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 #if 0
1298 static int
1299 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1300 {
1301         if (p1->lpni_txqnob < p2->lpni_txqnob)
1302                 return 1;
1303
1304         if (p1->lpni_txqnob > p2->lpni_txqnob)
1305                 return -1;
1306
1307         if (p1->lpni_txcredits > p2->lpni_txcredits)
1308                 return 1;
1309
1310         if (p1->lpni_txcredits < p2->lpni_txcredits)
1311                 return -1;
1312
1313         return 0;
1314 }
1315 #endif
1316
1317 static int
1318 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1319 {
1320         /* TODO re-implement gateway comparison
1321         struct lnet_peer_ni *p1 = r1->lr_gateway;
1322         struct lnet_peer_ni *p2 = r2->lr_gateway;
1323         */
1324         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1325         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1326         /*int rc;*/
1327
1328         if (r1->lr_priority < r2->lr_priority)
1329                 return 1;
1330
1331         if (r1->lr_priority > r2->lr_priority)
1332                 return -1;
1333
1334         if (r1_hops < r2_hops)
1335                 return 1;
1336
1337         if (r1_hops > r2_hops)
1338                 return -1;
1339
1340         /*
1341         rc = lnet_compare_peers(p1, p2);
1342         if (rc)
1343                 return rc;
1344         */
1345
1346         if (r1->lr_seq - r2->lr_seq <= 0)
1347                 return 1;
1348
1349         return -1;
1350 }
1351
1352 /* TODO: lnet_find_route_locked() needs to be reimplemented */
1353 static struct lnet_route *
1354 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1355                        lnet_nid_t rtr_nid, struct lnet_route **prev_route)
1356 {
1357         struct lnet_remotenet *rnet;
1358         struct lnet_route *route;
1359         struct lnet_route *best_route;
1360         struct lnet_route *last_route;
1361         struct lnet_peer *lp_best;
1362         struct lnet_peer *lp;
1363         int rc;
1364
1365         /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1366          * rtr_nid nid, otherwise find the best gateway I can use */
1367
1368         rnet = lnet_find_rnet_locked(remote_net);
1369         if (rnet == NULL)
1370                 return NULL;
1371
1372         lp_best = NULL;
1373         best_route = last_route = NULL;
1374         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1375                 lp = route->lr_gateway;
1376
1377                 if (!lnet_is_route_alive(route))
1378                         continue;
1379
1380                 if (lp_best == NULL) {
1381                         best_route = last_route = route;
1382                         lp_best = lp;
1383                         continue;
1384                 }
1385
1386                 /* no protection on below fields, but it's harmless */
1387                 if (last_route->lr_seq - route->lr_seq < 0)
1388                         last_route = route;
1389
1390                 rc = lnet_compare_routes(route, best_route);
1391                 if (rc < 0)
1392                         continue;
1393
1394                 best_route = route;
1395                 lp_best = lp;
1396         }
1397
1398         *prev_route = last_route;
1399
1400         return best_route;
1401 }
1402
1403 static struct lnet_ni *
1404 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1405                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1406                  int md_cpt)
1407 {
1408         struct lnet_ni *ni = NULL;
1409         unsigned int shortest_distance;
1410         int best_credits;
1411         int best_healthv;
1412
1413         /*
1414          * If there is no peer_ni that we can send to on this network,
1415          * then there is no point in looking for a new best_ni here.
1416         */
1417         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1418                 return best_ni;
1419
1420         if (best_ni == NULL) {
1421                 shortest_distance = UINT_MAX;
1422                 best_credits = INT_MIN;
1423                 best_healthv = 0;
1424         } else {
1425                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1426                                                      best_ni->ni_dev_cpt);
1427                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1428                 best_healthv = atomic_read(&best_ni->ni_healthv);
1429         }
1430
1431         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1432                 unsigned int distance;
1433                 int ni_credits;
1434                 int ni_healthv;
1435                 int ni_fatal;
1436
1437                 ni_credits = atomic_read(&ni->ni_tx_credits);
1438                 ni_healthv = atomic_read(&ni->ni_healthv);
1439                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1440
1441                 /*
1442                  * calculate the distance from the CPT on which
1443                  * the message memory is allocated to the CPT of
1444                  * the NI's physical device
1445                  */
1446                 distance = cfs_cpt_distance(lnet_cpt_table(),
1447                                             md_cpt,
1448                                             ni->ni_dev_cpt);
1449
1450                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1451                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1452                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1453                         : "not seleced", best_credits, shortest_distance,
1454                         (best_ni) ? best_ni->ni_seq : 0);
1455
1456                 /*
1457                  * All distances smaller than the NUMA range
1458                  * are treated equally.
1459                  */
1460                 if (distance < lnet_numa_range)
1461                         distance = lnet_numa_range;
1462
1463                 /*
1464                  * Select on health, shorter distance, available
1465                  * credits, then round-robin.
1466                  */
1467                 if (ni_fatal) {
1468                         continue;
1469                 } else if (ni_healthv < best_healthv) {
1470                         continue;
1471                 } else if (ni_healthv > best_healthv) {
1472                         best_healthv = ni_healthv;
1473                         /*
1474                          * If we're going to prefer this ni because it's
1475                          * the healthiest, then we should set the
1476                          * shortest_distance in the algorithm in case
1477                          * there are multiple NIs with the same health but
1478                          * different distances.
1479                          */
1480                         if (distance < shortest_distance)
1481                                 shortest_distance = distance;
1482                 } else if (distance > shortest_distance) {
1483                         continue;
1484                 } else if (distance < shortest_distance) {
1485                         shortest_distance = distance;
1486                 } else if (ni_credits < best_credits) {
1487                         continue;
1488                 } else if (ni_credits == best_credits) {
1489                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1490                                 continue;
1491                 }
1492                 best_ni = ni;
1493                 best_credits = ni_credits;
1494         }
1495
1496         CDEBUG(D_NET, "selected best_ni %s\n",
1497                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1498
1499         return best_ni;
1500 }
1501
1502 /*
1503  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1504  * because such traffic is required to perform discovery. We therefore
1505  * exclude all GET and PUT on that portal. We also exclude all ACK and
1506  * REPLY traffic, but that is because the portal is not tracked in the
1507  * message structure for these message types. We could restrict this
1508  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1509  */
1510 static bool
1511 lnet_msg_discovery(struct lnet_msg *msg)
1512 {
1513         if (msg->msg_type == LNET_MSG_PUT) {
1514                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1515                         return true;
1516         } else if (msg->msg_type == LNET_MSG_GET) {
1517                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1518                         return true;
1519         }
1520         return false;
1521 }
1522
1523 #define SRC_SPEC        0x0001
1524 #define SRC_ANY         0x0002
1525 #define LOCAL_DST       0x0004
1526 #define REMOTE_DST      0x0008
1527 #define MR_DST          0x0010
1528 #define NMR_DST         0x0020
1529 #define SND_RESP        0x0040
1530
1531 /* The following to defines are used for return codes */
1532 #define REPEAT_SEND     0x1000
1533 #define PASS_THROUGH    0x2000
1534
1535 /* The different cases lnet_select pathway needs to handle */
1536 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1537 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1538 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1539 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1540 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1541 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1542 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1543 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1544
1545 static int
1546 lnet_handle_lo_send(struct lnet_send_data *sd)
1547 {
1548         struct lnet_msg *msg = sd->sd_msg;
1549         int cpt = sd->sd_cpt;
1550
1551         /* No send credit hassles with LOLND */
1552         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1553         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1554         if (!msg->msg_routing)
1555                 msg->msg_hdr.src_nid =
1556                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1557         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1558         lnet_msg_commit(msg, cpt);
1559         msg->msg_txni = the_lnet.ln_loni;
1560
1561         return LNET_CREDIT_OK;
1562 }
1563
1564 static int
1565 lnet_handle_send(struct lnet_send_data *sd)
1566 {
1567         struct lnet_ni *best_ni = sd->sd_best_ni;
1568         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1569         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1570         struct lnet_msg *msg = sd->sd_msg;
1571         int cpt2;
1572         __u32 send_case = sd->sd_send_case;
1573         int rc;
1574         __u32 routing = send_case & REMOTE_DST;
1575          struct lnet_rsp_tracker *rspt;
1576
1577         /*
1578          * Increment sequence number of the selected peer so that we
1579          * pick the next one in Round Robin.
1580          */
1581         best_lpni->lpni_seq++;
1582
1583         /*
1584          * grab a reference on the peer_ni so it sticks around even if
1585          * we need to drop and relock the lnet_net_lock below.
1586          */
1587         lnet_peer_ni_addref_locked(best_lpni);
1588
1589         /*
1590          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1591          * message. This ensures that we get a CPT that is correct for
1592          * the NI when the NI has been restricted to a subset of all CPTs.
1593          * If the selected CPT differs from the one currently locked, we
1594          * must unlock and relock the lnet_net_lock(), and then check whether
1595          * the configuration has changed. We don't have a hold on the best_ni
1596          * yet, and it may have vanished.
1597          */
1598         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1599         if (sd->sd_cpt != cpt2) {
1600                 __u32 seq = lnet_get_dlc_seq_locked();
1601                 lnet_net_unlock(sd->sd_cpt);
1602                 sd->sd_cpt = cpt2;
1603                 lnet_net_lock(sd->sd_cpt);
1604                 if (seq != lnet_get_dlc_seq_locked()) {
1605                         lnet_peer_ni_decref_locked(best_lpni);
1606                         return REPEAT_SEND;
1607                 }
1608         }
1609
1610         /*
1611          * store the best_lpni in the message right away to avoid having
1612          * to do the same operation under different conditions
1613          */
1614         msg->msg_txpeer = best_lpni;
1615         msg->msg_txni = best_ni;
1616
1617         /*
1618          * grab a reference for the best_ni since now it's in use in this
1619          * send. The reference will be dropped in lnet_finalize()
1620          */
1621         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1622
1623         /*
1624          * Always set the target.nid to the best peer picked. Either the
1625          * NID will be one of the peer NIDs selected, or the same NID as
1626          * what was originally set in the target or it will be the NID of
1627          * a router if this message should be routed
1628          */
1629         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1630
1631         /*
1632          * lnet_msg_commit assigns the correct cpt to the message, which
1633          * is used to decrement the correct refcount on the ni when it's
1634          * time to return the credits
1635          */
1636         lnet_msg_commit(msg, sd->sd_cpt);
1637
1638         /*
1639          * If we are routing the message then we keep the src_nid that was
1640          * set by the originator. If we are not routing then we are the
1641          * originator and set it here.
1642          */
1643         if (!msg->msg_routing)
1644                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1645
1646         if (routing) {
1647                 msg->msg_target_is_router = 1;
1648                 msg->msg_target.pid = LNET_PID_LUSTRE;
1649                 /*
1650                  * since we're routing we want to ensure that the
1651                  * msg_hdr.dest_nid is set to the final destination. When
1652                  * the router receives this message it knows how to route
1653                  * it.
1654                  *
1655                  * final_dst_lpni is set at the beginning of the
1656                  * lnet_select_pathway() function and is never changed.
1657                  * It's safe to use it here.
1658                  */
1659                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1660         } else {
1661                 /*
1662                  * if we're not routing set the dest_nid to the best peer
1663                  * ni NID that we picked earlier in the algorithm.
1664                  */
1665                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1666         }
1667
1668         /*
1669          * if we have response tracker block update it with the next hop
1670          * nid
1671          */
1672         if (msg->msg_md) {
1673                 rspt = msg->msg_md->md_rspt_ptr;
1674                 if (rspt) {
1675                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1676                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1677                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1678                 }
1679         }
1680
1681         rc = lnet_post_send_locked(msg, 0);
1682
1683         if (!rc)
1684                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1685                        libcfs_nid2str(msg->msg_hdr.src_nid),
1686                        libcfs_nid2str(msg->msg_txni->ni_nid),
1687                        libcfs_nid2str(sd->sd_src_nid),
1688                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1689                        libcfs_nid2str(sd->sd_dst_nid),
1690                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1691                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1692
1693         return rc;
1694 }
1695
1696 static struct lnet_peer_ni *
1697 lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
1698                     struct lnet_peer_net *peer_net)
1699 {
1700         /*
1701          * Look at the peer NIs for the destination peer that connect
1702          * to the chosen net. If a peer_ni is preferred when using the
1703          * best_ni to communicate, we use that one. If there is no
1704          * preferred peer_ni, or there are multiple preferred peer_ni,
1705          * the available transmit credits are used. If the transmit
1706          * credits are equal, we round-robin over the peer_ni.
1707          */
1708         struct lnet_peer_ni *lpni = NULL;
1709         struct lnet_peer_ni *best_lpni = NULL;
1710         struct lnet_ni *best_ni = sd->sd_best_ni;
1711         lnet_nid_t dst_nid = sd->sd_dst_nid;
1712         int best_lpni_credits = INT_MIN;
1713         bool preferred = false;
1714         bool ni_is_pref;
1715         int best_lpni_healthv = 0;
1716         int lpni_healthv;
1717
1718         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1719                 /*
1720                  * if the best_ni we've chosen aleady has this lpni
1721                  * preferred, then let's use it
1722                  */
1723                 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1724                                                           best_ni->ni_nid);
1725
1726                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1727
1728                 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1729                        libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1730
1731                 if (best_lpni)
1732                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1733                                 libcfs_nid2str(lpni->lpni_nid),
1734                                 lpni->lpni_txcredits, best_lpni_credits,
1735                                 lpni->lpni_seq, best_lpni->lpni_seq);
1736
1737                 /* pick the healthiest peer ni */
1738                 if (lpni_healthv < best_lpni_healthv) {
1739                         continue;
1740                 } else if (lpni_healthv > best_lpni_healthv) {
1741                         best_lpni_healthv = lpni_healthv;
1742                 /* if this is a preferred peer use it */
1743                 } else if (!preferred && ni_is_pref) {
1744                         preferred = true;
1745                 } else if (preferred && !ni_is_pref) {
1746                         /*
1747                          * this is not the preferred peer so let's ignore
1748                          * it.
1749                          */
1750                         continue;
1751                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1752                         /*
1753                          * We already have a peer that has more credits
1754                          * available than this one. No need to consider
1755                          * this peer further.
1756                          */
1757                         continue;
1758                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1759                         /*
1760                          * The best peer found so far and the current peer
1761                          * have the same number of available credits let's
1762                          * make sure to select between them using Round
1763                          * Robin
1764                          */
1765                         if (best_lpni) {
1766                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1767                                         continue;
1768                         }
1769                 }
1770
1771                 best_lpni = lpni;
1772                 best_lpni_credits = lpni->lpni_txcredits;
1773         }
1774
1775         /* if we still can't find a peer ni then we can't reach it */
1776         if (!best_lpni) {
1777                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1778                         LNET_NIDNET(dst_nid);
1779                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1780                                 libcfs_net2str(net_id));
1781                 return NULL;
1782         }
1783
1784         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1785                libcfs_nid2str(best_lpni->lpni_nid));
1786
1787         return best_lpni;
1788 }
1789
1790 /*
1791  * Prerequisite: the best_ni should already be set in the sd
1792  */
1793 static inline struct lnet_peer_ni *
1794 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1795                            __u32 net_id)
1796 {
1797         struct lnet_peer_net *peer_net;
1798
1799         /*
1800          * The gateway is Multi-Rail capable so now we must select the
1801          * proper peer_ni
1802          */
1803         peer_net = lnet_peer_get_net_locked(peer, net_id);
1804
1805         if (!peer_net) {
1806                 CERROR("gateway peer %s has no NI on net %s\n",
1807                        libcfs_nid2str(peer->lp_primary_nid),
1808                        libcfs_net2str(net_id));
1809                 return NULL;
1810         }
1811
1812         return lnet_select_peer_ni(sd, peer, peer_net);
1813 }
1814
1815 static inline void
1816 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1817 {
1818         if (sd->sd_send_case & NMR_DST &&
1819             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1820             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1821             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1822                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1823                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1824                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1825                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1826                                                  sd->sd_best_ni->ni_nid);
1827         }
1828 }
1829
1830 /*
1831  * Source Specified
1832  * Local Destination
1833  * non-mr peer
1834  *
1835  * use the source and destination NIDs as the pathway
1836  */
1837 static int
1838 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1839 {
1840         /* the destination lpni is set before we get here. */
1841
1842         /* find local NI */
1843         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1844         if (!sd->sd_best_ni) {
1845                 CERROR("Can't send to %s: src %s is not a "
1846                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1847                                 libcfs_nid2str(sd->sd_src_nid));
1848                 return -EINVAL;
1849         }
1850
1851         /*
1852          * the preferred NID will only be set for NMR peers
1853          */
1854         lnet_set_non_mr_pref_nid(sd);
1855
1856         return lnet_handle_send(sd);
1857 }
1858
1859 /*
1860  * Source Specified
1861  * Local Destination
1862  * MR Peer
1863  *
1864  * Run the selection algorithm on the peer NIs unless we're sending
1865  * a response, in this case just send to the destination
1866  */
1867 static int
1868 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1869 {
1870         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1871         if (!sd->sd_best_ni) {
1872                 CERROR("Can't send to %s: src %s is not a "
1873                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1874                                 libcfs_nid2str(sd->sd_src_nid));
1875                 return -EINVAL;
1876         }
1877
1878         /*
1879          * only run the selection algorithm to pick the peer_ni if we're
1880          * sending a GET or a PUT. Responses are sent to the same
1881          * destination NID provided.
1882          */
1883         if (!(sd->sd_send_case & SND_RESP)) {
1884                 sd->sd_best_lpni =
1885                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1886                                              sd->sd_best_ni->ni_net->net_id);
1887         }
1888
1889         if (sd->sd_best_lpni &&
1890             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1891                 return lnet_handle_lo_send(sd);
1892         else if (sd->sd_best_lpni)
1893                 return lnet_handle_send(sd);
1894
1895         CERROR("can't send to %s. no NI on %s\n",
1896                libcfs_nid2str(sd->sd_dst_nid),
1897                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1898
1899         return -EHOSTUNREACH;
1900 }
1901
1902 struct lnet_ni *
1903 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1904                               struct lnet_peer *peer,
1905                               struct lnet_peer_net *peer_net,
1906                               int cpt,
1907                               bool incr_seq)
1908 {
1909         struct lnet_net *local_net;
1910         struct lnet_ni *best_ni;
1911
1912         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1913         if (!local_net)
1914                 return NULL;
1915
1916         /*
1917          * Iterate through the NIs in this local Net and select
1918          * the NI to send from. The selection is determined by
1919          * these 3 criterion in the following priority:
1920          *      1. NUMA
1921          *      2. NI available credits
1922          *      3. Round Robin
1923          */
1924         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1925                                    peer, peer_net, cpt);
1926
1927         if (incr_seq && best_ni)
1928                 best_ni->ni_seq++;
1929
1930         return best_ni;
1931 }
1932
1933 static int
1934 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1935                              struct lnet_msg *msg, lnet_nid_t rtr_nid,
1936                              int cpt)
1937 {
1938         struct lnet_peer *peer;
1939         lnet_nid_t primary_nid;
1940         int rc;
1941
1942         lnet_peer_ni_addref_locked(lpni);
1943
1944         peer = lpni->lpni_peer_net->lpn_peer;
1945
1946         if (lnet_peer_gw_discovery(peer)) {
1947                 lnet_peer_ni_decref_locked(lpni);
1948                 return 0;
1949         }
1950
1951         rc = lnet_discover_peer_locked(lpni, cpt, false);
1952         if (rc) {
1953                 lnet_peer_ni_decref_locked(lpni);
1954                 return rc;
1955         }
1956         /* The peer may have changed. */
1957         peer = lpni->lpni_peer_net->lpn_peer;
1958         /* queue message and return */
1959         msg->msg_rtr_nid_param = rtr_nid;
1960         msg->msg_sending = 0;
1961         msg->msg_txpeer = NULL;
1962         spin_lock(&peer->lp_lock);
1963         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1964         spin_unlock(&peer->lp_lock);
1965         lnet_peer_ni_decref_locked(lpni);
1966         primary_nid = peer->lp_primary_nid;
1967
1968         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1969                 msg, libcfs_nid2str(primary_nid));
1970
1971         return LNET_DC_WAIT;
1972 }
1973
1974 static int
1975 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1976                              lnet_nid_t dst_nid,
1977                              struct lnet_peer_ni **gw_lpni,
1978                              struct lnet_peer **gw_peer)
1979 {
1980         struct lnet_peer *gw;
1981         struct lnet_route *best_route;
1982         struct lnet_route *last_route;
1983         struct lnet_peer_ni *lpni = NULL;
1984         lnet_nid_t src_nid = sd->sd_src_nid;
1985
1986         best_route = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
1987                                             sd->sd_rtr_nid, &last_route);
1988         if (!best_route) {
1989                 CERROR("no route to %s from %s\n",
1990                        libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1991                 return -EHOSTUNREACH;
1992         }
1993
1994         gw = best_route->lr_gateway;
1995         *gw_peer = gw;
1996
1997         /*
1998          * Discover this gateway if it hasn't already been discovered.
1999          * This means we might delay the message until discovery has
2000          * completed
2001          */
2002 #if 0
2003         /* TODO: disable discovey for now */
2004         if (lnet_msg_discovery(sd->sd_msg) &&
2005             !lnet_peer_is_uptodate(*gw_peer)) {
2006                 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2007                 return lnet_initiate_peer_discovery(gw, sd->sd_msg,
2008                                                     sd->sd_rtr_nid, sd->sd_cpt);
2009         }
2010 #endif
2011
2012         if (!sd->sd_best_ni)
2013                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2014                                         lnet_peer_get_net_locked(gw,
2015                                                 best_route->lr_lnet),
2016                                         sd->sd_md_cpt,
2017                                         true);
2018
2019         if (!sd->sd_best_ni) {
2020                 CERROR("Internal Error. Expected local ni on %s "
2021                        "but non found :%s\n",
2022                        libcfs_net2str(best_route->lr_lnet),
2023                        libcfs_nid2str(sd->sd_src_nid));
2024                 return -EFAULT;
2025         }
2026
2027         /*
2028          * if gw is MR let's find its best peer_ni
2029          */
2030         if (lnet_peer_is_multi_rail(gw)) {
2031                 lpni = lnet_find_best_lpni_on_net(sd, gw,
2032                                 sd->sd_best_ni->ni_net->net_id);
2033                 /*
2034                  * We've already verified that the gw has an NI on that
2035                  * desired net, but we're not finding it. Something is
2036                  * wrong.
2037                  */
2038                 if (!lpni) {
2039                         CERROR("Internal Error. Route expected to %s from %s\n",
2040                                 libcfs_nid2str(dst_nid),
2041                                 libcfs_nid2str(src_nid));
2042                         return -EFAULT;
2043                 }
2044         } else {
2045                 struct lnet_peer_net *lpn;
2046                 lpn = lnet_peer_get_net_locked(gw, best_route->lr_lnet);
2047                 if (!lpn) {
2048                         CERROR("Internal Error. Route expected to %s from %s\n",
2049                                 libcfs_nid2str(dst_nid),
2050                                 libcfs_nid2str(src_nid));
2051                         return -EFAULT;
2052                 }
2053                 lpni = list_entry(lpn->lpn_peer_nis.next, struct lnet_peer_ni,
2054                                   lpni_peer_nis);
2055                 if (!lpni) {
2056                         CERROR("Internal Error. Route expected to %s from %s\n",
2057                                 libcfs_nid2str(dst_nid),
2058                                 libcfs_nid2str(src_nid));
2059                         return -EFAULT;
2060                 }
2061         }
2062
2063         *gw_lpni = lpni;
2064
2065         /*
2066          * increment the route sequence number since now we're sure we're
2067          * going to use it
2068          */
2069         LASSERT(best_route && last_route);
2070         best_route->lr_seq = last_route->lr_seq + 1;
2071
2072         return 0;
2073 }
2074
2075 /*
2076  * Handle two cases:
2077  *
2078  * Case 1:
2079  *  Source specified
2080  *  Remote destination
2081  *  Non-MR destination
2082  *
2083  * Case 2:
2084  *  Source specified
2085  *  Remote destination
2086  *  MR destination
2087  *
2088  * The handling of these two cases is similar. Even though the destination
2089  * can be MR or non-MR, we'll deal directly with the router.
2090  */
2091 static int
2092 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2093 {
2094         int rc;
2095         struct lnet_peer_ni *gw_lpni = NULL;
2096         struct lnet_peer *gw_peer = NULL;
2097
2098         /* find local NI */
2099         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2100         if (!sd->sd_best_ni) {
2101                 CERROR("Can't send to %s: src %s is not a "
2102                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2103                                 libcfs_nid2str(sd->sd_src_nid));
2104                 return -EINVAL;
2105         }
2106
2107         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2108                                      &gw_peer);
2109         if (rc)
2110                 return rc;
2111
2112         if (sd->sd_send_case & NMR_DST)
2113                 /*
2114                 * since the final destination is non-MR let's set its preferred
2115                 * NID before we send
2116                 */
2117                 lnet_set_non_mr_pref_nid(sd);
2118
2119         /*
2120          * We're going to send to the gw found so let's set its
2121          * info
2122          */
2123         sd->sd_peer = gw_peer;
2124         sd->sd_best_lpni = gw_lpni;
2125
2126         return lnet_handle_send(sd);
2127 }
2128
2129 struct lnet_ni *
2130 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2131 {
2132         struct lnet_peer_net *peer_net = NULL;
2133         struct lnet_ni *best_ni = NULL;
2134
2135         /*
2136          * The peer can have multiple interfaces, some of them can be on
2137          * the local network and others on a routed network. We should
2138          * prefer the local network. However if the local network is not
2139          * available then we need to try the routed network
2140          */
2141
2142         /* go through all the peer nets and find the best_ni */
2143         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2144                 /*
2145                  * The peer's list of nets can contain non-local nets. We
2146                  * want to only examine the local ones.
2147                  */
2148                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2149                         continue;
2150                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2151                                                    peer_net, md_cpt, false);
2152         }
2153
2154         if (best_ni)
2155                 /* increment sequence number so we can round robin */
2156                 best_ni->ni_seq++;
2157
2158         return best_ni;
2159 }
2160
2161 static struct lnet_ni *
2162 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2163 {
2164         struct lnet_ni *best_ni = NULL;
2165         struct lnet_peer_net *peer_net;
2166         struct lnet_peer *peer = sd->sd_peer;
2167         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2168         struct lnet_peer_ni *lpni;
2169         int cpt = sd->sd_cpt;
2170
2171         /*
2172          * We must use a consistent source address when sending to a
2173          * non-MR peer. However, a non-MR peer can have multiple NIDs
2174          * on multiple networks, and we may even need to talk to this
2175          * peer on multiple networks -- certain types of
2176          * load-balancing configuration do this.
2177          *
2178          * So we need to pick the NI the peer prefers for this
2179          * particular network.
2180          */
2181
2182         /* Get the target peer_ni */
2183         peer_net = lnet_peer_get_net_locked(peer,
2184                         LNET_NIDNET(best_lpni->lpni_nid));
2185         LASSERT(peer_net != NULL);
2186         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2187                                 lpni_peer_nis) {
2188                 if (lpni->lpni_pref_nnids == 0)
2189                         continue;
2190                 LASSERT(lpni->lpni_pref_nnids == 1);
2191                 best_ni = lnet_nid2ni_locked(
2192                                 lpni->lpni_pref.nid, cpt);
2193                 break;
2194         }
2195
2196         return best_ni;
2197 }
2198
2199 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2200 static int
2201 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2202 {
2203         struct lnet_ni *best_ni = NULL;
2204         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2205
2206         /*
2207          * We must use a consistent source address when sending to a
2208          * non-MR peer. However, a non-MR peer can have multiple NIDs
2209          * on multiple networks, and we may even need to talk to this
2210          * peer on multiple networks -- certain types of
2211          * load-balancing configuration do this.
2212          *
2213          * So we need to pick the NI the peer prefers for this
2214          * particular network.
2215          */
2216
2217         best_ni = lnet_find_existing_preferred_best_ni(sd);
2218
2219         /* if best_ni is still not set just pick one */
2220         if (!best_ni) {
2221                 best_ni =
2222                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2223                                                 sd->sd_best_lpni->lpni_peer_net,
2224                                                 sd->sd_md_cpt, true);
2225                 /* If there is no best_ni we don't have a route */
2226                 if (!best_ni) {
2227                         CERROR("no path to %s from net %s\n",
2228                                 libcfs_nid2str(best_lpni->lpni_nid),
2229                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2230                         return -EHOSTUNREACH;
2231                 }
2232         }
2233
2234         sd->sd_best_ni = best_ni;
2235
2236         /* Set preferred NI if necessary. */
2237         lnet_set_non_mr_pref_nid(sd);
2238
2239         return 0;
2240 }
2241
2242
2243 /*
2244  * Source not specified
2245  * Local destination
2246  * Non-MR Peer
2247  *
2248  * always use the same source NID for NMR peers
2249  * If we've talked to that peer before then we already have a preferred
2250  * source NI associated with it. Otherwise, we select a preferred local NI
2251  * and store it in the peer
2252  */
2253 static int
2254 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2255 {
2256         int rc;
2257
2258         /* sd->sd_best_lpni is already set to the final destination */
2259
2260         /*
2261          * At this point we should've created the peer ni and peer. If we
2262          * can't find it, then something went wrong. Instead of assert
2263          * output a relevant message and fail the send
2264          */
2265         if (!sd->sd_best_lpni) {
2266                 CERROR("Internal fault. Unable to send msg %s to %s. "
2267                        "NID not known\n",
2268                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2269                        libcfs_nid2str(sd->sd_dst_nid));
2270                 return -EFAULT;
2271         }
2272
2273         rc = lnet_select_preferred_best_ni(sd);
2274         if (!rc)
2275                 rc = lnet_handle_send(sd);
2276
2277         return rc;
2278 }
2279
2280 static int
2281 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2282 {
2283         /*
2284          * NOTE we've already handled the remote peer case. So we only
2285          * need to worry about the local case here.
2286          *
2287          * if we're sending a response, ACK or reply, we need to send it
2288          * to the destination NID given to us. At this point we already
2289          * have the peer_ni we're suppose to send to, so just find the
2290          * best_ni on the peer net and use that. Since we're sending to an
2291          * MR peer then we can just run the selection algorithm on our
2292          * local NIs and pick the best one.
2293          */
2294         if (sd->sd_send_case & SND_RESP) {
2295                 sd->sd_best_ni =
2296                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2297                                                 sd->sd_best_lpni->lpni_peer_net,
2298                                                 sd->sd_md_cpt, true);
2299
2300                 if (!sd->sd_best_ni) {
2301                         /*
2302                          * We're not going to deal with not able to send
2303                          * a response to the provided final destination
2304                          */
2305                         CERROR("Can't send response to %s. "
2306                                "No local NI available\n",
2307                                 libcfs_nid2str(sd->sd_dst_nid));
2308                         return -EHOSTUNREACH;
2309                 }
2310
2311                 return lnet_handle_send(sd);
2312         }
2313
2314         /*
2315          * If we get here that means we're sending a fresh request, PUT or
2316          * GET, so we need to run our standard selection algorithm.
2317          * First find the best local interface that's on any of the peer's
2318          * networks.
2319          */
2320         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2321                                                         sd->sd_md_cpt);
2322         if (sd->sd_best_ni) {
2323                 sd->sd_best_lpni =
2324                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2325                                              sd->sd_best_ni->ni_net->net_id);
2326
2327                 /*
2328                  * if we're successful in selecting a peer_ni on the local
2329                  * network, then send to it. Otherwise fall through and
2330                  * try and see if we can reach it over another routed
2331                  * network
2332                  */
2333                 if (sd->sd_best_lpni &&
2334                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2335                         /*
2336                          * in case we initially started with a routed
2337                          * destination, let's reset to local
2338                          */
2339                         sd->sd_send_case &= ~REMOTE_DST;
2340                         sd->sd_send_case |= LOCAL_DST;
2341                         return lnet_handle_lo_send(sd);
2342                 } else if (sd->sd_best_lpni) {
2343                         /*
2344                          * in case we initially started with a routed
2345                          * destination, let's reset to local
2346                          */
2347                         sd->sd_send_case &= ~REMOTE_DST;
2348                         sd->sd_send_case |= LOCAL_DST;
2349                         return lnet_handle_send(sd);
2350                 }
2351
2352                 CERROR("Internal Error. Expected to have a best_lpni: "
2353                        "%s -> %s\n",
2354                        libcfs_nid2str(sd->sd_src_nid),
2355                        libcfs_nid2str(sd->sd_dst_nid));
2356
2357                 return -EFAULT;
2358         }
2359
2360         /*
2361          * Peer doesn't have a local network. Let's see if there is
2362          * a remote network we can reach it on.
2363          */
2364         return PASS_THROUGH;
2365 }
2366
2367 /*
2368  * Case 1:
2369  *      Source NID not specified
2370  *      Local destination
2371  *      MR peer
2372  *
2373  * Case 2:
2374  *      Source NID not speified
2375  *      Remote destination
2376  *      MR peer
2377  *
2378  * In both of these cases if we're sending a response, ACK or REPLY, then
2379  * we need to send to the destination NID provided.
2380  *
2381  * In the remote case let's deal with MR routers.
2382  *
2383  */
2384
2385 static int
2386 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2387 {
2388         int rc = 0;
2389         struct lnet_peer *gw_peer = NULL;
2390         struct lnet_peer_ni *gw_lpni = NULL;
2391
2392         /*
2393          * handle sending a response to a remote peer here so we don't
2394          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2395          */
2396         if (sd->sd_send_case & REMOTE_DST &&
2397             sd->sd_send_case & SND_RESP) {
2398                 struct lnet_peer_ni *gw;
2399                 struct lnet_peer *gw_peer;
2400
2401                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2402                                                   &gw_peer);
2403                 if (rc < 0) {
2404                         CERROR("Can't send response to %s. "
2405                                "No route available\n",
2406                                 libcfs_nid2str(sd->sd_dst_nid));
2407                         return -EHOSTUNREACH;
2408                 } else if (rc > 0) {
2409                         return rc;
2410                 }
2411
2412                 sd->sd_best_lpni = gw;
2413                 sd->sd_peer = gw_peer;
2414
2415                 return lnet_handle_send(sd);
2416         }
2417
2418         /*
2419          * Even though the NID for the peer might not be on a local network,
2420          * since the peer is MR there could be other interfaces on the
2421          * local network. In that case we'd still like to prefer the local
2422          * network over the routed network. If we're unable to do that
2423          * then we select the best router among the different routed networks,
2424          * and if the router is MR then we can deal with it as such.
2425          */
2426         rc = lnet_handle_any_mr_dsta(sd);
2427         if (rc != PASS_THROUGH)
2428                 return rc;
2429
2430         /*
2431          * TODO; One possible enhancement is to run the selection
2432          * algorithm on the peer. However for remote peers the credits are
2433          * not decremented, so we'll be basically going over the peer NIs
2434          * in round robin. An MR router will run the selection algorithm
2435          * on the next-hop interfaces.
2436          */
2437         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2438                                           &gw_peer);
2439         if (rc)
2440                 return rc;
2441
2442         sd->sd_send_case &= ~LOCAL_DST;
2443         sd->sd_send_case |= REMOTE_DST;
2444
2445         sd->sd_peer = gw_peer;
2446         sd->sd_best_lpni = gw_lpni;
2447
2448         return lnet_handle_send(sd);
2449 }
2450
2451 /*
2452  * Source not specified
2453  * Remote destination
2454  * Non-MR peer
2455  *
2456  * Must send to the specified peer NID using the same source NID that
2457  * we've used before. If it's the first time to talk to that peer then
2458  * find the source NI and assign it as preferred to that peer
2459  */
2460 static int
2461 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2462 {
2463         int rc;
2464         struct lnet_peer_ni *gw_lpni = NULL;
2465         struct lnet_peer *gw_peer = NULL;
2466
2467         /*
2468          * Let's set if we have a preferred NI to talk to this NMR peer
2469          */
2470         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2471
2472         /*
2473          * find the router and that'll find the best NI if we didn't find
2474          * it already.
2475          */
2476         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2477                                           &gw_peer);
2478         if (rc)
2479                 return rc;
2480
2481         /*
2482          * set the best_ni we've chosen as the preferred one for
2483          * this peer
2484          */
2485         lnet_set_non_mr_pref_nid(sd);
2486
2487         /* we'll be sending to the gw */
2488         sd->sd_best_lpni = gw_lpni;
2489         sd->sd_peer = gw_peer;
2490
2491         return lnet_handle_send(sd);
2492 }
2493
2494 static int
2495 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2496 {
2497         /*
2498          * turn off the SND_RESP bit.
2499          * It will be checked in the case handling
2500          */
2501         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2502
2503         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2504                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2505                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2506                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2507                 libcfs_nid2str(sd->sd_dst_nid),
2508                 (send_case & LOCAL_DST) ? "local" : "routed");
2509
2510         switch (send_case) {
2511         /*
2512          * For all cases where the source is specified, we should always
2513          * use the destination NID, whether it's an MR destination or not,
2514          * since we're continuing a series of related messages for the
2515          * same RPC
2516          */
2517         case SRC_SPEC_LOCAL_NMR_DST:
2518                 return lnet_handle_spec_local_nmr_dst(sd);
2519         case SRC_SPEC_LOCAL_MR_DST:
2520                 return lnet_handle_spec_local_mr_dst(sd);
2521         case SRC_SPEC_ROUTER_NMR_DST:
2522         case SRC_SPEC_ROUTER_MR_DST:
2523                 return lnet_handle_spec_router_dst(sd);
2524         case SRC_ANY_LOCAL_NMR_DST:
2525                 return lnet_handle_any_local_nmr_dst(sd);
2526         case SRC_ANY_LOCAL_MR_DST:
2527         case SRC_ANY_ROUTER_MR_DST:
2528                 return lnet_handle_any_mr_dst(sd);
2529         case SRC_ANY_ROUTER_NMR_DST:
2530                 return lnet_handle_any_router_nmr_dst(sd);
2531         default:
2532                 CERROR("Unknown send case\n");
2533                 return -1;
2534         }
2535 }
2536
2537 static int
2538 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2539                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2540 {
2541         struct lnet_peer_ni     *lpni;
2542         struct lnet_peer        *peer;
2543         struct lnet_send_data   send_data;
2544         int                     cpt, rc;
2545         int                     md_cpt;
2546         __u32                   send_case = 0;
2547
2548         memset(&send_data, 0, sizeof(send_data));
2549
2550         /*
2551          * get an initial CPT to use for locking. The idea here is not to
2552          * serialize the calls to select_pathway, so that as many
2553          * operations can run concurrently as possible. To do that we use
2554          * the CPT where this call is being executed. Later on when we
2555          * determine the CPT to use in lnet_message_commit, we switch the
2556          * lock and check if there was any configuration change.  If none,
2557          * then we proceed, if there is, then we restart the operation.
2558          */
2559         cpt = lnet_net_lock_current();
2560
2561         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2562         if (md_cpt == CFS_CPT_ANY)
2563                 md_cpt = cpt;
2564
2565 again:
2566
2567         /*
2568          * If we're being asked to send to the loopback interface, there
2569          * is no need to go through any selection. We can just shortcut
2570          * the entire process and send over lolnd
2571          */
2572         send_data.sd_msg = msg;
2573         send_data.sd_cpt = cpt;
2574         if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2575                 rc = lnet_handle_lo_send(&send_data);
2576                 lnet_net_unlock(cpt);
2577                 return rc;
2578         }
2579
2580         /*
2581          * find an existing peer_ni, or create one and mark it as having been
2582          * created due to network traffic. This call will create the
2583          * peer->peer_net->peer_ni tree.
2584          */
2585         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2586         if (IS_ERR(lpni)) {
2587                 lnet_net_unlock(cpt);
2588                 return PTR_ERR(lpni);
2589         }
2590
2591         /*
2592          * Cache the original src_nid. If we need to resend the message
2593          * then we'll need to know whether the src_nid was originally
2594          * specified for this message. If it was originally specified,
2595          * then we need to keep using the same src_nid since it's
2596          * continuing the same sequence of messages.
2597          */
2598         msg->msg_src_nid_param = src_nid;
2599
2600         /*
2601          * Now that we have a peer_ni, check if we want to discover
2602          * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2603          * trigger discovery.
2604          */
2605         peer = lpni->lpni_peer_net->lpn_peer;
2606         if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2607                 rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2608                 lnet_peer_ni_decref_locked(lpni);
2609                 lnet_net_unlock(cpt);
2610                 return rc;
2611         }
2612         lnet_peer_ni_decref_locked(lpni);
2613
2614         /*
2615          * Identify the different send cases
2616          */
2617         if (src_nid == LNET_NID_ANY)
2618                 send_case |= SRC_ANY;
2619         else
2620                 send_case |= SRC_SPEC;
2621
2622         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2623                 send_case |= LOCAL_DST;
2624         else
2625                 send_case |= REMOTE_DST;
2626
2627         /*
2628          * if this is a non-MR peer or if we're recovering a peer ni then
2629          * let's consider this an NMR case so we can hit the destination
2630          * NID.
2631          */
2632         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2633                 send_case |= NMR_DST;
2634         else
2635                 send_case |= MR_DST;
2636
2637         if (msg->msg_type == LNET_MSG_REPLY ||
2638             msg->msg_type == LNET_MSG_ACK)
2639                 send_case |= SND_RESP;
2640
2641         /* assign parameters to the send_data */
2642         send_data.sd_rtr_nid = rtr_nid;
2643         send_data.sd_src_nid = src_nid;
2644         send_data.sd_dst_nid = dst_nid;
2645         send_data.sd_best_lpni = lpni;
2646         /*
2647          * keep a pointer to the final destination in case we're going to
2648          * route, so we'll need to access it later
2649          */
2650         send_data.sd_final_dst_lpni = lpni;
2651         send_data.sd_peer = peer;
2652         send_data.sd_md_cpt = md_cpt;
2653         send_data.sd_send_case = send_case;
2654
2655         rc = lnet_handle_send_case_locked(&send_data);
2656
2657         /*
2658          * Update the local cpt since send_data.sd_cpt might've been
2659          * updated as a result of calling lnet_handle_send_case_locked().
2660          */
2661         cpt = send_data.sd_cpt;
2662
2663         if (rc == REPEAT_SEND)
2664                 goto again;
2665
2666         lnet_net_unlock(cpt);
2667
2668         return rc;
2669 }
2670
2671 int
2672 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2673 {
2674         lnet_nid_t              dst_nid = msg->msg_target.nid;
2675         int                     rc;
2676
2677         /*
2678          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2679          * but we might want to use pre-determined router for ACK/REPLY
2680          * in the future
2681          */
2682         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2683         LASSERT(msg->msg_txpeer == NULL);
2684         LASSERT(msg->msg_txni == NULL);
2685         LASSERT(!msg->msg_sending);
2686         LASSERT(!msg->msg_target_is_router);
2687         LASSERT(!msg->msg_receiving);
2688
2689         msg->msg_sending = 1;
2690
2691         LASSERT(!msg->msg_tx_committed);
2692
2693         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2694         if (rc < 0)
2695                 return rc;
2696
2697         if (rc == LNET_CREDIT_OK)
2698                 lnet_ni_send(msg->msg_txni, msg);
2699
2700         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2701         return 0;
2702 }
2703
2704 enum lnet_mt_event_type {
2705         MT_TYPE_LOCAL_NI = 0,
2706         MT_TYPE_PEER_NI
2707 };
2708
2709 struct lnet_mt_event_info {
2710         enum lnet_mt_event_type mt_type;
2711         lnet_nid_t mt_nid;
2712 };
2713
2714 /* called with res_lock held */
2715 void
2716 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2717 {
2718         struct lnet_rsp_tracker *rspt;
2719
2720         /*
2721          * msg has a refcount on the MD so the MD is not going away.
2722          * The rspt queue for the cpt is protected by
2723          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2724          */
2725         if (!md->md_rspt_ptr)
2726                 return;
2727
2728         rspt = md->md_rspt_ptr;
2729         md->md_rspt_ptr = NULL;
2730
2731         /* debug code */
2732         LASSERT(rspt->rspt_cpt == cpt);
2733
2734         /*
2735          * invalidate the handle to indicate that a response has been
2736          * received, which will then lead the monitor thread to clean up
2737          * the rspt block.
2738          */
2739         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2740 }
2741
2742 static void
2743 lnet_finalize_expired_responses(bool force)
2744 {
2745         struct lnet_libmd *md;
2746         struct list_head local_queue;
2747         struct lnet_rsp_tracker *rspt, *tmp;
2748         int i;
2749
2750         if (the_lnet.ln_mt_rstq == NULL)
2751                 return;
2752
2753         cfs_cpt_for_each(i, lnet_cpt_table()) {
2754                 INIT_LIST_HEAD(&local_queue);
2755
2756                 lnet_net_lock(i);
2757                 if (!the_lnet.ln_mt_rstq[i]) {
2758                         lnet_net_unlock(i);
2759                         continue;
2760                 }
2761                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2762                 lnet_net_unlock(i);
2763
2764                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2765                         /*
2766                          * The rspt mdh will be invalidated when a response
2767                          * is received or whenever we want to discard the
2768                          * block the monitor thread will walk the queue
2769                          * and clean up any rsts with an invalid mdh.
2770                          * The monitor thread will walk the queue until
2771                          * the first unexpired rspt block. This means that
2772                          * some rspt blocks which received their
2773                          * corresponding responses will linger in the
2774                          * queue until they are cleaned up eventually.
2775                          */
2776                         lnet_res_lock(i);
2777                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2778                                 lnet_res_unlock(i);
2779                                 list_del_init(&rspt->rspt_on_list);
2780                                 lnet_rspt_free(rspt, i);
2781                                 continue;
2782                         }
2783
2784                         if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2785                             force) {
2786                                 struct lnet_peer_ni *lpni;
2787                                 lnet_nid_t nid;
2788
2789                                 md = lnet_handle2md(&rspt->rspt_mdh);
2790                                 if (!md) {
2791                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2792                                         lnet_res_unlock(i);
2793                                         list_del_init(&rspt->rspt_on_list);
2794                                         lnet_rspt_free(rspt, i);
2795                                         continue;
2796                                 }
2797                                 LASSERT(md->md_rspt_ptr == rspt);
2798                                 md->md_rspt_ptr = NULL;
2799                                 lnet_res_unlock(i);
2800
2801                                 lnet_net_lock(i);
2802                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2803                                 lnet_net_unlock(i);
2804
2805                                 list_del_init(&rspt->rspt_on_list);
2806
2807                                 nid = rspt->rspt_next_hop_nid;
2808
2809                                 CNETERR("Response timed out: md = %p: nid = %s\n",
2810                                         md, libcfs_nid2str(nid));
2811                                 LNetMDUnlink(rspt->rspt_mdh);
2812                                 lnet_rspt_free(rspt, i);
2813
2814                                 /*
2815                                  * If there is a timeout on the response
2816                                  * from the next hop decrement its health
2817                                  * value so that we don't use it
2818                                  */
2819                                 lnet_net_lock(0);
2820                                 lpni = lnet_find_peer_ni_locked(nid);
2821                                 if (lpni) {
2822                                         lnet_handle_remote_failure_locked(lpni);
2823                                         lnet_peer_ni_decref_locked(lpni);
2824                                 }
2825                                 lnet_net_unlock(0);
2826                         } else {
2827                                 lnet_res_unlock(i);
2828                                 break;
2829                         }
2830                 }
2831
2832                 lnet_net_lock(i);
2833                 if (!list_empty(&local_queue))
2834                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2835                 lnet_net_unlock(i);
2836         }
2837 }
2838
2839 static void
2840 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2841 {
2842         struct lnet_msg *msg;
2843
2844         while (!list_empty(resendq)) {
2845                 struct lnet_peer_ni *lpni;
2846
2847                 msg = list_entry(resendq->next, struct lnet_msg,
2848                                  msg_list);
2849
2850                 list_del_init(&msg->msg_list);
2851
2852                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2853                 if (!lpni) {
2854                         lnet_net_unlock(cpt);
2855                         CERROR("Expected that a peer is already created for %s\n",
2856                                libcfs_nid2str(msg->msg_hdr.dest_nid));
2857                         msg->msg_no_resend = true;
2858                         lnet_finalize(msg, -EFAULT);
2859                         lnet_net_lock(cpt);
2860                 } else {
2861                         struct lnet_peer *peer;
2862                         int rc;
2863                         lnet_nid_t src_nid = LNET_NID_ANY;
2864
2865                         /*
2866                          * if this message is not being routed and the
2867                          * peer is non-MR then we must use the same
2868                          * src_nid that was used in the original send.
2869                          * Otherwise if we're routing the message (IE
2870                          * we're a router) then we can use any of our
2871                          * local interfaces. It doesn't matter to the
2872                          * final destination.
2873                          */
2874                         peer = lpni->lpni_peer_net->lpn_peer;
2875                         if (!msg->msg_routing &&
2876                             !lnet_peer_is_multi_rail(peer))
2877                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2878
2879                         /*
2880                          * If we originally specified a src NID, then we
2881                          * must attempt to reuse it in the resend as well.
2882                          */
2883                         if (msg->msg_src_nid_param != LNET_NID_ANY)
2884                                 src_nid = msg->msg_src_nid_param;
2885                         lnet_peer_ni_decref_locked(lpni);
2886
2887                         lnet_net_unlock(cpt);
2888                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
2889                                libcfs_nid2str(src_nid),
2890                                libcfs_id2str(msg->msg_target),
2891                                lnet_msgtyp2str(msg->msg_type),
2892                                msg->msg_recovery,
2893                                msg->msg_retry_count);
2894                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2895                         if (rc) {
2896                                 CERROR("Error sending %s to %s: %d\n",
2897                                        lnet_msgtyp2str(msg->msg_type),
2898                                        libcfs_id2str(msg->msg_target), rc);
2899                                 msg->msg_no_resend = true;
2900                                 lnet_finalize(msg, rc);
2901                         }
2902                         lnet_net_lock(cpt);
2903                         if (!rc)
2904                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
2905                 }
2906         }
2907 }
2908
2909 static void
2910 lnet_resend_pending_msgs(void)
2911 {
2912         int i;
2913
2914         cfs_cpt_for_each(i, lnet_cpt_table()) {
2915                 lnet_net_lock(i);
2916                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2917                 lnet_net_unlock(i);
2918         }
2919 }
2920
2921 /* called with cpt and ni_lock held */
2922 static void
2923 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
2924 {
2925         struct lnet_handle_md recovery_mdh;
2926
2927         LNetInvalidateMDHandle(&recovery_mdh);
2928
2929         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
2930             force) {
2931                 recovery_mdh = ni->ni_ping_mdh;
2932                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2933         }
2934         lnet_ni_unlock(ni);
2935         lnet_net_unlock(cpt);
2936         if (!LNetMDHandleIsInvalid(recovery_mdh))
2937                 LNetMDUnlink(recovery_mdh);
2938         lnet_net_lock(cpt);
2939         lnet_ni_lock(ni);
2940 }
2941
2942 static void
2943 lnet_recover_local_nis(void)
2944 {
2945         struct lnet_mt_event_info *ev_info;
2946         struct list_head processed_list;
2947         struct list_head local_queue;
2948         struct lnet_handle_md mdh;
2949         struct lnet_ni *tmp;
2950         struct lnet_ni *ni;
2951         lnet_nid_t nid;
2952         int healthv;
2953         int rc;
2954
2955         INIT_LIST_HEAD(&local_queue);
2956         INIT_LIST_HEAD(&processed_list);
2957
2958         /*
2959          * splice the recovery queue on a local queue. We will iterate
2960          * through the local queue and update it as needed. Once we're
2961          * done with the traversal, we'll splice the local queue back on
2962          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
2963          * will be traversed in the next iteration.
2964          */
2965         lnet_net_lock(0);
2966         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
2967                          &local_queue);
2968         lnet_net_unlock(0);
2969
2970         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
2971                 /*
2972                  * if an NI is being deleted or it is now healthy, there
2973                  * is no need to keep it around in the recovery queue.
2974                  * The monitor thread is the only thread responsible for
2975                  * removing the NI from the recovery queue.
2976                  * Multiple threads can be adding NIs to the recovery
2977                  * queue.
2978                  */
2979                 healthv = atomic_read(&ni->ni_healthv);
2980
2981                 lnet_net_lock(0);
2982                 lnet_ni_lock(ni);
2983                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
2984                     healthv == LNET_MAX_HEALTH_VALUE) {
2985                         list_del_init(&ni->ni_recovery);
2986                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
2987                         lnet_ni_unlock(ni);
2988                         lnet_ni_decref_locked(ni, 0);
2989                         lnet_net_unlock(0);
2990                         continue;
2991                 }
2992
2993                 /*
2994                  * if the local NI failed recovery we must unlink the md.
2995                  * But we want to keep the local_ni on the recovery queue
2996                  * so we can continue the attempts to recover it.
2997                  */
2998                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
2999                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3000                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3001                 }
3002
3003                 lnet_ni_unlock(ni);
3004                 lnet_net_unlock(0);
3005
3006
3007                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3008                        libcfs_nid2str(ni->ni_nid));
3009
3010                 lnet_ni_lock(ni);
3011                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3012                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3013                         lnet_ni_unlock(ni);
3014
3015                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3016                         if (!ev_info) {
3017                                 CERROR("out of memory. Can't recover %s\n",
3018                                        libcfs_nid2str(ni->ni_nid));
3019                                 lnet_ni_lock(ni);
3020                                 ni->ni_recovery_state &=
3021                                   ~LNET_NI_RECOVERY_PENDING;
3022                                 lnet_ni_unlock(ni);
3023                                 continue;
3024                         }
3025
3026                         mdh = ni->ni_ping_mdh;
3027                         /*
3028                          * Invalidate the ni mdh in case it's deleted.
3029                          * We'll unlink the mdh in this case below.
3030                          */
3031                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3032                         nid = ni->ni_nid;
3033
3034                         /*
3035                          * remove the NI from the local queue and drop the
3036                          * reference count to it while we're recovering
3037                          * it. The reason for that, is that the NI could
3038                          * be deleted, and the way the code is structured
3039                          * is if we don't drop the NI, then the deletion
3040                          * code will enter a loop waiting for the
3041                          * reference count to be removed while holding the
3042                          * ln_mutex_lock(). When we look up the peer to
3043                          * send to in lnet_select_pathway() we will try to
3044                          * lock the ln_mutex_lock() as well, leading to
3045                          * a deadlock. By dropping the refcount and
3046                          * removing it from the list, we allow for the NI
3047                          * to be removed, then we use the cached NID to
3048                          * look it up again. If it's gone, then we just
3049                          * continue examining the rest of the queue.
3050                          */
3051                         lnet_net_lock(0);
3052                         list_del_init(&ni->ni_recovery);
3053                         lnet_ni_decref_locked(ni, 0);
3054                         lnet_net_unlock(0);
3055
3056                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3057                         ev_info->mt_nid = nid;
3058                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3059                                             ev_info, the_lnet.ln_mt_eqh, true);
3060                         /* lookup the nid again */
3061                         lnet_net_lock(0);
3062                         ni = lnet_nid2ni_locked(nid, 0);
3063                         if (!ni) {
3064                                 /*
3065                                  * the NI has been deleted when we dropped
3066                                  * the ref count
3067                                  */
3068                                 lnet_net_unlock(0);
3069                                 LNetMDUnlink(mdh);
3070                                 continue;
3071                         }
3072                         /*
3073                          * Same note as in lnet_recover_peer_nis(). When
3074                          * we're sending the ping, the NI is free to be
3075                          * deleted or manipulated. By this point it
3076                          * could've been added back on the recovery queue,
3077                          * and a refcount taken on it.
3078                          * So we can't just add it blindly again or we'll
3079                          * corrupt the queue. We must check under lock if
3080                          * it's not on any list and if not then add it
3081                          * to the processed list, which will eventually be
3082                          * spliced back on to the recovery queue.
3083                          */
3084                         ni->ni_ping_mdh = mdh;
3085                         if (list_empty(&ni->ni_recovery)) {
3086                                 list_add_tail(&ni->ni_recovery, &processed_list);
3087                                 lnet_ni_addref_locked(ni, 0);
3088                         }
3089                         lnet_net_unlock(0);
3090
3091                         lnet_ni_lock(ni);
3092                         if (rc)
3093                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3094                 }
3095                 lnet_ni_unlock(ni);
3096         }
3097
3098         /*
3099          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3100          * reexamined in the next iteration.
3101          */
3102         list_splice_init(&processed_list, &local_queue);
3103         lnet_net_lock(0);
3104         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3105         lnet_net_unlock(0);
3106 }
3107
3108 static struct list_head **
3109 lnet_create_array_of_queues(void)
3110 {
3111         struct list_head **qs;
3112         struct list_head *q;
3113         int i;
3114
3115         qs = cfs_percpt_alloc(lnet_cpt_table(),
3116                               sizeof(struct list_head));
3117         if (!qs) {
3118                 CERROR("Failed to allocate queues\n");
3119                 return NULL;
3120         }
3121
3122         cfs_percpt_for_each(q, i, qs)
3123                 INIT_LIST_HEAD(q);
3124
3125         return qs;
3126 }
3127
3128 static int
3129 lnet_resendqs_create(void)
3130 {
3131         struct list_head **resendqs;
3132         resendqs = lnet_create_array_of_queues();
3133
3134         if (!resendqs)
3135                 return -ENOMEM;
3136
3137         lnet_net_lock(LNET_LOCK_EX);
3138         the_lnet.ln_mt_resendqs = resendqs;
3139         lnet_net_unlock(LNET_LOCK_EX);
3140
3141         return 0;
3142 }
3143
3144 static void
3145 lnet_clean_local_ni_recoveryq(void)
3146 {
3147         struct lnet_ni *ni;
3148
3149         /* This is only called when the monitor thread has stopped */
3150         lnet_net_lock(0);
3151
3152         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3153                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3154                                 struct lnet_ni, ni_recovery);
3155                 list_del_init(&ni->ni_recovery);
3156                 lnet_ni_lock(ni);
3157                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3158                 lnet_ni_unlock(ni);
3159                 lnet_ni_decref_locked(ni, 0);
3160         }
3161
3162         lnet_net_unlock(0);
3163 }
3164
3165 static void
3166 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3167                                      bool force)
3168 {
3169         struct lnet_handle_md recovery_mdh;
3170
3171         LNetInvalidateMDHandle(&recovery_mdh);
3172
3173         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3174                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3175                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3176         }
3177         spin_unlock(&lpni->lpni_lock);
3178         lnet_net_unlock(cpt);
3179         if (!LNetMDHandleIsInvalid(recovery_mdh))
3180                 LNetMDUnlink(recovery_mdh);
3181         lnet_net_lock(cpt);
3182         spin_lock(&lpni->lpni_lock);
3183 }
3184
3185 static void
3186 lnet_clean_peer_ni_recoveryq(void)
3187 {
3188         struct lnet_peer_ni *lpni, *tmp;
3189
3190         lnet_net_lock(LNET_LOCK_EX);
3191
3192         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3193                                  lpni_recovery) {
3194                 list_del_init(&lpni->lpni_recovery);
3195                 spin_lock(&lpni->lpni_lock);
3196                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3197                 spin_unlock(&lpni->lpni_lock);
3198                 lnet_peer_ni_decref_locked(lpni);
3199         }
3200
3201         lnet_net_unlock(LNET_LOCK_EX);
3202 }
3203
3204 static void
3205 lnet_clean_resendqs(void)
3206 {
3207         struct lnet_msg *msg, *tmp;
3208         struct list_head msgs;
3209         int i;
3210
3211         INIT_LIST_HEAD(&msgs);
3212
3213         cfs_cpt_for_each(i, lnet_cpt_table()) {
3214                 lnet_net_lock(i);
3215                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3216                 lnet_net_unlock(i);
3217                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3218                         list_del_init(&msg->msg_list);
3219                         msg->msg_no_resend = true;
3220                         lnet_finalize(msg, -ESHUTDOWN);
3221                 }
3222         }
3223
3224         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3225 }
3226
3227 static void
3228 lnet_recover_peer_nis(void)
3229 {
3230         struct lnet_mt_event_info *ev_info;
3231         struct list_head processed_list;
3232         struct list_head local_queue;
3233         struct lnet_handle_md mdh;
3234         struct lnet_peer_ni *lpni;
3235         struct lnet_peer_ni *tmp;
3236         lnet_nid_t nid;
3237         int healthv;
3238         int rc;
3239
3240         INIT_LIST_HEAD(&local_queue);
3241         INIT_LIST_HEAD(&processed_list);
3242
3243         /*
3244          * Always use cpt 0 for locking across all interactions with
3245          * ln_mt_peerNIRecovq
3246          */
3247         lnet_net_lock(0);
3248         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3249                          &local_queue);
3250         lnet_net_unlock(0);
3251
3252         list_for_each_entry_safe(lpni, tmp, &local_queue,
3253                                  lpni_recovery) {
3254                 /*
3255                  * The same protection strategy is used here as is in the
3256                  * local recovery case.
3257                  */
3258                 lnet_net_lock(0);
3259                 healthv = atomic_read(&lpni->lpni_healthv);
3260                 spin_lock(&lpni->lpni_lock);
3261                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3262                     healthv == LNET_MAX_HEALTH_VALUE) {
3263                         list_del_init(&lpni->lpni_recovery);
3264                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3265                         spin_unlock(&lpni->lpni_lock);
3266                         lnet_peer_ni_decref_locked(lpni);
3267                         lnet_net_unlock(0);
3268                         continue;
3269                 }
3270
3271                 /*
3272                  * If the peer NI has failed recovery we must unlink the
3273                  * md. But we want to keep the peer ni on the recovery
3274                  * queue so we can try to continue recovering it
3275                  */
3276                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3277                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3278                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3279                 }
3280
3281                 spin_unlock(&lpni->lpni_lock);
3282                 lnet_net_unlock(0);
3283
3284                 /*
3285                  * NOTE: we're racing with peer deletion from user space.
3286                  * It's possible that a peer is deleted after we check its
3287                  * state. In this case the recovery can create a new peer
3288                  */
3289                 spin_lock(&lpni->lpni_lock);
3290                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3291                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3292                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3293                         spin_unlock(&lpni->lpni_lock);
3294
3295                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3296                         if (!ev_info) {
3297                                 CERROR("out of memory. Can't recover %s\n",
3298                                        libcfs_nid2str(lpni->lpni_nid));
3299                                 spin_lock(&lpni->lpni_lock);
3300                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3301                                 spin_unlock(&lpni->lpni_lock);
3302                                 continue;
3303                         }
3304
3305                         /* look at the comments in lnet_recover_local_nis() */
3306                         mdh = lpni->lpni_recovery_ping_mdh;
3307                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3308                         nid = lpni->lpni_nid;
3309                         lnet_net_lock(0);
3310                         list_del_init(&lpni->lpni_recovery);
3311                         lnet_peer_ni_decref_locked(lpni);
3312                         lnet_net_unlock(0);
3313
3314                         ev_info->mt_type = MT_TYPE_PEER_NI;
3315                         ev_info->mt_nid = nid;
3316                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3317                                             ev_info, the_lnet.ln_mt_eqh, true);
3318                         lnet_net_lock(0);
3319                         /*
3320                          * lnet_find_peer_ni_locked() grabs a refcount for
3321                          * us. No need to take it explicitly.
3322                          */
3323                         lpni = lnet_find_peer_ni_locked(nid);
3324                         if (!lpni) {
3325                                 lnet_net_unlock(0);
3326                                 LNetMDUnlink(mdh);
3327                                 continue;
3328                         }
3329
3330                         lpni->lpni_recovery_ping_mdh = mdh;
3331                         /*
3332                          * While we're unlocked the lpni could've been
3333                          * readded on the recovery queue. In this case we
3334                          * don't need to add it to the local queue, since
3335                          * it's already on there and the thread that added
3336                          * it would've incremented the refcount on the
3337                          * peer, which means we need to decref the refcount
3338                          * that was implicitly grabbed by find_peer_ni_locked.
3339                          * Otherwise, if the lpni is still not on
3340                          * the recovery queue, then we'll add it to the
3341                          * processed list.
3342                          */
3343                         if (list_empty(&lpni->lpni_recovery))
3344                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3345                         else
3346                                 lnet_peer_ni_decref_locked(lpni);
3347                         lnet_net_unlock(0);
3348
3349                         spin_lock(&lpni->lpni_lock);
3350                         if (rc)
3351                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3352                 }
3353                 spin_unlock(&lpni->lpni_lock);
3354         }
3355
3356         list_splice_init(&processed_list, &local_queue);
3357         lnet_net_lock(0);
3358         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3359         lnet_net_unlock(0);
3360 }
3361
3362 static int
3363 lnet_monitor_thread(void *arg)
3364 {
3365         time64_t recovery_timeout = 0;
3366         time64_t rsp_timeout = 0;
3367         int interval;
3368         time64_t now;
3369
3370         /*
3371          * The monitor thread takes care of the following:
3372          *  1. Checks the aliveness of routers
3373          *  2. Checks if there are messages on the resend queue to resend
3374          *     them.
3375          *  3. Check if there are any NIs on the local recovery queue and
3376          *     pings them
3377          *  4. Checks if there are any NIs on the remote recovery queue
3378          *     and pings them.
3379          */
3380         cfs_block_allsigs();
3381
3382         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3383                 now = ktime_get_real_seconds();
3384
3385                 if (lnet_router_checker_active())
3386                         lnet_check_routers();
3387
3388                 lnet_resend_pending_msgs();
3389
3390                 if (now >= rsp_timeout) {
3391                         lnet_finalize_expired_responses(false);
3392                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3393                 }
3394
3395                 if (now >= recovery_timeout) {
3396                         lnet_recover_local_nis();
3397                         lnet_recover_peer_nis();
3398                         recovery_timeout = now + lnet_recovery_interval;
3399                 }
3400
3401                 /*
3402                  * TODO do we need to check if we should sleep without
3403                  * timeout?  Technically, an active system will always
3404                  * have messages in flight so this check will always
3405                  * evaluate to false. And on an idle system do we care
3406                  * if we wake up every 1 second? Although, we've seen
3407                  * cases where we get a complaint that an idle thread
3408                  * is waking up unnecessarily.
3409                  */
3410                 interval = min(lnet_recovery_interval,
3411                                lnet_transaction_timeout / 2);
3412                 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
3413                                                 false,
3414                                                 cfs_time_seconds(interval));
3415         }
3416
3417         /* Shutting down */
3418         lnet_net_lock(LNET_LOCK_EX);
3419         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3420         lnet_net_unlock(LNET_LOCK_EX);
3421
3422         /* signal that the monitor thread is exiting */
3423         up(&the_lnet.ln_mt_signal);
3424
3425         return 0;
3426 }
3427
3428 /*
3429  * lnet_send_ping
3430  * Sends a ping.
3431  * Returns == 0 if success
3432  * Returns > 0 if LNetMDBind or prior fails
3433  * Returns < 0 if LNetGet fails
3434  */
3435 int
3436 lnet_send_ping(lnet_nid_t dest_nid,
3437                struct lnet_handle_md *mdh, int nnis,
3438                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3439 {
3440         struct lnet_md md = { NULL };
3441         struct lnet_process_id id;
3442         struct lnet_ping_buffer *pbuf;
3443         int rc;
3444
3445         if (dest_nid == LNET_NID_ANY) {
3446                 rc = -EHOSTUNREACH;
3447                 goto fail_error;
3448         }
3449
3450         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3451         if (!pbuf) {
3452                 rc = ENOMEM;
3453                 goto fail_error;
3454         }
3455
3456         /* initialize md content */
3457         md.start     = &pbuf->pb_info;
3458         md.length    = LNET_PING_INFO_SIZE(nnis);
3459         md.threshold = 2; /* GET/REPLY */
3460         md.max_size  = 0;
3461         md.options   = LNET_MD_TRUNCATE;
3462         md.user_ptr  = user_data;
3463         md.eq_handle = eqh;
3464
3465         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3466         if (rc) {
3467                 lnet_ping_buffer_decref(pbuf);
3468                 CERROR("Can't bind MD: %d\n", rc);
3469                 rc = -rc; /* change the rc to positive */
3470                 goto fail_error;
3471         }
3472         id.pid = LNET_PID_LUSTRE;
3473         id.nid = dest_nid;
3474
3475         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3476                      LNET_RESERVED_PORTAL,
3477                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3478
3479         if (rc)
3480                 goto fail_unlink_md;
3481
3482         return 0;
3483
3484 fail_unlink_md:
3485         LNetMDUnlink(*mdh);
3486         LNetInvalidateMDHandle(mdh);
3487 fail_error:
3488         return rc;
3489 }
3490
3491 static void
3492 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3493                            int status, bool unlink_event)
3494 {
3495         lnet_nid_t nid = ev_info->mt_nid;
3496
3497         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3498                 struct lnet_ni *ni;
3499
3500                 lnet_net_lock(0);
3501                 ni = lnet_nid2ni_locked(nid, 0);
3502                 if (!ni) {
3503                         lnet_net_unlock(0);
3504                         return;
3505                 }
3506                 lnet_ni_lock(ni);
3507                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3508                 if (status)
3509                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3510                 lnet_ni_unlock(ni);
3511                 lnet_net_unlock(0);
3512
3513                 if (status != 0) {
3514                         CERROR("local NI (%s) recovery failed with %d\n",
3515                                libcfs_nid2str(nid), status);
3516                         return;
3517                 }
3518                 /*
3519                  * need to increment healthv for the ni here, because in
3520                  * the lnet_finalize() path we don't have access to this
3521                  * NI. And in order to get access to it, we'll need to
3522                  * carry forward too much information.
3523                  * In the peer case, it'll naturally be incremented
3524                  */
3525                 if (!unlink_event)
3526                         lnet_inc_healthv(&ni->ni_healthv);
3527         } else {
3528                 struct lnet_peer_ni *lpni;
3529                 int cpt;
3530
3531                 cpt = lnet_net_lock_current();
3532                 lpni = lnet_find_peer_ni_locked(nid);
3533                 if (!lpni) {
3534                         lnet_net_unlock(cpt);
3535                         return;
3536                 }
3537                 spin_lock(&lpni->lpni_lock);
3538                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3539                 if (status)
3540                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3541                 spin_unlock(&lpni->lpni_lock);
3542                 lnet_peer_ni_decref_locked(lpni);
3543                 lnet_net_unlock(cpt);
3544
3545                 if (status != 0)
3546                         CERROR("peer NI (%s) recovery failed with %d\n",
3547                                libcfs_nid2str(nid), status);
3548         }
3549 }
3550
3551 void
3552 lnet_mt_event_handler(struct lnet_event *event)
3553 {
3554         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3555         struct lnet_ping_buffer *pbuf;
3556
3557         /* TODO: remove assert */
3558         LASSERT(event->type == LNET_EVENT_REPLY ||
3559                 event->type == LNET_EVENT_SEND ||
3560                 event->type == LNET_EVENT_UNLINK);
3561
3562         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3563                event->status);
3564
3565         switch (event->type) {
3566         case LNET_EVENT_UNLINK:
3567                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3568                        libcfs_nid2str(ev_info->mt_nid));
3569         case LNET_EVENT_REPLY:
3570                 lnet_handle_recovery_reply(ev_info, event->status,
3571                                            event->type == LNET_EVENT_UNLINK);
3572                 break;
3573         case LNET_EVENT_SEND:
3574                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3575                                libcfs_nid2str(ev_info->mt_nid),
3576                                (event->status) ? "unsuccessfully" :
3577                                "successfully", event->status);
3578                 break;
3579         default:
3580                 CERROR("Unexpected event: %d\n", event->type);
3581                 break;
3582         }
3583         if (event->unlinked) {
3584                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3585                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3586                 lnet_ping_buffer_decref(pbuf);
3587         }
3588 }
3589
3590 static int
3591 lnet_rsp_tracker_create(void)
3592 {
3593         struct list_head **rstqs;
3594         rstqs = lnet_create_array_of_queues();
3595
3596         if (!rstqs)
3597                 return -ENOMEM;
3598
3599         the_lnet.ln_mt_rstq = rstqs;
3600
3601         return 0;
3602 }
3603
3604 static void
3605 lnet_rsp_tracker_clean(void)
3606 {
3607         lnet_finalize_expired_responses(true);
3608
3609         cfs_percpt_free(the_lnet.ln_mt_rstq);
3610         the_lnet.ln_mt_rstq = NULL;
3611 }
3612
3613 int lnet_monitor_thr_start(void)
3614 {
3615         int rc = 0;
3616         struct task_struct *task;
3617
3618         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3619                 return -EALREADY;
3620
3621         rc = lnet_resendqs_create();
3622         if (rc)
3623                 return rc;
3624
3625         rc = lnet_rsp_tracker_create();
3626         if (rc)
3627                 goto clean_queues;
3628
3629         sema_init(&the_lnet.ln_mt_signal, 0);
3630
3631         lnet_net_lock(LNET_LOCK_EX);
3632         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3633         lnet_net_unlock(LNET_LOCK_EX);
3634         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3635         if (IS_ERR(task)) {
3636                 rc = PTR_ERR(task);
3637                 CERROR("Can't start monitor thread: %d\n", rc);
3638                 goto clean_thread;
3639         }
3640
3641         return 0;
3642
3643 clean_thread:
3644         lnet_net_lock(LNET_LOCK_EX);
3645         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3646         lnet_net_unlock(LNET_LOCK_EX);
3647         /* block until event callback signals exit */
3648         down(&the_lnet.ln_mt_signal);
3649         /* clean up */
3650         lnet_net_lock(LNET_LOCK_EX);
3651         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3652         lnet_net_unlock(LNET_LOCK_EX);
3653         lnet_rsp_tracker_clean();
3654         lnet_clean_local_ni_recoveryq();
3655         lnet_clean_peer_ni_recoveryq();
3656         lnet_clean_resendqs();
3657         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3658         return rc;
3659 clean_queues:
3660         lnet_rsp_tracker_clean();
3661         lnet_clean_local_ni_recoveryq();
3662         lnet_clean_peer_ni_recoveryq();
3663         lnet_clean_resendqs();
3664         return rc;
3665 }
3666
3667 void lnet_monitor_thr_stop(void)
3668 {
3669         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3670                 return;
3671
3672         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3673         lnet_net_lock(LNET_LOCK_EX);
3674         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3675         lnet_net_unlock(LNET_LOCK_EX);
3676
3677         /* tell the monitor thread that we're shutting down */
3678         wake_up(&the_lnet.ln_mt_waitq);
3679
3680         /* block until monitor thread signals that it's done */
3681         down(&the_lnet.ln_mt_signal);
3682         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3683
3684         /* perform cleanup tasks */
3685         lnet_rsp_tracker_clean();
3686         lnet_clean_local_ni_recoveryq();
3687         lnet_clean_peer_ni_recoveryq();
3688         lnet_clean_resendqs();
3689
3690         return;
3691 }
3692
3693 void
3694 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3695                   __u32 msg_type)
3696 {
3697         lnet_net_lock(cpt);
3698         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3699         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3700         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3701         lnet_net_unlock(cpt);
3702
3703         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3704 }
3705
3706 static void
3707 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3708 {
3709         struct lnet_hdr *hdr = &msg->msg_hdr;
3710
3711         if (msg->msg_wanted != 0)
3712                 lnet_setpayloadbuffer(msg);
3713
3714         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3715
3716         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3717          * it back into the ACK during lnet_finalize() */
3718         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3719                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3720
3721         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3722                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3723 }
3724
3725 static int
3726 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3727 {
3728         struct lnet_hdr         *hdr = &msg->msg_hdr;
3729         struct lnet_match_info  info;
3730         int                     rc;
3731         bool                    ready_delay;
3732
3733         /* Convert put fields to host byte order */
3734         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3735         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3736         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3737
3738         /* Primary peer NID. */
3739         info.mi_id.nid  = msg->msg_initiator;
3740         info.mi_id.pid  = hdr->src_pid;
3741         info.mi_opc     = LNET_MD_OP_PUT;
3742         info.mi_portal  = hdr->msg.put.ptl_index;
3743         info.mi_rlength = hdr->payload_length;
3744         info.mi_roffset = hdr->msg.put.offset;
3745         info.mi_mbits   = hdr->msg.put.match_bits;
3746         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3747
3748         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3749         ready_delay = msg->msg_rx_ready_delay;
3750
3751  again:
3752         rc = lnet_ptl_match_md(&info, msg);
3753         switch (rc) {
3754         default:
3755                 LBUG();
3756
3757         case LNET_MATCHMD_OK:
3758                 lnet_recv_put(ni, msg);
3759                 return 0;
3760
3761         case LNET_MATCHMD_NONE:
3762                 if (ready_delay)
3763                         /* no eager_recv or has already called it, should
3764                          * have been attached on delayed list */
3765                         return 0;
3766
3767                 rc = lnet_ni_eager_recv(ni, msg);
3768                 if (rc == 0) {
3769                         ready_delay = true;
3770                         goto again;
3771                 }
3772                 /* fall through */
3773
3774         case LNET_MATCHMD_DROP:
3775                 CNETERR("Dropping PUT from %s portal %d match %llu"
3776                         " offset %d length %d: %d\n",
3777                         libcfs_id2str(info.mi_id), info.mi_portal,
3778                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3779
3780                 return -ENOENT; /* -ve: OK but no match */
3781         }
3782 }
3783
3784 static int
3785 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3786 {
3787         struct lnet_match_info info;
3788         struct lnet_hdr *hdr = &msg->msg_hdr;
3789         struct lnet_process_id source_id;
3790         struct lnet_handle_wire reply_wmd;
3791         int rc;
3792
3793         /* Convert get fields to host byte order */
3794         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3795         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3796         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3797         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3798
3799         source_id.nid = hdr->src_nid;
3800         source_id.pid = hdr->src_pid;
3801         /* Primary peer NID */
3802         info.mi_id.nid  = msg->msg_initiator;
3803         info.mi_id.pid  = hdr->src_pid;
3804         info.mi_opc     = LNET_MD_OP_GET;
3805         info.mi_portal  = hdr->msg.get.ptl_index;
3806         info.mi_rlength = hdr->msg.get.sink_length;
3807         info.mi_roffset = hdr->msg.get.src_offset;
3808         info.mi_mbits   = hdr->msg.get.match_bits;
3809         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3810
3811         rc = lnet_ptl_match_md(&info, msg);
3812         if (rc == LNET_MATCHMD_DROP) {
3813                 CNETERR("Dropping GET from %s portal %d match %llu"
3814                         " offset %d length %d\n",
3815                         libcfs_id2str(info.mi_id), info.mi_portal,
3816                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3817                 return -ENOENT; /* -ve: OK but no match */
3818         }
3819
3820         LASSERT(rc == LNET_MATCHMD_OK);
3821
3822         lnet_build_msg_event(msg, LNET_EVENT_GET);
3823
3824         reply_wmd = hdr->msg.get.return_wmd;
3825
3826         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3827                        msg->msg_offset, msg->msg_wanted);
3828
3829         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3830
3831         if (rdma_get) {
3832                 /* The LND completes the REPLY from her recv procedure */
3833                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3834                              msg->msg_offset, msg->msg_len, msg->msg_len);
3835                 return 0;
3836         }
3837
3838         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3839         msg->msg_receiving = 0;
3840
3841         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3842         if (rc < 0) {
3843                 /* didn't get as far as lnet_ni_send() */
3844                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3845                        libcfs_nid2str(ni->ni_nid),
3846                        libcfs_id2str(info.mi_id), rc);
3847
3848                 lnet_finalize(msg, rc);
3849         }
3850
3851         return 0;
3852 }
3853
3854 static int
3855 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3856 {
3857         void *private = msg->msg_private;
3858         struct lnet_hdr *hdr = &msg->msg_hdr;
3859         struct lnet_process_id src = {0};
3860         struct lnet_libmd *md;
3861         int rlength;
3862         int mlength;
3863         int cpt;
3864
3865         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3866         lnet_res_lock(cpt);
3867
3868         src.nid = hdr->src_nid;
3869         src.pid = hdr->src_pid;
3870
3871         /* NB handles only looked up by creator (no flips) */
3872         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3873         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3874                 CNETERR("%s: Dropping REPLY from %s for %s "
3875                         "MD %#llx.%#llx\n",
3876                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3877                         (md == NULL) ? "invalid" : "inactive",
3878                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
3879                         hdr->msg.reply.dst_wmd.wh_object_cookie);
3880                 if (md != NULL && md->md_me != NULL)
3881                         CERROR("REPLY MD also attached to portal %d\n",
3882                                md->md_me->me_portal);
3883
3884                 lnet_res_unlock(cpt);
3885                 return -ENOENT; /* -ve: OK but no match */
3886         }
3887
3888         LASSERT(md->md_offset == 0);
3889
3890         rlength = hdr->payload_length;
3891         mlength = MIN(rlength, (int)md->md_length);
3892
3893         if (mlength < rlength &&
3894             (md->md_options & LNET_MD_TRUNCATE) == 0) {
3895                 CNETERR("%s: Dropping REPLY from %s length %d "
3896                         "for MD %#llx would overflow (%d)\n",
3897                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3898                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
3899                         mlength);
3900                 lnet_res_unlock(cpt);
3901                 return -ENOENT; /* -ve: OK but no match */
3902         }
3903
3904         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
3905                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3906                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
3907
3908         lnet_msg_attach_md(msg, md, 0, mlength);
3909
3910         if (mlength != 0)
3911                 lnet_setpayloadbuffer(msg);
3912
3913         lnet_res_unlock(cpt);
3914
3915         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3916
3917         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
3918         return 0;
3919 }
3920
3921 static int
3922 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
3923 {
3924         struct lnet_hdr *hdr = &msg->msg_hdr;
3925         struct lnet_process_id src = {0};
3926         struct lnet_libmd *md;
3927         int cpt;
3928
3929         src.nid = hdr->src_nid;
3930         src.pid = hdr->src_pid;
3931
3932         /* Convert ack fields to host byte order */
3933         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
3934         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
3935
3936         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
3937         lnet_res_lock(cpt);
3938
3939         /* NB handles only looked up by creator (no flips) */
3940         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
3941         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3942                 /* Don't moan; this is expected */
3943                 CDEBUG(D_NET,
3944                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
3945                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3946                        (md == NULL) ? "invalid" : "inactive",
3947                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
3948                        hdr->msg.ack.dst_wmd.wh_object_cookie);
3949                 if (md != NULL && md->md_me != NULL)
3950                         CERROR("Source MD also attached to portal %d\n",
3951                                md->md_me->me_portal);
3952
3953                 lnet_res_unlock(cpt);
3954                 return -ENOENT;                  /* -ve! */
3955         }
3956
3957         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
3958                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3959                hdr->msg.ack.dst_wmd.wh_object_cookie);
3960
3961         lnet_msg_attach_md(msg, md, 0, 0);
3962
3963         lnet_res_unlock(cpt);
3964
3965         lnet_build_msg_event(msg, LNET_EVENT_ACK);
3966
3967         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
3968         return 0;
3969 }
3970
3971 /**
3972  * \retval LNET_CREDIT_OK       If \a msg is forwarded
3973  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
3974  * \retval -ve                  error code
3975  */
3976 int
3977 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
3978 {
3979         int     rc = 0;
3980
3981         if (!the_lnet.ln_routing)
3982                 return -ECANCELED;
3983
3984         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
3985             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
3986                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
3987                         msg->msg_rx_ready_delay = 1;
3988                 } else {
3989                         lnet_net_unlock(msg->msg_rx_cpt);
3990                         rc = lnet_ni_eager_recv(ni, msg);
3991                         lnet_net_lock(msg->msg_rx_cpt);
3992                 }
3993         }
3994
3995         if (rc == 0)
3996                 rc = lnet_post_routed_recv_locked(msg, 0);
3997         return rc;
3998 }
3999
4000 int
4001 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4002 {
4003         int     rc;
4004
4005         switch (msg->msg_type) {
4006         case LNET_MSG_ACK:
4007                 rc = lnet_parse_ack(ni, msg);
4008                 break;
4009         case LNET_MSG_PUT:
4010                 rc = lnet_parse_put(ni, msg);
4011                 break;
4012         case LNET_MSG_GET:
4013                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4014                 break;
4015         case LNET_MSG_REPLY:
4016                 rc = lnet_parse_reply(ni, msg);
4017                 break;
4018         default: /* prevent an unused label if !kernel */
4019                 LASSERT(0);
4020                 return -EPROTO;
4021         }
4022
4023         LASSERT(rc == 0 || rc == -ENOENT);
4024         return rc;
4025 }
4026
4027 char *
4028 lnet_msgtyp2str (int type)
4029 {
4030         switch (type) {
4031         case LNET_MSG_ACK:
4032                 return ("ACK");
4033         case LNET_MSG_PUT:
4034                 return ("PUT");
4035         case LNET_MSG_GET:
4036                 return ("GET");
4037         case LNET_MSG_REPLY:
4038                 return ("REPLY");
4039         case LNET_MSG_HELLO:
4040                 return ("HELLO");
4041         default:
4042                 return ("<UNKNOWN>");
4043         }
4044 }
4045
4046 void
4047 lnet_print_hdr(struct lnet_hdr *hdr)
4048 {
4049         struct lnet_process_id src = {
4050                 .nid = hdr->src_nid,
4051                 .pid = hdr->src_pid,
4052         };
4053         struct lnet_process_id dst = {
4054                 .nid = hdr->dest_nid,
4055                 .pid = hdr->dest_pid,
4056         };
4057         char *type_str = lnet_msgtyp2str(hdr->type);
4058
4059         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4060         CWARN("    From %s\n", libcfs_id2str(src));
4061         CWARN("    To   %s\n", libcfs_id2str(dst));
4062
4063         switch (hdr->type) {
4064         default:
4065                 break;
4066
4067         case LNET_MSG_PUT:
4068                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4069                       "match bits %llu\n",
4070                       hdr->msg.put.ptl_index,
4071                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4072                       hdr->msg.put.ack_wmd.wh_object_cookie,
4073                       hdr->msg.put.match_bits);
4074                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4075                       hdr->payload_length, hdr->msg.put.offset,
4076                       hdr->msg.put.hdr_data);
4077                 break;
4078
4079         case LNET_MSG_GET:
4080                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4081                       "match bits %llu\n", hdr->msg.get.ptl_index,
4082                       hdr->msg.get.return_wmd.wh_interface_cookie,
4083                       hdr->msg.get.return_wmd.wh_object_cookie,
4084                       hdr->msg.get.match_bits);
4085                 CWARN("    Length %d, src offset %d\n",
4086                       hdr->msg.get.sink_length,
4087                       hdr->msg.get.src_offset);
4088                 break;
4089
4090         case LNET_MSG_ACK:
4091                 CWARN("    dst md %#llx.%#llx, "
4092                       "manipulated length %d\n",
4093                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4094                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4095                       hdr->msg.ack.mlength);
4096                 break;
4097
4098         case LNET_MSG_REPLY:
4099                 CWARN("    dst md %#llx.%#llx, "
4100                       "length %d\n",
4101                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4102                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4103                       hdr->payload_length);
4104         }
4105
4106 }
4107
4108 int
4109 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4110            void *private, int rdma_req)
4111 {
4112         int             rc = 0;
4113         int             cpt;
4114         int             for_me;
4115         struct lnet_msg *msg;
4116         lnet_pid_t     dest_pid;
4117         lnet_nid_t     dest_nid;
4118         lnet_nid_t     src_nid;
4119         struct lnet_peer_ni *lpni;
4120         __u32          payload_length;
4121         __u32          type;
4122
4123         LASSERT (!in_interrupt ());
4124
4125         type = le32_to_cpu(hdr->type);
4126         src_nid = le64_to_cpu(hdr->src_nid);
4127         dest_nid = le64_to_cpu(hdr->dest_nid);
4128         dest_pid = le32_to_cpu(hdr->dest_pid);
4129         payload_length = le32_to_cpu(hdr->payload_length);
4130
4131         for_me = (ni->ni_nid == dest_nid);
4132         cpt = lnet_cpt_of_nid(from_nid, ni);
4133
4134         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4135                 libcfs_nid2str(dest_nid),
4136                 libcfs_nid2str(ni->ni_nid),
4137                 libcfs_nid2str(src_nid),
4138                 lnet_msgtyp2str(type),
4139                 (for_me) ? "for me" : "routed");
4140
4141         switch (type) {
4142         case LNET_MSG_ACK:
4143         case LNET_MSG_GET:
4144                 if (payload_length > 0) {
4145                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4146                                libcfs_nid2str(from_nid),
4147                                libcfs_nid2str(src_nid),
4148                                lnet_msgtyp2str(type), payload_length);
4149                         return -EPROTO;
4150                 }
4151                 break;
4152
4153         case LNET_MSG_PUT:
4154         case LNET_MSG_REPLY:
4155                 if (payload_length >
4156                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4157                         CERROR("%s, src %s: bad %s payload %d "
4158                                "(%d max expected)\n",
4159                                libcfs_nid2str(from_nid),
4160                                libcfs_nid2str(src_nid),
4161                                lnet_msgtyp2str(type),
4162                                payload_length,
4163                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4164                         return -EPROTO;
4165                 }
4166                 break;
4167
4168         default:
4169                 CERROR("%s, src %s: Bad message type 0x%x\n",
4170                        libcfs_nid2str(from_nid),
4171                        libcfs_nid2str(src_nid), type);
4172                 return -EPROTO;
4173         }
4174
4175         if (the_lnet.ln_routing &&
4176             ni->ni_last_alive != ktime_get_real_seconds()) {
4177                 /* NB: so far here is the only place to set NI status to "up */
4178                 lnet_ni_lock(ni);
4179                 ni->ni_last_alive = ktime_get_real_seconds();
4180                 if (ni->ni_status != NULL &&
4181                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
4182                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4183                 lnet_ni_unlock(ni);
4184         }
4185
4186         /* Regard a bad destination NID as a protocol error.  Senders should
4187          * know what they're doing; if they don't they're misconfigured, buggy
4188          * or malicious so we chop them off at the knees :) */
4189
4190         if (!for_me) {
4191                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4192                         /* should have gone direct */
4193                         CERROR("%s, src %s: Bad dest nid %s "
4194                                "(should have been sent direct)\n",
4195                                 libcfs_nid2str(from_nid),
4196                                 libcfs_nid2str(src_nid),
4197                                 libcfs_nid2str(dest_nid));
4198                         return -EPROTO;
4199                 }
4200
4201                 if (lnet_islocalnid(dest_nid)) {
4202                         /* dest is another local NI; sender should have used
4203                          * this node's NID on its own network */
4204                         CERROR("%s, src %s: Bad dest nid %s "
4205                                "(it's my nid but on a different network)\n",
4206                                 libcfs_nid2str(from_nid),
4207                                 libcfs_nid2str(src_nid),
4208                                 libcfs_nid2str(dest_nid));
4209                         return -EPROTO;
4210                 }
4211
4212                 if (rdma_req && type == LNET_MSG_GET) {
4213                         CERROR("%s, src %s: Bad optimized GET for %s "
4214                                "(final destination must be me)\n",
4215                                 libcfs_nid2str(from_nid),
4216                                 libcfs_nid2str(src_nid),
4217                                 libcfs_nid2str(dest_nid));
4218                         return -EPROTO;
4219                 }
4220
4221                 if (!the_lnet.ln_routing) {
4222                         CERROR("%s, src %s: Dropping message for %s "
4223                                "(routing not enabled)\n",
4224                                 libcfs_nid2str(from_nid),
4225                                 libcfs_nid2str(src_nid),
4226                                 libcfs_nid2str(dest_nid));
4227                         goto drop;
4228                 }
4229         }
4230
4231         /* Message looks OK; we're not going to return an error, so we MUST
4232          * call back lnd_recv() come what may... */
4233
4234         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4235             fail_peer(src_nid, 0)) {                    /* shall we now? */
4236                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4237                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4238                        lnet_msgtyp2str(type));
4239                 goto drop;
4240         }
4241
4242         if (!list_empty(&the_lnet.ln_drop_rules) &&
4243             lnet_drop_rule_match(hdr, NULL)) {
4244                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4245                               "silent message loss\n",
4246                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4247                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4248                 goto drop;
4249         }
4250
4251         if (lnet_drop_asym_route && for_me &&
4252             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4253                 struct lnet_net *net;
4254                 struct lnet_remotenet *rnet;
4255                 bool found = true;
4256
4257                 /* we are dealing with a routed message,
4258                  * so see if route to reach src_nid goes through from_nid
4259                  */
4260                 lnet_net_lock(cpt);
4261                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4262                 if (!net) {
4263                         lnet_net_unlock(cpt);
4264                         CERROR("net %s not found\n",
4265                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4266                         return -EPROTO;
4267                 }
4268
4269                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4270                 if (rnet) {
4271                         struct lnet_peer *gw = NULL;
4272                         struct lnet_peer_ni *lpni = NULL;
4273                         struct lnet_route *route;
4274
4275                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4276                                 found = false;
4277                                 gw = route->lr_gateway;
4278                                 if (route->lr_lnet != net->net_id)
4279                                         continue;
4280                                 /*
4281                                  * if the nid is one of the gateway's NIDs
4282                                  * then this is a valid gateway
4283                                  */
4284                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4285                                                 NULL, lpni)) != NULL) {
4286                                         if (lpni->lpni_nid == from_nid) {
4287                                                 found = true;
4288                                                 break;
4289                                         }
4290                                 }
4291                         }
4292                 }
4293                 lnet_net_unlock(cpt);
4294                 if (!found) {
4295                         /* we would not use from_nid to route a message to
4296                          * src_nid
4297                          * => asymmetric routing detected but forbidden
4298                          */
4299                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4300                                libcfs_nid2str(from_nid),
4301                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4302                         goto drop;
4303                 }
4304         }
4305
4306         msg = lnet_msg_alloc();
4307         if (msg == NULL) {
4308                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4309                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4310                        lnet_msgtyp2str(type));
4311                 goto drop;
4312         }
4313
4314         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4315          * pointers NULL etc */
4316
4317         msg->msg_type = type;
4318         msg->msg_private = private;
4319         msg->msg_receiving = 1;
4320         msg->msg_rdma_get = rdma_req;
4321         msg->msg_len = msg->msg_wanted = payload_length;
4322         msg->msg_offset = 0;
4323         msg->msg_hdr = *hdr;
4324         /* for building message event */
4325         msg->msg_from = from_nid;
4326         if (!for_me) {
4327                 msg->msg_target.pid     = dest_pid;
4328                 msg->msg_target.nid     = dest_nid;
4329                 msg->msg_routing        = 1;
4330
4331         } else {
4332                 /* convert common msg->hdr fields to host byteorder */
4333                 msg->msg_hdr.type       = type;
4334                 msg->msg_hdr.src_nid    = src_nid;
4335                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4336                 msg->msg_hdr.dest_nid   = dest_nid;
4337                 msg->msg_hdr.dest_pid   = dest_pid;
4338                 msg->msg_hdr.payload_length = payload_length;
4339         }
4340
4341         lnet_net_lock(cpt);
4342         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4343         if (IS_ERR(lpni)) {
4344                 lnet_net_unlock(cpt);
4345                 CERROR("%s, src %s: Dropping %s "
4346                        "(error %ld looking up sender)\n",
4347                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4348                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4349                 lnet_msg_free(msg);
4350                 if (rc == -ESHUTDOWN)
4351                         /* We are shutting down.  Don't do anything more */
4352                         return 0;
4353                 goto drop;
4354         }
4355         msg->msg_rxpeer = lpni;
4356         msg->msg_rxni = ni;
4357         lnet_ni_addref_locked(ni, cpt);
4358         /* Multi-Rail: Primary NID of source. */
4359         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4360
4361         /*
4362          * mark the status of this lpni as UP since we received a message
4363          * from it. The ping response reports back the ns_status which is
4364          * marked on the remote as up or down and we cache it here.
4365          */
4366         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4367
4368         lnet_msg_commit(msg, cpt);
4369
4370         /* message delay simulation */
4371         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4372                      lnet_delay_rule_match_locked(hdr, msg))) {
4373                 lnet_net_unlock(cpt);
4374                 return 0;
4375         }
4376
4377         if (!for_me) {
4378                 rc = lnet_parse_forward_locked(ni, msg);
4379                 lnet_net_unlock(cpt);
4380
4381                 if (rc < 0)
4382                         goto free_drop;
4383
4384                 if (rc == LNET_CREDIT_OK) {
4385                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4386                                      0, payload_length, payload_length);
4387                 }
4388                 return 0;
4389         }
4390
4391         lnet_net_unlock(cpt);
4392
4393         rc = lnet_parse_local(ni, msg);
4394         if (rc != 0)
4395                 goto free_drop;
4396         return 0;
4397
4398  free_drop:
4399         LASSERT(msg->msg_md == NULL);
4400         lnet_finalize(msg, rc);
4401
4402  drop:
4403         lnet_drop_message(ni, cpt, private, payload_length, type);
4404         return 0;
4405 }
4406 EXPORT_SYMBOL(lnet_parse);
4407
4408 void
4409 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4410 {
4411         while (!list_empty(head)) {
4412                 struct lnet_process_id id = {0};
4413                 struct lnet_msg *msg;
4414
4415                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4416                 list_del(&msg->msg_list);
4417
4418                 id.nid = msg->msg_hdr.src_nid;
4419                 id.pid = msg->msg_hdr.src_pid;
4420
4421                 LASSERT(msg->msg_md == NULL);
4422                 LASSERT(msg->msg_rx_delayed);
4423                 LASSERT(msg->msg_rxpeer != NULL);
4424                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4425
4426                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4427                       " offset %d length %d: %s\n",
4428                       libcfs_id2str(id),
4429                       msg->msg_hdr.msg.put.ptl_index,
4430                       msg->msg_hdr.msg.put.match_bits,
4431                       msg->msg_hdr.msg.put.offset,
4432                       msg->msg_hdr.payload_length, reason);
4433
4434                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4435                  * called lnet_drop_message(), so I just hang onto msg as well
4436                  * until that's done */
4437
4438                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4439                                   msg->msg_private, msg->msg_len,
4440                                   msg->msg_type);
4441
4442                 msg->msg_no_resend = true;
4443                 /*
4444                  * NB: message will not generate event because w/o attached MD,
4445                  * but we still should give error code so lnet_msg_decommit()
4446                  * can skip counters operations and other checks.
4447                  */
4448                 lnet_finalize(msg, -ENOENT);
4449         }
4450 }
4451
4452 void
4453 lnet_recv_delayed_msg_list(struct list_head *head)
4454 {
4455         while (!list_empty(head)) {
4456                 struct lnet_msg *msg;
4457                 struct lnet_process_id id;
4458
4459                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4460                 list_del(&msg->msg_list);
4461
4462                 /* md won't disappear under me, since each msg
4463                  * holds a ref on it */
4464
4465                 id.nid = msg->msg_hdr.src_nid;
4466                 id.pid = msg->msg_hdr.src_pid;
4467
4468                 LASSERT(msg->msg_rx_delayed);
4469                 LASSERT(msg->msg_md != NULL);
4470                 LASSERT(msg->msg_rxpeer != NULL);
4471                 LASSERT(msg->msg_rxni != NULL);
4472                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4473
4474                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4475                        "match %llu offset %d length %d.\n",
4476                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4477                         msg->msg_hdr.msg.put.match_bits,
4478                         msg->msg_hdr.msg.put.offset,
4479                         msg->msg_hdr.payload_length);
4480
4481                 lnet_recv_put(msg->msg_rxni, msg);
4482         }
4483 }
4484
4485 static void
4486 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4487                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4488 {
4489         s64 timeout_ns;
4490         bool new_entry = true;
4491         struct lnet_rsp_tracker *local_rspt;
4492
4493         /*
4494          * MD has a refcount taken by message so it's not going away.
4495          * The MD however can be looked up. We need to secure the access
4496          * to the md_rspt_ptr by taking the res_lock.
4497          * The rspt can be accessed without protection up to when it gets
4498          * added to the list.
4499          */
4500
4501         lnet_res_lock(cpt);
4502         local_rspt = md->md_rspt_ptr;
4503         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4504         if (local_rspt != NULL) {
4505                 /*
4506                  * we already have an rspt attached to the md, so we'll
4507                  * update the deadline on that one.
4508                  */
4509                 LIBCFS_FREE(rspt, sizeof(*rspt));
4510                 new_entry = false;
4511         } else {
4512                 /* new md */
4513                 rspt->rspt_mdh = mdh;
4514                 rspt->rspt_cpt = cpt;
4515                 /* store the rspt so we can access it when we get the REPLY */
4516                 md->md_rspt_ptr = rspt;
4517                 local_rspt = rspt;
4518         }
4519         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4520
4521         /*
4522          * add to the list of tracked responses. It's added to tail of the
4523          * list in order to expire all the older entries first.
4524          */
4525         lnet_net_lock(cpt);
4526         if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4527                 list_del_init(&local_rspt->rspt_on_list);
4528         list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4529         lnet_net_unlock(cpt);
4530         lnet_res_unlock(cpt);
4531 }
4532
4533 /**
4534  * Initiate an asynchronous PUT operation.
4535  *
4536  * There are several events associated with a PUT: completion of the send on
4537  * the initiator node (LNET_EVENT_SEND), and when the send completes
4538  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4539  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4540  * used at the target node to indicate the completion of incoming data
4541  * delivery.
4542  *
4543  * The local events will be logged in the EQ associated with the MD pointed to
4544  * by \a mdh handle. Using a MD without an associated EQ results in these
4545  * events being discarded. In this case, the caller must have another
4546  * mechanism (e.g., a higher level protocol) for determining when it is safe
4547  * to modify the memory region associated with the MD.
4548  *
4549  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4550  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4551  *
4552  * \param self Indicates the NID of a local interface through which to send
4553  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4554  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4555  * must be "free floating" (See LNetMDBind()).
4556  * \param ack Controls whether an acknowledgment is requested.
4557  * Acknowledgments are only sent when they are requested by the initiating
4558  * process and the target MD enables them.
4559  * \param target A process identifier for the target process.
4560  * \param portal The index in the \a target's portal table.
4561  * \param match_bits The match bits to use for MD selection at the target
4562  * process.
4563  * \param offset The offset into the target MD (only used when the target
4564  * MD has the LNET_MD_MANAGE_REMOTE option set).
4565  * \param hdr_data 64 bits of user data that can be included in the message
4566  * header. This data is written to an event queue entry at the target if an
4567  * EQ is present on the matching MD.
4568  *
4569  * \retval  0      Success, and only in this case events will be generated
4570  * and logged to EQ (if it exists).
4571  * \retval -EIO    Simulated failure.
4572  * \retval -ENOMEM Memory allocation failure.
4573  * \retval -ENOENT Invalid MD object.
4574  *
4575  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4576  */
4577 int
4578 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4579         struct lnet_process_id target, unsigned int portal,
4580         __u64 match_bits, unsigned int offset,
4581         __u64 hdr_data)
4582 {
4583         struct lnet_msg *msg;
4584         struct lnet_libmd *md;
4585         int cpt;
4586         int rc;
4587         struct lnet_rsp_tracker *rspt = NULL;
4588
4589         LASSERT(the_lnet.ln_refcount > 0);
4590
4591         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4592             fail_peer(target.nid, 1)) {                 /* shall we now? */
4593                 CERROR("Dropping PUT to %s: simulated failure\n",
4594                        libcfs_id2str(target));
4595                 return -EIO;
4596         }
4597
4598         msg = lnet_msg_alloc();
4599         if (msg == NULL) {
4600                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4601                        libcfs_id2str(target));
4602                 return -ENOMEM;
4603         }
4604         msg->msg_vmflush = !!memory_pressure_get();
4605
4606         cpt = lnet_cpt_of_cookie(mdh.cookie);
4607
4608         if (ack == LNET_ACK_REQ) {
4609                 rspt = lnet_rspt_alloc(cpt);
4610                 if (!rspt) {
4611                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4612                                 libcfs_id2str(target));
4613                         return -ENOMEM;
4614                 }
4615                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4616         }
4617
4618         lnet_res_lock(cpt);
4619
4620         md = lnet_handle2md(&mdh);
4621         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4622                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4623                        match_bits, portal, libcfs_id2str(target),
4624                        md == NULL ? -1 : md->md_threshold);
4625                 if (md != NULL && md->md_me != NULL)
4626                         CERROR("Source MD also attached to portal %d\n",
4627                                md->md_me->me_portal);
4628                 lnet_res_unlock(cpt);
4629
4630                 LIBCFS_FREE(rspt, sizeof(*rspt));
4631                 lnet_msg_free(msg);
4632                 return -ENOENT;
4633         }
4634
4635         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4636
4637         lnet_msg_attach_md(msg, md, 0, 0);
4638
4639         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4640
4641         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4642         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4643         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4644         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4645
4646         /* NB handles only looked up by creator (no flips) */
4647         if (ack == LNET_ACK_REQ) {
4648                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4649                         the_lnet.ln_interface_cookie;
4650                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4651                         md->md_lh.lh_cookie;
4652         } else {
4653                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4654                         LNET_WIRE_HANDLE_COOKIE_NONE;
4655                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4656                         LNET_WIRE_HANDLE_COOKIE_NONE;
4657         }
4658
4659         lnet_res_unlock(cpt);
4660
4661         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4662
4663         if (ack == LNET_ACK_REQ)
4664                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4665
4666         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4667                                  CFS_FAIL_ONCE))
4668                 rc = -EIO;
4669         else
4670                 rc = lnet_send(self, msg, LNET_NID_ANY);
4671
4672         if (rc != 0) {
4673                 CNETERR("Error sending PUT to %s: %d\n",
4674                         libcfs_id2str(target), rc);
4675                 msg->msg_no_resend = true;
4676                 lnet_finalize(msg, rc);
4677         }
4678
4679         /* completion will be signalled by an event */
4680         return 0;
4681 }
4682 EXPORT_SYMBOL(LNetPut);
4683
4684 /*
4685  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4686  * returns a msg for the LND to pass to lnet_finalize() when the sink
4687  * data has been received.
4688  *
4689  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4690  * lnet_finalize() is called on it, so the LND must call this first
4691  */
4692 struct lnet_msg *
4693 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4694 {
4695         struct lnet_msg *msg = lnet_msg_alloc();
4696         struct lnet_libmd *getmd = getmsg->msg_md;
4697         struct lnet_process_id peer_id = getmsg->msg_target;
4698         int cpt;
4699
4700         LASSERT(!getmsg->msg_target_is_router);
4701         LASSERT(!getmsg->msg_routing);
4702
4703         if (msg == NULL) {
4704                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4705                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4706                 goto drop;
4707         }
4708
4709         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4710         lnet_res_lock(cpt);
4711
4712         LASSERT(getmd->md_refcount > 0);
4713
4714         if (getmd->md_threshold == 0) {
4715                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4716                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4717                         getmd);
4718                 lnet_res_unlock(cpt);
4719                 goto drop;
4720         }
4721
4722         LASSERT(getmd->md_offset == 0);
4723
4724         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4725                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4726
4727         /* setup information for lnet_build_msg_event */
4728         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4729         msg->msg_from = peer_id.nid;
4730         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4731         msg->msg_hdr.src_nid = peer_id.nid;
4732         msg->msg_hdr.payload_length = getmd->md_length;
4733         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4734
4735         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4736         lnet_res_unlock(cpt);
4737
4738         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4739
4740         lnet_net_lock(cpt);
4741         lnet_msg_commit(msg, cpt);
4742         lnet_net_unlock(cpt);
4743
4744         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4745
4746         return msg;
4747
4748  drop:
4749         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4750
4751         lnet_net_lock(cpt);
4752         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4753         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4754         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4755                 getmd->md_length;
4756         lnet_net_unlock(cpt);
4757
4758         if (msg != NULL)
4759                 lnet_msg_free(msg);
4760
4761         return NULL;
4762 }
4763 EXPORT_SYMBOL(lnet_create_reply_msg);
4764
4765 void
4766 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4767                        unsigned int len)
4768 {
4769         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4770          * completed and I know it. */
4771         LASSERT(reply != NULL);
4772         LASSERT(reply->msg_type == LNET_MSG_GET);
4773         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4774
4775         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4776          * the end of my buffer, I might as well be dead. */
4777         LASSERT(len <= reply->msg_ev.mlength);
4778
4779         reply->msg_ev.mlength = len;
4780 }
4781 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4782
4783 /**
4784  * Initiate an asynchronous GET operation.
4785  *
4786  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4787  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4788  * the target node in the REPLY has been written to local MD.
4789  *
4790  * On the target node, an LNET_EVENT_GET is logged when the GET request
4791  * arrives and is accepted into a MD.
4792  *
4793  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4794  * \param mdh A handle for the MD that describes the memory into which the
4795  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4796  *
4797  * \retval  0      Success, and only in this case events will be generated
4798  * and logged to EQ (if it exists) of the MD.
4799  * \retval -EIO    Simulated failure.
4800  * \retval -ENOMEM Memory allocation failure.
4801  * \retval -ENOENT Invalid MD object.
4802  */
4803 int
4804 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4805         struct lnet_process_id target, unsigned int portal,
4806         __u64 match_bits, unsigned int offset, bool recovery)
4807 {
4808         struct lnet_msg *msg;
4809         struct lnet_libmd *md;
4810         struct lnet_rsp_tracker *rspt;
4811         int cpt;
4812         int rc;
4813
4814         LASSERT(the_lnet.ln_refcount > 0);
4815
4816         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4817             fail_peer(target.nid, 1))                   /* shall we now? */
4818         {
4819                 CERROR("Dropping GET to %s: simulated failure\n",
4820                        libcfs_id2str(target));
4821                 return -EIO;
4822         }
4823
4824         msg = lnet_msg_alloc();
4825         if (!msg) {
4826                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4827                        libcfs_id2str(target));
4828                 return -ENOMEM;
4829         }
4830
4831         cpt = lnet_cpt_of_cookie(mdh.cookie);
4832
4833         rspt = lnet_rspt_alloc(cpt);
4834         if (!rspt) {
4835                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4836                        libcfs_id2str(target));
4837                 return -ENOMEM;
4838         }
4839         INIT_LIST_HEAD(&rspt->rspt_on_list);
4840
4841         msg->msg_recovery = recovery;
4842
4843         lnet_res_lock(cpt);
4844
4845         md = lnet_handle2md(&mdh);
4846         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4847                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4848                        match_bits, portal, libcfs_id2str(target),
4849                        md == NULL ? -1 : md->md_threshold);
4850                 if (md != NULL && md->md_me != NULL)
4851                         CERROR("REPLY MD also attached to portal %d\n",
4852                                md->md_me->me_portal);
4853
4854                 lnet_res_unlock(cpt);
4855
4856                 lnet_msg_free(msg);
4857                 LIBCFS_FREE(rspt, sizeof(*rspt));
4858                 return -ENOENT;
4859         }
4860
4861         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4862
4863         lnet_msg_attach_md(msg, md, 0, 0);
4864
4865         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
4866
4867         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
4868         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
4869         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
4870         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
4871
4872         /* NB handles only looked up by creator (no flips) */
4873         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
4874                 the_lnet.ln_interface_cookie;
4875         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
4876                 md->md_lh.lh_cookie;
4877
4878         lnet_res_unlock(cpt);
4879
4880         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4881
4882         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4883
4884         rc = lnet_send(self, msg, LNET_NID_ANY);
4885         if (rc < 0) {
4886                 CNETERR("Error sending GET to %s: %d\n",
4887                         libcfs_id2str(target), rc);
4888                 msg->msg_no_resend = true;
4889                 lnet_finalize(msg, rc);
4890         }
4891
4892         /* completion will be signalled by an event */
4893         return 0;
4894 }
4895 EXPORT_SYMBOL(LNetGet);
4896
4897 /**
4898  * Calculate distance to node at \a dstnid.
4899  *
4900  * \param dstnid Target NID.
4901  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
4902  * is saved here.
4903  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
4904  * here.
4905  *
4906  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
4907  * local_nid_dist_zero is set, which is the default.
4908  * \retval positives Distance to target NID, i.e. number of hops plus one.
4909  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
4910  */
4911 int
4912 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
4913 {
4914         struct list_head        *e;
4915         struct lnet_ni *ni = NULL;
4916         struct lnet_remotenet *rnet;
4917         __u32                   dstnet = LNET_NIDNET(dstnid);
4918         int                     hops;
4919         int                     cpt;
4920         __u32                   order = 2;
4921         struct list_head        *rn_list;
4922
4923         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
4924          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
4925          * keep order 0 free for 0@lo and order 1 free for a local NID
4926          * match */
4927
4928         LASSERT(the_lnet.ln_refcount > 0);
4929
4930         cpt = lnet_net_lock_current();
4931
4932         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
4933                 if (ni->ni_nid == dstnid) {
4934                         if (srcnidp != NULL)
4935                                 *srcnidp = dstnid;
4936                         if (orderp != NULL) {
4937                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
4938                                         *orderp = 0;
4939                                 else
4940                                         *orderp = 1;
4941                         }
4942                         lnet_net_unlock(cpt);
4943
4944                         return local_nid_dist_zero ? 0 : 1;
4945                 }
4946
4947                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
4948                         /* Check if ni was originally created in
4949                          * current net namespace.
4950                          * If not, assign order above 0xffff0000,
4951                          * to make this ni not a priority. */
4952                         if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
4953                                 order += 0xffff0000;
4954
4955                         if (srcnidp != NULL)
4956                                 *srcnidp = ni->ni_nid;
4957                         if (orderp != NULL)
4958                                 *orderp = order;
4959                         lnet_net_unlock(cpt);
4960                         return 1;
4961                 }
4962
4963                 order++;
4964         }
4965
4966         rn_list = lnet_net2rnethash(dstnet);
4967         list_for_each(e, rn_list) {
4968                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
4969
4970                 if (rnet->lrn_net == dstnet) {
4971                         struct lnet_route *route;
4972                         struct lnet_route *shortest = NULL;
4973                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
4974                         __u32 route_hops;
4975
4976                         LASSERT(!list_empty(&rnet->lrn_routes));
4977
4978                         list_for_each_entry(route, &rnet->lrn_routes,
4979                                             lr_list) {
4980                                 route_hops = route->lr_hops;
4981                                 if (route_hops == LNET_UNDEFINED_HOPS)
4982                                         route_hops = 1;
4983                                 if (shortest == NULL ||
4984                                     route_hops < shortest_hops) {
4985                                         shortest = route;
4986                                         shortest_hops = route_hops;
4987                                 }
4988                         }
4989
4990                         LASSERT(shortest != NULL);
4991                         hops = shortest_hops;
4992                         if (srcnidp != NULL) {
4993                                 struct lnet_net *net;
4994                                 net = lnet_get_net_locked(shortest->lr_lnet);
4995                                 LASSERT(net);
4996                                 ni = lnet_get_next_ni_locked(net, NULL);
4997                                 *srcnidp = ni->ni_nid;
4998                         }
4999                         if (orderp != NULL)
5000                                 *orderp = order;
5001                         lnet_net_unlock(cpt);
5002                         return hops + 1;
5003                 }
5004                 order++;
5005         }
5006
5007         lnet_net_unlock(cpt);
5008         return -EHOSTUNREACH;
5009 }
5010 EXPORT_SYMBOL(LNetDist);