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_move(&tp->tp_list, &cull);
 198                 }
 199         }
 200
 201         lnet_net_unlock(0);
 202
 203         while (!list_empty(&cull)) {
 204                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 205
 206                 list_del(&tp->tp_list);
 207                 LIBCFS_FREE(tp, sizeof(*tp));
 208         }
 209         return 0;
 210 }
 211
 212 static int
 213 fail_peer (lnet_nid_t nid, int outgoing)
 214 {
 215         struct lnet_test_peer *tp;
 216         struct list_head *el;
 217         struct list_head *next;
 218         struct list_head  cull;
 219         int               fail = 0;
 220
 221         INIT_LIST_HEAD(&cull);
 222
 223         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 224         lnet_net_lock(0);
 225
 226         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 227                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 228
 229                 if (tp->tp_threshold == 0) {
 230                         /* zombie entry */
 231                         if (outgoing) {
 232                                 /* only cull zombies on outgoing tests,
 233                                  * since we may be at interrupt priority on
 234                                  * incoming messages. */
 235                                 list_move(&tp->tp_list, &cull);
 236                         }
 237                         continue;
 238                 }
 239
 240                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
 241                     nid == tp->tp_nid) {                /* fail this peer */
 242                         fail = 1;
 243
 244                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 245                                 tp->tp_threshold--;
 246                                 if (outgoing &&
 247                                     tp->tp_threshold == 0) {
 248                                         /* see above */
 249                                         list_move(&tp->tp_list, &cull);
 250                                 }
 251                         }
 252                         break;
 253                 }
 254         }
 255
 256         lnet_net_unlock(0);
 257
 258         while (!list_empty(&cull)) {
 259                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 260                 list_del(&tp->tp_list);
 261
 262                 LIBCFS_FREE(tp, sizeof(*tp));
 263         }
 264
 265         return fail;
 266 }
 267
 268 unsigned int
 269 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 270 {
 271         unsigned int nob = 0;
 272
 273         LASSERT(niov == 0 || iov != NULL);
 274         while (niov-- > 0)
 275                 nob += (iov++)->iov_len;
 276
 277         return (nob);
 278 }
 279 EXPORT_SYMBOL(lnet_iov_nob);
 280
 281 void
 282 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 283                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 284                   unsigned int nob)
 285 {
 286         /* NB diov, siov are READ-ONLY */
 287         unsigned int  this_nob;
 288
 289         if (nob == 0)
 290                 return;
 291
 292         /* skip complete frags before 'doffset' */
 293         LASSERT(ndiov > 0);
 294         while (doffset >= diov->iov_len) {
 295                 doffset -= diov->iov_len;
 296                 diov++;
 297                 ndiov--;
 298                 LASSERT(ndiov > 0);
 299         }
 300
 301         /* skip complete frags before 'soffset' */
 302         LASSERT(nsiov > 0);
 303         while (soffset >= siov->iov_len) {
 304                 soffset -= siov->iov_len;
 305                 siov++;
 306                 nsiov--;
 307                 LASSERT(nsiov > 0);
 308         }
 309
 310         do {
 311                 LASSERT(ndiov > 0);
 312                 LASSERT(nsiov > 0);
 313                 this_nob = MIN(diov->iov_len - doffset,
 314                                siov->iov_len - soffset);
 315                 this_nob = MIN(this_nob, nob);
 316
 317                 memcpy((char *)diov->iov_base + doffset,
 318                        (char *)siov->iov_base + soffset, this_nob);
 319                 nob -= this_nob;
 320
 321                 if (diov->iov_len > doffset + this_nob) {
 322                         doffset += this_nob;
 323                 } else {
 324                         diov++;
 325                         ndiov--;
 326                         doffset = 0;
 327                 }
 328
 329                 if (siov->iov_len > soffset + this_nob) {
 330                         soffset += this_nob;
 331                 } else {
 332                         siov++;
 333                         nsiov--;
 334                         soffset = 0;
 335                 }
 336         } while (nob > 0);
 337 }
 338 EXPORT_SYMBOL(lnet_copy_iov2iov);
 339
 340 int
 341 lnet_extract_iov(int dst_niov, struct kvec *dst,
 342                  int src_niov, struct kvec *src,
 343                  unsigned int offset, unsigned int len)
 344 {
 345         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 346          * for exactly 'len' bytes, and return the number of entries.
 347          * NB not destructive to 'src' */
 348         unsigned int    frag_len;
 349         unsigned int    niov;
 350
 351         if (len == 0)                           /* no data => */
 352                 return (0);                     /* no frags */
 353
 354         LASSERT(src_niov > 0);
 355         while (offset >= src->iov_len) {      /* skip initial frags */
 356                 offset -= src->iov_len;
 357                 src_niov--;
 358                 src++;
 359                 LASSERT(src_niov > 0);
 360         }
 361
 362         niov = 1;
 363         for (;;) {
 364                 LASSERT(src_niov > 0);
 365                 LASSERT((int)niov <= dst_niov);
 366
 367                 frag_len = src->iov_len - offset;
 368                 dst->iov_base = ((char *)src->iov_base) + offset;
 369
 370                 if (len <= frag_len) {
 371                         dst->iov_len = len;
 372                         return (niov);
 373                 }
 374
 375                 dst->iov_len = frag_len;
 376
 377                 len -= frag_len;
 378                 dst++;
 379                 src++;
 380                 niov++;
 381                 src_niov--;
 382                 offset = 0;
 383         }
 384 }
 385 EXPORT_SYMBOL(lnet_extract_iov);
 386
 387
 388 unsigned int
 389 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
 390 {
 391         unsigned int  nob = 0;
 392
 393         LASSERT(niov == 0 || kiov != NULL);
 394         while (niov-- > 0)
 395                 nob += (kiov++)->kiov_len;
 396
 397         return (nob);
 398 }
 399 EXPORT_SYMBOL(lnet_kiov_nob);
 400
 401 void
 402 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
 403                     unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
 404                     unsigned int nob)
 405 {
 406         /* NB diov, siov are READ-ONLY */
 407         unsigned int    this_nob;
 408         char           *daddr = NULL;
 409         char           *saddr = NULL;
 410
 411         if (nob == 0)
 412                 return;
 413
 414         LASSERT (!in_interrupt ());
 415
 416         LASSERT (ndiov > 0);
 417         while (doffset >= diov->kiov_len) {
 418                 doffset -= diov->kiov_len;
 419                 diov++;
 420                 ndiov--;
 421                 LASSERT(ndiov > 0);
 422         }
 423
 424         LASSERT(nsiov > 0);
 425         while (soffset >= siov->kiov_len) {
 426                 soffset -= siov->kiov_len;
 427                 siov++;
 428                 nsiov--;
 429                 LASSERT(nsiov > 0);
 430         }
 431
 432         do {
 433                 LASSERT(ndiov > 0);
 434                 LASSERT(nsiov > 0);
 435                 this_nob = MIN(diov->kiov_len - doffset,
 436                                siov->kiov_len - soffset);
 437                 this_nob = MIN(this_nob, nob);
 438
 439                 if (daddr == NULL)
 440                         daddr = ((char *)kmap(diov->kiov_page)) +
 441                                 diov->kiov_offset + doffset;
 442                 if (saddr == NULL)
 443                         saddr = ((char *)kmap(siov->kiov_page)) +
 444                                 siov->kiov_offset + soffset;
 445
 446                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 447                  * However in practice at least one of the kiovs will be mapped
 448                  * kernel pages and the map/unmap will be NOOPs */
 449
 450                 memcpy (daddr, saddr, this_nob);
 451                 nob -= this_nob;
 452
 453                 if (diov->kiov_len > doffset + this_nob) {
 454                         daddr += this_nob;
 455                         doffset += this_nob;
 456                 } else {
 457                         kunmap(diov->kiov_page);
 458                         daddr = NULL;
 459                         diov++;
 460                         ndiov--;
 461                         doffset = 0;
 462                 }
 463
 464                 if (siov->kiov_len > soffset + this_nob) {
 465                         saddr += this_nob;
 466                         soffset += this_nob;
 467                 } else {
 468                         kunmap(siov->kiov_page);
 469                         saddr = NULL;
 470                         siov++;
 471                         nsiov--;
 472                         soffset = 0;
 473                 }
 474         } while (nob > 0);
 475
 476         if (daddr != NULL)
 477                 kunmap(diov->kiov_page);
 478         if (saddr != NULL)
 479                 kunmap(siov->kiov_page);
 480 }
 481 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 482
 483 void
 484 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 485                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 486                     unsigned int nob)
 487 {
 488         /* NB iov, kiov are READ-ONLY */
 489         unsigned int    this_nob;
 490         char           *addr = NULL;
 491
 492         if (nob == 0)
 493                 return;
 494
 495         LASSERT (!in_interrupt ());
 496
 497         LASSERT (niov > 0);
 498         while (iovoffset >= iov->iov_len) {
 499                 iovoffset -= iov->iov_len;
 500                 iov++;
 501                 niov--;
 502                 LASSERT(niov > 0);
 503         }
 504
 505         LASSERT(nkiov > 0);
 506         while (kiovoffset >= kiov->kiov_len) {
 507                 kiovoffset -= kiov->kiov_len;
 508                 kiov++;
 509                 nkiov--;
 510                 LASSERT(nkiov > 0);
 511         }
 512
 513         do {
 514                 LASSERT(niov > 0);
 515                 LASSERT(nkiov > 0);
 516                 this_nob = MIN(iov->iov_len - iovoffset,
 517                                kiov->kiov_len - kiovoffset);
 518                 this_nob = MIN(this_nob, nob);
 519
 520                 if (addr == NULL)
 521                         addr = ((char *)kmap(kiov->kiov_page)) +
 522                                 kiov->kiov_offset + kiovoffset;
 523
 524                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 525                 nob -= this_nob;
 526
 527                 if (iov->iov_len > iovoffset + this_nob) {
 528                         iovoffset += this_nob;
 529                 } else {
 530                         iov++;
 531                         niov--;
 532                         iovoffset = 0;
 533                 }
 534
 535                 if (kiov->kiov_len > kiovoffset + this_nob) {
 536                         addr += this_nob;
 537                         kiovoffset += this_nob;
 538                 } else {
 539                         kunmap(kiov->kiov_page);
 540                         addr = NULL;
 541                         kiov++;
 542                         nkiov--;
 543                         kiovoffset = 0;
 544                 }
 545
 546         } while (nob > 0);
 547
 548         if (addr != NULL)
 549                 kunmap(kiov->kiov_page);
 550 }
 551 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 552
 553 void
 554 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 555                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 556                    unsigned int nob)
 557 {
 558         /* NB kiov, iov are READ-ONLY */
 559         unsigned int    this_nob;
 560         char           *addr = NULL;
 561
 562         if (nob == 0)
 563                 return;
 564
 565         LASSERT (!in_interrupt ());
 566
 567         LASSERT (nkiov > 0);
 568         while (kiovoffset >= kiov->kiov_len) {
 569                 kiovoffset -= kiov->kiov_len;
 570                 kiov++;
 571                 nkiov--;
 572                 LASSERT(nkiov > 0);
 573         }
 574
 575         LASSERT(niov > 0);
 576         while (iovoffset >= iov->iov_len) {
 577                 iovoffset -= iov->iov_len;
 578                 iov++;
 579                 niov--;
 580                 LASSERT(niov > 0);
 581         }
 582
 583         do {
 584                 LASSERT(nkiov > 0);
 585                 LASSERT(niov > 0);
 586                 this_nob = MIN(kiov->kiov_len - kiovoffset,
 587                                iov->iov_len - iovoffset);
 588                 this_nob = MIN(this_nob, nob);
 589
 590                 if (addr == NULL)
 591                         addr = ((char *)kmap(kiov->kiov_page)) +
 592                                 kiov->kiov_offset + kiovoffset;
 593
 594                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 595                 nob -= this_nob;
 596
 597                 if (kiov->kiov_len > kiovoffset + this_nob) {
 598                         addr += this_nob;
 599                         kiovoffset += this_nob;
 600                 } else {
 601                         kunmap(kiov->kiov_page);
 602                         addr = NULL;
 603                         kiov++;
 604                         nkiov--;
 605                         kiovoffset = 0;
 606                 }
 607
 608                 if (iov->iov_len > iovoffset + this_nob) {
 609                         iovoffset += this_nob;
 610                 } else {
 611                         iov++;
 612                         niov--;
 613                         iovoffset = 0;
 614                 }
 615         } while (nob > 0);
 616
 617         if (addr != NULL)
 618                 kunmap(kiov->kiov_page);
 619 }
 620 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 621
 622 int
 623 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
 624                   int src_niov, lnet_kiov_t *src,
 625                   unsigned int offset, unsigned int len)
 626 {
 627         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 628          * for exactly 'len' bytes, and return the number of entries.
 629          * NB not destructive to 'src' */
 630         unsigned int    frag_len;
 631         unsigned int    niov;
 632
 633         if (len == 0)                           /* no data => */
 634                 return (0);                     /* no frags */
 635
 636         LASSERT(src_niov > 0);
 637         while (offset >= src->kiov_len) {      /* skip initial frags */
 638                 offset -= src->kiov_len;
 639                 src_niov--;
 640                 src++;
 641                 LASSERT(src_niov > 0);
 642         }
 643
 644         niov = 1;
 645         for (;;) {
 646                 LASSERT(src_niov > 0);
 647                 LASSERT((int)niov <= dst_niov);
 648
 649                 frag_len = src->kiov_len - offset;
 650                 dst->kiov_page = src->kiov_page;
 651                 dst->kiov_offset = src->kiov_offset + offset;
 652
 653                 if (len <= frag_len) {
 654                         dst->kiov_len = len;
 655                         LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 656                         return niov;
 657                 }
 658
 659                 dst->kiov_len = frag_len;
 660                 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 661
 662                 len -= frag_len;
 663                 dst++;
 664                 src++;
 665                 niov++;
 666                 src_niov--;
 667                 offset = 0;
 668         }
 669 }
 670 EXPORT_SYMBOL(lnet_extract_kiov);
 671
 672 void
 673 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 674              int delayed, unsigned int offset, unsigned int mlen,
 675              unsigned int rlen)
 676 {
 677         unsigned int  niov = 0;
 678         struct kvec *iov = NULL;
 679         lnet_kiov_t  *kiov = NULL;
 680         int           rc;
 681
 682         LASSERT (!in_interrupt ());
 683         LASSERT (mlen == 0 || msg != NULL);
 684
 685         if (msg != NULL) {
 686                 LASSERT(msg->msg_receiving);
 687                 LASSERT(!msg->msg_sending);
 688                 LASSERT(rlen == msg->msg_len);
 689                 LASSERT(mlen <= msg->msg_len);
 690                 LASSERT(msg->msg_offset == offset);
 691                 LASSERT(msg->msg_wanted == mlen);
 692
 693                 msg->msg_receiving = 0;
 694
 695                 if (mlen != 0) {
 696                         niov = msg->msg_niov;
 697                         iov  = msg->msg_iov;
 698                         kiov = msg->msg_kiov;
 699
 700                         LASSERT (niov > 0);
 701                         LASSERT ((iov == NULL) != (kiov == NULL));
 702                 }
 703         }
 704
 705         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 706                                              niov, iov, kiov, offset, mlen,
 707                                              rlen);
 708         if (rc < 0)
 709                 lnet_finalize(msg, rc);
 710 }
 711
 712 static void
 713 lnet_setpayloadbuffer(struct lnet_msg *msg)
 714 {
 715         struct lnet_libmd *md = msg->msg_md;
 716
 717         LASSERT(msg->msg_len > 0);
 718         LASSERT(!msg->msg_routing);
 719         LASSERT(md != NULL);
 720         LASSERT(msg->msg_niov == 0);
 721         LASSERT(msg->msg_iov == NULL);
 722         LASSERT(msg->msg_kiov == NULL);
 723
 724         msg->msg_niov = md->md_niov;
 725         if ((md->md_options & LNET_MD_KIOV) != 0)
 726                 msg->msg_kiov = md->md_iov.kiov;
 727         else
 728                 msg->msg_iov = md->md_iov.iov;
 729 }
 730
 731 void
 732 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 733                unsigned int offset, unsigned int len)
 734 {
 735         msg->msg_type = type;
 736         msg->msg_target = target;
 737         msg->msg_len = len;
 738         msg->msg_offset = offset;
 739
 740         if (len != 0)
 741                 lnet_setpayloadbuffer(msg);
 742
 743         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 744         msg->msg_hdr.type           = cpu_to_le32(type);
 745         /* dest_nid will be overwritten by lnet_select_pathway() */
 746         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 747         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 748         /* src_nid will be set later */
 749         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 750         msg->msg_hdr.payload_length = cpu_to_le32(len);
 751 }
 752
 753 static void
 754 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 755 {
 756         void   *priv = msg->msg_private;
 757         int rc;
 758
 759         LASSERT (!in_interrupt ());
 760         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
 761                  (msg->msg_txcredit && msg->msg_peertxcredit));
 762
 763         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 764         if (rc < 0) {
 765                 msg->msg_no_resend = true;
 766                 lnet_finalize(msg, rc);
 767         }
 768 }
 769
 770 static int
 771 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 772 {
 773         int     rc;
 774
 775         LASSERT(!msg->msg_sending);
 776         LASSERT(msg->msg_receiving);
 777         LASSERT(!msg->msg_rx_ready_delay);
 778         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 779
 780         msg->msg_rx_ready_delay = 1;
 781         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 782                                                   &msg->msg_private);
 783         if (rc != 0) {
 784                 CERROR("recv from %s / send to %s aborted: "
 785                        "eager_recv failed %d\n",
 786                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 787                        libcfs_id2str(msg->msg_target), rc);
 788                 LASSERT(rc < 0); /* required by my callers */
 789         }
 790
 791         return rc;
 792 }
 793
 794 static bool
 795 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 796 {
 797         time64_t deadline;
 798
 799         deadline = lpni->lpni_last_alive +
 800                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 801
 802         /*
 803          * assume peer_ni is alive as long as we're within the configured
 804          * peer timeout
 805          */
 806         if (deadline > now)
 807                 return false;
 808
 809         return true;
 810 }
 811
 812 /* NB: returns 1 when alive, 0 when dead, negative when error;
 813  *     may drop the lnet_net_lock */
 814 static int
 815 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 816                        struct lnet_msg *msg)
 817 {
 818         time64_t now = ktime_get_seconds();
 819
 820         if (!lnet_peer_aliveness_enabled(lpni))
 821                 return -ENODEV;
 822
 823         /*
 824          * If we're resending a message, let's attempt to send it even if
 825          * the peer is down to fulfill our resend quota on the message
 826          */
 827         if (msg->msg_retry_count > 0)
 828                 return 1;
 829
 830         /* try and send recovery messages irregardless */
 831         if (msg->msg_recovery)
 832                 return 1;
 833
 834         /* always send any responses */
 835         if (msg->msg_type == LNET_MSG_ACK ||
 836             msg->msg_type == LNET_MSG_REPLY)
 837                 return 1;
 838
 839         if (!lnet_is_peer_deadline_passed(lpni, now))
 840                 return true;
 841
 842         return lnet_is_peer_ni_alive(lpni);
 843 }
 844
 845 /**
 846  * \param msg The message to be sent.
 847  * \param do_send True if lnet_ni_send() should be called in this function.
 848  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 849  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 850  *
 851  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 852  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 853  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 854  * \retval -ECANCELED If the MD of the message has been unlinked.
 855  */
 856 static int
 857 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 858 {
 859         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 860         struct lnet_ni          *ni = msg->msg_txni;
 861         int                     cpt = msg->msg_tx_cpt;
 862         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 863
 864         /* non-lnet_send() callers have checked before */
 865         LASSERT(!do_send || msg->msg_tx_delayed);
 866         LASSERT(!msg->msg_receiving);
 867         LASSERT(msg->msg_tx_committed);
 868         /* can't get here if we're sending to the loopback interface */
 869         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
 870
 871         /* NB 'lp' is always the next hop */
 872         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 873             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 874                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 875                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 876                         msg->msg_len;
 877                 lnet_net_unlock(cpt);
 878                 if (msg->msg_txpeer)
 879                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 880                                         msg->msg_type,
 881                                         LNET_STATS_TYPE_DROP);
 882                 if (msg->msg_txni)
 883                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 884                                         msg->msg_type,
 885                                         LNET_STATS_TYPE_DROP);
 886
 887                 CNETERR("Dropping message for %s: peer not alive\n",
 888                         libcfs_id2str(msg->msg_target));
 889                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 890                 if (do_send)
 891                         lnet_finalize(msg, -EHOSTUNREACH);
 892
 893                 lnet_net_lock(cpt);
 894                 return -EHOSTUNREACH;
 895         }
 896
 897         if (msg->msg_md != NULL &&
 898             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 899                 lnet_net_unlock(cpt);
 900
 901                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 902                         "called on the MD/ME.\n",
 903                         libcfs_id2str(msg->msg_target));
 904                 if (do_send) {
 905                         msg->msg_no_resend = true;
 906                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 907                                msg, libcfs_id2str(msg->msg_target));
 908                         lnet_finalize(msg, -ECANCELED);
 909                 }
 910
 911                 lnet_net_lock(cpt);
 912                 return -ECANCELED;
 913         }
 914
 915         if (!msg->msg_peertxcredit) {
 916                 spin_lock(&lp->lpni_lock);
 917                 LASSERT((lp->lpni_txcredits < 0) ==
 918                         !list_empty(&lp->lpni_txq));
 919
 920                 msg->msg_peertxcredit = 1;
 921                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 922                 lp->lpni_txcredits--;
 923
 924                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 925                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 926
 927                 if (lp->lpni_txcredits < 0) {
 928                         msg->msg_tx_delayed = 1;
 929                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 930                         spin_unlock(&lp->lpni_lock);
 931                         return LNET_CREDIT_WAIT;
 932                 }
 933                 spin_unlock(&lp->lpni_lock);
 934         }
 935
 936         if (!msg->msg_txcredit) {
 937                 LASSERT((tq->tq_credits < 0) ==
 938                         !list_empty(&tq->tq_delayed));
 939
 940                 msg->msg_txcredit = 1;
 941                 tq->tq_credits--;
 942                 atomic_dec(&ni->ni_tx_credits);
 943
 944                 if (tq->tq_credits < tq->tq_credits_min)
 945                         tq->tq_credits_min = tq->tq_credits;
 946
 947                 if (tq->tq_credits < 0) {
 948                         msg->msg_tx_delayed = 1;
 949                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 950                         return LNET_CREDIT_WAIT;
 951                 }
 952         }
 953
 954         /* unset the tx_delay flag as we're going to send it now */
 955         msg->msg_tx_delayed = 0;
 956
 957         if (do_send) {
 958                 lnet_net_unlock(cpt);
 959                 lnet_ni_send(ni, msg);
 960                 lnet_net_lock(cpt);
 961         }
 962         return LNET_CREDIT_OK;
 963 }
 964
 965
 966 static struct lnet_rtrbufpool *
 967 lnet_msg2bufpool(struct lnet_msg *msg)
 968 {
 969         struct lnet_rtrbufpool  *rbp;
 970         int                     cpt;
 971
 972         LASSERT(msg->msg_rx_committed);
 973
 974         cpt = msg->msg_rx_cpt;
 975         rbp = &the_lnet.ln_rtrpools[cpt][0];
 976
 977         LASSERT(msg->msg_len <= LNET_MTU);
 978         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 979                 rbp++;
 980                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 981         }
 982
 983         return rbp;
 984 }
 985
 986 static int
 987 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 988 {
 989         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 990          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 991          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 992          * received or OK to receive */
 993         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 994         struct lnet_peer *lp;
 995         struct lnet_rtrbufpool *rbp;
 996         struct lnet_rtrbuf *rb;
 997
 998         LASSERT(msg->msg_iov == NULL);
 999         LASSERT(msg->msg_kiov == NULL);
1000         LASSERT(msg->msg_niov == 0);
1001         LASSERT(msg->msg_routing);
1002         LASSERT(msg->msg_receiving);
1003         LASSERT(!msg->msg_sending);
1004         LASSERT(lpni->lpni_peer_net);
1005         LASSERT(lpni->lpni_peer_net->lpn_peer);
1006
1007         lp = lpni->lpni_peer_net->lpn_peer;
1008
1009         /* non-lnet_parse callers only receive delayed messages */
1010         LASSERT(!do_recv || msg->msg_rx_delayed);
1011
1012         if (!msg->msg_peerrtrcredit) {
1013                 /* lpni_lock protects the credit manipulation */
1014                 spin_lock(&lpni->lpni_lock);
1015                 /* lp_lock protects the lp_rtrq */
1016                 spin_lock(&lp->lp_lock);
1017
1018                 msg->msg_peerrtrcredit = 1;
1019                 lpni->lpni_rtrcredits--;
1020                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1021                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1022
1023                 if (lpni->lpni_rtrcredits < 0) {
1024                         /* must have checked eager_recv before here */
1025                         LASSERT(msg->msg_rx_ready_delay);
1026                         msg->msg_rx_delayed = 1;
1027                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1028                         spin_unlock(&lp->lp_lock);
1029                         spin_unlock(&lpni->lpni_lock);
1030                         return LNET_CREDIT_WAIT;
1031                 }
1032                 spin_unlock(&lp->lp_lock);
1033                 spin_unlock(&lpni->lpni_lock);
1034         }
1035
1036         rbp = lnet_msg2bufpool(msg);
1037
1038         if (!msg->msg_rtrcredit) {
1039                 msg->msg_rtrcredit = 1;
1040                 rbp->rbp_credits--;
1041                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1042                         rbp->rbp_mincredits = rbp->rbp_credits;
1043
1044                 if (rbp->rbp_credits < 0) {
1045                         /* must have checked eager_recv before here */
1046                         LASSERT(msg->msg_rx_ready_delay);
1047                         msg->msg_rx_delayed = 1;
1048                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1049                         return LNET_CREDIT_WAIT;
1050                 }
1051         }
1052
1053         LASSERT(!list_empty(&rbp->rbp_bufs));
1054         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1055         list_del(&rb->rb_list);
1056
1057         msg->msg_niov = rbp->rbp_npages;
1058         msg->msg_kiov = &rb->rb_kiov[0];
1059
1060         /* unset the msg-rx_delayed flag since we're receiving the message */
1061         msg->msg_rx_delayed = 0;
1062
1063         if (do_recv) {
1064                 int cpt = msg->msg_rx_cpt;
1065
1066                 lnet_net_unlock(cpt);
1067                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1068                              0, msg->msg_len, msg->msg_len);
1069                 lnet_net_lock(cpt);
1070         }
1071         return LNET_CREDIT_OK;
1072 }
1073
1074 void
1075 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1076 {
1077         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1078         struct lnet_ni          *txni = msg->msg_txni;
1079         struct lnet_msg         *msg2;
1080
1081         if (msg->msg_txcredit) {
1082                 struct lnet_ni       *ni = msg->msg_txni;
1083                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1084
1085                 /* give back NI txcredits */
1086                 msg->msg_txcredit = 0;
1087
1088                 LASSERT((tq->tq_credits < 0) ==
1089                         !list_empty(&tq->tq_delayed));
1090
1091                 tq->tq_credits++;
1092                 atomic_inc(&ni->ni_tx_credits);
1093                 if (tq->tq_credits <= 0) {
1094                         msg2 = list_entry(tq->tq_delayed.next,
1095                                           struct lnet_msg, msg_list);
1096                         list_del(&msg2->msg_list);
1097
1098                         LASSERT(msg2->msg_txni == ni);
1099                         LASSERT(msg2->msg_tx_delayed);
1100                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1101
1102                         (void) lnet_post_send_locked(msg2, 1);
1103                 }
1104         }
1105
1106         if (msg->msg_peertxcredit) {
1107                 /* give back peer txcredits */
1108                 msg->msg_peertxcredit = 0;
1109
1110                 spin_lock(&txpeer->lpni_lock);
1111                 LASSERT((txpeer->lpni_txcredits < 0) ==
1112                         !list_empty(&txpeer->lpni_txq));
1113
1114                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1115                 LASSERT(txpeer->lpni_txqnob >= 0);
1116
1117                 txpeer->lpni_txcredits++;
1118                 if (txpeer->lpni_txcredits <= 0) {
1119                         int msg2_cpt;
1120
1121                         msg2 = list_entry(txpeer->lpni_txq.next,
1122                                               struct lnet_msg, msg_list);
1123                         list_del(&msg2->msg_list);
1124                         spin_unlock(&txpeer->lpni_lock);
1125
1126                         LASSERT(msg2->msg_txpeer == txpeer);
1127                         LASSERT(msg2->msg_tx_delayed);
1128
1129                         msg2_cpt = msg2->msg_tx_cpt;
1130
1131                         /*
1132                          * The msg_cpt can be different from the msg2_cpt
1133                          * so we need to make sure we lock the correct cpt
1134                          * for msg2.
1135                          * Once we call lnet_post_send_locked() it is no
1136                          * longer safe to access msg2, since it could've
1137                          * been freed by lnet_finalize(), but we still
1138                          * need to relock the correct cpt, so we cache the
1139                          * msg2_cpt for the purpose of the check that
1140                          * follows the call to lnet_pose_send_locked().
1141                          */
1142                         if (msg2_cpt != msg->msg_tx_cpt) {
1143                                 lnet_net_unlock(msg->msg_tx_cpt);
1144                                 lnet_net_lock(msg2_cpt);
1145                         }
1146                         (void) lnet_post_send_locked(msg2, 1);
1147                         if (msg2_cpt != msg->msg_tx_cpt) {
1148                                 lnet_net_unlock(msg2_cpt);
1149                                 lnet_net_lock(msg->msg_tx_cpt);
1150                         }
1151                 } else {
1152                         spin_unlock(&txpeer->lpni_lock);
1153                 }
1154         }
1155
1156         if (txni != NULL) {
1157                 msg->msg_txni = NULL;
1158                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1159         }
1160
1161         if (txpeer != NULL) {
1162                 msg->msg_txpeer = NULL;
1163                 lnet_peer_ni_decref_locked(txpeer);
1164         }
1165 }
1166
1167 void
1168 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1169 {
1170         struct lnet_msg *msg;
1171
1172         if (list_empty(&rbp->rbp_msgs))
1173                 return;
1174         msg = list_entry(rbp->rbp_msgs.next,
1175                          struct lnet_msg, msg_list);
1176         list_del(&msg->msg_list);
1177
1178         (void)lnet_post_routed_recv_locked(msg, 1);
1179 }
1180
1181 void
1182 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1183 {
1184         struct lnet_msg *msg;
1185         struct lnet_msg *tmp;
1186
1187         lnet_net_unlock(cpt);
1188
1189         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1190                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1191                              0, 0, 0, msg->msg_hdr.payload_length);
1192                 list_del_init(&msg->msg_list);
1193                 msg->msg_no_resend = true;
1194                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1195                 lnet_finalize(msg, -ECANCELED);
1196         }
1197
1198         lnet_net_lock(cpt);
1199 }
1200
1201 void
1202 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1203 {
1204         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1205         struct lnet_peer *lp;
1206         struct lnet_ni *rxni = msg->msg_rxni;
1207         struct lnet_msg *msg2;
1208
1209         if (msg->msg_rtrcredit) {
1210                 /* give back global router credits */
1211                 struct lnet_rtrbuf *rb;
1212                 struct lnet_rtrbufpool *rbp;
1213
1214                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1215                  * there until it gets one allocated, or aborts the wait
1216                  * itself */
1217                 LASSERT(msg->msg_kiov != NULL);
1218
1219                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1220                 rbp = rb->rb_pool;
1221
1222                 msg->msg_kiov = NULL;
1223                 msg->msg_rtrcredit = 0;
1224
1225                 LASSERT(rbp == lnet_msg2bufpool(msg));
1226
1227                 LASSERT((rbp->rbp_credits > 0) ==
1228                         !list_empty(&rbp->rbp_bufs));
1229
1230                 /* If routing is now turned off, we just drop this buffer and
1231                  * don't bother trying to return credits.  */
1232                 if (!the_lnet.ln_routing) {
1233                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1234                         goto routing_off;
1235                 }
1236
1237                 /* It is possible that a user has lowered the desired number of
1238                  * buffers in this pool.  Make sure we never put back
1239                  * more buffers than the stated number. */
1240                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1241                         /* Discard this buffer so we don't have too
1242                          * many. */
1243                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1244                         rbp->rbp_nbuffers--;
1245                 } else {
1246                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1247                         rbp->rbp_credits++;
1248                         if (rbp->rbp_credits <= 0)
1249                                 lnet_schedule_blocked_locked(rbp);
1250                 }
1251         }
1252
1253 routing_off:
1254         if (msg->msg_peerrtrcredit) {
1255                 LASSERT(rxpeerni);
1256                 LASSERT(rxpeerni->lpni_peer_net);
1257                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1258
1259                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1260
1261                 /* give back peer router credits */
1262                 msg->msg_peerrtrcredit = 0;
1263
1264                 spin_lock(&rxpeerni->lpni_lock);
1265                 spin_lock(&lp->lp_lock);
1266
1267                 rxpeerni->lpni_rtrcredits++;
1268
1269                 /* drop all messages which are queued to be routed on that
1270                  * peer. */
1271                 if (!the_lnet.ln_routing) {
1272                         struct list_head drop;
1273                         INIT_LIST_HEAD(&drop);
1274                         list_splice_init(&lp->lp_rtrq, &drop);
1275                         spin_unlock(&lp->lp_lock);
1276                         spin_unlock(&rxpeerni->lpni_lock);
1277                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1278                 } else if (!list_empty(&lp->lp_rtrq)) {
1279                         int msg2_cpt;
1280
1281                         msg2 = list_entry(lp->lp_rtrq.next,
1282                                           struct lnet_msg, msg_list);
1283                         list_del(&msg2->msg_list);
1284                         msg2_cpt = msg2->msg_rx_cpt;
1285                         spin_unlock(&lp->lp_lock);
1286                         spin_unlock(&rxpeerni->lpni_lock);
1287                         /*
1288                          * messages on the lp_rtrq can be from any NID in
1289                          * the peer, which means they might have different
1290                          * cpts. We need to make sure we lock the right
1291                          * one.
1292                          */
1293                         if (msg2_cpt != msg->msg_rx_cpt) {
1294                                 lnet_net_unlock(msg->msg_rx_cpt);
1295                                 lnet_net_lock(msg2_cpt);
1296                         }
1297                         (void) lnet_post_routed_recv_locked(msg2, 1);
1298                         if (msg2_cpt != msg->msg_rx_cpt) {
1299                                 lnet_net_unlock(msg2_cpt);
1300                                 lnet_net_lock(msg->msg_rx_cpt);
1301                         }
1302                 } else {
1303                         spin_unlock(&lp->lp_lock);
1304                         spin_unlock(&rxpeerni->lpni_lock);
1305                 }
1306         }
1307         if (rxni != NULL) {
1308                 msg->msg_rxni = NULL;
1309                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1310         }
1311         if (rxpeerni != NULL) {
1312                 msg->msg_rxpeer = NULL;
1313                 lnet_peer_ni_decref_locked(rxpeerni);
1314         }
1315 }
1316
1317 static int
1318 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1319 {
1320         if (p1->lpni_txqnob < p2->lpni_txqnob)
1321                 return 1;
1322
1323         if (p1->lpni_txqnob > p2->lpni_txqnob)
1324                 return -1;
1325
1326         if (p1->lpni_txcredits > p2->lpni_txcredits)
1327                 return 1;
1328
1329         if (p1->lpni_txcredits < p2->lpni_txcredits)
1330                 return -1;
1331
1332         return 0;
1333 }
1334
1335 static struct lnet_peer_ni *
1336 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1337                     struct lnet_peer *peer,
1338                     struct lnet_peer_net *peer_net)
1339 {
1340         /*
1341          * Look at the peer NIs for the destination peer that connect
1342          * to the chosen net. If a peer_ni is preferred when using the
1343          * best_ni to communicate, we use that one. If there is no
1344          * preferred peer_ni, or there are multiple preferred peer_ni,
1345          * the available transmit credits are used. If the transmit
1346          * credits are equal, we round-robin over the peer_ni.
1347          */
1348         struct lnet_peer_ni *lpni = NULL;
1349         struct lnet_peer_ni *best_lpni = NULL;
1350         int best_lpni_credits = INT_MIN;
1351         bool preferred = false;
1352         bool ni_is_pref;
1353         int best_lpni_healthv = 0;
1354         int lpni_healthv;
1355
1356         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1357                 /*
1358                  * if the best_ni we've chosen aleady has this lpni
1359                  * preferred, then let's use it
1360                  */
1361                 if (best_ni) {
1362                         ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1363                                                                 best_ni->ni_nid);
1364                         CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1365                                libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1366                 } else {
1367                         ni_is_pref = false;
1368                 }
1369
1370                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1371
1372                 if (best_lpni)
1373                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1374                                 libcfs_nid2str(lpni->lpni_nid),
1375                                 lpni->lpni_txcredits, best_lpni_credits,
1376                                 lpni->lpni_seq, best_lpni->lpni_seq);
1377
1378                 /* pick the healthiest peer ni */
1379                 if (lpni_healthv < best_lpni_healthv) {
1380                         continue;
1381                 } else if (lpni_healthv > best_lpni_healthv) {
1382                         best_lpni_healthv = lpni_healthv;
1383                 /* if this is a preferred peer use it */
1384                 } else if (!preferred && ni_is_pref) {
1385                         preferred = true;
1386                 } else if (preferred && !ni_is_pref) {
1387                         /*
1388                          * this is not the preferred peer so let's ignore
1389                          * it.
1390                          */
1391                         continue;
1392                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1393                         /*
1394                          * We already have a peer that has more credits
1395                          * available than this one. No need to consider
1396                          * this peer further.
1397                          */
1398                         continue;
1399                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1400                         /*
1401                          * The best peer found so far and the current peer
1402                          * have the same number of available credits let's
1403                          * make sure to select between them using Round
1404                          * Robin
1405                          */
1406                         if (best_lpni) {
1407                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1408                                         continue;
1409                         }
1410                 }
1411
1412                 best_lpni = lpni;
1413                 best_lpni_credits = lpni->lpni_txcredits;
1414         }
1415
1416         /* if we still can't find a peer ni then we can't reach it */
1417         if (!best_lpni) {
1418                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1419                         LNET_NIDNET(dst_nid);
1420                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1421                                 libcfs_net2str(net_id));
1422                 return NULL;
1423         }
1424
1425         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1426                libcfs_nid2str(best_lpni->lpni_nid));
1427
1428         return best_lpni;
1429 }
1430
1431 /*
1432  * Prerequisite: the best_ni should already be set in the sd
1433  */
1434 static inline struct lnet_peer_ni *
1435 lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
1436                            struct lnet_peer *peer, __u32 net_id)
1437 {
1438         struct lnet_peer_net *peer_net;
1439
1440         /*
1441          * The gateway is Multi-Rail capable so now we must select the
1442          * proper peer_ni
1443          */
1444         peer_net = lnet_peer_get_net_locked(peer, net_id);
1445
1446         if (!peer_net) {
1447                 CERROR("gateway peer %s has no NI on net %s\n",
1448                        libcfs_nid2str(peer->lp_primary_nid),
1449                        libcfs_net2str(net_id));
1450                 return NULL;
1451         }
1452
1453         return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
1454 }
1455
1456 static int
1457 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2,
1458                     struct lnet_peer_ni **best_lpni)
1459 {
1460         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1461         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1462         struct lnet_peer *lp1 = r1->lr_gateway;
1463         struct lnet_peer *lp2 = r2->lr_gateway;
1464         struct lnet_peer_ni *lpni1;
1465         struct lnet_peer_ni *lpni2;
1466         int rc;
1467
1468         lpni1 = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY, lp1,
1469                                            r1->lr_lnet);
1470         lpni2 = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY, lp2,
1471                                            r2->lr_lnet);
1472         LASSERT(lpni1 && lpni2);
1473
1474         if (r1->lr_priority < r2->lr_priority) {
1475                 *best_lpni = lpni1;
1476                 return 1;
1477         }
1478
1479         if (r1->lr_priority > r2->lr_priority) {
1480                 *best_lpni = lpni2;
1481                 return -1;
1482         }
1483
1484         if (r1_hops < r2_hops) {
1485                 *best_lpni = lpni1;
1486                 return 1;
1487         }
1488
1489         if (r1_hops > r2_hops) {
1490                 *best_lpni = lpni2;
1491                 return -1;
1492         }
1493
1494         rc = lnet_compare_peers(lpni1, lpni2);
1495         if (rc == 1) {
1496                 *best_lpni = lpni1;
1497                 return rc;
1498         } else if (rc == -1) {
1499                 *best_lpni = lpni2;
1500                 return rc;
1501         }
1502
1503         if (r1->lr_seq - r2->lr_seq <= 0) {
1504                 *best_lpni = lpni1;
1505                 return 1;
1506         }
1507
1508         *best_lpni = lpni2;
1509         return -1;
1510 }
1511
1512 static struct lnet_route *
1513 lnet_find_route_locked(struct lnet_remotenet *rnet,
1514                        struct lnet_route **prev_route,
1515                        struct lnet_peer_ni **gwni)
1516 {
1517         struct lnet_peer_ni *best_gw_ni = NULL;
1518         struct lnet_route *best_route;
1519         struct lnet_route *last_route;
1520         struct lnet_route *route;
1521         int rc;
1522
1523         best_route = last_route = NULL;
1524         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1525                 if (!lnet_is_route_alive(route))
1526                         continue;
1527
1528                 if (!best_route) {
1529                         best_route = last_route = route;
1530                         best_gw_ni = lnet_find_best_lpni_on_net(NULL,
1531                                                                 LNET_NID_ANY,
1532                                                                 route->lr_gateway,
1533                                                                 route->lr_lnet);
1534                         LASSERT(best_gw_ni);
1535                         continue;
1536                 }
1537
1538                 /* no protection on below fields, but it's harmless */
1539                 if (last_route->lr_seq - route->lr_seq < 0)
1540                         last_route = route;
1541
1542                 rc = lnet_compare_routes(route, best_route, &best_gw_ni);
1543                 if (rc < 0)
1544                         continue;
1545
1546                 best_route = route;
1547         }
1548
1549         *prev_route = last_route;
1550         *gwni = best_gw_ni;
1551
1552         return best_route;
1553 }
1554
1555 static struct lnet_ni *
1556 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1557                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1558                  int md_cpt)
1559 {
1560         struct lnet_ni *ni = NULL;
1561         unsigned int shortest_distance;
1562         int best_credits;
1563         int best_healthv;
1564
1565         /*
1566          * If there is no peer_ni that we can send to on this network,
1567          * then there is no point in looking for a new best_ni here.
1568         */
1569         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1570                 return best_ni;
1571
1572         if (best_ni == NULL) {
1573                 shortest_distance = UINT_MAX;
1574                 best_credits = INT_MIN;
1575                 best_healthv = 0;
1576         } else {
1577                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1578                                                      best_ni->ni_dev_cpt);
1579                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1580                 best_healthv = atomic_read(&best_ni->ni_healthv);
1581         }
1582
1583         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1584                 unsigned int distance;
1585                 int ni_credits;
1586                 int ni_healthv;
1587                 int ni_fatal;
1588
1589                 ni_credits = atomic_read(&ni->ni_tx_credits);
1590                 ni_healthv = atomic_read(&ni->ni_healthv);
1591                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1592
1593                 /*
1594                  * calculate the distance from the CPT on which
1595                  * the message memory is allocated to the CPT of
1596                  * the NI's physical device
1597                  */
1598                 distance = cfs_cpt_distance(lnet_cpt_table(),
1599                                             md_cpt,
1600                                             ni->ni_dev_cpt);
1601
1602                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1603                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1604                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1605                         : "not seleced", best_credits, shortest_distance,
1606                         (best_ni) ? best_ni->ni_seq : 0);
1607
1608                 /*
1609                  * All distances smaller than the NUMA range
1610                  * are treated equally.
1611                  */
1612                 if (distance < lnet_numa_range)
1613                         distance = lnet_numa_range;
1614
1615                 /*
1616                  * Select on health, shorter distance, available
1617                  * credits, then round-robin.
1618                  */
1619                 if (ni_fatal) {
1620                         continue;
1621                 } else if (ni_healthv < best_healthv) {
1622                         continue;
1623                 } else if (ni_healthv > best_healthv) {
1624                         best_healthv = ni_healthv;
1625                         /*
1626                          * If we're going to prefer this ni because it's
1627                          * the healthiest, then we should set the
1628                          * shortest_distance in the algorithm in case
1629                          * there are multiple NIs with the same health but
1630                          * different distances.
1631                          */
1632                         if (distance < shortest_distance)
1633                                 shortest_distance = distance;
1634                 } else if (distance > shortest_distance) {
1635                         continue;
1636                 } else if (distance < shortest_distance) {
1637                         shortest_distance = distance;
1638                 } else if (ni_credits < best_credits) {
1639                         continue;
1640                 } else if (ni_credits == best_credits) {
1641                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1642                                 continue;
1643                 }
1644                 best_ni = ni;
1645                 best_credits = ni_credits;
1646         }
1647
1648         CDEBUG(D_NET, "selected best_ni %s\n",
1649                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1650
1651         return best_ni;
1652 }
1653
1654 /*
1655  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1656  * because such traffic is required to perform discovery. We therefore
1657  * exclude all GET and PUT on that portal. We also exclude all ACK and
1658  * REPLY traffic, but that is because the portal is not tracked in the
1659  * message structure for these message types. We could restrict this
1660  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1661  */
1662 static bool
1663 lnet_msg_discovery(struct lnet_msg *msg)
1664 {
1665         if (msg->msg_type == LNET_MSG_PUT) {
1666                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1667                         return true;
1668         } else if (msg->msg_type == LNET_MSG_GET) {
1669                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1670                         return true;
1671         }
1672         return false;
1673 }
1674
1675 #define SRC_SPEC        0x0001
1676 #define SRC_ANY         0x0002
1677 #define LOCAL_DST       0x0004
1678 #define REMOTE_DST      0x0008
1679 #define MR_DST          0x0010
1680 #define NMR_DST         0x0020
1681 #define SND_RESP        0x0040
1682
1683 /* The following to defines are used for return codes */
1684 #define REPEAT_SEND     0x1000
1685 #define PASS_THROUGH    0x2000
1686
1687 /* The different cases lnet_select pathway needs to handle */
1688 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1689 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1690 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1691 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1692 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1693 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1694 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1695 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1696
1697 static int
1698 lnet_handle_lo_send(struct lnet_send_data *sd)
1699 {
1700         struct lnet_msg *msg = sd->sd_msg;
1701         int cpt = sd->sd_cpt;
1702
1703         /* No send credit hassles with LOLND */
1704         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1705         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1706         if (!msg->msg_routing)
1707                 msg->msg_hdr.src_nid =
1708                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1709         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1710         lnet_msg_commit(msg, cpt);
1711         msg->msg_txni = the_lnet.ln_loni;
1712
1713         return LNET_CREDIT_OK;
1714 }
1715
1716 static int
1717 lnet_handle_send(struct lnet_send_data *sd)
1718 {
1719         struct lnet_ni *best_ni = sd->sd_best_ni;
1720         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1721         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1722         struct lnet_msg *msg = sd->sd_msg;
1723         int cpt2;
1724         __u32 send_case = sd->sd_send_case;
1725         int rc;
1726         __u32 routing = send_case & REMOTE_DST;
1727          struct lnet_rsp_tracker *rspt;
1728
1729         /*
1730          * Increment sequence number of the selected peer so that we
1731          * pick the next one in Round Robin.
1732          */
1733         best_lpni->lpni_seq++;
1734
1735         /*
1736          * grab a reference on the peer_ni so it sticks around even if
1737          * we need to drop and relock the lnet_net_lock below.
1738          */
1739         lnet_peer_ni_addref_locked(best_lpni);
1740
1741         /*
1742          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1743          * message. This ensures that we get a CPT that is correct for
1744          * the NI when the NI has been restricted to a subset of all CPTs.
1745          * If the selected CPT differs from the one currently locked, we
1746          * must unlock and relock the lnet_net_lock(), and then check whether
1747          * the configuration has changed. We don't have a hold on the best_ni
1748          * yet, and it may have vanished.
1749          */
1750         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1751         if (sd->sd_cpt != cpt2) {
1752                 __u32 seq = lnet_get_dlc_seq_locked();
1753                 lnet_net_unlock(sd->sd_cpt);
1754                 sd->sd_cpt = cpt2;
1755                 lnet_net_lock(sd->sd_cpt);
1756                 if (seq != lnet_get_dlc_seq_locked()) {
1757                         lnet_peer_ni_decref_locked(best_lpni);
1758                         return REPEAT_SEND;
1759                 }
1760         }
1761
1762         /*
1763          * store the best_lpni in the message right away to avoid having
1764          * to do the same operation under different conditions
1765          */
1766         msg->msg_txpeer = best_lpni;
1767         msg->msg_txni = best_ni;
1768
1769         /*
1770          * grab a reference for the best_ni since now it's in use in this
1771          * send. The reference will be dropped in lnet_finalize()
1772          */
1773         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1774
1775         /*
1776          * Always set the target.nid to the best peer picked. Either the
1777          * NID will be one of the peer NIDs selected, or the same NID as
1778          * what was originally set in the target or it will be the NID of
1779          * a router if this message should be routed
1780          */
1781         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1782
1783         /*
1784          * lnet_msg_commit assigns the correct cpt to the message, which
1785          * is used to decrement the correct refcount on the ni when it's
1786          * time to return the credits
1787          */
1788         lnet_msg_commit(msg, sd->sd_cpt);
1789
1790         /*
1791          * If we are routing the message then we keep the src_nid that was
1792          * set by the originator. If we are not routing then we are the
1793          * originator and set it here.
1794          */
1795         if (!msg->msg_routing)
1796                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1797
1798         if (routing) {
1799                 msg->msg_target_is_router = 1;
1800                 msg->msg_target.pid = LNET_PID_LUSTRE;
1801                 /*
1802                  * since we're routing we want to ensure that the
1803                  * msg_hdr.dest_nid is set to the final destination. When
1804                  * the router receives this message it knows how to route
1805                  * it.
1806                  *
1807                  * final_dst_lpni is set at the beginning of the
1808                  * lnet_select_pathway() function and is never changed.
1809                  * It's safe to use it here.
1810                  */
1811                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1812         } else {
1813                 /*
1814                  * if we're not routing set the dest_nid to the best peer
1815                  * ni NID that we picked earlier in the algorithm.
1816                  */
1817                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1818         }
1819
1820         /*
1821          * if we have response tracker block update it with the next hop
1822          * nid
1823          */
1824         if (msg->msg_md) {
1825                 rspt = msg->msg_md->md_rspt_ptr;
1826                 if (rspt) {
1827                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1828                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1829                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1830                 }
1831         }
1832
1833         rc = lnet_post_send_locked(msg, 0);
1834
1835         if (!rc)
1836                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1837                        libcfs_nid2str(msg->msg_hdr.src_nid),
1838                        libcfs_nid2str(msg->msg_txni->ni_nid),
1839                        libcfs_nid2str(sd->sd_src_nid),
1840                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1841                        libcfs_nid2str(sd->sd_dst_nid),
1842                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1843                        libcfs_nid2str(sd->sd_rtr_nid),
1844                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1845
1846         return rc;
1847 }
1848
1849 static inline void
1850 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1851 {
1852         if (sd->sd_send_case & NMR_DST &&
1853             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1854             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1855             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1856                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1857                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1858                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1859                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1860                                                  sd->sd_best_ni->ni_nid);
1861         }
1862 }
1863
1864 /*
1865  * Source Specified
1866  * Local Destination
1867  * non-mr peer
1868  *
1869  * use the source and destination NIDs as the pathway
1870  */
1871 static int
1872 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1873 {
1874         /* the destination lpni is set before we get here. */
1875
1876         /* find local NI */
1877         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1878         if (!sd->sd_best_ni) {
1879                 CERROR("Can't send to %s: src %s is not a "
1880                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1881                                 libcfs_nid2str(sd->sd_src_nid));
1882                 return -EINVAL;
1883         }
1884
1885         /*
1886          * the preferred NID will only be set for NMR peers
1887          */
1888         lnet_set_non_mr_pref_nid(sd);
1889
1890         return lnet_handle_send(sd);
1891 }
1892
1893 /*
1894  * Source Specified
1895  * Local Destination
1896  * MR Peer
1897  *
1898  * Don't run the selection algorithm on the peer NIs. By specifying the
1899  * local NID, we're also saying that we should always use the destination NID
1900  * provided. This handles the case where we should be using the same
1901  * destination NID for the all the messages which belong to the same RPC
1902  * request.
1903  */
1904 static int
1905 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1906 {
1907         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1908         if (!sd->sd_best_ni) {
1909                 CERROR("Can't send to %s: src %s is not a "
1910                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1911                                 libcfs_nid2str(sd->sd_src_nid));
1912                 return -EINVAL;
1913         }
1914
1915         if (sd->sd_best_lpni &&
1916             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1917                 return lnet_handle_lo_send(sd);
1918         else if (sd->sd_best_lpni)
1919                 return lnet_handle_send(sd);
1920
1921         CERROR("can't send to %s. no NI on %s\n",
1922                libcfs_nid2str(sd->sd_dst_nid),
1923                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1924
1925         return -EHOSTUNREACH;
1926 }
1927
1928 struct lnet_ni *
1929 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1930                               struct lnet_peer *peer,
1931                               struct lnet_peer_net *peer_net,
1932                               int cpt,
1933                               bool incr_seq)
1934 {
1935         struct lnet_net *local_net;
1936         struct lnet_ni *best_ni;
1937
1938         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1939         if (!local_net)
1940                 return NULL;
1941
1942         /*
1943          * Iterate through the NIs in this local Net and select
1944          * the NI to send from. The selection is determined by
1945          * these 3 criterion in the following priority:
1946          *      1. NUMA
1947          *      2. NI available credits
1948          *      3. Round Robin
1949          */
1950         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1951                                    peer, peer_net, cpt);
1952
1953         if (incr_seq && best_ni)
1954                 best_ni->ni_seq++;
1955
1956         return best_ni;
1957 }
1958
1959 static int
1960 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1961                              struct lnet_msg *msg, lnet_nid_t rtr_nid,
1962                              int cpt)
1963 {
1964         struct lnet_peer *peer;
1965         lnet_nid_t primary_nid;
1966         int rc;
1967
1968         lnet_peer_ni_addref_locked(lpni);
1969
1970         peer = lpni->lpni_peer_net->lpn_peer;
1971
1972         if (lnet_peer_gw_discovery(peer)) {
1973                 lnet_peer_ni_decref_locked(lpni);
1974                 return 0;
1975         }
1976
1977         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1978                 lnet_peer_ni_decref_locked(lpni);
1979                 return 0;
1980         }
1981
1982         rc = lnet_discover_peer_locked(lpni, cpt, false);
1983         if (rc) {
1984                 lnet_peer_ni_decref_locked(lpni);
1985                 return rc;
1986         }
1987         /* The peer may have changed. */
1988         peer = lpni->lpni_peer_net->lpn_peer;
1989         spin_lock(&peer->lp_lock);
1990         if (lnet_peer_is_uptodate_locked(peer)) {
1991                 spin_unlock(&peer->lp_lock);
1992                 lnet_peer_ni_decref_locked(lpni);
1993                 return 0;
1994         }
1995         /* queue message and return */
1996         msg->msg_rtr_nid_param = rtr_nid;
1997         msg->msg_sending = 0;
1998         msg->msg_txpeer = NULL;
1999         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2000         primary_nid = peer->lp_primary_nid;
2001         spin_unlock(&peer->lp_lock);
2002
2003         lnet_peer_ni_decref_locked(lpni);
2004
2005         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2006                 msg, libcfs_nid2str(primary_nid));
2007
2008         return LNET_DC_WAIT;
2009 }
2010
2011 static int
2012 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2013                              lnet_nid_t dst_nid,
2014                              struct lnet_peer_ni **gw_lpni,
2015                              struct lnet_peer **gw_peer)
2016 {
2017         int rc;
2018         __u32 local_lnet;
2019         struct lnet_peer *gw;
2020         struct lnet_peer *lp;
2021         struct lnet_peer_net *lpn;
2022         struct lnet_peer_net *best_lpn = NULL;
2023         struct lnet_remotenet *rnet, *best_rnet = NULL;
2024         struct lnet_route *best_route = NULL;
2025         struct lnet_route *last_route = NULL;
2026         struct lnet_peer_ni *lpni = NULL;
2027         struct lnet_peer_ni *gwni = NULL;
2028         lnet_nid_t src_nid = sd->sd_src_nid;
2029
2030         /* If a router nid was specified then we are replying to a GET or
2031          * sending an ACK. In this case we use the gateway associated with the
2032          * specified router nid.
2033          */
2034         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2035                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2036                 if (!gwni) {
2037                         CERROR("No peer NI for gateway %s\n",
2038                                libcfs_nid2str(sd->sd_rtr_nid));
2039                         return -EHOSTUNREACH;
2040                 }
2041                 gw = gwni->lpni_peer_net->lpn_peer;
2042                 lnet_peer_ni_decref_locked(gwni);
2043                 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2044         } else {
2045                 /* we've already looked up the initial lpni using dst_nid */
2046                 lpni = sd->sd_best_lpni;
2047                 /* the peer tree must be in existence */
2048                 LASSERT(lpni && lpni->lpni_peer_net &&
2049                         lpni->lpni_peer_net->lpn_peer);
2050                 lp = lpni->lpni_peer_net->lpn_peer;
2051
2052                 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2053                         /* is this remote network reachable?  */
2054                         rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2055                         if (!rnet)
2056                                 continue;
2057
2058                         if (!best_lpn) {
2059                                 best_lpn = lpn;
2060                                 best_rnet = rnet;
2061                         }
2062
2063                         if (best_lpn->lpn_seq <= lpn->lpn_seq)
2064                                 continue;
2065
2066                         best_lpn = lpn;
2067                         best_rnet = rnet;
2068                 }
2069
2070                 if (!best_lpn) {
2071                         CERROR("peer %s has no available nets\n",
2072                                libcfs_nid2str(sd->sd_dst_nid));
2073                         return -EHOSTUNREACH;
2074                 }
2075
2076                 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
2077                                                               sd->sd_dst_nid,
2078                                                               lp,
2079                                                               best_lpn->lpn_net_id);
2080                 if (!sd->sd_best_lpni) {
2081                         CERROR("peer %s down\n",
2082                                libcfs_nid2str(sd->sd_dst_nid));
2083                         return -EHOSTUNREACH;
2084                 }
2085
2086                 best_route = lnet_find_route_locked(best_rnet, &last_route,
2087                                                     &gwni);
2088                 if (!best_route) {
2089                         CERROR("no route to %s from %s\n",
2090                                libcfs_nid2str(dst_nid),
2091                                libcfs_nid2str(src_nid));
2092                         return -EHOSTUNREACH;
2093                 }
2094
2095                 if (!gwni) {
2096                         CERROR("Internal Error. Route expected to %s from %s\n",
2097                                libcfs_nid2str(dst_nid),
2098                                libcfs_nid2str(src_nid));
2099                         return -EFAULT;
2100                 }
2101
2102                 gw = best_route->lr_gateway;
2103                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2104                 local_lnet = best_route->lr_lnet;
2105
2106         }
2107
2108         /*
2109          * Discover this gateway if it hasn't already been discovered.
2110          * This means we might delay the message until discovery has
2111          * completed
2112          */
2113         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2114         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2115                                           sd->sd_cpt);
2116         if (rc)
2117                 return rc;
2118
2119         if (!sd->sd_best_ni)
2120                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2121                                         lnet_peer_get_net_locked(gw,
2122                                                                  local_lnet),
2123                                         sd->sd_md_cpt,
2124                                         true);
2125
2126         if (!sd->sd_best_ni) {
2127                 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2128                        libcfs_net2str(local_lnet),
2129                        libcfs_nid2str(sd->sd_src_nid));
2130                 return -EFAULT;
2131         }
2132
2133         *gw_lpni = gwni;
2134         *gw_peer = gw;
2135
2136         /*
2137          * increment the sequence numbers since now we're sure we're
2138          * going to use this path
2139          */
2140         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2141                 LASSERT(best_route && last_route);
2142                 best_route->lr_seq = last_route->lr_seq + 1;
2143                 best_lpn->lpn_seq++;
2144         }
2145
2146         return 0;
2147 }
2148
2149 /*
2150  * Handle two cases:
2151  *
2152  * Case 1:
2153  *  Source specified
2154  *  Remote destination
2155  *  Non-MR destination
2156  *
2157  * Case 2:
2158  *  Source specified
2159  *  Remote destination
2160  *  MR destination
2161  *
2162  * The handling of these two cases is similar. Even though the destination
2163  * can be MR or non-MR, we'll deal directly with the router.
2164  */
2165 static int
2166 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2167 {
2168         int rc;
2169         struct lnet_peer_ni *gw_lpni = NULL;
2170         struct lnet_peer *gw_peer = NULL;
2171
2172         /* find local NI */
2173         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2174         if (!sd->sd_best_ni) {
2175                 CERROR("Can't send to %s: src %s is not a "
2176                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2177                                 libcfs_nid2str(sd->sd_src_nid));
2178                 return -EINVAL;
2179         }
2180
2181         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2182                                      &gw_peer);
2183         if (rc)
2184                 return rc;
2185
2186         if (sd->sd_send_case & NMR_DST)
2187                 /*
2188                 * since the final destination is non-MR let's set its preferred
2189                 * NID before we send
2190                 */
2191                 lnet_set_non_mr_pref_nid(sd);
2192
2193         /*
2194          * We're going to send to the gw found so let's set its
2195          * info
2196          */
2197         sd->sd_peer = gw_peer;
2198         sd->sd_best_lpni = gw_lpni;
2199
2200         return lnet_handle_send(sd);
2201 }
2202
2203 struct lnet_ni *
2204 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2205                                bool discovery)
2206 {
2207         struct lnet_peer_net *peer_net = NULL;
2208         struct lnet_ni *best_ni = NULL;
2209
2210         /*
2211          * The peer can have multiple interfaces, some of them can be on
2212          * the local network and others on a routed network. We should
2213          * prefer the local network. However if the local network is not
2214          * available then we need to try the routed network
2215          */
2216
2217         /* go through all the peer nets and find the best_ni */
2218         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2219                 /*
2220                  * The peer's list of nets can contain non-local nets. We
2221                  * want to only examine the local ones.
2222                  */
2223                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2224                         continue;
2225                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2226                                                    peer_net, md_cpt, false);
2227
2228                 /*
2229                  * if this is a discovery message and lp_disc_net_id is
2230                  * specified then use that net to send the discovery on.
2231                  */
2232                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2233                     discovery)
2234                         break;
2235         }
2236
2237         if (best_ni)
2238                 /* increment sequence number so we can round robin */
2239                 best_ni->ni_seq++;
2240
2241         return best_ni;
2242 }
2243
2244 static struct lnet_ni *
2245 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2246 {
2247         struct lnet_ni *best_ni = NULL;
2248         struct lnet_peer_net *peer_net;
2249         struct lnet_peer *peer = sd->sd_peer;
2250         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2251         struct lnet_peer_ni *lpni;
2252         int cpt = sd->sd_cpt;
2253
2254         /*
2255          * We must use a consistent source address when sending to a
2256          * non-MR peer. However, a non-MR peer can have multiple NIDs
2257          * on multiple networks, and we may even need to talk to this
2258          * peer on multiple networks -- certain types of
2259          * load-balancing configuration do this.
2260          *
2261          * So we need to pick the NI the peer prefers for this
2262          * particular network.
2263          */
2264
2265         /* Get the target peer_ni */
2266         peer_net = lnet_peer_get_net_locked(peer,
2267                         LNET_NIDNET(best_lpni->lpni_nid));
2268         LASSERT(peer_net != NULL);
2269         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2270                                 lpni_peer_nis) {
2271                 if (lpni->lpni_pref_nnids == 0)
2272                         continue;
2273                 LASSERT(lpni->lpni_pref_nnids == 1);
2274                 best_ni = lnet_nid2ni_locked(
2275                                 lpni->lpni_pref.nid, cpt);
2276                 break;
2277         }
2278
2279         return best_ni;
2280 }
2281
2282 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2283 static int
2284 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2285 {
2286         struct lnet_ni *best_ni = NULL;
2287         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2288
2289         /*
2290          * We must use a consistent source address when sending to a
2291          * non-MR peer. However, a non-MR peer can have multiple NIDs
2292          * on multiple networks, and we may even need to talk to this
2293          * peer on multiple networks -- certain types of
2294          * load-balancing configuration do this.
2295          *
2296          * So we need to pick the NI the peer prefers for this
2297          * particular network.
2298          */
2299
2300         best_ni = lnet_find_existing_preferred_best_ni(sd);
2301
2302         /* if best_ni is still not set just pick one */
2303         if (!best_ni) {
2304                 best_ni =
2305                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2306                                                 sd->sd_best_lpni->lpni_peer_net,
2307                                                 sd->sd_md_cpt, true);
2308                 /* If there is no best_ni we don't have a route */
2309                 if (!best_ni) {
2310                         CERROR("no path to %s from net %s\n",
2311                                 libcfs_nid2str(best_lpni->lpni_nid),
2312                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2313                         return -EHOSTUNREACH;
2314                 }
2315         }
2316
2317         sd->sd_best_ni = best_ni;
2318
2319         /* Set preferred NI if necessary. */
2320         lnet_set_non_mr_pref_nid(sd);
2321
2322         return 0;
2323 }
2324
2325
2326 /*
2327  * Source not specified
2328  * Local destination
2329  * Non-MR Peer
2330  *
2331  * always use the same source NID for NMR peers
2332  * If we've talked to that peer before then we already have a preferred
2333  * source NI associated with it. Otherwise, we select a preferred local NI
2334  * and store it in the peer
2335  */
2336 static int
2337 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2338 {
2339         int rc;
2340
2341         /* sd->sd_best_lpni is already set to the final destination */
2342
2343         /*
2344          * At this point we should've created the peer ni and peer. If we
2345          * can't find it, then something went wrong. Instead of assert
2346          * output a relevant message and fail the send
2347          */
2348         if (!sd->sd_best_lpni) {
2349                 CERROR("Internal fault. Unable to send msg %s to %s. "
2350                        "NID not known\n",
2351                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2352                        libcfs_nid2str(sd->sd_dst_nid));
2353                 return -EFAULT;
2354         }
2355
2356         rc = lnet_select_preferred_best_ni(sd);
2357         if (!rc)
2358                 rc = lnet_handle_send(sd);
2359
2360         return rc;
2361 }
2362
2363 static int
2364 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2365 {
2366         /*
2367          * NOTE we've already handled the remote peer case. So we only
2368          * need to worry about the local case here.
2369          *
2370          * if we're sending a response, ACK or reply, we need to send it
2371          * to the destination NID given to us. At this point we already
2372          * have the peer_ni we're suppose to send to, so just find the
2373          * best_ni on the peer net and use that. Since we're sending to an
2374          * MR peer then we can just run the selection algorithm on our
2375          * local NIs and pick the best one.
2376          */
2377         if (sd->sd_send_case & SND_RESP) {
2378                 sd->sd_best_ni =
2379                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2380                                                 sd->sd_best_lpni->lpni_peer_net,
2381                                                 sd->sd_md_cpt, true);
2382
2383                 if (!sd->sd_best_ni) {
2384                         /*
2385                          * We're not going to deal with not able to send
2386                          * a response to the provided final destination
2387                          */
2388                         CERROR("Can't send response to %s. "
2389                                "No local NI available\n",
2390                                 libcfs_nid2str(sd->sd_dst_nid));
2391                         return -EHOSTUNREACH;
2392                 }
2393
2394                 return lnet_handle_send(sd);
2395         }
2396
2397         /*
2398          * If we get here that means we're sending a fresh request, PUT or
2399          * GET, so we need to run our standard selection algorithm.
2400          * First find the best local interface that's on any of the peer's
2401          * networks.
2402          */
2403         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2404                                         sd->sd_md_cpt,
2405                                         lnet_msg_discovery(sd->sd_msg));
2406         if (sd->sd_best_ni) {
2407                 sd->sd_best_lpni =
2408                   lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
2409                                              sd->sd_peer,
2410                                              sd->sd_best_ni->ni_net->net_id);
2411
2412                 /*
2413                  * if we're successful in selecting a peer_ni on the local
2414                  * network, then send to it. Otherwise fall through and
2415                  * try and see if we can reach it over another routed
2416                  * network
2417                  */
2418                 if (sd->sd_best_lpni &&
2419                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2420                         /*
2421                          * in case we initially started with a routed
2422                          * destination, let's reset to local
2423                          */
2424                         sd->sd_send_case &= ~REMOTE_DST;
2425                         sd->sd_send_case |= LOCAL_DST;
2426                         return lnet_handle_lo_send(sd);
2427                 } else if (sd->sd_best_lpni) {
2428                         /*
2429                          * in case we initially started with a routed
2430                          * destination, let's reset to local
2431                          */
2432                         sd->sd_send_case &= ~REMOTE_DST;
2433                         sd->sd_send_case |= LOCAL_DST;
2434                         return lnet_handle_send(sd);
2435                 }
2436
2437                 CERROR("Internal Error. Expected to have a best_lpni: "
2438                        "%s -> %s\n",
2439                        libcfs_nid2str(sd->sd_src_nid),
2440                        libcfs_nid2str(sd->sd_dst_nid));
2441
2442                 return -EFAULT;
2443         }
2444
2445         /*
2446          * Peer doesn't have a local network. Let's see if there is
2447          * a remote network we can reach it on.
2448          */
2449         return PASS_THROUGH;
2450 }
2451
2452 /*
2453  * Case 1:
2454  *      Source NID not specified
2455  *      Local destination
2456  *      MR peer
2457  *
2458  * Case 2:
2459  *      Source NID not speified
2460  *      Remote destination
2461  *      MR peer
2462  *
2463  * In both of these cases if we're sending a response, ACK or REPLY, then
2464  * we need to send to the destination NID provided.
2465  *
2466  * In the remote case let's deal with MR routers.
2467  *
2468  */
2469
2470 static int
2471 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2472 {
2473         int rc = 0;
2474         struct lnet_peer *gw_peer = NULL;
2475         struct lnet_peer_ni *gw_lpni = NULL;
2476
2477         /*
2478          * handle sending a response to a remote peer here so we don't
2479          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2480          */
2481         if (sd->sd_send_case & REMOTE_DST &&
2482             sd->sd_send_case & SND_RESP) {
2483                 struct lnet_peer_ni *gw;
2484                 struct lnet_peer *gw_peer;
2485
2486                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2487                                                   &gw_peer);
2488                 if (rc < 0) {
2489                         CERROR("Can't send response to %s. "
2490                                "No route available\n",
2491                                 libcfs_nid2str(sd->sd_dst_nid));
2492                         return -EHOSTUNREACH;
2493                 } else if (rc > 0) {
2494                         return rc;
2495                 }
2496
2497                 sd->sd_best_lpni = gw;
2498                 sd->sd_peer = gw_peer;
2499
2500                 return lnet_handle_send(sd);
2501         }
2502
2503         /*
2504          * Even though the NID for the peer might not be on a local network,
2505          * since the peer is MR there could be other interfaces on the
2506          * local network. In that case we'd still like to prefer the local
2507          * network over the routed network. If we're unable to do that
2508          * then we select the best router among the different routed networks,
2509          * and if the router is MR then we can deal with it as such.
2510          */
2511         rc = lnet_handle_any_mr_dsta(sd);
2512         if (rc != PASS_THROUGH)
2513                 return rc;
2514
2515         /*
2516          * Now that we must route to the destination, we must consider the
2517          * MR case, where the destination has multiple interfaces, some of
2518          * which we can route to and others we do not. For this reason we
2519          * need to select the destination which we can route to and if
2520          * there are multiple, we need to round robin.
2521          */
2522         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2523                                           &gw_peer);
2524         if (rc)
2525                 return rc;
2526
2527         sd->sd_send_case &= ~LOCAL_DST;
2528         sd->sd_send_case |= REMOTE_DST;
2529
2530         sd->sd_peer = gw_peer;
2531         sd->sd_best_lpni = gw_lpni;
2532
2533         return lnet_handle_send(sd);
2534 }
2535
2536 /*
2537  * Source not specified
2538  * Remote destination
2539  * Non-MR peer
2540  *
2541  * Must send to the specified peer NID using the same source NID that
2542  * we've used before. If it's the first time to talk to that peer then
2543  * find the source NI and assign it as preferred to that peer
2544  */
2545 static int
2546 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2547 {
2548         int rc;
2549         struct lnet_peer_ni *gw_lpni = NULL;
2550         struct lnet_peer *gw_peer = NULL;
2551
2552         /*
2553          * Let's set if we have a preferred NI to talk to this NMR peer
2554          */
2555         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2556
2557         /*
2558          * find the router and that'll find the best NI if we didn't find
2559          * it already.
2560          */
2561         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2562                                           &gw_peer);
2563         if (rc)
2564                 return rc;
2565
2566         /*
2567          * set the best_ni we've chosen as the preferred one for
2568          * this peer
2569          */
2570         lnet_set_non_mr_pref_nid(sd);
2571
2572         /* we'll be sending to the gw */
2573         sd->sd_best_lpni = gw_lpni;
2574         sd->sd_peer = gw_peer;
2575
2576         return lnet_handle_send(sd);
2577 }
2578
2579 static int
2580 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2581 {
2582         /*
2583          * turn off the SND_RESP bit.
2584          * It will be checked in the case handling
2585          */
2586         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2587
2588         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2589                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2590                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2591                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2592                 libcfs_nid2str(sd->sd_dst_nid),
2593                 (send_case & LOCAL_DST) ? "local" : "routed");
2594
2595         switch (send_case) {
2596         /*
2597          * For all cases where the source is specified, we should always
2598          * use the destination NID, whether it's an MR destination or not,
2599          * since we're continuing a series of related messages for the
2600          * same RPC
2601          */
2602         case SRC_SPEC_LOCAL_NMR_DST:
2603                 return lnet_handle_spec_local_nmr_dst(sd);
2604         case SRC_SPEC_LOCAL_MR_DST:
2605                 return lnet_handle_spec_local_mr_dst(sd);
2606         case SRC_SPEC_ROUTER_NMR_DST:
2607         case SRC_SPEC_ROUTER_MR_DST:
2608                 return lnet_handle_spec_router_dst(sd);
2609         case SRC_ANY_LOCAL_NMR_DST:
2610                 return lnet_handle_any_local_nmr_dst(sd);
2611         case SRC_ANY_LOCAL_MR_DST:
2612         case SRC_ANY_ROUTER_MR_DST:
2613                 return lnet_handle_any_mr_dst(sd);
2614         case SRC_ANY_ROUTER_NMR_DST:
2615                 return lnet_handle_any_router_nmr_dst(sd);
2616         default:
2617                 CERROR("Unknown send case\n");
2618                 return -1;
2619         }
2620 }
2621
2622 static int
2623 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2624                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2625 {
2626         struct lnet_peer_ni     *lpni;
2627         struct lnet_peer        *peer;
2628         struct lnet_send_data   send_data;
2629         int                     cpt, rc;
2630         int                     md_cpt;
2631         __u32                   send_case = 0;
2632
2633         memset(&send_data, 0, sizeof(send_data));
2634
2635         /*
2636          * get an initial CPT to use for locking. The idea here is not to
2637          * serialize the calls to select_pathway, so that as many
2638          * operations can run concurrently as possible. To do that we use
2639          * the CPT where this call is being executed. Later on when we
2640          * determine the CPT to use in lnet_message_commit, we switch the
2641          * lock and check if there was any configuration change.  If none,
2642          * then we proceed, if there is, then we restart the operation.
2643          */
2644         cpt = lnet_net_lock_current();
2645
2646         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2647         if (md_cpt == CFS_CPT_ANY)
2648                 md_cpt = cpt;
2649
2650 again:
2651
2652         /*
2653          * If we're sending to ourselves then there is no need to go through
2654          * any selection. We can shortcut the entire process and send over
2655          * lolnd.
2656          *
2657          * However, we make two exceptions to this rule:
2658          * 1. If the src_nid is specified then our API defines that we must send
2659          *    via that interface.
2660          * 2. Recovery messages must be sent to the lnet_ni that is being
2661          *    recovered.
2662          */
2663         send_data.sd_msg = msg;
2664         send_data.sd_cpt = cpt;
2665         if (src_nid == LNET_NID_ANY && !msg->msg_recovery &&
2666             lnet_nid2ni_locked(dst_nid, cpt)) {
2667                 rc = lnet_handle_lo_send(&send_data);
2668                 lnet_net_unlock(cpt);
2669                 return rc;
2670         }
2671
2672         /*
2673          * find an existing peer_ni, or create one and mark it as having been
2674          * created due to network traffic. This call will create the
2675          * peer->peer_net->peer_ni tree.
2676          */
2677         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2678         if (IS_ERR(lpni)) {
2679                 lnet_net_unlock(cpt);
2680                 return PTR_ERR(lpni);
2681         }
2682
2683         /*
2684          * Cache the original src_nid. If we need to resend the message
2685          * then we'll need to know whether the src_nid was originally
2686          * specified for this message. If it was originally specified,
2687          * then we need to keep using the same src_nid since it's
2688          * continuing the same sequence of messages.
2689          */
2690         msg->msg_src_nid_param = src_nid;
2691
2692         /*
2693          * If necessary, perform discovery on the peer that owns this peer_ni.
2694          * Note, this can result in the ownership of this peer_ni changing
2695          * to another peer object.
2696          */
2697         rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2698         if (rc) {
2699                 lnet_peer_ni_decref_locked(lpni);
2700                 lnet_net_unlock(cpt);
2701                 return rc;
2702         }
2703         lnet_peer_ni_decref_locked(lpni);
2704
2705         peer = lpni->lpni_peer_net->lpn_peer;
2706
2707         /*
2708          * Identify the different send cases
2709          */
2710         if (src_nid == LNET_NID_ANY)
2711                 send_case |= SRC_ANY;
2712         else
2713                 send_case |= SRC_SPEC;
2714
2715         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2716                 send_case |= LOCAL_DST;
2717         else
2718                 send_case |= REMOTE_DST;
2719
2720         /*
2721          * if this is a non-MR peer or if we're recovering a peer ni then
2722          * let's consider this an NMR case so we can hit the destination
2723          * NID.
2724          */
2725         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2726                 send_case |= NMR_DST;
2727         else
2728                 send_case |= MR_DST;
2729
2730         if (msg->msg_type == LNET_MSG_REPLY ||
2731             msg->msg_type == LNET_MSG_ACK)
2732                 send_case |= SND_RESP;
2733
2734         /* assign parameters to the send_data */
2735         send_data.sd_rtr_nid = rtr_nid;
2736         send_data.sd_src_nid = src_nid;
2737         send_data.sd_dst_nid = dst_nid;
2738         send_data.sd_best_lpni = lpni;
2739         /*
2740          * keep a pointer to the final destination in case we're going to
2741          * route, so we'll need to access it later
2742          */
2743         send_data.sd_final_dst_lpni = lpni;
2744         send_data.sd_peer = peer;
2745         send_data.sd_md_cpt = md_cpt;
2746         send_data.sd_send_case = send_case;
2747
2748         rc = lnet_handle_send_case_locked(&send_data);
2749
2750         /*
2751          * Update the local cpt since send_data.sd_cpt might've been
2752          * updated as a result of calling lnet_handle_send_case_locked().
2753          */
2754         cpt = send_data.sd_cpt;
2755
2756         if (rc == REPEAT_SEND)
2757                 goto again;
2758
2759         lnet_net_unlock(cpt);
2760
2761         return rc;
2762 }
2763
2764 int
2765 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2766 {
2767         lnet_nid_t              dst_nid = msg->msg_target.nid;
2768         int                     rc;
2769
2770         /*
2771          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2772          * but we might want to use pre-determined router for ACK/REPLY
2773          * in the future
2774          */
2775         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2776         LASSERT(msg->msg_txpeer == NULL);
2777         LASSERT(msg->msg_txni == NULL);
2778         LASSERT(!msg->msg_sending);
2779         LASSERT(!msg->msg_target_is_router);
2780         LASSERT(!msg->msg_receiving);
2781
2782         msg->msg_sending = 1;
2783
2784         LASSERT(!msg->msg_tx_committed);
2785
2786         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2787         if (rc < 0) {
2788                 if (rc == -EHOSTUNREACH)
2789                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2790                 else
2791                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2792                 return rc;
2793         }
2794
2795         if (rc == LNET_CREDIT_OK)
2796                 lnet_ni_send(msg->msg_txni, msg);
2797
2798         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2799         return 0;
2800 }
2801
2802 enum lnet_mt_event_type {
2803         MT_TYPE_LOCAL_NI = 0,
2804         MT_TYPE_PEER_NI
2805 };
2806
2807 struct lnet_mt_event_info {
2808         enum lnet_mt_event_type mt_type;
2809         lnet_nid_t mt_nid;
2810 };
2811
2812 /* called with res_lock held */
2813 void
2814 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2815 {
2816         struct lnet_rsp_tracker *rspt;
2817
2818         /*
2819          * msg has a refcount on the MD so the MD is not going away.
2820          * The rspt queue for the cpt is protected by
2821          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2822          */
2823         if (!md->md_rspt_ptr)
2824                 return;
2825
2826         rspt = md->md_rspt_ptr;
2827
2828         /* debug code */
2829         LASSERT(rspt->rspt_cpt == cpt);
2830
2831         md->md_rspt_ptr = NULL;
2832
2833         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2834                 /*
2835                  * The monitor thread has invalidated this handle because the
2836                  * response timed out, but it failed to lookup the MD. That
2837                  * means this response tracker is on the zombie list. We can
2838                  * safely remove it under the resource lock (held by caller) and
2839                  * free the response tracker block.
2840                  */
2841                 list_del(&rspt->rspt_on_list);
2842                 lnet_rspt_free(rspt, cpt);
2843         } else {
2844                 /*
2845                  * invalidate the handle to indicate that a response has been
2846                  * received, which will then lead the monitor thread to clean up
2847                  * the rspt block.
2848                  */
2849                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2850         }
2851 }
2852
2853 void
2854 lnet_clean_zombie_rstqs(void)
2855 {
2856         struct lnet_rsp_tracker *rspt, *tmp;
2857         int i;
2858
2859         cfs_cpt_for_each(i, lnet_cpt_table()) {
2860                 list_for_each_entry_safe(rspt, tmp,
2861                                          the_lnet.ln_mt_zombie_rstqs[i],
2862                                          rspt_on_list) {
2863                         list_del(&rspt->rspt_on_list);
2864                         lnet_rspt_free(rspt, i);
2865                 }
2866         }
2867
2868         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2869 }
2870
2871 static void
2872 lnet_finalize_expired_responses(void)
2873 {
2874         struct lnet_libmd *md;
2875         struct list_head local_queue;
2876         struct lnet_rsp_tracker *rspt, *tmp;
2877         ktime_t now;
2878         int i;
2879
2880         if (the_lnet.ln_mt_rstq == NULL)
2881                 return;
2882
2883         cfs_cpt_for_each(i, lnet_cpt_table()) {
2884                 INIT_LIST_HEAD(&local_queue);
2885
2886                 lnet_net_lock(i);
2887                 if (!the_lnet.ln_mt_rstq[i]) {
2888                         lnet_net_unlock(i);
2889                         continue;
2890                 }
2891                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2892                 lnet_net_unlock(i);
2893
2894                 now = ktime_get();
2895
2896                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2897                         /*
2898                          * The rspt mdh will be invalidated when a response
2899                          * is received or whenever we want to discard the
2900                          * block the monitor thread will walk the queue
2901                          * and clean up any rsts with an invalid mdh.
2902                          * The monitor thread will walk the queue until
2903                          * the first unexpired rspt block. This means that
2904                          * some rspt blocks which received their
2905                          * corresponding responses will linger in the
2906                          * queue until they are cleaned up eventually.
2907                          */
2908                         lnet_res_lock(i);
2909                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2910                                 lnet_res_unlock(i);
2911                                 list_del(&rspt->rspt_on_list);
2912                                 lnet_rspt_free(rspt, i);
2913                                 continue;
2914                         }
2915
2916                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2917                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2918                                 struct lnet_peer_ni *lpni;
2919                                 lnet_nid_t nid;
2920
2921                                 md = lnet_handle2md(&rspt->rspt_mdh);
2922                                 if (!md) {
2923                                         /* MD has been queued for unlink, but
2924                                          * rspt hasn't been detached (Note we've
2925                                          * checked above that the rspt_mdh is
2926                                          * valid). Since we cannot lookup the MD
2927                                          * we're unable to detach the rspt
2928                                          * ourselves. Thus, move the rspt to the
2929                                          * zombie list where we'll wait for
2930                                          * either:
2931                                          *   1. The remaining operations on the
2932                                          *   MD to complete. In this case the
2933                                          *   final operation will result in
2934                                          *   lnet_msg_detach_md()->
2935                                          *   lnet_detach_rsp_tracker() where
2936                                          *   we will clean up this response
2937                                          *   tracker.
2938                                          *   2. LNet to shutdown. In this case
2939                                          *   we'll wait until after all LND Nets
2940                                          *   have shutdown and then we can
2941                                          *   safely free any remaining response
2942                                          *   tracker blocks on the zombie list.
2943                                          * Note: We need to hold the resource
2944                                          * lock when adding to the zombie list
2945                                          * because we may have concurrent access
2946                                          * with lnet_detach_rsp_tracker().
2947                                          */
2948                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2949                                         list_move(&rspt->rspt_on_list,
2950                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2951                                         lnet_res_unlock(i);
2952                                         continue;
2953                                 }
2954                                 LASSERT(md->md_rspt_ptr == rspt);
2955                                 md->md_rspt_ptr = NULL;
2956                                 lnet_res_unlock(i);
2957
2958                                 LNetMDUnlink(rspt->rspt_mdh);
2959
2960                                 nid = rspt->rspt_next_hop_nid;
2961
2962                                 list_del(&rspt->rspt_on_list);
2963                                 lnet_rspt_free(rspt, i);
2964
2965                                 /* If we're shutting down we just want to clean
2966                                  * up the rspt blocks
2967                                  */
2968                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2969                                         continue;
2970
2971                                 lnet_net_lock(i);
2972                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2973                                 lnet_net_unlock(i);
2974
2975                                 CDEBUG(D_NET,
2976                                        "Response timeout: md = %p: nid = %s\n",
2977                                        md, libcfs_nid2str(nid));
2978
2979                                 /*
2980                                  * If there is a timeout on the response
2981                                  * from the next hop decrement its health
2982                                  * value so that we don't use it
2983                                  */
2984                                 lnet_net_lock(0);
2985                                 lpni = lnet_find_peer_ni_locked(nid);
2986                                 if (lpni) {
2987                                         lnet_handle_remote_failure_locked(lpni);
2988                                         lnet_peer_ni_decref_locked(lpni);
2989                                 }
2990                                 lnet_net_unlock(0);
2991                         } else {
2992                                 lnet_res_unlock(i);
2993                                 break;
2994                         }
2995                 }
2996
2997                 if (!list_empty(&local_queue)) {
2998                         lnet_net_lock(i);
2999                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3000                         lnet_net_unlock(i);
3001                 }
3002         }
3003 }
3004
3005 static void
3006 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3007 {
3008         struct lnet_msg *msg;
3009
3010         while (!list_empty(resendq)) {
3011                 struct lnet_peer_ni *lpni;
3012
3013                 msg = list_entry(resendq->next, struct lnet_msg,
3014                                  msg_list);
3015
3016                 list_del_init(&msg->msg_list);
3017
3018                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3019                 if (!lpni) {
3020                         lnet_net_unlock(cpt);
3021                         CERROR("Expected that a peer is already created for %s\n",
3022                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3023                         msg->msg_no_resend = true;
3024                         lnet_finalize(msg, -EFAULT);
3025                         lnet_net_lock(cpt);
3026                 } else {
3027                         struct lnet_peer *peer;
3028                         int rc;
3029                         lnet_nid_t src_nid = LNET_NID_ANY;
3030
3031                         /*
3032                          * if this message is not being routed and the
3033                          * peer is non-MR then we must use the same
3034                          * src_nid that was used in the original send.
3035                          * Otherwise if we're routing the message (IE
3036                          * we're a router) then we can use any of our
3037                          * local interfaces. It doesn't matter to the
3038                          * final destination.
3039                          */
3040                         peer = lpni->lpni_peer_net->lpn_peer;
3041                         if (!msg->msg_routing &&
3042                             !lnet_peer_is_multi_rail(peer))
3043                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
3044
3045                         /*
3046                          * If we originally specified a src NID, then we
3047                          * must attempt to reuse it in the resend as well.
3048                          */
3049                         if (msg->msg_src_nid_param != LNET_NID_ANY)
3050                                 src_nid = msg->msg_src_nid_param;
3051                         lnet_peer_ni_decref_locked(lpni);
3052
3053                         lnet_net_unlock(cpt);
3054                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3055                                libcfs_nid2str(src_nid),
3056                                libcfs_id2str(msg->msg_target),
3057                                lnet_msgtyp2str(msg->msg_type),
3058                                msg->msg_recovery,
3059                                msg->msg_retry_count);
3060                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
3061                         if (rc) {
3062                                 CERROR("Error sending %s to %s: %d\n",
3063                                        lnet_msgtyp2str(msg->msg_type),
3064                                        libcfs_id2str(msg->msg_target), rc);
3065                                 msg->msg_no_resend = true;
3066                                 lnet_finalize(msg, rc);
3067                         }
3068                         lnet_net_lock(cpt);
3069                         if (!rc)
3070                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3071                 }
3072         }
3073 }
3074
3075 static void
3076 lnet_resend_pending_msgs(void)
3077 {
3078         int i;
3079
3080         cfs_cpt_for_each(i, lnet_cpt_table()) {
3081                 lnet_net_lock(i);
3082                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3083                 lnet_net_unlock(i);
3084         }
3085 }
3086
3087 /* called with cpt and ni_lock held */
3088 static void
3089 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3090 {
3091         struct lnet_handle_md recovery_mdh;
3092
3093         LNetInvalidateMDHandle(&recovery_mdh);
3094
3095         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3096             force) {
3097                 recovery_mdh = ni->ni_ping_mdh;
3098                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3099         }
3100         lnet_ni_unlock(ni);
3101         lnet_net_unlock(cpt);
3102         if (!LNetMDHandleIsInvalid(recovery_mdh))
3103                 LNetMDUnlink(recovery_mdh);
3104         lnet_net_lock(cpt);
3105         lnet_ni_lock(ni);
3106 }
3107
3108 static void
3109 lnet_recover_local_nis(void)
3110 {
3111         struct lnet_mt_event_info *ev_info;
3112         struct list_head processed_list;
3113         struct list_head local_queue;
3114         struct lnet_handle_md mdh;
3115         struct lnet_ni *tmp;
3116         struct lnet_ni *ni;
3117         lnet_nid_t nid;
3118         int healthv;
3119         int rc;
3120
3121         INIT_LIST_HEAD(&local_queue);
3122         INIT_LIST_HEAD(&processed_list);
3123
3124         /*
3125          * splice the recovery queue on a local queue. We will iterate
3126          * through the local queue and update it as needed. Once we're
3127          * done with the traversal, we'll splice the local queue back on
3128          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3129          * will be traversed in the next iteration.
3130          */
3131         lnet_net_lock(0);
3132         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3133                          &local_queue);
3134         lnet_net_unlock(0);
3135
3136         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3137                 /*
3138                  * if an NI is being deleted or it is now healthy, there
3139                  * is no need to keep it around in the recovery queue.
3140                  * The monitor thread is the only thread responsible for
3141                  * removing the NI from the recovery queue.
3142                  * Multiple threads can be adding NIs to the recovery
3143                  * queue.
3144                  */
3145                 healthv = atomic_read(&ni->ni_healthv);
3146
3147                 lnet_net_lock(0);
3148                 lnet_ni_lock(ni);
3149                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3150                     healthv == LNET_MAX_HEALTH_VALUE) {
3151                         list_del_init(&ni->ni_recovery);
3152                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3153                         lnet_ni_unlock(ni);
3154                         lnet_ni_decref_locked(ni, 0);
3155                         lnet_net_unlock(0);
3156                         continue;
3157                 }
3158
3159                 /*
3160                  * if the local NI failed recovery we must unlink the md.
3161                  * But we want to keep the local_ni on the recovery queue
3162                  * so we can continue the attempts to recover it.
3163                  */
3164                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3165                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3166                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3167                 }
3168
3169                 lnet_ni_unlock(ni);
3170                 lnet_net_unlock(0);
3171
3172
3173                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3174                        libcfs_nid2str(ni->ni_nid));
3175
3176                 lnet_ni_lock(ni);
3177                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3178                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3179                         lnet_ni_unlock(ni);
3180
3181                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3182                         if (!ev_info) {
3183                                 CERROR("out of memory. Can't recover %s\n",
3184                                        libcfs_nid2str(ni->ni_nid));
3185                                 lnet_ni_lock(ni);
3186                                 ni->ni_recovery_state &=
3187                                   ~LNET_NI_RECOVERY_PENDING;
3188                                 lnet_ni_unlock(ni);
3189                                 continue;
3190                         }
3191
3192                         mdh = ni->ni_ping_mdh;
3193                         /*
3194                          * Invalidate the ni mdh in case it's deleted.
3195                          * We'll unlink the mdh in this case below.
3196                          */
3197                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3198                         nid = ni->ni_nid;
3199
3200                         /*
3201                          * remove the NI from the local queue and drop the
3202                          * reference count to it while we're recovering
3203                          * it. The reason for that, is that the NI could
3204                          * be deleted, and the way the code is structured
3205                          * is if we don't drop the NI, then the deletion
3206                          * code will enter a loop waiting for the
3207                          * reference count to be removed while holding the
3208                          * ln_mutex_lock(). When we look up the peer to
3209                          * send to in lnet_select_pathway() we will try to
3210                          * lock the ln_mutex_lock() as well, leading to
3211                          * a deadlock. By dropping the refcount and
3212                          * removing it from the list, we allow for the NI
3213                          * to be removed, then we use the cached NID to
3214                          * look it up again. If it's gone, then we just
3215                          * continue examining the rest of the queue.
3216                          */
3217                         lnet_net_lock(0);
3218                         list_del_init(&ni->ni_recovery);
3219                         lnet_ni_decref_locked(ni, 0);
3220                         lnet_net_unlock(0);
3221
3222                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3223                         ev_info->mt_nid = nid;
3224                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3225                                             ev_info, the_lnet.ln_mt_eqh, true);
3226                         /* lookup the nid again */
3227                         lnet_net_lock(0);
3228                         ni = lnet_nid2ni_locked(nid, 0);
3229                         if (!ni) {
3230                                 /*
3231                                  * the NI has been deleted when we dropped
3232                                  * the ref count
3233                                  */
3234                                 lnet_net_unlock(0);
3235                                 LNetMDUnlink(mdh);
3236                                 continue;
3237                         }
3238                         /*
3239                          * Same note as in lnet_recover_peer_nis(). When
3240                          * we're sending the ping, the NI is free to be
3241                          * deleted or manipulated. By this point it
3242                          * could've been added back on the recovery queue,
3243                          * and a refcount taken on it.
3244                          * So we can't just add it blindly again or we'll
3245                          * corrupt the queue. We must check under lock if
3246                          * it's not on any list and if not then add it
3247                          * to the processed list, which will eventually be
3248                          * spliced back on to the recovery queue.
3249                          */
3250                         ni->ni_ping_mdh = mdh;
3251                         if (list_empty(&ni->ni_recovery)) {
3252                                 list_add_tail(&ni->ni_recovery, &processed_list);
3253                                 lnet_ni_addref_locked(ni, 0);
3254                         }
3255                         lnet_net_unlock(0);
3256
3257                         lnet_ni_lock(ni);
3258                         if (rc)
3259                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3260                 }
3261                 lnet_ni_unlock(ni);
3262         }
3263
3264         /*
3265          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3266          * reexamined in the next iteration.
3267          */
3268         list_splice_init(&processed_list, &local_queue);
3269         lnet_net_lock(0);
3270         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3271         lnet_net_unlock(0);
3272 }
3273
3274 static int
3275 lnet_resendqs_create(void)
3276 {
3277         struct list_head **resendqs;
3278         resendqs = lnet_create_array_of_queues();
3279
3280         if (!resendqs)
3281                 return -ENOMEM;
3282
3283         lnet_net_lock(LNET_LOCK_EX);
3284         the_lnet.ln_mt_resendqs = resendqs;
3285         lnet_net_unlock(LNET_LOCK_EX);
3286
3287         return 0;
3288 }
3289
3290 static void
3291 lnet_clean_local_ni_recoveryq(void)
3292 {
3293         struct lnet_ni *ni;
3294
3295         /* This is only called when the monitor thread has stopped */
3296         lnet_net_lock(0);
3297
3298         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3299                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3300                                 struct lnet_ni, ni_recovery);
3301                 list_del_init(&ni->ni_recovery);
3302                 lnet_ni_lock(ni);
3303                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3304                 lnet_ni_unlock(ni);
3305                 lnet_ni_decref_locked(ni, 0);
3306         }
3307
3308         lnet_net_unlock(0);
3309 }
3310
3311 static void
3312 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3313                                      bool force)
3314 {
3315         struct lnet_handle_md recovery_mdh;
3316
3317         LNetInvalidateMDHandle(&recovery_mdh);
3318
3319         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3320                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3321                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3322         }
3323         spin_unlock(&lpni->lpni_lock);
3324         lnet_net_unlock(cpt);
3325         if (!LNetMDHandleIsInvalid(recovery_mdh))
3326                 LNetMDUnlink(recovery_mdh);
3327         lnet_net_lock(cpt);
3328         spin_lock(&lpni->lpni_lock);
3329 }
3330
3331 static void
3332 lnet_clean_peer_ni_recoveryq(void)
3333 {
3334         struct lnet_peer_ni *lpni, *tmp;
3335
3336         lnet_net_lock(LNET_LOCK_EX);
3337
3338         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3339                                  lpni_recovery) {
3340                 list_del_init(&lpni->lpni_recovery);
3341                 spin_lock(&lpni->lpni_lock);
3342                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3343                 spin_unlock(&lpni->lpni_lock);
3344                 lnet_peer_ni_decref_locked(lpni);
3345         }
3346
3347         lnet_net_unlock(LNET_LOCK_EX);
3348 }
3349
3350 static void
3351 lnet_clean_resendqs(void)
3352 {
3353         struct lnet_msg *msg, *tmp;
3354         struct list_head msgs;
3355         int i;
3356
3357         INIT_LIST_HEAD(&msgs);
3358
3359         cfs_cpt_for_each(i, lnet_cpt_table()) {
3360                 lnet_net_lock(i);
3361                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3362                 lnet_net_unlock(i);
3363                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3364                         list_del_init(&msg->msg_list);
3365                         msg->msg_no_resend = true;
3366                         lnet_finalize(msg, -ESHUTDOWN);
3367                 }
3368         }
3369
3370         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3371 }
3372
3373 static void
3374 lnet_recover_peer_nis(void)
3375 {
3376         struct lnet_mt_event_info *ev_info;
3377         struct list_head processed_list;
3378         struct list_head local_queue;
3379         struct lnet_handle_md mdh;
3380         struct lnet_peer_ni *lpni;
3381         struct lnet_peer_ni *tmp;
3382         lnet_nid_t nid;
3383         int healthv;
3384         int rc;
3385
3386         INIT_LIST_HEAD(&local_queue);
3387         INIT_LIST_HEAD(&processed_list);
3388
3389         /*
3390          * Always use cpt 0 for locking across all interactions with
3391          * ln_mt_peerNIRecovq
3392          */
3393         lnet_net_lock(0);
3394         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3395                          &local_queue);
3396         lnet_net_unlock(0);
3397
3398         list_for_each_entry_safe(lpni, tmp, &local_queue,
3399                                  lpni_recovery) {
3400                 /*
3401                  * The same protection strategy is used here as is in the
3402                  * local recovery case.
3403                  */
3404                 lnet_net_lock(0);
3405                 healthv = atomic_read(&lpni->lpni_healthv);
3406                 spin_lock(&lpni->lpni_lock);
3407                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3408                     healthv == LNET_MAX_HEALTH_VALUE) {
3409                         list_del_init(&lpni->lpni_recovery);
3410                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3411                         spin_unlock(&lpni->lpni_lock);
3412                         lnet_peer_ni_decref_locked(lpni);
3413                         lnet_net_unlock(0);
3414                         continue;
3415                 }
3416
3417                 /*
3418                  * If the peer NI has failed recovery we must unlink the
3419                  * md. But we want to keep the peer ni on the recovery
3420                  * queue so we can try to continue recovering it
3421                  */
3422                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3423                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3424                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3425                 }
3426
3427                 spin_unlock(&lpni->lpni_lock);
3428                 lnet_net_unlock(0);
3429
3430                 /*
3431                  * NOTE: we're racing with peer deletion from user space.
3432                  * It's possible that a peer is deleted after we check its
3433                  * state. In this case the recovery can create a new peer
3434                  */
3435                 spin_lock(&lpni->lpni_lock);
3436                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3437                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3438                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3439                         spin_unlock(&lpni->lpni_lock);
3440
3441                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3442                         if (!ev_info) {
3443                                 CERROR("out of memory. Can't recover %s\n",
3444                                        libcfs_nid2str(lpni->lpni_nid));
3445                                 spin_lock(&lpni->lpni_lock);
3446                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3447                                 spin_unlock(&lpni->lpni_lock);
3448                                 continue;
3449                         }
3450
3451                         /* look at the comments in lnet_recover_local_nis() */
3452                         mdh = lpni->lpni_recovery_ping_mdh;
3453                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3454                         nid = lpni->lpni_nid;
3455                         lnet_net_lock(0);
3456                         list_del_init(&lpni->lpni_recovery);
3457                         lnet_peer_ni_decref_locked(lpni);
3458                         lnet_net_unlock(0);
3459
3460                         ev_info->mt_type = MT_TYPE_PEER_NI;
3461                         ev_info->mt_nid = nid;
3462                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3463                                             ev_info, the_lnet.ln_mt_eqh, true);
3464                         lnet_net_lock(0);
3465                         /*
3466                          * lnet_find_peer_ni_locked() grabs a refcount for
3467                          * us. No need to take it explicitly.
3468                          */
3469                         lpni = lnet_find_peer_ni_locked(nid);
3470                         if (!lpni) {
3471                                 lnet_net_unlock(0);
3472                                 LNetMDUnlink(mdh);
3473                                 continue;
3474                         }
3475
3476                         lpni->lpni_recovery_ping_mdh = mdh;
3477                         /*
3478                          * While we're unlocked the lpni could've been
3479                          * readded on the recovery queue. In this case we
3480                          * don't need to add it to the local queue, since
3481                          * it's already on there and the thread that added
3482                          * it would've incremented the refcount on the
3483                          * peer, which means we need to decref the refcount
3484                          * that was implicitly grabbed by find_peer_ni_locked.
3485                          * Otherwise, if the lpni is still not on
3486                          * the recovery queue, then we'll add it to the
3487                          * processed list.
3488                          */
3489                         if (list_empty(&lpni->lpni_recovery))
3490                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3491                         else
3492                                 lnet_peer_ni_decref_locked(lpni);
3493                         lnet_net_unlock(0);
3494
3495                         spin_lock(&lpni->lpni_lock);
3496                         if (rc)
3497                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3498                 }
3499                 spin_unlock(&lpni->lpni_lock);
3500         }
3501
3502         list_splice_init(&processed_list, &local_queue);
3503         lnet_net_lock(0);
3504         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3505         lnet_net_unlock(0);
3506 }
3507
3508 static int
3509 lnet_monitor_thread(void *arg)
3510 {
3511         time64_t recovery_timeout = 0;
3512         time64_t rsp_timeout = 0;
3513         int interval;
3514         time64_t now;
3515
3516         wait_for_completion(&the_lnet.ln_started);
3517         /*
3518          * The monitor thread takes care of the following:
3519          *  1. Checks the aliveness of routers
3520          *  2. Checks if there are messages on the resend queue to resend
3521          *     them.
3522          *  3. Check if there are any NIs on the local recovery queue and
3523          *     pings them
3524          *  4. Checks if there are any NIs on the remote recovery queue
3525          *     and pings them.
3526          */
3527         cfs_block_allsigs();
3528
3529         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3530                 now = ktime_get_real_seconds();
3531
3532                 if (lnet_router_checker_active())
3533                         lnet_check_routers();
3534
3535                 lnet_resend_pending_msgs();
3536
3537                 if (now >= rsp_timeout) {
3538                         lnet_finalize_expired_responses();
3539                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3540                 }
3541
3542                 if (now >= recovery_timeout) {
3543                         lnet_recover_local_nis();
3544                         lnet_recover_peer_nis();
3545                         recovery_timeout = now + lnet_recovery_interval;
3546                 }
3547
3548                 /*
3549                  * TODO do we need to check if we should sleep without
3550                  * timeout?  Technically, an active system will always
3551                  * have messages in flight so this check will always
3552                  * evaluate to false. And on an idle system do we care
3553                  * if we wake up every 1 second? Although, we've seen
3554                  * cases where we get a complaint that an idle thread
3555                  * is waking up unnecessarily.
3556                  *
3557                  * Take into account the current net_count when you wake
3558                  * up for alive router checking, since we need to check
3559                  * possibly as many networks as we have configured.
3560                  */
3561                 interval = min(lnet_recovery_interval,
3562                                min((unsigned int) alive_router_check_interval /
3563                                         lnet_current_net_count,
3564                                    lnet_transaction_timeout / 2));
3565                 wait_for_completion_interruptible_timeout(
3566                         &the_lnet.ln_mt_wait_complete,
3567                         cfs_time_seconds(interval));
3568                 /* Must re-init the completion before testing anything,
3569                  * including ln_mt_state.
3570                  */
3571                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3572         }
3573
3574         /* Shutting down */
3575         lnet_net_lock(LNET_LOCK_EX);
3576         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3577         lnet_net_unlock(LNET_LOCK_EX);
3578
3579         /* signal that the monitor thread is exiting */
3580         up(&the_lnet.ln_mt_signal);
3581
3582         return 0;
3583 }
3584
3585 /*
3586  * lnet_send_ping
3587  * Sends a ping.
3588  * Returns == 0 if success
3589  * Returns > 0 if LNetMDBind or prior fails
3590  * Returns < 0 if LNetGet fails
3591  */
3592 int
3593 lnet_send_ping(lnet_nid_t dest_nid,
3594                struct lnet_handle_md *mdh, int nnis,
3595                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3596 {
3597         struct lnet_md md = { NULL };
3598         struct lnet_process_id id;
3599         struct lnet_ping_buffer *pbuf;
3600         int rc;
3601
3602         if (dest_nid == LNET_NID_ANY) {
3603                 rc = -EHOSTUNREACH;
3604                 goto fail_error;
3605         }
3606
3607         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3608         if (!pbuf) {
3609                 rc = ENOMEM;
3610                 goto fail_error;
3611         }
3612
3613         /* initialize md content */
3614         md.start     = &pbuf->pb_info;
3615         md.length    = LNET_PING_INFO_SIZE(nnis);
3616         md.threshold = 2; /* GET/REPLY */
3617         md.max_size  = 0;
3618         md.options   = LNET_MD_TRUNCATE;
3619         md.user_ptr  = user_data;
3620         md.eq_handle = eqh;
3621
3622         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3623         if (rc) {
3624                 lnet_ping_buffer_decref(pbuf);
3625                 CERROR("Can't bind MD: %d\n", rc);
3626                 rc = -rc; /* change the rc to positive */
3627                 goto fail_error;
3628         }
3629         id.pid = LNET_PID_LUSTRE;
3630         id.nid = dest_nid;
3631
3632         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3633                      LNET_RESERVED_PORTAL,
3634                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3635
3636         if (rc)
3637                 goto fail_unlink_md;
3638
3639         return 0;
3640
3641 fail_unlink_md:
3642         LNetMDUnlink(*mdh);
3643         LNetInvalidateMDHandle(mdh);
3644 fail_error:
3645         return rc;
3646 }
3647
3648 static void
3649 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3650                            int status, bool unlink_event)
3651 {
3652         lnet_nid_t nid = ev_info->mt_nid;
3653
3654         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3655                 struct lnet_ni *ni;
3656
3657                 lnet_net_lock(0);
3658                 ni = lnet_nid2ni_locked(nid, 0);
3659                 if (!ni) {
3660                         lnet_net_unlock(0);
3661                         return;
3662                 }
3663                 lnet_ni_lock(ni);
3664                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3665                 if (status)
3666                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3667                 lnet_ni_unlock(ni);
3668                 lnet_net_unlock(0);
3669
3670                 if (status != 0) {
3671                         CERROR("local NI (%s) recovery failed with %d\n",
3672                                libcfs_nid2str(nid), status);
3673                         return;
3674                 }
3675                 /*
3676                  * need to increment healthv for the ni here, because in
3677                  * the lnet_finalize() path we don't have access to this
3678                  * NI. And in order to get access to it, we'll need to
3679                  * carry forward too much information.
3680                  * In the peer case, it'll naturally be incremented
3681                  */
3682                 if (!unlink_event)
3683                         lnet_inc_healthv(&ni->ni_healthv);
3684         } else {
3685                 struct lnet_peer_ni *lpni;
3686                 int cpt;
3687
3688                 cpt = lnet_net_lock_current();
3689                 lpni = lnet_find_peer_ni_locked(nid);
3690                 if (!lpni) {
3691                         lnet_net_unlock(cpt);
3692                         return;
3693                 }
3694                 spin_lock(&lpni->lpni_lock);
3695                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3696                 if (status)
3697                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3698                 spin_unlock(&lpni->lpni_lock);
3699                 lnet_peer_ni_decref_locked(lpni);
3700                 lnet_net_unlock(cpt);
3701
3702                 if (status != 0)
3703                         CERROR("peer NI (%s) recovery failed with %d\n",
3704                                libcfs_nid2str(nid), status);
3705         }
3706 }
3707
3708 void
3709 lnet_mt_event_handler(struct lnet_event *event)
3710 {
3711         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3712         struct lnet_ping_buffer *pbuf;
3713
3714         /* TODO: remove assert */
3715         LASSERT(event->type == LNET_EVENT_REPLY ||
3716                 event->type == LNET_EVENT_SEND ||
3717                 event->type == LNET_EVENT_UNLINK);
3718
3719         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3720                event->status);
3721
3722         switch (event->type) {
3723         case LNET_EVENT_UNLINK:
3724                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3725                        libcfs_nid2str(ev_info->mt_nid));
3726                 /* fallthrough */
3727         case LNET_EVENT_REPLY:
3728                 lnet_handle_recovery_reply(ev_info, event->status,
3729                                            event->type == LNET_EVENT_UNLINK);
3730                 break;
3731         case LNET_EVENT_SEND:
3732                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3733                                libcfs_nid2str(ev_info->mt_nid),
3734                                (event->status) ? "unsuccessfully" :
3735                                "successfully", event->status);
3736                 break;
3737         default:
3738                 CERROR("Unexpected event: %d\n", event->type);
3739                 break;
3740         }
3741         if (event->unlinked) {
3742                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3743                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3744                 lnet_ping_buffer_decref(pbuf);
3745         }
3746 }
3747
3748 static int
3749 lnet_rsp_tracker_create(void)
3750 {
3751         struct list_head **rstqs;
3752         rstqs = lnet_create_array_of_queues();
3753
3754         if (!rstqs)
3755                 return -ENOMEM;
3756
3757         the_lnet.ln_mt_rstq = rstqs;
3758
3759         return 0;
3760 }
3761
3762 static void
3763 lnet_rsp_tracker_clean(void)
3764 {
3765         lnet_finalize_expired_responses();
3766
3767         cfs_percpt_free(the_lnet.ln_mt_rstq);
3768         the_lnet.ln_mt_rstq = NULL;
3769 }
3770
3771 int lnet_monitor_thr_start(void)
3772 {
3773         int rc = 0;
3774         struct task_struct *task;
3775
3776         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3777                 return -EALREADY;
3778
3779         rc = lnet_resendqs_create();
3780         if (rc)
3781                 return rc;
3782
3783         rc = lnet_rsp_tracker_create();
3784         if (rc)
3785                 goto clean_queues;
3786
3787         sema_init(&the_lnet.ln_mt_signal, 0);
3788
3789         lnet_net_lock(LNET_LOCK_EX);
3790         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3791         lnet_net_unlock(LNET_LOCK_EX);
3792         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3793         if (IS_ERR(task)) {
3794                 rc = PTR_ERR(task);
3795                 CERROR("Can't start monitor thread: %d\n", rc);
3796                 goto clean_thread;
3797         }
3798
3799         return 0;
3800
3801 clean_thread:
3802         lnet_net_lock(LNET_LOCK_EX);
3803         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3804         lnet_net_unlock(LNET_LOCK_EX);
3805         /* block until event callback signals exit */
3806         down(&the_lnet.ln_mt_signal);
3807         /* clean up */
3808         lnet_net_lock(LNET_LOCK_EX);
3809         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3810         lnet_net_unlock(LNET_LOCK_EX);
3811         lnet_rsp_tracker_clean();
3812         lnet_clean_local_ni_recoveryq();
3813         lnet_clean_peer_ni_recoveryq();
3814         lnet_clean_resendqs();
3815         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3816         return rc;
3817 clean_queues:
3818         lnet_rsp_tracker_clean();
3819         lnet_clean_local_ni_recoveryq();
3820         lnet_clean_peer_ni_recoveryq();
3821         lnet_clean_resendqs();
3822         return rc;
3823 }
3824
3825 void lnet_monitor_thr_stop(void)
3826 {
3827         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3828                 return;
3829
3830         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3831         lnet_net_lock(LNET_LOCK_EX);
3832         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3833         lnet_net_unlock(LNET_LOCK_EX);
3834
3835         /* tell the monitor thread that we're shutting down */
3836         complete(&the_lnet.ln_mt_wait_complete);
3837
3838         /* block until monitor thread signals that it's done */
3839         down(&the_lnet.ln_mt_signal);
3840         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3841
3842         /* perform cleanup tasks */
3843         lnet_rsp_tracker_clean();
3844         lnet_clean_local_ni_recoveryq();
3845         lnet_clean_peer_ni_recoveryq();
3846         lnet_clean_resendqs();
3847 }
3848
3849 void
3850 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3851                   __u32 msg_type)
3852 {
3853         lnet_net_lock(cpt);
3854         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3855         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3856         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3857         lnet_net_unlock(cpt);
3858
3859         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3860 }
3861
3862 static void
3863 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3864 {
3865         struct lnet_hdr *hdr = &msg->msg_hdr;
3866
3867         if (msg->msg_wanted != 0)
3868                 lnet_setpayloadbuffer(msg);
3869
3870         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3871
3872         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3873          * it back into the ACK during lnet_finalize() */
3874         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3875                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3876
3877         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3878                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3879 }
3880
3881 static int
3882 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3883 {
3884         struct lnet_hdr         *hdr = &msg->msg_hdr;
3885         struct lnet_match_info  info;
3886         int                     rc;
3887         bool                    ready_delay;
3888
3889         /* Convert put fields to host byte order */
3890         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3891         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3892         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3893
3894         /* Primary peer NID. */
3895         info.mi_id.nid  = msg->msg_initiator;
3896         info.mi_id.pid  = hdr->src_pid;
3897         info.mi_opc     = LNET_MD_OP_PUT;
3898         info.mi_portal  = hdr->msg.put.ptl_index;
3899         info.mi_rlength = hdr->payload_length;
3900         info.mi_roffset = hdr->msg.put.offset;
3901         info.mi_mbits   = hdr->msg.put.match_bits;
3902         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3903
3904         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3905         ready_delay = msg->msg_rx_ready_delay;
3906
3907  again:
3908         rc = lnet_ptl_match_md(&info, msg);
3909         switch (rc) {
3910         default:
3911                 LBUG();
3912
3913         case LNET_MATCHMD_OK:
3914                 lnet_recv_put(ni, msg);
3915                 return 0;
3916
3917         case LNET_MATCHMD_NONE:
3918                 if (ready_delay)
3919                         /* no eager_recv or has already called it, should
3920                          * have been attached on delayed list */
3921                         return 0;
3922
3923                 rc = lnet_ni_eager_recv(ni, msg);
3924                 if (rc == 0) {
3925                         ready_delay = true;
3926                         goto again;
3927                 }
3928                 /* fall through */
3929
3930         case LNET_MATCHMD_DROP:
3931                 CNETERR("Dropping PUT from %s portal %d match %llu"
3932                         " offset %d length %d: %d\n",
3933                         libcfs_id2str(info.mi_id), info.mi_portal,
3934                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3935
3936                 return -ENOENT; /* -ve: OK but no match */
3937         }
3938 }
3939
3940 static int
3941 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3942 {
3943         struct lnet_match_info info;
3944         struct lnet_hdr *hdr = &msg->msg_hdr;
3945         struct lnet_process_id source_id;
3946         struct lnet_handle_wire reply_wmd;
3947         int rc;
3948
3949         /* Convert get fields to host byte order */
3950         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3951         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3952         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3953         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3954
3955         source_id.nid = hdr->src_nid;
3956         source_id.pid = hdr->src_pid;
3957         /* Primary peer NID */
3958         info.mi_id.nid  = msg->msg_initiator;
3959         info.mi_id.pid  = hdr->src_pid;
3960         info.mi_opc     = LNET_MD_OP_GET;
3961         info.mi_portal  = hdr->msg.get.ptl_index;
3962         info.mi_rlength = hdr->msg.get.sink_length;
3963         info.mi_roffset = hdr->msg.get.src_offset;
3964         info.mi_mbits   = hdr->msg.get.match_bits;
3965         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3966
3967         rc = lnet_ptl_match_md(&info, msg);
3968         if (rc == LNET_MATCHMD_DROP) {
3969                 CNETERR("Dropping GET from %s portal %d match %llu"
3970                         " offset %d length %d\n",
3971                         libcfs_id2str(info.mi_id), info.mi_portal,
3972                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3973                 return -ENOENT; /* -ve: OK but no match */
3974         }
3975
3976         LASSERT(rc == LNET_MATCHMD_OK);
3977
3978         lnet_build_msg_event(msg, LNET_EVENT_GET);
3979
3980         reply_wmd = hdr->msg.get.return_wmd;
3981
3982         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3983                        msg->msg_offset, msg->msg_wanted);
3984
3985         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3986
3987         if (rdma_get) {
3988                 /* The LND completes the REPLY from her recv procedure */
3989                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3990                              msg->msg_offset, msg->msg_len, msg->msg_len);
3991                 return 0;
3992         }
3993
3994         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3995         msg->msg_receiving = 0;
3996
3997         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3998         if (rc < 0) {
3999                 /* didn't get as far as lnet_ni_send() */
4000                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4001                        libcfs_nid2str(ni->ni_nid),
4002                        libcfs_id2str(info.mi_id), rc);
4003
4004                 lnet_finalize(msg, rc);
4005         }
4006
4007         return 0;
4008 }
4009
4010 static int
4011 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4012 {
4013         void *private = msg->msg_private;
4014         struct lnet_hdr *hdr = &msg->msg_hdr;
4015         struct lnet_process_id src = {0};
4016         struct lnet_libmd *md;
4017         int rlength;
4018         int mlength;
4019         int cpt;
4020
4021         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4022         lnet_res_lock(cpt);
4023
4024         src.nid = hdr->src_nid;
4025         src.pid = hdr->src_pid;
4026
4027         /* NB handles only looked up by creator (no flips) */
4028         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4029         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4030                 CNETERR("%s: Dropping REPLY from %s for %s "
4031                         "MD %#llx.%#llx\n",
4032                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4033                         (md == NULL) ? "invalid" : "inactive",
4034                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4035                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4036                 if (md != NULL && md->md_me != NULL)
4037                         CERROR("REPLY MD also attached to portal %d\n",
4038                                md->md_me->me_portal);
4039
4040                 lnet_res_unlock(cpt);
4041                 return -ENOENT; /* -ve: OK but no match */
4042         }
4043
4044         LASSERT(md->md_offset == 0);
4045
4046         rlength = hdr->payload_length;
4047         mlength = MIN(rlength, (int)md->md_length);
4048
4049         if (mlength < rlength &&
4050             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4051                 CNETERR("%s: Dropping REPLY from %s length %d "
4052                         "for MD %#llx would overflow (%d)\n",
4053                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4054                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4055                         mlength);
4056                 lnet_res_unlock(cpt);
4057                 return -ENOENT; /* -ve: OK but no match */
4058         }
4059
4060         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4061                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4062                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4063
4064         lnet_msg_attach_md(msg, md, 0, mlength);
4065
4066         if (mlength != 0)
4067                 lnet_setpayloadbuffer(msg);
4068
4069         lnet_res_unlock(cpt);
4070
4071         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4072
4073         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4074         return 0;
4075 }
4076
4077 static int
4078 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4079 {
4080         struct lnet_hdr *hdr = &msg->msg_hdr;
4081         struct lnet_process_id src = {0};
4082         struct lnet_libmd *md;
4083         int cpt;
4084
4085         src.nid = hdr->src_nid;
4086         src.pid = hdr->src_pid;
4087
4088         /* Convert ack fields to host byte order */
4089         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4090         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4091
4092         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4093         lnet_res_lock(cpt);
4094
4095         /* NB handles only looked up by creator (no flips) */
4096         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4097         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4098                 /* Don't moan; this is expected */
4099                 CDEBUG(D_NET,
4100                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4101                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4102                        (md == NULL) ? "invalid" : "inactive",
4103                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4104                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4105                 if (md != NULL && md->md_me != NULL)
4106                         CERROR("Source MD also attached to portal %d\n",
4107                                md->md_me->me_portal);
4108
4109                 lnet_res_unlock(cpt);
4110                 return -ENOENT;                  /* -ve! */
4111         }
4112
4113         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4114                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4115                hdr->msg.ack.dst_wmd.wh_object_cookie);
4116
4117         lnet_msg_attach_md(msg, md, 0, 0);
4118
4119         lnet_res_unlock(cpt);
4120
4121         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4122
4123         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4124         return 0;
4125 }
4126
4127 /**
4128  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4129  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4130  * \retval -ve                  error code
4131  */
4132 int
4133 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4134 {
4135         int     rc = 0;
4136
4137         if (!the_lnet.ln_routing)
4138                 return -ECANCELED;
4139
4140         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4141             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4142                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4143                         msg->msg_rx_ready_delay = 1;
4144                 } else {
4145                         lnet_net_unlock(msg->msg_rx_cpt);
4146                         rc = lnet_ni_eager_recv(ni, msg);
4147                         lnet_net_lock(msg->msg_rx_cpt);
4148                 }
4149         }
4150
4151         if (rc == 0)
4152                 rc = lnet_post_routed_recv_locked(msg, 0);
4153         return rc;
4154 }
4155
4156 int
4157 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4158 {
4159         int     rc;
4160
4161         switch (msg->msg_type) {
4162         case LNET_MSG_ACK:
4163                 rc = lnet_parse_ack(ni, msg);
4164                 break;
4165         case LNET_MSG_PUT:
4166                 rc = lnet_parse_put(ni, msg);
4167                 break;
4168         case LNET_MSG_GET:
4169                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4170                 break;
4171         case LNET_MSG_REPLY:
4172                 rc = lnet_parse_reply(ni, msg);
4173                 break;
4174         default: /* prevent an unused label if !kernel */
4175                 LASSERT(0);
4176                 return -EPROTO;
4177         }
4178
4179         LASSERT(rc == 0 || rc == -ENOENT);
4180         return rc;
4181 }
4182
4183 char *
4184 lnet_msgtyp2str (int type)
4185 {
4186         switch (type) {
4187         case LNET_MSG_ACK:
4188                 return ("ACK");
4189         case LNET_MSG_PUT:
4190                 return ("PUT");
4191         case LNET_MSG_GET:
4192                 return ("GET");
4193         case LNET_MSG_REPLY:
4194                 return ("REPLY");
4195         case LNET_MSG_HELLO:
4196                 return ("HELLO");
4197         default:
4198                 return ("<UNKNOWN>");
4199         }
4200 }
4201
4202 void
4203 lnet_print_hdr(struct lnet_hdr *hdr)
4204 {
4205         struct lnet_process_id src = {
4206                 .nid = hdr->src_nid,
4207                 .pid = hdr->src_pid,
4208         };
4209         struct lnet_process_id dst = {
4210                 .nid = hdr->dest_nid,
4211                 .pid = hdr->dest_pid,
4212         };
4213         char *type_str = lnet_msgtyp2str(hdr->type);
4214
4215         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4216         CWARN("    From %s\n", libcfs_id2str(src));
4217         CWARN("    To   %s\n", libcfs_id2str(dst));
4218
4219         switch (hdr->type) {
4220         default:
4221                 break;
4222
4223         case LNET_MSG_PUT:
4224                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4225                       "match bits %llu\n",
4226                       hdr->msg.put.ptl_index,
4227                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4228                       hdr->msg.put.ack_wmd.wh_object_cookie,
4229                       hdr->msg.put.match_bits);
4230                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4231                       hdr->payload_length, hdr->msg.put.offset,
4232                       hdr->msg.put.hdr_data);
4233                 break;
4234
4235         case LNET_MSG_GET:
4236                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4237                       "match bits %llu\n", hdr->msg.get.ptl_index,
4238                       hdr->msg.get.return_wmd.wh_interface_cookie,
4239                       hdr->msg.get.return_wmd.wh_object_cookie,
4240                       hdr->msg.get.match_bits);
4241                 CWARN("    Length %d, src offset %d\n",
4242                       hdr->msg.get.sink_length,
4243                       hdr->msg.get.src_offset);
4244                 break;
4245
4246         case LNET_MSG_ACK:
4247                 CWARN("    dst md %#llx.%#llx, "
4248                       "manipulated length %d\n",
4249                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4250                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4251                       hdr->msg.ack.mlength);
4252                 break;
4253
4254         case LNET_MSG_REPLY:
4255                 CWARN("    dst md %#llx.%#llx, "
4256                       "length %d\n",
4257                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4258                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4259                       hdr->payload_length);
4260         }
4261
4262 }
4263
4264 int
4265 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4266            void *private, int rdma_req)
4267 {
4268         struct lnet_peer_ni *lpni;
4269         struct lnet_msg *msg;
4270         __u32 payload_length;
4271         lnet_pid_t dest_pid;
4272         lnet_nid_t dest_nid;
4273         lnet_nid_t src_nid;
4274         bool push = false;
4275         int for_me;
4276         __u32 type;
4277         int rc = 0;
4278         int cpt;
4279
4280         LASSERT (!in_interrupt ());
4281
4282         type = le32_to_cpu(hdr->type);
4283         src_nid = le64_to_cpu(hdr->src_nid);
4284         dest_nid = le64_to_cpu(hdr->dest_nid);
4285         dest_pid = le32_to_cpu(hdr->dest_pid);
4286         payload_length = le32_to_cpu(hdr->payload_length);
4287
4288         for_me = (ni->ni_nid == dest_nid);
4289         cpt = lnet_cpt_of_nid(from_nid, ni);
4290
4291         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4292                 libcfs_nid2str(dest_nid),
4293                 libcfs_nid2str(ni->ni_nid),
4294                 libcfs_nid2str(src_nid),
4295                 lnet_msgtyp2str(type),
4296                 (for_me) ? "for me" : "routed");
4297
4298         switch (type) {
4299         case LNET_MSG_ACK:
4300         case LNET_MSG_GET:
4301                 if (payload_length > 0) {
4302                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4303                                libcfs_nid2str(from_nid),
4304                                libcfs_nid2str(src_nid),
4305                                lnet_msgtyp2str(type), payload_length);
4306                         return -EPROTO;
4307                 }
4308                 break;
4309
4310         case LNET_MSG_PUT:
4311         case LNET_MSG_REPLY:
4312                 if (payload_length >
4313                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4314                         CERROR("%s, src %s: bad %s payload %d "
4315                                "(%d max expected)\n",
4316                                libcfs_nid2str(from_nid),
4317                                libcfs_nid2str(src_nid),
4318                                lnet_msgtyp2str(type),
4319                                payload_length,
4320                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4321                         return -EPROTO;
4322                 }
4323                 break;
4324
4325         default:
4326                 CERROR("%s, src %s: Bad message type 0x%x\n",
4327                        libcfs_nid2str(from_nid),
4328                        libcfs_nid2str(src_nid), type);
4329                 return -EPROTO;
4330         }
4331
4332         if (the_lnet.ln_routing &&
4333             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4334                 lnet_ni_lock(ni);
4335                 spin_lock(&ni->ni_net->net_lock);
4336                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4337                 spin_unlock(&ni->ni_net->net_lock);
4338                 if (ni->ni_status != NULL &&
4339                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4340                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4341                         push = true;
4342                 }
4343                 lnet_ni_unlock(ni);
4344         }
4345
4346         if (push)
4347                 lnet_push_update_to_peers(1);
4348
4349         /* Regard a bad destination NID as a protocol error.  Senders should
4350          * know what they're doing; if they don't they're misconfigured, buggy
4351          * or malicious so we chop them off at the knees :) */
4352
4353         if (!for_me) {
4354                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4355                         /* should have gone direct */
4356                         CERROR("%s, src %s: Bad dest nid %s "
4357                                "(should have been sent direct)\n",
4358                                 libcfs_nid2str(from_nid),
4359                                 libcfs_nid2str(src_nid),
4360                                 libcfs_nid2str(dest_nid));
4361                         return -EPROTO;
4362                 }
4363
4364                 if (lnet_islocalnid(dest_nid)) {
4365                         /* dest is another local NI; sender should have used
4366                          * this node's NID on its own network */
4367                         CERROR("%s, src %s: Bad dest nid %s "
4368                                "(it's my nid but on a different network)\n",
4369                                 libcfs_nid2str(from_nid),
4370                                 libcfs_nid2str(src_nid),
4371                                 libcfs_nid2str(dest_nid));
4372                         return -EPROTO;
4373                 }
4374
4375                 if (rdma_req && type == LNET_MSG_GET) {
4376                         CERROR("%s, src %s: Bad optimized GET for %s "
4377                                "(final destination must be me)\n",
4378                                 libcfs_nid2str(from_nid),
4379                                 libcfs_nid2str(src_nid),
4380                                 libcfs_nid2str(dest_nid));
4381                         return -EPROTO;
4382                 }
4383
4384                 if (!the_lnet.ln_routing) {
4385                         CERROR("%s, src %s: Dropping message for %s "
4386                                "(routing not enabled)\n",
4387                                 libcfs_nid2str(from_nid),
4388                                 libcfs_nid2str(src_nid),
4389                                 libcfs_nid2str(dest_nid));
4390                         goto drop;
4391                 }
4392         }
4393
4394         /* Message looks OK; we're not going to return an error, so we MUST
4395          * call back lnd_recv() come what may... */
4396
4397         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4398             fail_peer(src_nid, 0)) {                    /* shall we now? */
4399                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4400                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4401                        lnet_msgtyp2str(type));
4402                 goto drop;
4403         }
4404
4405         if (!list_empty(&the_lnet.ln_drop_rules) &&
4406             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4407                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4408                               "silent message loss\n",
4409                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4410                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4411                 goto drop;
4412         }
4413
4414         if (lnet_drop_asym_route && for_me &&
4415             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4416                 struct lnet_net *net;
4417                 struct lnet_remotenet *rnet;
4418                 bool found = true;
4419
4420                 /* we are dealing with a routed message,
4421                  * so see if route to reach src_nid goes through from_nid
4422                  */
4423                 lnet_net_lock(cpt);
4424                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4425                 if (!net) {
4426                         lnet_net_unlock(cpt);
4427                         CERROR("net %s not found\n",
4428                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4429                         return -EPROTO;
4430                 }
4431
4432                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4433                 if (rnet) {
4434                         struct lnet_peer *gw = NULL;
4435                         struct lnet_peer_ni *lpni = NULL;
4436                         struct lnet_route *route;
4437
4438                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4439                                 found = false;
4440                                 gw = route->lr_gateway;
4441                                 if (route->lr_lnet != net->net_id)
4442                                         continue;
4443                                 /*
4444                                  * if the nid is one of the gateway's NIDs
4445                                  * then this is a valid gateway
4446                                  */
4447                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4448                                                 NULL, lpni)) != NULL) {
4449                                         if (lpni->lpni_nid == from_nid) {
4450                                                 found = true;
4451                                                 break;
4452                                         }
4453                                 }
4454                         }
4455                 }
4456                 lnet_net_unlock(cpt);
4457                 if (!found) {
4458                         /* we would not use from_nid to route a message to
4459                          * src_nid
4460                          * => asymmetric routing detected but forbidden
4461                          */
4462                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4463                                libcfs_nid2str(from_nid),
4464                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4465                         goto drop;
4466                 }
4467         }
4468
4469         msg = lnet_msg_alloc();
4470         if (msg == NULL) {
4471                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4472                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4473                        lnet_msgtyp2str(type));
4474                 goto drop;
4475         }
4476
4477         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4478          * pointers NULL etc */
4479
4480         msg->msg_type = type;
4481         msg->msg_private = private;
4482         msg->msg_receiving = 1;
4483         msg->msg_rdma_get = rdma_req;
4484         msg->msg_len = msg->msg_wanted = payload_length;
4485         msg->msg_offset = 0;
4486         msg->msg_hdr = *hdr;
4487         /* for building message event */
4488         msg->msg_from = from_nid;
4489         if (!for_me) {
4490                 msg->msg_target.pid     = dest_pid;
4491                 msg->msg_target.nid     = dest_nid;
4492                 msg->msg_routing        = 1;
4493
4494         } else {
4495                 /* convert common msg->hdr fields to host byteorder */
4496                 msg->msg_hdr.type       = type;
4497                 msg->msg_hdr.src_nid    = src_nid;
4498                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4499                 msg->msg_hdr.dest_nid   = dest_nid;
4500                 msg->msg_hdr.dest_pid   = dest_pid;
4501                 msg->msg_hdr.payload_length = payload_length;
4502         }
4503
4504         lnet_net_lock(cpt);
4505         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4506         if (IS_ERR(lpni)) {
4507                 lnet_net_unlock(cpt);
4508                 CERROR("%s, src %s: Dropping %s "
4509                        "(error %ld looking up sender)\n",
4510                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4511                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4512                 lnet_msg_free(msg);
4513                 if (rc == -ESHUTDOWN)
4514                         /* We are shutting down.  Don't do anything more */
4515                         return 0;
4516                 goto drop;
4517         }
4518
4519         if (the_lnet.ln_routing)
4520                 lpni->lpni_last_alive = ktime_get_seconds();
4521
4522         msg->msg_rxpeer = lpni;
4523         msg->msg_rxni = ni;
4524         lnet_ni_addref_locked(ni, cpt);
4525         /* Multi-Rail: Primary NID of source. */
4526         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4527
4528         /*
4529          * mark the status of this lpni as UP since we received a message
4530          * from it. The ping response reports back the ns_status which is
4531          * marked on the remote as up or down and we cache it here.
4532          */
4533         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4534
4535         lnet_msg_commit(msg, cpt);
4536
4537         /* message delay simulation */
4538         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4539                      lnet_delay_rule_match_locked(hdr, msg))) {
4540                 lnet_net_unlock(cpt);
4541                 return 0;
4542         }
4543
4544         if (!for_me) {
4545                 rc = lnet_parse_forward_locked(ni, msg);
4546                 lnet_net_unlock(cpt);
4547
4548                 if (rc < 0)
4549                         goto free_drop;
4550
4551                 if (rc == LNET_CREDIT_OK) {
4552                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4553                                      0, payload_length, payload_length);
4554                 }
4555                 return 0;
4556         }
4557
4558         lnet_net_unlock(cpt);
4559
4560         rc = lnet_parse_local(ni, msg);
4561         if (rc != 0)
4562                 goto free_drop;
4563         return 0;
4564
4565  free_drop:
4566         LASSERT(msg->msg_md == NULL);
4567         lnet_finalize(msg, rc);
4568
4569  drop:
4570         lnet_drop_message(ni, cpt, private, payload_length, type);
4571         return 0;
4572 }
4573 EXPORT_SYMBOL(lnet_parse);
4574
4575 void
4576 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4577 {
4578         while (!list_empty(head)) {
4579                 struct lnet_process_id id = {0};
4580                 struct lnet_msg *msg;
4581
4582                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4583                 list_del(&msg->msg_list);
4584
4585                 id.nid = msg->msg_hdr.src_nid;
4586                 id.pid = msg->msg_hdr.src_pid;
4587
4588                 LASSERT(msg->msg_md == NULL);
4589                 LASSERT(msg->msg_rx_delayed);
4590                 LASSERT(msg->msg_rxpeer != NULL);
4591                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4592
4593                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4594                       " offset %d length %d: %s\n",
4595                       libcfs_id2str(id),
4596                       msg->msg_hdr.msg.put.ptl_index,
4597                       msg->msg_hdr.msg.put.match_bits,
4598                       msg->msg_hdr.msg.put.offset,
4599                       msg->msg_hdr.payload_length, reason);
4600
4601                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4602                  * called lnet_drop_message(), so I just hang onto msg as well
4603                  * until that's done */
4604
4605                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4606                                   msg->msg_private, msg->msg_len,
4607                                   msg->msg_type);
4608
4609                 msg->msg_no_resend = true;
4610                 /*
4611                  * NB: message will not generate event because w/o attached MD,
4612                  * but we still should give error code so lnet_msg_decommit()
4613                  * can skip counters operations and other checks.
4614                  */
4615                 lnet_finalize(msg, -ENOENT);
4616         }
4617 }
4618
4619 void
4620 lnet_recv_delayed_msg_list(struct list_head *head)
4621 {
4622         while (!list_empty(head)) {
4623                 struct lnet_msg *msg;
4624                 struct lnet_process_id id;
4625
4626                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4627                 list_del(&msg->msg_list);
4628
4629                 /* md won't disappear under me, since each msg
4630                  * holds a ref on it */
4631
4632                 id.nid = msg->msg_hdr.src_nid;
4633                 id.pid = msg->msg_hdr.src_pid;
4634
4635                 LASSERT(msg->msg_rx_delayed);
4636                 LASSERT(msg->msg_md != NULL);
4637                 LASSERT(msg->msg_rxpeer != NULL);
4638                 LASSERT(msg->msg_rxni != NULL);
4639                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4640
4641                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4642                        "match %llu offset %d length %d.\n",
4643                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4644                         msg->msg_hdr.msg.put.match_bits,
4645                         msg->msg_hdr.msg.put.offset,
4646                         msg->msg_hdr.payload_length);
4647
4648                 lnet_recv_put(msg->msg_rxni, msg);
4649         }
4650 }
4651
4652 static void
4653 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4654                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4655 {
4656         s64 timeout_ns;
4657         bool new_entry = true;
4658         struct lnet_rsp_tracker *local_rspt;
4659
4660         /*
4661          * MD has a refcount taken by message so it's not going away.
4662          * The MD however can be looked up. We need to secure the access
4663          * to the md_rspt_ptr by taking the res_lock.
4664          * The rspt can be accessed without protection up to when it gets
4665          * added to the list.
4666          */
4667
4668         lnet_res_lock(cpt);
4669         local_rspt = md->md_rspt_ptr;
4670         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4671         if (local_rspt != NULL) {
4672                 /*
4673                  * we already have an rspt attached to the md, so we'll
4674                  * update the deadline on that one.
4675                  */
4676                 lnet_rspt_free(rspt, cpt);
4677                 new_entry = false;
4678         } else {
4679                 /* new md */
4680                 rspt->rspt_mdh = mdh;
4681                 rspt->rspt_cpt = cpt;
4682                 /* store the rspt so we can access it when we get the REPLY */
4683                 md->md_rspt_ptr = rspt;
4684                 local_rspt = rspt;
4685         }
4686         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4687
4688         /*
4689          * add to the list of tracked responses. It's added to tail of the
4690          * list in order to expire all the older entries first.
4691          */
4692         lnet_net_lock(cpt);
4693         if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4694                 list_del_init(&local_rspt->rspt_on_list);
4695         list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4696         lnet_net_unlock(cpt);
4697         lnet_res_unlock(cpt);
4698 }
4699
4700 /**
4701  * Initiate an asynchronous PUT operation.
4702  *
4703  * There are several events associated with a PUT: completion of the send on
4704  * the initiator node (LNET_EVENT_SEND), and when the send completes
4705  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4706  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4707  * used at the target node to indicate the completion of incoming data
4708  * delivery.
4709  *
4710  * The local events will be logged in the EQ associated with the MD pointed to
4711  * by \a mdh handle. Using a MD without an associated EQ results in these
4712  * events being discarded. In this case, the caller must have another
4713  * mechanism (e.g., a higher level protocol) for determining when it is safe
4714  * to modify the memory region associated with the MD.
4715  *
4716  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4717  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4718  *
4719  * \param self Indicates the NID of a local interface through which to send
4720  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4721  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4722  * must be "free floating" (See LNetMDBind()).
4723  * \param ack Controls whether an acknowledgment is requested.
4724  * Acknowledgments are only sent when they are requested by the initiating
4725  * process and the target MD enables them.
4726  * \param target A process identifier for the target process.
4727  * \param portal The index in the \a target's portal table.
4728  * \param match_bits The match bits to use for MD selection at the target
4729  * process.
4730  * \param offset The offset into the target MD (only used when the target
4731  * MD has the LNET_MD_MANAGE_REMOTE option set).
4732  * \param hdr_data 64 bits of user data that can be included in the message
4733  * header. This data is written to an event queue entry at the target if an
4734  * EQ is present on the matching MD.
4735  *
4736  * \retval  0      Success, and only in this case events will be generated
4737  * and logged to EQ (if it exists).
4738  * \retval -EIO    Simulated failure.
4739  * \retval -ENOMEM Memory allocation failure.
4740  * \retval -ENOENT Invalid MD object.
4741  *
4742  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4743  */
4744 int
4745 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4746         struct lnet_process_id target, unsigned int portal,
4747         __u64 match_bits, unsigned int offset,
4748         __u64 hdr_data)
4749 {
4750         struct lnet_msg *msg;
4751         struct lnet_libmd *md;
4752         int cpt;
4753         int rc;
4754         struct lnet_rsp_tracker *rspt = NULL;
4755
4756         LASSERT(the_lnet.ln_refcount > 0);
4757
4758         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4759             fail_peer(target.nid, 1)) {                 /* shall we now? */
4760                 CERROR("Dropping PUT to %s: simulated failure\n",
4761                        libcfs_id2str(target));
4762                 return -EIO;
4763         }
4764
4765         msg = lnet_msg_alloc();
4766         if (msg == NULL) {
4767                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4768                        libcfs_id2str(target));
4769                 return -ENOMEM;
4770         }
4771         msg->msg_vmflush = !!memory_pressure_get();
4772
4773         cpt = lnet_cpt_of_cookie(mdh.cookie);
4774
4775         if (ack == LNET_ACK_REQ) {
4776                 rspt = lnet_rspt_alloc(cpt);
4777                 if (!rspt) {
4778                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4779                                 libcfs_id2str(target));
4780                         return -ENOMEM;
4781                 }
4782                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4783         }
4784
4785         lnet_res_lock(cpt);
4786
4787         md = lnet_handle2md(&mdh);
4788         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4789                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4790                        match_bits, portal, libcfs_id2str(target),
4791                        md == NULL ? -1 : md->md_threshold);
4792                 if (md != NULL && md->md_me != NULL)
4793                         CERROR("Source MD also attached to portal %d\n",
4794                                md->md_me->me_portal);
4795                 lnet_res_unlock(cpt);
4796
4797                 lnet_rspt_free(rspt, cpt);
4798                 lnet_msg_free(msg);
4799                 return -ENOENT;
4800         }
4801
4802         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4803
4804         lnet_msg_attach_md(msg, md, 0, 0);
4805
4806         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4807
4808         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4809         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4810         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4811         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4812
4813         /* NB handles only looked up by creator (no flips) */
4814         if (ack == LNET_ACK_REQ) {
4815                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4816                         the_lnet.ln_interface_cookie;
4817                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4818                         md->md_lh.lh_cookie;
4819         } else {
4820                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4821                         LNET_WIRE_HANDLE_COOKIE_NONE;
4822                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4823                         LNET_WIRE_HANDLE_COOKIE_NONE;
4824         }
4825
4826         lnet_res_unlock(cpt);
4827
4828         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4829
4830         if (ack == LNET_ACK_REQ)
4831                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4832
4833         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4834                                  CFS_FAIL_ONCE))
4835                 rc = -EIO;
4836         else
4837                 rc = lnet_send(self, msg, LNET_NID_ANY);
4838
4839         if (rc != 0) {
4840                 CNETERR("Error sending PUT to %s: %d\n",
4841                         libcfs_id2str(target), rc);
4842                 msg->msg_no_resend = true;
4843                 lnet_finalize(msg, rc);
4844         }
4845
4846         /* completion will be signalled by an event */
4847         return 0;
4848 }
4849 EXPORT_SYMBOL(LNetPut);
4850
4851 /*
4852  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4853  * returns a msg for the LND to pass to lnet_finalize() when the sink
4854  * data has been received.
4855  *
4856  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4857  * lnet_finalize() is called on it, so the LND must call this first
4858  */
4859 struct lnet_msg *
4860 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4861 {
4862         struct lnet_msg *msg = lnet_msg_alloc();
4863         struct lnet_libmd *getmd = getmsg->msg_md;
4864         struct lnet_process_id peer_id = getmsg->msg_target;
4865         int cpt;
4866
4867         LASSERT(!getmsg->msg_target_is_router);
4868         LASSERT(!getmsg->msg_routing);
4869
4870         if (msg == NULL) {
4871                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4872                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4873                 goto drop;
4874         }
4875
4876         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4877         lnet_res_lock(cpt);
4878
4879         LASSERT(getmd->md_refcount > 0);
4880
4881         if (getmd->md_threshold == 0) {
4882                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4883                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4884                         getmd);
4885                 lnet_res_unlock(cpt);
4886                 goto drop;
4887         }
4888
4889         LASSERT(getmd->md_offset == 0);
4890
4891         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4892                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4893
4894         /* setup information for lnet_build_msg_event */
4895         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4896         msg->msg_from = peer_id.nid;
4897         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4898         msg->msg_hdr.src_nid = peer_id.nid;
4899         msg->msg_hdr.payload_length = getmd->md_length;
4900         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4901
4902         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4903         lnet_res_unlock(cpt);
4904
4905         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4906
4907         lnet_net_lock(cpt);
4908         lnet_msg_commit(msg, cpt);
4909         lnet_net_unlock(cpt);
4910
4911         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4912
4913         return msg;
4914
4915  drop:
4916         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4917
4918         lnet_net_lock(cpt);
4919         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4920         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4921         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4922                 getmd->md_length;
4923         lnet_net_unlock(cpt);
4924
4925         if (msg != NULL)
4926                 lnet_msg_free(msg);
4927
4928         return NULL;
4929 }
4930 EXPORT_SYMBOL(lnet_create_reply_msg);
4931
4932 void
4933 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4934                        unsigned int len)
4935 {
4936         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4937          * completed and I know it. */
4938         LASSERT(reply != NULL);
4939         LASSERT(reply->msg_type == LNET_MSG_GET);
4940         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4941
4942         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4943          * the end of my buffer, I might as well be dead. */
4944         LASSERT(len <= reply->msg_ev.mlength);
4945
4946         reply->msg_ev.mlength = len;
4947 }
4948 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4949
4950 /**
4951  * Initiate an asynchronous GET operation.
4952  *
4953  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4954  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4955  * the target node in the REPLY has been written to local MD.
4956  *
4957  * On the target node, an LNET_EVENT_GET is logged when the GET request
4958  * arrives and is accepted into a MD.
4959  *
4960  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4961  * \param mdh A handle for the MD that describes the memory into which the
4962  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4963  *
4964  * \retval  0      Success, and only in this case events will be generated
4965  * and logged to EQ (if it exists) of the MD.
4966  * \retval -EIO    Simulated failure.
4967  * \retval -ENOMEM Memory allocation failure.
4968  * \retval -ENOENT Invalid MD object.
4969  */
4970 int
4971 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4972         struct lnet_process_id target, unsigned int portal,
4973         __u64 match_bits, unsigned int offset, bool recovery)
4974 {
4975         struct lnet_msg *msg;
4976         struct lnet_libmd *md;
4977         struct lnet_rsp_tracker *rspt;
4978         int cpt;
4979         int rc;
4980
4981         LASSERT(the_lnet.ln_refcount > 0);
4982
4983         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4984             fail_peer(target.nid, 1))                   /* shall we now? */
4985         {
4986                 CERROR("Dropping GET to %s: simulated failure\n",
4987                        libcfs_id2str(target));
4988                 return -EIO;
4989         }
4990
4991         msg = lnet_msg_alloc();
4992         if (!msg) {
4993                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4994                        libcfs_id2str(target));
4995                 return -ENOMEM;
4996         }
4997
4998         cpt = lnet_cpt_of_cookie(mdh.cookie);
4999
5000         rspt = lnet_rspt_alloc(cpt);
5001         if (!rspt) {
5002                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5003                        libcfs_id2str(target));
5004                 return -ENOMEM;
5005         }
5006         INIT_LIST_HEAD(&rspt->rspt_on_list);
5007
5008         msg->msg_recovery = recovery;
5009
5010         lnet_res_lock(cpt);
5011
5012         md = lnet_handle2md(&mdh);
5013         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5014                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5015                        match_bits, portal, libcfs_id2str(target),
5016                        md == NULL ? -1 : md->md_threshold);
5017                 if (md != NULL && md->md_me != NULL)
5018                         CERROR("REPLY MD also attached to portal %d\n",
5019                                md->md_me->me_portal);
5020
5021                 lnet_res_unlock(cpt);
5022
5023                 lnet_msg_free(msg);
5024                 lnet_rspt_free(rspt, cpt);
5025                 return -ENOENT;
5026         }
5027
5028         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5029
5030         lnet_msg_attach_md(msg, md, 0, 0);
5031
5032         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5033
5034         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5035         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5036         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5037         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5038
5039         /* NB handles only looked up by creator (no flips) */
5040         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5041                 the_lnet.ln_interface_cookie;
5042         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5043                 md->md_lh.lh_cookie;
5044
5045         lnet_res_unlock(cpt);
5046
5047         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5048
5049         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5050
5051         rc = lnet_send(self, msg, LNET_NID_ANY);
5052         if (rc < 0) {
5053                 CNETERR("Error sending GET to %s: %d\n",
5054                         libcfs_id2str(target), rc);
5055                 msg->msg_no_resend = true;
5056                 lnet_finalize(msg, rc);
5057         }
5058
5059         /* completion will be signalled by an event */
5060         return 0;
5061 }
5062 EXPORT_SYMBOL(LNetGet);
5063
5064 /**
5065  * Calculate distance to node at \a dstnid.
5066  *
5067  * \param dstnid Target NID.
5068  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5069  * is saved here.
5070  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5071  * here.
5072  *
5073  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5074  * local_nid_dist_zero is set, which is the default.
5075  * \retval positives Distance to target NID, i.e. number of hops plus one.
5076  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5077  */
5078 int
5079 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5080 {
5081         struct list_head        *e;
5082         struct lnet_ni *ni = NULL;
5083         struct lnet_remotenet *rnet;
5084         __u32                   dstnet = LNET_NIDNET(dstnid);
5085         int                     hops;
5086         int                     cpt;
5087         __u32                   order = 2;
5088         struct list_head        *rn_list;
5089
5090         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5091          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5092          * keep order 0 free for 0@lo and order 1 free for a local NID
5093          * match */
5094
5095         LASSERT(the_lnet.ln_refcount > 0);
5096
5097         cpt = lnet_net_lock_current();
5098
5099         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5100                 if (ni->ni_nid == dstnid) {
5101                         if (srcnidp != NULL)
5102                                 *srcnidp = dstnid;
5103                         if (orderp != NULL) {
5104                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5105                                         *orderp = 0;
5106                                 else
5107                                         *orderp = 1;
5108                         }
5109                         lnet_net_unlock(cpt);
5110
5111                         return local_nid_dist_zero ? 0 : 1;
5112                 }
5113
5114                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5115                         /* Check if ni was originally created in
5116                          * current net namespace.
5117                          * If not, assign order above 0xffff0000,
5118                          * to make this ni not a priority. */
5119                         if (current->nsproxy &&
5120                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5121                                         order += 0xffff0000;
5122                         if (srcnidp != NULL)
5123                                 *srcnidp = ni->ni_nid;
5124                         if (orderp != NULL)
5125                                 *orderp = order;
5126                         lnet_net_unlock(cpt);
5127                         return 1;
5128                 }
5129
5130                 order++;
5131         }
5132
5133         rn_list = lnet_net2rnethash(dstnet);
5134         list_for_each(e, rn_list) {
5135                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5136
5137                 if (rnet->lrn_net == dstnet) {
5138                         struct lnet_route *route;
5139                         struct lnet_route *shortest = NULL;
5140                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5141                         __u32 route_hops;
5142
5143                         LASSERT(!list_empty(&rnet->lrn_routes));
5144
5145                         list_for_each_entry(route, &rnet->lrn_routes,
5146                                             lr_list) {
5147                                 route_hops = route->lr_hops;
5148                                 if (route_hops == LNET_UNDEFINED_HOPS)
5149                                         route_hops = 1;
5150                                 if (shortest == NULL ||
5151                                     route_hops < shortest_hops) {
5152                                         shortest = route;
5153                                         shortest_hops = route_hops;
5154                                 }
5155                         }
5156
5157                         LASSERT(shortest != NULL);
5158                         hops = shortest_hops;
5159                         if (srcnidp != NULL) {
5160                                 struct lnet_net *net;
5161                                 net = lnet_get_net_locked(shortest->lr_lnet);
5162                                 LASSERT(net);
5163                                 ni = lnet_get_next_ni_locked(net, NULL);
5164                                 *srcnidp = ni->ni_nid;
5165                         }
5166                         if (orderp != NULL)
5167                                 *orderp = order;
5168                         lnet_net_unlock(cpt);
5169                         return hops + 1;
5170                 }
5171                 order++;
5172         }
5173
5174         lnet_net_unlock(cpt);
5175         return -EHOSTUNREACH;
5176 }
5177 EXPORT_SYMBOL(LNetDist);