lnet/lnet/lib-move.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, 2016, 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-move.c
  33  *
  34  * Data movement routines
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <lnet/lib-lnet.h>
  40 #include <linux/nsproxy.h>
  41 #include <net/net_namespace.h>
  42
  43 static int local_nid_dist_zero = 1;
  44 module_param(local_nid_dist_zero, int, 0444);
  45 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
  46
  47 int
  48 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
  49 {
  50         lnet_test_peer_t *tp;
  51         struct list_head *el;
  52         struct list_head *next;
  53         struct list_head  cull;
  54
  55         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
  56         if (threshold != 0) {
  57                 /* Adding a new entry */
  58                 LIBCFS_ALLOC(tp, sizeof(*tp));
  59                 if (tp == NULL)
  60                         return -ENOMEM;
  61
  62                 tp->tp_nid = nid;
  63                 tp->tp_threshold = threshold;
  64
  65                 lnet_net_lock(0);
  66                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
  67                 lnet_net_unlock(0);
  68                 return 0;
  69         }
  70
  71         /* removing entries */
  72         INIT_LIST_HEAD(&cull);
  73
  74         lnet_net_lock(0);
  75
  76         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
  77                 tp = list_entry(el, lnet_test_peer_t, tp_list);
  78
  79                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
  80                     nid == LNET_NID_ANY ||      /* removing all entries */
  81                     tp->tp_nid == nid) {        /* matched this one */
  82                         list_del(&tp->tp_list);
  83                         list_add(&tp->tp_list, &cull);
  84                 }
  85         }
  86
  87         lnet_net_unlock(0);
  88
  89         while (!list_empty(&cull)) {
  90                 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
  91
  92                 list_del(&tp->tp_list);
  93                 LIBCFS_FREE(tp, sizeof(*tp));
  94         }
  95         return 0;
  96 }
  97
  98 static int
  99 fail_peer (lnet_nid_t nid, int outgoing)
 100 {
 101         lnet_test_peer_t *tp;
 102         struct list_head *el;
 103         struct list_head *next;
 104         struct list_head  cull;
 105         int               fail = 0;
 106
 107         INIT_LIST_HEAD(&cull);
 108
 109         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 110         lnet_net_lock(0);
 111
 112         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 113                 tp = list_entry(el, lnet_test_peer_t, tp_list);
 114
 115                 if (tp->tp_threshold == 0) {
 116                         /* zombie entry */
 117                         if (outgoing) {
 118                                 /* only cull zombies on outgoing tests,
 119                                  * since we may be at interrupt priority on
 120                                  * incoming messages. */
 121                                 list_del(&tp->tp_list);
 122                                 list_add(&tp->tp_list, &cull);
 123                         }
 124                         continue;
 125                 }
 126
 127                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
 128                     nid == tp->tp_nid) {                /* fail this peer */
 129                         fail = 1;
 130
 131                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 132                                 tp->tp_threshold--;
 133                                 if (outgoing &&
 134                                     tp->tp_threshold == 0) {
 135                                         /* see above */
 136                                         list_del(&tp->tp_list);
 137                                         list_add(&tp->tp_list, &cull);
 138                                 }
 139                         }
 140                         break;
 141                 }
 142         }
 143
 144         lnet_net_unlock(0);
 145
 146         while (!list_empty(&cull)) {
 147                 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
 148                 list_del(&tp->tp_list);
 149
 150                 LIBCFS_FREE(tp, sizeof(*tp));
 151         }
 152
 153         return fail;
 154 }
 155
 156 unsigned int
 157 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 158 {
 159         unsigned int nob = 0;
 160
 161         LASSERT(niov == 0 || iov != NULL);
 162         while (niov-- > 0)
 163                 nob += (iov++)->iov_len;
 164
 165         return (nob);
 166 }
 167 EXPORT_SYMBOL(lnet_iov_nob);
 168
 169 void
 170 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 171                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 172                   unsigned int nob)
 173 {
 174         /* NB diov, siov are READ-ONLY */
 175         unsigned int  this_nob;
 176
 177         if (nob == 0)
 178                 return;
 179
 180         /* skip complete frags before 'doffset' */
 181         LASSERT(ndiov > 0);
 182         while (doffset >= diov->iov_len) {
 183                 doffset -= diov->iov_len;
 184                 diov++;
 185                 ndiov--;
 186                 LASSERT(ndiov > 0);
 187         }
 188
 189         /* skip complete frags before 'soffset' */
 190         LASSERT(nsiov > 0);
 191         while (soffset >= siov->iov_len) {
 192                 soffset -= siov->iov_len;
 193                 siov++;
 194                 nsiov--;
 195                 LASSERT(nsiov > 0);
 196         }
 197
 198         do {
 199                 LASSERT(ndiov > 0);
 200                 LASSERT(nsiov > 0);
 201                 this_nob = MIN(diov->iov_len - doffset,
 202                                siov->iov_len - soffset);
 203                 this_nob = MIN(this_nob, nob);
 204
 205                 memcpy((char *)diov->iov_base + doffset,
 206                        (char *)siov->iov_base + soffset, this_nob);
 207                 nob -= this_nob;
 208
 209                 if (diov->iov_len > doffset + this_nob) {
 210                         doffset += this_nob;
 211                 } else {
 212                         diov++;
 213                         ndiov--;
 214                         doffset = 0;
 215                 }
 216
 217                 if (siov->iov_len > soffset + this_nob) {
 218                         soffset += this_nob;
 219                 } else {
 220                         siov++;
 221                         nsiov--;
 222                         soffset = 0;
 223                 }
 224         } while (nob > 0);
 225 }
 226 EXPORT_SYMBOL(lnet_copy_iov2iov);
 227
 228 int
 229 lnet_extract_iov(int dst_niov, struct kvec *dst,
 230                  int src_niov, struct kvec *src,
 231                  unsigned int offset, unsigned int len)
 232 {
 233         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 234          * for exactly 'len' bytes, and return the number of entries.
 235          * NB not destructive to 'src' */
 236         unsigned int    frag_len;
 237         unsigned int    niov;
 238
 239         if (len == 0)                           /* no data => */
 240                 return (0);                     /* no frags */
 241
 242         LASSERT(src_niov > 0);
 243         while (offset >= src->iov_len) {      /* skip initial frags */
 244                 offset -= src->iov_len;
 245                 src_niov--;
 246                 src++;
 247                 LASSERT(src_niov > 0);
 248         }
 249
 250         niov = 1;
 251         for (;;) {
 252                 LASSERT(src_niov > 0);
 253                 LASSERT((int)niov <= dst_niov);
 254
 255                 frag_len = src->iov_len - offset;
 256                 dst->iov_base = ((char *)src->iov_base) + offset;
 257
 258                 if (len <= frag_len) {
 259                         dst->iov_len = len;
 260                         return (niov);
 261                 }
 262
 263                 dst->iov_len = frag_len;
 264
 265                 len -= frag_len;
 266                 dst++;
 267                 src++;
 268                 niov++;
 269                 src_niov--;
 270                 offset = 0;
 271         }
 272 }
 273 EXPORT_SYMBOL(lnet_extract_iov);
 274
 275
 276 unsigned int
 277 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
 278 {
 279         unsigned int  nob = 0;
 280
 281         LASSERT(niov == 0 || kiov != NULL);
 282         while (niov-- > 0)
 283                 nob += (kiov++)->kiov_len;
 284
 285         return (nob);
 286 }
 287 EXPORT_SYMBOL(lnet_kiov_nob);
 288
 289 void
 290 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
 291                     unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
 292                     unsigned int nob)
 293 {
 294         /* NB diov, siov are READ-ONLY */
 295         unsigned int    this_nob;
 296         char           *daddr = NULL;
 297         char           *saddr = NULL;
 298
 299         if (nob == 0)
 300                 return;
 301
 302         LASSERT (!in_interrupt ());
 303
 304         LASSERT (ndiov > 0);
 305         while (doffset >= diov->kiov_len) {
 306                 doffset -= diov->kiov_len;
 307                 diov++;
 308                 ndiov--;
 309                 LASSERT(ndiov > 0);
 310         }
 311
 312         LASSERT(nsiov > 0);
 313         while (soffset >= siov->kiov_len) {
 314                 soffset -= siov->kiov_len;
 315                 siov++;
 316                 nsiov--;
 317                 LASSERT(nsiov > 0);
 318         }
 319
 320         do {
 321                 LASSERT(ndiov > 0);
 322                 LASSERT(nsiov > 0);
 323                 this_nob = MIN(diov->kiov_len - doffset,
 324                                siov->kiov_len - soffset);
 325                 this_nob = MIN(this_nob, nob);
 326
 327                 if (daddr == NULL)
 328                         daddr = ((char *)kmap(diov->kiov_page)) +
 329                                 diov->kiov_offset + doffset;
 330                 if (saddr == NULL)
 331                         saddr = ((char *)kmap(siov->kiov_page)) +
 332                                 siov->kiov_offset + soffset;
 333
 334                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 335                  * However in practice at least one of the kiovs will be mapped
 336                  * kernel pages and the map/unmap will be NOOPs */
 337
 338                 memcpy (daddr, saddr, this_nob);
 339                 nob -= this_nob;
 340
 341                 if (diov->kiov_len > doffset + this_nob) {
 342                         daddr += this_nob;
 343                         doffset += this_nob;
 344                 } else {
 345                         kunmap(diov->kiov_page);
 346                         daddr = NULL;
 347                         diov++;
 348                         ndiov--;
 349                         doffset = 0;
 350                 }
 351
 352                 if (siov->kiov_len > soffset + this_nob) {
 353                         saddr += this_nob;
 354                         soffset += this_nob;
 355                 } else {
 356                         kunmap(siov->kiov_page);
 357                         saddr = NULL;
 358                         siov++;
 359                         nsiov--;
 360                         soffset = 0;
 361                 }
 362         } while (nob > 0);
 363
 364         if (daddr != NULL)
 365                 kunmap(diov->kiov_page);
 366         if (saddr != NULL)
 367                 kunmap(siov->kiov_page);
 368 }
 369 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 370
 371 void
 372 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 373                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 374                     unsigned int nob)
 375 {
 376         /* NB iov, kiov are READ-ONLY */
 377         unsigned int    this_nob;
 378         char           *addr = NULL;
 379
 380         if (nob == 0)
 381                 return;
 382
 383         LASSERT (!in_interrupt ());
 384
 385         LASSERT (niov > 0);
 386         while (iovoffset >= iov->iov_len) {
 387                 iovoffset -= iov->iov_len;
 388                 iov++;
 389                 niov--;
 390                 LASSERT(niov > 0);
 391         }
 392
 393         LASSERT(nkiov > 0);
 394         while (kiovoffset >= kiov->kiov_len) {
 395                 kiovoffset -= kiov->kiov_len;
 396                 kiov++;
 397                 nkiov--;
 398                 LASSERT(nkiov > 0);
 399         }
 400
 401         do {
 402                 LASSERT(niov > 0);
 403                 LASSERT(nkiov > 0);
 404                 this_nob = MIN(iov->iov_len - iovoffset,
 405                                kiov->kiov_len - kiovoffset);
 406                 this_nob = MIN(this_nob, nob);
 407
 408                 if (addr == NULL)
 409                         addr = ((char *)kmap(kiov->kiov_page)) +
 410                                 kiov->kiov_offset + kiovoffset;
 411
 412                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 413                 nob -= this_nob;
 414
 415                 if (iov->iov_len > iovoffset + this_nob) {
 416                         iovoffset += this_nob;
 417                 } else {
 418                         iov++;
 419                         niov--;
 420                         iovoffset = 0;
 421                 }
 422
 423                 if (kiov->kiov_len > kiovoffset + this_nob) {
 424                         addr += this_nob;
 425                         kiovoffset += this_nob;
 426                 } else {
 427                         kunmap(kiov->kiov_page);
 428                         addr = NULL;
 429                         kiov++;
 430                         nkiov--;
 431                         kiovoffset = 0;
 432                 }
 433
 434         } while (nob > 0);
 435
 436         if (addr != NULL)
 437                 kunmap(kiov->kiov_page);
 438 }
 439 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 440
 441 void
 442 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 443                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 444                    unsigned int nob)
 445 {
 446         /* NB kiov, iov are READ-ONLY */
 447         unsigned int    this_nob;
 448         char           *addr = NULL;
 449
 450         if (nob == 0)
 451                 return;
 452
 453         LASSERT (!in_interrupt ());
 454
 455         LASSERT (nkiov > 0);
 456         while (kiovoffset >= kiov->kiov_len) {
 457                 kiovoffset -= kiov->kiov_len;
 458                 kiov++;
 459                 nkiov--;
 460                 LASSERT(nkiov > 0);
 461         }
 462
 463         LASSERT(niov > 0);
 464         while (iovoffset >= iov->iov_len) {
 465                 iovoffset -= iov->iov_len;
 466                 iov++;
 467                 niov--;
 468                 LASSERT(niov > 0);
 469         }
 470
 471         do {
 472                 LASSERT(nkiov > 0);
 473                 LASSERT(niov > 0);
 474                 this_nob = MIN(kiov->kiov_len - kiovoffset,
 475                                iov->iov_len - iovoffset);
 476                 this_nob = MIN(this_nob, nob);
 477
 478                 if (addr == NULL)
 479                         addr = ((char *)kmap(kiov->kiov_page)) +
 480                                 kiov->kiov_offset + kiovoffset;
 481
 482                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 483                 nob -= this_nob;
 484
 485                 if (kiov->kiov_len > kiovoffset + this_nob) {
 486                         addr += this_nob;
 487                         kiovoffset += this_nob;
 488                 } else {
 489                         kunmap(kiov->kiov_page);
 490                         addr = NULL;
 491                         kiov++;
 492                         nkiov--;
 493                         kiovoffset = 0;
 494                 }
 495
 496                 if (iov->iov_len > iovoffset + this_nob) {
 497                         iovoffset += this_nob;
 498                 } else {
 499                         iov++;
 500                         niov--;
 501                         iovoffset = 0;
 502                 }
 503         } while (nob > 0);
 504
 505         if (addr != NULL)
 506                 kunmap(kiov->kiov_page);
 507 }
 508 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 509
 510 int
 511 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
 512                   int src_niov, lnet_kiov_t *src,
 513                   unsigned int offset, unsigned int len)
 514 {
 515         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 516          * for exactly 'len' bytes, and return the number of entries.
 517          * NB not destructive to 'src' */
 518         unsigned int    frag_len;
 519         unsigned int    niov;
 520
 521         if (len == 0)                           /* no data => */
 522                 return (0);                     /* no frags */
 523
 524         LASSERT(src_niov > 0);
 525         while (offset >= src->kiov_len) {      /* skip initial frags */
 526                 offset -= src->kiov_len;
 527                 src_niov--;
 528                 src++;
 529                 LASSERT(src_niov > 0);
 530         }
 531
 532         niov = 1;
 533         for (;;) {
 534                 LASSERT(src_niov > 0);
 535                 LASSERT((int)niov <= dst_niov);
 536
 537                 frag_len = src->kiov_len - offset;
 538                 dst->kiov_page = src->kiov_page;
 539                 dst->kiov_offset = src->kiov_offset + offset;
 540
 541                 if (len <= frag_len) {
 542                         dst->kiov_len = len;
 543                         LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 544                         return niov;
 545                 }
 546
 547                 dst->kiov_len = frag_len;
 548                 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 549
 550                 len -= frag_len;
 551                 dst++;
 552                 src++;
 553                 niov++;
 554                 src_niov--;
 555                 offset = 0;
 556         }
 557 }
 558 EXPORT_SYMBOL(lnet_extract_kiov);
 559
 560 void
 561 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
 562              unsigned int offset, unsigned int mlen, unsigned int rlen)
 563 {
 564         unsigned int  niov = 0;
 565         struct kvec *iov = NULL;
 566         lnet_kiov_t  *kiov = NULL;
 567         int           rc;
 568
 569         LASSERT (!in_interrupt ());
 570         LASSERT (mlen == 0 || msg != NULL);
 571
 572         if (msg != NULL) {
 573                 LASSERT(msg->msg_receiving);
 574                 LASSERT(!msg->msg_sending);
 575                 LASSERT(rlen == msg->msg_len);
 576                 LASSERT(mlen <= msg->msg_len);
 577                 LASSERT(msg->msg_offset == offset);
 578                 LASSERT(msg->msg_wanted == mlen);
 579
 580                 msg->msg_receiving = 0;
 581
 582                 if (mlen != 0) {
 583                         niov = msg->msg_niov;
 584                         iov  = msg->msg_iov;
 585                         kiov = msg->msg_kiov;
 586
 587                         LASSERT (niov > 0);
 588                         LASSERT ((iov == NULL) != (kiov == NULL));
 589                 }
 590         }
 591
 592         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 593                                              niov, iov, kiov, offset, mlen,
 594                                              rlen);
 595         if (rc < 0)
 596                 lnet_finalize(ni, msg, rc);
 597 }
 598
 599 static void
 600 lnet_setpayloadbuffer(lnet_msg_t *msg)
 601 {
 602         lnet_libmd_t *md = msg->msg_md;
 603
 604         LASSERT(msg->msg_len > 0);
 605         LASSERT(!msg->msg_routing);
 606         LASSERT(md != NULL);
 607         LASSERT(msg->msg_niov == 0);
 608         LASSERT(msg->msg_iov == NULL);
 609         LASSERT(msg->msg_kiov == NULL);
 610
 611         msg->msg_niov = md->md_niov;
 612         if ((md->md_options & LNET_MD_KIOV) != 0)
 613                 msg->msg_kiov = md->md_iov.kiov;
 614         else
 615                 msg->msg_iov = md->md_iov.iov;
 616 }
 617
 618 void
 619 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
 620                unsigned int offset, unsigned int len)
 621 {
 622         msg->msg_type = type;
 623         msg->msg_target = target;
 624         msg->msg_len = len;
 625         msg->msg_offset = offset;
 626
 627         if (len != 0)
 628                 lnet_setpayloadbuffer(msg);
 629
 630         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 631         msg->msg_hdr.type           = cpu_to_le32(type);
 632         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 633         /* src_nid will be set later */
 634         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 635         msg->msg_hdr.payload_length = cpu_to_le32(len);
 636 }
 637
 638 static void
 639 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
 640 {
 641         void   *priv = msg->msg_private;
 642         int     rc;
 643
 644         LASSERT (!in_interrupt ());
 645         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
 646                  (msg->msg_txcredit && msg->msg_peertxcredit));
 647
 648         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 649         if (rc < 0)
 650                 lnet_finalize(ni, msg, rc);
 651 }
 652
 653 static int
 654 lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
 655 {
 656         int     rc;
 657
 658         LASSERT(!msg->msg_sending);
 659         LASSERT(msg->msg_receiving);
 660         LASSERT(!msg->msg_rx_ready_delay);
 661         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 662
 663         msg->msg_rx_ready_delay = 1;
 664         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 665                                                   &msg->msg_private);
 666         if (rc != 0) {
 667                 CERROR("recv from %s / send to %s aborted: "
 668                        "eager_recv failed %d\n",
 669                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 670                        libcfs_id2str(msg->msg_target), rc);
 671                 LASSERT(rc < 0); /* required by my callers */
 672         }
 673
 674         return rc;
 675 }
 676
 677 /* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
 678 static void
 679 lnet_ni_query_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
 680 {
 681         cfs_time_t last_alive = 0;
 682
 683         LASSERT(lnet_peer_aliveness_enabled(lp));
 684         LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
 685
 686         lnet_net_unlock(lp->lpni_cpt);
 687         (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
 688         lnet_net_lock(lp->lpni_cpt);
 689
 690         lp->lpni_last_query = cfs_time_current();
 691
 692         if (last_alive != 0) /* NI has updated timestamp */
 693                 lp->lpni_last_alive = last_alive;
 694 }
 695
 696 /* NB: always called with lnet_net_lock held */
 697 static inline int
 698 lnet_peer_is_alive (struct lnet_peer_ni *lp, cfs_time_t now)
 699 {
 700         int        alive;
 701         cfs_time_t deadline;
 702
 703         LASSERT (lnet_peer_aliveness_enabled(lp));
 704
 705         /*
 706          * Trust lnet_notify() if it has more recent aliveness news, but
 707          * ignore the initial assumed death (see lnet_peers_start_down()).
 708          */
 709         if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
 710             cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive))
 711                 return 0;
 712
 713         deadline =
 714           cfs_time_add(lp->lpni_last_alive,
 715                        cfs_time_seconds(lp->lpni_net->net_tunables.
 716                                         lct_peer_timeout));
 717         alive = cfs_time_after(deadline, now);
 718
 719         /*
 720          * Update obsolete lp_alive except for routers assumed to be dead
 721          * initially, because router checker would update aliveness in this
 722          * case, and moreover lpni_last_alive at peer creation is assumed.
 723          */
 724         if (alive && !lp->lpni_alive &&
 725             !(lnet_isrouter(lp) && lp->lpni_alive_count == 0))
 726                 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
 727
 728         return alive;
 729 }
 730
 731
 732 /* NB: returns 1 when alive, 0 when dead, negative when error;
 733  *     may drop the lnet_net_lock */
 734 static int
 735 lnet_peer_alive_locked (struct lnet_ni *ni, struct lnet_peer_ni *lp)
 736 {
 737         cfs_time_t now = cfs_time_current();
 738
 739         if (!lnet_peer_aliveness_enabled(lp))
 740                 return -ENODEV;
 741
 742         if (lnet_peer_is_alive(lp, now))
 743                 return 1;
 744
 745         /*
 746          * Peer appears dead, but we should avoid frequent NI queries (at
 747          * most once per lnet_queryinterval seconds).
 748          */
 749         if (lp->lpni_last_query != 0) {
 750                 static const int lnet_queryinterval = 1;
 751
 752                 cfs_time_t next_query =
 753                            cfs_time_add(lp->lpni_last_query,
 754                                         cfs_time_seconds(lnet_queryinterval));
 755
 756                 if (cfs_time_before(now, next_query)) {
 757                         if (lp->lpni_alive)
 758                                 CWARN("Unexpected aliveness of peer %s: "
 759                                       "%d < %d (%d/%d)\n",
 760                                       libcfs_nid2str(lp->lpni_nid),
 761                                       (int)now, (int)next_query,
 762                                       lnet_queryinterval,
 763                                       lp->lpni_net->net_tunables.lct_peer_timeout);
 764                         return 0;
 765                 }
 766         }
 767
 768         /* query NI for latest aliveness news */
 769         lnet_ni_query_locked(ni, lp);
 770
 771         if (lnet_peer_is_alive(lp, now))
 772                 return 1;
 773
 774         lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
 775         return 0;
 776 }
 777
 778 /**
 779  * \param msg The message to be sent.
 780  * \param do_send True if lnet_ni_send() should be called in this function.
 781  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 782  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 783  *
 784  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 785  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 786  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 787  * \retval -ECANCELED If the MD of the message has been unlinked.
 788  */
 789 static int
 790 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
 791 {
 792         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 793         struct lnet_ni          *ni = msg->msg_txni;
 794         int                     cpt = msg->msg_tx_cpt;
 795         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 796
 797         /* non-lnet_send() callers have checked before */
 798         LASSERT(!do_send || msg->msg_tx_delayed);
 799         LASSERT(!msg->msg_receiving);
 800         LASSERT(msg->msg_tx_committed);
 801
 802         /* NB 'lp' is always the next hop */
 803         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 804             lnet_peer_alive_locked(ni, lp) == 0) {
 805                 the_lnet.ln_counters[cpt]->drop_count++;
 806                 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
 807                 lnet_net_unlock(cpt);
 808
 809                 CNETERR("Dropping message for %s: peer not alive\n",
 810                         libcfs_id2str(msg->msg_target));
 811                 if (do_send)
 812                         lnet_finalize(ni, msg, -EHOSTUNREACH);
 813
 814                 lnet_net_lock(cpt);
 815                 return -EHOSTUNREACH;
 816         }
 817
 818         if (msg->msg_md != NULL &&
 819             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 820                 lnet_net_unlock(cpt);
 821
 822                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 823                         "called on the MD/ME.\n",
 824                         libcfs_id2str(msg->msg_target));
 825                 if (do_send)
 826                         lnet_finalize(ni, msg, -ECANCELED);
 827
 828                 lnet_net_lock(cpt);
 829                 return -ECANCELED;
 830         }
 831
 832         if (!msg->msg_peertxcredit) {
 833                 LASSERT((lp->lpni_txcredits < 0) ==
 834                         !list_empty(&lp->lpni_txq));
 835
 836                 msg->msg_peertxcredit = 1;
 837                 lp->lpni_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
 838                 lp->lpni_txcredits--;
 839
 840                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 841                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 842
 843                 if (lp->lpni_txcredits < 0) {
 844                         msg->msg_tx_delayed = 1;
 845                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 846                         return LNET_CREDIT_WAIT;
 847                 }
 848         }
 849
 850         if (!msg->msg_txcredit) {
 851                 LASSERT((tq->tq_credits < 0) ==
 852                         !list_empty(&tq->tq_delayed));
 853
 854                 msg->msg_txcredit = 1;
 855                 tq->tq_credits--;
 856
 857                 if (tq->tq_credits < tq->tq_credits_min)
 858                         tq->tq_credits_min = tq->tq_credits;
 859
 860                 if (tq->tq_credits < 0) {
 861                         msg->msg_tx_delayed = 1;
 862                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 863                         return LNET_CREDIT_WAIT;
 864                 }
 865         }
 866
 867         if (do_send) {
 868                 lnet_net_unlock(cpt);
 869                 lnet_ni_send(ni, msg);
 870                 lnet_net_lock(cpt);
 871         }
 872         return LNET_CREDIT_OK;
 873 }
 874
 875
 876 static lnet_rtrbufpool_t *
 877 lnet_msg2bufpool(lnet_msg_t *msg)
 878 {
 879         lnet_rtrbufpool_t       *rbp;
 880         int                     cpt;
 881
 882         LASSERT(msg->msg_rx_committed);
 883
 884         cpt = msg->msg_rx_cpt;
 885         rbp = &the_lnet.ln_rtrpools[cpt][0];
 886
 887         LASSERT(msg->msg_len <= LNET_MTU);
 888         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 889                 rbp++;
 890                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 891         }
 892
 893         return rbp;
 894 }
 895
 896 static int
 897 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
 898 {
 899         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 900          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 901          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 902          * received or OK to receive */
 903         struct lnet_peer_ni *lp = msg->msg_rxpeer;
 904         lnet_rtrbufpool_t   *rbp;
 905         lnet_rtrbuf_t       *rb;
 906
 907         LASSERT (msg->msg_iov == NULL);
 908         LASSERT (msg->msg_kiov == NULL);
 909         LASSERT (msg->msg_niov == 0);
 910         LASSERT (msg->msg_routing);
 911         LASSERT (msg->msg_receiving);
 912         LASSERT (!msg->msg_sending);
 913
 914         /* non-lnet_parse callers only receive delayed messages */
 915         LASSERT(!do_recv || msg->msg_rx_delayed);
 916
 917         if (!msg->msg_peerrtrcredit) {
 918                 LASSERT((lp->lpni_rtrcredits < 0) ==
 919                         !list_empty(&lp->lpni_rtrq));
 920
 921                 msg->msg_peerrtrcredit = 1;
 922                 lp->lpni_rtrcredits--;
 923                 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
 924                         lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
 925
 926                 if (lp->lpni_rtrcredits < 0) {
 927                         /* must have checked eager_recv before here */
 928                         LASSERT(msg->msg_rx_ready_delay);
 929                         msg->msg_rx_delayed = 1;
 930                         list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
 931                         return LNET_CREDIT_WAIT;
 932                 }
 933         }
 934
 935         rbp = lnet_msg2bufpool(msg);
 936
 937         if (!msg->msg_rtrcredit) {
 938                 msg->msg_rtrcredit = 1;
 939                 rbp->rbp_credits--;
 940                 if (rbp->rbp_credits < rbp->rbp_mincredits)
 941                         rbp->rbp_mincredits = rbp->rbp_credits;
 942
 943                 if (rbp->rbp_credits < 0) {
 944                         /* must have checked eager_recv before here */
 945                         LASSERT(msg->msg_rx_ready_delay);
 946                         msg->msg_rx_delayed = 1;
 947                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
 948                         return LNET_CREDIT_WAIT;
 949                 }
 950         }
 951
 952         LASSERT(!list_empty(&rbp->rbp_bufs));
 953         rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
 954         list_del(&rb->rb_list);
 955
 956         msg->msg_niov = rbp->rbp_npages;
 957         msg->msg_kiov = &rb->rb_kiov[0];
 958
 959         if (do_recv) {
 960                 int cpt = msg->msg_rx_cpt;
 961
 962                 lnet_net_unlock(cpt);
 963                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
 964                              0, msg->msg_len, msg->msg_len);
 965                 lnet_net_lock(cpt);
 966         }
 967         return LNET_CREDIT_OK;
 968 }
 969
 970 void
 971 lnet_return_tx_credits_locked(lnet_msg_t *msg)
 972 {
 973         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
 974         struct lnet_ni          *txni = msg->msg_txni;
 975         lnet_msg_t              *msg2;
 976
 977         if (msg->msg_txcredit) {
 978                 struct lnet_ni       *ni = msg->msg_txni;
 979                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
 980
 981                 /* give back NI txcredits */
 982                 msg->msg_txcredit = 0;
 983
 984                 LASSERT((tq->tq_credits < 0) ==
 985                         !list_empty(&tq->tq_delayed));
 986
 987                 tq->tq_credits++;
 988                 if (tq->tq_credits <= 0) {
 989                         msg2 = list_entry(tq->tq_delayed.next,
 990                                           lnet_msg_t, msg_list);
 991                         list_del(&msg2->msg_list);
 992
 993                         LASSERT(msg2->msg_txni == ni);
 994                         LASSERT(msg2->msg_tx_delayed);
 995
 996                         (void) lnet_post_send_locked(msg2, 1);
 997                 }
 998         }
 999
1000         if (msg->msg_peertxcredit) {
1001                 /* give back peer txcredits */
1002                 msg->msg_peertxcredit = 0;
1003
1004                 LASSERT((txpeer->lpni_txcredits < 0) ==
1005                         !list_empty(&txpeer->lpni_txq));
1006
1007                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1008                 LASSERT (txpeer->lpni_txqnob >= 0);
1009
1010                 txpeer->lpni_txcredits++;
1011                 if (txpeer->lpni_txcredits <= 0) {
1012                         msg2 = list_entry(txpeer->lpni_txq.next,
1013                                               lnet_msg_t, msg_list);
1014                         list_del(&msg2->msg_list);
1015
1016                         LASSERT(msg2->msg_txpeer == txpeer);
1017                         LASSERT(msg2->msg_tx_delayed);
1018
1019                         (void) lnet_post_send_locked(msg2, 1);
1020                 }
1021         }
1022
1023         if (txni != NULL) {
1024                 msg->msg_txni = NULL;
1025                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1026         }
1027
1028         if (txpeer != NULL) {
1029                 /*
1030                  * TODO:
1031                  * Once the patch for the health comes in we need to set
1032                  * the health of the peer ni to bad when we fail to send
1033                  * a message.
1034                  * int status = msg->msg_ev.status;
1035                  * if (status != 0)
1036                  *      lnet_set_peer_ni_health_locked(txpeer, false)
1037                  */
1038                 msg->msg_txpeer = NULL;
1039                 lnet_peer_ni_decref_locked(txpeer);
1040         }
1041 }
1042
1043 void
1044 lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
1045 {
1046         lnet_msg_t      *msg;
1047
1048         if (list_empty(&rbp->rbp_msgs))
1049                 return;
1050         msg = list_entry(rbp->rbp_msgs.next,
1051                          lnet_msg_t, msg_list);
1052         list_del(&msg->msg_list);
1053
1054         (void)lnet_post_routed_recv_locked(msg, 1);
1055 }
1056
1057 void
1058 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1059 {
1060         lnet_msg_t       *msg;
1061         lnet_msg_t       *tmp;
1062         struct list_head drop;
1063
1064         INIT_LIST_HEAD(&drop);
1065
1066         list_splice_init(list, &drop);
1067
1068         lnet_net_unlock(cpt);
1069
1070         list_for_each_entry_safe(msg, tmp, &drop, msg_list) {
1071                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1072                              0, 0, 0, msg->msg_hdr.payload_length);
1073                 list_del_init(&msg->msg_list);
1074                 lnet_finalize(NULL, msg, -ECANCELED);
1075         }
1076
1077         lnet_net_lock(cpt);
1078 }
1079
1080 void
1081 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1082 {
1083         struct lnet_peer_ni     *rxpeer = msg->msg_rxpeer;
1084         struct lnet_ni          *rxni = msg->msg_rxni;
1085         lnet_msg_t              *msg2;
1086
1087         if (msg->msg_rtrcredit) {
1088                 /* give back global router credits */
1089                 lnet_rtrbuf_t     *rb;
1090                 lnet_rtrbufpool_t *rbp;
1091
1092                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1093                  * there until it gets one allocated, or aborts the wait
1094                  * itself */
1095                 LASSERT(msg->msg_kiov != NULL);
1096
1097                 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1098                 rbp = rb->rb_pool;
1099
1100                 msg->msg_kiov = NULL;
1101                 msg->msg_rtrcredit = 0;
1102
1103                 LASSERT(rbp == lnet_msg2bufpool(msg));
1104
1105                 LASSERT((rbp->rbp_credits > 0) ==
1106                         !list_empty(&rbp->rbp_bufs));
1107
1108                 /* If routing is now turned off, we just drop this buffer and
1109                  * don't bother trying to return credits.  */
1110                 if (!the_lnet.ln_routing) {
1111                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1112                         goto routing_off;
1113                 }
1114
1115                 /* It is possible that a user has lowered the desired number of
1116                  * buffers in this pool.  Make sure we never put back
1117                  * more buffers than the stated number. */
1118                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1119                         /* Discard this buffer so we don't have too
1120                          * many. */
1121                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1122                         rbp->rbp_nbuffers--;
1123                 } else {
1124                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1125                         rbp->rbp_credits++;
1126                         if (rbp->rbp_credits <= 0)
1127                                 lnet_schedule_blocked_locked(rbp);
1128                 }
1129         }
1130
1131 routing_off:
1132         if (msg->msg_peerrtrcredit) {
1133                 /* give back peer router credits */
1134                 msg->msg_peerrtrcredit = 0;
1135
1136                 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1137                         !list_empty(&rxpeer->lpni_rtrq));
1138
1139                 rxpeer->lpni_rtrcredits++;
1140
1141                 /* drop all messages which are queued to be routed on that
1142                  * peer. */
1143                 if (!the_lnet.ln_routing) {
1144                         lnet_drop_routed_msgs_locked(&rxpeer->lpni_rtrq,
1145                                                      msg->msg_rx_cpt);
1146                 } else if (rxpeer->lpni_rtrcredits <= 0) {
1147                         msg2 = list_entry(rxpeer->lpni_rtrq.next,
1148                                           lnet_msg_t, msg_list);
1149                         list_del(&msg2->msg_list);
1150
1151                         (void) lnet_post_routed_recv_locked(msg2, 1);
1152                 }
1153         }
1154         if (rxni != NULL) {
1155                 msg->msg_rxni = NULL;
1156                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1157         }
1158         if (rxpeer != NULL) {
1159                 msg->msg_rxpeer = NULL;
1160                 lnet_peer_ni_decref_locked(rxpeer);
1161         }
1162 }
1163
1164 static int
1165 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1166 {
1167         if (p1->lpni_txqnob < p2->lpni_txqnob)
1168                 return 1;
1169
1170         if (p1->lpni_txqnob > p2->lpni_txqnob)
1171                 return -1;
1172
1173         if (p1->lpni_txcredits > p2->lpni_txcredits)
1174                 return 1;
1175
1176         if (p1->lpni_txcredits < p2->lpni_txcredits)
1177                 return -1;
1178
1179         return 0;
1180 }
1181
1182 static int
1183 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1184 {
1185         struct lnet_peer_ni *p1 = r1->lr_gateway;
1186         struct lnet_peer_ni *p2 = r2->lr_gateway;
1187         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1188         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1189         int rc;
1190
1191         if (r1->lr_priority < r2->lr_priority)
1192                 return 1;
1193
1194         if (r1->lr_priority > r2->lr_priority)
1195                 return -1;
1196
1197         if (r1_hops < r2_hops)
1198                 return 1;
1199
1200         if (r1_hops > r2_hops)
1201                 return -1;
1202
1203         rc = lnet_compare_peers(p1, p2);
1204         if (rc)
1205                 return rc;
1206
1207         if (r1->lr_seq - r2->lr_seq <= 0)
1208                 return 1;
1209
1210         return -1;
1211 }
1212
1213 static struct lnet_peer_ni *
1214 lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
1215                        lnet_nid_t rtr_nid)
1216 {
1217         lnet_remotenet_t        *rnet;
1218         lnet_route_t            *route;
1219         lnet_route_t            *best_route;
1220         lnet_route_t            *last_route;
1221         struct lnet_peer_ni     *lpni_best;
1222         struct lnet_peer_ni     *lp;
1223         int                     rc;
1224
1225         /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1226          * rtr_nid nid, otherwise find the best gateway I can use */
1227
1228         rnet = lnet_find_rnet_locked(LNET_NIDNET(target));
1229         if (rnet == NULL)
1230                 return NULL;
1231
1232         lpni_best = NULL;
1233         best_route = last_route = NULL;
1234         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1235                 lp = route->lr_gateway;
1236
1237                 if (!lnet_is_route_alive(route))
1238                         continue;
1239
1240                 if (net != NULL && lp->lpni_net != net)
1241                         continue;
1242
1243                 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1244                         return lp;
1245
1246                 if (lpni_best == NULL) {
1247                         best_route = last_route = route;
1248                         lpni_best = lp;
1249                         continue;
1250                 }
1251
1252                 /* no protection on below fields, but it's harmless */
1253                 if (last_route->lr_seq - route->lr_seq < 0)
1254                         last_route = route;
1255
1256                 rc = lnet_compare_routes(route, best_route);
1257                 if (rc < 0)
1258                         continue;
1259
1260                 best_route = route;
1261                 lpni_best = lp;
1262         }
1263
1264         /* set sequence number on the best router to the latest sequence + 1
1265          * so we can round-robin all routers, it's race and inaccurate but
1266          * harmless and functional  */
1267         if (best_route != NULL)
1268                 best_route->lr_seq = last_route->lr_seq + 1;
1269         return lpni_best;
1270 }
1271
1272 static int
1273 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
1274                     struct lnet_msg *msg, lnet_nid_t rtr_nid, bool *lo_sent)
1275 {
1276         struct lnet_ni          *best_ni = NULL;
1277         struct lnet_peer_ni     *best_lpni = NULL;
1278         struct lnet_peer_ni     *net_gw = NULL;
1279         struct lnet_peer_ni     *best_gw = NULL;
1280         struct lnet_peer_ni     *lpni;
1281         struct lnet_peer        *peer = NULL;
1282         struct lnet_peer_net    *peer_net;
1283         struct lnet_net         *local_net;
1284         struct lnet_ni          *ni = NULL;
1285         int                     cpt, cpt2, rc;
1286         bool                    routing = false;
1287         bool                    ni_is_pref = false;
1288         bool                    preferred = false;
1289         int                     best_credits = 0;
1290         __u32                   seq, seq2;
1291         int                     best_lpni_credits = INT_MIN;
1292         int                     md_cpt = 0;
1293         int                     shortest_distance = INT_MAX;
1294         int                     distance = 0;
1295         bool                    found_ir = false;
1296
1297 again:
1298         /*
1299          * get an initial CPT to use for locking. The idea here is not to
1300          * serialize the calls to select_pathway, so that as many
1301          * operations can run concurrently as possible. To do that we use
1302          * the CPT where this call is being executed. Later on when we
1303          * determine the CPT to use in lnet_message_commit, we switch the
1304          * lock and check if there was any configuration changes, if none,
1305          * then we proceed, if there is, then we'll need to update the cpt
1306          * and redo the operation.
1307          */
1308         cpt = lnet_net_lock_current();
1309
1310         best_gw = NULL;
1311         routing = false;
1312         local_net = NULL;
1313         best_ni = NULL;
1314         shortest_distance = INT_MAX;
1315         found_ir = false;
1316
1317         if (the_lnet.ln_shutdown) {
1318                 lnet_net_unlock(cpt);
1319                 return -ESHUTDOWN;
1320         }
1321
1322         if (msg->msg_md != NULL)
1323                 /* get the cpt of the MD, used during NUMA based selection */
1324                 md_cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
1325         else
1326                 md_cpt = CFS_CPT_ANY;
1327
1328         /*
1329          * initialize the variables which could be reused if we go to
1330          * again
1331          */
1332         lpni = NULL;
1333         seq = lnet_get_dlc_seq_locked();
1334
1335         rc = lnet_find_or_create_peer_locked(dst_nid, cpt, &peer);
1336         if (rc != 0) {
1337                 lnet_net_unlock(cpt);
1338                 return rc;
1339         }
1340
1341         /* If peer is not healthy then can not send anything to it */
1342         if (!lnet_is_peer_healthy_locked(peer)) {
1343                 lnet_net_unlock(cpt);
1344                 return -EHOSTUNREACH;
1345         }
1346
1347         /*
1348          * STEP 1: first jab at determineing best_ni
1349          * if src_nid is explicitly specified, then best_ni is already
1350          * pre-determiend for us. Otherwise we need to select the best
1351          * one to use later on
1352          */
1353         if (src_nid != LNET_NID_ANY) {
1354                 best_ni = lnet_nid2ni_locked(src_nid, cpt);
1355                 if (!best_ni) {
1356                         lnet_net_unlock(cpt);
1357                         LCONSOLE_WARN("Can't send to %s: src %s is not a "
1358                                       "local nid\n", libcfs_nid2str(dst_nid),
1359                                       libcfs_nid2str(src_nid));
1360                         return -EINVAL;
1361                 }
1362
1363                 if (best_ni->ni_net->net_id != LNET_NIDNET(dst_nid)) {
1364                         lnet_net_unlock(cpt);
1365                         LCONSOLE_WARN("No route to %s via from %s\n",
1366                                       libcfs_nid2str(dst_nid),
1367                                       libcfs_nid2str(src_nid));
1368                         return -EINVAL;
1369                 }
1370         }
1371
1372         if (best_ni == the_lnet.ln_loni) {
1373                 /* No send credit hassles with LOLND */
1374                 msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
1375                 if (!msg->msg_routing)
1376                         msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
1377                 msg->msg_target.nid = best_ni->ni_nid;
1378                 lnet_msg_commit(msg, cpt);
1379
1380                 lnet_ni_addref_locked(best_ni, cpt);
1381                 lnet_net_unlock(cpt);
1382                 msg->msg_txni = best_ni;
1383                 lnet_ni_send(best_ni, msg);
1384
1385                 *lo_sent = true;
1386                 return 0;
1387         }
1388
1389         if (best_ni)
1390                 goto pick_peer;
1391
1392         /*
1393          * Decide whether we need to route to peer_ni.
1394          * Get the local net that I need to be on to be able to directly
1395          * send to that peer.
1396          *
1397          * a. Find the peer which the dst_nid belongs to.
1398          * b. Iterate through each of the peer_nets/nis to decide
1399          * the best peer/local_ni pair to use
1400          */
1401         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_on_peer_list) {
1402                 if (!lnet_is_peer_net_healthy_locked(peer_net))
1403                         continue;
1404
1405                 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1406                 if (!local_net) {
1407                         /*
1408                          * go through each peer_ni on that peer_net and
1409                          * determine the best possible gw to go through
1410                          */
1411                         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
1412                                             lpni_on_peer_net_list) {
1413                                 net_gw = lnet_find_route_locked(NULL,
1414                                                                 lpni->lpni_nid,
1415                                                                 rtr_nid);
1416
1417                                 /*
1418                                  * if no route is found for that network then
1419                                  * move onto the next peer_ni in the peer
1420                                  */
1421                                 if (!net_gw)
1422                                         continue;
1423
1424                                 if (!best_gw) {
1425                                         best_gw = net_gw;
1426                                         best_lpni = lpni;
1427                                 } else  {
1428                                         rc = lnet_compare_peers(net_gw,
1429                                                                 best_gw);
1430                                         if (rc > 0) {
1431                                                 best_gw = net_gw;
1432                                                 best_lpni = lpni;
1433                                         }
1434                                 }
1435                         }
1436
1437                         if (!best_gw)
1438                                 continue;
1439
1440                         local_net = lnet_get_net_locked
1441                                         (LNET_NIDNET(best_gw->lpni_nid));
1442                         routing = true;
1443                 } else {
1444                         routing = false;
1445                         best_gw = NULL;
1446                 }
1447
1448                 /* no routable net found go on to a different net */
1449                 if (!local_net)
1450                         continue;
1451
1452                 /*
1453                  * Iterate through the NIs in this local Net and select
1454                  * the NI to send from. The selection is determined by
1455                  * these 3 criterion in the following priority:
1456                  *      1. NUMA
1457                  *      2. NI available credits
1458                  *      3. Round Robin
1459                  */
1460                 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1461                         if (!lnet_is_ni_healthy_locked(ni))
1462                                 continue;
1463
1464                         /*
1465                          * calculate the distance from the cpt on which
1466                          * the message memory is allocated to the CPT of
1467                          * the NI's physical device
1468                          */
1469                         distance = cfs_cpt_distance(lnet_cpt_table(),
1470                                                     md_cpt,
1471                                                     ni->dev_cpt);
1472
1473                         /*
1474                          * If we already have a closer NI within the NUMA
1475                          * range provided, then there is no need to
1476                          * consider the current NI. Move on to the next
1477                          * one.
1478                          */
1479                         if (distance > shortest_distance &&
1480                             distance > lnet_get_numa_range())
1481                                 continue;
1482
1483                         if (distance < shortest_distance &&
1484                             distance > lnet_get_numa_range()) {
1485                                 /*
1486                                  * The current NI is the closest one that we
1487                                  * have found, even though it's not in the
1488                                  * NUMA range specified. This occurs if
1489                                  * the NUMA range is less than the least
1490                                  * of the distances in the system.
1491                                  * In effect NUMA range consideration is
1492                                  * turned off.
1493                                  */
1494                                 shortest_distance = distance;
1495                         } else if ((distance <= shortest_distance &&
1496                                     distance < lnet_get_numa_range()) ||
1497                                    distance == shortest_distance) {
1498                                 /*
1499                                  * This NI is either within range or it's
1500                                  * equidistant. In both of these cases we
1501                                  * would want to select the NI based on
1502                                  * its available credits first, and then
1503                                  * via Round Robin.
1504                                  */
1505                                 if (distance <= shortest_distance &&
1506                                     distance < lnet_get_numa_range()) {
1507                                         /*
1508                                          * If this is the first NI that's
1509                                          * within range, then set the
1510                                          * shortest distance to the range
1511                                          * specified by the user. In
1512                                          * effect we're saying that all
1513                                          * NIs that fall within this NUMA
1514                                          * range shall be dealt with as
1515                                          * having equal NUMA weight. Which
1516                                          * will mean that we should select
1517                                          * through that set by their
1518                                          * available credits first
1519                                          * followed by Round Robin.
1520                                          *
1521                                          * And since this is the first NI
1522                                          * in the range, let's just set it
1523                                          * as our best_ni for now. The
1524                                          * following NIs found in the
1525                                          * range will be dealt with as
1526                                          * mentioned previously.
1527                                          */
1528                                         shortest_distance = lnet_get_numa_range();
1529                                         if (!found_ir) {
1530                                                 found_ir = true;
1531                                                 goto set_ni;
1532                                         }
1533                                 }
1534                                 /*
1535                                  * This NI is NUMA equidistant let's
1536                                  * select using credits followed by Round
1537                                  * Robin.
1538                                  */
1539                                 if (ni->ni_tx_queues[cpt]->tq_credits <
1540                                         best_credits) {
1541                                         continue;
1542                                 } else if (ni->ni_tx_queues[cpt]->tq_credits ==
1543                                                 best_credits) {
1544                                         if (best_ni) {
1545                                                 if (best_ni->ni_seq <= ni->ni_seq)
1546                                                         continue;
1547                                         }
1548                                 }
1549                         }
1550 set_ni:
1551                         best_ni = ni;
1552                         best_credits = ni->ni_tx_queues[cpt]->tq_credits;
1553                 }
1554         }
1555         /*
1556          * Now that we selected the NI to use increment its sequence
1557          * number so the Round Robin algorithm will detect that it has
1558          * been used and pick the next NI.
1559          */
1560         best_ni->ni_seq++;
1561
1562         if (!best_ni) {
1563                 lnet_net_unlock(cpt);
1564                 LCONSOLE_WARN("No local ni found to send from to %s\n",
1565                         libcfs_nid2str(dst_nid));
1566                 return -EINVAL;
1567         }
1568
1569         if (routing)
1570                 goto send;
1571
1572 pick_peer:
1573         lpni = NULL;
1574
1575         if (msg->msg_type == LNET_MSG_REPLY ||
1576             msg->msg_type == LNET_MSG_ACK) {
1577                 /*
1578                  * for replies we want to respond on the same peer_ni we
1579                  * received the message on if possible. If not, then pick
1580                  * a peer_ni to send to
1581                  */
1582                 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1583                 if (best_lpni) {
1584                         lnet_peer_ni_decref_locked(best_lpni);
1585                         goto send;
1586                 } else {
1587                         CDEBUG(D_NET, "unable to send msg_type %d to "
1588                               "originating %s\n", msg->msg_type,
1589                               libcfs_nid2str(dst_nid));
1590                 }
1591         }
1592
1593         peer_net = lnet_peer_get_net_locked(peer,
1594                                             best_ni->ni_net->net_id);
1595         /*
1596          * peer_net is not available or the src_nid is explicitly defined
1597          * and the peer_net for that src_nid is unhealthy. find a route to
1598          * the destination nid.
1599          */
1600         if (!peer_net ||
1601             (src_nid != LNET_NID_ANY &&
1602              !lnet_is_peer_net_healthy_locked(peer_net))) {
1603                 best_gw = lnet_find_route_locked(best_ni->ni_net,
1604                                                  dst_nid,
1605                                                  rtr_nid);
1606                 /*
1607                  * if no route is found for that network then
1608                  * move onto the next peer_ni in the peer
1609                  */
1610                 if (!best_gw) {
1611                         lnet_net_unlock(cpt);
1612                         LCONSOLE_WARN("No route to peer from %s\n",
1613                                 libcfs_nid2str(best_ni->ni_nid));
1614                         return -EHOSTUNREACH;
1615                 }
1616
1617                 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1618                         libcfs_nid2str(lpni->lpni_nid),
1619                         libcfs_nid2str(best_gw->lpni_nid),
1620                         lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1621
1622                 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1623                 LASSERT(best_lpni != NULL);
1624                 lnet_peer_ni_decref_locked(best_lpni);
1625
1626                 routing = true;
1627
1628                 goto send;
1629         } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
1630                 /*
1631                  * this peer_net is unhealthy but we still have an opportunity
1632                  * to find another peer_net that we can use
1633                  */
1634                 __u32 net_id = peer_net->lpn_net_id;
1635                 lnet_net_unlock(cpt);
1636                 if (!best_lpni)
1637                         LCONSOLE_WARN("peer net %s unhealthy\n",
1638                                       libcfs_net2str(net_id));
1639                 goto again;
1640         }
1641
1642         best_lpni = NULL;
1643         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1644                 /*
1645                  * if this peer ni is not healthy just skip it, no point in
1646                  * examining it further
1647                  */
1648                 if (!lnet_is_peer_ni_healthy_locked(lpni))
1649                         continue;
1650                 ni_is_pref = lnet_peer_is_ni_pref_locked(lpni, best_ni);
1651
1652                 /* if this is a preferred peer use it */
1653                 if (!preferred && ni_is_pref) {
1654                         preferred = true;
1655                 } else if (preferred && !ni_is_pref) {
1656                         /*
1657                          * this is not the preferred peer so let's ignore
1658                          * it.
1659                          */
1660                         continue;
1661                 } if (lpni->lpni_txcredits < best_lpni_credits)
1662                         /*
1663                          * We already have a peer that has more credits
1664                          * available than this one. No need to consider
1665                          * this peer further.
1666                          */
1667                         continue;
1668                 else if (lpni->lpni_txcredits == best_lpni_credits) {
1669                         /*
1670                          * The best peer found so far and the current peer
1671                          * have the same number of available credits let's
1672                          * make sure to select between them using Round
1673                          * Robin
1674                          */
1675                         if (best_lpni) {
1676                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1677                                         continue;
1678                         }
1679                 }
1680
1681                 best_lpni = lpni;
1682                 best_lpni_credits = lpni->lpni_txcredits;
1683         }
1684
1685         /*
1686          * Increment sequence number of the peer selected so that we can
1687          * pick the next one in Round Robin.
1688          */
1689         best_lpni->lpni_seq++;
1690
1691         /* if we still can't find a peer ni then we can't reach it */
1692         if (!best_lpni) {
1693                 __u32 net_id = peer_net->lpn_net_id;
1694                 lnet_net_unlock(cpt);
1695                 LCONSOLE_WARN("no peer_ni found on peer net %s\n",
1696                                 libcfs_net2str(net_id));
1697                 return -EHOSTUNREACH;
1698         }
1699
1700 send:
1701         /*
1702          * determine the cpt to use and if it has changed then
1703          * lock the new cpt and check if the config has changed.
1704          * If it has changed then repeat the algorithm since the
1705          * ni or peer list could have changed and the algorithm
1706          * would endup picking a different ni/peer_ni pair.
1707          */
1708         cpt2 = best_lpni->lpni_cpt;
1709         if (cpt != cpt2) {
1710                 lnet_net_unlock(cpt);
1711                 cpt = cpt2;
1712                 lnet_net_lock(cpt);
1713                 seq2 = lnet_get_dlc_seq_locked();
1714                 if (seq2 != seq) {
1715                         lnet_net_unlock(cpt);
1716                         goto again;
1717                 }
1718         }
1719
1720         /*
1721          * store the best_lpni in the message right away to avoid having
1722          * to do the same operation under different conditions
1723          */
1724         msg->msg_txpeer = (routing) ? best_gw : best_lpni;
1725         msg->msg_txni = best_ni;
1726         /*
1727          * grab a reference for the best_ni since now it's in use in this
1728          * send. the reference will need to be dropped when the message is
1729          * finished in lnet_finalize()
1730          */
1731         lnet_ni_addref_locked(msg->msg_txni, cpt);
1732         lnet_peer_ni_addref_locked(msg->msg_txpeer);
1733
1734         /*
1735          * set the destination nid in the message here because it's
1736          * possible that we'd be sending to a different nid than the one
1737          * originaly given.
1738          */
1739         msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1740
1741         /*
1742          * Always set the target.nid to the best peer picked. Either the
1743          * nid will be one of the preconfigured NIDs, or the same NID as
1744          * what was originaly set in the target or it will be the NID of
1745          * a router if this message should be routed
1746          */
1747         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1748
1749         /*
1750          * lnet_msg_commit assigns the correct cpt to the message, which
1751          * is used to decrement the correct refcount on the ni when it's
1752          * time to return the credits
1753          */
1754         lnet_msg_commit(msg, cpt);
1755
1756         /*
1757          * If we are routing the message then we don't need to overwrite
1758          * the src_nid since it would've been set at the origin. Otherwise
1759          * we are the originator so we need to set it.
1760          */
1761         if (!msg->msg_routing)
1762                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1763
1764         if (routing) {
1765                 msg->msg_target_is_router = 1;
1766                 msg->msg_target.pid = LNET_PID_LUSTRE;
1767         }
1768
1769         rc = lnet_post_send_locked(msg, 0);
1770
1771         lnet_net_unlock(cpt);
1772
1773         return rc;
1774 }
1775
1776 int
1777 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1778 {
1779         lnet_nid_t              dst_nid = msg->msg_target.nid;
1780         int                     rc;
1781         bool                    lo_sent = false;
1782
1783         /*
1784          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1785          * but we might want to use pre-determined router for ACK/REPLY
1786          * in the future
1787          */
1788         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1789         LASSERT (msg->msg_txpeer == NULL);
1790         LASSERT (!msg->msg_sending);
1791         LASSERT (!msg->msg_target_is_router);
1792         LASSERT (!msg->msg_receiving);
1793
1794         msg->msg_sending = 1;
1795
1796         LASSERT(!msg->msg_tx_committed);
1797
1798         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid, &lo_sent);
1799         if (rc < 0 || lo_sent)
1800                 return rc;
1801
1802         if (rc == LNET_CREDIT_OK)
1803                 lnet_ni_send(msg->msg_txni, msg);
1804
1805         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
1806         return 0;
1807 }
1808
1809 void
1810 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1811 {
1812         lnet_net_lock(cpt);
1813         the_lnet.ln_counters[cpt]->drop_count++;
1814         the_lnet.ln_counters[cpt]->drop_length += nob;
1815         lnet_net_unlock(cpt);
1816
1817         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1818 }
1819
1820 static void
1821 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1822 {
1823         lnet_hdr_t      *hdr = &msg->msg_hdr;
1824
1825         if (msg->msg_wanted != 0)
1826                 lnet_setpayloadbuffer(msg);
1827
1828         lnet_build_msg_event(msg, LNET_EVENT_PUT);
1829
1830         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
1831          * it back into the ACK during lnet_finalize() */
1832         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1833                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1834
1835         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1836                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1837 }
1838
1839 static int
1840 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1841 {
1842         lnet_hdr_t              *hdr = &msg->msg_hdr;
1843         struct lnet_match_info  info;
1844         int                     rc;
1845         bool                    ready_delay;
1846
1847         /* Convert put fields to host byte order */
1848         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1849         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
1850         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
1851
1852         info.mi_id.nid  = hdr->src_nid;
1853         info.mi_id.pid  = hdr->src_pid;
1854         info.mi_opc     = LNET_MD_OP_PUT;
1855         info.mi_portal  = hdr->msg.put.ptl_index;
1856         info.mi_rlength = hdr->payload_length;
1857         info.mi_roffset = hdr->msg.put.offset;
1858         info.mi_mbits   = hdr->msg.put.match_bits;
1859         info.mi_cpt     = msg->msg_rxpeer->lpni_cpt;
1860
1861         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
1862         ready_delay = msg->msg_rx_ready_delay;
1863
1864  again:
1865         rc = lnet_ptl_match_md(&info, msg);
1866         switch (rc) {
1867         default:
1868                 LBUG();
1869
1870         case LNET_MATCHMD_OK:
1871                 lnet_recv_put(ni, msg);
1872                 return 0;
1873
1874         case LNET_MATCHMD_NONE:
1875                 if (ready_delay)
1876                         /* no eager_recv or has already called it, should
1877                          * have been attached on delayed list */
1878                         return 0;
1879
1880                 rc = lnet_ni_eager_recv(ni, msg);
1881                 if (rc == 0) {
1882                         ready_delay = true;
1883                         goto again;
1884                 }
1885                 /* fall through */
1886
1887         case LNET_MATCHMD_DROP:
1888                 CNETERR("Dropping PUT from %s portal %d match %llu"
1889                         " offset %d length %d: %d\n",
1890                         libcfs_id2str(info.mi_id), info.mi_portal,
1891                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1892
1893                 return -ENOENT; /* -ve: OK but no match */
1894         }
1895 }
1896
1897 static int
1898 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1899 {
1900         struct lnet_match_info  info;
1901         lnet_hdr_t              *hdr = &msg->msg_hdr;
1902         struct lnet_handle_wire reply_wmd;
1903         int                     rc;
1904
1905         /* Convert get fields to host byte order */
1906         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
1907         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
1908         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
1909         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
1910
1911         info.mi_id.nid  = hdr->src_nid;
1912         info.mi_id.pid  = hdr->src_pid;
1913         info.mi_opc     = LNET_MD_OP_GET;
1914         info.mi_portal  = hdr->msg.get.ptl_index;
1915         info.mi_rlength = hdr->msg.get.sink_length;
1916         info.mi_roffset = hdr->msg.get.src_offset;
1917         info.mi_mbits   = hdr->msg.get.match_bits;
1918
1919         rc = lnet_ptl_match_md(&info, msg);
1920         if (rc == LNET_MATCHMD_DROP) {
1921                 CNETERR("Dropping GET from %s portal %d match %llu"
1922                         " offset %d length %d\n",
1923                         libcfs_id2str(info.mi_id), info.mi_portal,
1924                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
1925                 return -ENOENT; /* -ve: OK but no match */
1926         }
1927
1928         LASSERT(rc == LNET_MATCHMD_OK);
1929
1930         lnet_build_msg_event(msg, LNET_EVENT_GET);
1931
1932         reply_wmd = hdr->msg.get.return_wmd;
1933
1934         lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
1935                        msg->msg_offset, msg->msg_wanted);
1936
1937         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1938
1939         if (rdma_get) {
1940                 /* The LND completes the REPLY from her recv procedure */
1941                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1942                              msg->msg_offset, msg->msg_len, msg->msg_len);
1943                 return 0;
1944         }
1945
1946         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1947         msg->msg_receiving = 0;
1948
1949         rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1950         if (rc < 0) {
1951                 /* didn't get as far as lnet_ni_send() */
1952                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1953                        libcfs_nid2str(ni->ni_nid),
1954                        libcfs_id2str(info.mi_id), rc);
1955
1956                 lnet_finalize(ni, msg, rc);
1957         }
1958
1959         return 0;
1960 }
1961
1962 static int
1963 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1964 {
1965         void             *private = msg->msg_private;
1966         lnet_hdr_t       *hdr = &msg->msg_hdr;
1967         lnet_process_id_t src = {0};
1968         lnet_libmd_t     *md;
1969         int               rlength;
1970         int               mlength;
1971         int                     cpt;
1972
1973         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1974         lnet_res_lock(cpt);
1975
1976         src.nid = hdr->src_nid;
1977         src.pid = hdr->src_pid;
1978
1979         /* NB handles only looked up by creator (no flips) */
1980         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1981         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1982                 CNETERR("%s: Dropping REPLY from %s for %s "
1983                         "MD %#llx.%#llx\n",
1984                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1985                         (md == NULL) ? "invalid" : "inactive",
1986                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
1987                         hdr->msg.reply.dst_wmd.wh_object_cookie);
1988                 if (md != NULL && md->md_me != NULL)
1989                         CERROR("REPLY MD also attached to portal %d\n",
1990                                md->md_me->me_portal);
1991
1992                 lnet_res_unlock(cpt);
1993                 return -ENOENT; /* -ve: OK but no match */
1994         }
1995
1996         LASSERT(md->md_offset == 0);
1997
1998         rlength = hdr->payload_length;
1999         mlength = MIN(rlength, (int)md->md_length);
2000
2001         if (mlength < rlength &&
2002             (md->md_options & LNET_MD_TRUNCATE) == 0) {
2003                 CNETERR("%s: Dropping REPLY from %s length %d "
2004                         "for MD %#llx would overflow (%d)\n",
2005                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2006                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2007                         mlength);
2008                 lnet_res_unlock(cpt);
2009                 return -ENOENT; /* -ve: OK but no match */
2010         }
2011
2012         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
2013                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2014                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2015
2016         lnet_msg_attach_md(msg, md, 0, mlength);
2017
2018         if (mlength != 0)
2019                 lnet_setpayloadbuffer(msg);
2020
2021         lnet_res_unlock(cpt);
2022
2023         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2024
2025         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2026         return 0;
2027 }
2028
2029 static int
2030 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
2031 {
2032         lnet_hdr_t       *hdr = &msg->msg_hdr;
2033         lnet_process_id_t src = {0};
2034         lnet_libmd_t     *md;
2035         int                     cpt;
2036
2037         src.nid = hdr->src_nid;
2038         src.pid = hdr->src_pid;
2039
2040         /* Convert ack fields to host byte order */
2041         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2042         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2043
2044         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
2045         lnet_res_lock(cpt);
2046
2047         /* NB handles only looked up by creator (no flips) */
2048         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2049         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2050                 /* Don't moan; this is expected */
2051                 CDEBUG(D_NET,
2052                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
2053                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2054                        (md == NULL) ? "invalid" : "inactive",
2055                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
2056                        hdr->msg.ack.dst_wmd.wh_object_cookie);
2057                 if (md != NULL && md->md_me != NULL)
2058                         CERROR("Source MD also attached to portal %d\n",
2059                                md->md_me->me_portal);
2060
2061                 lnet_res_unlock(cpt);
2062                 return -ENOENT;                  /* -ve! */
2063         }
2064
2065         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
2066                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2067                hdr->msg.ack.dst_wmd.wh_object_cookie);
2068
2069         lnet_msg_attach_md(msg, md, 0, 0);
2070
2071         lnet_res_unlock(cpt);
2072
2073         lnet_build_msg_event(msg, LNET_EVENT_ACK);
2074
2075         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2076         return 0;
2077 }
2078
2079 /**
2080  * \retval LNET_CREDIT_OK       If \a msg is forwarded
2081  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
2082  * \retval -ve                  error code
2083  */
2084 int
2085 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
2086 {
2087         int     rc = 0;
2088
2089         if (!the_lnet.ln_routing)
2090                 return -ECANCELED;
2091
2092         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
2093             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2094                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
2095                         msg->msg_rx_ready_delay = 1;
2096                 } else {
2097                         lnet_net_unlock(msg->msg_rx_cpt);
2098                         rc = lnet_ni_eager_recv(ni, msg);
2099                         lnet_net_lock(msg->msg_rx_cpt);
2100                 }
2101         }
2102
2103         if (rc == 0)
2104                 rc = lnet_post_routed_recv_locked(msg, 0);
2105         return rc;
2106 }
2107
2108 int
2109 lnet_parse_local(lnet_ni_t *ni, lnet_msg_t *msg)
2110 {
2111         int     rc;
2112
2113         switch (msg->msg_type) {
2114         case LNET_MSG_ACK:
2115                 rc = lnet_parse_ack(ni, msg);
2116                 break;
2117         case LNET_MSG_PUT:
2118                 rc = lnet_parse_put(ni, msg);
2119                 break;
2120         case LNET_MSG_GET:
2121                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
2122                 break;
2123         case LNET_MSG_REPLY:
2124                 rc = lnet_parse_reply(ni, msg);
2125                 break;
2126         default: /* prevent an unused label if !kernel */
2127                 LASSERT(0);
2128                 return -EPROTO;
2129         }
2130
2131         LASSERT(rc == 0 || rc == -ENOENT);
2132         return rc;
2133 }
2134
2135 char *
2136 lnet_msgtyp2str (int type)
2137 {
2138         switch (type) {
2139         case LNET_MSG_ACK:
2140                 return ("ACK");
2141         case LNET_MSG_PUT:
2142                 return ("PUT");
2143         case LNET_MSG_GET:
2144                 return ("GET");
2145         case LNET_MSG_REPLY:
2146                 return ("REPLY");
2147         case LNET_MSG_HELLO:
2148                 return ("HELLO");
2149         default:
2150                 return ("<UNKNOWN>");
2151         }
2152 }
2153
2154 void
2155 lnet_print_hdr(lnet_hdr_t * hdr)
2156 {
2157         lnet_process_id_t src = {0};
2158         lnet_process_id_t dst = {0};
2159         char *type_str = lnet_msgtyp2str(hdr->type);
2160
2161         src.nid = hdr->src_nid;
2162         src.pid = hdr->src_pid;
2163
2164         dst.nid = hdr->dest_nid;
2165         dst.pid = hdr->dest_pid;
2166
2167         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2168         CWARN("    From %s\n", libcfs_id2str(src));
2169         CWARN("    To   %s\n", libcfs_id2str(dst));
2170
2171         switch (hdr->type) {
2172         default:
2173                 break;
2174
2175         case LNET_MSG_PUT:
2176                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
2177                       "match bits %llu\n",
2178                       hdr->msg.put.ptl_index,
2179                       hdr->msg.put.ack_wmd.wh_interface_cookie,
2180                       hdr->msg.put.ack_wmd.wh_object_cookie,
2181                       hdr->msg.put.match_bits);
2182                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
2183                       hdr->payload_length, hdr->msg.put.offset,
2184                       hdr->msg.put.hdr_data);
2185                 break;
2186
2187         case LNET_MSG_GET:
2188                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
2189                       "match bits %llu\n", hdr->msg.get.ptl_index,
2190                       hdr->msg.get.return_wmd.wh_interface_cookie,
2191                       hdr->msg.get.return_wmd.wh_object_cookie,
2192                       hdr->msg.get.match_bits);
2193                 CWARN("    Length %d, src offset %d\n",
2194                       hdr->msg.get.sink_length,
2195                       hdr->msg.get.src_offset);
2196                 break;
2197
2198         case LNET_MSG_ACK:
2199                 CWARN("    dst md %#llx.%#llx, "
2200                       "manipulated length %d\n",
2201                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
2202                       hdr->msg.ack.dst_wmd.wh_object_cookie,
2203                       hdr->msg.ack.mlength);
2204                 break;
2205
2206         case LNET_MSG_REPLY:
2207                 CWARN("    dst md %#llx.%#llx, "
2208                       "length %d\n",
2209                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
2210                       hdr->msg.reply.dst_wmd.wh_object_cookie,
2211                       hdr->payload_length);
2212         }
2213
2214 }
2215
2216 int
2217 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
2218            void *private, int rdma_req)
2219 {
2220         int             rc = 0;
2221         int             cpt;
2222         int             for_me;
2223         struct lnet_msg *msg;
2224         lnet_pid_t     dest_pid;
2225         lnet_nid_t     dest_nid;
2226         lnet_nid_t     src_nid;
2227         __u32          payload_length;
2228         __u32          type;
2229
2230         LASSERT (!in_interrupt ());
2231
2232         type = le32_to_cpu(hdr->type);
2233         src_nid = le64_to_cpu(hdr->src_nid);
2234         dest_nid = le64_to_cpu(hdr->dest_nid);
2235         dest_pid = le32_to_cpu(hdr->dest_pid);
2236         payload_length = le32_to_cpu(hdr->payload_length);
2237
2238         for_me = (ni->ni_nid == dest_nid);
2239         cpt = lnet_cpt_of_nid(from_nid, ni);
2240
2241         switch (type) {
2242         case LNET_MSG_ACK:
2243         case LNET_MSG_GET:
2244                 if (payload_length > 0) {
2245                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2246                                libcfs_nid2str(from_nid),
2247                                libcfs_nid2str(src_nid),
2248                                lnet_msgtyp2str(type), payload_length);
2249                         return -EPROTO;
2250                 }
2251                 break;
2252
2253         case LNET_MSG_PUT:
2254         case LNET_MSG_REPLY:
2255                 if (payload_length >
2256                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2257                         CERROR("%s, src %s: bad %s payload %d "
2258                                "(%d max expected)\n",
2259                                libcfs_nid2str(from_nid),
2260                                libcfs_nid2str(src_nid),
2261                                lnet_msgtyp2str(type),
2262                                payload_length,
2263                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2264                         return -EPROTO;
2265                 }
2266                 break;
2267
2268         default:
2269                 CERROR("%s, src %s: Bad message type 0x%x\n",
2270                        libcfs_nid2str(from_nid),
2271                        libcfs_nid2str(src_nid), type);
2272                 return -EPROTO;
2273         }
2274
2275         if (the_lnet.ln_routing &&
2276             ni->ni_last_alive != ktime_get_real_seconds()) {
2277                 /* NB: so far here is the only place to set NI status to "up */
2278                 lnet_ni_lock(ni);
2279                 ni->ni_last_alive = ktime_get_real_seconds();
2280                 if (ni->ni_status != NULL &&
2281                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2282                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2283                 lnet_ni_unlock(ni);
2284         }
2285
2286         /* Regard a bad destination NID as a protocol error.  Senders should
2287          * know what they're doing; if they don't they're misconfigured, buggy
2288          * or malicious so we chop them off at the knees :) */
2289
2290         if (!for_me) {
2291                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2292                         /* should have gone direct */
2293                         CERROR("%s, src %s: Bad dest nid %s "
2294                                "(should have been sent direct)\n",
2295                                 libcfs_nid2str(from_nid),
2296                                 libcfs_nid2str(src_nid),
2297                                 libcfs_nid2str(dest_nid));
2298                         return -EPROTO;
2299                 }
2300
2301                 if (lnet_islocalnid(dest_nid)) {
2302                         /* dest is another local NI; sender should have used
2303                          * this node's NID on its own network */
2304                         CERROR("%s, src %s: Bad dest nid %s "
2305                                "(it's my nid but on a different network)\n",
2306                                 libcfs_nid2str(from_nid),
2307                                 libcfs_nid2str(src_nid),
2308                                 libcfs_nid2str(dest_nid));
2309                         return -EPROTO;
2310                 }
2311
2312                 if (rdma_req && type == LNET_MSG_GET) {
2313                         CERROR("%s, src %s: Bad optimized GET for %s "
2314                                "(final destination must be me)\n",
2315                                 libcfs_nid2str(from_nid),
2316                                 libcfs_nid2str(src_nid),
2317                                 libcfs_nid2str(dest_nid));
2318                         return -EPROTO;
2319                 }
2320
2321                 if (!the_lnet.ln_routing) {
2322                         CERROR("%s, src %s: Dropping message for %s "
2323                                "(routing not enabled)\n",
2324                                 libcfs_nid2str(from_nid),
2325                                 libcfs_nid2str(src_nid),
2326                                 libcfs_nid2str(dest_nid));
2327                         goto drop;
2328                 }
2329         }
2330
2331         /* Message looks OK; we're not going to return an error, so we MUST
2332          * call back lnd_recv() come what may... */
2333
2334         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
2335             fail_peer(src_nid, 0)) {                    /* shall we now? */
2336                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2337                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2338                        lnet_msgtyp2str(type));
2339                 goto drop;
2340         }
2341
2342         if (!list_empty(&the_lnet.ln_drop_rules) &&
2343             lnet_drop_rule_match(hdr)) {
2344                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
2345                               "silent message loss\n",
2346                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2347                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
2348                 goto drop;
2349         }
2350
2351
2352         msg = lnet_msg_alloc();
2353         if (msg == NULL) {
2354                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2355                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2356                        lnet_msgtyp2str(type));
2357                 goto drop;
2358         }
2359
2360         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
2361          * pointers NULL etc */
2362
2363         msg->msg_type = type;
2364         msg->msg_private = private;
2365         msg->msg_receiving = 1;
2366         msg->msg_rdma_get = rdma_req;
2367         msg->msg_len = msg->msg_wanted = payload_length;
2368         msg->msg_offset = 0;
2369         msg->msg_hdr = *hdr;
2370         /* for building message event */
2371         msg->msg_from = from_nid;
2372         if (!for_me) {
2373                 msg->msg_target.pid     = dest_pid;
2374                 msg->msg_target.nid     = dest_nid;
2375                 msg->msg_routing        = 1;
2376
2377         } else {
2378                 /* convert common msg->hdr fields to host byteorder */
2379                 msg->msg_hdr.type       = type;
2380                 msg->msg_hdr.src_nid    = src_nid;
2381                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
2382                 msg->msg_hdr.dest_nid   = dest_nid;
2383                 msg->msg_hdr.dest_pid   = dest_pid;
2384                 msg->msg_hdr.payload_length = payload_length;
2385         }
2386
2387         lnet_net_lock(cpt);
2388         rc = lnet_nid2peerni_locked(&msg->msg_rxpeer, from_nid, cpt);
2389         if (rc != 0) {
2390                 lnet_net_unlock(cpt);
2391                 CERROR("%s, src %s: Dropping %s "
2392                        "(error %d looking up sender)\n",
2393                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2394                        lnet_msgtyp2str(type), rc);
2395                 lnet_msg_free(msg);
2396                 if (rc == -ESHUTDOWN)
2397                         /* We are shutting down.  Don't do anything more */
2398                         return 0;
2399                 goto drop;
2400         }
2401         msg->msg_rxni = ni;
2402         lnet_ni_addref_locked(ni, cpt);
2403
2404         if (lnet_isrouter(msg->msg_rxpeer)) {
2405                 lnet_peer_set_alive(msg->msg_rxpeer);
2406                 if (avoid_asym_router_failure &&
2407                     LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2408                         /* received a remote message from router, update
2409                          * remote NI status on this router.
2410                          * NB: multi-hop routed message will be ignored.
2411                          */
2412                         lnet_router_ni_update_locked(msg->msg_rxpeer,
2413                                                      LNET_NIDNET(src_nid));
2414                 }
2415         }
2416
2417         lnet_msg_commit(msg, cpt);
2418
2419         /* message delay simulation */
2420         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2421                      lnet_delay_rule_match_locked(hdr, msg))) {
2422                 lnet_net_unlock(cpt);
2423                 return 0;
2424         }
2425
2426         if (!for_me) {
2427                 rc = lnet_parse_forward_locked(ni, msg);
2428                 lnet_net_unlock(cpt);
2429
2430                 if (rc < 0)
2431                         goto free_drop;
2432
2433                 if (rc == LNET_CREDIT_OK) {
2434                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
2435                                      0, payload_length, payload_length);
2436                 }
2437                 return 0;
2438         }
2439
2440         lnet_net_unlock(cpt);
2441
2442         rc = lnet_parse_local(ni, msg);
2443         if (rc != 0)
2444                 goto free_drop;
2445         return 0;
2446
2447  free_drop:
2448         LASSERT(msg->msg_md == NULL);
2449         lnet_finalize(ni, msg, rc);
2450
2451  drop:
2452         lnet_drop_message(ni, cpt, private, payload_length);
2453         return 0;
2454 }
2455 EXPORT_SYMBOL(lnet_parse);
2456
2457 void
2458 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2459 {
2460         while (!list_empty(head)) {
2461                 lnet_process_id_t       id = {0};
2462                 lnet_msg_t              *msg;
2463
2464                 msg = list_entry(head->next, lnet_msg_t, msg_list);
2465                 list_del(&msg->msg_list);
2466
2467                 id.nid = msg->msg_hdr.src_nid;
2468                 id.pid = msg->msg_hdr.src_pid;
2469
2470                 LASSERT(msg->msg_md == NULL);
2471                 LASSERT(msg->msg_rx_delayed);
2472                 LASSERT(msg->msg_rxpeer != NULL);
2473                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2474
2475                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
2476                       " offset %d length %d: %s\n",
2477                       libcfs_id2str(id),
2478                       msg->msg_hdr.msg.put.ptl_index,
2479                       msg->msg_hdr.msg.put.match_bits,
2480                       msg->msg_hdr.msg.put.offset,
2481                       msg->msg_hdr.payload_length, reason);
2482
2483                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2484                  * called lnet_drop_message(), so I just hang onto msg as well
2485                  * until that's done */
2486
2487                 lnet_drop_message(msg->msg_rxni,
2488                                   msg->msg_rxpeer->lpni_cpt,
2489                                   msg->msg_private, msg->msg_len);
2490                 /*
2491                  * NB: message will not generate event because w/o attached MD,
2492                  * but we still should give error code so lnet_msg_decommit()
2493                  * can skip counters operations and other checks.
2494                  */
2495                 lnet_finalize(msg->msg_rxni, msg, -ENOENT);
2496         }
2497 }
2498
2499 void
2500 lnet_recv_delayed_msg_list(struct list_head *head)
2501 {
2502         while (!list_empty(head)) {
2503                 lnet_msg_t        *msg;
2504                 lnet_process_id_t  id;
2505
2506                 msg = list_entry(head->next, lnet_msg_t, msg_list);
2507                 list_del(&msg->msg_list);
2508
2509                 /* md won't disappear under me, since each msg
2510                  * holds a ref on it */
2511
2512                 id.nid = msg->msg_hdr.src_nid;
2513                 id.pid = msg->msg_hdr.src_pid;
2514
2515                 LASSERT(msg->msg_rx_delayed);
2516                 LASSERT(msg->msg_md != NULL);
2517                 LASSERT(msg->msg_rxpeer != NULL);
2518                 LASSERT(msg->msg_rxni != NULL);
2519                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2520
2521                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2522                        "match %llu offset %d length %d.\n",
2523                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2524                         msg->msg_hdr.msg.put.match_bits,
2525                         msg->msg_hdr.msg.put.offset,
2526                         msg->msg_hdr.payload_length);
2527
2528                 lnet_recv_put(msg->msg_rxni, msg);
2529         }
2530 }
2531
2532 /**
2533  * Initiate an asynchronous PUT operation.
2534  *
2535  * There are several events associated with a PUT: completion of the send on
2536  * the initiator node (LNET_EVENT_SEND), and when the send completes
2537  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2538  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2539  * used at the target node to indicate the completion of incoming data
2540  * delivery.
2541  *
2542  * The local events will be logged in the EQ associated with the MD pointed to
2543  * by \a mdh handle. Using a MD without an associated EQ results in these
2544  * events being discarded. In this case, the caller must have another
2545  * mechanism (e.g., a higher level protocol) for determining when it is safe
2546  * to modify the memory region associated with the MD.
2547  *
2548  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2549  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2550  *
2551  * \param self Indicates the NID of a local interface through which to send
2552  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2553  * \param mdh A handle for the MD that describes the memory to be sent. The MD
2554  * must be "free floating" (See LNetMDBind()).
2555  * \param ack Controls whether an acknowledgment is requested.
2556  * Acknowledgments are only sent when they are requested by the initiating
2557  * process and the target MD enables them.
2558  * \param target A process identifier for the target process.
2559  * \param portal The index in the \a target's portal table.
2560  * \param match_bits The match bits to use for MD selection at the target
2561  * process.
2562  * \param offset The offset into the target MD (only used when the target
2563  * MD has the LNET_MD_MANAGE_REMOTE option set).
2564  * \param hdr_data 64 bits of user data that can be included in the message
2565  * header. This data is written to an event queue entry at the target if an
2566  * EQ is present on the matching MD.
2567  *
2568  * \retval  0      Success, and only in this case events will be generated
2569  * and logged to EQ (if it exists).
2570  * \retval -EIO    Simulated failure.
2571  * \retval -ENOMEM Memory allocation failure.
2572  * \retval -ENOENT Invalid MD object.
2573  *
2574  * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2575  */
2576 int
2577 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2578         lnet_process_id_t target, unsigned int portal,
2579         __u64 match_bits, unsigned int offset,
2580         __u64 hdr_data)
2581 {
2582         struct lnet_msg         *msg;
2583         struct lnet_libmd       *md;
2584         int                     cpt;
2585         int                     rc;
2586
2587         LASSERT(the_lnet.ln_refcount > 0);
2588
2589         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
2590             fail_peer(target.nid, 1)) {                 /* shall we now? */
2591                 CERROR("Dropping PUT to %s: simulated failure\n",
2592                        libcfs_id2str(target));
2593                 return -EIO;
2594         }
2595
2596         msg = lnet_msg_alloc();
2597         if (msg == NULL) {
2598                 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2599                        libcfs_id2str(target));
2600                 return -ENOMEM;
2601         }
2602         msg->msg_vmflush = !!memory_pressure_get();
2603
2604         cpt = lnet_cpt_of_cookie(mdh.cookie);
2605         lnet_res_lock(cpt);
2606
2607         md = lnet_handle2md(&mdh);
2608         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2609                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
2610                        match_bits, portal, libcfs_id2str(target),
2611                        md == NULL ? -1 : md->md_threshold);
2612                 if (md != NULL && md->md_me != NULL)
2613                         CERROR("Source MD also attached to portal %d\n",
2614                                md->md_me->me_portal);
2615                 lnet_res_unlock(cpt);
2616
2617                 lnet_msg_free(msg);
2618                 return -ENOENT;
2619         }
2620
2621         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2622
2623         lnet_msg_attach_md(msg, md, 0, 0);
2624
2625         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2626
2627         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2628         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2629         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2630         msg->msg_hdr.msg.put.hdr_data = hdr_data;
2631
2632         /* NB handles only looked up by creator (no flips) */
2633         if (ack == LNET_ACK_REQ) {
2634                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2635                         the_lnet.ln_interface_cookie;
2636                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2637                         md->md_lh.lh_cookie;
2638         } else {
2639                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2640                         LNET_WIRE_HANDLE_COOKIE_NONE;
2641                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2642                         LNET_WIRE_HANDLE_COOKIE_NONE;
2643         }
2644
2645         lnet_res_unlock(cpt);
2646
2647         lnet_build_msg_event(msg, LNET_EVENT_SEND);
2648
2649         rc = lnet_send(self, msg, LNET_NID_ANY);
2650         if (rc != 0) {
2651                 CNETERR("Error sending PUT to %s: %d\n",
2652                         libcfs_id2str(target), rc);
2653                 lnet_finalize(NULL, msg, rc);
2654         }
2655
2656         /* completion will be signalled by an event */
2657         return 0;
2658 }
2659 EXPORT_SYMBOL(LNetPut);
2660
2661 lnet_msg_t *
2662 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2663 {
2664         /* The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
2665          * returns a msg for the LND to pass to lnet_finalize() when the sink
2666          * data has been received.
2667          *
2668          * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2669          * lnet_finalize() is called on it, so the LND must call this first */
2670
2671         struct lnet_msg         *msg = lnet_msg_alloc();
2672         struct lnet_libmd       *getmd = getmsg->msg_md;
2673         lnet_process_id_t       peer_id = getmsg->msg_target;
2674         int                     cpt;
2675
2676         LASSERT(!getmsg->msg_target_is_router);
2677         LASSERT(!getmsg->msg_routing);
2678
2679         if (msg == NULL) {
2680                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2681                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2682                 goto drop;
2683         }
2684
2685         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2686         lnet_res_lock(cpt);
2687
2688         LASSERT(getmd->md_refcount > 0);
2689
2690         if (getmd->md_threshold == 0) {
2691                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
2692                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2693                         getmd);
2694                 lnet_res_unlock(cpt);
2695                 goto drop;
2696         }
2697
2698         LASSERT(getmd->md_offset == 0);
2699
2700         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2701                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2702
2703         /* setup information for lnet_build_msg_event */
2704         msg->msg_from = peer_id.nid;
2705         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2706         msg->msg_hdr.src_nid = peer_id.nid;
2707         msg->msg_hdr.payload_length = getmd->md_length;
2708         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2709
2710         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2711         lnet_res_unlock(cpt);
2712
2713         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2714
2715         lnet_net_lock(cpt);
2716         lnet_msg_commit(msg, cpt);
2717         lnet_net_unlock(cpt);
2718
2719         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2720
2721         return msg;
2722
2723  drop:
2724         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2725
2726         lnet_net_lock(cpt);
2727         the_lnet.ln_counters[cpt]->drop_count++;
2728         the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2729         lnet_net_unlock(cpt);
2730
2731         if (msg != NULL)
2732                 lnet_msg_free(msg);
2733
2734         return NULL;
2735 }
2736 EXPORT_SYMBOL(lnet_create_reply_msg);
2737
2738 void
2739 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2740 {
2741         /* Set the REPLY length, now the RDMA that elides the REPLY message has
2742          * completed and I know it. */
2743         LASSERT(reply != NULL);
2744         LASSERT(reply->msg_type == LNET_MSG_GET);
2745         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
2746
2747         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
2748          * the end of my buffer, I might as well be dead. */
2749         LASSERT(len <= reply->msg_ev.mlength);
2750
2751         reply->msg_ev.mlength = len;
2752 }
2753 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2754
2755 /**
2756  * Initiate an asynchronous GET operation.
2757  *
2758  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2759  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2760  * the target node in the REPLY has been written to local MD.
2761  *
2762  * On the target node, an LNET_EVENT_GET is logged when the GET request
2763  * arrives and is accepted into a MD.
2764  *
2765  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2766  * \param mdh A handle for the MD that describes the memory into which the
2767  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2768  *
2769  * \retval  0      Success, and only in this case events will be generated
2770  * and logged to EQ (if it exists) of the MD.
2771  * \retval -EIO    Simulated failure.
2772  * \retval -ENOMEM Memory allocation failure.
2773  * \retval -ENOENT Invalid MD object.
2774  */
2775 int
2776 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2777         lnet_process_id_t target, unsigned int portal,
2778         __u64 match_bits, unsigned int offset)
2779 {
2780         struct lnet_msg         *msg;
2781         struct lnet_libmd       *md;
2782         int                     cpt;
2783         int                     rc;
2784
2785         LASSERT(the_lnet.ln_refcount > 0);
2786
2787         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
2788             fail_peer(target.nid, 1))                   /* shall we now? */
2789         {
2790                 CERROR("Dropping GET to %s: simulated failure\n",
2791                        libcfs_id2str(target));
2792                 return -EIO;
2793         }
2794
2795         msg = lnet_msg_alloc();
2796         if (msg == NULL) {
2797                 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2798                        libcfs_id2str(target));
2799                 return -ENOMEM;
2800         }
2801
2802         cpt = lnet_cpt_of_cookie(mdh.cookie);
2803         lnet_res_lock(cpt);
2804
2805         md = lnet_handle2md(&mdh);
2806         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2807                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
2808                        match_bits, portal, libcfs_id2str(target),
2809                        md == NULL ? -1 : md->md_threshold);
2810                 if (md != NULL && md->md_me != NULL)
2811                         CERROR("REPLY MD also attached to portal %d\n",
2812                                md->md_me->me_portal);
2813
2814                 lnet_res_unlock(cpt);
2815
2816                 lnet_msg_free(msg);
2817                 return -ENOENT;
2818         }
2819
2820         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2821
2822         lnet_msg_attach_md(msg, md, 0, 0);
2823
2824         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2825
2826         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2827         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2828         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2829         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2830
2831         /* NB handles only looked up by creator (no flips) */
2832         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2833                 the_lnet.ln_interface_cookie;
2834         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2835                 md->md_lh.lh_cookie;
2836
2837         lnet_res_unlock(cpt);
2838
2839         lnet_build_msg_event(msg, LNET_EVENT_SEND);
2840
2841         rc = lnet_send(self, msg, LNET_NID_ANY);
2842         if (rc < 0) {
2843                 CNETERR("Error sending GET to %s: %d\n",
2844                         libcfs_id2str(target), rc);
2845                 lnet_finalize(NULL, msg, rc);
2846         }
2847
2848         /* completion will be signalled by an event */
2849         return 0;
2850 }
2851 EXPORT_SYMBOL(LNetGet);
2852
2853 /**
2854  * Calculate distance to node at \a dstnid.
2855  *
2856  * \param dstnid Target NID.
2857  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2858  * is saved here.
2859  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2860  * here.
2861  *
2862  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2863  * local_nid_dist_zero is set, which is the default.
2864  * \retval positives Distance to target NID, i.e. number of hops plus one.
2865  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2866  */
2867 int
2868 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2869 {
2870         struct list_head        *e;
2871         struct lnet_ni          *ni = NULL;
2872         lnet_remotenet_t        *rnet;
2873         __u32                   dstnet = LNET_NIDNET(dstnid);
2874         int                     hops;
2875         int                     cpt;
2876         __u32                   order = 2;
2877         struct list_head        *rn_list;
2878
2879         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2880          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2881          * keep order 0 free for 0@lo and order 1 free for a local NID
2882          * match */
2883
2884         LASSERT(the_lnet.ln_refcount > 0);
2885
2886         cpt = lnet_net_lock_current();
2887
2888         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
2889                 if (ni->ni_nid == dstnid) {
2890                         if (srcnidp != NULL)
2891                                 *srcnidp = dstnid;
2892                         if (orderp != NULL) {
2893                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2894                                         *orderp = 0;
2895                                 else
2896                                         *orderp = 1;
2897                         }
2898                         lnet_net_unlock(cpt);
2899
2900                         return local_nid_dist_zero ? 0 : 1;
2901                 }
2902
2903                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2904                         /* Check if ni was originally created in
2905                          * current net namespace.
2906                          * If not, assign order above 0xffff0000,
2907                          * to make this ni not a priority. */
2908                         if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
2909                                 order += 0xffff0000;
2910
2911                         if (srcnidp != NULL)
2912                                 *srcnidp = ni->ni_nid;
2913                         if (orderp != NULL)
2914                                 *orderp = order;
2915                         lnet_net_unlock(cpt);
2916                         return 1;
2917                 }
2918
2919                 order++;
2920         }
2921
2922         rn_list = lnet_net2rnethash(dstnet);
2923         list_for_each(e, rn_list) {
2924                 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2925
2926                 if (rnet->lrn_net == dstnet) {
2927                         lnet_route_t *route;
2928                         lnet_route_t *shortest = NULL;
2929                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
2930                         __u32 route_hops;
2931
2932                         LASSERT(!list_empty(&rnet->lrn_routes));
2933
2934                         list_for_each_entry(route, &rnet->lrn_routes,
2935                                             lr_list) {
2936                                 route_hops = route->lr_hops;
2937                                 if (route_hops == LNET_UNDEFINED_HOPS)
2938                                         route_hops = 1;
2939                                 if (shortest == NULL ||
2940                                     route_hops < shortest_hops) {
2941                                         shortest = route;
2942                                         shortest_hops = route_hops;
2943                                 }
2944                         }
2945
2946                         LASSERT(shortest != NULL);
2947                         hops = shortest_hops;
2948                         if (srcnidp != NULL) {
2949                                 ni = lnet_get_next_ni_locked(
2950                                         shortest->lr_gateway->lpni_net,
2951                                         NULL);
2952                                 *srcnidp = ni->ni_nid;
2953                         }
2954                         if (orderp != NULL)
2955                                 *orderp = order;
2956                         lnet_net_unlock(cpt);
2957                         return hops + 1;
2958                 }
2959                 order++;
2960         }
2961
2962         lnet_net_unlock(cpt);
2963         return -EHOSTUNREACH;
2964 }
2965 EXPORT_SYMBOL(LNetDist);