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, 2014, 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(lnet_libmd_t *md, lnet_event_t *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(lnet_msg_t *msg, lnet_event_kind_t ev_type)
  61 {
  62         lnet_hdr_t      *hdr = &msg->msg_hdr;
  63         lnet_event_t    *ev  = &msg->msg_ev;
  64
  65         LASSERT(!msg->msg_routing);
  66
  67         ev->type = ev_type;
  68
  69         if (ev_type == LNET_EVENT_SEND) {
  70                 /* event for active message */
  71                 ev->target.nid    = le64_to_cpu(hdr->dest_nid);
  72                 ev->target.pid    = le32_to_cpu(hdr->dest_pid);
  73                 ev->initiator.nid = LNET_NID_ANY;
  74                 ev->initiator.pid = the_lnet.ln_pid;
  75                 ev->source.nid    = LNET_NID_ANY;
  76                 ev->source.pid    = the_lnet.ln_pid;
  77                 ev->sender        = LNET_NID_ANY;
  78
  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(lnet_msg_t *msg, int cpt)
 143 {
 144         struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
 145         lnet_counters_t           *counters  = the_lnet.ln_counters[cpt];
 146
 147         /* routed message can be committed for both receiving and sending */
 148         LASSERT(!msg->msg_tx_committed);
 149
 150         if (msg->msg_sending) {
 151                 LASSERT(!msg->msg_receiving);
 152
 153                 msg->msg_tx_cpt = cpt;
 154                 msg->msg_tx_committed = 1;
 155                 if (msg->msg_rx_committed) { /* routed message REPLY */
 156                         LASSERT(msg->msg_onactivelist);
 157                         return;
 158                 }
 159         } else {
 160                 LASSERT(!msg->msg_sending);
 161                 msg->msg_rx_cpt = cpt;
 162                 msg->msg_rx_committed = 1;
 163         }
 164
 165         LASSERT(!msg->msg_onactivelist);
 166         msg->msg_onactivelist = 1;
 167         list_add(&msg->msg_activelist, &container->msc_active);
 168
 169         counters->msgs_alloc++;
 170         if (counters->msgs_alloc > counters->msgs_max)
 171                 counters->msgs_max = counters->msgs_alloc;
 172 }
 173
 174 static void
 175 lnet_msg_decommit_tx(lnet_msg_t *msg, int status)
 176 {
 177         lnet_counters_t *counters;
 178         lnet_event_t    *ev = &msg->msg_ev;
 179
 180         LASSERT(msg->msg_tx_committed);
 181         if (status != 0)
 182                 goto out;
 183
 184         counters = the_lnet.ln_counters[msg->msg_tx_cpt];
 185         switch (ev->type) {
 186         default: /* routed message */
 187                 LASSERT(msg->msg_routing);
 188                 LASSERT(msg->msg_rx_committed);
 189                 LASSERT(ev->type == 0);
 190
 191                 counters->route_length += msg->msg_len;
 192                 counters->route_count++;
 193                 goto out;
 194
 195         case LNET_EVENT_PUT:
 196                 /* should have been decommitted */
 197                 LASSERT(!msg->msg_rx_committed);
 198                 /* overwritten while sending ACK */
 199                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 200                 msg->msg_type = LNET_MSG_PUT; /* fix type */
 201                 break;
 202
 203         case LNET_EVENT_SEND:
 204                 LASSERT(!msg->msg_rx_committed);
 205                 if (msg->msg_type == LNET_MSG_PUT)
 206                         counters->send_length += msg->msg_len;
 207                 break;
 208
 209         case LNET_EVENT_GET:
 210                 LASSERT(msg->msg_rx_committed);
 211                 /* overwritten while sending reply, we should never be
 212                  * here for optimized GET */
 213                 LASSERT(msg->msg_type == LNET_MSG_REPLY);
 214                 msg->msg_type = LNET_MSG_GET; /* fix type */
 215                 break;
 216         }
 217
 218         counters->send_count++;
 219  out:
 220         lnet_return_tx_credits_locked(msg);
 221         msg->msg_tx_committed = 0;
 222 }
 223
 224 static void
 225 lnet_msg_decommit_rx(lnet_msg_t *msg, int status)
 226 {
 227         lnet_counters_t *counters;
 228         lnet_event_t    *ev = &msg->msg_ev;
 229
 230         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 231         LASSERT(msg->msg_rx_committed);
 232
 233         if (status != 0)
 234                 goto out;
 235
 236         counters = the_lnet.ln_counters[msg->msg_rx_cpt];
 237         switch (ev->type) {
 238         default:
 239                 LASSERT(ev->type == 0);
 240                 LASSERT(msg->msg_routing);
 241                 goto out;
 242
 243         case LNET_EVENT_ACK:
 244                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 245                 break;
 246
 247         case LNET_EVENT_GET:
 248                 /* type is "REPLY" if it's an optimized GET on passive side,
 249                  * because optimized GET will never be committed for sending,
 250                  * so message type wouldn't be changed back to "GET" by
 251                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 252                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 253                         msg->msg_type == LNET_MSG_GET);
 254                 counters->send_length += msg->msg_wanted;
 255                 break;
 256
 257         case LNET_EVENT_PUT:
 258                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 259                 break;
 260
 261         case LNET_EVENT_REPLY:
 262                 /* type is "GET" if it's an optimized GET on active side,
 263                  * see details in lnet_create_reply_msg() */
 264                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 265                         msg->msg_type == LNET_MSG_REPLY);
 266                 break;
 267         }
 268
 269         counters->recv_count++;
 270         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 271                 counters->recv_length += msg->msg_wanted;
 272
 273  out:
 274         lnet_return_rx_credits_locked(msg);
 275         msg->msg_rx_committed = 0;
 276 }
 277
 278 void
 279 lnet_msg_decommit(lnet_msg_t *msg, int cpt, int status)
 280 {
 281         int     cpt2 = cpt;
 282
 283         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 284         LASSERT(msg->msg_onactivelist);
 285
 286         if (msg->msg_tx_committed) { /* always decommit for sending first */
 287                 LASSERT(cpt == msg->msg_tx_cpt);
 288                 lnet_msg_decommit_tx(msg, status);
 289         }
 290
 291         if (msg->msg_rx_committed) {
 292                 /* forwarding msg committed for both receiving and sending */
 293                 if (cpt != msg->msg_rx_cpt) {
 294                         lnet_net_unlock(cpt);
 295                         cpt2 = msg->msg_rx_cpt;
 296                         lnet_net_lock(cpt2);
 297                 }
 298                 lnet_msg_decommit_rx(msg, status);
 299         }
 300
 301         list_del(&msg->msg_activelist);
 302         msg->msg_onactivelist = 0;
 303
 304         the_lnet.ln_counters[cpt2]->msgs_alloc--;
 305
 306         if (cpt2 != cpt) {
 307                 lnet_net_unlock(cpt2);
 308                 lnet_net_lock(cpt);
 309         }
 310 }
 311
 312 void
 313 lnet_msg_attach_md(lnet_msg_t *msg, lnet_libmd_t *md,
 314                    unsigned int offset, unsigned int mlen)
 315 {
 316         /* NB: @offset and @len are only useful for receiving */
 317         /* Here, we attach the MD on lnet_msg and mark it busy and
 318          * decrementing its threshold. Come what may, the lnet_msg "owns"
 319          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 320          * signals completion. */
 321         LASSERT(!msg->msg_routing);
 322
 323         msg->msg_md = md;
 324         if (msg->msg_receiving) { /* committed for receiving */
 325                 msg->msg_offset = offset;
 326                 msg->msg_wanted = mlen;
 327         }
 328
 329         md->md_refcount++;
 330         if (md->md_threshold != LNET_MD_THRESH_INF) {
 331                 LASSERT(md->md_threshold > 0);
 332                 md->md_threshold--;
 333         }
 334
 335         /* build umd in event */
 336         lnet_md2handle(&msg->msg_ev.md_handle, md);
 337         lnet_md_deconstruct(md, &msg->msg_ev.md);
 338 }
 339
 340 void
 341 lnet_msg_detach_md(lnet_msg_t *msg, int status)
 342 {
 343         lnet_libmd_t    *md = msg->msg_md;
 344         int             unlink;
 345
 346         /* Now it's safe to drop my caller's ref */
 347         md->md_refcount--;
 348         LASSERT(md->md_refcount >= 0);
 349
 350         unlink = lnet_md_unlinkable(md);
 351         if (md->md_eq != NULL) {
 352                 msg->msg_ev.status   = status;
 353                 msg->msg_ev.unlinked = unlink;
 354                 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
 355         }
 356
 357         if (unlink)
 358                 lnet_md_unlink(md);
 359
 360         msg->msg_md = NULL;
 361 }
 362
 363 static int
 364 lnet_complete_msg_locked(lnet_msg_t *msg, int cpt)
 365 {
 366         struct lnet_handle_wire ack_wmd;
 367         int                rc;
 368         int                status = msg->msg_ev.status;
 369
 370         LASSERT(msg->msg_onactivelist);
 371
 372         if (status == 0 && msg->msg_ack) {
 373                 /* Only send an ACK if the PUT completed successfully */
 374
 375                 lnet_msg_decommit(msg, cpt, 0);
 376
 377                 msg->msg_ack = 0;
 378                 lnet_net_unlock(cpt);
 379
 380                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 381                 LASSERT(!msg->msg_routing);
 382
 383                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 384
 385                 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
 386
 387                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 388                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 389                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 390
 391                 /* NB: we probably want to use NID of msg::msg_from as 3rd
 392                  * parameter (router NID) if it's routed message */
 393                 rc = lnet_send(msg->msg_ev.target.nid, msg, LNET_NID_ANY);
 394
 395                 lnet_net_lock(cpt);
 396                 /*
 397                  * NB: message is committed for sending, we should return
 398                  * on success because LND will finalize this message later.
 399                  *
 400                  * Also, there is possibility that message is committed for
 401                  * sending and also failed before delivering to LND,
 402                  * i.e: ENOMEM, in that case we can't fall through either
 403                  * because CPT for sending can be different with CPT for
 404                  * receiving, so we should return back to lnet_finalize()
 405                  * to make sure we are locking the correct partition.
 406                  */
 407                 return rc;
 408
 409         } else if (status == 0 &&       /* OK so far */
 410                    (msg->msg_routing && !msg->msg_sending)) {
 411                 /* not forwarded */
 412                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 413                 lnet_net_unlock(cpt);
 414
 415                 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
 416
 417                 lnet_net_lock(cpt);
 418                 /*
 419                  * NB: message is committed for sending, we should return
 420                  * on success because LND will finalize this message later.
 421                  *
 422                  * Also, there is possibility that message is committed for
 423                  * sending and also failed before delivering to LND,
 424                  * i.e: ENOMEM, in that case we can't fall through either:
 425                  * - The rule is message must decommit for sending first if
 426                  *   the it's committed for both sending and receiving
 427                  * - CPT for sending can be different with CPT for receiving,
 428                  *   so we should return back to lnet_finalize() to make
 429                  *   sure we are locking the correct partition.
 430                  */
 431                 return rc;
 432         }
 433
 434         lnet_msg_decommit(msg, cpt, status);
 435         lnet_msg_free(msg);
 436         return 0;
 437 }
 438
 439 void
 440 lnet_finalize(lnet_ni_t *ni, lnet_msg_t *msg, int status)
 441 {
 442         struct lnet_msg_container       *container;
 443         int                             my_slot;
 444         int                             cpt;
 445         int                             rc;
 446         int                             i;
 447
 448         LASSERT(!in_interrupt());
 449
 450         if (msg == NULL)
 451                 return;
 452
 453         msg->msg_ev.status = status;
 454
 455         if (msg->msg_md != NULL) {
 456                 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
 457
 458                 lnet_res_lock(cpt);
 459                 lnet_msg_detach_md(msg, status);
 460                 lnet_res_unlock(cpt);
 461         }
 462
 463  again:
 464         rc = 0;
 465         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
 466                 /* not committed to network yet */
 467                 LASSERT(!msg->msg_onactivelist);
 468                 lnet_msg_free(msg);
 469                 return;
 470         }
 471
 472         /*
 473          * NB: routed message can be committed for both receiving and sending,
 474          * we should finalize in LIFO order and keep counters correct.
 475          * (finalize sending first then finalize receiving)
 476          */
 477         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
 478         lnet_net_lock(cpt);
 479
 480         container = the_lnet.ln_msg_containers[cpt];
 481         list_add_tail(&msg->msg_list, &container->msc_finalizing);
 482
 483         /* Recursion breaker.  Don't complete the message here if I am (or
 484          * enough other threads are) already completing messages */
 485
 486         my_slot = -1;
 487         for (i = 0; i < container->msc_nfinalizers; i++) {
 488                 if (container->msc_finalizers[i] == current)
 489                         break;
 490
 491                 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
 492                         my_slot = i;
 493         }
 494
 495         if (i < container->msc_nfinalizers || my_slot < 0) {
 496                 lnet_net_unlock(cpt);
 497                 return;
 498         }
 499
 500         container->msc_finalizers[my_slot] = current;
 501
 502         while (!list_empty(&container->msc_finalizing)) {
 503                 msg = list_entry(container->msc_finalizing.next,
 504                                  lnet_msg_t, msg_list);
 505
 506                 list_del(&msg->msg_list);
 507
 508                 /* NB drops and regains the lnet lock if it actually does
 509                  * anything, so my finalizing friends can chomp along too */
 510                 rc = lnet_complete_msg_locked(msg, cpt);
 511                 if (rc != 0)
 512                         break;
 513         }
 514
 515         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
 516                 lnet_net_unlock(cpt);
 517                 lnet_delay_rule_check();
 518                 lnet_net_lock(cpt);
 519         }
 520
 521         container->msc_finalizers[my_slot] = NULL;
 522         lnet_net_unlock(cpt);
 523
 524         if (rc != 0)
 525                 goto again;
 526 }
 527 EXPORT_SYMBOL(lnet_finalize);
 528
 529 void
 530 lnet_msg_container_cleanup(struct lnet_msg_container *container)
 531 {
 532         int     count = 0;
 533
 534         if (container->msc_init == 0)
 535                 return;
 536
 537         while (!list_empty(&container->msc_active)) {
 538                 lnet_msg_t *msg = list_entry(container->msc_active.next,
 539                                              lnet_msg_t, msg_activelist);
 540
 541                 LASSERT(msg->msg_onactivelist);
 542                 msg->msg_onactivelist = 0;
 543                 list_del(&msg->msg_activelist);
 544                 lnet_msg_free(msg);
 545                 count++;
 546         }
 547
 548         if (count > 0)
 549                 CERROR("%d active msg on exit\n", count);
 550
 551         if (container->msc_finalizers != NULL) {
 552                 LIBCFS_FREE(container->msc_finalizers,
 553                             container->msc_nfinalizers *
 554                             sizeof(*container->msc_finalizers));
 555                 container->msc_finalizers = NULL;
 556         }
 557         container->msc_init = 0;
 558 }
 559
 560 int
 561 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
 562 {
 563         int     rc;
 564
 565         container->msc_init = 1;
 566
 567         INIT_LIST_HEAD(&container->msc_active);
 568         INIT_LIST_HEAD(&container->msc_finalizing);
 569
 570         rc = 0;
 571         /* number of CPUs */
 572         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
 573
 574         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
 575                          container->msc_nfinalizers *
 576                          sizeof(*container->msc_finalizers));
 577
 578         if (container->msc_finalizers == NULL) {
 579                 CERROR("Failed to allocate message finalizers\n");
 580                 lnet_msg_container_cleanup(container);
 581                 return -ENOMEM;
 582         }
 583
 584         return rc;
 585 }
 586
 587 void
 588 lnet_msg_containers_destroy(void)
 589 {
 590         struct lnet_msg_container *container;
 591         int     i;
 592
 593         if (the_lnet.ln_msg_containers == NULL)
 594                 return;
 595
 596         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
 597                 lnet_msg_container_cleanup(container);
 598
 599         cfs_percpt_free(the_lnet.ln_msg_containers);
 600         the_lnet.ln_msg_containers = NULL;
 601 }
 602
 603 int
 604 lnet_msg_containers_create(void)
 605 {
 606         struct lnet_msg_container *container;
 607         int     rc;
 608         int     i;
 609
 610         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
 611                                                       sizeof(*container));
 612
 613         if (the_lnet.ln_msg_containers == NULL) {
 614                 CERROR("Failed to allocate cpu-partition data for network\n");
 615                 return -ENOMEM;
 616         }
 617
 618         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
 619                 rc = lnet_msg_container_setup(container, i);
 620                 if (rc != 0) {
 621                         lnet_msg_containers_destroy();
 622                         return rc;
 623                 }
 624         }
 625
 626         return 0;
 627 }