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->sender        = LNET_NID_ANY;
  76
  77         } else {
  78                 /* event for passive message */
  79                 ev->target.pid    = hdr->dest_pid;
  80                 ev->target.nid    = hdr->dest_nid;
  81                 ev->initiator.pid = hdr->src_pid;
  82                 ev->initiator.nid = hdr->src_nid;
  83                 ev->rlength       = hdr->payload_length;
  84                 ev->sender        = msg->msg_from;
  85                 ev->mlength       = msg->msg_wanted;
  86                 ev->offset        = msg->msg_offset;
  87         }
  88
  89         switch (ev_type) {
  90         default:
  91                 LBUG();
  92
  93         case LNET_EVENT_PUT: /* passive PUT */
  94                 ev->pt_index   = hdr->msg.put.ptl_index;
  95                 ev->match_bits = hdr->msg.put.match_bits;
  96                 ev->hdr_data   = hdr->msg.put.hdr_data;
  97                 return;
  98
  99         case LNET_EVENT_GET: /* passive GET */
 100                 ev->pt_index   = hdr->msg.get.ptl_index;
 101                 ev->match_bits = hdr->msg.get.match_bits;
 102                 ev->hdr_data   = 0;
 103                 return;
 104
 105         case LNET_EVENT_ACK: /* ACK */
 106                 ev->match_bits = hdr->msg.ack.match_bits;
 107                 ev->mlength    = hdr->msg.ack.mlength;
 108                 return;
 109
 110         case LNET_EVENT_REPLY: /* REPLY */
 111                 return;
 112
 113         case LNET_EVENT_SEND: /* active message */
 114                 if (msg->msg_type == LNET_MSG_PUT) {
 115                         ev->pt_index   = le32_to_cpu(hdr->msg.put.ptl_index);
 116                         ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
 117                         ev->offset     = le32_to_cpu(hdr->msg.put.offset);
 118                         ev->mlength    =
 119                         ev->rlength    = le32_to_cpu(hdr->payload_length);
 120                         ev->hdr_data   = le64_to_cpu(hdr->msg.put.hdr_data);
 121
 122                 } else {
 123                         LASSERT(msg->msg_type == LNET_MSG_GET);
 124                         ev->pt_index   = le32_to_cpu(hdr->msg.get.ptl_index);
 125                         ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
 126                         ev->mlength    =
 127                         ev->rlength    = le32_to_cpu(hdr->msg.get.sink_length);
 128                         ev->offset     = le32_to_cpu(hdr->msg.get.src_offset);
 129                         ev->hdr_data   = 0;
 130                 }
 131                 return;
 132         }
 133 }
 134
 135 void
 136 lnet_msg_commit(lnet_msg_t *msg, int cpt)
 137 {
 138         struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
 139         lnet_counters_t           *counters  = the_lnet.ln_counters[cpt];
 140
 141         /* routed message can be committed for both receiving and sending */
 142         LASSERT(!msg->msg_tx_committed);
 143
 144         if (msg->msg_sending) {
 145                 LASSERT(!msg->msg_receiving);
 146
 147                 msg->msg_tx_cpt = cpt;
 148                 msg->msg_tx_committed = 1;
 149                 if (msg->msg_rx_committed) { /* routed message REPLY */
 150                         LASSERT(msg->msg_onactivelist);
 151                         return;
 152                 }
 153         } else {
 154                 LASSERT(!msg->msg_sending);
 155                 msg->msg_rx_cpt = cpt;
 156                 msg->msg_rx_committed = 1;
 157         }
 158
 159         LASSERT(!msg->msg_onactivelist);
 160         msg->msg_onactivelist = 1;
 161         list_add(&msg->msg_activelist, &container->msc_active);
 162
 163         counters->msgs_alloc++;
 164         if (counters->msgs_alloc > counters->msgs_max)
 165                 counters->msgs_max = counters->msgs_alloc;
 166 }
 167
 168 static void
 169 lnet_msg_decommit_tx(lnet_msg_t *msg, int status)
 170 {
 171         lnet_counters_t *counters;
 172         lnet_event_t    *ev = &msg->msg_ev;
 173
 174         LASSERT(msg->msg_tx_committed);
 175         if (status != 0)
 176                 goto out;
 177
 178         counters = the_lnet.ln_counters[msg->msg_tx_cpt];
 179         switch (ev->type) {
 180         default: /* routed message */
 181                 LASSERT(msg->msg_routing);
 182                 LASSERT(msg->msg_rx_committed);
 183                 LASSERT(ev->type == 0);
 184
 185                 counters->route_length += msg->msg_len;
 186                 counters->route_count++;
 187                 goto out;
 188
 189         case LNET_EVENT_PUT:
 190                 /* should have been decommitted */
 191                 LASSERT(!msg->msg_rx_committed);
 192                 /* overwritten while sending ACK */
 193                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 194                 msg->msg_type = LNET_MSG_PUT; /* fix type */
 195                 break;
 196
 197         case LNET_EVENT_SEND:
 198                 LASSERT(!msg->msg_rx_committed);
 199                 if (msg->msg_type == LNET_MSG_PUT)
 200                         counters->send_length += msg->msg_len;
 201                 break;
 202
 203         case LNET_EVENT_GET:
 204                 LASSERT(msg->msg_rx_committed);
 205                 /* overwritten while sending reply, we should never be
 206                  * here for optimized GET */
 207                 LASSERT(msg->msg_type == LNET_MSG_REPLY);
 208                 msg->msg_type = LNET_MSG_GET; /* fix type */
 209                 break;
 210         }
 211
 212         counters->send_count++;
 213  out:
 214         lnet_return_tx_credits_locked(msg);
 215         msg->msg_tx_committed = 0;
 216 }
 217
 218 static void
 219 lnet_msg_decommit_rx(lnet_msg_t *msg, int status)
 220 {
 221         lnet_counters_t *counters;
 222         lnet_event_t    *ev = &msg->msg_ev;
 223
 224         LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
 225         LASSERT(msg->msg_rx_committed);
 226
 227         if (status != 0)
 228                 goto out;
 229
 230         counters = the_lnet.ln_counters[msg->msg_rx_cpt];
 231         switch (ev->type) {
 232         default:
 233                 LASSERT(ev->type == 0);
 234                 LASSERT(msg->msg_routing);
 235                 goto out;
 236
 237         case LNET_EVENT_ACK:
 238                 LASSERT(msg->msg_type == LNET_MSG_ACK);
 239                 break;
 240
 241         case LNET_EVENT_GET:
 242                 /* type is "REPLY" if it's an optimized GET on passive side,
 243                  * because optimized GET will never be committed for sending,
 244                  * so message type wouldn't be changed back to "GET" by
 245                  * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
 246                 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
 247                         msg->msg_type == LNET_MSG_GET);
 248                 counters->send_length += msg->msg_wanted;
 249                 break;
 250
 251         case LNET_EVENT_PUT:
 252                 LASSERT(msg->msg_type == LNET_MSG_PUT);
 253                 break;
 254
 255         case LNET_EVENT_REPLY:
 256                 /* type is "GET" if it's an optimized GET on active side,
 257                  * see details in lnet_create_reply_msg() */
 258                 LASSERT(msg->msg_type == LNET_MSG_GET ||
 259                         msg->msg_type == LNET_MSG_REPLY);
 260                 break;
 261         }
 262
 263         counters->recv_count++;
 264         if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
 265                 counters->recv_length += msg->msg_wanted;
 266
 267  out:
 268         lnet_return_rx_credits_locked(msg);
 269         msg->msg_rx_committed = 0;
 270 }
 271
 272 void
 273 lnet_msg_decommit(lnet_msg_t *msg, int cpt, int status)
 274 {
 275         int     cpt2 = cpt;
 276
 277         LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
 278         LASSERT(msg->msg_onactivelist);
 279
 280         if (msg->msg_tx_committed) { /* always decommit for sending first */
 281                 LASSERT(cpt == msg->msg_tx_cpt);
 282                 lnet_msg_decommit_tx(msg, status);
 283         }
 284
 285         if (msg->msg_rx_committed) {
 286                 /* forwarding msg committed for both receiving and sending */
 287                 if (cpt != msg->msg_rx_cpt) {
 288                         lnet_net_unlock(cpt);
 289                         cpt2 = msg->msg_rx_cpt;
 290                         lnet_net_lock(cpt2);
 291                 }
 292                 lnet_msg_decommit_rx(msg, status);
 293         }
 294
 295         list_del(&msg->msg_activelist);
 296         msg->msg_onactivelist = 0;
 297
 298         the_lnet.ln_counters[cpt2]->msgs_alloc--;
 299
 300         if (cpt2 != cpt) {
 301                 lnet_net_unlock(cpt2);
 302                 lnet_net_lock(cpt);
 303         }
 304 }
 305
 306 void
 307 lnet_msg_attach_md(lnet_msg_t *msg, lnet_libmd_t *md,
 308                    unsigned int offset, unsigned int mlen)
 309 {
 310         /* NB: @offset and @len are only useful for receiving */
 311         /* Here, we attach the MD on lnet_msg and mark it busy and
 312          * decrementing its threshold. Come what may, the lnet_msg "owns"
 313          * the MD until a call to lnet_msg_detach_md or lnet_finalize()
 314          * signals completion. */
 315         LASSERT(!msg->msg_routing);
 316
 317         msg->msg_md = md;
 318         if (msg->msg_receiving) { /* committed for receiving */
 319                 msg->msg_offset = offset;
 320                 msg->msg_wanted = mlen;
 321         }
 322
 323         md->md_refcount++;
 324         if (md->md_threshold != LNET_MD_THRESH_INF) {
 325                 LASSERT(md->md_threshold > 0);
 326                 md->md_threshold--;
 327         }
 328
 329         /* build umd in event */
 330         lnet_md2handle(&msg->msg_ev.md_handle, md);
 331         lnet_md_deconstruct(md, &msg->msg_ev.md);
 332 }
 333
 334 void
 335 lnet_msg_detach_md(lnet_msg_t *msg, int status)
 336 {
 337         lnet_libmd_t    *md = msg->msg_md;
 338         int             unlink;
 339
 340         /* Now it's safe to drop my caller's ref */
 341         md->md_refcount--;
 342         LASSERT(md->md_refcount >= 0);
 343
 344         unlink = lnet_md_unlinkable(md);
 345         if (md->md_eq != NULL) {
 346                 msg->msg_ev.status   = status;
 347                 msg->msg_ev.unlinked = unlink;
 348                 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
 349         }
 350
 351         if (unlink)
 352                 lnet_md_unlink(md);
 353
 354         msg->msg_md = NULL;
 355 }
 356
 357 static int
 358 lnet_complete_msg_locked(lnet_msg_t *msg, int cpt)
 359 {
 360         lnet_handle_wire_t ack_wmd;
 361         int                rc;
 362         int                status = msg->msg_ev.status;
 363
 364         LASSERT(msg->msg_onactivelist);
 365
 366         if (status == 0 && msg->msg_ack) {
 367                 /* Only send an ACK if the PUT completed successfully */
 368
 369                 lnet_msg_decommit(msg, cpt, 0);
 370
 371                 msg->msg_ack = 0;
 372                 lnet_net_unlock(cpt);
 373
 374                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
 375                 LASSERT(!msg->msg_routing);
 376
 377                 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
 378
 379                 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.initiator, 0, 0);
 380
 381                 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
 382                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
 383                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 384
 385                 /* NB: we probably want to use NID of msg::msg_from as 3rd
 386                  * parameter (router NID) if it's routed message */
 387                 rc = lnet_send(msg->msg_ev.target.nid, msg, LNET_NID_ANY);
 388
 389                 lnet_net_lock(cpt);
 390                 /*
 391                  * NB: message is committed for sending, we should return
 392                  * on success because LND will finalize this message later.
 393                  *
 394                  * Also, there is possibility that message is committed for
 395                  * sending and also failed before delivering to LND,
 396                  * i.e: ENOMEM, in that case we can't fall through either
 397                  * because CPT for sending can be different with CPT for
 398                  * receiving, so we should return back to lnet_finalize()
 399                  * to make sure we are locking the correct partition.
 400                  */
 401                 return rc;
 402
 403         } else if (status == 0 &&       /* OK so far */
 404                    (msg->msg_routing && !msg->msg_sending)) {
 405                 /* not forwarded */
 406                 LASSERT(!msg->msg_receiving);   /* called back recv already */
 407                 lnet_net_unlock(cpt);
 408
 409                 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
 410
 411                 lnet_net_lock(cpt);
 412                 /*
 413                  * NB: message is committed for sending, we should return
 414                  * on success because LND will finalize this message later.
 415                  *
 416                  * Also, there is possibility that message is committed for
 417                  * sending and also failed before delivering to LND,
 418                  * i.e: ENOMEM, in that case we can't fall through either:
 419                  * - The rule is message must decommit for sending first if
 420                  *   the it's committed for both sending and receiving
 421                  * - CPT for sending can be different with CPT for receiving,
 422                  *   so we should return back to lnet_finalize() to make
 423                  *   sure we are locking the correct partition.
 424                  */
 425                 return rc;
 426         }
 427
 428         lnet_msg_decommit(msg, cpt, status);
 429         lnet_msg_free(msg);
 430         return 0;
 431 }
 432
 433 void
 434 lnet_finalize(lnet_ni_t *ni, lnet_msg_t *msg, int status)
 435 {
 436         struct lnet_msg_container       *container;
 437         int                             my_slot;
 438         int                             cpt;
 439         int                             rc;
 440         int                             i;
 441
 442         LASSERT(!in_interrupt());
 443
 444         if (msg == NULL)
 445                 return;
 446
 447         msg->msg_ev.status = status;
 448
 449         if (msg->msg_md != NULL) {
 450                 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
 451
 452                 lnet_res_lock(cpt);
 453                 lnet_msg_detach_md(msg, status);
 454                 lnet_res_unlock(cpt);
 455         }
 456
 457  again:
 458         rc = 0;
 459         if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
 460                 /* not committed to network yet */
 461                 LASSERT(!msg->msg_onactivelist);
 462                 lnet_msg_free(msg);
 463                 return;
 464         }
 465
 466         /*
 467          * NB: routed message can be committed for both receiving and sending,
 468          * we should finalize in LIFO order and keep counters correct.
 469          * (finalize sending first then finalize receiving)
 470          */
 471         cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
 472         lnet_net_lock(cpt);
 473
 474         container = the_lnet.ln_msg_containers[cpt];
 475         list_add_tail(&msg->msg_list, &container->msc_finalizing);
 476
 477         /* Recursion breaker.  Don't complete the message here if I am (or
 478          * enough other threads are) already completing messages */
 479
 480         my_slot = -1;
 481         for (i = 0; i < container->msc_nfinalizers; i++) {
 482                 if (container->msc_finalizers[i] == current)
 483                         break;
 484
 485                 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
 486                         my_slot = i;
 487         }
 488
 489         if (i < container->msc_nfinalizers || my_slot < 0) {
 490                 lnet_net_unlock(cpt);
 491                 return;
 492         }
 493
 494         container->msc_finalizers[my_slot] = current;
 495
 496         while (!list_empty(&container->msc_finalizing)) {
 497                 msg = list_entry(container->msc_finalizing.next,
 498                                  lnet_msg_t, msg_list);
 499
 500                 list_del(&msg->msg_list);
 501
 502                 /* NB drops and regains the lnet lock if it actually does
 503                  * anything, so my finalizing friends can chomp along too */
 504                 rc = lnet_complete_msg_locked(msg, cpt);
 505                 if (rc != 0)
 506                         break;
 507         }
 508
 509         if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
 510                 lnet_net_unlock(cpt);
 511                 lnet_delay_rule_check();
 512                 lnet_net_lock(cpt);
 513         }
 514
 515         container->msc_finalizers[my_slot] = NULL;
 516         lnet_net_unlock(cpt);
 517
 518         if (rc != 0)
 519                 goto again;
 520 }
 521 EXPORT_SYMBOL(lnet_finalize);
 522
 523 void
 524 lnet_msg_container_cleanup(struct lnet_msg_container *container)
 525 {
 526         int     count = 0;
 527
 528         if (container->msc_init == 0)
 529                 return;
 530
 531         while (!list_empty(&container->msc_active)) {
 532                 lnet_msg_t *msg = list_entry(container->msc_active.next,
 533                                              lnet_msg_t, msg_activelist);
 534
 535                 LASSERT(msg->msg_onactivelist);
 536                 msg->msg_onactivelist = 0;
 537                 list_del(&msg->msg_activelist);
 538                 lnet_msg_free(msg);
 539                 count++;
 540         }
 541
 542         if (count > 0)
 543                 CERROR("%d active msg on exit\n", count);
 544
 545         if (container->msc_finalizers != NULL) {
 546                 LIBCFS_FREE(container->msc_finalizers,
 547                             container->msc_nfinalizers *
 548                             sizeof(*container->msc_finalizers));
 549                 container->msc_finalizers = NULL;
 550         }
 551         container->msc_init = 0;
 552 }
 553
 554 int
 555 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
 556 {
 557         int     rc;
 558
 559         container->msc_init = 1;
 560
 561         INIT_LIST_HEAD(&container->msc_active);
 562         INIT_LIST_HEAD(&container->msc_finalizing);
 563
 564         rc = 0;
 565         /* number of CPUs */
 566         container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
 567
 568         LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
 569                          container->msc_nfinalizers *
 570                          sizeof(*container->msc_finalizers));
 571
 572         if (container->msc_finalizers == NULL) {
 573                 CERROR("Failed to allocate message finalizers\n");
 574                 lnet_msg_container_cleanup(container);
 575                 return -ENOMEM;
 576         }
 577
 578         return rc;
 579 }
 580
 581 void
 582 lnet_msg_containers_destroy(void)
 583 {
 584         struct lnet_msg_container *container;
 585         int     i;
 586
 587         if (the_lnet.ln_msg_containers == NULL)
 588                 return;
 589
 590         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
 591                 lnet_msg_container_cleanup(container);
 592
 593         cfs_percpt_free(the_lnet.ln_msg_containers);
 594         the_lnet.ln_msg_containers = NULL;
 595 }
 596
 597 int
 598 lnet_msg_containers_create(void)
 599 {
 600         struct lnet_msg_container *container;
 601         int     rc;
 602         int     i;
 603
 604         the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
 605                                                       sizeof(*container));
 606
 607         if (the_lnet.ln_msg_containers == NULL) {
 608                 CERROR("Failed to allocate cpu-partition data for network\n");
 609                 return -ENOMEM;
 610         }
 611
 612         cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
 613                 rc = lnet_msg_container_setup(container, i);
 614                 if (rc != 0) {
 615                         lnet_msg_containers_destroy();
 616                         return rc;
 617                 }
 618         }
 619
 620         return 0;
 621 }