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(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
  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(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
 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(struct lnet_msg *msg, int status)
 176 {
 177         struct lnet_counters *counters;
 178         struct lnet_event *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         if (msg->msg_txpeer)
 220                 atomic_inc(&msg->msg_txpeer->lpni_stats.send_count);
 221         if (msg->msg_txni)
 222                 atomic_inc(&msg->msg_txni->ni_stats.send_count);
 223  out:
 224         lnet_return_tx_credits_locked(msg);
 225         msg->msg_tx_committed = 0;
 226 }
 227
 228 static void
 229 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
 230 {
 231         struct lnet_counters *counters;
 232         struct lnet_event *ev = &msg->msg_ev;
 233
 234         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 235         LASSERT(msg->msg_rx_committed);
 236
 237         if (status != 0)
 238                 goto out;
 239
 240         counters = the_lnet.ln_counters[msg->msg_rx_cpt];
 241         switch (ev->type) {
 242         default:
 243                 LASSERT(ev->type == 0);
 244                 LASSERT(msg->msg_routing);
 245                 goto out;
 246
 247         case LNET_EVENT_ACK:
 248                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 249                 break;
 250
 251         case LNET_EVENT_GET:
 252                 /* type is "REPLY" if it's an optimized GET on passive side,
 253                  * because optimized GET will never be committed for sending,
 254                  * so message type wouldn't be changed back to "GET" by
 255                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 256                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 257                         msg->msg_type == LNET_MSG_GET);
 258                 counters->send_length += msg->msg_wanted;
 259                 break;
 260
 261         case LNET_EVENT_PUT:
 262                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 263                 break;
 264
 265         case LNET_EVENT_REPLY:
 266                 /* type is "GET" if it's an optimized GET on active side,
 267                  * see details in lnet_create_reply_msg() */
 268                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 269                         msg->msg_type == LNET_MSG_REPLY);
 270                 break;
 271         }
 272
 273         counters->recv_count++;
 274         if (msg->msg_rxpeer)
 275                 atomic_inc(&msg->msg_rxpeer->lpni_stats.recv_count);
 276         if (msg->msg_rxni)
 277                 atomic_inc(&msg->msg_rxni->ni_stats.recv_count);
 278         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 279                 counters->recv_length += msg->msg_wanted;
 280
 281  out:
 282         lnet_return_rx_credits_locked(msg);
 283         msg->msg_rx_committed = 0;
 284 }
 285
 286 void
 287 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
 288 {
 289         int     cpt2 = cpt;
 290
 291         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 292         LASSERT(msg->msg_onactivelist);
 293
 294         if (msg->msg_tx_committed) { /* always decommit for sending first */
 295                 LASSERT(cpt == msg->msg_tx_cpt);
 296                 lnet_msg_decommit_tx(msg, status);
 297         }
 298
 299         if (msg->msg_rx_committed) {
 300                 /* forwarding msg committed for both receiving and sending */
 301                 if (cpt != msg->msg_rx_cpt) {
 302                         lnet_net_unlock(cpt);
 303                         cpt2 = msg->msg_rx_cpt;
 304                         lnet_net_lock(cpt2);
 305                 }
 306                 lnet_msg_decommit_rx(msg, status);
 307         }
 308
 309         list_del(&msg->msg_activelist);
 310         msg->msg_onactivelist = 0;
 311
 312         the_lnet.ln_counters[cpt2]->msgs_alloc--;
 313
 314         if (cpt2 != cpt) {
 315                 lnet_net_unlock(cpt2);
 316                 lnet_net_lock(cpt);
 317         }
 318 }
 319
 320 void
 321 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
 322                    unsigned int offset, unsigned int mlen)
 323 {
 324         /* NB: @offset and @len are only useful for receiving */
 325         /* Here, we attach the MD on lnet_msg and mark it busy and
 326          * decrementing its threshold. Come what may, the lnet_msg "owns"
 327          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 328          * signals completion. */
 329         LASSERT(!msg->msg_routing);
 330
 331         msg->msg_md = md;
 332         if (msg->msg_receiving) { /* committed for receiving */
 333                 msg->msg_offset = offset;
 334                 msg->msg_wanted = mlen;
 335         }
 336
 337         md->md_refcount++;
 338         if (md->md_threshold != LNET_MD_THRESH_INF) {
 339                 LASSERT(md->md_threshold > 0);
 340                 md->md_threshold--;
 341         }
 342
 343         /* build umd in event */
 344         lnet_md2handle(&msg->msg_ev.md_handle, md);
 345         lnet_md_deconstruct(md, &msg->msg_ev.md);
 346 }
 347
 348 void
 349 lnet_msg_detach_md(struct lnet_msg *msg, int status)
 350 {
 351         struct lnet_libmd *md = msg->msg_md;
 352         int unlink;
 353
 354         /* Now it's safe to drop my caller's ref */
 355         md->md_refcount--;
 356         LASSERT(md->md_refcount >= 0);
 357
 358         unlink = lnet_md_unlinkable(md);
 359         if (md->md_eq != NULL) {
 360                 msg->msg_ev.status   = status;
 361                 msg->msg_ev.unlinked = unlink;
 362                 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
 363         }
 364
 365         if (unlink)
 366                 lnet_md_unlink(md);
 367
 368         msg->msg_md = NULL;
 369 }
 370
 371 static int
 372 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
 373 {
 374         struct lnet_handle_wire ack_wmd;
 375         int                rc;
 376         int                status = msg->msg_ev.status;
 377
 378         LASSERT(msg->msg_onactivelist);
 379
 380         if (status == 0 && msg->msg_ack) {
 381                 /* Only send an ACK if the PUT completed successfully */
 382
 383                 lnet_msg_decommit(msg, cpt, 0);
 384
 385                 msg->msg_ack = 0;
 386                 lnet_net_unlock(cpt);
 387
 388                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 389                 LASSERT(!msg->msg_routing);
 390
 391                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 392
 393                 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
 394
 395                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 396                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 397                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 398
 399                 /* NB: we probably want to use NID of msg::msg_from as 3rd
 400                  * parameter (router NID) if it's routed message */
 401                 rc = lnet_send(msg->msg_ev.target.nid, msg, LNET_NID_ANY);
 402
 403                 lnet_net_lock(cpt);
 404                 /*
 405                  * NB: message is committed for sending, we should return
 406                  * on success because LND will finalize this message later.
 407                  *
 408                  * Also, there is possibility that message is committed for
 409                  * sending and also failed before delivering to LND,
 410                  * i.e: ENOMEM, in that case we can't fall through either
 411                  * because CPT for sending can be different with CPT for
 412                  * receiving, so we should return back to lnet_finalize()
 413                  * to make sure we are locking the correct partition.
 414                  */
 415                 return rc;
 416
 417         } else if (status == 0 &&       /* OK so far */
 418                    (msg->msg_routing && !msg->msg_sending)) {
 419                 /* not forwarded */
 420                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 421                 lnet_net_unlock(cpt);
 422
 423                 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
 424
 425                 lnet_net_lock(cpt);
 426                 /*
 427                  * NB: message is committed for sending, we should return
 428                  * on success because LND will finalize this message later.
 429                  *
 430                  * Also, there is possibility that message is committed for
 431                  * sending and also failed before delivering to LND,
 432                  * i.e: ENOMEM, in that case we can't fall through either:
 433                  * - The rule is message must decommit for sending first if
 434                  *   the it's committed for both sending and receiving
 435                  * - CPT for sending can be different with CPT for receiving,
 436                  *   so we should return back to lnet_finalize() to make
 437                  *   sure we are locking the correct partition.
 438                  */
 439                 return rc;
 440         }
 441
 442         lnet_msg_decommit(msg, cpt, status);
 443         lnet_msg_free(msg);
 444         return 0;
 445 }
 446
 447 void
 448 lnet_finalize(struct lnet_msg *msg, int status)
 449 {
 450         struct lnet_msg_container       *container;
 451         int                             my_slot;
 452         int                             cpt;
 453         int                             rc;
 454         int                             i;
 455
 456         LASSERT(!in_interrupt());
 457
 458         if (msg == NULL)
 459                 return;
 460
 461         msg->msg_ev.status = status;
 462
 463         if (msg->msg_md != NULL) {
 464                 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
 465
 466                 lnet_res_lock(cpt);
 467                 lnet_msg_detach_md(msg, status);
 468                 lnet_res_unlock(cpt);
 469         }
 470
 471  again:
 472         rc = 0;
 473         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
 474                 /* not committed to network yet */
 475                 LASSERT(!msg->msg_onactivelist);
 476                 lnet_msg_free(msg);
 477                 return;
 478         }
 479
 480         /*
 481          * NB: routed message can be committed for both receiving and sending,
 482          * we should finalize in LIFO order and keep counters correct.
 483          * (finalize sending first then finalize receiving)
 484          */
 485         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
 486         lnet_net_lock(cpt);
 487
 488         container = the_lnet.ln_msg_containers[cpt];
 489         list_add_tail(&msg->msg_list, &container->msc_finalizing);
 490
 491         /* Recursion breaker.  Don't complete the message here if I am (or
 492          * enough other threads are) already completing messages */
 493
 494         my_slot = -1;
 495         for (i = 0; i < container->msc_nfinalizers; i++) {
 496                 if (container->msc_finalizers[i] == current)
 497                         break;
 498
 499                 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
 500                         my_slot = i;
 501         }
 502
 503         if (i < container->msc_nfinalizers || my_slot < 0) {
 504                 lnet_net_unlock(cpt);
 505                 return;
 506         }
 507
 508         container->msc_finalizers[my_slot] = current;
 509
 510         while (!list_empty(&container->msc_finalizing)) {
 511                 msg = list_entry(container->msc_finalizing.next,
 512                                  struct lnet_msg, msg_list);
 513
 514                 list_del(&msg->msg_list);
 515
 516                 /* NB drops and regains the lnet lock if it actually does
 517                  * anything, so my finalizing friends can chomp along too */
 518                 rc = lnet_complete_msg_locked(msg, cpt);
 519                 if (rc != 0)
 520                         break;
 521         }
 522
 523         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
 524                 lnet_net_unlock(cpt);
 525                 lnet_delay_rule_check();
 526                 lnet_net_lock(cpt);
 527         }
 528
 529         container->msc_finalizers[my_slot] = NULL;
 530         lnet_net_unlock(cpt);
 531
 532         if (rc != 0)
 533                 goto again;
 534 }
 535 EXPORT_SYMBOL(lnet_finalize);
 536
 537 void
 538 lnet_msg_container_cleanup(struct lnet_msg_container *container)
 539 {
 540         int     count = 0;
 541
 542         if (container->msc_init == 0)
 543                 return;
 544
 545         while (!list_empty(&container->msc_active)) {
 546                 struct lnet_msg *msg;
 547
 548                 msg  = list_entry(container->msc_active.next,
 549                                   struct lnet_msg, msg_activelist);
 550                 LASSERT(msg->msg_onactivelist);
 551                 msg->msg_onactivelist = 0;
 552                 list_del(&msg->msg_activelist);
 553                 lnet_msg_free(msg);
 554                 count++;
 555         }
 556
 557         if (count > 0)
 558                 CERROR("%d active msg on exit\n", count);
 559
 560         if (container->msc_finalizers != NULL) {
 561                 LIBCFS_FREE(container->msc_finalizers,
 562                             container->msc_nfinalizers *
 563                             sizeof(*container->msc_finalizers));
 564                 container->msc_finalizers = NULL;
 565         }
 566         container->msc_init = 0;
 567 }
 568
 569 int
 570 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
 571 {
 572         int     rc;
 573
 574         container->msc_init = 1;
 575
 576         INIT_LIST_HEAD(&container->msc_active);
 577         INIT_LIST_HEAD(&container->msc_finalizing);
 578
 579         rc = 0;
 580         /* number of CPUs */
 581         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
 582
 583         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
 584                          container->msc_nfinalizers *
 585                          sizeof(*container->msc_finalizers));
 586
 587         if (container->msc_finalizers == NULL) {
 588                 CERROR("Failed to allocate message finalizers\n");
 589                 lnet_msg_container_cleanup(container);
 590                 return -ENOMEM;
 591         }
 592
 593         return rc;
 594 }
 595
 596 void
 597 lnet_msg_containers_destroy(void)
 598 {
 599         struct lnet_msg_container *container;
 600         int     i;
 601
 602         if (the_lnet.ln_msg_containers == NULL)
 603                 return;
 604
 605         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
 606                 lnet_msg_container_cleanup(container);
 607
 608         cfs_percpt_free(the_lnet.ln_msg_containers);
 609         the_lnet.ln_msg_containers = NULL;
 610 }
 611
 612 int
 613 lnet_msg_containers_create(void)
 614 {
 615         struct lnet_msg_container *container;
 616         int     rc;
 617         int     i;
 618
 619         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
 620                                                       sizeof(*container));
 621
 622         if (the_lnet.ln_msg_containers == NULL) {
 623                 CERROR("Failed to allocate cpu-partition data for network\n");
 624                 return -ENOMEM;
 625         }
 626
 627         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
 628                 rc = lnet_msg_container_setup(container, i);
 629                 if (rc != 0) {
 630                         lnet_msg_containers_destroy();
 631                         return rc;
 632                 }
 633         }
 634
 635         return 0;
 636 }