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