink || (md->md_refcount == 0 &&
 928                        md->md_threshold == LNET_MD_THRESH_INF))
 929                 lnet_detach_rsp_tracker(md, cpt);
 930
 931         if (unlink)
 932                 lnet_md_unlink(md);
 933
 934         msg->msg_md = NULL;
 935 }
 936
 937 static bool
 938 lnet_is_health_check(struct lnet_msg *msg)
 939 {
 940         bool hc = true;
 941         int status = msg->msg_ev.status;
 942
 943         if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
 944             !msg->msg_onactivelist) {
 945                 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
 946                        msg);
 947                 return false;
 948         }
 949
 950         if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
 951             (msg->msg_rx_committed && !msg->msg_rxpeer)) {
 952                 /* The optimized GET case does not set msg_rxpeer, but status
 953                  * could be zero. Only print the error message if we have a
 954                  * non-zero status.
 955                  */
 956                 if (status)
 957                         CDEBUG(D_NET, "msg %p status %d cannot retry\n", msg,
 958                                status);
 959                 return false;
 960         }
 961
 962         /* Check for status inconsistencies */
 963         if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
 964              (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
 965                 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
 966                        "checking (%d, %d)\n", msg, status,
 967                        msg->msg_health_status);
 968                 hc = false;
 969         }
 970
 971         CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
 972                hc, status, msg->msg_health_status);
 973
 974         return hc;
 975 }
 976
 977 char *
 978 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
 979 {
 980         switch (hstatus) {
 981         case LNET_MSG_STATUS_LOCAL_INTERRUPT:
 982                 return "LOCAL_INTERRUPT";
 983         case LNET_MSG_STATUS_LOCAL_DROPPED:
 984                 return "LOCAL_DROPPED";
 985         case LNET_MSG_STATUS_LOCAL_ABORTED:
 986                 return "LOCAL_ABORTED";
 987         case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
 988                 return "LOCAL_NO_ROUTE";
 989         case LNET_MSG_STATUS_LOCAL_TIMEOUT:
 990                 return "LOCAL_TIMEOUT";
 991         case LNET_MSG_STATUS_LOCAL_ERROR:
 992                 return "LOCAL_ERROR";
 993         case LNET_MSG_STATUS_REMOTE_DROPPED:
 994                 return "REMOTE_DROPPED";
 995         case LNET_MSG_STATUS_REMOTE_ERROR:
 996                 return "REMOTE_ERROR";
 997         case LNET_MSG_STATUS_REMOTE_TIMEOUT:
 998                 return "REMOTE_TIMEOUT";
 999         case LNET_MSG_STATUS_NETWORK_TIMEOUT:
1000                 return "NETWORK_TIMEOUT";
1001         case LNET_MSG_STATUS_OK:
1002                 return "OK";
1003         default:
1004                 return "<UNKNOWN>";
1005         }
1006 }
1007
1008 bool
1009 lnet_send_error_simulation(struct lnet_msg *msg,
1010                            enum lnet_msg_hstatus *hstatus)
1011 {
1012         if (!msg)
1013                 return false;
1014
1015         if (list_empty(&the_lnet.ln_drop_rules))
1016             return false;
1017
1018         /* match only health rules */
1019         if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
1020                                   hstatus))
1021                 return false;
1022
1023         CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
1024                 libcfs_nid2str(msg->msg_hdr.src_nid),
1025                 libcfs_nid2str(msg->msg_txni->ni_nid),
1026                 libcfs_nid2str(msg->msg_hdr.dest_nid),
1027                 lnet_msgtyp2str(msg->msg_type),
1028                 lnet_health_error2str(*hstatus));
1029
1030         return true;
1031 }
1032 EXPORT_SYMBOL(lnet_send_error_simulation);
1033
1034 void
1035 lnet_finalize(struct lnet_msg *msg, int status)
1036 {
1037         struct lnet_msg_container *container;
1038         int my_slot;
1039         int cpt;
1040         int rc;
1041
1042         LASSERT(!in_interrupt());
1043
1044         if (msg == NULL)
1045                 return;
1046
1047         msg->msg_ev.status = status;
1048
1049         if (lnet_is_health_check(msg)) {
1050                 /*
1051                  * Check the health status of the message. If it has one
1052                  * of the errors that we're supposed to handle, and it has
1053                  * not timed out, then
1054                  *      1. Decrement the appropriate health_value
1055                  *      2. queue the message on the resend queue
1056
1057                  * if the message send is success, timed out or failed in the
1058                  * health check for any reason then we'll just finalize the
1059                  * message. Otherwise just return since the message has been
1060                  * put on the resend queue.
1061                  */
1062                 if (!lnet_health_check(msg))
1063                         return;
1064         }
1065
1066         /*
1067          * We're not going to resend this message so detach its MD and invoke
1068          * the appropriate callbacks
1069          */
1070         if (msg->msg_md != NULL) {
1071                 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
1072                 lnet_res_lock(cpt);
1073                 lnet_msg_detach_md(msg, cpt, status);
1074                 lnet_res_unlock(cpt);
1075         }
1076
1077 again:
1078         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
1079                 /* not committed to network yet */
1080                 LASSERT(!msg->msg_onactivelist);
1081                 lnet_msg_free(msg);
1082                 return;
1083         }
1084
1085         /*
1086          * NB: routed message can be committed for both receiving and sending,
1087          * we should finalize in LIFO order and keep counters correct.
1088          * (finalize sending first then finalize receiving)
1089          */
1090         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
1091         lnet_net_lock(cpt);
1092
1093         container = the_lnet.ln_msg_containers[cpt];
1094
1095         /* Recursion breaker.  Don't complete the message here if I am (or
1096          * enough other threads are) already completing messages */
1097         my_slot = lnet_check_finalize_recursion_locked(msg,
1098                                                 &container->msc_finalizing,
1099                                                 container->msc_nfinalizers,
1100                                                 container->msc_finalizers);
1101
1102         /* enough threads are resending */
1103         if (my_slot == -1) {
1104                 lnet_net_unlock(cpt);
1105                 return;
1106         }
1107
1108         rc = 0;
1109         while (!list_empty(&container->msc_finalizing)) {
1110                 msg = list_entry(container->msc_finalizing.next,
1111                                  struct lnet_msg, msg_list);
1112
1113                 list_del_init(&msg->msg_list);
1114
1115                 /* NB drops and regains the lnet lock if it actually does
1116                  * anything, so my finalizing friends can chomp along too */
1117                 rc = lnet_complete_msg_locked(msg, cpt);
1118                 if (rc != 0)
1119                         break;
1120         }
1121
1122         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1123                 lnet_net_unlock(cpt);
1124                 lnet_delay_rule_check();
1125                 lnet_net_lock(cpt);
1126         }
1127
1128         container->msc_finalizers[my_slot] = NULL;
1129         lnet_net_unlock(cpt);
1130
1131         if (rc != 0)
1132                 goto again;
1133 }
1134 EXPORT_SYMBOL(lnet_finalize);
1135
1136 void
1137 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1138 {
1139         int     count = 0;
1140
1141         if (container->msc_init == 0)
1142                 return;
1143
1144         while (!list_empty(&container->msc_active)) {
1145                 struct lnet_msg *msg;
1146
1147                 msg  = list_entry(container->msc_active.next,
1148                                   struct lnet_msg, msg_activelist);
1149                 LASSERT(msg->msg_onactivelist);
1150                 msg->msg_onactivelist = 0;
1151                 list_del_init(&msg->msg_activelist);
1152                 lnet_msg_free(msg);
1153                 count++;
1154         }
1155
1156         if (count > 0)
1157                 CERROR("%d active msg on exit\n", count);
1158
1159         if (container->msc_finalizers != NULL) {
1160                 LIBCFS_FREE(container->msc_finalizers,
1161                             container->msc_nfinalizers *
1162                             sizeof(*container->msc_finalizers));
1163                 container->msc_finalizers = NULL;
1164         }
1165
1166         if (container->msc_resenders != NULL) {
1167                 LIBCFS_FREE(container->msc_resenders,
1168                             container->msc_nfinalizers *
1169                             sizeof(*container->msc_resenders));
1170                 container->msc_resenders = NULL;
1171         }
1172         container->msc_init = 0;
1173 }
1174
1175 int
1176 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1177 {
1178         int rc = 0;
1179
1180         container->msc_init = 1;
1181
1182         INIT_LIST_HEAD(&container->msc_active);
1183         INIT_LIST_HEAD(&container->msc_finalizing);
1184         INIT_LIST_HEAD(&container->msc_resending);
1185
1186         /* number of CPUs */
1187         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1188         if (container->msc_nfinalizers == 0)
1189                 container->msc_nfinalizers = 1;
1190
1191         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1192                          container->msc_nfinalizers *
1193                          sizeof(*container->msc_finalizers));
1194
1195         if (container->msc_finalizers == NULL) {
1196                 CERROR("Failed to allocate message finalizers\n");
1197                 lnet_msg_container_cleanup(container);
1198                 return -ENOMEM;
1199         }
1200
1201         LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
1202                          container->msc_nfinalizers *
1203                          sizeof(*container->msc_resenders));
1204
1205         if (container->msc_resenders == NULL) {
1206                 CERROR("Failed to allocate message resenders\n");
1207                 lnet_msg_container_cleanup(container);
1208                 return -ENOMEM;
1209         }
1210
1211         return rc;
1212 }
1213
1214 void
1215 lnet_msg_containers_destroy(void)
1216 {
1217         struct lnet_msg_container *container;
1218         int     i;
1219
1220         if (the_lnet.ln_msg_containers == NULL)
1221                 return;
1222
1223         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1224                 lnet_msg_container_cleanup(container);
1225
1226         cfs_percpt_free(the_lnet.ln_msg_containers);
1227         the_lnet.ln_msg_containers = NULL;
1228 }
1229
1230 int
1231 lnet_msg_containers_create(void)
1232 {
1233         struct lnet_msg_container *container;
1234         int     rc;
1235         int     i;
1236
1237         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1238                                                       sizeof(*container));
1239
1240         if (the_lnet.ln_msg_containers == NULL) {
1241                 CERROR("Failed to allocate cpu-partition data for network\n");
1242                 return -ENOMEM;
1243         }
1244
1245         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1246                 rc = lnet_msg_container_setup(container, i);
1247                 if (rc != 0) {
1248                         lnet_msg_containers_destroy();
1249                         return rc;
1250                 }
1251         }
1252
1253         return 0;
1254 }