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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/lib-move.c
  33  *
  34  * Data movement routines
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <linux/pagemap.h>
  40
  41 #include <lnet/lib-lnet.h>
  42 #include <linux/nsproxy.h>
  43 #include <net/net_namespace.h>
  44
  45 extern unsigned int lnet_current_net_count;
  46
  47 static int local_nid_dist_zero = 1;
  48 module_param(local_nid_dist_zero, int, 0444);
  49 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
  50
  51 struct lnet_send_data {
  52         struct lnet_ni *sd_best_ni;
  53         struct lnet_peer_ni *sd_best_lpni;
  54         struct lnet_peer_ni *sd_final_dst_lpni;
  55         struct lnet_peer *sd_peer;
  56         struct lnet_peer *sd_gw_peer;
  57         struct lnet_peer_ni *sd_gw_lpni;
  58         struct lnet_peer_net *sd_peer_net;
  59         struct lnet_msg *sd_msg;
  60         lnet_nid_t sd_dst_nid;
  61         lnet_nid_t sd_src_nid;
  62         lnet_nid_t sd_rtr_nid;
  63         int sd_cpt;
  64         int sd_md_cpt;
  65         __u32 sd_send_case;
  66 };
  67
  68 static inline struct lnet_comm_count *
  69 get_stats_counts(struct lnet_element_stats *stats,
  70                  enum lnet_stats_type stats_type)
  71 {
  72         switch (stats_type) {
  73         case LNET_STATS_TYPE_SEND:
  74                 return &stats->el_send_stats;
  75         case LNET_STATS_TYPE_RECV:
  76                 return &stats->el_recv_stats;
  77         case LNET_STATS_TYPE_DROP:
  78                 return &stats->el_drop_stats;
  79         default:
  80                 CERROR("Unknown stats type\n");
  81         }
  82
  83         return NULL;
  84 }
  85
  86 void lnet_incr_stats(struct lnet_element_stats *stats,
  87                      enum lnet_msg_type msg_type,
  88                      enum lnet_stats_type stats_type)
  89 {
  90         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
  91         if (!counts)
  92                 return;
  93
  94         switch (msg_type) {
  95         case LNET_MSG_ACK:
  96                 atomic_inc(&counts->co_ack_count);
  97                 break;
  98         case LNET_MSG_PUT:
  99                 atomic_inc(&counts->co_put_count);
 100                 break;
 101         case LNET_MSG_GET:
 102                 atomic_inc(&counts->co_get_count);
 103                 break;
 104         case LNET_MSG_REPLY:
 105                 atomic_inc(&counts->co_reply_count);
 106                 break;
 107         case LNET_MSG_HELLO:
 108                 atomic_inc(&counts->co_hello_count);
 109                 break;
 110         default:
 111                 CERROR("There is a BUG in the code. Unknown message type\n");
 112                 break;
 113         }
 114 }
 115
 116 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
 117                      enum lnet_stats_type stats_type)
 118 {
 119         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 120         if (!counts)
 121                 return 0;
 122
 123         return (atomic_read(&counts->co_ack_count) +
 124                 atomic_read(&counts->co_put_count) +
 125                 atomic_read(&counts->co_get_count) +
 126                 atomic_read(&counts->co_reply_count) +
 127                 atomic_read(&counts->co_hello_count));
 128 }
 129
 130 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
 131                                 struct lnet_comm_count *counts)
 132 {
 133         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
 134         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
 135         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
 136         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
 137         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
 138 }
 139
 140 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
 141                               struct lnet_element_stats *stats)
 142 {
 143         struct lnet_comm_count *counts;
 144
 145         LASSERT(msg_stats);
 146         LASSERT(stats);
 147
 148         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
 149         if (!counts)
 150                 return;
 151         assign_stats(&msg_stats->im_send_stats, counts);
 152
 153         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
 154         if (!counts)
 155                 return;
 156         assign_stats(&msg_stats->im_recv_stats, counts);
 157
 158         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
 159         if (!counts)
 160                 return;
 161         assign_stats(&msg_stats->im_drop_stats, counts);
 162 }
 163
 164 int
 165 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
 166 {
 167         struct lnet_test_peer *tp;
 168         struct list_head *el;
 169         struct list_head *next;
 170         struct list_head  cull;
 171
 172         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 173         if (threshold != 0) {
 174                 /* Adding a new entry */
 175                 LIBCFS_ALLOC(tp, sizeof(*tp));
 176                 if (tp == NULL)
 177                         return -ENOMEM;
 178
 179                 tp->tp_nid = nid;
 180                 tp->tp_threshold = threshold;
 181
 182                 lnet_net_lock(0);
 183                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
 184                 lnet_net_unlock(0);
 185                 return 0;
 186         }
 187
 188         /* removing entries */
 189         INIT_LIST_HEAD(&cull);
 190
 191         lnet_net_lock(0);
 192
 193         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 194                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 195
 196                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 197                     nid == LNET_NID_ANY ||      /* removing all entries */
 198                     tp->tp_nid == nid) {        /* matched this one */
 199                         list_del(&tp->tp_list);
 200                         list_add(&tp->tp_list, &cull);
 201                 }
 202         }
 203
 204         lnet_net_unlock(0);
 205
 206         while (!list_empty(&cull)) {
 207                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 208
 209                 list_del(&tp->tp_list);
 210                 LIBCFS_FREE(tp, sizeof(*tp));
 211         }
 212         return 0;
 213 }
 214
 215 static int
 216 fail_peer (lnet_nid_t nid, int outgoing)
 217 {
 218         struct lnet_test_peer *tp;
 219         struct list_head *el;
 220         struct list_head *next;
 221         struct list_head  cull;
 222         int               fail = 0;
 223
 224         INIT_LIST_HEAD(&cull);
 225
 226         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 227         lnet_net_lock(0);
 228
 229         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 230                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 231
 232                 if (tp->tp_threshold == 0) {
 233                         /* zombie entry */
 234                         if (outgoing) {
 235                                 /* only cull zombies on outgoing tests,
 236                                  * since we may be at interrupt priority on
 237                                  * incoming messages. */
 238                                 list_del(&tp->tp_list);
 239                                 list_add(&tp->tp_list, &cull);
 240                         }
 241                         continue;
 242                 }
 243
 244                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
 245                     nid == tp->tp_nid) {                /* fail this peer */
 246                         fail = 1;
 247
 248                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 249                                 tp->tp_threshold--;
 250                                 if (outgoing &&
 251                                     tp->tp_threshold == 0) {
 252                                         /* see above */
 253                                         list_del(&tp->tp_list);
 254                                         list_add(&tp->tp_list, &cull);
 255                                 }
 256                         }
 257                         break;
 258                 }
 259         }
 260
 261         lnet_net_unlock(0);
 262
 263         while (!list_empty(&cull)) {
 264                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 265                 list_del(&tp->tp_list);
 266
 267                 LIBCFS_FREE(tp, sizeof(*tp));
 268         }
 269
 270         return fail;
 271 }
 272
 273 unsigned int
 274 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 275 {
 276         unsigned int nob = 0;
 277
 278         LASSERT(niov == 0 || iov != NULL);
 279         while (niov-- > 0)
 280                 nob += (iov++)->iov_len;
 281
 282         return (nob);
 283 }
 284 EXPORT_SYMBOL(lnet_iov_nob);
 285
 286 void
 287 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 288                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 289                   unsigned int nob)
 290 {
 291         /* NB diov, siov are READ-ONLY */
 292         unsigned int  this_nob;
 293
 294         if (nob == 0)
 295                 return;
 296
 297         /* skip complete frags before 'doffset' */
 298         LASSERT(ndiov > 0);
 299         while (doffset >= diov->iov_len) {
 300                 doffset -= diov->iov_len;
 301                 diov++;
 302                 ndiov--;
 303                 LASSERT(ndiov > 0);
 304         }
 305
 306         /* skip complete frags before 'soffset' */
 307         LASSERT(nsiov > 0);
 308         while (soffset >= siov->iov_len) {
 309                 soffset -= siov->iov_len;
 310                 siov++;
 311                 nsiov--;
 312                 LASSERT(nsiov > 0);
 313         }
 314
 315         do {
 316                 LASSERT(ndiov > 0);
 317                 LASSERT(nsiov > 0);
 318                 this_nob = MIN(diov->iov_len - doffset,
 319                                siov->iov_len - soffset);
 320                 this_nob = MIN(this_nob, nob);
 321
 322                 memcpy((char *)diov->iov_base + doffset,
 323                        (char *)siov->iov_base + soffset, this_nob);
 324                 nob -= this_nob;
 325
 326                 if (diov->iov_len > doffset + this_nob) {
 327                         doffset += this_nob;
 328                 } else {
 329                         diov++;
 330                         ndiov--;
 331                         doffset = 0;
 332                 }
 333
 334                 if (siov->iov_len > soffset + this_nob) {
 335                         soffset += this_nob;
 336                 } else {
 337                         siov++;
 338                         nsiov--;
 339                         soffset = 0;
 340                 }
 341         } while (nob > 0);
 342 }
 343 EXPORT_SYMBOL(lnet_copy_iov2iov);
 344
 345 int
 346 lnet_extract_iov(int dst_niov, struct kvec *dst,
 347                  int src_niov, struct kvec *src,
 348                  unsigned int offset, unsigned int len)
 349 {
 350         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 351          * for exactly 'len' bytes, and return the number of entries.
 352          * NB not destructive to 'src' */
 353         unsigned int    frag_len;
 354         unsigned int    niov;
 355
 356         if (len == 0)                           /* no data => */
 357                 return (0);                     /* no frags */
 358
 359         LASSERT(src_niov > 0);
 360         while (offset >= src->iov_len) {      /* skip initial frags */
 361                 offset -= src->iov_len;
 362                 src_niov--;
 363                 src++;
 364                 LASSERT(src_niov > 0);
 365         }
 366
 367         niov = 1;
 368         for (;;) {
 369                 LASSERT(src_niov > 0);
 370                 LASSERT((int)niov <= dst_niov);
 371
 372                 frag_len = src->iov_len - offset;
 373                 dst->iov_base = ((char *)src->iov_base) + offset;
 374
 375                 if (len <= frag_len) {
 376                         dst->iov_len = len;
 377                         return (niov);
 378                 }
 379
 380                 dst->iov_len = frag_len;
 381
 382                 len -= frag_len;
 383                 dst++;
 384                 src++;
 385                 niov++;
 386                 src_niov--;
 387                 offset = 0;
 388         }
 389 }
 390 EXPORT_SYMBOL(lnet_extract_iov);
 391
 392
 393 unsigned int
 394 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
 395 {
 396         unsigned int  nob = 0;
 397
 398         LASSERT(niov == 0 || kiov != NULL);
 399         while (niov-- > 0)
 400                 nob += (kiov++)->kiov_len;
 401
 402         return (nob);
 403 }
 404 EXPORT_SYMBOL(lnet_kiov_nob);
 405
 406 void
 407 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
 408                     unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
 409                     unsigned int nob)
 410 {
 411         /* NB diov, siov are READ-ONLY */
 412         unsigned int    this_nob;
 413         char           *daddr = NULL;
 414         char           *saddr = NULL;
 415
 416         if (nob == 0)
 417                 return;
 418
 419         LASSERT (!in_interrupt ());
 420
 421         LASSERT (ndiov > 0);
 422         while (doffset >= diov->kiov_len) {
 423                 doffset -= diov->kiov_len;
 424                 diov++;
 425                 ndiov--;
 426                 LASSERT(ndiov > 0);
 427         }
 428
 429         LASSERT(nsiov > 0);
 430         while (soffset >= siov->kiov_len) {
 431                 soffset -= siov->kiov_len;
 432                 siov++;
 433                 nsiov--;
 434                 LASSERT(nsiov > 0);
 435         }
 436
 437         do {
 438                 LASSERT(ndiov > 0);
 439                 LASSERT(nsiov > 0);
 440                 this_nob = MIN(diov->kiov_len - doffset,
 441                                siov->kiov_len - soffset);
 442                 this_nob = MIN(this_nob, nob);
 443
 444                 if (daddr == NULL)
 445                         daddr = ((char *)kmap(diov->kiov_page)) +
 446                                 diov->kiov_offset + doffset;
 447                 if (saddr == NULL)
 448                         saddr = ((char *)kmap(siov->kiov_page)) +
 449                                 siov->kiov_offset + soffset;
 450
 451                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 452                  * However in practice at least one of the kiovs will be mapped
 453                  * kernel pages and the map/unmap will be NOOPs */
 454
 455                 memcpy (daddr, saddr, this_nob);
 456                 nob -= this_nob;
 457
 458                 if (diov->kiov_len > doffset + this_nob) {
 459                         daddr += this_nob;
 460                         doffset += this_nob;
 461                 } else {
 462                         kunmap(diov->kiov_page);
 463                         daddr = NULL;
 464                         diov++;
 465                         ndiov--;
 466                         doffset = 0;
 467                 }
 468
 469                 if (siov->kiov_len > soffset + this_nob) {
 470                         saddr += this_nob;
 471                         soffset += this_nob;
 472                 } else {
 473                         kunmap(siov->kiov_page);
 474                         saddr = NULL;
 475                         siov++;
 476                         nsiov--;
 477                         soffset = 0;
 478                 }
 479         } while (nob > 0);
 480
 481         if (daddr != NULL)
 482                 kunmap(diov->kiov_page);
 483         if (saddr != NULL)
 484                 kunmap(siov->kiov_page);
 485 }
 486 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 487
 488 void
 489 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 490                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 491                     unsigned int nob)
 492 {
 493         /* NB iov, kiov are READ-ONLY */
 494         unsigned int    this_nob;
 495         char           *addr = NULL;
 496
 497         if (nob == 0)
 498                 return;
 499
 500         LASSERT (!in_interrupt ());
 501
 502         LASSERT (niov > 0);
 503         while (iovoffset >= iov->iov_len) {
 504                 iovoffset -= iov->iov_len;
 505                 iov++;
 506                 niov--;
 507                 LASSERT(niov > 0);
 508         }
 509
 510         LASSERT(nkiov > 0);
 511         while (kiovoffset >= kiov->kiov_len) {
 512                 kiovoffset -= kiov->kiov_len;
 513                 kiov++;
 514                 nkiov--;
 515                 LASSERT(nkiov > 0);
 516         }
 517
 518         do {
 519                 LASSERT(niov > 0);
 520                 LASSERT(nkiov > 0);
 521                 this_nob = MIN(iov->iov_len - iovoffset,
 522                                kiov->kiov_len - kiovoffset);
 523                 this_nob = MIN(this_nob, nob);
 524
 525                 if (addr == NULL)
 526                         addr = ((char *)kmap(kiov->kiov_page)) +
 527                                 kiov->kiov_offset + kiovoffset;
 528
 529                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 530                 nob -= this_nob;
 531
 532                 if (iov->iov_len > iovoffset + this_nob) {
 533                         iovoffset += this_nob;
 534                 } else {
 535                         iov++;
 536                         niov--;
 537                         iovoffset = 0;
 538                 }
 539
 540                 if (kiov->kiov_len > kiovoffset + this_nob) {
 541                         addr += this_nob;
 542                         kiovoffset += this_nob;
 543                 } else {
 544                         kunmap(kiov->kiov_page);
 545                         addr = NULL;
 546                         kiov++;
 547                         nkiov--;
 548                         kiovoffset = 0;
 549                 }
 550
 551         } while (nob > 0);
 552
 553         if (addr != NULL)
 554                 kunmap(kiov->kiov_page);
 555 }
 556 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 557
 558 void
 559 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 560                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 561                    unsigned int nob)
 562 {
 563         /* NB kiov, iov are READ-ONLY */
 564         unsigned int    this_nob;
 565         char           *addr = NULL;
 566
 567         if (nob == 0)
 568                 return;
 569
 570         LASSERT (!in_interrupt ());
 571
 572         LASSERT (nkiov > 0);
 573         while (kiovoffset >= kiov->kiov_len) {
 574                 kiovoffset -= kiov->kiov_len;
 575                 kiov++;
 576                 nkiov--;
 577                 LASSERT(nkiov > 0);
 578         }
 579
 580         LASSERT(niov > 0);
 581         while (iovoffset >= iov->iov_len) {
 582                 iovoffset -= iov->iov_len;
 583                 iov++;
 584                 niov--;
 585                 LASSERT(niov > 0);
 586         }
 587
 588         do {
 589                 LASSERT(nkiov > 0);
 590                 LASSERT(niov > 0);
 591                 this_nob = MIN(kiov->kiov_len - kiovoffset,
 592                                iov->iov_len - iovoffset);
 593                 this_nob = MIN(this_nob, nob);
 594
 595                 if (addr == NULL)
 596                         addr = ((char *)kmap(kiov->kiov_page)) +
 597                                 kiov->kiov_offset + kiovoffset;
 598
 599                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 600                 nob -= this_nob;
 601
 602                 if (kiov->kiov_len > kiovoffset + this_nob) {
 603                         addr += this_nob;
 604                         kiovoffset += this_nob;
 605                 } else {
 606                         kunmap(kiov->kiov_page);
 607                         addr = NULL;
 608                         kiov++;
 609                         nkiov--;
 610                         kiovoffset = 0;
 611                 }
 612
 613                 if (iov->iov_len > iovoffset + this_nob) {
 614                         iovoffset += this_nob;
 615                 } else {
 616                         iov++;
 617                         niov--;
 618                         iovoffset = 0;
 619                 }
 620         } while (nob > 0);
 621
 622         if (addr != NULL)
 623                 kunmap(kiov->kiov_page);
 624 }
 625 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 626
 627 int
 628 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
 629                   int src_niov, lnet_kiov_t *src,
 630                   unsigned int offset, unsigned int len)
 631 {
 632         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 633          * for exactly 'len' bytes, and return the number of entries.
 634          * NB not destructive to 'src' */
 635         unsigned int    frag_len;
 636         unsigned int    niov;
 637
 638         if (len == 0)                           /* no data => */
 639                 return (0);                     /* no frags */
 640
 641         LASSERT(src_niov > 0);
 642         while (offset >= src->kiov_len) {      /* skip initial frags */
 643                 offset -= src->kiov_len;
 644                 src_niov--;
 645                 src++;
 646                 LASSERT(src_niov > 0);
 647         }
 648
 649         niov = 1;
 650         for (;;) {
 651                 LASSERT(src_niov > 0);
 652                 LASSERT((int)niov <= dst_niov);
 653
 654                 frag_len = src->kiov_len - offset;
 655                 dst->kiov_page = src->kiov_page;
 656                 dst->kiov_offset = src->kiov_offset + offset;
 657
 658                 if (len <= frag_len) {
 659                         dst->kiov_len = len;
 660                         LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 661                         return niov;
 662                 }
 663
 664                 dst->kiov_len = frag_len;
 665                 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 666
 667                 len -= frag_len;
 668                 dst++;
 669                 src++;
 670                 niov++;
 671                 src_niov--;
 672                 offset = 0;
 673         }
 674 }
 675 EXPORT_SYMBOL(lnet_extract_kiov);
 676
 677 void
 678 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 679              int delayed, unsigned int offset, unsigned int mlen,
 680              unsigned int rlen)
 681 {
 682         unsigned int  niov = 0;
 683         struct kvec *iov = NULL;
 684         lnet_kiov_t  *kiov = NULL;
 685         int           rc;
 686
 687         LASSERT (!in_interrupt ());
 688         LASSERT (mlen == 0 || msg != NULL);
 689
 690         if (msg != NULL) {
 691                 LASSERT(msg->msg_receiving);
 692                 LASSERT(!msg->msg_sending);
 693                 LASSERT(rlen == msg->msg_len);
 694                 LASSERT(mlen <= msg->msg_len);
 695                 LASSERT(msg->msg_offset == offset);
 696                 LASSERT(msg->msg_wanted == mlen);
 697
 698                 msg->msg_receiving = 0;
 699
 700                 if (mlen != 0) {
 701                         niov = msg->msg_niov;
 702                         iov  = msg->msg_iov;
 703                         kiov = msg->msg_kiov;
 704
 705                         LASSERT (niov > 0);
 706                         LASSERT ((iov == NULL) != (kiov == NULL));
 707                 }
 708         }
 709
 710         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 711                                              niov, iov, kiov, offset, mlen,
 712                                              rlen);
 713         if (rc < 0)
 714                 lnet_finalize(msg, rc);
 715 }
 716
 717 static void
 718 lnet_setpayloadbuffer(struct lnet_msg *msg)
 719 {
 720         struct lnet_libmd *md = msg->msg_md;
 721
 722         LASSERT(msg->msg_len > 0);
 723         LASSERT(!msg->msg_routing);
 724         LASSERT(md != NULL);
 725         LASSERT(msg->msg_niov == 0);
 726         LASSERT(msg->msg_iov == NULL);
 727         LASSERT(msg->msg_kiov == NULL);
 728
 729         msg->msg_niov = md->md_niov;
 730         if ((md->md_options & LNET_MD_KIOV) != 0)
 731                 msg->msg_kiov = md->md_iov.kiov;
 732         else
 733                 msg->msg_iov = md->md_iov.iov;
 734 }
 735
 736 void
 737 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 738                unsigned int offset, unsigned int len)
 739 {
 740         msg->msg_type = type;
 741         msg->msg_target = target;
 742         msg->msg_len = len;
 743         msg->msg_offset = offset;
 744
 745         if (len != 0)
 746                 lnet_setpayloadbuffer(msg);
 747
 748         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 749         msg->msg_hdr.type           = cpu_to_le32(type);
 750         /* dest_nid will be overwritten by lnet_select_pathway() */
 751         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 752         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 753         /* src_nid will be set later */
 754         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 755         msg->msg_hdr.payload_length = cpu_to_le32(len);
 756 }
 757
 758 static void
 759 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 760 {
 761         void   *priv = msg->msg_private;
 762         int rc;
 763
 764         LASSERT (!in_interrupt ());
 765         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
 766                  (msg->msg_txcredit && msg->msg_peertxcredit));
 767
 768         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 769         if (rc < 0) {
 770                 msg->msg_no_resend = true;
 771                 lnet_finalize(msg, rc);
 772         }
 773 }
 774
 775 static int
 776 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 777 {
 778         int     rc;
 779
 780         LASSERT(!msg->msg_sending);
 781         LASSERT(msg->msg_receiving);
 782         LASSERT(!msg->msg_rx_ready_delay);
 783         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 784
 785         msg->msg_rx_ready_delay = 1;
 786         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 787                                                   &msg->msg_private);
 788         if (rc != 0) {
 789                 CERROR("recv from %s / send to %s aborted: "
 790                        "eager_recv failed %d\n",
 791                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 792                        libcfs_id2str(msg->msg_target), rc);
 793                 LASSERT(rc < 0); /* required by my callers */
 794         }
 795
 796         return rc;
 797 }
 798
 799 static bool
 800 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 801 {
 802         time64_t deadline;
 803
 804         deadline = lpni->lpni_last_alive +
 805                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 806
 807         /*
 808          * assume peer_ni is alive as long as we're within the configured
 809          * peer timeout
 810          */
 811         if (deadline > now)
 812                 return false;
 813
 814         return true;
 815 }
 816
 817 /* NB: returns 1 when alive, 0 when dead, negative when error;
 818  *     may drop the lnet_net_lock */
 819 static int
 820 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 821                        struct lnet_msg *msg)
 822 {
 823         time64_t now = ktime_get_seconds();
 824
 825         if (!lnet_peer_aliveness_enabled(lpni))
 826                 return -ENODEV;
 827
 828         /*
 829          * If we're resending a message, let's attempt to send it even if
 830          * the peer is down to fulfill our resend quota on the message
 831          */
 832         if (msg->msg_retry_count > 0)
 833                 return 1;
 834
 835         /* try and send recovery messages irregardless */
 836         if (msg->msg_recovery)
 837                 return 1;
 838
 839         /* always send any responses */
 840         if (msg->msg_type == LNET_MSG_ACK ||
 841             msg->msg_type == LNET_MSG_REPLY)
 842                 return 1;
 843
 844         if (!lnet_is_peer_deadline_passed(lpni, now))
 845                 return true;
 846
 847         return lnet_is_peer_ni_alive(lpni);
 848 }
 849
 850 /**
 851  * \param msg The message to be sent.
 852  * \param do_send True if lnet_ni_send() should be called in this function.
 853  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 854  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 855  *
 856  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 857  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 858  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 859  * \retval -ECANCELED If the MD of the message has been unlinked.
 860  */
 861 static int
 862 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 863 {
 864         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 865         struct lnet_ni          *ni = msg->msg_txni;
 866         int                     cpt = msg->msg_tx_cpt;
 867         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 868
 869         /* non-lnet_send() callers have checked before */
 870         LASSERT(!do_send || msg->msg_tx_delayed);
 871         LASSERT(!msg->msg_receiving);
 872         LASSERT(msg->msg_tx_committed);
 873         /* can't get here if we're sending to the loopback interface */
 874         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
 875
 876         /* NB 'lp' is always the next hop */
 877         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 878             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 879                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 880                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 881                         msg->msg_len;
 882                 lnet_net_unlock(cpt);
 883                 if (msg->msg_txpeer)
 884                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 885                                         msg->msg_type,
 886                                         LNET_STATS_TYPE_DROP);
 887                 if (msg->msg_txni)
 888                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 889                                         msg->msg_type,
 890                                         LNET_STATS_TYPE_DROP);
 891
 892                 CNETERR("Dropping message for %s: peer not alive\n",
 893                         libcfs_id2str(msg->msg_target));
 894                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 895                 if (do_send)
 896                         lnet_finalize(msg, -EHOSTUNREACH);
 897
 898                 lnet_net_lock(cpt);
 899                 return -EHOSTUNREACH;
 900         }
 901
 902         if (msg->msg_md != NULL &&
 903             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 904                 lnet_net_unlock(cpt);
 905
 906                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 907                         "called on the MD/ME.\n",
 908                         libcfs_id2str(msg->msg_target));
 909                 if (do_send) {
 910                         msg->msg_no_resend = true;
 911                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 912                                msg, libcfs_id2str(msg->msg_target));
 913                         lnet_finalize(msg, -ECANCELED);
 914                 }
 915
 916                 lnet_net_lock(cpt);
 917                 return -ECANCELED;
 918         }
 919
 920         if (!msg->msg_peertxcredit) {
 921                 spin_lock(&lp->lpni_lock);
 922                 LASSERT((lp->lpni_txcredits < 0) ==
 923                         !list_empty(&lp->lpni_txq));
 924
 925                 msg->msg_peertxcredit = 1;
 926                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 927                 lp->lpni_txcredits--;
 928
 929                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 930                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 931
 932                 if (lp->lpni_txcredits < 0) {
 933                         msg->msg_tx_delayed = 1;
 934                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 935                         spin_unlock(&lp->lpni_lock);
 936                         return LNET_CREDIT_WAIT;
 937                 }
 938                 spin_unlock(&lp->lpni_lock);
 939         }
 940
 941         if (!msg->msg_txcredit) {
 942                 LASSERT((tq->tq_credits < 0) ==
 943                         !list_empty(&tq->tq_delayed));
 944
 945                 msg->msg_txcredit = 1;
 946                 tq->tq_credits--;
 947                 atomic_dec(&ni->ni_tx_credits);
 948
 949                 if (tq->tq_credits < tq->tq_credits_min)
 950                         tq->tq_credits_min = tq->tq_credits;
 951
 952                 if (tq->tq_credits < 0) {
 953                         msg->msg_tx_delayed = 1;
 954                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 955                         return LNET_CREDIT_WAIT;
 956                 }
 957         }
 958
 959         /* unset the tx_delay flag as we're going to send it now */
 960         msg->msg_tx_delayed = 0;
 961
 962         if (do_send) {
 963                 lnet_net_unlock(cpt);
 964                 lnet_ni_send(ni, msg);
 965                 lnet_net_lock(cpt);
 966         }
 967         return LNET_CREDIT_OK;
 968 }
 969
 970
 971 static struct lnet_rtrbufpool *
 972 lnet_msg2bufpool(struct lnet_msg *msg)
 973 {
 974         struct lnet_rtrbufpool  *rbp;
 975         int                     cpt;
 976
 977         LASSERT(msg->msg_rx_committed);
 978
 979         cpt = msg->msg_rx_cpt;
 980         rbp = &the_lnet.ln_rtrpools[cpt][0];
 981
 982         LASSERT(msg->msg_len <= LNET_MTU);
 983         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 984                 rbp++;
 985                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 986         }
 987
 988         return rbp;
 989 }
 990
 991 static int
 992 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 993 {
 994         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 995          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 996          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 997          * received or OK to receive */
 998         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 999         struct lnet_peer *lp;
1000         struct lnet_rtrbufpool *rbp;
1001         struct lnet_rtrbuf *rb;
1002
1003         LASSERT(msg->msg_iov == NULL);
1004         LASSERT(msg->msg_kiov == NULL);
1005         LASSERT(msg->msg_niov == 0);
1006         LASSERT(msg->msg_routing);
1007         LASSERT(msg->msg_receiving);
1008         LASSERT(!msg->msg_sending);
1009         LASSERT(lpni->lpni_peer_net);
1010         LASSERT(lpni->lpni_peer_net->lpn_peer);
1011
1012         lp = lpni->lpni_peer_net->lpn_peer;
1013
1014         /* non-lnet_parse callers only receive delayed messages */
1015         LASSERT(!do_recv || msg->msg_rx_delayed);
1016
1017         if (!msg->msg_peerrtrcredit) {
1018                 /* lpni_lock protects the credit manipulation */
1019                 spin_lock(&lpni->lpni_lock);
1020                 /* lp_lock protects the lp_rtrq */
1021                 spin_lock(&lp->lp_lock);
1022
1023                 msg->msg_peerrtrcredit = 1;
1024                 lpni->lpni_rtrcredits--;
1025                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1026                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1027
1028                 if (lpni->lpni_rtrcredits < 0) {
1029                         /* must have checked eager_recv before here */
1030                         LASSERT(msg->msg_rx_ready_delay);
1031                         msg->msg_rx_delayed = 1;
1032                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1033                         spin_unlock(&lp->lp_lock);
1034                         spin_unlock(&lpni->lpni_lock);
1035                         return LNET_CREDIT_WAIT;
1036                 }
1037                 spin_unlock(&lp->lp_lock);
1038                 spin_unlock(&lpni->lpni_lock);
1039         }
1040
1041         rbp = lnet_msg2bufpool(msg);
1042
1043         if (!msg->msg_rtrcredit) {
1044                 msg->msg_rtrcredit = 1;
1045                 rbp->rbp_credits--;
1046                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1047                         rbp->rbp_mincredits = rbp->rbp_credits;
1048
1049                 if (rbp->rbp_credits < 0) {
1050                         /* must have checked eager_recv before here */
1051                         LASSERT(msg->msg_rx_ready_delay);
1052                         msg->msg_rx_delayed = 1;
1053                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1054                         return LNET_CREDIT_WAIT;
1055                 }
1056         }
1057
1058         LASSERT(!list_empty(&rbp->rbp_bufs));
1059         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1060         list_del(&rb->rb_list);
1061
1062         msg->msg_niov = rbp->rbp_npages;
1063         msg->msg_kiov = &rb->rb_kiov[0];
1064
1065         /* unset the msg-rx_delayed flag since we're receiving the message */
1066         msg->msg_rx_delayed = 0;
1067
1068         if (do_recv) {
1069                 int cpt = msg->msg_rx_cpt;
1070
1071                 lnet_net_unlock(cpt);
1072                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1073                              0, msg->msg_len, msg->msg_len);
1074                 lnet_net_lock(cpt);
1075         }
1076         return LNET_CREDIT_OK;
1077 }
1078
1079 void
1080 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1081 {
1082         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1083         struct lnet_ni          *txni = msg->msg_txni;
1084         struct lnet_msg         *msg2;
1085
1086         if (msg->msg_txcredit) {
1087                 struct lnet_ni       *ni = msg->msg_txni;
1088                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1089
1090                 /* give back NI txcredits */
1091                 msg->msg_txcredit = 0;
1092
1093                 LASSERT((tq->tq_credits < 0) ==
1094                         !list_empty(&tq->tq_delayed));
1095
1096                 tq->tq_credits++;
1097                 atomic_inc(&ni->ni_tx_credits);
1098                 if (tq->tq_credits <= 0) {
1099                         msg2 = list_entry(tq->tq_delayed.next,
1100                                           struct lnet_msg, msg_list);
1101                         list_del(&msg2->msg_list);
1102
1103                         LASSERT(msg2->msg_txni == ni);
1104                         LASSERT(msg2->msg_tx_delayed);
1105                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1106
1107                         (void) lnet_post_send_locked(msg2, 1);
1108                 }
1109         }
1110
1111         if (msg->msg_peertxcredit) {
1112                 /* give back peer txcredits */
1113                 msg->msg_peertxcredit = 0;
1114
1115                 spin_lock(&txpeer->lpni_lock);
1116                 LASSERT((txpeer->lpni_txcredits < 0) ==
1117                         !list_empty(&txpeer->lpni_txq));
1118
1119                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1120                 LASSERT(txpeer->lpni_txqnob >= 0);
1121
1122                 txpeer->lpni_txcredits++;
1123                 if (txpeer->lpni_txcredits <= 0) {
1124                         int msg2_cpt;
1125
1126                         msg2 = list_entry(txpeer->lpni_txq.next,
1127                                               struct lnet_msg, msg_list);
1128                         list_del(&msg2->msg_list);
1129                         spin_unlock(&txpeer->lpni_lock);
1130
1131                         LASSERT(msg2->msg_txpeer == txpeer);
1132                         LASSERT(msg2->msg_tx_delayed);
1133
1134                         msg2_cpt = msg2->msg_tx_cpt;
1135
1136                         /*
1137                          * The msg_cpt can be different from the msg2_cpt
1138                          * so we need to make sure we lock the correct cpt
1139                          * for msg2.
1140                          * Once we call lnet_post_send_locked() it is no
1141                          * longer safe to access msg2, since it could've
1142                          * been freed by lnet_finalize(), but we still
1143                          * need to relock the correct cpt, so we cache the
1144                          * msg2_cpt for the purpose of the check that
1145                          * follows the call to lnet_pose_send_locked().
1146                          */
1147                         if (msg2_cpt != msg->msg_tx_cpt) {
1148                                 lnet_net_unlock(msg->msg_tx_cpt);
1149                                 lnet_net_lock(msg2_cpt);
1150                         }
1151                         (void) lnet_post_send_locked(msg2, 1);
1152                         if (msg2_cpt != msg->msg_tx_cpt) {
1153                                 lnet_net_unlock(msg2_cpt);
1154                                 lnet_net_lock(msg->msg_tx_cpt);
1155                         }
1156                 } else {
1157                         spin_unlock(&txpeer->lpni_lock);
1158                 }
1159         }
1160
1161         if (txni != NULL) {
1162                 msg->msg_txni = NULL;
1163                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1164         }
1165
1166         if (txpeer != NULL) {
1167                 msg->msg_txpeer = NULL;
1168                 lnet_peer_ni_decref_locked(txpeer);
1169         }
1170 }
1171
1172 void
1173 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1174 {
1175         struct lnet_msg *msg;
1176
1177         if (list_empty(&rbp->rbp_msgs))
1178                 return;
1179         msg = list_entry(rbp->rbp_msgs.next,
1180                          struct lnet_msg, msg_list);
1181         list_del(&msg->msg_list);
1182
1183         (void)lnet_post_routed_recv_locked(msg, 1);
1184 }
1185
1186 void
1187 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1188 {
1189         struct lnet_msg *msg;
1190         struct lnet_msg *tmp;
1191
1192         lnet_net_unlock(cpt);
1193
1194         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1195                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1196                              0, 0, 0, msg->msg_hdr.payload_length);
1197                 list_del_init(&msg->msg_list);
1198                 msg->msg_no_resend = true;
1199                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1200                 lnet_finalize(msg, -ECANCELED);
1201         }
1202
1203         lnet_net_lock(cpt);
1204 }
1205
1206 void
1207 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1208 {
1209         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1210         struct lnet_peer *lp;
1211         struct lnet_ni *rxni = msg->msg_rxni;
1212         struct lnet_msg *msg2;
1213
1214         if (msg->msg_rtrcredit) {
1215                 /* give back global router credits */
1216                 struct lnet_rtrbuf *rb;
1217                 struct lnet_rtrbufpool *rbp;
1218
1219                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1220                  * there until it gets one allocated, or aborts the wait
1221                  * itself */
1222                 LASSERT(msg->msg_kiov != NULL);
1223
1224                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1225                 rbp = rb->rb_pool;
1226
1227                 msg->msg_kiov = NULL;
1228                 msg->msg_rtrcredit = 0;
1229
1230                 LASSERT(rbp == lnet_msg2bufpool(msg));
1231
1232                 LASSERT((rbp->rbp_credits > 0) ==
1233                         !list_empty(&rbp->rbp_bufs));
1234
1235                 /* If routing is now turned off, we just drop this buffer and
1236                  * don't bother trying to return credits.  */
1237                 if (!the_lnet.ln_routing) {
1238                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1239                         goto routing_off;
1240                 }
1241
1242                 /* It is possible that a user has lowered the desired number of
1243                  * buffers in this pool.  Make sure we never put back
1244                  * more buffers than the stated number. */
1245                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1246                         /* Discard this buffer so we don't have too
1247                          * many. */
1248                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1249                         rbp->rbp_nbuffers--;
1250                 } else {
1251                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1252                         rbp->rbp_credits++;
1253                         if (rbp->rbp_credits <= 0)
1254                                 lnet_schedule_blocked_locked(rbp);
1255                 }
1256         }
1257
1258 routing_off:
1259         if (msg->msg_peerrtrcredit) {
1260                 LASSERT(rxpeerni);
1261                 LASSERT(rxpeerni->lpni_peer_net);
1262                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1263
1264                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1265
1266                 /* give back peer router credits */
1267                 msg->msg_peerrtrcredit = 0;
1268
1269                 spin_lock(&rxpeerni->lpni_lock);
1270                 spin_lock(&lp->lp_lock);
1271
1272                 rxpeerni->lpni_rtrcredits++;
1273
1274                 /* drop all messages which are queued to be routed on that
1275                  * peer. */
1276                 if (!the_lnet.ln_routing) {
1277                         struct list_head drop;
1278                         INIT_LIST_HEAD(&drop);
1279                         list_splice_init(&lp->lp_rtrq, &drop);
1280                         spin_unlock(&lp->lp_lock);
1281                         spin_unlock(&rxpeerni->lpni_lock);
1282                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1283                 } else if (!list_empty(&lp->lp_rtrq)) {
1284                         int msg2_cpt;
1285
1286                         msg2 = list_entry(lp->lp_rtrq.next,
1287                                           struct lnet_msg, msg_list);
1288                         list_del(&msg2->msg_list);
1289                         msg2_cpt = msg2->msg_rx_cpt;
1290                         spin_unlock(&lp->lp_lock);
1291                         spin_unlock(&rxpeerni->lpni_lock);
1292                         /*
1293                          * messages on the lp_rtrq can be from any NID in
1294                          * the peer, which means they might have different
1295                          * cpts. We need to make sure we lock the right
1296                          * one.
1297                          */
1298                         if (msg2_cpt != msg->msg_rx_cpt) {
1299                                 lnet_net_unlock(msg->msg_rx_cpt);
1300                                 lnet_net_lock(msg2_cpt);
1301                         }
1302                         (void) lnet_post_routed_recv_locked(msg2, 1);
1303                         if (msg2_cpt != msg->msg_rx_cpt) {
1304                                 lnet_net_unlock(msg2_cpt);
1305                                 lnet_net_lock(msg->msg_rx_cpt);
1306                         }
1307                 } else {
1308                         spin_unlock(&lp->lp_lock);
1309                         spin_unlock(&rxpeerni->lpni_lock);
1310                 }
1311         }
1312         if (rxni != NULL) {
1313                 msg->msg_rxni = NULL;
1314                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1315         }
1316         if (rxpeerni != NULL) {
1317                 msg->msg_rxpeer = NULL;
1318                 lnet_peer_ni_decref_locked(rxpeerni);
1319         }
1320 }
1321
1322 static int
1323 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1324 {
1325         if (p1->lpni_txqnob < p2->lpni_txqnob)
1326                 return 1;
1327
1328         if (p1->lpni_txqnob > p2->lpni_txqnob)
1329                 return -1;
1330
1331         if (p1->lpni_txcredits > p2->lpni_txcredits)
1332                 return 1;
1333
1334         if (p1->lpni_txcredits < p2->lpni_txcredits)
1335                 return -1;
1336
1337         return 0;
1338 }
1339
1340 static struct lnet_peer_ni *
1341 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1342                     struct lnet_peer *peer,
1343                     struct lnet_peer_net *peer_net)
1344 {
1345         /*
1346          * Look at the peer NIs for the destination peer that connect
1347          * to the chosen net. If a peer_ni is preferred when using the
1348          * best_ni to communicate, we use that one. If there is no
1349          * preferred peer_ni, or there are multiple preferred peer_ni,
1350          * the available transmit credits are used. If the transmit
1351          * credits are equal, we round-robin over the peer_ni.
1352          */
1353         struct lnet_peer_ni *lpni = NULL;
1354         struct lnet_peer_ni *best_lpni = NULL;
1355         int best_lpni_credits = INT_MIN;
1356         bool preferred = false;
1357         bool ni_is_pref;
1358         int best_lpni_healthv = 0;
1359         int lpni_healthv;
1360
1361         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1362                 /*
1363                  * if the best_ni we've chosen aleady has this lpni
1364                  * preferred, then let's use it
1365                  */
1366                 if (best_ni) {
1367                         ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1368                                                                 best_ni->ni_nid);
1369                         CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1370                                libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1371                 } else {
1372                         ni_is_pref = false;
1373                 }
1374
1375                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1376
1377                 if (best_lpni)
1378                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1379                                 libcfs_nid2str(lpni->lpni_nid),
1380                                 lpni->lpni_txcredits, best_lpni_credits,
1381                                 lpni->lpni_seq, best_lpni->lpni_seq);
1382
1383                 /* pick the healthiest peer ni */
1384                 if (lpni_healthv < best_lpni_healthv) {
1385                         continue;
1386                 } else if (lpni_healthv > best_lpni_healthv) {
1387                         best_lpni_healthv = lpni_healthv;
1388                 /* if this is a preferred peer use it */
1389                 } else if (!preferred && ni_is_pref) {
1390                         preferred = true;
1391                 } else if (preferred && !ni_is_pref) {
1392                         /*
1393                          * this is not the preferred peer so let's ignore
1394                          * it.
1395                          */
1396                         continue;
1397                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1398                         /*
1399                          * We already have a peer that has more credits
1400                          * available than this one. No need to consider
1401                          * this peer further.
1402                          */
1403                         continue;
1404                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1405                         /*
1406                          * The best peer found so far and the current peer
1407                          * have the same number of available credits let's
1408                          * make sure to select between them using Round
1409                          * Robin
1410                          */
1411                         if (best_lpni) {
1412                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1413                                         continue;
1414                         }
1415                 }
1416
1417                 best_lpni = lpni;
1418                 best_lpni_credits = lpni->lpni_txcredits;
1419         }
1420
1421         /* if we still can't find a peer ni then we can't reach it */
1422         if (!best_lpni) {
1423                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1424                         LNET_NIDNET(dst_nid);
1425                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1426                                 libcfs_net2str(net_id));
1427                 return NULL;
1428         }
1429
1430         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1431                libcfs_nid2str(best_lpni->lpni_nid));
1432
1433         return best_lpni;
1434 }
1435
1436 /*
1437  * Prerequisite: the best_ni should already be set in the sd
1438  */
1439 static inline struct lnet_peer_ni *
1440 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1441                            __u32 net_id)
1442 {
1443         struct lnet_peer_net *peer_net;
1444
1445         /*
1446          * The gateway is Multi-Rail capable so now we must select the
1447          * proper peer_ni
1448          */
1449         peer_net = lnet_peer_get_net_locked(peer, net_id);
1450
1451         if (!peer_net) {
1452                 CERROR("gateway peer %s has no NI on net %s\n",
1453                        libcfs_nid2str(peer->lp_primary_nid),
1454                        libcfs_net2str(net_id));
1455                 return NULL;
1456         }
1457
1458         return lnet_select_peer_ni(sd->sd_best_ni, sd->sd_dst_nid,
1459                                    peer, peer_net);
1460 }
1461
1462 static int
1463 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2,
1464                     struct lnet_peer_ni **best_lpni)
1465 {
1466         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1467         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1468         struct lnet_peer *lp1 = r1->lr_gateway;
1469         struct lnet_peer *lp2 = r2->lr_gateway;
1470         struct lnet_peer_ni *lpni1;
1471         struct lnet_peer_ni *lpni2;
1472         struct lnet_send_data sd;
1473         int rc;
1474
1475         sd.sd_best_ni = NULL;
1476         sd.sd_dst_nid = LNET_NID_ANY;
1477         lpni1 = lnet_find_best_lpni_on_net(&sd, lp1, r1->lr_lnet);
1478         lpni2 = lnet_find_best_lpni_on_net(&sd, lp2, r2->lr_lnet);
1479         LASSERT(lpni1 && lpni2);
1480
1481         if (r1->lr_priority < r2->lr_priority) {
1482                 *best_lpni = lpni1;
1483                 return 1;
1484         }
1485
1486         if (r1->lr_priority > r2->lr_priority) {
1487                 *best_lpni = lpni2;
1488                 return -1;
1489         }
1490
1491         if (r1_hops < r2_hops) {
1492                 *best_lpni = lpni1;
1493                 return 1;
1494         }
1495
1496         if (r1_hops > r2_hops) {
1497                 *best_lpni = lpni2;
1498                 return -1;
1499         }
1500
1501         rc = lnet_compare_peers(lpni1, lpni2);
1502         if (rc == 1) {
1503                 *best_lpni = lpni1;
1504                 return rc;
1505         } else if (rc == -1) {
1506                 *best_lpni = lpni2;
1507                 return rc;
1508         }
1509
1510         if (r1->lr_seq - r2->lr_seq <= 0) {
1511                 *best_lpni = lpni1;
1512                 return 1;
1513         }
1514
1515         *best_lpni = lpni2;
1516         return -1;
1517 }
1518
1519 static struct lnet_route *
1520 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1521                        lnet_nid_t rtr_nid, struct lnet_route **prev_route,
1522                        struct lnet_peer_ni **gwni)
1523 {
1524         struct lnet_peer_ni *best_gw_ni = NULL;
1525         struct lnet_route *best_route;
1526         struct lnet_route *last_route;
1527         struct lnet_remotenet *rnet;
1528         struct lnet_peer *lp_best;
1529         struct lnet_route *route;
1530         struct lnet_peer *lp;
1531         int rc;
1532
1533         /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1534          * rtr_nid nid, otherwise find the best gateway I can use */
1535
1536         rnet = lnet_find_rnet_locked(remote_net);
1537         if (rnet == NULL)
1538                 return NULL;
1539
1540         lp_best = NULL;
1541         best_route = last_route = NULL;
1542         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1543                 lp = route->lr_gateway;
1544
1545                 if (!lnet_is_route_alive(route))
1546                         continue;
1547
1548                 if (lp_best == NULL) {
1549                         best_route = last_route = route;
1550                         lp_best = lp;
1551                 }
1552
1553                 /* no protection on below fields, but it's harmless */
1554                 if (last_route->lr_seq - route->lr_seq < 0)
1555                         last_route = route;
1556
1557                 rc = lnet_compare_routes(route, best_route, &best_gw_ni);
1558                 if (rc < 0)
1559                         continue;
1560
1561                 best_route = route;
1562                 lp_best = lp;
1563         }
1564
1565         *prev_route = last_route;
1566         *gwni = best_gw_ni;
1567
1568         return best_route;
1569 }
1570
1571 static struct lnet_ni *
1572 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1573                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1574                  int md_cpt)
1575 {
1576         struct lnet_ni *ni = NULL;
1577         unsigned int shortest_distance;
1578         int best_credits;
1579         int best_healthv;
1580
1581         /*
1582          * If there is no peer_ni that we can send to on this network,
1583          * then there is no point in looking for a new best_ni here.
1584         */
1585         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1586                 return best_ni;
1587
1588         if (best_ni == NULL) {
1589                 shortest_distance = UINT_MAX;
1590                 best_credits = INT_MIN;
1591                 best_healthv = 0;
1592         } else {
1593                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1594                                                      best_ni->ni_dev_cpt);
1595                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1596                 best_healthv = atomic_read(&best_ni->ni_healthv);
1597         }
1598
1599         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1600                 unsigned int distance;
1601                 int ni_credits;
1602                 int ni_healthv;
1603                 int ni_fatal;
1604
1605                 ni_credits = atomic_read(&ni->ni_tx_credits);
1606                 ni_healthv = atomic_read(&ni->ni_healthv);
1607                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1608
1609                 /*
1610                  * calculate the distance from the CPT on which
1611                  * the message memory is allocated to the CPT of
1612                  * the NI's physical device
1613                  */
1614                 distance = cfs_cpt_distance(lnet_cpt_table(),
1615                                             md_cpt,
1616                                             ni->ni_dev_cpt);
1617
1618                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1619                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1620                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1621                         : "not seleced", best_credits, shortest_distance,
1622                         (best_ni) ? best_ni->ni_seq : 0);
1623
1624                 /*
1625                  * All distances smaller than the NUMA range
1626                  * are treated equally.
1627                  */
1628                 if (distance < lnet_numa_range)
1629                         distance = lnet_numa_range;
1630
1631                 /*
1632                  * Select on health, shorter distance, available
1633                  * credits, then round-robin.
1634                  */
1635                 if (ni_fatal) {
1636                         continue;
1637                 } else if (ni_healthv < best_healthv) {
1638                         continue;
1639                 } else if (ni_healthv > best_healthv) {
1640                         best_healthv = ni_healthv;
1641                         /*
1642                          * If we're going to prefer this ni because it's
1643                          * the healthiest, then we should set the
1644                          * shortest_distance in the algorithm in case
1645                          * there are multiple NIs with the same health but
1646                          * different distances.
1647                          */
1648                         if (distance < shortest_distance)
1649                                 shortest_distance = distance;
1650                 } else if (distance > shortest_distance) {
1651                         continue;
1652                 } else if (distance < shortest_distance) {
1653                         shortest_distance = distance;
1654                 } else if (ni_credits < best_credits) {
1655                         continue;
1656                 } else if (ni_credits == best_credits) {
1657                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1658                                 continue;
1659                 }
1660                 best_ni = ni;
1661                 best_credits = ni_credits;
1662         }
1663
1664         CDEBUG(D_NET, "selected best_ni %s\n",
1665                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1666
1667         return best_ni;
1668 }
1669
1670 /*
1671  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1672  * because such traffic is required to perform discovery. We therefore
1673  * exclude all GET and PUT on that portal. We also exclude all ACK and
1674  * REPLY traffic, but that is because the portal is not tracked in the
1675  * message structure for these message types. We could restrict this
1676  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1677  */
1678 static bool
1679 lnet_msg_discovery(struct lnet_msg *msg)
1680 {
1681         if (msg->msg_type == LNET_MSG_PUT) {
1682                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1683                         return true;
1684         } else if (msg->msg_type == LNET_MSG_GET) {
1685                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1686                         return true;
1687         }
1688         return false;
1689 }
1690
1691 #define SRC_SPEC        0x0001
1692 #define SRC_ANY         0x0002
1693 #define LOCAL_DST       0x0004
1694 #define REMOTE_DST      0x0008
1695 #define MR_DST          0x0010
1696 #define NMR_DST         0x0020
1697 #define SND_RESP        0x0040
1698
1699 /* The following to defines are used for return codes */
1700 #define REPEAT_SEND     0x1000
1701 #define PASS_THROUGH    0x2000
1702
1703 /* The different cases lnet_select pathway needs to handle */
1704 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1705 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1706 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1707 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1708 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1709 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1710 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1711 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1712
1713 static int
1714 lnet_handle_lo_send(struct lnet_send_data *sd)
1715 {
1716         struct lnet_msg *msg = sd->sd_msg;
1717         int cpt = sd->sd_cpt;
1718
1719         /* No send credit hassles with LOLND */
1720         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1721         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1722         if (!msg->msg_routing)
1723                 msg->msg_hdr.src_nid =
1724                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1725         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1726         lnet_msg_commit(msg, cpt);
1727         msg->msg_txni = the_lnet.ln_loni;
1728
1729         return LNET_CREDIT_OK;
1730 }
1731
1732 static int
1733 lnet_handle_send(struct lnet_send_data *sd)
1734 {
1735         struct lnet_ni *best_ni = sd->sd_best_ni;
1736         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1737         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1738         struct lnet_msg *msg = sd->sd_msg;
1739         int cpt2;
1740         __u32 send_case = sd->sd_send_case;
1741         int rc;
1742         __u32 routing = send_case & REMOTE_DST;
1743          struct lnet_rsp_tracker *rspt;
1744
1745         /*
1746          * Increment sequence number of the selected peer so that we
1747          * pick the next one in Round Robin.
1748          */
1749         best_lpni->lpni_seq++;
1750
1751         /*
1752          * grab a reference on the peer_ni so it sticks around even if
1753          * we need to drop and relock the lnet_net_lock below.
1754          */
1755         lnet_peer_ni_addref_locked(best_lpni);
1756
1757         /*
1758          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1759          * message. This ensures that we get a CPT that is correct for
1760          * the NI when the NI has been restricted to a subset of all CPTs.
1761          * If the selected CPT differs from the one currently locked, we
1762          * must unlock and relock the lnet_net_lock(), and then check whether
1763          * the configuration has changed. We don't have a hold on the best_ni
1764          * yet, and it may have vanished.
1765          */
1766         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1767         if (sd->sd_cpt != cpt2) {
1768                 __u32 seq = lnet_get_dlc_seq_locked();
1769                 lnet_net_unlock(sd->sd_cpt);
1770                 sd->sd_cpt = cpt2;
1771                 lnet_net_lock(sd->sd_cpt);
1772                 if (seq != lnet_get_dlc_seq_locked()) {
1773                         lnet_peer_ni_decref_locked(best_lpni);
1774                         return REPEAT_SEND;
1775                 }
1776         }
1777
1778         /*
1779          * store the best_lpni in the message right away to avoid having
1780          * to do the same operation under different conditions
1781          */
1782         msg->msg_txpeer = best_lpni;
1783         msg->msg_txni = best_ni;
1784
1785         /*
1786          * grab a reference for the best_ni since now it's in use in this
1787          * send. The reference will be dropped in lnet_finalize()
1788          */
1789         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1790
1791         /*
1792          * Always set the target.nid to the best peer picked. Either the
1793          * NID will be one of the peer NIDs selected, or the same NID as
1794          * what was originally set in the target or it will be the NID of
1795          * a router if this message should be routed
1796          */
1797         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1798
1799         /*
1800          * lnet_msg_commit assigns the correct cpt to the message, which
1801          * is used to decrement the correct refcount on the ni when it's
1802          * time to return the credits
1803          */
1804         lnet_msg_commit(msg, sd->sd_cpt);
1805
1806         /*
1807          * If we are routing the message then we keep the src_nid that was
1808          * set by the originator. If we are not routing then we are the
1809          * originator and set it here.
1810          */
1811         if (!msg->msg_routing)
1812                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1813
1814         if (routing) {
1815                 msg->msg_target_is_router = 1;
1816                 msg->msg_target.pid = LNET_PID_LUSTRE;
1817                 /*
1818                  * since we're routing we want to ensure that the
1819                  * msg_hdr.dest_nid is set to the final destination. When
1820                  * the router receives this message it knows how to route
1821                  * it.
1822                  *
1823                  * final_dst_lpni is set at the beginning of the
1824                  * lnet_select_pathway() function and is never changed.
1825                  * It's safe to use it here.
1826                  */
1827                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1828         } else {
1829                 /*
1830                  * if we're not routing set the dest_nid to the best peer
1831                  * ni NID that we picked earlier in the algorithm.
1832                  */
1833                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1834         }
1835
1836         /*
1837          * if we have response tracker block update it with the next hop
1838          * nid
1839          */
1840         if (msg->msg_md) {
1841                 rspt = msg->msg_md->md_rspt_ptr;
1842                 if (rspt) {
1843                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1844                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1845                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1846                 }
1847         }
1848
1849         rc = lnet_post_send_locked(msg, 0);
1850
1851         if (!rc)
1852                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1853                        libcfs_nid2str(msg->msg_hdr.src_nid),
1854                        libcfs_nid2str(msg->msg_txni->ni_nid),
1855                        libcfs_nid2str(sd->sd_src_nid),
1856                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1857                        libcfs_nid2str(sd->sd_dst_nid),
1858                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1859                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1860
1861         return rc;
1862 }
1863
1864 static inline void
1865 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1866 {
1867         if (sd->sd_send_case & NMR_DST &&
1868             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1869             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1870             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1871                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1872                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1873                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1874                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1875                                                  sd->sd_best_ni->ni_nid);
1876         }
1877 }
1878
1879 /*
1880  * Source Specified
1881  * Local Destination
1882  * non-mr peer
1883  *
1884  * use the source and destination NIDs as the pathway
1885  */
1886 static int
1887 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1888 {
1889         /* the destination lpni is set before we get here. */
1890
1891         /* find local NI */
1892         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1893         if (!sd->sd_best_ni) {
1894                 CERROR("Can't send to %s: src %s is not a "
1895                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1896                                 libcfs_nid2str(sd->sd_src_nid));
1897                 return -EINVAL;
1898         }
1899
1900         /*
1901          * the preferred NID will only be set for NMR peers
1902          */
1903         lnet_set_non_mr_pref_nid(sd);
1904
1905         return lnet_handle_send(sd);
1906 }
1907
1908 /*
1909  * Source Specified
1910  * Local Destination
1911  * MR Peer
1912  *
1913  * Don't run the selection algorithm on the peer NIs. By specifying the
1914  * local NID, we're also saying that we should always use the destination NID
1915  * provided. This handles the case where we should be using the same
1916  * destination NID for the all the messages which belong to the same RPC
1917  * request.
1918  */
1919 static int
1920 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1921 {
1922         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1923         if (!sd->sd_best_ni) {
1924                 CERROR("Can't send to %s: src %s is not a "
1925                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1926                                 libcfs_nid2str(sd->sd_src_nid));
1927                 return -EINVAL;
1928         }
1929
1930         if (sd->sd_best_lpni &&
1931             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1932                 return lnet_handle_lo_send(sd);
1933         else if (sd->sd_best_lpni)
1934                 return lnet_handle_send(sd);
1935
1936         CERROR("can't send to %s. no NI on %s\n",
1937                libcfs_nid2str(sd->sd_dst_nid),
1938                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1939
1940         return -EHOSTUNREACH;
1941 }
1942
1943 struct lnet_ni *
1944 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1945                               struct lnet_peer *peer,
1946                               struct lnet_peer_net *peer_net,
1947                               int cpt,
1948                               bool incr_seq)
1949 {
1950         struct lnet_net *local_net;
1951         struct lnet_ni *best_ni;
1952
1953         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1954         if (!local_net)
1955                 return NULL;
1956
1957         /*
1958          * Iterate through the NIs in this local Net and select
1959          * the NI to send from. The selection is determined by
1960          * these 3 criterion in the following priority:
1961          *      1. NUMA
1962          *      2. NI available credits
1963          *      3. Round Robin
1964          */
1965         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1966                                    peer, peer_net, cpt);
1967
1968         if (incr_seq && best_ni)
1969                 best_ni->ni_seq++;
1970
1971         return best_ni;
1972 }
1973
1974 static int
1975 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1976                              struct lnet_msg *msg, lnet_nid_t rtr_nid,
1977                              int cpt)
1978 {
1979         struct lnet_peer *peer;
1980         lnet_nid_t primary_nid;
1981         int rc;
1982
1983         lnet_peer_ni_addref_locked(lpni);
1984
1985         peer = lpni->lpni_peer_net->lpn_peer;
1986
1987         if (lnet_peer_gw_discovery(peer)) {
1988                 lnet_peer_ni_decref_locked(lpni);
1989                 return 0;
1990         }
1991
1992         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1993                 lnet_peer_ni_decref_locked(lpni);
1994                 return 0;
1995         }
1996
1997         rc = lnet_discover_peer_locked(lpni, cpt, false);
1998         if (rc) {
1999                 lnet_peer_ni_decref_locked(lpni);
2000                 return rc;
2001         }
2002         /* The peer may have changed. */
2003         peer = lpni->lpni_peer_net->lpn_peer;
2004         spin_lock(&peer->lp_lock);
2005         if (lnet_peer_is_uptodate_locked(peer)) {
2006                 spin_unlock(&peer->lp_lock);
2007                 lnet_peer_ni_decref_locked(lpni);
2008                 return 0;
2009         }
2010         /* queue message and return */
2011         msg->msg_rtr_nid_param = rtr_nid;
2012         msg->msg_sending = 0;
2013         msg->msg_txpeer = NULL;
2014         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2015         primary_nid = peer->lp_primary_nid;
2016         spin_unlock(&peer->lp_lock);
2017
2018         lnet_peer_ni_decref_locked(lpni);
2019
2020         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2021                 msg, libcfs_nid2str(primary_nid));
2022
2023         return LNET_DC_WAIT;
2024 }
2025
2026 static int
2027 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2028                              lnet_nid_t dst_nid,
2029                              struct lnet_peer_ni **gw_lpni,
2030                              struct lnet_peer **gw_peer)
2031 {
2032         int rc;
2033         struct lnet_peer *gw;
2034         struct lnet_peer *lp;
2035         struct lnet_peer_net *lpn;
2036         struct lnet_peer_net *best_lpn = NULL;
2037         struct lnet_remotenet *rnet;
2038         struct lnet_route *best_route;
2039         struct lnet_route *last_route;
2040         struct lnet_peer_ni *lpni = NULL;
2041         struct lnet_peer_ni *gwni = NULL;
2042         lnet_nid_t src_nid = sd->sd_src_nid;
2043
2044         /* we've already looked up the initial lpni using dst_nid */
2045         lpni = sd->sd_best_lpni;
2046         /* the peer tree must be in existence */
2047         LASSERT(lpni && lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
2048         lp = lpni->lpni_peer_net->lpn_peer;
2049
2050         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2051                 /* is this remote network reachable?  */
2052                 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2053                 if (!rnet)
2054                         continue;
2055
2056                 if (!best_lpn)
2057                         best_lpn = lpn;
2058
2059                 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2060                         continue;
2061
2062                 best_lpn = lpn;
2063         }
2064
2065         if (!best_lpn) {
2066                 CERROR("peer %s has no available nets \n",
2067                        libcfs_nid2str(sd->sd_dst_nid));
2068                 return -EHOSTUNREACH;
2069         }
2070
2071         sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd, lp, best_lpn->lpn_net_id);
2072         if (!sd->sd_best_lpni) {
2073                 CERROR("peer %s down\n", libcfs_nid2str(sd->sd_dst_nid));
2074                 return -EHOSTUNREACH;
2075         }
2076
2077         best_route = lnet_find_route_locked(NULL, best_lpn->lpn_net_id,
2078                                             sd->sd_rtr_nid, &last_route,
2079                                             &gwni);
2080         if (!best_route) {
2081                 CERROR("no route to %s from %s\n",
2082                        libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
2083                 return -EHOSTUNREACH;
2084         }
2085
2086         if (!gwni) {
2087                 CERROR("Internal Error. Route expected to %s from %s\n",
2088                         libcfs_nid2str(dst_nid),
2089                         libcfs_nid2str(src_nid));
2090                 return -EFAULT;
2091         }
2092
2093         gw = best_route->lr_gateway;
2094         LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2095
2096         /*
2097          * Discover this gateway if it hasn't already been discovered.
2098          * This means we might delay the message until discovery has
2099          * completed
2100          */
2101         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2102         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2103                                           sd->sd_cpt);
2104         if (rc)
2105                 return rc;
2106
2107         if (!sd->sd_best_ni)
2108                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2109                                         lnet_peer_get_net_locked(gw,
2110                                                 best_route->lr_lnet),
2111                                         sd->sd_md_cpt,
2112                                         true);
2113
2114         if (!sd->sd_best_ni) {
2115                 CERROR("Internal Error. Expected local ni on %s "
2116                        "but non found :%s\n",
2117                        libcfs_net2str(best_route->lr_lnet),
2118                        libcfs_nid2str(sd->sd_src_nid));
2119                 return -EFAULT;
2120         }
2121
2122         *gw_lpni = gwni;
2123         *gw_peer = gw;
2124
2125         /*
2126          * increment the sequence numbers since now we're sure we're
2127          * going to use this path
2128          */
2129         LASSERT(best_route && last_route);
2130         best_route->lr_seq = last_route->lr_seq + 1;
2131         best_lpn->lpn_seq++;
2132
2133         return 0;
2134 }
2135
2136 /*
2137  * Handle two cases:
2138  *
2139  * Case 1:
2140  *  Source specified
2141  *  Remote destination
2142  *  Non-MR destination
2143  *
2144  * Case 2:
2145  *  Source specified
2146  *  Remote destination
2147  *  MR destination
2148  *
2149  * The handling of these two cases is similar. Even though the destination
2150  * can be MR or non-MR, we'll deal directly with the router.
2151  */
2152 static int
2153 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2154 {
2155         int rc;
2156         struct lnet_peer_ni *gw_lpni = NULL;
2157         struct lnet_peer *gw_peer = NULL;
2158
2159         /* find local NI */
2160         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2161         if (!sd->sd_best_ni) {
2162                 CERROR("Can't send to %s: src %s is not a "
2163                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2164                                 libcfs_nid2str(sd->sd_src_nid));
2165                 return -EINVAL;
2166         }
2167
2168         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2169                                      &gw_peer);
2170         if (rc)
2171                 return rc;
2172
2173         if (sd->sd_send_case & NMR_DST)
2174                 /*
2175                 * since the final destination is non-MR let's set its preferred
2176                 * NID before we send
2177                 */
2178                 lnet_set_non_mr_pref_nid(sd);
2179
2180         /*
2181          * We're going to send to the gw found so let's set its
2182          * info
2183          */
2184         sd->sd_peer = gw_peer;
2185         sd->sd_best_lpni = gw_lpni;
2186
2187         return lnet_handle_send(sd);
2188 }
2189
2190 struct lnet_ni *
2191 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2192                                bool discovery)
2193 {
2194         struct lnet_peer_net *peer_net = NULL;
2195         struct lnet_ni *best_ni = NULL;
2196
2197         /*
2198          * The peer can have multiple interfaces, some of them can be on
2199          * the local network and others on a routed network. We should
2200          * prefer the local network. However if the local network is not
2201          * available then we need to try the routed network
2202          */
2203
2204         /* go through all the peer nets and find the best_ni */
2205         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2206                 /*
2207                  * The peer's list of nets can contain non-local nets. We
2208                  * want to only examine the local ones.
2209                  */
2210                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2211                         continue;
2212                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2213                                                    peer_net, md_cpt, false);
2214
2215                 /*
2216                  * if this is a discovery message and lp_disc_net_id is
2217                  * specified then use that net to send the discovery on.
2218                  */
2219                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2220                     discovery)
2221                         break;
2222         }
2223
2224         if (best_ni)
2225                 /* increment sequence number so we can round robin */
2226                 best_ni->ni_seq++;
2227
2228         return best_ni;
2229 }
2230
2231 static struct lnet_ni *
2232 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2233 {
2234         struct lnet_ni *best_ni = NULL;
2235         struct lnet_peer_net *peer_net;
2236         struct lnet_peer *peer = sd->sd_peer;
2237         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2238         struct lnet_peer_ni *lpni;
2239         int cpt = sd->sd_cpt;
2240
2241         /*
2242          * We must use a consistent source address when sending to a
2243          * non-MR peer. However, a non-MR peer can have multiple NIDs
2244          * on multiple networks, and we may even need to talk to this
2245          * peer on multiple networks -- certain types of
2246          * load-balancing configuration do this.
2247          *
2248          * So we need to pick the NI the peer prefers for this
2249          * particular network.
2250          */
2251
2252         /* Get the target peer_ni */
2253         peer_net = lnet_peer_get_net_locked(peer,
2254                         LNET_NIDNET(best_lpni->lpni_nid));
2255         LASSERT(peer_net != NULL);
2256         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2257                                 lpni_peer_nis) {
2258                 if (lpni->lpni_pref_nnids == 0)
2259                         continue;
2260                 LASSERT(lpni->lpni_pref_nnids == 1);
2261                 best_ni = lnet_nid2ni_locked(
2262                                 lpni->lpni_pref.nid, cpt);
2263                 break;
2264         }
2265
2266         return best_ni;
2267 }
2268
2269 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2270 static int
2271 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2272 {
2273         struct lnet_ni *best_ni = NULL;
2274         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2275
2276         /*
2277          * We must use a consistent source address when sending to a
2278          * non-MR peer. However, a non-MR peer can have multiple NIDs
2279          * on multiple networks, and we may even need to talk to this
2280          * peer on multiple networks -- certain types of
2281          * load-balancing configuration do this.
2282          *
2283          * So we need to pick the NI the peer prefers for this
2284          * particular network.
2285          */
2286
2287         best_ni = lnet_find_existing_preferred_best_ni(sd);
2288
2289         /* if best_ni is still not set just pick one */
2290         if (!best_ni) {
2291                 best_ni =
2292                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2293                                                 sd->sd_best_lpni->lpni_peer_net,
2294                                                 sd->sd_md_cpt, true);
2295                 /* If there is no best_ni we don't have a route */
2296                 if (!best_ni) {
2297                         CERROR("no path to %s from net %s\n",
2298                                 libcfs_nid2str(best_lpni->lpni_nid),
2299                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2300                         return -EHOSTUNREACH;
2301                 }
2302         }
2303
2304         sd->sd_best_ni = best_ni;
2305
2306         /* Set preferred NI if necessary. */
2307         lnet_set_non_mr_pref_nid(sd);
2308
2309         return 0;
2310 }
2311
2312
2313 /*
2314  * Source not specified
2315  * Local destination
2316  * Non-MR Peer
2317  *
2318  * always use the same source NID for NMR peers
2319  * If we've talked to that peer before then we already have a preferred
2320  * source NI associated with it. Otherwise, we select a preferred local NI
2321  * and store it in the peer
2322  */
2323 static int
2324 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2325 {
2326         int rc;
2327
2328         /* sd->sd_best_lpni is already set to the final destination */
2329
2330         /*
2331          * At this point we should've created the peer ni and peer. If we
2332          * can't find it, then something went wrong. Instead of assert
2333          * output a relevant message and fail the send
2334          */
2335         if (!sd->sd_best_lpni) {
2336                 CERROR("Internal fault. Unable to send msg %s to %s. "
2337                        "NID not known\n",
2338                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2339                        libcfs_nid2str(sd->sd_dst_nid));
2340                 return -EFAULT;
2341         }
2342
2343         rc = lnet_select_preferred_best_ni(sd);
2344         if (!rc)
2345                 rc = lnet_handle_send(sd);
2346
2347         return rc;
2348 }
2349
2350 static int
2351 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2352 {
2353         /*
2354          * NOTE we've already handled the remote peer case. So we only
2355          * need to worry about the local case here.
2356          *
2357          * if we're sending a response, ACK or reply, we need to send it
2358          * to the destination NID given to us. At this point we already
2359          * have the peer_ni we're suppose to send to, so just find the
2360          * best_ni on the peer net and use that. Since we're sending to an
2361          * MR peer then we can just run the selection algorithm on our
2362          * local NIs and pick the best one.
2363          */
2364         if (sd->sd_send_case & SND_RESP) {
2365                 sd->sd_best_ni =
2366                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2367                                                 sd->sd_best_lpni->lpni_peer_net,
2368                                                 sd->sd_md_cpt, true);
2369
2370                 if (!sd->sd_best_ni) {
2371                         /*
2372                          * We're not going to deal with not able to send
2373                          * a response to the provided final destination
2374                          */
2375                         CERROR("Can't send response to %s. "
2376                                "No local NI available\n",
2377                                 libcfs_nid2str(sd->sd_dst_nid));
2378                         return -EHOSTUNREACH;
2379                 }
2380
2381                 return lnet_handle_send(sd);
2382         }
2383
2384         /*
2385          * If we get here that means we're sending a fresh request, PUT or
2386          * GET, so we need to run our standard selection algorithm.
2387          * First find the best local interface that's on any of the peer's
2388          * networks.
2389          */
2390         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2391                                         sd->sd_md_cpt,
2392                                         lnet_msg_discovery(sd->sd_msg));
2393         if (sd->sd_best_ni) {
2394                 sd->sd_best_lpni =
2395                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2396                                              sd->sd_best_ni->ni_net->net_id);
2397
2398                 /*
2399                  * if we're successful in selecting a peer_ni on the local
2400                  * network, then send to it. Otherwise fall through and
2401                  * try and see if we can reach it over another routed
2402                  * network
2403                  */
2404                 if (sd->sd_best_lpni &&
2405                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2406                         /*
2407                          * in case we initially started with a routed
2408                          * destination, let's reset to local
2409                          */
2410                         sd->sd_send_case &= ~REMOTE_DST;
2411                         sd->sd_send_case |= LOCAL_DST;
2412                         return lnet_handle_lo_send(sd);
2413                 } else if (sd->sd_best_lpni) {
2414                         /*
2415                          * in case we initially started with a routed
2416                          * destination, let's reset to local
2417                          */
2418                         sd->sd_send_case &= ~REMOTE_DST;
2419                         sd->sd_send_case |= LOCAL_DST;
2420                         return lnet_handle_send(sd);
2421                 }
2422
2423                 CERROR("Internal Error. Expected to have a best_lpni: "
2424                        "%s -> %s\n",
2425                        libcfs_nid2str(sd->sd_src_nid),
2426                        libcfs_nid2str(sd->sd_dst_nid));
2427
2428                 return -EFAULT;
2429         }
2430
2431         /*
2432          * Peer doesn't have a local network. Let's see if there is
2433          * a remote network we can reach it on.
2434          */
2435         return PASS_THROUGH;
2436 }
2437
2438 /*
2439  * Case 1:
2440  *      Source NID not specified
2441  *      Local destination
2442  *      MR peer
2443  *
2444  * Case 2:
2445  *      Source NID not speified
2446  *      Remote destination
2447  *      MR peer
2448  *
2449  * In both of these cases if we're sending a response, ACK or REPLY, then
2450  * we need to send to the destination NID provided.
2451  *
2452  * In the remote case let's deal with MR routers.
2453  *
2454  */
2455
2456 static int
2457 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2458 {
2459         int rc = 0;
2460         struct lnet_peer *gw_peer = NULL;
2461         struct lnet_peer_ni *gw_lpni = NULL;
2462
2463         /*
2464          * handle sending a response to a remote peer here so we don't
2465          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2466          */
2467         if (sd->sd_send_case & REMOTE_DST &&
2468             sd->sd_send_case & SND_RESP) {
2469                 struct lnet_peer_ni *gw;
2470                 struct lnet_peer *gw_peer;
2471
2472                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2473                                                   &gw_peer);
2474                 if (rc < 0) {
2475                         CERROR("Can't send response to %s. "
2476                                "No route available\n",
2477                                 libcfs_nid2str(sd->sd_dst_nid));
2478                         return -EHOSTUNREACH;
2479                 } else if (rc > 0) {
2480                         return rc;
2481                 }
2482
2483                 sd->sd_best_lpni = gw;
2484                 sd->sd_peer = gw_peer;
2485
2486                 return lnet_handle_send(sd);
2487         }
2488
2489         /*
2490          * Even though the NID for the peer might not be on a local network,
2491          * since the peer is MR there could be other interfaces on the
2492          * local network. In that case we'd still like to prefer the local
2493          * network over the routed network. If we're unable to do that
2494          * then we select the best router among the different routed networks,
2495          * and if the router is MR then we can deal with it as such.
2496          */
2497         rc = lnet_handle_any_mr_dsta(sd);
2498         if (rc != PASS_THROUGH)
2499                 return rc;
2500
2501         /*
2502          * Now that we must route to the destination, we must consider the
2503          * MR case, where the destination has multiple interfaces, some of
2504          * which we can route to and others we do not. For this reason we
2505          * need to select the destination which we can route to and if
2506          * there are multiple, we need to round robin.
2507          */
2508         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2509                                           &gw_peer);
2510         if (rc)
2511                 return rc;
2512
2513         sd->sd_send_case &= ~LOCAL_DST;
2514         sd->sd_send_case |= REMOTE_DST;
2515
2516         sd->sd_peer = gw_peer;
2517         sd->sd_best_lpni = gw_lpni;
2518
2519         return lnet_handle_send(sd);
2520 }
2521
2522 /*
2523  * Source not specified
2524  * Remote destination
2525  * Non-MR peer
2526  *
2527  * Must send to the specified peer NID using the same source NID that
2528  * we've used before. If it's the first time to talk to that peer then
2529  * find the source NI and assign it as preferred to that peer
2530  */
2531 static int
2532 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2533 {
2534         int rc;
2535         struct lnet_peer_ni *gw_lpni = NULL;
2536         struct lnet_peer *gw_peer = NULL;
2537
2538         /*
2539          * Let's set if we have a preferred NI to talk to this NMR peer
2540          */
2541         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2542
2543         /*
2544          * find the router and that'll find the best NI if we didn't find
2545          * it already.
2546          */
2547         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2548                                           &gw_peer);
2549         if (rc)
2550                 return rc;
2551
2552         /*
2553          * set the best_ni we've chosen as the preferred one for
2554          * this peer
2555          */
2556         lnet_set_non_mr_pref_nid(sd);
2557
2558         /* we'll be sending to the gw */
2559         sd->sd_best_lpni = gw_lpni;
2560         sd->sd_peer = gw_peer;
2561
2562         return lnet_handle_send(sd);
2563 }
2564
2565 static int
2566 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2567 {
2568         /*
2569          * turn off the SND_RESP bit.
2570          * It will be checked in the case handling
2571          */
2572         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2573
2574         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2575                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2576                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2577                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2578                 libcfs_nid2str(sd->sd_dst_nid),
2579                 (send_case & LOCAL_DST) ? "local" : "routed");
2580
2581         switch (send_case) {
2582         /*
2583          * For all cases where the source is specified, we should always
2584          * use the destination NID, whether it's an MR destination or not,
2585          * since we're continuing a series of related messages for the
2586          * same RPC
2587          */
2588         case SRC_SPEC_LOCAL_NMR_DST:
2589                 return lnet_handle_spec_local_nmr_dst(sd);
2590         case SRC_SPEC_LOCAL_MR_DST:
2591                 return lnet_handle_spec_local_mr_dst(sd);
2592         case SRC_SPEC_ROUTER_NMR_DST:
2593         case SRC_SPEC_ROUTER_MR_DST:
2594                 return lnet_handle_spec_router_dst(sd);
2595         case SRC_ANY_LOCAL_NMR_DST:
2596                 return lnet_handle_any_local_nmr_dst(sd);
2597         case SRC_ANY_LOCAL_MR_DST:
2598         case SRC_ANY_ROUTER_MR_DST:
2599                 return lnet_handle_any_mr_dst(sd);
2600         case SRC_ANY_ROUTER_NMR_DST:
2601                 return lnet_handle_any_router_nmr_dst(sd);
2602         default:
2603                 CERROR("Unknown send case\n");
2604                 return -1;
2605         }
2606 }
2607
2608 static int
2609 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2610                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2611 {
2612         struct lnet_peer_ni     *lpni;
2613         struct lnet_peer        *peer;
2614         struct lnet_send_data   send_data;
2615         int                     cpt, rc;
2616         int                     md_cpt;
2617         __u32                   send_case = 0;
2618
2619         memset(&send_data, 0, sizeof(send_data));
2620
2621         /*
2622          * get an initial CPT to use for locking. The idea here is not to
2623          * serialize the calls to select_pathway, so that as many
2624          * operations can run concurrently as possible. To do that we use
2625          * the CPT where this call is being executed. Later on when we
2626          * determine the CPT to use in lnet_message_commit, we switch the
2627          * lock and check if there was any configuration change.  If none,
2628          * then we proceed, if there is, then we restart the operation.
2629          */
2630         cpt = lnet_net_lock_current();
2631
2632         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2633         if (md_cpt == CFS_CPT_ANY)
2634                 md_cpt = cpt;
2635
2636 again:
2637
2638         /*
2639          * If we're being asked to send to the loopback interface, there
2640          * is no need to go through any selection. We can just shortcut
2641          * the entire process and send over lolnd
2642          */
2643         send_data.sd_msg = msg;
2644         send_data.sd_cpt = cpt;
2645         if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2646                 rc = lnet_handle_lo_send(&send_data);
2647                 lnet_net_unlock(cpt);
2648                 return rc;
2649         }
2650
2651         /*
2652          * find an existing peer_ni, or create one and mark it as having been
2653          * created due to network traffic. This call will create the
2654          * peer->peer_net->peer_ni tree.
2655          */
2656         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2657         if (IS_ERR(lpni)) {
2658                 lnet_net_unlock(cpt);
2659                 return PTR_ERR(lpni);
2660         }
2661
2662         /*
2663          * Cache the original src_nid. If we need to resend the message
2664          * then we'll need to know whether the src_nid was originally
2665          * specified for this message. If it was originally specified,
2666          * then we need to keep using the same src_nid since it's
2667          * continuing the same sequence of messages.
2668          */
2669         msg->msg_src_nid_param = src_nid;
2670
2671         /*
2672          * If necessary, perform discovery on the peer that owns this peer_ni.
2673          * Note, this can result in the ownership of this peer_ni changing
2674          * to another peer object.
2675          */
2676         rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2677         if (rc) {
2678                 lnet_peer_ni_decref_locked(lpni);
2679                 lnet_net_unlock(cpt);
2680                 return rc;
2681         }
2682         lnet_peer_ni_decref_locked(lpni);
2683
2684         peer = lpni->lpni_peer_net->lpn_peer;
2685
2686         /*
2687          * Identify the different send cases
2688          */
2689         if (src_nid == LNET_NID_ANY)
2690                 send_case |= SRC_ANY;
2691         else
2692                 send_case |= SRC_SPEC;
2693
2694         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2695                 send_case |= LOCAL_DST;
2696         else
2697                 send_case |= REMOTE_DST;
2698
2699         /*
2700          * if this is a non-MR peer or if we're recovering a peer ni then
2701          * let's consider this an NMR case so we can hit the destination
2702          * NID.
2703          */
2704         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2705                 send_case |= NMR_DST;
2706         else
2707                 send_case |= MR_DST;
2708
2709         if (msg->msg_type == LNET_MSG_REPLY ||
2710             msg->msg_type == LNET_MSG_ACK)
2711                 send_case |= SND_RESP;
2712
2713         /* assign parameters to the send_data */
2714         send_data.sd_rtr_nid = rtr_nid;
2715         send_data.sd_src_nid = src_nid;
2716         send_data.sd_dst_nid = dst_nid;
2717         send_data.sd_best_lpni = lpni;
2718         /*
2719          * keep a pointer to the final destination in case we're going to
2720          * route, so we'll need to access it later
2721          */
2722         send_data.sd_final_dst_lpni = lpni;
2723         send_data.sd_peer = peer;
2724         send_data.sd_md_cpt = md_cpt;
2725         send_data.sd_send_case = send_case;
2726
2727         rc = lnet_handle_send_case_locked(&send_data);
2728
2729         /*
2730          * Update the local cpt since send_data.sd_cpt might've been
2731          * updated as a result of calling lnet_handle_send_case_locked().
2732          */
2733         cpt = send_data.sd_cpt;
2734
2735         if (rc == REPEAT_SEND)
2736                 goto again;
2737
2738         lnet_net_unlock(cpt);
2739
2740         return rc;
2741 }
2742
2743 int
2744 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2745 {
2746         lnet_nid_t              dst_nid = msg->msg_target.nid;
2747         int                     rc;
2748
2749         /*
2750          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2751          * but we might want to use pre-determined router for ACK/REPLY
2752          * in the future
2753          */
2754         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2755         LASSERT(msg->msg_txpeer == NULL);
2756         LASSERT(msg->msg_txni == NULL);
2757         LASSERT(!msg->msg_sending);
2758         LASSERT(!msg->msg_target_is_router);
2759         LASSERT(!msg->msg_receiving);
2760
2761         msg->msg_sending = 1;
2762
2763         LASSERT(!msg->msg_tx_committed);
2764
2765         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2766         if (rc < 0) {
2767                 if (rc == -EHOSTUNREACH)
2768                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2769                 else
2770                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2771                 return rc;
2772         }
2773
2774         if (rc == LNET_CREDIT_OK)
2775                 lnet_ni_send(msg->msg_txni, msg);
2776
2777         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2778         return 0;
2779 }
2780
2781 enum lnet_mt_event_type {
2782         MT_TYPE_LOCAL_NI = 0,
2783         MT_TYPE_PEER_NI
2784 };
2785
2786 struct lnet_mt_event_info {
2787         enum lnet_mt_event_type mt_type;
2788         lnet_nid_t mt_nid;
2789 };
2790
2791 /* called with res_lock held */
2792 void
2793 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2794 {
2795         struct lnet_rsp_tracker *rspt;
2796
2797         /*
2798          * msg has a refcount on the MD so the MD is not going away.
2799          * The rspt queue for the cpt is protected by
2800          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2801          */
2802         if (!md->md_rspt_ptr)
2803                 return;
2804
2805         rspt = md->md_rspt_ptr;
2806
2807         /* debug code */
2808         LASSERT(rspt->rspt_cpt == cpt);
2809
2810         md->md_rspt_ptr = NULL;
2811
2812         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2813                 /*
2814                  * The monitor thread has invalidated this handle because the
2815                  * response timed out, but it failed to lookup the MD. That
2816                  * means this response tracker is on the zombie list. We can
2817                  * safely remove it under the resource lock (held by caller) and
2818                  * free the response tracker block.
2819                  */
2820                 list_del(&rspt->rspt_on_list);
2821                 lnet_rspt_free(rspt, cpt);
2822         } else {
2823                 /*
2824                  * invalidate the handle to indicate that a response has been
2825                  * received, which will then lead the monitor thread to clean up
2826                  * the rspt block.
2827                  */
2828                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2829         }
2830 }
2831
2832 void
2833 lnet_clean_zombie_rstqs(void)
2834 {
2835         struct lnet_rsp_tracker *rspt, *tmp;
2836         int i;
2837
2838         cfs_cpt_for_each(i, lnet_cpt_table()) {
2839                 list_for_each_entry_safe(rspt, tmp,
2840                                          the_lnet.ln_mt_zombie_rstqs[i],
2841                                          rspt_on_list) {
2842                         list_del(&rspt->rspt_on_list);
2843                         lnet_rspt_free(rspt, i);
2844                 }
2845         }
2846
2847         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2848 }
2849
2850 static void
2851 lnet_finalize_expired_responses(void)
2852 {
2853         struct lnet_libmd *md;
2854         struct list_head local_queue;
2855         struct lnet_rsp_tracker *rspt, *tmp;
2856         ktime_t now;
2857         int i;
2858
2859         if (the_lnet.ln_mt_rstq == NULL)
2860                 return;
2861
2862         cfs_cpt_for_each(i, lnet_cpt_table()) {
2863                 INIT_LIST_HEAD(&local_queue);
2864
2865                 lnet_net_lock(i);
2866                 if (!the_lnet.ln_mt_rstq[i]) {
2867                         lnet_net_unlock(i);
2868                         continue;
2869                 }
2870                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2871                 lnet_net_unlock(i);
2872
2873                 now = ktime_get();
2874
2875                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2876                         /*
2877                          * The rspt mdh will be invalidated when a response
2878                          * is received or whenever we want to discard the
2879                          * block the monitor thread will walk the queue
2880                          * and clean up any rsts with an invalid mdh.
2881                          * The monitor thread will walk the queue until
2882                          * the first unexpired rspt block. This means that
2883                          * some rspt blocks which received their
2884                          * corresponding responses will linger in the
2885                          * queue until they are cleaned up eventually.
2886                          */
2887                         lnet_res_lock(i);
2888                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2889                                 lnet_res_unlock(i);
2890                                 list_del(&rspt->rspt_on_list);
2891                                 lnet_rspt_free(rspt, i);
2892                                 continue;
2893                         }
2894
2895                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2896                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2897                                 struct lnet_peer_ni *lpni;
2898                                 lnet_nid_t nid;
2899
2900                                 md = lnet_handle2md(&rspt->rspt_mdh);
2901                                 if (!md) {
2902                                         /* MD has been queued for unlink, but
2903                                          * rspt hasn't been detached (Note we've
2904                                          * checked above that the rspt_mdh is
2905                                          * valid). Since we cannot lookup the MD
2906                                          * we're unable to detach the rspt
2907                                          * ourselves. Thus, move the rspt to the
2908                                          * zombie list where we'll wait for
2909                                          * either:
2910                                          *   1. The remaining operations on the
2911                                          *   MD to complete. In this case the
2912                                          *   final operation will result in
2913                                          *   lnet_msg_detach_md()->
2914                                          *   lnet_detach_rsp_tracker() where
2915                                          *   we will clean up this response
2916                                          *   tracker.
2917                                          *   2. LNet to shutdown. In this case
2918                                          *   we'll wait until after all LND Nets
2919                                          *   have shutdown and then we can
2920                                          *   safely free any remaining response
2921                                          *   tracker blocks on the zombie list.
2922                                          * Note: We need to hold the resource
2923                                          * lock when adding to the zombie list
2924                                          * because we may have concurrent access
2925                                          * with lnet_detach_rsp_tracker().
2926                                          */
2927                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2928                                         list_move(&rspt->rspt_on_list,
2929                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2930                                         lnet_res_unlock(i);
2931                                         continue;
2932                                 }
2933                                 LASSERT(md->md_rspt_ptr == rspt);
2934                                 md->md_rspt_ptr = NULL;
2935                                 lnet_res_unlock(i);
2936
2937                                 LNetMDUnlink(rspt->rspt_mdh);
2938
2939                                 nid = rspt->rspt_next_hop_nid;
2940
2941                                 list_del(&rspt->rspt_on_list);
2942                                 lnet_rspt_free(rspt, i);
2943
2944                                 /* If we're shutting down we just want to clean
2945                                  * up the rspt blocks
2946                                  */
2947                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2948                                         continue;
2949
2950                                 lnet_net_lock(i);
2951                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2952                                 lnet_net_unlock(i);
2953
2954                                 CDEBUG(D_NET,
2955                                        "Response timeout: md = %p: nid = %s\n",
2956                                        md, libcfs_nid2str(nid));
2957
2958                                 /*
2959                                  * If there is a timeout on the response
2960                                  * from the next hop decrement its health
2961                                  * value so that we don't use it
2962                                  */
2963                                 lnet_net_lock(0);
2964                                 lpni = lnet_find_peer_ni_locked(nid);
2965                                 if (lpni) {
2966                                         lnet_handle_remote_failure_locked(lpni);
2967                                         lnet_peer_ni_decref_locked(lpni);
2968                                 }
2969                                 lnet_net_unlock(0);
2970                         } else {
2971                                 lnet_res_unlock(i);
2972                                 break;
2973                         }
2974                 }
2975
2976                 if (!list_empty(&local_queue)) {
2977                         lnet_net_lock(i);
2978                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2979                         lnet_net_unlock(i);
2980                 }
2981         }
2982 }
2983
2984 static void
2985 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2986 {
2987         struct lnet_msg *msg;
2988
2989         while (!list_empty(resendq)) {
2990                 struct lnet_peer_ni *lpni;
2991
2992                 msg = list_entry(resendq->next, struct lnet_msg,
2993                                  msg_list);
2994
2995                 list_del_init(&msg->msg_list);
2996
2997                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2998                 if (!lpni) {
2999                         lnet_net_unlock(cpt);
3000                         CERROR("Expected that a peer is already created for %s\n",
3001                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3002                         msg->msg_no_resend = true;
3003                         lnet_finalize(msg, -EFAULT);
3004                         lnet_net_lock(cpt);
3005                 } else {
3006                         struct lnet_peer *peer;
3007                         int rc;
3008                         lnet_nid_t src_nid = LNET_NID_ANY;
3009
3010                         /*
3011                          * if this message is not being routed and the
3012                          * peer is non-MR then we must use the same
3013                          * src_nid that was used in the original send.
3014                          * Otherwise if we're routing the message (IE
3015                          * we're a router) then we can use any of our
3016                          * local interfaces. It doesn't matter to the
3017                          * final destination.
3018                          */
3019                         peer = lpni->lpni_peer_net->lpn_peer;
3020                         if (!msg->msg_routing &&
3021                             !lnet_peer_is_multi_rail(peer))
3022                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
3023
3024                         /*
3025                          * If we originally specified a src NID, then we
3026                          * must attempt to reuse it in the resend as well.
3027                          */
3028                         if (msg->msg_src_nid_param != LNET_NID_ANY)
3029                                 src_nid = msg->msg_src_nid_param;
3030                         lnet_peer_ni_decref_locked(lpni);
3031
3032                         lnet_net_unlock(cpt);
3033                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3034                                libcfs_nid2str(src_nid),
3035                                libcfs_id2str(msg->msg_target),
3036                                lnet_msgtyp2str(msg->msg_type),
3037                                msg->msg_recovery,
3038                                msg->msg_retry_count);
3039                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
3040                         if (rc) {
3041                                 CERROR("Error sending %s to %s: %d\n",
3042                                        lnet_msgtyp2str(msg->msg_type),
3043                                        libcfs_id2str(msg->msg_target), rc);
3044                                 msg->msg_no_resend = true;
3045                                 lnet_finalize(msg, rc);
3046                         }
3047                         lnet_net_lock(cpt);
3048                         if (!rc)
3049                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3050                 }
3051         }
3052 }
3053
3054 static void
3055 lnet_resend_pending_msgs(void)
3056 {
3057         int i;
3058
3059         cfs_cpt_for_each(i, lnet_cpt_table()) {
3060                 lnet_net_lock(i);
3061                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3062                 lnet_net_unlock(i);
3063         }
3064 }
3065
3066 /* called with cpt and ni_lock held */
3067 static void
3068 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3069 {
3070         struct lnet_handle_md recovery_mdh;
3071
3072         LNetInvalidateMDHandle(&recovery_mdh);
3073
3074         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3075             force) {
3076                 recovery_mdh = ni->ni_ping_mdh;
3077                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3078         }
3079         lnet_ni_unlock(ni);
3080         lnet_net_unlock(cpt);
3081         if (!LNetMDHandleIsInvalid(recovery_mdh))
3082                 LNetMDUnlink(recovery_mdh);
3083         lnet_net_lock(cpt);
3084         lnet_ni_lock(ni);
3085 }
3086
3087 static void
3088 lnet_recover_local_nis(void)
3089 {
3090         struct lnet_mt_event_info *ev_info;
3091         struct list_head processed_list;
3092         struct list_head local_queue;
3093         struct lnet_handle_md mdh;
3094         struct lnet_ni *tmp;
3095         struct lnet_ni *ni;
3096         lnet_nid_t nid;
3097         int healthv;
3098         int rc;
3099
3100         INIT_LIST_HEAD(&local_queue);
3101         INIT_LIST_HEAD(&processed_list);
3102
3103         /*
3104          * splice the recovery queue on a local queue. We will iterate
3105          * through the local queue and update it as needed. Once we're
3106          * done with the traversal, we'll splice the local queue back on
3107          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3108          * will be traversed in the next iteration.
3109          */
3110         lnet_net_lock(0);
3111         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3112                          &local_queue);
3113         lnet_net_unlock(0);
3114
3115         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3116                 /*
3117                  * if an NI is being deleted or it is now healthy, there
3118                  * is no need to keep it around in the recovery queue.
3119                  * The monitor thread is the only thread responsible for
3120                  * removing the NI from the recovery queue.
3121                  * Multiple threads can be adding NIs to the recovery
3122                  * queue.
3123                  */
3124                 healthv = atomic_read(&ni->ni_healthv);
3125
3126                 lnet_net_lock(0);
3127                 lnet_ni_lock(ni);
3128                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3129                     healthv == LNET_MAX_HEALTH_VALUE) {
3130                         list_del_init(&ni->ni_recovery);
3131                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3132                         lnet_ni_unlock(ni);
3133                         lnet_ni_decref_locked(ni, 0);
3134                         lnet_net_unlock(0);
3135                         continue;
3136                 }
3137
3138                 /*
3139                  * if the local NI failed recovery we must unlink the md.
3140                  * But we want to keep the local_ni on the recovery queue
3141                  * so we can continue the attempts to recover it.
3142                  */
3143                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3144                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3145                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3146                 }
3147
3148                 lnet_ni_unlock(ni);
3149                 lnet_net_unlock(0);
3150
3151
3152                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3153                        libcfs_nid2str(ni->ni_nid));
3154
3155                 lnet_ni_lock(ni);
3156                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3157                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3158                         lnet_ni_unlock(ni);
3159
3160                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3161                         if (!ev_info) {
3162                                 CERROR("out of memory. Can't recover %s\n",
3163                                        libcfs_nid2str(ni->ni_nid));
3164                                 lnet_ni_lock(ni);
3165                                 ni->ni_recovery_state &=
3166                                   ~LNET_NI_RECOVERY_PENDING;
3167                                 lnet_ni_unlock(ni);
3168                                 continue;
3169                         }
3170
3171                         mdh = ni->ni_ping_mdh;
3172                         /*
3173                          * Invalidate the ni mdh in case it's deleted.
3174                          * We'll unlink the mdh in this case below.
3175                          */
3176                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3177                         nid = ni->ni_nid;
3178
3179                         /*
3180                          * remove the NI from the local queue and drop the
3181                          * reference count to it while we're recovering
3182                          * it. The reason for that, is that the NI could
3183                          * be deleted, and the way the code is structured
3184                          * is if we don't drop the NI, then the deletion
3185                          * code will enter a loop waiting for the
3186                          * reference count to be removed while holding the
3187                          * ln_mutex_lock(). When we look up the peer to
3188                          * send to in lnet_select_pathway() we will try to
3189                          * lock the ln_mutex_lock() as well, leading to
3190                          * a deadlock. By dropping the refcount and
3191                          * removing it from the list, we allow for the NI
3192                          * to be removed, then we use the cached NID to
3193                          * look it up again. If it's gone, then we just
3194                          * continue examining the rest of the queue.
3195                          */
3196                         lnet_net_lock(0);
3197                         list_del_init(&ni->ni_recovery);
3198                         lnet_ni_decref_locked(ni, 0);
3199                         lnet_net_unlock(0);
3200
3201                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3202                         ev_info->mt_nid = nid;
3203                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3204                                             ev_info, the_lnet.ln_mt_eqh, true);
3205                         /* lookup the nid again */
3206                         lnet_net_lock(0);
3207                         ni = lnet_nid2ni_locked(nid, 0);
3208                         if (!ni) {
3209                                 /*
3210                                  * the NI has been deleted when we dropped
3211                                  * the ref count
3212                                  */
3213                                 lnet_net_unlock(0);
3214                                 LNetMDUnlink(mdh);
3215                                 continue;
3216                         }
3217                         /*
3218                          * Same note as in lnet_recover_peer_nis(). When
3219                          * we're sending the ping, the NI is free to be
3220                          * deleted or manipulated. By this point it
3221                          * could've been added back on the recovery queue,
3222                          * and a refcount taken on it.
3223                          * So we can't just add it blindly again or we'll
3224                          * corrupt the queue. We must check under lock if
3225                          * it's not on any list and if not then add it
3226                          * to the processed list, which will eventually be
3227                          * spliced back on to the recovery queue.
3228                          */
3229                         ni->ni_ping_mdh = mdh;
3230                         if (list_empty(&ni->ni_recovery)) {
3231                                 list_add_tail(&ni->ni_recovery, &processed_list);
3232                                 lnet_ni_addref_locked(ni, 0);
3233                         }
3234                         lnet_net_unlock(0);
3235
3236                         lnet_ni_lock(ni);
3237                         if (rc)
3238                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3239                 }
3240                 lnet_ni_unlock(ni);
3241         }
3242
3243         /*
3244          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3245          * reexamined in the next iteration.
3246          */
3247         list_splice_init(&processed_list, &local_queue);
3248         lnet_net_lock(0);
3249         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3250         lnet_net_unlock(0);
3251 }
3252
3253 static int
3254 lnet_resendqs_create(void)
3255 {
3256         struct list_head **resendqs;
3257         resendqs = lnet_create_array_of_queues();
3258
3259         if (!resendqs)
3260                 return -ENOMEM;
3261
3262         lnet_net_lock(LNET_LOCK_EX);
3263         the_lnet.ln_mt_resendqs = resendqs;
3264         lnet_net_unlock(LNET_LOCK_EX);
3265
3266         return 0;
3267 }
3268
3269 static void
3270 lnet_clean_local_ni_recoveryq(void)
3271 {
3272         struct lnet_ni *ni;
3273
3274         /* This is only called when the monitor thread has stopped */
3275         lnet_net_lock(0);
3276
3277         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3278                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3279                                 struct lnet_ni, ni_recovery);
3280                 list_del_init(&ni->ni_recovery);
3281                 lnet_ni_lock(ni);
3282                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3283                 lnet_ni_unlock(ni);
3284                 lnet_ni_decref_locked(ni, 0);
3285         }
3286
3287         lnet_net_unlock(0);
3288 }
3289
3290 static void
3291 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3292                                      bool force)
3293 {
3294         struct lnet_handle_md recovery_mdh;
3295
3296         LNetInvalidateMDHandle(&recovery_mdh);
3297
3298         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3299                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3300                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3301         }
3302         spin_unlock(&lpni->lpni_lock);
3303         lnet_net_unlock(cpt);
3304         if (!LNetMDHandleIsInvalid(recovery_mdh))
3305                 LNetMDUnlink(recovery_mdh);
3306         lnet_net_lock(cpt);
3307         spin_lock(&lpni->lpni_lock);
3308 }
3309
3310 static void
3311 lnet_clean_peer_ni_recoveryq(void)
3312 {
3313         struct lnet_peer_ni *lpni, *tmp;
3314
3315         lnet_net_lock(LNET_LOCK_EX);
3316
3317         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3318                                  lpni_recovery) {
3319                 list_del_init(&lpni->lpni_recovery);
3320                 spin_lock(&lpni->lpni_lock);
3321                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3322                 spin_unlock(&lpni->lpni_lock);
3323                 lnet_peer_ni_decref_locked(lpni);
3324         }
3325
3326         lnet_net_unlock(LNET_LOCK_EX);
3327 }
3328
3329 static void
3330 lnet_clean_resendqs(void)
3331 {
3332         struct lnet_msg *msg, *tmp;
3333         struct list_head msgs;
3334         int i;
3335
3336         INIT_LIST_HEAD(&msgs);
3337
3338         cfs_cpt_for_each(i, lnet_cpt_table()) {
3339                 lnet_net_lock(i);
3340                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3341                 lnet_net_unlock(i);
3342                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3343                         list_del_init(&msg->msg_list);
3344                         msg->msg_no_resend = true;
3345                         lnet_finalize(msg, -ESHUTDOWN);
3346                 }
3347         }
3348
3349         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3350 }
3351
3352 static void
3353 lnet_recover_peer_nis(void)
3354 {
3355         struct lnet_mt_event_info *ev_info;
3356         struct list_head processed_list;
3357         struct list_head local_queue;
3358         struct lnet_handle_md mdh;
3359         struct lnet_peer_ni *lpni;
3360         struct lnet_peer_ni *tmp;
3361         lnet_nid_t nid;
3362         int healthv;
3363         int rc;
3364
3365         INIT_LIST_HEAD(&local_queue);
3366         INIT_LIST_HEAD(&processed_list);
3367
3368         /*
3369          * Always use cpt 0 for locking across all interactions with
3370          * ln_mt_peerNIRecovq
3371          */
3372         lnet_net_lock(0);
3373         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3374                          &local_queue);
3375         lnet_net_unlock(0);
3376
3377         list_for_each_entry_safe(lpni, tmp, &local_queue,
3378                                  lpni_recovery) {
3379                 /*
3380                  * The same protection strategy is used here as is in the
3381                  * local recovery case.
3382                  */
3383                 lnet_net_lock(0);
3384                 healthv = atomic_read(&lpni->lpni_healthv);
3385                 spin_lock(&lpni->lpni_lock);
3386                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3387                     healthv == LNET_MAX_HEALTH_VALUE) {
3388                         list_del_init(&lpni->lpni_recovery);
3389                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3390                         spin_unlock(&lpni->lpni_lock);
3391                         lnet_peer_ni_decref_locked(lpni);
3392                         lnet_net_unlock(0);
3393                         continue;
3394                 }
3395
3396                 /*
3397                  * If the peer NI has failed recovery we must unlink the
3398                  * md. But we want to keep the peer ni on the recovery
3399                  * queue so we can try to continue recovering it
3400                  */
3401                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3402                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3403                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3404                 }
3405
3406                 spin_unlock(&lpni->lpni_lock);
3407                 lnet_net_unlock(0);
3408
3409                 /*
3410                  * NOTE: we're racing with peer deletion from user space.
3411                  * It's possible that a peer is deleted after we check its
3412                  * state. In this case the recovery can create a new peer
3413                  */
3414                 spin_lock(&lpni->lpni_lock);
3415                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3416                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3417                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3418                         spin_unlock(&lpni->lpni_lock);
3419
3420                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3421                         if (!ev_info) {
3422                                 CERROR("out of memory. Can't recover %s\n",
3423                                        libcfs_nid2str(lpni->lpni_nid));
3424                                 spin_lock(&lpni->lpni_lock);
3425                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3426                                 spin_unlock(&lpni->lpni_lock);
3427                                 continue;
3428                         }
3429
3430                         /* look at the comments in lnet_recover_local_nis() */
3431                         mdh = lpni->lpni_recovery_ping_mdh;
3432                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3433                         nid = lpni->lpni_nid;
3434                         lnet_net_lock(0);
3435                         list_del_init(&lpni->lpni_recovery);
3436                         lnet_peer_ni_decref_locked(lpni);
3437                         lnet_net_unlock(0);
3438
3439                         ev_info->mt_type = MT_TYPE_PEER_NI;
3440                         ev_info->mt_nid = nid;
3441                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3442                                             ev_info, the_lnet.ln_mt_eqh, true);
3443                         lnet_net_lock(0);
3444                         /*
3445                          * lnet_find_peer_ni_locked() grabs a refcount for
3446                          * us. No need to take it explicitly.
3447                          */
3448                         lpni = lnet_find_peer_ni_locked(nid);
3449                         if (!lpni) {
3450                                 lnet_net_unlock(0);
3451                                 LNetMDUnlink(mdh);
3452                                 continue;
3453                         }
3454
3455                         lpni->lpni_recovery_ping_mdh = mdh;
3456                         /*
3457                          * While we're unlocked the lpni could've been
3458                          * readded on the recovery queue. In this case we
3459                          * don't need to add it to the local queue, since
3460                          * it's already on there and the thread that added
3461                          * it would've incremented the refcount on the
3462                          * peer, which means we need to decref the refcount
3463                          * that was implicitly grabbed by find_peer_ni_locked.
3464                          * Otherwise, if the lpni is still not on
3465                          * the recovery queue, then we'll add it to the
3466                          * processed list.
3467                          */
3468                         if (list_empty(&lpni->lpni_recovery))
3469                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3470                         else
3471                                 lnet_peer_ni_decref_locked(lpni);
3472                         lnet_net_unlock(0);
3473
3474                         spin_lock(&lpni->lpni_lock);
3475                         if (rc)
3476                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3477                 }
3478                 spin_unlock(&lpni->lpni_lock);
3479         }
3480
3481         list_splice_init(&processed_list, &local_queue);
3482         lnet_net_lock(0);
3483         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3484         lnet_net_unlock(0);
3485 }
3486
3487 static int
3488 lnet_monitor_thread(void *arg)
3489 {
3490         time64_t recovery_timeout = 0;
3491         time64_t rsp_timeout = 0;
3492         int interval;
3493         time64_t now;
3494
3495         wait_for_completion(&the_lnet.ln_started);
3496         /*
3497          * The monitor thread takes care of the following:
3498          *  1. Checks the aliveness of routers
3499          *  2. Checks if there are messages on the resend queue to resend
3500          *     them.
3501          *  3. Check if there are any NIs on the local recovery queue and
3502          *     pings them
3503          *  4. Checks if there are any NIs on the remote recovery queue
3504          *     and pings them.
3505          */
3506         cfs_block_allsigs();
3507
3508         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3509                 now = ktime_get_real_seconds();
3510
3511                 if (lnet_router_checker_active())
3512                         lnet_check_routers();
3513
3514                 lnet_resend_pending_msgs();
3515
3516                 if (now >= rsp_timeout) {
3517                         lnet_finalize_expired_responses();
3518                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3519                 }
3520
3521                 if (now >= recovery_timeout) {
3522                         lnet_recover_local_nis();
3523                         lnet_recover_peer_nis();
3524                         recovery_timeout = now + lnet_recovery_interval;
3525                 }
3526
3527                 /*
3528                  * TODO do we need to check if we should sleep without
3529                  * timeout?  Technically, an active system will always
3530                  * have messages in flight so this check will always
3531                  * evaluate to false. And on an idle system do we care
3532                  * if we wake up every 1 second? Although, we've seen
3533                  * cases where we get a complaint that an idle thread
3534                  * is waking up unnecessarily.
3535                  *
3536                  * Take into account the current net_count when you wake
3537                  * up for alive router checking, since we need to check
3538                  * possibly as many networks as we have configured.
3539                  */
3540                 interval = min(lnet_recovery_interval,
3541                                min((unsigned int) alive_router_check_interval /
3542                                         lnet_current_net_count,
3543                                    lnet_transaction_timeout / 2));
3544                 wait_for_completion_interruptible_timeout(
3545                         &the_lnet.ln_mt_wait_complete,
3546                         cfs_time_seconds(interval));
3547                 /* Must re-init the completion before testing anything,
3548                  * including ln_mt_state.
3549                  */
3550                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3551         }
3552
3553         /* Shutting down */
3554         lnet_net_lock(LNET_LOCK_EX);
3555         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3556         lnet_net_unlock(LNET_LOCK_EX);
3557
3558         /* signal that the monitor thread is exiting */
3559         up(&the_lnet.ln_mt_signal);
3560
3561         return 0;
3562 }
3563
3564 /*
3565  * lnet_send_ping
3566  * Sends a ping.
3567  * Returns == 0 if success
3568  * Returns > 0 if LNetMDBind or prior fails
3569  * Returns < 0 if LNetGet fails
3570  */
3571 int
3572 lnet_send_ping(lnet_nid_t dest_nid,
3573                struct lnet_handle_md *mdh, int nnis,
3574                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3575 {
3576         struct lnet_md md = { NULL };
3577         struct lnet_process_id id;
3578         struct lnet_ping_buffer *pbuf;
3579         int rc;
3580
3581         if (dest_nid == LNET_NID_ANY) {
3582                 rc = -EHOSTUNREACH;
3583                 goto fail_error;
3584         }
3585
3586         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3587         if (!pbuf) {
3588                 rc = ENOMEM;
3589                 goto fail_error;
3590         }
3591
3592         /* initialize md content */
3593         md.start     = &pbuf->pb_info;
3594         md.length    = LNET_PING_INFO_SIZE(nnis);
3595         md.threshold = 2; /* GET/REPLY */
3596         md.max_size  = 0;
3597         md.options   = LNET_MD_TRUNCATE;
3598         md.user_ptr  = user_data;
3599         md.eq_handle = eqh;
3600
3601         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3602         if (rc) {
3603                 lnet_ping_buffer_decref(pbuf);
3604                 CERROR("Can't bind MD: %d\n", rc);
3605                 rc = -rc; /* change the rc to positive */
3606                 goto fail_error;
3607         }
3608         id.pid = LNET_PID_LUSTRE;
3609         id.nid = dest_nid;
3610
3611         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3612                      LNET_RESERVED_PORTAL,
3613                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3614
3615         if (rc)
3616                 goto fail_unlink_md;
3617
3618         return 0;
3619
3620 fail_unlink_md:
3621         LNetMDUnlink(*mdh);
3622         LNetInvalidateMDHandle(mdh);
3623 fail_error:
3624         return rc;
3625 }
3626
3627 static void
3628 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3629                            int status, bool unlink_event)
3630 {
3631         lnet_nid_t nid = ev_info->mt_nid;
3632
3633         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3634                 struct lnet_ni *ni;
3635
3636                 lnet_net_lock(0);
3637                 ni = lnet_nid2ni_locked(nid, 0);
3638                 if (!ni) {
3639                         lnet_net_unlock(0);
3640                         return;
3641                 }
3642                 lnet_ni_lock(ni);
3643                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3644                 if (status)
3645                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3646                 lnet_ni_unlock(ni);
3647                 lnet_net_unlock(0);
3648
3649                 if (status != 0) {
3650                         CERROR("local NI (%s) recovery failed with %d\n",
3651                                libcfs_nid2str(nid), status);
3652                         return;
3653                 }
3654                 /*
3655                  * need to increment healthv for the ni here, because in
3656                  * the lnet_finalize() path we don't have access to this
3657                  * NI. And in order to get access to it, we'll need to
3658                  * carry forward too much information.
3659                  * In the peer case, it'll naturally be incremented
3660                  */
3661                 if (!unlink_event)
3662                         lnet_inc_healthv(&ni->ni_healthv);
3663         } else {
3664                 struct lnet_peer_ni *lpni;
3665                 int cpt;
3666
3667                 cpt = lnet_net_lock_current();
3668                 lpni = lnet_find_peer_ni_locked(nid);
3669                 if (!lpni) {
3670                         lnet_net_unlock(cpt);
3671                         return;
3672                 }
3673                 spin_lock(&lpni->lpni_lock);
3674                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3675                 if (status)
3676                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3677                 spin_unlock(&lpni->lpni_lock);
3678                 lnet_peer_ni_decref_locked(lpni);
3679                 lnet_net_unlock(cpt);
3680
3681                 if (status != 0)
3682                         CERROR("peer NI (%s) recovery failed with %d\n",
3683                                libcfs_nid2str(nid), status);
3684         }
3685 }
3686
3687 void
3688 lnet_mt_event_handler(struct lnet_event *event)
3689 {
3690         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3691         struct lnet_ping_buffer *pbuf;
3692
3693         /* TODO: remove assert */
3694         LASSERT(event->type == LNET_EVENT_REPLY ||
3695                 event->type == LNET_EVENT_SEND ||
3696                 event->type == LNET_EVENT_UNLINK);
3697
3698         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3699                event->status);
3700
3701         switch (event->type) {
3702         case LNET_EVENT_UNLINK:
3703                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3704                        libcfs_nid2str(ev_info->mt_nid));
3705                 /* fallthrough */
3706         case LNET_EVENT_REPLY:
3707                 lnet_handle_recovery_reply(ev_info, event->status,
3708                                            event->type == LNET_EVENT_UNLINK);
3709                 break;
3710         case LNET_EVENT_SEND:
3711                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3712                                libcfs_nid2str(ev_info->mt_nid),
3713                                (event->status) ? "unsuccessfully" :
3714                                "successfully", event->status);
3715                 break;
3716         default:
3717                 CERROR("Unexpected event: %d\n", event->type);
3718                 break;
3719         }
3720         if (event->unlinked) {
3721                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3722                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3723                 lnet_ping_buffer_decref(pbuf);
3724         }
3725 }
3726
3727 static int
3728 lnet_rsp_tracker_create(void)
3729 {
3730         struct list_head **rstqs;
3731         rstqs = lnet_create_array_of_queues();
3732
3733         if (!rstqs)
3734                 return -ENOMEM;
3735
3736         the_lnet.ln_mt_rstq = rstqs;
3737
3738         return 0;
3739 }
3740
3741 static void
3742 lnet_rsp_tracker_clean(void)
3743 {
3744         lnet_finalize_expired_responses();
3745
3746         cfs_percpt_free(the_lnet.ln_mt_rstq);
3747         the_lnet.ln_mt_rstq = NULL;
3748 }
3749
3750 int lnet_monitor_thr_start(void)
3751 {
3752         int rc = 0;
3753         struct task_struct *task;
3754
3755         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3756                 return -EALREADY;
3757
3758         rc = lnet_resendqs_create();
3759         if (rc)
3760                 return rc;
3761
3762         rc = lnet_rsp_tracker_create();
3763         if (rc)
3764                 goto clean_queues;
3765
3766         sema_init(&the_lnet.ln_mt_signal, 0);
3767
3768         lnet_net_lock(LNET_LOCK_EX);
3769         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3770         lnet_net_unlock(LNET_LOCK_EX);
3771         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3772         if (IS_ERR(task)) {
3773                 rc = PTR_ERR(task);
3774                 CERROR("Can't start monitor thread: %d\n", rc);
3775                 goto clean_thread;
3776         }
3777
3778         return 0;
3779
3780 clean_thread:
3781         lnet_net_lock(LNET_LOCK_EX);
3782         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3783         lnet_net_unlock(LNET_LOCK_EX);
3784         /* block until event callback signals exit */
3785         down(&the_lnet.ln_mt_signal);
3786         /* clean up */
3787         lnet_net_lock(LNET_LOCK_EX);
3788         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3789         lnet_net_unlock(LNET_LOCK_EX);
3790         lnet_rsp_tracker_clean();
3791         lnet_clean_local_ni_recoveryq();
3792         lnet_clean_peer_ni_recoveryq();
3793         lnet_clean_resendqs();
3794         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3795         return rc;
3796 clean_queues:
3797         lnet_rsp_tracker_clean();
3798         lnet_clean_local_ni_recoveryq();
3799         lnet_clean_peer_ni_recoveryq();
3800         lnet_clean_resendqs();
3801         return rc;
3802 }
3803
3804 void lnet_monitor_thr_stop(void)
3805 {
3806         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3807                 return;
3808
3809         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3810         lnet_net_lock(LNET_LOCK_EX);
3811         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3812         lnet_net_unlock(LNET_LOCK_EX);
3813
3814         /* tell the monitor thread that we're shutting down */
3815         complete(&the_lnet.ln_mt_wait_complete);
3816
3817         /* block until monitor thread signals that it's done */
3818         down(&the_lnet.ln_mt_signal);
3819         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3820
3821         /* perform cleanup tasks */
3822         lnet_rsp_tracker_clean();
3823         lnet_clean_local_ni_recoveryq();
3824         lnet_clean_peer_ni_recoveryq();
3825         lnet_clean_resendqs();
3826 }
3827
3828 void
3829 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3830                   __u32 msg_type)
3831 {
3832         lnet_net_lock(cpt);
3833         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3834         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3835         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3836         lnet_net_unlock(cpt);
3837
3838         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3839 }
3840
3841 static void
3842 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3843 {
3844         struct lnet_hdr *hdr = &msg->msg_hdr;
3845
3846         if (msg->msg_wanted != 0)
3847                 lnet_setpayloadbuffer(msg);
3848
3849         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3850
3851         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3852          * it back into the ACK during lnet_finalize() */
3853         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3854                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3855
3856         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3857                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3858 }
3859
3860 static int
3861 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3862 {
3863         struct lnet_hdr         *hdr = &msg->msg_hdr;
3864         struct lnet_match_info  info;
3865         int                     rc;
3866         bool                    ready_delay;
3867
3868         /* Convert put fields to host byte order */
3869         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3870         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3871         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3872
3873         /* Primary peer NID. */
3874         info.mi_id.nid  = msg->msg_initiator;
3875         info.mi_id.pid  = hdr->src_pid;
3876         info.mi_opc     = LNET_MD_OP_PUT;
3877         info.mi_portal  = hdr->msg.put.ptl_index;
3878         info.mi_rlength = hdr->payload_length;
3879         info.mi_roffset = hdr->msg.put.offset;
3880         info.mi_mbits   = hdr->msg.put.match_bits;
3881         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3882
3883         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3884         ready_delay = msg->msg_rx_ready_delay;
3885
3886  again:
3887         rc = lnet_ptl_match_md(&info, msg);
3888         switch (rc) {
3889         default:
3890                 LBUG();
3891
3892         case LNET_MATCHMD_OK:
3893                 lnet_recv_put(ni, msg);
3894                 return 0;
3895
3896         case LNET_MATCHMD_NONE:
3897                 if (ready_delay)
3898                         /* no eager_recv or has already called it, should
3899                          * have been attached on delayed list */
3900                         return 0;
3901
3902                 rc = lnet_ni_eager_recv(ni, msg);
3903                 if (rc == 0) {
3904                         ready_delay = true;
3905                         goto again;
3906                 }
3907                 /* fall through */
3908
3909         case LNET_MATCHMD_DROP:
3910                 CNETERR("Dropping PUT from %s portal %d match %llu"
3911                         " offset %d length %d: %d\n",
3912                         libcfs_id2str(info.mi_id), info.mi_portal,
3913                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3914
3915                 return -ENOENT; /* -ve: OK but no match */
3916         }
3917 }
3918
3919 static int
3920 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3921 {
3922         struct lnet_match_info info;
3923         struct lnet_hdr *hdr = &msg->msg_hdr;
3924         struct lnet_process_id source_id;
3925         struct lnet_handle_wire reply_wmd;
3926         int rc;
3927
3928         /* Convert get fields to host byte order */
3929         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3930         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3931         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3932         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3933
3934         source_id.nid = hdr->src_nid;
3935         source_id.pid = hdr->src_pid;
3936         /* Primary peer NID */
3937         info.mi_id.nid  = msg->msg_initiator;
3938         info.mi_id.pid  = hdr->src_pid;
3939         info.mi_opc     = LNET_MD_OP_GET;
3940         info.mi_portal  = hdr->msg.get.ptl_index;
3941         info.mi_rlength = hdr->msg.get.sink_length;
3942         info.mi_roffset = hdr->msg.get.src_offset;
3943         info.mi_mbits   = hdr->msg.get.match_bits;
3944         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3945
3946         rc = lnet_ptl_match_md(&info, msg);
3947         if (rc == LNET_MATCHMD_DROP) {
3948                 CNETERR("Dropping GET from %s portal %d match %llu"
3949                         " offset %d length %d\n",
3950                         libcfs_id2str(info.mi_id), info.mi_portal,
3951                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3952                 return -ENOENT; /* -ve: OK but no match */
3953         }
3954
3955         LASSERT(rc == LNET_MATCHMD_OK);
3956
3957         lnet_build_msg_event(msg, LNET_EVENT_GET);
3958
3959         reply_wmd = hdr->msg.get.return_wmd;
3960
3961         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3962                        msg->msg_offset, msg->msg_wanted);
3963
3964         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3965
3966         if (rdma_get) {
3967                 /* The LND completes the REPLY from her recv procedure */
3968                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3969                              msg->msg_offset, msg->msg_len, msg->msg_len);
3970                 return 0;
3971         }
3972
3973         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3974         msg->msg_receiving = 0;
3975
3976         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3977         if (rc < 0) {
3978                 /* didn't get as far as lnet_ni_send() */
3979                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3980                        libcfs_nid2str(ni->ni_nid),
3981                        libcfs_id2str(info.mi_id), rc);
3982
3983                 lnet_finalize(msg, rc);
3984         }
3985
3986         return 0;
3987 }
3988
3989 static int
3990 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3991 {
3992         void *private = msg->msg_private;
3993         struct lnet_hdr *hdr = &msg->msg_hdr;
3994         struct lnet_process_id src = {0};
3995         struct lnet_libmd *md;
3996         int rlength;
3997         int mlength;
3998         int cpt;
3999
4000         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4001         lnet_res_lock(cpt);
4002
4003         src.nid = hdr->src_nid;
4004         src.pid = hdr->src_pid;
4005
4006         /* NB handles only looked up by creator (no flips) */
4007         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4008         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4009                 CNETERR("%s: Dropping REPLY from %s for %s "
4010                         "MD %#llx.%#llx\n",
4011                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4012                         (md == NULL) ? "invalid" : "inactive",
4013                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4014                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4015                 if (md != NULL && md->md_me != NULL)
4016                         CERROR("REPLY MD also attached to portal %d\n",
4017                                md->md_me->me_portal);
4018
4019                 lnet_res_unlock(cpt);
4020                 return -ENOENT; /* -ve: OK but no match */
4021         }
4022
4023         LASSERT(md->md_offset == 0);
4024
4025         rlength = hdr->payload_length;
4026         mlength = MIN(rlength, (int)md->md_length);
4027
4028         if (mlength < rlength &&
4029             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4030                 CNETERR("%s: Dropping REPLY from %s length %d "
4031                         "for MD %#llx would overflow (%d)\n",
4032                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4033                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4034                         mlength);
4035                 lnet_res_unlock(cpt);
4036                 return -ENOENT; /* -ve: OK but no match */
4037         }
4038
4039         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4040                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4041                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4042
4043         lnet_msg_attach_md(msg, md, 0, mlength);
4044
4045         if (mlength != 0)
4046                 lnet_setpayloadbuffer(msg);
4047
4048         lnet_res_unlock(cpt);
4049
4050         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4051
4052         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4053         return 0;
4054 }
4055
4056 static int
4057 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4058 {
4059         struct lnet_hdr *hdr = &msg->msg_hdr;
4060         struct lnet_process_id src = {0};
4061         struct lnet_libmd *md;
4062         int cpt;
4063
4064         src.nid = hdr->src_nid;
4065         src.pid = hdr->src_pid;
4066
4067         /* Convert ack fields to host byte order */
4068         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4069         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4070
4071         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4072         lnet_res_lock(cpt);
4073
4074         /* NB handles only looked up by creator (no flips) */
4075         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4076         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4077                 /* Don't moan; this is expected */
4078                 CDEBUG(D_NET,
4079                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4080                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4081                        (md == NULL) ? "invalid" : "inactive",
4082                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4083                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4084                 if (md != NULL && md->md_me != NULL)
4085                         CERROR("Source MD also attached to portal %d\n",
4086                                md->md_me->me_portal);
4087
4088                 lnet_res_unlock(cpt);
4089                 return -ENOENT;                  /* -ve! */
4090         }
4091
4092         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4093                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4094                hdr->msg.ack.dst_wmd.wh_object_cookie);
4095
4096         lnet_msg_attach_md(msg, md, 0, 0);
4097
4098         lnet_res_unlock(cpt);
4099
4100         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4101
4102         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4103         return 0;
4104 }
4105
4106 /**
4107  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4108  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4109  * \retval -ve                  error code
4110  */
4111 int
4112 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4113 {
4114         int     rc = 0;
4115
4116         if (!the_lnet.ln_routing)
4117                 return -ECANCELED;
4118
4119         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4120             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4121                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4122                         msg->msg_rx_ready_delay = 1;
4123                 } else {
4124                         lnet_net_unlock(msg->msg_rx_cpt);
4125                         rc = lnet_ni_eager_recv(ni, msg);
4126                         lnet_net_lock(msg->msg_rx_cpt);
4127                 }
4128         }
4129
4130         if (rc == 0)
4131                 rc = lnet_post_routed_recv_locked(msg, 0);
4132         return rc;
4133 }
4134
4135 int
4136 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4137 {
4138         int     rc;
4139
4140         switch (msg->msg_type) {
4141         case LNET_MSG_ACK:
4142                 rc = lnet_parse_ack(ni, msg);
4143                 break;
4144         case LNET_MSG_PUT:
4145                 rc = lnet_parse_put(ni, msg);
4146                 break;
4147         case LNET_MSG_GET:
4148                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4149                 break;
4150         case LNET_MSG_REPLY:
4151                 rc = lnet_parse_reply(ni, msg);
4152                 break;
4153         default: /* prevent an unused label if !kernel */
4154                 LASSERT(0);
4155                 return -EPROTO;
4156         }
4157
4158         LASSERT(rc == 0 || rc == -ENOENT);
4159         return rc;
4160 }
4161
4162 char *
4163 lnet_msgtyp2str (int type)
4164 {
4165         switch (type) {
4166         case LNET_MSG_ACK:
4167                 return ("ACK");
4168         case LNET_MSG_PUT:
4169                 return ("PUT");
4170         case LNET_MSG_GET:
4171                 return ("GET");
4172         case LNET_MSG_REPLY:
4173                 return ("REPLY");
4174         case LNET_MSG_HELLO:
4175                 return ("HELLO");
4176         default:
4177                 return ("<UNKNOWN>");
4178         }
4179 }
4180
4181 void
4182 lnet_print_hdr(struct lnet_hdr *hdr)
4183 {
4184         struct lnet_process_id src = {
4185                 .nid = hdr->src_nid,
4186                 .pid = hdr->src_pid,
4187         };
4188         struct lnet_process_id dst = {
4189                 .nid = hdr->dest_nid,
4190                 .pid = hdr->dest_pid,
4191         };
4192         char *type_str = lnet_msgtyp2str(hdr->type);
4193
4194         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4195         CWARN("    From %s\n", libcfs_id2str(src));
4196         CWARN("    To   %s\n", libcfs_id2str(dst));
4197
4198         switch (hdr->type) {
4199         default:
4200                 break;
4201
4202         case LNET_MSG_PUT:
4203                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4204                       "match bits %llu\n",
4205                       hdr->msg.put.ptl_index,
4206                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4207                       hdr->msg.put.ack_wmd.wh_object_cookie,
4208                       hdr->msg.put.match_bits);
4209                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4210                       hdr->payload_length, hdr->msg.put.offset,
4211                       hdr->msg.put.hdr_data);
4212                 break;
4213
4214         case LNET_MSG_GET:
4215                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4216                       "match bits %llu\n", hdr->msg.get.ptl_index,
4217                       hdr->msg.get.return_wmd.wh_interface_cookie,
4218                       hdr->msg.get.return_wmd.wh_object_cookie,
4219                       hdr->msg.get.match_bits);
4220                 CWARN("    Length %d, src offset %d\n",
4221                       hdr->msg.get.sink_length,
4222                       hdr->msg.get.src_offset);
4223                 break;
4224
4225         case LNET_MSG_ACK:
4226                 CWARN("    dst md %#llx.%#llx, "
4227                       "manipulated length %d\n",
4228                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4229                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4230                       hdr->msg.ack.mlength);
4231                 break;
4232
4233         case LNET_MSG_REPLY:
4234                 CWARN("    dst md %#llx.%#llx, "
4235                       "length %d\n",
4236                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4237                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4238                       hdr->payload_length);
4239         }
4240
4241 }
4242
4243 int
4244 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4245            void *private, int rdma_req)
4246 {
4247         struct lnet_peer_ni *lpni;
4248         struct lnet_msg *msg;
4249         __u32 payload_length;
4250         lnet_pid_t dest_pid;
4251         lnet_nid_t dest_nid;
4252         lnet_nid_t src_nid;
4253         bool push = false;
4254         int for_me;
4255         __u32 type;
4256         int rc = 0;
4257         int cpt;
4258
4259         LASSERT (!in_interrupt ());
4260
4261         type = le32_to_cpu(hdr->type);
4262         src_nid = le64_to_cpu(hdr->src_nid);
4263         dest_nid = le64_to_cpu(hdr->dest_nid);
4264         dest_pid = le32_to_cpu(hdr->dest_pid);
4265         payload_length = le32_to_cpu(hdr->payload_length);
4266
4267         for_me = (ni->ni_nid == dest_nid);
4268         cpt = lnet_cpt_of_nid(from_nid, ni);
4269
4270         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4271                 libcfs_nid2str(dest_nid),
4272                 libcfs_nid2str(ni->ni_nid),
4273                 libcfs_nid2str(src_nid),
4274                 lnet_msgtyp2str(type),
4275                 (for_me) ? "for me" : "routed");
4276
4277         switch (type) {
4278         case LNET_MSG_ACK:
4279         case LNET_MSG_GET:
4280                 if (payload_length > 0) {
4281                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4282                                libcfs_nid2str(from_nid),
4283                                libcfs_nid2str(src_nid),
4284                                lnet_msgtyp2str(type), payload_length);
4285                         return -EPROTO;
4286                 }
4287                 break;
4288
4289         case LNET_MSG_PUT:
4290         case LNET_MSG_REPLY:
4291                 if (payload_length >
4292                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4293                         CERROR("%s, src %s: bad %s payload %d "
4294                                "(%d max expected)\n",
4295                                libcfs_nid2str(from_nid),
4296                                libcfs_nid2str(src_nid),
4297                                lnet_msgtyp2str(type),
4298                                payload_length,
4299                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4300                         return -EPROTO;
4301                 }
4302                 break;
4303
4304         default:
4305                 CERROR("%s, src %s: Bad message type 0x%x\n",
4306                        libcfs_nid2str(from_nid),
4307                        libcfs_nid2str(src_nid), type);
4308                 return -EPROTO;
4309         }
4310
4311         if (the_lnet.ln_routing &&
4312             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4313                 lnet_ni_lock(ni);
4314                 spin_lock(&ni->ni_net->net_lock);
4315                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4316                 spin_unlock(&ni->ni_net->net_lock);
4317                 if (ni->ni_status != NULL &&
4318                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4319                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4320                         push = true;
4321                 }
4322                 lnet_ni_unlock(ni);
4323         }
4324
4325         if (push)
4326                 lnet_push_update_to_peers(1);
4327
4328         /* Regard a bad destination NID as a protocol error.  Senders should
4329          * know what they're doing; if they don't they're misconfigured, buggy
4330          * or malicious so we chop them off at the knees :) */
4331
4332         if (!for_me) {
4333                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4334                         /* should have gone direct */
4335                         CERROR("%s, src %s: Bad dest nid %s "
4336                                "(should have been sent direct)\n",
4337                                 libcfs_nid2str(from_nid),
4338                                 libcfs_nid2str(src_nid),
4339                                 libcfs_nid2str(dest_nid));
4340                         return -EPROTO;
4341                 }
4342
4343                 if (lnet_islocalnid(dest_nid)) {
4344                         /* dest is another local NI; sender should have used
4345                          * this node's NID on its own network */
4346                         CERROR("%s, src %s: Bad dest nid %s "
4347                                "(it's my nid but on a different network)\n",
4348                                 libcfs_nid2str(from_nid),
4349                                 libcfs_nid2str(src_nid),
4350                                 libcfs_nid2str(dest_nid));
4351                         return -EPROTO;
4352                 }
4353
4354                 if (rdma_req && type == LNET_MSG_GET) {
4355                         CERROR("%s, src %s: Bad optimized GET for %s "
4356                                "(final destination must be me)\n",
4357                                 libcfs_nid2str(from_nid),
4358                                 libcfs_nid2str(src_nid),
4359                                 libcfs_nid2str(dest_nid));
4360                         return -EPROTO;
4361                 }
4362
4363                 if (!the_lnet.ln_routing) {
4364                         CERROR("%s, src %s: Dropping message for %s "
4365                                "(routing not enabled)\n",
4366                                 libcfs_nid2str(from_nid),
4367                                 libcfs_nid2str(src_nid),
4368                                 libcfs_nid2str(dest_nid));
4369                         goto drop;
4370                 }
4371         }
4372
4373         /* Message looks OK; we're not going to return an error, so we MUST
4374          * call back lnd_recv() come what may... */
4375
4376         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4377             fail_peer(src_nid, 0)) {                    /* shall we now? */
4378                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4379                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4380                        lnet_msgtyp2str(type));
4381                 goto drop;
4382         }
4383
4384         if (!list_empty(&the_lnet.ln_drop_rules) &&
4385             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4386                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4387                               "silent message loss\n",
4388                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4389                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4390                 goto drop;
4391         }
4392
4393         if (lnet_drop_asym_route && for_me &&
4394             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4395                 struct lnet_net *net;
4396                 struct lnet_remotenet *rnet;
4397                 bool found = true;
4398
4399                 /* we are dealing with a routed message,
4400                  * so see if route to reach src_nid goes through from_nid
4401                  */
4402                 lnet_net_lock(cpt);
4403                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4404                 if (!net) {
4405                         lnet_net_unlock(cpt);
4406                         CERROR("net %s not found\n",
4407                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4408                         return -EPROTO;
4409                 }
4410
4411                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4412                 if (rnet) {
4413                         struct lnet_peer *gw = NULL;
4414                         struct lnet_peer_ni *lpni = NULL;
4415                         struct lnet_route *route;
4416
4417                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4418                                 found = false;
4419                                 gw = route->lr_gateway;
4420                                 if (route->lr_lnet != net->net_id)
4421                                         continue;
4422                                 /*
4423                                  * if the nid is one of the gateway's NIDs
4424                                  * then this is a valid gateway
4425                                  */
4426                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4427                                                 NULL, lpni)) != NULL) {
4428                                         if (lpni->lpni_nid == from_nid) {
4429                                                 found = true;
4430                                                 break;
4431                                         }
4432                                 }
4433                         }
4434                 }
4435                 lnet_net_unlock(cpt);
4436                 if (!found) {
4437                         /* we would not use from_nid to route a message to
4438                          * src_nid
4439                          * => asymmetric routing detected but forbidden
4440                          */
4441                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4442                                libcfs_nid2str(from_nid),
4443                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4444                         goto drop;
4445                 }
4446         }
4447
4448         msg = lnet_msg_alloc();
4449         if (msg == NULL) {
4450                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4451                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4452                        lnet_msgtyp2str(type));
4453                 goto drop;
4454         }
4455
4456         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4457          * pointers NULL etc */
4458
4459         msg->msg_type = type;
4460         msg->msg_private = private;
4461         msg->msg_receiving = 1;
4462         msg->msg_rdma_get = rdma_req;
4463         msg->msg_len = msg->msg_wanted = payload_length;
4464         msg->msg_offset = 0;
4465         msg->msg_hdr = *hdr;
4466         /* for building message event */
4467         msg->msg_from = from_nid;
4468         if (!for_me) {
4469                 msg->msg_target.pid     = dest_pid;
4470                 msg->msg_target.nid     = dest_nid;
4471                 msg->msg_routing        = 1;
4472
4473         } else {
4474                 /* convert common msg->hdr fields to host byteorder */
4475                 msg->msg_hdr.type       = type;
4476                 msg->msg_hdr.src_nid    = src_nid;
4477                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4478                 msg->msg_hdr.dest_nid   = dest_nid;
4479                 msg->msg_hdr.dest_pid   = dest_pid;
4480                 msg->msg_hdr.payload_length = payload_length;
4481         }
4482
4483         lnet_net_lock(cpt);
4484         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4485         if (IS_ERR(lpni)) {
4486                 lnet_net_unlock(cpt);
4487                 CERROR("%s, src %s: Dropping %s "
4488                        "(error %ld looking up sender)\n",
4489                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4490                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4491                 lnet_msg_free(msg);
4492                 if (rc == -ESHUTDOWN)
4493                         /* We are shutting down.  Don't do anything more */
4494                         return 0;
4495                 goto drop;
4496         }
4497
4498         if (the_lnet.ln_routing)
4499                 lpni->lpni_last_alive = ktime_get_seconds();
4500
4501         msg->msg_rxpeer = lpni;
4502         msg->msg_rxni = ni;
4503         lnet_ni_addref_locked(ni, cpt);
4504         /* Multi-Rail: Primary NID of source. */
4505         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4506
4507         /*
4508          * mark the status of this lpni as UP since we received a message
4509          * from it. The ping response reports back the ns_status which is
4510          * marked on the remote as up or down and we cache it here.
4511          */
4512         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4513
4514         lnet_msg_commit(msg, cpt);
4515
4516         /* message delay simulation */
4517         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4518                      lnet_delay_rule_match_locked(hdr, msg))) {
4519                 lnet_net_unlock(cpt);
4520                 return 0;
4521         }
4522
4523         if (!for_me) {
4524                 rc = lnet_parse_forward_locked(ni, msg);
4525                 lnet_net_unlock(cpt);
4526
4527                 if (rc < 0)
4528                         goto free_drop;
4529
4530                 if (rc == LNET_CREDIT_OK) {
4531                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4532                                      0, payload_length, payload_length);
4533                 }
4534                 return 0;
4535         }
4536
4537         lnet_net_unlock(cpt);
4538
4539         rc = lnet_parse_local(ni, msg);
4540         if (rc != 0)
4541                 goto free_drop;
4542         return 0;
4543
4544  free_drop:
4545         LASSERT(msg->msg_md == NULL);
4546         lnet_finalize(msg, rc);
4547
4548  drop:
4549         lnet_drop_message(ni, cpt, private, payload_length, type);
4550         return 0;
4551 }
4552 EXPORT_SYMBOL(lnet_parse);
4553
4554 void
4555 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4556 {
4557         while (!list_empty(head)) {
4558                 struct lnet_process_id id = {0};
4559                 struct lnet_msg *msg;
4560
4561                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4562                 list_del(&msg->msg_list);
4563
4564                 id.nid = msg->msg_hdr.src_nid;
4565                 id.pid = msg->msg_hdr.src_pid;
4566
4567                 LASSERT(msg->msg_md == NULL);
4568                 LASSERT(msg->msg_rx_delayed);
4569                 LASSERT(msg->msg_rxpeer != NULL);
4570                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4571
4572                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4573                       " offset %d length %d: %s\n",
4574                       libcfs_id2str(id),
4575                       msg->msg_hdr.msg.put.ptl_index,
4576                       msg->msg_hdr.msg.put.match_bits,
4577                       msg->msg_hdr.msg.put.offset,
4578                       msg->msg_hdr.payload_length, reason);
4579
4580                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4581                  * called lnet_drop_message(), so I just hang onto msg as well
4582                  * until that's done */
4583
4584                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4585                                   msg->msg_private, msg->msg_len,
4586                                   msg->msg_type);
4587
4588                 msg->msg_no_resend = true;
4589                 /*
4590                  * NB: message will not generate event because w/o attached MD,
4591                  * but we still should give error code so lnet_msg_decommit()
4592                  * can skip counters operations and other checks.
4593                  */
4594                 lnet_finalize(msg, -ENOENT);
4595         }
4596 }
4597
4598 void
4599 lnet_recv_delayed_msg_list(struct list_head *head)
4600 {
4601         while (!list_empty(head)) {
4602                 struct lnet_msg *msg;
4603                 struct lnet_process_id id;
4604
4605                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4606                 list_del(&msg->msg_list);
4607
4608                 /* md won't disappear under me, since each msg
4609                  * holds a ref on it */
4610
4611                 id.nid = msg->msg_hdr.src_nid;
4612                 id.pid = msg->msg_hdr.src_pid;
4613
4614                 LASSERT(msg->msg_rx_delayed);
4615                 LASSERT(msg->msg_md != NULL);
4616                 LASSERT(msg->msg_rxpeer != NULL);
4617                 LASSERT(msg->msg_rxni != NULL);
4618                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4619
4620                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4621                        "match %llu offset %d length %d.\n",
4622                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4623                         msg->msg_hdr.msg.put.match_bits,
4624                         msg->msg_hdr.msg.put.offset,
4625                         msg->msg_hdr.payload_length);
4626
4627                 lnet_recv_put(msg->msg_rxni, msg);
4628         }
4629 }
4630
4631 static void
4632 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4633                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4634 {
4635         s64 timeout_ns;
4636         bool new_entry = true;
4637         struct lnet_rsp_tracker *local_rspt;
4638
4639         /*
4640          * MD has a refcount taken by message so it's not going away.
4641          * The MD however can be looked up. We need to secure the access
4642          * to the md_rspt_ptr by taking the res_lock.
4643          * The rspt can be accessed without protection up to when it gets
4644          * added to the list.
4645          */
4646
4647         lnet_res_lock(cpt);
4648         local_rspt = md->md_rspt_ptr;
4649         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4650         if (local_rspt != NULL) {
4651                 /*
4652                  * we already have an rspt attached to the md, so we'll
4653                  * update the deadline on that one.
4654                  */
4655                 LIBCFS_FREE(rspt, sizeof(*rspt));
4656                 new_entry = false;
4657         } else {
4658                 /* new md */
4659                 rspt->rspt_mdh = mdh;
4660                 rspt->rspt_cpt = cpt;
4661                 /* store the rspt so we can access it when we get the REPLY */
4662                 md->md_rspt_ptr = rspt;
4663                 local_rspt = rspt;
4664         }
4665         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4666
4667         /*
4668          * add to the list of tracked responses. It's added to tail of the
4669          * list in order to expire all the older entries first.
4670          */
4671         lnet_net_lock(cpt);
4672         if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4673                 list_del_init(&local_rspt->rspt_on_list);
4674         list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4675         lnet_net_unlock(cpt);
4676         lnet_res_unlock(cpt);
4677 }
4678
4679 /**
4680  * Initiate an asynchronous PUT operation.
4681  *
4682  * There are several events associated with a PUT: completion of the send on
4683  * the initiator node (LNET_EVENT_SEND), and when the send completes
4684  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4685  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4686  * used at the target node to indicate the completion of incoming data
4687  * delivery.
4688  *
4689  * The local events will be logged in the EQ associated with the MD pointed to
4690  * by \a mdh handle. Using a MD without an associated EQ results in these
4691  * events being discarded. In this case, the caller must have another
4692  * mechanism (e.g., a higher level protocol) for determining when it is safe
4693  * to modify the memory region associated with the MD.
4694  *
4695  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4696  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4697  *
4698  * \param self Indicates the NID of a local interface through which to send
4699  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4700  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4701  * must be "free floating" (See LNetMDBind()).
4702  * \param ack Controls whether an acknowledgment is requested.
4703  * Acknowledgments are only sent when they are requested by the initiating
4704  * process and the target MD enables them.
4705  * \param target A process identifier for the target process.
4706  * \param portal The index in the \a target's portal table.
4707  * \param match_bits The match bits to use for MD selection at the target
4708  * process.
4709  * \param offset The offset into the target MD (only used when the target
4710  * MD has the LNET_MD_MANAGE_REMOTE option set).
4711  * \param hdr_data 64 bits of user data that can be included in the message
4712  * header. This data is written to an event queue entry at the target if an
4713  * EQ is present on the matching MD.
4714  *
4715  * \retval  0      Success, and only in this case events will be generated
4716  * and logged to EQ (if it exists).
4717  * \retval -EIO    Simulated failure.
4718  * \retval -ENOMEM Memory allocation failure.
4719  * \retval -ENOENT Invalid MD object.
4720  *
4721  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4722  */
4723 int
4724 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4725         struct lnet_process_id target, unsigned int portal,
4726         __u64 match_bits, unsigned int offset,
4727         __u64 hdr_data)
4728 {
4729         struct lnet_msg *msg;
4730         struct lnet_libmd *md;
4731         int cpt;
4732         int rc;
4733         struct lnet_rsp_tracker *rspt = NULL;
4734
4735         LASSERT(the_lnet.ln_refcount > 0);
4736
4737         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4738             fail_peer(target.nid, 1)) {                 /* shall we now? */
4739                 CERROR("Dropping PUT to %s: simulated failure\n",
4740                        libcfs_id2str(target));
4741                 return -EIO;
4742         }
4743
4744         msg = lnet_msg_alloc();
4745         if (msg == NULL) {
4746                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4747                        libcfs_id2str(target));
4748                 return -ENOMEM;
4749         }
4750         msg->msg_vmflush = !!memory_pressure_get();
4751
4752         cpt = lnet_cpt_of_cookie(mdh.cookie);
4753
4754         if (ack == LNET_ACK_REQ) {
4755                 rspt = lnet_rspt_alloc(cpt);
4756                 if (!rspt) {
4757                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4758                                 libcfs_id2str(target));
4759                         return -ENOMEM;
4760                 }
4761                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4762         }
4763
4764         lnet_res_lock(cpt);
4765
4766         md = lnet_handle2md(&mdh);
4767         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4768                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4769                        match_bits, portal, libcfs_id2str(target),
4770                        md == NULL ? -1 : md->md_threshold);
4771                 if (md != NULL && md->md_me != NULL)
4772                         CERROR("Source MD also attached to portal %d\n",
4773                                md->md_me->me_portal);
4774                 lnet_res_unlock(cpt);
4775
4776                 LIBCFS_FREE(rspt, sizeof(*rspt));
4777                 lnet_msg_free(msg);
4778                 return -ENOENT;
4779         }
4780
4781         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4782
4783         lnet_msg_attach_md(msg, md, 0, 0);
4784
4785         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4786
4787         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4788         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4789         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4790         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4791
4792         /* NB handles only looked up by creator (no flips) */
4793         if (ack == LNET_ACK_REQ) {
4794                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4795                         the_lnet.ln_interface_cookie;
4796                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4797                         md->md_lh.lh_cookie;
4798         } else {
4799                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4800                         LNET_WIRE_HANDLE_COOKIE_NONE;
4801                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4802                         LNET_WIRE_HANDLE_COOKIE_NONE;
4803         }
4804
4805         lnet_res_unlock(cpt);
4806
4807         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4808
4809         if (ack == LNET_ACK_REQ)
4810                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4811
4812         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4813                                  CFS_FAIL_ONCE))
4814                 rc = -EIO;
4815         else
4816                 rc = lnet_send(self, msg, LNET_NID_ANY);
4817
4818         if (rc != 0) {
4819                 CNETERR("Error sending PUT to %s: %d\n",
4820                         libcfs_id2str(target), rc);
4821                 msg->msg_no_resend = true;
4822                 lnet_finalize(msg, rc);
4823         }
4824
4825         /* completion will be signalled by an event */
4826         return 0;
4827 }
4828 EXPORT_SYMBOL(LNetPut);
4829
4830 /*
4831  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4832  * returns a msg for the LND to pass to lnet_finalize() when the sink
4833  * data has been received.
4834  *
4835  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4836  * lnet_finalize() is called on it, so the LND must call this first
4837  */
4838 struct lnet_msg *
4839 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4840 {
4841         struct lnet_msg *msg = lnet_msg_alloc();
4842         struct lnet_libmd *getmd = getmsg->msg_md;
4843         struct lnet_process_id peer_id = getmsg->msg_target;
4844         int cpt;
4845
4846         LASSERT(!getmsg->msg_target_is_router);
4847         LASSERT(!getmsg->msg_routing);
4848
4849         if (msg == NULL) {
4850                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4851                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4852                 goto drop;
4853         }
4854
4855         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4856         lnet_res_lock(cpt);
4857
4858         LASSERT(getmd->md_refcount > 0);
4859
4860         if (getmd->md_threshold == 0) {
4861                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4862                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4863                         getmd);
4864                 lnet_res_unlock(cpt);
4865                 goto drop;
4866         }
4867
4868         LASSERT(getmd->md_offset == 0);
4869
4870         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4871                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4872
4873         /* setup information for lnet_build_msg_event */
4874         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4875         msg->msg_from = peer_id.nid;
4876         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4877         msg->msg_hdr.src_nid = peer_id.nid;
4878         msg->msg_hdr.payload_length = getmd->md_length;
4879         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4880
4881         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4882         lnet_res_unlock(cpt);
4883
4884         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4885
4886         lnet_net_lock(cpt);
4887         lnet_msg_commit(msg, cpt);
4888         lnet_net_unlock(cpt);
4889
4890         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4891
4892         return msg;
4893
4894  drop:
4895         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4896
4897         lnet_net_lock(cpt);
4898         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4899         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4900         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4901                 getmd->md_length;
4902         lnet_net_unlock(cpt);
4903
4904         if (msg != NULL)
4905                 lnet_msg_free(msg);
4906
4907         return NULL;
4908 }
4909 EXPORT_SYMBOL(lnet_create_reply_msg);
4910
4911 void
4912 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4913                        unsigned int len)
4914 {
4915         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4916          * completed and I know it. */
4917         LASSERT(reply != NULL);
4918         LASSERT(reply->msg_type == LNET_MSG_GET);
4919         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4920
4921         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4922          * the end of my buffer, I might as well be dead. */
4923         LASSERT(len <= reply->msg_ev.mlength);
4924
4925         reply->msg_ev.mlength = len;
4926 }
4927 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4928
4929 /**
4930  * Initiate an asynchronous GET operation.
4931  *
4932  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4933  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4934  * the target node in the REPLY has been written to local MD.
4935  *
4936  * On the target node, an LNET_EVENT_GET is logged when the GET request
4937  * arrives and is accepted into a MD.
4938  *
4939  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4940  * \param mdh A handle for the MD that describes the memory into which the
4941  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4942  *
4943  * \retval  0      Success, and only in this case events will be generated
4944  * and logged to EQ (if it exists) of the MD.
4945  * \retval -EIO    Simulated failure.
4946  * \retval -ENOMEM Memory allocation failure.
4947  * \retval -ENOENT Invalid MD object.
4948  */
4949 int
4950 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4951         struct lnet_process_id target, unsigned int portal,
4952         __u64 match_bits, unsigned int offset, bool recovery)
4953 {
4954         struct lnet_msg *msg;
4955         struct lnet_libmd *md;
4956         struct lnet_rsp_tracker *rspt;
4957         int cpt;
4958         int rc;
4959
4960         LASSERT(the_lnet.ln_refcount > 0);
4961
4962         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4963             fail_peer(target.nid, 1))                   /* shall we now? */
4964         {
4965                 CERROR("Dropping GET to %s: simulated failure\n",
4966                        libcfs_id2str(target));
4967                 return -EIO;
4968         }
4969
4970         msg = lnet_msg_alloc();
4971         if (!msg) {
4972                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4973                        libcfs_id2str(target));
4974                 return -ENOMEM;
4975         }
4976
4977         cpt = lnet_cpt_of_cookie(mdh.cookie);
4978
4979         rspt = lnet_rspt_alloc(cpt);
4980         if (!rspt) {
4981                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4982                        libcfs_id2str(target));
4983                 return -ENOMEM;
4984         }
4985         INIT_LIST_HEAD(&rspt->rspt_on_list);
4986
4987         msg->msg_recovery = recovery;
4988
4989         lnet_res_lock(cpt);
4990
4991         md = lnet_handle2md(&mdh);
4992         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4993                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4994                        match_bits, portal, libcfs_id2str(target),
4995                        md == NULL ? -1 : md->md_threshold);
4996                 if (md != NULL && md->md_me != NULL)
4997                         CERROR("REPLY MD also attached to portal %d\n",
4998                                md->md_me->me_portal);
4999
5000                 lnet_res_unlock(cpt);
5001
5002                 lnet_msg_free(msg);
5003                 LIBCFS_FREE(rspt, sizeof(*rspt));
5004                 return -ENOENT;
5005         }
5006
5007         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5008
5009         lnet_msg_attach_md(msg, md, 0, 0);
5010
5011         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5012
5013         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5014         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5015         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5016         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5017
5018         /* NB handles only looked up by creator (no flips) */
5019         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5020                 the_lnet.ln_interface_cookie;
5021         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5022                 md->md_lh.lh_cookie;
5023
5024         lnet_res_unlock(cpt);
5025
5026         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5027
5028         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5029
5030         rc = lnet_send(self, msg, LNET_NID_ANY);
5031         if (rc < 0) {
5032                 CNETERR("Error sending GET to %s: %d\n",
5033                         libcfs_id2str(target), rc);
5034                 msg->msg_no_resend = true;
5035                 lnet_finalize(msg, rc);
5036         }
5037
5038         /* completion will be signalled by an event */
5039         return 0;
5040 }
5041 EXPORT_SYMBOL(LNetGet);
5042
5043 /**
5044  * Calculate distance to node at \a dstnid.
5045  *
5046  * \param dstnid Target NID.
5047  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5048  * is saved here.
5049  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5050  * here.
5051  *
5052  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5053  * local_nid_dist_zero is set, which is the default.
5054  * \retval positives Distance to target NID, i.e. number of hops plus one.
5055  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5056  */
5057 int
5058 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5059 {
5060         struct list_head        *e;
5061         struct lnet_ni *ni = NULL;
5062         struct lnet_remotenet *rnet;
5063         __u32                   dstnet = LNET_NIDNET(dstnid);
5064         int                     hops;
5065         int                     cpt;
5066         __u32                   order = 2;
5067         struct list_head        *rn_list;
5068
5069         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5070          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5071          * keep order 0 free for 0@lo and order 1 free for a local NID
5072          * match */
5073
5074         LASSERT(the_lnet.ln_refcount > 0);
5075
5076         cpt = lnet_net_lock_current();
5077
5078         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5079                 if (ni->ni_nid == dstnid) {
5080                         if (srcnidp != NULL)
5081                                 *srcnidp = dstnid;
5082                         if (orderp != NULL) {
5083                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5084                                         *orderp = 0;
5085                                 else
5086                                         *orderp = 1;
5087                         }
5088                         lnet_net_unlock(cpt);
5089
5090                         return local_nid_dist_zero ? 0 : 1;
5091                 }
5092
5093                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5094                         /* Check if ni was originally created in
5095                          * current net namespace.
5096                          * If not, assign order above 0xffff0000,
5097                          * to make this ni not a priority. */
5098                         if (current->nsproxy &&
5099                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5100                                         order += 0xffff0000;
5101                         if (srcnidp != NULL)
5102                                 *srcnidp = ni->ni_nid;
5103                         if (orderp != NULL)
5104                                 *orderp = order;
5105                         lnet_net_unlock(cpt);
5106                         return 1;
5107                 }
5108
5109                 order++;
5110         }
5111
5112         rn_list = lnet_net2rnethash(dstnet);
5113         list_for_each(e, rn_list) {
5114                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5115
5116                 if (rnet->lrn_net == dstnet) {
5117                         struct lnet_route *route;
5118                         struct lnet_route *shortest = NULL;
5119                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5120                         __u32 route_hops;
5121
5122                         LASSERT(!list_empty(&rnet->lrn_routes));
5123
5124                         list_for_each_entry(route, &rnet->lrn_routes,
5125                                             lr_list) {
5126                                 route_hops = route->lr_hops;
5127                                 if (route_hops == LNET_UNDEFINED_HOPS)
5128                                         route_hops = 1;
5129                                 if (shortest == NULL ||
5130                                     route_hops < shortest_hops) {
5131                                         shortest = route;
5132                                         shortest_hops = route_hops;
5133                                 }
5134                         }
5135
5136                         LASSERT(shortest != NULL);
5137                         hops = shortest_hops;
5138                         if (srcnidp != NULL) {
5139                                 struct lnet_net *net;
5140                                 net = lnet_get_net_locked(shortest->lr_lnet);
5141                                 LASSERT(net);
5142                                 ni = lnet_get_next_ni_locked(net, NULL);
5143                                 *srcnidp = ni->ni_nid;
5144                         }
5145                         if (orderp != NULL)
5146                                 *orderp = order;
5147                         lnet_net_unlock(cpt);
5148                         return hops + 1;
5149                 }
5150                 order++;
5151         }
5152
5153         lnet_net_unlock(cpt);
5154         return -EHOSTUNREACH;
5155 }
5156 EXPORT_SYMBOL(LNetDist);